Move into version control

197 files changed, 68723 insertions, 0 deletions

diff --git a/src/EigenUnsupported/AdolcForward b/src/EigenUnsupported/AdolcForward
new file mode 100644
index 0000000..56caeae
--- /dev/null
+++ b/src/EigenUnsupported/AdolcForward
@@ -0,0 +1,159 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <g.gael@free.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ADLOC_FORWARD
+#define EIGEN_ADLOC_FORWARD
+
+//--------------------------------------------------------------------------------
+//
+// This file provides support for adolc's adouble type in forward mode.
+// ADOL-C is a C++ automatic differentiation library,
+// see https://projects.coin-or.org/ADOL-C for more information.
+//
+// Note that the maximal number of directions is controlled by
+// the preprocessor token NUMBER_DIRECTIONS. The default is 2.
+//
+//--------------------------------------------------------------------------------
+
+#define ADOLC_TAPELESS
+#ifndef NUMBER_DIRECTIONS
+# define NUMBER_DIRECTIONS 2
+#endif
+#include <adolc/adtl.h>
+
+// adolc defines some very stupid macros:
+#if defined(malloc)
+# undef malloc
+#endif
+
+#if defined(calloc)
+# undef calloc
+#endif
+
+#if defined(realloc)
+# undef realloc
+#endif
+
+#include "../../Eigen/Core"
+
+namespace Eigen {
+
+/**
+  * \defgroup AdolcForward_Module Adolc forward module
+  * This module provides support for adolc's adouble type in forward mode.
+  * ADOL-C is a C++ automatic differentiation library,
+  * see https://projects.coin-or.org/ADOL-C for more information.
+  * It mainly consists in:
+  *  - a struct Eigen::NumTraits<adtl::adouble> specialization
+  *  - overloads of internal::* math function for adtl::adouble type.
+  *
+  * Note that the maximal number of directions is controlled by
+  * the preprocessor token NUMBER_DIRECTIONS. The default is 2.
+  *
+  * \code
+  * #include <unsupported/Eigen/AdolcSupport>
+  * \endcode
+  */
+  //@{
+
+} // namespace Eigen
+
+// Eigen's require a few additional functions which must be defined in the same namespace
+// than the custom scalar type own namespace
+namespace adtl {
+
+inline const adouble& conj(const adouble& x)  { return x; }
+inline const adouble& real(const adouble& x)  { return x; }
+inline adouble imag(const adouble&)    { return 0.; }
+inline adouble abs(const adouble&  x)  { return fabs(x); }
+inline adouble abs2(const adouble& x)  { return x*x; }
+
+inline bool (isinf)(const adouble& x) { return (Eigen::numext::isinf)(x.getValue()); }
+inline bool (isnan)(const adouble& x) { return (Eigen::numext::isnan)(x.getValue()); }
+
+}
+
+namespace Eigen {
+
+template<> struct NumTraits<adtl::adouble>
+    : NumTraits<double>
+{
+  typedef adtl::adouble Real;
+  typedef adtl::adouble NonInteger;
+  typedef adtl::adouble Nested;
+  enum {
+    IsComplex = 0,
+    IsInteger = 0,
+    IsSigned = 1,
+    RequireInitialization = 1,
+    ReadCost = 1,
+    AddCost = 1,
+    MulCost = 1
+  };
+};
+
+template<typename Functor> class AdolcForwardJacobian : public Functor
+{
+  typedef adtl::adouble ActiveScalar;
+public:
+
+  AdolcForwardJacobian() : Functor() {}
+  AdolcForwardJacobian(const Functor& f) : Functor(f) {}
+
+  // forward constructors
+  template<typename T0>
+  AdolcForwardJacobian(const T0& a0) : Functor(a0) {}
+  template<typename T0, typename T1>
+  AdolcForwardJacobian(const T0& a0, const T1& a1) : Functor(a0, a1) {}
+  template<typename T0, typename T1, typename T2>
+  AdolcForwardJacobian(const T0& a0, const T1& a1, const T1& a2) : Functor(a0, a1, a2) {}
+
+  typedef typename Functor::InputType InputType;
+  typedef typename Functor::ValueType ValueType;
+  typedef typename Functor::JacobianType JacobianType;
+
+  typedef Matrix<ActiveScalar, InputType::SizeAtCompileTime, 1> ActiveInput;
+  typedef Matrix<ActiveScalar, ValueType::SizeAtCompileTime, 1> ActiveValue;
+
+  void operator() (const InputType& x, ValueType* v, JacobianType* _jac) const
+  {
+    eigen_assert(v!=0);
+    if (!_jac)
+    {
+      Functor::operator()(x, v);
+      return;
+    }
+
+    JacobianType& jac = *_jac;
+
+    ActiveInput ax = x.template cast<ActiveScalar>();
+    ActiveValue av(jac.rows());
+
+    for (int j=0; j<jac.cols(); j++)
+      for (int i=0; i<jac.cols(); i++)
+        ax[i].setADValue(j, i==j ? 1 : 0);
+
+    Functor::operator()(ax, &av);
+
+    for (int i=0; i<jac.rows(); i++)
+    {
+      (*v)[i] = av[i].getValue();
+      for (int j=0; j<jac.cols(); j++)
+        jac.coeffRef(i,j) = av[i].getADValue(j);
+    }
+  }
+protected:
+
+};
+
+//@}
+
+}
+
+#endif // EIGEN_ADLOC_FORWARD
diff --git a/src/EigenUnsupported/AlignedVector3 b/src/EigenUnsupported/AlignedVector3
new file mode 100644
index 0000000..4fa1842
--- /dev/null
+++ b/src/EigenUnsupported/AlignedVector3
@@ -0,0 +1,234 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <g.gael@free.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ALIGNED_VECTOR3
+#define EIGEN_ALIGNED_VECTOR3
+
+#include "../../Eigen/Geometry"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+namespace Eigen {
+
+/**
+  * \defgroup AlignedVector3_Module Aligned vector3 module
+  *
+  * \code
+  * #include <unsupported/Eigen/AlignedVector3>
+  * \endcode
+  */
+  //@{
+
+
+/** \class AlignedVector3
+  *
+  * \brief A vectorization friendly 3D vector
+  *
+  * This class represents a 3D vector internally using a 4D vector
+  * such that vectorization can be seamlessly enabled. Of course,
+  * the same result can be achieved by directly using a 4D vector.
+  * This class makes this process simpler.
+  *
+  */
+// TODO specialize Cwise
+template<typename _Scalar> class AlignedVector3;
+
+namespace internal {
+template<typename _Scalar> struct traits<AlignedVector3<_Scalar> >
+  : traits<Matrix<_Scalar,3,1,0,4,1> >
+{
+};
+}
+
+template<typename _Scalar> class AlignedVector3
+  : public MatrixBase<AlignedVector3<_Scalar> >
+{
+    typedef Matrix<_Scalar,4,1> CoeffType;
+    CoeffType m_coeffs;
+  public:
+
+    typedef MatrixBase<AlignedVector3<_Scalar> > Base;	
+    EIGEN_DENSE_PUBLIC_INTERFACE(AlignedVector3)
+    using Base::operator*;
+
+    inline Index rows() const { return 3; }
+    inline Index cols() const { return 1; }
+    
+    Scalar* data() { return m_coeffs.data(); }
+    const Scalar* data() const { return m_coeffs.data(); }
+    Index innerStride() const { return 1; }
+    Index outerStride() const { return 3; }
+
+    inline const Scalar& coeff(Index row, Index col) const
+    { return m_coeffs.coeff(row, col); }
+
+    inline Scalar& coeffRef(Index row, Index col)
+    { return m_coeffs.coeffRef(row, col); }
+
+    inline const Scalar& coeff(Index index) const
+    { return m_coeffs.coeff(index); }
+
+    inline Scalar& coeffRef(Index index)
+    { return m_coeffs.coeffRef(index);}
+
+
+    inline AlignedVector3()
+    {}
+
+    inline AlignedVector3(const Scalar& x, const Scalar& y, const Scalar& z)
+      : m_coeffs(x, y, z, Scalar(0))
+    {}
+
+    inline AlignedVector3(const AlignedVector3& other)
+      : Base(), m_coeffs(other.m_coeffs)
+    {}
+
+    template<typename XprType, int Size=XprType::SizeAtCompileTime>
+    struct generic_assign_selector {};
+
+    template<typename XprType> struct generic_assign_selector<XprType,4>
+    {
+      inline static void run(AlignedVector3& dest, const XprType& src)
+      {
+        dest.m_coeffs = src;
+      }
+    };
+
+    template<typename XprType> struct generic_assign_selector<XprType,3>
+    {
+      inline static void run(AlignedVector3& dest, const XprType& src)
+      {
+        dest.m_coeffs.template head<3>() = src;
+        dest.m_coeffs.w() = Scalar(0);
+      }
+    };
+
+    template<typename Derived>
+    inline AlignedVector3(const MatrixBase<Derived>& other)
+    {
+      generic_assign_selector<Derived>::run(*this,other.derived());
+    }
+
+    inline AlignedVector3& operator=(const AlignedVector3& other)
+    { m_coeffs = other.m_coeffs; return *this; }
+
+    template <typename Derived>
+    inline AlignedVector3& operator=(const MatrixBase<Derived>& other)
+    {
+      generic_assign_selector<Derived>::run(*this,other.derived());
+      return *this;
+    }
+
+    inline AlignedVector3 operator+(const AlignedVector3& other) const
+    { return AlignedVector3(m_coeffs + other.m_coeffs); }
+
+    inline AlignedVector3& operator+=(const AlignedVector3& other)
+    { m_coeffs += other.m_coeffs; return *this; }
+
+    inline AlignedVector3 operator-(const AlignedVector3& other) const
+    { return AlignedVector3(m_coeffs - other.m_coeffs); }
+
+    inline AlignedVector3 operator-() const
+    { return AlignedVector3(-m_coeffs); }
+
+    inline AlignedVector3 operator-=(const AlignedVector3& other)
+    { m_coeffs -= other.m_coeffs; return *this; }
+
+    inline AlignedVector3 operator*(const Scalar& s) const
+    { return AlignedVector3(m_coeffs * s); }
+
+    inline friend AlignedVector3 operator*(const Scalar& s,const AlignedVector3& vec)
+    { return AlignedVector3(s * vec.m_coeffs); }
+
+    inline AlignedVector3& operator*=(const Scalar& s)
+    { m_coeffs *= s; return *this; }
+
+    inline AlignedVector3 operator/(const Scalar& s) const
+    { return AlignedVector3(m_coeffs / s); }
+
+    inline AlignedVector3& operator/=(const Scalar& s)
+    { m_coeffs /= s; return *this; }
+
+    inline Scalar dot(const AlignedVector3& other) const
+    {
+      eigen_assert(m_coeffs.w()==Scalar(0));
+      eigen_assert(other.m_coeffs.w()==Scalar(0));
+      return m_coeffs.dot(other.m_coeffs);
+    }
+
+    inline void normalize()
+    {
+      m_coeffs /= norm();
+    }
+
+    inline AlignedVector3 normalized() const
+    {
+      return AlignedVector3(m_coeffs / norm());
+    }
+
+    inline Scalar sum() const
+    {
+      eigen_assert(m_coeffs.w()==Scalar(0));
+      return m_coeffs.sum();
+    }
+
+    inline Scalar squaredNorm() const
+    {
+      eigen_assert(m_coeffs.w()==Scalar(0));
+      return m_coeffs.squaredNorm();
+    }
+
+    inline Scalar norm() const
+    {
+      using std::sqrt;
+      return sqrt(squaredNorm());
+    }
+
+    inline AlignedVector3 cross(const AlignedVector3& other) const
+    {
+      return AlignedVector3(m_coeffs.cross3(other.m_coeffs));
+    }
+
+    template<typename Derived>
+    inline bool isApprox(const MatrixBase<Derived>& other, const RealScalar& eps=NumTraits<Scalar>::dummy_precision()) const
+    {
+      return m_coeffs.template head<3>().isApprox(other,eps);
+    }
+    
+    CoeffType& coeffs() { return m_coeffs; }
+    const CoeffType& coeffs() const { return m_coeffs; }
+};
+
+namespace internal {
+
+template<typename _Scalar>
+struct eval<AlignedVector3<_Scalar>, Dense>
+{
+ typedef const AlignedVector3<_Scalar>& type;
+};
+
+template<typename Scalar>
+struct evaluator<AlignedVector3<Scalar> >
+  : evaluator<Matrix<Scalar,4,1> >
+{
+  typedef AlignedVector3<Scalar> XprType;
+  typedef evaluator<Matrix<Scalar,4,1> > Base;
+  
+  evaluator(const XprType &m) : Base(m.coeffs()) {}  
+};
+
+}
+
+//@}
+
+}
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_ALIGNED_VECTOR3
diff --git a/src/EigenUnsupported/ArpackSupport b/src/EigenUnsupported/ArpackSupport
new file mode 100644
index 0000000..67c4ac8
--- /dev/null
+++ b/src/EigenUnsupported/ArpackSupport
@@ -0,0 +1,30 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ARPACKSUPPORT_MODULE_H
+#define EIGEN_ARPACKSUPPORT_MODULE_H
+
+#include "../../Eigen/Core"
+
+/** \defgroup ArpackSupport_Module Arpack support module
+  *
+  * This module provides a wrapper to Arpack, a library for sparse eigenvalue decomposition.
+  *
+  * \code
+  * #include <Eigen/ArpackSupport>
+  * \endcode
+  */
+
+#include "../../Eigen/SparseCholesky"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+#include "src/Eigenvalues/ArpackSelfAdjointEigenSolver.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_ARPACKSUPPORT_MODULE_H
diff --git a/src/EigenUnsupported/AutoDiff b/src/EigenUnsupported/AutoDiff
new file mode 100644
index 0000000..7a4ff46
--- /dev/null
+++ b/src/EigenUnsupported/AutoDiff
@@ -0,0 +1,46 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <g.gael@free.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_AUTODIFF_MODULE
+#define EIGEN_AUTODIFF_MODULE
+
+namespace Eigen {
+
+/**
+  * \defgroup AutoDiff_Module Auto Diff module
+  *
+  * This module features forward automatic differentation via a simple
+  * templated scalar type wrapper AutoDiffScalar.
+  *
+  * Warning : this should NOT be confused with numerical differentiation, which
+  * is a different method and has its own module in Eigen : \ref NumericalDiff_Module.
+  *
+  * \code
+  * #include <unsupported/Eigen/AutoDiff>
+  * \endcode
+  */
+//@{
+
+}
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+
+#include "src/AutoDiff/AutoDiffScalar.h"
+// #include "src/AutoDiff/AutoDiffVector.h"
+#include "src/AutoDiff/AutoDiffJacobian.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+
+
+namespace Eigen {
+//@}
+}
+
+#endif // EIGEN_AUTODIFF_MODULE
diff --git a/src/EigenUnsupported/BVH b/src/EigenUnsupported/BVH
new file mode 100644
index 0000000..666c983
--- /dev/null
+++ b/src/EigenUnsupported/BVH
@@ -0,0 +1,95 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Ilya Baran <ibaran@mit.edu>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BVH_MODULE_H
+#define EIGEN_BVH_MODULE_H
+
+#include "../../Eigen/Core"
+#include "../../Eigen/Geometry"
+#include "../../Eigen/StdVector"
+#include <algorithm>
+#include <queue>
+
+namespace Eigen {
+
+/**
+  * \defgroup BVH_Module BVH module
+  * \brief This module provides generic bounding volume hierarchy algorithms
+  * and reference tree implementations.
+  *
+  *
+  * \code
+  * #include <unsupported/Eigen/BVH>
+  * \endcode
+  *
+  * A bounding volume hierarchy (BVH) can accelerate many geometric queries.  This module provides a generic implementation
+  * of the two basic algorithms over a BVH: intersection of a query object against all objects in the hierarchy and minimization
+  * of a function over the objects in the hierarchy.  It also provides intersection and minimization over a cartesian product of
+  * two BVH's.  A BVH accelerates intersection by using the fact that if a query object does not intersect a volume, then it cannot
+  * intersect any object contained in that volume.  Similarly, a BVH accelerates minimization because the minimum of a function
+  * over a volume is no greater than the minimum of a function over any object contained in it.
+  *
+  * Some sample queries that can be written in terms of intersection are:
+  *   - Determine all points where a ray intersects a triangle mesh
+  *   - Given a set of points, determine which are contained in a query sphere
+  *   - Given a set of spheres, determine which contain the query point
+  *   - Given a set of disks, determine if any is completely contained in a query rectangle (represent each 2D disk as a point \f$(x,y,r)\f$
+  *     in 3D and represent the rectangle as a pyramid based on the original rectangle and shrinking in the \f$r\f$ direction)
+  *   - Given a set of points, count how many pairs are \f$d\pm\epsilon\f$ apart (done by looking at the cartesian product of the set
+  *     of points with itself)
+  *
+  * Some sample queries that can be written in terms of function minimization over a set of objects are:
+  *   - Find the intersection between a ray and a triangle mesh closest to the ray origin (function is infinite off the ray)
+  *   - Given a polyline and a query point, determine the closest point on the polyline to the query
+  *   - Find the diameter of a point cloud (done by looking at the cartesian product and using negative distance as the function)
+  *   - Determine how far two meshes are from colliding (this is also a cartesian product query)
+  *
+  * This implementation decouples the basic algorithms both from the type of hierarchy (and the types of the bounding volumes) and
+  * from the particulars of the query.  To enable abstraction from the BVH, the BVH is required to implement a generic mechanism
+  * for traversal.  To abstract from the query, the query is responsible for keeping track of results.
+  *
+  * To be used in the algorithms, a hierarchy must implement the following traversal mechanism (see KdBVH for a sample implementation): \code
+      typedef Volume  //the type of bounding volume
+      typedef Object  //the type of object in the hierarchy
+      typedef Index   //a reference to a node in the hierarchy--typically an int or a pointer
+      typedef VolumeIterator //an iterator type over node children--returns Index
+      typedef ObjectIterator //an iterator over object (leaf) children--returns const Object &
+      Index getRootIndex() const //returns the index of the hierarchy root
+      const Volume &getVolume(Index index) const //returns the bounding volume of the node at given index
+      void getChildren(Index index, VolumeIterator &outVBegin, VolumeIterator &outVEnd,
+                      ObjectIterator &outOBegin, ObjectIterator &outOEnd) const
+      //getChildren takes a node index and makes [outVBegin, outVEnd) range over its node children
+      //and [outOBegin, outOEnd) range over its object children
+    \endcode
+  *
+  * To use the hierarchy, call BVIntersect or BVMinimize, passing it a BVH (or two, for cartesian product) and a minimizer or intersector.
+  * For an intersection query on a single BVH, the intersector encapsulates the query and must provide two functions:
+  * \code
+      bool intersectVolume(const Volume &volume) //returns true if the query intersects the volume
+      bool intersectObject(const Object &object) //returns true if the intersection search should terminate immediately
+    \endcode
+  * The guarantee that BVIntersect provides is that intersectObject will be called on every object whose bounding volume
+  * intersects the query (but possibly on other objects too) unless the search is terminated prematurely.  It is the
+  * responsibility of the intersectObject function to keep track of the results in whatever manner is appropriate.
+  * The cartesian product intersection and the BVMinimize queries are similar--see their individual documentation.
+  *
+  * The following is a simple but complete example for how to use the BVH to accelerate the search for a closest red-blue point pair:
+  * \include BVH_Example.cpp
+  * Output: \verbinclude BVH_Example.out
+  */
+}
+
+//@{
+
+#include "src/BVH/BVAlgorithms.h"
+#include "src/BVH/KdBVH.h"
+
+//@}
+
+#endif // EIGEN_BVH_MODULE_H
diff --git a/src/EigenUnsupported/CMakeLists.txt b/src/EigenUnsupported/CMakeLists.txt
new file mode 100644
index 0000000..631a060
--- /dev/null
+++ b/src/EigenUnsupported/CMakeLists.txt
@@ -0,0 +1,32 @@
+set(Eigen_HEADERS 
+  AdolcForward
+  AlignedVector3
+  ArpackSupport
+  AutoDiff
+  BVH
+  EulerAngles
+  FFT
+  IterativeSolvers 
+  KroneckerProduct
+  LevenbergMarquardt
+  MatrixFunctions 
+  MoreVectorization
+  MPRealSupport
+  NonLinearOptimization
+  NumericalDiff
+  OpenGLSupport
+  Polynomials
+  Skyline 
+  SparseExtra
+  SpecialFunctions
+  Splines
+  )
+
+install(FILES
+  ${Eigen_HEADERS}
+  DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen COMPONENT Devel
+  )
+
+install(DIRECTORY src DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen COMPONENT Devel FILES_MATCHING PATTERN "*.h")
+
+add_subdirectory(CXX11)
diff --git a/src/EigenUnsupported/CXX11/CMakeLists.txt b/src/EigenUnsupported/CXX11/CMakeLists.txt
new file mode 100644
index 0000000..385ed24
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/CMakeLists.txt
@@ -0,0 +1,8 @@
+set(Eigen_CXX11_HEADERS Tensor TensorSymmetry ThreadPool)
+
+install(FILES
+  ${Eigen_CXX11_HEADERS}
+  DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11 COMPONENT Devel
+  )
+
+install(DIRECTORY src DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/CXX11 COMPONENT Devel FILES_MATCHING PATTERN "*.h")
diff --git a/src/EigenUnsupported/CXX11/Tensor b/src/EigenUnsupported/CXX11/Tensor
new file mode 100644
index 0000000..0938bb5
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/Tensor
@@ -0,0 +1,137 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+//#ifndef EIGEN_CXX11_TENSOR_MODULE
+//#define EIGEN_CXX11_TENSOR_MODULE
+
+#include "../../../Eigen/Core"
+
+#if EIGEN_HAS_CXX11
+
+#include "../SpecialFunctions"
+
+#include "../../../Eigen/src/Core/util/DisableStupidWarnings.h"
+#include "src/util/CXX11Meta.h"
+#include "src/util/MaxSizeVector.h"
+
+/** \defgroup CXX11_Tensor_Module Tensor Module
+  *
+  * This module provides a Tensor class for storing arbitrarily indexed
+  * objects.
+  *
+  * \code
+  * #include <Eigen/CXX11/Tensor>
+  * \endcode
+  *
+  * Much of the documentation can be found \ref eigen_tensors "here".
+  */
+
+#include <atomic>
+#include <chrono>
+#include <cmath>
+#include <cstddef>
+#include <cstring>
+#include <random>
+#include <thread>
+
+#if defined(EIGEN_USE_THREADS) || defined(EIGEN_USE_SYCL)
+#include "ThreadPool"
+#endif
+
+#ifdef EIGEN_USE_GPU
+  #include <iostream>
+  #if defined(EIGEN_USE_HIP)
+    #include <hip/hip_runtime.h>
+  #else
+    #include <cuda_runtime.h>
+  #endif
+#endif
+
+#include "src/Tensor/TensorMacros.h"
+#include "src/Tensor/TensorForwardDeclarations.h"
+#include "src/Tensor/TensorMeta.h"
+#include "src/Tensor/TensorFunctors.h"
+#include "src/Tensor/TensorCostModel.h"
+#include "src/Tensor/TensorDeviceDefault.h"
+#include "src/Tensor/TensorDeviceThreadPool.h"
+#include "src/Tensor/TensorDeviceGpu.h"
+#ifndef gpu_assert
+#define gpu_assert(x)
+#endif
+#include "src/Tensor/TensorDeviceSycl.h"
+#include "src/Tensor/TensorIndexList.h"
+#include "src/Tensor/TensorDimensionList.h"
+#include "src/Tensor/TensorDimensions.h"
+#include "src/Tensor/TensorInitializer.h"
+#include "src/Tensor/TensorTraits.h"
+#include "src/Tensor/TensorRandom.h"
+#include "src/Tensor/TensorUInt128.h"
+#include "src/Tensor/TensorIntDiv.h"
+#include "src/Tensor/TensorGlobalFunctions.h"
+
+#include "src/Tensor/TensorBase.h"
+#include "src/Tensor/TensorBlock.h"
+
+#include "src/Tensor/TensorEvaluator.h"
+#include "src/Tensor/TensorExpr.h"
+#include "src/Tensor/TensorReduction.h"
+#include "src/Tensor/TensorReductionGpu.h"
+#include "src/Tensor/TensorArgMax.h"
+#include "src/Tensor/TensorConcatenation.h"
+#include "src/Tensor/TensorContractionMapper.h"
+#include "src/Tensor/TensorContractionBlocking.h"
+#include "src/Tensor/TensorContraction.h"
+#include "src/Tensor/TensorContractionThreadPool.h"
+#include "src/Tensor/TensorContractionGpu.h"
+#include "src/Tensor/TensorConversion.h"
+#include "src/Tensor/TensorConvolution.h"
+#include "src/Tensor/TensorFFT.h"
+#include "src/Tensor/TensorPatch.h"
+#include "src/Tensor/TensorImagePatch.h"
+#include "src/Tensor/TensorVolumePatch.h"
+#include "src/Tensor/TensorBroadcasting.h"
+#include "src/Tensor/TensorChipping.h"
+#include "src/Tensor/TensorInflation.h"
+#include "src/Tensor/TensorLayoutSwap.h"
+#include "src/Tensor/TensorMorphing.h"
+#include "src/Tensor/TensorPadding.h"
+#include "src/Tensor/TensorReverse.h"
+#include "src/Tensor/TensorShuffling.h"
+#include "src/Tensor/TensorStriding.h"
+#include "src/Tensor/TensorCustomOp.h"
+#include "src/Tensor/TensorEvalTo.h"
+#include "src/Tensor/TensorForcedEval.h"
+#include "src/Tensor/TensorGenerator.h"
+#include "src/Tensor/TensorAssign.h"
+#include "src/Tensor/TensorScan.h"
+#include "src/Tensor/TensorTrace.h"
+
+#ifdef EIGEN_USE_SYCL
+#include "src/Tensor/TensorReductionSycl.h"
+#include "src/Tensor/TensorConvolutionSycl.h"
+#include "src/Tensor/TensorContractionSycl.h"
+#include "src/Tensor/TensorScanSycl.h"
+#endif
+
+#include "src/Tensor/TensorExecutor.h"
+#include "src/Tensor/TensorDevice.h"
+
+#include "src/Tensor/TensorStorage.h"
+#include "src/Tensor/Tensor.h"
+#include "src/Tensor/TensorFixedSize.h"
+#include "src/Tensor/TensorMap.h"
+#include "src/Tensor/TensorRef.h"
+
+#include "src/Tensor/TensorIO.h"
+
+#include "../../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif  // EIGEN_HAS_CXX11
+//#endif // EIGEN_CXX11_TENSOR_MODULE
diff --git a/src/EigenUnsupported/CXX11/TensorSymmetry b/src/EigenUnsupported/CXX11/TensorSymmetry
new file mode 100644
index 0000000..b09c5e4
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/TensorSymmetry
@@ -0,0 +1,42 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSORSYMMETRY_MODULE
+#define EIGEN_CXX11_TENSORSYMMETRY_MODULE
+
+#include "Tensor"
+
+#include "../../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#include "src/util/CXX11Meta.h"
+
+/** \defgroup CXX11_TensorSymmetry_Module Tensor Symmetry Module
+  *
+  * This module provides a classes that allow for the definition of
+  * symmetries w.r.t. tensor indices.
+  *
+  * Including this module will implicitly include the Tensor module.
+  *
+  * \code
+  * #include <Eigen/TensorSymmetry>
+  * \endcode
+  */
+
+#include "src/TensorSymmetry/util/TemplateGroupTheory.h"
+#include "src/TensorSymmetry/Symmetry.h"
+#include "src/TensorSymmetry/StaticSymmetry.h"
+#include "src/TensorSymmetry/DynamicSymmetry.h"
+
+#include "../../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_CXX11_TENSORSYMMETRY_MODULE
+
+/*
+ * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle;
+ */
diff --git a/src/EigenUnsupported/CXX11/ThreadPool b/src/EigenUnsupported/CXX11/ThreadPool
new file mode 100644
index 0000000..c5cafb2
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/ThreadPool
@@ -0,0 +1,74 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_MODULE
+#define EIGEN_CXX11_THREADPOOL_MODULE
+
+#include "../../../Eigen/Core"
+
+#include "../../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+/** \defgroup CXX11_ThreadPool_Module C++11 ThreadPool Module
+  *
+  * This module provides 2 threadpool implementations
+  *  - a simple reference implementation
+  *  - a faster non blocking implementation
+  *
+  * This module requires C++11.
+  *
+  * \code
+  * #include <Eigen/CXX11/ThreadPool>
+  * \endcode
+  */
+
+
+// The code depends on CXX11, so only include the module if the
+// compiler supports it.
+#if (EIGEN_COMP_CXXVER >= 11)
+#include <cstddef>
+#include <cstring>
+#include <time.h>
+
+#include <vector>
+#include <atomic>
+#include <condition_variable>
+#include <deque>
+#include <mutex>
+#include <thread>
+#include <functional>
+#include <memory>
+#include <utility>
+
+// There are non-parenthesized calls to "max" in the  <unordered_map> header,
+// which trigger a check in test/main.h causing compilation to fail.
+// We work around the check here by removing the check for max in
+// the case where we have to emulate thread_local.
+#ifdef max
+#undef max
+#endif
+#include <unordered_map>
+
+#include "src/util/CXX11Meta.h"
+#include "src/util/MaxSizeVector.h"
+
+#include "src/ThreadPool/ThreadLocal.h"
+#include "src/ThreadPool/ThreadYield.h"
+#include "src/ThreadPool/ThreadCancel.h"
+#include "src/ThreadPool/EventCount.h"
+#include "src/ThreadPool/RunQueue.h"
+#include "src/ThreadPool/ThreadPoolInterface.h"
+#include "src/ThreadPool/ThreadEnvironment.h"
+#include "src/ThreadPool/Barrier.h"
+#include "src/ThreadPool/NonBlockingThreadPool.h"
+
+#endif
+
+#include "../../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_CXX11_THREADPOOL_MODULE
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/README.md b/src/EigenUnsupported/CXX11/src/Tensor/README.md
new file mode 100644
index 0000000..2f65b1b
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/README.md
@@ -0,0 +1,1815 @@
+# Eigen Tensors {#eigen_tensors}
+
+Tensors are multidimensional arrays of elements. Elements are typically scalars,
+but more complex types such as strings are also supported.
+
+## Tensor Classes
+
+You can manipulate a tensor with one of the following classes.  They all are in
+the namespace `::Eigen.`
+
+
+### Class Tensor<data_type, rank>
+
+This is the class to use to create a tensor and allocate memory for it.  The
+class is templatized with the tensor datatype, such as float or int, and the
+tensor rank.  The rank is the number of dimensions, for example rank 2 is a
+matrix.
+
+Tensors of this class are resizable.  For example, if you assign a tensor of a
+different size to a Tensor, that tensor is resized to match its new value.
+
+#### Constructor Tensor<data_type, rank>(size0, size1, ...)
+
+Constructor for a Tensor.  The constructor must be passed `rank` integers
+indicating the sizes of the instance along each of the the `rank`
+dimensions.
+
+    // Create a tensor of rank 3 of sizes 2, 3, 4.  This tensor owns
+    // memory to hold 24 floating point values (24 = 2 x 3 x 4).
+    Tensor<float, 3> t_3d(2, 3, 4);
+
+    // Resize t_3d by assigning a tensor of different sizes, but same rank.
+    t_3d = Tensor<float, 3>(3, 4, 3);
+
+#### Constructor Tensor<data_type, rank>(size_array)
+
+Constructor where the sizes for the constructor are specified as an array of
+values instead of an explicitly list of parameters.  The array type to use is
+`Eigen::array<Eigen::Index>`.  The array can be constructed automatically
+from an initializer list.
+
+    // Create a tensor of strings of rank 2 with sizes 5, 7.
+    Tensor<string, 2> t_2d({5, 7});
+
+
+### Class TensorFixedSize<data_type, Sizes<size0, size1, ...>>
+
+Class to use for tensors of fixed size, where the size is known at compile
+time.  Fixed sized tensors can provide very fast computations because all their
+dimensions are known by the compiler.  FixedSize tensors are not resizable.
+
+If the total number of elements in a fixed size tensor is small enough the
+tensor data is held onto the stack and does not cause heap allocation and free.
+
+    // Create a 4 x 3 tensor of floats.
+    TensorFixedSize<float, Sizes<4, 3>> t_4x3;
+
+### Class TensorMap<Tensor<data_type, rank>>
+
+This is the class to use to create a tensor on top of memory allocated and
+owned by another part of your code.  It allows to view any piece of allocated
+memory as a Tensor.  Instances of this class do not own the memory where the
+data are stored.
+
+A TensorMap is not resizable because it does not own the memory where its data
+are stored.
+
+#### Constructor TensorMap<Tensor<data_type, rank>>(data, size0, size1, ...)
+
+Constructor for a Tensor.  The constructor must be passed a pointer to the
+storage for the data, and "rank" size attributes.  The storage has to be
+large enough to hold all the data.
+
+    // Map a tensor of ints on top of stack-allocated storage.
+    int storage[128];  // 2 x 4 x 2 x 8 = 128
+    TensorMap<Tensor<int, 4>> t_4d(storage, 2, 4, 2, 8);
+
+    // The same storage can be viewed as a different tensor.
+    // You can also pass the sizes as an array.
+    TensorMap<Tensor<int, 2>> t_2d(storage, 16, 8);
+
+    // You can also map fixed-size tensors.  Here we get a 1d view of
+    // the 2d fixed-size tensor.
+    TensorFixedSize<float, Sizes<4, 3>> t_4x3;
+    TensorMap<Tensor<float, 1>> t_12(t_4x3.data(), 12);
+
+
+#### Class TensorRef
+
+See Assigning to a TensorRef below.
+
+## Accessing Tensor Elements
+
+#### <data_type> tensor(index0, index1...)
+
+Return the element at position `(index0, index1...)` in tensor
+`tensor`.  You must pass as many parameters as the rank of `tensor`.
+The expression can be used as an l-value to set the value of the element at the
+specified position.  The value returned is of the datatype of the tensor.
+
+    // Set the value of the element at position (0, 1, 0);
+    Tensor<float, 3> t_3d(2, 3, 4);
+    t_3d(0, 1, 0) = 12.0f;
+
+    // Initialize all elements to random values.
+    for (int i = 0; i < 2; ++i) {
+      for (int j = 0; j < 3; ++j) {
+        for (int k = 0; k < 4; ++k) {
+          t_3d(i, j, k) = ...some random value...;
+        }
+      }
+    }
+
+    // Print elements of a tensor.
+    for (int i = 0; i < 2; ++i) {
+      LOG(INFO) << t_3d(i, 0, 0);
+    }
+
+
+## TensorLayout
+
+The tensor library supports 2 layouts: `ColMajor` (the default) and
+`RowMajor`.  Only the default column major layout is currently fully
+supported, and it is therefore not recommended to attempt to use the row major
+layout at the moment.
+
+The layout of a tensor is optionally specified as part of its type. If not
+specified explicitly column major is assumed.
+
+    Tensor<float, 3, ColMajor> col_major;  // equivalent to Tensor<float, 3>
+    TensorMap<Tensor<float, 3, RowMajor> > row_major(data, ...);
+
+All the arguments to an expression must use the same layout. Attempting to mix
+different layouts will result in a compilation error.
+
+It is possible to change the layout of a tensor or an expression using the
+`swap_layout()` method.  Note that this will also reverse the order of the
+dimensions.
+
+    Tensor<float, 2, ColMajor> col_major(2, 4);
+    Tensor<float, 2, RowMajor> row_major(2, 4);
+
+    Tensor<float, 2> col_major_result = col_major;  // ok, layouts match
+    Tensor<float, 2> col_major_result = row_major;  // will not compile
+
+    // Simple layout swap
+    col_major_result = row_major.swap_layout();
+    eigen_assert(col_major_result.dimension(0) == 4);
+    eigen_assert(col_major_result.dimension(1) == 2);
+
+    // Swap the layout and preserve the order of the dimensions
+    array<int, 2> shuffle(1, 0);
+    col_major_result = row_major.swap_layout().shuffle(shuffle);
+    eigen_assert(col_major_result.dimension(0) == 2);
+    eigen_assert(col_major_result.dimension(1) == 4);
+
+
+## Tensor Operations
+
+The Eigen Tensor library provides a vast library of operations on Tensors:
+numerical operations such as addition and multiplication, geometry operations
+such as slicing and shuffling, etc.  These operations are available as methods
+of the Tensor classes, and in some cases as operator overloads.  For example
+the following code computes the elementwise addition of two tensors:
+
+    Tensor<float, 3> t1(2, 3, 4);
+    ...set some values in t1...
+    Tensor<float, 3> t2(2, 3, 4);
+    ...set some values in t2...
+    // Set t3 to the element wise sum of t1 and t2
+    Tensor<float, 3> t3 = t1 + t2;
+
+While the code above looks easy enough, it is important to understand that the
+expression `t1 + t2` is not actually adding the values of the tensors.  The
+expression instead constructs a "tensor operator" object of the class
+TensorCwiseBinaryOp<scalar_sum>, which has references to the tensors
+`t1` and `t2`.  This is a small C++ object that knows how to add
+`t1` and `t2`.  It is only when the value of the expression is assigned
+to the tensor `t3` that the addition is actually performed.  Technically,
+this happens through the overloading of `operator=()` in the Tensor class.
+
+This mechanism for computing tensor expressions allows for lazy evaluation and
+optimizations which are what make the tensor library very fast.
+
+Of course, the tensor operators do nest, and the expression `t1 + t2 * 0.3f`
+is actually represented with the (approximate) tree of operators:
+
+    TensorCwiseBinaryOp<scalar_sum>(t1, TensorCwiseUnaryOp<scalar_mul>(t2, 0.3f))
+
+
+### Tensor Operations and C++ "auto"
+
+Because Tensor operations create tensor operators, the C++ `auto` keyword
+does not have its intuitive meaning.  Consider these 2 lines of code:
+
+    Tensor<float, 3> t3 = t1 + t2;
+    auto t4 = t1 + t2;
+
+In the first line we allocate the tensor `t3` and it will contain the
+result of the addition of `t1` and `t2`.  In the second line, `t4`
+is actually the tree of tensor operators that will compute the addition of
+`t1` and `t2`.  In fact, `t4` is *not* a tensor and you cannot get
+the values of its elements:
+
+    Tensor<float, 3> t3 = t1 + t2;
+    cout << t3(0, 0, 0);  // OK prints the value of t1(0, 0, 0) + t2(0, 0, 0)
+
+    auto t4 = t1 + t2;
+    cout << t4(0, 0, 0);  // Compilation error!
+
+When you use `auto` you do not get a Tensor as a result but instead a
+non-evaluated expression.  So only use `auto` to delay evaluation.
+
+Unfortunately, there is no single underlying concrete type for holding
+non-evaluated expressions, hence you have to use auto in the case when you do
+want to hold non-evaluated expressions.
+
+When you need the results of set of tensor computations you have to assign the
+result to a Tensor that will be capable of holding onto them.  This can be
+either a normal Tensor, a fixed size Tensor, or a TensorMap on an existing
+piece of memory.  All the following will work:
+
+    auto t4 = t1 + t2;
+
+    Tensor<float, 3> result = t4;  // Could also be: result(t4);
+    cout << result(0, 0, 0);
+
+    TensorMap<float, 4> result(<a float* with enough space>, <size0>, ...) = t4;
+    cout << result(0, 0, 0);
+
+    TensorFixedSize<float, Sizes<size0, ...>> result = t4;
+    cout << result(0, 0, 0);
+
+Until you need the results, you can keep the operation around, and even reuse
+it for additional operations.  As long as you keep the expression as an
+operation, no computation is performed.
+
+    // One way to compute exp((t1 + t2) * 0.2f);
+    auto t3 = t1 + t2;
+    auto t4 = t3 * 0.2f;
+    auto t5 = t4.exp();
+    Tensor<float, 3> result = t5;
+
+    // Another way, exactly as efficient as the previous one:
+    Tensor<float, 3> result = ((t1 + t2) * 0.2f).exp();
+
+### Controlling When Expression are Evaluated
+
+There are several ways to control when expressions are evaluated:
+
+*   Assignment to a Tensor, TensorFixedSize, or TensorMap.
+*   Use of the eval() method.
+*   Assignment to a TensorRef.
+
+#### Assigning to a Tensor, TensorFixedSize, or TensorMap.
+
+The most common way to evaluate an expression is to assign it to a Tensor.  In
+the example below, the `auto` declarations make the intermediate values
+"Operations", not Tensors, and do not cause the expressions to be evaluated.
+The assignment to the Tensor `result` causes the evaluation of all the
+operations.
+
+    auto t3 = t1 + t2;             // t3 is an Operation.
+    auto t4 = t3 * 0.2f;           // t4 is an Operation.
+    auto t5 = t4.exp();            // t5 is an Operation.
+    Tensor<float, 3> result = t5;  // The operations are evaluated.
+
+If you know the ranks and sizes of the Operation value you can assign the
+Operation to a TensorFixedSize instead of a Tensor, which is a bit more
+efficient.
+
+    // We know that the result is a 4x4x2 tensor!
+    TensorFixedSize<float, Sizes<4, 4, 2>> result = t5;
+
+Simiarly, assigning an expression to a TensorMap causes its evaluation.  Like
+tensors of type TensorFixedSize, TensorMaps cannot be resized so they have to
+have the rank and sizes of the expression that are assigned to them.
+
+#### Calling eval().
+
+When you compute large composite expressions, you sometimes want to tell Eigen
+that an intermediate value in the expression tree is worth evaluating ahead of
+time.  This is done by inserting a call to the `eval()` method of the
+expression Operation.
+
+    // The previous example could have been written:
+    Tensor<float, 3> result = ((t1 + t2) * 0.2f).exp();
+
+    // If you want to compute (t1 + t2) once ahead of time you can write:
+    Tensor<float, 3> result = ((t1 + t2).eval() * 0.2f).exp();
+
+Semantically, calling `eval()` is equivalent to materializing the value of
+the expression in a temporary Tensor of the right size.  The code above in
+effect does:
+
+    // .eval() knows the size!
+    TensorFixedSize<float, Sizes<4, 4, 2>> tmp = t1 + t2;
+    Tensor<float, 3> result = (tmp * 0.2f).exp();
+
+Note that the return value of `eval()` is itself an Operation, so the
+following code does not do what you may think:
+
+    // Here t3 is an evaluation Operation.  t3 has not been evaluated yet.
+    auto t3 = (t1 + t2).eval();
+
+    // You can use t3 in another expression.  Still no evaluation.
+    auto t4 = (t3 * 0.2f).exp();
+
+    // The value is evaluated when you assign the Operation to a Tensor, using
+    // an intermediate tensor to represent t3.x
+    Tensor<float, 3> result = t4;
+
+While in the examples above calling `eval()` does not make a difference in
+performance, in other cases it can make a huge difference.  In the expression
+below the `broadcast()` expression causes the `X.maximum()` expression
+to be evaluated many times:
+
+    Tensor<...> X ...;
+    Tensor<...> Y = ((X - X.maximum(depth_dim).reshape(dims2d).broadcast(bcast))
+                     * beta).exp();
+
+Inserting a call to `eval()` between the `maximum()` and
+`reshape()` calls guarantees that maximum() is only computed once and
+greatly speeds-up execution:
+
+    Tensor<...> Y =
+      ((X - X.maximum(depth_dim).eval().reshape(dims2d).broadcast(bcast))
+        * beta).exp();
+
+In the other example below, the tensor `Y` is both used in the expression
+and its assignment.  This is an aliasing problem and if the evaluation is not
+done in the right order Y will be updated incrementally during the evaluation
+resulting in bogus results:
+
+     Tensor<...> Y ...;
+     Y = Y / (Y.sum(depth_dim).reshape(dims2d).broadcast(bcast));
+
+Inserting a call to `eval()` between the `sum()` and `reshape()`
+expressions ensures that the sum is computed before any updates to `Y` are
+done.
+
+     Y = Y / (Y.sum(depth_dim).eval().reshape(dims2d).broadcast(bcast));
+
+Note that an eval around the full right hand side expression is not needed
+because the generated has to compute the i-th value of the right hand side
+before assigning it to the left hand side.
+
+However, if you were assigning the expression value to a shuffle of `Y`
+then you would need to force an eval for correctness by adding an `eval()`
+call for the right hand side:
+
+     Y.shuffle(...) =
+        (Y / (Y.sum(depth_dim).eval().reshape(dims2d).broadcast(bcast))).eval();
+
+
+#### Assigning to a TensorRef.
+
+If you need to access only a few elements from the value of an expression you
+can avoid materializing the value in a full tensor by using a TensorRef.
+
+A TensorRef is a small wrapper class for any Eigen Operation.  It provides
+overloads for the `()` operator that let you access individual values in
+the expression.  TensorRef is convenient, because the Operation themselves do
+not provide a way to access individual elements.
+
+    // Create a TensorRef for the expression.  The expression is not
+    // evaluated yet.
+    TensorRef<Tensor<float, 3> > ref = ((t1 + t2) * 0.2f).exp();
+
+    // Use "ref" to access individual elements.  The expression is evaluated
+    // on the fly.
+    float at_0 = ref(0, 0, 0);
+    cout << ref(0, 1, 0);
+
+Only use TensorRef when you need a subset of the values of the expression.
+TensorRef only computes the values you access.  However note that if you are
+going to access all the values it will be much faster to materialize the
+results in a Tensor first.
+
+In some cases, if the full Tensor result would be very large, you may save
+memory by accessing it as a TensorRef.  But not always.  So don't count on it.
+
+
+### Controlling How Expressions Are Evaluated
+
+The tensor library provides several implementations of the various operations
+such as contractions and convolutions.  The implementations are optimized for
+different environments: single threaded on CPU, multi threaded on CPU, or on a
+GPU using cuda.  Additional implementations may be added later.
+
+You can choose which implementation to use with the `device()` call.  If
+you do not choose an implementation explicitly the default implementation that
+uses a single thread on the CPU is used.
+
+The default implementation has been optimized for recent Intel CPUs, taking
+advantage of SSE, AVX, and FMA instructions.  Work is ongoing to tune the
+library on ARM CPUs.  Note that you need to pass compiler-dependent flags
+to enable the use of SSE, AVX, and other instructions.
+
+For example, the following code adds two tensors using the default
+single-threaded CPU implementation:
+
+    Tensor<float, 2> a(30, 40);
+    Tensor<float, 2> b(30, 40);
+    Tensor<float, 2> c = a + b;
+
+To choose a different implementation you have to insert a `device()` call
+before the assignment of the result.  For technical C++ reasons this requires
+that the Tensor for the result be declared on its own.  This means that you
+have to know the size of the result.
+
+    Eigen::Tensor<float, 2> c(30, 40);
+    c.device(...) = a + b;
+
+The call to `device()` must be the last call on the left of the operator=.
+
+You must pass to the `device()` call an Eigen device object.  There are
+presently three devices you can use: DefaultDevice, ThreadPoolDevice and
+GpuDevice.
+
+
+#### Evaluating With the DefaultDevice
+
+This is exactly the same as not inserting a `device()` call.
+
+    DefaultDevice my_device;
+    c.device(my_device) = a + b;
+
+#### Evaluating with a Thread Pool
+
+    // Create the Eigen ThreadPool
+    Eigen::ThreadPool pool(8 /* number of threads in pool */)
+
+    // Create the Eigen ThreadPoolDevice.
+    Eigen::ThreadPoolDevice my_device(&pool, 4 /* number of threads to use */);
+
+    // Now just use the device when evaluating expressions.
+    Eigen::Tensor<float, 2> c(30, 50);
+    c.device(my_device) = a.contract(b, dot_product_dims);
+
+
+#### Evaluating On GPU
+
+This is presently a bit more complicated than just using a thread pool device.
+You need to create a GPU device but you also need to explicitly allocate the
+memory for tensors with cuda.
+
+
+## API Reference
+
+### Datatypes
+
+In the documentation of the tensor methods and Operation we mention datatypes
+that are tensor-type specific:
+
+#### <Tensor-Type>::Dimensions
+
+Acts like an array of ints.  Has an `int size` attribute, and can be
+indexed like an array to access individual values.  Used to represent the
+dimensions of a tensor.  See `dimensions()`.
+
+#### <Tensor-Type>::Index
+
+Acts like an `int`.  Used for indexing tensors along their dimensions.  See
+`operator()`, `dimension()`, and `size()`.
+
+#### <Tensor-Type>::Scalar
+
+Represents the datatype of individual tensor elements.  For example, for a
+`Tensor<float>`, `Scalar` is the type `float`.  See
+`setConstant()`.
+
+#### <Operation>
+
+We use this pseudo type to indicate that a tensor Operation is returned by a
+method.  We indicate in the text the type and dimensions of the tensor that the
+Operation returns after evaluation.
+
+The Operation will have to be evaluated, for example by assigning it to a
+tensor, before you can access the values of the resulting tensor.  You can also
+access the values through a TensorRef.
+
+
+## Built-in Tensor Methods
+
+These are usual C++ methods that act on tensors immediately.  They are not
+Operations which provide delayed evaluation of their results.  Unless specified
+otherwise, all the methods listed below are available on all tensor classes:
+Tensor, TensorFixedSize, and TensorMap.
+
+## Metadata
+
+### int NumDimensions
+
+Constant value indicating the number of dimensions of a Tensor.  This is also
+known as the tensor "rank".
+
+      Eigen::Tensor<float, 2> a(3, 4);
+      cout << "Dims " << a.NumDimensions;
+      => Dims 2
+
+### Dimensions dimensions()
+
+Returns an array-like object representing the dimensions of the tensor.
+The actual type of the `dimensions()` result is `<Tensor-Type>::``Dimensions`.
+
+    Eigen::Tensor<float, 2> a(3, 4);
+    const Eigen::Tensor<float, 2>::Dimensions& d = a.dimensions();
+    cout << "Dim size: " << d.size << ", dim 0: " << d[0]
+         << ", dim 1: " << d[1];
+    => Dim size: 2, dim 0: 3, dim 1: 4
+
+If you use a C++11 compiler, you can use `auto` to simplify the code:
+
+    const auto& d = a.dimensions();
+    cout << "Dim size: " << d.size << ", dim 0: " << d[0]
+         << ", dim 1: " << d[1];
+    => Dim size: 2, dim 0: 3, dim 1: 4
+
+### Index dimension(Index n)
+
+Returns the n-th dimension of the tensor.  The actual type of the
+`dimension()` result is `<Tensor-Type>::``Index`, but you can
+always use it like an int.
+
+      Eigen::Tensor<float, 2> a(3, 4);
+      int dim1 = a.dimension(1);
+      cout << "Dim 1: " << dim1;
+      => Dim 1: 4
+
+### Index size()
+
+Returns the total number of elements in the tensor.  This is the product of all
+the tensor dimensions.  The actual type of the `size()` result is
+`<Tensor-Type>::``Index`, but you can always use it like an int.
+
+    Eigen::Tensor<float, 2> a(3, 4);
+    cout << "Size: " << a.size();
+    => Size: 12
+
+
+### Getting Dimensions From An Operation
+
+A few operations provide `dimensions()` directly,
+e.g. `TensorReslicingOp`.  Most operations defer calculating dimensions
+until the operation is being evaluated.  If you need access to the dimensions
+of a deferred operation, you can wrap it in a TensorRef (see Assigning to a
+TensorRef above), which provides `dimensions()` and `dimension()` as
+above.
+
+TensorRef can also wrap the plain Tensor types, so this is a useful idiom in
+templated contexts where the underlying object could be either a raw Tensor
+or some deferred operation (e.g. a slice of a Tensor).  In this case, the
+template code can wrap the object in a TensorRef and reason about its
+dimensionality while remaining agnostic to the underlying type.
+
+
+## Constructors
+
+### Tensor
+
+Creates a tensor of the specified size. The number of arguments must be equal
+to the rank of the tensor. The content of the tensor is not initialized.
+
+    Eigen::Tensor<float, 2> a(3, 4);
+    cout << "NumRows: " << a.dimension(0) << " NumCols: " << a.dimension(1) << endl;
+    => NumRows: 3 NumCols: 4
+
+### TensorFixedSize
+
+Creates a tensor of the specified size. The number of arguments in the Sizes<>
+template parameter determines the rank of the tensor. The content of the tensor
+is not initialized.
+
+    Eigen::TensorFixedSize<float, Sizes<3, 4>> a;
+    cout << "Rank: " << a.rank() << endl;
+    => Rank: 2
+    cout << "NumRows: " << a.dimension(0) << " NumCols: " << a.dimension(1) << endl;
+    => NumRows: 3 NumCols: 4
+
+### TensorMap
+
+Creates a tensor mapping an existing array of data. The data must not be freed
+until the TensorMap is discarded, and the size of the data must be large enough
+to accommodate the coefficients of the tensor.
+
+    float data[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11};
+    Eigen::TensorMap<Tensor<float, 2>> a(data, 3, 4);
+    cout << "NumRows: " << a.dimension(0) << " NumCols: " << a.dimension(1) << endl;
+    => NumRows: 3 NumCols: 4
+    cout << "a(1, 2): " << a(1, 2) << endl;
+    => a(1, 2): 7
+
+
+## Contents Initialization
+
+When a new Tensor or a new TensorFixedSize are created, memory is allocated to
+hold all the tensor elements, but the memory is not initialized.  Similarly,
+when a new TensorMap is created on top of non-initialized memory the memory its
+contents are not initialized.
+
+You can use one of the methods below to initialize the tensor memory.  These
+have an immediate effect on the tensor and return the tensor itself as a
+result.  These are not tensor Operations which delay evaluation.
+
+### <Tensor-Type> setConstant(const Scalar& val)
+
+Sets all elements of the tensor to the constant value `val`.  `Scalar`
+is the type of data stored in the tensor.  You can pass any value that is
+convertible to that type.
+
+Returns the tensor itself in case you want to chain another call.
+
+    a.setConstant(12.3f);
+    cout << "Constant: " << endl << a << endl << endl;
+    =>
+    Constant:
+    12.3 12.3 12.3 12.3
+    12.3 12.3 12.3 12.3
+    12.3 12.3 12.3 12.3
+
+Note that `setConstant()` can be used on any tensor where the element type
+has a copy constructor and an `operator=()`:
+
+    Eigen::Tensor<string, 2> a(2, 3);
+    a.setConstant("yolo");
+    cout << "String tensor: " << endl << a << endl << endl;
+    =>
+    String tensor:
+    yolo yolo yolo
+    yolo yolo yolo
+
+
+### <Tensor-Type> setZero()
+
+Fills the tensor with zeros.  Equivalent to `setConstant(Scalar(0))`.
+Returns the tensor itself in case you want to chain another call.
+
+    a.setZero();
+    cout << "Zeros: " << endl << a << endl << endl;
+    =>
+    Zeros:
+    0 0 0 0
+    0 0 0 0
+    0 0 0 0
+
+
+### <Tensor-Type> setValues({..initializer_list})
+
+Fills the tensor with explicit values specified in a std::initializer_list.
+The type of the initializer list depends on the type and rank of the tensor.
+
+If the tensor has rank N, the initializer list must be nested N times.  The
+most deeply nested lists must contains P scalars of the Tensor type where P is
+the size of the last dimension of the Tensor.
+
+For example, for a `TensorFixedSize<float, 2, 3>` the initializer list must
+contains 2 lists of 3 floats each.
+
+`setValues()` returns the tensor itself in case you want to chain another
+call.
+
+    Eigen::Tensor<float, 2> a(2, 3);
+    a.setValues({{0.0f, 1.0f, 2.0f}, {3.0f, 4.0f, 5.0f}});
+    cout << "a" << endl << a << endl << endl;
+    =>
+    a
+    0 1 2
+    3 4 5
+
+If a list is too short, the corresponding elements of the tensor will not be
+changed.  This is valid at each level of nesting.  For example the following
+code only sets the values of the first row of the tensor.
+
+    Eigen::Tensor<int, 2> a(2, 3);
+    a.setConstant(1000);
+    a.setValues({{10, 20, 30}});
+    cout << "a" << endl << a << endl << endl;
+    =>
+    a
+    10   20   30
+    1000 1000 1000
+
+### <Tensor-Type> setRandom()
+
+Fills the tensor with random values.  Returns the tensor itself in case you
+want to chain another call.
+
+    a.setRandom();
+    cout << "Random: " << endl << a << endl << endl;
+    =>
+    Random:
+      0.680375    0.59688  -0.329554    0.10794
+     -0.211234   0.823295   0.536459 -0.0452059
+      0.566198  -0.604897  -0.444451   0.257742
+
+You can customize `setRandom()` by providing your own random number
+generator as a template argument:
+
+    a.setRandom<MyRandomGenerator>();
+
+Here, `MyRandomGenerator` must be a struct with the following member
+functions, where Scalar and Index are the same as `<Tensor-Type>::``Scalar`
+and `<Tensor-Type>::``Index`.
+
+See `struct UniformRandomGenerator` in TensorFunctors.h for an example.
+
+    // Custom number generator for use with setRandom().
+    struct MyRandomGenerator {
+      // Default and copy constructors. Both are needed
+      MyRandomGenerator() { }
+      MyRandomGenerator(const MyRandomGenerator& ) { }
+
+      // Return a random value to be used.  "element_location" is the
+      // location of the entry to set in the tensor, it can typically
+      // be ignored.
+      Scalar operator()(Eigen::DenseIndex element_location,
+                        Eigen::DenseIndex /*unused*/ = 0) const {
+        return <randomly generated value of type T>;
+      }
+
+      // Same as above but generates several numbers at a time.
+      typename internal::packet_traits<Scalar>::type packetOp(
+          Eigen::DenseIndex packet_location, Eigen::DenseIndex /*unused*/ = 0) const {
+        return <a packet of randomly generated values>;
+      }
+    };
+
+You can also use one of the 2 random number generators that are part of the
+tensor library:
+*   UniformRandomGenerator
+*   NormalRandomGenerator
+
+
+## Data Access
+
+The Tensor, TensorFixedSize, and TensorRef classes provide the following
+accessors to access the tensor coefficients:
+
+    const Scalar& operator()(const array<Index, NumIndices>& indices)
+    const Scalar& operator()(Index firstIndex, IndexTypes... otherIndices)
+    Scalar& operator()(const array<Index, NumIndices>& indices)
+    Scalar& operator()(Index firstIndex, IndexTypes... otherIndices)
+
+The number of indices must be equal to the rank of the tensor. Moreover, these
+accessors are not available on tensor expressions. In order to access the
+values of a tensor expression, the expression must either be evaluated or
+wrapped in a TensorRef.
+
+
+### Scalar* data() and const Scalar* data() const
+
+Returns a pointer to the storage for the tensor.  The pointer is const if the
+tensor was const.  This allows direct access to the data.  The layout of the
+data depends on the tensor layout: RowMajor or ColMajor.
+
+This access is usually only needed for special cases, for example when mixing
+Eigen Tensor code with other libraries.
+
+Scalar is the type of data stored in the tensor.
+
+    Eigen::Tensor<float, 2> a(3, 4);
+    float* a_data = a.data();
+    a_data[0] = 123.45f;
+    cout << "a(0, 0): " << a(0, 0);
+    => a(0, 0): 123.45
+
+
+## Tensor Operations
+
+All the methods documented below return non evaluated tensor `Operations`.
+These can be chained: you can apply another Tensor Operation to the value
+returned by the method.
+
+The chain of Operation is evaluated lazily, typically when it is assigned to a
+tensor.  See "Controlling when Expression are Evaluated" for more details about
+their evaluation.
+
+### <Operation> constant(const Scalar& val)
+
+Returns a tensor of the same type and dimensions as the original tensor but
+where all elements have the value `val`.
+
+This is useful, for example, when you want to add or subtract a constant from a
+tensor, or multiply every element of a tensor by a scalar.
+
+    Eigen::Tensor<float, 2> a(2, 3);
+    a.setConstant(1.0f);
+    Eigen::Tensor<float, 2> b = a + a.constant(2.0f);
+    Eigen::Tensor<float, 2> c = b * b.constant(0.2f);
+    cout << "a" << endl << a << endl << endl;
+    cout << "b" << endl << b << endl << endl;
+    cout << "c" << endl << c << endl << endl;
+    =>
+    a
+    1 1 1
+    1 1 1
+
+    b
+    3 3 3
+    3 3 3
+
+    c
+    0.6 0.6 0.6
+    0.6 0.6 0.6
+
+### <Operation> random()
+
+Returns a tensor of the same type and dimensions as the current tensor
+but where all elements have random values.
+
+This is for example useful to add random values to an existing tensor.
+The generation of random values can be customized in the same manner
+as for `setRandom()`.
+
+    Eigen::Tensor<float, 2> a(2, 3);
+    a.setConstant(1.0f);
+    Eigen::Tensor<float, 2> b = a + a.random();
+    cout << "a" << endl << a << endl << endl;
+    cout << "b" << endl << b << endl << endl;
+    =>
+    a
+    1 1 1
+    1 1 1
+
+    b
+    1.68038   1.5662  1.82329
+    0.788766  1.59688 0.395103
+
+
+## Unary Element Wise Operations
+
+All these operations take a single input tensor as argument and return a tensor
+of the same type and dimensions as the tensor to which they are applied.  The
+requested operations are applied to each element independently.
+
+### <Operation> operator-()
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the opposite values of the original tensor.
+
+    Eigen::Tensor<float, 2> a(2, 3);
+    a.setConstant(1.0f);
+    Eigen::Tensor<float, 2> b = -a;
+    cout << "a" << endl << a << endl << endl;
+    cout << "b" << endl << b << endl << endl;
+    =>
+    a
+    1 1 1
+    1 1 1
+
+    b
+    -1 -1 -1
+    -1 -1 -1
+
+### <Operation> sqrt()
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the square roots of the original tensor.
+
+### <Operation> rsqrt()
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the inverse square roots of the original tensor.
+
+### <Operation> square()
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the squares of the original tensor values.
+
+### <Operation> inverse()
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the inverse of the original tensor values.
+
+### <Operation> exp()
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the exponential of the original tensor.
+
+### <Operation> log()
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the natural logarithms of the original tensor.
+
+### <Operation> abs()
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the absolute values of the original tensor.
+
+### <Operation> pow(Scalar exponent)
+
+Returns a tensor of the same type and dimensions as the original tensor
+containing the coefficients of the original tensor to the power of the
+exponent.
+
+The type of the exponent, Scalar, is always the same as the type of the
+tensor coefficients.  For example, only integer exponents can be used in
+conjuntion with tensors of integer values.
+
+You can use cast() to lift this restriction.  For example this computes
+cubic roots of an int Tensor:
+
+    Eigen::Tensor<int, 2> a(2, 3);
+    a.setValues({{0, 1, 8}, {27, 64, 125}});
+    Eigen::Tensor<double, 2> b = a.cast<double>().pow(1.0 / 3.0);
+    cout << "a" << endl << a << endl << endl;
+    cout << "b" << endl << b << endl << endl;
+    =>
+    a
+    0   1   8
+    27  64 125
+
+    b
+    0 1 2
+    3 4 5
+
+### <Operation>  operator * (Scalar scale)
+
+Multiplies all the coefficients of the input tensor by the provided scale.
+
+### <Operation>  cwiseMax(Scalar threshold)
+TODO
+
+### <Operation>  cwiseMin(Scalar threshold)
+TODO
+
+### <Operation>  unaryExpr(const CustomUnaryOp& func)
+TODO
+
+
+## Binary Element Wise Operations
+
+These operations take two input tensors as arguments. The 2 input tensors should
+be of the same type and dimensions. The result is a tensor of the same
+dimensions as the tensors to which they are applied, and unless otherwise
+specified it is also of the same type. The requested operations are applied to
+each pair of elements independently.
+
+### <Operation> operator+(const OtherDerived& other)
+
+Returns a tensor of the same type and dimensions as the input tensors
+containing the coefficient wise sums of the inputs.
+
+### <Operation> operator-(const OtherDerived& other)
+
+Returns a tensor of the same type and dimensions as the input tensors
+containing the coefficient wise differences of the inputs.
+
+### <Operation> operator*(const OtherDerived& other)
+
+Returns a tensor of the same type and dimensions as the input tensors
+containing the coefficient wise products of the inputs.
+
+### <Operation> operator/(const OtherDerived& other)
+
+Returns a tensor of the same type and dimensions as the input tensors
+containing the coefficient wise quotients of the inputs.
+
+This operator is not supported for integer types.
+
+### <Operation> cwiseMax(const OtherDerived& other)
+
+Returns a tensor of the same type and dimensions as the input tensors
+containing the coefficient wise maximums of the inputs.
+
+### <Operation> cwiseMin(const OtherDerived& other)
+
+Returns a tensor of the same type and dimensions as the input tensors
+containing the coefficient wise mimimums of the inputs.
+
+### <Operation> Logical operators
+
+The following logical operators are supported as well:
+
+*   operator&&(const OtherDerived& other)
+*   operator||(const OtherDerived& other)
+*   operator<(const OtherDerived& other)
+*   operator<=(const OtherDerived& other)
+*   operator>(const OtherDerived& other)
+*   operator>=(const OtherDerived& other)
+*   operator==(const OtherDerived& other)
+*   operator!=(const OtherDerived& other)
+
+They all return a tensor of boolean values.
+
+
+## Selection (select(const ThenDerived& thenTensor, const ElseDerived& elseTensor)
+
+Selection is a coefficient-wise ternary operator that is the tensor equivalent
+to the if-then-else operation.
+
+    Tensor<bool, 3> if = ...;
+    Tensor<float, 3> then = ...;
+    Tensor<float, 3> else = ...;
+    Tensor<float, 3> result = if.select(then, else);
+
+The 3 arguments must be of the same dimensions, which will also be the dimension
+of the result.  The 'if' tensor must be of type boolean, the 'then' and the
+'else' tensor must be of the same type, which will also be the type of the
+result.
+
+Each coefficient in the result is equal to the corresponding coefficient in the
+'then' tensor if the corresponding value in the 'if' tensor is true. If not, the
+resulting coefficient will come from the 'else' tensor.
+
+
+## Contraction
+
+Tensor *contractions* are a generalization of the matrix product to the
+multidimensional case.
+
+    // Create 2 matrices using tensors of rank 2
+    Eigen::Tensor<int, 2> a(2, 3);
+    a.setValues({{1, 2, 3}, {6, 5, 4}});
+    Eigen::Tensor<int, 2> b(3, 2);
+    b.setValues({{1, 2}, {4, 5}, {5, 6}});
+
+    // Compute the traditional matrix product
+    Eigen::array<Eigen::IndexPair<int>, 1> product_dims = { Eigen::IndexPair<int>(1, 0) };
+    Eigen::Tensor<int, 2> AB = a.contract(b, product_dims);
+
+    // Compute the product of the transpose of the matrices
+    Eigen::array<Eigen::IndexPair<int>, 1> transposed_product_dims = { Eigen::IndexPair<int>(0, 1) };
+    Eigen::Tensor<int, 2> AtBt = a.contract(b, transposed_product_dims);
+
+    // Contraction to scalar value using a double contraction.
+    // First coordinate of both tensors are contracted as well as both second coordinates, i.e., this computes the sum of the squares of the elements.
+    Eigen::array<Eigen::IndexPair<int>, 2> double_contraction_product_dims = { Eigen::IndexPair<int>(0, 0), Eigen::IndexPair<int>(1, 1) };
+    Eigen::Tensor<int, 0> AdoubleContractedA = a.contract(a, double_contraction_product_dims);
+
+    // Extracting the scalar value of the tensor contraction for further usage
+    int value = AdoubleContractedA(0);
+
+## Reduction Operations
+
+A *Reduction* operation returns a tensor with fewer dimensions than the
+original tensor.  The values in the returned tensor are computed by applying a
+*reduction operator* to slices of values from the original tensor.  You specify
+the dimensions along which the slices are made.
+
+The Eigen Tensor library provides a set of predefined reduction operators such
+as `maximum()` and `sum()` and lets you define additional operators by
+implementing a few methods from a reductor template.
+
+### Reduction Dimensions
+
+All reduction operations take a single parameter of type
+`<TensorType>::``Dimensions` which can always be specified as an array of
+ints.  These are called the "reduction dimensions."  The values are the indices
+of the dimensions of the input tensor over which the reduction is done.  The
+parameter can have at most as many element as the rank of the input tensor;
+each element must be less than the tensor rank, as it indicates one of the
+dimensions to reduce.
+
+Each dimension of the input tensor should occur at most once in the reduction
+dimensions as the implementation does not remove duplicates.
+
+The order of the values in the reduction dimensions does not affect the
+results, but the code may execute faster if you list the dimensions in
+increasing order.
+
+Example: Reduction along one dimension.
+
+    // Create a tensor of 2 dimensions
+    Eigen::Tensor<int, 2> a(2, 3);
+    a.setValues({{1, 2, 3}, {6, 5, 4}});
+    // Reduce it along the second dimension (1)...
+    Eigen::array<int, 1> dims({1 /* dimension to reduce */});
+    // ...using the "maximum" operator.
+    // The result is a tensor with one dimension.  The size of
+    // that dimension is the same as the first (non-reduced) dimension of a.
+    Eigen::Tensor<int, 1> b = a.maximum(dims);
+    cout << "a" << endl << a << endl << endl;
+    cout << "b" << endl << b << endl << endl;
+    =>
+    a
+    1 2 3
+    6 5 4
+
+    b
+    3
+    6
+
+Example: Reduction along two dimensions.
+
+    Eigen::Tensor<float, 3, Eigen::ColMajor> a(2, 3, 4);
+    a.setValues({{{0.0f, 1.0f, 2.0f, 3.0f},
+                  {7.0f, 6.0f, 5.0f, 4.0f},
+                  {8.0f, 9.0f, 10.0f, 11.0f}},
+                 {{12.0f, 13.0f, 14.0f, 15.0f},
+                  {19.0f, 18.0f, 17.0f, 16.0f},
+                  {20.0f, 21.0f, 22.0f, 23.0f}}});
+    // The tensor a has 3 dimensions.  We reduce along the
+    // first 2, resulting in a tensor with a single dimension
+    // of size 4 (the last dimension of a.)
+    // Note that we pass the array of reduction dimensions
+    // directly to the maximum() call.
+    Eigen::Tensor<float, 1, Eigen::ColMajor> b =
+        a.maximum(Eigen::array<int, 2>({0, 1}));
+    cout << "b" << endl << b << endl << endl;
+    =>
+    b
+    20
+    21
+    22
+    23
+
+#### Reduction along all dimensions
+
+As a special case, if you pass no parameter to a reduction operation the
+original tensor is reduced along *all* its dimensions.  The result is a
+scalar, represented as a zero-dimension tensor.
+
+    Eigen::Tensor<float, 3> a(2, 3, 4);
+    a.setValues({{{0.0f, 1.0f, 2.0f, 3.0f},
+                  {7.0f, 6.0f, 5.0f, 4.0f},
+                  {8.0f, 9.0f, 10.0f, 11.0f}},
+                 {{12.0f, 13.0f, 14.0f, 15.0f},
+                  {19.0f, 18.0f, 17.0f, 16.0f},
+                  {20.0f, 21.0f, 22.0f, 23.0f}}});
+    // Reduce along all dimensions using the sum() operator.
+    Eigen::Tensor<float, 0> b = a.sum();
+    cout << "b" << endl << b << endl << endl;
+    =>
+    b
+    276
+
+
+### <Operation> sum(const Dimensions& new_dims)
+### <Operation> sum()
+
+Reduce a tensor using the sum() operator.  The resulting values
+are the sum of the reduced values.
+
+### <Operation> mean(const Dimensions& new_dims)
+### <Operation> mean()
+
+Reduce a tensor using the mean() operator.  The resulting values
+are the mean of the reduced values.
+
+### <Operation> maximum(const Dimensions& new_dims)
+### <Operation> maximum()
+
+Reduce a tensor using the maximum() operator.  The resulting values are the
+largest of the reduced values.
+
+### <Operation> minimum(const Dimensions& new_dims)
+### <Operation> minimum()
+
+Reduce a tensor using the minimum() operator.  The resulting values
+are the smallest of the reduced values.
+
+### <Operation> prod(const Dimensions& new_dims)
+### <Operation> prod()
+
+Reduce a tensor using the prod() operator.  The resulting values
+are the product of the reduced values.
+
+### <Operation> all(const Dimensions& new_dims)
+### <Operation> all()
+Reduce a tensor using the all() operator.  Casts tensor to bool and then checks
+whether all elements are true.  Runs through all elements rather than
+short-circuiting, so may be significantly inefficient.
+
+### <Operation> any(const Dimensions& new_dims)
+### <Operation> any()
+Reduce a tensor using the any() operator.  Casts tensor to bool and then checks
+whether any element is true.  Runs through all elements rather than
+short-circuiting, so may be significantly inefficient.
+
+
+### <Operation> reduce(const Dimensions& new_dims, const Reducer& reducer)
+
+Reduce a tensor using a user-defined reduction operator.  See `SumReducer`
+in TensorFunctors.h for information on how to implement a reduction operator.
+
+
+## Trace
+
+A *Trace* operation returns a tensor with fewer dimensions than the original
+tensor. It returns a tensor whose elements are the sum of the elements of the
+original tensor along the main diagonal for a list of specified dimensions, the
+"trace dimensions". Similar to the `Reduction Dimensions`, the trace dimensions
+are passed as an input parameter to the operation, are of type `<TensorType>::``Dimensions`
+, and have the same requirements when passed as an input parameter. In addition,
+the trace dimensions must have the same size.
+
+Example: Trace along 2 dimensions.
+
+    // Create a tensor of 3 dimensions
+    Eigen::Tensor<int, 3> a(2, 2, 3);
+    a.setValues({{{1, 2, 3}, {4, 5, 6}}, {{7, 8, 9}, {10, 11, 12}}});
+    // Specify the dimensions along which the trace will be computed.
+    // In this example, the trace can only be computed along the dimensions
+    // with indices 0 and 1
+    Eigen::array<int, 2> dims({0, 1});
+    // The output tensor contains all but the trace dimensions.
+    Tensor<int, 1> a_trace = a.trace(dims);
+    cout << "a_trace:" << endl;
+    cout << a_trace << endl;
+    =>
+    a_trace:
+    11
+    13
+    15
+
+
+### <Operation> trace(const Dimensions& new_dims)
+### <Operation> trace()
+
+As a special case, if no parameter is passed to the operation, trace is computed
+along *all* dimensions of the input tensor.
+
+Example: Trace along all dimensions.
+
+    // Create a tensor of 3 dimensions, with all dimensions having the same size.
+    Eigen::Tensor<int, 3> a(3, 3, 3);
+    a.setValues({{{1, 2, 3}, {4, 5, 6}, {7, 8, 9}},
+                {{10, 11, 12}, {13, 14, 15}, {16, 17, 18}},
+                {{19, 20, 21}, {22, 23, 24}, {25, 26, 27}}});
+    // Result is a zero dimension tensor
+    Tensor<int, 0> a_trace = a.trace();
+    cout<<"a_trace:"<<endl;
+    cout<<a_trace<<endl;
+    =>
+    a_trace:
+    42
+
+
+## Scan Operations
+
+A *Scan* operation returns a tensor with the same dimensions as the original
+tensor. The operation performs an inclusive scan along the specified
+axis, which means it computes a running total along the axis for a given
+reduction operation.
+If the reduction operation corresponds to summation, then this computes the
+prefix sum of the tensor along the given axis.
+
+Example:
+dd a comment to this line
+
+    // Create a tensor of 2 dimensions
+    Eigen::Tensor<int, 2> a(2, 3);
+    a.setValues({{1, 2, 3}, {4, 5, 6}});
+    // Scan it along the second dimension (1) using summation
+    Eigen::Tensor<int, 2> b = a.cumsum(1);
+    // The result is a tensor with the same size as the input
+    cout << "a" << endl << a << endl << endl;
+    cout << "b" << endl << b << endl << endl;
+    =>
+    a
+    1 2 3
+    4 5 6
+
+    b
+    1  3  6
+    4  9 15
+
+### <Operation> cumsum(const Index& axis)
+
+Perform a scan by summing consecutive entries.
+
+### <Operation> cumprod(const Index& axis)
+
+Perform a scan by multiplying consecutive entries.
+
+
+## Convolutions
+
+### <Operation> convolve(const Kernel& kernel, const Dimensions& dims)
+
+Returns a tensor that is the output of the convolution of the input tensor with the kernel,
+along the specified dimensions of the input tensor. The dimension size for dimensions of the output tensor
+which were part of the convolution will be reduced by the formula:
+output_dim_size = input_dim_size - kernel_dim_size + 1 (requires: input_dim_size >= kernel_dim_size).
+The dimension sizes for dimensions that were not part of the convolution will remain the same.
+Performance of the convolution can depend on the length of the stride(s) of the input tensor dimension(s) along which the
+convolution is computed (the first dimension has the shortest stride for ColMajor, whereas RowMajor's shortest stride is
+for the last dimension).
+
+    // Compute convolution along the second and third dimension.
+    Tensor<float, 4, DataLayout> input(3, 3, 7, 11);
+    Tensor<float, 2, DataLayout> kernel(2, 2);
+    Tensor<float, 4, DataLayout> output(3, 2, 6, 11);
+    input.setRandom();
+    kernel.setRandom();
+
+    Eigen::array<ptrdiff_t, 2> dims({1, 2});  // Specify second and third dimension for convolution.
+    output = input.convolve(kernel, dims);
+
+    for (int i = 0; i < 3; ++i) {
+      for (int j = 0; j < 2; ++j) {
+        for (int k = 0; k < 6; ++k) {
+          for (int l = 0; l < 11; ++l) {
+            const float result = output(i,j,k,l);
+            const float expected = input(i,j+0,k+0,l) * kernel(0,0) +
+                                   input(i,j+1,k+0,l) * kernel(1,0) +
+                                   input(i,j+0,k+1,l) * kernel(0,1) +
+                                   input(i,j+1,k+1,l) * kernel(1,1);
+            VERIFY_IS_APPROX(result, expected);
+          }
+        }
+      }
+    }
+
+
+## Geometrical Operations
+
+These operations return a Tensor with different dimensions than the original
+Tensor.  They can be used to access slices of tensors, see them with different
+dimensions, or pad tensors with additional data.
+
+### <Operation> reshape(const Dimensions& new_dims)
+
+Returns a view of the input tensor that has been reshaped to the specified
+new dimensions.  The argument new_dims is an array of Index values.  The
+rank of the resulting tensor is equal to the number of elements in new_dims.
+
+The product of all the sizes in the new dimension array must be equal to
+the number of elements in the input tensor.
+
+    // Increase the rank of the input tensor by introducing a new dimension
+    // of size 1.
+    Tensor<float, 2> input(7, 11);
+    array<int, 3> three_dims{{7, 11, 1}};
+    Tensor<float, 3> result = input.reshape(three_dims);
+
+    // Decrease the rank of the input tensor by merging 2 dimensions;
+    array<int, 1> one_dim{{7 * 11}};
+    Tensor<float, 1> result = input.reshape(one_dim);
+
+This operation does not move any data in the input tensor, so the resulting
+contents of a reshaped Tensor depend on the data layout of the original Tensor.
+
+For example this is what happens when you `reshape()` a 2D ColMajor tensor
+to one dimension:
+
+    Eigen::Tensor<float, 2, Eigen::ColMajor> a(2, 3);
+    a.setValues({{0.0f, 100.0f, 200.0f}, {300.0f, 400.0f, 500.0f}});
+    Eigen::array<Eigen::DenseIndex, 1> one_dim({3 * 2});
+    Eigen::Tensor<float, 1, Eigen::ColMajor> b = a.reshape(one_dim);
+    cout << "b" << endl << b << endl;
+    =>
+    b
+      0
+    300
+    100
+    400
+    200
+    500
+
+This is what happens when the 2D Tensor is RowMajor:
+
+    Eigen::Tensor<float, 2, Eigen::RowMajor> a(2, 3);
+    a.setValues({{0.0f, 100.0f, 200.0f}, {300.0f, 400.0f, 500.0f}});
+    Eigen::array<Eigen::DenseIndex, 1> one_dim({3 * 2});
+    Eigen::Tensor<float, 1, Eigen::RowMajor> b = a.reshape(one_dim);
+    cout << "b" << endl << b << endl;
+    =>
+    b
+      0
+    100
+    200
+    300
+    400
+    500
+
+The reshape operation is a lvalue. In other words, it can be used on the left
+side of the assignment operator.
+
+The previous example can be rewritten as follow:
+
+    Eigen::Tensor<float, 2, Eigen::ColMajor> a(2, 3);
+    a.setValues({{0.0f, 100.0f, 200.0f}, {300.0f, 400.0f, 500.0f}});
+    Eigen::array<Eigen::DenseIndex, 2> two_dim({2, 3});
+    Eigen::Tensor<float, 1, Eigen::ColMajor> b(6);
+    b.reshape(two_dim) = a;
+    cout << "b" << endl << b << endl;
+    =>
+    b
+      0
+    300
+    100
+    400
+    200
+    500
+
+Note that "b" itself was not reshaped but that instead the assignment is done to
+the reshape view of b.
+
+
+### <Operation> shuffle(const Shuffle& shuffle)
+
+Returns a copy of the input tensor whose dimensions have been
+reordered according to the specified permutation. The argument shuffle
+is an array of Index values. Its size is the rank of the input
+tensor. It must contain a permutation of 0, 1, ..., rank - 1. The i-th
+dimension of the output tensor equals to the size of the shuffle[i]-th
+dimension of the input tensor. For example:
+
+    // Shuffle all dimensions to the left by 1.
+    Tensor<float, 3> input(20, 30, 50);
+    // ... set some values in input.
+    Tensor<float, 3> output = input.shuffle({1, 2, 0})
+
+    eigen_assert(output.dimension(0) == 30);
+    eigen_assert(output.dimension(1) == 50);
+    eigen_assert(output.dimension(2) == 20);
+
+Indices into the output tensor are shuffled accordingly to formulate
+indices into the input tensor. For example, one can assert in the above
+code snippet that:
+
+    eigen_assert(output(3, 7, 11) == input(11, 3, 7));
+
+In general, one can assert that
+
+    eigen_assert(output(..., indices[shuffle[i]], ...) ==
+                 input(..., indices[i], ...))
+
+The shuffle operation results in a lvalue, which means that it can be assigned
+to. In other words, it can be used on the left side of the assignment operator.
+
+Let's rewrite the previous example to take advantage of this feature:
+
+    // Shuffle all dimensions to the left by 1.
+    Tensor<float, 3> input(20, 30, 50);
+    // ... set some values in input.
+    Tensor<float, 3> output(30, 50, 20);
+    output.shuffle({2, 0, 1}) = input;
+
+
+### <Operation> stride(const Strides& strides)
+
+Returns a view of the input tensor that strides (skips stride-1
+elements) along each of the dimensions.  The argument strides is an
+array of Index values.  The dimensions of the resulting tensor are
+ceil(input_dimensions[i] / strides[i]).
+
+For example this is what happens when you `stride()` a 2D tensor:
+
+    Eigen::Tensor<int, 2> a(4, 3);
+    a.setValues({{0, 100, 200}, {300, 400, 500}, {600, 700, 800}, {900, 1000, 1100}});
+    Eigen::array<Eigen::DenseIndex, 2> strides({3, 2});
+    Eigen::Tensor<int, 2> b = a.stride(strides);
+    cout << "b" << endl << b << endl;
+    =>
+    b
+       0   200
+     900  1100
+
+It is possible to assign a tensor to a stride:
+    Tensor<float, 3> input(20, 30, 50);
+    // ... set some values in input.
+    Tensor<float, 3> output(40, 90, 200);
+    output.stride({2, 3, 4}) = input;
+
+
+### <Operation> slice(const StartIndices& offsets, const Sizes& extents)
+
+Returns a sub-tensor of the given tensor. For each dimension i, the slice is
+made of the coefficients stored between offset[i] and offset[i] + extents[i] in
+the input tensor.
+
+    Eigen::Tensor<int, 2> a(4, 3);
+    a.setValues({{0, 100, 200}, {300, 400, 500},
+                 {600, 700, 800}, {900, 1000, 1100}});
+    Eigen::array<int, 2> offsets = {1, 0};
+    Eigen::array<int, 2> extents = {2, 2};
+    Eigen::Tensor<int, 1> slice = a.slice(offsets, extents);
+    cout << "a" << endl << a << endl;
+    =>
+    a
+       0   100   200
+     300   400   500
+     600   700   800
+     900  1000  1100
+    cout << "slice" << endl << slice << endl;
+    =>
+    slice
+     300   400
+     600   700
+
+
+### <Operation> chip(const Index offset, const Index dim)
+
+A chip is a special kind of slice. It is the subtensor at the given offset in
+the dimension dim. The returned tensor has one fewer dimension than the input
+tensor: the dimension dim is removed.
+
+For example, a matrix chip would be either a row or a column of the input
+matrix.
+
+    Eigen::Tensor<int, 2> a(4, 3);
+    a.setValues({{0, 100, 200}, {300, 400, 500},
+                 {600, 700, 800}, {900, 1000, 1100}});
+    Eigen::Tensor<int, 1> row_3 = a.chip(2, 0);
+    Eigen::Tensor<int, 1> col_2 = a.chip(1, 1);
+    cout << "a" << endl << a << endl;
+    =>
+    a
+       0   100   200
+     300   400   500
+     600   700   800
+     900  1000  1100
+    cout << "row_3" << endl << row_3 << endl;
+    =>
+    row_3
+       600   700   800
+    cout << "col_2" << endl << col_2 << endl;
+    =>
+    col_2
+       100   400   700    1000
+
+It is possible to assign values to a tensor chip since the chip operation is a
+lvalue. For example:
+
+    Eigen::Tensor<int, 1> a(3);
+    a.setValues({{100, 200, 300}});
+    Eigen::Tensor<int, 2> b(2, 3);
+    b.setZero();
+    b.chip(0, 0) = a;
+    cout << "a" << endl << a << endl;
+    =>
+    a
+     100
+     200
+     300
+    cout << "b" << endl << b << endl;
+    =>
+    b
+       100   200   300
+         0     0     0
+
+
+### <Operation> reverse(const ReverseDimensions& reverse)
+
+Returns a view of the input tensor that reverses the order of the coefficients
+along a subset of the dimensions.  The argument reverse is an array of boolean
+values that indicates whether or not the order of the coefficients should be
+reversed along each of the dimensions.  This operation preserves the dimensions
+of the input tensor.
+
+For example this is what happens when you `reverse()` the first dimension
+of a 2D tensor:
+
+    Eigen::Tensor<int, 2> a(4, 3);
+    a.setValues({{0, 100, 200}, {300, 400, 500},
+                {600, 700, 800}, {900, 1000, 1100}});
+    Eigen::array<bool, 2> reverse({true, false});
+    Eigen::Tensor<int, 2> b = a.reverse(reverse);
+    cout << "a" << endl << a << endl << "b" << endl << b << endl;
+    =>
+    a
+       0   100   200
+     300   400   500
+     600   700   800
+     900  1000  1100
+    b
+     900  1000  1100
+     600   700   800
+     300   400   500
+       0   100   200
+
+
+### <Operation> broadcast(const Broadcast& broadcast)
+
+Returns a view of the input tensor in which the input is replicated one to many
+times.
+The broadcast argument specifies how many copies of the input tensor need to be
+made in each of the dimensions.
+
+    Eigen::Tensor<int, 2> a(2, 3);
+    a.setValues({{0, 100, 200}, {300, 400, 500}});
+    Eigen::array<int, 2> bcast({3, 2});
+    Eigen::Tensor<int, 2> b = a.broadcast(bcast);
+    cout << "a" << endl << a << endl << "b" << endl << b << endl;
+    =>
+    a
+       0   100   200
+     300   400   500
+    b
+       0   100   200    0   100   200
+     300   400   500  300   400   500
+       0   100   200    0   100   200
+     300   400   500  300   400   500
+       0   100   200    0   100   200
+     300   400   500  300   400   500
+
+### <Operation> concatenate(const OtherDerived& other, Axis axis)
+
+TODO
+
+### <Operation>  pad(const PaddingDimensions& padding)
+
+Returns a view of the input tensor in which the input is padded with zeros.
+
+    Eigen::Tensor<int, 2> a(2, 3);
+    a.setValues({{0, 100, 200}, {300, 400, 500}});
+    Eigen::array<pair<int, int>, 2> paddings;
+    paddings[0] = make_pair(0, 1);
+    paddings[1] = make_pair(2, 3);
+    Eigen::Tensor<int, 2> b = a.pad(paddings);
+    cout << "a" << endl << a << endl << "b" << endl << b << endl;
+    =>
+    a
+       0   100   200
+     300   400   500
+    b
+       0     0     0    0
+       0     0     0    0
+       0   100   200    0
+     300   400   500    0
+       0     0     0    0
+       0     0     0    0
+       0     0     0    0
+
+
+### <Operation>  extract_patches(const PatchDims& patch_dims)
+
+Returns a tensor of coefficient patches extracted from the input tensor, where
+each patch is of dimension specified by 'patch_dims'. The returned tensor has
+one greater dimension than the input tensor, which is used to index each patch.
+The patch index in the output tensor depends on the data layout of the input
+tensor: the patch index is the last dimension ColMajor layout, and the first
+dimension in RowMajor layout.
+
+For example, given the following input tensor:
+
+    Eigen::Tensor<float, 2, DataLayout> tensor(3,4);
+    tensor.setValues({{0.0f, 1.0f, 2.0f, 3.0f},
+                      {4.0f, 5.0f, 6.0f, 7.0f},
+                      {8.0f, 9.0f, 10.0f, 11.0f}});
+
+    cout << "tensor: " << endl << tensor << endl;
+    =>
+    tensor:
+     0   1   2   3
+     4   5   6   7
+     8   9  10  11
+
+Six 2x2 patches can be extracted and indexed using the following code:
+
+    Eigen::Tensor<float, 3, DataLayout> patch;
+    Eigen::array<ptrdiff_t, 2> patch_dims;
+    patch_dims[0] = 2;
+    patch_dims[1] = 2;
+    patch = tensor.extract_patches(patch_dims);
+    for (int k = 0; k < 6; ++k) {
+      cout << "patch index: " << k << endl;
+      for (int i = 0; i < 2; ++i) {
+    	for (int j = 0; j < 2; ++j) {
+    	  if (DataLayout == ColMajor) {
+    		cout << patch(i, j, k) << " ";
+    	  } else {
+    		cout << patch(k, i, j) << " ";
+    	  }
+    	}
+    	cout << endl;
+      }
+    }
+
+This code results in the following output when the data layout is ColMajor:
+
+    patch index: 0
+    0 1
+    4 5
+    patch index: 1
+    4 5
+    8 9
+    patch index: 2
+    1 2
+    5 6
+    patch index: 3
+    5 6
+    9 10
+    patch index: 4
+    2 3
+    6 7
+    patch index: 5
+    6 7
+    10 11
+
+This code results in the following output when the data layout is RowMajor:
+(NOTE: the set of patches is the same as in ColMajor, but are indexed differently).
+
+    patch index: 0
+    0 1
+    4 5
+    patch index: 1
+    1 2
+    5 6
+    patch index: 2
+    2 3
+    6 7
+    patch index: 3
+    4 5
+    8 9
+    patch index: 4
+    5 6
+    9 10
+    patch index: 5
+    6 7
+    10 11
+
+### <Operation>  extract_image_patches(const Index patch_rows, const Index patch_cols, const Index row_stride, const Index col_stride, const PaddingType padding_type)
+
+Returns a tensor of coefficient image patches extracted from the input tensor,
+which is expected to have dimensions ordered as follows (depending on the data
+layout of the input tensor, and the number of additional dimensions 'N'):
+
+*) ColMajor
+1st dimension: channels (of size d)
+2nd dimension: rows (of size r)
+3rd dimension: columns (of size c)
+4th-Nth dimension: time (for video) or batch (for bulk processing).
+
+*) RowMajor (reverse order of ColMajor)
+1st-Nth dimension: time (for video) or batch (for bulk processing).
+N+1'th dimension: columns (of size c)
+N+2'th dimension: rows (of size r)
+N+3'th dimension: channels (of size d)
+
+The returned tensor has one greater dimension than the input tensor, which is
+used to index each patch. The patch index in the output tensor depends on the
+data layout of the input tensor: the patch index is the 4'th dimension in
+ColMajor layout, and the 4'th from the last dimension in RowMajor layout.
+
+For example, given the following input tensor with the following dimension
+sizes:
+ *) depth:   2
+ *) rows:    3
+ *) columns: 5
+ *) batch:   7
+
+    Tensor<float, 4> tensor(2,3,5,7);
+    Tensor<float, 4, RowMajor> tensor_row_major = tensor.swap_layout();
+
+2x2 image patches can be extracted and indexed using the following code:
+
+*) 2D patch: ColMajor (patch indexed by second-to-last dimension)
+
+    Tensor<float, 5> twod_patch;
+    twod_patch = tensor.extract_image_patches<2, 2>();
+    // twod_patch.dimension(0) == 2
+    // twod_patch.dimension(1) == 2
+    // twod_patch.dimension(2) == 2
+    // twod_patch.dimension(3) == 3*5
+    // twod_patch.dimension(4) == 7
+
+*) 2D patch: RowMajor (patch indexed by the second dimension)
+
+    Tensor<float, 5, RowMajor> twod_patch_row_major;
+    twod_patch_row_major = tensor_row_major.extract_image_patches<2, 2>();
+    // twod_patch_row_major.dimension(0) == 7
+    // twod_patch_row_major.dimension(1) == 3*5
+    // twod_patch_row_major.dimension(2) == 2
+    // twod_patch_row_major.dimension(3) == 2
+    // twod_patch_row_major.dimension(4) == 2
+
+## Special Operations
+
+### <Operation> cast<T>()
+
+Returns a tensor of type T with the same dimensions as the original tensor.
+The returned tensor contains the values of the original tensor converted to
+type T.
+
+    Eigen::Tensor<float, 2> a(2, 3);
+    Eigen::Tensor<int, 2> b = a.cast<int>();
+
+This can be useful for example if you need to do element-wise division of
+Tensors of integers.  This is not currently supported by the Tensor library
+but you can easily cast the tensors to floats to do the division:
+
+    Eigen::Tensor<int, 2> a(2, 3);
+    a.setValues({{0, 1, 2}, {3, 4, 5}});
+    Eigen::Tensor<int, 2> b =
+        (a.cast<float>() / a.constant(2).cast<float>()).cast<int>();
+    cout << "a" << endl << a << endl << endl;
+    cout << "b" << endl << b << endl << endl;
+    =>
+    a
+    0 1 2
+    3 4 5
+
+    b
+    0 0 1
+    1 2 2
+
+
+### <Operation>     eval()
+
+TODO
+
+
+## Representation of scalar values
+
+Scalar values are often represented by tensors of size 1 and rank 0.For example
+Tensor<T, N>::maximum() currently returns a Tensor<T, 0>. Similarly, the inner
+product of 2 1d tensors (through contractions) returns a 0d tensor.
+
+## Limitations
+
+*   The number of tensor dimensions is currently limited to 250 when using a
+    compiler that supports cxx11. It is limited to only 5 for older compilers.
+*   The IndexList class requires a cxx11 compliant compiler. You can use an
+    array of indices instead if you don't have access to a modern compiler.
+*   On GPUs only floating point values are properly tested and optimized for.
+*   Complex and integer values are known to be broken on GPUs. If you try to use
+    them you'll most likely end up triggering a static assertion failure such as
+    EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
+
+
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/Tensor.h b/src/EigenUnsupported/CXX11/src/Tensor/Tensor.h
new file mode 100644
index 0000000..8cac2bb
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/Tensor.h
@@ -0,0 +1,554 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_H
+#define EIGEN_CXX11_TENSOR_TENSOR_H
+
+namespace Eigen {
+
+/** \class Tensor
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief The tensor class.
+  *
+  * The %Tensor class is the work-horse for all \em dense tensors within Eigen.
+  *
+  * The %Tensor class encompasses only dynamic-size objects so far.
+  *
+  * The first two template parameters are required:
+  * \tparam Scalar_  Numeric type, e.g. float, double, int or `std::complex<float>`.
+  *                 User defined scalar types are supported as well (see \ref user_defined_scalars "here").
+  * \tparam NumIndices_ Number of indices (i.e. rank of the tensor)
+  *
+  * The remaining template parameters are optional -- in most cases you don't have to worry about them.
+  * \tparam Options_  A combination of either \b #RowMajor or \b #ColMajor, and of either
+  *                 \b #AutoAlign or \b #DontAlign.
+  *                 The former controls \ref TopicStorageOrders "storage order", and defaults to column-major. The latter controls alignment, which is required
+  *                 for vectorization. It defaults to aligning tensors. Note that tensors currently do not support any operations that profit from vectorization.
+  *                 Support for such operations (i.e. adding two tensors etc.) is planned.
+  *
+  * You can access elements of tensors using normal subscripting:
+  *
+  * \code
+  * Eigen::Tensor<double, 4> t(10, 10, 10, 10);
+  * t(0, 1, 2, 3) = 42.0;
+  * \endcode
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_TENSOR_PLUGIN.
+  *
+  * <i><b>Some notes:</b></i>
+  *
+  * <dl>
+  * <dt><b>Relation to other parts of Eigen:</b></dt>
+  * <dd>The midterm development goal for this class is to have a similar hierarchy as Eigen uses for matrices, so that
+  * taking blocks or using tensors in expressions is easily possible, including an interface with the vector/matrix code
+  * by providing .asMatrix() and .asVector() (or similar) methods for rank 2 and 1 tensors. However, currently, the %Tensor
+  * class does not provide any of these features and is only available as a stand-alone class that just allows for
+  * coefficient access. Also, when fixed-size tensors are implemented, the number of template arguments is likely to
+  * change dramatically.</dd>
+  * </dl>
+  *
+  * \ref TopicStorageOrders
+  */
+
+template<typename Scalar_, int NumIndices_, int Options_, typename IndexType_>
+class Tensor : public TensorBase<Tensor<Scalar_, NumIndices_, Options_, IndexType_> >
+{
+  public:
+    typedef Tensor<Scalar_, NumIndices_, Options_, IndexType_> Self;
+    typedef TensorBase<Tensor<Scalar_, NumIndices_, Options_, IndexType_> > Base;
+    typedef typename Eigen::internal::nested<Self>::type Nested;
+    typedef typename internal::traits<Self>::StorageKind StorageKind;
+    typedef typename internal::traits<Self>::Index Index;
+    typedef Scalar_ Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef typename Base::CoeffReturnType CoeffReturnType;
+
+    enum {
+      IsAligned = bool(EIGEN_MAX_ALIGN_BYTES>0) & !(Options_&DontAlign),
+      Layout = Options_ & RowMajor ? RowMajor : ColMajor,
+      CoordAccess = true,
+      RawAccess = true
+    };
+
+    static const int Options = Options_;
+    static const int NumIndices = NumIndices_;
+    typedef DSizes<Index, NumIndices_> Dimensions;
+
+  protected:
+    TensorStorage<Scalar, Dimensions, Options> m_storage;
+
+#ifdef EIGEN_HAS_SFINAE
+    template<typename CustomIndices>
+    struct isOfNormalIndex{
+      static const bool is_array = internal::is_base_of<array<Index, NumIndices>, CustomIndices>::value;
+      static const bool is_int = NumTraits<CustomIndices>::IsInteger;
+      static const bool value = is_array | is_int;
+    };
+#endif
+
+  public:
+    // Metadata
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index                         rank()                   const { return NumIndices; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index                         dimension(std::size_t n) const { return m_storage.dimensions()[n]; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions&             dimensions()             const { return m_storage.dimensions(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index                         size()                   const { return m_storage.size(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar                        *data()                        { return m_storage.data(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar                  *data()                  const { return m_storage.data(); }
+
+    // This makes EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    // work, because that uses base().coeffRef() - and we don't yet
+    // implement a similar class hierarchy
+    inline Self& base()             { return *this; }
+    inline const Self& base() const { return *this; }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const
+    {
+      // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return coeff(array<Index, NumIndices>{{firstIndex, secondIndex, otherIndices...}});
+    }
+#endif
+
+    // normal indices
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(const array<Index, NumIndices>& indices) const
+    {
+      eigen_internal_assert(checkIndexRange(indices));
+      return m_storage.data()[linearizedIndex(indices)];
+    }
+
+    // custom indices
+#ifdef EIGEN_HAS_SFINAE
+    template<typename CustomIndices,
+             EIGEN_SFINAE_ENABLE_IF( !(isOfNormalIndex<CustomIndices>::value) )
+    >
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(CustomIndices& indices) const
+    {
+        return coeff(internal::customIndices2Array<Index,NumIndices>(indices));
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff() const
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return m_storage.data()[0];
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const
+    {
+      eigen_internal_assert(index >= 0 && index < size());
+      return m_storage.data()[index];
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    inline Scalar& coeffRef(Index firstIndex, Index secondIndex, IndexTypes... otherIndices)
+    {
+      // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return coeffRef(array<Index, NumIndices>{{firstIndex, secondIndex, otherIndices...}});
+    }
+#endif
+
+    // normal indices
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(const array<Index, NumIndices>& indices)
+    {
+      eigen_internal_assert(checkIndexRange(indices));
+      return m_storage.data()[linearizedIndex(indices)];
+    }
+
+    // custom indices
+#ifdef EIGEN_HAS_SFINAE
+    template<typename CustomIndices,
+             EIGEN_SFINAE_ENABLE_IF( !(isOfNormalIndex<CustomIndices>::value) )
+             >
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(CustomIndices& indices)
+    {
+        return coeffRef(internal::customIndices2Array<Index,NumIndices>(indices));
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef()
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return m_storage.data()[0];
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
+    {
+      eigen_internal_assert(index >= 0 && index < size());
+      return m_storage.data()[index];
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    inline const Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const
+    {
+      // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return this->operator()(array<Index, NumIndices>{{firstIndex, secondIndex, otherIndices...}});
+    }
+#else
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1) const
+    {
+      return coeff(array<Index, 2>(i0, i1));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2) const
+    {
+      return coeff(array<Index, 3>(i0, i1, i2));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2, Index i3) const
+    {
+      return coeff(array<Index, 4>(i0, i1, i2, i3));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2, Index i3, Index i4) const
+    {
+      return coeff(array<Index, 5>(i0, i1, i2, i3, i4));
+    }
+#endif
+
+    // custom indices
+#ifdef EIGEN_HAS_SFINAE
+    template<typename CustomIndices,
+             EIGEN_SFINAE_ENABLE_IF( !(isOfNormalIndex<CustomIndices>::value) )
+    >
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()(CustomIndices& indices) const
+    {
+        return coeff(internal::customIndices2Array<Index,NumIndices>(indices));
+    }
+#endif
+
+    // normal indices
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()(const array<Index, NumIndices>& indices) const
+    {
+      return coeff(indices);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()(Index index) const
+    {
+      eigen_internal_assert(index >= 0 && index < size());
+      return coeff(index);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()() const
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return coeff();
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator[](Index index) const
+    {
+      // The bracket operator is only for vectors, use the parenthesis operator instead.
+      EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return coeff(index);
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    inline Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices)
+    {
+      // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return operator()(array<Index, NumIndices>{{firstIndex, secondIndex, otherIndices...}});
+    }
+#else
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1)
+    {
+      return coeffRef(array<Index, 2>(i0, i1));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2)
+    {
+      return coeffRef(array<Index, 3>(i0, i1, i2));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2, Index i3)
+    {
+      return coeffRef(array<Index, 4>(i0, i1, i2, i3));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2, Index i3, Index i4)
+    {
+      return coeffRef(array<Index, 5>(i0, i1, i2, i3, i4));
+    }
+#endif
+
+    // normal indices
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()(const array<Index, NumIndices>& indices)
+    {
+      return coeffRef(indices);
+    }
+
+    // custom indices
+#ifdef EIGEN_HAS_SFINAE
+    template<typename CustomIndices,
+             EIGEN_SFINAE_ENABLE_IF( !(isOfNormalIndex<CustomIndices>::value) )
+    >
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()(CustomIndices& indices)
+    {
+      return coeffRef(internal::customIndices2Array<Index,NumIndices>(indices));
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()(Index index)
+    {
+      eigen_assert(index >= 0 && index < size());
+      return coeffRef(index);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()()
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return coeffRef();
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator[](Index index)
+    {
+      // The bracket operator is only for vectors, use the parenthesis operator instead
+      EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return coeffRef(index);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor()
+      : m_storage()
+    {
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor(const Self& other)
+      : m_storage(other.m_storage)
+    {
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor(Index firstDimension, IndexTypes... otherDimensions)
+        : m_storage(firstDimension, otherDimensions...)
+    {
+      // The number of dimensions used to construct a tensor must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+#else
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit Tensor(Index dim1)
+      : m_storage(dim1, array<Index, 1>(dim1))
+    {
+      EIGEN_STATIC_ASSERT(1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor(Index dim1, Index dim2)
+      : m_storage(dim1*dim2, array<Index, 2>(dim1, dim2))
+    {
+      EIGEN_STATIC_ASSERT(2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor(Index dim1, Index dim2, Index dim3)
+      : m_storage(dim1*dim2*dim3, array<Index, 3>(dim1, dim2, dim3))
+    {
+      EIGEN_STATIC_ASSERT(3 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor(Index dim1, Index dim2, Index dim3, Index dim4)
+      : m_storage(dim1*dim2*dim3*dim4, array<Index, 4>(dim1, dim2, dim3, dim4))
+    {
+      EIGEN_STATIC_ASSERT(4 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor(Index dim1, Index dim2, Index dim3, Index dim4, Index dim5)
+      : m_storage(dim1*dim2*dim3*dim4*dim5, array<Index, 5>(dim1, dim2, dim3, dim4, dim5))
+    {
+      EIGEN_STATIC_ASSERT(5 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+#endif
+
+    /** Normal Dimension */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit Tensor(const array<Index, NumIndices>& dimensions)
+        : m_storage(internal::array_prod(dimensions), dimensions)
+    {
+      EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    }
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor(const TensorBase<OtherDerived, ReadOnlyAccessors>& other)
+    {
+      typedef TensorAssignOp<Tensor, const OtherDerived> Assign;
+      Assign assign(*this, other.derived());
+      resize(TensorEvaluator<const Assign, DefaultDevice>(assign, DefaultDevice()).dimensions());
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
+    }
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor(const TensorBase<OtherDerived, WriteAccessors>& other)
+    {
+      typedef TensorAssignOp<Tensor, const OtherDerived> Assign;
+      Assign assign(*this, other.derived());
+      resize(TensorEvaluator<const Assign, DefaultDevice>(assign, DefaultDevice()).dimensions());
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
+    }
+
+    #if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor(Self&& other)
+      : m_storage(std::move(other.m_storage))
+    {
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor& operator=(Self&& other)
+    {
+      m_storage = std::move(other.m_storage);
+      return *this;
+    }
+    #endif
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor& operator=(const Tensor& other)
+    {
+      typedef TensorAssignOp<Tensor, const Tensor> Assign;
+      Assign assign(*this, other);
+      resize(TensorEvaluator<const Assign, DefaultDevice>(assign, DefaultDevice()).dimensions());
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
+      return *this;
+    }
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Tensor& operator=(const OtherDerived& other)
+    {
+      typedef TensorAssignOp<Tensor, const OtherDerived> Assign;
+      Assign assign(*this, other);
+      resize(TensorEvaluator<const Assign, DefaultDevice>(assign, DefaultDevice()).dimensions());
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
+      return *this;
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes> EIGEN_DEVICE_FUNC
+    void resize(Index firstDimension, IndexTypes... otherDimensions)
+    {
+      // The number of dimensions used to resize a tensor must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      resize(array<Index, NumIndices>{{firstDimension, otherDimensions...}});
+    }
+#endif
+
+    /** Normal Dimension */
+    EIGEN_DEVICE_FUNC void resize(const array<Index, NumIndices>& dimensions)
+    {
+      int i;
+      Index size = Index(1);
+      for (i = 0; i < NumIndices; i++) {
+        internal::check_rows_cols_for_overflow<Dynamic>::run(size, dimensions[i]);
+        size *= dimensions[i];
+      }
+      #ifdef EIGEN_INITIALIZE_COEFFS
+        bool size_changed = size != this->size();
+        m_storage.resize(size, dimensions);
+        if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+      #else
+        m_storage.resize(size, dimensions);
+      #endif
+    }
+
+    // Why this overload, DSizes is derived from array ??? //
+    EIGEN_DEVICE_FUNC void resize(const DSizes<Index, NumIndices>& dimensions) {
+      array<Index, NumIndices> dims;
+      for (int i = 0; i < NumIndices; ++i) {
+        dims[i] = dimensions[i];
+      }
+      resize(dims);
+    }
+
+    EIGEN_DEVICE_FUNC
+    void resize()
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      // Nothing to do: rank 0 tensors have fixed size
+    }
+
+#ifdef EIGEN_HAS_INDEX_LIST
+    template <typename FirstType, typename... OtherTypes>
+    EIGEN_DEVICE_FUNC
+    void resize(const Eigen::IndexList<FirstType, OtherTypes...>& dimensions) {
+      array<Index, NumIndices> dims;
+      for (int i = 0; i < NumIndices; ++i) {
+        dims[i] = static_cast<Index>(dimensions[i]);
+      }
+      resize(dims);
+    }
+#endif
+
+    /** Custom Dimension */
+#ifdef EIGEN_HAS_SFINAE
+    template<typename CustomDimension,
+             EIGEN_SFINAE_ENABLE_IF( !(isOfNormalIndex<CustomDimension>::value) )
+    >
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(CustomDimension& dimensions)
+    {
+      resize(internal::customIndices2Array<Index,NumIndices>(dimensions));
+    }
+#endif
+
+#ifndef EIGEN_EMULATE_CXX11_META_H
+    template <typename std::ptrdiff_t... Indices>
+    EIGEN_DEVICE_FUNC
+    void resize(const Sizes<Indices...>& dimensions) {
+      array<Index, NumIndices> dims;
+      for (int i = 0; i < NumIndices; ++i) {
+        dims[i] = static_cast<Index>(dimensions[i]);
+      }
+      resize(dims);
+    }
+#else
+    template <std::size_t V1, std::size_t V2, std::size_t V3, std::size_t V4, std::size_t V5>
+    EIGEN_DEVICE_FUNC
+    void resize(const Sizes<V1, V2, V3, V4, V5>& dimensions) {
+      array<Index, NumIndices> dims;
+      for (int i = 0; i < NumIndices; ++i) {
+        dims[i] = static_cast<Index>(dimensions[i]);
+      }
+      resize(dims);
+    }
+#endif
+
+  protected:
+
+    bool checkIndexRange(const array<Index, NumIndices>& indices) const
+    {
+      using internal::array_apply_and_reduce;
+      using internal::array_zip_and_reduce;
+      using internal::greater_equal_zero_op;
+      using internal::logical_and_op;
+      using internal::lesser_op;
+
+      return
+        // check whether the indices are all >= 0
+        array_apply_and_reduce<logical_and_op, greater_equal_zero_op>(indices) &&
+        // check whether the indices fit in the dimensions
+        array_zip_and_reduce<logical_and_op, lesser_op>(indices, m_storage.dimensions());
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index linearizedIndex(const array<Index, NumIndices>& indices) const
+    {
+      if (Options&RowMajor) {
+        return m_storage.dimensions().IndexOfRowMajor(indices);
+      } else {
+        return m_storage.dimensions().IndexOfColMajor(indices);
+      }
+    }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorArgMax.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorArgMax.h
new file mode 100644
index 0000000..8b8fb92
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorArgMax.h
@@ -0,0 +1,329 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Eugene Brevdo <ebrevdo@gmail.com>
+//                    Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_ARG_MAX_H
+#define EIGEN_CXX11_TENSOR_TENSOR_ARG_MAX_H
+
+namespace Eigen {
+namespace internal {
+
+/** \class TensorIndexTuple
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor + Index Tuple class.
+  *
+  *
+  */
+template<typename XprType>
+struct traits<TensorIndexTupleOp<XprType> > : public traits<XprType>
+{
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef Tuple<Index, typename XprTraits::Scalar> Scalar;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+};
+
+template<typename XprType>
+struct eval<TensorIndexTupleOp<XprType>, Eigen::Dense>
+{
+  typedef const TensorIndexTupleOp<XprType>EIGEN_DEVICE_REF type;
+};
+
+template<typename XprType>
+struct nested<TensorIndexTupleOp<XprType>, 1,
+              typename eval<TensorIndexTupleOp<XprType> >::type>
+{
+  typedef TensorIndexTupleOp<XprType> type;
+};
+
+}  // end namespace internal
+
+template<typename XprType>
+class TensorIndexTupleOp : public TensorBase<TensorIndexTupleOp<XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorIndexTupleOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename Eigen::internal::nested<TensorIndexTupleOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorIndexTupleOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorIndexTupleOp>::Index Index;
+  typedef Tuple<Index, typename XprType::CoeffReturnType> CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIndexTupleOp(const XprType& expr)
+      : m_xpr(expr) {}
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename XprType::Nested>::type&
+  expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+};
+
+// Eval as rvalue
+template<typename ArgType, typename Device>
+struct TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device>
+{
+  typedef TensorIndexTupleOp<ArgType> XprType;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
+  static const int NumDims = internal::array_size<Dimensions>::value;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
+    PacketAccess = /*TensorEvaluator<ArgType, Device>::PacketAccess*/ false,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device) { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
+    return m_impl.dimensions();
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return CoeffReturnType(index, m_impl.coeff(index));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, 1);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ protected:
+  TensorEvaluator<ArgType, Device> m_impl;
+};
+
+namespace internal {
+
+/** \class TensorTupleIndex
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Converts to Tensor<Tuple<Index, Scalar> > and reduces to Tensor<Index>.
+  *
+  */
+template<typename ReduceOp, typename Dims, typename XprType>
+struct traits<TensorTupleReducerOp<ReduceOp, Dims, XprType> > : public traits<XprType>
+{
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef Index Scalar;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions - array_size<Dims>::value;
+  static const int Layout = XprTraits::Layout;
+};
+
+template<typename ReduceOp, typename Dims, typename XprType>
+struct eval<TensorTupleReducerOp<ReduceOp, Dims, XprType>, Eigen::Dense>
+{
+  typedef const TensorTupleReducerOp<ReduceOp, Dims, XprType>EIGEN_DEVICE_REF type;
+};
+
+template<typename ReduceOp, typename Dims, typename XprType>
+struct nested<TensorTupleReducerOp<ReduceOp, Dims, XprType>, 1,
+              typename eval<TensorTupleReducerOp<ReduceOp, Dims, XprType> >::type>
+{
+  typedef TensorTupleReducerOp<ReduceOp, Dims, XprType> type;
+};
+
+}  // end namespace internal
+
+template<typename ReduceOp, typename Dims, typename XprType>
+class TensorTupleReducerOp : public TensorBase<TensorTupleReducerOp<ReduceOp, Dims, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorTupleReducerOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename Eigen::internal::nested<TensorTupleReducerOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorTupleReducerOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorTupleReducerOp>::Index Index;
+  typedef Index CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorTupleReducerOp(const XprType& expr,
+                                                          const ReduceOp& reduce_op,
+                                                          const Index return_dim,
+                                                          const Dims& reduce_dims)
+      : m_xpr(expr), m_reduce_op(reduce_op), m_return_dim(return_dim), m_reduce_dims(reduce_dims) {}
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename XprType::Nested>::type&
+  expression() const { return m_xpr; }
+
+  EIGEN_DEVICE_FUNC
+  const ReduceOp& reduce_op() const { return m_reduce_op; }
+
+  EIGEN_DEVICE_FUNC
+  const Dims& reduce_dims() const { return m_reduce_dims; }
+
+  EIGEN_DEVICE_FUNC
+  Index return_dim() const { return m_return_dim; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const ReduceOp m_reduce_op;
+    const Index m_return_dim;
+    const Dims m_reduce_dims;
+};
+
+// Eval as rvalue
+template<typename ReduceOp, typename Dims, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Device>
+{
+  typedef TensorTupleReducerOp<ReduceOp, Dims, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename TensorIndexTupleOp<ArgType>::CoeffReturnType TupleType;
+  typedef typename TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device>::Dimensions Dimensions;
+  typedef typename TensorEvaluator<const TensorIndexTupleOp<ArgType> , Device>::Dimensions InputDimensions;
+  static const int NumDims = internal::array_size<InputDimensions>::value;
+  typedef array<Index, NumDims> StrideDims;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  typedef StorageMemory<TupleType, Device> TupleStorageMem;
+
+  enum {
+    IsAligned         = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
+    PacketAccess      = /*TensorEvaluator<ArgType, Device>::PacketAccess*/ false,
+    BlockAccess       = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout            = TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_orig_impl(op.expression(), device),
+        m_impl(op.expression().index_tuples().reduce(op.reduce_dims(), op.reduce_op()), device),
+        m_return_dim(op.return_dim())
+  {
+    gen_strides(m_orig_impl.dimensions(), m_strides);
+    if (Layout == static_cast<int>(ColMajor)) {
+      const Index total_size = internal::array_prod(m_orig_impl.dimensions());
+      m_stride_mod = (m_return_dim < NumDims - 1) ? m_strides[m_return_dim + 1] : total_size;
+    } else {
+      const Index total_size = internal::array_prod(m_orig_impl.dimensions());
+      m_stride_mod = (m_return_dim > 0) ? m_strides[m_return_dim - 1] : total_size;
+    }    
+    // If m_return_dim is not a valid index, returns 1 or this can crash on Windows.
+    m_stride_div = ((m_return_dim >= 0) &&
+                    (m_return_dim < static_cast<Index>(m_strides.size())))
+                   ? m_strides[m_return_dim] : 1;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
+    return m_impl.dimensions();
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    const TupleType v = m_impl.coeff(index);
+    return (m_return_dim < 0) ? v.first : (v.first % m_stride_mod) / m_stride_div;
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+#ifdef EIGEN_USE_SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+    m_orig_impl.bind(cgh);
+  }
+#endif
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    const double compute_cost = 1.0 +
+        (m_return_dim < 0 ? 0.0 : (TensorOpCost::ModCost<Index>() + TensorOpCost::DivCost<Index>()));
+    return m_orig_impl.costPerCoeff(vectorized) +
+           m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, compute_cost);
+  }
+
+ private:
+  EIGEN_DEVICE_FUNC void gen_strides(const InputDimensions& dims, StrideDims& strides) {
+    if (m_return_dim < 0) {
+      return;  // Won't be using the strides.
+    }
+    eigen_assert(m_return_dim < NumDims &&
+                 "Asking to convert index to a dimension outside of the rank");
+
+    // Calculate m_stride_div and m_stride_mod, which are used to
+    // calculate the value of an index w.r.t. the m_return_dim.
+    if (Layout == static_cast<int>(ColMajor)) {
+      strides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        strides[i] = strides[i-1] * dims[i-1];
+      }
+    } else {
+      strides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        strides[i] = strides[i+1] * dims[i+1];
+      }
+    }
+  }
+
+ protected:
+  TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device> m_orig_impl;
+  TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device> m_impl;
+  const Index m_return_dim;
+  StrideDims m_strides;
+  Index m_stride_mod;
+  Index m_stride_div;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_ARG_MAX_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorAssign.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorAssign.h
new file mode 100644
index 0000000..e5811d6
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorAssign.h
@@ -0,0 +1,247 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H
+#define EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H
+
+namespace Eigen {
+
+/** \class TensorAssign
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief The tensor assignment class.
+  *
+  * This class is represents the assignment of the values resulting from the evaluation of
+  * the rhs expression to the memory locations denoted by the lhs expression.
+  */
+namespace internal {
+template<typename LhsXprType, typename RhsXprType>
+struct traits<TensorAssignOp<LhsXprType, RhsXprType> >
+{
+  typedef typename LhsXprType::Scalar Scalar;
+  typedef typename traits<LhsXprType>::StorageKind StorageKind;
+  typedef typename promote_index_type<typename traits<LhsXprType>::Index,
+                                      typename traits<RhsXprType>::Index>::type Index;
+  typedef typename LhsXprType::Nested LhsNested;
+  typedef typename RhsXprType::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+  static const std::size_t NumDimensions = internal::traits<LhsXprType>::NumDimensions;
+  static const int Layout = internal::traits<LhsXprType>::Layout;
+  typedef typename traits<LhsXprType>::PointerType PointerType;
+
+  enum {
+    Flags = 0
+  };
+};
+
+template<typename LhsXprType, typename RhsXprType>
+struct eval<TensorAssignOp<LhsXprType, RhsXprType>, Eigen::Dense>
+{
+  typedef const TensorAssignOp<LhsXprType, RhsXprType>& type;
+};
+
+template<typename LhsXprType, typename RhsXprType>
+struct nested<TensorAssignOp<LhsXprType, RhsXprType>, 1, typename eval<TensorAssignOp<LhsXprType, RhsXprType> >::type>
+{
+  typedef TensorAssignOp<LhsXprType, RhsXprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename LhsXprType, typename RhsXprType>
+class TensorAssignOp : public TensorBase<TensorAssignOp<LhsXprType, RhsXprType> >
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorAssignOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename LhsXprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorAssignOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorAssignOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorAssignOp>::Index Index;
+
+  static const int NumDims = Eigen::internal::traits<TensorAssignOp>::NumDimensions;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorAssignOp(LhsXprType& lhs, const RhsXprType& rhs)
+      : m_lhs_xpr(lhs), m_rhs_xpr(rhs) {}
+
+    /** \returns the nested expressions */
+    EIGEN_DEVICE_FUNC
+    typename internal::remove_all<typename LhsXprType::Nested>::type&
+    lhsExpression() const { return *((typename internal::remove_all<typename LhsXprType::Nested>::type*)&m_lhs_xpr); }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename RhsXprType::Nested>::type&
+    rhsExpression() const { return m_rhs_xpr; }
+
+  protected:
+    typename internal::remove_all<typename LhsXprType::Nested>::type& m_lhs_xpr;
+    const typename internal::remove_all<typename RhsXprType::Nested>::type& m_rhs_xpr;
+};
+
+
+template<typename LeftArgType, typename RightArgType, typename Device>
+struct TensorEvaluator<const TensorAssignOp<LeftArgType, RightArgType>, Device>
+{
+  typedef TensorAssignOp<LeftArgType, RightArgType> XprType;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef typename TensorEvaluator<RightArgType, Device>::Dimensions Dimensions;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  static const int NumDims = XprType::NumDims;
+
+  enum {
+    IsAligned         = int(TensorEvaluator<LeftArgType, Device>::IsAligned) &
+                        int(TensorEvaluator<RightArgType, Device>::IsAligned),
+    PacketAccess      = int(TensorEvaluator<LeftArgType, Device>::PacketAccess) &
+                        int(TensorEvaluator<RightArgType, Device>::PacketAccess),
+    BlockAccess       = int(TensorEvaluator<LeftArgType, Device>::BlockAccess) &
+                        int(TensorEvaluator<RightArgType, Device>::BlockAccess),
+    PreferBlockAccess = int(TensorEvaluator<LeftArgType, Device>::PreferBlockAccess) |
+                        int(TensorEvaluator<RightArgType, Device>::PreferBlockAccess),
+    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
+    RawAccess         = TensorEvaluator<LeftArgType, Device>::RawAccess
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename TensorEvaluator<const RightArgType, Device>::TensorBlock
+      RightTensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  TensorEvaluator(const XprType& op, const Device& device) :
+      m_leftImpl(op.lhsExpression(), device),
+      m_rightImpl(op.rhsExpression(), device)
+  {
+    EIGEN_STATIC_ASSERT(
+        (static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) ==
+         static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout)),
+        YOU_MADE_A_PROGRAMMING_MISTAKE);
+  }
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
+  {
+    // The dimensions of the lhs and the rhs tensors should be equal to prevent
+    // overflows and ensure the result is fully initialized.
+    // TODO: use left impl instead if right impl dimensions are known at compile time.
+    return m_rightImpl.dimensions();
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    eigen_assert(dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions()));
+    m_leftImpl.evalSubExprsIfNeeded(NULL);
+    // If the lhs provides raw access to its storage area (i.e. if m_leftImpl.data() returns a non
+    // null value), attempt to evaluate the rhs expression in place. Returns true iff in place
+    // evaluation isn't supported and the caller still needs to manually assign the values generated
+    // by the rhs to the lhs.
+    return m_rightImpl.evalSubExprsIfNeeded(m_leftImpl.data());
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    m_leftImpl.evalSubExprsIfNeededAsync(nullptr, [this, done](bool) {
+      m_rightImpl.evalSubExprsIfNeededAsync(
+          m_leftImpl.data(), [done](bool need_assign) { done(need_assign); });
+    });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_leftImpl.cleanup();
+    m_rightImpl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalScalar(Index i) {
+    m_leftImpl.coeffRef(i) = m_rightImpl.coeff(i);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalPacket(Index i) {
+
+    const int LhsStoreMode = TensorEvaluator<LeftArgType, Device>::IsAligned ? Aligned : Unaligned;
+    const int RhsLoadMode = TensorEvaluator<RightArgType, Device>::IsAligned ? Aligned : Unaligned;
+    m_leftImpl.template writePacket<LhsStoreMode>(i, m_rightImpl.template packet<RhsLoadMode>(i));
+  }
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const
+  {
+    return m_leftImpl.coeff(index);
+  }
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const
+  {
+    return m_leftImpl.template packet<LoadMode>(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    // We assume that evalPacket or evalScalar is called to perform the
+    // assignment and account for the cost of the write here, but reduce left
+    // cost by one load because we are using m_leftImpl.coeffRef.
+    TensorOpCost left = m_leftImpl.costPerCoeff(vectorized);
+    return m_rightImpl.costPerCoeff(vectorized) +
+           TensorOpCost(
+               numext::maxi(0.0, left.bytes_loaded() - sizeof(CoeffReturnType)),
+               left.bytes_stored(), left.compute_cycles()) +
+           TensorOpCost(0, sizeof(CoeffReturnType), 0, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    return internal::TensorBlockResourceRequirements::merge(
+        m_leftImpl.getResourceRequirements(),
+        m_rightImpl.getResourceRequirements());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalBlock(
+      TensorBlockDesc& desc, TensorBlockScratch& scratch) {
+    if (TensorEvaluator<LeftArgType, Device>::RawAccess &&
+        m_leftImpl.data() != NULL) {
+      // If destination has raw data access, we pass it as a potential
+      // destination for a block descriptor evaluation.
+      desc.template AddDestinationBuffer<Layout>(
+          /*dst_base=*/m_leftImpl.data() + desc.offset(),
+          /*dst_strides=*/internal::strides<Layout>(m_leftImpl.dimensions()));
+    }
+
+    RightTensorBlock block = m_rightImpl.block(desc, scratch, /*root_of_expr_ast=*/true);
+    // If block was evaluated into a destination, there is no need to do assignment.
+    if (block.kind() != internal::TensorBlockKind::kMaterializedInOutput) {
+      m_leftImpl.writeBlock(desc, block);
+    }
+    block.cleanup();
+  }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_leftImpl.bind(cgh);
+    m_rightImpl.bind(cgh);
+  }
+#endif
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_leftImpl.data(); }
+
+ private:
+  TensorEvaluator<LeftArgType, Device> m_leftImpl;
+  TensorEvaluator<RightArgType, Device> m_rightImpl;
+};
+
+}
+
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorBase.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorBase.h
new file mode 100644
index 0000000..35b6458
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorBase.h
@@ -0,0 +1,1176 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_BASE_H
+#define EIGEN_CXX11_TENSOR_TENSOR_BASE_H
+
+// clang-format off
+
+namespace Eigen {
+
+/** \class TensorBase
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief The tensor base class.
+  *
+  * This class is the common parent of the Tensor and TensorMap class, thus
+  * making it possible to use either class interchangeably in expressions.
+  */
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+// FIXME Doxygen does not like the inheritance with different template parameters
+// Since there is no doxygen documentation inside, we disable it for now
+template<typename Derived>
+class TensorBase<Derived, ReadOnlyAccessors>
+{
+  public:
+    typedef internal::traits<Derived> DerivedTraits;
+    typedef typename DerivedTraits::Scalar Scalar;
+    typedef typename DerivedTraits::Index Index;
+    typedef typename internal::remove_const<Scalar>::type CoeffReturnType;
+    static const int NumDimensions = DerivedTraits::NumDimensions;
+
+    // Generic nullary operation support.
+    template <typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseNullaryOp<CustomNullaryOp, const Derived>
+    nullaryExpr(const CustomNullaryOp& func) const {
+      return TensorCwiseNullaryOp<CustomNullaryOp, const Derived>(derived(), func);
+    }
+
+    // Coefficient-wise nullary operators
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived>
+    constant(const Scalar& value) const {
+      return nullaryExpr(internal::scalar_constant_op<Scalar>(value));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseNullaryOp<internal::UniformRandomGenerator<Scalar>, const Derived>
+    random() const {
+      return nullaryExpr(internal::UniformRandomGenerator<Scalar>());
+    }
+    template <typename RandomGenerator> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseNullaryOp<RandomGenerator, const Derived>
+    random(const RandomGenerator& gen = RandomGenerator()) const {
+      return nullaryExpr(gen);
+    }
+
+    // Tensor generation
+    template <typename Generator> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorGeneratorOp<Generator, const Derived>
+    generate(const Generator& generator) const {
+      return TensorGeneratorOp<Generator, const Derived>(derived(), generator);
+    }
+
+    // Generic unary operation support.
+    template <typename CustomUnaryOp> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<CustomUnaryOp, const Derived>
+    unaryExpr(const CustomUnaryOp& func) const {
+      return TensorCwiseUnaryOp<CustomUnaryOp, const Derived>(derived(), func);
+    }
+
+    // Coefficient-wise unary operators
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_opposite_op<Scalar>, const Derived>
+    operator-() const {
+      return unaryExpr(internal::scalar_opposite_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_sqrt_op<Scalar>, const Derived>
+    sqrt() const {
+      return unaryExpr(internal::scalar_sqrt_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_sign_op<Scalar>, const Derived>
+    sign() const {
+      return unaryExpr(internal::scalar_sign_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_rsqrt_op<Scalar>, const Derived>
+    rsqrt() const {
+      return unaryExpr(internal::scalar_rsqrt_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_square_op<Scalar>, const Derived>
+    square() const {
+      return unaryExpr(internal::scalar_square_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_cube_op<Scalar>, const Derived>
+    cube() const {
+      return unaryExpr(internal::scalar_cube_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const Derived>
+    inverse() const {
+      return unaryExpr(internal::scalar_inverse_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_tanh_op<Scalar>, const Derived>
+    tanh() const {
+      return unaryExpr(internal::scalar_tanh_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_lgamma_op<Scalar>, const Derived>
+    lgamma() const {
+      return unaryExpr(internal::scalar_lgamma_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_digamma_op<Scalar>, const Derived>
+    digamma() const {
+      return unaryExpr(internal::scalar_digamma_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_i0_op<Scalar>, const Derived>
+    bessel_i0() const {
+      return unaryExpr(internal::scalar_bessel_i0_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_i0e_op<Scalar>, const Derived>
+    bessel_i0e() const {
+      return unaryExpr(internal::scalar_bessel_i0e_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_i1_op<Scalar>, const Derived>
+    bessel_i1() const {
+      return unaryExpr(internal::scalar_bessel_i1_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_i1e_op<Scalar>, const Derived>
+    bessel_i1e() const {
+      return unaryExpr(internal::scalar_bessel_i1e_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_j0_op<Scalar>, const Derived>
+    bessel_j0() const {
+      return unaryExpr(internal::scalar_bessel_j0_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_y0_op<Scalar>, const Derived>
+    bessel_y0() const {
+      return unaryExpr(internal::scalar_bessel_y0_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_j1_op<Scalar>, const Derived>
+    bessel_j1() const {
+      return unaryExpr(internal::scalar_bessel_j1_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_y1_op<Scalar>, const Derived>
+    bessel_y1() const {
+      return unaryExpr(internal::scalar_bessel_y1_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_k0_op<Scalar>, const Derived>
+    bessel_k0() const {
+      return unaryExpr(internal::scalar_bessel_k0_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_k0e_op<Scalar>, const Derived>
+    bessel_k0e() const {
+      return unaryExpr(internal::scalar_bessel_k0e_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_k1_op<Scalar>, const Derived>
+    bessel_k1() const {
+      return unaryExpr(internal::scalar_bessel_k1_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_bessel_k1e_op<Scalar>, const Derived>
+    bessel_k1e() const {
+      return unaryExpr(internal::scalar_bessel_k1e_op<Scalar>());
+    }
+
+    // igamma(a = this, x = other)
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_igamma_op<Scalar>, const Derived, const OtherDerived>
+    igamma(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_igamma_op<Scalar>());
+    }
+
+    // igamma_der_a(a = this, x = other)
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_igamma_der_a_op<Scalar>, const Derived, const OtherDerived>
+    igamma_der_a(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_igamma_der_a_op<Scalar>());
+    }
+
+    // gamma_sample_der_alpha(alpha = this, sample = other)
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_gamma_sample_der_alpha_op<Scalar>, const Derived, const OtherDerived>
+    gamma_sample_der_alpha(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_gamma_sample_der_alpha_op<Scalar>());
+    }
+
+    // igammac(a = this, x = other)
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_igammac_op<Scalar>, const Derived, const OtherDerived>
+    igammac(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_igammac_op<Scalar>());
+    }
+
+    // zeta(x = this, q = other)
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_zeta_op<Scalar>, const Derived, const OtherDerived>
+    zeta(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_zeta_op<Scalar>());
+    }
+
+    // polygamma(n = this, x = other)
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_polygamma_op<Scalar>, const Derived, const OtherDerived>
+    polygamma(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_polygamma_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_erf_op<Scalar>, const Derived>
+    erf() const {
+      return unaryExpr(internal::scalar_erf_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_erfc_op<Scalar>, const Derived>
+    erfc() const {
+      return unaryExpr(internal::scalar_erfc_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_ndtri_op<Scalar>, const Derived>
+    ndtri() const {
+      return unaryExpr(internal::scalar_ndtri_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_logistic_op<Scalar>, const Derived>
+    sigmoid() const {
+      return unaryExpr(internal::scalar_logistic_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_exp_op<Scalar>, const Derived>
+    exp() const {
+      return unaryExpr(internal::scalar_exp_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_expm1_op<Scalar>, const Derived>
+    expm1() const {
+      return unaryExpr(internal::scalar_expm1_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_log_op<Scalar>, const Derived>
+    log() const {
+      return unaryExpr(internal::scalar_log_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_log1p_op<Scalar>, const Derived>
+    log1p() const {
+      return unaryExpr(internal::scalar_log1p_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_log2_op<Scalar>, const Derived>
+    log2() const {
+      return unaryExpr(internal::scalar_log2_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_abs_op<Scalar>, const Derived>
+    abs() const {
+      return unaryExpr(internal::scalar_abs_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_clamp_op<Scalar>, const Derived>
+    clip(Scalar min, Scalar max) const {
+      return unaryExpr(internal::scalar_clamp_op<Scalar>(min, max));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const typename internal::conditional<NumTraits<CoeffReturnType>::IsComplex,
+                                                             TensorCwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, const Derived>,
+                                                             Derived>::type
+    conjugate() const {
+      return choose(Cond<NumTraits<CoeffReturnType>::IsComplex>(), unaryExpr(internal::scalar_conjugate_op<Scalar>()), derived());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::bind2nd_op<internal::scalar_pow_op<Scalar,Scalar> >, const Derived>
+    pow(Scalar exponent) const {
+      return unaryExpr(internal::bind2nd_op<internal::scalar_pow_op<Scalar,Scalar> >(exponent));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_real_op<Scalar>, const Derived>
+    real() const {
+      return unaryExpr(internal::scalar_real_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_imag_op<Scalar>, const Derived>
+    imag() const {
+      return unaryExpr(internal::scalar_imag_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::bind2nd_op<internal::scalar_sum_op<Scalar,Scalar> >, const Derived>
+    operator+ (Scalar rhs) const {
+      return unaryExpr(internal::bind2nd_op<internal::scalar_sum_op<Scalar,Scalar> >(rhs));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE friend
+    const TensorCwiseUnaryOp<internal::bind1st_op<internal::scalar_sum_op<Scalar> >, const Derived>
+    operator+ (Scalar lhs, const Derived& rhs) {
+      return rhs.unaryExpr(internal::bind1st_op<internal::scalar_sum_op<Scalar> >(lhs));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::bind2nd_op<internal::scalar_difference_op<Scalar,Scalar> >, const Derived>
+    operator- (Scalar rhs) const {
+      EIGEN_STATIC_ASSERT((NumTraits<Scalar>::IsSigned || internal::is_same<Scalar, const std::complex<float> >::value), YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return unaryExpr(internal::bind2nd_op<internal::scalar_difference_op<Scalar,Scalar> >(rhs));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE friend
+    const TensorCwiseUnaryOp<internal::bind1st_op<internal::scalar_difference_op<Scalar> >, const Derived>
+    operator- (Scalar lhs, const Derived& rhs) {
+      return rhs.unaryExpr(internal::bind1st_op<internal::scalar_difference_op<Scalar> >(lhs));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::bind2nd_op<internal::scalar_product_op<Scalar,Scalar> >, const Derived>
+    operator* (Scalar rhs) const {
+      return unaryExpr(internal::bind2nd_op<internal::scalar_product_op<Scalar,Scalar> >(rhs));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE friend
+    const TensorCwiseUnaryOp<internal::bind1st_op<internal::scalar_product_op<Scalar> >, const Derived>
+    operator* (Scalar lhs, const Derived& rhs) {
+      return rhs.unaryExpr(internal::bind1st_op<internal::scalar_product_op<Scalar> >(lhs));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::bind2nd_op<internal::scalar_quotient_op<Scalar,Scalar> >, const Derived>
+    operator/ (Scalar rhs) const {
+      return unaryExpr(internal::bind2nd_op<internal::scalar_quotient_op<Scalar,Scalar> >(rhs));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE friend
+    const TensorCwiseUnaryOp<internal::bind1st_op<internal::scalar_quotient_op<Scalar> >, const Derived>
+    operator/ (Scalar lhs, const Derived& rhs) {
+      return rhs.unaryExpr(internal::bind1st_op<internal::scalar_quotient_op<Scalar> >(lhs));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_mod_op<Scalar>, const Derived>
+    operator% (Scalar rhs) const {
+      EIGEN_STATIC_ASSERT(NumTraits<Scalar>::IsInteger, YOU_MADE_A_PROGRAMMING_MISTAKE_TRY_MOD);
+      return unaryExpr(internal::scalar_mod_op<Scalar>(rhs));
+    }
+
+    template <int NanPropagation=PropagateFast>
+    EIGEN_DEVICE_FUNC
+        EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_max_op<Scalar,Scalar,NanPropagation>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+    cwiseMax(Scalar threshold) const {
+      return cwiseMax<NanPropagation>(constant(threshold));
+    }
+
+    template <int NanPropagation=PropagateFast>
+    EIGEN_DEVICE_FUNC
+        EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_min_op<Scalar,Scalar,NanPropagation>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+    cwiseMin(Scalar threshold) const {
+      return cwiseMin<NanPropagation>(constant(threshold));
+    }
+
+    template<typename NewType>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const typename internal::conditional<internal::is_same<NewType, CoeffReturnType>::value,
+                                                             Derived,
+                                                             TensorConversionOp<NewType, const Derived> >::type
+    cast() const {
+      return choose(Cond<internal::is_same<NewType, CoeffReturnType>::value>(), derived(), TensorConversionOp<NewType, const Derived>(derived()));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_round_op<Scalar>, const Derived>
+    round() const {
+      return unaryExpr(internal::scalar_round_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_rint_op<Scalar>, const Derived>
+    rint() const {
+      return unaryExpr(internal::scalar_rint_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_ceil_op<Scalar>, const Derived>
+    ceil() const {
+      return unaryExpr(internal::scalar_ceil_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_floor_op<Scalar>, const Derived>
+    floor() const {
+      return unaryExpr(internal::scalar_floor_op<Scalar>());
+    }
+
+    // Generic binary operation support.
+    template <typename CustomBinaryOp, typename OtherDerived> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<CustomBinaryOp, const Derived, const OtherDerived>
+    binaryExpr(const OtherDerived& other, const CustomBinaryOp& func) const {
+      return TensorCwiseBinaryOp<CustomBinaryOp, const Derived, const OtherDerived>(derived(), other, func);
+    }
+
+    // Coefficient-wise binary operators.
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_sum_op<Scalar>, const Derived, const OtherDerived>
+    operator+(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_sum_op<Scalar>());
+    }
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_difference_op<Scalar>, const Derived, const OtherDerived>
+    operator-(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_difference_op<Scalar>());
+    }
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_product_op<Scalar>, const Derived, const OtherDerived>
+    operator*(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_product_op<Scalar>());
+    }
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_quotient_op<Scalar>, const Derived, const OtherDerived>
+    operator/(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_quotient_op<Scalar>());
+    }
+
+  template<int NaNPropagation=PropagateFast, typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+      const TensorCwiseBinaryOp<internal::scalar_max_op<Scalar,Scalar, NaNPropagation>, const Derived, const OtherDerived>
+    cwiseMax(const OtherDerived& other) const {
+    return binaryExpr(other.derived(), internal::scalar_max_op<Scalar,Scalar, NaNPropagation>());
+    }
+
+  template<int NaNPropagation=PropagateFast, typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_min_op<Scalar,Scalar, NaNPropagation>, const Derived, const OtherDerived>
+    cwiseMin(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_min_op<Scalar,Scalar, NaNPropagation>());
+    }
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_boolean_and_op, const Derived, const OtherDerived>
+    operator&&(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_boolean_and_op());
+    }
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_boolean_or_op, const Derived, const OtherDerived>
+    operator||(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_boolean_or_op());
+    }
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_boolean_xor_op, const Derived, const OtherDerived>
+    operator^(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_boolean_xor_op());
+    }
+
+    // Comparisons and tests.
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_LT>, const Derived, const OtherDerived>
+    operator<(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_LT>());
+    }
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_LE>, const Derived, const OtherDerived>
+    operator<=(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_LE>());
+    }
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_GT>, const Derived, const OtherDerived>
+    operator>(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_GT>());
+    }
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_GE>, const Derived, const OtherDerived>
+    operator>=(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_GE>());
+    }
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_EQ>, const Derived, const OtherDerived>
+    operator==(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_EQ>());
+    }
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_NEQ>, const Derived, const OtherDerived>
+    operator!=(const OtherDerived& other) const {
+      return binaryExpr(other.derived(), internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_NEQ>());
+    }
+
+    // comparisons and tests for Scalars
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_LT>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+    operator<(Scalar threshold) const {
+      return operator<(constant(threshold));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_LE>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+    operator<=(Scalar threshold) const {
+      return operator<=(constant(threshold));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_GT>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+    operator>(Scalar threshold) const {
+      return operator>(constant(threshold));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_GE>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+    operator>=(Scalar threshold) const {
+      return operator>=(constant(threshold));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_EQ>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+    operator==(Scalar threshold) const {
+      return operator==(constant(threshold));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_cmp_op<Scalar, Scalar, internal::cmp_NEQ>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+    operator!=(Scalar threshold) const {
+      return operator!=(constant(threshold));
+    }
+
+    // Checks
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_isnan_op<Scalar>, const Derived>
+    (isnan)() const {
+      return unaryExpr(internal::scalar_isnan_op<Scalar>());
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_isinf_op<Scalar>, const Derived>
+    (isinf)() const {
+      return unaryExpr(internal::scalar_isinf_op<Scalar>());
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_isfinite_op<Scalar>, const Derived>
+    (isfinite)() const {
+      return unaryExpr(internal::scalar_isfinite_op<Scalar>());
+    }
+
+    // Coefficient-wise ternary operators.
+    template<typename ThenDerived, typename ElseDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorSelectOp<const Derived, const ThenDerived, const ElseDerived>
+    select(const ThenDerived& thenTensor, const ElseDerived& elseTensor) const {
+      return TensorSelectOp<const Derived, const ThenDerived, const ElseDerived>(derived(), thenTensor.derived(), elseTensor.derived());
+    }
+
+    // Contractions.
+    typedef Eigen::IndexPair<Index> DimensionPair;
+
+    template<typename OtherDerived, typename Dimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorContractionOp<const Dimensions, const Derived, const OtherDerived, const NoOpOutputKernel>
+    contract(const OtherDerived& other, const Dimensions& dims) const {
+      return TensorContractionOp<const Dimensions, const Derived, const OtherDerived, const NoOpOutputKernel>(derived(), other.derived(), dims);
+    }
+
+    template<typename OtherDerived, typename Dimensions, typename OutputKernel> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorContractionOp<const Dimensions, const Derived, const OtherDerived, const OutputKernel>
+    contract(const OtherDerived& other, const Dimensions& dims, const OutputKernel& output_kernel) const {
+      return TensorContractionOp<const Dimensions, const Derived, const OtherDerived, const OutputKernel>(derived(), other.derived(), dims, output_kernel);
+    }
+
+    // Convolutions.
+    template<typename KernelDerived, typename Dimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorConvolutionOp<const Dimensions, const Derived, const KernelDerived>
+    convolve(const KernelDerived& kernel, const Dimensions& dims) const {
+      return TensorConvolutionOp<const Dimensions, const Derived, const KernelDerived>(derived(), kernel.derived(), dims);
+    }
+
+    // Fourier transforms
+    template <int FFTDataType, int FFTDirection, typename FFT> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorFFTOp<const FFT, const Derived, FFTDataType, FFTDirection>
+    fft(const FFT& dims) const {
+      return TensorFFTOp<const FFT, const Derived, FFTDataType, FFTDirection>(derived(), dims);
+    }
+
+    // Scan.
+    typedef TensorScanOp<internal::SumReducer<CoeffReturnType>, const Derived> TensorScanSumOp;
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorScanSumOp
+    cumsum(const Index& axis, bool exclusive = false) const {
+      return TensorScanSumOp(derived(), axis, exclusive);
+    }
+
+    typedef TensorScanOp<internal::ProdReducer<CoeffReturnType>, const Derived> TensorScanProdOp;
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorScanProdOp
+    cumprod(const Index& axis, bool exclusive = false) const {
+      return TensorScanProdOp(derived(), axis, exclusive);
+    }
+
+    template <typename Reducer>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorScanOp<Reducer, const Derived>
+    scan(const Index& axis, const Reducer& reducer, bool exclusive = false) const {
+      return TensorScanOp<Reducer, const Derived>(derived(), axis, exclusive, reducer);
+    }
+
+    // Reductions.
+    template <typename Dims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::SumReducer<CoeffReturnType>, const Dims, const Derived>
+    sum(const Dims& dims) const {
+      return TensorReductionOp<internal::SumReducer<CoeffReturnType>, const Dims, const Derived>(derived(), dims, internal::SumReducer<CoeffReturnType>());
+    }
+
+    const TensorReductionOp<internal::SumReducer<CoeffReturnType>, const DimensionList<Index, NumDimensions>, const Derived>
+    sum() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return TensorReductionOp<internal::SumReducer<CoeffReturnType>, const DimensionList<Index, NumDimensions>, const Derived>(derived(), in_dims, internal::SumReducer<CoeffReturnType>());
+    }
+
+    template <typename Dims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::MeanReducer<CoeffReturnType>, const Dims, const Derived>
+    mean(const Dims& dims) const {
+      return TensorReductionOp<internal::MeanReducer<CoeffReturnType>, const Dims, const Derived>(derived(), dims, internal::MeanReducer<CoeffReturnType>());
+    }
+
+    const TensorReductionOp<internal::MeanReducer<CoeffReturnType>, const DimensionList<Index, NumDimensions>, const Derived>
+    mean() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return TensorReductionOp<internal::MeanReducer<CoeffReturnType>, const DimensionList<Index, NumDimensions>, const Derived>(derived(), in_dims, internal::MeanReducer<CoeffReturnType>());
+    }
+
+    template <typename Dims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::ProdReducer<CoeffReturnType>, const Dims, const Derived>
+    prod(const Dims& dims) const {
+      return TensorReductionOp<internal::ProdReducer<CoeffReturnType>, const Dims, const Derived>(derived(), dims, internal::ProdReducer<CoeffReturnType>());
+    }
+
+    const TensorReductionOp<internal::ProdReducer<CoeffReturnType>, const DimensionList<Index, NumDimensions>, const Derived>
+    prod() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return TensorReductionOp<internal::ProdReducer<CoeffReturnType>, const DimensionList<Index, NumDimensions>, const Derived>(derived(), in_dims, internal::ProdReducer<CoeffReturnType>());
+    }
+
+    template <typename Dims,int NanPropagation=PropagateFast> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::MaxReducer<CoeffReturnType,NanPropagation>, const Dims, const Derived>
+    maximum(const Dims& dims) const {
+      return TensorReductionOp<internal::MaxReducer<CoeffReturnType,NanPropagation>, const Dims, const Derived>(derived(), dims, internal::MaxReducer<CoeffReturnType,NanPropagation>());
+    }
+
+    template <int NanPropagation=PropagateFast>
+    const TensorReductionOp<internal::MaxReducer<CoeffReturnType,NanPropagation>, const DimensionList<Index, NumDimensions>, const Derived>
+    maximum() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return TensorReductionOp<internal::MaxReducer<CoeffReturnType,NanPropagation>, const DimensionList<Index, NumDimensions>, const Derived>(derived(), in_dims, internal::MaxReducer<CoeffReturnType,NanPropagation>());
+    }
+
+    template <typename Dims,int NanPropagation=PropagateFast> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::MinReducer<CoeffReturnType,NanPropagation>, const Dims, const Derived>
+    minimum(const Dims& dims) const {
+      return TensorReductionOp<internal::MinReducer<CoeffReturnType,NanPropagation>, const Dims, const Derived>(derived(), dims, internal::MinReducer<CoeffReturnType,NanPropagation>());
+    }
+
+    template <int NanPropagation=PropagateFast>
+    const TensorReductionOp<internal::MinReducer<CoeffReturnType,NanPropagation>, const DimensionList<Index, NumDimensions>, const Derived>
+    minimum() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return TensorReductionOp<internal::MinReducer<CoeffReturnType,NanPropagation>, const DimensionList<Index, NumDimensions>, const Derived>(derived(), in_dims, internal::MinReducer<CoeffReturnType,NanPropagation>());
+    }
+
+    template <typename Dims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::AndReducer, const Dims, const typename internal::conditional<internal::is_same<bool, CoeffReturnType>::value, Derived, TensorConversionOp<bool, const Derived> >::type >
+    all(const Dims& dims) const {
+      return cast<bool>().reduce(dims, internal::AndReducer());
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::AndReducer, const DimensionList<Index, NumDimensions>, const typename internal::conditional<internal::is_same<bool, CoeffReturnType>::value, Derived, TensorConversionOp<bool, const Derived> >::type >
+    all() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return cast<bool>().reduce(in_dims, internal::AndReducer());
+    }
+
+    template <typename Dims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::OrReducer, const Dims, const typename internal::conditional<internal::is_same<bool, CoeffReturnType>::value, Derived, TensorConversionOp<bool, const Derived> >::type >
+    any(const Dims& dims) const {
+      return cast<bool>().reduce(dims, internal::OrReducer());
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<internal::OrReducer, const DimensionList<Index, NumDimensions>, const typename internal::conditional<internal::is_same<bool, CoeffReturnType>::value, Derived, TensorConversionOp<bool, const Derived> >::type >
+    any() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return cast<bool>().reduce(in_dims, internal::OrReducer());
+    }
+
+   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorTupleReducerOp<
+      internal::ArgMaxTupleReducer<Tuple<Index, CoeffReturnType> >,
+      const array<Index, NumDimensions>, const Derived>
+    argmax() const {
+      array<Index, NumDimensions> in_dims;
+      for (Index d = 0; d < NumDimensions; ++d) in_dims[d] = d;
+      return TensorTupleReducerOp<
+        internal::ArgMaxTupleReducer<Tuple<Index, CoeffReturnType> >,
+        const array<Index, NumDimensions>,
+        const Derived>(derived(), internal::ArgMaxTupleReducer<Tuple<Index, CoeffReturnType> >(), -1, in_dims);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorTupleReducerOp<
+      internal::ArgMinTupleReducer<Tuple<Index, CoeffReturnType> >,
+      const array<Index, NumDimensions>, const Derived>
+    argmin() const {
+      array<Index, NumDimensions> in_dims;
+      for (Index d = 0; d < NumDimensions; ++d) in_dims[d] = d;
+      return TensorTupleReducerOp<
+        internal::ArgMinTupleReducer<Tuple<Index, CoeffReturnType> >,
+        const array<Index, NumDimensions>,
+        const Derived>(derived(), internal::ArgMinTupleReducer<Tuple<Index, CoeffReturnType> >(), -1, in_dims);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorTupleReducerOp<
+      internal::ArgMaxTupleReducer<Tuple<Index, CoeffReturnType> >,
+      const array<Index, 1>, const Derived>
+    argmax(const Index return_dim) const {
+      array<Index, 1> in_dims;
+      in_dims[0] = return_dim;
+      return TensorTupleReducerOp<
+        internal::ArgMaxTupleReducer<Tuple<Index, CoeffReturnType> >,
+        const array<Index, 1>,
+        const Derived>(derived(), internal::ArgMaxTupleReducer<Tuple<Index, CoeffReturnType> >(), return_dim, in_dims);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorTupleReducerOp<
+      internal::ArgMinTupleReducer<Tuple<Index, CoeffReturnType> >,
+      const array<Index, 1>, const Derived>
+    argmin(const Index return_dim) const {
+      array<Index, 1> in_dims;
+      in_dims[0] = return_dim;
+      return TensorTupleReducerOp<
+        internal::ArgMinTupleReducer<Tuple<Index, CoeffReturnType> >,
+        const array<Index, 1>,
+        const Derived>(derived(), internal::ArgMinTupleReducer<Tuple<Index, CoeffReturnType> >(), return_dim, in_dims);
+    }
+
+    template <typename Reducer, typename Dims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReductionOp<Reducer, const Dims, const Derived>
+    reduce(const Dims& dims, const Reducer& reducer) const {
+      return TensorReductionOp<Reducer, const Dims, const Derived>(derived(), dims, reducer);
+    }
+
+    template <typename Dims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorTraceOp<const Dims, const Derived>
+    trace(const Dims& dims) const {
+      return TensorTraceOp<const Dims, const Derived>(derived(), dims);
+    }
+
+    const TensorTraceOp<const DimensionList<Index, NumDimensions>, const Derived>
+    trace() const {
+      DimensionList<Index, NumDimensions> in_dims;
+      return TensorTraceOp<const DimensionList<Index, NumDimensions>, const Derived>(derived(), in_dims);
+    }
+
+    template <typename Broadcast> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorBroadcastingOp<const Broadcast, const Derived>
+    broadcast(const Broadcast& bcast) const {
+      return TensorBroadcastingOp<const Broadcast, const Derived>(derived(), bcast);
+    }
+
+    template <typename Axis, typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorConcatenationOp<Axis, const Derived, const OtherDerived>
+    concatenate(const OtherDerived& other, Axis axis) const {
+      return TensorConcatenationOp<Axis, const Derived, const OtherDerived>(derived(), other.derived(), axis);
+    }
+
+    template <typename PatchDims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorPatchOp<const PatchDims, const Derived>
+    extract_patches(const PatchDims& patch_dims) const {
+      return TensorPatchOp<const PatchDims, const Derived>(derived(), patch_dims);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorImagePatchOp<Dynamic, Dynamic, const Derived>
+    extract_image_patches(const Index patch_rows = 1, const Index patch_cols = 1,
+                          const Index row_stride = 1, const Index col_stride = 1,
+                          const Index in_row_stride = 1, const Index in_col_stride = 1,
+                          const PaddingType padding_type = PADDING_SAME, const Scalar padding_value = Scalar(0)) const {
+      return TensorImagePatchOp<Dynamic, Dynamic, const Derived>(derived(), patch_rows, patch_cols, row_stride, col_stride,
+                                                                 in_row_stride, in_col_stride, 1, 1, padding_type, padding_value);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorImagePatchOp<Dynamic, Dynamic, const Derived>
+    extract_image_patches(const Index patch_rows, const Index patch_cols,
+                          const Index row_stride, const Index col_stride,
+                          const Index in_row_stride, const Index in_col_stride,
+                          const Index row_inflate_stride, const Index col_inflate_stride,
+                          const Index padding_top, const Index padding_bottom,
+                          const Index padding_left,const Index padding_right,
+                          const Scalar padding_value) const {
+      return TensorImagePatchOp<Dynamic, Dynamic, const Derived>(derived(), patch_rows, patch_cols, row_stride, col_stride,
+                                                                 in_row_stride, in_col_stride, row_inflate_stride, col_inflate_stride,
+                                                                 padding_top, padding_bottom, padding_left, padding_right, padding_value);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorVolumePatchOp<Dynamic, Dynamic, Dynamic, const Derived>
+    extract_volume_patches(const Index patch_planes, const Index patch_rows, const Index patch_cols,
+                           const Index plane_stride = 1, const Index row_stride = 1, const Index col_stride = 1,
+                           const PaddingType padding_type = PADDING_SAME, const Scalar padding_value = Scalar(0)) const {
+      return TensorVolumePatchOp<Dynamic, Dynamic, Dynamic, const Derived>(derived(), patch_planes, patch_rows, patch_cols, plane_stride, row_stride, col_stride, 1, 1, 1, 1, 1, 1, padding_type, padding_value);
+    }
+
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorVolumePatchOp<Dynamic, Dynamic, Dynamic, const Derived>
+    extract_volume_patches(const Index patch_planes, const Index patch_rows, const Index patch_cols,
+                           const Index plane_stride, const Index row_stride, const Index col_stride,
+                           const Index plane_inflate_stride, const Index row_inflate_stride, const Index col_inflate_stride,
+                           const Index padding_top_z, const Index padding_bottom_z,
+                           const Index padding_top, const Index padding_bottom,
+                           const Index padding_left, const Index padding_right, const Scalar padding_value = Scalar(0)) const {
+      return TensorVolumePatchOp<Dynamic, Dynamic, Dynamic, const Derived>(derived(), patch_planes, patch_rows, patch_cols, plane_stride, row_stride, col_stride, 1, 1, 1, plane_inflate_stride, row_inflate_stride, col_inflate_stride, padding_top_z, padding_bottom_z, padding_top, padding_bottom, padding_left, padding_right, padding_value);
+    }
+
+    // Morphing operators.
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorLayoutSwapOp<const Derived>
+    swap_layout() const {
+      return TensorLayoutSwapOp<const Derived>(derived());
+    }
+    template <typename NewDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReshapingOp<const NewDimensions, const Derived>
+    reshape(const NewDimensions& newDimensions) const {
+      return TensorReshapingOp<const NewDimensions, const Derived>(derived(), newDimensions);
+    }
+    template <typename StartIndices, typename Sizes> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorSlicingOp<const StartIndices, const Sizes, const Derived>
+    slice(const StartIndices& startIndices, const Sizes& sizes) const {
+      return TensorSlicingOp<const StartIndices, const Sizes, const Derived>(derived(), startIndices, sizes);
+    }
+    template <typename StartIndices, typename StopIndices, typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorStridingSlicingOp<const StartIndices, const StopIndices, const Strides, const Derived>
+    stridedSlice(const StartIndices& startIndices, const StopIndices& stopIndices, const Strides& strides) const {
+      return TensorStridingSlicingOp<const StartIndices, const StopIndices, const Strides,
+                                const Derived>(derived(), startIndices, stopIndices, strides);
+    }
+    template <Index DimId> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorChippingOp<DimId, const Derived>
+    chip(const Index offset) const {
+      return TensorChippingOp<DimId, const Derived>(derived(), offset, DimId);
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorChippingOp<Dynamic, const Derived>
+    chip(const Index offset, const Index dim) const {
+      return TensorChippingOp<Dynamic, const Derived>(derived(), offset, dim);
+    }
+    template <typename ReverseDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReverseOp<const ReverseDimensions, const Derived>
+    reverse(const ReverseDimensions& rev) const {
+      return TensorReverseOp<const ReverseDimensions, const Derived>(derived(), rev);
+    }
+    template <typename PaddingDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorPaddingOp<const PaddingDimensions, const Derived>
+    pad(const PaddingDimensions& padding) const {
+      return TensorPaddingOp<const PaddingDimensions, const Derived>(derived(), padding, internal::scalar_cast_op<int, Scalar>()(0));
+    }
+    template <typename PaddingDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorPaddingOp<const PaddingDimensions, const Derived>
+    pad(const PaddingDimensions& padding, const Scalar padding_value) const {
+      return TensorPaddingOp<const PaddingDimensions, const Derived>(derived(), padding, padding_value);
+    }
+    template <typename Shuffle> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorShufflingOp<const Shuffle, const Derived>
+    shuffle(const Shuffle& shfl) const {
+      return TensorShufflingOp<const Shuffle, const Derived>(derived(), shfl);
+    }
+    template <typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorStridingOp<const Strides, const Derived>
+    stride(const Strides& strides) const {
+      return TensorStridingOp<const Strides, const Derived>(derived(), strides);
+    }
+    template <typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorInflationOp<const Strides, const Derived>
+    inflate(const Strides& strides) const {
+      return TensorInflationOp<const Strides, const Derived>(derived(), strides);
+    }
+
+    // Returns a tensor containing index/value tuples
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorIndexTupleOp<const Derived>
+    index_tuples() const {
+      return TensorIndexTupleOp<const Derived>(derived());
+    }
+
+    // Support for custom unary and binary operations
+    template <typename CustomUnaryFunc>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCustomUnaryOp<const CustomUnaryFunc, const Derived> customOp(const CustomUnaryFunc& op) const {
+      return TensorCustomUnaryOp<const CustomUnaryFunc, const Derived>(derived(), op);
+    }
+    template <typename OtherDerived, typename CustomBinaryFunc>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorCustomBinaryOp<const CustomBinaryFunc, const Derived, const OtherDerived> customOp(const OtherDerived& other, const CustomBinaryFunc& op) const {
+      return TensorCustomBinaryOp<const CustomBinaryFunc, const Derived, const OtherDerived>(derived(), other, op);
+    }
+
+    // Force the evaluation of the expression.
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorForcedEvalOp<const Derived> eval() const {
+      return TensorForcedEvalOp<const Derived>(derived());
+    }
+
+  protected:
+    template <typename Scalar, int NumIndices, int Options, typename IndexType> friend class Tensor;
+    template <typename Scalar, typename Dimensions, int Option, typename IndexTypes> friend class TensorFixedSize;
+    // the Eigen:: prefix is required to workaround a compilation issue with nvcc 9.0
+    template <typename OtherDerived, int AccessLevel> friend class Eigen::TensorBase;
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Derived& derived() const { return *static_cast<const Derived*>(this); }
+};
+
+template<typename Derived, int AccessLevel = internal::accessors_level<Derived>::value>
+class TensorBase : public TensorBase<Derived, ReadOnlyAccessors> {
+ public:
+    typedef TensorBase<Derived, ReadOnlyAccessors> Base;
+    typedef internal::traits<Derived> DerivedTraits;
+    typedef typename DerivedTraits::Scalar Scalar;
+    typedef typename DerivedTraits::Index Index;
+    typedef Scalar CoeffReturnType;
+    static const int NumDimensions = DerivedTraits::NumDimensions;
+
+    template <typename Scalar, int NumIndices, int Options, typename IndexType> friend class Tensor;
+    template <typename Scalar, typename Dimensions, int Option, typename IndexTypes> friend class TensorFixedSize;
+    // the Eigen:: prefix is required to workaround a compilation issue with nvcc 9.0
+    template <typename OtherDerived, int OtherAccessLevel> friend class Eigen::TensorBase;
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& setZero() {
+      return setConstant(Scalar(0));
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& setConstant(const Scalar& val) {
+      return derived() = this->constant(val);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& setRandom() {
+      return derived() = this->random();
+    }
+    template <typename RandomGenerator> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& setRandom() {
+      return derived() = this->template random<RandomGenerator>();
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& setValues(
+        const typename internal::Initializer<Derived, NumDimensions>::InitList& vals) {
+      TensorEvaluator<Derived, DefaultDevice> eval(derived(), DefaultDevice());
+      internal::initialize_tensor<Derived, NumDimensions>(eval, vals);
+      return derived();
+    }
+#endif  // EIGEN_HAS_VARIADIC_TEMPLATES
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator+=(const OtherDerived& other) {
+      return derived() = derived() + other.derived();
+    }
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator-=(const OtherDerived& other) {
+      return derived() = derived() - other.derived();
+    }
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator*=(const OtherDerived& other) {
+      return derived() = derived() * other.derived();
+    }
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator/=(const OtherDerived& other) {
+      return derived() = derived() / other.derived();
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorLayoutSwapOp<const Derived>
+    swap_layout() const {
+      return TensorLayoutSwapOp<const Derived>(derived());
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorLayoutSwapOp<Derived>
+    swap_layout() {
+      return TensorLayoutSwapOp<Derived>(derived());
+    }
+
+    template <typename Axis, typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorConcatenationOp<const Axis, const Derived, const OtherDerived>
+    concatenate(const OtherDerived& other, const Axis& axis) const {
+      return TensorConcatenationOp<const Axis, const Derived, const OtherDerived>(derived(), other, axis);
+    }
+    template <typename Axis, typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorConcatenationOp<const Axis, Derived, OtherDerived>
+    concatenate(const OtherDerived& other, const Axis& axis) {
+      return TensorConcatenationOp<const Axis, Derived, OtherDerived>(derived(), other, axis);
+    }
+
+    template <typename NewDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReshapingOp<const NewDimensions, const Derived>
+    reshape(const NewDimensions& newDimensions) const {
+      return TensorReshapingOp<const NewDimensions, const Derived>(derived(), newDimensions);
+    }
+    template <typename NewDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorReshapingOp<const NewDimensions, Derived>
+    reshape(const NewDimensions& newDimensions) {
+      return TensorReshapingOp<const NewDimensions, Derived>(derived(), newDimensions);
+    }
+
+    template <typename StartIndices, typename Sizes> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorSlicingOp<const StartIndices, const Sizes, const Derived>
+    slice(const StartIndices& startIndices, const Sizes& sizes) const {
+      return TensorSlicingOp<const StartIndices, const Sizes, const Derived>(derived(), startIndices, sizes);
+    }
+    template <typename StartIndices, typename Sizes> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorSlicingOp<const StartIndices, const Sizes, Derived>
+    slice(const StartIndices& startIndices, const Sizes& sizes) {
+      return TensorSlicingOp<const StartIndices, const Sizes, Derived>(derived(), startIndices, sizes);
+    }
+
+    template <typename StartIndices, typename StopIndices, typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorStridingSlicingOp<const StartIndices, const StopIndices, const Strides, const Derived>
+    stridedSlice(const StartIndices& startIndices, const StopIndices& stopIndices, const Strides& strides) const {
+      return TensorStridingSlicingOp<const StartIndices, const StopIndices, const Strides,
+                                const Derived>(derived(), startIndices, stopIndices, strides);
+    }
+    template <typename StartIndices, typename StopIndices, typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorStridingSlicingOp<const StartIndices, const StopIndices, const Strides, Derived>
+    stridedSlice(const StartIndices& startIndices, const StopIndices& stopIndices, const Strides& strides) {
+      return TensorStridingSlicingOp<const StartIndices, const StopIndices, const Strides,
+                                Derived>(derived(), startIndices, stopIndices, strides);
+    }
+
+    template <DenseIndex DimId> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorChippingOp<DimId, const Derived>
+    chip(const Index offset) const {
+      return TensorChippingOp<DimId, const Derived>(derived(), offset, DimId);
+    }
+    template <Index DimId> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorChippingOp<DimId, Derived>
+    chip(const Index offset) {
+      return TensorChippingOp<DimId, Derived>(derived(), offset, DimId);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorChippingOp<Dynamic, const Derived>
+    chip(const Index offset, const Index dim) const {
+      return TensorChippingOp<Dynamic, const Derived>(derived(), offset, dim);
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorChippingOp<Dynamic, Derived>
+    chip(const Index offset, const Index dim) {
+      return TensorChippingOp<Dynamic, Derived>(derived(), offset, dim);
+    }
+
+    template <typename ReverseDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorReverseOp<const ReverseDimensions, const Derived>
+    reverse(const ReverseDimensions& rev) const {
+      return TensorReverseOp<const ReverseDimensions, const Derived>(derived(), rev);
+    }
+    template <typename ReverseDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorReverseOp<const ReverseDimensions, Derived>
+    reverse(const ReverseDimensions& rev) {
+      return TensorReverseOp<const ReverseDimensions, Derived>(derived(), rev);
+    }
+
+    template <typename Shuffle> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorShufflingOp<const Shuffle, const Derived>
+    shuffle(const Shuffle& shfl) const {
+      return TensorShufflingOp<const Shuffle, const Derived>(derived(), shfl);
+    }
+    template <typename Shuffle> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorShufflingOp<const Shuffle, Derived>
+    shuffle(const Shuffle& shfl) {
+      return TensorShufflingOp<const Shuffle, Derived>(derived(), shfl);
+    }
+
+    template <typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorStridingOp<const Strides, const Derived>
+    stride(const Strides& strides) const {
+      return TensorStridingOp<const Strides, const Derived>(derived(), strides);
+    }
+    template <typename Strides> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorStridingOp<const Strides, Derived>
+    stride(const Strides& strides) {
+      return TensorStridingOp<const Strides, Derived>(derived(), strides);
+    }
+
+    // Select the device on which to evaluate the expression.
+    template <typename DeviceType>
+    TensorDevice<Derived, DeviceType> device(const DeviceType& dev) {
+      return TensorDevice<Derived, DeviceType>(dev, derived());
+    }
+
+    // Select the async device on which to evaluate the expression.
+    template <typename DeviceType, typename DoneCallback>
+    TensorAsyncDevice<Derived, DeviceType, DoneCallback> device(const DeviceType& dev, DoneCallback done) {
+      return TensorAsyncDevice<Derived, DeviceType, DoneCallback>(dev, derived(), std::move(done));
+    }
+
+ protected:
+    EIGEN_DEFAULT_EMPTY_CONSTRUCTOR_AND_DESTRUCTOR(TensorBase)
+    EIGEN_DEFAULT_COPY_CONSTRUCTOR(TensorBase)
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& operator=(const OtherDerived& other)
+    {
+      typedef TensorAssignOp<Derived, const OtherDerived> Assign;
+      Assign assign(derived(), other.derived());
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
+      return derived();
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& derived() { return *static_cast<Derived*>(this); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Derived& derived() const { return *static_cast<const Derived*>(this); }
+};
+#endif // EIGEN_PARSED_BY_DOXYGEN
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_BASE_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorBlock.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorBlock.h
new file mode 100644
index 0000000..1e55d12
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorBlock.h
@@ -0,0 +1,1559 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_BLOCK_H
+#define EIGEN_CXX11_TENSOR_TENSOR_BLOCK_H
+
+namespace Eigen {
+namespace internal {
+
+// -------------------------------------------------------------------------- //
+// Forward declarations for templates defined below.
+template <typename Scalar, typename IndexType, int NumDims, int Layout>
+class TensorBlockIO;
+
+// -------------------------------------------------------------------------- //
+// Helper function to compute strides for densely stored buffer of given
+// dimensions.
+
+// TODO(ezhulenev): We compute strides 1000 times in different evaluators, use
+// this function instead everywhere.
+template <int Layout, typename IndexType, int NumDims>
+EIGEN_ALWAYS_INLINE DSizes<IndexType, NumDims> strides(
+    const DSizes<IndexType, NumDims>& dimensions) {
+  DSizes<IndexType, NumDims> strides;
+  if (NumDims == 0) return strides;
+
+  // TODO(ezhulenev): Use templates to unroll this loop (similar to
+  // h_array_reduce in CXX11meta.h)? Benchmark it.
+  if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+    strides[0] = 1;
+    for (int i = 1; i < NumDims; ++i) {
+      strides[i] = strides[i - 1] * dimensions[i - 1];
+    }
+  } else {
+    strides[NumDims - 1] = 1;
+    for (int i = NumDims - 2; i >= 0; --i) {
+      strides[i] = strides[i + 1] * dimensions[i + 1];
+    }
+  }
+
+  return strides;
+}
+
+template <int Layout, typename IndexType, size_t NumDims>
+EIGEN_ALWAYS_INLINE DSizes<IndexType, NumDims> strides(
+    const Eigen::array<IndexType, NumDims>& dimensions) {
+  return strides<Layout>(DSizes<IndexType, NumDims>(dimensions));
+}
+
+template <int Layout, std::ptrdiff_t... Indices>
+EIGEN_STRONG_INLINE DSizes<std::ptrdiff_t, sizeof...(Indices)> strides(
+    const Sizes<Indices...>& sizes) {
+  return strides<Layout>(DSizes<std::ptrdiff_t, sizeof...(Indices)>(sizes));
+}
+
+// -------------------------------------------------------------------------- //
+
+// Tensor block shape type defines what are the shape preference for the blocks
+// extracted from the larger tensor.
+//
+// Example: blocks of 100 elements from the large 100x100 tensor:
+// - tensor: 100x100
+// - target_block_size: 100
+//
+// TensorBlockShapeType:
+//  - kUniformAllDims: 100 blocks of size 10x10
+//  - kSkewedInnerDims: 100 blocks of size 100x1 (or 1x100 depending on a column
+//                      or row major layout)
+enum class TensorBlockShapeType { kUniformAllDims, kSkewedInnerDims };
+
+struct TensorBlockResourceRequirements {
+  TensorBlockShapeType shape_type;  // target block shape
+  size_t size;                      // target block size
+  TensorOpCost cost_per_coeff;      // cost of computing a single block element
+
+#ifdef EIGEN_HIPCC
+  // For HIPCC, we need to explicitly declare as a "device fun", the constructor
+  // which is implicitly invoked in the "merge" / "any" routines. else HIPCC
+  // errors out complaining about the lack of a matching constructor
+  EIGEN_DEVICE_FUNC
+  TensorBlockResourceRequirements(TensorBlockShapeType shape_type_, size_t size_,
+				  TensorOpCost cost_)
+    : shape_type(shape_type_), size(size_), cost_per_coeff(cost_)
+  {}
+#endif
+
+  template <typename Scalar>
+  EIGEN_DEVICE_FUNC static TensorBlockResourceRequirements withShapeAndSize(
+      TensorBlockShapeType shape_type, size_t size_in_bytes,
+      TensorOpCost cost) {
+    const size_t size = numext::maxi(size_t(1), size_in_bytes / sizeof(Scalar));
+    return {shape_type, size, cost};
+  }
+
+  template <typename Scalar>
+  EIGEN_DEVICE_FUNC static TensorBlockResourceRequirements withShapeAndSize(
+      TensorBlockShapeType shape_type, size_t size_in_bytes) {
+    // This default cost per coefficient is valid for most materialized tensor
+    // block evaluation implementations, because they typically just read
+    // coefficients from the underlying tensor storage, and write to the tensor
+    // block buffer (scratch or destination memory, reads and writes have linear
+    // access pattern). We ignore the fixed cost of block evaluation, because in
+    // practice it should negligible.
+    //
+    // Lazy block evaluation adds the cost of calling a functor for each
+    // coefficient.
+    //
+    // All non-trivial block evaluation implementations must provide their own
+    // cost approximation (e.g. shuffling inner dimension has a much higher cost
+    // because it reads memory randomly, although the total number of moved
+    // bytes is the same).
+    return withShapeAndSize<Scalar>(shape_type, size_in_bytes,
+                                    {/*bytes_loaded=*/sizeof(Scalar),
+                                     /*bytes_stored=*/sizeof(Scalar),
+                                     /*compute_cycles=*/0});
+  }
+
+  template <typename Scalar>
+  EIGEN_DEVICE_FUNC static TensorBlockResourceRequirements skewed(
+      size_t size_in_bytes) {
+    return withShapeAndSize<Scalar>(TensorBlockShapeType::kSkewedInnerDims,
+                                    size_in_bytes);
+  }
+
+  template <typename Scalar>
+  EIGEN_DEVICE_FUNC static TensorBlockResourceRequirements uniform(
+      size_t size_in_bytes) {
+    return withShapeAndSize<Scalar>(TensorBlockShapeType::kUniformAllDims,
+                                    size_in_bytes);
+  }
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE TensorBlockResourceRequirements
+  merge(const TensorBlockResourceRequirements& lhs,
+        const TensorBlockResourceRequirements& rhs) {
+    return {merge(lhs.shape_type, rhs.shape_type),           // shape_type
+            merge(lhs.size, rhs.size),                       // size
+            merge(lhs.cost_per_coeff, rhs.cost_per_coeff)};  // cost_per_coeff
+  }
+
+  EIGEN_DEVICE_FUNC TensorBlockResourceRequirements& addCostPerCoeff(
+      TensorOpCost cost) {
+    cost_per_coeff += cost;
+    return *this;
+  }
+
+  // This is a resource requirement that should be returned from expressions
+  // that do not have any block evaluation preference (e.g. default tensor
+  // expression with raw buffer access).
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE TensorBlockResourceRequirements any() {
+    return {TensorBlockShapeType::kUniformAllDims, 1, {0, 0, 0}};
+  }
+
+ private:
+  using Requirements = TensorBlockResourceRequirements;
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE size_t merge(size_t lhs_size, size_t rhs_size) {
+    return numext::maxi(lhs_size, rhs_size);
+  }
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE TensorBlockShapeType
+  merge(TensorBlockShapeType lhs, TensorBlockShapeType rhs) {
+    return (lhs == TensorBlockShapeType::kSkewedInnerDims ||
+            rhs == TensorBlockShapeType::kSkewedInnerDims)
+               ? TensorBlockShapeType::kSkewedInnerDims
+               : TensorBlockShapeType::kUniformAllDims;
+  }
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE TensorOpCost merge(TensorOpCost lhs_cost,
+                                                TensorOpCost rhs_cost) {
+    return lhs_cost + rhs_cost;
+  }
+};
+
+// -------------------------------------------------------------------------- //
+// TensorBlockDescriptor specifies a block offset within a tensor and the block
+// sizes along each of the tensor dimensions.
+
+template <int NumDims, typename IndexType = Eigen::Index>
+class TensorBlockDescriptor {
+ public:
+  typedef DSizes<IndexType, NumDims> Dimensions;
+
+  // If we evaluate a Tensor assignment, and expression on the left, already has
+  // a memory buffer, then we might do performance optimization, and evaluate
+  // the root expression directly into the final output memory. Some time it's
+  // possible to reuse it for materializing subexpressions inside an expression
+  // tree, to to avoid dynamic memory allocation.
+  //
+  // The pointer type of the underlying storage is erased, because passing
+  // Scalar type through all the expression evaluation layers is way too many
+  // templates. In practice destination buffer type should always match the
+  // evaluated expression scalar type.
+  class DestinationBuffer {
+   public:
+    enum DestinationBufferKind : int {
+      // The above explicit specification of "int" as the enum basetype is
+      // needed to get around a HIPCC link error ("the field type is not
+      // amp-compatible")
+      // which is issued for class members with the enum type.
+      // TODO(rocm):
+      // remove the "int" basetype once HIPCC has been fixed to not error out
+      // in the above scenario.
+
+      // Destination buffer is not defined (`m_data` == nullptr).
+      kEmpty,
+
+      // Tensor block defined by an owning tensor block descriptor can fit
+      // contiguously into the destination buffer. In this case it's safe to
+      // materialize tensor block in the destination buffer, wrap it in a
+      // TensorMap, and use to build Eigen expression on top of it.
+      kContiguous,
+
+      // Destination buffer strides do not match strides of the contiguously
+      // stored block, and it's impossible to define a TensorMap over this
+      // buffer. However if we are evaluating a root of an expression tree, we
+      // still can materialize an output into this destination, because we can
+      // guarantee that no one will ever access it through block API.
+      //
+      // In theory it is possible to build valid TensorStriding<TensorMap>
+      // expression on top of this destination buffer, however it has
+      // inefficient coeff/packet access, and defeats the purpose of fast block
+      // evaluation API.
+      kStrided
+    };
+
+    template <typename Scalar>
+    Scalar* data() const {
+      eigen_assert(m_data_type_size == sizeof(Scalar));
+      return static_cast<Scalar*>(m_data);
+    }
+
+    const Dimensions& strides() const { return m_strides; }
+    const DestinationBufferKind& kind() const { return m_kind; }
+
+   private:
+    friend class TensorBlockDescriptor;
+
+    DestinationBuffer() : m_data(NULL), m_data_type_size(0), m_kind(kEmpty) {}
+
+    template <typename Scalar>
+    DestinationBuffer(Scalar* data, const Dimensions& strides,
+                      DestinationBufferKind kind)
+        : m_data(static_cast<void*>(data)),
+          m_data_type_size(sizeof(Scalar)),
+          m_strides(strides),
+          m_kind(kind) {}
+
+    template <int Layout, typename Scalar>
+    static DestinationBuffer make(const TensorBlockDescriptor& desc,
+                                  Scalar* data, const Dimensions& strides) {
+      return DestinationBuffer(data, strides, kind<Layout>(desc, strides));
+    }
+
+    template <int Layout>
+    static DestinationBufferKind kind(const TensorBlockDescriptor& desc,
+                                      const Dimensions& strides) {
+      const Dimensions& desc_dims = desc.dimensions();
+      const Dimensions& desc_strides = internal::strides<Layout>(desc_dims);
+      for (int i = 0; i < NumDims; ++i) {
+        if (desc_dims[i] == 1) continue;
+        if (desc_strides[i] != strides[i]) return kStrided;
+      }
+      return kContiguous;
+    }
+
+    // Storage pointer is type erased, to reduce template bloat, but we still
+    // keep the size of the underlying element type for error checking.
+    void* m_data;
+    size_t m_data_type_size;
+
+    // Destination buffer dimensions always match the dimensions of a tensor
+    // block descriptor it belongs to, however strides might be different.
+    Dimensions m_strides;
+
+    DestinationBufferKind m_kind;
+  };
+
+  TensorBlockDescriptor(const IndexType offset, const Dimensions& dimensions,
+                        const DestinationBuffer& destination)
+      : m_offset(offset),
+        m_dimensions(dimensions),
+        m_destination(destination) {}
+
+  TensorBlockDescriptor(const IndexType offset, const Dimensions& dimensions)
+      : m_offset(offset),
+        m_dimensions(dimensions),
+        m_destination(DestinationBuffer()) {}
+
+  IndexType offset() const { return m_offset; }
+  const Dimensions& dimensions() const { return m_dimensions; }
+  IndexType dimension(int index) const { return m_dimensions[index]; }
+  IndexType size() const { return array_prod<IndexType>(m_dimensions); }
+
+  const DestinationBuffer& destination() const { return m_destination; }
+
+  template <int Layout, typename Scalar>
+  void AddDestinationBuffer(Scalar* dst_base, const Dimensions& dst_strides) {
+    eigen_assert(dst_base != NULL);
+    m_destination =
+        DestinationBuffer::template make<Layout>(*this, dst_base, dst_strides);
+  }
+
+  template <int Layout, typename Scalar, typename DstStridesIndexType>
+  void AddDestinationBuffer(
+      Scalar* dst_base,
+      const DSizes<DstStridesIndexType, NumDims>& dst_strides) {
+    // DSizes constructor will do index type promotion if it's safe.
+    AddDestinationBuffer<Layout>(dst_base, Dimensions(dst_strides));
+  }
+
+  TensorBlockDescriptor& DropDestinationBuffer() {
+    m_destination.m_data = NULL;
+    m_destination.m_kind = DestinationBuffer::kEmpty;
+    return *this;
+  }
+
+  bool HasDestinationBuffer() const {
+    return m_destination.kind() != DestinationBuffer::kEmpty;
+  }
+
+  // Returns a copy of `*this` with updated offset.
+  TensorBlockDescriptor WithOffset(IndexType offset) const {
+    return TensorBlockDescriptor(offset, m_dimensions, m_destination);
+  }
+
+ private:
+  // Offset and dimensions are immutable after construction. Block descriptor
+  // can only be mutated by adding or dropping destination.
+  const IndexType m_offset;
+  const Dimensions m_dimensions;
+  DestinationBuffer m_destination;
+};
+
+// -------------------------------------------------------------------------- //
+// TensorBlockMapper is responsible for iterating over the blocks of a tensor.
+
+template <int NumDims, int Layout, typename IndexType = Eigen::Index>
+class TensorBlockMapper {
+  typedef TensorBlockDescriptor<NumDims, IndexType> BlockDescriptor;
+
+ public:
+  typedef DSizes<IndexType, NumDims> Dimensions;
+
+  TensorBlockMapper() = default;
+  TensorBlockMapper(const DSizes<IndexType, NumDims>& dimensions,
+                    const TensorBlockResourceRequirements& requirements)
+      : m_tensor_dimensions(dimensions), m_requirements(requirements) {
+    // Compute block dimensions and the total number of blocks.
+    InitializeBlockDimensions();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE IndexType blockCount() const {
+    return m_total_block_count;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE IndexType blockTotalSize() const {
+    return m_block_dimensions.TotalSize();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const DSizes<IndexType, NumDims>&
+  blockDimensions() const {
+    return m_block_dimensions;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE BlockDescriptor
+  blockDescriptor(IndexType block_index) const {
+    static const bool isColMajor = Layout == static_cast<int>(ColMajor);
+
+    IndexType offset = 0;
+    DSizes<IndexType, NumDims> dimensions;
+
+    if (NumDims == 0) return BlockDescriptor(offset, dimensions);
+
+    // Iterate outer -> inner dimensions.
+    for (int i = NumDims - 1; i >= 0; --i) {
+      const int dim = isColMajor ? i : NumDims - i - 1;
+
+      const IndexType idx = block_index / m_block_strides[dim];
+      block_index -= idx * m_block_strides[dim];
+
+      const IndexType coord = idx * m_block_dimensions[dim];
+      dimensions[dim] = numext::mini(m_tensor_dimensions[dim] - coord,
+                                     m_block_dimensions[dim]);
+      offset += coord * m_tensor_strides[dim];
+    }
+
+    return {offset, dimensions};
+  }
+
+ private:
+  void InitializeBlockDimensions() {
+    // Requested block shape and size.
+    const TensorBlockShapeType shape_type = m_requirements.shape_type;
+    IndexType target_block_size =
+        numext::maxi<IndexType>(1, static_cast<IndexType>(m_requirements.size));
+
+    IndexType tensor_size = m_tensor_dimensions.TotalSize();
+
+    // Corner case: one of the dimensions is zero. Logic below is too complex
+    // to handle this case on a general basis, just use unit block size.
+    // Note: we must not yield blocks with zero dimensions (recipe for
+    // overflows/underflows, divisions by zero and NaNs later).
+    if (tensor_size == 0) {
+      for (int i = 0; i < NumDims; ++i) {
+        m_block_dimensions[i] = 1;
+      }
+      m_total_block_count = 0;
+      return;
+    }
+
+    // If tensor fits into a target block size, evaluate it as a single block.
+    if (tensor_size <= target_block_size) {
+      m_block_dimensions = m_tensor_dimensions;
+      m_total_block_count = 1;
+      // The only valid block index is `0`, and in this case we do not need
+      // to compute real strides for tensor or blocks (see blockDescriptor).
+      for (int i = 0; i < NumDims; ++i) {
+        m_tensor_strides[i] = 0;
+        m_block_strides[i] = 1;
+      }
+      return;
+    }
+
+    static const bool isColMajor = Layout == static_cast<int>(ColMajor);
+
+    // Block shape skewed towards inner dimension.
+    if (shape_type == TensorBlockShapeType::kSkewedInnerDims) {
+      IndexType coeff_to_allocate = target_block_size;
+
+      for (int i = 0; i < NumDims; ++i) {
+        const int dim = isColMajor ? i : NumDims - i - 1;
+        m_block_dimensions[dim] =
+            numext::mini(coeff_to_allocate, m_tensor_dimensions[dim]);
+        coeff_to_allocate = divup(
+            coeff_to_allocate,
+            numext::maxi(static_cast<IndexType>(1), m_block_dimensions[dim]));
+      }
+      eigen_assert(coeff_to_allocate == 1);
+
+    } else if (shape_type == TensorBlockShapeType::kUniformAllDims) {
+      // Tensor will not fit within 'target_block_size' budget: calculate tensor
+      // block dimension sizes based on "square" dimension size target.
+      const IndexType dim_size_target = convert_index<IndexType>(
+          std::pow(static_cast<float>(target_block_size),
+                   1.0f / static_cast<float>(m_block_dimensions.rank())));
+
+      for (int i = 0; i < NumDims; ++i) {
+        // TODO(andydavis) Adjust the inner most 'block_dim_size' to make it
+        // a multiple of the packet size. Note that reducing
+        // 'block_dim_size' in this manner can increase the number of
+        // blocks, and so will amplify any per-block overhead.
+        m_block_dimensions[i] =
+            numext::mini(dim_size_target, m_tensor_dimensions[i]);
+      }
+
+      // Add any un-allocated coefficients to inner dimension(s).
+      IndexType total_size = m_block_dimensions.TotalSize();
+      for (int i = 0; i < NumDims; ++i) {
+        const int dim = isColMajor ? i : NumDims - i - 1;
+
+        if (m_block_dimensions[dim] < m_tensor_dimensions[dim]) {
+          const IndexType total_size_other_dims =
+              total_size / m_block_dimensions[dim];
+          const IndexType alloc_avail =
+              divup<IndexType>(target_block_size, total_size_other_dims);
+          if (alloc_avail == m_block_dimensions[dim]) {
+            // Insufficient excess coefficients to allocate.
+            break;
+          }
+          m_block_dimensions[dim] =
+              numext::mini(m_tensor_dimensions[dim], alloc_avail);
+          total_size = total_size_other_dims * m_block_dimensions[dim];
+        }
+      }
+
+    } else {
+      eigen_assert(false);  // unknown block shape
+    }
+
+    eigen_assert(m_block_dimensions.TotalSize() >=
+                 numext::mini<IndexType>(target_block_size,
+                                         m_tensor_dimensions.TotalSize()));
+
+    // Calculate block counts by dimension and total block count.
+    DSizes<IndexType, NumDims> block_count;
+    for (int i = 0; i < NumDims; ++i) {
+      block_count[i] = divup(m_tensor_dimensions[i], m_block_dimensions[i]);
+    }
+    m_total_block_count = array_prod(block_count);
+
+    // Calculate block strides (used for enumerating blocks).
+    m_tensor_strides = strides<Layout>(m_tensor_dimensions);
+    m_block_strides = strides<Layout>(block_count);
+  }
+
+  DSizes<IndexType, NumDims> m_tensor_dimensions;
+  TensorBlockResourceRequirements m_requirements;
+
+  DSizes<IndexType, NumDims> m_block_dimensions;
+  IndexType m_total_block_count;
+
+  DSizes<IndexType, NumDims> m_tensor_strides;
+  DSizes<IndexType, NumDims> m_block_strides;
+};
+
+// -------------------------------------------------------------------------- //
+// TensorBlockScratchAllocator is responsible for allocating temporary buffers
+// for block evaluation (output or input block materialization). Given that
+// Eigen expression traversal order is deterministic, all temporary allocations
+// are happening in the same order, and usually have exactly the same size.
+// Scratch allocator keeps a trace of all dynamic allocations, and after the
+// first block evaluation is completed, we should be able to reuse all the
+// temporary buffers for the next block evaluation.
+
+template <typename Device>
+class TensorBlockScratchAllocator {
+ public:
+  explicit TensorBlockScratchAllocator(const Device& device)
+      : m_device(device), m_allocation_index(0) {}
+
+  ~TensorBlockScratchAllocator() {
+    for (size_t i = 0; i < m_allocations.size(); ++i) {
+      m_device.deallocate(m_allocations[i].ptr);
+    }
+  }
+
+  void* allocate(size_t size) {
+    // TODO(ezhulenev): Remove when replaced with inlined vector.
+    if (m_allocations.capacity() == 0) m_allocations.reserve(8);
+
+    // Check if we already have an existing allocation att current index.
+    const int num_allocations = static_cast<int>(m_allocations.size());
+    const bool has_allocation = m_allocation_index < num_allocations;
+
+    // Allocation index can't be larger than the number of allocations.
+    eigen_assert(m_allocation_index <= num_allocations);
+
+    // If we have existing allocation, and its size is larger or equal to
+    // requested size, we do nothing.
+
+    // If current allocation can't fit requested size, we deallocate it, and
+    // replace with a larger allocation.
+    if (has_allocation && m_allocations[m_allocation_index].size < size) {
+      m_device.deallocate(m_allocations[m_allocation_index].ptr);
+      m_allocations[m_allocation_index].ptr = m_device.allocate(size);
+      m_allocations[m_allocation_index].size = size;
+    }
+
+    // Make a new allocation if we don't have and existing one.
+    if (!has_allocation) {
+      Allocation allocation;
+      allocation.ptr = m_device.allocate(size);
+      allocation.size = size;
+      m_allocations.push_back(allocation);
+    }
+
+    eigen_assert(m_allocations[m_allocation_index].ptr != NULL);
+    eigen_assert(m_allocations[m_allocation_index].size >= size);
+
+    return m_allocations[m_allocation_index++].ptr;
+  }
+
+  void reset() { m_allocation_index = 0; }
+
+ private:
+  struct Allocation {
+    void* ptr;
+    size_t size;
+  };
+
+  const Device& m_device;
+  int m_allocation_index;
+  // TODO(ezhulenev): This should be an inlined vector.
+  std::vector<Allocation> m_allocations;
+};
+
+// -------------------------------------------------------------------------- //
+// TensorBlockKind represents all possible block kinds, that can be produced by
+// TensorEvaluator::evalBlock function.
+enum TensorBlockKind {
+  // Tensor block that is a lazy expression that must be assigned to a
+  // destination using TensorBlockAssign.
+  kExpr,
+
+  // Tensor block that is a view into a memory buffer owned by an underlying
+  // Tensor expression (e.g. it can be a view into a Tensor buffer).
+  kView,
+
+  // Tensor block that was materialized in a scratch memory buffer, allocated
+  // with TensorBlockScratchAllocator. This block must be copied to a
+  // destination, similar to a block of `kExpr` type.
+  kMaterializedInScratch,
+
+  // Tensor block that was materialized directly into the final output memory
+  // buffer. For example if the left side of an assignment is a Tensor, we can
+  // directly materialize the block in the destination memory.
+  //
+  // If strides in the output buffer do not match tensor block strides, the
+  // Tensor expression will be invalid, and should not be used by
+  // TensorBlockAssign or for constructing another block expression.
+  kMaterializedInOutput
+};
+
+// -------------------------------------------------------------------------- //
+// TensorBlockNotImplemented should be used to defined TensorBlock typedef in
+// TensorEvaluators that do not support block evaluation.
+
+class TensorBlockNotImplemented {
+ public:
+  typedef void XprType;
+};
+
+// -------------------------------------------------------------------------- //
+// XprScalar extracts Scalar type from the Eigen expressions (if expression type
+// is not void). It's required to be able to define lazy block expression for
+// argument types, that do not support block evaluation.
+
+template <typename XprType>
+struct XprScalar {
+  typedef typename XprType::Scalar type;
+};
+template <>
+struct XprScalar<void> {
+  typedef void type;
+};
+
+// -------------------------------------------------------------------------- //
+// TensorMaterializedBlock is a fully evaluated block of the original tensor,
+// and XprType is just a TensorMap over the data. This block type is typically
+// used to materialize blocks of tensor expressions, that can't be efficiently
+// represented as lazy Tensor expressions with fast coeff/packet operations,
+// e.g. we materialize all broadcasts into evaluated blocks.
+//
+// TensorMaterializedBlock does not own its memory buffer, it's either a memory
+// buffer that backs the original expression (e.g. block is just a view into a
+// Tensor), or a memory buffer allocated with scratch allocator, and in this
+// case the scratch allocator will deallocate it at the end of block based
+// expression execution.
+//
+// If the block was evaluated directly into the output buffer, and strides in
+// the output buffer do not match block strides, the TensorMap expression will
+// be invalid, and should never be used in block assignment or any other tensor
+// expression.
+
+template <typename Scalar, int NumDims, int Layout,
+          typename IndexType = Eigen::Index>
+class TensorMaterializedBlock {
+ public:
+  typedef DSizes<IndexType, NumDims> Dimensions;
+  typedef TensorMap<const Tensor<Scalar, NumDims, Layout> > XprType;
+
+  TensorMaterializedBlock(TensorBlockKind kind, const Scalar* data,
+                          const Dimensions& dimensions, bool valid_expr = true)
+      : m_kind(kind),
+        m_data(data),
+        m_dimensions(dimensions),
+        m_expr(m_data, m_dimensions),
+        m_valid_expr(valid_expr) {
+    eigen_assert(m_kind == internal::TensorBlockKind::kView ||
+                 m_kind == internal::TensorBlockKind::kMaterializedInScratch ||
+                 m_kind == internal::TensorBlockKind::kMaterializedInOutput);
+  }
+
+  TensorBlockKind kind() const { return m_kind; }
+  // NOTE(ezhulenev): Returning XprType by value like in other block types
+  // causes asan failures. The theory is that XprType::Nested doesn't work
+  // properly for TensorMap.
+  const XprType& expr() const {
+    eigen_assert(m_valid_expr);
+    return m_expr;
+  }
+  const Scalar* data() const { return m_data; }
+  void cleanup() {}
+
+  typedef internal::TensorBlockDescriptor<NumDims, IndexType> TensorBlockDesc;
+
+  // TensorMaterializedBlock can be backed by different types of storage:
+  //
+  //   (1) Contiguous block of memory allocated with scratch allocator.
+  //   (2) Contiguous block of memory reused from tensor block descriptor
+  //       destination buffer.
+  //   (3) Strided block of memory reused from tensor block descriptor
+  //       destination buffer.
+  //
+  class Storage {
+   public:
+    Scalar* data() const { return m_data; }
+    const Dimensions& dimensions() const { return m_dimensions; }
+    const Dimensions& strides() const { return m_strides; }
+
+    TensorMaterializedBlock AsTensorMaterializedBlock() const {
+      return TensorMaterializedBlock(
+          m_materialized_in_output
+              ? internal::TensorBlockKind::kMaterializedInOutput
+              : internal::TensorBlockKind::kMaterializedInScratch,
+          m_data, m_dimensions, !m_strided_storage);
+    }
+
+   private:
+    friend class TensorMaterializedBlock;
+
+    Storage(Scalar* data, const Dimensions& dimensions,
+            const Dimensions& strides, bool materialized_in_output,
+            bool strided_storage)
+        : m_data(data),
+          m_dimensions(dimensions),
+          m_strides(strides),
+          m_materialized_in_output(materialized_in_output),
+          m_strided_storage(strided_storage) {}
+
+    Scalar* m_data;
+    Dimensions m_dimensions;
+    Dimensions m_strides;
+    bool m_materialized_in_output;
+    bool m_strided_storage;
+  };
+
+  // Creates a storage for materialized block either from the block descriptor
+  // destination buffer, or allocates a new buffer with scratch allocator.
+  template <typename TensorBlockScratch>
+  EIGEN_STRONG_INLINE static Storage prepareStorage(
+      TensorBlockDesc& desc, TensorBlockScratch& scratch,
+      bool allow_strided_storage = false) {
+    // Try to reuse destination as an output block buffer.
+    typedef typename TensorBlockDesc::DestinationBuffer DestinationBuffer;
+
+    if (desc.destination().kind() == DestinationBuffer::kContiguous) {
+      Scalar* buffer = desc.destination().template data<Scalar>();
+      desc.DropDestinationBuffer();
+      return Storage(buffer, desc.dimensions(),
+                     internal::strides<Layout>(desc.dimensions()),
+                     /*materialized_in_output=*/true,
+                     /*strided_storage=*/false);
+
+    } else if (desc.destination().kind() == DestinationBuffer::kStrided &&
+               allow_strided_storage) {
+      Scalar* buffer = desc.destination().template data<Scalar>();
+      desc.DropDestinationBuffer();
+      return Storage(buffer, desc.dimensions(), desc.destination().strides(),
+                     /*materialized_in_output=*/true, /*strided_storage=*/true);
+
+    } else {
+      void* mem = scratch.allocate(desc.size() * sizeof(Scalar));
+      return Storage(static_cast<Scalar*>(mem), desc.dimensions(),
+                     internal::strides<Layout>(desc.dimensions()),
+                     /*materialized_in_output=*/false,
+                     /*strided_storage=*/false);
+    }
+  }
+
+  // Creates a materialized block for the given descriptor from a memory buffer.
+  template <typename DataDimensions, typename TensorBlockScratch>
+  EIGEN_STRONG_INLINE static TensorMaterializedBlock materialize(
+      const Scalar* data, const DataDimensions& data_dims,
+      TensorBlockDesc& desc, TensorBlockScratch& scratch) {
+    eigen_assert(array_size<DataDimensions>::value == desc.dimensions().size());
+
+    // If a tensor block dimensions covers a contiguous block of the underlying
+    // memory, we can skip block buffer memory allocation, and construct a block
+    // from existing `data` memory buffer.
+    //
+    // Example: (RowMajor layout)
+    //   data_dims:          [11, 12, 13, 14]
+    //   desc.dimensions():  [1,   1,  3, 14]
+    //
+    // In this case we can construct a TensorBlock starting at
+    // `data + desc.offset()`, with a `desc.dimensions()` block sizes.
+    static const bool is_col_major = Layout == ColMajor;
+
+    // Find out how many inner dimensions have a matching size.
+    int num_matching_inner_dims = 0;
+    for (int i = 0; i < NumDims; ++i) {
+      int dim = is_col_major ? i : NumDims - i - 1;
+      if (data_dims[dim] != desc.dimensions()[dim]) break;
+      ++num_matching_inner_dims;
+    }
+
+    // All the outer dimensions must be of size `1`, except a single dimension
+    // before the matching inner dimension (`3` in the example above).
+    bool can_use_direct_access = true;
+    for (int i = num_matching_inner_dims + 1; i < NumDims; ++i) {
+      int dim = is_col_major ? i : NumDims - i - 1;
+      if (desc.dimension(dim) != 1) {
+        can_use_direct_access = false;
+        break;
+      }
+    }
+
+    if (can_use_direct_access) {
+      const Scalar* block_start = data + desc.offset();
+      return TensorMaterializedBlock(internal::TensorBlockKind::kView,
+                                     block_start, desc.dimensions());
+
+    } else {
+      // Reuse destination buffer or allocate new buffer with scratch allocator.
+      const Storage storage = prepareStorage(desc, scratch);
+
+      typedef internal::TensorBlockIO<Scalar, IndexType, NumDims, Layout>
+          TensorBlockIO;
+      typedef typename TensorBlockIO::Dst TensorBlockIODst;
+      typedef typename TensorBlockIO::Src TensorBlockIOSrc;
+
+      TensorBlockIOSrc src(internal::strides<Layout>(Dimensions(data_dims)),
+                           data, desc.offset());
+      TensorBlockIODst dst(storage.dimensions(), storage.strides(),
+                           storage.data());
+
+      TensorBlockIO::Copy(dst, src);
+      return storage.AsTensorMaterializedBlock();
+    }
+  }
+
+ private:
+  TensorBlockKind m_kind;
+  const Scalar* m_data;
+  Dimensions m_dimensions;
+  XprType m_expr;
+  bool m_valid_expr;
+};
+
+// -------------------------------------------------------------------------- //
+// TensorCwiseUnaryBlock is a lazy tensor expression block that applies UnaryOp
+// functor to the blocks produced by the underlying Tensor expression.
+
+template <typename UnaryOp, typename ArgTensorBlock>
+class TensorCwiseUnaryBlock {
+  static const bool NoArgBlockAccess =
+      internal::is_void<typename ArgTensorBlock::XprType>::value;
+
+ public:
+  typedef typename conditional<
+      NoArgBlockAccess, void,
+      TensorCwiseUnaryOp<UnaryOp, const typename ArgTensorBlock::XprType> >::
+      type XprType;
+
+  typedef typename XprScalar<XprType>::type Scalar;
+
+  TensorCwiseUnaryBlock(const ArgTensorBlock& arg_block, const UnaryOp& functor)
+      : m_arg_block(arg_block), m_functor(functor) {}
+
+  TensorBlockKind kind() const { return internal::TensorBlockKind::kExpr; }
+
+  XprType expr() const { return XprType(m_arg_block.expr(), m_functor); }
+  const Scalar* data() const { return NULL; }
+  void cleanup() { m_arg_block.cleanup(); }
+
+ private:
+  ArgTensorBlock m_arg_block;
+  UnaryOp m_functor;
+};
+
+// -------------------------------------------------------------------------- //
+// TensorCwiseUnaryBlock is a lazy tensor expression block that applies BinaryOp
+// functor to the blocks produced by the underlying Tensor expression.
+
+template <typename BinaryOp, typename LhsTensorBlock, typename RhsTensorBlock>
+class TensorCwiseBinaryBlock {
+  static const bool NoArgBlockAccess =
+      internal::is_void<typename LhsTensorBlock::XprType>::value ||
+      internal::is_void<typename RhsTensorBlock::XprType>::value;
+
+ public:
+  typedef typename conditional<
+      NoArgBlockAccess, void,
+      TensorCwiseBinaryOp<BinaryOp, const typename LhsTensorBlock::XprType,
+                          const typename RhsTensorBlock::XprType> >::type
+      XprType;
+
+  typedef typename XprScalar<XprType>::type Scalar;
+
+  TensorCwiseBinaryBlock(const LhsTensorBlock& left_block,
+                         const RhsTensorBlock& right_block,
+                         const BinaryOp& functor)
+      : m_left_block(left_block),
+        m_right_block(right_block),
+        m_functor(functor) {}
+
+  TensorBlockKind kind() const { return internal::TensorBlockKind::kExpr; }
+
+  XprType expr() const {
+    return XprType(m_left_block.expr(), m_right_block.expr(), m_functor);
+  }
+
+  const Scalar* data() const { return NULL; }
+
+  void cleanup() {
+    m_left_block.cleanup();
+    m_right_block.cleanup();
+  }
+
+ private:
+  LhsTensorBlock m_left_block;
+  RhsTensorBlock m_right_block;
+  BinaryOp m_functor;
+};
+
+// -------------------------------------------------------------------------- //
+// TensorUnaryExprBlock is a lazy tensor expression block that can construct
+// an arbitrary tensor expression from a block of the underlying type (this is a
+// generalization of the TensorCwiseUnaryBlock for arbitrary expressions).
+
+template <typename BlockFactory, typename ArgTensorBlock>
+class TensorUnaryExprBlock {
+  typedef typename ArgTensorBlock::XprType ArgXprType;
+  static const bool NoArgBlockAccess = internal::is_void<ArgXprType>::value;
+
+ public:
+  typedef typename conditional<
+      NoArgBlockAccess, void,
+      typename BlockFactory::template XprType<ArgXprType>::type>::type XprType;
+
+  typedef typename XprScalar<XprType>::type Scalar;
+
+  TensorUnaryExprBlock(const ArgTensorBlock& arg_block,
+                       const BlockFactory& factory)
+      : m_arg_block(arg_block), m_factory(factory) {}
+
+  TensorBlockKind kind() const { return internal::TensorBlockKind::kExpr; }
+  XprType expr() const { return m_factory.expr(m_arg_block.expr()); }
+  const Scalar* data() const { return NULL; }
+  void cleanup() { m_arg_block.cleanup(); }
+
+ private:
+  ArgTensorBlock m_arg_block;
+  BlockFactory m_factory;
+};
+
+// -------------------------------------------------------------------------- //
+// TensorTernaryExprBlock is a lazy tensor expression block that can construct
+// an arbitrary tensor expression from three blocks of the underlying type.
+
+template <typename BlockFactory, typename Arg1TensorBlock,
+          typename Arg2TensorBlock, typename Arg3TensorBlock>
+class TensorTernaryExprBlock {
+  typedef typename Arg1TensorBlock::XprType Arg1XprType;
+  typedef typename Arg2TensorBlock::XprType Arg2XprType;
+  typedef typename Arg3TensorBlock::XprType Arg3XprType;
+
+  static const bool NoArgBlockAccess = internal::is_void<Arg1XprType>::value ||
+                                       internal::is_void<Arg2XprType>::value ||
+                                       internal::is_void<Arg3XprType>::value;
+
+ public:
+  typedef typename conditional<
+      NoArgBlockAccess, void,
+      typename BlockFactory::template XprType<Arg1XprType, Arg2XprType,
+                                              Arg3XprType>::type>::type XprType;
+
+  typedef typename XprScalar<XprType>::type Scalar;
+
+  TensorTernaryExprBlock(const Arg1TensorBlock& arg1_block,
+                         const Arg2TensorBlock& arg2_block,
+                         const Arg3TensorBlock& arg3_block,
+                         const BlockFactory& factory)
+      : m_arg1_block(arg1_block),
+        m_arg2_block(arg2_block),
+        m_arg3_block(arg3_block),
+        m_factory(factory) {}
+
+  TensorBlockKind kind() const { return internal::TensorBlockKind::kExpr; }
+  XprType expr() const {
+    return m_factory.expr(m_arg1_block.expr(), m_arg2_block.expr(),
+                          m_arg3_block.expr());
+  }
+  const Scalar* data() const { return NULL; }
+  void cleanup() {
+    m_arg1_block.cleanup();
+    m_arg2_block.cleanup();
+    m_arg3_block.cleanup();
+  }
+
+ private:
+  Arg1TensorBlock m_arg1_block;
+  Arg2TensorBlock m_arg2_block;
+  Arg3TensorBlock m_arg3_block;
+  BlockFactory m_factory;
+};
+
+// -------------------------------------------------------------------------- //
+// StridedLinearBufferCopy provides a method to copy data between two linear
+// buffers with different strides, with optimized paths for scatter/gather.
+
+template <typename Scalar, typename IndexType>
+class StridedLinearBufferCopy {
+  typedef typename packet_traits<Scalar>::type Packet;
+  enum {
+    Vectorizable = packet_traits<Scalar>::Vectorizable,
+    PacketSize = packet_traits<Scalar>::size
+  };
+
+ public:
+  // Specifying linear copy kind statically gives ~30% speedup for small sizes.
+  enum class Kind {
+    Linear = 0,       // src_stride == 1 && dst_stride == 1
+    Scatter = 1,      // src_stride == 1 && dst_stride != 1
+    FillLinear = 2,   // src_stride == 0 && dst_stride == 1
+    FillScatter = 3,  // src_stride == 0 && dst_stride != 1
+    Gather = 4,       // dst_stride == 1
+    Random = 5        // everything else
+  };
+
+  struct Dst {
+    Dst(IndexType o, IndexType s, Scalar* d) : offset(o), stride(s), data(d) {}
+
+    IndexType offset;
+    IndexType stride;
+    Scalar* data;
+  };
+
+  struct Src {
+    Src(IndexType o, IndexType s, const Scalar* d)
+        : offset(o), stride(s), data(d) {}
+
+    IndexType offset;
+    IndexType stride;
+    const Scalar* data;
+  };
+
+  template <typename StridedLinearBufferCopy::Kind kind>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(const Dst& dst,
+                                                        const Src& src,
+                                                        const size_t count) {
+    Run<kind>(count, dst.offset, dst.stride, dst.data, src.offset, src.stride,
+              src.data);
+  }
+
+ private:
+  template <typename StridedLinearBufferCopy::Kind kind>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const IndexType count, const IndexType dst_offset,
+      const IndexType dst_stride, Scalar* EIGEN_RESTRICT dst_data,
+      const IndexType src_offset, const IndexType src_stride,
+      const Scalar* EIGEN_RESTRICT src_data) {
+    const Scalar* src = &src_data[src_offset];
+    Scalar* dst = &dst_data[dst_offset];
+
+    if (!Vectorizable) {
+      for (Index i = 0; i < count; ++i) {
+        dst[i * dst_stride] = src[i * src_stride];
+      }
+      return;
+    }
+
+    const IndexType vectorized_size = count - PacketSize;
+    IndexType i = 0;
+
+    if (kind == StridedLinearBufferCopy::Kind::Linear) {
+      // ******************************************************************** //
+      // Linear copy from `src` to `dst`.
+      const IndexType unrolled_size = count - 4 * PacketSize;
+      eigen_assert(src_stride == 1 && dst_stride == 1);
+      for (; i <= unrolled_size; i += 4 * PacketSize) {
+        for (int j = 0; j < 4; ++j) {
+          Packet p = ploadu<Packet>(src + i + j * PacketSize);
+          pstoreu<Scalar, Packet>(dst + i + j * PacketSize, p);
+        }
+      }
+      for (; i <= vectorized_size; i += PacketSize) {
+        Packet p = ploadu<Packet>(src + i);
+        pstoreu<Scalar, Packet>(dst + i, p);
+      }
+      for (; i < count; ++i) {
+        dst[i] = src[i];
+      }
+      // ******************************************************************** //
+    } else if (kind == StridedLinearBufferCopy::Kind::Scatter) {
+      // Scatter from `src` to `dst`.
+      eigen_assert(src_stride == 1 && dst_stride != 1);
+      for (; i <= vectorized_size; i += PacketSize) {
+        Packet p = ploadu<Packet>(src + i);
+        pscatter<Scalar, Packet>(dst + i * dst_stride, p, dst_stride);
+      }
+      for (; i < count; ++i) {
+        dst[i * dst_stride] = src[i];
+      }
+      // ******************************************************************** //
+    } else if (kind == StridedLinearBufferCopy::Kind::FillLinear) {
+      // Fill `dst` with value at `*src`.
+      eigen_assert(src_stride == 0 && dst_stride == 1);
+      const IndexType unrolled_size = count - 4 * PacketSize;
+      Packet p = pload1<Packet>(src);
+      for (; i <= unrolled_size; i += 4 * PacketSize) {
+        for (int j = 0; j < 4; ++j) {
+          pstoreu<Scalar, Packet>(dst + i + j * PacketSize, p);
+        }
+      }
+      for (; i <= vectorized_size; i += PacketSize) {
+        pstoreu<Scalar, Packet>(dst + i, p);
+      }
+      for (; i < count; ++i) {
+        dst[i] = *src;
+      }
+      // ******************************************************************** //
+    } else if (kind == StridedLinearBufferCopy::Kind::FillScatter) {
+      // Scatter `*src` into `dst`.
+      eigen_assert(src_stride == 0 && dst_stride != 1);
+      Packet p = pload1<Packet>(src);
+      for (; i <= vectorized_size; i += PacketSize) {
+        pscatter<Scalar, Packet>(dst + i * dst_stride, p, dst_stride);
+      }
+      for (; i < count; ++i) {
+        dst[i * dst_stride] = *src;
+      }
+      // ******************************************************************** //
+    } else if (kind == StridedLinearBufferCopy::Kind::Gather) {
+      // Gather from `src` into `dst`.
+      eigen_assert(dst_stride == 1);
+      for (; i <= vectorized_size; i += PacketSize) {
+        Packet p = pgather<Scalar, Packet>(src + i * src_stride, src_stride);
+        pstoreu<Scalar, Packet>(dst + i, p);
+      }
+      for (; i < count; ++i) {
+        dst[i] = src[i * src_stride];
+      }
+      // ******************************************************************** //
+    } else if (kind == StridedLinearBufferCopy::Kind::Random) {
+      // Random.
+      for (; i < count; ++i) {
+        dst[i * dst_stride] = src[i * src_stride];
+      }
+    } else {
+      eigen_assert(false);
+    }
+  }
+};
+
+// -------------------------------------------------------------------------- //
+// TensorBlockIO copies data from `src` tensor block, to the `dst` tensor block.
+// It's possible to specify src->dst dimension mapping for the copy operation.
+// Dimensions of `dst` specify how many elements have to be copied, for the
+// `src` we need to know only stride to navigate through source memory buffer.
+
+template <typename Scalar, typename IndexType, int NumDims, int Layout>
+class TensorBlockIO {
+  static const bool IsColMajor = (Layout == ColMajor);
+
+  typedef StridedLinearBufferCopy<Scalar, IndexType> LinCopy;
+
+ public:
+  typedef DSizes<IndexType, NumDims> Dimensions;
+  typedef DSizes<int, NumDims> DimensionsMap;
+
+  struct Dst {
+    Dst(const Dimensions& dst_dims, const Dimensions& dst_strides, Scalar* dst,
+        IndexType dst_offset = 0)
+        : dims(dst_dims), strides(dst_strides), data(dst), offset(dst_offset) {}
+
+    Dimensions dims;
+    Dimensions strides;
+    Scalar* data;
+    IndexType offset;
+  };
+
+  struct Src {
+    Src(const Dimensions& src_strides, const Scalar* src,
+        IndexType src_offset = 0)
+        : strides(src_strides), data(src), offset(src_offset) {}
+
+    Dimensions strides;
+    const Scalar* data;
+    IndexType offset;
+  };
+
+  // Copies data to `dst` from `src`, using provided dimensions mapping:
+  //
+  //   src_dimension_index = dst_to_src_dim_map[dst_dimension_index]
+  //
+  // Returns the number of copied elements.
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE IndexType Copy(
+      const Dst& dst, const Src& src, const DimensionsMap& dst_to_src_dim_map) {
+    // Copy single scalar value from `src` to `dst`.
+    if (NumDims == 0) {
+      *(dst.data + dst.offset) = *(src.data + src.offset);
+      return 1;
+    }
+
+    // Both `dst` and `src` must have contiguous innermost dimension. We also
+    // accept the special case with stride '0', because it's used as a trick to
+    // implement broadcasting.
+    {
+      int inner_dim = IsColMajor ? 0 : NumDims - 1;
+      EIGEN_UNUSED_VARIABLE(inner_dim);
+      eigen_assert(dst.strides[inner_dim] == 1 || dst.strides[inner_dim] == 0);
+      eigen_assert(src.strides[inner_dim] == 1 || src.strides[inner_dim] == 0);
+    }
+
+    // Give a shorter name to `dst_to_src_dim_map`.
+    const DimensionsMap& dim_map = dst_to_src_dim_map;
+
+    // Do not squeeze reordered inner dimensions.
+    int num_squeezable_dims = NumSqueezableInnerDims(dim_map);
+
+    // NOTE: We find the innermost dimension (contiguous in memory) in the dst
+    // block, and we write data linearly into that dimension, reading it from
+    // the src. If dimensions are reordered, we might end up reading data from
+    // the src with `stride != 1`.
+    //
+    // NOTE: Random-Read/Linear-Write can be up to ~2X faster than
+    // Linear-Read/Random-Write: https://stackoverflow.com/a/54935680
+
+    // Find the innermost dimension in the dst whose size is not 1. This is the
+    // effective inner dim.
+    int num_size_one_inner_dims = 0;
+    for (int i = 0; i < num_squeezable_dims; ++i) {
+      const int dst_dim = IsColMajor ? i : NumDims - i - 1;
+      if (dst.dims[dst_dim] != 1) break;
+      num_size_one_inner_dims++;
+    }
+
+    // If all dimensions are of size 1, just copy a scalar from `src` to `dst`.
+    if (num_size_one_inner_dims == NumDims) {
+      *(dst.data + dst.offset) = *(src.data + src.offset);
+      return 1;
+    }
+
+    // Outermost dimension in the dst with `stride == 1` (contiguous in memory).
+    const int dst_stride1_dim = IsColMajor
+                                    ? num_size_one_inner_dims
+                                    : NumDims - num_size_one_inner_dims - 1;
+
+    // Dimension in the src that corresponds to the dst innermost dimension.
+    const int src_dim_for_dst_stride1_dim =
+        NumDims == 0 ? 1 : dim_map[dst_stride1_dim];
+
+    // Size of the innermost dimension (length of contiguous blocks of memory).
+    IndexType dst_inner_dim_size = NumDims == 0 ? 1 : dst.dims[dst_stride1_dim];
+
+    // Squeeze multiple inner dims into one if they are contiguous in `dst` and
+    // `src` memory, so we can do less linear copy calls.
+    for (int i = num_size_one_inner_dims + 1; i < num_squeezable_dims; ++i) {
+      const int dst_dim = IsColMajor ? i : NumDims - i - 1;
+      const IndexType dst_stride = dst.strides[dst_dim];
+      const IndexType src_stride = src.strides[dim_map[dst_dim]];
+      if (dst_inner_dim_size == dst_stride && dst_stride == src_stride) {
+        dst_inner_dim_size *= dst.dims[dst_dim];
+        ++num_size_one_inner_dims;
+      } else {
+        break;
+      }
+    }
+
+    // Setup strides to read data from `src` and write to `dst`.
+    IndexType input_offset = src.offset;
+    IndexType output_offset = dst.offset;
+    IndexType input_stride =
+        NumDims == 0 ? 1 : src.strides[src_dim_for_dst_stride1_dim];
+    IndexType output_stride = NumDims == 0 ? 1 : dst.strides[dst_stride1_dim];
+
+    const int at_least_1_dim = NumDims <= 1 ? 1 : NumDims - 1;
+    array<BlockIteratorState, at_least_1_dim> it;
+
+    // Initialize block iterator state. Squeeze away any dimension of size 1.
+    int idx = 0;  // currently initialized iterator state index
+    for (int i = num_size_one_inner_dims; i < NumDims - 1; ++i) {
+      const int dst_dim = IsColMajor ? i + 1 : NumDims - i - 2;
+      if (dst.dims[dst_dim] == 1) continue;
+
+      it[idx].size = dst.dims[dst_dim];
+      it[idx].input_stride = src.strides[dim_map[dst_dim]];
+      it[idx].output_stride = dst.strides[dst_dim];
+
+      it[idx].input_span = it[idx].input_stride * (it[idx].size - 1);
+      it[idx].output_span = it[idx].output_stride * (it[idx].size - 1);
+
+      idx++;
+    }
+
+    // Iterate copying data from src to dst.
+    const IndexType block_total_size = NumDims == 0 ? 1 : dst.dims.TotalSize();
+
+#define COPY_INNER_DIM(KIND)                                           \
+  IndexType num_copied = 0;                                            \
+  for (num_copied = 0; num_copied < block_total_size;                  \
+       num_copied += dst_inner_dim_size) {                             \
+    LinCopy::template Run<KIND>(                                       \
+        typename LinCopy::Dst(output_offset, output_stride, dst.data), \
+        typename LinCopy::Src(input_offset, input_stride, src.data),   \
+        dst_inner_dim_size);                                           \
+                                                                       \
+    for (int j = 0; j < idx; ++j) {                                    \
+      if (++it[j].count < it[j].size) {                                \
+        input_offset += it[j].input_stride;                            \
+        output_offset += it[j].output_stride;                          \
+        break;                                                         \
+      }                                                                \
+      it[j].count = 0;                                                 \
+      input_offset -= it[j].input_span;                                \
+      output_offset -= it[j].output_span;                              \
+    }                                                                  \
+  }                                                                    \
+  return num_copied;
+
+    if (input_stride == 1 && output_stride == 1) {
+      COPY_INNER_DIM(LinCopy::Kind::Linear);
+    } else if (input_stride == 1 && output_stride != 1) {
+      COPY_INNER_DIM(LinCopy::Kind::Scatter);
+    } else if (input_stride == 0 && output_stride == 1) {
+      COPY_INNER_DIM(LinCopy::Kind::FillLinear);
+    } else if (input_stride == 0 && output_stride != 1) {
+      COPY_INNER_DIM(LinCopy::Kind::FillScatter);
+    } else if (output_stride == 1) {
+      COPY_INNER_DIM(LinCopy::Kind::Gather);
+    } else {
+      COPY_INNER_DIM(LinCopy::Kind::Random);
+    }
+
+#undef COPY_INNER_DIM
+  }
+
+  // Copy from `src` to `dst` with an identity src->dst dimension map. Returns
+  // the number of copied elements.
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexType Copy(const Dst& dst,
+                                                              const Src& src) {
+    DimensionsMap dst_to_src_map;
+    for (int i = 0; i < NumDims; ++i) dst_to_src_map[i] = i;
+    return Copy(dst, src, dst_to_src_map);
+  }
+
+ private:
+  struct BlockIteratorState {
+    BlockIteratorState()
+        : size(0),
+          count(0),
+          input_stride(0),
+          output_stride(0),
+          input_span(0),
+          output_span(0) {}
+
+    IndexType size;
+    IndexType count;
+    IndexType input_stride;
+    IndexType output_stride;
+    IndexType input_span;
+    IndexType output_span;
+  };
+
+  // Compute how many inner dimensions it's allowed to squeeze when doing IO
+  // between two tensor blocks. It's safe to squeeze inner dimensions, only
+  // if they are not reordered.
+  static int NumSqueezableInnerDims(const DimensionsMap& dim_map) {
+    int num_squeezable_dims = 0;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+      if (dim_map[dim] != dim) break;
+      num_squeezable_dims++;
+    }
+    return num_squeezable_dims;
+  }
+};
+
+// -------------------------------------------------------------------------- //
+// TensorBlockAssignment assigns a block expression of type `TensorBlockExpr` to
+// a Tensor block defined by `desc`, backed by a memory buffer at `target`.
+//
+// Currently there is no way to write from a Tensor expression to a block of
+// memory, if dimensions are reordered. If you need to do that, you should
+// materialize a Tensor block expression into a memory buffer, and then use
+// TensorBlockIO to copy data between two memory buffers with a custom
+// `target->src` dimension map (see definition above).
+//
+// Also currently the innermost dimension of `target` must have a stride '1'
+// (contiguous in memory). This restriction could be lifted with a `pscatter`,
+// but in practice it's never needed, and there is a similar TensorBlockIO
+// workaround for that.
+//
+// TODO(ezhulenev): TensorBlockAssignment is a special case of TensorBlockIO
+// where `src` is a tensor expression. Explore if it is possible to rewrite IO
+// to use expressions instead of pointers, and after that TensorBlockAssignment
+// will become an alias to IO.
+template <typename Scalar, int NumDims, typename TensorBlockExpr,
+          typename IndexType = Eigen::Index>
+class TensorBlockAssignment {
+  // We will use coeff/packet path to evaluate block expressions.
+  typedef TensorEvaluator<const TensorBlockExpr, DefaultDevice>
+      TensorBlockEvaluator;
+
+  typedef DSizes<IndexType, NumDims> Dimensions;
+
+  enum {
+    Vectorizable = packet_traits<Scalar>::Vectorizable,
+    PacketSize = packet_traits<Scalar>::size
+  };
+
+  template <bool Vectorizable, typename Evaluator>
+  struct InnerDimAssign {
+    EIGEN_ALWAYS_INLINE static void Run(Scalar* target, IndexType count,
+                                        const Evaluator& eval,
+                                        IndexType eval_offset) {
+      for (IndexType i = 0; i < count; ++i) {
+        target[i] = eval.coeff(eval_offset + i);
+      }
+    }
+  };
+
+  template <typename Evaluator>
+  struct InnerDimAssign<true, Evaluator> {
+    EIGEN_ALWAYS_INLINE static void Run(Scalar* target, IndexType count,
+                                        const Evaluator& eval,
+                                        IndexType eval_offset) {
+      typedef typename packet_traits<Scalar>::type Packet;
+
+      const IndexType unrolled_size = count - 4 * PacketSize;
+      const IndexType vectorized_size = count - PacketSize;
+      IndexType i = 0;
+
+      for (; i <= unrolled_size; i += 4 * PacketSize) {
+        for (int j = 0; j < 4; ++j) {
+          const IndexType idx = eval_offset + i + j * PacketSize;
+          Packet p = eval.template packet<Unaligned>(idx);
+          pstoreu<Scalar>(target + i + j * PacketSize, p);
+        }
+      }
+
+      for (; i <= vectorized_size; i += PacketSize) {
+        Packet p = eval.template packet<Unaligned>(eval_offset + i);
+        pstoreu<Scalar>(target + i, p);
+      }
+
+      for (; i < count; ++i) {
+        target[i] = eval.coeff(eval_offset + i);
+      }
+    }
+  };
+
+ public:
+  struct Target {
+    Target(const Dimensions& target_dims, const Dimensions& target_strides,
+           Scalar* target_data, IndexType target_offset = 0)
+        : dims(target_dims),
+          strides(target_strides),
+          data(target_data),
+          offset(target_offset) {}
+
+    Dimensions dims;
+    Dimensions strides;
+    Scalar* data;
+    IndexType offset;
+  };
+
+  static Target target(const Dimensions& target_dims,
+                       const Dimensions& target_strides, Scalar* target_data,
+                       IndexType target_offset = 0) {
+    return Target(target_dims, target_strides, target_data, target_offset);
+  }
+
+  template <typename TargetDimsIndexType, typename TargetStridesIndexType>
+  static Target target(
+      const DSizes<TargetDimsIndexType, NumDims>& target_dims,
+      const DSizes<TargetStridesIndexType, NumDims>& target_strides,
+      Scalar* target_data, IndexType target_offset = 0) {
+    // DSizes constructor will do index type promotion if it's safe.
+    return Target(Dimensions(target_dims), Dimensions(target_strides),
+                  target_data, target_offset);
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Target& target, const TensorBlockExpr& expr) {
+    // Prepare evaluator for block expression.
+    DefaultDevice default_device;
+    TensorBlockEvaluator eval(expr, default_device);
+
+    // Tensor block expression dimension should match destination dimensions.
+    eigen_assert(dimensions_match(target.dims, eval.dimensions()));
+
+    static const int Layout = TensorBlockEvaluator::Layout;
+    static const bool is_col_major = Layout == ColMajor;
+
+    // Initialize output inner dimension size based on a layout.
+    const IndexType output_size = NumDims == 0 ? 1 : target.dims.TotalSize();
+    const int inner_dim_idx = is_col_major ? 0 : NumDims - 1;
+    IndexType output_inner_dim_size = target.dims[inner_dim_idx];
+
+    // Target inner dimension stride must be '1'.
+    eigen_assert(target.strides[inner_dim_idx] == 1);
+
+    // Squeeze multiple inner dims into one if they are contiguous in `target`.
+    IndexType num_squeezed_dims = 0;
+    for (Index i = 1; i < NumDims; ++i) {
+      const Index dim = is_col_major ? i : NumDims - i - 1;
+      const IndexType target_stride = target.strides[dim];
+
+      if (output_inner_dim_size == target_stride) {
+        output_inner_dim_size *= target.dims[dim];
+        num_squeezed_dims++;
+      } else {
+        break;
+      }
+    }
+
+    // Initialize output block iterator state. Dimension in this array are
+    // always in inner_most -> outer_most order (col major layout).
+    array<BlockIteratorState, NumDims> it;
+
+    int idx = 0;  // currently initialized iterator state index
+    for (Index i = num_squeezed_dims; i < NumDims - 1; ++i) {
+      const Index dim = is_col_major ? i + 1 : NumDims - i - 2;
+
+      it[idx].count = 0;
+      it[idx].size = target.dims[dim];
+      it[idx].output_stride = target.strides[dim];
+      it[idx].output_span = it[idx].output_stride * (it[idx].size - 1);
+      idx++;
+    }
+
+    // We read block expression from the beginning, and start writing data to
+    // `target` at given offset.
+    IndexType input_offset = 0;
+    IndexType output_offset = target.offset;
+
+    // Iterate copying data from `eval` to `target`.
+    for (IndexType i = 0; i < output_size; i += output_inner_dim_size) {
+      // Assign to `target` at current offset.
+      InnerDimAssign<Vectorizable && TensorBlockEvaluator::PacketAccess,
+                     TensorBlockEvaluator>::Run(target.data + output_offset,
+                                                output_inner_dim_size, eval,
+                                                input_offset);
+
+      // Move input offset forward by the number of assigned coefficients.
+      input_offset += output_inner_dim_size;
+
+      // Update index.
+      for (int j = 0; j < idx; ++j) {
+        if (++it[j].count < it[j].size) {
+          output_offset += it[j].output_stride;
+          break;
+        }
+        it[j].count = 0;
+        output_offset -= it[j].output_span;
+      }
+    }
+  }
+
+ private:
+  struct BlockIteratorState {
+    BlockIteratorState()
+        : count(0), size(0), output_stride(0), output_span(0) {}
+
+    IndexType count;
+    IndexType size;
+    IndexType output_stride;
+    IndexType output_span;
+  };
+};
+
+// -------------------------------------------------------------------------- //
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_BLOCK_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorBroadcasting.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorBroadcasting.h
new file mode 100644
index 0000000..a354132
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorBroadcasting.h
@@ -0,0 +1,1093 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H
+#define EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H
+
+namespace Eigen {
+
+/** \class TensorBroadcasting
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor broadcasting class.
+  *
+  *
+  */
+namespace internal {
+template<typename Broadcast, typename XprType>
+struct traits<TensorBroadcastingOp<Broadcast, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename Broadcast, typename XprType>
+struct eval<TensorBroadcastingOp<Broadcast, XprType>, Eigen::Dense>
+{
+  typedef const TensorBroadcastingOp<Broadcast, XprType> EIGEN_DEVICE_REF type;
+};
+
+template<typename Broadcast, typename XprType>
+struct nested<TensorBroadcastingOp<Broadcast, XprType>, 1, typename eval<TensorBroadcastingOp<Broadcast, XprType> >::type>
+{
+  typedef TensorBroadcastingOp<Broadcast, XprType> type;
+};
+
+template <typename Dims>
+struct is_input_scalar {
+  static const bool value = false;
+};
+template <>
+struct is_input_scalar<Sizes<> > {
+  static const bool value = true;
+};
+#ifndef EIGEN_EMULATE_CXX11_META_H
+template <typename std::ptrdiff_t... Indices>
+struct is_input_scalar<Sizes<Indices...> > {
+  static const bool value = (Sizes<Indices...>::total_size == 1);
+};
+#endif
+
+}  // end namespace internal
+
+
+
+template<typename Broadcast, typename XprType>
+class TensorBroadcastingOp : public TensorBase<TensorBroadcastingOp<Broadcast, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorBroadcastingOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorBroadcastingOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorBroadcastingOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorBroadcastingOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBroadcastingOp(const XprType& expr, const Broadcast& broadcast)
+      : m_xpr(expr), m_broadcast(broadcast) {}
+
+    EIGEN_DEVICE_FUNC
+    const Broadcast& broadcast() const { return m_broadcast; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Broadcast m_broadcast;
+};
+
+
+// Eval as rvalue
+template<typename Broadcast, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
+{
+  typedef TensorBroadcastingOp<Broadcast, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  protected: //  all the non-static fields must have the same access control, otherwise the TensorEvaluator wont be standard layout;
+  bool isCopy, nByOne, oneByN;
+  public:
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned         = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess         = false
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  // We do block based broadcasting using a trick with 2x tensor rank and 0
+  // strides. See block method implementation for details.
+  typedef DSizes<Index, 2 * NumDims> BroadcastDimensions;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+ typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename TensorEvaluator<const ArgType, Device>::TensorBlock
+      ArgTensorBlock;
+
+  typedef typename internal::TensorMaterializedBlock<ScalarNoConst, NumDims,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : isCopy(false), nByOne(false), oneByN(false),
+        m_device(device), m_broadcast(op.broadcast()), m_impl(op.expression(), device)
+  {
+
+    // The broadcasting op doesn't change the rank of the tensor. One can't broadcast a scalar
+    // and store the result in a scalar. Instead one should reshape the scalar into a a N-D
+    // tensor with N >= 1 of 1 element first and then broadcast.
+    EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    const InputDimensions& input_dims = m_impl.dimensions();
+    isCopy = true;
+    for (int i = 0; i < NumDims; ++i) {
+      eigen_assert(input_dims[i] > 0);
+      m_dimensions[i] = input_dims[i] * m_broadcast[i];
+      if (m_broadcast[i] != 1) {
+        isCopy = false;
+      }
+    }
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputStrides[0] = 1;
+      m_outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+        m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
+      }
+    } else {
+      m_inputStrides[NumDims-1] = 1;
+      m_outputStrides[NumDims-1] = 1;
+      for (int i = NumDims-2; i >= 0; --i) {
+        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
+        m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
+      }
+    }
+
+    if (input_dims[0] == 1) {
+      oneByN = true;
+      for (int i = 1; i < NumDims; ++i) {
+        if (m_broadcast[i] != 1) {
+          oneByN = false;
+          break;
+        }
+      }
+    } else if (input_dims[NumDims-1] == 1) {
+      nByOne = true;
+      for (int i = 0; i < NumDims-1; ++i) {
+        if (m_broadcast[i] != 1) {
+          nByOne = false;
+          break;
+        }
+      }
+    }
+
+    // Handle special format like NCHW, its input shape is '[1, N..., 1]' and
+    // broadcast shape is '[N, 1..., N]'
+    if (!oneByN && !nByOne) {
+      if (input_dims[0] == 1 && input_dims[NumDims-1] == 1 && NumDims > 2) {
+        nByOne = true;
+        oneByN = true;
+        for (int i = 1; i < NumDims-1; ++i) {
+          if (m_broadcast[i] != 1) {
+            nByOne = false;
+            oneByN = false;
+            break;
+          }
+        }
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffReturnType coeff(Index index) const
+  {
+    if (internal::is_input_scalar<typename internal::remove_all<InputDimensions>::type>::value) {
+      return m_impl.coeff(0);
+    }
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      if (isCopy) {
+        return m_impl.coeff(index);
+      } else {
+        return coeffColMajor(index);
+      }
+    } else {
+      if (isCopy) {
+        return m_impl.coeff(index);
+      } else {
+        return coeffRowMajor(index);
+      }
+    }
+  }
+
+  // TODO: attempt to speed this up. The integer divisions and modulo are slow
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index indexColMajor(Index index) const {
+    Index inputIndex = 0;
+    EIGEN_UNROLL_LOOP
+    for (int i = NumDims - 1; i > 0; --i) {
+      const Index idx = index / m_outputStrides[i];
+      if (internal::index_statically_eq<Broadcast>(i, 1)) {
+        eigen_assert(idx < m_impl.dimensions()[i]);
+        inputIndex += idx * m_inputStrides[i];
+      } else {
+        if (internal::index_statically_eq<InputDimensions>(i, 1)) {
+          eigen_assert(idx % m_impl.dimensions()[i] == 0);
+        } else {
+          inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
+        }
+      }
+      index -= idx * m_outputStrides[i];
+    }
+    if (internal::index_statically_eq<Broadcast>(0, 1)) {
+      eigen_assert(index < m_impl.dimensions()[0]);
+      inputIndex += index;
+    } else {
+      if (internal::index_statically_eq<InputDimensions>(0, 1)) {
+        eigen_assert(index % m_impl.dimensions()[0] == 0);
+      } else {
+        inputIndex += (index % m_impl.dimensions()[0]);
+      }
+    }
+    return inputIndex;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffColMajor(Index index) const
+  {
+    return m_impl.coeff(indexColMajor(index));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index indexRowMajor(Index index) const {
+    Index inputIndex = 0;
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < NumDims - 1; ++i) {
+      const Index idx = index / m_outputStrides[i];
+      if (internal::index_statically_eq<Broadcast>(i, 1)) {
+        eigen_assert(idx < m_impl.dimensions()[i]);
+        inputIndex += idx * m_inputStrides[i];
+      } else {
+        if (internal::index_statically_eq<InputDimensions>(i, 1)) {
+          eigen_assert(idx % m_impl.dimensions()[i] == 0);
+        } else {
+          inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
+        }
+      }
+      index -= idx * m_outputStrides[i];
+    }
+    if (internal::index_statically_eq<Broadcast>(NumDims - 1, 1)) {
+      eigen_assert(index < m_impl.dimensions()[NumDims - 1]);
+      inputIndex += index;
+    } else {
+      if (internal::index_statically_eq<InputDimensions>(NumDims - 1, 1)) {
+        eigen_assert(index % m_impl.dimensions()[NumDims - 1] == 0);
+      } else {
+        inputIndex += (index % m_impl.dimensions()[NumDims - 1]);
+      }
+    }
+    return inputIndex;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffRowMajor(Index index) const
+  {
+    return m_impl.coeff(indexRowMajor(index));
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketReturnType packet(Index index) const
+  {
+    if (internal::is_input_scalar<typename internal::remove_all<InputDimensions>::type>::value) {
+      return internal::pset1<PacketReturnType>(m_impl.coeff(0));
+    }
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      if (isCopy) {
+        #ifdef EIGEN_GPU_COMPILE_PHASE
+        // See PR 437: on NVIDIA P100 and K20m we observed a x3-4 speed up by enforcing
+        // unaligned loads here. The reason is unclear though.
+        return m_impl.template packet<Unaligned>(index);
+        #else
+        return m_impl.template packet<LoadMode>(index);
+        #endif
+      } else if (oneByN && !nByOne) {
+        return packetNByOne<LoadMode>(index);
+      } else if (!oneByN && nByOne) {
+        return packetOneByN<LoadMode>(index);
+      } else if (oneByN && nByOne) {
+        return packetOneByNByOne<LoadMode>(index);
+      } else {
+        return packetColMajor<LoadMode>(index);
+      }
+    } else {
+      if (isCopy) {
+        #ifdef EIGEN_GPU_COMPILE_PHASE
+        // See above.
+        return m_impl.template packet<Unaligned>(index);
+        #else
+        return m_impl.template packet<LoadMode>(index);
+        #endif
+      } else if (oneByN && !nByOne) {
+        return packetOneByN<LoadMode>(index);
+      } else if (!oneByN && nByOne) {
+        return packetNByOne<LoadMode>(index);
+      } else if (oneByN && nByOne) {
+        return packetOneByNByOne<LoadMode>(index);
+      } else {
+        return packetRowMajor<LoadMode>(index);
+      }
+    }
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetOneByNByOne
+  (Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    Index startDim, endDim;
+    Index inputIndex, outputOffset, batchedIndex;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      startDim = NumDims - 1;
+      endDim = 1;
+    } else {
+      startDim = 0;
+      endDim = NumDims - 2;
+    }
+
+    batchedIndex = index % m_outputStrides[startDim];
+    inputIndex   = batchedIndex / m_outputStrides[endDim];
+    outputOffset = batchedIndex % m_outputStrides[endDim];
+
+    if (outputOffset + PacketSize <= m_outputStrides[endDim]) {
+      values[0] = m_impl.coeff(inputIndex);
+      return internal::pload1<PacketReturnType>(values);
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0, cur = 0; i < PacketSize; ++i, ++cur) {
+        if (outputOffset + cur < m_outputStrides[endDim]) {
+          values[i] = m_impl.coeff(inputIndex);
+        } else {
+          ++inputIndex;
+          inputIndex = (inputIndex == m_inputStrides[startDim] ? 0 : inputIndex);
+          values[i] = m_impl.coeff(inputIndex);
+          outputOffset = 0;
+          cur = 0;
+        }
+      }
+      return internal::pload<PacketReturnType>(values);
+    }
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetOneByN(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    Index dim, inputIndex;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      dim = NumDims - 1;
+    } else {
+      dim = 0;
+    }
+
+    inputIndex = index % m_inputStrides[dim];
+    if (inputIndex + PacketSize <= m_inputStrides[dim]) {
+      return m_impl.template packet<Unaligned>(inputIndex);
+    } else {
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < PacketSize; ++i) {
+        if (inputIndex > m_inputStrides[dim]-1) {
+          inputIndex = 0;
+        }
+        values[i] = m_impl.coeff(inputIndex++);
+      }
+      return internal::pload<PacketReturnType>(values);
+    }
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetNByOne(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    Index dim, inputIndex, outputOffset;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      dim = 1;
+    } else {
+      dim = NumDims - 2;
+    }
+
+    inputIndex   = index / m_outputStrides[dim];
+    outputOffset = index % m_outputStrides[dim];
+    if (outputOffset + PacketSize <= m_outputStrides[dim]) {
+      values[0] = m_impl.coeff(inputIndex);
+      return internal::pload1<PacketReturnType>(values);
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0, cur = 0; i < PacketSize; ++i, ++cur) {
+        if (outputOffset + cur < m_outputStrides[dim]) {
+          values[i] = m_impl.coeff(inputIndex);
+        } else {
+          values[i] = m_impl.coeff(++inputIndex);
+          outputOffset = 0;
+          cur = 0;
+        }
+      }
+      return internal::pload<PacketReturnType>(values);
+    }
+  }
+
+  // Ignore the LoadMode and always use unaligned loads since we can't guarantee
+  // the alignment at compile time.
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetColMajor(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    const Index originalIndex = index;
+
+    Index inputIndex = 0;
+    EIGEN_UNROLL_LOOP
+    for (int i = NumDims - 1; i > 0; --i) {
+      const Index idx = index / m_outputStrides[i];
+      if (internal::index_statically_eq<Broadcast>(i, 1)) {
+        eigen_assert(idx < m_impl.dimensions()[i]);
+        inputIndex += idx * m_inputStrides[i];
+      } else {
+        if (internal::index_statically_eq<InputDimensions>(i, 1)) {
+          eigen_assert(idx % m_impl.dimensions()[i] == 0);
+        } else {
+          inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
+        }
+      }
+      index -= idx * m_outputStrides[i];
+    }
+    Index innermostLoc;
+    if (internal::index_statically_eq<Broadcast>(0, 1)) {
+      eigen_assert(index < m_impl.dimensions()[0]);
+      innermostLoc = index;
+    } else {
+      if (internal::index_statically_eq<InputDimensions>(0, 1)) {
+        eigen_assert(index % m_impl.dimensions()[0] == 0);
+        innermostLoc = 0;
+      } else {
+        innermostLoc = index % m_impl.dimensions()[0];
+      }
+    }
+    inputIndex += innermostLoc;
+
+    // Todo: this could be extended to the second dimension if we're not
+    // broadcasting alongside the first dimension, and so on.
+    if (innermostLoc + PacketSize <= m_impl.dimensions()[0]) {
+      return m_impl.template packet<Unaligned>(inputIndex);
+    } else {
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      values[0] = m_impl.coeff(inputIndex);
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < PacketSize; ++i) {
+        if (innermostLoc + i < m_impl.dimensions()[0]) {
+          values[i] = m_impl.coeff(inputIndex+i);
+        } else {
+          values[i] = coeffColMajor(originalIndex+i);
+        }
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    }
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetRowMajor(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    const Index originalIndex = index;
+
+    Index inputIndex = 0;
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < NumDims - 1; ++i) {
+      const Index idx = index / m_outputStrides[i];
+      if (internal::index_statically_eq<Broadcast>(i, 1)) {
+        eigen_assert(idx < m_impl.dimensions()[i]);
+        inputIndex += idx * m_inputStrides[i];
+      } else {
+        if (internal::index_statically_eq<InputDimensions>(i, 1)) {
+          eigen_assert(idx % m_impl.dimensions()[i] == 0);
+        } else {
+          inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
+        }
+      }
+      index -= idx * m_outputStrides[i];
+    }
+    Index innermostLoc;
+    if (internal::index_statically_eq<Broadcast>(NumDims-1, 1)) {
+      eigen_assert(index < m_impl.dimensions()[NumDims-1]);
+      innermostLoc = index;
+    } else {
+      if (internal::index_statically_eq<InputDimensions>(NumDims-1, 1)) {
+        eigen_assert(index % m_impl.dimensions()[NumDims-1] == 0);
+        innermostLoc = 0;
+      } else {
+        innermostLoc = index % m_impl.dimensions()[NumDims-1];
+      }
+    }
+    inputIndex += innermostLoc;
+
+    // Todo: this could be extended to the second dimension if we're not
+    // broadcasting alongside the first dimension, and so on.
+    if (innermostLoc + PacketSize <= m_impl.dimensions()[NumDims-1]) {
+      return m_impl.template packet<Unaligned>(inputIndex);
+    } else {
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      values[0] = m_impl.coeff(inputIndex);
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < PacketSize; ++i) {
+        if (innermostLoc + i < m_impl.dimensions()[NumDims-1]) {
+          values[i] = m_impl.coeff(inputIndex+i);
+        } else {
+          values[i] = coeffRowMajor(originalIndex+i);
+        }
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    double compute_cost = TensorOpCost::AddCost<Index>();
+    if (!isCopy && NumDims > 0) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        compute_cost += TensorOpCost::DivCost<Index>();
+        if (internal::index_statically_eq<Broadcast>(i, 1)) {
+          compute_cost +=
+              TensorOpCost::MulCost<Index>() + TensorOpCost::AddCost<Index>();
+        } else {
+          if (!internal::index_statically_eq<InputDimensions>(i, 1)) {
+            compute_cost += TensorOpCost::MulCost<Index>() +
+                            TensorOpCost::ModCost<Index>() +
+                            TensorOpCost::AddCost<Index>();
+          }
+        }
+        compute_cost +=
+            TensorOpCost::MulCost<Index>() + TensorOpCost::AddCost<Index>();
+      }
+    }
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    // TODO(wuke): Targeting L1 size is 30% faster than targeting L{-1} on large
+    // tensors. But this might need further tuning.
+    const size_t target_size = m_device.firstLevelCacheSize();
+    return internal::TensorBlockResourceRequirements::merge(
+        m_impl.getResourceRequirements(),
+        internal::TensorBlockResourceRequirements::skewed<Scalar>(target_size));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    BlockBroadcastingParams params = blockBroadcastingParams(desc);
+
+    if (params.inner_dim_size == 0 || params.bcast_dim_size == 0) {
+      return emptyBlock();
+    }
+
+    // Prepare storage for the materialized broadcasting result.
+    const typename TensorBlock::Storage block_storage =
+        TensorBlock::prepareStorage(desc, scratch);
+    ScalarNoConst* materialized_output = block_storage.data();
+
+    // We potentially will need to materialize input blocks.
+    size_t materialized_input_size = 0;
+    ScalarNoConst* materialized_input = NULL;
+
+    // Initialize block broadcating iterator state for outer dimensions (outer
+    // with regard to bcast dimension). Dimension in this array are always in
+    // inner_most -> outer_most order (col major layout).
+    array<BlockBroadcastingIteratorState, NumDims> it;
+    int idx = 0;
+
+    for (int i = params.inner_dim_count + 1; i < NumDims; ++i) {
+      const Index dim = IsColMajor ? i : NumDims - 1 - i;
+      it[idx].size = params.output_dims[dim];
+      it[idx].count = 0;
+      it[idx].output_stride = m_outputStrides[dim];
+      it[idx].output_span = it[idx].output_stride * (it[idx].size - 1);
+      idx++;
+    }
+
+    // Write output into the beginning of `materialized_output`.
+    Index output_offset = 0;
+
+    // We will fill output block by broadcasting along the bcast dim, and
+    // iterating over outer dimension.
+    const Index output_size = NumDims == 0 ? 1 : params.output_dims.TotalSize();
+
+    for (Index num_output_coeffs = 0; num_output_coeffs < output_size;) {
+      ScalarNoConst* bcast_output = materialized_output + num_output_coeffs;
+      Index bcast_offset = desc.offset() + output_offset;
+
+      // Broadcast along the bcast dimension.
+      num_output_coeffs += BroadcastBlockAlongBcastDim(
+          params, bcast_offset, scratch, bcast_output, &materialized_input,
+          &materialized_input_size);
+
+      // Switch to the next outer dimension.
+      for (int j = 0; j < idx; ++j) {
+        if (++it[j].count < it[j].size) {
+          output_offset += it[j].output_stride;
+          break;
+        }
+        it[j].count = 0;
+        output_offset -= it[j].output_span;
+      }
+    }
+
+    return block_storage.AsTensorMaterializedBlock();
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+
+  Broadcast functor() const { return m_broadcast; }
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(
+      cl::sycl::handler& cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+ private:
+  static const bool IsColMajor =
+      static_cast<int>(Layout) == static_cast<int>(ColMajor);
+
+  // We will build a general case block broadcasting on top of broadcasting
+  // primitive that will do broadcasting only for the inner dimension(s) along
+  // the first dimension smaller than the input size (it's called `bcast_dim`).
+  //
+  // Example:
+  //           dim:  0  1  2   (ColMajor)
+  //    input size: [9, 3, 6]
+  //    block size: [9, 2, 6]
+  //
+  // We will compute broadcasted block by iterating over the outer dimensions
+  // before `bcast_dim` (only dimension `2` in this example) and computing
+  // broadcasts along the `bcast_dim` (dimension `1` in this example).
+
+  // BlockBroadcastingParams holds precomputed parameters for broadcasting a
+  // single block along the broadcasting dimension. Sizes and strides along the
+  // `bcast_dim` might be invalid, they will be adjusted later in
+  // `BroadcastBlockAlongBcastDim`.
+  struct BlockBroadcastingParams {
+    Dimensions input_dims;      // input expression dimensions
+    Dimensions output_dims;     // output block sizes
+    Dimensions output_strides;  // output block strides
+
+    int inner_dim_count;   // count inner dimensions matching in size
+    int bcast_dim;         // broadcasting dimension index
+    Index bcast_dim_size;  // broadcasting dimension size
+    Index inner_dim_size;  // inner dimensions size
+
+    // Block sizes and strides for the input block where all dimensions before
+    // `bcast_dim` are equal to `1`.
+    Dimensions input_block_sizes;
+    Dimensions input_block_strides;
+
+    // Block sizes and strides for blocks with extra dimensions and strides `0`.
+    BroadcastDimensions bcast_block_sizes;
+    BroadcastDimensions bcast_block_strides;
+    BroadcastDimensions bcast_input_strides;
+  };
+
+  struct BlockBroadcastingIteratorState {
+    Index size;
+    Index count;
+    Index output_stride;
+    Index output_span;
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlockBroadcastingParams
+  blockBroadcastingParams(TensorBlockDesc& desc) const {
+    BlockBroadcastingParams params;
+
+    params.input_dims = Dimensions(m_impl.dimensions());
+
+    // Output block sizes and strides.
+    params.output_dims = desc.dimensions();
+    params.output_strides = internal::strides<Layout>(params.output_dims);
+
+    // Find the broadcasting dimension (first dimension with output size smaller
+    // that the input size).
+    params.bcast_dim = 0;
+    params.bcast_dim_size = 1;
+    params.inner_dim_size = 1;
+
+    // Count the number of inner dimensions that have the same size in the block
+    // and in the broadcast expression.
+    params.inner_dim_count = 0;
+
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+
+      if (params.output_dims[dim] == m_dimensions[dim]) {
+        params.inner_dim_size *= params.output_dims[dim];
+        ++params.inner_dim_count;
+        continue;
+      }
+
+      // First non-matching dimension is the broadcasting dimension.
+      eigen_assert(params.output_dims[dim] < m_dimensions[dim]);
+      params.bcast_dim = dim;
+      params.bcast_dim_size = params.output_dims[dim];
+      break;
+    }
+
+    // Calculate the input block size for looking into the input.
+    for (int i = 0; i < params.inner_dim_count; ++i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+      params.input_block_sizes[dim] = params.input_dims[dim];
+    }
+    for (int i = params.inner_dim_count; i < NumDims; ++i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+      params.input_block_sizes[dim] = 1;
+    }
+    params.input_block_strides =
+        internal::strides<Layout>(params.input_block_sizes);
+
+    // Broadcast with the 0-stride trick: Create 1 extra dim for each
+    // broadcast, set the input stride to 0.
+    //
+    // When ColMajor:
+    //
+    // - bcast_block_sizes:
+    //   [d_0, b_0, d_1, b_1, ...]
+    //
+    // - bcast_block_strides:
+    //   [output_block_strides[0], output_block_strides[0] * d_0,
+    //    output_block_strides[1], output_block_strides[1] * d_1,
+    //   ...]
+    //
+    // - bcast_input_strides:
+    //   [input_block_strides[0], 0,
+    //    input_block_strides[1], 0,
+    //   ...].
+    //
+    for (int i = 0; i < params.inner_dim_count; ++i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+
+      const int copy_dim = IsColMajor ? 2 * i : 2 * NumDims - 2 * i - 1;
+      const int broadcast_dim = IsColMajor ? copy_dim + 1 : copy_dim - 1;
+
+      params.bcast_block_sizes[copy_dim] = params.input_dims[dim];
+      params.bcast_block_sizes[broadcast_dim] = m_broadcast[dim];
+      params.bcast_block_strides[copy_dim] = params.output_strides[dim];
+      params.bcast_block_strides[broadcast_dim] =
+          params.output_strides[dim] * params.input_dims[dim];
+      params.bcast_input_strides[copy_dim] = params.input_block_strides[dim];
+      params.bcast_input_strides[broadcast_dim] = 0;
+    }
+
+    for (int i = 2 * params.inner_dim_count; i < 2 * NumDims; ++i) {
+      const int dim = IsColMajor ? i : 2 * NumDims - i - 1;
+      params.bcast_block_sizes[dim] = 1;
+      params.bcast_block_strides[dim] = 0;
+      params.bcast_input_strides[dim] = 0;
+    }
+
+    return params;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock emptyBlock() const {
+    DSizes<Index, NumDims> dimensions;
+    for (int i = 0; i < NumDims; ++i) dimensions[i] = 0;
+    return TensorBlock(internal::TensorBlockKind::kView, NULL, dimensions);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index BroadcastBlockAlongBcastDim(
+      BlockBroadcastingParams params, Index bcast_offset,
+      TensorBlockScratch& scratch, ScalarNoConst* materialized_output,
+      ScalarNoConst** materialized_input,
+      size_t* materialized_input_size) const {
+    if (params.bcast_dim_size == 1) {
+      // We just need one block read using the ready-set values above.
+      return BroadcastBlock(
+          params.input_block_sizes, params.input_block_strides,
+          params.bcast_block_sizes, params.bcast_block_strides,
+          params.bcast_input_strides, bcast_offset, 0, scratch,
+          materialized_output, materialized_input, materialized_input_size);
+
+    } else if (params.input_dims[params.bcast_dim] == 1) {
+      // Broadcast bcast dimension (< NumDims) by bcast_dim_size.
+      const int broadcast_bcast_dim =
+          IsColMajor ? 2 * params.inner_dim_count + 1
+                     : 2 * NumDims - 2 * params.inner_dim_count - 2;
+
+      params.bcast_block_sizes[broadcast_bcast_dim] = params.bcast_dim_size;
+      params.bcast_input_strides[broadcast_bcast_dim] = 0;
+      params.bcast_block_strides[broadcast_bcast_dim] =
+          params.output_strides[params.bcast_dim];
+
+      return BroadcastBlock(
+          params.input_block_sizes, params.input_block_strides,
+          params.bcast_block_sizes, params.bcast_block_strides,
+          params.bcast_input_strides, bcast_offset, 0, scratch,
+          materialized_output, materialized_input, materialized_input_size);
+
+    } else {
+      // Keep track of the total number of the coefficients written to the
+      // output block.
+      Index num_output_coeffs = 0;
+
+      // The general case. Let's denote the output block as
+      //
+      //   x[..., a:a+bcast_dim_size, :, ..., :]
+      //
+      // where a:a+bcast_dim_size is a slice on the bcast_dim dimension
+      // (< NumDims). We need to split the a:a+bcast_dim_size into possibly 3
+      // sub-blocks:
+      //
+      // (1) a:b, where b is the smallest multiple of
+      //     input_dims[bcast_dim_start] in [a, a+bcast_dim_size].
+      //
+      // (2) b:c, where c is the largest multiple of input_dims[bcast_dim_start]
+      //     in [a, a+bcast_dim_size].
+      //
+      // (3) c:a+bcast_dim_size .
+      //
+      // Or, when b and c do not exist, we just need to process the whole block
+      // together.
+
+      // Find a.
+      const Index bcast_dim_left_index =
+          bcast_offset / m_outputStrides[params.bcast_dim];
+
+      // Find b and c.
+      const Index input_bcast_dim_size = params.input_dims[params.bcast_dim];
+
+      // First multiple after a. This is b when <= bcast_dim_left_index +
+      // bcast_dim_size.
+      const Index first_multiple =
+          divup<Index>(bcast_dim_left_index, input_bcast_dim_size) *
+          input_bcast_dim_size;
+
+      if (first_multiple <= bcast_dim_left_index + params.bcast_dim_size) {
+        // b exists, so does c. Find it.
+        const Index last_multiple =
+            (bcast_dim_left_index + params.bcast_dim_size) /
+            input_bcast_dim_size * input_bcast_dim_size;
+        const int copy_bcast_dim =
+            IsColMajor ? 2 * params.inner_dim_count
+                       : 2 * NumDims - 2 * params.inner_dim_count - 1;
+        const int broadcast_bcast_dim =
+            IsColMajor ? 2 * params.inner_dim_count + 1
+                       : 2 * NumDims - 2 * params.inner_dim_count - 2;
+
+        if (first_multiple > bcast_dim_left_index) {
+          const Index head_size = first_multiple - bcast_dim_left_index;
+          params.input_block_sizes[params.bcast_dim] = head_size;
+          params.bcast_block_sizes[copy_bcast_dim] = head_size;
+          params.bcast_input_strides[copy_bcast_dim] =
+              params.input_block_strides[params.bcast_dim];
+          params.bcast_block_strides[copy_bcast_dim] =
+              params.output_strides[params.bcast_dim];
+          params.bcast_block_sizes[broadcast_bcast_dim] = 1;
+          params.bcast_input_strides[broadcast_bcast_dim] = 0;
+          params.bcast_block_strides[broadcast_bcast_dim] =
+              params.output_strides[params.bcast_dim] *
+              params.input_dims[params.bcast_dim];
+
+          num_output_coeffs += BroadcastBlock(
+              params.input_block_sizes, params.input_block_strides,
+              params.bcast_block_sizes, params.bcast_block_strides,
+              params.bcast_input_strides, bcast_offset, 0, scratch,
+              materialized_output, materialized_input, materialized_input_size);
+        }
+        if (first_multiple < last_multiple) {
+          params.input_block_sizes[params.bcast_dim] = input_bcast_dim_size;
+          params.bcast_block_sizes[copy_bcast_dim] = input_bcast_dim_size;
+          params.bcast_input_strides[copy_bcast_dim] =
+              params.input_block_strides[params.bcast_dim];
+          params.bcast_block_strides[copy_bcast_dim] =
+              params.output_strides[params.bcast_dim];
+          params.bcast_block_sizes[broadcast_bcast_dim] =
+              (last_multiple - first_multiple) / input_bcast_dim_size;
+          params.bcast_input_strides[broadcast_bcast_dim] = 0;
+          params.bcast_block_strides[broadcast_bcast_dim] =
+              params.output_strides[params.bcast_dim] *
+              params.input_dims[params.bcast_dim];
+          const Index offset = (first_multiple - bcast_dim_left_index) *
+                               m_outputStrides[params.bcast_dim];
+
+          num_output_coeffs += BroadcastBlock(
+              params.input_block_sizes, params.input_block_strides,
+              params.bcast_block_sizes, params.bcast_block_strides,
+              params.bcast_input_strides, bcast_offset, offset, scratch,
+              materialized_output, materialized_input, materialized_input_size);
+        }
+        if (last_multiple < bcast_dim_left_index + params.bcast_dim_size) {
+          const Index tail_size =
+              bcast_dim_left_index + params.bcast_dim_size - last_multiple;
+          params.input_block_sizes[params.bcast_dim] = tail_size;
+          params.bcast_block_sizes[copy_bcast_dim] = tail_size;
+          params.bcast_input_strides[copy_bcast_dim] =
+              params.input_block_strides[params.bcast_dim];
+          params.bcast_block_strides[copy_bcast_dim] =
+              params.output_strides[params.bcast_dim];
+          params.bcast_block_sizes[broadcast_bcast_dim] = 1;
+          params.bcast_input_strides[broadcast_bcast_dim] = 0;
+          params.bcast_block_strides[broadcast_bcast_dim] =
+              params.output_strides[params.bcast_dim] *
+              params.input_dims[params.bcast_dim];
+          const Index offset = (last_multiple - bcast_dim_left_index) *
+                               m_outputStrides[params.bcast_dim];
+
+          num_output_coeffs += BroadcastBlock(
+              params.input_block_sizes, params.input_block_strides,
+              params.bcast_block_sizes, params.bcast_block_strides,
+              params.bcast_input_strides, bcast_offset, offset, scratch,
+              materialized_output, materialized_input, materialized_input_size);
+        }
+      } else {
+        // b and c do not exist.
+        const int copy_bcast_dim =
+            IsColMajor ? 2 * params.inner_dim_count
+                       : 2 * NumDims - 2 * params.inner_dim_count - 1;
+        params.input_block_sizes[params.bcast_dim] = params.bcast_dim_size;
+        params.bcast_block_sizes[copy_bcast_dim] = params.bcast_dim_size;
+        params.bcast_input_strides[copy_bcast_dim] =
+            params.input_block_strides[params.bcast_dim];
+        params.bcast_block_strides[copy_bcast_dim] =
+            params.output_strides[params.bcast_dim];
+
+        num_output_coeffs += BroadcastBlock(
+            params.input_block_sizes, params.input_block_strides,
+            params.bcast_block_sizes, params.bcast_block_strides,
+            params.bcast_input_strides, bcast_offset, 0, scratch,
+            materialized_output, materialized_input, materialized_input_size);
+      }
+
+      return num_output_coeffs;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index BroadcastBlock(
+      const Dimensions& input_block_sizes,
+      const Dimensions& input_block_strides,
+      const BroadcastDimensions& bcast_block_sizes,
+      const BroadcastDimensions& bcast_block_strides,
+      const BroadcastDimensions& bcast_input_strides, Index bcast_offset,
+      Index offset, TensorBlockScratch& scratch,
+      ScalarNoConst* materialized_output, ScalarNoConst** materialized_input,
+      size_t* materialized_input_size) const {
+    // ---------------------------------------------------------------------- //
+    // Tensor block descriptor for reading block from the input.
+    const Index input_offset = bcast_offset + offset;
+    TensorBlockDesc input_desc(
+        IsColMajor ? indexColMajor(input_offset) : indexRowMajor(input_offset),
+        input_block_sizes);
+
+    ArgTensorBlock input_block = m_impl.block(input_desc, scratch);
+
+    // ---------------------------------------------------------------------- //
+    // Materialize input block into a temporary memory buffer only if it's not
+    // already available in the arg block.
+    const ScalarNoConst* input_buffer = NULL;
+
+    if (input_block.data() != NULL) {
+      // Input block already has raw data, there is no need to materialize it.
+      input_buffer = input_block.data();
+
+    } else {
+      // Otherwise we have to do block assignment into a temporary buffer.
+
+      // Maybe reuse previously allocated buffer, or allocate a new one with a
+      // scratch allocator.
+      const size_t input_total_size = input_block_sizes.TotalSize();
+      if (*materialized_input == NULL ||
+          *materialized_input_size < input_total_size) {
+        *materialized_input_size = input_total_size;
+        void* mem = scratch.allocate(*materialized_input_size * sizeof(Scalar));
+        *materialized_input = static_cast<ScalarNoConst*>(mem);
+      }
+
+      typedef internal::TensorBlockAssignment<
+          ScalarNoConst, NumDims, typename ArgTensorBlock::XprType, Index>
+          TensorBlockAssignment;
+
+      TensorBlockAssignment::Run(
+          TensorBlockAssignment::target(input_block_sizes, input_block_strides,
+                                        *materialized_input),
+          input_block.expr());
+
+      input_buffer = *materialized_input;
+    }
+
+    // ---------------------------------------------------------------------- //
+    // Copy data from materialized input block to the materialized output, using
+    // given broadcast strides (strides with zeroes).
+    typedef internal::TensorBlockIO<ScalarNoConst, Index, 2 * NumDims, Layout>
+        TensorBlockIO;
+
+    typename TensorBlockIO::Src src(bcast_input_strides, input_buffer);
+    typename TensorBlockIO::Dst dst(bcast_block_sizes, bcast_block_strides,
+                                      materialized_output + offset);
+
+    return TensorBlockIO::Copy(dst, src);
+  }
+
+protected:
+  const Device EIGEN_DEVICE_REF m_device;
+  const typename internal::remove_reference<Broadcast>::type m_broadcast;
+  Dimensions m_dimensions;
+  array<Index, NumDims> m_outputStrides;
+  array<Index, NumDims> m_inputStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorChipping.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorChipping.h
new file mode 100644
index 0000000..3764573
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorChipping.h
@@ -0,0 +1,518 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H
+
+namespace Eigen {
+
+/** \class TensorKChippingReshaping
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief A chip is a thin slice, corresponding to a column or a row in a 2-d tensor.
+  *
+  *
+  */
+
+namespace internal {
+template<DenseIndex DimId, typename XprType>
+struct traits<TensorChippingOp<DimId, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions - 1;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<DenseIndex DimId, typename XprType>
+struct eval<TensorChippingOp<DimId, XprType>, Eigen::Dense>
+{
+  typedef const TensorChippingOp<DimId, XprType> EIGEN_DEVICE_REF type;
+};
+
+template<DenseIndex DimId, typename XprType>
+struct nested<TensorChippingOp<DimId, XprType>, 1, typename eval<TensorChippingOp<DimId, XprType> >::type>
+{
+  typedef TensorChippingOp<DimId, XprType> type;
+};
+
+template <DenseIndex DimId>
+struct DimensionId
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DimensionId(DenseIndex dim) {
+    EIGEN_UNUSED_VARIABLE(dim);
+    eigen_assert(dim == DimId);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim() const {
+    return DimId;
+  }
+};
+template <>
+struct DimensionId<Dynamic>
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DimensionId(DenseIndex dim) : actual_dim(dim) {
+    eigen_assert(dim >= 0);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim() const {
+    return actual_dim;
+  }
+ private:
+  const DenseIndex actual_dim;
+};
+
+
+}  // end namespace internal
+
+
+
+template<DenseIndex DimId, typename XprType>
+class TensorChippingOp : public TensorBase<TensorChippingOp<DimId, XprType> >
+{
+  public:
+    typedef TensorBase<TensorChippingOp<DimId, XprType> > Base;
+    typedef typename Eigen::internal::traits<TensorChippingOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename XprType::CoeffReturnType CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorChippingOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorChippingOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorChippingOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp(const XprType& expr, const Index offset, const Index dim)
+        : m_xpr(expr), m_offset(offset), m_dim(dim) {
+    }
+
+    EIGEN_DEVICE_FUNC
+    const Index offset() const { return m_offset; }
+    EIGEN_DEVICE_FUNC
+    const Index dim() const { return m_dim.actualDim(); }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorChippingOp)
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Index m_offset;
+    const internal::DimensionId<DimId> m_dim;
+};
+
+
+// Eval as rvalue
+template<DenseIndex DimId, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
+{
+  typedef TensorChippingOp<DimId, ArgType> XprType;
+  static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  static const int NumDims = NumInputDims-1;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    // Alignment can't be guaranteed at compile time since it depends on the
+    // slice offsets.
+    IsAligned         = false,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    // Chipping of outer-most dimension is a trivial operation, because we can
+    // read and write directly from the underlying tensor using single offset.
+    IsOuterChipping   = (static_cast<int>(Layout) == ColMajor && DimId == NumInputDims - 1) ||
+                        (static_cast<int>(Layout) == RowMajor && DimId == 0),
+    // Chipping inner-most dimension.
+    IsInnerChipping   = (static_cast<int>(Layout) == ColMajor && DimId == 0) ||
+                        (static_cast<int>(Layout) == RowMajor && DimId == NumInputDims - 1),
+    // Prefer block access if the underlying expression prefers it, otherwise
+    // only if chipping is not trivial.
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess ||
+                        !IsOuterChipping,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef internal::TensorBlockDescriptor<NumInputDims, Index>
+      ArgTensorBlockDesc;
+  typedef typename TensorEvaluator<const ArgType, Device>::TensorBlock
+      ArgTensorBlock;
+
+  typedef typename internal::TensorMaterializedBlock<ScalarNoConst, NumDims,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_dim(op.dim()), m_device(device)
+  {
+    EIGEN_STATIC_ASSERT((NumInputDims >= 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    eigen_assert(NumInputDims > m_dim.actualDim());
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    eigen_assert(op.offset() < input_dims[m_dim.actualDim()]);
+
+    int j = 0;
+    for (int i = 0; i < NumInputDims; ++i) {
+      if (i != m_dim.actualDim()) {
+        m_dimensions[j] = input_dims[i];
+        ++j;
+      }
+    }
+
+    m_stride = 1;
+    m_inputStride = 1;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < m_dim.actualDim(); ++i) {
+        m_stride *= input_dims[i];
+        m_inputStride *= input_dims[i];
+      }
+    } else {
+      for (int i = NumInputDims-1; i > m_dim.actualDim(); --i) {
+        m_stride *= input_dims[i];
+        m_inputStride *= input_dims[i];
+      }
+    }
+    m_inputStride *= input_dims[m_dim.actualDim()];
+    m_inputOffset = m_stride * op.offset();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return m_impl.coeff(srcCoeff(index));
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    if (isInnerChipping()) {
+      // m_stride is equal to 1, so let's avoid the integer division.
+      eigen_assert(m_stride == 1);
+      Index inputIndex = index * m_inputStride + m_inputOffset;
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < PacketSize; ++i) {
+        values[i] = m_impl.coeff(inputIndex);
+        inputIndex += m_inputStride;
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    } else if (isOuterChipping()) {
+      // m_stride is always greater than index, so let's avoid the integer division.
+      eigen_assert(m_stride > index);
+      return m_impl.template packet<LoadMode>(index + m_inputOffset);
+    } else {
+      const Index idx = index / m_stride;
+      const Index rem = index - idx * m_stride;
+      if (rem + PacketSize <= m_stride) {
+        Index inputIndex = idx * m_inputStride + m_inputOffset + rem;
+        return m_impl.template packet<LoadMode>(inputIndex);
+      } else {
+        // Cross the stride boundary. Fallback to slow path.
+        EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+       EIGEN_UNROLL_LOOP
+        for (int i = 0; i < PacketSize; ++i) {
+          values[i] = coeff(index);
+          ++index;
+        }
+        PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+        return rslt;
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    double cost = 0;
+    if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) &&
+         m_dim.actualDim() == 0) ||
+        (static_cast<int>(Layout) == static_cast<int>(RowMajor) &&
+         m_dim.actualDim() == NumInputDims - 1)) {
+      cost += TensorOpCost::MulCost<Index>() + TensorOpCost::AddCost<Index>();
+    } else if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) &&
+                m_dim.actualDim() == NumInputDims - 1) ||
+               (static_cast<int>(Layout) == static_cast<int>(RowMajor) &&
+                m_dim.actualDim() == 0)) {
+      cost += TensorOpCost::AddCost<Index>();
+    } else {
+      cost += 3 * TensorOpCost::MulCost<Index>() + TensorOpCost::DivCost<Index>() +
+              3 * TensorOpCost::AddCost<Index>();
+    }
+
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    const size_t target_size = m_device.lastLevelCacheSize();
+    return internal::TensorBlockResourceRequirements::merge(
+        internal::TensorBlockResourceRequirements::skewed<Scalar>(target_size),
+        m_impl.getResourceRequirements());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool root_of_expr_ast = false) const {
+    const Index chip_dim = m_dim.actualDim();
+
+    DSizes<Index, NumInputDims> input_block_dims;
+    for (int i = 0; i < NumInputDims; ++i) {
+      input_block_dims[i]
+            = i < chip_dim ? desc.dimension(i)
+            : i > chip_dim ? desc.dimension(i - 1)
+            : 1;
+    }
+
+    ArgTensorBlockDesc arg_desc(srcCoeff(desc.offset()), input_block_dims);
+
+    // Try to reuse destination buffer for materializing argument block.
+    if (desc.HasDestinationBuffer()) {
+      DSizes<Index, NumInputDims> arg_destination_strides;
+      for (int i = 0; i < NumInputDims; ++i) {
+      arg_destination_strides[i]
+            = i < chip_dim ? desc.destination().strides()[i]
+            : i > chip_dim ? desc.destination().strides()[i - 1]
+            : 0; // for dimensions of size `1` stride should never be used.
+      }
+
+      arg_desc.template AddDestinationBuffer<Layout>(
+          desc.destination().template data<ScalarNoConst>(),
+          arg_destination_strides);
+    }
+
+    ArgTensorBlock arg_block = m_impl.block(arg_desc, scratch, root_of_expr_ast);
+    if (!arg_desc.HasDestinationBuffer()) desc.DropDestinationBuffer();
+
+    if (arg_block.data() != NULL) {
+      // Forward argument block buffer if possible.
+      return TensorBlock(arg_block.kind(), arg_block.data(),
+                           desc.dimensions());
+
+    } else {
+      // Assign argument block expression to a buffer.
+
+      // Prepare storage for the materialized chipping result.
+      const typename TensorBlock::Storage block_storage =
+          TensorBlock::prepareStorage(desc, scratch);
+
+      typedef internal::TensorBlockAssignment<
+          ScalarNoConst, NumInputDims, typename ArgTensorBlock::XprType, Index>
+          TensorBlockAssignment;
+
+      TensorBlockAssignment::Run(
+          TensorBlockAssignment::target(
+              arg_desc.dimensions(),
+              internal::strides<Layout>(arg_desc.dimensions()),
+              block_storage.data()),
+          arg_block.expr());
+
+      return block_storage.AsTensorMaterializedBlock();
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Storage::Type data() const {
+    typename Storage::Type result = constCast(m_impl.data());
+    if (isOuterChipping() && result) {
+      return result + m_inputOffset;
+    } else {
+      return NULL;
+    }
+  }
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const
+  {
+    Index inputIndex;
+    if (isInnerChipping()) {
+      // m_stride is equal to 1, so let's avoid the integer division.
+      eigen_assert(m_stride == 1);
+      inputIndex = index * m_inputStride + m_inputOffset;
+    } else if (isOuterChipping()) {
+      // m_stride is always greater than index, so let's avoid the integer
+      // division.
+      eigen_assert(m_stride > index);
+      inputIndex = index + m_inputOffset;
+    } else {
+      const Index idx = index / m_stride;
+      inputIndex = idx * m_inputStride + m_inputOffset;
+      index -= idx * m_stride;
+      inputIndex += index;
+    }
+    return inputIndex;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool isInnerChipping() const {
+    return IsInnerChipping ||
+           (static_cast<int>(Layout) == ColMajor && m_dim.actualDim() == 0) ||
+           (static_cast<int>(Layout) == RowMajor && m_dim.actualDim() == NumInputDims - 1);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool isOuterChipping() const {
+    return IsOuterChipping ||
+           (static_cast<int>(Layout) == ColMajor && m_dim.actualDim() == NumInputDims-1) ||
+           (static_cast<int>(Layout) == RowMajor && m_dim.actualDim() == 0);
+  }
+
+  Dimensions m_dimensions;
+  Index m_stride;
+  Index m_inputOffset;
+  Index m_inputStride;
+  TensorEvaluator<ArgType, Device> m_impl;
+  const internal::DimensionId<DimId> m_dim;
+  const Device EIGEN_DEVICE_REF m_device;
+};
+
+
+// Eval as lvalue
+template<DenseIndex DimId, typename ArgType, typename Device>
+struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
+  : public TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
+{
+  typedef TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device> Base;
+  typedef TensorChippingOp<DimId, ArgType> XprType;
+  static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  static const int NumDims = NumInputDims-1;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+
+  enum {
+    IsAligned     = false,
+    PacketAccess  = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess   = TensorEvaluator<ArgType, Device>::RawAccess,
+    Layout        = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess     = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : Base(op, device)
+    { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
+  {
+    return this->m_impl.coeffRef(this->srcCoeff(index));
+  }
+
+  template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x)
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+
+    if (this->isInnerChipping()) {
+      // m_stride is equal to 1, so let's avoid the integer division.
+      eigen_assert(this->m_stride == 1);
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
+      Index inputIndex = index * this->m_inputStride + this->m_inputOffset;
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < PacketSize; ++i) {
+        this->m_impl.coeffRef(inputIndex) = values[i];
+        inputIndex += this->m_inputStride;
+      }
+    } else if (this->isOuterChipping()) {
+      // m_stride is always greater than index, so let's avoid the integer division.
+      eigen_assert(this->m_stride > index);
+      this->m_impl.template writePacket<StoreMode>(index + this->m_inputOffset, x);
+    } else {
+      const Index idx = index / this->m_stride;
+      const Index rem = index - idx * this->m_stride;
+      if (rem + PacketSize <= this->m_stride) {
+        const Index inputIndex = idx * this->m_inputStride + this->m_inputOffset + rem;
+        this->m_impl.template writePacket<StoreMode>(inputIndex, x);
+      } else {
+        // Cross stride boundary. Fallback to slow path.
+        EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+        internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
+        EIGEN_UNROLL_LOOP
+        for (int i = 0; i < PacketSize; ++i) {
+          this->coeffRef(index) = values[i];
+          ++index;
+        }
+      }
+    }
+  }
+
+  template <typename TensorBlock>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const TensorBlockDesc& desc, const TensorBlock& block) {
+    assert(this->m_impl.data() != NULL);
+
+    const Index chip_dim = this->m_dim.actualDim();
+
+    DSizes<Index, NumInputDims> input_block_dims;
+    for (int i = 0; i < NumInputDims; ++i) {
+      input_block_dims[i] = i < chip_dim ? desc.dimension(i)
+                          : i > chip_dim ? desc.dimension(i - 1)
+                          : 1;
+    }
+
+    typedef TensorReshapingOp<const DSizes<Index, NumInputDims>,
+                              const typename TensorBlock::XprType>
+        TensorBlockExpr;
+
+    typedef internal::TensorBlockAssignment<Scalar, NumInputDims,
+                                            TensorBlockExpr, Index>
+        TensorBlockAssign;
+
+    TensorBlockAssign::Run(
+        TensorBlockAssign::target(
+            input_block_dims,
+            internal::strides<Layout>(this->m_impl.dimensions()),
+            this->m_impl.data(), this->srcCoeff(desc.offset())),
+        block.expr().reshape(input_block_dims));
+  }
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorConcatenation.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorConcatenation.h
new file mode 100644
index 0000000..5235a8e
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorConcatenation.h
@@ -0,0 +1,377 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONCATENATION_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONCATENATION_H
+
+namespace Eigen {
+
+/** \class TensorConcatenationOp
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor concatenation class.
+  *
+  *
+  */
+namespace internal {
+template<typename Axis, typename LhsXprType, typename RhsXprType>
+struct traits<TensorConcatenationOp<Axis, LhsXprType, RhsXprType> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs are different.
+  typedef typename promote_storage_type<typename LhsXprType::Scalar,
+                                        typename RhsXprType::Scalar>::ret Scalar;
+  typedef typename promote_storage_type<typename traits<LhsXprType>::StorageKind,
+                                        typename traits<RhsXprType>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<LhsXprType>::Index,
+                                      typename traits<RhsXprType>::Index>::type Index;
+  typedef typename LhsXprType::Nested LhsNested;
+  typedef typename RhsXprType::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+  static const int NumDimensions = traits<LhsXprType>::NumDimensions;
+  static const int Layout = traits<LhsXprType>::Layout;
+  enum { Flags = 0 };
+  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                               typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
+};
+
+template<typename Axis, typename LhsXprType, typename RhsXprType>
+struct eval<TensorConcatenationOp<Axis, LhsXprType, RhsXprType>, Eigen::Dense>
+{
+  typedef const TensorConcatenationOp<Axis, LhsXprType, RhsXprType>& type;
+};
+
+template<typename Axis, typename LhsXprType, typename RhsXprType>
+struct nested<TensorConcatenationOp<Axis, LhsXprType, RhsXprType>, 1, typename eval<TensorConcatenationOp<Axis, LhsXprType, RhsXprType> >::type>
+{
+  typedef TensorConcatenationOp<Axis, LhsXprType, RhsXprType> type;
+};
+
+}  // end namespace internal
+
+
+template<typename Axis, typename LhsXprType, typename RhsXprType>
+class TensorConcatenationOp : public TensorBase<TensorConcatenationOp<Axis, LhsXprType, RhsXprType>, WriteAccessors>
+{
+  public:
+    typedef TensorBase<TensorConcatenationOp<Axis, LhsXprType, RhsXprType>, WriteAccessors> Base;
+    typedef typename internal::traits<TensorConcatenationOp>::Scalar Scalar;
+    typedef typename internal::traits<TensorConcatenationOp>::StorageKind StorageKind;
+    typedef typename internal::traits<TensorConcatenationOp>::Index Index;
+    typedef typename internal::nested<TensorConcatenationOp>::type Nested;
+    typedef typename internal::promote_storage_type<typename LhsXprType::CoeffReturnType,
+                                                    typename RhsXprType::CoeffReturnType>::ret CoeffReturnType;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorConcatenationOp(const LhsXprType& lhs, const RhsXprType& rhs, Axis axis)
+        : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_axis(axis) {}
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename LhsXprType::Nested>::type&
+    lhsExpression() const { return m_lhs_xpr; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename RhsXprType::Nested>::type&
+    rhsExpression() const { return m_rhs_xpr; }
+
+    EIGEN_DEVICE_FUNC const Axis& axis() const { return m_axis; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorConcatenationOp)
+  protected:
+    typename LhsXprType::Nested m_lhs_xpr;
+    typename RhsXprType::Nested m_rhs_xpr;
+    const Axis m_axis;
+};
+
+
+// Eval as rvalue
+template<typename Axis, typename LeftArgType, typename RightArgType, typename Device>
+struct TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgType>, Device>
+{
+  typedef TensorConcatenationOp<Axis, LeftArgType, RightArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<LeftArgType, Device>::Dimensions>::value;
+  static const int RightNumDims = internal::array_size<typename TensorEvaluator<RightArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  enum {
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<LeftArgType, Device>::PacketAccess &&
+                        TensorEvaluator<RightArgType, Device>::PacketAccess,
+    BlockAccess       = false,
+    PreferBlockAccess = TensorEvaluator<LeftArgType, Device>::PreferBlockAccess ||
+                        TensorEvaluator<RightArgType, Device>::PreferBlockAccess,
+    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
+    RawAccess         = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device), m_axis(op.axis())
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) == static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout) || NumDims == 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT((NumDims == RightNumDims), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    eigen_assert(0 <= m_axis && m_axis < NumDims);
+    const Dimensions& lhs_dims = m_leftImpl.dimensions();
+    const Dimensions& rhs_dims = m_rightImpl.dimensions();
+    {
+      int i = 0;
+      for (; i < m_axis; ++i) {
+        eigen_assert(lhs_dims[i] > 0);
+        eigen_assert(lhs_dims[i] == rhs_dims[i]);
+        m_dimensions[i] = lhs_dims[i];
+      }
+      eigen_assert(lhs_dims[i] > 0);  // Now i == m_axis.
+      eigen_assert(rhs_dims[i] > 0);
+      m_dimensions[i] = lhs_dims[i] + rhs_dims[i];
+      for (++i; i < NumDims; ++i) {
+        eigen_assert(lhs_dims[i] > 0);
+        eigen_assert(lhs_dims[i] == rhs_dims[i]);
+        m_dimensions[i] = lhs_dims[i];
+      }
+    }
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_leftStrides[0] = 1;
+      m_rightStrides[0] = 1;
+      m_outputStrides[0] = 1;
+
+      for (int j = 1; j < NumDims; ++j) {
+        m_leftStrides[j] = m_leftStrides[j-1] * lhs_dims[j-1];
+        m_rightStrides[j] = m_rightStrides[j-1] * rhs_dims[j-1];
+        m_outputStrides[j] = m_outputStrides[j-1] * m_dimensions[j-1];
+      }
+    } else {
+      m_leftStrides[NumDims - 1] = 1;
+      m_rightStrides[NumDims - 1] = 1;
+      m_outputStrides[NumDims - 1] = 1;
+
+      for (int j = NumDims - 2; j >= 0; --j) {
+        m_leftStrides[j] = m_leftStrides[j+1] * lhs_dims[j+1];
+        m_rightStrides[j] = m_rightStrides[j+1] * rhs_dims[j+1];
+        m_outputStrides[j] = m_outputStrides[j+1] * m_dimensions[j+1];
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  // TODO(phli): Add short-circuit memcpy evaluation if underlying data are linear?
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType)
+  {
+    m_leftImpl.evalSubExprsIfNeeded(NULL);
+    m_rightImpl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_STRONG_INLINE void cleanup()
+  {
+    m_leftImpl.cleanup();
+    m_rightImpl.cleanup();
+  }
+
+  // TODO(phli): attempt to speed this up. The integer divisions and modulo are slow.
+  // See CL/76180724 comments for more ideas.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    // Collect dimension-wise indices (subs).
+    array<Index, NumDims> subs;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        subs[i] = index / m_outputStrides[i];
+        index -= subs[i] * m_outputStrides[i];
+      }
+      subs[0] = index;
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        subs[i] = index / m_outputStrides[i];
+        index -= subs[i] * m_outputStrides[i];
+      }
+      subs[NumDims - 1] = index;
+    }
+
+    const Dimensions& left_dims = m_leftImpl.dimensions();
+    if (subs[m_axis] < left_dims[m_axis]) {
+      Index left_index;
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        left_index = subs[0];
+        EIGEN_UNROLL_LOOP
+        for (int i = 1; i < NumDims; ++i) {
+          left_index += (subs[i] % left_dims[i]) * m_leftStrides[i];
+        }
+      } else {
+        left_index = subs[NumDims - 1];
+        EIGEN_UNROLL_LOOP
+        for (int i = NumDims - 2; i >= 0; --i) {
+          left_index += (subs[i] % left_dims[i]) * m_leftStrides[i];
+        }
+      }
+      return m_leftImpl.coeff(left_index);
+    } else {
+      subs[m_axis] -= left_dims[m_axis];
+      const Dimensions& right_dims = m_rightImpl.dimensions();
+      Index right_index;
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        right_index = subs[0];
+        EIGEN_UNROLL_LOOP
+        for (int i = 1; i < NumDims; ++i) {
+          right_index += (subs[i] % right_dims[i]) * m_rightStrides[i];
+        }
+      } else {
+        right_index = subs[NumDims - 1];
+        EIGEN_UNROLL_LOOP
+        for (int i = NumDims - 2; i >= 0; --i) {
+          right_index += (subs[i] % right_dims[i]) * m_rightStrides[i];
+        }
+      }
+      return m_rightImpl.coeff(right_index);
+    }
+  }
+
+  // TODO(phli): Add a real vectorization.
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
+    EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index + packetSize - 1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX CoeffReturnType values[packetSize];
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < packetSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    const double compute_cost = NumDims * (2 * TensorOpCost::AddCost<Index>() +
+                                           2 * TensorOpCost::MulCost<Index>() +
+                                           TensorOpCost::DivCost<Index>() +
+                                           TensorOpCost::ModCost<Index>());
+    const double lhs_size = m_leftImpl.dimensions().TotalSize();
+    const double rhs_size = m_rightImpl.dimensions().TotalSize();
+    return (lhs_size / (lhs_size + rhs_size)) *
+               m_leftImpl.costPerCoeff(vectorized) +
+           (rhs_size / (lhs_size + rhs_size)) *
+               m_rightImpl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, compute_cost);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+  #ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_leftImpl.bind(cgh);
+    m_rightImpl.bind(cgh);
+  }
+  #endif
+
+  protected:
+    Dimensions m_dimensions;
+    array<Index, NumDims> m_outputStrides;
+    array<Index, NumDims> m_leftStrides;
+    array<Index, NumDims> m_rightStrides;
+    TensorEvaluator<LeftArgType, Device> m_leftImpl;
+    TensorEvaluator<RightArgType, Device> m_rightImpl;
+    const Axis m_axis;
+};
+
+// Eval as lvalue
+template<typename Axis, typename LeftArgType, typename RightArgType, typename Device>
+  struct TensorEvaluator<TensorConcatenationOp<Axis, LeftArgType, RightArgType>, Device>
+  : public TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgType>, Device>
+{
+  typedef TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgType>, Device> Base;
+  typedef TensorConcatenationOp<Axis, LeftArgType, RightArgType> XprType;
+  typedef typename Base::Dimensions Dimensions;
+  enum {
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<LeftArgType, Device>::PacketAccess &&
+                        TensorEvaluator<RightArgType, Device>::PacketAccess,
+    BlockAccess       = false,
+    PreferBlockAccess = TensorEvaluator<LeftArgType, Device>::PreferBlockAccess ||
+                        TensorEvaluator<RightArgType, Device>::PreferBlockAccess,
+    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
+    RawAccess         = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(XprType& op, const Device& device)
+    : Base(op, device)
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(Layout) == static_cast<int>(ColMajor)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+  }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
+  {
+    // Collect dimension-wise indices (subs).
+    array<Index, Base::NumDims> subs;
+    for (int i = Base::NumDims - 1; i > 0; --i) {
+      subs[i] = index / this->m_outputStrides[i];
+      index -= subs[i] * this->m_outputStrides[i];
+    }
+    subs[0] = index;
+
+    const Dimensions& left_dims = this->m_leftImpl.dimensions();
+    if (subs[this->m_axis] < left_dims[this->m_axis]) {
+      Index left_index = subs[0];
+      for (int i = 1; i < Base::NumDims; ++i) {
+        left_index += (subs[i] % left_dims[i]) * this->m_leftStrides[i];
+      }
+      return this->m_leftImpl.coeffRef(left_index);
+    } else {
+      subs[this->m_axis] -= left_dims[this->m_axis];
+      const Dimensions& right_dims = this->m_rightImpl.dimensions();
+      Index right_index = subs[0];
+      for (int i = 1; i < Base::NumDims; ++i) {
+        right_index += (subs[i] % right_dims[i]) * this->m_rightStrides[i];
+      }
+      return this->m_rightImpl.coeffRef(right_index);
+    }
+  }
+
+  template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x)
+  {
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
+    EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index + packetSize - 1 < this->dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX CoeffReturnType values[packetSize];
+    internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
+    for (int i = 0; i < packetSize; ++i) {
+      coeffRef(index+i) = values[i];
+    }
+  }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONCATENATION_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorContraction.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorContraction.h
new file mode 100644
index 0000000..8b35f79
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorContraction.h
@@ -0,0 +1,1023 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H
+
+namespace Eigen {
+
+/** \class TensorContraction
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor contraction class.
+  *
+  *
+  */
+namespace internal {
+
+template<typename Dimensions, typename LhsXprType, typename RhsXprType, typename OutputKernelType>
+struct traits<TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs are different.
+  typedef typename gebp_traits<typename remove_const<typename LhsXprType::Scalar>::type,
+                               typename remove_const<typename RhsXprType::Scalar>::type>::ResScalar Scalar;
+
+  typedef typename promote_storage_type<typename traits<LhsXprType>::StorageKind,
+                                        typename traits<RhsXprType>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<LhsXprType>::Index,
+                                      typename traits<RhsXprType>::Index>::type Index;
+  typedef typename LhsXprType::Nested LhsNested;
+  typedef typename RhsXprType::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+
+  // From NumDims below.
+  static const int NumDimensions = traits<LhsXprType>::NumDimensions + traits<RhsXprType>::NumDimensions - 2 * array_size<Dimensions>::value;
+  static const int Layout = traits<LhsXprType>::Layout;
+  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                               typename traits<LhsXprType>::PointerType,
+                               typename traits<RhsXprType>::PointerType>::type
+      PointerType;
+
+  enum {
+    Flags = 0
+  };
+};
+
+template<typename Dimensions, typename LhsXprType, typename RhsXprType, typename OutputKernelType>
+struct eval<TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType>, Eigen::Dense>
+{
+  typedef const TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType>& type;
+};
+
+template<typename Dimensions, typename LhsXprType, typename RhsXprType, typename OutputKernelType>
+struct nested<TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType>, 1, typename eval<TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType> >::type>
+{
+  typedef TensorContractionOp<Dimensions, LhsXprType, RhsXprType, OutputKernelType> type;
+};
+
+template<typename Indices_, typename LeftArgType_, typename RightArgType_, typename OutputKernelType_, typename Device_>
+struct traits<TensorEvaluator<const TensorContractionOp<Indices_, LeftArgType_, RightArgType_, OutputKernelType_>, Device_> > {
+  typedef Indices_ Indices;
+  typedef LeftArgType_ LeftArgType;
+  typedef RightArgType_ RightArgType;
+  typedef OutputKernelType_ OutputKernelType;
+  typedef Device_ Device;
+
+  // From NumDims below.
+  static const int NumDimensions = traits<LeftArgType_>::NumDimensions + traits<RightArgType_>::NumDimensions - 2 * array_size<Indices_>::value;
+};
+
+// Helper class to allocate and deallocate temporary memory for packed buffers.
+template <typename LhsScalar, typename RhsScalar>
+struct TensorContractionBlockMemAllocator {
+  typedef void* BlockMemHandle;
+
+  template <typename Device>
+  EIGEN_DEVICE_FUNC static BlockMemHandle allocate(Device& d, const Index bm,
+                                                   const Index bk,
+                                                   const Index bn,
+                                                   LhsScalar** lhs_block,
+                                                   RhsScalar** rhs_block) {
+    eigen_assert(lhs_block);
+    eigen_assert(rhs_block);
+    BlockSizes sz = ComputeLhsRhsBlockSizes(bm, bk, bn);
+    char* block_mem = static_cast<char*>(d.allocate(sz.lhs_size + sz.rhs_size));
+    eigen_assert(block_mem);
+    *lhs_block = reinterpret_cast<LhsScalar*>(block_mem);
+    *rhs_block = reinterpret_cast<RhsScalar*>(block_mem + sz.lhs_size);
+    return block_mem;
+  }
+
+  template <typename Device>
+  EIGEN_DEVICE_FUNC static BlockMemHandle allocateSlices(
+      Device& d, const Index bm, const Index bk, const Index bn,
+      const Index num_lhs, const Index num_rhs, const Index num_slices,
+      std::vector<LhsScalar*>* lhs_blocks,
+      std::vector<RhsScalar*>* rhs_blocks) {
+    eigen_assert(num_slices > 0);
+    eigen_assert(num_lhs >= 0 && num_rhs >= 0);
+    eigen_assert(num_lhs == 0 || lhs_blocks);
+    eigen_assert(num_rhs == 0 || rhs_blocks);
+    BlockSizes sz = ComputeLhsRhsBlockSizes(bm, bk, bn);
+    void* block_mem = d.allocate(
+        (num_lhs * sz.lhs_size + num_rhs * sz.rhs_size) * num_slices);
+    eigen_assert(block_mem);
+    char* mem = static_cast<char*>(block_mem);
+
+    for (Index x = 0; x < num_slices; x++) {
+      if (num_lhs > 0) lhs_blocks[x].resize(num_lhs);
+      for (Index m = 0; m < num_lhs; m++) {
+        lhs_blocks[x][m] = reinterpret_cast<LhsScalar*>(mem);
+        mem += sz.lhs_size;
+      }
+      if (num_rhs > 0) rhs_blocks[x].resize(num_rhs);
+      for (Index n = 0; n < num_rhs; n++) {
+        rhs_blocks[x][n] = reinterpret_cast<RhsScalar*>(mem);
+        mem += sz.rhs_size;
+      }
+    }
+
+    return block_mem;
+  }
+
+  template <typename Device>
+  EIGEN_DEVICE_FUNC static void deallocate(Device& d, BlockMemHandle handle) {
+    d.deallocate(handle);
+  }
+
+ private:
+  struct BlockSizes {
+    Index lhs_size;
+    Index rhs_size;
+  };
+  EIGEN_DEVICE_FUNC static BlockSizes ComputeLhsRhsBlockSizes(const Index bm,
+                                                              const Index bk,
+                                                              const Index bn) {
+    Index align = numext::maxi(EIGEN_MAX_ALIGN_BYTES, 1);
+    BlockSizes sz;
+    sz.lhs_size = divup<Index>(bm * bk * sizeof(LhsScalar), align) * align;
+    sz.rhs_size = divup<Index>(bn * bk * sizeof(RhsScalar), align) * align;
+    return sz;
+  }
+};
+
+// WARNING: In this code we assume that Lhs and Rhs tensor expressions are in
+// ColMajor storage order. This property is guaranteed by the
+// TensorContractionOp evaluator. TensorContractionKernel specifies how we pack
+// blocks of Lhs and Rhs tensor expressions, and how we invoke matrix
+// multiplication for these blocks. Default tensor contraction uses
+// gemm_pack_rhs, gemm_pack_lhs and gebp_kernel from Eigen Core (see
+// GeneralBlocPanelKernel.h for details).
+//
+// By specializing contraction kernels we can use other low level libraries to
+// perform matrix multiplication, and still rely on Eigen contraction evaluator.
+// This also includes full support in TensorContractionThreadPool, assuming that
+// underlying gemm do not use it's own threading.
+//
+// - ResScalar/LhsScalar/RhsScalar - scalar type for the result of
+//   multiplication, lhs tensor and rhs tensor respectively.
+//
+// - StorageIndex - index type for the tensor expressions. In practice almost
+//   always is Eigen::Index.
+//
+// - OutputMapper provides access to the memory of the output matrix. In
+//   practice it's always column major blas_data_mapper (it must be of ResScalar
+//   type).
+//
+// - LhsMapper/RhsMapper similarly to blas_data_mapper provide a two dimensional
+//   view into the Lhs/Rhs tensor expressions. In practice it's
+//   TensorContractionInputMapper, or some specialization of it based on the
+//   type of tensor expression (e.g. TensorImagePatchOp has optimized input
+//   mapper).
+template <typename ResScalar, typename LhsScalar, typename RhsScalar,
+    typename StorageIndex, typename OutputMapper, typename LhsMapper,
+    typename RhsMapper>
+struct TensorContractionKernel {
+  // True if `invoke()` supports `beta` in `C <- alpha * A * B + beta * C`
+  // (otherwise beta should be always equal to 1).
+  enum { HasBeta = false };
+
+  EIGEN_DEVICE_FUNC
+  TensorContractionKernel(StorageIndex m_, StorageIndex k_, StorageIndex n_,
+                          StorageIndex bm_, StorageIndex bk_, StorageIndex bn_)
+      : m(m_), k(k_), n(n_), bm(bm_), bk(bk_), bn(bn_) {}
+
+  // Pack blocks of Lhs and Rhs into contiguous blocks in memory.
+  typedef LhsScalar* LhsBlock;
+  typedef RhsScalar* RhsBlock;
+
+  // Packed Lhs/Rhs block memory allocator.
+  typedef TensorContractionBlockMemAllocator<LhsScalar, RhsScalar>
+      BlockMemAllocator;
+  typedef typename BlockMemAllocator::BlockMemHandle BlockMemHandle;
+
+  typedef typename internal::gebp_traits<LhsScalar, RhsScalar> Traits;
+
+  typedef internal::gemm_pack_lhs<
+      LhsScalar, StorageIndex, typename LhsMapper::SubMapper, Traits::mr,
+      Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor>
+      LhsPacker;
+
+  typedef internal::gemm_pack_rhs<RhsScalar, StorageIndex,
+                                  typename RhsMapper::SubMapper, Traits::nr,
+                                  ColMajor>
+      RhsPacker;
+
+  typedef internal::gebp_kernel<LhsScalar, RhsScalar, StorageIndex,
+                                OutputMapper, Traits::mr, Traits::nr,
+      /*ConjugateLhs*/ false, /*ConjugateRhs*/ false>
+      GebpKernel;
+
+  template <typename Device>
+  EIGEN_DEVICE_FUNC BlockMemHandle allocate(Device& d, LhsBlock* lhs_block,
+                                            RhsBlock* rhs_block) {
+    return BlockMemAllocator::allocate(d, bm, bk, bn, lhs_block, rhs_block);
+  }
+
+  template <typename Device>
+  EIGEN_DEVICE_FUNC BlockMemHandle allocateSlices(
+      Device& d, const StorageIndex num_lhs, const StorageIndex num_rhs,
+      const StorageIndex num_slices, std::vector<LhsBlock>* lhs_blocks,
+      std::vector<RhsBlock>* rhs_blocks) {
+    return BlockMemAllocator::allocateSlices(
+        d, bm, bk, bn, num_lhs, num_rhs, num_slices, lhs_blocks, rhs_blocks);
+  }
+
+  template <typename Device>
+  EIGEN_DEVICE_FUNC static void deallocate(Device& d, BlockMemHandle handle) {
+    BlockMemAllocator::deallocate(d, handle);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void packLhs(
+      LhsBlock* lhsBlock, const typename LhsMapper::SubMapper& data_mapper,
+      const StorageIndex depth, const StorageIndex rows) {
+    LhsPacker()(*lhsBlock, data_mapper, depth, rows, /*stride*/ 0,
+        /*offset*/ 0);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void packRhs(
+      RhsBlock* rhsBlock, const typename RhsMapper::SubMapper& data_mapper,
+      const StorageIndex depth, const StorageIndex cols) {
+    RhsPacker()(*rhsBlock, data_mapper, depth, cols);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void invoke(
+      const OutputMapper& output_mapper, const LhsBlock& lhsBlock,
+      const RhsBlock& rhsBlock, const StorageIndex rows,
+      const StorageIndex depth, const StorageIndex cols,
+      const ResScalar alpha, const ResScalar beta) {
+    // Default GEBP kernel does not support beta.
+    eigen_assert(beta == ResScalar(1));
+    static const int kComputeStrideFromBlockDimensions = -1;
+    GebpKernel()(output_mapper, lhsBlock, rhsBlock, rows, depth, cols, alpha,
+        /*strideA*/ kComputeStrideFromBlockDimensions,
+        /*strideB*/ kComputeStrideFromBlockDimensions,
+        /*offsetA*/ 0, /*offsetB*/ 0);
+  }
+
+ private:
+  // These are dimensions of the original Tensors, and selected block sizes. The
+  // actual block sizes passed to all function above might be smaller because of
+  // the partial blocks at the end.
+  const StorageIndex m;
+  const StorageIndex k;
+  const StorageIndex n;
+  const StorageIndex bm;
+  const StorageIndex bk;
+  const StorageIndex bn;
+};
+
+}  // end namespace internal
+
+// Tensor contraction params that should enable to get from output matrix
+// 2-dimensional coordinates to the output tensor dimensions.
+struct TensorContractionParams {
+  // TensorContraction evaluator assumes that both tensors are in ColMajor
+  // layout, if tensors are in RowMajor evaluator swap lhs with rhs.
+  bool swapped_arguments;
+};
+
+// Output kernel allows to fuse operations into the tensor contraction.
+//
+// Examples:
+//   1. Elementwise Relu transformation following Conv2D.
+//   2. AddBias to the Conv2D output channels dimension.
+//
+// The NoOpOutputKernel implements an output kernel that does absolutely nothing.
+struct NoOpOutputKernel {
+  /**
+   * Tensor contraction evaluator calls this kernel after finishing each block
+   * of output matrix. Output blocks belong to the 2-dimensional output tensor.
+   *
+   * TensorContractionParams contains contraction dimensions information
+   * required to map output 2-d space into the expected output tensor space
+   * (potentially higher dimensional).
+   *
+   * \param[in] output_mapper Access to output tensor memory
+   * \param[in] params   Tensor contraction parameters
+   * \param[in] i        Index of a first row available through output_mapper
+   * \param[in] j        Index of a first column available through output_mapper
+   * \param[in] num_rows Number of available rows
+   * \param[in] num_cols Number of available columns
+   */
+  template <typename Index, typename Scalar>
+  EIGEN_ALWAYS_INLINE void operator()(
+      const internal::blas_data_mapper<Scalar, Index, ColMajor>& output_mapper,
+      const TensorContractionParams& params, Index i,
+      Index j, Index num_rows, Index num_cols) const {
+    EIGEN_UNUSED_VARIABLE(output_mapper);
+    EIGEN_UNUSED_VARIABLE(params);
+    EIGEN_UNUSED_VARIABLE(i);
+    EIGEN_UNUSED_VARIABLE(j);
+    EIGEN_UNUSED_VARIABLE(num_rows);
+    EIGEN_UNUSED_VARIABLE(num_cols);
+  }
+};
+
+template<typename Indices, typename LhsXprType, typename RhsXprType, typename OutputKernelType = const NoOpOutputKernel>
+class TensorContractionOp : public TensorBase<TensorContractionOp<Indices, LhsXprType, RhsXprType, OutputKernelType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorContractionOp>::Scalar Scalar;
+  typedef typename internal::gebp_traits<typename LhsXprType::CoeffReturnType,
+                                         typename RhsXprType::CoeffReturnType>::ResScalar CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorContractionOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorContractionOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorContractionOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorContractionOp(
+      const LhsXprType& lhs, const RhsXprType& rhs, const Indices& dims,
+      const OutputKernelType& output_kernel = OutputKernelType())
+      : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_indices(dims),
+        m_output_kernel(output_kernel) {}
+
+  EIGEN_DEVICE_FUNC
+  const Indices& indices() const { return m_indices; }
+
+  /** \returns the nested expressions */
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename LhsXprType::Nested>::type&
+  lhsExpression() const { return m_lhs_xpr; }
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename RhsXprType::Nested>::type&
+  rhsExpression() const { return m_rhs_xpr; }
+
+  EIGEN_DEVICE_FUNC
+  const OutputKernelType& outputKernel() const { return m_output_kernel; }
+
+  protected:
+    typename LhsXprType::Nested m_lhs_xpr;
+    typename RhsXprType::Nested m_rhs_xpr;
+    const Indices m_indices;
+    const OutputKernelType m_output_kernel;
+};
+
+
+template<typename Derived>
+struct TensorContractionEvaluatorBase : internal::no_assignment_operator
+{
+  typedef typename internal::traits<Derived>::Indices Indices;
+  typedef typename internal::traits<Derived>::LeftArgType LeftArgType;
+  typedef typename internal::traits<Derived>::RightArgType RightArgType;
+  typedef typename internal::traits<Derived>::OutputKernelType OutputKernelType;
+  typedef typename internal::traits<Derived>::Device Device;
+
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef StorageMemory<Scalar, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned         = true,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = false,
+    PreferBlockAccess = false,
+    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = true
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  // Most of the code is assuming that both input tensors are ColMajor. If the
+  // inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
+  // If we want to compute A * B = C, where A is LHS and B is RHS, the code
+  // will pretend B is LHS and A is RHS.
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType;
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType;
+
+  typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluatorType;
+  typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluatorType;
+
+  static const int LDims =
+      internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
+  static const int RDims =
+      internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
+  static const int ContractDims = internal::array_size<Indices>::value;
+  static const int NumDims = LDims + RDims - 2 * ContractDims;
+
+  typedef array<Index, ContractDims> contract_t;
+  typedef array<Index, LDims - ContractDims> left_nocontract_t;
+  typedef array<Index, RDims - ContractDims> right_nocontract_t;
+
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  EIGEN_STRONG_INLINE
+  TensorContractionEvaluatorBase(const XprType& op, const Device& device)
+      : m_leftImpl(choose(Cond<static_cast<int>(Layout) == static_cast<int>(ColMajor)>(),
+                          op.lhsExpression(), op.rhsExpression()), device),
+        m_rightImpl(choose(Cond<static_cast<int>(Layout) == static_cast<int>(ColMajor)>(),
+                           op.rhsExpression(), op.lhsExpression()), device),
+        m_device(device),
+        m_output_kernel(op.outputKernel()),
+        m_result(NULL) {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) ==
+         static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout)),
+                        YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+
+    DSizes<Index, LDims> eval_left_dims;
+    DSizes<Index, RDims> eval_right_dims;
+    array<IndexPair<Index>, ContractDims> eval_op_indices;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      // For ColMajor, we keep using the existing dimensions
+      for (int i = 0; i < LDims; i++) {
+        eval_left_dims[i] = m_leftImpl.dimensions()[i];
+      }
+      for (int i = 0; i < RDims; i++) {
+        eval_right_dims[i] = m_rightImpl.dimensions()[i];
+      }
+      // We keep the pairs of contracting indices.
+      for (int i = 0; i < ContractDims; i++) {
+        eval_op_indices[i].first = op.indices()[i].first;
+        eval_op_indices[i].second = op.indices()[i].second;
+      }
+    } else {
+      // For RowMajor, we need to reverse the existing dimensions
+      for (int i = 0; i < LDims; i++) {
+        eval_left_dims[i] = m_leftImpl.dimensions()[LDims - i - 1];
+      }
+      for (int i = 0; i < RDims; i++) {
+        eval_right_dims[i] = m_rightImpl.dimensions()[RDims - i - 1];
+      }
+      // We need to flip all the pairs of contracting indices as well as
+      // reversing the dimensions.
+      for (int i = 0; i < ContractDims; i++) {
+        eval_op_indices[i].first = LDims - 1 - op.indices()[ContractDims - 1 - i].second;
+        eval_op_indices[i].second = RDims - 1 - op.indices()[ContractDims - 1 - i].first;
+      }
+    }
+
+    // Check for duplicate axes and make sure the first index in eval_op_indices
+    // is increasing. Using O(n^2) sorting is OK since ContractDims is small
+    for (int i = 0; i < ContractDims; i++) {
+      for (int j = i + 1; j < ContractDims; j++) {
+        eigen_assert(eval_op_indices[j].first != eval_op_indices[i].first &&
+                     eval_op_indices[j].second != eval_op_indices[i].second &&
+                     "contraction axes should be unique");
+        if (eval_op_indices[j].first < eval_op_indices[i].first) {
+          numext::swap(eval_op_indices[j], eval_op_indices[i]);
+        }
+      }
+    }
+
+    array<Index, LDims> lhs_strides;
+    lhs_strides[0] = 1;
+    for (int i = 0; i < LDims-1; ++i) {
+      lhs_strides[i+1] = lhs_strides[i] * eval_left_dims[i];
+    }
+
+    array<Index, RDims> rhs_strides;
+    rhs_strides[0] = 1;
+    for (int i = 0; i < RDims-1; ++i) {
+      rhs_strides[i+1] = rhs_strides[i] * eval_right_dims[i];
+    }
+
+    if (m_i_strides.size() > 0) m_i_strides[0] = 1;
+    if (m_j_strides.size() > 0) m_j_strides[0] = 1;
+    if (m_k_strides.size() > 0) m_k_strides[0] = 1;
+
+    m_i_size = 1;
+    m_j_size = 1;
+    m_k_size = 1;
+
+    // To compute the dimension, we simply concatenate the non-contracting
+    // dimensions of the left and then the right tensor. Additionally, we also
+    // compute the strides corresponding to the left non-contracting
+    // dimensions and right non-contracting dimensions.
+    m_lhs_inner_dim_contiguous = true;
+    int dim_idx = 0;
+    Index nocontract_idx = 0;
+
+    for (int i = 0; i < LDims; i++) {
+      // find if we are contracting on index i of left tensor
+      bool contracting = false;
+      for (int j = 0; j < ContractDims; j++) {
+        if (eval_op_indices[j].first == i) {
+          contracting = true;
+          break;
+        }
+      }
+      if (!contracting) {
+        // add dimension size to output dimensions
+        m_dimensions[dim_idx] = eval_left_dims[i];
+        m_left_nocontract_strides[nocontract_idx] = lhs_strides[i];
+        if (dim_idx != i) {
+          m_lhs_inner_dim_contiguous = false;
+        }
+        if (nocontract_idx+1 < internal::array_size<left_nocontract_t>::value) {
+          m_i_strides[nocontract_idx+1] =
+              m_i_strides[nocontract_idx] * eval_left_dims[i];
+        } else {
+          m_i_size = m_i_strides[nocontract_idx] * eval_left_dims[i];
+        }
+        dim_idx++;
+        nocontract_idx++;
+      }
+    }
+
+    nocontract_idx = 0;
+    for (int i = 0; i < RDims; i++) {
+      bool contracting = false;
+      // find if we are contracting on index i of right tensor
+      for (int j = 0; j < ContractDims; j++) {
+        if (eval_op_indices[j].second == i) {
+          contracting = true;
+          break;
+        }
+      }
+      if (!contracting) {
+        m_dimensions[dim_idx] = eval_right_dims[i];
+        if (nocontract_idx+1 < internal::array_size<right_nocontract_t>::value) {
+          m_j_strides[nocontract_idx+1] =
+              m_j_strides[nocontract_idx] * eval_right_dims[i];
+        } else {
+          m_j_size = m_j_strides[nocontract_idx] * eval_right_dims[i];
+        }
+        m_right_nocontract_strides[nocontract_idx] = rhs_strides[i];
+        dim_idx++;
+        nocontract_idx++;
+      }
+    }
+
+    // Now compute the strides corresponding to the contracting dimensions. We
+    // assumed above that non-contracting axes are represented in the same order
+    // in the matrix as they are in the tensor. This is not the case for
+    // contracting axes. As the contracting axes must be of the same size in
+    // each tensor, we'll only look at the first tensor here.
+    m_rhs_inner_dim_contiguous = true;
+    m_rhs_inner_dim_reordered = false;
+    for (int i = 0; i < ContractDims; i++) {
+      Index left = eval_op_indices[i].first;
+      Index right = eval_op_indices[i].second;
+
+      Index size = eval_left_dims[left];
+      eigen_assert(size == eval_right_dims[right] &&
+                   "Contraction axes must be same size");
+
+      if (i+1 < static_cast<int>(internal::array_size<contract_t>::value)) {
+        m_k_strides[i+1] = m_k_strides[i] * size;
+      } else {
+        m_k_size = m_k_strides[i] * size;
+      }
+      m_left_contracting_strides[i] = lhs_strides[left];
+      m_right_contracting_strides[i] = rhs_strides[right];
+
+      if (i > 0 && right < eval_op_indices[i-1].second) {
+        m_rhs_inner_dim_reordered = true;
+      }
+      if (right != i) {
+        m_rhs_inner_dim_contiguous = false;
+      }
+    }
+
+    // If the layout is RowMajor, we need to reverse the m_dimensions
+    if (static_cast<int>(Layout) == static_cast<int>(RowMajor)) {
+      for (int i = 0, j = NumDims - 1; i < j; i++, j--) {
+        numext::swap(m_dimensions[i], m_dimensions[j]);
+      }
+    }
+
+    // A set of parameters that will allow output kernel to get from output
+    // tensor dimensions (i, j) into the original tensor dimensions.
+    // TODO(ezhulenev): Add parameters required to infer output tensor index for
+    // more complex contractions than 2x2 on internal dimension.
+    m_tensor_contraction_params.swapped_arguments = static_cast<int>(Layout) == RowMajor;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    m_leftImpl.evalSubExprsIfNeeded(NULL);
+    m_rightImpl.evalSubExprsIfNeeded(NULL);
+    if (data) {
+      evalTo(data);
+      return false;
+    } else {
+      m_result = static_cast<EvaluatorPointerType>(m_device.allocate(dimensions().TotalSize() * sizeof(Scalar)));
+      evalTo(m_result);
+      return true;
+    }
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType dest, EvalSubExprsCallback done) {
+    m_leftImpl.evalSubExprsIfNeededAsync(nullptr, [this, done, dest](bool) {
+      m_rightImpl.evalSubExprsIfNeededAsync(nullptr, [this, done, dest](bool) {
+        if (dest) {
+          evalToAsync(dest, [done]() { done(false); });
+        } else {
+          m_result = static_cast<EvaluatorPointerType>(
+              m_device.allocate(dimensions().TotalSize() * sizeof(Scalar)));
+          evalToAsync(m_result, [done]() { done(true); });
+        }
+      });
+    });
+  }
+#endif  // EIGEN_USE_THREADS
+
+#ifndef TENSOR_CONTRACTION_DISPATCH
+#define TENSOR_CONTRACTION_DISPATCH(METHOD, ALIGNMENT, ARGS) \
+  if (this->m_lhs_inner_dim_contiguous) {                    \
+    if (this->m_rhs_inner_dim_contiguous) {                  \
+      if (this->m_rhs_inner_dim_reordered) {                 \
+        METHOD<true, true, true, ALIGNMENT> ARGS;            \
+      } else {                                               \
+        METHOD<true, true, false, ALIGNMENT> ARGS;           \
+      }                                                      \
+    } else {                                                 \
+      if (this->m_rhs_inner_dim_reordered) {                 \
+        METHOD<true, false, true, ALIGNMENT> ARGS;           \
+      } else {                                               \
+        METHOD<true, false, false, ALIGNMENT> ARGS;          \
+      }                                                      \
+    }                                                        \
+  } else {                                                   \
+    if (this->m_rhs_inner_dim_contiguous) {                  \
+      if (this->m_rhs_inner_dim_reordered) {                 \
+        METHOD<false, true, true, ALIGNMENT> ARGS;           \
+      } else {                                               \
+        METHOD<false, true, false, ALIGNMENT> ARGS;          \
+      }                                                      \
+    } else {                                                 \
+      if (this->m_rhs_inner_dim_reordered) {                 \
+        METHOD<false, false, true, ALIGNMENT> ARGS;          \
+      } else {                                               \
+        METHOD<false, false, false, ALIGNMENT> ARGS;         \
+      }                                                      \
+    }                                                        \
+  }
+#endif
+
+#ifndef TENSOR_CONTRACTION_ASYNC_DISPATCH
+#define TENSOR_CONTRACTION_ASYNC_DISPATCH(METHOD, DONE, ALIGNMENT, ARGS, FN) \
+  if (this->m_lhs_inner_dim_contiguous) {                                    \
+    if (this->m_rhs_inner_dim_contiguous) {                                  \
+      if (this->m_rhs_inner_dim_reordered) {                                 \
+        (new METHOD<DONE, true, true, true, ALIGNMENT> ARGS)->FN;            \
+      } else {                                                               \
+        (new METHOD<DONE, true, true, false, ALIGNMENT> ARGS)->FN;           \
+      }                                                                      \
+    } else {                                                                 \
+      if (this->m_rhs_inner_dim_reordered) {                                 \
+        (new METHOD<DONE, true, false, true, ALIGNMENT> ARGS)->FN;           \
+      } else {                                                               \
+        (new METHOD<DONE, true, false, false, ALIGNMENT> ARGS)->FN;          \
+      }                                                                      \
+    }                                                                        \
+  } else {                                                                   \
+    if (this->m_rhs_inner_dim_contiguous) {                                  \
+      if (this->m_rhs_inner_dim_reordered) {                                 \
+        (new METHOD<DONE, false, true, true, ALIGNMENT> ARGS)->FN;           \
+      } else {                                                               \
+        (new METHOD<DONE, false, true, false, ALIGNMENT> ARGS)->FN;          \
+      }                                                                      \
+    } else {                                                                 \
+      if (this->m_rhs_inner_dim_reordered) {                                 \
+        (new METHOD<DONE, false, false, true, ALIGNMENT> ARGS)->FN;          \
+      } else {                                                               \
+        (new METHOD<DONE, false, false, false, ALIGNMENT> ARGS)->FN;         \
+      }                                                                      \
+    }                                                                        \
+  }
+#endif
+
+  EIGEN_DEVICE_FUNC void evalTo(Scalar* buffer) const {
+   static_cast<const Derived*>(this)->template evalProduct<Unaligned>(buffer);
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalToCallback>
+  void evalToAsync(Scalar* buffer, EvalToCallback done) const {
+    static_cast<const Derived*>(this)
+        ->template evalProductAsync<EvalToCallback, Unaligned>(buffer,
+                                                               std::move(done));
+  }
+#endif  // EIGEN_USE_THREADS
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous,
+            bool rhs_inner_dim_reordered, int Alignment>
+  void evalProductSequential(Scalar* buffer) const {
+    if (this->m_j_size == 1) {
+      this->template evalGemv<lhs_inner_dim_contiguous,
+                              rhs_inner_dim_contiguous, rhs_inner_dim_reordered,
+                              Alignment>(buffer);
+    } else {
+      this->template evalGemm<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous,
+                              rhs_inner_dim_reordered, Alignment>(buffer);
+    }
+  }
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
+  #if !defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
+  void evalGemv(Scalar* buffer) const {
+    const Index rows = m_i_size;
+    const Index cols = m_k_size;
+
+    typedef typename internal::remove_const<typename EvalLeftArgType::Scalar>::type LhsScalar;
+    typedef typename internal::remove_const<typename EvalRightArgType::Scalar>::type RhsScalar;
+    typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
+    typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
+    const Index lhs_packet_size = internal::unpacket_traits<typename LeftEvaluator::PacketReturnType>::size;
+    const Index rhs_packet_size = internal::unpacket_traits<typename RightEvaluator::PacketReturnType>::size;
+    const int lhs_alignment = LeftEvaluator::IsAligned ? Aligned : Unaligned;
+    const int rhs_alignment = RightEvaluator::IsAligned ? Aligned : Unaligned;
+    typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs,
+                                                   LeftEvaluator, left_nocontract_t,
+                                                   contract_t, lhs_packet_size,
+                                                   lhs_inner_dim_contiguous,
+                                                   false, lhs_alignment> LhsMapper;
+
+    typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs,
+                                                   RightEvaluator, right_nocontract_t,
+                                                   contract_t, rhs_packet_size,
+                                                   rhs_inner_dim_contiguous,
+                                                   rhs_inner_dim_reordered, rhs_alignment> RhsMapper;
+
+    LhsMapper lhs(m_leftImpl, m_left_nocontract_strides, m_i_strides,
+                  m_left_contracting_strides, m_k_strides);
+    RhsMapper rhs(m_rightImpl, m_right_nocontract_strides, m_j_strides,
+                  m_right_contracting_strides, m_k_strides);
+
+    const Scalar alpha(1);
+    const Index resIncr(1);
+
+    // zero out the result buffer (which must be of size at least rows * sizeof(Scalar)
+    m_device.memset(buffer, 0, rows * sizeof(Scalar));
+
+    internal::general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,false,RhsScalar,RhsMapper,false>::run(
+        rows, cols, lhs, rhs,
+        buffer, resIncr, alpha);
+
+    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
+    m_output_kernel(OutputMapper(buffer, rows), m_tensor_contraction_params,
+                    static_cast<Index>(0), static_cast<Index>(0), rows,
+                    static_cast<Index>(1));
+  }
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
+  #if !defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
+  void evalGemm(Scalar* buffer) const {
+    // columns in left side, rows in right side
+    const Index k = this->m_k_size;
+    this->template evalGemmPartial<lhs_inner_dim_contiguous,
+                                   rhs_inner_dim_contiguous,
+                                   rhs_inner_dim_reordered,
+                                   Alignment, true>(buffer, 0, k, 1);
+  }
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous,
+      bool rhs_inner_dim_reordered, int Alignment>
+  EIGEN_DEVICE_FUNC void evalGemmPartialWithoutOutputKernel(
+      Scalar* buffer, Index k_start, Index k_end, int num_threads) const {
+    evalGemmPartial<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous,
+                    rhs_inner_dim_reordered, Alignment,
+        /*use_output_kernel*/ false>(buffer, k_start, k_end,
+                                     num_threads);
+  }
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment, bool use_output_kernel>
+  EIGEN_DEVICE_FUNC void evalGemmPartial(Scalar* buffer, Index k_start, Index k_end, int num_threads) const {
+    eigen_assert(k_end >= k_start && k_start >= 0 && k_end <= this->m_k_size);
+    // columns in slice on left side, rows on right side
+    const Index k_slice = k_end - k_start;
+
+    // rows in left side
+    const Index m = this->m_i_size;
+
+    // columns in right side
+    const Index n = this->m_j_size;
+
+    // define data mappers for Lhs and Rhs
+    typedef typename internal::remove_const<typename EvalLeftArgType::Scalar>::type LhsScalar;
+    typedef typename internal::remove_const<typename EvalRightArgType::Scalar>::type RhsScalar;
+
+    typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
+    typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
+
+    const Index lhs_packet_size = internal::unpacket_traits<typename LeftEvaluator::PacketReturnType>::size;
+    const Index rhs_packet_size = internal::unpacket_traits<typename RightEvaluator::PacketReturnType>::size;
+
+    typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs,
+                                                   LeftEvaluator, left_nocontract_t,
+                                                   contract_t, lhs_packet_size,
+                                                   lhs_inner_dim_contiguous,
+                                                   false, Unaligned> LhsMapper;
+
+    typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs,
+                                                   RightEvaluator, right_nocontract_t,
+                                                   contract_t, rhs_packet_size,
+                                                   rhs_inner_dim_contiguous,
+                                                   rhs_inner_dim_reordered, Unaligned> RhsMapper;
+
+    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
+
+    typedef internal::TensorContractionKernel<
+        Scalar, LhsScalar, RhsScalar, Index, OutputMapper, LhsMapper, RhsMapper>
+        TensorContractionKernel;
+
+    // initialize data mappers
+    LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides,
+                  this->m_left_contracting_strides, this->m_k_strides);
+
+    RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides,
+                  this->m_right_contracting_strides, this->m_k_strides);
+
+    OutputMapper output(buffer, m);
+
+    // Sizes of the blocks to load in cache. See the Goto paper for details.
+    internal::TensorContractionBlocking<Scalar, LhsScalar, RhsScalar,
+                                        Index, internal::ShardByCol>
+        blocking(k_slice, m, n, num_threads);
+    const Index kc = blocking.kc();
+    const Index mc = numext::mini(m, blocking.mc());
+    const Index nc = numext::mini(n, blocking.nc());
+
+    typedef typename TensorContractionKernel::LhsBlock LhsBlock;
+    typedef typename TensorContractionKernel::RhsBlock RhsBlock;
+
+    LhsBlock blockA;
+    RhsBlock blockB;
+
+    TensorContractionKernel kernel(m, k_slice, n, mc, kc, nc);
+
+    typedef typename TensorContractionKernel::BlockMemHandle BlockMemHandle;
+    const BlockMemHandle packed_mem =
+        kernel.allocate(this->m_device, &blockA, &blockB);
+
+    // If a contraction kernel does not support beta, explicitly initialize
+    // output buffer with zeroes.
+    if (!TensorContractionKernel::HasBeta) {
+      this->m_device.memset(buffer, 0, m * n * sizeof(Scalar));
+    }
+
+    for(Index i2=0; i2<m; i2+=mc)
+    {
+      const Index actual_mc = numext::mini(i2+mc,m)-i2;
+      for (Index k2 = k_start; k2 < k_end; k2 += kc) {
+        // make sure we don't overshoot right edge of left matrix, then pack vertical panel
+        const Index actual_kc = numext::mini(k2 + kc, k_end) - k2;
+        kernel.packLhs(&blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
+
+        // If kernel supports beta, there is no need to initialize output
+        // buffer with zeroes.
+        const Scalar alpha = Scalar(1);
+        const Scalar beta = (TensorContractionKernel::HasBeta && k2 == k_start)
+                                ? Scalar(0)
+                                : Scalar(1);
+
+        // series of horizontal blocks
+        for (Index j2 = 0; j2 < n; j2 += nc) {
+          // make sure we don't overshoot right edge of right matrix, then pack block
+          const Index actual_nc = numext::mini(j2 + nc, n) - j2;
+          kernel.packRhs(&blockB, rhs.getSubMapper(k2, j2), actual_kc,
+                         actual_nc);
+
+          // call gebp (matrix kernel)
+          // The parameters here are copied from Eigen's GEMM implementation
+          const OutputMapper output_mapper = output.getSubMapper(i2, j2);
+          kernel.invoke(output_mapper, blockA, blockB, actual_mc, actual_kc,
+                        actual_nc, alpha, beta);
+
+          // We are done with this [i2, j2] output block.
+          if (use_output_kernel && k2 + kc >= k_end) {
+            m_output_kernel(output_mapper, m_tensor_contraction_params, i2, j2,
+                            actual_mc, actual_nc);
+          }
+        }
+      }
+    }
+
+    kernel.deallocate(this->m_device, packed_mem);
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_leftImpl.cleanup();
+    m_rightImpl.cleanup();
+
+    if (m_result != NULL) {
+      m_device.deallocate(m_result);
+      m_result = NULL;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    return m_result[index];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool) const {
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_result + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE EvaluatorPointerType data() const { return m_result; }
+
+protected:
+  Dimensions m_dimensions;
+
+  contract_t m_k_strides;
+  contract_t m_left_contracting_strides;
+  contract_t m_right_contracting_strides;
+
+  bool m_lhs_inner_dim_contiguous;
+  bool m_rhs_inner_dim_contiguous;
+  bool m_rhs_inner_dim_reordered;
+
+  left_nocontract_t m_i_strides;
+  right_nocontract_t m_j_strides;
+  left_nocontract_t m_left_nocontract_strides;
+  right_nocontract_t m_right_nocontract_strides;
+
+  Index m_i_size;
+  Index m_j_size;
+  Index m_k_size;
+
+  TensorContractionParams m_tensor_contraction_params;
+
+  TensorEvaluator<EvalLeftArgType, Device> m_leftImpl;
+  TensorEvaluator<EvalRightArgType, Device> m_rightImpl;
+  const Device EIGEN_DEVICE_REF m_device;
+  OutputKernelType m_output_kernel;
+  EvaluatorPointerType m_result;
+};
+
+
+// evaluator for default device
+template<typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType, typename Device>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> :
+    public TensorContractionEvaluatorBase<
+      TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> > {
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
+  typedef TensorContractionEvaluatorBase<Self> Base;
+
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+
+  enum {
+    Layout = TensorEvaluator<LeftArgType, Device>::Layout
+  };
+
+  // Most of the code is assuming that both input tensors are ColMajor. If the
+  // inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
+  // If we want to compute A * B = C, where A is LHS and B is RHS, the code
+  // will pretend B is LHS and A is RHS.
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType;
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType;
+
+  static const int LDims =
+      internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
+  static const int RDims =
+      internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
+  static const int ContractDims = internal::array_size<Indices>::value;
+
+  typedef array<Index, ContractDims> contract_t;
+  typedef array<Index, LDims - ContractDims> left_nocontract_t;
+  typedef array<Index, RDims - ContractDims> right_nocontract_t;
+
+  static const int NumDims = LDims + RDims - 2 * ContractDims;
+
+  // Could we use NumDimensions here?
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  TensorEvaluator(const XprType& op, const Device& device) :
+      Base(op, device) { }
+
+  template <int Alignment>
+  void evalProduct(Scalar* buffer) const {
+    TENSOR_CONTRACTION_DISPATCH(this->template evalProductSequential, Alignment, (buffer));
+  }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionBlocking.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionBlocking.h
new file mode 100644
index 0000000..974feb0
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionBlocking.h
@@ -0,0 +1,73 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_BLOCKING_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_BLOCKING_H
+
+
+namespace Eigen {
+namespace internal {
+
+enum {
+  ShardByRow = 0,
+  ShardByCol = 1
+};
+
+
+// Default Blocking Strategy
+template<typename ResScalar, typename LhsScalar, typename RhsScalar, typename StorageIndex, int ShardingType = ShardByCol>
+class TensorContractionBlocking {
+ public:
+
+ /*
+   adding EIGEN_DEVICE_FUNC unconditionally to 'TensorContractionBlocking' constructor in `TensorContractionBlocking.h`
+     requires adding EIGEN_DEVICE_FUNC to `computeProductBlockingSizes` in `GeneralBlockPanelKernel.h`
+     which in turn, requires adding EIGEN_DEVICE_FUNC to `evaluateProductBlockingSizesHeuristic` in `GeneralBlockPanelKernel.h`
+     which in turn, requires adding EIGEN_DEVICE_FUNC to `manage_caching_sizes` in `GeneralBlockPanelKernel.h`
+     (else HIPCC will error out)
+
+   However adding EIGEN_DEVICE_FUNC to `manage_caching_sizes` in `GeneralBlockPanelKernel.h`
+   results in NVCC erroring out with the following error
+
+   ../Eigen/src/Core/products/GeneralBlockPanelKernel.h(57): error #2901:
+      dynamic initialization is not supported for function-scope static variables within a __device__/__global__ function
+ */
+
+  #if !defined(EIGEN_HIPCC)
+  EIGEN_DEVICE_FUNC
+  #endif
+ TensorContractionBlocking(StorageIndex k, StorageIndex m, StorageIndex n, StorageIndex num_threads = 1) :
+      kc_(k), mc_(m), nc_(n)
+  {
+    if (ShardingType == ShardByCol) {
+      computeProductBlockingSizes<LhsScalar, RhsScalar, 1>(kc_, mc_, nc_, num_threads);
+    }
+    else {
+      computeProductBlockingSizes<LhsScalar, RhsScalar, 1>(kc_, nc_, mc_, num_threads);
+    }
+
+    const int rhs_packet_size = internal::packet_traits<RhsScalar>::size;
+    kc_ = (rhs_packet_size <= 8 || kc_ <= rhs_packet_size) ?
+      kc_ : (kc_ / rhs_packet_size) * rhs_packet_size;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE StorageIndex kc() const { return kc_; }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE StorageIndex mc() const { return mc_; }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE StorageIndex nc() const { return nc_; }
+
+ private:
+  StorageIndex kc_;
+  StorageIndex mc_;
+  StorageIndex nc_;
+};
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_BLOCKING_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionCuda.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionCuda.h
new file mode 100644
index 0000000..3f315fe
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionCuda.h
@@ -0,0 +1,6 @@
+
+#if defined(__clang__) || defined(__GNUC__)
+#warning "Deprecated header file, please either include the main Eigen/CXX11/Tensor header or the respective TensorContractionGpu.h file"
+#endif
+
+#include "TensorContractionGpu.h"
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionGpu.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionGpu.h
new file mode 100644
index 0000000..c818038
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionGpu.h
@@ -0,0 +1,1413 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014-2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2015 Navdeep Jaitly <ndjaitly@google.com>
+// Copyright (C) 2014 Eric Martin <eric@ericmart.in>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H
+
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
+
+namespace Eigen {
+
+template<typename Scalar, typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper, bool needs_edge_check>
+__device__ EIGEN_STRONG_INLINE void
+EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
+                               const OutputMapper output, Scalar* lhs_shmem, Scalar* rhs_shmem,
+                       const Index m_size, const Index n_size, const Index k_size) {
+
+  const Index m_block_idx = blockIdx.x;
+  const Index n_block_idx = blockIdx.y;
+
+  const Index base_m = 64 * m_block_idx;
+  const Index base_n = 64 * n_block_idx;
+
+  // declare and initialize 64 registers for output 8x8 block
+
+  // prefetch registers
+  Scalar lhs_pf0;
+  Scalar lhs_pf1;
+  Scalar lhs_pf2;
+  Scalar lhs_pf3;
+  Scalar lhs_pf4;
+  Scalar lhs_pf5;
+  Scalar lhs_pf6;
+  Scalar lhs_pf7;
+
+  Scalar rhs_pf0;
+  Scalar rhs_pf1;
+  Scalar rhs_pf2;
+  Scalar rhs_pf3;
+  Scalar rhs_pf4;
+  Scalar rhs_pf5;
+  Scalar rhs_pf6;
+  Scalar rhs_pf7;
+
+  // shared memory is formatted
+  // (contract idx in block, nocontract idx in block, block idx)
+  // where block idx is column major. This transposition limits the number of
+  // bank conflicts when reading the LHS. The core idea is that since the contracting
+  // index is shared by both sides, then the contracting index should be in threadIdx.x.
+
+  // On the LHS, we pad each row inside of each block with an extra element. This makes
+  // each block 8 rows of 9 elements, which is 72 elements. This gives no bank conflicts
+  // on writes and very few 2-way conflicts on reads. There is an 8x8 grid of these blocks.
+
+  // On the RHS we just add 8 padding elements to the end of each block. This gives no bank
+  // conflicts on writes and also none on reads.
+
+  // storage indices
+  const Index lhs_store_idx_base = threadIdx.y * 72 + threadIdx.x * 9 + threadIdx.z;
+  const Index rhs_store_idx_base = threadIdx.y * 72 + threadIdx.z * 8 + threadIdx.x;
+
+  const Index lhs_store_idx_0 = lhs_store_idx_base + 576 * 0;
+  const Index lhs_store_idx_1 = lhs_store_idx_base + 576 * 1;
+  const Index lhs_store_idx_2 = lhs_store_idx_base + 576 * 2;
+  const Index lhs_store_idx_3 = lhs_store_idx_base + 576 * 3;
+  const Index lhs_store_idx_4 = lhs_store_idx_base + 576 * 4;
+  const Index lhs_store_idx_5 = lhs_store_idx_base + 576 * 5;
+  const Index lhs_store_idx_6 = lhs_store_idx_base + 576 * 6;
+  const Index lhs_store_idx_7 = lhs_store_idx_base + 576 * 7;
+
+  const Index rhs_store_idx_0 = rhs_store_idx_base + 576 * 0;
+  const Index rhs_store_idx_1 = rhs_store_idx_base + 576 * 1;
+  const Index rhs_store_idx_2 = rhs_store_idx_base + 576 * 2;
+  const Index rhs_store_idx_3 = rhs_store_idx_base + 576 * 3;
+  const Index rhs_store_idx_4 = rhs_store_idx_base + 576 * 4;
+  const Index rhs_store_idx_5 = rhs_store_idx_base + 576 * 5;
+  const Index rhs_store_idx_6 = rhs_store_idx_base + 576 * 6;
+  const Index rhs_store_idx_7 = rhs_store_idx_base + 576 * 7;
+
+  // in the loading code, the following variables are important:
+  // threadIdx.x: the vertical position in an 8x8 block
+  // threadIdx.y: the vertical index of the 8x8 block in the grid
+  // threadIdx.z: the horizontal position in an 8x8 block
+  // k: the horizontal index of the 8x8 block in the grid
+  //
+  // The k parameter is implicit (it was the loop counter for a loop that went
+  // from 0 to <8, but now that loop is unrolled in the below code.
+
+  const Index load_idx_vert = threadIdx.x + 8 * threadIdx.y;
+  const Index lhs_vert = base_m + load_idx_vert;
+
+#define prefetchIntoRegisters(base_k)                           \
+  {                                                             \
+    lhs_pf0 = conv(0);                                          \
+    lhs_pf1 = conv(0);                                          \
+    lhs_pf2 = conv(0);                                          \
+    lhs_pf3 = conv(0);                                          \
+    lhs_pf4 = conv(0);                                          \
+    lhs_pf5 = conv(0);                                          \
+    lhs_pf6 = conv(0);                                          \
+    lhs_pf7 = conv(0);                                          \
+                                                                \
+    rhs_pf0 = conv(0);                                          \
+    rhs_pf1 = conv(0);                                          \
+    rhs_pf2 = conv(0);                                          \
+    rhs_pf3 = conv(0);                                          \
+    rhs_pf4 = conv(0);                                          \
+    rhs_pf5 = conv(0);                                          \
+    rhs_pf6 = conv(0);                                          \
+    rhs_pf7 = conv(0);                                          \
+                                                                \
+    if (!needs_edge_check || lhs_vert < m_size) {               \
+      const Index lhs_horiz_0 = base_k + threadIdx.z + 0 * 8;   \
+      const Index lhs_horiz_1 = base_k + threadIdx.z + 1 * 8;   \
+      const Index lhs_horiz_2 = base_k + threadIdx.z + 2 * 8;   \
+      const Index lhs_horiz_3 = base_k + threadIdx.z + 3 * 8;   \
+      const Index lhs_horiz_4 = base_k + threadIdx.z + 4 * 8;   \
+      const Index lhs_horiz_5 = base_k + threadIdx.z + 5 * 8;   \
+      const Index lhs_horiz_6 = base_k + threadIdx.z + 6 * 8;   \
+      const Index lhs_horiz_7 = base_k + threadIdx.z + 7 * 8;   \
+                                                                \
+      if (!needs_edge_check || lhs_horiz_7 < k_size) {          \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
+        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
+        lhs_pf6 = lhs(lhs_vert, lhs_horiz_6);                   \
+        lhs_pf7 = lhs(lhs_vert, lhs_horiz_7);                   \
+      } else if (lhs_horiz_6 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
+        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
+        lhs_pf6 = lhs(lhs_vert, lhs_horiz_6);                   \
+      } else if (lhs_horiz_5 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
+        lhs_pf5 = lhs(lhs_vert, lhs_horiz_5);                   \
+      } else if (lhs_horiz_4 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+        lhs_pf4 = lhs(lhs_vert, lhs_horiz_4);                   \
+      } else if (lhs_horiz_3 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+        lhs_pf3 = lhs(lhs_vert, lhs_horiz_3);                   \
+      } else if (lhs_horiz_2 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+        lhs_pf2 = lhs(lhs_vert, lhs_horiz_2);                   \
+      } else if (lhs_horiz_1 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+        lhs_pf1 = lhs(lhs_vert, lhs_horiz_1);                   \
+      } else if (lhs_horiz_0 < k_size) {                        \
+        lhs_pf0 = lhs(lhs_vert, lhs_horiz_0);                   \
+      }                                                         \
+    }                                                           \
+                                                                \
+    const Index rhs_vert = base_k + load_idx_vert;              \
+    if (!needs_edge_check || rhs_vert < k_size) {               \
+      const Index rhs_horiz_0 = base_n + threadIdx.z + 0 * 8;   \
+      const Index rhs_horiz_1 = base_n + threadIdx.z + 1 * 8;   \
+      const Index rhs_horiz_2 = base_n + threadIdx.z + 2 * 8;   \
+      const Index rhs_horiz_3 = base_n + threadIdx.z + 3 * 8;   \
+      const Index rhs_horiz_4 = base_n + threadIdx.z + 4 * 8;   \
+      const Index rhs_horiz_5 = base_n + threadIdx.z + 5 * 8;   \
+      const Index rhs_horiz_6 = base_n + threadIdx.z + 6 * 8;   \
+      const Index rhs_horiz_7 = base_n + threadIdx.z + 7 * 8;   \
+                                                                \
+      if (rhs_horiz_7 < n_size) {                               \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
+        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
+        rhs_pf6 = rhs(rhs_vert, rhs_horiz_6);                   \
+        rhs_pf7 = rhs(rhs_vert, rhs_horiz_7);                   \
+      } else if (rhs_horiz_6 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
+        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
+        rhs_pf6 = rhs(rhs_vert, rhs_horiz_6);                   \
+      } else if (rhs_horiz_5 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
+        rhs_pf5 = rhs(rhs_vert, rhs_horiz_5);                   \
+      } else if (rhs_horiz_4 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+        rhs_pf4 = rhs(rhs_vert, rhs_horiz_4);                   \
+      } else if (rhs_horiz_3 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+        rhs_pf3 = rhs(rhs_vert, rhs_horiz_3);                   \
+      } else if (rhs_horiz_2 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+        rhs_pf2 = rhs(rhs_vert, rhs_horiz_2);                   \
+      } else if (rhs_horiz_1 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+        rhs_pf1 = rhs(rhs_vert, rhs_horiz_1);                   \
+      } else if (rhs_horiz_0 < n_size) {                        \
+        rhs_pf0 = rhs(rhs_vert, rhs_horiz_0);                   \
+      }                                                         \
+    }                                                           \
+  }                                                             \
+
+#define writeRegToShmem(_)                      \
+  lhs_shmem[lhs_store_idx_0] = lhs_pf0;         \
+  rhs_shmem[rhs_store_idx_0] = rhs_pf0;         \
+                                                \
+  lhs_shmem[lhs_store_idx_1] = lhs_pf1;         \
+  rhs_shmem[rhs_store_idx_1] = rhs_pf1;         \
+                                                \
+  lhs_shmem[lhs_store_idx_2] = lhs_pf2;         \
+  rhs_shmem[rhs_store_idx_2] = rhs_pf2;         \
+                                                \
+  lhs_shmem[lhs_store_idx_3] = lhs_pf3;         \
+  rhs_shmem[rhs_store_idx_3] = rhs_pf3;         \
+                                                \
+  lhs_shmem[lhs_store_idx_4] = lhs_pf4;         \
+  rhs_shmem[rhs_store_idx_4] = rhs_pf4;         \
+                                                \
+  lhs_shmem[lhs_store_idx_5] = lhs_pf5;         \
+  rhs_shmem[rhs_store_idx_5] = rhs_pf5;         \
+                                                \
+  lhs_shmem[lhs_store_idx_6] = lhs_pf6;         \
+  rhs_shmem[rhs_store_idx_6] = rhs_pf6;         \
+                                                \
+  lhs_shmem[lhs_store_idx_7] = lhs_pf7;         \
+  rhs_shmem[rhs_store_idx_7] = rhs_pf7;         \
+
+  // declare and initialize result array
+#define res(i, j) _res_##i##j
+#define initResultRow(i)                        \
+  Scalar res(i, 0) = conv(0);                   \
+  Scalar res(i, 1) = conv(0);                   \
+  Scalar res(i, 2) = conv(0);                   \
+  Scalar res(i, 3) = conv(0);                   \
+  Scalar res(i, 4) = conv(0);                   \
+  Scalar res(i, 5) = conv(0);                   \
+  Scalar res(i, 6) = conv(0);                   \
+  Scalar res(i, 7) = conv(0);                   \
+
+  internal::scalar_cast_op<int, Scalar> conv;
+  initResultRow(0);
+  initResultRow(1);
+  initResultRow(2);
+  initResultRow(3);
+  initResultRow(4);
+  initResultRow(5);
+  initResultRow(6);
+  initResultRow(7);
+#undef initResultRow
+
+  for (Index base_k = 0; base_k < k_size; base_k += 64) {
+    // wait for previous iteration to finish with shmem. Despite common sense,
+    // the code is a bit faster with this here then at bottom of loop
+    __syncthreads();
+
+    prefetchIntoRegisters(base_k);
+    writeRegToShmem();
+
+    #undef prefetchIntoRegisters
+    #undef writeRegToShmem
+
+    // wait for shared mem packing to be done before starting computation
+    __syncthreads();
+
+    // compute 8x8 matrix product by outer product. This involves packing one column
+    // of LHS and one row of RHS into registers (takes 16 registers).
+
+#define lcol(i) _lcol##i
+    Scalar lcol(0);
+    Scalar lcol(1);
+    Scalar lcol(2);
+    Scalar lcol(3);
+    Scalar lcol(4);
+    Scalar lcol(5);
+    Scalar lcol(6);
+    Scalar lcol(7);
+
+#define rrow(j) _rrow##j
+    Scalar rrow(0);
+    Scalar rrow(1);
+    Scalar rrow(2);
+    Scalar rrow(3);
+    Scalar rrow(4);
+    Scalar rrow(5);
+    Scalar rrow(6);
+    Scalar rrow(7);
+
+    // Now x corresponds to k, y to m, and z to n
+    const Scalar* lhs_block = &lhs_shmem[threadIdx.x + 9 * threadIdx.y];
+    const Scalar* rhs_block = &rhs_shmem[threadIdx.x + 8 * threadIdx.z];
+
+#define lhs_element(i, j) lhs_block[72 * ((i) + 8 * (j))]
+#define rhs_element(i, j) rhs_block[72 * ((i) + 8 * (j))]
+
+#define loadData(i, j)                          \
+    lcol(0) = lhs_element(0, j);               \
+    rrow(0) = rhs_element(i, 0);               \
+    lcol(1) = lhs_element(1, j);               \
+    rrow(1) = rhs_element(i, 1);               \
+    lcol(2) = lhs_element(2, j);               \
+    rrow(2) = rhs_element(i, 2);               \
+    lcol(3) = lhs_element(3, j);               \
+    rrow(3) = rhs_element(i, 3);               \
+    lcol(4) = lhs_element(4, j);               \
+    rrow(4) = rhs_element(i, 4);               \
+    lcol(5) = lhs_element(5, j);               \
+    rrow(5) = rhs_element(i, 5);               \
+    lcol(6) = lhs_element(6, j);               \
+    rrow(6) = rhs_element(i, 6);               \
+    lcol(7) = lhs_element(7, j);               \
+    rrow(7) = rhs_element(i, 7);               \
+
+#define computeCol(j)                           \
+    res(0, j) += lcol(0) * rrow(j);             \
+    res(1, j) += lcol(1) * rrow(j);             \
+    res(2, j) += lcol(2) * rrow(j);             \
+    res(3, j) += lcol(3) * rrow(j);             \
+    res(4, j) += lcol(4) * rrow(j);             \
+    res(5, j) += lcol(5) * rrow(j);             \
+    res(6, j) += lcol(6) * rrow(j);             \
+    res(7, j) += lcol(7) * rrow(j);             \
+
+#define computePass(i)                          \
+    loadData(i, i);                             \
+                                                \
+    computeCol(0);                              \
+    computeCol(1);                              \
+    computeCol(2);                              \
+    computeCol(3);                              \
+    computeCol(4);                              \
+    computeCol(5);                              \
+    computeCol(6);                              \
+    computeCol(7);                              \
+
+    computePass(0);
+    computePass(1);
+    computePass(2);
+    computePass(3);
+    computePass(4);
+    computePass(5);
+    computePass(6);
+    computePass(7);
+
+#undef lcol
+#undef rrow
+#undef lhs_element
+#undef rhs_element
+#undef loadData
+#undef computeCol
+#undef computePass
+  } // end loop over k
+
+  // we've now iterated over all of the large (ie width 64) k blocks and
+  // accumulated results in registers. At this point thread (x, y, z) contains
+  // the sum across all big k blocks of the product of little k block of index (x, y)
+  // with block of index (y, z). To compute the final output, we need to reduce
+  // the 8 threads over y by summation.
+#if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000)
+#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask)
+#else
+#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor_sync(0xFFFFFFFF, res(i, j), mask)
+#endif
+
+#define reduceRow(i, mask)                      \
+  shuffleInc(i, 0, mask);                       \
+  shuffleInc(i, 1, mask);                       \
+  shuffleInc(i, 2, mask);                       \
+  shuffleInc(i, 3, mask);                       \
+  shuffleInc(i, 4, mask);                       \
+  shuffleInc(i, 5, mask);                       \
+  shuffleInc(i, 6, mask);                       \
+  shuffleInc(i, 7, mask);                       \
+
+#define reduceMatrix(mask)                      \
+  reduceRow(0, mask);                           \
+  reduceRow(1, mask);                           \
+  reduceRow(2, mask);                           \
+  reduceRow(3, mask);                           \
+  reduceRow(4, mask);                           \
+  reduceRow(5, mask);                           \
+  reduceRow(6, mask);                           \
+  reduceRow(7, mask);                           \
+
+  // actually perform the reduction, now each thread of index (_, y, z)
+  // contains the correct values in its registers that belong in the output
+  // block
+  reduceMatrix(1);
+  reduceMatrix(2);
+  reduceMatrix(4);
+
+#undef shuffleInc
+#undef reduceRow
+#undef reduceMatrix
+
+  // now we need to copy the 64 values into main memory. We can't split work
+  // among threads because all variables are in registers. There's 2 ways
+  // to do this:
+  // (1) have 1 thread do 64 writes from registers into global memory
+  // (2) have 1 thread do 64 writes into shared memory, and then 8 threads
+  //     each do 8 writes into global memory. We can just overwrite the shared
+  //     memory from the problem we just solved.
+  // (2) is slightly faster than (1) due to less branching and more ILP
+
+  // TODO: won't yield much gain, but could just use currently unused shared mem
+  //       and then we won't have to sync
+  // wait for shared mem to be out of use
+  __syncthreads();
+
+#define writeResultShmem(i, j)                                          \
+  lhs_shmem[i + 8 * threadIdx.y + 64 * threadIdx.z + 512 * j] = res(i, j); \
+
+#define writeRow(i)                             \
+  writeResultShmem(i, 0);                       \
+  writeResultShmem(i, 1);                       \
+  writeResultShmem(i, 2);                       \
+  writeResultShmem(i, 3);                       \
+  writeResultShmem(i, 4);                       \
+  writeResultShmem(i, 5);                       \
+  writeResultShmem(i, 6);                       \
+  writeResultShmem(i, 7);                       \
+
+  if (threadIdx.x == 0) {
+    writeRow(0);
+    writeRow(1);
+    writeRow(2);
+    writeRow(3);
+    writeRow(4);
+    writeRow(5);
+    writeRow(6);
+    writeRow(7);
+  }
+#undef writeResultShmem
+#undef writeRow
+
+  const int max_i_write = numext::mini((int)((m_size - base_m - threadIdx.y + 7) / 8), 8);
+  const int max_j_write = numext::mini((int)((n_size - base_n - threadIdx.z + 7) / 8), 8);
+
+  if (threadIdx.x < max_i_write) {
+    if (max_j_write == 8) {
+      // TODO: can i trade bank conflicts for coalesced writes?
+      Scalar val0 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 0];
+      Scalar val1 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 1];
+      Scalar val2 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 2];
+      Scalar val3 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 3];
+      Scalar val4 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 4];
+      Scalar val5 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 5];
+      Scalar val6 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 6];
+      Scalar val7 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 7];
+
+      output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 0) = val0;
+      output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 1) = val1;
+      output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 2) = val2;
+      output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 3) = val3;
+      output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 4) = val4;
+      output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 5) = val5;
+      output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 6) = val6;
+      output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 7) = val7;
+    } else {
+#pragma unroll 7
+      for (int j = 0; j < max_j_write; j++) {
+        Scalar val = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * j];
+        output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * j) = val;
+      }
+    }
+  }
+#undef res
+}
+
+
+template<typename Scalar, typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper>
+__global__ void
+#if defined(EIGEN_HIPCC)
+__launch_bounds__(512, 1)
+#else
+__launch_bounds__(512)
+#endif
+EigenContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output,
+                       const Index m_size, const Index n_size, const Index k_size) {
+  __shared__ Scalar lhs_shmem[72 * 64];
+  __shared__ Scalar rhs_shmem[72 * 64];
+
+  const Index m_block_idx = blockIdx.x;
+  const Index n_block_idx = blockIdx.y;
+
+  const Index base_m = 64 * m_block_idx;
+  const Index base_n = 64 * n_block_idx;
+
+  if (base_m + 63 < m_size && base_n + 63 < n_size) {
+    EigenContractionKernelInternal<Scalar, Index, LhsMapper, RhsMapper, OutputMapper, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size);
+  } else {
+    EigenContractionKernelInternal<Scalar, Index, LhsMapper, RhsMapper, OutputMapper, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size);
+  }
+}
+
+
+template<typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper, bool CHECK_LHS_BOUNDARY,
+         bool CHECK_RHS_BOUNDARY>
+__device__ __forceinline__ void
+EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output, float2 lhs_shmem2[][16],
+                       float2 rhs_shmem2[][8], const Index m_size,
+                       const Index n_size, const Index k_size,
+                       const Index base_m, const Index base_n) {
+
+  // prefetch registers
+  float4 lhs_pf0, rhs_pf0;
+
+  float4 results[4];
+  for (int i=0; i < 4; i++) {
+    results[i].x = results[i].y = results[i].z = results[i].w = 0;
+  }
+
+#define prefetch_lhs(reg, row, col)                            \
+    if (!CHECK_LHS_BOUNDARY) {                                 \
+      if (col < k_size) {                                      \
+        reg =lhs.template loadPacket<float4,Unaligned>(row, col);     \
+      }                                                        \
+    } else {                                                   \
+      if (col < k_size) {                                      \
+        if (row + 3 < m_size) {                                \
+          reg =lhs.template loadPacket<float4,Unaligned>(row, col);   \
+        } else if (row + 2 < m_size) {                         \
+          reg.x =lhs(row + 0, col);                            \
+          reg.y =lhs(row + 1, col);                            \
+          reg.z =lhs(row + 2, col);                            \
+        } else if (row + 1 < m_size) {                         \
+          reg.x =lhs(row + 0, col);                            \
+          reg.y =lhs(row + 1, col);                            \
+        } else if (row  < m_size) {                            \
+          reg.x =lhs(row + 0, col);                            \
+        }                                                      \
+      }                                                        \
+    }							       \
+
+  Index lhs_vert = base_m+threadIdx.x*4;
+
+  for (Index k = 0; k < k_size; k += 16) {
+
+    lhs_pf0 = internal::pset1<float4>(0);
+    rhs_pf0 = internal::pset1<float4>(0);
+
+    Index lhs_horiz = threadIdx.y+k;
+    prefetch_lhs(lhs_pf0, lhs_vert, lhs_horiz)
+
+    Index rhs_vert = k+(threadIdx.x%4)*4;
+    Index rhs_horiz0 = (threadIdx.x>>2)+threadIdx.y*4+base_n;
+
+    if (!CHECK_RHS_BOUNDARY) {
+      if ((rhs_vert + 3) < k_size) {
+        // just CHECK_RHS_BOUNDARY
+        rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+      } else if (rhs_vert + 2 < k_size) {
+        // just CHECK_RHS_BOUNDARY
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+      } else if (rhs_vert + 1 < k_size) {
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+      } else if (rhs_vert  < k_size) {
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+      }
+    } else {
+      if (rhs_horiz0 < n_size) {
+        if ((rhs_vert + 3) < k_size) {
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+        } else if ((rhs_vert + 2) < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+        } else if ((rhs_vert + 1) < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        } else if (rhs_vert  < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        }
+      }
+    }
+    float x1, x2 ;
+    // the following can be a bitwise operation..... some day.
+    if((threadIdx.x%8) < 4) {
+      x1 = rhs_pf0.y;
+      x2 = rhs_pf0.w;
+    } else {
+      x1 = rhs_pf0.x;
+      x2 = rhs_pf0.z;
+    }
+    #if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000)
+    x1 = __shfl_xor(x1, 4);
+    x2 = __shfl_xor(x2, 4);
+    #else
+    x1 = __shfl_xor_sync(0xFFFFFFFF, x1, 4);
+    x2 = __shfl_xor_sync(0xFFFFFFFF, x2, 4);
+    #endif
+    if((threadIdx.x%8) < 4) {
+      rhs_pf0.y = x1;
+      rhs_pf0.w = x2;
+    } else {
+      rhs_pf0.x = x1;
+      rhs_pf0.z = x2;
+    }
+
+    // We have 64 features.
+    // Row 0 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 0, 1.
+    // Row 1 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 2, 3.
+    // ...
+    // Row 31 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 62, 63
+    // Row 32 -> times (2, 6, 10, 14, 3, 7, 11, 15) for features 0, 1
+    // ...
+    rhs_shmem2[(threadIdx.x>>3)+ threadIdx.y*2][threadIdx.x%8] = make_float2(rhs_pf0.x, rhs_pf0.y);
+    rhs_shmem2[(threadIdx.x>>3)+ threadIdx.y*2+32][threadIdx.x%8] = make_float2(rhs_pf0.z, rhs_pf0.w);
+
+    // Row 0 (time 0) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
+    // Row 1 (time 1) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
+    // ...
+    // Row 15 (time 15) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61)
+    // Row 16 (time 0) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63)
+    // ...
+
+    lhs_shmem2[threadIdx.y][threadIdx.x] = make_float2(lhs_pf0.x, lhs_pf0.y);
+    lhs_shmem2[threadIdx.y+16][threadIdx.x] = make_float2(lhs_pf0.z, lhs_pf0.w);
+
+
+#define add_vals(fl1, fl2, fr1, fr2)\
+    results[0].x += fl1.x * fr1.x;\
+    results[0].y += fl1.y * fr1.x;\
+    results[0].z += fl2.x * fr1.x;\
+    results[0].w += fl2.y * fr1.x;\
+\
+    results[1].x += fl1.x * fr1.y;\
+    results[1].y += fl1.y * fr1.y;\
+    results[1].z += fl2.x * fr1.y;\
+    results[1].w += fl2.y * fr1.y;\
+\
+    results[2].x += fl1.x * fr2.x;\
+    results[2].y += fl1.y * fr2.x;\
+    results[2].z += fl2.x * fr2.x;\
+    results[2].w += fl2.y * fr2.x;\
+\
+    results[3].x += fl1.x * fr2.y;\
+    results[3].y += fl1.y * fr2.y;\
+    results[3].z += fl2.x * fr2.y;\
+    results[3].w += fl2.y * fr2.y;\
+
+    __syncthreads();
+
+    // Do the multiplies.
+    #pragma unroll
+    for (int koff = 0; koff < 16; koff ++) {
+      // 32 x threads.
+      float2 fl1 = lhs_shmem2[koff][threadIdx.x];
+      float2 fl2 = lhs_shmem2[koff + 16][threadIdx.x];
+
+      int start_feature = threadIdx.y * 4;
+      float2 fr1 = rhs_shmem2[(start_feature>>1) + 32*((koff%4)/2)][koff/4 + (koff%2)*4];
+      float2 fr2 = rhs_shmem2[(start_feature>>1) + 1 + 32*((koff%4)/2)][koff/4 + (koff%2)*4];
+
+      add_vals(fl1, fl2, fr1, fr2)
+    }
+    __syncthreads();
+  }
+
+#undef prefetch_lhs
+#undef add_vals
+
+  Index horiz_base = threadIdx.y*4+base_n;
+  if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
+    for (int i = 0; i < 4; i++) {
+      output(lhs_vert, horiz_base + i) = results[i].x;
+      output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      output(lhs_vert + 3, horiz_base + i) = results[i].w;
+    }
+  } else if (!CHECK_RHS_BOUNDARY) {
+    // CHECK LHS
+    if (lhs_vert + 3 < m_size) {
+      for (int i = 0; i < 4; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    } else if (lhs_vert + 2 < m_size) {
+      for (int i = 0; i < 4; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      }
+    } else if (lhs_vert + 1 < m_size) {
+      for (int i = 0; i < 4; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      }
+    } else if (lhs_vert  < m_size) {
+      for (int i = 0; i < 4; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+      }
+    }
+  } else if (!CHECK_LHS_BOUNDARY) {
+    // CHECK RHS
+    /*
+    int ncols_rem = fminf(n_size- horiz_base, 4);
+    for (int i = 0; i < ncols_rem; i++) {
+      output(lhs_vert, horiz_base + i) = results[i].x;
+      output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      output(lhs_vert + 3, horiz_base + i) = results[i].w;
+    }*/
+    for (int i = 0; i < 4; i++) {
+      if (horiz_base+i < n_size) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        output(lhs_vert + 3, horiz_base + i) = results[i].w;
+       }
+    }
+  } else {
+    // CHECK both boundaries.
+    for (int i = 0; i < 4; i++) {
+      if (horiz_base+i < n_size) {
+        if (lhs_vert < m_size)
+          output(lhs_vert, horiz_base + i) = results[i].x;
+        if (lhs_vert + 1 < m_size)
+          output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        if (lhs_vert + 2 < m_size)
+          output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        if (lhs_vert + 3 < m_size)
+          output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    }
+  }
+}
+
+
+template<typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper, bool CHECK_LHS_BOUNDARY,
+         bool CHECK_RHS_BOUNDARY>
+__device__ __forceinline__ void
+EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output, float2 lhs_shmem2[][32],
+                       float2 rhs_shmem2[][8], const Index m_size,
+                       const Index n_size, const Index k_size,
+                       const Index base_m, const Index base_n) {
+
+  // prefetch registers
+  float4 lhs_pf0, lhs_pf1, lhs_pf2, lhs_pf3;
+  float4 rhs_pf0, rhs_pf1;
+
+  float4 results[8];
+  for (int i=0; i < 8; i++) {
+    results[i].x = results[i].y = results[i].z = results[i].w = 0;
+  }
+
+  Index lhs_vert = base_m+threadIdx.x*4+(threadIdx.y%4)*32;
+  for (Index k = 0; k < k_size; k += 32) {
+    lhs_pf0 = internal::pset1<float4>(0);
+    lhs_pf1 = internal::pset1<float4>(0);
+    lhs_pf2 = internal::pset1<float4>(0);
+    lhs_pf3 = internal::pset1<float4>(0);
+
+    rhs_pf0 = internal::pset1<float4>(0);
+    rhs_pf1 = internal::pset1<float4>(0);
+
+     if (!CHECK_LHS_BOUNDARY) {
+      if ((threadIdx.y/4+k+24) < k_size) {
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+        lhs_pf3 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+      } else if ((threadIdx.y/4+k+16) < k_size) {
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+      } else if ((threadIdx.y/4+k+8) < k_size) {
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+      } else if ((threadIdx.y/4+k) < k_size) {
+        lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+      }
+    } else {
+      // just CHECK_LHS_BOUNDARY
+      if (lhs_vert + 3 < m_size) {
+        if ((threadIdx.y/4+k+24) < k_size) {
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+          lhs_pf3 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+24));
+        } else if ((threadIdx.y/4+k+16) < k_size) {
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+          lhs_pf2 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+16));
+        } else if ((threadIdx.y/4+k+8) < k_size) {
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+          lhs_pf1 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k+8));
+        } else if ((threadIdx.y/4+k) < k_size) {
+          lhs_pf0 =lhs.template loadPacket<float4,Unaligned>(lhs_vert, (threadIdx.y/4+k));
+        }
+      } else if (lhs_vert + 2 < m_size) {
+        if ((threadIdx.y/4+k+24) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (threadIdx.y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (threadIdx.y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+16));
+          lhs_pf2.z =lhs(lhs_vert + 2, (threadIdx.y/4+k+16));
+          lhs_pf3.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+24));
+          lhs_pf3.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+24));
+          lhs_pf3.z =lhs(lhs_vert + 2, (threadIdx.y/4+k+24));
+        } else if ((threadIdx.y/4+k+16) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (threadIdx.y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (threadIdx.y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+16));
+          lhs_pf2.z =lhs(lhs_vert + 2, (threadIdx.y/4+k+16));
+        } else if ((threadIdx.y/4+k+8) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (threadIdx.y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+8));
+          lhs_pf1.z =lhs(lhs_vert + 2, (threadIdx.y/4+k+8));
+        } else if ((threadIdx.y/4+k) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (threadIdx.y/4+k));
+          lhs_pf0.z =lhs(lhs_vert + 2, (threadIdx.y/4+k));
+        }
+      } else if (lhs_vert + 1 < m_size) {
+        if ((threadIdx.y/4+k+24) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+16));
+          lhs_pf3.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+24));
+          lhs_pf3.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+24));
+        } else if ((threadIdx.y/4+k+16) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+16));
+          lhs_pf2.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+16));
+        } else if ((threadIdx.y/4+k+8) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+          lhs_pf1.y =lhs(lhs_vert + 1, (threadIdx.y/4+k+8));
+        } else if ((threadIdx.y/4+k) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf0.y =lhs(lhs_vert + 1, (threadIdx.y/4+k));
+        }
+      } else if (lhs_vert < m_size) {
+        if ((threadIdx.y/4+k+24) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+16));
+          lhs_pf3.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+24));
+        } else if ((threadIdx.y/4+k+16) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+          lhs_pf2.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+16));
+        } else if ((threadIdx.y/4+k+8) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+          lhs_pf1.x =lhs(lhs_vert + 0, (threadIdx.y/4+k+8));
+        } else if ((threadIdx.y/4+k) < k_size) {
+          lhs_pf0.x =lhs(lhs_vert + 0, (threadIdx.y/4+k));
+        }
+      }
+    }
+    __syncthreads();
+    Index rhs_vert = k+threadIdx.x*4;
+    Index rhs_horiz0 = threadIdx.y*2+base_n;
+    Index rhs_horiz1 = threadIdx.y*2+1+base_n;
+    if (!CHECK_RHS_BOUNDARY) {
+      if ((rhs_vert + 3) < k_size) {
+        // just CHECK_RHS_BOUNDARY
+        rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+        rhs_pf1 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz1);
+      } else if (rhs_vert + 2 < k_size) {
+        // just CHECK_RHS_BOUNDARY
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+        rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+        rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
+      } else if (rhs_vert + 1 < k_size) {
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+        rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+      } else if (rhs_vert  < k_size) {
+        rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+      }
+    } else {
+      if (rhs_horiz1 < n_size) {
+        if ((rhs_vert + 3) < k_size) {
+          // just CHECK_RHS_BOUNDARY
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+          rhs_pf1 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz1);
+        } else if (rhs_vert + 2 < k_size) {
+          // just CHECK_RHS_BOUNDARY
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+          rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+          rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
+        } else if (k+threadIdx.x*4 + 1 < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+          rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
+        } else if (k+threadIdx.x*4  < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
+        }
+      } else if (rhs_horiz0 < n_size) {
+        if ((rhs_vert + 3) < k_size) {
+          // just CHECK_RHS_BOUNDARY
+          rhs_pf0 = rhs.template loadPacket<float4,Unaligned>(rhs_vert, rhs_horiz0);
+        } else if ((rhs_vert + 2) < k_size) {
+          // just CHECK_RHS_BOUNDARY
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+          rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
+        } else if ((rhs_vert + 1) < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+          rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
+        } else if (rhs_vert  < k_size) {
+          rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
+        }
+      }
+    }
+    __syncthreads();
+    // Loaded. Do computation
+    // Row 0 -> times (0, 4, 8, .. 28) for features 0, 1.
+    // Row 1 -> times (0, 4, 8, .. 28) for features 2, 3.
+    // ..
+    // Row 31 -> times (0, 4, 8, .. 28) for features 62, 63
+    rhs_shmem2[threadIdx.y][threadIdx.x] = make_float2(rhs_pf0.x, rhs_pf1.x);
+    // Row 32 -> times (1, 5, 9, .. 29) for features 0, 1.
+    // Row 33 -> times (1, 5, 9, .. 29) for features 2, 3.
+    // ..
+    rhs_shmem2[threadIdx.y+32][threadIdx.x] = make_float2(rhs_pf0.y, rhs_pf1.y);
+    // Row 64 -> times (2, 6, 10, .. 30) for features 0, 1.
+    // Row 65 -> times (2, 6, 10, .. 30) for features 2, 3.
+    rhs_shmem2[threadIdx.y+64][threadIdx.x] = make_float2(rhs_pf0.z, rhs_pf1.z);
+    // Row 96 -> times (3, 7, 11, .. 31) for features 0, 1.
+    // Row 97 -> times (3, 7, 11, .. 31) for features 2, 3.
+    rhs_shmem2[threadIdx.y+96][threadIdx.x] = make_float2(rhs_pf0.w, rhs_pf1.w);
+
+    // LHS.
+    // Row 0 (time 0) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61) .. (124, 125)
+    // Row 1 (time 1) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), ..  (60, 61) .. (124, 125)
+    // ...
+    // Row 8 (time 0) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63) .. (126, 127)
+    // Row 15 (time 7) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), ..  (62, 63) .. (126, 127)
+
+
+#define add_vals(a_feat1, a_feat2, f1, f2, f3, f4)\
+      results[0].x += a_feat1.x * f1.x;\
+      results[1].x += a_feat1.x * f1.y;\
+      results[2].x += a_feat1.x * f2.x;\
+      results[3].x += a_feat1.x * f2.y;\
+      results[4].x += a_feat1.x * f3.x;\
+      results[5].x += a_feat1.x * f3.y;\
+      results[6].x += a_feat1.x * f4.x;\
+      results[7].x += a_feat1.x * f4.y;\
+\
+      results[0].y += a_feat1.y * f1.x;\
+      results[1].y += a_feat1.y * f1.y;\
+      results[2].y += a_feat1.y * f2.x;\
+      results[3].y += a_feat1.y * f2.y;\
+      results[4].y += a_feat1.y * f3.x;\
+      results[5].y += a_feat1.y * f3.y;\
+      results[6].y += a_feat1.y * f4.x;\
+      results[7].y += a_feat1.y * f4.y;\
+\
+      results[0].z += a_feat2.x * f1.x;\
+      results[1].z += a_feat2.x * f1.y;\
+      results[2].z += a_feat2.x * f2.x;\
+      results[3].z += a_feat2.x * f2.y;\
+      results[4].z += a_feat2.x * f3.x;\
+      results[5].z += a_feat2.x * f3.y;\
+      results[6].z += a_feat2.x * f4.x;\
+      results[7].z += a_feat2.x * f4.y;\
+\
+      results[0].w += a_feat2.y * f1.x;\
+      results[1].w += a_feat2.y * f1.y;\
+      results[2].w += a_feat2.y * f2.x;\
+      results[3].w += a_feat2.y * f2.y;\
+      results[4].w += a_feat2.y * f3.x;\
+      results[5].w += a_feat2.y * f3.y;\
+      results[6].w += a_feat2.y * f4.x;\
+      results[7].w += a_feat2.y * f4.y;\
+
+    lhs_shmem2[threadIdx.y/4][threadIdx.x+(threadIdx.y%4)*8] = make_float2(lhs_pf0.x, lhs_pf0.y);
+    lhs_shmem2[threadIdx.y/4+8][threadIdx.x+(threadIdx.y%4)*8] = make_float2(lhs_pf1.x, lhs_pf1.y);
+    lhs_shmem2[threadIdx.y/4+16][threadIdx.x+(threadIdx.y%4)*8] = make_float2(lhs_pf2.x, lhs_pf2.y);
+    lhs_shmem2[threadIdx.y/4+24][threadIdx.x+(threadIdx.y%4)*8] = make_float2(lhs_pf3.x, lhs_pf3.y);
+
+    lhs_shmem2[threadIdx.y/4 + 32][threadIdx.x+(threadIdx.y%4)*8] = make_float2(lhs_pf0.z, lhs_pf0.w);
+    lhs_shmem2[threadIdx.y/4 + 40][threadIdx.x+(threadIdx.y%4)*8] = make_float2(lhs_pf1.z, lhs_pf1.w);
+    lhs_shmem2[threadIdx.y/4 + 48][threadIdx.x+(threadIdx.y%4)*8] = make_float2(lhs_pf2.z, lhs_pf2.w);
+    lhs_shmem2[threadIdx.y/4 + 56][threadIdx.x+(threadIdx.y%4)*8] = make_float2(lhs_pf3.z, lhs_pf3.w);
+
+    __syncthreads();
+
+    // Do the multiplies.
+    #pragma unroll
+    for (int koff = 0; koff < 32; koff ++) {
+      float2 a3 = lhs_shmem2[koff][threadIdx.x + (threadIdx.y % 4) * 8];
+      float2 a4 = lhs_shmem2[koff + 32][threadIdx.x + (threadIdx.y % 4) * 8];
+
+      // first feature is at (threadIdx.y/4) * 8 last is at start + 8.
+      int start_feature = (threadIdx.y / 4) * 8;
+
+      float2 br1 = rhs_shmem2[start_feature/2 +     (koff % 4) * 32][koff/4];
+      float2 br2 = rhs_shmem2[start_feature/2 + 1 + (koff % 4) * 32][koff/4];
+      float2 br3 = rhs_shmem2[start_feature/2 + 2 + (koff % 4) * 32][koff/4];
+      float2 br4 = rhs_shmem2[start_feature/2 + 3 + (koff % 4) * 32][koff/4];
+
+      add_vals(a3, a4, br1, br2, br3, br4)
+    }
+    __syncthreads();
+  } // end loop over k
+
+  __syncthreads();
+  Index horiz_base = (threadIdx.y/4)*8+base_n;
+  if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
+    for (int i = 0; i < 8; i++) {
+      output(lhs_vert, horiz_base + i) = results[i].x;
+      output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      output(lhs_vert + 3, horiz_base + i) = results[i].w;
+    }
+  } else if (!CHECK_RHS_BOUNDARY) {
+    if (lhs_vert + 3 < m_size) {
+      for (int i = 0; i < 8; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    } else if (lhs_vert + 2 < m_size) {
+      for (int i = 0; i < 8; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+      }
+    } else if (lhs_vert + 1 < m_size) {
+      for (int i = 0; i < 8; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+      }
+    } else if (lhs_vert  < m_size) {
+      for (int i = 0; i < 8; i++) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+      }
+    }
+  } else if (!CHECK_LHS_BOUNDARY) {
+    // CHECK BOUNDARY_B
+    for (int i = 0; i < 8; i++) {
+      if (horiz_base + i < n_size) {
+        output(lhs_vert, horiz_base + i) = results[i].x;
+        output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    }
+  } else {
+    // CHECK both boundaries.
+    for (int i = 0; i < 8; i++) {
+      if (horiz_base + i < n_size) {
+        if (lhs_vert < m_size)
+          output(lhs_vert, horiz_base + i) = results[i].x;
+        if (lhs_vert + 1 < m_size)
+          output(lhs_vert + 1, horiz_base + i) = results[i].y;
+        if (lhs_vert + 2 < m_size)
+          output(lhs_vert + 2, horiz_base + i) = results[i].z;
+        if (lhs_vert + 3 < m_size)
+          output(lhs_vert + 3, horiz_base + i) = results[i].w;
+      }
+    }
+  }
+}
+
+
+template<typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper>
+__global__ void
+#if defined(EIGEN_HIPCC)
+__launch_bounds__(256, 1)
+#else
+__launch_bounds__(256)
+#endif
+EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output,
+                       const Index m_size, const Index n_size, const Index k_size) {
+  __shared__ float2 lhs_shmem[64*32];
+  __shared__ float2 rhs_shmem[128*8];
+
+  typedef float2 LHS_MEM[64][32];
+  typedef float2 RHS_MEM[128][8];
+
+  const Index m_block_idx = blockIdx.x;
+  const Index n_block_idx = blockIdx.y;
+
+  const Index base_m = 128 * m_block_idx;
+  const Index base_n = 64 * n_block_idx;
+
+  bool check_rhs = (base_n + 63) >= n_size;
+  bool check_lhs128 = (base_m + 127) >= m_size;
+
+  if (!check_rhs) {
+    if (!check_lhs128) {
+      // >= 128 rows left
+      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, false, false>(
+                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
+    } else {
+      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, true, false>(
+                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
+    }
+  } else {
+    if (!check_lhs128) {
+      // >= 128 rows left
+      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, false, true>(
+                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
+    } else {
+      EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, true, true>(
+                     lhs, rhs, output, *((LHS_MEM *) lhs_shmem), *((RHS_MEM *) rhs_shmem), m_size, n_size, k_size, base_m, base_n);
+    }
+  }
+}
+
+template<typename Index, typename LhsMapper,
+         typename RhsMapper, typename OutputMapper>
+__global__ void
+#if defined(EIGEN_HIPCC)
+__launch_bounds__(256, 1)
+#else
+__launch_bounds__(256)
+#endif
+EigenFloatContractionKernel16x16(const LhsMapper lhs, const RhsMapper rhs,
+                       const OutputMapper output,
+                       const Index m_size, const Index n_size, const Index k_size) {
+  __shared__ float2 lhs_shmem[32][16];
+  __shared__ float2 rhs_shmem[64][8];
+
+  const Index m_block_idx = blockIdx.x;
+  const Index n_block_idx = blockIdx.y;
+
+  const Index base_m = 64 * m_block_idx;
+  const Index base_n = 64 * n_block_idx;
+
+  if (base_m + 63 < m_size) {
+    if (base_n + 63 < n_size) {
+      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, false, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
+    } else {
+      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, false, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
+    }
+  } else {
+    if (base_n + 63 < n_size) {
+      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, true, false>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
+    } else {
+      EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, true, true>(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
+    }
+  }
+}
+
+
+template<typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, GpuDevice> :
+    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, GpuDevice> > {
+
+  typedef GpuDevice Device;
+
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
+  typedef TensorContractionEvaluatorBase<Self> Base;
+
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, GpuDevice>::type PacketReturnType;
+
+  enum {
+    Layout = TensorEvaluator<LeftArgType, Device>::Layout,
+  };
+
+  // Most of the code is assuming that both input tensors are ColMajor. If the
+  // inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
+  // If we want to compute A * B = C, where A is LHS and B is RHS, the code
+  // will pretend B is LHS and A is RHS.
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType;
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType;
+
+  static const int LDims =
+      internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
+  static const int RDims =
+      internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
+  static const int ContractDims = internal::array_size<Indices>::value;
+
+  typedef array<Index, LDims> left_dim_mapper_t;
+  typedef array<Index, RDims> right_dim_mapper_t;
+
+  typedef array<Index, ContractDims> contract_t;
+  typedef array<Index, LDims - ContractDims> left_nocontract_t;
+  typedef array<Index, RDims - ContractDims> right_nocontract_t;
+
+  static const int NumDims = LDims + RDims - 2 * ContractDims;
+
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  // typedefs needed in evalTo
+  typedef typename internal::remove_const<typename EvalLeftArgType::Scalar>::type LhsScalar;
+  typedef typename internal::remove_const<typename EvalRightArgType::Scalar>::type RhsScalar;
+
+  typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
+  typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
+
+  typedef typename LeftEvaluator::Dimensions LeftDimensions;
+  typedef typename RightEvaluator::Dimensions RightDimensions;
+
+  TensorEvaluator(const XprType& op, const Device& device) :
+      Base(op, device)
+  {
+    EIGEN_STATIC_ASSERT( (internal::is_same<OutputKernelType, const NoOpOutputKernel>::value),
+                          GPU_TENSOR_CONTRACTION_DOES_NOT_SUPPORT_OUTPUT_KERNELS);
+  }
+
+  // We need to redefine this method to make nvcc happy
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
+    this->m_leftImpl.evalSubExprsIfNeeded(NULL);
+    this->m_rightImpl.evalSubExprsIfNeeded(NULL);
+    if (data) {
+      evalTo(data);
+      return false;
+    } else {
+      this->m_result = static_cast<Scalar *>(this->m_device.allocate(this->dimensions().TotalSize() * sizeof(Scalar)));
+      evalTo(this->m_result);
+      return true;
+    }
+  }
+
+  void evalTo(Scalar* buffer) const {
+    if (this->m_lhs_inner_dim_contiguous) {
+      if (this->m_rhs_inner_dim_contiguous) {
+        if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<true, true, true, Unaligned>(buffer);
+        }
+        else {
+          evalTyped<true, true, false, Unaligned>(buffer);
+        }
+      }
+      else {
+       if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<true, false, true, Unaligned>(buffer);
+        }
+        else {
+          evalTyped<true, false, false, Unaligned>(buffer);
+        }
+      }
+    }
+    else {
+      if (this->m_rhs_inner_dim_contiguous) {
+        if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<false, true, true, Unaligned>(buffer);
+        }
+        else {
+          evalTyped<false, true, false, Unaligned>(buffer);
+        }
+      }
+      else {
+       if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<false, false, true, Unaligned>(buffer);
+        }
+        else {
+          evalTyped<false, false, false, Unaligned>(buffer);
+        }
+      }
+    }
+  }
+
+  template <typename LhsScalar, typename RhsScalar, typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper> struct LaunchKernels {
+    static void Run(const LhsMapper& lhs, const RhsMapper& rhs, const OutputMapper& output, Index m, Index n, Index k, const GpuDevice& device) {
+    const Index m_blocks = (m + 63) / 64;
+    const Index n_blocks = (n + 63) / 64;
+    const dim3 num_blocks(m_blocks, n_blocks, 1);
+    const dim3 block_size(8, 8, 8);
+    LAUNCH_GPU_KERNEL((EigenContractionKernel<Scalar, Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
+    }
+  };
+
+  template <typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper> struct LaunchKernels<float, float, Index, LhsMapper, RhsMapper, OutputMapper> {
+    static void Run(const LhsMapper& lhs, const RhsMapper& rhs, const OutputMapper& output, Index m, Index n, Index k, const GpuDevice& device) {
+      if (m < 768 || n < 768) {
+        const Index m_blocks = (m + 63) / 64;
+        const Index n_blocks = (n + 63) / 64;
+        const dim3 num_blocks(m_blocks, n_blocks, 1);
+        const dim3 block_size(16, 16, 1);
+        LAUNCH_GPU_KERNEL((EigenFloatContractionKernel16x16<Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
+      } else {
+        const Index m_blocks = (m + 127) / 128;
+        const Index n_blocks = (n + 63) / 64;
+        const dim3 num_blocks(m_blocks, n_blocks, 1);
+        const dim3 block_size(8, 32, 1);
+        LAUNCH_GPU_KERNEL((EigenFloatContractionKernel<Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks, block_size, 0, device, lhs, rhs, output, m, n, k);
+      }
+    }
+  };
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
+  void evalTyped(Scalar* buffer) const {
+    // columns in left side, rows in right side
+    const Index k = this->m_k_size;
+    EIGEN_UNUSED_VARIABLE(k)
+
+    // rows in left side
+    const Index m = this->m_i_size;
+
+    // columns in right side
+    const Index n = this->m_j_size;
+
+    // zero out the result buffer (which must be of size at least m * n * sizeof(Scalar)
+    this->m_device.memset(buffer, 0, m * n * sizeof(Scalar));
+
+    typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs,
+                                                   LeftEvaluator, left_nocontract_t,
+                                                   contract_t, 4,
+                                                   lhs_inner_dim_contiguous,
+                                                   false, Unaligned> LhsMapper;
+
+    typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs,
+                                                   RightEvaluator, right_nocontract_t,
+                                                   contract_t, 4,
+                                                   rhs_inner_dim_contiguous,
+                                                   rhs_inner_dim_reordered, Unaligned> RhsMapper;
+
+    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
+
+
+    // initialize data mappers
+    LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides,
+                  this->m_left_contracting_strides, this->m_k_strides);
+
+    RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides,
+                  this->m_right_contracting_strides, this->m_k_strides);
+
+    OutputMapper output(buffer, m);
+
+#if defined(EIGEN_USE_HIP)
+    setGpuSharedMemConfig(hipSharedMemBankSizeEightByte);
+#else
+    setGpuSharedMemConfig(cudaSharedMemBankSizeEightByte);
+#endif
+
+    LaunchKernels<LhsScalar, RhsScalar, Index, LhsMapper, RhsMapper, OutputMapper>::Run(lhs, rhs, output,  m, n, k, this->m_device);
+  }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_USE_GPU and EIGEN_GPUCC
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionMapper.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionMapper.h
new file mode 100644
index 0000000..9ab900b
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionMapper.h
@@ -0,0 +1,575 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_MAPPER_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_MAPPER_H
+
+namespace Eigen {
+
+namespace internal {
+
+enum {
+  Rhs = 0,
+  Lhs = 1
+};
+
+/*
+ * Implementation of the Eigen blas_data_mapper class for tensors.
+ */
+/// The make pointer class is used by sycl in order to build the mapper class on the device. For other platform the default make pointer is used which
+/// is scalar * for CoeffLoader.
+template <typename Tensor, bool HasRawAccess, template <class> class MakePointer_ = MakePointer>
+struct CoeffLoader;
+
+template <typename Scalar, typename Index, int side, typename Tensor,
+          typename nocontract_t, typename contract_t, int packet_size,
+          bool inner_dim_contiguous, bool inner_dim_reordered, int Alignment,
+          template <class> class MakePointer_ = MakePointer>
+class BaseTensorContractionMapper;
+
+template <typename Tensor, bool HasRawAccess, template <class> class MakePointer_>
+struct CoeffLoader {
+  enum {
+    DirectOffsets = false
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffLoader(const Tensor& tensor) : m_tensor(tensor) { }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void offsetBuffer(typename Tensor::Index) {
+    eigen_assert(false && "unsupported");
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE const typename MakePointer_<const typename Tensor::Scalar>::Type
+  data() const {
+    eigen_assert(false && "unsupported");
+    return NULL;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE typename Tensor::Scalar coeff(typename Tensor::Index index) const { return m_tensor.coeff(index); }
+
+ template<int LoadMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+ typename Tensor::PacketReturnType packet(typename Tensor::Index index) const
+  {
+    return m_tensor.template packet<LoadMode>(index);
+  }
+
+  #ifdef EIGEN_USE_SYCL
+  // The placeholder accessors require to be bound to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_tensor.bind(cgh);
+  }
+  #endif
+
+ private:
+  const Tensor m_tensor;
+};
+
+template <typename Tensor, template <class> class MakePointer_>
+struct CoeffLoader<Tensor, true, MakePointer_> {
+  enum {
+    DirectOffsets = true
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffLoader(const Tensor& tensor) : m_data(tensor.data()) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void offsetBuffer(typename Tensor::Index offset) {
+    m_data += offset;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE const typename MakePointer_<const typename Tensor::Scalar>::Type
+  data() const {
+    return m_data;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE typename Tensor::Scalar coeff(typename Tensor::Index index) const { return loadConstant(m_data+index); }
+
+ template<int LoadMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+ typename Tensor::PacketReturnType packet(typename Tensor::Index index) const
+  {
+    return internal::ploadt_ro<typename Tensor::PacketReturnType, LoadMode>(m_data + index);
+  }
+
+  #ifdef EIGEN_USE_SYCL
+  // The placeholder accessors require to be bound to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_data.bind(cgh);
+  }
+  #endif
+ private:
+  typedef typename Tensor::Scalar Scalar;
+
+  typename MakePointer_<const Scalar>::Type m_data;
+};
+
+template<typename Scalar, typename Index, int side,
+         typename Tensor,
+         typename nocontract_t, typename contract_t,
+         int packet_size, bool inner_dim_contiguous, int Alignment, template <class> class MakePointer_ = MakePointer>
+class SimpleTensorContractionMapper {
+  public:
+  EIGEN_DEVICE_FUNC
+  SimpleTensorContractionMapper(const Tensor& tensor,
+                                const nocontract_t& nocontract_strides,
+                                const nocontract_t& ij_strides,
+                                const contract_t& contract_strides,
+                                const contract_t& k_strides) :
+      m_tensor(tensor),
+      m_nocontract_strides(nocontract_strides),
+      m_ij_strides(ij_strides),
+      m_contract_strides(contract_strides),
+      m_k_strides(k_strides) { }
+
+  enum {
+    DirectOffsets = CoeffLoader<Tensor, Tensor::RawAccess, MakePointer_>::DirectOffsets
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void offsetBuffer(typename Tensor::Index offset) {
+    m_tensor.offsetBuffer(offset);
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE void prefetch(Index /*i*/) { }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE Scalar operator()(Index row) const {
+    // column major assumption
+    return operator()(row, 0);
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE Scalar operator()(Index row, Index col) const {
+    return m_tensor.coeff(computeIndex(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE Index computeIndex(Index row, Index col) const {
+    const bool left = (side == Lhs);
+    EIGEN_UNUSED_VARIABLE(left); // annoying bug in g++8.1: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85963
+    Index nocontract_val = left ? row : col;
+    Index linidx = 0;
+    EIGEN_UNROLL_LOOP
+    for (int i = static_cast<int>(array_size<nocontract_t>::value) - 1; i > 0; i--) {
+      const Index idx = nocontract_val / m_ij_strides[i];
+      linidx += idx * m_nocontract_strides[i];
+      nocontract_val -= idx * m_ij_strides[i];
+    }
+    if (array_size<typename Tensor::Dimensions>::value > array_size<contract_t>::value) {
+      if (side == Lhs && inner_dim_contiguous) {
+        eigen_assert(m_nocontract_strides[0] == 1);
+        linidx += nocontract_val;
+      } else {
+        linidx += nocontract_val * m_nocontract_strides[0];
+      }
+    }
+
+    Index contract_val = left ? col : row;
+    if(array_size<contract_t>::value > 0) {
+      EIGEN_UNROLL_LOOP
+      for (int i = static_cast<int>(array_size<contract_t>::value) - 1; i > 0; i--) {
+        const Index idx = contract_val / m_k_strides[i];
+        linidx += idx * m_contract_strides[i];
+        contract_val -= idx * m_k_strides[i];
+      }
+
+      if (side == Rhs && inner_dim_contiguous) {
+        eigen_assert(m_contract_strides[0] == 1);
+        linidx += contract_val;
+      } else {
+        linidx += contract_val * m_contract_strides[0];
+      }
+    }
+
+    return linidx;
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE IndexPair<Index> computeIndexPair(Index row, Index col, const Index distance) const {
+    const bool left = (side == Lhs);
+    EIGEN_UNUSED_VARIABLE(left); // annoying bug in g++8.1: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85963
+    Index nocontract_val[2] = {left ? row : col, left ? row + distance : col};
+    Index linidx[2] = {0, 0};
+    if (array_size<typename Tensor::Dimensions>::value > array_size<contract_t>::value) {
+      EIGEN_UNROLL_LOOP
+      for (int i = static_cast<int>(array_size<nocontract_t>::value) - 1; i > 0; i--) {
+        const Index idx0 = nocontract_val[0] / m_ij_strides[i];
+        const Index idx1 = nocontract_val[1] / m_ij_strides[i];
+        linidx[0] += idx0 * m_nocontract_strides[i];
+        linidx[1] += idx1 * m_nocontract_strides[i];
+        nocontract_val[0] -= idx0 * m_ij_strides[i];
+        nocontract_val[1] -= idx1 * m_ij_strides[i];
+      }
+      if (side == Lhs && inner_dim_contiguous) {
+        eigen_assert(m_nocontract_strides[0] == 1);
+        linidx[0] += nocontract_val[0];
+        linidx[1] += nocontract_val[1];
+      } else {
+        linidx[0] += nocontract_val[0] * m_nocontract_strides[0];
+        linidx[1] += nocontract_val[1] * m_nocontract_strides[0];
+      }
+    }
+
+    Index contract_val[2] = {left ? col : row, left ? col : row + distance};
+    if (array_size<contract_t>::value> 0) {
+      EIGEN_UNROLL_LOOP
+      for (int i = static_cast<int>(array_size<contract_t>::value) - 1; i > 0; i--) {
+        const Index idx0 = contract_val[0] / m_k_strides[i];
+        const Index idx1 = contract_val[1] / m_k_strides[i];
+        linidx[0] += idx0 * m_contract_strides[i];
+        linidx[1] += idx1 * m_contract_strides[i];
+        contract_val[0] -= idx0 * m_k_strides[i];
+        contract_val[1] -= idx1 * m_k_strides[i];
+      }
+
+      if (side == Rhs && inner_dim_contiguous) {
+        eigen_assert(m_contract_strides[0] == 1);
+        linidx[0] += contract_val[0];
+        linidx[1] += contract_val[1];
+      } else {
+        linidx[0] += contract_val[0] * m_contract_strides[0];
+        linidx[1] += contract_val[1] * m_contract_strides[0];
+      }
+    }
+    return IndexPair<Index>(linidx[0], linidx[1]);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Index firstAligned(Index size) const {
+    // Only claim alignment when we can compute the actual stride (ie when we're
+    // dealing with the lhs with inner_dim_contiguous. This is because the
+    // matrix-vector product relies on the stride when dealing with aligned inputs.
+    return (Alignment == Aligned) && (side == Lhs) && inner_dim_contiguous ? 0 : size;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Index stride() const {
+    return ((side == Lhs) && inner_dim_contiguous && array_size<contract_t>::value > 0) ? m_contract_strides[0] : 1;
+  }
+
+  #ifdef EIGEN_USE_SYCL
+  // The placeholder accessors require to be bound to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_tensor.bind(cgh);
+  }
+  #endif
+
+  const CoeffLoader<Tensor, Tensor::RawAccess, MakePointer_>& tensor() const {
+    return m_tensor;
+  }
+
+  const nocontract_t& nocontract_strides() const {
+    return m_nocontract_strides;
+  }
+  const nocontract_t& ij_strides() const { return m_ij_strides; }
+  const contract_t& contract_strides() const { return m_contract_strides; }
+  const contract_t& k_strides() const { return m_k_strides; }
+
+ protected:
+  CoeffLoader<Tensor, Tensor::RawAccess, MakePointer_> m_tensor;
+  const nocontract_t m_nocontract_strides;
+  const nocontract_t m_ij_strides;
+  const contract_t m_contract_strides;
+  const contract_t m_k_strides;
+};
+
+template<typename Scalar, typename Index, int side,
+         typename Tensor,
+         typename nocontract_t, typename contract_t,
+         int packet_size, bool inner_dim_contiguous,
+         bool inner_dim_reordered, int Alignment, template <class> class MakePointer_>
+class BaseTensorContractionMapper : public SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, Alignment, MakePointer_>
+{
+ public:
+  typedef SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, Alignment, MakePointer_> ParentMapper;
+
+  EIGEN_DEVICE_FUNC
+  BaseTensorContractionMapper(const Tensor& tensor,
+                              const nocontract_t& nocontract_strides,
+                              const nocontract_t& ij_strides,
+                              const contract_t& contract_strides,
+                              const contract_t& k_strides) :
+  ParentMapper(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { }
+
+  template <typename PacketT,int AlignmentType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename internal::enable_if<internal::unpacket_traits<PacketT>::size==packet_size,PacketT>::type
+  load(Index i, Index j) const
+  {
+    // whole method makes column major assumption
+
+    // don't need to add offsets for now (because operator handles that)
+    // current code assumes packet size must be a multiple of 2
+    EIGEN_STATIC_ASSERT(packet_size % 2 == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    if (Tensor::PacketAccess && inner_dim_contiguous && !inner_dim_reordered) {
+      const Index index = this->computeIndex(i, j);
+      eigen_assert(this->computeIndex(i+packet_size-1, j) == index + packet_size-1);
+      return this->m_tensor.template packet<AlignmentType>(index);
+    }
+
+    const IndexPair<Index> indexPair = this->computeIndexPair(i, j, packet_size - 1);
+    const Index first = indexPair.first;
+    const Index lastIdx = indexPair.second;
+
+    // We can always do optimized packet reads from left hand side right now, because
+    // the vertical matrix dimension on the left hand side is never contracting.
+    // On the right hand side we need to check if the contracting dimensions may have
+    // been shuffled first.
+    if (Tensor::PacketAccess &&
+        (side == Lhs || internal::array_size<contract_t>::value <= 1 || !inner_dim_reordered) &&
+        (lastIdx - first) == (packet_size - 1)) {
+
+      return this->m_tensor.template packet<AlignmentType>(first);
+    }
+
+    EIGEN_ALIGN_MAX Scalar data[packet_size];
+
+    data[0] = this->m_tensor.coeff(first);
+    EIGEN_UNROLL_LOOP
+    for (Index k = 1; k < packet_size - 1; k += 2) {
+      const IndexPair<Index> internal_pair = this->computeIndexPair(i + k, j, 1);
+      data[k] = this->m_tensor.coeff(internal_pair.first);
+      data[k + 1] = this->m_tensor.coeff(internal_pair.second);
+    }
+    data[packet_size - 1] = this->m_tensor.coeff(lastIdx);
+
+    return pload<PacketT>(data);
+  }
+
+  template <typename PacketT,int AlignmentType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename internal::enable_if<internal::unpacket_traits<PacketT>::size!=packet_size,PacketT>::type
+  load(Index i, Index j) const
+  {
+    const Index requested_packet_size = internal::unpacket_traits<PacketT>::size;
+    EIGEN_ALIGN_MAX Scalar data[requested_packet_size];
+
+    const IndexPair<Index> indexPair = this->computeIndexPair(i, j, requested_packet_size - 1);
+    const Index first = indexPair.first;
+    const Index lastIdx = indexPair.second;
+
+    data[0] = this->m_tensor.coeff(first);
+    for (Index k = 1; k < requested_packet_size - 1; k += 2) {
+      const IndexPair<Index> internal_pair = this->computeIndexPair(i + k, j, 1);
+      data[k] = this->m_tensor.coeff(internal_pair.first);
+      data[k + 1] = this->m_tensor.coeff(internal_pair.second);
+    }
+    data[requested_packet_size - 1] = this->m_tensor.coeff(lastIdx);
+
+    return pload<PacketT>(data);
+  }
+
+  template <typename PacketT,int AlignmentType>
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE PacketT loadPacket(Index i, Index j) const {
+    return this->load<PacketT,AlignmentType>(i,j);
+  }
+};
+
+
+template<typename Scalar, typename Index, int side,
+         typename Tensor,
+         typename nocontract_t, typename contract_t,
+         bool inner_dim_contiguous,
+         bool inner_dim_reordered, int Alignment, template <class> class MakePointer_>
+class BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_>
+  : public SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, Alignment, MakePointer_>
+{
+ public:
+  typedef SimpleTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, 1, inner_dim_contiguous, Alignment, MakePointer_> ParentMapper;
+
+  EIGEN_DEVICE_FUNC
+  BaseTensorContractionMapper(const Tensor& tensor,
+                              const nocontract_t& nocontract_strides,
+                              const nocontract_t& ij_strides,
+                              const contract_t& contract_strides,
+                              const contract_t& k_strides) :
+  ParentMapper(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { }
+
+  template <typename PacketT,int> EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE PacketT loadPacket(Index i, Index j) const {
+    EIGEN_ALIGN_MAX Scalar data[1];
+    data[0] = this->m_tensor.coeff(this->computeIndex(i, j));
+    return pload<PacketT>(data);
+  }
+  template <typename PacketT,int> EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE PacketT load(Index i, Index j) const {
+    EIGEN_ALIGN_MAX Scalar data[1];
+    data[0] = this->m_tensor.coeff(this->computeIndex(i, j));
+    return pload<PacketT>(data);
+  }
+};
+
+
+template<typename Scalar, typename Index, int side,
+         typename Tensor,
+         typename nocontract_t, typename contract_t,
+         int packet_size,
+         bool inner_dim_contiguous, bool inner_dim_reordered, int Alignment, template <class> class MakePointer_=MakePointer>
+class TensorContractionSubMapper {
+ public:
+
+  typedef BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> ParentMapper;
+  typedef TensorContractionSubMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> Self;
+  typedef Self LinearMapper;
+
+  enum {
+    // We can use direct offsets iff the parent mapper supports then and we can compute the strides.
+    // TODO: we should also enable direct offsets for the Rhs case.
+    UseDirectOffsets = ParentMapper::DirectOffsets && (side == Lhs) && inner_dim_contiguous && (array_size<contract_t>::value > 0)
+  };
+
+  EIGEN_DEVICE_FUNC TensorContractionSubMapper(const ParentMapper& base_mapper, Index vert_offset, Index horiz_offset)
+      : m_base_mapper(base_mapper), m_vert_offset(vert_offset), m_horiz_offset(horiz_offset) {
+    // Bake the offsets into the buffer used by the base mapper whenever possible. This avoids the need to recompute
+    // this offset every time we attempt to access a coefficient.
+    if (UseDirectOffsets) {
+      Index stride = m_base_mapper.stride();
+      m_base_mapper.offsetBuffer(vert_offset + horiz_offset * stride);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar operator()(Index i) const {
+    if (UseDirectOffsets) {
+      return m_base_mapper(i, 0);
+    }
+    return m_base_mapper(i + m_vert_offset, m_horiz_offset);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar operator()(Index i, Index j) const {
+    if (UseDirectOffsets) {
+      return m_base_mapper(i, j);
+    }
+    return m_base_mapper(i + m_vert_offset, j + m_horiz_offset);
+  }
+
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT loadPacket(Index i) const {
+    if (UseDirectOffsets) {
+      return m_base_mapper.template loadPacket<PacketT,Alignment>(i, 0);
+    }
+    return m_base_mapper.template loadPacket<PacketT,Alignment>(i + m_vert_offset, m_horiz_offset);
+  }
+
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT loadPacket(Index i, Index j) const {
+    if (UseDirectOffsets) {
+      return m_base_mapper.template loadPacket<PacketT,Alignment>(i, j);
+    }
+    return m_base_mapper.template loadPacket<PacketT,Alignment>(i + m_vert_offset, j + m_horiz_offset);
+  }
+
+  template <typename PacketT, int AlignmentType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT loadPacket(Index i, Index j) const {
+    if (UseDirectOffsets) {
+      return m_base_mapper.template load<PacketT,AlignmentType>(i, j);
+    }
+    return m_base_mapper.template loadPacket<PacketT,AlignmentType>(i + m_vert_offset, j + m_horiz_offset);
+  }
+
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const PacketT& p) const {
+    if (UseDirectOffsets) {
+      m_base_mapper.storePacket(i, 0, p);
+    }
+    m_base_mapper.storePacket(i + m_vert_offset, m_horiz_offset, p);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE LinearMapper getLinearMapper(Index i, Index j) const {
+    if (UseDirectOffsets) {
+      return LinearMapper(m_base_mapper, i, j);
+    }
+    return LinearMapper(m_base_mapper, i + m_vert_offset, j + m_horiz_offset);
+  }
+
+  template <typename PacketT, int AlignmentType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketT load(Index i) const {
+    EIGEN_STATIC_ASSERT((internal::is_same<PacketT, PacketT>::value), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    const int ActualAlignment = (AlignmentType == Aligned) && (Alignment == Aligned) ? Aligned : Unaligned;
+    if (UseDirectOffsets) {
+     return m_base_mapper.template loadPacket<PacketT,ActualAlignment>(i, 0);
+    }
+    return m_base_mapper.template loadPacket<PacketT,ActualAlignment>(i + m_vert_offset, m_horiz_offset);
+  }
+
+  template <typename PacketT>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool aligned(Index) const {
+    return false;
+  }
+
+  #ifdef EIGEN_USE_SYCL
+  // The placeholder accessors require to be bound to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_base_mapper.bind(cgh);
+  }
+  #endif
+
+  const ParentMapper& base_mapper() const { return m_base_mapper; }
+  Index vert_offset() const { return m_vert_offset; }
+  Index horiz_offset() const { return m_horiz_offset; }
+
+ private:
+  ParentMapper m_base_mapper;
+  const Index m_vert_offset;
+  const Index m_horiz_offset;
+};
+
+
+template<typename Scalar_, typename Index, int side,
+         typename Tensor,
+         typename nocontract_t, typename contract_t,
+         int packet_size,
+         bool inner_dim_contiguous, bool inner_dim_reordered, int Alignment,  template <class> class MakePointer_=MakePointer>
+class TensorContractionInputMapper
+  : public BaseTensorContractionMapper<Scalar_, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> {
+
+ public:
+  typedef Scalar_ Scalar;
+  typedef BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> Base;
+  typedef TensorContractionSubMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment, MakePointer_> SubMapper;
+  typedef SubMapper VectorMapper;
+
+  EIGEN_DEVICE_FUNC TensorContractionInputMapper(const Tensor& tensor,
+                               const nocontract_t& nocontract_strides,
+                               const nocontract_t& ij_strides,
+                               const contract_t& contract_strides,
+                               const contract_t& k_strides)
+      : Base(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE SubMapper getSubMapper(Index i, Index j) const {
+    return SubMapper(*this, i, j);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE VectorMapper getVectorMapper(Index i, Index j) const {
+    return VectorMapper(*this, i, j);
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE const CoeffLoader<Tensor, Tensor::RawAccess, MakePointer_>& get_tensor() const {
+    return Base::m_tensor;
+  }
+};
+
+
+template <typename T> struct TensorContractionInputMapperTrait;
+
+template<typename Scalar_, typename Index_, int side_,
+         typename Tensor_,
+         typename nocontract_t_, typename contract_t_,
+         int packet_size_,
+         bool inner_dim_contiguous_, bool inner_dim_reordered_, int Alignment_,  template <class> class MakePointer_>
+struct TensorContractionInputMapperTrait<TensorContractionInputMapper<Scalar_, Index_, side_, Tensor_, 
+                                                    nocontract_t_, contract_t_, packet_size_, inner_dim_contiguous_, 
+                                                    inner_dim_reordered_, Alignment_, MakePointer_> > {
+
+      typedef Tensor_ XprType;
+      static const bool  inner_dim_contiguous = inner_dim_contiguous_;
+      static const bool  inner_dim_reordered = inner_dim_reordered_;
+  };  
+
+
+}  // end namespace internal
+}  // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_MAPPER_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionSycl.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionSycl.h
new file mode 100755
index 0000000..473c228
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionSycl.h
@@ -0,0 +1,1650 @@
+// This file is part of Eigen, a lightweight C++ template library for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla Public License v. 2.0. If a copy of the MPL was not
+// distributed with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/*****************************************************************
+ * TensorContractionSycl.h
+ *
+ * \brief:
+ *  TensorContractionSycl.h, provides various tensor contraction kernel for SYCL backend
+ *
+ *****************************************************************/
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_SYCL_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_SYCL_H
+
+namespace Eigen {
+
+namespace TensorSycl {
+namespace internal {
+
+#ifndef EIGEN_SYCL_DISABLE_GEMV
+/*!
+ * \brief TVPanelSize, a template class used for setting the panel size required for launching General TensorVector
+ * contraction kernel on various hardware devices.
+ *
+ * \tparam Scalar: determines the element type of the tensor/vector
+ *
+ * \tparam StorageIndex  determines the Index type.
+ *
+ * \tparam NCWindow: determines the number of non-contracting element to be process by each work-group
+ *
+ * \tparam CFactor: determines the number of contracting element to be process by each thread
+ *
+ * \tparam NCFactor: determines the number of non-contracting element to be process by each thread
+ */
+template <typename Scalar, typename StorageIndex, StorageIndex NCWindow, StorageIndex CFactor, StorageIndex NCFactor>
+struct TVPanelSize {
+  // LocalThreadSizeC: determines total number of thread per workgroup for the contracting dimension
+  static EIGEN_CONSTEXPR StorageIndex LocalThreadSizeC = EIGEN_SYCL_LOCAL_THREAD_DIM0;
+  // LocalThreadSizeNC: determines total number of thread per workgroup for the non-contracting dimension
+  static EIGEN_CONSTEXPR StorageIndex LocalThreadSizeNC = EIGEN_SYCL_LOCAL_THREAD_DIM1;
+  // TileSizeDimNC: determines the tile size for the non-contracting dimension
+  static EIGEN_CONSTEXPR StorageIndex TileSizeDimNC = NCWindow / NCFactor;
+  // TileSizeDimC: determines the tile size for the contracting dimension
+  static EIGEN_CONSTEXPR StorageIndex TileSizeDimC = CFactor * LocalThreadSizeNC * LocalThreadSizeC;
+  // WorkLoadPerThreadNC : determines workload per thread for loading the non-contracting dimension
+  static EIGEN_CONSTEXPR StorageIndex WorkLoadPerThreadNC = TileSizeDimNC / LocalThreadSizeNC;
+  // WorkLoadPerThreadC: determines workload per thread for loading the non-contracting dimension
+  static EIGEN_CONSTEXPR StorageIndex WorkLoadPerThreadC = TileSizeDimC / LocalThreadSizeC;
+  // BC : determines if supporting bank conflict is required
+  static EIGEN_CONSTEXPR bool BC = false;
+};
+#endif
+
+/*!
+ * \brief TTPanelSize, a template class used for setting the panel size required for launching General Tensor Tensor
+ contraction kernel on various hardware devices.
+ *
+ * \tparam Scalar: determines the element type of the tensor
+ *
+ * \tparam StorageIndex: determines the Index type.
+ *
+ * \tparam REG_SIZE_M: determines workload per thread for loading the M dimension This can be varied based on the
+ available register on a chosen device(can be controlled by EIGEN_SYCL_REG_M macro).
+ *
+ * \tparam REG_SIZE_N: determines workload per thread for loading the N dimension This can be varied based on the
+ available register on a chosen device(can be controlled by EIGEN_SYCL_REG_N macro).
+ *
+ * \tparam TSDK: determines Tile size for dimension K. The packet size is assumed to be considered
+ */
+
+template <typename Scalar, typename StorageIndex, StorageIndex REG_SIZE_M, StorageIndex REG_SIZE_N, StorageIndex TSDK>
+struct TTPanelSize {
+  // TileSizeDimK: determines Tile size for dimension K. The packet size is assumed to be considered
+  static EIGEN_CONSTEXPR StorageIndex TileSizeDimK = TSDK;
+  // WorkLoadPerThreadM : determines workload per thread for loading the M dimension This can be varied based on the
+  // available register on a chosen device(can be controlled by EIGEN_SYCL_REG_M macro//
+#ifndef EIGEN_SYCL_REG_M
+  static EIGEN_CONSTEXPR StorageIndex WorkLoadPerThreadM = REG_SIZE_M;
+#else
+  static EIGEN_CONSTEXPR StorageIndex WorkLoadPerThreadM = EIGEN_SYCL_REG_M;
+#endif
+// WorkLoadPerThreadN : determines workload per thread for loading the N dimension This can be varied based on the
+// available register on a chosen device(can be controlled by EIGEN_SYCL_REG_N macro
+#ifndef EIGEN_SYCL_REG_N
+  static EIGEN_CONSTEXPR StorageIndex WorkLoadPerThreadN = REG_SIZE_N;
+#else
+  static EIGEN_CONSTEXPR StorageIndex WorkLoadPerThreadN = EIGEN_SYCL_REG_N;
+#endif
+  // LocalThreadSizeM: determines total number of thread per workgroup for the m dimension
+  static EIGEN_CONSTEXPR StorageIndex LocalThreadSizeM = EIGEN_SYCL_LOCAL_THREAD_DIM0;
+  // LocalThreadSizeN: determines total number of thread per workgroup for the n dimension
+  static EIGEN_CONSTEXPR StorageIndex LocalThreadSizeN = EIGEN_SYCL_LOCAL_THREAD_DIM1;
+  // TileSizeDimM: determines the tile size for the m dimension
+  static EIGEN_CONSTEXPR StorageIndex TileSizeDimM = LocalThreadSizeM * WorkLoadPerThreadM;
+  // TileSizeDimN: determines the tile size for the n dimension
+  static EIGEN_CONSTEXPR StorageIndex TileSizeDimN = LocalThreadSizeN * WorkLoadPerThreadN;
+  // LoadPerThreadLhs: determines workload per thread for loading Lhs Tensor. This must be divisable by packetsize
+  static EIGEN_CONSTEXPR StorageIndex LoadPerThreadLhs =
+      ((TileSizeDimK * WorkLoadPerThreadM * WorkLoadPerThreadN) / (TileSizeDimN));
+  // LoadPerThreadRhs: determines workload per thread for loading Rhs Tensor. This must be divisable by packetsize
+  static EIGEN_CONSTEXPR StorageIndex LoadPerThreadRhs =
+      ((TileSizeDimK * WorkLoadPerThreadM * WorkLoadPerThreadN) / (TileSizeDimM));
+  // BC : determines if supporting bank conflict is required
+  static EIGEN_CONSTEXPR bool BC = true;
+  // DoubleBuffer: determines if double buffering technique should be used (This can be disabled by
+  // EIGEN_SYCL_DISABLE_DOUBLE_BUFFER macro when the device doesnot have sufficient  local memory)
+  static EIGEN_CONSTEXPR bool DoubleBuffer =
+#ifdef EIGEN_SYCL_DISABLE_DOUBLE_BUFFER
+      false;
+#else
+      true;
+#endif
+};
+
+/* !
+ * \brief contraction_type: an enum class representing the Tensor Contraction implementation algorithm. This is used to
+ * specialize the contraction algorithm based on device support for dedicated local memory.
+ */
+enum class contraction_type { local, no_local };
+/* !
+ * \brief data_source an enum class determining the location of the data in a memory hierarchy (global, local, private).
+ */
+enum class data_source { global_mem, local_mem, private_mem };
+
+/*!
+ * \brief read, a template function used for loading the data from global
+ memory. This function is used to guarantee coalesced and vectorized load whenever possible
+ *
+ * \tparam PacketLoad: determines if the each element of this tensor block should be loaded in a packet mode
+ *
+ * \param is_coalesced_layout: determines whether or not the Tensor data in a memory can be access coalesced and
+ vectorized when possible. Coalesced memory access is a key factor in Kernel performance. When a tensor is 2d and the
+ contracting dimension is 1, it is always possible to accessed tensor data coalesced and vectorized. This is the case
+ when RHS(right hand side) Tensor is transposed or when LHS(left hand side) Tensor is not transposed.
+ *
+ * \tparam PacketType:  determines the type of packet
+ *
+ * \tparam TensorMapper: determines the input tensor mapper type
+ *
+ * \tparam StorageIndex: determines the Index type
+
+ * \param tensorMapper: is the input tensor
+ *
+ * \param NCIndex: is the non-contracting dim index
+ *
+ * \param CIndex is the contracting dim index
+ *
+ * \param ld: is the leading dimension of the flattened tensor
+ */
+template <bool PacketLoad, bool is_coalesced_layout, bool, typename PacketType, typename TensorMapper,
+          typename StorageIndex>
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<PacketLoad, PacketType>::type read(
+    const TensorMapper &tensorMapper, const StorageIndex &NCIndex, const StorageIndex &CIndex, const StorageIndex &ld) {
+  const StorageIndex row = (is_coalesced_layout) ? NCIndex : CIndex;
+  const StorageIndex col = (is_coalesced_layout) ? CIndex : NCIndex;
+  return tensorMapper.get_tensor().template packet<Unaligned>(row + (col * ld));
+}
+
+/*!
+ * \brief read, special overload of read function, when the read access is not vectorized
+ *
+ * \tparam PacketLoad: determines if the each element of this tensor block should be loaded in a packet mode
+ *
+ * \param is_coalesced_layout: determines whether or not the Tensor data in a memory can be access coalesced and
+  vectorized when possible. Coalesced memory access is a key factor in Kernel performance. When a tensor is 2d and the
+  contracting dimension is 1, it is always possible to accessed tensor data coalesced and vectorized. This is the case
+  when RHS(right hand side) Tensor is transposed or when LHS(left hand side) Tensor is not transposed.
+ *
+ * \tparam PacketType: determines the type of packet
+ *
+ * \tparam TensorMapper: determines the input tensor mapper type
+ *
+ * \tparam StorageIndex: determines the Index type
+
+ * \param tensorMapper: is the input tensor
+ *
+ * \param NCIndex: is the non-contracting dim index
+ *
+ * \param CIndex: is the contracting dim index
+ */
+template <bool PacketLoad, bool, bool IsRhs, typename PacketType, typename TensorMapper, typename StorageIndex>
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<!PacketLoad, PacketType>::type read(
+    const TensorMapper &tensorMapper, const StorageIndex &NCIndex, const StorageIndex &CIndex, const StorageIndex &) {
+  const StorageIndex row = (IsRhs) ? CIndex : NCIndex;
+  const StorageIndex col = (IsRhs) ? NCIndex : CIndex;
+  return tensorMapper(row, col);
+}
+
+/*!
+ * \brief write, a template function used for storing the data to local memory. This function is used to guarantee
+ * coalesced and vectorized store whenever possible.
+ *
+ * \tparam StorageIndex: determines the Index type
+ *
+ * \param ld is the leading dimension of the local memory. ld is a compile time value for the local memory
+ *
+ * \tparam data_source: an enum value representing if the location of the data in a memory hierarchy.
+ *
+ * \tparam PacketType:  determines the type of packet
+ *
+ * \tparam DataScalar: determines the output data type
+ *
+ * \param packet_data: the data to be written in the local memory
+ *
+ * \param ptr: a pointer to the local memory
+ *
+ * \param CIndex is the contracting dim index
+ */
+
+template <typename StorageIndex, StorageIndex ld, data_source dt, typename PacketType, typename DataScalar>
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    typename ::Eigen::internal::enable_if<dt != data_source::global_mem, void>::type
+    write(PacketType &packet_data, DataScalar ptr) {
+  EIGEN_CONSTEXPR int PacketSize = Eigen::internal::unpacket_traits<PacketType>::size;
+  EIGEN_UNROLL_LOOP
+  for (int i = 0; i < PacketSize; i++) {
+    *ptr = PacketWrapper<PacketType, PacketSize>::scalarize(i, packet_data);
+    ptr += ld;
+  }
+}
+
+/*!
+ * \brief Overloading the write function for storing the data to global memory, when vectorization enabled This function
+ * is used to guarantee coalesced and vectorized store whenever possible.
+ *
+ * \tparam data_source: an enum value representing if the location of the data in a memory hierarchy.
+ *
+ * \tparam PacketType:  determines the type of packet
+ *
+ * \tparam DataScalar: determines the output data type
+ *
+ * \param packet_data: the data to be written in the local memory
+ *
+ * \param ptr: a pointer to the local memory
+ */
+
+template <data_source dt, typename PacketType, typename DataScalar>
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<
+    Eigen::internal::unpacket_traits<PacketType>::size != 1 && dt == data_source::global_mem, void>::type
+write(PacketType &packet_data, DataScalar *ptr) {
+  ::Eigen::internal::pstoreu<DataScalar, PacketType>(ptr, packet_data);
+}
+
+/*!
+ * \brief Overloading the write function for storing the data to global memory, when vectorization is disabled.
+ *
+ * \tparam data_source: an enum value representing if the location of the data in a memory hierarchy.
+ *
+ * \tparam PacketType:  determines the type of packet
+ *
+ * \tparam DataScalar: determines the output data type
+ *
+ * \param packet_data: the data to be written in the local memory
+ *
+ * \param ptr: a pointer to the local memory
+ */
+template <data_source dt, typename PacketType, typename DataScalar>
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<
+    Eigen::internal::unpacket_traits<PacketType>::size == 1 && dt == data_source::global_mem, void>::type
+write(PacketType &packet_data, DataScalar *ptr) {
+  *ptr = packet_data;
+}
+
+/*!
+ * \brief check_boundary: is used to check the edge condition for non-internal blocks.
+ *
+ * \tparam is_internal: determines if the block is internal
+ */
+template <bool is_internal>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool check_boundary(bool) {
+  return true;
+}
+
+/*!
+ * \brief check_boundary: specialization of the check_boundary for non-internal blocks.
+ *
+ * \param cond: true when the data is in range. Otherwise false
+ */
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool check_boundary<false>(bool cond) {
+  return cond;
+}
+
+/*!
+ * \brief BlockProperties is a template class that provides different characteristic of a block of each Tensor processed
+ * by each workgroup.
+ *
+ * \tparam is_transposed: iff true, determines whether or not the block of the Tensor is transposed
+ *
+ * \tparam packet_load_: determines if the each element of this tensor block should be loaded in a packet mode
+ *
+ * \tparam PacketType:  determines the type of packet
+ *
+ * \tparam OutType: determines the type of each element for this block of tensor. If packet load is true, it will be
+ * packetType; Otherwise it will be scalar Type
+ *
+ * \param elements_per_access determines the size of each element based on OutType
+ *
+ * \param is_coalesced_layout  determines whether or not the Tensor data in a memory can be access coalesced and
+ * vectorized when possible. Coalesced memory access is a key factor in Kernel performance. When a tensor is 2d and the
+ * contracting dimension is 1, it is always possible to accessed tensor data coalesced and vectorized. This is the case
+ * when RHS(right hand side) Tensor is transposed or when LHS(left hand side) Tensor is not transposed.
+ *
+ * \param nc_stride determines the stride of non-contracting dimension to access the next adjustment element within the
+ * Tensor Block for each workgroup
+ *
+ * \param c_stride  determines the stride of contracting dimension to access the next adjustment element within the
+ * Tensor Block for each workgroup
+ */
+template <bool is_transposed, bool is_rhs_, bool packet_load_, typename PacketType>
+struct BlockProperties {
+  static EIGEN_CONSTEXPR bool packet_load = packet_load_;
+  typedef typename Eigen::internal::unpacket_traits<PacketType>::type OutScalar;
+  static EIGEN_CONSTEXPR bool is_rhs = is_rhs_;
+  typedef typename Eigen::internal::conditional<packet_load, PacketType, OutScalar>::type OutType;
+  static EIGEN_CONSTEXPR int elements_per_access = Eigen::internal::unpacket_traits<OutType>::size;
+  static EIGEN_CONSTEXPR bool is_coalesced_layout = !(is_transposed ^ is_rhs);
+  static EIGEN_CONSTEXPR int nc_stride = (is_coalesced_layout ? elements_per_access : 1);
+  static EIGEN_CONSTEXPR int c_stride = (is_coalesced_layout ? 1 : elements_per_access);
+};
+
+/*!
+ * \brief ThreadProperties is a template class that provides each thread's properties within a workgroup.  Please see
+ * the sycl-1.2.1 specification (https://www.khronos.org/registry/SYCL/specs/sycl-1.2.1.pdf) for the workgroup,
+ * work-items
+ *
+ * \tparam StorageIndex: determines the StorageIndex Type
+ *
+ * \param linearLocalThreadId: determines the linearized location of a thread within a work-group
+ *
+ * \param kGroupId: determines the logical group id in a k dimension of the flattened tensor. It will be > 1 when
+ * tall/skinny algorithm is used
+ *
+ * \param mGroupOffset: determines the logical start position of all thread within a workgroup for the m dimension of
+ * the flattened tensor.
+ *
+ * \param kGroupOffset determines the logical start position of all thread within a workgroup for the k dimension of the
+ * flattened tensor. It will be > 1 when tall/skinny algorithm is used.
+ *
+ * \param mLocalOffset: determines the logical start position of each thread within a workgroup for the m dimension of a
+ * flattened tensor. The position determines the distance of each thread within the workgroup from each other
+ * independent from their global position.
+ *
+ * \param nLocalOffset: determines the logical start position of each thread within a workgroup for the n dimension of a
+ * flattened tensor. The position determines the distance of each thread within the workgroup from each other
+ * independent from their global position.
+ *
+ * \param mGlobalOffset: determines the logical start position of each thread a thread for the m dimension on a
+ * flattened tensor
+ *
+ * \param nGlobalOffset: determines the logical start position of each thread a thread for the n dimension on a
+ * flattened tensor
+ *
+ * \param kSize : determine the number of the k elements of the flattened Tensor to be processed by each thread for the
+ * given tensor block. This is !=K dimension of Flattened Tensor when Tall/Skinny matrix is used.
+ *
+ * \param is_internal : this will determined if the thread within the work-group computes an internal block of tensor or
+ * the edge blocks. When it is internal, there is no need to check the boundaries and all the if stantement can be
+ * resolve by compiler.
+ */
+template <typename StorageIndex>
+struct ThreadProperties {
+  const StorageIndex linearLocalThreadId;
+  const StorageIndex kGroupId;
+  const StorageIndex mGroupOffset;
+  const StorageIndex nGroupOffset;
+  const StorageIndex kGroupOffset;
+  const StorageIndex mLocalOffset;
+  const StorageIndex nLocalOffset;
+  const StorageIndex mGlobalOffset;
+  const StorageIndex nGlobalOffset;
+  StorageIndex kSize;
+  const bool is_internal;
+  // this is used to adjust the last block
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ThreadProperties(
+      const StorageIndex linearLocalThreadId_, const StorageIndex kGroupId_, const StorageIndex mGroupOffset_,
+      const StorageIndex nGroupOffset_, const StorageIndex kGroupOffset_, const StorageIndex mLocalOffset_,
+      const StorageIndex nLocalOffset_, const StorageIndex mGlobalOffset_, const StorageIndex nGlobalOffset_,
+      StorageIndex kSize_, const bool is_internal_)
+      : linearLocalThreadId(linearLocalThreadId_),
+        kGroupId(kGroupId_),
+        mGroupOffset(mGroupOffset_),
+        nGroupOffset(nGroupOffset_),
+        kGroupOffset(kGroupOffset_),
+        mLocalOffset(mLocalOffset_),
+        nLocalOffset(nLocalOffset_),
+        mGlobalOffset(mGlobalOffset_),
+        nGlobalOffset(nGlobalOffset_),
+        kSize(kSize_),
+        is_internal(is_internal_) {}
+};
+
+/*!
+ * \brief TensorContractionKernel is a template class that provides Tensor -Tensor contraction operation.
+ *
+ * \tparam OutScalar: determines the output scalar type
+ *
+ * \tparam LhsScalar: determines the left-hand-side scalar type
+ *
+ * \tparam RhsScalar: determines the right-hand-side scalar type
+ *
+ * \tparam OutAccessor: determines the sycl accessor type for out put (please see the sycl-1.2.1 specification
+ (https://www.khronos.org/registry/SYCL/specs/sycl-1.2.1.pdf) for accessor definition)
+ *
+ * \tparam LhsMapper determines the tensor contraction mapper type for left-hand-side matrix
+ *
+ * \tparam RhsMapper determines the tensor contraction mapper type for right-hand-side matrix
+ *
+ * \tparam StorageIndex: determines the StorageIndex Type
+ *
+ * \tparam Properties: determines the Contraction Panel properties
+ *
+ * \tparam TripleDim: determines the M, K, N dimensions for the flatten tensors in order to treat them as a matrix
+ *
+ * \tparam Vectorizable: determines whether or not the vectorization is enabled for the Eigen expression.
+ *
+ * \tparam input_mapper_properties : determine if the input tensors are matrix. If they are matrix, special memory
+ access is used to guarantee that always the memory access are coalesced.
+ *
+ * \tptaram IsFinal : determine if this is the final kernel. If so, the result will be written in a final output.
+ Otherwise, the result of contraction will be written iin a temporary buffer. This is the case when Tall/Skinny
+ contraction is used. So in this case, a final reduction step is required to compute final output.
+
+ * \tparam contraction_tp: it is an enum value representing whether the local memroy/no local memory implementation of
+ the algorithm to be used
+ *
+ * \param scratch: local memory containing tiles of LHS and RHS tensors for each work-group
+ *
+ * \param lhs: determines the left-hand-side flattened tensor (tensor mapper)
+ *
+ * \param rhs: determines the right-hand-side flattened tensor (tensor mapper)
+ *
+ * \param out_res: determines the output tensor containing the contraction result
+ *
+ * \param groupSizeM: a logical number determining the number of work-group for m dimension
+ *
+ * \param groupSizeN: a logical number determining the number of work-group for n dimension
+ *
+ * \param numTiles: determines total number of tiles on the k dimension
+ *
+ * \param TripleDim: determines the M, K, N dimensions for the flatten tensors in order to treat them as a matrix
+ */
+template <typename OutScalar, typename LhsScalar, typename RhsScalar, typename OutAccessor, typename LhsMapper,
+          typename RhsMapper, typename StorageIndex, typename Properties, typename TripleDim, bool Vectorizable,
+          typename input_mapper_properties, bool IsFinal, contraction_type contraction_tp>
+class TensorContractionKernel {
+ public:
+  typedef typename Eigen::TensorSycl::internal::Vectorise<OutScalar, Eigen::SyclDevice, Vectorizable>::PacketReturnType
+      PacketReturnType;
+  static EIGEN_CONSTEXPR int PacketSize =
+      Eigen::TensorSycl::internal::Vectorise<OutScalar, Eigen::SyclDevice, Vectorizable>::PacketSize;
+  static EIGEN_CONSTEXPR bool is_lhs_transposed =
+      !::Eigen::internal::TensorContractionInputMapperTrait<LhsMapper>::inner_dim_contiguous;
+  static EIGEN_CONSTEXPR bool is_rhs_transposed =
+      !::Eigen::internal::TensorContractionInputMapperTrait<RhsMapper>::inner_dim_contiguous;
+
+  typedef BlockProperties<is_lhs_transposed, false, input_mapper_properties::is_lhs_matrix && Vectorizable,
+                          PacketReturnType>
+      LHSBlockProperties;
+
+  typedef BlockProperties<is_rhs_transposed, true, input_mapper_properties::is_rhs_matrix && Vectorizable,
+                          PacketReturnType>
+      RHSBlockProperties;
+
+  static EIGEN_CONSTEXPR StorageIndex NStride =
+      contraction_tp == contraction_type::local ? Properties::WorkLoadPerThreadN : RHSBlockProperties::nc_stride;
+
+  typedef cl::sycl::accessor<OutScalar, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local> Scratch;
+  typedef cl::sycl::multi_ptr<OutScalar, cl::sycl::access::address_space::local_space> local_ptr;
+  typedef OutScalar * /*cl::sycl::multi_ptr<OutScalar, cl::sycl::access::address_space::private_space>*/ private_ptr;
+  typedef
+      typename ::Eigen::internal::conditional<contraction_tp == contraction_type::local, local_ptr, private_ptr>::type
+          tile_ptr;
+  static EIGEN_CONSTEXPR StorageIndex LSDL = contraction_tp == contraction_type::local
+                                                 ? Properties::TileSizeDimM + Properties::BC
+                                                 : Properties::WorkLoadPerThreadM;
+  static EIGEN_CONSTEXPR StorageIndex LSDR = contraction_tp == contraction_type::local
+                                                 ? Properties::TileSizeDimN + Properties::BC
+                                                 : Properties::WorkLoadPerThreadN;
+  static EIGEN_CONSTEXPR StorageIndex LocalOffset = Properties::LocalThreadSizeM * Properties::LocalThreadSizeN;
+
+  /**
+   * \brief MemHolder this is a place holder struct for creating memory hierarchy in SYCL. Inside SYCL kernel it is not
+   * allowed to have dynamic memory allocation. While the local memory is created outside of the kernel and passed to
+   * the kernel as an accessor, the private memory can only allowed to be allocated statically. Since we are abstracting
+   * the TiledMemory for both local and private memory, the MemHolder structs is used as a helper to abstract out
+   * different type of memory needed when local/no_local memory computation is called.
+   *
+   * \tparam contraction_type: it is an enum value representing whether the local memroy/no local memory implementation
+   of the algorithm to be used
+   * \tparam the private memory size
+   * \param ptr the tile memory pointer type
+   */
+  template <contraction_type, StorageIndex>
+  struct MemHolder {
+    tile_ptr ptr;
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE MemHolder(local_ptr block_start_ptr) : ptr(block_start_ptr) {}
+  };
+  /**
+   * \brief specialization of memHolder class when no local memory kernel is used.
+   */
+  template <StorageIndex MemSize>
+  struct MemHolder<contraction_type::no_local, MemSize> {
+    OutScalar ptr[MemSize] = {OutScalar{0}};
+  };
+  /**
+   * \brief TiledMemory: contains required memory pointer for loading  each tile of the TensorContraction panel from
+   * global memory to local/private memory when local/no_local algorithm used.
+   *
+   * \param lhs_scratch_extract : determines the LHS tile memory. It is either private or local memory based on the
+   * selected contraction_type.
+   *
+   * \param rhs_scratch_extract : determines the RHS tile memory. It is either private or local memory based on the
+   * selected contraction_type.
+   *
+   * \param lhs_extract_index: determins the position of each thread on a local memory for lhs input. When private
+   * memory is used this is set to zero as this is not applicable in case of private memory.
+   *
+   * \param rhs_extract_index: determins the position of each thread on a local memory for rhs input. When private
+   * memory is used this is set to zero as this is not applicable in case of private memory.
+   *
+   * \param lhs_scratch_compute : determines the  location to load for computation for lhs_local memory. This is the
+   * same as lhs_scratch_extract for private memory.
+   *
+   * \param rhs_scratch_compute : determines the  location to load for computation for rhs_local memory. This is the
+   * same as rhs_scratch_extract for private memory.
+   */
+  struct TiledMemory {
+    MemHolder<contraction_tp, Properties::WorkLoadPerThreadM * Properties::TileSizeDimK> lhs_scratch_extract;
+    MemHolder<contraction_tp, Properties::WorkLoadPerThreadN * Properties::TileSizeDimK> rhs_scratch_extract;
+    tile_ptr lhs_scratch_ptr_compute;
+    tile_ptr rhs_scratch_ptr_compute;
+    const std::pair<StorageIndex, StorageIndex> lhs_extract_index;
+    const std::pair<StorageIndex, StorageIndex> rhs_extract_index;
+    template <contraction_type tp = contraction_tp>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TiledMemory(const ThreadProperties<StorageIndex> &, local_ptr,
+                typename ::Eigen::internal::enable_if<tp == contraction_type::no_local>::type * = 0)
+        : lhs_scratch_extract{},
+          rhs_scratch_extract{},
+          lhs_scratch_ptr_compute(lhs_scratch_extract.ptr),
+          rhs_scratch_ptr_compute(rhs_scratch_extract.ptr),
+          lhs_extract_index(std::pair<StorageIndex, StorageIndex>(StorageIndex{0}, StorageIndex{0})),
+          rhs_extract_index(std::pair<StorageIndex, StorageIndex>(StorageIndex{0}, StorageIndex{0})) {}
+
+    template <contraction_type tp = contraction_tp>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TiledMemory(const ThreadProperties<StorageIndex> &thread_properties, local_ptr block_start_ptr,
+                typename ::Eigen::internal::enable_if<tp == contraction_type::local>::type * = 0)
+        : lhs_scratch_extract{block_start_ptr},
+          rhs_scratch_extract{lhs_scratch_extract.ptr +
+                              ((Properties::DoubleBuffer + 1) * LSDL * Properties::TileSizeDimK)},
+          lhs_scratch_ptr_compute(lhs_scratch_extract.ptr + thread_properties.mLocalOffset),
+          rhs_scratch_ptr_compute(rhs_scratch_extract.ptr + thread_properties.nLocalOffset),
+          lhs_extract_index(
+              local_id_extract<LHSBlockProperties, Properties::TileSizeDimM>(thread_properties.linearLocalThreadId)),
+          rhs_extract_index(
+              local_id_extract<RHSBlockProperties, Properties::TileSizeDimN>(thread_properties.linearLocalThreadId)) {}
+  };
+
+  Scratch scratch;
+  const LhsMapper lhs;
+  const RhsMapper rhs;
+  OutAccessor out_res;
+  const StorageIndex groupSizeM;
+  const StorageIndex groupSizeN;
+  const StorageIndex numTiles;
+  const TripleDim triple_dim;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorContractionKernel(Scratch scratch_, const LhsMapper lhs_,
+                                                                const RhsMapper rhs_, OutAccessor out_res_,
+                                                                const StorageIndex groupSizeM_,
+                                                                const StorageIndex groupSizeN_,
+                                                                const StorageIndex numTiles_,
+                                                                const TripleDim triple_dim_)
+      : scratch(scratch_),
+        lhs(lhs_),
+        rhs(rhs_),
+        out_res(out_res_),
+        groupSizeM(groupSizeM_),
+        groupSizeN(groupSizeN_),
+        numTiles(numTiles_),
+        triple_dim(triple_dim_) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorContractionKernel(Scratch scratch_, const LhsMapper lhs_,
+                                                                const RhsMapper rhs_, OutAccessor out_res_,
+                                                                const StorageIndex groupSizeM_,
+                                                                const StorageIndex numTiles_,
+                                                                const TripleDim triple_dim_)
+      : TensorContractionKernel(scratch_, lhs_, rhs_, out_res_, groupSizeM_, 1, numTiles_, triple_dim_) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(cl::sycl::nd_item<1> itemID) {
+    const StorageIndex linearLocalThreadId = itemID.get_local_id(0);
+    const StorageIndex nLocalThreadId = linearLocalThreadId / Properties::LocalThreadSizeM;
+    const StorageIndex mLocalThreadId = linearLocalThreadId % Properties::LocalThreadSizeM;
+    const StorageIndex mGroupId = itemID.get_group(0) % groupSizeM;
+    const StorageIndex tmp = itemID.get_group(0) / groupSizeM;
+    const StorageIndex nGroupId = IsFinal ? tmp : tmp % groupSizeN;
+    const StorageIndex kGroupId = IsFinal ? 0 : tmp / groupSizeN;
+    const StorageIndex mGroupOffset = mGroupId * Properties::TileSizeDimM;
+    const StorageIndex nGroupOffset = nGroupId * Properties::TileSizeDimN;
+    const StorageIndex mLocalOffset = PacketSize * mLocalThreadId;
+    const StorageIndex nLocalOffset = NStride * nLocalThreadId;
+    const StorageIndex mGlobalOffset = mGroupOffset + mLocalOffset;
+    const StorageIndex nGlobalOffset = nGroupOffset + nLocalOffset;
+
+    const StorageIndex kSizePerWG = IsFinal ? triple_dim.K : numTiles * Properties::TileSizeDimK;
+    StorageIndex kGroupOffset = kGroupId * kSizePerWG;
+    const bool is_internal = triple_dim.M - mGroupOffset >= Properties::TileSizeDimM &&
+                             triple_dim.N - nGroupOffset >= Properties::TileSizeDimN &&
+                             triple_dim.K - kGroupOffset >= kSizePerWG;
+    // this is used to adjust the last block
+    StorageIndex kSize = IsFinal ? triple_dim.K : std::min(kSizePerWG, triple_dim.K - kGroupOffset);
+    // This is used to find out the lats K offset so that kGroupOffset -kSize can compute the coffset for loading to
+    // tile
+    kGroupOffset += kSize;
+
+    auto thread_properties =
+        ThreadProperties<StorageIndex>(linearLocalThreadId, kGroupId, mGroupOffset, nGroupOffset, kGroupOffset,
+                                       mLocalOffset, nLocalOffset, mGlobalOffset, nGlobalOffset, kSize, is_internal);
+
+    auto out_ptr = out_res.get_pointer() + (IsFinal ? 0 : thread_properties.kGroupId * triple_dim.M * triple_dim.N);
+
+    (thread_properties.is_internal) ? compute_panel<true>(itemID, thread_properties, out_ptr)
+                                    : compute_panel<false>(itemID, thread_properties, out_ptr);
+  }
+  // The compute block computes the contraction operation private block for each thread and store the resutl in the
+  // privateRes memory of Each computation the compute block function is independent of local and no local concepts as
+  // it only compute the block on each thread's private memory space
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void compute_block_per_tile(OutScalar *lhs_block_ptr, OutScalar *rhs_block_ptr,
+                                                                    PacketReturnType *privateRes) {
+    StorageIndex idx = 0;
+    EIGEN_CONSTEXPR StorageIndex lhs_stride =
+        contraction_tp == contraction_type::local ? (PacketSize * Properties::LocalThreadSizeM) : 1;
+    EIGEN_UNROLL_LOOP
+    for (StorageIndex wLPTN = 0; wLPTN < Properties::WorkLoadPerThreadN; wLPTN++) {
+      auto rhsPacket = PacketReturnType{*(rhs_block_ptr + wLPTN)};
+      StorageIndex lhs_index = 0;
+      EIGEN_UNROLL_LOOP
+      for (StorageIndex wLPTM = 0; wLPTM < Properties::WorkLoadPerThreadM / PacketSize; wLPTM++) {
+        PacketReturnType lhsPack{};
+        Eigen::TensorSycl::internal::PacketWrapper<PacketReturnType, PacketSize>::set_packet(lhsPack,
+                                                                                             lhs_block_ptr + lhs_index);
+        privateRes[idx] = ::Eigen::internal::pmadd(lhsPack, rhsPacket, privateRes[idx]);
+
+        lhs_index += lhs_stride;
+        idx++;
+      }
+    }
+  }
+  // The store function write the computed contraction operation in the private memory of each thread to the global
+  // memory. The store function is independent of local and no local concepts s that it can be abstract out in the base
+  // class.
+  template <bool is_internal_block, StorageIndex PrivateNStride, typename OutPtr>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void store(OutPtr *out_ptr, PacketReturnType *privateRes,
+                                                   StorageIndex mGlobalOffset, StorageIndex nGlobalOffset) {
+    auto chk_bound = [&](const StorageIndex &mIndex, const StorageIndex &nIndex) EIGEN_DEVICE_FUNC {
+      return (mIndex + PacketSize - 1 < triple_dim.M && nGlobalOffset + nIndex < triple_dim.N);
+    };
+    // when local memory is not used M and N are both accessed in a coalesced way. However, when local memory is
+    // available the k*N is transposed in the local to N*K therefore, each blocks operates on blockId*
+    // WorkLoadPerThreadN slice of N
+    EIGEN_CONSTEXPR StorageIndex GlobalNStride =
+        contraction_tp == contraction_type::local ? 1 : Properties::LocalThreadSizeN;
+    EIGEN_UNROLL_LOOP
+    for (StorageIndex wLPTN = 0; wLPTN < Properties::WorkLoadPerThreadN / PrivateNStride; wLPTN++) {
+      // output leading dimension
+      StorageIndex outputLD = 0;
+      // When local memory is used the PrivateNstride is always 1 because the coalesed access on N is loaded into Local
+      // memory and extracting from local to global is the same as no transposed version. However, when local memory is
+      // not used and RHS is transposed we packetize the load for RHS.
+      EIGEN_UNROLL_LOOP
+      for (StorageIndex nId = 0; nId < PrivateNStride; nId++) {
+        StorageIndex globalRow = mGlobalOffset;
+        EIGEN_UNROLL_LOOP
+        for (StorageIndex wLPTM = 0; wLPTM < Properties::WorkLoadPerThreadM / PacketSize; wLPTM++) {
+          PacketReturnType privetOut = privateRes[wLPTM];
+          if (check_boundary<is_internal_block>(chk_bound(globalRow, nId))) {
+            // Store the final results in C. The C matrix has always M as a first StorageIndex and N as a second
+            // StorageIndex Therefore it is always coalesced layout
+            write<data_source::global_mem>(privetOut, out_ptr + outputLD + globalRow);
+          } else {
+            EIGEN_UNROLL_LOOP
+            for (StorageIndex mId = 0; mId < PacketSize; mId++) {
+              StorageIndex mOffset = globalRow + mId;
+              if (mOffset < triple_dim.M && (nGlobalOffset + nId < triple_dim.N)) {
+                out_ptr[mOffset + outputLD] =
+                    Eigen::TensorSycl::internal::PacketWrapper<PacketReturnType, PacketSize>::scalarize(mId, privetOut);
+              }
+            }
+          }
+          globalRow += (PacketSize * Properties::LocalThreadSizeM);
+        }
+        outputLD += triple_dim.M;
+        privateRes += Properties::WorkLoadPerThreadM / PacketSize;
+      }
+      out_ptr += (GlobalNStride * outputLD);
+
+      nGlobalOffset += (PrivateNStride * GlobalNStride);
+    }
+  }
+  // when no local memory is used the following extract_block will be enabled
+  template <typename InputBlockProperties, bool is_internal_block, typename Input, typename PrivateReg,
+            contraction_type contract_tp = contraction_tp>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+      typename ::Eigen::internal::enable_if<contract_tp == contraction_type::no_local>::type
+      extract_block(const Input &inpt, PrivateReg private_ptr, const std::pair<StorageIndex, StorageIndex> &,
+                    const StorageIndex &ncOffset, const StorageIndex cOffset) {
+    EIGEN_CONSTEXPR StorageIndex LocalThreadSizeNC =
+        InputBlockProperties::is_rhs ? Properties::LocalThreadSizeN : Properties::LocalThreadSizeM;
+    EIGEN_CONSTEXPR StorageIndex WorkLoadPerThreadNC =
+        InputBlockProperties::is_rhs ? Properties::WorkLoadPerThreadN : Properties::WorkLoadPerThreadM;
+    const StorageIndex &NC = InputBlockProperties::is_rhs ? triple_dim.N : triple_dim.M;
+
+    auto chk_bound = [&](const StorageIndex &CIndex, const StorageIndex &NCIndex) EIGEN_DEVICE_FUNC {
+      return ((CIndex + InputBlockProperties::c_stride - 1 < triple_dim.K) &&
+              (NCIndex + InputBlockProperties::nc_stride - 1 < NC));
+    };
+    const StorageIndex ld = InputBlockProperties::is_coalesced_layout ? NC : triple_dim.K;
+    StorageIndex cIndex = cOffset;
+
+    EIGEN_UNROLL_LOOP
+    for (StorageIndex cId = 0; cId < Properties::TileSizeDimK / InputBlockProperties::c_stride; cId++) {
+      StorageIndex ncIndex = ncOffset;
+      EIGEN_UNROLL_LOOP
+      for (StorageIndex ncId = 0; ncId < WorkLoadPerThreadNC / InputBlockProperties::nc_stride; ncId++) {
+        if (check_boundary<is_internal_block>(chk_bound(cIndex, ncIndex))) {
+          auto val =
+              read<InputBlockProperties::packet_load, InputBlockProperties::is_coalesced_layout,
+                   InputBlockProperties::is_rhs, typename InputBlockProperties::OutType>(inpt, ncIndex, cIndex, ld);
+
+          write<StorageIndex, (InputBlockProperties::is_coalesced_layout ? 1 : WorkLoadPerThreadNC),
+                data_source::private_mem>(val, private_ptr);
+        } else {
+          EIGEN_UNROLL_LOOP
+          for (StorageIndex i = 0; i < InputBlockProperties::elements_per_access; i++) {
+            const StorageIndex ncInd = ncIndex + (InputBlockProperties::is_coalesced_layout ? i : 0);
+            const StorageIndex cInd = cIndex + (InputBlockProperties::is_coalesced_layout ? 0 : i);
+            OutScalar val =
+                (ncInd < NC && cInd < triple_dim.K)
+                    ? read<false, InputBlockProperties::is_coalesced_layout, InputBlockProperties::is_rhs, OutScalar>(
+                          inpt, ncInd, cInd, ld)
+                    : OutScalar(0);
+            write<StorageIndex, (InputBlockProperties::is_coalesced_layout ? 1 : WorkLoadPerThreadNC),
+                  data_source::private_mem>(
+                val, private_ptr + (InputBlockProperties::is_coalesced_layout ? i : 0) +
+                         ((InputBlockProperties::is_coalesced_layout ? 0 : i) * WorkLoadPerThreadNC));
+          }
+        }
+
+        // if it is lhs we have to load it packetised when the packet size is > 1, because the output is coalesced. So
+        // even if M is not accessed in a coalesced mode, we have to load packet_size number of m per thread.
+        ncIndex = (!InputBlockProperties::is_rhs && InputBlockProperties::nc_stride == 1 && PacketSize != 1)
+                      ? ncOffset + (ncId + 1) % PacketSize + ((ncId + 1) / PacketSize) * LocalThreadSizeNC
+                      : (ncIndex + InputBlockProperties::nc_stride * LocalThreadSizeNC);
+        private_ptr += InputBlockProperties::nc_stride;
+      }
+      // the previous for loop ( private_ptr += (ncId * nc_stride)) has already moved ptr with one WorkLoadPerThreadNC
+      private_ptr += (InputBlockProperties::c_stride - 1) * WorkLoadPerThreadNC;
+      cIndex += InputBlockProperties::c_stride;
+    }
+  }
+  template <typename InputBlockProperties, StorageIndex TileSizeDimNC>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::pair<StorageIndex, StorageIndex> local_id_extract(
+      const StorageIndex &linearLocalThreadId) {
+    const StorageIndex localThreadNC =
+        (InputBlockProperties::is_coalesced_layout)
+            ? linearLocalThreadId % (TileSizeDimNC / InputBlockProperties::nc_stride)
+            : linearLocalThreadId / (Properties::TileSizeDimK / InputBlockProperties::c_stride);
+    const StorageIndex localThreadC =
+        (InputBlockProperties::is_coalesced_layout)
+            ? linearLocalThreadId / (TileSizeDimNC / InputBlockProperties::nc_stride)
+            : linearLocalThreadId % (Properties::TileSizeDimK / InputBlockProperties::c_stride);
+    return std::pair<StorageIndex, StorageIndex>(localThreadNC, localThreadC);
+  }
+
+  template <bool db = Properties::DoubleBuffer, contraction_type ctp = contraction_tp>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+      typename ::Eigen::internal::enable_if<db && ctp == contraction_type::local>::type
+      sync_mem(const cl::sycl::nd_item<1> &, bool &db_offset) noexcept {
+    db_offset = !db_offset;
+  }
+
+  template <bool db = Properties::DoubleBuffer, contraction_type ctp = contraction_tp>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+      typename ::Eigen::internal::enable_if<!db && ctp == contraction_type::local>::type
+      sync_mem(const cl::sycl::nd_item<1> &itemID, bool &) noexcept {
+    itemID.barrier(cl::sycl::access::fence_space::local_space);
+  }
+
+  template <contraction_type ctp = contraction_tp>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+      typename ::Eigen::internal::enable_if<ctp == contraction_type::no_local>::type
+      sync_mem(const cl::sycl::nd_item<1> &, bool &) noexcept {
+    return;
+  }
+
+  template <bool need_sync, contraction_type ctp = contraction_tp>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+      typename ::Eigen::internal::enable_if<need_sync && ctp == contraction_type::no_local>::type
+      sync_thread(const cl::sycl::nd_item<1> &
+#ifdef EIGEN_SYCL_ARM_GPU_CACHE_OPTIMISATION
+                      itemID
+#endif
+                  ) noexcept {
+#ifdef EIGEN_SYCL_ARM_GPU_CACHE_OPTIMISATION
+    itemID.barrier(cl::sycl::access::fence_spacce::local_space);
+#else
+    return;
+#endif
+  }
+  template <bool need_sync, contraction_type ctp = contraction_tp>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+      typename ::Eigen::internal::enable_if<need_sync && ctp == contraction_type::local>::type
+      sync_thread(const cl::sycl::nd_item<1> &itemID) {
+    itemID.barrier(cl::sycl::access::fence_space::local_space);
+  }
+  template <bool need_sync>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<!need_sync>::type sync_thread(
+      const cl::sycl::nd_item<1> &) {
+    return;
+  }
+
+  template <bool is_internal_block>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void compute_tile_per_panel(const cl::sycl::nd_item<1> &itemID,
+                                                                    ThreadProperties<StorageIndex> &thread_properties,
+                                                                    TiledMemory &tiled_input_block,
+                                                                    PacketReturnType *privateRes, bool &db_offset) {
+    // Tiling the Rhs block from global to local memory
+    extract_block<RHSBlockProperties, is_internal_block>(
+        rhs, tiled_input_block.rhs_scratch_extract.ptr + (db_offset * Properties::TileSizeDimK * LSDR),
+        tiled_input_block.rhs_extract_index,
+        contraction_tp == contraction_type::local ? thread_properties.nGroupOffset : thread_properties.nGlobalOffset,
+        thread_properties.kGroupOffset - thread_properties.kSize);
+
+    sync_thread<contraction_tp == contraction_type::no_local>(itemID);
+
+    // Tiling the Lhs block from global to local memory
+    extract_block<LHSBlockProperties, is_internal_block>(
+        lhs, tiled_input_block.lhs_scratch_extract.ptr + (db_offset * LSDL * Properties::TileSizeDimK),
+        tiled_input_block.lhs_extract_index,
+        contraction_tp == contraction_type::local ? thread_properties.mGroupOffset : thread_properties.mGlobalOffset,
+        thread_properties.kGroupOffset - thread_properties.kSize);
+
+    // itemID.barrier(cl::sycl::access::fence_space::local_space);
+    sync_thread<contraction_tp == contraction_type::local>(itemID);
+    // switch to compute mede
+    StorageIndex lhs_offset = (db_offset * LSDL * Properties::TileSizeDimK);
+    StorageIndex rhs_offset = (db_offset * Properties::TileSizeDimK * LSDR);
+    // Loop over the values of a single tile
+    for (StorageIndex k = 0; k < Properties::TileSizeDimK; k++) {
+      compute_block_per_tile(tiled_input_block.lhs_scratch_ptr_compute + lhs_offset,
+                             tiled_input_block.rhs_scratch_ptr_compute + rhs_offset, privateRes);
+      lhs_offset += LSDL;
+      rhs_offset += LSDR;
+    }
+    // computing the K index for the next tile
+    thread_properties.kSize -= Properties::TileSizeDimK;
+    sync_mem(itemID, db_offset);
+  }
+
+  // when local memory is available the following compute_panel will be enabled
+  template <bool is_internal_block, typename OutPtr>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void compute_panel(const cl::sycl::nd_item<1> &itemID,
+                                                           ThreadProperties<StorageIndex> &thread_properties,
+                                                           OutPtr out_ptr) {
+    auto tiled_input_block = TiledMemory{thread_properties, scratch.get_pointer()};
+    // Allocate register space
+    PacketReturnType privateRes[Properties::WorkLoadPerThreadM * Properties::WorkLoadPerThreadN / PacketSize] = {
+        PacketReturnType{0}};
+    bool db_offset = 0;
+
+    while (thread_properties.kSize >= Properties::TileSizeDimK) {
+      compute_tile_per_panel<is_internal_block>(itemID, thread_properties, tiled_input_block, privateRes, db_offset);
+    }
+    if (thread_properties.kSize > 0) {
+      compute_tile_per_panel<false>(itemID, thread_properties, tiled_input_block, privateRes, db_offset);
+    }
+
+    // Storing the final results in the output
+    store<is_internal_block,
+          contraction_tp == contraction_type::local ? static_cast<StorageIndex>(1) : RHSBlockProperties::nc_stride>(
+        out_ptr + thread_properties.nGlobalOffset * triple_dim.M, privateRes, thread_properties.mGlobalOffset,
+        thread_properties.nGlobalOffset);
+  }
+  // When local memory is available the following extract_block will be enabled
+  template <typename InputBlockProperties, bool is_internal_block, typename Input, typename Local,
+            contraction_type contract_tp = contraction_tp>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+      typename ::Eigen::internal::enable_if<contract_tp == contraction_type::local>::type
+      extract_block(const Input &inpt, Local local_ptr, const std::pair<StorageIndex, StorageIndex>& local_index,
+                    const StorageIndex &ncOffset, const StorageIndex cOffset) {
+    EIGEN_CONSTEXPR StorageIndex TileSizeDimNC =
+        InputBlockProperties::is_rhs ? Properties::TileSizeDimN : Properties::TileSizeDimM;
+    EIGEN_CONSTEXPR StorageIndex LoadPerThread =
+        InputBlockProperties::is_rhs ? Properties::LoadPerThreadRhs : Properties::LoadPerThreadLhs;
+    EIGEN_CONSTEXPR StorageIndex LSD = InputBlockProperties::is_rhs ? LSDR : LSDL;
+    static_assert(((LocalOffset % (TileSizeDimNC / InputBlockProperties::nc_stride) == 0) &&
+                   (LocalOffset % (Properties::TileSizeDimK / InputBlockProperties::c_stride) == 0)),
+                  " LocalOffset must be divisable by stride");
+    const StorageIndex &NC = InputBlockProperties::is_rhs ? triple_dim.N : triple_dim.M;
+    StorageIndex localThreadNC = local_index.first;
+    StorageIndex localThreadC = local_index.second;
+    auto chk_bound = [&](const StorageIndex &CIndex, const StorageIndex &NCIndex) EIGEN_DEVICE_FUNC {
+      return ((CIndex + InputBlockProperties::c_stride - 1 < triple_dim.K) &&
+              (NCIndex + InputBlockProperties::nc_stride - 1 < NC));
+    };
+    EIGEN_UNROLL_LOOP
+    for (StorageIndex lPT = 0; lPT < LoadPerThread / InputBlockProperties::elements_per_access; lPT++) {
+      const StorageIndex CIndex = cOffset + (InputBlockProperties::c_stride * localThreadC);
+      const StorageIndex NCIndex = ncOffset + (InputBlockProperties::nc_stride * localThreadNC);
+      const StorageIndex ld = InputBlockProperties::is_coalesced_layout ? NC : triple_dim.K;
+      if (check_boundary<is_internal_block>(chk_bound(CIndex, NCIndex))) {
+        auto val =
+            read<InputBlockProperties::packet_load, InputBlockProperties::is_coalesced_layout,
+                 InputBlockProperties::is_rhs, typename InputBlockProperties::OutType>(inpt, NCIndex, CIndex, ld);
+        write<StorageIndex, (InputBlockProperties::is_coalesced_layout ? 1 : LSD), data_source::local_mem>(
+            val, local_ptr + (InputBlockProperties::nc_stride * localThreadNC) +
+                     (InputBlockProperties::c_stride * localThreadC * LSD));
+      } else {
+        EIGEN_UNROLL_LOOP
+        for (StorageIndex i = 0; i < InputBlockProperties::elements_per_access; i++) {
+          const StorageIndex nCInd = NCIndex + (InputBlockProperties::is_coalesced_layout ? i : 0);
+          const StorageIndex cInd = CIndex + (InputBlockProperties::is_coalesced_layout ? 0 : i);
+          OutScalar val =
+              (nCInd < NC && cInd < triple_dim.K)
+                  ? read<false, InputBlockProperties::is_coalesced_layout, InputBlockProperties::is_rhs, OutScalar>(
+                        inpt, nCInd, cInd, ld)
+                  : OutScalar(0);
+
+          write<StorageIndex, (InputBlockProperties::is_coalesced_layout ? 1 : LSD), data_source::local_mem>(
+              val, local_ptr + (InputBlockProperties::nc_stride * localThreadNC) +
+                       (InputBlockProperties::is_coalesced_layout ? i : 0) +
+                       ((InputBlockProperties::c_stride * localThreadC +
+                         (InputBlockProperties::is_coalesced_layout ? 0 : i)) *
+                        LSD));
+        }
+      }
+      localThreadNC += (InputBlockProperties::is_coalesced_layout)
+                           ? LocalOffset % (TileSizeDimNC / InputBlockProperties::nc_stride)
+                           : LocalOffset / (Properties::TileSizeDimK / InputBlockProperties::c_stride);
+      localThreadC += (InputBlockProperties::is_coalesced_layout)
+                          ? LocalOffset / (TileSizeDimNC / InputBlockProperties::nc_stride)
+                          : LocalOffset % (Properties::TileSizeDimK / InputBlockProperties::c_stride);
+    }
+  }
+};
+
+#ifndef EIGEN_SYCL_DISABLE_GEMV
+
+/*!
+ * \brief GeneralVectorTensor is a template class that provides Tensor -vector contraction operation, which is a special
+ * case of Tensor Tensor contraction.
+ *
+ * \tparam OutScalar: determines the output scalar type
+ *
+ * \tparam OutAccessor: determines the sycl accessor type for out put (please see the sycl-1.2.1 specification
+ * (https://www.khronos.org/registry/SYCL/specs/sycl-1.2.1.pdf) for accessor definition)
+ *
+ * \tparam VectorMapper: determines the tensor contraction mapper for the vector input (can be lhs or rhs)
+ *
+ * \tparam TensorMapper: determines the tensor contraction mapper for the tensor input (can be lhs or rhs)
+ *
+ * \tparam StorageIndex: determines the StorageIndex Type
+ *
+ * \tparam Properties: determines the Contraction Panel properties
+ *
+ * \tparam KFactor: determines the number of elements in K dimension in a Tile
+ *
+ * \tparam Vectorizable: determines whether or not the vectorization is enabled for the Eigen expression.
+ *
+ * \tparam is_lhs_vec: determines whether lhs is a vector or rhs is a vector
+ *
+ * \tparam IsFinal: determine if this is the final kernel. If so, the result will be written in a final output.
+ * Otherwise, the result of contraction will be written iin a temporary buffer.
+ *
+ * \param scratch: determines the local memory containing the vector block for each work-group
+ *
+ * \param vec: determines the vector input (tensor mapper)
+ *
+ * \param mat: determines the tensor input (tensor mapper)
+ *
+ * \param out_res: determines the output vector containing the contraction result
+ *
+ * \param nonContractGroupSize: a logical number determining the number of work-group for non-contracting dimension
+ *
+ * \param nonContractDim: determines the size of non contracting dimension for the flattened tensor
+ *
+ * \param contractDim: determines the size of non contracting dimension for the flattened tensor
+ *
+ */
+template <typename OutScalar, typename OutAccessor, typename VectorMapper, typename TensorMapper, typename StorageIndex,
+          typename Properties, StorageIndex KFactor, bool Vectorizable, bool is_lhs_vec, bool IsFinal>
+struct GeneralVectorTensor {
+  typedef typename Eigen::TensorSycl::internal::Vectorise<OutScalar, Eigen::SyclDevice, Vectorizable>::PacketReturnType
+      PacketReturnType;
+  static EIGEN_CONSTEXPR int PacketSize =
+      Eigen::TensorSycl::internal::Vectorise<OutScalar, Eigen::SyclDevice, Vectorizable>::PacketSize;
+  typedef cl::sycl::accessor<OutScalar, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local> Scratch;
+
+  static EIGEN_CONSTEXPR StorageIndex OutScratchOffset =
+      KFactor * Properties::LocalThreadSizeC * Properties::LocalThreadSizeNC;
+
+  // Since the access layout for a vector can always be coalesced, when LHS is a vector, we pass false and false to make
+  // sure that the !^ is true When RHS is a vector, we pass true and true to make sure that the !^ is true.
+  typedef BlockProperties<is_lhs_vec ? false : true, is_lhs_vec ? false : true, Vectorizable, PacketReturnType>
+      VecBlockProperties;
+
+  Scratch scratch;
+  const VectorMapper vec;
+  const TensorMapper mat;
+  OutAccessor out_res;
+  const StorageIndex nonContractGroupSize;
+  const StorageIndex nonContractDim;
+  const StorageIndex contractDim;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE GeneralVectorTensor(Scratch scratch_, const VectorMapper vec_,
+                                                            const TensorMapper mat_, OutAccessor out_res_,
+                                                            const StorageIndex nonContractGroupSize_,
+                                                            const StorageIndex nonContractDim_,
+                                                            const StorageIndex contractDim_)
+      : scratch(scratch_),
+        vec(vec_),
+        mat(mat_),
+        out_res(out_res_),
+        nonContractGroupSize(nonContractGroupSize_),
+        nonContractDim(nonContractDim_),
+        contractDim(contractDim_) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(cl::sycl::nd_item<1> itemID) {
+    auto scratch_ptr = scratch.get_pointer();
+    const StorageIndex linearLocalThreadId = itemID.get_local_id(0);
+    StorageIndex nonContractId = is_lhs_vec ? linearLocalThreadId / Properties::LocalThreadSizeC
+                                            : linearLocalThreadId % Properties::LocalThreadSizeNC;
+    StorageIndex contractId = is_lhs_vec ? linearLocalThreadId % Properties::LocalThreadSizeC
+                                         : linearLocalThreadId / Properties::LocalThreadSizeNC;
+    const StorageIndex cGroupSize = itemID.get_group_range(0) / nonContractGroupSize;
+    const StorageIndex nonContractGroupId =
+        is_lhs_vec ? itemID.get_group(0) / cGroupSize : itemID.get_group(0) % nonContractGroupSize;
+    const StorageIndex contractGroupId =
+        is_lhs_vec ? itemID.get_group(0) % cGroupSize : itemID.get_group(0) / nonContractGroupSize;
+    auto out_ptr = out_res.get_pointer() + (IsFinal ? 0 : contractGroupId * nonContractDim);
+
+    const StorageIndex nonContractGroupOffset = nonContractGroupId * Properties::TileSizeDimNC;
+    const StorageIndex contractGroupOffset = contractGroupId * Properties::TileSizeDimC;
+    auto outScratchIndex = nonContractId + contractId * Properties::LocalThreadSizeNC;
+    const StorageIndex globalNonContractDimOffset = nonContractGroupOffset + nonContractId;
+    const StorageIndex globalContractDimOffset = contractGroupOffset + contractId;
+    auto local_output = scratch_ptr + OutScratchOffset;
+    const bool is_internal = nonContractDim - nonContractGroupOffset >= Properties::TileSizeDimNC &&
+                             contractDim - contractGroupOffset >= Properties::TileSizeDimC;
+    is_internal
+        ? compute_panel<true>(itemID, vec, mat, local_output, out_ptr,
+#ifdef EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON
+                              scratch_ptr, contractGroupOffset,
+#endif
+                              nonContractGroupOffset, linearLocalThreadId, contractDim, nonContractDim, contractId,
+                              nonContractId, globalContractDimOffset, globalNonContractDimOffset, outScratchIndex)
+        : compute_panel<false>(itemID, vec, mat, local_output, out_ptr,
+#ifdef EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON
+                               scratch_ptr, contractGroupOffset,
+#endif
+                               nonContractGroupOffset, linearLocalThreadId, contractDim, nonContractDim, contractId,
+                               nonContractId, globalContractDimOffset, globalNonContractDimOffset, outScratchIndex);
+  }
+  template <bool is_internal_block, typename OutPtr>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void compute_panel(
+      const cl::sycl::nd_item<1> &itemID, const VectorMapper &vec, const TensorMapper &mat, OutScalar *local_output,
+      OutPtr out_ptr,
+#ifdef EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON
+      OutScalar *scratch_ptr, const StorageIndex contractGroupOffset,
+#endif
+      const StorageIndex nonContractGroupOffset, const StorageIndex linearLocalThreadId, StorageIndex contractDim,
+      StorageIndex nonContractDim, StorageIndex contractId, StorageIndex nonContractId,
+      StorageIndex globalContractDimOffset, StorageIndex globalNonContractDimOffset, StorageIndex outScratchIndex) {
+    OutScalar outScalar[Properties::WorkLoadPerThreadNC] = {OutScalar(0)};
+    // Reading the vector
+#ifdef EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON
+    const StorageIndex vectorOffset = contractGroupOffset + linearLocalThreadId;
+    extract_block<VecBlockProperties, is_internal_block, KFactor,
+                  Properties::LocalThreadSizeNC * Properties::LocalThreadSizeC>(vec, scratch_ptr, linearLocalThreadId,
+                                                                                vectorOffset, contractDim);
+
+    itemID.barrier(cl::sycl::access::fence_space::local_space);
+    auto in_scratch_ptr = scratch_ptr + contractId;
+#endif
+
+    StorageIndex privateOffsetC = 0;
+    EIGEN_UNROLL_LOOP
+    for (StorageIndex i = 0; i < Properties::WorkLoadPerThreadC; i++) {
+      StorageIndex privateOffsetNC = 0;
+      bool contract_conds = ((globalContractDimOffset + privateOffsetC) < contractDim);
+#ifdef EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON
+      auto vecScalar = *in_scratch_ptr;
+#else
+      auto vecScalar = (check_boundary<is_internal_block>(contract_conds))
+                           ? vec(is_lhs_vec ? StorageIndex(0) : globalContractDimOffset + privateOffsetC,
+                                 is_lhs_vec ? globalContractDimOffset + privateOffsetC : StorageIndex(0))
+                           : OutScalar(0);
+#endif
+      EIGEN_UNROLL_LOOP
+      for (StorageIndex j = 0; j < Properties::WorkLoadPerThreadNC; j++) {
+        auto matScalar = (check_boundary<is_internal_block>(
+                             contract_conds && ((globalNonContractDimOffset + privateOffsetNC) < nonContractDim)))
+                             ? mat(is_lhs_vec ? globalContractDimOffset + privateOffsetC
+                                              : globalNonContractDimOffset + privateOffsetNC,
+                                   is_lhs_vec ? globalNonContractDimOffset + privateOffsetNC
+                                              : globalContractDimOffset + privateOffsetC)
+                             : OutScalar(0);
+
+        outScalar[j] = cl::sycl::mad(matScalar, vecScalar, outScalar[j]);
+        privateOffsetNC += Properties::LocalThreadSizeNC;
+      }
+      privateOffsetC += Properties::LocalThreadSizeC;
+#ifdef EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON
+      in_scratch_ptr += Properties::LocalThreadSizeC;
+#endif
+    }
+
+    auto out_scratch_ptr = local_output + outScratchIndex;
+    // Each block of 16*16 element in shared memory should reduce to 16*1
+    EIGEN_UNROLL_LOOP
+    for (StorageIndex j = 0; j < Properties::WorkLoadPerThreadNC; j++) {
+      *out_scratch_ptr = outScalar[j];
+
+      out_scratch_ptr += (Properties::LocalThreadSizeNC * Properties::LocalThreadSizeC);
+    }
+    if (is_lhs_vec) {
+      nonContractId = linearLocalThreadId % Properties::LocalThreadSizeNC;
+      contractId = linearLocalThreadId / Properties::LocalThreadSizeNC;
+      outScratchIndex = nonContractId + contractId * Properties::LocalThreadSizeNC;
+    }
+
+    out_scratch_ptr = local_output + outScratchIndex;
+    EIGEN_UNROLL_LOOP
+    for (StorageIndex j = 0; j < Properties::WorkLoadPerThreadNC; j++) {
+      EIGEN_UNROLL_LOOP
+      for (StorageIndex offset = Properties::LocalThreadSizeC >> 1; offset > 0; offset >>= 1) {
+        itemID.barrier(cl::sycl::access::fence_space::local_space);
+        if (contractId < offset) {
+          StorageIndex myNeigbourId = (Properties::LocalThreadSizeNC * offset);
+          *out_scratch_ptr += out_scratch_ptr[myNeigbourId];
+        }
+      }
+      // moving to next 16 by 16 block
+      out_scratch_ptr += (Properties::LocalThreadSizeNC * Properties::LocalThreadSizeC);
+    }
+
+    if (contractId == 0) {
+      out_scratch_ptr = local_output + nonContractId;
+      StorageIndex global_final_offset = nonContractGroupOffset + nonContractId;
+      out_ptr += global_final_offset;
+      EIGEN_UNROLL_LOOP
+      for (StorageIndex j = 0; j < Properties::WorkLoadPerThreadNC; j++) {
+        if (check_boundary<is_internal_block>(global_final_offset < nonContractDim)) {
+          auto res = *out_scratch_ptr;
+
+          *out_ptr = res;
+          out_ptr += Properties::LocalThreadSizeNC;
+        }
+        // moving to next 16 by 16 block to ge the next 16 reduced elements
+        out_scratch_ptr += (Properties::LocalThreadSizeNC * Properties::LocalThreadSizeC);
+        if (!(is_internal_block)) global_final_offset += Properties::LocalThreadSizeNC;
+      }
+    }
+  }
+
+  template <typename InputBlockProperties, bool is_internal_block, int CFactor, int GroupSize, typename Input,
+            typename Local>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void extract_block(const Input &inpt, Local *local_ptr,
+                                                                  const StorageIndex &linearLocalThreadId,
+                                                                  const StorageIndex &cOffset, const StorageIndex &C) {
+    local_ptr += InputBlockProperties::c_stride * linearLocalThreadId;
+    StorageIndex cIndex = cOffset;
+    for (StorageIndex cId = 0; cId < CFactor / InputBlockProperties::c_stride; cId++) {
+      if (check_boundary<is_internal_block>(cIndex + InputBlockProperties::c_stride - 1 < C)) {
+        auto val = read<InputBlockProperties::packet_load, InputBlockProperties::is_coalesced_layout,
+                        InputBlockProperties::is_rhs, typename InputBlockProperties::OutType>(inpt, StorageIndex(0),
+                                                                                              cIndex, StorageIndex(1));
+        write<StorageIndex, 1, data_source::local_mem>(val, local_ptr);
+      } else {
+        EIGEN_UNROLL_LOOP
+        for (StorageIndex i = 0; i < InputBlockProperties::elements_per_access; i++) {
+          OutScalar val =
+              (cIndex + i < C)
+                  ? read<false, InputBlockProperties::is_coalesced_layout, InputBlockProperties::is_rhs, OutScalar>(
+                        inpt, StorageIndex(0), cIndex + i, StorageIndex(1))
+                  : OutScalar(0);
+          write<StorageIndex, 1, data_source::local_mem>(val, local_ptr + i);
+        }
+      }
+      local_ptr += InputBlockProperties::c_stride * GroupSize;
+      cIndex += InputBlockProperties::c_stride * GroupSize;
+    }
+  }
+};
+#endif
+
+#ifndef EIGEN_SYCL_DISABLE_SCALAR
+
+/*!
+ * \brief GeneralScalarContraction is a template class that provides the scalar value of Tensor -Tensor contraction
+ * operation, when all the dimensions are contracting dimensions. This Kernel reduces two tensors to an scalar
+ *
+ * \tparam OutScalar: determines the output scalar type
+ *
+ * \tparam LhsScalar: determines the left-hand-side scalar type
+ *
+ * \tparam RhsScalar: determines the right-hand-side scalar type
+ *
+ * \tparam OutAccessor: determines the sycl accessor type for out put (please see the sycl-1.2.1 specification
+ * (https://www.khronos.org/registry/SYCL/specs/sycl-1.2.1.pdf) for accessor definition)
+ *
+ * \tparam LhsMapper: determines the tensor contraction mapper type for left-hand-side matrix
+ *
+ * \tparam RhsMapper: determines the tensor contraction mapper type for right-hand-side matrix
+ *
+ * \tparam StorageIndex: determines the StorageIndex Type
+ *
+ * \tparam Vectorizable: determines whether or not the vectorization is enabled for the Eigen expression.
+ *
+ * \param scratch: local memory containing tiles of LHS and RHS tensors for each work-group
+ *
+ * \param lhs: determines the left-hand-side flattened tensor (tensor mapper)
+ *
+ * \param rhs: determines the right-hand-side flattened tensor (tensor mapper)
+ *
+ * \param out_res: determines the output tensor containing the contraction result
+ *
+ * \param rng: determins the total input data size
+ */
+template <typename OutScalar, typename LhsScalar, typename RhsScalar, typename OutAccessor, typename LhsMapper,
+          typename RhsMapper, typename StorageIndex, bool Vectorizable>
+struct GeneralScalarContraction {
+  typedef cl::sycl::accessor<OutScalar, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local> Scratch;
+  Scratch scratch;
+  const LhsMapper lhs;
+  const RhsMapper rhs;
+  OutAccessor out_res;
+  const StorageIndex rng;
+
+  EIGEN_DEVICE_FUNC
+  GeneralScalarContraction(Scratch scratch_, const LhsMapper lhs_, const RhsMapper rhs_, OutAccessor out_res_,
+                           const StorageIndex rng_)
+      : scratch(scratch_), lhs(lhs_), rhs(rhs_), out_res(out_res_), rng(rng_) {}
+
+  EIGEN_DEVICE_FUNC void operator()(cl::sycl::nd_item<1> itemID) {
+    auto out_ptr = out_res.get_pointer();
+    auto scratch_ptr = scratch.get_pointer().get();
+
+    StorageIndex globalid = itemID.get_global_id(0);
+    StorageIndex localid = itemID.get_local_id(0);
+    OutScalar accumulator = OutScalar(0);
+    for (StorageIndex i = globalid; i < rng; i += itemID.get_global_range(0)) {
+      accumulator = cl::sycl::mad(lhs(0, i), rhs(i, 0), accumulator);
+    }
+    auto out_scratch_ptr = scratch_ptr + localid;
+    *out_scratch_ptr = accumulator;
+    for (StorageIndex offset = itemID.get_local_range(0) >> 1; offset > 0; offset >>= 1) {
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+      if (localid < offset) {
+        *out_scratch_ptr = (accumulator += out_scratch_ptr[offset]);
+      }
+    }
+    if (localid == 0) {
+      out_ptr[itemID.get_group(0)] = accumulator;
+    }
+  }
+};
+#endif
+
+}  // namespace internal
+}  // namespace TensorSycl
+
+template <typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>,
+                       Eigen::SyclDevice>
+    : public TensorContractionEvaluatorBase<TensorEvaluator<
+          const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Eigen::SyclDevice>> {
+  static_assert(std::is_same<OutputKernelType, const NoOpOutputKernel>::value,
+                "SYCL tensor contraction does not support output kernels.");
+
+  typedef Eigen::SyclDevice Device;
+
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
+  typedef TensorContractionEvaluatorBase<Self> Base;
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef typename XprType::Index StorageIndex;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef typename Base::Storage Storage;
+  typedef typename Base::EvaluatorPointerType EvaluatorPointerType;
+  struct TripleDim {
+    const StorageIndex M;
+    const StorageIndex N;
+    const StorageIndex K;
+    TripleDim(const StorageIndex M_, const StorageIndex N_, const StorageIndex K_) : M(M_), N(N_), K(K_) {}
+  };
+  enum {
+    Layout = TensorEvaluator<LeftArgType, Device>::Layout,
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess = false,
+  };
+
+  static EIGEN_CONSTEXPR int LDims = Base::LDims;
+  static EIGEN_CONSTEXPR int RDims = Base::RDims;
+  static EIGEN_CONSTEXPR int ContractDims = Base::ContractDims;
+
+  typedef array<StorageIndex, LDims> left_dim_mapper_t;
+  typedef array<StorageIndex, RDims> right_dim_mapper_t;
+
+  typedef array<StorageIndex, ContractDims> contract_t;
+  typedef array<StorageIndex, LDims - ContractDims> left_nocontract_t;
+  typedef array<StorageIndex, RDims - ContractDims> right_nocontract_t;
+
+  static const int NumDims = LDims + RDims - 2 * ContractDims;
+
+  typedef DSizes<StorageIndex, NumDims> Dimensions;
+
+  typedef TensorEvaluator<typename Base::EvalLeftArgType, Device> LeftEvaluator;
+  typedef TensorEvaluator<typename Base::EvalRightArgType, Device> RightEvaluator;
+  typedef typename Eigen::internal::remove_const<typename LeftEvaluator::CoeffReturnType>::type LhsScalar;
+  typedef typename Eigen::internal::remove_const<typename RightEvaluator::CoeffReturnType>::type RhsScalar;
+
+  typedef typename LeftEvaluator::Dimensions LeftDimensions;
+  typedef typename RightEvaluator::Dimensions RightDimensions;
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered>
+  struct input_mapper_propertis {
+    static EIGEN_CONSTEXPR bool is_lhs_matrix = (LDims == 2 && ContractDims == 1) || lhs_inner_dim_contiguous;
+    static EIGEN_CONSTEXPR bool is_rhs_matrix =
+        (RDims == 2 && ContractDims == 1) || (rhs_inner_dim_contiguous && !rhs_inner_dim_reordered);
+  };
+
+  TensorEvaluator(const XprType &op, const Device &device) : Base(op, device) {}
+
+  // We need to redefine this method to make nvcc happy
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(typename Base::EvaluatorPointerType data) {
+    this->m_leftImpl.evalSubExprsIfNeeded(NULL);
+    this->m_rightImpl.evalSubExprsIfNeeded(NULL);
+    if (!data) {
+      this->m_result = this->m_device.get(
+          static_cast<Scalar *>(this->m_device.allocate_temp(this->dimensions().TotalSize() * sizeof(Scalar))));
+      data = this->m_result;
+    }
+    evalToSycl(data);
+    return (this->m_result != NULL);
+  }
+  const Eigen::SyclDevice &device() const { return this->m_device; }
+  void evalToSycl(typename Base::EvaluatorPointerType buffer) const {
+    if (this->m_lhs_inner_dim_contiguous) {
+      if (this->m_rhs_inner_dim_contiguous) {
+        if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<true, true, true, Unaligned>(buffer);
+        } else {
+          evalTyped<true, true, false, Unaligned>(buffer);
+        }
+      } else {
+        if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<true, false, true, Unaligned>(buffer);
+        } else {
+          evalTyped<true, false, false, Unaligned>(buffer);
+        }
+      }
+    } else {
+      if (this->m_rhs_inner_dim_contiguous) {
+        if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<false, true, true, Unaligned>(buffer);
+        } else {
+          evalTyped<false, true, false, Unaligned>(buffer);
+        }
+      } else {
+        if (this->m_rhs_inner_dim_reordered) {
+          evalTyped<false, false, true, Unaligned>(buffer);
+        } else {
+          evalTyped<false, false, false, Unaligned>(buffer);
+        }
+      }
+    }
+  }
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
+  void evalTyped(typename Base::EvaluatorPointerType buffer) const {
+    const auto triple_dim = TripleDim{this->m_i_size, this->m_j_size, this->m_k_size};
+    typedef internal::TensorContractionInputMapper<
+        LhsScalar, StorageIndex, internal::Lhs, LeftEvaluator, left_nocontract_t, contract_t,
+        PacketType<CoeffReturnType, Device>::size, lhs_inner_dim_contiguous, false, Unaligned, MakeSYCLPointer>
+        LhsMapper;
+
+    typedef internal::TensorContractionInputMapper<RhsScalar, StorageIndex, internal::Rhs, RightEvaluator,
+                                                   right_nocontract_t, contract_t,
+                                                   PacketType<CoeffReturnType, Device>::size, rhs_inner_dim_contiguous,
+                                                   rhs_inner_dim_reordered, Unaligned, MakeSYCLPointer>
+        RhsMapper;
+
+    // initialize data mappers
+    LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides,
+                  this->m_left_contracting_strides, this->m_k_strides);
+
+    RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides,
+                  this->m_right_contracting_strides, this->m_k_strides);
+
+#ifndef EIGEN_SYCL_DISABLE_SCALAR
+    if (triple_dim.M == 1 && triple_dim.N == 1) {
+      launchSC(buffer, lhs, rhs, triple_dim.K);
+    } else
+#endif
+#ifndef EIGEN_SYCL_DISABLE_GEMV
+        if (triple_dim.M != 1 && triple_dim.N == 1) {
+      LaunchVT<false>(buffer, rhs, lhs, triple_dim.M, triple_dim.K);
+    } else if (triple_dim.M == 1 && triple_dim.N != 1) {
+      LaunchVT<true>(buffer, lhs, rhs, triple_dim.N, triple_dim.K);
+    } else  // This is equivalent of if (m!=1 && n!=1)
+#endif
+    {
+      typedef input_mapper_propertis<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered>
+          inpt_mapper_properties;
+#ifndef EIGEN_SYCL_DISABLE_SKINNY
+      bool skinny = false;
+      auto platform_name = this->device().getPlatformName();
+      // This is based on empirical calculation for AMD r9-nano and Fiji
+      if (platform_name.find("AMD") == 0) {
+        skinny = (triple_dim.M < triple_dim.K || triple_dim.N < triple_dim.K) &&
+                 ((triple_dim.M < 1024 && triple_dim.N < 1024) ||
+                  (uint64_t(triple_dim.M * triple_dim.N) < uint64_t(triple_dim.K)));
+      } else {
+        skinny = (((std::max(triple_dim.K, triple_dim.N) / std::min(triple_dim.K, triple_dim.N)) > 100) ||
+                  ((std::max(triple_dim.K, triple_dim.M) / std::min(triple_dim.K, triple_dim.M)) > 100) ||
+                  ((std::max(triple_dim.N, triple_dim.M) / std::min(triple_dim.N, triple_dim.M)) > 100));
+      }
+      if (skinny)
+        adjustTT<true, inpt_mapper_properties>(buffer, lhs, rhs, triple_dim);
+      else
+#endif  // EIGEN_SYCL_DISABLE_SKINNY
+        adjustTT<false, inpt_mapper_properties>(buffer, lhs, rhs, triple_dim);
+    }
+  }
+
+  template <bool skinny, typename input_mapper_properties, typename LhsMapper, typename RhsMapper>
+  void EIGEN_ALWAYS_INLINE adjustTT(EvaluatorPointerType buffer, const LhsMapper &lhs, const RhsMapper &rhs,
+                                    const TripleDim &triple_dim) const {
+#ifdef EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON
+    if (device().has_local_memory()) {
+      typedef TensorSycl::internal::TTPanelSize<CoeffReturnType, StorageIndex, 4, 4, 16> PanelParameters;
+      launchTT<TensorSycl::internal::contraction_type::local, skinny, input_mapper_properties, PanelParameters>(
+          buffer, lhs, rhs, triple_dim);
+    }
+#endif
+#ifdef EIGEN_SYCL_LOCAL_MEM_UNSET_OR_OFF
+    if (!(device().has_local_memory())) {
+      typedef TensorSycl::internal::TTPanelSize<CoeffReturnType, StorageIndex, 4, 4, 4> PanelParameters;
+      launchTT<TensorSycl::internal::contraction_type::no_local, skinny, input_mapper_properties, PanelParameters>(
+          buffer, lhs, rhs, triple_dim);
+    }
+#endif
+  }
+
+  template <TensorSycl::internal::contraction_type ct, bool skinny, typename input_mapper_properties,
+            typename Properties, typename LhsMapper, typename RhsMapper>
+  void launchTT(EvaluatorPointerType buffer, const LhsMapper &lhs, const RhsMapper &rhs,
+                const TripleDim &triple_dim) const {
+    const StorageIndex roundUpM = Eigen::TensorSycl::internal::roundUp(triple_dim.M, Properties::TileSizeDimM);
+    const StorageIndex roundUpN = Eigen::TensorSycl::internal::roundUp(triple_dim.N, Properties::TileSizeDimN);
+    const StorageIndex groupSizeM = roundUpM / Properties::TileSizeDimM;
+    const StorageIndex groupSizeN = roundUpN / Properties::TileSizeDimN;
+
+    const StorageIndex roundUpK = Eigen::TensorSycl::internal::roundUp(triple_dim.K, Properties::TileSizeDimK);
+    StorageIndex totalTilesK = roundUpK / Properties::TileSizeDimK;
+    StorageIndex groupSizeK =
+        skinny
+            ? std::max(std::min(totalTilesK,
+                                (StorageIndex)(device().getPowerOfTwo(device().getNumSyclMultiProcessors(), true) * 4) /
+                                    (groupSizeM * groupSizeN)),
+                       StorageIndex(1))
+            : StorageIndex(1);
+
+    const StorageIndex numTilesPerGroup = Eigen::TensorSycl::internal::roundUp(totalTilesK, groupSizeK) / groupSizeK;
+
+    const StorageIndex totalGroupSize = groupSizeM * groupSizeN * groupSizeK;
+
+    const StorageIndex localRange = Properties::LocalThreadSizeM * Properties::LocalThreadSizeN;
+    const StorageIndex globalRange = totalGroupSize * localRange;
+
+    const StorageIndex scratchSize = (ct == TensorSycl::internal::contraction_type::local)
+                                         ? ((Properties::DoubleBuffer + 1) *
+                                            (Properties::TileSizeDimM + Properties::BC) * (Properties::TileSizeDimK)) +
+                                               ((Properties::DoubleBuffer + 1) * (Properties::TileSizeDimK) *
+                                                (Properties::TileSizeDimN + Properties::BC))
+                                         : StorageIndex(1);
+
+    auto thread_range = cl::sycl::nd_range<1>(cl::sycl::range<1>(globalRange), cl::sycl::range<1>(localRange));
+    if (groupSizeK == 1) {
+      typedef TensorSycl::internal::TensorContractionKernel<CoeffReturnType, LhsScalar, RhsScalar, EvaluatorPointerType,
+                                                            LhsMapper, RhsMapper, StorageIndex, Properties, TripleDim,
+                                                            PacketAccess, input_mapper_properties, true, ct>
+          ContractKernelName;
+      device().template binary_kernel_launcher<CoeffReturnType, ContractKernelName>(
+          lhs, rhs, buffer, thread_range, scratchSize, groupSizeM, groupSizeN, numTilesPerGroup, triple_dim);
+    } else {
+      typedef TensorSycl::internal::TensorContractionKernel<CoeffReturnType, LhsScalar, RhsScalar, EvaluatorPointerType,
+                                                            LhsMapper, RhsMapper, StorageIndex, Properties, TripleDim,
+                                                            PacketAccess, input_mapper_properties, false, ct>
+          ContractKernelName;
+      CoeffReturnType *temp_pointer = static_cast<CoeffReturnType *>(
+          device().allocate_temp(triple_dim.M * triple_dim.N * groupSizeK * sizeof(CoeffReturnType)));
+      EvaluatorPointerType tmp_global_accessor = device().get(temp_pointer);
+
+      device().template binary_kernel_launcher<CoeffReturnType, ContractKernelName>(
+          lhs, rhs, tmp_global_accessor, thread_range, scratchSize, groupSizeM, groupSizeN, numTilesPerGroup,
+          triple_dim);
+
+      typedef Eigen::internal::SumReducer<CoeffReturnType> Op;
+      auto op = Op();
+      typedef TensorSycl::internal::SecondStepPartialReduction<CoeffReturnType, StorageIndex, EvaluatorPointerType,
+                                                               EvaluatorPointerType, Op>
+          ReductionKernel;
+
+      device().template unary_kernel_launcher<CoeffReturnType, ReductionKernel>(
+          tmp_global_accessor, buffer,
+          cl::sycl::nd_range<1>(cl::sycl::range<1>(StorageIndex(
+                                    Eigen::TensorSycl::internal::roundUp(triple_dim.M * triple_dim.N, localRange))),
+                                cl::sycl::range<1>(localRange)),
+          StorageIndex(1), op, StorageIndex(triple_dim.M * triple_dim.N), groupSizeK);
+
+      device().deallocate_temp(temp_pointer);
+    }
+  }
+
+#ifndef EIGEN_SYCL_DISABLE_GEMV
+  template <bool is_lhs_vec, typename VectorMapper, typename TensorMapper, typename StorageIndex>
+  void EIGEN_ALWAYS_INLINE LaunchVT(EvaluatorPointerType buffer, const VectorMapper &vec, const TensorMapper &mat,
+                                    StorageIndex NC, StorageIndex C) const {
+    const StorageIndex nonContractDim = NC;
+    EIGEN_CONSTEXPR StorageIndex NCFactor = 1;
+    EIGEN_CONSTEXPR StorageIndex CFactor = 1;
+    EIGEN_CONSTEXPR StorageIndex NCWindow = 16;
+    typedef Eigen::TensorSycl::internal::TVPanelSize<CoeffReturnType, StorageIndex, NCWindow, CFactor, NCFactor>
+        Properties;
+    const StorageIndex roundUpC = Eigen::TensorSycl::internal::roundUp(C, Properties::TileSizeDimC);
+    const StorageIndex cNumGroups = roundUpC / (Properties::LocalThreadSizeC * Properties::WorkLoadPerThreadC);
+    const StorageIndex roundUpNC = Eigen::TensorSycl::internal::roundUp(nonContractDim, Properties::TileSizeDimNC);
+    const StorageIndex nCNumGroups = roundUpNC / (Properties::LocalThreadSizeNC * Properties::WorkLoadPerThreadNC);
+    const StorageIndex globalRange =
+        (roundUpNC / (Properties::WorkLoadPerThreadNC)) * (roundUpC / (Properties::WorkLoadPerThreadC));
+    const StorageIndex localRange = Properties::LocalThreadSizeNC * Properties::LocalThreadSizeC;
+    const StorageIndex scratchSize =
+        (Properties::WorkLoadPerThreadNC + CFactor) * Properties::LocalThreadSizeC * Properties::LocalThreadSizeNC;
+    auto thread_range = cl::sycl::nd_range<1>(cl::sycl::range<1>(globalRange), cl::sycl::range<1>(localRange));
+    if (cNumGroups > 1) {
+      typedef Eigen::TensorSycl::internal::GeneralVectorTensor<CoeffReturnType, EvaluatorPointerType, VectorMapper,
+                                                               TensorMapper, StorageIndex, Properties, CFactor, false,
+                                                               is_lhs_vec, false>
+          ContractKernelName;
+      CoeffReturnType *temp_pointer =
+          static_cast<CoeffReturnType *>(device().allocate_temp(nonContractDim * cNumGroups * sizeof(CoeffReturnType)));
+      EvaluatorPointerType tmp_global_accessor = device().get(temp_pointer);
+
+      device().template binary_kernel_launcher<CoeffReturnType, ContractKernelName>(
+          vec, mat, tmp_global_accessor, thread_range, scratchSize, nCNumGroups, nonContractDim, C);
+
+      typedef Eigen::internal::SumReducer<CoeffReturnType> Op;
+      typedef TensorSycl::internal::SecondStepPartialReduction<CoeffReturnType, StorageIndex, EvaluatorPointerType,
+                                                               EvaluatorPointerType, Op>
+          ReductionKernel;
+
+      device().template unary_kernel_launcher<CoeffReturnType, ReductionKernel>(
+          tmp_global_accessor, buffer,
+          cl::sycl::nd_range<1>(cl::sycl::range<1>(Eigen::TensorSycl::internal::roundUp(nonContractDim, localRange)),
+                                cl::sycl::range<1>(localRange)),
+          StorageIndex(1), Op(), nonContractDim, cNumGroups);
+
+      device().deallocate_temp(temp_pointer);
+    } else {
+      typedef Eigen::TensorSycl::internal::GeneralVectorTensor<CoeffReturnType, EvaluatorPointerType, VectorMapper,
+                                                               TensorMapper, StorageIndex, Properties, CFactor, false,
+                                                               is_lhs_vec, true>
+          ContractKernelName;
+      device().template binary_kernel_launcher<CoeffReturnType, ContractKernelName>(
+          vec, mat, buffer, thread_range, scratchSize, nCNumGroups, nonContractDim, C);
+    }
+  }
+#endif
+
+#ifndef EIGEN_SYCL_DISABLE_SCALAR
+  template <typename LhsMapper, typename RhsMapper>
+  EIGEN_ALWAYS_INLINE void launchSC(EvaluatorPointerType buffer, const LhsMapper &lhs, const RhsMapper &rhs,
+                                    StorageIndex K) const {
+    EIGEN_STATIC_ASSERT(!((EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1) &
+                          (EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1 - 1)),
+                        "The Local thread size must be a power of 2 for the reduction "
+                        "operation");
+    EIGEN_CONSTEXPR StorageIndex local_range = EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1;
+
+    // Here we force the code not to be more than 2-step reduction: Our empirical research shows that if each thread
+    // reduces at least 512 elementss individually, we get better performance.
+    const StorageIndex num_work_group = ((K + (512 * local_range - 1)) / (512 * local_range) > 1 ? local_range : 1);
+    const StorageIndex global_range = num_work_group * local_range;
+
+    typedef Eigen::TensorSycl::internal::GeneralScalarContraction<
+        CoeffReturnType, LhsScalar, RhsScalar, EvaluatorPointerType, LhsMapper, RhsMapper, StorageIndex, false>
+        ContractKernelName;
+    auto thread_range = cl::sycl::nd_range<1>(cl::sycl::range<1>(global_range), cl::sycl::range<1>(local_range));
+    if (num_work_group > 1) {
+      CoeffReturnType *temp_pointer =
+          static_cast<CoeffReturnType *>(device().allocate_temp(num_work_group * sizeof(CoeffReturnType)));
+      EvaluatorPointerType tmp_global_accessor = device().get(temp_pointer);
+      device().template binary_kernel_launcher<CoeffReturnType, ContractKernelName>(lhs, rhs, tmp_global_accessor,
+                                                                                    thread_range, local_range, K);
+      typedef Eigen::internal::SumReducer<CoeffReturnType> Op;
+      typedef TensorSycl::internal::SecondStepFullReducer<CoeffReturnType, Op, EvaluatorPointerType,
+                                                          EvaluatorPointerType, StorageIndex, local_range>
+          GenericRKernel;
+      device().template unary_kernel_launcher<CoeffReturnType, GenericRKernel>(
+          tmp_global_accessor, buffer,
+          cl::sycl::nd_range<1>(cl::sycl::range<1>(local_range), cl::sycl::range<1>(local_range)), local_range, Op());
+
+      device().deallocate_temp(temp_pointer);
+    } else {
+      device().template binary_kernel_launcher<CoeffReturnType, ContractKernelName>(lhs, rhs, buffer, thread_range,
+                                                                                    local_range, K);
+    }
+  }
+#endif
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    this->m_leftImpl.cleanup();
+    this->m_rightImpl.cleanup();
+
+    if (this->m_result) {
+      this->m_device.deallocate_temp(this->m_result);
+      this->m_result = NULL;
+    }
+  }
+  // The placeholder accessors must bound to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    this->m_leftImpl.bind(cgh);
+    this->m_rightImpl.bind(cgh);
+    this->m_result.bind(cgh);
+  }
+};
+}  // namespace Eigen
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_SYCL_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionThreadPool.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionThreadPool.h
new file mode 100644
index 0000000..21be6ea
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -0,0 +1,1679 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_THREAD_POOL_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_THREAD_POOL_H
+
+// evaluator for thread pool device
+#ifdef EIGEN_USE_THREADS
+
+namespace Eigen {
+
+template<typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType>
+struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, ThreadPoolDevice> :
+    public TensorContractionEvaluatorBase<TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, ThreadPoolDevice> > {
+
+  typedef ThreadPoolDevice Device;
+
+  typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
+  typedef TensorContractionEvaluatorBase<Self> Base;
+
+  typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+
+  enum {
+    Layout = TensorEvaluator<LeftArgType, Device>::Layout,
+  };
+
+  // Most of the code is assuming that both input tensors are ColMajor. If the
+  // inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
+  // If we want to compute A * B = C, where A is LHS and B is RHS, the code
+  // will pretend B is LHS and A is RHS.
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType;
+  typedef typename internal::conditional<
+    static_cast<int>(Layout) == static_cast<int>(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType;
+
+  static const int LDims =
+      internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
+  static const int RDims =
+      internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
+  static const int ContractDims = internal::array_size<Indices>::value;
+
+  typedef array<Index, LDims> left_dim_mapper_t;
+  typedef array<Index, RDims> right_dim_mapper_t;
+
+  typedef array<Index, ContractDims> contract_t;
+  typedef array<Index, LDims - ContractDims> left_nocontract_t;
+  typedef array<Index, RDims - ContractDims> right_nocontract_t;
+
+  static const int NumDims = LDims + RDims - 2 * ContractDims;
+
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  // typedefs needed in evalTo
+  typedef typename internal::remove_const<typename EvalLeftArgType::Scalar>::type LhsScalar;
+  typedef typename internal::remove_const<typename EvalRightArgType::Scalar>::type RhsScalar;
+  typedef typename internal::gebp_traits<LhsScalar, RhsScalar> Traits;
+
+  typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
+  typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
+
+  TensorEvaluator(const XprType& op, const Device& device) :
+      Base(op, device) {}
+
+  template <int Alignment>
+  void evalProduct(Scalar* buffer) const {
+    evalProductImpl<NoCallback, Alignment>(buffer, NoCallback());
+  }
+
+  template <typename EvalToCallback, int Alignment>
+  void evalProductAsync(Scalar* buffer, EvalToCallback done) const {
+    evalProductImpl<EvalToCallback, Alignment>(buffer, std::move(done));
+  }
+
+  template <typename DoneCallback, int Alignment>
+  void evalProductImpl(Scalar* buffer, DoneCallback done) const {
+    // This function computes a lot of heuristics in multiple steps, and it
+    // also has multiple exit points. To keep it sane, readable and all in one
+    // place, sync/async execution decision is made at runtime at the very end.
+    //
+    // (1) In sync mode we allocate Context on the stack, submit computations
+    //     to the device thread pool, and block on a barrier until it is
+    //     completed.
+    //
+    // (2) In async mode we allocate Context on the heap, and after all tasks
+    //     are finished, we call provided the done callback, and delete a
+    //     context from the heap.
+    //
+    // (*) EvalParallelContext & EvalShardedByInnerDimContext owns all the state
+    // and temporary buffers, requried for executing the tensor contraction.
+    // They are responsible for cleaning it up after contraction is done.
+    static const bool IsEvalInSyncMode =
+        std::is_same<DoneCallback, NoCallback>::value;
+
+    const Index m = this->m_i_size;
+    const Index n = this->m_j_size;
+    const Index k = this->m_k_size;
+    if (m == 0 || n == 0 || k == 0) return;
+
+    // Compute a set of algorithm parameters:
+    // - kernel block sizes (bm, bn, bk)
+    // - task grain sizes (number of kernels executed per task: gm, gn)
+    // - number of threads
+    // - sharding by row/column
+    // - parallel packing or first lhs then rhs
+    // and some derived parameters:
+    // - number of tasks (nm, nn, nk)
+    // - number of kernels (nm0, nn0)
+    // Unfortunately, all these parameters are tightly interdependent.
+    // So in some cases we first compute approximate values, then compute other
+    // values based on these approximations and then refine the approximations.
+
+    // There are lots of heuristics here. There is some reasoning behind them,
+    // but ultimately they are just tuned on contraction benchmarks for
+    // different input configurations, thread counts and instruction sets.
+    // So feel free to question any of them.
+
+    // Compute whether we want to shard by row or by column.
+    // This is a first approximation, it will be refined later. Since we don't
+    // know number of threads yet we use 2, because what's we are most
+    // interested in at this point is whether it makes sense to use
+    // parallelization at all or not.
+    bool shard_by_col = shardByCol(m, n, 2);
+
+    // First approximation of kernel blocking sizes.
+    // Again, we don't know number of threads yet, so we use 2.
+    Index bm, bn, bk;
+    if (shard_by_col) {
+      internal::TensorContractionBlocking<Scalar, LhsScalar, RhsScalar, Index,
+                                          internal::ShardByCol>
+          blocking(k, m, n, 2);
+      bm = blocking.mc();
+      bn = blocking.nc();
+      bk = blocking.kc();
+    } else {
+      internal::TensorContractionBlocking<Scalar, LhsScalar, RhsScalar, Index,
+                                          internal::ShardByRow>
+          blocking(k, m, n, 2);
+      bm = blocking.mc();
+      bn = blocking.nc();
+      bk = blocking.kc();
+    }
+
+    // Compute optimal number of threads.
+    // Note: we use bk instead of k here because we are interested in amount of
+    // _parallelizable_ computations, and computations are not parallelizable
+    // across k dimension.
+    const TensorOpCost cost =
+        contractionCost(m, n, bm, bn, bk, shard_by_col, false);
+    int num_threads = TensorCostModel<ThreadPoolDevice>::numThreads(
+        static_cast<double>(n) * m, cost, this->m_device.numThreads());
+    int num_threads_by_k = numThreadsInnerDim(m, n, k);
+    if (shardByInnerDim(m, n, k, num_threads, num_threads_by_k)) {
+      // We are in the scenario where it is more effective to shard by the
+      // inner dimension.
+      if (IsEvalInSyncMode) {
+        EvalShardedByInnerDimContext<DoneCallback> ctx(
+            this, num_threads_by_k, buffer, m, n, k, std::move(done));
+        ctx.template run<Alignment>();
+      } else {
+        auto* ctx = new EvalShardedByInnerDimContext<DoneCallback>(
+            this, num_threads_by_k, buffer, m, n, k, std::move(done));
+        ctx->template runAsync<Alignment>();
+      }
+
+      return;
+    }
+
+    // TODO(dvyukov): this is a stop-gap to prevent regressions while the cost
+    // model is not tuned. Remove this when the cost model is tuned.
+    if (n == 1) num_threads = 1;
+
+    if (num_threads == 1) {
+      TENSOR_CONTRACTION_DISPATCH(this->template evalProductSequential,
+                                  Unaligned, (buffer));
+      if (!IsEvalInSyncMode) done();
+      return;
+    }
+
+    // Now that we know number of threads, recalculate sharding and blocking.
+    shard_by_col = shardByCol(m, n, num_threads);
+    if (shard_by_col) {
+      internal::TensorContractionBlocking<Scalar, LhsScalar, RhsScalar, Index,
+                                          internal::ShardByCol>
+          blocking(k, m, n, num_threads);
+      bm = blocking.mc();
+      bn = blocking.nc();
+      bk = blocking.kc();
+    } else {
+      internal::TensorContractionBlocking<Scalar, LhsScalar, RhsScalar, Index,
+                                          internal::ShardByRow>
+          blocking(k, m, n, num_threads);
+      bm = blocking.mc();
+      bn = blocking.nc();
+      bk = blocking.kc();
+    }
+
+    // Number of kernels for each dimension.
+    Index nm0 = divup(m, bm);
+    Index nn0 = divup(n, bn);
+    Index nk = divup(k, bk);
+
+    // Calculate task grain size (number of kernels executed per task).
+    // This task size coarsening serves two purposes:
+    // 1. It reduces per-task overheads including synchronization overheads.
+    // 2. It allows to use caches better (reuse the same packed rhs in several
+    // consecutive kernels).
+    Index gm = 1;
+    Index gn = 1;
+    // If we are sharding by column, then we prefer to reduce rows first.
+    if (shard_by_col) {
+      gm = coarsenM(m, n, bm, bn, bk, gn, num_threads, shard_by_col);
+      gn = coarsenN(m, n, bm, bn, bk, gm, num_threads, shard_by_col);
+    } else {
+      gn = coarsenN(m, n, bm, bn, bk, gm, num_threads, shard_by_col);
+      gm = coarsenM(m, n, bm, bn, bk, gn, num_threads, shard_by_col);
+    }
+    // Number of tasks in each dimension.
+    Index nm = divup(nm0, gm);
+    Index nn = divup(nn0, gn);
+
+    // If there is enough concurrency in the sharding dimension, we choose not
+    // to paralellize by the other dimension, and execute all kernels in sync
+    // mode. This reduces parallelism from the nm x nn down to nn
+    // (shard_by_col==true) or nm (shard_by_col==false).
+    const Index sharding_dim_tasks = shard_by_col ? nn : nm;
+    const int num_worker_threads = this->m_device.numThreadsInPool();
+
+    // With small number of threads we want to make sure that we do not reduce
+    // parallelism too much. With large number of threads we trade maximum
+    // parallelism for better memory locality.
+    const float oversharding_factor =
+        num_worker_threads <= 4  ? 8.0 :
+        num_worker_threads <= 8  ? 4.0 :
+        num_worker_threads <= 16 ? 2.0 :
+        num_worker_threads <= 32 ? 1.0 :
+        num_worker_threads <= 64 ? 0.8 : /* num_worker_threads > 64 */ 0.6;
+
+    const bool parallelize_by_sharding_dim_only =
+        sharding_dim_tasks >= oversharding_factor * num_worker_threads;
+
+    // Last by not least, decide whether we want to issue both lhs and rhs
+    // packing in parallel; or issue lhs packing first, and then issue rhs
+    // packing when lhs packing completes (for !shard_by_col lhs and rhs are
+    // swapped). Parallel packing allows more parallelism (for both packing and
+    // kernels), while sequential packing provides better locality (once
+    // a thread finishes rhs packing it proceed to kernels with that rhs).
+    // First, we are interested in parallel packing if there are few tasks.
+    bool parallel_pack = num_threads >= nm * nn;
+    // Also do parallel packing if all data fits into L2$.
+    if (m * bk * Index(sizeof(LhsScalar)) + n * bk * Index(sizeof(RhsScalar)) <=
+        l2CacheSize() * num_threads)
+      parallel_pack = true;
+    // But don't do it if we will use each rhs only once. Locality seems to be
+    // more important in this case.
+    if ((shard_by_col ? nm : nn) == 1) parallel_pack = false;
+    // Also don't get in the way of parallelize_by_sharding_dim_only
+    // optimization.
+    if (parallelize_by_sharding_dim_only) parallel_pack = false;
+
+    // TODO(ezhulnev): With if contexpr we don't need SyncEvalParallelContext.
+    if (IsEvalInSyncMode) {
+#define CONTEXT_ARGS                                                        \
+  (this, num_threads, buffer, m, n, k, bm, bn, bk, nm, nn, nk, gm, gn, nm0, \
+   nn0, shard_by_col, parallel_pack, parallelize_by_sharding_dim_only,      \
+   NoCallback())                                                            \
+      .run()
+      TENSOR_CONTRACTION_DISPATCH(SyncEvalParallelContext, Alignment,
+                                  CONTEXT_ARGS);
+#undef CONTEXT_ARGS
+
+    } else {
+#define CONTEXT_ARGS                                                        \
+  (this, num_threads, buffer, m, n, k, bm, bn, bk, nm, nn, nk, gm, gn, nm0, \
+   nn0, shard_by_col, parallel_pack, parallelize_by_sharding_dim_only,      \
+   std::move(done))
+      TENSOR_CONTRACTION_ASYNC_DISPATCH(EvalParallelContext, DoneCallback,
+                                        Alignment, CONTEXT_ARGS, run());
+#undef CONTEXT_ARGS
+    }
+  }
+
+  // ------------------------------------------------------------------------ //
+
+  // Dummy struct to represent an empty DoneCallback.
+
+  struct NoCallback {
+    void operator()() {
+      eigen_assert(false && "NoCallback should never be called");
+    }
+  };
+
+  // ------------------------------------------------------------------------ //
+
+  template <typename DoneCallback, typename Context>
+  class EvalParallelNotification;
+
+  // Synchronous evaluation notification that blocks caller thread in Wait().
+  template <typename Context>
+  class EvalParallelNotification<NoCallback, Context> {
+   public:
+    EvalParallelNotification(Context*, NoCallback) {}
+    void Notify() { done_.Notify(); }
+    void Wait() { done_.Wait(); }
+   private:
+    Eigen::Notification done_;
+  };
+
+  // Asynchronous evaluation notification that does not block in Wait().
+  template <typename DoneCallback, typename Context>
+  class EvalParallelNotification {
+   public:
+    EvalParallelNotification(Context* ctx, DoneCallback done)
+        : ctx_(ctx), done_(std::move(done)) {}
+
+    void Notify() {
+      // Make a copy of done callback, because it will be destructed when we
+      // will delete context in the next line (EvalParallelNotification is a
+      // data member of EvalParallelContext class).
+      DoneCallback done_copy = std::move(done_);
+
+      // Delete parallel evaluation context.
+      delete ctx_;
+
+      // Now safely call the done callback.
+      done_copy();
+    }
+
+    void Wait() {}
+
+   private:
+    Context* ctx_;
+    DoneCallback done_;
+  };
+
+  // Context orchestrates sync/async parallel contraction evaluation. When it is
+  // executed in asynchronous mode, it owns all the shared state that might be
+  // accessible by block packing and kernel tasks.
+
+  template <typename DoneCallback, bool lhs_inner_dim_contiguous,
+            bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered,
+            int Alignment>
+  class EvalParallelContext {
+   public:
+    typedef internal::TensorContractionInputMapper<
+        LhsScalar, Index, internal::Lhs, LeftEvaluator, left_nocontract_t,
+        contract_t, internal::packet_traits<LhsScalar>::size,
+        lhs_inner_dim_contiguous, false, Unaligned>
+        LhsMapper;
+    typedef internal::TensorContractionInputMapper<
+        RhsScalar, Index, internal::Rhs, RightEvaluator, right_nocontract_t,
+        contract_t, internal::packet_traits<RhsScalar>::size,
+        rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Unaligned>
+        RhsMapper;
+
+    typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
+
+    typedef internal::TensorContractionKernel<
+        Scalar, LhsScalar, RhsScalar, Index, OutputMapper, LhsMapper, RhsMapper>
+        TensorContractionKernel;
+
+    typedef typename TensorContractionKernel::LhsBlock LhsBlock;
+    typedef typename TensorContractionKernel::RhsBlock RhsBlock;
+    typedef typename TensorContractionKernel::BlockMemHandle BlockMemHandle;
+
+    EvalParallelContext(const Self* self, int num_threads, Scalar* buffer,
+                        Index tm, Index tn, Index tk, Index bm, Index bn,
+                        Index bk, Index nm, Index nn, Index nk, Index gm,
+                        Index gn, Index nm0, Index nn0, bool shard_by_col,
+                        bool parallel_pack,
+                        bool parallelize_by_sharding_dim_only,
+                        DoneCallback done)
+        : created_by_thread_id_(std::this_thread::get_id()),
+          done_(this, std::move(done)),
+          device_(self->m_device),
+          lhs_(self->m_leftImpl, self->m_left_nocontract_strides,
+               self->m_i_strides, self->m_left_contracting_strides,
+               self->m_k_strides),
+          rhs_(self->m_rightImpl, self->m_right_nocontract_strides,
+               self->m_j_strides, self->m_right_contracting_strides,
+               self->m_k_strides),
+          buffer_(buffer),
+          output_(buffer, tm),
+          output_kernel_(self->m_output_kernel),
+          tensor_contraction_params_(self->m_tensor_contraction_params),
+          num_threads_(num_threads),
+          shard_by_col_(shard_by_col),
+          parallel_pack_(parallel_pack),
+          parallelize_by_sharding_dim_only_(parallelize_by_sharding_dim_only),
+          m_(tm),
+          n_(tn),
+          k_(tk),
+          bm_(bm),
+          bn_(bn),
+          bk_(bk),
+          nm_(nm),
+          nn_(nn),
+          nk_(nk),
+          gm_(gm),
+          gn_(gn),
+          nm0_(nm0),
+          nn0_(nn0),
+          kernel_(m_, k_, n_, bm_, bk_, bn_),
+          num_thread_local_allocations_(0),
+          // We reserve 2X more capacity for a thread local values, than the
+          // number of threads in the pool to efficiently handle task stealing
+          // by threads that are not managed by the pool.
+          thread_local_capacity(2 * (parallelize_by_sharding_dim_only_
+                                         ? device_.numThreadsInPool()
+                                         : 0)),
+          // We will use only one of the Lhs/Rhs thread local storage depending
+          // on the shard_by_col value and we parallelize by sharding dim ONLY.
+          lhs_thread_local_blocks_(shard_by_col_ ? 0 : thread_local_capacity,
+                                   {*this}, {*this}),
+          rhs_thread_local_blocks_(shard_by_col_ ? thread_local_capacity : 0,
+                                   {*this}, {*this}) {
+      // These two options are mutually exclusive.
+      eigen_assert(!(parallel_pack && parallelize_by_sharding_dim_only));
+
+      for (Index x = 0; x < P; x++) {
+        // Normal number of notifications for k slice switch is
+        // nm_ + nn_ + nm_ * nn_. However, first P - 1 slices will receive only
+        // nm_ + nn_ notifications, because they will not receive notifications
+        // from preceding kernels.
+        state_switch_[x] =
+            x == 0
+                ? 1
+                : (parallel_pack_ ? nn_ + nm_ : (shard_by_col_ ? nn_ : nm_)) +
+                      (x == P - 1 ? nm_ * nn_ : 0);
+        state_packing_ready_[x] =
+            parallel_pack_ ? 0 : (shard_by_col_ ? nm_ : nn_);
+        state_kernel_[x] = new std::atomic<uint8_t>*[nm_];
+        for (Index m = 0; m < nm_; m++) {
+          state_kernel_[x][m] = new std::atomic<uint8_t>[nn_];
+          // Kernels generally receive 3 notifications (previous kernel + 2
+          // packing), but the first slice won't get notifications from previous
+          // kernels.
+          for (Index n = 0; n < nn_; n++)
+            state_kernel_[x][m][n].store(
+                (x == 0 ? 0 : 1) + (parallel_pack_ ? 2 : 1),
+                std::memory_order_relaxed);
+        }
+      }
+
+      // Allocate memory for packed rhs/lhs matrices.
+      packed_mem_ = kernel_.allocateSlices(            //
+          device_,                                     //
+          /*num_lhs=*/nm0_,                            //
+          /*num_rhs=*/nn0_,                            //
+          /*num_slices=*/std::min<Index>(nk_, P - 1),  //
+          packed_lhs_, packed_rhs_);
+
+      if (parallelize_by_sharding_dim_only_) {
+        const int num_worker_threads = device_.numThreadsInPool();
+
+        if (shard_by_col) {
+          can_use_thread_local_packed_ = new std::atomic<bool>[nn_];
+          for (int i = 0; i < nn_; ++i)
+            can_use_thread_local_packed_[i].store(true,
+                                                  std::memory_order_relaxed);
+
+          Index num_blocks = num_worker_threads * gn_;
+          thread_local_pre_alocated_mem_ = kernel_.allocateSlices(  //
+              device_,                                              //
+              /*num_lhs=*/0,                                        //
+              /*num_rhs=*/num_blocks,                               //
+              /*num_slices=*/1,                                     //
+              /*lhs_blocks=*/nullptr, &rhs_thread_local_pre_allocated_);
+
+        } else {
+          can_use_thread_local_packed_ = new std::atomic<bool>[nm_];
+          for (int i = 0; i < nm_; ++i)
+            can_use_thread_local_packed_[i].store(true,
+                                                  std::memory_order_relaxed);
+
+          Index num_blocks = num_worker_threads * gm_;
+          thread_local_pre_alocated_mem_ = kernel_.allocateSlices(  //
+              device_,                                              //
+              /*num_lhs=*/num_blocks,                               //
+              /*num_rhs=*/0,                                        //
+              /*num_slices=*/1, &lhs_thread_local_pre_allocated_,   //
+              /*rhs_blocks=*/nullptr);
+        }
+      }
+    }
+
+    ~EvalParallelContext() {
+      for (Index x = 0; x < P; x++) {
+        for (Index m = 0; m < nm_; m++) delete[] state_kernel_[x][m];
+        delete[] state_kernel_[x];
+      }
+      kernel_.deallocate(device_, packed_mem_);
+      if (parallelize_by_sharding_dim_only_) {
+        kernel_.deallocate(device_, thread_local_pre_alocated_mem_);
+        delete[] can_use_thread_local_packed_;
+      }
+    }
+
+    void run() {
+      // Kick off packing of the first slice.
+      signal_switch(0, 1);
+
+      // Wait for overall completion.
+      //
+      // If parallel evaluation is executed in async mode, this is a no-op, and
+      // Wait() will return immediately. In synchronous mode it will block the
+      // caller thread until it will receive notification from last task.
+      //
+      // In async mode, last task when completed will call done callback from
+      // the same thread, and will delete this context.
+      //
+      // TODO(dvyukov): This wait can lead to deadlock if contraction is
+      // evaluated in synchronous mode. If nthreads contractions are
+      // concurrently submitted from worker threads, this wait will block all
+      // worker threads and the system will deadlock.
+      done_.Wait();
+    }
+
+   private:
+    std::thread::id created_by_thread_id_;
+
+    // This notification is specialized on the type of DoneCallback and can be
+    // blocking or non-blocking.
+    EvalParallelNotification<DoneCallback, EvalParallelContext> done_;
+
+    const Device& device_;
+    LhsMapper lhs_;
+    RhsMapper rhs_;
+    Scalar* const buffer_;
+    OutputMapper output_;
+    OutputKernelType output_kernel_;
+    TensorContractionParams tensor_contraction_params_;
+    const int num_threads_;
+    const bool shard_by_col_;
+    const bool parallel_pack_;
+    const bool parallelize_by_sharding_dim_only_;
+    // Matrix sizes.
+    const Index m_;
+    const Index n_;
+    const Index k_;
+    // Block sizes.
+    const Index bm_;
+    const Index bn_;
+    const Index bk_;
+    // Number of tasks.
+    const Index nm_;
+    const Index nn_;
+    const Index nk_;
+    // Task grain sizes (number of kernels executed per task).
+    const Index gm_;
+    const Index gn_;
+    // Number of blocks (this is different from ni_/nn_ because of task size
+    // coarsening).
+    const Index nm0_;
+    const Index nn0_;
+    // Tensor contraction kernel.
+    TensorContractionKernel kernel_;
+
+    // Parallelization strategy.
+    //
+    // Blocks related to the same k block can run in parallel because they write
+    // to different output blocks. So we parallelize within k slices, this
+    // gives us parallelism level of m x n. Before we can start any kernels
+    // related to k-th slice, we need to issue m lhs packing tasks and n rhs
+    // packing tasks.
+    //
+    // However, there is a bottleneck when we are finishing kernels for k-th
+    // slice (at the very end there is only 1 runnable kernel). To mitigate this
+    // bottleneck we allow kernels from k-th and k+1-th slices to run in
+    // parallel. Note that (m, n, k) and (m, n, k+1) kernels write to the same
+    // output block, so they must not run in parallel.
+    //
+    // This gives us the following dependency graph.
+    // On each k slice we have m x n kernel tasks, m lhs paking tasks and n rhs
+    // packing tasks.
+    // Kernel (m, n, k) can start when:
+    //  - kernel (m, n, k-1) has finished
+    //  - lhs packing (m, k) has finished
+    //  - rhs packing (n, k) has finished
+    // Lhs/rhs packing can start when:
+    //  - all k-1 packing has finished (artificially imposed to limit amount of
+    //  parallel packing)
+    //
+    // On top of that we limit runnable tasks to two consecutive k slices.
+    // This is done to limit amount of memory we need for packed lhs/rhs
+    // (for each k slice we need m*bk + n*bk memory in packed_lhs_/packed_rhs_).
+    //
+    // state_switch_ tracks when we are ready to switch to the next k slice.
+    // state_kernel_[m][n] tracks when we are ready to kick off kernel (m, n).
+    // These variable are rolling over 3 consecutive k slices: first two we are
+    // actively executing + one to track completion of kernels in the second
+    // slice.
+    static const Index P = 3;
+
+    // Handle to the allocated temporary storage for Lhs/Rhs blocks.
+    BlockMemHandle packed_mem_;
+    std::vector<LhsBlock> packed_lhs_[P - 1];
+    std::vector<RhsBlock> packed_rhs_[P - 1];
+
+    // If we choose to parallelize only by the sharding dimension, each thread
+    // will have it's own "thead local" (not a c++ thread local storage) memory
+    // for packed_lhs or packed_rhs (shard_by_col = false of true). This memory
+    // can't be passed to a kernel that might execute on a different thread.
+    //
+    // In practice when we are ready to pack memory for the sharding dimension
+    // (rhs if shard_by_col==true) of the K-th slice, all kernels for K-1 slice
+    // already computed (99% of the time), and we can pack data into the thread
+    // local storage, and guarantee that all the kernels will be executed
+    // immediately in the same thread. This significantly increases L1 cache hit
+    // ratio and reduces pressure on the memory bus.
+    //
+    // It's still possible that kernel for the K-th slice will be ready before
+    // completion of the K-1 kernel, so we have to allocate "global" packed_lhs_
+    // and packed_rhs_ to allow kernels to be executed later on a thread
+    // different from the thread that was used for packing.
+
+    // Handle for pre-allocated thread local memory buffers.
+    BlockMemHandle thread_local_pre_alocated_mem_;
+
+    // Only one of these will be initialized depending on shard_by_col value
+    // (the size will be `num_worker_threads * num_grains_in_the_sharding_dim`).
+    std::vector<LhsBlock> lhs_thread_local_pre_allocated_;
+    std::vector<RhsBlock> rhs_thread_local_pre_allocated_;
+
+    // How many thread local blocks were already allocated.
+    std::atomic<int> num_thread_local_allocations_;
+    const int thread_local_capacity;
+
+    // We will use pre-allocated Lhs/Rhs blocks defined above, if the number of
+    // unique threads in a system is below or equal to the number of threads in
+    // a thread pool. We will fallback on dynamic memory allocation after that.
+
+    // ThreadLocalBlocks is a container for Lhs or Rhs thread local buffers. Its
+    // size is equal to the grain size in Lhs/Rhs sharding dimension.
+    template <typename BlockType>
+    class ThreadLocalBlocks {
+     public:
+      ThreadLocalBlocks() = default;
+
+      ThreadLocalBlocks(BlockType* base, size_t grain_size)
+          : is_pre_allocated_(true),
+            thread_local_pre_allocated_base_(base),
+            grain_size_(grain_size) {}
+
+      ThreadLocalBlocks(BlockMemHandle mem_handle,
+                        std::vector<BlockType> blocks)
+          : is_pre_allocated_(false),
+            mem_handle_(std::move(mem_handle)),
+            blocks_(std::move(blocks)) {}
+
+      BlockType& block(int grain_index) {
+        eigen_assert(grain_index >= 0);
+        eigen_assert(static_cast<size_t>(grain_index) < size());
+        return is_pre_allocated_ ? thread_local_pre_allocated_base_[grain_index]
+                                 : blocks_[grain_index];
+      }
+
+      void Release(EvalParallelContext& ctx) const {
+        if (!is_pre_allocated_) {
+          ctx.kernel_.deallocate(ctx.device_, mem_handle_);
+        }
+      }
+
+      size_t size() const {
+        return is_pre_allocated_ ? grain_size_ : blocks_.size();
+      }
+
+     private:
+      bool is_pre_allocated_;
+
+      // Reuse pre-allocated thread local buffers.
+      BlockType* thread_local_pre_allocated_base_ = nullptr;
+      size_t grain_size_ = 0;
+
+      // These will be initialized only if `is_pre_allocated == false`.
+      BlockMemHandle mem_handle_{};
+      std::vector<BlockType> blocks_;
+    };
+
+    // ThreadLocalBlocksInitialize callable does custom thread local blocks
+    // initialization, and will reuse pre-allocated buffers if possible, or will
+    // dynamically allocate new memory.
+    //
+    // Lhs/Rhs blocks might be of the same type, so we have to pass explicitly
+    // for what side do we plan to do block allocation.
+    template <typename BlockType, bool is_rhs>
+    class ThreadLocalBlocksInitialize {
+      static constexpr bool kIsLhs =
+          !is_rhs && std::is_same<BlockType, LhsBlock>::value;
+      static const bool kIsRhs =
+          is_rhs && std::is_same<BlockType, RhsBlock>::value;
+      static_assert(kIsLhs || kIsRhs, "Unkown block type");
+
+      using Blocks = ThreadLocalBlocks<BlockType>;
+
+     public:
+      ThreadLocalBlocksInitialize(EvalParallelContext& ctx)
+          : ctx_(ctx),
+            num_worker_threads_(ctx_.device_.numThreadsInPool()) {}
+
+      void operator()(Blocks& blocks) {
+        const int n = ctx_.num_thread_local_allocations_.fetch_add(
+            1, std::memory_order_relaxed);
+
+        if (n >= num_worker_threads_) {
+          ThreadLocalBlocksAllocator<is_rhs>::allocate(ctx_, blocks);
+        } else {
+          ThreadLocalBlocksAllocator<is_rhs>::reuse(ctx_, n, blocks);
+        }
+      }
+
+     private:
+      // NOTE(ezhulenev): Without 'if constexpr' we have to put calls to
+      // TensorContractionKernel::allocateSlices into template specializations.
+      // Also explicit specializations are not allowed at class scope in C++03,
+      // EvalCtx type parameter is just a workaround for that limitation.
+      template <bool pack_rhs, typename EvalCtx = EvalParallelContext>
+      struct ThreadLocalBlocksAllocator;
+
+      template <typename EvalCtx>
+      struct ThreadLocalBlocksAllocator</*pack_rhs=*/true, EvalCtx> {
+        static void allocate(EvalCtx& ctx, Blocks& blocks) {
+          std::vector<RhsBlock> rhs_blocks;
+          BlockMemHandle mem_handle = ctx.kernel_.allocateSlices(
+              ctx.device_,
+              /*num_lhs=*/0,
+              /*num_rhs=*/ctx.gn_,
+              /*num_slices=*/1,
+              /*lhs_blocks=*/nullptr, /*rhs_blocks=*/&rhs_blocks);
+
+          blocks = ThreadLocalBlocks<RhsBlock>(std::move(mem_handle),
+                                               std::move(rhs_blocks));
+        }
+
+        static void reuse(EvalCtx& ctx, int index, Blocks& blocks) {
+          RhsBlock* ptr = &ctx.rhs_thread_local_pre_allocated_[ctx.gn_ * index];
+          blocks = ThreadLocalBlocks<RhsBlock>(ptr, ctx.gn_);
+        }
+      };
+
+      template <typename EvalCtx>
+      struct ThreadLocalBlocksAllocator</*pack_rhs=*/false, EvalCtx> {
+        static void allocate(EvalCtx& ctx, Blocks& blocks) {
+          std::vector<LhsBlock> lhs_blocks;
+          BlockMemHandle mem_handle = ctx.kernel_.allocateSlices(
+              ctx.device_,
+              /*num_lhs=*/ctx.gm_,
+              /*num_rhs=*/0,
+              /*num_slices=*/1,
+              /*lhs_blocks=*/&lhs_blocks, /*rhs_blocks=*/nullptr);
+
+          blocks = ThreadLocalBlocks<LhsBlock>(std::move(mem_handle),
+                                               std::move(lhs_blocks));
+        }
+
+        static void reuse(EvalCtx& ctx, int index, Blocks& blocks) {
+          LhsBlock* ptr = &ctx.lhs_thread_local_pre_allocated_[ctx.gm_ * index];
+          blocks = ThreadLocalBlocks<LhsBlock>(ptr, ctx.gm_);
+        }
+      };
+
+      EvalParallelContext& ctx_;
+      const int num_worker_threads_;
+    };
+
+    template <typename BlockType>
+    class ThreadLocalBlocksRelease {
+     public:
+      using Blocks = ThreadLocalBlocks<BlockType>;
+      ThreadLocalBlocksRelease(EvalParallelContext& ctx) : ctx_(ctx) {}
+      void operator()(Blocks& blocks) { blocks.Release(ctx_); }
+
+     private:
+      EvalParallelContext& ctx_;
+    };
+
+    // ThreadLocalBlocks initialization callables.
+    using ThreadLocalLhsInit =
+        ThreadLocalBlocksInitialize<LhsBlock, /*is_rhs=*/false>;
+    using ThreadLocalRhsInit =
+        ThreadLocalBlocksInitialize<RhsBlock, /*is_rhs=*/true>;
+
+    // ThreadLocalBlocks release callables.
+    using ThreadLocalLhsRelease = ThreadLocalBlocksRelease<LhsBlock>;
+    using ThreadLocalRhsRelease = ThreadLocalBlocksRelease<RhsBlock>;
+
+    // Thread local containers for Lhs/Rhs block packs. In practice only one of
+    // them will be used, depending on the shard_by_col value.
+    Eigen::ThreadLocal<ThreadLocalBlocks<LhsBlock>, ThreadLocalLhsInit,
+                       ThreadLocalLhsRelease>
+        lhs_thread_local_blocks_;
+    Eigen::ThreadLocal<ThreadLocalBlocks<RhsBlock>, ThreadLocalRhsInit,
+                       ThreadLocalRhsRelease>
+        rhs_thread_local_blocks_;
+
+    // After a particular shard for Kth slice missed thread local execution
+    // opportunity (K-1 slice didn't complete kernels execution), we can no
+    // longer schedule K+1 and following slices in thread local mode, because
+    // there is no more guarantee that previous kernels were executed
+    // sequentially in the same thread (size is nn_ or nm_).
+    std::atomic<bool>* can_use_thread_local_packed_;
+
+    std::atomic<uint8_t>** state_kernel_[P];
+    // state_switch_ is frequently modified by worker threads, while other
+    // fields are read-only after constructor. Let's move it to a separate cache
+    // line to reduce cache-coherency traffic.
+    char pad_[128];
+    std::atomic<Index> state_packing_ready_[P];
+    std::atomic<Index> state_switch_[P];
+
+    LhsBlock& packed_lhs(Index m, Index k, Index m1, bool use_thread_local) {
+      if (use_thread_local) {
+        eigen_assert(!shard_by_col_);
+        ThreadLocalBlocks<LhsBlock>& blocks = lhs_thread_local_blocks_.local();
+
+        Index grain_index = m1 - m * gm_;
+        return blocks.block(internal::convert_index<int>(grain_index)); // FIXME better make ThreadLocalBlocks use Eigen::Index?
+      } else {
+        return packed_lhs_[k % (P - 1)][m1];
+      }
+    }
+
+    RhsBlock& packed_rhs(Index n, Index k, Index n1, bool use_thread_local) {
+      if (use_thread_local) {
+        eigen_assert(shard_by_col_);
+        ThreadLocalBlocks<RhsBlock>& blocks = rhs_thread_local_blocks_.local();
+
+        Index grain_index = n1 - n * gn_;
+        return blocks.block(internal::convert_index<int>(grain_index)); // FIXME better make ThreadLocalBlocks use Eigen::Index?
+      } else {
+        return packed_rhs_[k % (P - 1)][n1];
+      }
+    }
+
+    // In following two methods (pack_lhs and pack_rhs), if we know for sure
+    // that we'll be able to immediately call a kernel with packed data, and do
+    // not submit it to the thread pool, we can use thread local memory for
+    // packed data.
+    //
+    // We can only reliably check it if we are running all kernels in sync mode
+    // (parallelize only by sharding dim). If kernel for m==0 (n==0) is ready to
+    // run, it's guaranteed that all kernels with larger values of m (n) are
+    // also ready, because we execute them in the same order for all K slices.
+
+    void pack_lhs(Index m, Index k) {
+      bool use_thread_local = false;
+
+      if (parallelize_by_sharding_dim_only_ && !shard_by_col_ &&
+          can_use_thread_local_packed_[m].load(std::memory_order_relaxed)) {
+        if (state_kernel_[k % P][m][0].load(std::memory_order_relaxed) == 1) {
+          use_thread_local = true;
+        } else {
+          // If we can't guarantee that all kernels in `k` slice will be
+          // executed sequentially in current thread, it's no longer safe to use
+          // thread local memory in following slices along the k dimensions.
+          eigen_assert(k > 0);
+          can_use_thread_local_packed_[m].store(false,
+                                                std::memory_order_relaxed);
+        }
+      }
+
+      const Index mend = m * gm_ + gm(m);
+      for (Index m1 = m * gm_; m1 < mend; m1++)
+        kernel_.packLhs(&packed_lhs(m, k, m1, use_thread_local),
+                        lhs_.getSubMapper(m1 * bm_, k * bk_), bk(k), bm(m1));
+
+      if (!parallel_pack_ && shard_by_col_) {
+        assert(!use_thread_local);
+        signal_packing(k);
+      } else {
+        signal_switch(k + 1);
+        for (Index n = nn_ - 1; n >= 0; n--) {
+          bool sync = parallelize_by_sharding_dim_only_ || n == 0;
+          signal_kernel(m, n, k, sync, use_thread_local);
+        }
+      }
+    }
+
+    void pack_rhs(Index n, Index k) {
+      bool use_thread_local = false;
+
+      if (parallelize_by_sharding_dim_only_ && shard_by_col_ &&
+          can_use_thread_local_packed_[n].load(std::memory_order_relaxed)) {
+        if (state_kernel_[k % P][0][n].load(std::memory_order_relaxed) == 1) {
+          use_thread_local = true;
+        } else {
+          // If we can't guarantee that all kernels in `k` slice will be
+          // executed sequentially in current thread, it's no longer safe to use
+          // thread local memory in followig slices along the k dimensions.
+          eigen_assert(k > 0);
+          can_use_thread_local_packed_[n].store(false,
+                                                std::memory_order_relaxed);
+        }
+      }
+
+      const Index nend = n * gn_ + gn(n);
+      for (Index n1 = n * gn_; n1 < nend; n1++) {
+        if (!TensorContractionKernel::HasBeta && k == 0) {
+          // Zero the output memory in parallel, only if contraction kernel does
+          // not support `beta`. Otherwise we will pass beta 0.0 to the first
+          // call to the `TensorContractionKernel::invoke()`.
+          //
+          // On 10000x2x10000 mm zeroing can easily take half of time. Zero (bn
+          // x m) row. Safe to do here because all kernels that will write to
+          // this memory depend on completion of this task. Note: don't call
+          // device_.memset() here. device_.memset() blocks on thread pool
+          // worker thread, which can lead to underutilization and deadlocks.
+          memset(buffer_ + n1 * bn_ * m_, 0, bn(n1) * m_ * sizeof(Scalar));
+        }
+        kernel_.packRhs(&packed_rhs(n, k, n1, use_thread_local),
+                        rhs_.getSubMapper(k * bk_, n1 * bn_), bk(k), bn(n1));
+      }
+
+      if (parallel_pack_ || shard_by_col_) {
+        signal_switch(k + 1);
+        for (Index m = nm_ - 1; m >= 0; m--) {
+          bool sync = parallelize_by_sharding_dim_only_ || m == 0;
+          signal_kernel(m, n, k, sync, use_thread_local);
+        }
+      } else {
+        assert(!use_thread_local);
+        signal_packing(k);
+      }
+    }
+
+    void kernel(Index m, Index n, Index k, bool use_thread_local) {
+      // Note: order of iteration matters here. Iteration over m is innermost
+      // because we want to reuse the same packed rhs in consecutive tasks
+      // (rhs fits into L2$ while lhs only into L3$).
+      const Index nend = n * gn_ + gn(n);
+      const Index mend = m * gm_ + gm(m);
+
+      // NOTE: output = alpha * LHS * RHS + beta * output.
+      const Scalar alpha = Scalar(1);
+      const Scalar beta =
+          (TensorContractionKernel::HasBeta && k == 0) ? Scalar(0) : Scalar(1);
+
+      if (shard_by_col_) {
+        for (Index n1 = n * gn_; n1 < nend; n1++) {
+          for (Index m1 = m * gm_; m1 < mend; m1++) {
+            const auto output_mapper = output_.getSubMapper(m1 * bm_, n1 * bn_);
+            kernel_.invoke(
+                output_mapper,
+                packed_lhs(m, k, m1, !shard_by_col_ && use_thread_local),
+                packed_rhs(n, k, n1, shard_by_col_ && use_thread_local), bm(m1),
+                bk(k), bn(n1), alpha, beta);
+
+            // We are done with the last task for the [m1, n1] block.
+            if (k + 1 == nk_) {
+              output_kernel_(output_mapper, tensor_contraction_params_,
+                             m1 * bm_, n1 * bn_, bm(m1), bn(n1));
+            }
+          }
+        }
+      } else {
+        for (Index m1 = m * gm_; m1 < mend; m1++)
+          for (Index n1 = n * gn_; n1 < nend; n1++) {
+            const auto output_mapper = output_.getSubMapper(m1 * bm_, n1 * bn_);
+            kernel_.invoke(
+                output_mapper,
+                packed_lhs(m, k, m1, !shard_by_col_ && use_thread_local),
+                packed_rhs(n, k, n1, shard_by_col_ && use_thread_local), bm(m1),
+                bk(k), bn(n1), alpha, beta);
+
+            // We are done with the last task for the [m1, n1] block.
+            if (k + 1 == nk_) {
+              output_kernel_(output_mapper, tensor_contraction_params_,
+                             m1 * bm_, n1 * bn_, bm(m1), bn(n1));
+            }
+          }
+      }
+      signal_kernel(m, n, k + 1, /*sync=*/false, /*use_thread_local=*/false);
+      signal_switch(k + 2);
+    }
+
+    void signal_packing(Index k) {
+      eigen_assert(!parallel_pack_);
+      Index s = state_packing_ready_[k % P].fetch_sub(1);
+      eigen_assert(s > 0);
+      if (s != 1) return;
+      state_packing_ready_[k % P] = shard_by_col_ ? nm_ : nn_;
+      enqueue_packing(k, shard_by_col_);
+    }
+
+    void signal_kernel(Index m, Index n, Index k, bool sync,
+                       bool use_thread_local) {
+      std::atomic<uint8_t>* state = &state_kernel_[k % P][m][n];
+      Index s = state->load();
+      eigen_assert(s > 0);
+      if (s != 1 && state->fetch_sub(1) != 1) {
+        eigen_assert(!use_thread_local);
+        return;
+      }
+      state->store(parallel_pack_ ? 3 : 2, std::memory_order_relaxed);
+      if (sync) {
+        kernel(m, n, k, use_thread_local);
+      } else {
+        eigen_assert(!use_thread_local);
+        device_.enqueueNoNotification(
+            [=]() { kernel(m, n, k, use_thread_local); });
+      }
+    }
+
+    void signal_switch(Index k, Index v = 1) {
+      Index s = state_switch_[k % P].fetch_sub(v);
+      eigen_assert(s >= v);
+      if (s != v) return;
+
+      // Ready to switch to the next k slice.
+      // Reset counter for the next iteration.
+      state_switch_[k % P] =
+          (parallel_pack_ ? nm_ + nn_ : (shard_by_col_ ? nn_ : nm_)) +
+          nm_ * nn_;
+      if (k < nk_) {
+        // Issue lhs/rhs packing. Their completion will in turn kick off
+        // kernels.
+        if (parallel_pack_) {
+          enqueue_packing(k, !shard_by_col_);
+          enqueue_packing(k, shard_by_col_);
+        } else if (shard_by_col_) {
+          enqueue_packing(k, false);
+        } else {
+          enqueue_packing(k, true);
+        }
+
+        // Termination handling.
+        // Because kernel completion signals k + 2 switch, we need to finish nk
+        // + 2 slices without issuing any tasks on nk + 1 slice. So here we
+        // pretend that all nk + 1 packing tasks just finish instantly; so that
+        // nk + 2 switch only waits for completion of nk kernels.
+      } else if (k == nk_) {
+        signal_switch(k + 1,
+                      parallel_pack_ ? nm_ + nn_ : (shard_by_col_ ? nn_ : nm_));
+      } else {
+        done_.Notify();
+      }
+    }
+
+    // Enqueue all rhs/lhs packing for k-th slice.
+    void enqueue_packing(Index k, bool rhs) {
+      enqueue_packing_helper(0, rhs ? nn_ : nm_, k, rhs);
+    }
+
+    void enqueue_packing_helper(Index start, Index end, Index k, bool rhs) {
+      if (end - start == 1) {
+        if (rhs)
+          pack_rhs(start, k);
+        else
+          pack_lhs(start, k);
+      } else {
+        while (end - start > 1) {
+          Index mid = (start + end) / 2;
+          device_.enqueueNoNotification(
+              [=]() { enqueue_packing_helper(mid, end, k, rhs); });
+          end = mid;
+        }
+
+        // Decide if we want to run first packing task (start == 0) in
+        // async mode if we parallelize only by sharding dim:
+        // (1) pack_lhs and pack_rhs call signal_switch before completing
+        //     all calls to signal_kernel, which in sync mode might lead
+        //     to the execution of the first kernel of the k+1 slice, before
+        //     completing a call to the last kernel of the k slice.
+        // (2) all pack tasks for sharded dim must be executed in a thread
+        //     pool to get pre-allocated thead local buffers.
+        bool pack_async =
+          (start == 0) &&
+          (parallelize_by_sharding_dim_only_&& shard_by_col_ == rhs) &&
+          (k > 0 || std::this_thread::get_id() == created_by_thread_id_);
+
+        if (pack_async) {
+          device_.enqueueNoNotification(
+              [=]() { enqueue_packing_helper(start, end, k, rhs); });
+        } else {
+          enqueue_packing_helper(start, end, k, rhs);
+        }
+      }
+    }
+
+    // Block sizes with accounting for potentially incomplete last block.
+    Index bm(Index m) const { return m + 1 < nm0_ ? bm_ : m_ + bm_ - bm_ * nm0_; }
+    Index bn(Index n) const { return n + 1 < nn0_ ? bn_ : n_ + bn_ - bn_ * nn0_; }
+    Index bk(Index k) const { return k + 1 < nk_ ? bk_ : k_ + bk_ - bk_ * nk_; }
+    // Task grain sizes accounting for potentially incomplete last task.
+    Index gm(Index m) const { return m + 1 < nm_ ? gm_ : nm0_ + gm_ - gm_ * nm_; }
+    Index gn(Index n) const { return n + 1 < nn_ ? gn_ : nn0_ + gn_ - gn_ * nn_; }
+
+    EvalParallelContext(const EvalParallelContext&) = delete;
+    void operator=(const EvalParallelContext&) = delete;
+  };
+
+  template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous,
+            bool rhs_inner_dim_reordered, int Alignment>
+  using SyncEvalParallelContext =
+      EvalParallelContext<NoCallback, lhs_inner_dim_contiguous,
+                          rhs_inner_dim_contiguous, rhs_inner_dim_reordered,
+                          Alignment>;
+
+  // ------------------------------------------------------------------------ //
+
+  // EvalShardedByInnerDimContext orchestrates sync/async contraction
+  // evaluation, when we shard by inner dimension. When it is executed in
+  // asynchronous mode, it owns all the shared state that might be accessible by
+  // block processing tasks.
+
+  template <typename DoneCallback>
+  struct EvalShardedByInnerDimContext {
+    EvalShardedByInnerDimContext(const Self* self, int num_threads,
+                                 Scalar* result_buffer,
+                                 Index m_size, Index n_size, Index k_size,
+                                 DoneCallback done_callback)
+        : evaluator(self),
+          m_lhs_inner_dim_contiguous(evaluator->m_lhs_inner_dim_contiguous),
+          m_rhs_inner_dim_contiguous(evaluator->m_rhs_inner_dim_contiguous),
+          m_rhs_inner_dim_reordered(evaluator->m_rhs_inner_dim_reordered),
+          result(result_buffer),
+          m(m_size),
+          n(n_size),
+          k(k_size),
+          done(std::move(done_callback)),
+          buffer_size_bytes(m * n * sizeof(Scalar)),
+          block_size(blockSize(k, num_threads)),
+          num_blocks(divup<Index>(k, block_size)),
+          num_pending_blocks(internal::convert_index<int>(num_blocks)),
+          l0_ranges(divup<Index>(num_blocks, l0_size)),
+          l0_state(l0_ranges),
+          block_buffers(num_blocks) {
+      // Keep count of pending gemm tasks for each l0 range.
+      for (int i = 0; i < l0_ranges; ++i) {
+        const Index num_pending_tasks = actualRangeSize(l0_ranges, l0_size, i);
+        l0_state.emplace_back(internal::convert_index<int>(num_pending_tasks));
+      }
+
+      // Allocate temporary buffers for each block.
+      for (Index block_idx = 0; block_idx < num_blocks; ++block_idx) {
+        Scalar* buf = block_idx == 0
+                          ? result
+                          : static_cast<Scalar*>(evaluator->m_device.allocate(
+                                buffer_size_bytes));
+        block_buffers.emplace_back(buf);
+      }
+    }
+
+    ~EvalShardedByInnerDimContext() {
+      for (Index i = 1; i < num_blocks; ++i) {
+        evaluator->m_device.deallocate(block_buffers[i]);
+      }
+    }
+
+    template <int Alignment>
+    void run() {
+      Barrier barrier(internal::convert_index<int>(num_blocks));
+      eval<Alignment>(barrier, 0, num_blocks);
+      barrier.Wait();
+
+      // Aggregate partial sums from l0 ranges.
+      aggregateL0Blocks<Alignment>();
+
+      // Apply output kernel.
+      applyOutputKernel();
+    }
+
+    template <int Alignment>
+    void runAsync() {
+      evalAsync<Alignment>(0, num_blocks);
+    }
+
+   private:
+    // The underlying GEMM kernel assumes that k is a multiple of
+    // the packet size and subtle breakage occurs if this is violated.
+    static const Index packet_size = internal::packet_traits<RhsScalar>::size;
+
+    const Self* evaluator;  // TensorContraction evaluator
+
+    // These fields required fromTENSOR_CONTRACTION_DISPATCH macro.
+    bool m_lhs_inner_dim_contiguous;
+    bool m_rhs_inner_dim_contiguous;
+    bool m_rhs_inner_dim_reordered;
+
+    Scalar* result;
+
+    Index m;
+    Index n;
+    Index k;
+
+    DoneCallback done;
+
+    // ----------------------------------------------------------------------//
+    // Algorithm parameters.
+
+    // We will compute partial results into the buffers of this size.
+    Index buffer_size_bytes;
+
+    Index block_size;
+    Index num_blocks;
+
+    // Keep track of pending tasks when evaluate in async mode.
+    std::atomic<int> num_pending_blocks;
+
+    // We compute partial gemm results in parallel, and to get the final result
+    // we need to add them all together. For the large number of threads (>= 48)
+    // this adds a very expensive sequential step at the end.
+    //
+    // We split the [0, num_blocks) into small ranges, and when a task for the
+    // block finishes its partial gemm computation, it checks if it was the last
+    // gemm in the range, and if so, it will add all blocks of the range.
+    //
+    // After all tasks done, we need to add only these pre-aggregated blocks.
+
+    // For now we use just a single level of ranges to compute pre-aggregated
+    // partial sums, but in general we can use more layers to compute tree
+    // aggregation in parallel and reduce the size of the sequential step.
+    //
+    // TODO(ezhulenev): Add multilevel tree aggregation? Probably will make
+    // sense only if number of threads >= ~128?
+    static const Index l0_size = 4;
+    Index l0_ranges;
+
+    // Keep count of pending gemm tasks for each l0 range.
+    MaxSizeVector<std::atomic<int>> l0_state;  // [0, l0_ranges)
+
+    // Buffers allocated for each temporary block computation.
+    MaxSizeVector<Scalar*> block_buffers;  // [0, num_blocks)
+
+    template <int Alignment>
+    void processBlock(Index block_idx, Index begin, Index end) {
+      Scalar* buf = block_buffers[block_idx];
+
+      TENSOR_CONTRACTION_DISPATCH(
+          evaluator->template evalGemmPartialWithoutOutputKernel, Alignment,
+          (buf, begin, end,
+           /*num_threads=*/internal::convert_index<int>(num_blocks)));
+
+      // Check if it was the last task in l0 range.
+      const Index l0_index = block_idx / l0_size;
+      const int v = l0_state[l0_index].fetch_sub(1);
+      eigen_assert(v >= 1);
+
+      // If we processed the last block of the range, we can aggregate all
+      // partial results into the first block of the range.
+      if (v == 1) {
+        const Index rng_size = actualRangeSize(l0_ranges, l0_size, l0_index);
+        const Index dst_block_idx = l0_index * l0_size;
+
+        if (rng_size == l0_size) {
+          addAllToBuffer<Alignment>(
+              m * n,
+              /*src_buf0=*/block_buffers[dst_block_idx + 1],
+              /*src_buf1=*/block_buffers[dst_block_idx + 2],
+              /*src_buf2=*/block_buffers[dst_block_idx + 3],
+              /*dst_buf= */ block_buffers[dst_block_idx]);
+        } else {
+          // Aggregate blocks of potentially incomplete last range.
+          for (int i = 1; i < rng_size; ++i) {
+            addToBuffer<Alignment>(m * n,
+                                   /*src_buf=*/block_buffers[dst_block_idx + i],
+                                   /*dst_buf=*/block_buffers[dst_block_idx]);
+          }
+        }
+      }
+    }
+
+    // Aggregate partial sums from l0 ranges.
+    template <int Alignment>
+    void aggregateL0Blocks() const {
+      Index l0_index = 1;
+
+      for (; l0_index + 2 < l0_ranges; l0_index += 3) {
+        addAllToBuffer<Alignment>(
+            m * n,
+            /*src_buf0=*/block_buffers[(l0_index + 0) * l0_size],
+            /*src_buf1=*/block_buffers[(l0_index + 1) * l0_size],
+            /*src_buf2=*/block_buffers[(l0_index + 2) * l0_size],
+            /*dst_buf= */ block_buffers[0]);
+      }
+
+      for (; l0_index < l0_ranges; ++l0_index) {
+        addToBuffer<Alignment>(m * n, block_buffers[l0_index * l0_size],
+                               block_buffers[0]);
+      }
+    }
+
+    void applyOutputKernel() const {
+      typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
+      evaluator->m_output_kernel(
+          OutputMapper(result, m), evaluator->m_tensor_contraction_params,
+          static_cast<Eigen::Index>(0), static_cast<Eigen::Index>(0), m, n);
+    }
+
+    // Compute block size with accounting for potentially incomplete last block.
+    Index actualBlockSize(Index block_idx) const {
+      return block_idx + 1 < num_blocks
+                 ? block_size
+                 : k + block_size - block_size * num_blocks;
+    };
+
+    // Compute range size with accounting for potentially incomplete last range.
+    Index actualRangeSize(Index num_ranges, Index range_size,
+                          Index range_idx) const {
+      eigen_assert(range_idx < num_ranges);
+      return range_idx + 1 < num_ranges
+                 ? range_size
+                 : num_blocks + range_size - range_size * num_ranges;
+    };
+
+    template <int Alignment>
+    EIGEN_STRONG_INLINE static void addToBuffer(size_t n, const Scalar* src_buf,
+                                                Scalar* tgt_buf) {
+      const int output_packet_size =
+          internal::unpacket_traits<PacketReturnType>::size;
+      size_t i = 0;
+      const size_t num_packets = n / output_packet_size;
+      for (; i < output_packet_size * num_packets; i += output_packet_size) {
+        const PacketReturnType src_val =
+            internal::pload<PacketReturnType>(src_buf + i);
+        const PacketReturnType tgt_val =
+            internal::ploadt<PacketReturnType, Alignment>(tgt_buf + i);
+        const PacketReturnType sum = internal::padd(src_val, tgt_val);
+        internal::pstoret<Scalar, PacketReturnType, Alignment>(tgt_buf + i,
+                                                               sum);
+      }
+      for (; i < n; ++i) {
+        tgt_buf[i] += src_buf[i];
+      }
+    }
+
+    template <int Alignment>
+    EIGEN_STRONG_INLINE static void addAllToBuffer(size_t n,
+                                                   const Scalar* src_buf0,
+                                                   const Scalar* src_buf1,
+                                                   const Scalar* src_buf2,
+                                                   Scalar* dst_buf) {
+      using ::Eigen::internal::padd;
+      using ::Eigen::internal::pload;
+      using ::Eigen::internal::ploadt;
+      using ::Eigen::internal::pstoret;
+
+      const int output_packet_size =
+          internal::unpacket_traits<PacketReturnType>::size;
+
+      size_t i = 0;
+      const size_t num_packets = n / output_packet_size;
+      for (; i < output_packet_size * num_packets; i += output_packet_size) {
+        const auto src_val0 = pload<PacketReturnType>(src_buf0 + i);
+        const auto src_val1 = pload<PacketReturnType>(src_buf1 + i);
+        const auto src_val2 = pload<PacketReturnType>(src_buf2 + i);
+
+        const auto dst_val = ploadt<PacketReturnType, Alignment>(dst_buf + i);
+        const auto sum =
+            padd(padd(dst_val, src_val0), padd(src_val1, src_val2));
+
+        pstoret<Scalar, PacketReturnType, Alignment>(dst_buf + i, sum);
+      }
+      for (; i < n; ++i) {
+        dst_buf[i] += src_buf0[i] + src_buf1[i] + src_buf2[i];
+      }
+    }
+
+    template <int Alignment>
+    void eval(Barrier& barrier, Index start_block_idx, Index end_block_idx) {
+      while (end_block_idx - start_block_idx > 1) {
+        Index mid_block_idx = (start_block_idx + end_block_idx) / 2;
+        evaluator->m_device.enqueueNoNotification(
+            [this, &barrier, mid_block_idx, end_block_idx]() {
+              eval<Alignment>(barrier, mid_block_idx, end_block_idx);
+            });
+        end_block_idx = mid_block_idx;
+      }
+
+      Index block_idx = start_block_idx;
+      Index block_start = block_idx * block_size;
+      Index block_end = block_start + actualBlockSize(block_idx);
+
+      processBlock<Alignment>(block_idx, block_start, block_end);
+      barrier.Notify();
+    }
+
+    template <int Alignment>
+    void evalAsync(Index start_block_idx, Index end_block_idx) {
+      while (end_block_idx - start_block_idx > 1) {
+        Index mid_block_idx = (start_block_idx + end_block_idx) / 2;
+        evaluator->m_device.enqueueNoNotification(
+            [this, mid_block_idx, end_block_idx]() {
+              evalAsync<Alignment>(mid_block_idx, end_block_idx);
+            });
+        end_block_idx = mid_block_idx;
+      }
+
+      Index block_idx = start_block_idx;
+
+      Index block_start = block_idx * block_size;
+      Index block_end = block_start + actualBlockSize(block_idx);
+
+      processBlock<Alignment>(block_idx, block_start, block_end);
+
+      int v = num_pending_blocks.fetch_sub(1);
+      eigen_assert(v >= 1);
+
+      if (v == 1) {
+        // Aggregate partial sums from l0 ranges.
+        aggregateL0Blocks<Alignment>();
+
+        // Apply output kernel.
+        applyOutputKernel();
+
+        // NOTE: If we call `done` callback before deleting this (context),
+        // it might deallocate Self* pointer captured by context, and we'll
+        // fail in destructor trying to deallocate temporary buffers.
+
+        // Move done call back from context before it will be destructed.
+        DoneCallback done_copy = std::move(done);
+
+        // We are confident that we are the last one who touches context.
+        delete this;
+
+        // Now safely call the done callback.
+        done_copy();
+      }
+    }
+
+    // Cost model doesn't capture well the cost associated with constructing
+    // tensor contraction mappers and computing loop bounds in gemm_pack_lhs
+    // and gemm_pack_rhs, so we specify minimum desired block size.
+    static Index blockSize(Index k, int num_threads) {
+      const auto round_up = [=](Index index) -> Index {
+        const Index kmultiple = packet_size <= 8 ? 8 : packet_size;
+        return divup<Index>(index, kmultiple) * kmultiple;
+      };
+
+      const Index target_block_size = round_up(divup<Index>(k, num_threads));
+      const Index desired_min_block_size = 12 * packet_size;
+
+      return numext::mini<Index>(
+          k, numext::maxi<Index>(desired_min_block_size, target_block_size));
+    }
+
+    EvalShardedByInnerDimContext(const EvalShardedByInnerDimContext&) = delete;
+    void operator=(const EvalShardedByInnerDimContext&) = delete;
+  };
+
+  // ------------------------------------------------------------------------ //
+
+  // Below are the function used by evalProductImpl heuristics, trying to select
+  // optimcal parameters for parallelization algorithm.
+
+  // Decide whether we want to shard m x n contraction by columns or by rows.
+  static bool shardByCol(Index m, Index n, Index num_threads) {
+    // Note: we are comparing both n and m against Traits::nr, it is not
+    // a mistake. We are trying to figure out how both n and m will fit into
+    // the main sharding dimension.
+
+    // Sharding by column is the default
+    // ... unless there is enough data for vectorization over rows
+    if (m / num_threads >= Traits::nr &&
+        // and not enough data for vectorization over columns
+        (n / num_threads < Traits::nr ||
+         // ... or barely enough data for vectorization over columns,
+         // but it is not evenly dividable across threads
+         (n / num_threads < 4 * Traits::nr &&
+          (n % (num_threads * Traits::nr)) != 0 &&
+          // ... and it is evenly dividable across threads for rows
+          ((m % (num_threads * Traits::nr)) == 0 ||
+           // .. or it is not evenly dividable for both dimensions but
+           // there is much more data over rows so that corner effects are
+           // mitigated.
+           (m / n >= 6)))))
+      return false;
+    // Wait, or if matrices are just substantially prolonged over the other
+    // dimension.
+    if (n / num_threads < 16 * Traits::nr && m > n * 32) return false;
+    return true;
+  }
+
+  Index coarsenM(Index m, Index n, Index bm, Index bn, Index bk, Index gn,
+                 int num_threads, bool shard_by_col) const {
+    Index gm = 1;
+    Index gm1 = 1;
+    Index nm0 = divup(m, bm);
+    Index nm1 = nm0;
+    for (;;) {
+      // Find the next candidate for m grain size. It needs to result in
+      // different number of blocks. E.g. if we have 10 kernels, we want to try
+      // 5 and 10, but not 6, 7, 8 and 9.
+      while (gm1 <= nm0 && nm1 == divup(nm0, gm1)) gm1++;
+      if (gm1 > nm0) break;
+      // Check the candidate.
+      int res = checkGrain(m, n, bm, bn, bk, gm1, gn, gm, gn, num_threads,
+                           shard_by_col);
+      if (res < 0) break;
+      nm1 = divup(nm0, gm1);
+      if (res == 0) continue;
+      // Commit new grain size.
+      gm = gm1;
+    }
+    return gm;
+  }
+
+  Index coarsenN(Index m, Index n, Index bm, Index bn, Index bk, Index gm,
+                 int num_threads, bool shard_by_col) const {
+    Index gn = 1;
+    Index gn1 = 1;
+    Index nn0 = divup(n, bn);
+    Index nn1 = nn0;
+    for (;;) {
+      while (gn1 <= nn0 && nn1 == divup(nn0, gn1)) gn1++;
+      if (gn1 > nn0) break;
+      int res = checkGrain(m, n, bm, bn, bk, gm, gn1, gm, gn, num_threads,
+                           shard_by_col);
+      if (res < 0) break;
+      nn1 = divup(nn0, gn1);
+      if (res == 0) continue;
+      gn = gn1;
+    }
+    return gn;
+  }
+
+  // checkGrain checks whether grain (gm, gn) is suitable and is better than
+  // (oldgm, oldgn).
+  int checkGrain(Index m, Index n, Index bm, Index bn, Index bk, Index gm,
+                 Index gn, Index oldgm, Index oldgn, int num_threads,
+                 bool shard_by_col) const {
+    const TensorOpCost cost =
+        contractionCost(bm * gm, bn * gn, bm, bn, bk, shard_by_col, true);
+    double taskSize = TensorCostModel<ThreadPoolDevice>::taskSize(
+        static_cast<double>(bm) * gm * bn * gn, cost);
+    // If the task is too small, then we agree on it regardless of anything
+    // else. Otherwise synchronization overheads will dominate.
+    if (taskSize < 1) return 1;
+    // If it is too large, then we reject it and all larger tasks.
+    if (taskSize > 2) return -1;
+    // Now we are in presumably good task size range.
+    // The main deciding factor here is parallelism. Consider that we have 12
+    // kernels and 4 threads. Grains of 2, 3 and 4 all yield good task sizes.
+    // But 2/4 yield 6/3 tasks, which gives us parallelism of 0.75 (at most 3/4
+    // of cores will be busy). While grain size 3 gives us 4 tasks, which gives
+    // us parallelism of 1 (we can load all cores).
+    Index nm0 = divup(m, bm);
+    Index nn0 = divup(n, bn);
+    Index new_tasks = divup(nm0, gm) * divup(nn0, gn);
+    double new_parallelism = static_cast<double>(new_tasks) /
+                             (divup<int>(new_tasks, num_threads) * num_threads);
+    Index old_tasks = divup(nm0, oldgm) * divup(nn0, oldgn);
+    double old_parallelism = static_cast<double>(old_tasks) /
+                             (divup<int>(old_tasks, num_threads) * num_threads);
+    if (new_parallelism > old_parallelism || new_parallelism == 1) return 1;
+    return 0;
+  }
+
+  TensorOpCost contractionCost(Index m, Index n, Index bm, Index bn, Index bk,
+                               bool shard_by_col, bool prepacked) const {
+    const int packed_size = std::min<int>(PacketType<LhsScalar, Device>::size,
+                                          PacketType<RhsScalar, Device>::size);
+    const int output_packet_size = internal::unpacket_traits<PacketReturnType>::size;
+    const double kd = static_cast<double>(bk);
+    double compute_bandwidth = computeBandwidth(false, bm, bn, bk);
+    // Computations.
+    TensorOpCost cost = TensorOpCost(0, 0, kd * compute_bandwidth, true, packed_size);
+    // Output stores.
+    cost += TensorOpCost(0, sizeof(CoeffReturnType), 0, true, output_packet_size);
+    if (prepacked) {
+      // Packing and kernels are executed in different tasks. When we calculate
+      // task grain size we look only at kernel cost assuming that kernel
+      // is more expensive than packing.
+      return cost;
+    }
+    // Lhs/rhs loads + computations.
+    TensorOpCost lhsCost = this->m_leftImpl.costPerCoeff(true) * (kd / n);
+    TensorOpCost rhsCost = this->m_rightImpl.costPerCoeff(true) * (kd / m);
+    // Lhs packing memory cost does not contribute considerably to overall
+    // execution time because lhs is prefetched early and accessed sequentially.
+    if (shard_by_col)
+      lhsCost.dropMemoryCost();
+    else
+      rhsCost.dropMemoryCost();
+    return cost + lhsCost + rhsCost;
+  }
+
+  // Decide whether we want to shard m x k x n contraction over the inner
+  // (contraction) dimension (k).
+  static bool shardByInnerDim(Index m, Index n, Index k, int num_threads,
+                              int num_threads_by_k) {
+    std::ptrdiff_t bufsize = m * n * sizeof(Scalar);
+    bool shard_by_k = false;
+    if (n == 1 ||                // If mat*vec or...
+        num_threads_by_k < 2 ||  // running single threaded or...
+        num_threads_by_k <
+            num_threads ||  // sharding by k gives less parallelism or...
+        bufsize > l3CacheSize() / num_threads_by_k ||  // need more buffer space
+        // than L3 cache or...
+        k / num_threads_by_k < 2 * Traits::nr) {  // k per thread is tiny.
+      shard_by_k = false;
+    } else if (numext::maxi(m, n) / num_threads <
+                   Traits::nr ||  // both other dimensions are tiny or...
+               // k per thread is not small and...
+               (k / num_threads_by_k > 8 * Traits::nr &&
+                // one of the outer dimensions is tiny or sharding by k offers
+                // more parallelism.
+                (numext::mini(m, n) < 2 * Traits::nr ||
+                 num_threads_by_k > num_threads))) {
+      shard_by_k = true;
+    }
+    return shard_by_k;
+  }
+
+  TensorOpCost contractionCostPerInnerDim(Index m, Index n, Index k) const {
+    // Compute cost.
+    const int output_packet_size = internal::unpacket_traits<PacketReturnType>::size;
+    TensorOpCost cost(0, 0, (computeBandwidth(true, m, n, k) * m) * n, true, output_packet_size);
+    // Output stores.
+    cost += TensorOpCost(0, sizeof(CoeffReturnType), 0, true, output_packet_size);
+    TensorOpCost lhsCost = this->m_leftImpl.costPerCoeff(true) * m;
+    TensorOpCost rhsCost = this->m_rightImpl.costPerCoeff(true) * n;
+    // Since the inner gemm kernel is always sharded by column, the lhs
+    // load cost is negligible.
+    lhsCost.dropMemoryCost();
+    return cost + lhsCost + rhsCost;
+  }
+
+  int numThreadsInnerDim(Index m, Index n, Index k) const {
+    const int output_packet_size = internal::unpacket_traits<PacketReturnType>::size;
+    TensorOpCost cost = contractionCostPerInnerDim(m, n, k);
+    double total_parallel_cost =
+        TensorCostModel<ThreadPoolDevice>::totalCost(k, cost);
+    // Cost of reduction step accumulating the m*n per-thread buffers into the
+    // result.
+    double reduction_cost = TensorCostModel<ThreadPoolDevice>::totalCost(
+        m * n, TensorOpCost(2, 1, 1, true, output_packet_size));
+    int num_threads = 1;
+    double min_cost = total_parallel_cost;
+    double kPerThreadOverHead = 3000;
+    double kFixedOverHead = 100000;
+    for (int nt = 2; nt <= this->m_device.numThreads(); nt += 2) {
+      double sequential_cost =
+          kFixedOverHead + nt * (reduction_cost + kPerThreadOverHead);
+      double parallel_cost = total_parallel_cost / nt + sequential_cost;
+      if (parallel_cost < min_cost) {
+        num_threads = nt;
+        min_cost = parallel_cost;
+      }
+    }
+    return num_threads;
+  }
+
+  double computeBandwidth(bool shard_by_col, Index bm, Index bn,
+                          Index bk) const {
+    // Peak VFMA bandwidth is 0.5. However if we have not enough data for
+    // vectorization bandwidth drops. The 4.0 and 2.0 bandwidth is determined
+    // experimentally.
+    double computeBandwidth =
+        bk == 1 ? 4.0
+                : (shard_by_col ? bn : bm) < Traits::nr ||
+                          (shard_by_col ? bm : bn) < Traits::mr
+                      ? 2.0
+                      : 0.5;
+#ifndef EIGEN_VECTORIZE_FMA
+    // Bandwidth of all of VFMA/MULPS/ADDPS is 0.5 on latest Intel processors.
+    // However for MULPS/ADDPS we have dependent sequence of 2 such
+    // instructions,
+    // so overall bandwidth is 1.0.
+    if (computeBandwidth == 0.5) computeBandwidth = 1.0;
+#endif
+    return computeBandwidth;
+  }
+
+};
+
+} // end namespace Eigen
+
+#endif  // EIGEN_USE_THREADS
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_THREAD_POOL_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorConversion.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorConversion.h
new file mode 100644
index 0000000..09d2da9
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorConversion.h
@@ -0,0 +1,456 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONVERSION_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONVERSION_H
+
+namespace Eigen {
+
+/** \class TensorConversionOp
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor conversion class. This class makes it possible to vectorize
+  * type casting operations when the number of scalars per packet in the source
+  * and the destination type differ
+  */
+namespace internal {
+template<typename TargetType, typename XprType>
+struct traits<TensorConversionOp<TargetType, XprType> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs are different.
+  typedef TargetType Scalar;
+  typedef typename traits<XprType>::StorageKind StorageKind;
+  typedef typename traits<XprType>::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = traits<XprType>::NumDimensions;
+  static const int Layout = traits<XprType>::Layout;
+  enum { Flags = 0 };
+  typedef typename TypeConversion<Scalar, typename traits<XprType>::PointerType>::type PointerType;
+};
+
+template<typename TargetType, typename XprType>
+struct eval<TensorConversionOp<TargetType, XprType>, Eigen::Dense>
+{
+  typedef const TensorConversionOp<TargetType, XprType>& type;
+};
+
+template<typename TargetType, typename XprType>
+struct nested<TensorConversionOp<TargetType, XprType>, 1, typename eval<TensorConversionOp<TargetType, XprType> >::type>
+{
+  typedef TensorConversionOp<TargetType, XprType> type;
+};
+
+}  // end namespace internal
+
+
+template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket, int SrcCoeffRatio, int TgtCoeffRatio>
+struct PacketConverter;
+
+template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket>
+struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 1, 1> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketConverter(const TensorEvaluator& impl)
+      : m_impl(impl) {}
+
+  template<int LoadMode, typename Index>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TgtPacket packet(Index index) const {
+    return internal::pcast<SrcPacket, TgtPacket>(m_impl.template packet<LoadMode>(index));
+  }
+
+ private:
+  const TensorEvaluator& m_impl;
+};
+
+
+template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket>
+struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 2, 1> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketConverter(const TensorEvaluator& impl)
+      : m_impl(impl) {}
+
+  template<int LoadMode, typename Index>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TgtPacket packet(Index index) const {
+    const int SrcPacketSize = internal::unpacket_traits<SrcPacket>::size;
+
+    SrcPacket src1 = m_impl.template packet<LoadMode>(index);
+    SrcPacket src2 = m_impl.template packet<LoadMode>(index + SrcPacketSize);
+    TgtPacket result = internal::pcast<SrcPacket, TgtPacket>(src1, src2);
+    return result;
+  }
+
+ private:
+  const TensorEvaluator& m_impl;
+};
+
+template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket>
+struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 4, 1> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketConverter(const TensorEvaluator& impl)
+      : m_impl(impl) {}
+
+  template<int LoadMode, typename Index>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TgtPacket packet(Index index) const {
+    const int SrcPacketSize = internal::unpacket_traits<SrcPacket>::size;
+
+    SrcPacket src1 = m_impl.template packet<LoadMode>(index);
+    SrcPacket src2 = m_impl.template packet<LoadMode>(index + SrcPacketSize);
+    SrcPacket src3 = m_impl.template packet<LoadMode>(index + 2 * SrcPacketSize);
+    SrcPacket src4 = m_impl.template packet<LoadMode>(index + 3 * SrcPacketSize);
+    TgtPacket result = internal::pcast<SrcPacket, TgtPacket>(src1, src2, src3, src4);
+    return result;
+  }
+
+ private:
+  const TensorEvaluator& m_impl;
+};
+
+template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket>
+struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 8, 1> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketConverter(const TensorEvaluator& impl)
+      : m_impl(impl) {}
+
+  template<int LoadMode, typename Index>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TgtPacket packet(Index index) const {
+    const int SrcPacketSize = internal::unpacket_traits<SrcPacket>::size;
+
+    SrcPacket src1 = m_impl.template packet<LoadMode>(index);
+    SrcPacket src2 = m_impl.template packet<LoadMode>(index + 1 * SrcPacketSize);
+    SrcPacket src3 = m_impl.template packet<LoadMode>(index + 2 * SrcPacketSize);
+    SrcPacket src4 = m_impl.template packet<LoadMode>(index + 3 * SrcPacketSize);
+    SrcPacket src5 = m_impl.template packet<LoadMode>(index + 4 * SrcPacketSize);
+    SrcPacket src6 = m_impl.template packet<LoadMode>(index + 5 * SrcPacketSize);
+    SrcPacket src7 = m_impl.template packet<LoadMode>(index + 6 * SrcPacketSize);
+    SrcPacket src8 = m_impl.template packet<LoadMode>(index + 7 * SrcPacketSize);
+    TgtPacket result = internal::pcast<SrcPacket, TgtPacket>(src1, src2, src3, src4, src5, src6, src7, src8);
+    return result;
+  }
+
+ private:
+  const TensorEvaluator& m_impl;
+};
+
+template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket, int TgtCoeffRatio>
+struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 1, TgtCoeffRatio> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketConverter(const TensorEvaluator& impl)
+      : m_impl(impl), m_maxIndex(impl.dimensions().TotalSize()) {}
+
+  template<int LoadMode, typename Index>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TgtPacket packet(Index index) const {
+    const int SrcPacketSize = internal::unpacket_traits<SrcPacket>::size;
+    // Only call m_impl.packet() when we have direct access to the underlying data. This
+    // ensures that we don't compute the subexpression twice. We may however load some
+    // coefficients twice, but in practice this doesn't negatively impact performance.
+    if (m_impl.data() && (index + SrcPacketSize < m_maxIndex)) {
+      // Force unaligned memory loads since we can't ensure alignment anymore
+      return internal::pcast<SrcPacket, TgtPacket>(m_impl.template packet<Unaligned>(index));
+    } else {
+      const int TgtPacketSize = internal::unpacket_traits<TgtPacket>::size;
+      typedef typename internal::unpacket_traits<SrcPacket>::type SrcType;
+      typedef typename internal::unpacket_traits<TgtPacket>::type TgtType;
+      internal::scalar_cast_op<SrcType, TgtType> converter;
+      EIGEN_ALIGN_MAX typename internal::unpacket_traits<TgtPacket>::type values[TgtPacketSize];
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < TgtPacketSize; ++i) {
+        values[i] = converter(m_impl.coeff(index+i));
+      }
+      TgtPacket rslt = internal::pload<TgtPacket>(values);
+      return rslt;
+    }
+  }
+
+ private:
+  const TensorEvaluator& m_impl;
+  const typename TensorEvaluator::Index m_maxIndex;
+};
+
+template<typename TargetType, typename XprType>
+class TensorConversionOp : public TensorBase<TensorConversionOp<TargetType, XprType>, ReadOnlyAccessors>
+{
+  public:
+    typedef typename internal::traits<TensorConversionOp>::Scalar Scalar;
+    typedef typename internal::traits<TensorConversionOp>::StorageKind StorageKind;
+    typedef typename internal::traits<TensorConversionOp>::Index Index;
+    typedef typename internal::nested<TensorConversionOp>::type Nested;
+    typedef Scalar CoeffReturnType;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorConversionOp(const XprType& xpr)
+        : m_xpr(xpr) {}
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+};
+
+template <bool SameType, typename Eval, typename EvalPointerType> struct ConversionSubExprEval {
+  static EIGEN_STRONG_INLINE bool run(Eval& impl, EvalPointerType) {
+    impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+};
+
+template <typename Eval, typename EvalPointerType> struct ConversionSubExprEval<true, Eval, EvalPointerType> {
+  static EIGEN_STRONG_INLINE bool run(Eval& impl, EvalPointerType data) {
+    return impl.evalSubExprsIfNeeded(data);
+  }
+};
+
+#ifdef EIGEN_USE_THREADS
+template <bool SameType, typename Eval, typename EvalPointerType,
+          typename EvalSubExprsCallback>
+struct ConversionSubExprEvalAsync {
+  static EIGEN_STRONG_INLINE void run(Eval& impl, EvalPointerType, EvalSubExprsCallback done) {
+    impl.evalSubExprsIfNeededAsync(nullptr, std::move(done));
+  }
+};
+
+template <typename Eval, typename EvalPointerType,
+          typename EvalSubExprsCallback>
+struct ConversionSubExprEvalAsync<true, Eval, EvalPointerType,
+                                  EvalSubExprsCallback> {
+  static EIGEN_STRONG_INLINE void run(Eval& impl, EvalPointerType data, EvalSubExprsCallback done) {
+    impl.evalSubExprsIfNeededAsync(data, std::move(done));
+  }
+};
+#endif
+
+namespace internal {
+
+template <typename SrcType, typename TargetType, bool IsSameT>
+struct CoeffConv {
+  template <typename ArgType, typename Device>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetType run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
+    internal::scalar_cast_op<SrcType, TargetType> converter;
+    return converter(impl.coeff(index));
+  }
+};
+
+template <typename SrcType, typename TargetType>
+struct CoeffConv<SrcType, TargetType, true> {
+  template <typename ArgType, typename Device>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetType run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
+    return impl.coeff(index);
+  }
+};
+
+template <typename SrcPacket, typename TargetPacket, int LoadMode, bool ActuallyVectorize, bool IsSameT>
+struct PacketConv {
+  typedef typename internal::unpacket_traits<SrcPacket>::type SrcType;
+  typedef typename internal::unpacket_traits<TargetPacket>::type TargetType;
+
+  static const int PacketSize = internal::unpacket_traits<TargetPacket>::size;
+
+  template <typename ArgType, typename Device>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
+    internal::scalar_cast_op<SrcType, TargetType> converter;
+    EIGEN_ALIGN_MAX typename internal::remove_const<TargetType>::type values[PacketSize];
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      values[i] = converter(impl.coeff(index+i));
+    }
+    TargetPacket rslt = internal::pload<TargetPacket>(values);
+    return rslt;
+  }
+};
+
+template <typename SrcPacket, typename TargetPacket, int LoadMode, bool IsSameT>
+struct PacketConv<SrcPacket, TargetPacket, LoadMode, true, IsSameT> {
+  typedef typename internal::unpacket_traits<SrcPacket>::type SrcType;
+  typedef typename internal::unpacket_traits<TargetPacket>::type TargetType;
+
+  template <typename ArgType, typename Device>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
+    const int SrcCoeffRatio = internal::type_casting_traits<SrcType, TargetType>::SrcCoeffRatio;
+    const int TgtCoeffRatio = internal::type_casting_traits<SrcType, TargetType>::TgtCoeffRatio;
+    PacketConverter<TensorEvaluator<ArgType, Device>, SrcPacket, TargetPacket,
+                    SrcCoeffRatio, TgtCoeffRatio> converter(impl);
+    return converter.template packet<LoadMode>(index);
+  }
+};
+
+template <typename SrcPacket, typename TargetPacket, int LoadMode>
+struct PacketConv<SrcPacket, TargetPacket, LoadMode, /*ActuallyVectorize=*/false, /*IsSameT=*/true> {
+  typedef typename internal::unpacket_traits<TargetPacket>::type TargetType;
+  static const int PacketSize = internal::unpacket_traits<TargetPacket>::size;
+
+  template <typename ArgType, typename Device>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
+    EIGEN_ALIGN_MAX typename internal::remove_const<TargetType>::type values[PacketSize];
+    for (int i = 0; i < PacketSize; ++i) values[i] = impl.coeff(index+i);
+    return internal::pload<TargetPacket>(values);
+  }
+};
+
+template <typename SrcPacket, typename TargetPacket, int LoadMode>
+struct PacketConv<SrcPacket, TargetPacket, LoadMode, /*ActuallyVectorize=*/true, /*IsSameT=*/true> {
+  template <typename ArgType, typename Device>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
+    return impl.template packet<LoadMode>(index);
+  }
+};
+
+}  // namespace internal
+
+// Eval as rvalue
+template<typename TargetType, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
+{
+  typedef TensorConversionOp<TargetType, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
+  typedef TargetType Scalar;
+  typedef TargetType CoeffReturnType;
+  typedef typename internal::remove_all<typename internal::traits<ArgType>::Scalar>::type SrcType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef typename PacketType<SrcType, Device>::type PacketSourceType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  static const bool IsSameType = internal::is_same<TargetType, SrcType>::value;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned         = false,
+    PacketAccess      =
+    #ifndef EIGEN_USE_SYCL
+                        true,
+    #else
+                        TensorEvaluator<ArgType, Device>::PacketAccess &
+                        internal::type_casting_traits<SrcType, TargetType>::VectorizedCast,
+    #endif
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess         = false
+  };
+
+  static const int NumDims = internal::array_size<Dimensions>::value;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename TensorEvaluator<const ArgType, Device>::TensorBlock
+      ArgTensorBlock;
+
+  struct TensorConversionOpBlockFactory {
+    template <typename ArgXprType>
+    struct XprType {
+      typedef TensorConversionOp<TargetType, const ArgXprType> type;
+    };
+
+    template <typename ArgXprType>
+    typename XprType<ArgXprType>::type expr(const ArgXprType& expr) const {
+      return typename XprType<ArgXprType>::type(expr);
+    }
+  };
+
+  typedef internal::TensorUnaryExprBlock<TensorConversionOpBlockFactory,
+                                         ArgTensorBlock>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : m_impl(op.expression(), device)
+  {
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_impl.dimensions(); }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data)
+  {
+    return ConversionSubExprEval<IsSameType, TensorEvaluator<ArgType, Device>, EvaluatorPointerType>::run(m_impl, data);
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType data, EvalSubExprsCallback done) {
+    ConversionSubExprEvalAsync<IsSameType, TensorEvaluator<ArgType, Device>,
+                               EvaluatorPointerType,
+        EvalSubExprsCallback>::run(m_impl, data, std::move(done));
+  }
+#endif
+
+  EIGEN_STRONG_INLINE void cleanup()
+  {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return internal::CoeffConv<SrcType, TargetType, IsSameType>::run(m_impl,index);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType
+  packet(Index index) const {
+    // If we are not going to do the cast, we just need to check that base
+    // TensorEvaluator has packet access. Otherwise we also need to make sure,
+    // that we have an implementation of vectorized cast.
+    const bool Vectorizable =
+        IsSameType
+        ? TensorEvaluator<ArgType, Device>::PacketAccess
+        : int(TensorEvaluator<ArgType, Device>::PacketAccess) &
+          int(internal::type_casting_traits<SrcType, TargetType>::VectorizedCast);
+
+    return internal::PacketConv<PacketSourceType, PacketReturnType, LoadMode,
+                                Vectorizable, IsSameType>::run(m_impl, index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    const double cast_cost = TensorOpCost::CastCost<SrcType, TargetType>();
+    if (vectorized) {
+      const double SrcCoeffRatio =
+          internal::type_casting_traits<SrcType, TargetType>::SrcCoeffRatio;
+      const double TgtCoeffRatio =
+          internal::type_casting_traits<SrcType, TargetType>::TgtCoeffRatio;
+      return m_impl.costPerCoeff(vectorized) * (SrcCoeffRatio / PacketSize) +
+          TensorOpCost(0, 0, TgtCoeffRatio * (cast_cost / PacketSize));
+    } else {
+      return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, cast_cost);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    return m_impl.getResourceRequirements();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    return TensorBlock(m_impl.block(desc, scratch),
+                         TensorConversionOpBlockFactory());
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+  /// required by sycl in order to extract the sycl accessor
+  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ protected:
+  TensorEvaluator<ArgType, Device> m_impl;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONVERSION_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorConvolution.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorConvolution.h
new file mode 100644
index 0000000..b20f80b
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorConvolution.h
@@ -0,0 +1,1132 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
+
+namespace Eigen {
+
+/** \class TensorConvolution
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor convolution class.
+  *
+  *
+  */
+namespace internal {
+
+template <typename Index, typename InputDims, int NumKernelDims, int Layout>
+class IndexMapper {
+ public:
+  IndexMapper(const InputDims& input_dims, const array<Index, NumKernelDims>& kernel_dims,
+              const array<Index, NumKernelDims>& indices) {
+
+    array<Index, NumDims> dimensions = input_dims;
+    for (int i = 0; i < NumKernelDims; ++i) {
+      const Index index = indices[i];
+      const Index input_dim = input_dims[index];
+      const Index kernel_dim = kernel_dims[i];
+      const Index result_dim = input_dim - kernel_dim + 1;
+      dimensions[index] = result_dim;
+    }
+
+    array<Index, NumDims> inputStrides;
+    array<Index, NumDims> outputStrides;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      inputStrides[0] = 1;
+      outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        inputStrides[i] = inputStrides[i-1] * input_dims[i-1];
+        outputStrides[i] = outputStrides[i-1] * dimensions[i-1];
+      }
+    } else {
+      inputStrides[NumDims - 1] = 1;
+      outputStrides[NumDims - 1] = 1;
+      for (int i = static_cast<int>(NumDims) - 2; i >= 0; --i) {
+        inputStrides[i] = inputStrides[i + 1] * input_dims[i + 1];
+        outputStrides[i] = outputStrides[i + 1] * dimensions[i + 1];
+      }
+    }
+
+    array<Index, NumDims> gpuInputDimensions;
+    array<Index, NumDims> gpuOutputDimensions;
+    array<Index, NumDims> tmp = dimensions;
+    array<Index, NumDims> ordering;
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    for (int i = 0; i < NumKernelDims; ++i) {
+      const Index index = i + offset;
+      ordering[index] = indices[i];
+      tmp[indices[i]] = -1;
+      gpuInputDimensions[index] = input_dims[indices[i]];
+      gpuOutputDimensions[index] = dimensions[indices[i]];
+    }
+
+    int written = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                      ? NumKernelDims
+                      : 0;
+    for (int i = 0; i < NumDims; ++i) {
+      if (tmp[i] >= 0) {
+        ordering[written] = i;
+        gpuInputDimensions[written] = input_dims[i];
+        gpuOutputDimensions[written] = dimensions[i];
+        ++written;
+      }
+    }
+
+    for (int i = 0; i < NumDims; ++i) {
+      m_inputStrides[i] = inputStrides[ordering[i]];
+      m_outputStrides[i] = outputStrides[ordering[i]];
+    }
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < NumDims; ++i) {
+        if (i > NumKernelDims) {
+          m_gpuInputStrides[i] =
+              m_gpuInputStrides[i - 1] * gpuInputDimensions[i - 1];
+          m_gpuOutputStrides[i] =
+              m_gpuOutputStrides[i - 1] * gpuOutputDimensions[i - 1];
+        } else {
+          m_gpuInputStrides[i] = 1;
+          m_gpuOutputStrides[i] = 1;
+        }
+      }
+    } else {
+      for (int i = NumDims - 1; i >= 0; --i) {
+        if (static_cast<size_t>(i + 1) < offset) {
+          m_gpuInputStrides[i] =
+              m_gpuInputStrides[i + 1] * gpuInputDimensions[i + 1];
+          m_gpuOutputStrides[i] =
+              m_gpuOutputStrides[i + 1] * gpuOutputDimensions[i + 1];
+        } else {
+          m_gpuInputStrides[i] = 1;
+          m_gpuOutputStrides[i] = 1;
+        }
+      }
+    }
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuInputPlaneToTensorInputOffset(Index p) const {
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int d = NumDims - 1; d > NumKernelDims; --d) {
+        const Index idx = p / m_gpuInputStrides[d];
+        inputIndex += idx * m_inputStrides[d];
+        p -= idx * m_gpuInputStrides[d];
+      }
+      inputIndex += p * m_inputStrides[NumKernelDims];
+    } else {
+      std::ptrdiff_t limit = 0;
+      if (NumKernelDims < NumDims) {
+        limit = NumDims - NumKernelDims - 1;
+      }
+      for (int d = 0; d < limit; ++d) {
+        const Index idx = p / m_gpuInputStrides[d];
+        inputIndex += idx * m_inputStrides[d];
+        p -= idx * m_gpuInputStrides[d];
+      }
+      inputIndex += p * m_inputStrides[limit];
+    }
+    return inputIndex;
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuOutputPlaneToTensorOutputOffset(Index p) const {
+    Index outputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int d = NumDims - 1; d > NumKernelDims; --d) {
+        const Index idx = p / m_gpuOutputStrides[d];
+        outputIndex += idx * m_outputStrides[d];
+        p -= idx * m_gpuOutputStrides[d];
+      }
+      outputIndex += p * m_outputStrides[NumKernelDims];
+    } else {
+      std::ptrdiff_t limit = 0;
+      if (NumKernelDims < NumDims) {
+        limit = NumDims - NumKernelDims - 1;
+      }
+      for (int d = 0; d < limit; ++d) {
+        const Index idx = p / m_gpuOutputStrides[d];
+        outputIndex += idx * m_outputStrides[d];
+        p -= idx * m_gpuOutputStrides[d];
+      }
+      outputIndex += p * m_outputStrides[limit];
+    }
+    return outputIndex;
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuInputKernelToTensorInputOffset(Index i) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_inputStrides[offset];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuOutputKernelToTensorOutputOffset(Index i) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_outputStrides[offset];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuInputKernelToTensorInputOffset(Index i, Index j) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuOutputKernelToTensorOutputOffset(Index i, Index j) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_outputStrides[offset] + j * m_outputStrides[offset + 1];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuInputKernelToTensorInputOffset(Index i, Index j, Index k) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1] +
+           k * m_inputStrides[offset + 2];
+  }
+
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapGpuOutputKernelToTensorOutputOffset(Index i, Index j, Index k) const {
+    const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                              ? 0
+                              : NumDims - NumKernelDims;
+    return i * m_outputStrides[offset] + j * m_outputStrides[offset + 1] +
+           k * m_outputStrides[offset + 2];
+  }
+
+ private:
+  static const int NumDims = internal::array_size<InputDims>::value;
+  array<Index, NumDims> m_inputStrides;
+  array<Index, NumDims> m_outputStrides;
+  array<Index, NumDims> m_gpuInputStrides;
+  array<Index, NumDims> m_gpuOutputStrides;
+};
+
+
+
+template<typename Dimensions, typename InputXprType, typename KernelXprType>
+struct traits<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs are different.
+  typedef typename promote_storage_type<typename InputXprType::Scalar,
+                                        typename KernelXprType::Scalar>::ret Scalar;
+  typedef typename promote_storage_type<typename traits<InputXprType>::StorageKind,
+                                        typename traits<KernelXprType>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<InputXprType>::Index,
+                                      typename traits<KernelXprType>::Index>::type Index;
+  typedef typename InputXprType::Nested LhsNested;
+  typedef typename KernelXprType::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+  static const int NumDimensions = traits<InputXprType>::NumDimensions;
+  static const int Layout = traits<InputXprType>::Layout;
+  typedef typename conditional<Pointer_type_promotion<typename InputXprType::Scalar, Scalar>::val,
+  typename traits<InputXprType>::PointerType, typename traits<KernelXprType>::PointerType>::type PointerType;
+
+  enum {
+    Flags = 0
+  };
+};
+
+template<typename Dimensions, typename InputXprType, typename KernelXprType>
+struct eval<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>, Eigen::Dense>
+{
+  typedef const TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>& type;
+};
+
+template<typename Dimensions, typename InputXprType, typename KernelXprType>
+struct nested<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType>, 1, typename eval<TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> >::type>
+{
+  typedef TensorConvolutionOp<Dimensions, InputXprType, KernelXprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename Indices, typename InputXprType, typename KernelXprType>
+class TensorConvolutionOp : public TensorBase<TensorConvolutionOp<Indices, InputXprType, KernelXprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorConvolutionOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename internal::promote_storage_type<typename InputXprType::CoeffReturnType,
+                                                  typename KernelXprType::CoeffReturnType>::ret CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorConvolutionOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorConvolutionOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorConvolutionOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorConvolutionOp(const InputXprType& input, const KernelXprType& kernel, const Indices& dims)
+      : m_input_xpr(input), m_kernel_xpr(kernel), m_indices(dims) {}
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const Indices& indices() const { return m_indices; }
+
+    /** \returns the nested expressions */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const typename internal::remove_all<typename InputXprType::Nested>::type&
+    inputExpression() const { return m_input_xpr; }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const typename internal::remove_all<typename KernelXprType::Nested>::type&
+    kernelExpression() const { return m_kernel_xpr; }
+
+  protected:
+    typename InputXprType::Nested m_input_xpr;
+    typename KernelXprType::Nested m_kernel_xpr;
+    const Indices m_indices;
+};
+
+
+template<typename Indices, typename InputArgType, typename KernelArgType, typename Device>
+struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelArgType>, Device>
+{
+  typedef TensorConvolutionOp<Indices, InputArgType, KernelArgType> XprType;
+
+  static const int NumDims = internal::array_size<typename TensorEvaluator<InputArgType, Device>::Dimensions>::value;
+  static const int NumKernelDims = internal::array_size<Indices>::value;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<Scalar, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = int(TensorEvaluator<InputArgType, Device>::IsAligned) & int(TensorEvaluator<KernelArgType, Device>::IsAligned),
+    PacketAccess = int(TensorEvaluator<InputArgType, Device>::PacketAccess) & int(TensorEvaluator<KernelArgType, Device>::PacketAccess),
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<InputArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device), m_kernelArg(op.kernelExpression()), m_kernel(NULL), m_local_kernel(false), m_device(device)
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<InputArgType, Device>::Layout) == static_cast<int>(TensorEvaluator<KernelArgType, Device>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    const typename TensorEvaluator<InputArgType, Device>::Dimensions& input_dims = m_inputImpl.dimensions();
+    const typename TensorEvaluator<KernelArgType, Device>::Dimensions& kernel_dims = m_kernelImpl.dimensions();
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputStride[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_inputStride[i] = m_inputStride[i - 1] * input_dims[i - 1];
+      }
+    } else {
+      m_inputStride[NumDims - 1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_inputStride[i] = m_inputStride[i + 1] * input_dims[i + 1];
+      }
+    }
+
+    m_dimensions = m_inputImpl.dimensions();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < NumKernelDims; ++i) {
+        const Index index = op.indices()[i];
+        const Index input_dim = input_dims[index];
+        const Index kernel_dim = kernel_dims[i];
+        const Index result_dim = input_dim - kernel_dim + 1;
+        m_dimensions[index] = result_dim;
+        if (i > 0) {
+          m_kernelStride[i] = m_kernelStride[i - 1] * kernel_dims[i - 1];
+        } else {
+          m_kernelStride[0] = 1;
+        }
+        m_indexStride[i] = m_inputStride[index];
+      }
+
+      m_outputStride[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStride[i] = m_outputStride[i - 1] * m_dimensions[i - 1];
+      }
+    } else {
+      for (int i = NumKernelDims - 1; i >= 0; --i) {
+        const Index index = op.indices()[i];
+        const Index input_dim = input_dims[index];
+        const Index kernel_dim = kernel_dims[i];
+        const Index result_dim = input_dim - kernel_dim + 1;
+        m_dimensions[index] = result_dim;
+        if (i < NumKernelDims - 1) {
+          m_kernelStride[i] = m_kernelStride[i + 1] * kernel_dims[i + 1];
+        } else {
+          m_kernelStride[NumKernelDims - 1] = 1;
+        }
+        m_indexStride[i] = m_inputStride[index];
+      }
+
+      m_outputStride[NumDims - 1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_outputStride[i] = m_outputStride[i + 1] * m_dimensions[i + 1];
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) {
+    m_inputImpl.evalSubExprsIfNeeded(NULL);
+    preloadKernel();
+    return true;
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_inputImpl.cleanup();
+    if (m_local_kernel) {
+      m_device.deallocate((void*)m_kernel);
+      m_local_kernel = false;
+    }
+    m_kernel = NULL;
+  }
+
+  void evalTo(typename XprType::Scalar* buffer) {
+    evalSubExprsIfNeeded(NULL);
+    for (int i = 0; i < dimensions().TotalSize(); ++i) {
+      buffer[i] += coeff(i);
+    }
+    cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    CoeffReturnType result = CoeffReturnType(0);
+    convolve(firstInput(index), 0, NumKernelDims-1, result);
+    return result;
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(const Index index) const
+  {
+    Index indices[2] = {index, index+PacketSize-1};
+    Index startInputs[2] = {0, 0};
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx0 = indices[0] / m_outputStride[i];
+        const Index idx1 = indices[1] / m_outputStride[i];
+        startInputs[0] += idx0 * m_inputStride[i];
+        startInputs[1] += idx1 * m_inputStride[i];
+        indices[0] -= idx0 * m_outputStride[i];
+        indices[1] -= idx1 * m_outputStride[i];
+      }
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx0 = indices[0] / m_outputStride[i];
+        const Index idx1 = indices[1] / m_outputStride[i];
+        startInputs[0] += idx0 * m_inputStride[i];
+        startInputs[1] += idx1 * m_inputStride[i];
+        indices[0] -= idx0 * m_outputStride[i];
+        indices[1] -= idx1 * m_outputStride[i];
+      }
+    }
+    startInputs[0] += indices[0];
+    startInputs[1] += indices[1];
+
+    if (startInputs[1]-startInputs[0] == PacketSize-1) {
+      PacketReturnType result = internal::pset1<PacketReturnType>(0);
+      convolvePacket(startInputs[0], 0, NumKernelDims-1, result);
+      return result;
+    } else {
+      EIGEN_ALIGN_MAX Scalar data[PacketSize];
+      data[0] = Scalar(0);
+      convolve(startInputs[0], 0, NumKernelDims-1, data[0]);
+      for (int i = 1; i < PacketSize-1; ++i) {
+        data[i] = Scalar(0);
+        convolve(firstInput(index+i), 0, NumKernelDims-1, data[i]);
+      }
+      data[PacketSize-1] = Scalar(0);
+      convolve(startInputs[1], 0, NumKernelDims-1, data[PacketSize-1]);
+      return internal::pload<PacketReturnType>(data);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    const double kernel_size = m_kernelImpl.dimensions().TotalSize();
+    // We ignore the use of fused multiply-add.
+    const double convolve_compute_cost =
+        TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
+    const double firstIndex_compute_cost =
+        NumDims *
+        (2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() +
+         TensorOpCost::DivCost<Index>());
+    return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
+           kernel_size * (m_inputImpl.costPerCoeff(vectorized) +
+                          m_kernelImpl.costPerCoeff(vectorized) +
+                          TensorOpCost(0, 0, convolve_compute_cost, vectorized,
+                                       PacketSize));
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+ private:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const {
+    Index startInput = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_outputStride[i];
+        startInput += idx * m_inputStride[i];
+        index -= idx * m_outputStride[i];
+      }
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_outputStride[i];
+        startInput += idx * m_inputStride[i];
+        index -= idx * m_outputStride[i];
+      }
+    }
+    startInput += index;
+    return startInput;
+  }
+
+  EIGEN_DEVICE_FUNC void convolve(Index firstIndex, Index firstKernel, int DimIndex, CoeffReturnType& accum) const {
+    for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) {
+      const Index input = firstIndex + j * m_indexStride[DimIndex];
+      const Index kernel = firstKernel + j * m_kernelStride[DimIndex];
+      if (DimIndex > 0) {
+        convolve(input, kernel, DimIndex-1, accum);
+      } else {
+        accum += m_inputImpl.coeff(input) * m_kernel[kernel];
+      }
+    }
+  }
+
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC void convolvePacket(Index firstIndex, Index firstKernel, int DimIndex, Packet& accum) const {
+    for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) {
+      const Index input = firstIndex + j * m_indexStride[DimIndex];
+      const Index kernel = firstKernel + j * m_kernelStride[DimIndex];
+      if (DimIndex > 0) {
+        convolvePacket(input, kernel, DimIndex-1, accum);
+      } else {
+        accum = internal::pmadd<Packet>(m_inputImpl.template packet<Unaligned>(input), internal::pset1<Packet>(m_kernel[kernel]), accum);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void preloadKernel() {
+    // Don't make a local copy of the kernel unless we have to (i.e. it's an
+    // expression that needs to be evaluated)
+    const Scalar* in_place = m_kernelImpl.data();
+    if (in_place) {
+      m_kernel = in_place;
+      m_local_kernel = false;
+    } else {
+      size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
+      Scalar* local = (Scalar*)m_device.allocate_temp(kernel_sz);
+      typedef TensorEvalToOp<const KernelArgType> EvalTo;
+      EvalTo evalToTmp(local, m_kernelArg);
+      const bool Vectorize = internal::IsVectorizable<Device, KernelArgType>::value;
+      internal::TensorExecutor<const EvalTo, Device, Vectorize>::run(evalToTmp, m_device);
+
+      m_kernel = local;
+      m_local_kernel = true;
+    }
+  }
+
+  array<Index, NumDims> m_inputStride;
+  array<Index, NumDims> m_outputStride;
+
+  array<Index, NumKernelDims> m_indexStride;
+  array<Index, NumKernelDims> m_kernelStride;
+  TensorEvaluator<InputArgType, Device> m_inputImpl;
+  TensorEvaluator<KernelArgType, Device> m_kernelImpl;
+  Dimensions m_dimensions;
+
+  KernelArgType m_kernelArg;
+  const Scalar* m_kernel;
+  bool m_local_kernel;
+  const Device EIGEN_DEVICE_REF m_device;
+};
+
+
+
+
+// Use an optimized implementation of the evaluation code for GPUs whenever possible.
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
+
+template <int StaticKernelSize>
+struct GetKernelSize {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int /*kernelSize*/) const {
+    return StaticKernelSize;
+  }
+};
+template <>
+struct GetKernelSize<Dynamic> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int kernelSize) const {
+    return kernelSize;
+  }
+};
+
+template <typename InputEvaluator, typename Index, typename InputDims,
+          int StaticKernelSize>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void EigenConvolutionKernel1D(
+    InputEvaluator eval,
+    const internal::IndexMapper<Index, InputDims, 1, InputEvaluator::Layout>
+        indexMapper,
+    const float* __restrict kernel, const int numPlanes, const int numX,
+    const int maxX, const int kernelSize, float* buffer) {
+#if defined(EIGEN_HIPCC)
+  HIP_DYNAMIC_SHARED(float, s)
+#else
+  extern __shared__ float s[];
+#endif
+
+  const int first_x = blockIdx.x * maxX;
+  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
+  const int num_x_input = last_x - first_x + GetKernelSize<StaticKernelSize>()(kernelSize);
+  const int num_x_output = last_x - first_x + 1;
+
+  const int first_plane = blockIdx.y * blockDim.y;
+  const int plane_stride = blockDim.y * gridDim.y;
+
+  for (int p = first_plane + threadIdx.y; p < numPlanes; p += plane_stride) {
+    // Load inputs to shared memory
+    const int plane_input_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(p);
+    const int plane_kernel_offset = threadIdx.y * num_x_input;
+    #pragma unroll
+    for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) {
+      const int tensor_index = plane_input_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(i+first_x);
+      s[i + plane_kernel_offset] = eval.coeff(tensor_index);
+    }
+
+    __syncthreads();
+
+    // Compute the convolution
+    const int plane_output_offset = indexMapper.mapGpuOutputPlaneToTensorOutputOffset(p);
+
+    #pragma unroll
+    for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) {
+      const int kernel_offset = plane_kernel_offset + i;
+      float result = 0.0f;
+      #pragma unroll
+      for (int k = 0; k < GetKernelSize<StaticKernelSize>()(kernelSize); ++k) {
+        result += s[k + kernel_offset] * kernel[k];
+      }
+      const int tensor_index = plane_output_offset + indexMapper.mapGpuOutputKernelToTensorOutputOffset(i+first_x);
+      buffer[tensor_index] = result;
+    }
+    __syncthreads();
+  }
+};
+
+template <typename InputEvaluator, typename Index, typename InputDims,
+          int StaticKernelSizeX, int StaticKernelSizeY>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void EigenConvolutionKernel2D(
+    InputEvaluator eval,
+    const internal::IndexMapper<Index, InputDims, 2, InputEvaluator::Layout>
+        indexMapper,
+    const float* __restrict kernel, const int numPlanes, const int numX,
+    const int maxX, const int numY, const int maxY, const int kernelSizeX,
+    const int kernelSizeY, float* buffer) {
+#if defined(EIGEN_HIPCC)
+  HIP_DYNAMIC_SHARED(float, s)
+#else
+  extern __shared__ float s[];
+#endif
+
+  const int first_x = blockIdx.x * maxX;
+  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
+  const int num_x_input = last_x - first_x + GetKernelSize<StaticKernelSizeX>()(kernelSizeX);
+  const int num_x_output = last_x - first_x + 1;
+
+  const int first_y = blockIdx.y * maxY;
+  const int last_y = (first_y + maxY < numY ? first_y + maxY : numY) - 1;
+  const int num_y_input = last_y - first_y + GetKernelSize<StaticKernelSizeY>()(kernelSizeY);
+  const int num_y_output = last_y - first_y + 1;
+
+  const int first_plane = blockIdx.z * blockDim.z;
+  const int plane_stride = blockDim.z * gridDim.z;
+
+  for (int p = first_plane + threadIdx.z; p < numPlanes; p += plane_stride) {
+
+    const int plane_input_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(p);
+    const int plane_kernel_offset = threadIdx.z * num_y_input;
+
+    // Load inputs to shared memory
+    #pragma unroll
+    for (int j = threadIdx.y; j < num_y_input; j += blockDim.y) {
+      const int input_offset = num_x_input * (j + plane_kernel_offset);
+      #pragma unroll
+      for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) {
+        const int tensor_index = plane_input_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(i+first_x, j+first_y);
+        s[i + input_offset] = eval.coeff(tensor_index);
+      }
+    }
+
+    __syncthreads();
+
+    // Convolution
+    const int plane_output_offset = indexMapper.mapGpuOutputPlaneToTensorOutputOffset(p);
+
+    #pragma unroll
+    for (int j = threadIdx.y; j < num_y_output; j += blockDim.y) {
+      #pragma unroll
+      for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) {
+        float result = 0.0f;
+        #pragma unroll
+        for (int l = 0; l < GetKernelSize<StaticKernelSizeY>()(kernelSizeY); ++l) {
+          const int kernel_offset = kernelSizeX * l;
+          const int input_offset = i + num_x_input * (j + l + plane_kernel_offset);
+          #pragma unroll
+          for (int k = 0; k < GetKernelSize<StaticKernelSizeX>()(kernelSizeX); ++k) {
+            result += s[k + input_offset] * kernel[k + kernel_offset];
+          }
+        }
+        const int tensor_index = plane_output_offset + indexMapper.mapGpuOutputKernelToTensorOutputOffset(i+first_x, j+first_y);
+        buffer[tensor_index] = result;
+      }
+    }
+
+    __syncthreads();
+  }
+};
+
+template <typename InputEvaluator, typename Index, typename InputDims>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void EigenConvolutionKernel3D(
+    InputEvaluator eval,
+    const internal::IndexMapper<Index, InputDims, 3, InputEvaluator::Layout>
+        indexMapper,
+    const float* __restrict kernel, const size_t numPlanes, const size_t numX,
+    const size_t maxX, const size_t numY, const size_t maxY, const size_t numZ,
+    const size_t maxZ, const size_t kernelSizeX, const size_t kernelSizeY,
+    const size_t kernelSizeZ, float* buffer) {
+#if defined(EIGEN_HIPCC)
+  HIP_DYNAMIC_SHARED(float, s)
+#else
+  extern __shared__ float s[];
+#endif
+
+  // Load inputs to shared memory
+  const int first_x = blockIdx.x * maxX;
+  const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1;
+  const int num_x_input = last_x - first_x + kernelSizeX;
+
+  const int first_y = blockIdx.y * maxY;
+  const int last_y = (first_y + maxY < numY ? first_y + maxY : numY) - 1;
+  const int num_y_input = last_y - first_y + kernelSizeY;
+
+  const int first_z = blockIdx.z * maxZ;
+  const int last_z = (first_z + maxZ < numZ ? first_z + maxZ : numZ) - 1;
+  const int num_z_input = last_z - first_z + kernelSizeZ;
+
+  for (int p = 0; p < numPlanes; ++p) {
+
+    const int plane_input_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(p);
+    const int plane_kernel_offset = 0;
+
+    for (int k = threadIdx.z; k < num_z_input; k += blockDim.z) {
+      for (int j = threadIdx.y; j < num_y_input; j += blockDim.y) {
+        for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) {
+          const int tensor_index = plane_input_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(i+first_x, j+first_y, k+first_z);
+          s[i + num_x_input * (j + num_y_input * (k + plane_kernel_offset))] = eval.coeff(tensor_index);
+        }
+      }
+    }
+
+    __syncthreads();
+
+    // Convolution
+    const int num_z_output = last_z - first_z + 1;
+    const int num_y_output = last_y - first_y + 1;
+    const int num_x_output = last_x - first_x + 1;
+    const int plane_output_offset = indexMapper.mapGpuOutputPlaneToTensorOutputOffset(p);
+
+    for (int k = threadIdx.z; k < num_z_output; k += blockDim.z) {
+      for (int j = threadIdx.y; j < num_y_output; j += blockDim.y) {
+        for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) {
+          float result = 0.0f;
+          for (int n = 0; n < kernelSizeZ; ++n) {
+            for (int m = 0; m < kernelSizeY; ++m) {
+              for (int l = 0; l < kernelSizeX; ++l) {
+                result += s[i + l + num_x_input * (j + m + num_y_input * (k + n + plane_kernel_offset))] * kernel[l + kernelSizeX * (m + kernelSizeY * n)];
+              }
+            }
+          }
+          const int tensor_index = plane_output_offset + indexMapper.mapGpuOutputKernelToTensorOutputOffset(i+first_x, j+first_y, k+first_z);
+          buffer[tensor_index] = result;
+        }
+      }
+    }
+    __syncthreads();
+  }
+};
+
+
+
+template<typename Indices, typename InputArgType, typename KernelArgType>
+struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelArgType>, GpuDevice>
+{
+  typedef TensorConvolutionOp<Indices, InputArgType, KernelArgType> XprType;
+
+  static const int NumDims =  internal::array_size<typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions>::value;
+  static const int NumKernelDims = internal::array_size<Indices>::value;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename TensorEvaluator<KernelArgType, GpuDevice>::Dimensions KernelDimensions;
+
+  enum {
+    IsAligned = TensorEvaluator<InputArgType, GpuDevice>::IsAligned & TensorEvaluator<KernelArgType, GpuDevice>::IsAligned,
+    PacketAccess = false,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<InputArgType, GpuDevice>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  TensorEvaluator(const XprType& op, const GpuDevice& device)
+      : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device), m_kernelArg(op.kernelExpression()), m_indices(op.indices()), m_buf(NULL), m_kernel(NULL), m_local_kernel(false), m_device(device)
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<InputArgType, GpuDevice>::Layout) == static_cast<int>(TensorEvaluator<KernelArgType, GpuDevice>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    const typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions& input_dims = m_inputImpl.dimensions();
+    const typename TensorEvaluator<KernelArgType, GpuDevice>::Dimensions& kernel_dims = m_kernelImpl.dimensions();
+
+    m_dimensions = m_inputImpl.dimensions();
+    for (int i = 0; i < NumKernelDims; ++i) {
+      const Index index = op.indices()[i];
+      const Index input_dim = input_dims[index];
+      const Index kernel_dim = kernel_dims[i];
+      const Index result_dim = input_dim - kernel_dim + 1;
+      m_dimensions[index] = result_dim;
+    }
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, GpuDevice>::type PacketReturnType;
+  typedef typename InputArgType::Scalar Scalar;
+  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
+    preloadKernel();
+    m_inputImpl.evalSubExprsIfNeeded(NULL);
+    if (data) {
+      executeEval(data);
+      return false;
+    } else {
+      m_buf = (Scalar*)m_device.allocate(dimensions().TotalSize() * sizeof(Scalar));
+      executeEval(m_buf);
+      return true;
+    }
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_inputImpl.cleanup();
+    if (m_buf) {
+      m_device.deallocate(m_buf);
+      m_buf = NULL;
+    }
+    if (m_local_kernel) {
+      m_device.deallocate((void*)m_kernel);
+      m_local_kernel = false;
+    }
+    m_kernel = NULL;
+  }
+
+  EIGEN_STRONG_INLINE void preloadKernel() {
+    // Don't make a local copy of the kernel unless we have to (i.e. it's an
+    // expression that needs to be evaluated)
+    const Scalar* in_place = m_kernelImpl.data();
+    if (in_place) {
+      m_kernel = in_place;
+      m_local_kernel = false;
+    } else {
+      size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
+      Scalar* local = (Scalar*)m_device.allocate(kernel_sz);
+      typedef TensorEvalToOp<const KernelArgType> EvalTo;
+      EvalTo evalToTmp(local, m_kernelArg);
+      const bool PacketAccess = internal::IsVectorizable<GpuDevice, KernelArgType>::value;
+      internal::TensorExecutor<const EvalTo, GpuDevice, PacketAccess>::run(evalToTmp, m_device);
+
+      m_kernel = local;
+      m_local_kernel = true;
+    }
+  }
+
+  static unsigned int ceil(unsigned int num, unsigned int denom) {
+    const unsigned int rounded_toward_zero = num / denom;
+    if (num > rounded_toward_zero * denom) {
+      return rounded_toward_zero + 1;
+    }
+    return rounded_toward_zero;
+  }
+
+  void executeEval(Scalar* data) const {
+    typedef typename TensorEvaluator<InputArgType, GpuDevice>::Dimensions InputDims;
+
+    const int maxSharedMem = m_device.sharedMemPerBlock();
+    const int maxThreadsPerBlock = m_device.maxGpuThreadsPerBlock();
+    const int maxBlocksPerProcessor = m_device.maxGpuThreadsPerMultiProcessor() / maxThreadsPerBlock;
+    const int numMultiProcessors = m_device.getNumGpuMultiProcessors();
+    const int warpSize = 32;
+
+    switch (NumKernelDims) {
+      case 1: {
+        const int kernel_size = m_kernelImpl.dimensions().TotalSize();
+
+        const int numX = dimensions()[m_indices[0]];
+        const int numP = dimensions().TotalSize() / numX;
+        int maxX;
+        dim3 block_size;
+
+        const int single_stride_dim =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor)
+                ? 0
+                : m_inputImpl.dimensions().rank() - 1;
+        if (m_indices[0] == single_stride_dim) {
+          // Maximum the reuse
+          const int inner_dim = ((maxSharedMem / (sizeof(Scalar)) - kernel_size + 1 + 31) / 32) * 32;
+          maxX = numext::mini<int>(inner_dim, numX);
+          const int maxP = numext::mini<int>(maxSharedMem / ((kernel_size - 1 + maxX) * sizeof(Scalar)), numP);
+          block_size.x = numext::mini(maxThreadsPerBlock, maxX);
+          block_size.y = numext::mini<int>(maxThreadsPerBlock / block_size.x, maxP);
+        }
+        else {
+          // Read as much as possible alongside the inner most dimension, that is the plane
+          const int inner_dim = maxSharedMem / ((warpSize + kernel_size) * sizeof(Scalar));
+          const int maxP = numext::mini<int>(inner_dim, numP);
+          maxX = numext::mini<int>(maxSharedMem / (inner_dim * sizeof(Scalar)) - kernel_size + 1, numX);
+
+          block_size.x = numext::mini(warpSize, maxX);
+          block_size.y = numext::mini<int>(maxThreadsPerBlock/block_size.x, maxP);
+        }
+
+        const int shared_mem = block_size.y * (maxX + kernel_size - 1) * sizeof(Scalar);
+        gpu_assert(shared_mem <= maxSharedMem);
+
+        const int num_x_blocks = ceil(numX, maxX);
+        const int blocksPerProcessor = numext::mini(maxBlocksPerProcessor, maxSharedMem / shared_mem);
+        const int num_y_blocks = ceil(numMultiProcessors * blocksPerProcessor, num_x_blocks);
+
+        dim3 num_blocks(num_x_blocks, numext::mini<int>(num_y_blocks, ceil(numP, block_size.y)));
+
+
+        //cout << "launching 1D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " maxX: " << maxX << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
+
+        const array<Index, 1> indices(m_indices[0]);
+        const array<Index, 1> kernel_dims(m_kernelImpl.dimensions()[0]);
+        internal::IndexMapper<Index, InputDims, 1, Layout> indexMapper(
+            m_inputImpl.dimensions(), kernel_dims, indices);
+        switch(kernel_size) {
+          case 4: {
+            LAUNCH_GPU_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 4, data);
+            break;
+          }
+          case 7: {
+            LAUNCH_GPU_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 7, data);
+            break;
+          }
+          default: {
+            LAUNCH_GPU_KERNEL((EigenConvolutionKernel1D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, kernel_size, data);
+          }
+        }
+        break;
+      }
+
+      case 2: {
+        const int idxX =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : 1;
+        const int idxY =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 1 : 0;
+        const int kernel_size_x = m_kernelImpl.dimensions()[idxX];
+        const int kernel_size_y = m_kernelImpl.dimensions()[idxY];
+
+        const int numX = dimensions()[m_indices[idxX]];
+        const int numY = dimensions()[m_indices[idxY]];
+        const int numP = dimensions().TotalSize() / (numX*numY);
+
+        const float scaling_factor = sqrtf(static_cast<float>(maxSharedMem) / (sizeof(Scalar) * kernel_size_y * kernel_size_x));
+
+        // Snap maxX to warp size
+        int inner_dim = ((static_cast<int>(scaling_factor * kernel_size_x) - kernel_size_x + 1 + 32) / 32) * 32;
+        const int maxX = numext::mini<int>(inner_dim, numX);
+        const int maxY = numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1)) - kernel_size_y + 1, numY);
+        const int maxP = numext::mini<int>(maxSharedMem / ((kernel_size_x - 1 + maxX) * (kernel_size_y - 1 + maxY) * sizeof(Scalar)), numP);
+
+        dim3 block_size;
+        block_size.x = numext::mini(1024, maxX);
+        block_size.y = numext::mini<int>(1024/block_size.x, maxY);
+        block_size.z = numext::mini<int>(1024/(block_size.x*block_size.y), maxP);
+
+        const int shared_mem = block_size.z * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * sizeof(Scalar);
+        gpu_assert(shared_mem <= maxSharedMem);
+
+        const int num_x_blocks = ceil(numX, maxX);
+        const int num_y_blocks = ceil(numY, maxY);
+        const int blocksPerProcessor = numext::mini(maxBlocksPerProcessor, maxSharedMem / shared_mem);
+        const int num_z_blocks = ceil(numMultiProcessors * blocksPerProcessor, num_x_blocks * num_y_blocks);
+
+        dim3 num_blocks(num_x_blocks, num_y_blocks, numext::mini<int>(num_z_blocks, ceil(numP, block_size.z)));
+
+
+        //cout << "launching 2D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y  << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z << " maxX: " << maxX << " maxY: " << maxY << " maxP: " << maxP << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
+
+        const array<Index, 2> indices(m_indices[idxX], m_indices[idxY]);
+        const array<Index, 2> kernel_dims(m_kernelImpl.dimensions()[idxX],
+                                          m_kernelImpl.dimensions()[idxY]);
+        internal::IndexMapper<Index, InputDims, 2, Layout> indexMapper(
+            m_inputImpl.dimensions(), kernel_dims, indices);
+        switch (kernel_size_x) {
+          case 4: {
+            switch (kernel_size_y) {
+              case 7: {
+                LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, 7>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, 7, data);
+                break;
+              }
+              default: {
+                LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 4, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, kernel_size_y, data);
+                break;
+              }
+            }
+            break;
+          }
+          case 7: {
+            switch (kernel_size_y) {
+              case 4: {
+                LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, 4>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, 4, data);
+                break;
+              }
+              default: {
+                LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, 7, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, kernel_size_y, data);
+                break;
+              }
+            }
+            break;
+          }
+          default: {
+            LAUNCH_GPU_KERNEL((EigenConvolutionKernel2D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims, Dynamic, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, kernel_size_x, kernel_size_y, data);
+            break;
+          }
+        }
+        break;
+      }
+
+      case 3: {
+        const int idxX =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : 2;
+        const int idxY =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 1 : 1;
+        const int idxZ =
+            static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 2 : 0;
+
+        const int kernel_size_x = m_kernelImpl.dimensions()[idxX];
+        const int kernel_size_y = m_kernelImpl.dimensions()[idxY];
+        const int kernel_size_z = m_kernelImpl.dimensions()[idxZ];
+
+        const int numX = dimensions()[m_indices[idxX]];
+        const int numY = dimensions()[m_indices[idxY]];
+        const int numZ = dimensions()[m_indices[idxZ]];
+        const int numP = dimensions().TotalSize() / (numX*numY*numZ);
+
+        const int maxX = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * kernel_size_y * kernel_size_z) - kernel_size_x + 1, numX));
+        const int maxY = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1) * kernel_size_z) - kernel_size_y + 1, numY));
+        const int maxZ = numext::mini<int>(128, numext::mini<int>(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1)) - kernel_size_z + 1, numZ));
+
+        dim3 block_size;
+        block_size.x = numext::mini(32, maxX);
+        block_size.y = numext::mini(32, maxY);
+        block_size.z = numext::mini<int>(1024/(block_size.x*block_size.y), maxZ);
+        dim3 num_blocks(ceil(numX, maxX), ceil(numY, maxY), ceil(numZ, maxZ));
+
+        const int shared_mem = (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * (maxZ + kernel_size_z - 1) * sizeof(Scalar);
+        gpu_assert(shared_mem <= maxSharedMem);
+
+        //cout << "launching 3D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y  << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z  << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl;
+        const array<Index, 3> indices(m_indices[idxX], m_indices[idxY],
+                                      m_indices[idxZ]);
+        const array<Index, 3> kernel_dims(m_kernelImpl.dimensions()[idxX],
+                                          m_kernelImpl.dimensions()[idxY],
+                                          m_kernelImpl.dimensions()[idxZ]);
+        internal::IndexMapper<Index, InputDims, 3, Layout> indexMapper(
+            m_inputImpl.dimensions(), kernel_dims, indices);
+
+        LAUNCH_GPU_KERNEL((EigenConvolutionKernel3D<TensorEvaluator<InputArgType, GpuDevice>, Index, InputDims>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, numZ, maxZ, kernel_size_x, kernel_size_y, kernel_size_z, data);
+        break;
+      }
+
+      default: {
+        EIGEN_STATIC_ASSERT((NumKernelDims >= 1 && NumKernelDims <= 3), THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    eigen_assert(m_buf);
+    eigen_assert(index < m_dimensions.TotalSize());
+    return m_buf[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(const Index index) const
+  {
+    eigen_assert(m_buf);
+    eigen_assert(index < m_dimensions.TotalSize());
+    return internal::ploadt<PacketReturnType, LoadMode>(m_buf+index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    // TODO(rmlarsen): FIXME: For now, this is just a copy of the CPU cost
+    // model.
+    const double kernel_size = m_kernelImpl.dimensions().TotalSize();
+    // We ignore the use of fused multiply-add.
+    const double convolve_compute_cost =
+        TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
+    const double firstIndex_compute_cost =
+        NumDims *
+        (2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() +
+         TensorOpCost::DivCost<Index>());
+    return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
+           kernel_size * (m_inputImpl.costPerCoeff(vectorized) +
+                          m_kernelImpl.costPerCoeff(vectorized) +
+                          TensorOpCost(0, 0, convolve_compute_cost, vectorized,
+                                       PacketSize));
+  }
+
+ private:
+  // No assignment (copies are needed by the kernels)
+  TensorEvaluator& operator = (const TensorEvaluator&);
+
+  TensorEvaluator<InputArgType, GpuDevice> m_inputImpl;
+  TensorEvaluator<KernelArgType, GpuDevice> m_kernelImpl;
+  KernelArgType m_kernelArg;
+  Indices m_indices;
+  Dimensions m_dimensions;
+  Scalar* m_buf;
+  const Scalar* m_kernel;
+  bool m_local_kernel;
+
+  const GpuDevice& m_device;
+};
+#endif
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorConvolutionSycl.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorConvolutionSycl.h
new file mode 100644
index 0000000..033318f
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorConvolutionSycl.h
@@ -0,0 +1,544 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_SYCL_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_SYCL_H
+
+namespace Eigen {
+
+/** \class TensorConvolution
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor convolution class.
+ *
+ *
+ */
+
+enum class convolution_type { CONV1D, CONV2D, CONV3D };
+template <typename Evaluator, typename CoeffReturnType, typename KernelType, typename Index, typename InputDims,
+          typename Kernel_accessor, typename Buffer_accessor, convolution_type Conv_Dim>
+struct EigenConvolutionKernel;
+template <typename Evaluator, typename CoeffReturnType, typename KernelType, typename Index, typename InputDims,
+          typename Kernel_accessor, typename Buffer_accessor>
+struct EigenConvolutionKernel<Evaluator, CoeffReturnType, KernelType, Index, InputDims, Kernel_accessor,
+                              Buffer_accessor, convolution_type::CONV1D> {
+  typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      Local_accessor;
+  Local_accessor local_acc;
+  Evaluator device_evaluator;
+  Kernel_accessor kernel_filter;
+  Buffer_accessor buffer_acc;
+  internal::IndexMapper<Index, InputDims, 1, Evaluator::Layout> indexMapper;
+  const size_t kernelSize;
+  const cl::sycl::range<2> input_range;
+  EigenConvolutionKernel(Local_accessor local_acc_, Evaluator device_evaluator_, Kernel_accessor kernel_filter_,
+                         Buffer_accessor buffer_acc_,
+                         internal::IndexMapper<Index, InputDims, 1, Evaluator::Layout> indexMapper_,
+                         const size_t kernelSize_, const cl::sycl::range<2> input_range_)
+      : local_acc(local_acc_),
+        device_evaluator(device_evaluator_),
+        kernel_filter(kernel_filter_),
+        buffer_acc(buffer_acc_),
+        indexMapper(indexMapper_),
+        kernelSize(kernelSize_),
+        input_range(input_range_) {}
+
+  template <typename BooleanDim2>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool boundary_check(const BooleanDim2 boolean_check) {
+    return (boolean_check[0] && boolean_check[1]);
+  }
+  void operator()(cl::sycl::nd_item<2> itemID) {
+    auto buffer_ptr = buffer_acc.get_pointer();
+    auto kernel_ptr = kernel_filter.get_pointer();
+    // the required row to be calculated for the for each plane in shered memory
+    const size_t num_input = (itemID.get_local_range()[0] + kernelSize - 1);
+    const size_t plane_kernel_offset = itemID.get_local_id(1) * num_input;
+    const size_t input_offset = itemID.get_group(0) * itemID.get_local_range()[0];
+    const size_t plane_tensor_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(itemID.get_global_id(1));
+    /// fill the shared memory
+    for (size_t i = itemID.get_local_id(0); i < num_input; i += itemID.get_local_range()[0]) {
+      const size_t local_index = i + plane_kernel_offset;
+      const size_t tensor_index =
+          plane_tensor_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(i + input_offset);
+
+      local_acc[local_index] =
+          (((i + input_offset) < (input_range[0] + kernelSize - 1)) && itemID.get_global_id(1) < input_range[1])
+              ? device_evaluator.coeff(tensor_index)
+              : CoeffReturnType(0);
+    }
+
+    itemID.barrier(cl::sycl::access::fence_space::local_space);
+
+    // calculate the convolution // output start x
+    const size_t first_output_start = itemID.get_group(0) * (itemID.get_local_range()[0]);
+    if (boundary_check(itemID.get_global_id() < input_range)) {
+      CoeffReturnType result = static_cast<CoeffReturnType>(0);
+      const size_t index = plane_kernel_offset + itemID.get_local_id(0);
+      for (size_t k = 0; k < kernelSize; ++k) {
+        result += (local_acc[k + index] * kernel_ptr[k]);
+      }
+      const size_t tensor_index =
+          indexMapper.mapGpuOutputPlaneToTensorOutputOffset(itemID.get_global_id(1)) +
+          indexMapper.mapGpuOutputKernelToTensorOutputOffset(itemID.get_local_id(0) + first_output_start);
+      buffer_ptr[tensor_index] = result;
+    }
+  }
+};
+
+template <typename Evaluator, typename CoeffReturnType, typename KernelType, typename Index, typename InputDims,
+          typename Kernel_accessor, typename Buffer_accessor>
+struct EigenConvolutionKernel<Evaluator, CoeffReturnType, KernelType, Index, InputDims, Kernel_accessor,
+                              Buffer_accessor, convolution_type::CONV2D> {
+  typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      Local_accessor;
+  Local_accessor local_acc;
+  Evaluator device_evaluator;
+  Kernel_accessor kernel_filter;
+  Buffer_accessor buffer_acc;
+  internal::IndexMapper<Index, InputDims, 2, Evaluator::Layout> indexMapper;
+  const cl::sycl::range<2> kernel_size;
+  const cl::sycl::range<3> input_range;
+  EigenConvolutionKernel(Local_accessor local_acc_, Evaluator device_evaluator_, Kernel_accessor kernel_filter_,
+                         Buffer_accessor buffer_acc_,
+                         internal::IndexMapper<Index, InputDims, 2, Evaluator::Layout> indexMapper_,
+                         const cl::sycl::range<2> kernel_size_, const cl::sycl::range<3> input_range_)
+      : local_acc(local_acc_),
+        device_evaluator(device_evaluator_),
+        kernel_filter(kernel_filter_),
+        buffer_acc(buffer_acc_),
+        indexMapper(indexMapper_),
+        kernel_size(kernel_size_),
+        input_range(input_range_) {}
+  template <typename BooleanDim3>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool boundary_check(const BooleanDim3 boolean_check) {
+    return (boolean_check[0] && boolean_check[1] && boolean_check[2]);
+  }
+
+  void operator()(cl::sycl::nd_item<3> itemID) {
+    auto buffer_ptr = buffer_acc.get_pointer();
+    auto kernel_ptr = kernel_filter.get_pointer();
+    // the required row to be calculated for the for each plane in shered memory
+    const auto num_input = cl::sycl::range<2>{
+        (cl::sycl::range<2>(itemID.get_local_range()[0], itemID.get_local_range()[1]) + kernel_size - 1)};
+
+    const size_t plane_input_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(itemID.get_global_id(2));
+    const size_t plane_kernel_offset = itemID.get_local_id(2) * num_input[1];
+
+    const auto input_offset = cl::sycl::range<2>{itemID.get_group(0) * itemID.get_local_range()[0],
+                                                 itemID.get_group(1) * itemID.get_local_range()[1]};
+      
+    // fill the local memory
+    bool in_range_dim2 = itemID.get_global_id(2) < input_range[2];
+    for (size_t j = itemID.get_local_id(1); j < num_input[1]; j += itemID.get_local_range()[1]) {
+      const size_t local_input_offset = num_input[0] * (j + plane_kernel_offset);
+      bool in_range_dim1 = ((j + input_offset[1]) < (input_range[1] + kernel_size[1] - 1)); 
+      for (size_t i = itemID.get_local_id(0); i < num_input[0]; i += itemID.get_local_range()[0]) {
+        const size_t local_index = i + local_input_offset;
+        const size_t tensor_index = plane_input_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(
+                                                             i + input_offset[0], j + input_offset[1]);
+        local_acc[local_index] = (((i + input_offset[0]) < (input_range[0] + kernel_size[0] - 1)) &&
+                                  in_range_dim1 && in_range_dim2)
+                                     ? device_evaluator.coeff(tensor_index)
+                                     : CoeffReturnType(0);
+      }
+    }
+
+    itemID.barrier(cl::sycl::access::fence_space::local_space);
+
+    // output offset start for each thread
+    const auto output_offset = cl::sycl::range<2>{itemID.get_group(0) * itemID.get_local_range()[0],
+                                                  itemID.get_group(1) * itemID.get_local_range()[1]};
+
+    if (boundary_check(itemID.get_global_id() < input_range)) {
+      CoeffReturnType result = static_cast<CoeffReturnType>(0);
+
+      for (size_t j = 0; j < kernel_size[1]; j++) {
+        size_t kernel_offset = kernel_size[0] * j;
+        const size_t index =
+            (num_input[0] * (plane_kernel_offset + j + itemID.get_local_id(1))) + itemID.get_local_id(0);
+        for (size_t i = 0; i < kernel_size[0]; i++) {
+          result += (local_acc[i + index] * kernel_ptr[i + kernel_offset]);
+        }
+      }
+      const size_t tensor_index =
+          indexMapper.mapGpuOutputPlaneToTensorOutputOffset(itemID.get_global_id(2)) +
+          indexMapper.mapGpuOutputKernelToTensorOutputOffset(itemID.get_local_id(0) + output_offset[0],
+                                                             itemID.get_local_id(1) + output_offset[1]);
+
+      buffer_ptr[tensor_index] = result;
+    }
+  }
+};
+
+template <typename Evaluator, typename CoeffReturnType, typename KernelType, typename Index, typename InputDims,
+          typename Kernel_accessor, typename Buffer_accessor>
+struct EigenConvolutionKernel<Evaluator, CoeffReturnType, KernelType, Index, InputDims, Kernel_accessor,
+                              Buffer_accessor, convolution_type::CONV3D> {
+  typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      Local_accessor;
+  Local_accessor local_acc;
+  Evaluator device_evaluator;
+  Kernel_accessor kernel_filter;
+  Buffer_accessor buffer_acc;
+  internal::IndexMapper<Index, InputDims, 3, Evaluator::Layout> indexMapper;
+  const cl::sycl::range<3> kernel_size;
+  const cl::sycl::range<3> input_range;
+  const size_t numP;
+
+  EigenConvolutionKernel(Local_accessor local_acc_, Evaluator device_evaluator_, Kernel_accessor kernel_filter_,
+                         Buffer_accessor buffer_acc_,
+                         internal::IndexMapper<Index, InputDims, 3, Evaluator::Layout> indexMapper_,
+                         const cl::sycl::range<3> kernel_size_, const cl::sycl::range<3> input_range_,
+                         const size_t numP_)
+      : local_acc(local_acc_),
+        device_evaluator(device_evaluator_),
+        kernel_filter(kernel_filter_),
+        buffer_acc(buffer_acc_),
+        indexMapper(indexMapper_),
+        kernel_size(kernel_size_),
+        input_range(input_range_),
+        numP(numP_) {}
+  template <typename BooleanDim3>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool boundary_check(const BooleanDim3 boolean_check) {
+    return (boolean_check[0] && boolean_check[1] && boolean_check[2]);
+  }
+  void operator()(cl::sycl::nd_item<3> itemID) {
+    auto buffer_ptr = buffer_acc.get_pointer();
+    auto kernel_ptr = kernel_filter.get_pointer();
+    const auto num_input = cl::sycl::range<3>{itemID.get_local_range() + kernel_size - 1};
+
+    const auto input_offset = cl::sycl::range<3>{itemID.get_group().get_id() * itemID.get_local_range()};
+
+    const auto output_offset =
+          cl::sycl::range<3>{itemID.get_group().get_id() * itemID.get_local_range() + itemID.get_local_id()};
+
+    for (size_t p = 0; p < numP; p++) {
+      /// fill the shared memory
+      const size_t plane_input_offset = indexMapper.mapGpuInputPlaneToTensorInputOffset(p);
+      for (size_t k = itemID.get_local_id(2); k < num_input[2]; k += itemID.get_local_range()[2]) {
+        size_t local_index_dim2 = num_input[0] * num_input[1] * k;
+        bool cond_k_dim = (k + input_offset[2] < (input_range[2] + kernel_size[2] - 1));
+        for (size_t j = itemID.get_local_id(1); j < num_input[1]; j += itemID.get_local_range()[1]) {
+          bool cond_j_dim = cond_k_dim && (j + input_offset[1] < (input_range[1] + kernel_size[1] - 1));
+          size_t local_index_dim1 = (num_input[0] * j)  + local_index_dim2;
+          for (size_t i = itemID.get_local_id(0); i < num_input[0]; i += itemID.get_local_range()[0]) {
+            bool conds = cond_j_dim && (i + input_offset[0] < (input_range[0] + kernel_size[0] - 1));
+            const size_t local_index = local_index_dim1 + i;
+            const size_t tensor_index =
+                plane_input_offset + indexMapper.mapGpuInputKernelToTensorInputOffset(
+                                         i + input_offset[0], j + input_offset[1], k + input_offset[2]);
+            local_acc[local_index] = conds ? device_evaluator.coeff(tensor_index) : CoeffReturnType(0);
+          }
+        }
+      }
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+
+      // calculate the convolution
+
+      if (boundary_check(itemID.get_global_id() < input_range)) {
+        CoeffReturnType result = static_cast<CoeffReturnType>(0);
+        for (size_t k = 0; k < kernel_size[2]; k++) {
+          for (size_t j = 0; j < kernel_size[1]; j++) {
+            for (size_t i = 0; i < kernel_size[0]; i++) {
+              const size_t kernel_index = i + kernel_size[0] * (j + kernel_size[1] * k);
+              const size_t local_index =
+                  ((i + itemID.get_local_id(0)) +
+                   num_input[0] * ((j + itemID.get_local_id(1)) + num_input[1] * (k + itemID.get_local_id(2))));
+
+              result += (local_acc[local_index] * kernel_ptr[kernel_index]);
+            }
+          }
+        }
+        const size_t tensor_index =
+            indexMapper.mapGpuOutputPlaneToTensorOutputOffset(p) +
+            indexMapper.mapGpuOutputKernelToTensorOutputOffset(output_offset[0], output_offset[1], output_offset[2]);
+        buffer_ptr[tensor_index] = result;
+      }
+
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+    }
+  }
+};
+
+template <typename Indices, typename InputArgType, typename KernelArgType>
+struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelArgType>, Eigen::SyclDevice> {
+  typedef TensorConvolutionOp<Indices, InputArgType, KernelArgType> XprType;
+
+  static const int NumDims =
+      internal::array_size<typename TensorEvaluator<InputArgType, Eigen::SyclDevice>::Dimensions>::value;
+  static const int NumKernelDims = internal::array_size<Indices>::value;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename TensorEvaluator<KernelArgType, Eigen::SyclDevice>::Dimensions KernelDimensions;
+  typedef const Eigen::SyclDevice Device;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Eigen::SyclDevice>::type PacketReturnType;
+  typedef typename InputArgType::Scalar Scalar;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Eigen::SyclDevice> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  typedef StorageMemory<const CoeffReturnType, Eigen::SyclDevice> KernelStorage;
+
+  enum {
+    IsAligned = TensorEvaluator<InputArgType, Eigen::SyclDevice>::IsAligned &
+                TensorEvaluator<KernelArgType, Eigen::SyclDevice>::IsAligned,
+    PacketAccess = false,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<InputArgType, Eigen::SyclDevice>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  TensorEvaluator(const XprType &op, const Eigen::SyclDevice &device)
+      : m_inputImpl(op.inputExpression(), device),
+        m_kernelArg(op.kernelExpression()),
+        m_kernelImpl(op.kernelExpression(), device),
+        m_indices(op.indices()),
+        m_buf(NULL),
+        m_kernel(NULL),
+        m_local_kernel(false),
+        m_device(device) {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<InputArgType, Eigen::SyclDevice>::Layout) ==
+                         static_cast<int>(TensorEvaluator<KernelArgType, Eigen::SyclDevice>::Layout)),
+                        YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    const typename TensorEvaluator<InputArgType, Eigen::SyclDevice>::Dimensions &input_dims = m_inputImpl.dimensions();
+    const typename TensorEvaluator<KernelArgType, Eigen::SyclDevice>::Dimensions &kernel_dims =
+        m_kernelImpl.dimensions();
+
+    m_dimensions = m_inputImpl.dimensions();
+    for (int i = 0; i < NumKernelDims; ++i) {
+      const Index index = op.indices()[i];
+      const Index input_dim = input_dims[index];
+      const Index kernel_dim = kernel_dims[i];
+      const Index result_dim = input_dim - kernel_dim + 1;
+      m_dimensions[index] = result_dim;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC const Dimensions &dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    preloadKernel();
+    m_inputImpl.evalSubExprsIfNeeded(NULL);
+    if (data) {
+      executeEval(data);
+      return false;
+    } else {
+      m_buf = (EvaluatorPointerType)m_device.get(
+          (Scalar *)m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
+      executeEval(m_buf);
+      return true;
+    }
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_inputImpl.cleanup();
+    if (m_buf) {
+      m_device.deallocate_temp(m_buf);
+      m_buf = NULL;
+    }
+    if (m_local_kernel) {
+      m_device.deallocate_temp(m_kernel);
+      m_local_kernel = false;
+    }
+    m_kernel = NULL;
+  }
+  /// used by sycl in order to build the sycl buffer
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device &device() const { return m_device; }
+  /// used by sycl in order to build the sycl buffer
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE EvaluatorPointerType data() const { return m_buf; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void preloadKernel() {
+    // Don't make a local copy of the kernel unless we have to (i.e. it's an
+    // expression that needs to be evaluated)
+    typename KernelStorage::Type in_place = m_kernelImpl.data();
+    if (in_place) {
+      m_kernel = in_place;
+      m_local_kernel = false;
+    } else {
+      ptrdiff_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar);
+      EvaluatorPointerType local = (EvaluatorPointerType)m_device.get((Scalar *)m_device.allocate_temp(kernel_sz));
+      typedef TensorEvalToOp<const KernelArgType> EvalTo;
+      EvalTo evalToTmp(m_device.get(local), m_kernelArg);
+      const bool PacketAccess = internal::IsVectorizable<Eigen::SyclDevice, KernelArgType>::value;
+      internal::TensorExecutor<const EvalTo, Eigen::SyclDevice, PacketAccess>::run(evalToTmp, m_device);
+      m_kernel = local;
+      m_local_kernel = true;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void executeEval(EvaluatorPointerType data) const {
+    typedef TensorEvaluator<InputArgType, Eigen::SyclDevice> InputEvaluator;
+    typedef typename InputEvaluator::Dimensions InputDims;
+    switch (NumKernelDims) {
+      case 1: {
+        const size_t numX = dimensions()[m_indices[0]];
+        const size_t numP = dimensions().TotalSize() / numX;
+        const auto input_dim = std::array<size_t, 2>{numX, numP};
+        auto global_range = cl::sycl::range<2>{};
+        auto local_range = cl::sycl::range<2>{};
+        const size_t kernel_size = m_kernelImpl.dimensions().TotalSize();
+
+        m_device.parallel_for_setup(input_dim, global_range, local_range);
+        const size_t local_memory_size = (local_range[0] + kernel_size - 1) * (local_range[1]);
+        gpu_assert(static_cast<unsigned long>(local_memory_size) <= m_device.sharedMemPerBlock());
+        const array<Index, 1> indices{{m_indices[0]}};
+        const array<Index, 1> kernel_dims{{m_kernelImpl.dimensions()[0]}};
+        internal::IndexMapper<Index, InputDims, 1, Layout> indexMapper(m_inputImpl.dimensions(), kernel_dims, indices);
+
+        typedef EigenConvolutionKernel<InputEvaluator, CoeffReturnType, Scalar, Index, InputDims,
+                                       typename KernelStorage::Type, EvaluatorPointerType, convolution_type::CONV1D>
+            ConvKernel;
+
+        m_device.template binary_kernel_launcher<CoeffReturnType, ConvKernel>(
+            m_inputImpl, m_kernel, data, cl::sycl::nd_range<2>(global_range, local_range), local_memory_size,
+            indexMapper, kernel_size, cl::sycl::range<2>(input_dim[0], input_dim[1]));
+        break;
+      }
+
+      case 2: {
+        auto kernel_index = std::array<size_t, 2>{static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : 1,
+                                                  static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 1 : 0};
+        auto kernel_size = cl::sycl::range<2>{(size_t)m_kernelImpl.dimensions()[kernel_index[0]],
+                                              (size_t)m_kernelImpl.dimensions()[kernel_index[1]]};
+        const size_t numX = dimensions()[m_indices[kernel_index[0]]];
+        const size_t numY = dimensions()[m_indices[kernel_index[1]]];
+        const size_t numP = dimensions().TotalSize() / (numX * numY);
+        auto input_dim = std::array<size_t, 3>{numX, numY, numP};
+
+        auto global_range = cl::sycl::range<3>{};
+        auto local_range = cl::sycl::range<3>{};
+
+        m_device.parallel_for_setup(input_dim, global_range, local_range);
+
+        const size_t local_memory_size =
+            (local_range[0] + kernel_size[0] - 1) * (local_range[1] + kernel_size[1] - 1) * local_range[2];
+        gpu_assert(static_cast<unsigned long>(local_memory_size) <= m_device.sharedMemPerBlock());
+        const array<Index, 2> indices{{m_indices[kernel_index[0]], m_indices[kernel_index[1]]}};
+        const array<Index, 2> kernel_dims{
+            {m_kernelImpl.dimensions()[kernel_index[0]], m_kernelImpl.dimensions()[kernel_index[1]]}};
+        internal::IndexMapper<Index, InputDims, 2, Layout> indexMapper(m_inputImpl.dimensions(), kernel_dims, indices);
+        typedef EigenConvolutionKernel<InputEvaluator, CoeffReturnType, Scalar, Index, InputDims,
+                                       typename KernelStorage::Type, EvaluatorPointerType, convolution_type::CONV2D>
+            ConvKernel;
+        m_device.template binary_kernel_launcher<CoeffReturnType, ConvKernel>(
+            m_inputImpl, m_kernel, data, cl::sycl::nd_range<3>(global_range, local_range), local_memory_size,
+            indexMapper, kernel_size, cl::sycl::range<3>{input_dim[0], input_dim[1], input_dim[2]});
+        break;
+      }
+
+      case 3: {
+        auto kernel_index = std::array<size_t, 3>{static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : 2,
+                                                  static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 1 : 1,
+                                                  static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 2 : 0};
+
+        auto kernel_size = cl::sycl::range<3>{(size_t)m_kernelImpl.dimensions()[kernel_index[0]],
+                                              (size_t)m_kernelImpl.dimensions()[kernel_index[1]],
+                                              (size_t)m_kernelImpl.dimensions()[kernel_index[2]]};
+
+        const size_t numX = dimensions()[m_indices[kernel_index[0]]];
+        const size_t numY = dimensions()[m_indices[kernel_index[1]]];
+        const size_t numZ = dimensions()[m_indices[kernel_index[2]]];
+        auto input_dim = std::array<size_t, 3>{numX, numY, numZ};
+        const size_t numP = dimensions().TotalSize() / (numX * numY * numZ);
+
+        const array<Index, 3> indices{
+            {m_indices[kernel_index[0]], m_indices[kernel_index[1]], m_indices[kernel_index[2]]}};
+        const array<Index, 3> kernel_dims{{m_kernelImpl.dimensions()[kernel_index[0]],
+                                           m_kernelImpl.dimensions()[kernel_index[1]],
+                                           m_kernelImpl.dimensions()[kernel_index[2]]}};
+
+        internal::IndexMapper<Index, InputDims, 3, Layout> indexMapper(m_inputImpl.dimensions(), kernel_dims, indices);
+
+        auto global_range = cl::sycl::range<3>{};
+        auto local_range = cl::sycl::range<3>{};
+
+        m_device.parallel_for_setup(input_dim, global_range, local_range);
+        auto local_memory_range = (local_range + kernel_size - 1);
+        const size_t local_memory_size = local_memory_range[0] * local_memory_range[1] * local_memory_range[2];
+
+        gpu_assert(static_cast<unsigned long>(local_memory_size) <= m_device.sharedMemPerBlock());
+        typedef EigenConvolutionKernel<InputEvaluator, CoeffReturnType, Scalar, Index, InputDims,
+                                       typename KernelStorage::Type, EvaluatorPointerType, convolution_type::CONV3D>
+            ConvKernel;
+        m_device.template binary_kernel_launcher<CoeffReturnType, ConvKernel>(
+            m_inputImpl, m_kernel, data, cl::sycl::nd_range<3>(global_range, local_range), local_memory_size,
+            indexMapper, kernel_size, cl::sycl::range<3>(input_dim[0], input_dim[1], input_dim[2]), numP);
+        break;
+      }
+
+      default: {
+        EIGEN_STATIC_ASSERT((NumKernelDims >= 1 && NumKernelDims <= 3),
+                            THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    eigen_assert(m_buf != NULL);
+    eigen_assert(index < m_dimensions.TotalSize());
+    return m_buf[index];
+  }
+
+  template <int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(const Index index) const {
+    eigen_assert(m_buf != NULL);
+    eigen_assert(index < m_dimensions.TotalSize());
+    return internal::ploadt<PacketReturnType, LoadMode>(m_buf + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    // TODO(rmlarsen): FIXME: For now, this is just a copy of the CPU cost
+    // model.
+    const double kernel_size = m_kernelImpl.dimensions().TotalSize();
+    // We ignore the use of fused multiply-add.
+    const double convolve_compute_cost = TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
+    const double firstIndex_compute_cost =
+        NumDims *
+        (2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() + TensorOpCost::DivCost<Index>());
+    return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
+           kernel_size * (m_inputImpl.costPerCoeff(vectorized) + m_kernelImpl.costPerCoeff(vectorized) +
+                          TensorOpCost(0, 0, convolve_compute_cost, vectorized, PacketSize));
+  }
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_kernelImpl.bind(cgh);
+    m_inputImpl.bind(cgh);
+    m_buf.bind(cgh);
+    m_kernel.bind(cgh);
+  }
+
+ private:
+  // No assignment (copies are needed by the kernels)
+  TensorEvaluator &operator=(const TensorEvaluator &);
+  TensorEvaluator<InputArgType, Eigen::SyclDevice> m_inputImpl;
+  KernelArgType m_kernelArg;
+  TensorEvaluator<KernelArgType, Eigen::SyclDevice> m_kernelImpl;
+  Indices m_indices;
+  Dimensions m_dimensions;
+  EvaluatorPointerType m_buf;
+  typename KernelStorage::Type m_kernel;
+  bool m_local_kernel;
+  const Eigen::SyclDevice EIGEN_DEVICE_REF m_device;
+};  // namespace Eigen
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorCostModel.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorCostModel.h
new file mode 100644
index 0000000..195267c
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorCostModel.h
@@ -0,0 +1,214 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Rasmus Munk Larsen <rmlarsen@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H
+#define EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H
+
+namespace Eigen {
+
+/** \class TensorEvaluator
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief A cost model used to limit the number of threads used for evaluating
+  * tensor expression.
+  *
+  */
+
+// Class storing the cost of evaluating a tensor expression in terms of the
+// estimated number of operand bytes loads, bytes stored, and compute cycles.
+class TensorOpCost {
+ public:
+  // TODO(rmlarsen): Fix the scalar op costs in Eigen proper. Even a simple
+  // model based on minimal reciprocal throughput numbers from Intel or
+  // Agner Fog's tables would be better than what is there now.
+  template <typename ArgType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int MulCost() {
+    return internal::functor_traits<
+        internal::scalar_product_op<ArgType, ArgType> >::Cost;
+  }
+  template <typename ArgType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int AddCost() {
+    return internal::functor_traits<internal::scalar_sum_op<ArgType> >::Cost;
+  }
+  template <typename ArgType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int DivCost() {
+    return internal::functor_traits<
+        internal::scalar_quotient_op<ArgType, ArgType> >::Cost;
+  }
+  template <typename ArgType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int ModCost() {
+    return internal::functor_traits<internal::scalar_mod_op<ArgType> >::Cost;
+  }
+  template <typename SrcType, typename TargetType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int CastCost() {
+    return internal::functor_traits<
+        internal::scalar_cast_op<SrcType, TargetType> >::Cost;
+  }
+
+  EIGEN_DEVICE_FUNC
+  TensorOpCost() : bytes_loaded_(0), bytes_stored_(0), compute_cycles_(0) {}
+  EIGEN_DEVICE_FUNC
+  TensorOpCost(double bytes_loaded, double bytes_stored, double compute_cycles)
+      : bytes_loaded_(bytes_loaded),
+        bytes_stored_(bytes_stored),
+        compute_cycles_(compute_cycles) {}
+
+  EIGEN_DEVICE_FUNC
+  TensorOpCost(double bytes_loaded, double bytes_stored, double compute_cycles,
+               bool vectorized, double packet_size)
+      : bytes_loaded_(bytes_loaded),
+        bytes_stored_(bytes_stored),
+        compute_cycles_(vectorized ? compute_cycles / packet_size
+                                   : compute_cycles) {
+    eigen_assert(bytes_loaded >= 0 && (numext::isfinite)(bytes_loaded));
+    eigen_assert(bytes_stored >= 0 && (numext::isfinite)(bytes_stored));
+    eigen_assert(compute_cycles >= 0 && (numext::isfinite)(compute_cycles));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double bytes_loaded() const {
+    return bytes_loaded_;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double bytes_stored() const {
+    return bytes_stored_;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double compute_cycles() const {
+    return compute_cycles_;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double total_cost(
+      double load_cost, double store_cost, double compute_cost) const {
+    return load_cost * bytes_loaded_ + store_cost * bytes_stored_ +
+           compute_cost * compute_cycles_;
+  }
+
+  // Drop memory access component. Intended for cases when memory accesses are
+  // sequential or are completely masked by computations.
+  EIGEN_DEVICE_FUNC void dropMemoryCost() {
+    bytes_loaded_ = 0;
+    bytes_stored_ = 0;
+  }
+
+  // TODO(rmlarsen): Define min in terms of total cost, not elementwise.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost cwiseMin(
+      const TensorOpCost& rhs) const {
+    double bytes_loaded = numext::mini(bytes_loaded_, rhs.bytes_loaded());
+    double bytes_stored = numext::mini(bytes_stored_, rhs.bytes_stored());
+    double compute_cycles = numext::mini(compute_cycles_, rhs.compute_cycles());
+    return TensorOpCost(bytes_loaded, bytes_stored, compute_cycles);
+  }
+
+  // TODO(rmlarsen): Define max in terms of total cost, not elementwise.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost cwiseMax(
+      const TensorOpCost& rhs) const {
+    double bytes_loaded = numext::maxi(bytes_loaded_, rhs.bytes_loaded());
+    double bytes_stored = numext::maxi(bytes_stored_, rhs.bytes_stored());
+    double compute_cycles = numext::maxi(compute_cycles_, rhs.compute_cycles());
+    return TensorOpCost(bytes_loaded, bytes_stored, compute_cycles);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost& operator+=(
+      const TensorOpCost& rhs) {
+    bytes_loaded_ += rhs.bytes_loaded();
+    bytes_stored_ += rhs.bytes_stored();
+    compute_cycles_ += rhs.compute_cycles();
+    return *this;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost& operator*=(double rhs) {
+    bytes_loaded_ *= rhs;
+    bytes_stored_ *= rhs;
+    compute_cycles_ *= rhs;
+    return *this;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE friend TensorOpCost operator+(
+      TensorOpCost lhs, const TensorOpCost& rhs) {
+    lhs += rhs;
+    return lhs;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE friend TensorOpCost operator*(
+      TensorOpCost lhs, double rhs) {
+    lhs *= rhs;
+    return lhs;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE friend TensorOpCost operator*(
+      double lhs, TensorOpCost rhs) {
+    rhs *= lhs;
+    return rhs;
+  }
+
+  friend std::ostream& operator<<(std::ostream& os, const TensorOpCost& tc) {
+    return os << "[bytes_loaded = " << tc.bytes_loaded()
+              << ", bytes_stored = " << tc.bytes_stored()
+              << ", compute_cycles = " << tc.compute_cycles() << "]";
+  }
+
+ private:
+  double bytes_loaded_;
+  double bytes_stored_;
+  double compute_cycles_;
+};
+
+// TODO(rmlarsen): Implement a policy that chooses an "optimal" number of theads
+// in [1:max_threads] instead of just switching multi-threading off for small
+// work units.
+template <typename Device>
+class TensorCostModel {
+ public:
+  // Scaling from Eigen compute cost to device cycles.
+  static const int kDeviceCyclesPerComputeCycle = 1;
+
+ // Costs in device cycles.
+  static const int kStartupCycles = 100000;
+  static const int kPerThreadCycles = 100000;
+  static const int kTaskSize = 40000;
+
+  // Returns the number of threads in [1:max_threads] to use for
+  // evaluating an expression with the given output size and cost per
+  // coefficient.
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int numThreads(
+      double output_size, const TensorOpCost& cost_per_coeff, int max_threads) {
+    double cost = totalCost(output_size, cost_per_coeff);
+    double threads = (cost - kStartupCycles) / kPerThreadCycles + 0.9;
+    // Make sure we don't invoke undefined behavior when we convert to an int.
+    threads = numext::mini<double>(threads, GenericNumTraits<int>::highest());
+    return numext::mini(max_threads,
+                        numext::maxi<int>(1, static_cast<int>(threads)));
+  }
+
+  // taskSize assesses parallel task size.
+  // Value of 1.0 means ideal parallel task size. Values < 1.0 mean that task
+  // granularity needs to be increased to mitigate parallelization overheads.
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double taskSize(
+      double output_size, const TensorOpCost& cost_per_coeff) {
+    return totalCost(output_size, cost_per_coeff) / kTaskSize;
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double totalCost(
+      double output_size, const TensorOpCost& cost_per_coeff) {
+    // Cost of memory fetches from L2 cache. 64 is typical cache line size.
+    // 11 is L2 cache latency on Haswell.
+    // We don't know whether data is in L1, L2 or L3. But we are most interested
+    // in single-threaded computational time around 100us-10ms (smaller time
+    // is too small for parallelization, larger time is not interesting
+    // either because we are probably using all available threads already).
+    // And for the target time range, L2 seems to be what matters. Data set
+    // fitting into L1 is too small to take noticeable time. Data set fitting
+    // only into L3 presumably will take more than 10ms to load and process.
+    const double kLoadCycles = 1.0 / 64 * 11;
+    const double kStoreCycles = 1.0 / 64 * 11;
+    // Scaling from Eigen compute cost to device cycles.
+    return output_size *
+        cost_per_coeff.total_cost(kLoadCycles, kStoreCycles,
+                                  kDeviceCyclesPerComputeCycle);
+  }
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorCustomOp.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorCustomOp.h
new file mode 100644
index 0000000..95a8a84
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorCustomOp.h
@@ -0,0 +1,347 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CUSTOM_OP_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CUSTOM_OP_H
+
+namespace Eigen {
+
+/** \class TensorCustomUnaryOp
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor custom class.
+  *
+  *
+  */
+namespace internal {
+template<typename CustomUnaryFunc, typename XprType>
+struct traits<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::StorageKind StorageKind;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = traits<XprType>::NumDimensions;
+  static const int Layout = traits<XprType>::Layout;
+  typedef typename traits<XprType>::PointerType PointerType;
+};
+
+template<typename CustomUnaryFunc, typename XprType>
+struct eval<TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Eigen::Dense>
+{
+  typedef const TensorCustomUnaryOp<CustomUnaryFunc, XprType>EIGEN_DEVICE_REF type;
+};
+
+template<typename CustomUnaryFunc, typename XprType>
+struct nested<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
+{
+  typedef TensorCustomUnaryOp<CustomUnaryFunc, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename CustomUnaryFunc, typename XprType>
+class TensorCustomUnaryOp : public TensorBase<TensorCustomUnaryOp<CustomUnaryFunc, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename internal::traits<TensorCustomUnaryOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename internal::nested<TensorCustomUnaryOp>::type Nested;
+  typedef typename internal::traits<TensorCustomUnaryOp>::StorageKind StorageKind;
+  typedef typename internal::traits<TensorCustomUnaryOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCustomUnaryOp(const XprType& expr, const CustomUnaryFunc& func)
+      : m_expr(expr), m_func(func) {}
+
+  EIGEN_DEVICE_FUNC
+  const CustomUnaryFunc& func() const { return m_func; }
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename XprType::Nested>::type&
+  expression() const { return m_expr; }
+
+  protected:
+    typename XprType::Nested m_expr;
+    const CustomUnaryFunc m_func;
+};
+
+
+// Eval as rvalue
+template<typename CustomUnaryFunc, typename XprType, typename Device>
+struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Device>
+{
+  typedef TensorCustomUnaryOp<CustomUnaryFunc, XprType> ArgType;
+  typedef typename internal::traits<ArgType>::Index Index;
+  static const int NumDims = internal::traits<ArgType>::NumDimensions;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename internal::remove_const<typename ArgType::Scalar>::type Scalar;
+  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename Eigen::internal::traits<XprType>::PointerType TensorPointerType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<XprType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const ArgType& op, const Device& device)
+      : m_op(op), m_device(device), m_result(NULL)
+  {
+    m_dimensions = op.func().dimensions(op.expression());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    if (data) {
+      evalTo(data);
+      return false;
+    } else {
+      m_result = static_cast<EvaluatorPointerType>(m_device.get( (CoeffReturnType*)
+          m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar))));
+      evalTo(m_result);
+      return true;
+    }
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    if (m_result) {
+      m_device.deallocate_temp(m_result);
+      m_result = NULL;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    return m_result[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_result + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    // TODO(rmlarsen): Extend CustomOp API to return its cost estimate.
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_result; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_result.bind(cgh);
+  }
+#endif
+
+ protected:
+  void evalTo(EvaluatorPointerType data) {
+    TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(m_device.get(data), m_dimensions);
+    m_op.func().eval(m_op.expression(), result, m_device);
+  }
+
+  Dimensions m_dimensions;
+  const ArgType m_op;
+  const Device EIGEN_DEVICE_REF m_device;
+  EvaluatorPointerType m_result;
+};
+
+
+
+/** \class TensorCustomBinaryOp
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor custom class.
+  *
+  *
+  */
+namespace internal {
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
+struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
+{
+  typedef typename internal::promote_storage_type<typename LhsXprType::Scalar,
+                                                  typename RhsXprType::Scalar>::ret Scalar;
+  typedef typename internal::promote_storage_type<typename LhsXprType::CoeffReturnType,
+                                                  typename RhsXprType::CoeffReturnType>::ret CoeffReturnType;
+  typedef typename promote_storage_type<typename traits<LhsXprType>::StorageKind,
+                                        typename traits<RhsXprType>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<LhsXprType>::Index,
+                                      typename traits<RhsXprType>::Index>::type Index;
+  typedef typename LhsXprType::Nested LhsNested;
+  typedef typename RhsXprType::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+  static const int NumDimensions = traits<LhsXprType>::NumDimensions;
+  static const int Layout = traits<LhsXprType>::Layout;
+  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                                typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
+};
+
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
+struct eval<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Eigen::Dense>
+{
+  typedef const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>& type;
+};
+
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
+struct nested<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
+{
+  typedef TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
+class TensorCustomBinaryOp : public TensorBase<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename internal::traits<TensorCustomBinaryOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename internal::traits<TensorCustomBinaryOp>::CoeffReturnType CoeffReturnType;
+  typedef typename internal::nested<TensorCustomBinaryOp>::type Nested;
+  typedef typename internal::traits<TensorCustomBinaryOp>::StorageKind StorageKind;
+  typedef typename internal::traits<TensorCustomBinaryOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCustomBinaryOp(const LhsXprType& lhs, const RhsXprType& rhs, const CustomBinaryFunc& func)
+
+      : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_func(func) {}
+
+  EIGEN_DEVICE_FUNC
+  const CustomBinaryFunc& func() const { return m_func; }
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename LhsXprType::Nested>::type&
+  lhsExpression() const { return m_lhs_xpr; }
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename RhsXprType::Nested>::type&
+  rhsExpression() const { return m_rhs_xpr; }
+
+  protected:
+    typename LhsXprType::Nested m_lhs_xpr;
+    typename RhsXprType::Nested m_rhs_xpr;
+    const CustomBinaryFunc m_func;
+};
+
+
+// Eval as rvalue
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, typename Device>
+struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Device>
+{
+  typedef TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> XprType;
+  typedef typename internal::traits<XprType>::Index Index;
+  static const int NumDims = internal::traits<XprType>::NumDimensions;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+
+  typedef typename Eigen::internal::traits<XprType>::PointerType TensorPointerType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<LhsXprType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_op(op), m_device(device), m_result(NULL)
+  {
+    m_dimensions = op.func().dimensions(op.lhsExpression(), op.rhsExpression());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    if (data) {
+      evalTo(data);
+      return false;
+    } else {
+      m_result = static_cast<EvaluatorPointerType>(m_device.get( (CoeffReturnType*)
+        m_device.allocate_temp(dimensions().TotalSize() * sizeof(CoeffReturnType))));
+      evalTo(m_result);
+      return true;
+    }
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    if (m_result != NULL) {
+      m_device.deallocate_temp(m_result);
+      m_result = NULL;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    return m_result[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_result + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    // TODO(rmlarsen): Extend CustomOp API to return its cost estimate.
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_result; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_result.bind(cgh);
+  }
+#endif
+
+ protected:
+  void evalTo(EvaluatorPointerType data) {
+    TensorMap<Tensor<CoeffReturnType, NumDims, Layout> > result(m_device.get(data), m_dimensions);
+    m_op.func().eval(m_op.lhsExpression(), m_op.rhsExpression(), result, m_device);
+  }
+
+  Dimensions m_dimensions;
+  const XprType m_op;
+  const Device EIGEN_DEVICE_REF m_device;
+  EvaluatorPointerType m_result;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CUSTOM_OP_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorDevice.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorDevice.h
new file mode 100644
index 0000000..96fa46c
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorDevice.h
@@ -0,0 +1,137 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H
+#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H
+
+namespace Eigen {
+
+/** \class TensorDevice
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Pseudo expression providing an operator = that will evaluate its argument
+  * on the specified computing 'device' (GPU, thread pool, ...)
+  *
+  * Example:
+  *    C.device(EIGEN_GPU) = A + B;
+  *
+  * Todo: operator *= and /=.
+  */
+
+template <typename ExpressionType, typename DeviceType> class TensorDevice {
+  public:
+    TensorDevice(const DeviceType& device, ExpressionType& expression) : m_device(device), m_expression(expression) {}
+
+    EIGEN_DEFAULT_COPY_CONSTRUCTOR(TensorDevice)
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) {
+      typedef TensorAssignOp<ExpressionType, const OtherDerived> Assign;
+      Assign assign(m_expression, other);
+      internal::TensorExecutor<const Assign, DeviceType>::run(assign, m_device);
+      return *this;
+    }
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE TensorDevice& operator+=(const OtherDerived& other) {
+      typedef typename OtherDerived::Scalar Scalar;
+      typedef TensorCwiseBinaryOp<internal::scalar_sum_op<Scalar>, const ExpressionType, const OtherDerived> Sum;
+      Sum sum(m_expression, other);
+      typedef TensorAssignOp<ExpressionType, const Sum> Assign;
+      Assign assign(m_expression, sum);
+      internal::TensorExecutor<const Assign, DeviceType>::run(assign, m_device);
+      return *this;
+    }
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE TensorDevice& operator-=(const OtherDerived& other) {
+      typedef typename OtherDerived::Scalar Scalar;
+      typedef TensorCwiseBinaryOp<internal::scalar_difference_op<Scalar>, const ExpressionType, const OtherDerived> Difference;
+      Difference difference(m_expression, other);
+      typedef TensorAssignOp<ExpressionType, const Difference> Assign;
+      Assign assign(m_expression, difference);
+      internal::TensorExecutor<const Assign, DeviceType>::run(assign, m_device);
+      return *this;
+    }
+
+  protected:
+    const DeviceType& m_device;
+    ExpressionType& m_expression;
+};
+
+/** \class TensorAsyncDevice
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Pseudo expression providing an operator = that will evaluate its
+ * argument asynchronously on the specified device. Currently only
+ * ThreadPoolDevice implements proper asynchronous execution, while the default
+ * and GPU devices just run the expression synchronously and call m_done() on
+ * completion..
+ *
+ * Example:
+ *    auto done = []() { ... expression evaluation done ... };
+ *    C.device(thread_pool_device, std::move(done)) = A + B;
+ */
+
+template <typename ExpressionType, typename DeviceType, typename DoneCallback>
+class TensorAsyncDevice {
+ public:
+  TensorAsyncDevice(const DeviceType& device, ExpressionType& expression,
+                    DoneCallback done)
+      : m_device(device), m_expression(expression), m_done(std::move(done)) {}
+
+  template <typename OtherDerived>
+  EIGEN_STRONG_INLINE TensorAsyncDevice& operator=(const OtherDerived& other) {
+    typedef TensorAssignOp<ExpressionType, const OtherDerived> Assign;
+    typedef internal::TensorExecutor<const Assign, DeviceType> Executor;
+
+    Assign assign(m_expression, other);
+    Executor::run(assign, m_device);
+    m_done();
+
+    return *this;
+  }
+
+ protected:
+  const DeviceType& m_device;
+  ExpressionType& m_expression;
+  DoneCallback m_done;
+};
+
+
+#ifdef EIGEN_USE_THREADS
+template <typename ExpressionType, typename DoneCallback>
+class TensorAsyncDevice<ExpressionType, ThreadPoolDevice, DoneCallback> {
+ public:
+  TensorAsyncDevice(const ThreadPoolDevice& device, ExpressionType& expression,
+                    DoneCallback done)
+      : m_device(device), m_expression(expression), m_done(std::move(done)) {}
+
+  template <typename OtherDerived>
+  EIGEN_STRONG_INLINE TensorAsyncDevice& operator=(const OtherDerived& other) {
+    typedef TensorAssignOp<ExpressionType, const OtherDerived> Assign;
+    typedef internal::TensorAsyncExecutor<const Assign, ThreadPoolDevice, DoneCallback> Executor;
+
+    // WARNING: After assignment 'm_done' callback will be in undefined state.
+    Assign assign(m_expression, other);
+    Executor::runAsync(assign, m_device, std::move(m_done));
+
+    return *this;
+  }
+
+ protected:
+  const ThreadPoolDevice& m_device;
+  ExpressionType& m_expression;
+  DoneCallback m_done;
+};
+#endif
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceCuda.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceCuda.h
new file mode 100644
index 0000000..f779239
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceCuda.h
@@ -0,0 +1,6 @@
+
+#if defined(__clang__) || defined(__GNUC__)
+#warning "Deprecated header file, please either include the main Eigen/CXX11/Tensor header or the respective TensorDeviceGpu.h file"
+#endif
+
+#include "TensorDeviceGpu.h"
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceDefault.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceDefault.h
new file mode 100644
index 0000000..46b9d3a
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceDefault.h
@@ -0,0 +1,104 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H
+#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H
+
+
+namespace Eigen {
+
+// Default device for the machine (typically a single cpu core)
+struct DefaultDevice {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
+    return internal::aligned_malloc(num_bytes);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
+    internal::aligned_free(buffer);
+  }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
+    return allocate(num_bytes);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
+    deallocate(buffer);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
+    ::memcpy(dst, src, n);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
+    memcpy(dst, src, n);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const {
+    memcpy(dst, src, n);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
+    ::memset(buffer, c, n);
+  }
+  template<typename Type>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Type get(Type data) const { 
+    return data;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t numThreads() const {
+#if !defined(EIGEN_GPU_COMPILE_PHASE)
+    // Running on the host CPU
+    return 1;
+#elif defined(EIGEN_HIP_DEVICE_COMPILE)
+    // Running on a HIP device
+    return 64;
+#else
+    // Running on a CUDA device
+    return 32;
+#endif
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY)
+    // Running on the host CPU
+    return l1CacheSize();
+#elif defined(EIGEN_HIP_DEVICE_COMPILE)
+    // Running on a HIP device
+    return 48*1024; // FIXME : update this number for HIP
+#else
+    // Running on a CUDA device, return the amount of shared memory available.
+    return 48*1024;
+#endif
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY)
+    // Running single threaded on the host CPU
+    return l3CacheSize();
+#elif defined(EIGEN_HIP_DEVICE_COMPILE)
+    // Running on a HIP device
+    return firstLevelCacheSize(); // FIXME : update this number for HIP
+#else
+    // Running on a CUDA device
+    return firstLevelCacheSize();
+#endif
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const {
+#if !defined(EIGEN_GPU_COMPILE_PHASE)
+    // Running single threaded on the host CPU
+    // Should return an enum that encodes the ISA supported by the CPU
+    return 1;
+#elif defined(EIGEN_HIP_DEVICE_COMPILE)
+    // Running on a HIP device
+    // return 1 as major for HIP
+    return 1;
+#else
+    // Running on a CUDA device
+    return EIGEN_CUDA_ARCH / 100;
+#endif
+  }
+};
+
+}  // namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceGpu.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceGpu.h
new file mode 100644
index 0000000..ec2e3cb
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceGpu.h
@@ -0,0 +1,389 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#if defined(EIGEN_USE_GPU) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_GPU_H)
+#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_GPU_H
+
+// This header file container defines fo gpu* macros which will resolve to
+// their equivalent hip* or cuda* versions depending on the compiler in use
+// A separate header (included at the end of this file) will undefine all 
+#include "TensorGpuHipCudaDefines.h"
+
+namespace Eigen {
+
+static const int kGpuScratchSize = 1024;
+
+// This defines an interface that GPUDevice can take to use
+// HIP / CUDA streams underneath.
+class StreamInterface {
+ public:
+  virtual ~StreamInterface() {}
+
+  virtual const gpuStream_t& stream() const = 0;
+  virtual const gpuDeviceProp_t& deviceProperties() const = 0;
+
+  // Allocate memory on the actual device where the computation will run
+  virtual void* allocate(size_t num_bytes) const = 0;
+  virtual void deallocate(void* buffer) const = 0;
+
+  // Return a scratchpad buffer of size 1k
+  virtual void* scratchpad() const = 0;
+
+  // Return a semaphore. The semaphore is initially initialized to 0, and
+  // each kernel using it is responsible for resetting to 0 upon completion
+  // to maintain the invariant that the semaphore is always equal to 0 upon
+  // each kernel start.
+  virtual unsigned int* semaphore() const = 0;
+};
+
+class GpuDeviceProperties {
+ public:
+  GpuDeviceProperties() : 
+      initialized_(false), first_(true), device_properties_(nullptr) {}
+ 
+  ~GpuDeviceProperties() {
+    if (device_properties_) {
+      delete[] device_properties_;
+    }
+  }
+  
+  EIGEN_STRONG_INLINE const gpuDeviceProp_t& get(int device) const {
+    return device_properties_[device];
+  }
+
+  EIGEN_STRONG_INLINE bool isInitialized() const {
+    return initialized_;
+  }
+
+  void initialize() {
+    if (!initialized_) {
+      // Attempts to ensure proper behavior in the case of multiple threads
+      // calling this function simultaneously. This would be trivial to
+      // implement if we could use std::mutex, but unfortunately mutex don't
+      // compile with nvcc, so we resort to atomics and thread fences instead.
+      // Note that if the caller uses a compiler that doesn't support c++11 we
+      // can't ensure that the initialization is thread safe.
+      if (first_.exchange(false)) {
+        // We're the first thread to reach this point.
+        int num_devices;
+        gpuError_t status = gpuGetDeviceCount(&num_devices);
+        if (status != gpuSuccess) {
+          std::cerr << "Failed to get the number of GPU devices: "
+                    << gpuGetErrorString(status)
+                    << std::endl;
+          gpu_assert(status == gpuSuccess);
+        }
+        device_properties_ = new gpuDeviceProp_t[num_devices];
+        for (int i = 0; i < num_devices; ++i) {
+          status = gpuGetDeviceProperties(&device_properties_[i], i);
+          if (status != gpuSuccess) {
+            std::cerr << "Failed to initialize GPU device #"
+                      << i
+                      << ": "
+                      << gpuGetErrorString(status)
+                      << std::endl;
+            gpu_assert(status == gpuSuccess);
+          }
+        }
+
+        std::atomic_thread_fence(std::memory_order_release);
+        initialized_ = true;
+      } else {
+        // Wait for the other thread to inititialize the properties.
+        while (!initialized_) {
+          std::atomic_thread_fence(std::memory_order_acquire);
+          std::this_thread::sleep_for(std::chrono::milliseconds(1000));
+        }
+      }
+    }
+  }
+
+ private:
+  volatile bool initialized_;
+  std::atomic<bool> first_;
+  gpuDeviceProp_t* device_properties_;
+};
+
+EIGEN_ALWAYS_INLINE const GpuDeviceProperties& GetGpuDeviceProperties() {
+  static GpuDeviceProperties* deviceProperties = new GpuDeviceProperties();
+  if (!deviceProperties->isInitialized()) {
+    deviceProperties->initialize();
+  }
+  return *deviceProperties;
+}
+
+EIGEN_ALWAYS_INLINE const gpuDeviceProp_t& GetGpuDeviceProperties(int device) {
+  return GetGpuDeviceProperties().get(device);
+}
+
+static const gpuStream_t default_stream = gpuStreamDefault;
+
+class GpuStreamDevice : public StreamInterface {
+ public:
+  // Use the default stream on the current device
+  GpuStreamDevice() : stream_(&default_stream), scratch_(NULL), semaphore_(NULL) {
+    gpuGetDevice(&device_);
+  }
+  // Use the default stream on the specified device
+  GpuStreamDevice(int device) : stream_(&default_stream), device_(device), scratch_(NULL), semaphore_(NULL) {}
+  // Use the specified stream. Note that it's the
+  // caller responsibility to ensure that the stream can run on
+  // the specified device. If no device is specified the code
+  // assumes that the stream is associated to the current gpu device.
+  GpuStreamDevice(const gpuStream_t* stream, int device = -1)
+      : stream_(stream), device_(device), scratch_(NULL), semaphore_(NULL) {
+    if (device < 0) {
+      gpuGetDevice(&device_);
+    } else {
+      int num_devices;
+      gpuError_t err = gpuGetDeviceCount(&num_devices);
+      EIGEN_UNUSED_VARIABLE(err)
+      gpu_assert(err == gpuSuccess);
+      gpu_assert(device < num_devices);
+      device_ = device;
+    }
+  }
+
+  virtual ~GpuStreamDevice() {
+    if (scratch_) {
+      deallocate(scratch_);
+    }
+  }
+
+  const gpuStream_t& stream() const { return *stream_; }
+  const gpuDeviceProp_t& deviceProperties() const {
+    return GetGpuDeviceProperties(device_);
+  }
+  virtual void* allocate(size_t num_bytes) const {
+    gpuError_t err = gpuSetDevice(device_);
+    EIGEN_UNUSED_VARIABLE(err)
+    gpu_assert(err == gpuSuccess);
+    void* result;
+    err = gpuMalloc(&result, num_bytes);
+    gpu_assert(err == gpuSuccess);
+    gpu_assert(result != NULL);
+    return result;
+  }
+  virtual void deallocate(void* buffer) const {
+    gpuError_t err = gpuSetDevice(device_);
+    EIGEN_UNUSED_VARIABLE(err)
+    gpu_assert(err == gpuSuccess);
+    gpu_assert(buffer != NULL);
+    err = gpuFree(buffer);
+    gpu_assert(err == gpuSuccess);
+  }
+
+  virtual void* scratchpad() const {
+    if (scratch_ == NULL) {
+      scratch_ = allocate(kGpuScratchSize + sizeof(unsigned int));
+    }
+    return scratch_;
+  }
+
+  virtual unsigned int* semaphore() const {
+    if (semaphore_ == NULL) {
+      char* scratch = static_cast<char*>(scratchpad()) + kGpuScratchSize;
+      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
+      gpuError_t err = gpuMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_);
+      EIGEN_UNUSED_VARIABLE(err)
+      gpu_assert(err == gpuSuccess);
+    }
+    return semaphore_;
+  }
+
+ private:
+  const gpuStream_t* stream_;
+  int device_;
+  mutable void* scratch_;
+  mutable unsigned int* semaphore_;
+};
+
+struct GpuDevice {
+  // The StreamInterface is not owned: the caller is
+  // responsible for its initialization and eventual destruction.
+  explicit GpuDevice(const StreamInterface* stream) : stream_(stream), max_blocks_(INT_MAX) {
+    eigen_assert(stream);
+  }
+  explicit GpuDevice(const StreamInterface* stream, int num_blocks) : stream_(stream), max_blocks_(num_blocks) {
+    eigen_assert(stream);
+  }
+  // TODO(bsteiner): This is an internal API, we should not expose it.
+  EIGEN_STRONG_INLINE const gpuStream_t& stream() const {
+    return stream_->stream();
+  }
+
+  EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
+    return stream_->allocate(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
+    stream_->deallocate(buffer);
+  }
+
+  EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
+    return stream_->allocate(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
+    stream_->deallocate(buffer);
+  }
+
+  template<typename Type>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Type get(Type data) const { 
+    return data;
+  }
+
+  EIGEN_STRONG_INLINE void* scratchpad() const {
+    return stream_->scratchpad();
+  }
+
+  EIGEN_STRONG_INLINE unsigned int* semaphore() const {
+    return stream_->semaphore();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
+#ifndef EIGEN_GPU_COMPILE_PHASE
+    gpuError_t err = gpuMemcpyAsync(dst, src, n, gpuMemcpyDeviceToDevice,
+                                      stream_->stream());
+    EIGEN_UNUSED_VARIABLE(err)
+    gpu_assert(err == gpuSuccess);
+#else
+    EIGEN_UNUSED_VARIABLE(dst);
+    EIGEN_UNUSED_VARIABLE(src);
+    EIGEN_UNUSED_VARIABLE(n);
+    eigen_assert(false && "The default device should be used instead to generate kernel code");
+#endif
+  }
+
+  EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
+    gpuError_t err =
+        gpuMemcpyAsync(dst, src, n, gpuMemcpyHostToDevice, stream_->stream());
+    EIGEN_UNUSED_VARIABLE(err)
+    gpu_assert(err == gpuSuccess);
+  }
+
+  EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const {
+    gpuError_t err =
+        gpuMemcpyAsync(dst, src, n, gpuMemcpyDeviceToHost, stream_->stream());
+    EIGEN_UNUSED_VARIABLE(err)
+    gpu_assert(err == gpuSuccess);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
+#ifndef EIGEN_GPU_COMPILE_PHASE
+    gpuError_t err = gpuMemsetAsync(buffer, c, n, stream_->stream());
+    EIGEN_UNUSED_VARIABLE(err)
+    gpu_assert(err == gpuSuccess);
+#else
+  eigen_assert(false && "The default device should be used instead to generate kernel code");
+#endif
+  }
+
+  EIGEN_STRONG_INLINE size_t numThreads() const {
+    // FIXME
+    return 32;
+  }
+
+  EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
+    // FIXME
+    return 48*1024;
+  }
+
+  EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
+    // We won't try to take advantage of the l2 cache for the time being, and
+    // there is no l3 cache on hip/cuda devices.
+    return firstLevelCacheSize();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void synchronize() const {
+#ifndef EIGEN_GPU_COMPILE_PHASE
+    gpuError_t err = gpuStreamSynchronize(stream_->stream());
+    if (err != gpuSuccess) {
+      std::cerr << "Error detected in GPU stream: "
+                << gpuGetErrorString(err)
+                << std::endl;
+      gpu_assert(err == gpuSuccess);
+    }
+#else
+    gpu_assert(false && "The default device should be used instead to generate kernel code");
+#endif
+  }
+
+  EIGEN_STRONG_INLINE int getNumGpuMultiProcessors() const {
+    return stream_->deviceProperties().multiProcessorCount;
+  }
+  EIGEN_STRONG_INLINE int maxGpuThreadsPerBlock() const {
+    return stream_->deviceProperties().maxThreadsPerBlock;
+  }
+  EIGEN_STRONG_INLINE int maxGpuThreadsPerMultiProcessor() const {
+    return stream_->deviceProperties().maxThreadsPerMultiProcessor;
+  }
+  EIGEN_STRONG_INLINE int sharedMemPerBlock() const {
+    return stream_->deviceProperties().sharedMemPerBlock;
+  }
+  EIGEN_STRONG_INLINE int majorDeviceVersion() const {
+    return stream_->deviceProperties().major;
+  }
+  EIGEN_STRONG_INLINE int minorDeviceVersion() const {
+    return stream_->deviceProperties().minor;
+  }
+
+  EIGEN_STRONG_INLINE int maxBlocks() const {
+    return max_blocks_;
+  }
+
+  // This function checks if the GPU runtime recorded an error for the
+  // underlying stream device.
+  inline bool ok() const {
+#ifdef EIGEN_GPUCC
+    gpuError_t error = gpuStreamQuery(stream_->stream());
+    return (error == gpuSuccess) || (error == gpuErrorNotReady);
+#else
+    return false;
+#endif
+  }
+
+ private:
+  const StreamInterface* stream_;
+  int max_blocks_;
+};
+
+#if defined(EIGEN_HIPCC)
+
+#define LAUNCH_GPU_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
+  hipLaunchKernelGGL(kernel, dim3(gridsize), dim3(blocksize), (sharedmem), (device).stream(), __VA_ARGS__); \
+  gpu_assert(hipGetLastError() == hipSuccess);
+
+#else
+ 
+#define LAUNCH_GPU_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...)             \
+  (kernel) <<< (gridsize), (blocksize), (sharedmem), (device).stream() >>> (__VA_ARGS__);   \
+  gpu_assert(cudaGetLastError() == cudaSuccess);
+
+#endif
+ 
+// FIXME: Should be device and kernel specific.
+#ifdef EIGEN_GPUCC
+static EIGEN_DEVICE_FUNC inline void setGpuSharedMemConfig(gpuSharedMemConfig config) {
+#ifndef EIGEN_GPU_COMPILE_PHASE
+  gpuError_t status = gpuDeviceSetSharedMemConfig(config);
+  EIGEN_UNUSED_VARIABLE(status)
+  gpu_assert(status == gpuSuccess);
+#else
+  EIGEN_UNUSED_VARIABLE(config)
+#endif
+}
+#endif
+
+}  // end namespace Eigen
+
+// undefine all the gpu* macros we defined at the beginning of the file
+#include "TensorGpuHipCudaUndefines.h"
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_GPU_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceSycl.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceSycl.h
new file mode 100644
index 0000000..df591c2
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -0,0 +1,1048 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#if defined(EIGEN_USE_SYCL) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H)
+#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
+#include <unordered_set>
+
+namespace Eigen {
+
+namespace TensorSycl {
+namespace internal {
+
+/// Cache all the device information needed
+struct SyclDeviceInfo {
+  SyclDeviceInfo(cl::sycl::queue queue)
+      : local_mem_type(
+            queue.get_device()
+                .template get_info<cl::sycl::info::device::local_mem_type>()),
+        max_work_item_sizes(
+            queue.get_device()
+                .template get_info<
+                    cl::sycl::info::device::max_work_item_sizes>()),
+        max_mem_alloc_size(
+            queue.get_device()
+                .template get_info<
+                    cl::sycl::info::device::max_mem_alloc_size>()),
+        max_compute_units(queue.get_device()
+                              .template get_info<
+                                  cl::sycl::info::device::max_compute_units>()),
+        max_work_group_size(
+            queue.get_device()
+                .template get_info<
+                    cl::sycl::info::device::max_work_group_size>()),
+        local_mem_size(
+            queue.get_device()
+                .template get_info<cl::sycl::info::device::local_mem_size>()),
+        platform_name(queue.get_device()
+                          .get_platform()
+                          .template get_info<cl::sycl::info::platform::name>()),
+        device_name(queue.get_device()
+                        .template get_info<cl::sycl::info::device::name>()),
+        device_vendor(
+            queue.get_device()
+                .template get_info<cl::sycl::info::device::vendor>()) {}
+
+  cl::sycl::info::local_mem_type local_mem_type;
+  cl::sycl::id<3> max_work_item_sizes;
+  unsigned long max_mem_alloc_size;
+  unsigned long max_compute_units;
+  unsigned long max_work_group_size;
+  size_t local_mem_size;
+  std::string platform_name;
+  std::string device_name;
+  std::string device_vendor;
+};
+
+}  // end namespace internal
+}  // end namespace TensorSycl
+
+typedef TensorSycl::internal::buffer_data_type_t buffer_scalar_t;
+// All devices (even AMD CPU with intel OpenCL runtime) that support OpenCL and
+// can consume SPIR or SPIRV can use the Eigen SYCL backend and consequently
+// TensorFlow via the Eigen SYCL Backend.
+EIGEN_STRONG_INLINE auto get_sycl_supported_devices()
+    -> decltype(cl::sycl::device::get_devices()) {
+#ifdef EIGEN_SYCL_USE_DEFAULT_SELECTOR
+  return {cl::sycl::device(cl::sycl::default_selector())};
+#else
+  std::vector<cl::sycl::device> supported_devices;
+  auto platform_list = cl::sycl::platform::get_platforms();
+  for (const auto &platform : platform_list) {
+    auto device_list = platform.get_devices();
+    auto platform_name =
+        platform.template get_info<cl::sycl::info::platform::name>();
+    std::transform(platform_name.begin(), platform_name.end(),
+                   platform_name.begin(), ::tolower);
+    for (const auto &device : device_list) {
+      auto vendor = device.template get_info<cl::sycl::info::device::vendor>();
+      std::transform(vendor.begin(), vendor.end(), vendor.begin(), ::tolower);
+      bool unsupported_condition =
+          (device.is_cpu() && platform_name.find("amd") != std::string::npos &&
+           vendor.find("apu") == std::string::npos) ||
+          (platform_name.find("experimental") != std::string::npos) ||
+          device.is_host();
+      if (!unsupported_condition) {
+        supported_devices.push_back(device);
+      }
+    }
+  }
+  return supported_devices;
+#endif
+}
+
+class QueueInterface {
+ public:
+  /// Creating device by using cl::sycl::selector or cl::sycl::device.
+  template <typename DeviceOrSelector>
+  explicit QueueInterface(
+      const DeviceOrSelector &dev_or_sel, cl::sycl::async_handler handler,
+      unsigned num_threads = std::thread::hardware_concurrency())
+      : m_queue(dev_or_sel, handler),
+#ifdef EIGEN_SYCL_USE_PROGRAM_CLASS
+        m_prog(m_queue.get_context(), get_sycl_supported_devices()),
+#endif
+        m_thread_pool(num_threads),
+        m_device_info(m_queue) {
+#ifdef EIGEN_SYCL_USE_PROGRAM_CLASS
+    m_prog.build_with_kernel_type<DeviceOrSelector>();
+    auto f = [&](cl::sycl::handler &cgh) {
+      cgh.single_task<DeviceOrSelector>(m_prog.get_kernel<DeviceOrSelector>(),
+                                        [=]() {})
+    };
+    EIGEN_SYCL_TRY_CATCH(m_queue.submit(f));
+#endif
+  }
+
+  template <typename DeviceOrSelector>
+  explicit QueueInterface(
+      const DeviceOrSelector &dev_or_sel,
+      unsigned num_threads = std::thread::hardware_concurrency())
+      : QueueInterface(dev_or_sel,
+                       [this](cl::sycl::exception_list l) {
+                         this->exception_caught_ = this->sycl_async_handler(l);
+                       },
+                       num_threads) {}
+
+#ifdef EIGEN_SYCL_USE_PROGRAM_CLASS
+  EIGEN_STRONG_INLINE cl::sycl::program &program() const { return m_prog; }
+#endif
+
+  /// Attach an existing buffer to the pointer map, Eigen will not reuse it
+  EIGEN_STRONG_INLINE void *attach_buffer(
+      cl::sycl::buffer<buffer_scalar_t, 1> &buf) const {
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    return static_cast<void *>(pMapper.add_pointer(buf));
+  }
+
+  /// Detach previously attached buffer
+  EIGEN_STRONG_INLINE void detach_buffer(void *p) const {
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    TensorSycl::internal::SYCLfree<false>(p, pMapper);
+  }
+
+  /// Allocating device pointer. This pointer is actually an 8 bytes host
+  /// pointer used as key to access the sycl device buffer. The reason is that
+  /// we cannot use device buffer as a pointer as a m_data in Eigen leafNode
+  /// expressions. So we create a key pointer to be used in Eigen expression
+  /// construction. When we convert the Eigen construction into the sycl
+  /// construction we use this pointer as a key in our buffer_map and we make
+  /// sure that we dedicate only one buffer only for this pointer. The device
+  /// pointer would be deleted by calling deallocate function.
+  EIGEN_STRONG_INLINE void *allocate(size_t num_bytes) const {
+#if EIGEN_MAX_ALIGN_BYTES > 0
+    size_t align = num_bytes % EIGEN_MAX_ALIGN_BYTES;
+    if (align > 0) {
+      num_bytes += EIGEN_MAX_ALIGN_BYTES - align;
+    }
+#endif
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    return TensorSycl::internal::SYCLmalloc(num_bytes, pMapper);
+  }
+
+  EIGEN_STRONG_INLINE void *allocate_temp(size_t num_bytes) const {
+#if EIGEN_MAX_ALIGN_BYTES > 0
+    size_t align = num_bytes % EIGEN_MAX_ALIGN_BYTES;
+    if (align > 0) {
+      num_bytes += EIGEN_MAX_ALIGN_BYTES - align;
+    }
+#endif
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+#ifndef EIGEN_SYCL_NO_REUSE_BUFFERS
+    if (scratch_buffers.empty()) {
+      return TensorSycl::internal::SYCLmalloc(num_bytes, pMapper);
+      ;
+    } else {
+      for (auto it = scratch_buffers.begin(); it != scratch_buffers.end();) {
+        auto buff = pMapper.get_buffer(*it);
+        if (buff.get_size() >= num_bytes) {
+          auto ptr = *it;
+          scratch_buffers.erase(it);
+          return ptr;
+        } else {
+          ++it;
+        }
+      }
+      return TensorSycl::internal::SYCLmalloc(num_bytes, pMapper);
+    }
+#else
+    return TensorSycl::internal::SYCLmalloc(num_bytes, pMapper);
+#endif
+  }
+  template <typename data_t>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorSycl::internal::RangeAccess<
+      cl::sycl::access::mode::read_write, data_t>
+  get(data_t *data) const {
+    return get_range_accessor<cl::sycl::access::mode::read_write, data_t>(data);
+  }
+  template <typename data_t>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE data_t *get(
+      TensorSycl::internal::RangeAccess<cl::sycl::access::mode::read_write,
+                                        data_t>
+          data) const {
+    return static_cast<data_t *>(data.get_virtual_pointer());
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_temp(void *p) const {
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+#ifndef EIGEN_SYCL_NO_REUSE_BUFFERS
+    scratch_buffers.insert(p);
+#else
+    TensorSycl::internal::SYCLfree(p, pMapper);
+#endif
+  }
+  template <cl::sycl::access::mode AcMd, typename T>
+  EIGEN_STRONG_INLINE void deallocate_temp(
+      const TensorSycl::internal::RangeAccess<AcMd, T> &p) const {
+    deallocate_temp(p.get_virtual_pointer());
+  }
+
+  /// This is used to deallocate the device pointer. p is used as a key inside
+  /// the map to find the device buffer and delete it.
+  EIGEN_STRONG_INLINE void deallocate(void *p) const {
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    TensorSycl::internal::SYCLfree(p, pMapper);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_all() const {
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    TensorSycl::internal::SYCLfreeAll(pMapper);
+#ifndef EIGEN_SYCL_NO_REUSE_BUFFERS
+    scratch_buffers.clear();
+#endif
+  }
+
+  /// The memcpyHostToDevice is used to copy the data from host to device
+  /// The destination pointer could be deleted before the copy happend which is
+  /// why a callback function is needed. By default if none is provided, the
+  /// function is blocking.
+  EIGEN_STRONG_INLINE void memcpyHostToDevice(
+      void *dst, const void *src, size_t n,
+      std::function<void()> callback) const {
+    static const auto write_mode = cl::sycl::access::mode::discard_write;
+    static const auto global_access = cl::sycl::access::target::global_buffer;
+    typedef cl::sycl::accessor<buffer_scalar_t, 1, write_mode, global_access>
+        write_accessor;
+    if (n == 0) {
+      if (callback) callback();
+      return;
+    }
+    n /= sizeof(buffer_scalar_t);
+    auto f = [&](cl::sycl::handler &cgh) {
+      write_accessor dst_acc = get_range_accessor<write_mode>(cgh, dst, n);
+      buffer_scalar_t const *ptr = static_cast<buffer_scalar_t const *>(src);
+      auto non_deleter = [](buffer_scalar_t const *) {};
+      std::shared_ptr<const buffer_scalar_t> s_ptr(ptr, non_deleter);
+      cgh.copy(s_ptr, dst_acc);
+    };
+    cl::sycl::event e;
+    EIGEN_SYCL_TRY_CATCH(e = m_queue.submit(f));
+    synchronize_and_callback(e, callback);
+  }
+
+  /// The memcpyDeviceToHost is used to copy the data from device to host.
+  /// The source pointer could be deleted before the copy happend which is
+  /// why a callback function is needed. By default if none is provided, the
+  /// function is blocking.
+  EIGEN_STRONG_INLINE void memcpyDeviceToHost(
+      void *dst, const void *src, size_t n,
+      std::function<void()> callback) const {
+    static const auto read_mode = cl::sycl::access::mode::read;
+    static const auto global_access = cl::sycl::access::target::global_buffer;
+    typedef cl::sycl::accessor<buffer_scalar_t, 1, read_mode, global_access>
+        read_accessor;
+    if (n == 0) {
+      if (callback) callback();
+      return;
+    }
+    n /= sizeof(buffer_scalar_t);
+    auto f = [&](cl::sycl::handler &cgh) {
+      read_accessor src_acc = get_range_accessor<read_mode>(cgh, src, n);
+      buffer_scalar_t *ptr = static_cast<buffer_scalar_t *>(dst);
+      auto non_deleter = [](buffer_scalar_t *) {};
+      std::shared_ptr<buffer_scalar_t> s_ptr(ptr, non_deleter);
+      cgh.copy(src_acc, s_ptr);
+    };
+    cl::sycl::event e;
+    EIGEN_SYCL_TRY_CATCH(e = m_queue.submit(f));
+    synchronize_and_callback(e, callback);
+  }
+
+  /// The memcpy function.
+  /// No callback is required here as both arguments are on the device
+  /// and SYCL can handle the dependency.
+  EIGEN_STRONG_INLINE void memcpy(void *dst, const void *src, size_t n) const {
+    static const auto read_mode = cl::sycl::access::mode::read;
+    static const auto write_mode = cl::sycl::access::mode::discard_write;
+    if (n == 0) {
+      return;
+    }
+    n /= sizeof(buffer_scalar_t);
+    auto f = [&](cl::sycl::handler &cgh) {
+      auto src_acc = get_range_accessor<read_mode>(cgh, src, n);
+      auto dst_acc = get_range_accessor<write_mode>(cgh, dst, n);
+      cgh.copy(src_acc, dst_acc);
+    };
+    cl::sycl::event e;
+    EIGEN_SYCL_TRY_CATCH(e = m_queue.submit(f));
+    async_synchronize(e);
+  }
+
+  /// the memset function.
+  /// No callback is required here as both arguments are on the device
+  /// and SYCL can handle the dependency.
+  EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
+    static const auto write_mode = cl::sycl::access::mode::discard_write;
+    if (n == 0) {
+      return;
+    }
+    n /= sizeof(buffer_scalar_t);
+    auto f = [&](cl::sycl::handler &cgh) {
+      auto dst_acc = get_range_accessor<write_mode>(cgh, data, n);
+      // The cast to uint8_t is here to match the behaviour of the standard
+      // memset. The cast to buffer_scalar_t is needed to match the type of the
+      // accessor (in case buffer_scalar_t is not uint8_t)
+      cgh.fill(dst_acc, static_cast<buffer_scalar_t>(static_cast<uint8_t>(c)));
+    };
+    cl::sycl::event e;
+    EIGEN_SYCL_TRY_CATCH(e = m_queue.submit(f));
+    async_synchronize(e);
+  }
+
+  /// Get a range accessor to the virtual pointer's device memory. This range
+  /// accessor will allow access to the memory from the pointer to the end of
+  /// the buffer.
+  ///
+  /// NOTE: Inside a kernel the range accessor will always be indexed from the
+  /// start of the buffer, so the offset in the accessor is only used by
+  /// methods like handler::copy and will not be available inside a kernel.
+  template <cl::sycl::access::mode AcMd, typename T>
+  EIGEN_STRONG_INLINE TensorSycl::internal::RangeAccess<AcMd, T>
+  get_range_accessor(const void *ptr) const {
+    static const auto global_access = cl::sycl::access::target::global_buffer;
+    static const auto is_place_holder = cl::sycl::access::placeholder::true_t;
+    typedef TensorSycl::internal::RangeAccess<AcMd, T> ret_type;
+    typedef const TensorSycl::internal::buffer_data_type_t *internal_ptr_t;
+
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+
+    auto original_buffer = pMapper.get_buffer(ptr);
+    const ptrdiff_t offset = pMapper.get_offset(ptr);
+    const ptrdiff_t typed_offset = offset / sizeof(T);
+    eigen_assert(typed_offset >= 0);
+    const auto typed_size = original_buffer.get_size() / sizeof(T);
+    auto buffer = original_buffer.template reinterpret<
+        typename Eigen::internal::remove_const<T>::type>(
+        cl::sycl::range<1>(typed_size));
+    const ptrdiff_t size = buffer.get_count() - typed_offset;
+    eigen_assert(size >= 0);
+    typedef cl::sycl::accessor<typename Eigen::internal::remove_const<T>::type,
+                               1, AcMd, global_access, is_place_holder>
+        placeholder_accessor_t;
+    const auto start_ptr = static_cast<internal_ptr_t>(ptr) - offset;
+    return ret_type(placeholder_accessor_t(buffer, cl::sycl::range<1>(size),
+                                           cl::sycl::id<1>(typed_offset)),
+                    static_cast<size_t>(typed_offset),
+                    reinterpret_cast<std::intptr_t>(start_ptr));
+  }
+
+  /// Get a range accessor to the virtual pointer's device memory with a
+  /// specified size.
+  template <cl::sycl::access::mode AcMd, typename Index>
+  EIGEN_STRONG_INLINE cl::sycl::accessor<
+      buffer_scalar_t, 1, AcMd, cl::sycl::access::target::global_buffer>
+  get_range_accessor(cl::sycl::handler &cgh, const void *ptr,
+                     const Index n_bytes) const {
+    static const auto global_access = cl::sycl::access::target::global_buffer;
+    eigen_assert(n_bytes >= 0);
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    auto buffer = pMapper.get_buffer(ptr);
+    const ptrdiff_t offset = pMapper.get_offset(ptr);
+    eigen_assert(offset >= 0);
+    eigen_assert(offset + n_bytes <= buffer.get_size());
+    return buffer.template get_access<AcMd, global_access>(
+        cgh, cl::sycl::range<1>(n_bytes), cl::sycl::id<1>(offset));
+  }
+
+  /// Creation of sycl accessor for a buffer. This function first tries to find
+  /// the buffer in the buffer_map. If found it gets the accessor from it, if
+  /// not, the function then adds an entry by creating a sycl buffer for that
+  /// particular pointer.
+  template <cl::sycl::access::mode AcMd>
+  EIGEN_STRONG_INLINE cl::sycl::accessor<
+      buffer_scalar_t, 1, AcMd, cl::sycl::access::target::global_buffer>
+  get_sycl_accessor(cl::sycl::handler &cgh, const void *ptr) const {
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    return pMapper.get_buffer(ptr)
+        .template get_access<AcMd, cl::sycl::access::target::global_buffer>(
+            cgh);
+  }
+
+  EIGEN_STRONG_INLINE cl::sycl::buffer<buffer_scalar_t, 1> get_sycl_buffer(
+      const void *ptr) const {
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    return pMapper.get_buffer(ptr);
+  }
+
+  EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
+    std::lock_guard<std::mutex> lock(pmapper_mutex_);
+    return pMapper.get_offset(ptr);
+  }
+
+  template <typename OutScalar, typename sycl_kernel, typename Lhs,
+            typename Rhs, typename OutPtr, typename Range, typename Index,
+            typename... T>
+  EIGEN_ALWAYS_INLINE void binary_kernel_launcher(const Lhs &lhs,
+                                                  const Rhs &rhs, OutPtr outptr,
+                                                  Range thread_range,
+                                                  Index scratchSize,
+                                                  T... var) const {
+    auto kernel_functor = [=](cl::sycl::handler &cgh) {
+      // binding the placeholder accessors to a commandgroup handler
+      lhs.bind(cgh);
+      rhs.bind(cgh);
+      outptr.bind(cgh);
+      typedef cl::sycl::accessor<OutScalar, 1,
+                                 cl::sycl::access::mode::read_write,
+                                 cl::sycl::access::target::local>
+          LocalAccessor;
+
+      LocalAccessor scratch(cl::sycl::range<1>(scratchSize), cgh);
+      cgh.parallel_for(
+#ifdef EIGEN_SYCL_USE_PROGRAM_CLASS
+          program().template get_kernel<sycl_kernel>(),
+#endif
+          thread_range, sycl_kernel(scratch, lhs, rhs, outptr, var...));
+    };
+    cl::sycl::event e;
+    EIGEN_SYCL_TRY_CATCH(e = m_queue.submit(kernel_functor));
+    async_synchronize(e);
+  }
+
+  template <typename OutScalar, typename sycl_kernel, typename InPtr,
+            typename OutPtr, typename Range, typename Index, typename... T>
+  EIGEN_ALWAYS_INLINE void unary_kernel_launcher(const InPtr &inptr,
+                                                 OutPtr &outptr,
+                                                 Range thread_range,
+                                                 Index scratchSize,
+                                                 T... var) const {
+    auto kernel_functor = [=](cl::sycl::handler &cgh) {
+      // binding the placeholder accessors to a commandgroup handler
+      inptr.bind(cgh);
+      outptr.bind(cgh);
+      typedef cl::sycl::accessor<OutScalar, 1,
+                                 cl::sycl::access::mode::read_write,
+                                 cl::sycl::access::target::local>
+          LocalAccessor;
+
+      LocalAccessor scratch(cl::sycl::range<1>(scratchSize), cgh);
+      cgh.parallel_for(
+#ifdef EIGEN_SYCL_USE_PROGRAM_CLASS
+          program().template get_kernel<sycl_kernel>(),
+#endif
+          thread_range, sycl_kernel(scratch, inptr, outptr, var...));
+    };
+    cl::sycl::event e;
+    EIGEN_SYCL_TRY_CATCH(e = m_queue.submit(kernel_functor));
+    async_synchronize(e);
+  }
+
+    template <typename OutScalar, typename sycl_kernel, typename InPtr,
+           typename Range, typename Index, typename... T>
+  EIGEN_ALWAYS_INLINE void nullary_kernel_launcher(const InPtr &inptr,
+                                                 Range thread_range,
+                                                 Index scratchSize,
+                                                 T... var) const {
+    auto kernel_functor = [=](cl::sycl::handler &cgh) {
+      // binding the placeholder accessors to a commandgroup handler
+      inptr.bind(cgh);
+      typedef cl::sycl::accessor<OutScalar, 1,
+                                 cl::sycl::access::mode::read_write,
+                                 cl::sycl::access::target::local>
+          LocalAccessor;
+
+      LocalAccessor scratch(cl::sycl::range<1>(scratchSize), cgh);
+      cgh.parallel_for(
+#ifdef EIGEN_SYCL_USE_PROGRAM_CLASS
+          program().template get_kernel<sycl_kernel>(),
+#endif
+          thread_range, sycl_kernel(scratch, inptr, var...));
+    };
+    cl::sycl::event e;
+    EIGEN_SYCL_TRY_CATCH(e = m_queue.submit(kernel_functor));
+    async_synchronize(e);
+  }
+
+
+  EIGEN_STRONG_INLINE void synchronize() const {
+#ifdef EIGEN_EXCEPTIONS
+    m_queue.wait_and_throw();
+#else
+    m_queue.wait();
+#endif
+  }
+
+
+  EIGEN_STRONG_INLINE void async_synchronize(cl::sycl::event e) const {
+    set_latest_event(e);
+#ifndef EIGEN_SYCL_ASYNC_EXECUTION
+    synchronize();
+#endif
+  }
+
+  template <typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(Index n, Index &tileSize,
+                                              Index &rng, Index &GRange) const {
+    tileSize = static_cast<Index>(getNearestPowerOfTwoWorkGroupSize());
+    tileSize = std::min(static_cast<Index>(EIGEN_SYCL_LOCAL_THREAD_DIM0 *
+                                           EIGEN_SYCL_LOCAL_THREAD_DIM1),
+                        static_cast<Index>(tileSize));
+    rng = n;
+    if (rng == 0) rng = static_cast<Index>(1);
+    GRange = rng;
+    if (tileSize > GRange)
+      tileSize = GRange;
+    else if (GRange > tileSize) {
+      Index xMode = static_cast<Index>(GRange % tileSize);
+      if (xMode != 0) GRange += static_cast<Index>(tileSize - xMode);
+    }
+  }
+
+  /// This is used to prepare the number of threads and also the number of
+  /// threads per block for sycl kernels
+  template <typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(
+      const std::array<Index, 2> &input_dim, cl::sycl::range<2> &global_range,
+      cl::sycl::range<2> &local_range) const {
+    std::array<Index, 2> input_range = input_dim;
+    Index max_workgroup_Size =
+        static_cast<Index>(getNearestPowerOfTwoWorkGroupSize());
+    max_workgroup_Size =
+        std::min(static_cast<Index>(EIGEN_SYCL_LOCAL_THREAD_DIM0 *
+                                    EIGEN_SYCL_LOCAL_THREAD_DIM1),
+                 static_cast<Index>(max_workgroup_Size));
+    Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
+    local_range[1] =
+        static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2 / 2)));
+    input_range[1] = input_dim[1];
+    if (input_range[1] == 0) input_range[1] = static_cast<Index>(1);
+    global_range[1] = input_range[1];
+    if (local_range[1] > global_range[1])
+      local_range[1] = global_range[1];
+    else if (global_range[1] > local_range[1]) {
+      Index xMode = static_cast<Index>(global_range[1] % local_range[1]);
+      if (xMode != 0)
+        global_range[1] += static_cast<Index>(local_range[1] - xMode);
+    }
+    local_range[0] = static_cast<Index>(max_workgroup_Size / local_range[1]);
+    input_range[0] = input_dim[0];
+    if (input_range[0] == 0) input_range[0] = static_cast<Index>(1);
+    global_range[0] = input_range[0];
+    if (local_range[0] > global_range[0])
+      local_range[0] = global_range[0];
+    else if (global_range[0] > local_range[0]) {
+      Index xMode = static_cast<Index>(global_range[0] % local_range[0]);
+      if (xMode != 0)
+        global_range[0] += static_cast<Index>(local_range[0] - xMode);
+    }
+  }
+
+  /// This is used to prepare the number of threads and also the number of
+  /// threads per block for sycl kernels
+  template <typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(
+      const std::array<Index, 3> &input_dim, cl::sycl::range<3> &global_range,
+      cl::sycl::range<3> &local_range) const {
+    std::array<Index, 3> input_range = input_dim;
+    Index max_workgroup_Size =
+        static_cast<Index>(getNearestPowerOfTwoWorkGroupSize());
+    max_workgroup_Size =
+        std::min(static_cast<Index>(EIGEN_SYCL_LOCAL_THREAD_DIM0 *
+                                    EIGEN_SYCL_LOCAL_THREAD_DIM1),
+                 static_cast<Index>(max_workgroup_Size));
+    Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
+    local_range[2] =
+        static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2 / 3)));
+    input_range[2] = input_dim[2];
+    if (input_range[2] == 0) input_range[1] = static_cast<Index>(1);
+    global_range[2] = input_range[2];
+    if (local_range[2] > global_range[2])
+      local_range[2] = global_range[2];
+    else if (global_range[2] > local_range[2]) {
+      Index xMode = static_cast<Index>(global_range[2] % local_range[2]);
+      if (xMode != 0)
+        global_range[2] += static_cast<Index>(local_range[2] - xMode);
+    }
+    pow_of_2 = static_cast<Index>(
+        std::log2(static_cast<Index>(max_workgroup_Size / local_range[2])));
+    local_range[1] =
+        static_cast<Index>(std::pow(2, static_cast<Index>(pow_of_2 / 2)));
+    input_range[1] = input_dim[1];
+    if (input_range[1] == 0) input_range[1] = static_cast<Index>(1);
+    global_range[1] = input_range[1];
+    if (local_range[1] > global_range[1])
+      local_range[1] = global_range[1];
+    else if (global_range[1] > local_range[1]) {
+      Index xMode = static_cast<Index>(global_range[1] % local_range[1]);
+      if (xMode != 0)
+        global_range[1] += static_cast<Index>(local_range[1] - xMode);
+    }
+    local_range[0] = static_cast<Index>(max_workgroup_Size /
+                                        (local_range[1] * local_range[2]));
+    input_range[0] = input_dim[0];
+    if (input_range[0] == 0) input_range[0] = static_cast<Index>(1);
+    global_range[0] = input_range[0];
+    if (local_range[0] > global_range[0])
+      local_range[0] = global_range[0];
+    else if (global_range[0] > local_range[0]) {
+      Index xMode = static_cast<Index>(global_range[0] % local_range[0]);
+      if (xMode != 0)
+        global_range[0] += static_cast<Index>(local_range[0] - xMode);
+    }
+  }
+
+  EIGEN_STRONG_INLINE bool has_local_memory() const {
+#if !defined(EIGEN_SYCL_LOCAL_MEM) && defined(EIGEN_SYCL_NO_LOCAL_MEM)
+    return false;
+#elif defined(EIGEN_SYCL_LOCAL_MEM) && !defined(EIGEN_SYCL_NO_LOCAL_MEM)
+    return true;
+#else
+    return m_device_info.local_mem_type ==
+           cl::sycl::info::local_mem_type::local;
+#endif
+  }
+
+  EIGEN_STRONG_INLINE unsigned long max_buffer_size() const {
+    return m_device_info.max_mem_alloc_size;
+  }
+
+  EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
+    return m_device_info.max_compute_units;
+  }
+
+  EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
+    return m_device_info.max_work_group_size;
+  }
+
+  EIGEN_STRONG_INLINE cl::sycl::id<3> maxWorkItemSizes() const {
+    return m_device_info.max_work_item_sizes;
+  }
+
+  /// No need for sycl it should act the same as CPU version
+  EIGEN_STRONG_INLINE int majorDeviceVersion() const { return 1; }
+
+  EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
+    // OpenCL doesnot have such concept
+    return 2;
+  }
+
+  EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
+    return m_device_info.local_mem_size;
+  }
+
+  // This function returns the nearest power of 2 Work-group size which is <=
+  // maximum device workgroup size.
+  EIGEN_STRONG_INLINE size_t getNearestPowerOfTwoWorkGroupSize() const {
+    return getPowerOfTwo(m_device_info.max_work_group_size, false);
+  }
+
+  EIGEN_STRONG_INLINE std::string getPlatformName() const {
+    return m_device_info.platform_name;
+  }
+
+  EIGEN_STRONG_INLINE std::string getDeviceName() const {
+    return m_device_info.device_name;
+  }
+
+  EIGEN_STRONG_INLINE std::string getDeviceVendor() const {
+    return m_device_info.device_vendor;
+  }
+
+  // This function returns the nearest power of 2
+  // if roundup is true returns result>=wgsize
+  // else it return result <= wgsize
+  EIGEN_STRONG_INLINE size_t getPowerOfTwo(size_t wGSize, bool roundUp) const {
+    if (roundUp) --wGSize;
+    wGSize |= (wGSize >> 1);
+    wGSize |= (wGSize >> 2);
+    wGSize |= (wGSize >> 4);
+    wGSize |= (wGSize >> 8);
+    wGSize |= (wGSize >> 16);
+#if EIGEN_ARCH_x86_64 || EIGEN_ARCH_ARM64 || EIGEN_OS_WIN64
+    wGSize |= (wGSize >> 32);
+#endif
+    return ((!roundUp) ? (wGSize - (wGSize >> 1)) : ++wGSize);
+  }
+
+  EIGEN_STRONG_INLINE cl::sycl::queue &sycl_queue() const { return m_queue; }
+
+  // This function checks if the runtime recorded an error for the
+  // underlying stream device.
+  EIGEN_STRONG_INLINE bool ok() const {
+    if (!exception_caught_) {
+      synchronize();
+    }
+    return !exception_caught_;
+  }
+
+  EIGEN_STRONG_INLINE cl::sycl::event get_latest_event() const {
+#ifdef EIGEN_SYCL_STORE_LATEST_EVENT
+    std::lock_guard<std::mutex> lock(event_mutex_);
+    return latest_events_[std::this_thread::get_id()];
+#else
+    eigen_assert(false);
+    return cl::sycl::event();
+#endif
+  }
+
+  // destructor
+  ~QueueInterface() {
+    pMapper.clear();
+#ifndef EIGEN_SYCL_NO_REUSE_BUFFERS
+    scratch_buffers.clear();
+#endif
+  }
+
+ protected:
+  EIGEN_STRONG_INLINE void set_latest_event(cl::sycl::event e) const {
+#ifdef EIGEN_SYCL_STORE_LATEST_EVENT
+    std::lock_guard<std::mutex> lock(event_mutex_);
+    latest_events_[std::this_thread::get_id()] = e;
+#else
+    EIGEN_UNUSED_VARIABLE(e);
+#endif
+  }
+
+  void synchronize_and_callback(cl::sycl::event e,
+                                const std::function<void()> &callback) const {
+    set_latest_event(e);
+    if (callback) {
+      auto callback_ = [=]() {
+#ifdef EIGEN_EXCEPTIONS
+        cl::sycl::event(e).wait_and_throw();
+#else
+        cl::sycl::event(e).wait();
+#endif
+        callback();
+      };
+      m_thread_pool.Schedule(std::move(callback_));
+    } else {
+#ifdef EIGEN_EXCEPTIONS
+      m_queue.wait_and_throw();
+#else
+      m_queue.wait();
+#endif
+    }
+  }
+
+  bool sycl_async_handler(cl::sycl::exception_list exceptions) const {
+    bool exception_caught = false;
+    for (const auto &e : exceptions) {
+      if (e) {
+        exception_caught = true;
+        EIGEN_THROW_X(e);
+      }
+    }
+    return exception_caught;
+  }
+
+  /// class members:
+  bool exception_caught_ = false;
+
+  mutable std::mutex pmapper_mutex_;
+
+#ifdef EIGEN_SYCL_STORE_LATEST_EVENT
+  mutable std::mutex event_mutex_;
+  mutable std::unordered_map<std::thread::id, cl::sycl::event> latest_events_;
+#endif
+
+  /// std::map is the container used to make sure that we create only one buffer
+  /// per pointer. The lifespan of the buffer now depends on the lifespan of
+  /// SyclDevice. If a non-read-only pointer is needed to be accessed on the
+  /// host we should manually deallocate it.
+  mutable TensorSycl::internal::PointerMapper pMapper;
+#ifndef EIGEN_SYCL_NO_REUSE_BUFFERS
+  mutable std::unordered_set<void *> scratch_buffers;
+#endif
+  /// sycl queue
+  mutable cl::sycl::queue m_queue;
+#ifdef EIGEN_SYCL_USE_PROGRAM_CLASS
+  mutable cl::sycl::program m_prog;
+#endif
+
+  /// The thread pool is used to wait on events and call callbacks
+  /// asynchronously
+  mutable Eigen::ThreadPool m_thread_pool;
+
+  const TensorSycl::internal::SyclDeviceInfo m_device_info;
+};
+
+struct SyclDeviceBase {
+  /// QueueInterface is not owned. it is the caller's responsibility to destroy
+  /// it
+  const QueueInterface *m_queue_stream;
+  explicit SyclDeviceBase(const QueueInterface *queue_stream)
+      : m_queue_stream(queue_stream) {}
+  EIGEN_STRONG_INLINE const QueueInterface *queue_stream() const {
+    return m_queue_stream;
+  }
+};
+
+// Here is a sycl device struct which accept the sycl queue interface
+// as an input
+struct SyclDevice : public SyclDeviceBase {
+  explicit SyclDevice(const QueueInterface *queue_stream)
+      : SyclDeviceBase(queue_stream) {}
+
+  // this is the accessor used to construct the evaluator
+  template <cl::sycl::access::mode AcMd, typename T>
+  EIGEN_STRONG_INLINE TensorSycl::internal::RangeAccess<AcMd, T>
+  get_range_accessor(const void *ptr) const {
+    return queue_stream()->template get_range_accessor<AcMd, T>(ptr);
+  }
+
+  // get sycl accessor
+  template <cl::sycl::access::mode AcMd>
+  EIGEN_STRONG_INLINE cl::sycl::accessor<
+      buffer_scalar_t, 1, AcMd, cl::sycl::access::target::global_buffer>
+  get_sycl_accessor(cl::sycl::handler &cgh, const void *ptr) const {
+    return queue_stream()->template get_sycl_accessor<AcMd>(cgh, ptr);
+  }
+
+  /// Accessing the created sycl device buffer for the device pointer
+  EIGEN_STRONG_INLINE cl::sycl::buffer<buffer_scalar_t, 1> get_sycl_buffer(
+      const void *ptr) const {
+    return queue_stream()->get_sycl_buffer(ptr);
+  }
+
+  /// This is used to prepare the number of threads and also the number of
+  /// threads per block for sycl kernels
+  template <typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(Index n, Index &tileSize,
+                                              Index &rng, Index &GRange) const {
+    queue_stream()->parallel_for_setup(n, tileSize, rng, GRange);
+  }
+
+  /// This is used to prepare the number of threads and also the number of
+  /// threads per block for sycl kernels
+  template <typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(
+      const std::array<Index, 2> &input_dim, cl::sycl::range<2> &global_range,
+      cl::sycl::range<2> &local_range) const {
+    queue_stream()->parallel_for_setup(input_dim, global_range, local_range);
+  }
+
+  /// This is used to prepare the number of threads and also the number of
+  /// threads per block for sycl kernels
+  template <typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(
+      const std::array<Index, 3> &input_dim, cl::sycl::range<3> &global_range,
+      cl::sycl::range<3> &local_range) const {
+    queue_stream()->parallel_for_setup(input_dim, global_range, local_range);
+  }
+
+  /// allocate device memory
+  EIGEN_STRONG_INLINE void *allocate(size_t num_bytes) const {
+    return queue_stream()->allocate(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void *allocate_temp(size_t num_bytes) const {
+    return queue_stream()->allocate_temp(num_bytes);
+  }
+
+  /// deallocate device memory
+  EIGEN_STRONG_INLINE void deallocate(void *p) const {
+    queue_stream()->deallocate(p);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_temp(void *buffer) const {
+    queue_stream()->deallocate_temp(buffer);
+  }
+  template <cl::sycl::access::mode AcMd, typename T>
+  EIGEN_STRONG_INLINE void deallocate_temp(
+      const TensorSycl::internal::RangeAccess<AcMd, T> &buffer) const {
+    queue_stream()->deallocate_temp(buffer);
+  }
+  EIGEN_STRONG_INLINE void deallocate_all() const {
+    queue_stream()->deallocate_all();
+  }
+
+  template <typename data_t>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorSycl::internal::RangeAccess<
+      cl::sycl::access::mode::read_write, data_t>
+  get(data_t *data) const {
+    return queue_stream()->get(data);
+  }
+  template <typename data_t>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE data_t *get(
+      TensorSycl::internal::RangeAccess<cl::sycl::access::mode::read_write,
+                                        data_t>
+          data) const {
+    return queue_stream()->get(data);
+  }
+
+  /// attach existing buffer
+  EIGEN_STRONG_INLINE void *attach_buffer(
+      cl::sycl::buffer<buffer_scalar_t, 1> &buf) const {
+    return queue_stream()->attach_buffer(buf);
+  }
+  /// detach buffer
+  EIGEN_STRONG_INLINE void detach_buffer(void *p) const {
+    queue_stream()->detach_buffer(p);
+  }
+  EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
+    return queue_stream()->get_offset(ptr);
+  }
+
+  // some runtime conditions that can be applied here
+  EIGEN_STRONG_INLINE bool isDeviceSuitable() const { return true; }
+
+  /// memcpyHostToDevice
+  template <typename Index>
+  EIGEN_STRONG_INLINE void memcpyHostToDevice(
+      Index *dst, const Index *src, size_t n,
+      std::function<void()> callback = {}) const {
+    queue_stream()->memcpyHostToDevice(dst, src, n, callback);
+  }
+  /// memcpyDeviceToHost
+  template <typename Index>
+  EIGEN_STRONG_INLINE void memcpyDeviceToHost(
+      void *dst, const Index *src, size_t n,
+      std::function<void()> callback = {}) const {
+    queue_stream()->memcpyDeviceToHost(dst, src, n, callback);
+  }
+  /// the memcpy function
+  template <typename Index>
+  EIGEN_STRONG_INLINE void memcpy(void *dst, const Index *src, size_t n) const {
+    queue_stream()->memcpy(dst, src, n);
+  }
+  /// the memset function
+  EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
+    queue_stream()->memset(data, c, n);
+  }
+  /// returning the sycl queue
+  EIGEN_STRONG_INLINE cl::sycl::queue &sycl_queue() const {
+    return queue_stream()->sycl_queue();
+  }
+#ifdef EIGEN_SYCL_USE_PROGRAM_CLASS
+  EIGEN_STRONG_INLINE cl::sycl::program &program() const {
+    return queue_stream()->program();
+  }
+#endif
+
+  EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const { return 48 * 1024; }
+
+  EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
+    // We won't try to take advantage of the l2 cache for the time being, and
+    // there is no l3 cache on sycl devices.
+    return firstLevelCacheSize();
+  }
+  EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
+    return queue_stream()->getNumSyclMultiProcessors();
+  }
+  EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
+    return queue_stream()->maxSyclThreadsPerBlock();
+  }
+  EIGEN_STRONG_INLINE cl::sycl::id<3> maxWorkItemSizes() const {
+    return queue_stream()->maxWorkItemSizes();
+  }
+  EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
+    // OpenCL doesnot have such concept
+    return queue_stream()->maxSyclThreadsPerMultiProcessor();
+  }
+  EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
+    return queue_stream()->sharedMemPerBlock();
+  }
+  EIGEN_STRONG_INLINE size_t getNearestPowerOfTwoWorkGroupSize() const {
+    return queue_stream()->getNearestPowerOfTwoWorkGroupSize();
+  }
+
+  EIGEN_STRONG_INLINE size_t getPowerOfTwo(size_t val, bool roundUp) const {
+    return queue_stream()->getPowerOfTwo(val, roundUp);
+  }
+  /// No need for sycl it should act the same as CPU version
+  EIGEN_STRONG_INLINE int majorDeviceVersion() const {
+    return queue_stream()->majorDeviceVersion();
+  }
+
+  EIGEN_STRONG_INLINE void synchronize() const {
+    queue_stream()->synchronize();
+  }
+  EIGEN_STRONG_INLINE void async_synchronize(
+      cl::sycl::event e = cl::sycl::event()) const {
+    queue_stream()->async_synchronize(e);
+  }
+  EIGEN_STRONG_INLINE cl::sycl::event get_latest_event() const {
+    return queue_stream()->get_latest_event();
+  }
+
+  // This function checks if the runtime recorded an error for the
+  // underlying stream device.
+  EIGEN_STRONG_INLINE bool ok() const { return queue_stream()->ok(); }
+
+  EIGEN_STRONG_INLINE bool has_local_memory() const {
+    return queue_stream()->has_local_memory();
+  }
+  EIGEN_STRONG_INLINE long max_buffer_size() const {
+    return queue_stream()->max_buffer_size();
+  }
+  EIGEN_STRONG_INLINE std::string getPlatformName() const {
+    return queue_stream()->getPlatformName();
+  }
+  EIGEN_STRONG_INLINE std::string getDeviceName() const {
+    return queue_stream()->getDeviceName();
+  }
+  EIGEN_STRONG_INLINE std::string getDeviceVendor() const {
+    return queue_stream()->getDeviceVendor();
+  }
+  template <typename OutScalar, typename KernelType, typename... T>
+  EIGEN_ALWAYS_INLINE void binary_kernel_launcher(T... var) const {
+    queue_stream()->template binary_kernel_launcher<OutScalar, KernelType>(
+        var...);
+  }
+  template <typename OutScalar, typename KernelType, typename... T>
+  EIGEN_ALWAYS_INLINE void unary_kernel_launcher(T... var) const {
+    queue_stream()->template unary_kernel_launcher<OutScalar, KernelType>(
+        var...);
+  }
+
+  template <typename OutScalar, typename KernelType, typename... T>
+  EIGEN_ALWAYS_INLINE void nullary_kernel_launcher(T... var) const {
+    queue_stream()->template nullary_kernel_launcher<OutScalar, KernelType>(
+        var...);
+  }
+};
+}  // end namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceThreadPool.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceThreadPool.h
new file mode 100644
index 0000000..e524b53
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorDeviceThreadPool.h
@@ -0,0 +1,409 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#if defined(EIGEN_USE_THREADS) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_THREAD_POOL_H)
+#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_THREAD_POOL_H
+
+namespace Eigen {
+
+// Runs an arbitrary function and then calls Notify() on the passed in
+// Notification.
+template <typename Function, typename... Args> struct FunctionWrapperWithNotification
+{
+  static void run(Notification* n, Function f, Args... args) {
+    f(args...);
+    if (n) {
+      n->Notify();
+    }
+  }
+};
+
+template <typename Function, typename... Args> struct FunctionWrapperWithBarrier
+{
+  static void run(Barrier* b, Function f, Args... args) {
+    f(args...);
+    if (b) {
+      b->Notify();
+    }
+  }
+};
+
+template <typename SyncType>
+static EIGEN_STRONG_INLINE void wait_until_ready(SyncType* n) {
+  if (n) {
+    n->Wait();
+  }
+}
+
+// An abstract interface to a device specific memory allocator.
+class Allocator {
+ public:
+  virtual ~Allocator() {}
+  virtual void* allocate(size_t num_bytes) const = 0;
+  virtual void deallocate(void* buffer) const = 0;
+};
+
+// Build a thread pool device on top the an existing pool of threads.
+struct ThreadPoolDevice {
+  // The ownership of the thread pool remains with the caller.
+  ThreadPoolDevice(ThreadPoolInterface* pool, int num_cores, Allocator* allocator = nullptr)
+      : pool_(pool), num_threads_(num_cores), allocator_(allocator) { }
+
+  EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
+    return allocator_ ? allocator_->allocate(num_bytes)
+        : internal::aligned_malloc(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
+    if (allocator_) {
+      allocator_->deallocate(buffer);
+    } else {
+      internal::aligned_free(buffer);
+    }
+  }
+
+    EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
+    return allocate(num_bytes);
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
+    deallocate(buffer);
+  }
+
+  template<typename Type>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Type get(Type data) const {
+    return data;
+  }
+
+  EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
+#ifdef __ANDROID__
+    ::memcpy(dst, src, n);
+#else
+    // TODO(rmlarsen): Align blocks on cache lines.
+    // We have observed that going beyond 4 threads usually just wastes
+    // CPU cycles due to the threads competing for memory bandwidth, so we
+    // statically schedule at most 4 block copies here.
+    const size_t kMinBlockSize = 32768;
+    const size_t num_threads = CostModel::numThreads(n, TensorOpCost(1.0, 1.0, 0), 4);
+    if (n <= kMinBlockSize || num_threads < 2) {
+      ::memcpy(dst, src, n);
+    } else {
+      const char* src_ptr = static_cast<const char*>(src);
+      char* dst_ptr = static_cast<char*>(dst);
+      const size_t blocksize = (n + (num_threads - 1)) / num_threads;
+      Barrier barrier(static_cast<int>(num_threads - 1));
+      // Launch the last 3 blocks on worker threads.
+      for (size_t i = 1; i < num_threads; ++i) {
+        enqueue_with_barrier(&barrier, [n, i, src_ptr, dst_ptr, blocksize] {
+          ::memcpy(dst_ptr + i * blocksize, src_ptr + i * blocksize,
+                   numext::mini(blocksize, n - (i * blocksize)));
+        });
+      }
+      // Launch the first block on the main thread.
+      ::memcpy(dst_ptr, src_ptr, blocksize);
+      barrier.Wait();
+    }
+#endif
+  }
+  EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
+    memcpy(dst, src, n);
+  }
+  EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const {
+    memcpy(dst, src, n);
+  }
+
+  EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
+    ::memset(buffer, c, n);
+  }
+
+  EIGEN_STRONG_INLINE int numThreads() const {
+    return num_threads_;
+  }
+
+  // Number of theads available in the underlying thread pool. This number can
+  // be different from the value returned by numThreads().
+  EIGEN_STRONG_INLINE int numThreadsInPool() const {
+    return pool_->NumThreads();
+  }
+
+  EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
+    return l1CacheSize();
+  }
+
+  EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
+    // The l3 cache size is shared between all the cores.
+    return l3CacheSize() / num_threads_;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const {
+    // Should return an enum that encodes the ISA supported by the CPU
+    return 1;
+  }
+
+  template <class Function, class... Args>
+  EIGEN_STRONG_INLINE Notification* enqueue(Function&& f,
+                                            Args&&... args) const {
+    Notification* n = new Notification();
+    pool_->Schedule(
+        std::bind(&FunctionWrapperWithNotification<Function, Args...>::run, n,
+                  std::move(f), args...));
+    return n;
+  }
+
+  template <class Function, class... Args>
+  EIGEN_STRONG_INLINE void enqueue_with_barrier(Barrier* b, Function&& f,
+                                                Args&&... args) const {
+    pool_->Schedule(
+        std::bind(&FunctionWrapperWithBarrier<Function, Args...>::run, b,
+                  std::move(f), args...));
+  }
+
+  template <class Function, class... Args>
+  EIGEN_STRONG_INLINE void enqueueNoNotification(Function&& f,
+                                                 Args&&... args) const {
+    if (sizeof...(args) > 0) {
+      pool_->Schedule(std::bind(std::move(f), args...));
+    } else {
+      pool_->Schedule(std::move(f));
+    }
+  }
+
+  // Returns a logical thread index between 0 and pool_->NumThreads() - 1 if
+  // called from one of the threads in pool_. Returns -1 otherwise.
+  EIGEN_STRONG_INLINE int currentThreadId() const {
+    return pool_->CurrentThreadId();
+  }
+
+  // WARNING: This function is synchronous and will block the calling thread.
+  //
+  // Synchronous parallelFor executes f with [0, n) arguments in parallel and
+  // waits for completion. F accepts a half-open interval [first, last). Block
+  // size is chosen based on the iteration cost and resulting parallel
+  // efficiency. If block_align is not nullptr, it is called to round up the
+  // block size.
+  void parallelFor(Index n, const TensorOpCost& cost,
+                   std::function<Index(Index)> block_align,
+                   std::function<void(Index, Index)> f) const {
+    if (EIGEN_PREDICT_FALSE(n <= 0)){
+      return;
+    // Compute small problems directly in the caller thread.
+    } else if (n == 1 || numThreads() == 1 ||
+               CostModel::numThreads(n, cost, static_cast<int>(numThreads())) == 1) {
+      f(0, n);
+      return;
+    }
+
+    // Compute block size and total count of blocks.
+    ParallelForBlock block = CalculateParallelForBlock(n, cost, block_align);
+
+    // Recursively divide size into halves until we reach block_size.
+    // Division code rounds mid to block_size, so we are guaranteed to get
+    // block_count leaves that do actual computations.
+    Barrier barrier(static_cast<unsigned int>(block.count));
+    std::function<void(Index, Index)> handleRange;
+    handleRange = [=, &handleRange, &barrier, &f](Index firstIdx,
+                                                  Index lastIdx) {
+      while (lastIdx - firstIdx > block.size) {
+        // Split into halves and schedule the second half on a different thread.
+        const Index midIdx = firstIdx + divup((lastIdx - firstIdx) / 2, block.size) * block.size;
+        pool_->Schedule([=, &handleRange]() { handleRange(midIdx, lastIdx); });
+        lastIdx = midIdx;
+      }
+      // Single block or less, execute directly.
+      f(firstIdx, lastIdx);
+      barrier.Notify();
+    };
+
+    if (block.count <= numThreads()) {
+      // Avoid a thread hop by running the root of the tree and one block on the
+      // main thread.
+      handleRange(0, n);
+    } else {
+      // Execute the root in the thread pool to avoid running work on more than
+      // numThreads() threads.
+      pool_->Schedule([=, &handleRange]() { handleRange(0, n); });
+    }
+
+    barrier.Wait();
+  }
+
+  // Convenience wrapper for parallelFor that does not align blocks.
+  void parallelFor(Index n, const TensorOpCost& cost,
+                   std::function<void(Index, Index)> f) const {
+    parallelFor(n, cost, nullptr, std::move(f));
+  }
+
+  // WARNING: This function is asynchronous and will not block the calling thread.
+  //
+  // Asynchronous parallelFor executes f with [0, n) arguments in parallel
+  // without waiting for completion. When the last block finished, it will call
+  // 'done' callback. F accepts a half-open interval [first, last). Block size
+  // is chosen based on the iteration cost and resulting parallel efficiency. If
+  // block_align is not nullptr, it is called to round up the block size.
+  void parallelForAsync(Index n, const TensorOpCost& cost,
+                        std::function<Index(Index)> block_align,
+                        std::function<void(Index, Index)> f,
+                        std::function<void()> done) const {
+    // Compute small problems directly in the caller thread.
+    if (n <= 1 || numThreads() == 1 ||
+        CostModel::numThreads(n, cost, static_cast<int>(numThreads())) == 1) {
+      f(0, n);
+      done();
+      return;
+    }
+
+    // Compute block size and total count of blocks.
+    ParallelForBlock block = CalculateParallelForBlock(n, cost, block_align);
+
+    ParallelForAsyncContext* const ctx =
+        new ParallelForAsyncContext(block.count, std::move(f), std::move(done));
+
+    // Recursively divide size into halves until we reach block_size.
+    // Division code rounds mid to block_size, so we are guaranteed to get
+    // block_count leaves that do actual computations.
+    ctx->handle_range = [this, ctx, block](Index firstIdx, Index lastIdx) {
+      while (lastIdx - firstIdx > block.size) {
+        // Split into halves and schedule the second half on a different thread.
+        const Index midIdx = firstIdx + divup((lastIdx - firstIdx) / 2, block.size) * block.size;
+        pool_->Schedule(
+            [ctx, midIdx, lastIdx]() { ctx->handle_range(midIdx, lastIdx); });
+        lastIdx = midIdx;
+      }
+
+      // Single block or less, execute directly.
+      ctx->f(firstIdx, lastIdx);
+
+      // Delete async context if it was the last block.
+      if (ctx->count.fetch_sub(1) == 1) delete ctx;
+    };
+
+    if (block.count <= numThreads()) {
+      // Avoid a thread hop by running the root of the tree and one block on the
+      // main thread.
+      ctx->handle_range(0, n);
+    } else {
+      // Execute the root in the thread pool to avoid running work on more than
+      // numThreads() threads.
+      pool_->Schedule([ctx, n]() { ctx->handle_range(0, n); });
+    }
+  }
+
+  // Convenience wrapper for parallelForAsync that does not align blocks.
+  void parallelForAsync(Index n, const TensorOpCost& cost,
+                        std::function<void(Index, Index)> f,
+                        std::function<void()> done) const {
+    parallelForAsync(n, cost, nullptr, std::move(f), std::move(done));
+  }
+
+  // Thread pool accessor.
+  ThreadPoolInterface* getPool() const { return pool_; }
+
+  // Allocator accessor.
+  Allocator* allocator() const { return allocator_; }
+
+ private:
+  typedef TensorCostModel<ThreadPoolDevice> CostModel;
+
+  // For parallelForAsync we must keep passed in closures on the heap, and
+  // delete them only after `done` callback finished.
+  struct ParallelForAsyncContext {
+    ParallelForAsyncContext(Index block_count,
+                            std::function<void(Index, Index)> block_f,
+                            std::function<void()> done_callback)
+        : count(block_count),
+          f(std::move(block_f)),
+          done(std::move(done_callback)) {}
+    ~ParallelForAsyncContext() { done(); }
+
+    std::atomic<Index> count;
+    std::function<void(Index, Index)> f;
+    std::function<void()> done;
+
+    std::function<void(Index, Index)> handle_range;
+  };
+
+  struct ParallelForBlock {
+    Index size;   // block size
+    Index count;  // number of blocks
+  };
+
+  // Calculates block size based on (1) the iteration cost and (2) parallel
+  // efficiency. We want blocks to be not too small to mitigate parallelization
+  // overheads; not too large to mitigate tail effect and potential load
+  // imbalance and we also want number of blocks to be evenly dividable across
+  // threads.
+  ParallelForBlock CalculateParallelForBlock(
+      const Index n, const TensorOpCost& cost,
+      std::function<Index(Index)> block_align) const {
+    const double block_size_f = 1.0 / CostModel::taskSize(1, cost);
+    const Index max_oversharding_factor = 4;
+    Index block_size = numext::mini(
+        n, numext::maxi<Index>(
+               divup<Index>(n, max_oversharding_factor * numThreads()),
+               block_size_f));
+    const Index max_block_size = numext::mini(n, 2 * block_size);
+
+    if (block_align) {
+      Index new_block_size = block_align(block_size);
+      eigen_assert(new_block_size >= block_size);
+      block_size = numext::mini(n, new_block_size);
+    }
+
+    Index block_count = divup(n, block_size);
+
+    // Calculate parallel efficiency as fraction of total CPU time used for
+    // computations:
+    double max_efficiency =
+        static_cast<double>(block_count) /
+        (divup<int>(block_count, numThreads()) * numThreads());
+
+    // Now try to increase block size up to max_block_size as long as it
+    // doesn't decrease parallel efficiency.
+    for (Index prev_block_count = block_count;
+         max_efficiency < 1.0 && prev_block_count > 1;) {
+      // This is the next block size that divides size into a smaller number
+      // of blocks than the current block_size.
+      Index coarser_block_size = divup(n, prev_block_count - 1);
+      if (block_align) {
+        Index new_block_size = block_align(coarser_block_size);
+        eigen_assert(new_block_size >= coarser_block_size);
+        coarser_block_size = numext::mini(n, new_block_size);
+      }
+      if (coarser_block_size > max_block_size) {
+        break;  // Reached max block size. Stop.
+      }
+      // Recalculate parallel efficiency.
+      const Index coarser_block_count = divup(n, coarser_block_size);
+      eigen_assert(coarser_block_count < prev_block_count);
+      prev_block_count = coarser_block_count;
+      const double coarser_efficiency =
+          static_cast<double>(coarser_block_count) /
+          (divup<int>(coarser_block_count, numThreads()) * numThreads());
+      if (coarser_efficiency + 0.01 >= max_efficiency) {
+        // Taking it.
+        block_size = coarser_block_size;
+        block_count = coarser_block_count;
+        if (max_efficiency < coarser_efficiency) {
+          max_efficiency = coarser_efficiency;
+        }
+      }
+    }
+
+    return {block_size, block_count};
+  }
+
+  ThreadPoolInterface* pool_;
+  int num_threads_;
+  Allocator* allocator_;
+};
+
+
+}  // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_THREAD_POOL_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorDimensionList.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorDimensionList.h
new file mode 100644
index 0000000..1a30e45
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorDimensionList.h
@@ -0,0 +1,236 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_DIMENSION_LIST_H
+#define EIGEN_CXX11_TENSOR_TENSOR_DIMENSION_LIST_H
+
+namespace Eigen {
+
+/** \internal
+  *
+  * \class TensorDimensionList
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Special case of tensor index list used to list all the dimensions of a tensor of rank n.
+  *
+  * \sa Tensor
+  */
+
+template <typename Index, std::size_t Rank> struct DimensionList {
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  const Index operator[] (const Index i) const { return i; }
+};
+
+namespace internal {
+
+template<typename Index, std::size_t Rank> struct array_size<DimensionList<Index, Rank> > {
+  static const size_t value = Rank;
+};
+template<typename Index, std::size_t Rank> struct array_size<const DimensionList<Index, Rank> > {
+  static const size_t value = Rank;
+};
+
+template<DenseIndex n, typename Index, std::size_t Rank> const Index array_get(DimensionList<Index, Rank>&) {
+  return n;
+}
+template<DenseIndex n, typename Index, std::size_t Rank> const Index array_get(const DimensionList<Index, Rank>&) {
+  return n;
+}
+
+
+#if EIGEN_HAS_CONSTEXPR
+template <typename Index, std::size_t Rank>
+struct index_known_statically_impl<DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex) {
+    return true;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_known_statically_impl<const DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex) {
+    return true;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct all_indices_known_statically_impl<DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return true;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct all_indices_known_statically_impl<const DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return true;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct indices_statically_known_to_increase_impl<DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return true;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct indices_statically_known_to_increase_impl<const DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return true;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct index_statically_eq_impl<DimensionList<Index, Rank> > {
+  static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+    return i == value;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_statically_eq_impl<const DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+    return i == value;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct index_statically_ne_impl<DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+    return i != value;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_statically_ne_impl<const DimensionList<Index, Rank> > {
+  static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+    return i != value;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct index_statically_gt_impl<DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+    return i > value;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_statically_gt_impl<const DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+    return i > value;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct index_statically_lt_impl<DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+    return i < value;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_statically_lt_impl<const DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const DenseIndex i, const DenseIndex value) {
+    return i < value;
+  }
+};
+
+#else
+template <typename Index, std::size_t Rank>
+struct index_known_statically_impl<DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static EIGEN_ALWAYS_INLINE bool run(const DenseIndex) {
+    return true;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_known_statically_impl<const DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static EIGEN_ALWAYS_INLINE bool run(const DenseIndex) {
+    return true;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct all_indices_known_statically_impl<DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static EIGEN_ALWAYS_INLINE bool run() {
+    return true;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct all_indices_known_statically_impl<const DimensionList<Index, Rank> > {
+  EIGEN_DEVICE_FUNC static EIGEN_ALWAYS_INLINE bool run() {
+    return true;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct indices_statically_known_to_increase_impl<DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run() {
+    return true;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct indices_statically_known_to_increase_impl<const DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run() {
+    return true;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct index_statically_eq_impl<DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex, const DenseIndex) {
+    return false;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_statically_eq_impl<const DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex, const DenseIndex) {
+    return false;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct index_statically_ne_impl<DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex, const DenseIndex){
+    return false;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_statically_ne_impl<const DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex, const DenseIndex) {
+    return false;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct index_statically_gt_impl<DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex, const DenseIndex) {
+    return false;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_statically_gt_impl<const DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex, const DenseIndex) {
+    return false;
+  }
+};
+
+template <typename Index, std::size_t Rank>
+struct index_statically_lt_impl<DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex, const DenseIndex) {
+    return false;
+  }
+};
+template <typename Index, std::size_t Rank>
+struct index_statically_lt_impl<const DimensionList<Index, Rank> > {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const DenseIndex, const DenseIndex) {
+    return false;
+  }
+};
+#endif
+
+}  // end namespace internal
+}  // end namespace Eigen
+
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_DIMENSION_LIST_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorDimensions.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorDimensions.h
new file mode 100644
index 0000000..f0f1e83
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorDimensions.h
@@ -0,0 +1,490 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_DIMENSIONS_H
+#define EIGEN_CXX11_TENSOR_TENSOR_DIMENSIONS_H
+
+
+namespace Eigen {
+
+/** \internal
+  *
+  * \class TensorDimensions
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Set of classes used to encode and store the dimensions of a Tensor.
+  *
+  * The Sizes class encodes as part of the type the number of dimensions and the
+  * sizes corresponding to each dimension. It uses no storage space since it is
+  * entirely known at compile time.
+  * The DSizes class is its dynamic sibling: the number of dimensions is known
+  * at compile time but the sizes are set during execution.
+  *
+  * \sa Tensor
+  */
+
+// Boilerplate code
+namespace internal {
+
+template<std::ptrdiff_t n, typename Dimension> struct dget {
+  static const std::ptrdiff_t value = get<n, Dimension>::value;
+};
+
+
+template<typename Index, std::ptrdiff_t NumIndices, std::ptrdiff_t n, bool RowMajor>
+struct fixed_size_tensor_index_linearization_helper
+{
+  template <typename Dimensions> EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Index run(array<Index, NumIndices> const& indices,
+                          const Dimensions& dimensions)
+  {
+    return array_get<RowMajor ? n - 1 : (NumIndices - n)>(indices) +
+        dget<RowMajor ? n - 1 : (NumIndices - n), Dimensions>::value *
+        fixed_size_tensor_index_linearization_helper<Index, NumIndices, n - 1, RowMajor>::run(indices, dimensions);
+  }
+};
+
+template<typename Index, std::ptrdiff_t NumIndices, bool RowMajor>
+struct fixed_size_tensor_index_linearization_helper<Index, NumIndices, 0, RowMajor>
+{
+  template <typename Dimensions> EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Index run(array<Index, NumIndices> const&, const Dimensions&)
+  {
+    return 0;
+  }
+};
+
+template<typename Index, std::ptrdiff_t n>
+struct fixed_size_tensor_index_extraction_helper
+{
+  template <typename Dimensions> EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Index run(const Index index,
+                          const Dimensions& dimensions)
+  {
+    const Index mult = (index == n-1) ? 1 : 0;
+    return array_get<n-1>(dimensions) * mult +
+        fixed_size_tensor_index_extraction_helper<Index, n - 1>::run(index, dimensions);
+  }
+};
+
+template<typename Index>
+struct fixed_size_tensor_index_extraction_helper<Index, 0>
+{
+  template <typename Dimensions> EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Index run(const Index,
+                          const Dimensions&)
+  {
+    return 0;
+  }
+  };
+
+}  // end namespace internal
+
+
+// Fixed size
+#ifndef EIGEN_EMULATE_CXX11_META_H
+template <typename std::ptrdiff_t... Indices>
+struct Sizes {
+  typedef internal::numeric_list<std::ptrdiff_t, Indices...> Base;
+  const Base t = Base();
+  static const std::ptrdiff_t total_size = internal::arg_prod(Indices...);
+  static const ptrdiff_t count = Base::count;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t rank() const {
+    return Base::count;
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t TotalSize() {
+    return internal::arg_prod(Indices...);
+  }
+
+  EIGEN_DEVICE_FUNC Sizes() { }
+  template <typename DenseIndex>
+  explicit EIGEN_DEVICE_FUNC Sizes(const array<DenseIndex, Base::count>& /*indices*/) {
+    // todo: add assertion
+  }
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+  template <typename... DenseIndex> EIGEN_DEVICE_FUNC Sizes(DenseIndex...) { }
+  explicit EIGEN_DEVICE_FUNC Sizes(std::initializer_list<std::ptrdiff_t> /*l*/) {
+    // todo: add assertion
+  }
+#endif
+
+  template <typename T> Sizes& operator = (const T& /*other*/) {
+    // add assertion failure if the size of other is different
+    return *this;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t operator[] (const std::ptrdiff_t index) const {
+    return internal::fixed_size_tensor_index_extraction_helper<std::ptrdiff_t, Base::count>::run(index, t);
+  }
+
+  template <typename DenseIndex> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  ptrdiff_t IndexOfColMajor(const array<DenseIndex, Base::count>& indices) const {
+    return internal::fixed_size_tensor_index_linearization_helper<DenseIndex, Base::count, Base::count, false>::run(indices, t);
+  }
+  template <typename DenseIndex> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  ptrdiff_t IndexOfRowMajor(const array<DenseIndex, Base::count>& indices) const {
+    return internal::fixed_size_tensor_index_linearization_helper<DenseIndex, Base::count, Base::count, true>::run(indices, t);
+  }
+};
+
+namespace internal {
+template <typename std::ptrdiff_t... Indices>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_prod(const Sizes<Indices...>&) {
+  return Sizes<Indices...>::total_size;
+}
+}
+
+#else
+
+template <std::ptrdiff_t n>
+struct non_zero_size {
+  typedef internal::type2val<std::ptrdiff_t, n> type;
+};
+template <>
+struct non_zero_size<0> {
+  typedef internal::null_type type;
+};
+
+template <std::ptrdiff_t V1=0, std::ptrdiff_t V2=0, std::ptrdiff_t V3=0, std::ptrdiff_t V4=0, std::ptrdiff_t V5=0> struct Sizes {
+  typedef typename internal::make_type_list<typename non_zero_size<V1>::type, typename non_zero_size<V2>::type, typename non_zero_size<V3>::type, typename non_zero_size<V4>::type, typename non_zero_size<V5>::type >::type Base;
+  static const std::ptrdiff_t count = Base::count;
+  static const std::ptrdiff_t total_size = internal::arg_prod<Base>::value;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ptrdiff_t rank() const {
+    return count;
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ptrdiff_t TotalSize() {
+    return internal::arg_prod<Base>::value;
+  }
+
+  Sizes() { }
+  template <typename DenseIndex>
+  explicit Sizes(const array<DenseIndex, Base::count>& /*indices*/) {
+    // todo: add assertion
+  }
+  template <typename T> Sizes& operator = (const T& /*other*/) {
+    // add assertion failure if the size of other is different
+    return *this;
+  }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+  template <typename... DenseIndex> Sizes(DenseIndex... /*indices*/) { }
+  explicit Sizes(std::initializer_list<std::ptrdiff_t>) {
+    // todo: add assertion
+  }
+#else
+  EIGEN_DEVICE_FUNC explicit Sizes(const DenseIndex) {
+  }
+  EIGEN_DEVICE_FUNC Sizes(const DenseIndex, const DenseIndex) {
+  }
+  EIGEN_DEVICE_FUNC Sizes(const DenseIndex, const DenseIndex, const DenseIndex) {
+  }
+  EIGEN_DEVICE_FUNC Sizes(const DenseIndex, const DenseIndex, const DenseIndex, const DenseIndex) {
+  }
+  EIGEN_DEVICE_FUNC Sizes(const DenseIndex, const DenseIndex, const DenseIndex, const DenseIndex, const DenseIndex) {
+  }
+#endif
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index operator[] (const Index index) const {
+    switch (index) {
+      case 0:
+        return internal::get<0, Base>::value;
+      case 1:
+        return internal::get<1, Base>::value;
+      case 2:
+        return internal::get<2, Base>::value;
+      case 3:
+        return internal::get<3, Base>::value;
+      case 4:
+        return internal::get<4, Base>::value;
+      default:
+        eigen_assert(false && "index overflow");
+        return static_cast<Index>(-1);
+    }
+  }
+
+  template <typename DenseIndex> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  ptrdiff_t IndexOfColMajor(const array<DenseIndex, Base::count>& indices) const {
+    return internal::fixed_size_tensor_index_linearization_helper<DenseIndex, Base::count, Base::count, false>::run(indices, *reinterpret_cast<const Base*>(this));
+  }
+  template <typename DenseIndex> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  ptrdiff_t IndexOfRowMajor(const array<DenseIndex, Base::count>& indices) const {
+    return internal::fixed_size_tensor_index_linearization_helper<DenseIndex, Base::count, Base::count, true>::run(indices, *reinterpret_cast<const Base*>(this));
+  }
+};
+
+namespace internal {
+template <std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_prod(const Sizes<V1, V2, V3, V4, V5>&) {
+  return Sizes<V1, V2, V3, V4, V5>::total_size;
+}
+}
+
+#endif
+
+// Boilerplate
+namespace internal {
+template<typename Index, std::ptrdiff_t NumIndices, std::ptrdiff_t n, bool RowMajor>
+struct tensor_index_linearization_helper
+{
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Index run(array<Index, NumIndices> const& indices, array<Index, NumIndices> const& dimensions)
+  {
+    return array_get<RowMajor ? n : (NumIndices - n - 1)>(indices) +
+      array_get<RowMajor ? n : (NumIndices - n - 1)>(dimensions) *
+        tensor_index_linearization_helper<Index, NumIndices, n - 1, RowMajor>::run(indices, dimensions);
+  }
+};
+
+template<typename Index, std::ptrdiff_t NumIndices, bool RowMajor>
+struct tensor_index_linearization_helper<Index, NumIndices, 0, RowMajor>
+{
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Index run(array<Index, NumIndices> const& indices, array<Index, NumIndices> const&)
+  {
+    return array_get<RowMajor ? 0 : NumIndices - 1>(indices);
+  }
+};
+}  // end namespace internal
+
+
+
+// Dynamic size
+template <typename DenseIndex, int NumDims>
+struct DSizes : array<DenseIndex, NumDims> {
+  typedef array<DenseIndex, NumDims> Base;
+  static const int count = NumDims;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rank() const {
+    return NumDims;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex TotalSize() const {
+    return (NumDims == 0) ? 1 : internal::array_prod(*static_cast<const Base*>(this));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DSizes() {
+    for (int i = 0 ; i < NumDims; ++i) {
+      (*this)[i] = 0;
+    }
+  }
+  EIGEN_DEVICE_FUNC explicit DSizes(const array<DenseIndex, NumDims>& a) : Base(a) { }
+
+  EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0) {
+    eigen_assert(NumDims == 1);
+    (*this)[0] = i0;
+  }
+
+  EIGEN_DEVICE_FUNC DSizes(const DimensionList<DenseIndex, NumDims>& a) {
+    for (int i = 0 ; i < NumDims; ++i) {
+      (*this)[i] = a[i];
+    }
+  }
+
+  // Enable DSizes index type promotion only if we are promoting to the
+  // larger type, e.g. allow to promote dimensions of type int to long.
+  template<typename OtherIndex>
+  EIGEN_DEVICE_FUNC
+  explicit DSizes(const array<OtherIndex, NumDims>& other,
+                  // Default template parameters require c++11.
+                  typename internal::enable_if<
+                     internal::is_same<
+                         DenseIndex,
+                         typename internal::promote_index_type<
+                             DenseIndex,
+                             OtherIndex
+                         >::type
+                     >::value, void*>::type = 0) {
+    for (int i = 0; i < NumDims; ++i) {
+      (*this)[i] = static_cast<DenseIndex>(other[i]);
+    }
+  }
+
+#ifdef EIGEN_HAS_INDEX_LIST
+  template <typename FirstType, typename... OtherTypes>
+  EIGEN_DEVICE_FUNC
+  explicit DSizes(const Eigen::IndexList<FirstType, OtherTypes...>& dimensions) {
+    for (int i = 0; i < dimensions.count; ++i) {
+      (*this)[i] = dimensions[i];
+    }
+  }
+#endif
+
+#ifndef EIGEN_EMULATE_CXX11_META_H
+  template <typename std::ptrdiff_t... Indices>
+  EIGEN_DEVICE_FUNC DSizes(const Sizes<Indices...>& a) {
+    for (int i = 0 ; i < NumDims; ++i) {
+      (*this)[i] = a[i];
+    }
+  }
+#else
+  template <std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5>
+  EIGEN_DEVICE_FUNC DSizes(const Sizes<V1, V2, V3, V4, V5>& a) {
+    for (int i = 0 ; i < NumDims; ++i) {
+      (*this)[i] = a[i];
+    }
+  }
+#endif
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+  template<typename... IndexTypes> EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE explicit DSizes(DenseIndex firstDimension, DenseIndex secondDimension, IndexTypes... otherDimensions) : Base({{firstDimension, secondDimension, otherDimensions...}}) {
+    EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 2 == NumDims, YOU_MADE_A_PROGRAMMING_MISTAKE)
+  }
+#else
+  EIGEN_DEVICE_FUNC DSizes(const DenseIndex i0, const DenseIndex i1) {
+    eigen_assert(NumDims == 2);
+    (*this)[0] = i0;
+    (*this)[1] = i1;
+  }
+  EIGEN_DEVICE_FUNC DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2) {
+    eigen_assert(NumDims == 3);
+    (*this)[0] = i0;
+    (*this)[1] = i1;
+    (*this)[2] = i2;
+  }
+  EIGEN_DEVICE_FUNC DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3) {
+    eigen_assert(NumDims == 4);
+    (*this)[0] = i0;
+    (*this)[1] = i1;
+    (*this)[2] = i2;
+    (*this)[3] = i3;
+  }
+  EIGEN_DEVICE_FUNC DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3, const DenseIndex i4) {
+    eigen_assert(NumDims == 5);
+    (*this)[0] = i0;
+    (*this)[1] = i1;
+    (*this)[2] = i2;
+    (*this)[3] = i3;
+    (*this)[4] = i4;
+  }
+#endif
+
+  EIGEN_DEVICE_FUNC DSizes& operator = (const array<DenseIndex, NumDims>& other) {
+    *static_cast<Base*>(this) = other;
+    return *this;
+  }
+
+  // A constexpr would be so much better here
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex IndexOfColMajor(const array<DenseIndex, NumDims>& indices) const {
+    return internal::tensor_index_linearization_helper<DenseIndex, NumDims, NumDims - 1, false>::run(indices, *static_cast<const Base*>(this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex IndexOfRowMajor(const array<DenseIndex, NumDims>& indices) const {
+    return internal::tensor_index_linearization_helper<DenseIndex, NumDims, NumDims - 1, true>::run(indices, *static_cast<const Base*>(this));
+  }
+};
+
+template <typename IndexType, int NumDims>
+std::ostream& operator<<(std::ostream& os,
+                         const DSizes<IndexType, NumDims>& dims) {
+  os << "[";
+  for (int i = 0; i < NumDims; ++i) {
+    if (i > 0) os << ", ";
+    os << dims[i];
+  }
+  os << "]";
+  return os;
+}
+
+// Boilerplate
+namespace internal {
+template<typename Index, std::ptrdiff_t NumIndices, std::ptrdiff_t n, bool RowMajor>
+struct tensor_vsize_index_linearization_helper
+{
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Index run(array<Index, NumIndices> const& indices, std::vector<DenseIndex> const& dimensions)
+  {
+    return array_get<RowMajor ? n : (NumIndices - n - 1)>(indices) +
+      array_get<RowMajor ? n : (NumIndices - n - 1)>(dimensions) *
+        tensor_vsize_index_linearization_helper<Index, NumIndices, n - 1, RowMajor>::run(indices, dimensions);
+  }
+};
+
+template<typename Index, std::ptrdiff_t NumIndices, bool RowMajor>
+struct tensor_vsize_index_linearization_helper<Index, NumIndices, 0, RowMajor>
+{
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Index run(array<Index, NumIndices> const& indices, std::vector<DenseIndex> const&)
+  {
+    return array_get<RowMajor ? 0 : NumIndices - 1>(indices);
+  }
+};
+}  // end namespace internal
+
+
+namespace internal {
+
+template <typename DenseIndex, int NumDims> struct array_size<const DSizes<DenseIndex, NumDims> > {
+  static const ptrdiff_t value = NumDims;
+};
+template <typename DenseIndex, int NumDims> struct array_size<DSizes<DenseIndex, NumDims> > {
+  static const ptrdiff_t value = NumDims;
+};
+#ifndef EIGEN_EMULATE_CXX11_META_H
+template <typename std::ptrdiff_t... Indices> struct array_size<const Sizes<Indices...> > {
+static const std::ptrdiff_t value = Sizes<Indices...>::count;
+};
+template <typename std::ptrdiff_t... Indices> struct array_size<Sizes<Indices...> > {
+static const std::ptrdiff_t value = Sizes<Indices...>::count;
+};
+template <std::ptrdiff_t n, typename std::ptrdiff_t... Indices> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_get(const Sizes<Indices...>&) {
+  return get<n, internal::numeric_list<std::ptrdiff_t, Indices...> >::value;
+}
+template <std::ptrdiff_t n> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_get(const Sizes<>&) {
+  eigen_assert(false && "should never be called");
+  return -1;
+}
+#else
+template <std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5> struct array_size<const Sizes<V1,V2,V3,V4,V5> > {
+  static const ptrdiff_t value = Sizes<V1,V2,V3,V4,V5>::count;
+};
+template <std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5> struct array_size<Sizes<V1,V2,V3,V4,V5> > {
+  static const ptrdiff_t value = Sizes<V1,V2,V3,V4,V5>::count;
+};
+template <std::ptrdiff_t n, std::ptrdiff_t V1, std::ptrdiff_t V2, std::ptrdiff_t V3, std::ptrdiff_t V4, std::ptrdiff_t V5> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::ptrdiff_t array_get(const Sizes<V1,V2,V3,V4,V5>&) {
+  return get<n, typename Sizes<V1,V2,V3,V4,V5>::Base>::value;
+}
+
+#endif
+
+
+template <typename Dims1, typename Dims2, ptrdiff_t n, ptrdiff_t m>
+struct sizes_match_below_dim {
+  static EIGEN_DEVICE_FUNC  EIGEN_STRONG_INLINE bool run(Dims1&, Dims2&) {
+    return false;
+  }
+};
+template <typename Dims1, typename Dims2, ptrdiff_t n>
+struct sizes_match_below_dim<Dims1, Dims2, n, n> {
+  static EIGEN_DEVICE_FUNC  EIGEN_STRONG_INLINE bool run(Dims1& dims1, Dims2& dims2) {
+    return (array_get<n-1>(dims1) == array_get<n-1>(dims2)) &&
+        sizes_match_below_dim<Dims1, Dims2, n-1, n-1>::run(dims1, dims2);
+  }
+};
+template <typename Dims1, typename Dims2>
+struct sizes_match_below_dim<Dims1, Dims2, 0, 0> {
+  static EIGEN_DEVICE_FUNC  EIGEN_STRONG_INLINE bool run(Dims1&, Dims2&) {
+    return true;
+  }
+};
+
+} // end namespace internal
+
+
+template <typename Dims1, typename Dims2>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool dimensions_match(Dims1 dims1, Dims2 dims2) {
+  return internal::sizes_match_below_dim<Dims1, Dims2, internal::array_size<Dims1>::value, internal::array_size<Dims2>::value>::run(dims1, dims2);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_DIMENSIONS_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorEvalTo.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorEvalTo.h
new file mode 100644
index 0000000..a48d035
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorEvalTo.h
@@ -0,0 +1,236 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_EVAL_TO_H
+#define EIGEN_CXX11_TENSOR_TENSOR_EVAL_TO_H
+
+namespace Eigen {
+
+/** \class TensorForcedEval
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor reshaping class.
+  *
+  *
+  */
+namespace internal {
+template<typename XprType, template <class> class MakePointer_>
+struct traits<TensorEvalToOp<XprType, MakePointer_> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs are different.
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename MakePointer_<Scalar>::Type PointerType;
+
+  enum {
+    Flags = 0
+  };
+  template <class T>
+  struct MakePointer {
+    // Intermediate typedef to workaround MSVC issue.
+    typedef MakePointer_<T> MakePointerT;
+    typedef typename MakePointerT::Type Type;
+
+
+  };
+};
+
+template<typename XprType, template <class> class MakePointer_>
+struct eval<TensorEvalToOp<XprType, MakePointer_>, Eigen::Dense>
+{
+  typedef const TensorEvalToOp<XprType, MakePointer_>& type;
+};
+
+template<typename XprType, template <class> class MakePointer_>
+struct nested<TensorEvalToOp<XprType, MakePointer_>, 1, typename eval<TensorEvalToOp<XprType, MakePointer_> >::type>
+{
+  typedef TensorEvalToOp<XprType, MakePointer_> type;
+};
+
+}  // end namespace internal
+
+
+
+
+template<typename XprType, template <class> class MakePointer_>
+class TensorEvalToOp : public TensorBase<TensorEvalToOp<XprType, MakePointer_>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorEvalToOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename MakePointer_<CoeffReturnType>::Type PointerType;
+  typedef typename Eigen::internal::nested<TensorEvalToOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorEvalToOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorEvalToOp>::Index Index;
+
+  static const int NumDims = Eigen::internal::traits<TensorEvalToOp>::NumDimensions;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvalToOp(PointerType buffer, const XprType& expr)
+      : m_xpr(expr), m_buffer(buffer) {}
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_DEVICE_FUNC PointerType buffer() const { return m_buffer; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    PointerType m_buffer;
+};
+
+
+
+template<typename ArgType, typename Device, template <class> class MakePointer_>
+struct TensorEvaluator<const TensorEvalToOp<ArgType, MakePointer_>, Device>
+{
+  typedef TensorEvalToOp<ArgType, MakePointer_> XprType;
+  typedef typename ArgType::Scalar Scalar;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
+  typedef typename XprType::Index Index;
+  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename Eigen::internal::traits<XprType>::PointerType TensorPointerType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  enum {
+    IsAligned         = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = true,
+    PreferBlockAccess = false,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = true
+  };
+
+  static const int NumDims = internal::traits<ArgType>::NumDimensions;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename TensorEvaluator<const ArgType, Device>::TensorBlock
+      ArgTensorBlock;
+
+  typedef internal::TensorBlockAssignment<
+      CoeffReturnType, NumDims, typename ArgTensorBlock::XprType, Index>
+      TensorBlockAssignment;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_buffer(device.get(op.buffer())), m_expression(op.expression()){}
+
+
+  EIGEN_STRONG_INLINE ~TensorEvaluator() {
+  }
+
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType scalar) {
+    EIGEN_UNUSED_VARIABLE(scalar);
+    eigen_assert(scalar == NULL);
+    return m_impl.evalSubExprsIfNeeded(m_buffer);
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType scalar, EvalSubExprsCallback done) {
+    EIGEN_UNUSED_VARIABLE(scalar);
+    eigen_assert(scalar == NULL);
+    m_impl.evalSubExprsIfNeededAsync(m_buffer, std::move(done));
+  }
+#endif
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalScalar(Index i) {
+    m_buffer[i] = m_impl.coeff(i);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalPacket(Index i) {
+    internal::pstoret<CoeffReturnType, PacketReturnType, Aligned>(m_buffer + i, m_impl.template packet<TensorEvaluator<ArgType, Device>::IsAligned ? Aligned : Unaligned>(i));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    return m_impl.getResourceRequirements();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalBlock(
+      TensorBlockDesc& desc, TensorBlockScratch& scratch) {
+    // Add `m_buffer` as destination buffer to the block descriptor.
+    desc.template AddDestinationBuffer<Layout>(
+        /*dst_base=*/m_buffer + desc.offset(),
+        /*dst_strides=*/internal::strides<Layout>(m_impl.dimensions()));
+
+    ArgTensorBlock block =
+        m_impl.block(desc, scratch, /*root_of_expr_ast=*/true);
+
+    // If block was evaluated into a destination buffer, there is no need to do
+    // an assignment.
+    if (block.kind() != internal::TensorBlockKind::kMaterializedInOutput) {
+      TensorBlockAssignment::Run(
+          TensorBlockAssignment::target(
+              desc.dimensions(), internal::strides<Layout>(m_impl.dimensions()),
+              m_buffer, desc.offset()),
+          block.expr());
+    }
+    block.cleanup();
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return m_buffer[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_buffer + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    // We assume that evalPacket or evalScalar is called to perform the
+    // assignment and account for the cost of the write here.
+    return m_impl.costPerCoeff(vectorized) +
+        TensorOpCost(0, sizeof(CoeffReturnType), 0, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_buffer; }
+  ArgType expression() const { return m_expression; }
+  #ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+    m_buffer.bind(cgh);
+  }
+  #endif
+
+
+ private:
+  TensorEvaluator<ArgType, Device> m_impl;
+  EvaluatorPointerType m_buffer;
+  const ArgType m_expression;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_EVAL_TO_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorEvaluator.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorEvaluator.h
new file mode 100644
index 0000000..3aff7fa
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorEvaluator.h
@@ -0,0 +1,983 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_EVALUATOR_H
+#define EIGEN_CXX11_TENSOR_TENSOR_EVALUATOR_H
+
+namespace Eigen {
+
+/** \class TensorEvaluator
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief The tensor evaluator classes.
+  *
+  * These classes are responsible for the evaluation of the tensor expression.
+  *
+  * TODO: add support for more types of expressions, in particular expressions
+  * leading to lvalues (slicing, reshaping, etc...)
+  */
+
+// Generic evaluator
+template<typename Derived, typename Device>
+struct TensorEvaluator
+{
+  typedef typename Derived::Index Index;
+  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef typename Derived::Dimensions Dimensions;
+  typedef Derived XprType;
+  static const int PacketSize =  PacketType<CoeffReturnType, Device>::size;
+  typedef typename internal::traits<Derived>::template MakePointer<Scalar>::Type TensorPointerType;
+  typedef StorageMemory<Scalar, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  // NumDimensions is -1 for variable dim tensors
+  static const int NumCoords = internal::traits<Derived>::NumDimensions > 0 ?
+                               internal::traits<Derived>::NumDimensions : 0;
+
+  enum {
+    IsAligned          = Derived::IsAligned,
+    PacketAccess       = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess        = internal::is_arithmetic<typename internal::remove_const<Scalar>::type>::value,
+    PreferBlockAccess  = false,
+    Layout             = Derived::Layout,
+    CoordAccess        = NumCoords > 0,
+    RawAccess          = true
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumCoords, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename internal::TensorMaterializedBlock<ScalarNoConst, NumCoords,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const Derived& m, const Device& device)
+      : m_data(device.get((const_cast<TensorPointerType>(m.data())))),
+        m_dims(m.dimensions()),
+        m_device(device)
+  { }
+
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dims; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType dest) {
+    if (!NumTraits<typename internal::remove_const<Scalar>::type>::RequireInitialization && dest) {
+      m_device.memcpy((void*)(m_device.get(dest)), m_device.get(m_data), m_dims.TotalSize() * sizeof(Scalar));
+      return false;
+    }
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType dest, EvalSubExprsCallback done) {
+    // TODO(ezhulenev): ThreadPoolDevice memcpy is blockign operation.
+    done(evalSubExprsIfNeeded(dest));
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    eigen_assert(m_data != NULL);
+    return m_data[index];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) {
+    eigen_assert(m_data != NULL);
+    return m_data[index];
+  }
+
+  template<int LoadMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketReturnType packet(Index index) const
+  {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_data + index);
+  }
+
+  // Return a packet starting at `index` where `umask` specifies which elements
+  // have to be loaded. Type/size of mask depends on PacketReturnType, e.g. for
+  // Packet16f, `umask` is of type uint16_t and if a bit is 1, corresponding
+  // float element will be loaded, otherwise 0 will be loaded.
+  // Function has been templatized to enable Sfinae.
+  template <typename PacketReturnTypeT> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename internal::enable_if<internal::unpacket_traits<PacketReturnTypeT>::masked_load_available, PacketReturnTypeT>::type
+  partialPacket(Index index, typename internal::unpacket_traits<PacketReturnTypeT>::mask_t umask) const
+  {
+    return internal::ploadu<PacketReturnTypeT>(m_data + index, umask);
+  }
+
+  template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x)
+  {
+    return internal::pstoret<Scalar, PacketReturnType, StoreMode>(m_data + index, x);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array<DenseIndex, NumCoords>& coords) const {
+    eigen_assert(m_data != NULL);
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      return m_data[m_dims.IndexOfColMajor(coords)];
+    } else {
+      return m_data[m_dims.IndexOfRowMajor(coords)];
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType&
+  coeffRef(const array<DenseIndex, NumCoords>& coords) {
+    eigen_assert(m_data != NULL);
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      return m_data[m_dims.IndexOfColMajor(coords)];
+    } else {
+      return m_data[m_dims.IndexOfRowMajor(coords)];
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized,
+                        PacketType<CoeffReturnType, Device>::size);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    return internal::TensorBlockResourceRequirements::any();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    assert(m_data != NULL);
+    return TensorBlock::materialize(m_data, m_dims, desc, scratch);
+  }
+
+  template<typename TensorBlock>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const TensorBlockDesc& desc, const TensorBlock& block) {
+    assert(m_data != NULL);
+
+    typedef typename TensorBlock::XprType TensorBlockExpr;
+    typedef internal::TensorBlockAssignment<Scalar, NumCoords, TensorBlockExpr,
+                                            Index>
+        TensorBlockAssign;
+
+    TensorBlockAssign::Run(
+        TensorBlockAssign::target(desc.dimensions(),
+                                  internal::strides<Layout>(m_dims), m_data,
+                                  desc.offset()),
+        block.expr());
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_data; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_data.bind(cgh);
+  }
+#endif
+ protected:
+  EvaluatorPointerType m_data;
+  Dimensions m_dims;
+  const Device EIGEN_DEVICE_REF m_device;
+};
+
+namespace {
+template <typename T> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T loadConstant(const T* address) {
+  return *address;
+}
+// Use the texture cache on CUDA devices whenever possible
+#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float loadConstant(const float* address) {
+  return __ldg(address);
+}
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double loadConstant(const double* address) {
+  return __ldg(address);
+}
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+Eigen::half loadConstant(const Eigen::half* address) {
+  return Eigen::half(half_impl::raw_uint16_to_half(__ldg(&address->x)));
+}
+#endif
+#ifdef EIGEN_USE_SYCL
+// overload of load constant should be implemented here based on range access
+template <cl::sycl::access::mode AcMd, typename T>
+T &loadConstant(const Eigen::TensorSycl::internal::RangeAccess<AcMd, T> &address) {
+  return *address;
+}
+#endif
+}
+
+
+// Default evaluator for rvalues
+template<typename Derived, typename Device>
+struct TensorEvaluator<const Derived, Device>
+{
+  typedef typename Derived::Index Index;
+  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef typename Derived::Dimensions Dimensions;
+  typedef const Derived XprType;
+  typedef typename internal::traits<Derived>::template MakePointer<const Scalar>::Type TensorPointerType;
+  typedef StorageMemory<const Scalar, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  // NumDimensions is -1 for variable dim tensors
+  static const int NumCoords = internal::traits<Derived>::NumDimensions > 0 ?
+                               internal::traits<Derived>::NumDimensions : 0;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+
+  enum {
+    IsAligned         = Derived::IsAligned,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = internal::is_arithmetic<ScalarNoConst>::value,
+    PreferBlockAccess = false,
+    Layout            = Derived::Layout,
+    CoordAccess       = NumCoords > 0,
+    RawAccess         = true
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumCoords, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename internal::TensorMaterializedBlock<ScalarNoConst, NumCoords,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const Derived& m, const Device& device)
+      : m_data(device.get(m.data())), m_dims(m.dimensions()), m_device(device)
+  { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dims; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    if (!NumTraits<typename internal::remove_const<Scalar>::type>::RequireInitialization && data) {
+      m_device.memcpy((void*)(m_device.get(data)),m_device.get(m_data), m_dims.TotalSize() * sizeof(Scalar));
+      return false;
+    }
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType dest, EvalSubExprsCallback done) {
+    // TODO(ezhulenev): ThreadPoolDevice memcpy is a blockign operation.
+    done(evalSubExprsIfNeeded(dest));
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    eigen_assert(m_data != NULL);
+    return loadConstant(m_data+index);
+  }
+
+  template<int LoadMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketReturnType packet(Index index) const
+  {
+    return internal::ploadt_ro<PacketReturnType, LoadMode>(m_data + index);
+  }
+
+  // Return a packet starting at `index` where `umask` specifies which elements
+  // have to be loaded. Type/size of mask depends on PacketReturnType, e.g. for
+  // Packet16f, `umask` is of type uint16_t and if a bit is 1, corresponding
+  // float element will be loaded, otherwise 0 will be loaded.
+  // Function has been templatized to enable Sfinae.
+  template <typename PacketReturnTypeT> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename internal::enable_if<internal::unpacket_traits<PacketReturnTypeT>::masked_load_available, PacketReturnTypeT>::type
+  partialPacket(Index index, typename internal::unpacket_traits<PacketReturnTypeT>::mask_t umask) const
+  {
+    return internal::ploadu<PacketReturnTypeT>(m_data + index, umask);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array<DenseIndex, NumCoords>& coords) const {
+    eigen_assert(m_data != NULL);
+    const Index index = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? m_dims.IndexOfColMajor(coords)
+                        : m_dims.IndexOfRowMajor(coords);
+    return loadConstant(m_data+index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized,
+                        PacketType<CoeffReturnType, Device>::size);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    return internal::TensorBlockResourceRequirements::any();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    assert(m_data != NULL);
+    return TensorBlock::materialize(m_data, m_dims, desc, scratch);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_data; }
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_data.bind(cgh);
+  }
+#endif
+ protected:
+  EvaluatorPointerType m_data;
+  Dimensions m_dims;
+  const Device EIGEN_DEVICE_REF m_device;
+};
+
+
+
+
+// -------------------- CwiseNullaryOp --------------------
+
+template<typename NullaryOp, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorCwiseNullaryOp<NullaryOp, ArgType>, Device>
+{
+  typedef TensorCwiseNullaryOp<NullaryOp, ArgType> XprType;
+
+  TensorEvaluator(const XprType& op, const Device& device)
+      : m_functor(op.functor()), m_argImpl(op.nestedExpression(), device), m_wrapper()
+  { }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = true,
+    PacketAccess = internal::functor_traits<NullaryOp>::PacketAccess
+    #ifdef EIGEN_USE_SYCL
+    &&  (PacketType<CoeffReturnType, Device>::size >1)
+    #endif
+    ,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) { return true; }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    done(true);
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() { }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const
+  {
+    return m_wrapper(m_functor, index);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return m_wrapper.template packetOp<PacketReturnType, Index>(m_functor, index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized,
+                        PacketType<CoeffReturnType, Device>::size);
+  }
+
+  EIGEN_DEVICE_FUNC  EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+   // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_argImpl.bind(cgh);
+  }
+#endif
+
+ private:
+  const NullaryOp m_functor;
+  TensorEvaluator<ArgType, Device> m_argImpl;
+  const internal::nullary_wrapper<CoeffReturnType,NullaryOp> m_wrapper;
+};
+
+
+
+// -------------------- CwiseUnaryOp --------------------
+
+template<typename UnaryOp, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorCwiseUnaryOp<UnaryOp, ArgType>, Device>
+{
+  typedef TensorCwiseUnaryOp<UnaryOp, ArgType> XprType;
+
+  enum {
+    IsAligned          = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess       = int(TensorEvaluator<ArgType, Device>::PacketAccess) &
+                         int(internal::functor_traits<UnaryOp>::PacketAccess),
+    BlockAccess        = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess  = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout             = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess        = false,  // to be implemented
+    RawAccess          = false
+  };
+
+  TensorEvaluator(const XprType& op, const Device& device)
+    : m_device(device),
+      m_functor(op.functor()),
+      m_argImpl(op.nestedExpression(), device)
+  { }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+  typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  static const int NumDims = internal::array_size<Dimensions>::value;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename TensorEvaluator<const ArgType, Device>::TensorBlock
+      ArgTensorBlock;
+
+  typedef internal::TensorCwiseUnaryBlock<UnaryOp, ArgTensorBlock>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_argImpl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    m_argImpl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_argImpl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const
+  {
+    return m_functor(m_argImpl.coeff(index));
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return m_functor.packetOp(m_argImpl.template packet<LoadMode>(index));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    const double functor_cost = internal::functor_traits<UnaryOp>::Cost;
+    return m_argImpl.costPerCoeff(vectorized) +
+        TensorOpCost(0, 0, functor_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    static const double functor_cost = internal::functor_traits<UnaryOp>::Cost;
+    return m_argImpl.getResourceRequirements().addCostPerCoeff(
+        {0, 0, functor_cost / PacketSize});
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    return TensorBlock(m_argImpl.block(desc, scratch), m_functor);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const{
+    m_argImpl.bind(cgh);
+  }
+#endif
+
+
+ private:
+  const Device EIGEN_DEVICE_REF m_device;
+  const UnaryOp m_functor;
+  TensorEvaluator<ArgType, Device> m_argImpl;
+};
+
+
+// -------------------- CwiseBinaryOp --------------------
+
+template<typename BinaryOp, typename LeftArgType, typename RightArgType, typename Device>
+struct TensorEvaluator<const TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType>, Device>
+{
+  typedef TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType> XprType;
+
+  enum {
+    IsAligned         = int(TensorEvaluator<LeftArgType, Device>::IsAligned) &
+                        int(TensorEvaluator<RightArgType, Device>::IsAligned),
+    PacketAccess      = int(TensorEvaluator<LeftArgType, Device>::PacketAccess) &
+                        int(TensorEvaluator<RightArgType, Device>::PacketAccess) &
+                        int(internal::functor_traits<BinaryOp>::PacketAccess),
+    BlockAccess       = int(TensorEvaluator<LeftArgType, Device>::BlockAccess) &
+                        int(TensorEvaluator<RightArgType, Device>::BlockAccess),
+    PreferBlockAccess = int(TensorEvaluator<LeftArgType, Device>::PreferBlockAccess) |
+                        int(TensorEvaluator<RightArgType, Device>::PreferBlockAccess),
+    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
+  };
+
+  TensorEvaluator(const XprType& op, const Device& device)
+    : m_device(device),
+      m_functor(op.functor()),
+      m_leftImpl(op.lhsExpression(), device),
+      m_rightImpl(op.rhsExpression(), device)
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) == static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout) || internal::traits<XprType>::NumDimensions <= 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    eigen_assert(dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions()));
+  }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename TensorEvaluator<LeftArgType, Device>::Dimensions Dimensions;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  static const int NumDims = internal::array_size<
+      typename TensorEvaluator<LeftArgType, Device>::Dimensions>::value;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename TensorEvaluator<const LeftArgType, Device>::TensorBlock
+      LeftTensorBlock;
+  typedef typename TensorEvaluator<const RightArgType, Device>::TensorBlock
+      RightTensorBlock;
+
+  typedef internal::TensorCwiseBinaryBlock<BinaryOp, LeftTensorBlock,
+                                           RightTensorBlock>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
+  {
+    // TODO: use right impl instead if right impl dimensions are known at compile time.
+    return m_leftImpl.dimensions();
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_leftImpl.evalSubExprsIfNeeded(NULL);
+    m_rightImpl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    // TODO(ezhulenev): Evaluate two expression in parallel?
+    m_leftImpl.evalSubExprsIfNeededAsync(nullptr, [this, done](bool) {
+      m_rightImpl.evalSubExprsIfNeededAsync(nullptr,
+                                            [done](bool) { done(true); });
+    });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_leftImpl.cleanup();
+    m_rightImpl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const
+  {
+    return m_functor(m_leftImpl.coeff(index), m_rightImpl.coeff(index));
+  }
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return m_functor.packetOp(m_leftImpl.template packet<LoadMode>(index), m_rightImpl.template packet<LoadMode>(index));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    const double functor_cost = internal::functor_traits<BinaryOp>::Cost;
+    return m_leftImpl.costPerCoeff(vectorized) +
+           m_rightImpl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, functor_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    static const double functor_cost = internal::functor_traits<BinaryOp>::Cost;
+    return internal::TensorBlockResourceRequirements::merge(
+               m_leftImpl.getResourceRequirements(),
+               m_rightImpl.getResourceRequirements())
+        .addCostPerCoeff({0, 0, functor_cost / PacketSize});
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    desc.DropDestinationBuffer();
+    return TensorBlock(m_leftImpl.block(desc, scratch),
+                         m_rightImpl.block(desc, scratch), m_functor);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+  #ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_leftImpl.bind(cgh);
+    m_rightImpl.bind(cgh);
+  }
+  #endif
+ private:
+  const Device EIGEN_DEVICE_REF m_device;
+  const BinaryOp m_functor;
+  TensorEvaluator<LeftArgType, Device> m_leftImpl;
+  TensorEvaluator<RightArgType, Device> m_rightImpl;
+};
+
+// -------------------- CwiseTernaryOp --------------------
+
+template<typename TernaryOp, typename Arg1Type, typename Arg2Type, typename Arg3Type, typename Device>
+struct TensorEvaluator<const TensorCwiseTernaryOp<TernaryOp, Arg1Type, Arg2Type, Arg3Type>, Device>
+{
+  typedef TensorCwiseTernaryOp<TernaryOp, Arg1Type, Arg2Type, Arg3Type> XprType;
+
+  enum {
+    IsAligned = TensorEvaluator<Arg1Type, Device>::IsAligned & TensorEvaluator<Arg2Type, Device>::IsAligned & TensorEvaluator<Arg3Type, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<Arg1Type, Device>::PacketAccess &&
+                        TensorEvaluator<Arg2Type, Device>::PacketAccess &&
+                        TensorEvaluator<Arg3Type, Device>::PacketAccess &&
+                        internal::functor_traits<TernaryOp>::PacketAccess,
+    BlockAccess       = false,
+    PreferBlockAccess = TensorEvaluator<Arg1Type, Device>::PreferBlockAccess ||
+                        TensorEvaluator<Arg2Type, Device>::PreferBlockAccess ||
+                        TensorEvaluator<Arg3Type, Device>::PreferBlockAccess,
+    Layout            = TensorEvaluator<Arg1Type, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
+  };
+
+  TensorEvaluator(const XprType& op, const Device& device)
+    : m_functor(op.functor()),
+      m_arg1Impl(op.arg1Expression(), device),
+      m_arg2Impl(op.arg2Expression(), device),
+      m_arg3Impl(op.arg3Expression(), device)
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<Arg1Type, Device>::Layout) == static_cast<int>(TensorEvaluator<Arg3Type, Device>::Layout) || internal::traits<XprType>::NumDimensions <= 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    EIGEN_STATIC_ASSERT((internal::is_same<typename internal::traits<Arg1Type>::StorageKind,
+                         typename internal::traits<Arg2Type>::StorageKind>::value),
+                        STORAGE_KIND_MUST_MATCH)
+    EIGEN_STATIC_ASSERT((internal::is_same<typename internal::traits<Arg1Type>::StorageKind,
+                         typename internal::traits<Arg3Type>::StorageKind>::value),
+                        STORAGE_KIND_MUST_MATCH)
+    EIGEN_STATIC_ASSERT((internal::is_same<typename internal::traits<Arg1Type>::Index,
+                         typename internal::traits<Arg2Type>::Index>::value),
+                        STORAGE_INDEX_MUST_MATCH)
+    EIGEN_STATIC_ASSERT((internal::is_same<typename internal::traits<Arg1Type>::Index,
+                         typename internal::traits<Arg3Type>::Index>::value),
+                        STORAGE_INDEX_MUST_MATCH)
+
+    eigen_assert(dimensions_match(m_arg1Impl.dimensions(), m_arg2Impl.dimensions()) && dimensions_match(m_arg1Impl.dimensions(), m_arg3Impl.dimensions()));
+  }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename TensorEvaluator<Arg1Type, Device>::Dimensions Dimensions;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
+  {
+    // TODO: use arg2 or arg3 dimensions if they are known at compile time.
+    return m_arg1Impl.dimensions();
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_arg1Impl.evalSubExprsIfNeeded(NULL);
+    m_arg2Impl.evalSubExprsIfNeeded(NULL);
+    m_arg3Impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_arg1Impl.cleanup();
+    m_arg2Impl.cleanup();
+    m_arg3Impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const
+  {
+    return m_functor(m_arg1Impl.coeff(index), m_arg2Impl.coeff(index), m_arg3Impl.coeff(index));
+  }
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return m_functor.packetOp(m_arg1Impl.template packet<LoadMode>(index),
+                              m_arg2Impl.template packet<LoadMode>(index),
+                              m_arg3Impl.template packet<LoadMode>(index));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    const double functor_cost = internal::functor_traits<TernaryOp>::Cost;
+    return m_arg1Impl.costPerCoeff(vectorized) +
+           m_arg2Impl.costPerCoeff(vectorized) +
+           m_arg3Impl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, functor_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+   // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_arg1Impl.bind(cgh);
+    m_arg2Impl.bind(cgh);
+    m_arg3Impl.bind(cgh);
+  }
+#endif
+
+ private:
+  const TernaryOp m_functor;
+  TensorEvaluator<Arg1Type, Device> m_arg1Impl;
+  TensorEvaluator<Arg2Type, Device> m_arg2Impl;
+  TensorEvaluator<Arg3Type, Device> m_arg3Impl;
+};
+
+
+// -------------------- SelectOp --------------------
+
+template<typename IfArgType, typename ThenArgType, typename ElseArgType, typename Device>
+struct TensorEvaluator<const TensorSelectOp<IfArgType, ThenArgType, ElseArgType>, Device>
+{
+  typedef TensorSelectOp<IfArgType, ThenArgType, ElseArgType> XprType;
+  typedef typename XprType::Scalar Scalar;
+
+  enum {
+    IsAligned         = TensorEvaluator<ThenArgType, Device>::IsAligned &
+                        TensorEvaluator<ElseArgType, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<ThenArgType, Device>::PacketAccess &
+                        TensorEvaluator<ElseArgType, Device>::PacketAccess &
+                        PacketType<Scalar, Device>::HasBlend,
+    BlockAccess       = TensorEvaluator<IfArgType, Device>::BlockAccess &&
+                        TensorEvaluator<ThenArgType, Device>::BlockAccess &&
+                        TensorEvaluator<ElseArgType, Device>::BlockAccess,
+    PreferBlockAccess = TensorEvaluator<IfArgType, Device>::PreferBlockAccess ||
+                        TensorEvaluator<ThenArgType, Device>::PreferBlockAccess ||
+                        TensorEvaluator<ElseArgType, Device>::PreferBlockAccess,
+    Layout            = TensorEvaluator<IfArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
+  };
+
+  TensorEvaluator(const XprType& op, const Device& device)
+    : m_condImpl(op.ifExpression(), device),
+      m_thenImpl(op.thenExpression(), device),
+      m_elseImpl(op.elseExpression(), device)
+  {
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<IfArgType, Device>::Layout) == static_cast<int>(TensorEvaluator<ThenArgType, Device>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<IfArgType, Device>::Layout) == static_cast<int>(TensorEvaluator<ElseArgType, Device>::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    eigen_assert(dimensions_match(m_condImpl.dimensions(), m_thenImpl.dimensions()));
+    eigen_assert(dimensions_match(m_thenImpl.dimensions(), m_elseImpl.dimensions()));
+  }
+
+  typedef typename XprType::Index Index;
+  typedef typename internal::traits<XprType>::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename TensorEvaluator<IfArgType, Device>::Dimensions Dimensions;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  static const int NumDims = internal::array_size<Dimensions>::value;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+    typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename TensorEvaluator<const IfArgType, Device>::TensorBlock
+      IfArgTensorBlock;
+  typedef typename TensorEvaluator<const ThenArgType, Device>::TensorBlock
+      ThenArgTensorBlock;
+  typedef typename TensorEvaluator<const ElseArgType, Device>::TensorBlock
+      ElseArgTensorBlock;
+
+  struct TensorSelectOpBlockFactory {
+    template <typename IfArgXprType, typename ThenArgXprType, typename ElseArgXprType>
+    struct XprType {
+      typedef TensorSelectOp<const IfArgXprType, const ThenArgXprType, const ElseArgXprType> type;
+    };
+
+    template <typename IfArgXprType, typename ThenArgXprType, typename ElseArgXprType>
+    typename XprType<IfArgXprType, ThenArgXprType, ElseArgXprType>::type expr(
+        const IfArgXprType& if_expr, const ThenArgXprType& then_expr, const ElseArgXprType& else_expr) const {
+      return typename XprType<IfArgXprType, ThenArgXprType, ElseArgXprType>::type(if_expr, then_expr, else_expr);
+    }
+  };
+
+  typedef internal::TensorTernaryExprBlock<TensorSelectOpBlockFactory,
+                                           IfArgTensorBlock, ThenArgTensorBlock,
+                                           ElseArgTensorBlock>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
+  {
+    // TODO: use then or else impl instead if they happen to be known at compile time.
+    return m_condImpl.dimensions();
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_condImpl.evalSubExprsIfNeeded(NULL);
+    m_thenImpl.evalSubExprsIfNeeded(NULL);
+    m_elseImpl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    m_condImpl.evalSubExprsIfNeeded(nullptr, [this, done](bool) {
+      m_thenImpl.evalSubExprsIfNeeded(nullptr, [this, done](bool) {
+        m_elseImpl.evalSubExprsIfNeeded(nullptr, [done](bool) { done(true); });
+      });
+    });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_condImpl.cleanup();
+    m_thenImpl.cleanup();
+    m_elseImpl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const
+  {
+    return m_condImpl.coeff(index) ? m_thenImpl.coeff(index) : m_elseImpl.coeff(index);
+  }
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const
+  {
+     internal::Selector<PacketSize> select;
+     EIGEN_UNROLL_LOOP
+     for (Index i = 0; i < PacketSize; ++i) {
+       select.select[i] = m_condImpl.coeff(index+i);
+     }
+     return internal::pblend(select,
+                             m_thenImpl.template packet<LoadMode>(index),
+                             m_elseImpl.template packet<LoadMode>(index));
+
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    return m_condImpl.costPerCoeff(vectorized) +
+           m_thenImpl.costPerCoeff(vectorized)
+        .cwiseMax(m_elseImpl.costPerCoeff(vectorized));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    auto then_req = m_thenImpl.getResourceRequirements();
+    auto else_req = m_elseImpl.getResourceRequirements();
+
+    auto merged_req =
+        internal::TensorBlockResourceRequirements::merge(then_req, else_req);
+    merged_req.cost_per_coeff =
+        then_req.cost_per_coeff.cwiseMax(else_req.cost_per_coeff);
+
+    return internal::TensorBlockResourceRequirements::merge(
+        m_condImpl.getResourceRequirements(), merged_req);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    // It's unsafe to pass destination buffer to underlying expressions, because
+    // output might be aliased with one of the inputs.
+    desc.DropDestinationBuffer();
+
+    return TensorBlock(
+        m_condImpl.block(desc, scratch), m_thenImpl.block(desc, scratch),
+        m_elseImpl.block(desc, scratch), TensorSelectOpBlockFactory());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+ // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_condImpl.bind(cgh);
+    m_thenImpl.bind(cgh);
+    m_elseImpl.bind(cgh);
+  }
+#endif
+ private:
+  TensorEvaluator<IfArgType, Device> m_condImpl;
+  TensorEvaluator<ThenArgType, Device> m_thenImpl;
+  TensorEvaluator<ElseArgType, Device> m_elseImpl;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_EVALUATOR_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorExecutor.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorExecutor.h
new file mode 100644
index 0000000..c52fb77
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorExecutor.h
@@ -0,0 +1,703 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H
+#define EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H
+
+namespace Eigen {
+
+/**
+ * \class TensorExecutor
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief The tensor executor class.
+ *
+ * This class is responsible for launch the evaluation of the expression on
+ * the specified computing device.
+ *
+ * @tparam Vectorizable can use packet math (SSE/AVX/etc... registers and
+ *                      instructions)
+ * @tparam Tiling       can use block based tensor evaluation
+ *                      (see TensorBlock.h)
+ */
+namespace internal {
+
+/**
+ * Evaluating TensorBroadcastingOp via coefficient of packet path is extremely
+ * expensive. If expression has at least one broadcast op in it, and it supports
+ * block based evaluation, we always prefer it, even for the small tensors. For
+ * all other tileable ops, block evaluation overhead for small tensors (fits
+ * into L1) is too large, and we fallback on vectorized evaluation.
+ */
+
+// TODO(ezhulenev): Add specializations for all other types of Tensor ops.
+
+template<typename Expression>
+struct ExpressionHasTensorBroadcastingOp {
+  enum { value = false };
+};
+
+template<typename LhsXprType, typename RhsXprType>
+struct ExpressionHasTensorBroadcastingOp<
+    const TensorAssignOp<LhsXprType, RhsXprType> > {
+  enum { value = ExpressionHasTensorBroadcastingOp<RhsXprType>::value };
+};
+
+template<typename UnaryOp, typename XprType>
+struct ExpressionHasTensorBroadcastingOp<
+    const TensorCwiseUnaryOp<UnaryOp, XprType> > {
+  enum { value = ExpressionHasTensorBroadcastingOp<XprType>::value };
+};
+
+template<typename BinaryOp, typename LhsXprType, typename RhsXprType>
+struct ExpressionHasTensorBroadcastingOp<
+    const TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType> > {
+  enum {
+    value = ExpressionHasTensorBroadcastingOp<LhsXprType>::value ||
+        ExpressionHasTensorBroadcastingOp<RhsXprType>::value
+  };
+};
+
+template<typename Broadcast, typename XprType>
+struct ExpressionHasTensorBroadcastingOp<
+    const TensorBroadcastingOp<Broadcast, XprType> > {
+  enum { value = true };
+};
+
+// -------------------------------------------------------------------------- //
+
+/**
+ * Default strategy: the expression is evaluated sequentially with a single cpu
+ * thread, without vectorization and block evaluation.
+ */
+template <typename Expression, typename Device, bool Vectorizable,
+          TiledEvaluation Tiling>
+class TensorExecutor {
+ public:
+  typedef typename Expression::Index StorageIndex;
+
+  // Including `unsupported/Eigen/CXX11/Tensor` in different translation units
+  // with/without `EIGEN_USE_THREADS` or `EIGEN_USE_GPU` is a potential ODR
+  // violation. If this template is instantiated with a non-default device, it
+  // means that this header file was included without defining
+  // `EIGEN_USE_THREADS`, `EIGEN_USE_GPU` or `EIGEN_USE_SYCL`.
+  static_assert(std::is_same<Device, DefaultDevice>::value,
+                "Default executor instantiated with non-default device. "
+                "You must #define EIGEN_USE_THREADS, EIGEN_USE_GPU or "
+                "EIGEN_USE_SYCL before including Eigen headers.");
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
+                                      const Device& device = Device()) {
+    TensorEvaluator<Expression, Device> evaluator(expr, device);
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
+    if (needs_assign) {
+      const StorageIndex size = array_prod(evaluator.dimensions());
+      for (StorageIndex i = 0; i < size; ++i) {
+        evaluator.evalScalar(i);
+      }
+    }
+    evaluator.cleanup();
+  }
+};
+
+/**
+ * Default async execution strategy is not implemented. Currently it's only
+ * available for ThreadPoolDevice (see definition below).
+ */
+template <typename Expression, typename Device, typename DoneCallback,
+          bool Vectorizable, TiledEvaluation Tiling>
+class TensorAsyncExecutor {};
+
+/**
+ * Process all the data with a single cpu thread, using vectorized instructions.
+ */
+template <typename Expression>
+class TensorExecutor<Expression, DefaultDevice, /*Vectorizable=*/true,
+                     /*Tiling=*/TiledEvaluation::Off> {
+ public:
+  typedef typename Expression::Index StorageIndex;
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(
+      const Expression& expr, const DefaultDevice& device = DefaultDevice()) {
+    TensorEvaluator<Expression, DefaultDevice> evaluator(expr, device);
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
+    if (needs_assign) {
+      const StorageIndex size = array_prod(evaluator.dimensions());
+      const int PacketSize = unpacket_traits<typename TensorEvaluator<
+          Expression, DefaultDevice>::PacketReturnType>::size;
+
+      // Give compiler a strong possibility to unroll the loop. But don't insist
+      // on unrolling, because if the function is expensive compiler should not
+      // unroll the loop at the expense of inlining.
+      const StorageIndex UnrolledSize =
+          (size / (4 * PacketSize)) * 4 * PacketSize;
+      for (StorageIndex i = 0; i < UnrolledSize; i += 4 * PacketSize) {
+        for (StorageIndex j = 0; j < 4; j++) {
+          evaluator.evalPacket(i + j * PacketSize);
+        }
+      }
+      const StorageIndex VectorizedSize = (size / PacketSize) * PacketSize;
+      for (StorageIndex i = UnrolledSize; i < VectorizedSize; i += PacketSize) {
+        evaluator.evalPacket(i);
+      }
+      for (StorageIndex i = VectorizedSize; i < size; ++i) {
+        evaluator.evalScalar(i);
+      }
+    }
+    evaluator.cleanup();
+  }
+};
+
+/**
+ * Process all the data with a single cpu thread, using blocks of data. By
+ * sizing a block to fit L1 cache we get better cache performance.
+ */
+template <typename Expression, bool Vectorizable>
+class TensorExecutor<Expression, DefaultDevice, Vectorizable,
+                     /*Tiling=*/TiledEvaluation::On> {
+ public:
+  typedef typename traits<Expression>::Scalar Scalar;
+  typedef typename remove_const<Scalar>::type ScalarNoConst;
+
+  typedef TensorEvaluator<Expression, DefaultDevice> Evaluator;
+  typedef typename traits<Expression>::Index StorageIndex;
+
+  static const int NumDims = traits<Expression>::NumDimensions;
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
+                         const DefaultDevice& device = DefaultDevice()) {
+    typedef TensorBlockMapper<NumDims, Evaluator::Layout, StorageIndex>
+        TensorBlockMapper;
+
+    typedef internal::TensorBlockDescriptor<NumDims, StorageIndex>
+        TensorBlockDesc;
+    typedef internal::TensorBlockScratchAllocator<DefaultDevice>
+        TensorBlockScratch;
+
+    Evaluator evaluator(expr, device);
+
+    // TODO(ezhulenev): Do not use tiling for small tensors?
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
+
+    if (needs_assign) {
+      // Query expression tree for desired block size/shape.
+      const TensorBlockResourceRequirements requirements =
+          evaluator.getResourceRequirements();
+
+      const TensorBlockMapper block_mapper(
+          typename TensorBlockDesc::Dimensions(evaluator.dimensions()),
+          requirements);
+
+      // Share scratch memory allocator between all blocks.
+      TensorBlockScratch scratch(device);
+
+      const StorageIndex total_block_count = block_mapper.blockCount();
+      for (StorageIndex i = 0; i < total_block_count; ++i) {
+        TensorBlockDesc desc = block_mapper.blockDescriptor(i);
+        evaluator.evalBlock(desc, scratch);
+        scratch.reset();
+      }
+    }
+    evaluator.cleanup();
+  }
+};
+
+/**
+ * Multicore strategy: the index space is partitioned and each partition is
+ * executed on a single core.
+ *
+ * (1) TensorExecutor will submit work to the ThreadPoolDevice managed thread
+ *     pool, and will block the caller thread until all tasks are finished.
+ *
+ * (2) TensorAsyncExecutor is a non-blocking version, that will submit work to
+ *     the ThreadPoolDevice managed thread pool, and will return immediately.
+ *     It will call 'done' callback after all tasks are finished.
+ */
+#ifdef EIGEN_USE_THREADS
+
+template <typename TensorBlockMapper>
+struct TensorExecutorTilingContext {
+  TensorExecutorTilingContext() = default;
+  TensorExecutorTilingContext(const TensorBlockMapper& b_mapper,
+                              const TensorOpCost& b_cost, size_t b_aligned_size)
+      : block_mapper(b_mapper),
+        cost(b_cost),
+        aligned_blocksize(b_aligned_size) {}
+
+  TensorBlockMapper block_mapper;  // navigate through blocks
+  TensorOpCost cost;               // cost of computing a single block
+  size_t aligned_blocksize;        // block size after memory alignment
+};
+
+// Computes a block evaluation parameters, and allocates temporary memory buffer
+// for blocks. See TensorExecutor/TensorAsyncExecutor (Tiling=On) below.
+template <typename Evaluator, typename TensorBlockMapper, bool Vectorizable>
+TensorExecutorTilingContext<TensorBlockMapper> GetTensorExecutorTilingContext(
+    const Evaluator& evaluator) {
+  // Query expression tree for desired block size/shape.
+  TensorBlockResourceRequirements requirements =
+      evaluator.getResourceRequirements();
+
+  // Update target block size based on cost model.
+  double taskSize = TensorCostModel<ThreadPoolDevice>::taskSize(
+      1, requirements.cost_per_coeff);
+  requirements.size = static_cast<size_t>(1.0 / taskSize);
+
+  TensorBlockMapper block_mapper(
+      typename TensorBlockMapper::Dimensions(evaluator.dimensions()),
+      requirements);
+
+  size_t block_size = block_mapper.blockTotalSize();
+  const size_t align = numext::maxi(EIGEN_MAX_ALIGN_BYTES, 1);
+  const size_t aligned_blocksize =
+      align *
+      divup<size_t>(block_size * sizeof(typename Evaluator::Scalar), align);
+
+  return {block_mapper, requirements.cost_per_coeff * block_size,
+          aligned_blocksize};
+}
+
+template <typename Evaluator, typename StorageIndex, bool Vectorizable>
+struct EvalRange {
+  static void run(Evaluator* evaluator_in, const StorageIndex firstIdx,
+                  const StorageIndex lastIdx) {
+    Evaluator evaluator = *evaluator_in;
+    eigen_assert(lastIdx >= firstIdx);
+    for (StorageIndex i = firstIdx; i < lastIdx; ++i) {
+      evaluator.evalScalar(i);
+    }
+  }
+
+  static StorageIndex alignBlockSize(StorageIndex size) { return size; }
+};
+
+template <typename Evaluator, typename StorageIndex>
+struct EvalRange<Evaluator, StorageIndex, /*Vectorizable*/ true> {
+  static const int PacketSize =
+      unpacket_traits<typename Evaluator::PacketReturnType>::size;
+
+  static void run(Evaluator* evaluator_in, const StorageIndex firstIdx,
+                  const StorageIndex lastIdx) {
+    Evaluator evaluator = *evaluator_in;
+    eigen_assert(lastIdx >= firstIdx);
+    StorageIndex i = firstIdx;
+    if (lastIdx - firstIdx >= PacketSize) {
+      eigen_assert(firstIdx % PacketSize == 0);
+      StorageIndex last_chunk_offset = lastIdx - 4 * PacketSize;
+      // Give compiler a strong possibility to unroll the loop. But don't insist
+      // on unrolling, because if the function is expensive compiler should not
+      // unroll the loop at the expense of inlining.
+      for (; i <= last_chunk_offset; i += 4 * PacketSize) {
+        for (StorageIndex j = 0; j < 4; j++) {
+          evaluator.evalPacket(i + j * PacketSize);
+        }
+      }
+      last_chunk_offset = lastIdx - PacketSize;
+      for (; i <= last_chunk_offset; i += PacketSize) {
+        evaluator.evalPacket(i);
+      }
+    }
+    for (; i < lastIdx; ++i) {
+      evaluator.evalScalar(i);
+    }
+  }
+
+  static StorageIndex alignBlockSize(StorageIndex size) {
+    // Align block size to packet size and account for unrolling in run above.
+    if (size >= 16 * PacketSize) {
+      return (size + 4 * PacketSize - 1) & ~(4 * PacketSize - 1);
+    }
+    // Aligning to 4 * PacketSize would increase block size by more than 25%.
+    return (size + PacketSize - 1) & ~(PacketSize - 1);
+  }
+};
+
+template <typename Expression, bool Vectorizable, TiledEvaluation Tiling>
+class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tiling> {
+ public:
+  typedef typename Expression::Index StorageIndex;
+
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
+                         const ThreadPoolDevice& device) {
+    typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
+    typedef EvalRange<Evaluator, StorageIndex, Vectorizable> EvalRange;
+
+    Evaluator evaluator(expr, device);
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
+    if (needs_assign) {
+      const StorageIndex size = array_prod(evaluator.dimensions());
+      device.parallelFor(size, evaluator.costPerCoeff(Vectorizable),
+                         EvalRange::alignBlockSize,
+                         [&evaluator](StorageIndex firstIdx, StorageIndex lastIdx) {
+                           EvalRange::run(&evaluator, firstIdx, lastIdx);
+                         });
+    }
+    evaluator.cleanup();
+  }
+};
+
+template <typename Expression, bool Vectorizable>
+class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable,
+                     /*Tiling=*/TiledEvaluation::On> {
+ public:
+  typedef typename traits<Expression>::Index IndexType;
+  typedef typename traits<Expression>::Scalar Scalar;
+  typedef typename remove_const<Scalar>::type ScalarNoConst;
+
+  static const int NumDims = traits<Expression>::NumDimensions;
+
+  typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
+  typedef TensorBlockMapper<NumDims, Evaluator::Layout, IndexType> BlockMapper;
+  typedef TensorExecutorTilingContext<BlockMapper> TilingContext;
+
+  typedef internal::TensorBlockDescriptor<NumDims, IndexType>
+      TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<ThreadPoolDevice>
+      TensorBlockScratch;
+
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
+                                      const ThreadPoolDevice& device) {
+    Evaluator evaluator(expr, device);
+
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
+    if (needs_assign) {
+      const TilingContext tiling =
+          internal::GetTensorExecutorTilingContext<Evaluator, BlockMapper,
+                                                   Vectorizable>(evaluator);
+
+      auto eval_block = [&device, &evaluator, &tiling](IndexType firstBlockIdx,
+                                                       IndexType lastBlockIdx) {
+        TensorBlockScratch scratch(device);
+
+        for (IndexType block_idx = firstBlockIdx; block_idx < lastBlockIdx;
+             ++block_idx) {
+          TensorBlockDesc desc = tiling.block_mapper.blockDescriptor(block_idx);
+          evaluator.evalBlock(desc, scratch);
+          scratch.reset();
+        }
+      };
+
+      // Evaluate small expressions directly as a single block.
+      if (tiling.block_mapper.blockCount() == 1) {
+        TensorBlockScratch scratch(device);
+        TensorBlockDesc desc(0, tiling.block_mapper.blockDimensions());
+        evaluator.evalBlock(desc, scratch);
+      } else {
+        device.parallelFor(tiling.block_mapper.blockCount(), tiling.cost,
+                           eval_block);
+      }
+    }
+    evaluator.cleanup();
+  }
+};
+
+template <typename Expression, typename DoneCallback, bool Vectorizable,
+          TiledEvaluation Tiling>
+class TensorAsyncExecutor<Expression, ThreadPoolDevice, DoneCallback,
+                          Vectorizable, Tiling> {
+ public:
+  typedef typename Expression::Index StorageIndex;
+  typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
+
+  static EIGEN_STRONG_INLINE void runAsync(const Expression& expr,
+                                           const ThreadPoolDevice& device,
+                                           DoneCallback done) {
+    TensorAsyncExecutorContext* const ctx =
+        new TensorAsyncExecutorContext(expr, device, std::move(done));
+
+    const auto on_eval_subexprs = [ctx, &device](bool need_assign) -> void {
+      if (!need_assign) {
+        delete ctx;
+        return;
+      }
+
+      typedef EvalRange<Evaluator, StorageIndex, Vectorizable> EvalRange;
+      const StorageIndex size = array_prod(ctx->evaluator.dimensions());
+      device.parallelForAsync(
+          size, ctx->evaluator.costPerCoeff(Vectorizable),
+          EvalRange::alignBlockSize,
+          [ctx](StorageIndex firstIdx, StorageIndex lastIdx) {
+            EvalRange::run(&ctx->evaluator, firstIdx, lastIdx);
+          },
+          [ctx]() { delete ctx; });
+    };
+
+    ctx->evaluator.evalSubExprsIfNeededAsync(nullptr, on_eval_subexprs);
+  }
+
+ private:
+  struct TensorAsyncExecutorContext {
+    TensorAsyncExecutorContext(const Expression& expr,
+                               const ThreadPoolDevice& thread_pool,
+                               DoneCallback done)
+        : evaluator(expr, thread_pool), on_done(std::move(done)) {}
+
+    ~TensorAsyncExecutorContext() {
+      evaluator.cleanup();
+      on_done();
+    }
+
+    Evaluator evaluator;
+
+   private:
+    DoneCallback on_done;
+  };
+};
+
+template <typename Expression, typename DoneCallback, bool Vectorizable>
+class TensorAsyncExecutor<Expression, ThreadPoolDevice, DoneCallback,
+                          Vectorizable, /*Tileable*/ TiledEvaluation::On> {
+ public:
+  typedef typename traits<Expression>::Index IndexType;
+  typedef typename traits<Expression>::Scalar Scalar;
+  typedef typename remove_const<Scalar>::type ScalarNoConst;
+
+  static const int NumDims = traits<Expression>::NumDimensions;
+
+  typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
+  typedef TensorBlockMapper<NumDims, Evaluator::Layout, IndexType> BlockMapper;
+  typedef TensorExecutorTilingContext<BlockMapper> TilingContext;
+
+  typedef internal::TensorBlockDescriptor<NumDims, IndexType> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<ThreadPoolDevice>
+      TensorBlockScratch;
+
+  static EIGEN_STRONG_INLINE void runAsync(const Expression& expr,
+                                           const ThreadPoolDevice& device,
+                                           DoneCallback done) {
+
+    TensorAsyncExecutorContext* const ctx =
+        new TensorAsyncExecutorContext(expr, device, std::move(done));
+
+    const auto on_eval_subexprs = [ctx](bool need_assign) -> void {
+      if (!need_assign) {
+        delete ctx;
+        return;
+      }
+
+      ctx->tiling = internal::GetTensorExecutorTilingContext<
+          Evaluator, BlockMapper, Vectorizable>(ctx->evaluator);
+
+      auto eval_block = [ctx](IndexType firstBlockIdx, IndexType lastBlockIdx) {
+        TensorBlockScratch scratch(ctx->device);
+
+        for (IndexType block_idx = firstBlockIdx; block_idx < lastBlockIdx;
+             ++block_idx) {
+          TensorBlockDesc desc =
+              ctx->tiling.block_mapper.blockDescriptor(block_idx);
+          ctx->evaluator.evalBlock(desc, scratch);
+          scratch.reset();
+        }
+      };
+
+      // Evaluate small expressions directly as a single block.
+      if (ctx->tiling.block_mapper.blockCount() == 1) {
+        TensorBlockScratch scratch(ctx->device);
+        TensorBlockDesc desc(0, ctx->tiling.block_mapper.blockDimensions());
+        ctx->evaluator.evalBlock(desc, scratch);
+        delete ctx;
+      } else {
+        ctx->device.parallelForAsync(ctx->tiling.block_mapper.blockCount(),
+                                     ctx->tiling.cost, eval_block,
+                                     [ctx]() { delete ctx; });
+      }
+    };
+
+    ctx->evaluator.evalSubExprsIfNeededAsync(nullptr, on_eval_subexprs);
+  }
+
+ private:
+  struct TensorAsyncExecutorContext {
+    TensorAsyncExecutorContext(const Expression& expr,
+                               const ThreadPoolDevice& thread_pool,
+                               DoneCallback done)
+        : device(thread_pool),
+          evaluator(expr, thread_pool),
+          on_done(std::move(done)) {}
+
+    ~TensorAsyncExecutorContext() {
+      evaluator.cleanup();
+      on_done();
+    }
+
+    const ThreadPoolDevice& device;
+    Evaluator evaluator;
+    TilingContext tiling;
+
+   private:
+    DoneCallback on_done;
+  };
+};
+
+#endif  // EIGEN_USE_THREADS
+
+// GPU: the evaluation of the expression is offloaded to a GPU.
+#if defined(EIGEN_USE_GPU)
+
+template <typename Expression, bool Vectorizable, TiledEvaluation Tiling>
+class TensorExecutor<Expression, GpuDevice, Vectorizable, Tiling> {
+ public:
+  typedef typename Expression::Index StorageIndex;
+  static void run(const Expression& expr, const GpuDevice& device);
+};
+
+#if defined(EIGEN_GPUCC)
+template <typename Evaluator, typename StorageIndex, bool Vectorizable>
+struct EigenMetaKernelEval {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  void run(Evaluator& eval, StorageIndex firstIdx, StorageIndex lastIdx, StorageIndex step_size) {
+    for (StorageIndex i = firstIdx; i < lastIdx; i += step_size) {
+      eval.evalScalar(i);
+    }
+  }
+};
+
+template <typename Evaluator, typename StorageIndex>
+struct EigenMetaKernelEval<Evaluator, StorageIndex, true> {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  void run(Evaluator& eval, StorageIndex firstIdx, StorageIndex lastIdx, StorageIndex step_size) {
+    const StorageIndex PacketSize = unpacket_traits<typename Evaluator::PacketReturnType>::size;
+    const StorageIndex vectorized_size = (lastIdx / PacketSize) * PacketSize;
+    const StorageIndex vectorized_step_size = step_size * PacketSize;
+
+    // Use the vector path
+    for (StorageIndex i = firstIdx * PacketSize; i < vectorized_size;
+         i += vectorized_step_size) {
+      eval.evalPacket(i);
+    }
+    for (StorageIndex i = vectorized_size + firstIdx; i < lastIdx; i += step_size) {
+      eval.evalScalar(i);
+    }
+  }
+};
+
+template <typename Evaluator, typename StorageIndex>
+__global__ void
+__launch_bounds__(1024)
+EigenMetaKernel(Evaluator eval, StorageIndex size) {
+
+  const StorageIndex first_index = blockIdx.x * blockDim.x + threadIdx.x;
+  const StorageIndex step_size = blockDim.x * gridDim.x;
+
+  const bool vectorizable = Evaluator::PacketAccess & Evaluator::IsAligned;
+  EigenMetaKernelEval<Evaluator, StorageIndex, vectorizable>::run(eval, first_index, size, step_size);
+}
+
+/*static*/
+template <typename Expression, bool Vectorizable, TiledEvaluation Tiling>
+EIGEN_STRONG_INLINE void TensorExecutor<Expression, GpuDevice, Vectorizable, Tiling>::run(
+    const Expression& expr, const GpuDevice& device) {
+  TensorEvaluator<Expression, GpuDevice> evaluator(expr, device);
+  const bool needs_assign = evaluator.evalSubExprsIfNeeded(nullptr);
+  if (needs_assign) {
+
+    const int block_size = device.maxGpuThreadsPerBlock();
+    const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / block_size;
+    const StorageIndex size = array_prod(evaluator.dimensions());
+    // Create a least one block to ensure we won't crash when tensorflow calls with tensors of size 0.
+    const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, divup<int>(size, block_size)), 1);
+
+    LAUNCH_GPU_KERNEL(
+        (EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, StorageIndex>),
+        num_blocks, block_size, 0, device, evaluator, size);
+  }
+  evaluator.cleanup();
+}
+
+#endif  // EIGEN_GPUCC
+#endif  // EIGEN_USE_GPU
+
+// SYCL Executor policy
+#ifdef EIGEN_USE_SYCL
+
+template <typename Evaluator>
+struct ExecExprFunctorKernel {
+  typedef typename Evaluator::Index Index;
+  Evaluator evaluator;
+  const Index range;
+  template <typename Scratch>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE ExecExprFunctorKernel(
+      const Scratch, Evaluator evaluator_, const Index range_)
+      : evaluator(evaluator_), range(range_) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void operator()(
+      cl::sycl::nd_item<1> itemID) {
+    compute(itemID);
+  }
+  template <bool is_vec = Evaluator::PacketAccess>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE typename std::enable_if<!is_vec>::type
+  compute(const cl::sycl::nd_item<1>& itemID) {
+    Index gId = static_cast<Index>(itemID.get_global_linear_id());
+    Index total_threads = itemID.get_global_range(0);
+
+    for (Index i = gId; i < range; i += total_threads) {
+      evaluator.evalScalar(i);
+    }
+  }
+  template <bool is_vec = Evaluator::PacketAccess>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE typename std::enable_if<is_vec>::type
+  compute(const cl::sycl::nd_item<1>& itemID) {
+    const Index vectorizedRange =
+        (range / Evaluator::PacketSize) * Evaluator::PacketSize;
+    Index gId = static_cast<Index>(itemID.get_global_linear_id());
+    const Index step = Evaluator::PacketSize * itemID.get_global_range(0);
+    const Index start = Evaluator::PacketSize * gId;
+    for (Index i = start; i < vectorizedRange; i += step) {
+      evaluator.evalPacket(i);
+    }
+    gId += vectorizedRange;
+    for (Index i = gId; i < range; i += itemID.get_global_range(0)) {
+      evaluator.evalScalar(i);
+    }
+  }
+};
+
+template <typename Expression, bool Vectorizable, TiledEvaluation Tiling>
+class TensorExecutor<Expression, Eigen::SyclDevice, Vectorizable, Tiling> {
+ public:
+  typedef typename Expression::Index Index;
+  static EIGEN_STRONG_INLINE void run(const Expression& expr,
+                                      const Eigen::SyclDevice& dev) {
+    typedef Eigen::TensorEvaluator<Expression, Eigen::SyclDevice> Evaluator;
+    Evaluator evaluator(expr, dev);
+    const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
+    if (needs_assign) {
+      Index range, GRange, tileSize;
+      Index total_size = ::Eigen::internal::array_prod(evaluator.dimensions());
+      total_size = (total_size == 0) ? 1 : total_size;
+      const int PacketSize =
+          Eigen::PacketType<typename Evaluator::CoeffReturnType,
+                            Eigen::SyclDevice>::size;
+      Index vectorizable_threads = static_cast<Index>(total_size / PacketSize);
+      dev.parallel_for_setup(vectorizable_threads, tileSize, range, GRange);
+      range = total_size;
+
+      dev.template nullary_kernel_launcher<
+          typename Evaluator::CoeffReturnType,
+          ExecExprFunctorKernel<Evaluator> >(
+          evaluator,
+          cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange),
+                                cl::sycl::range<1>(tileSize)),
+          Index(1), range);
+    }
+    evaluator.cleanup();
+  }
+};
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorExpr.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorExpr.h
new file mode 100644
index 0000000..c9bccfc
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorExpr.h
@@ -0,0 +1,388 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_EXPR_H
+#define EIGEN_CXX11_TENSOR_TENSOR_EXPR_H
+
+namespace Eigen {
+
+/** \class TensorExpr
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor expression classes.
+  *
+  * The TensorCwiseNullaryOp class applies a nullary operators to an expression.
+  * This is typically used to generate constants.
+  *
+  * The TensorCwiseUnaryOp class represents an expression where a unary operator
+  * (e.g. cwiseSqrt) is applied to an expression.
+  *
+  * The TensorCwiseBinaryOp class represents an expression where a binary
+  * operator (e.g. addition) is applied to a lhs and a rhs expression.
+  *
+  */
+namespace internal {
+template<typename NullaryOp, typename XprType>
+struct traits<TensorCwiseNullaryOp<NullaryOp, XprType> >
+    : traits<XprType>
+{
+  typedef traits<XprType> XprTraits;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::Nested XprTypeNested;
+  typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+  enum {
+    Flags = 0
+  };
+};
+
+}  // end namespace internal
+
+
+
+template<typename NullaryOp, typename XprType>
+class TensorCwiseNullaryOp : public TensorBase<TensorCwiseNullaryOp<NullaryOp, XprType>, ReadOnlyAccessors>
+{
+  public:
+    typedef typename Eigen::internal::traits<TensorCwiseNullaryOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename XprType::CoeffReturnType CoeffReturnType;
+    typedef TensorCwiseNullaryOp<NullaryOp, XprType> Nested;
+    typedef typename Eigen::internal::traits<TensorCwiseNullaryOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorCwiseNullaryOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseNullaryOp(const XprType& xpr, const NullaryOp& func = NullaryOp())
+        : m_xpr(xpr), m_functor(func) {}
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    nestedExpression() const { return m_xpr; }
+
+    EIGEN_DEVICE_FUNC
+    const NullaryOp& functor() const { return m_functor; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const NullaryOp m_functor;
+};
+
+
+
+namespace internal {
+template<typename UnaryOp, typename XprType>
+struct traits<TensorCwiseUnaryOp<UnaryOp, XprType> >
+    : traits<XprType>
+{
+  // TODO(phli): Add InputScalar, InputPacket.  Check references to
+  // current Scalar/Packet to see if the intent is Input or Output.
+  typedef typename result_of<UnaryOp(typename XprType::Scalar)>::type Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprType::Nested XprTypeNested;
+  typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename TypeConversion<Scalar, 
+                                  typename XprTraits::PointerType
+                                  >::type 
+                                  PointerType;
+};
+
+template<typename UnaryOp, typename XprType>
+struct eval<TensorCwiseUnaryOp<UnaryOp, XprType>, Eigen::Dense>
+{
+  typedef const TensorCwiseUnaryOp<UnaryOp, XprType>& type;
+};
+
+template<typename UnaryOp, typename XprType>
+struct nested<TensorCwiseUnaryOp<UnaryOp, XprType>, 1, typename eval<TensorCwiseUnaryOp<UnaryOp, XprType> >::type>
+{
+  typedef TensorCwiseUnaryOp<UnaryOp, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename UnaryOp, typename XprType>
+class TensorCwiseUnaryOp : public TensorBase<TensorCwiseUnaryOp<UnaryOp, XprType>, ReadOnlyAccessors>
+{
+  public:
+    // TODO(phli): Add InputScalar, InputPacket.  Check references to
+    // current Scalar/Packet to see if the intent is Input or Output.
+    typedef typename Eigen::internal::traits<TensorCwiseUnaryOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef Scalar CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorCwiseUnaryOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorCwiseUnaryOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorCwiseUnaryOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp())
+      : m_xpr(xpr), m_functor(func) {}
+
+    EIGEN_DEVICE_FUNC
+    const UnaryOp& functor() const { return m_functor; }
+
+    /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    nestedExpression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const UnaryOp m_functor;
+};
+
+
+namespace internal {
+template<typename BinaryOp, typename LhsXprType, typename RhsXprType>
+struct traits<TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs
+  // are different.
+  // TODO(phli): Add Lhs/RhsScalar, Lhs/RhsPacket.  Check references to
+  // current Scalar/Packet to see if the intent is Inputs or Output.
+  typedef typename result_of<
+      BinaryOp(typename LhsXprType::Scalar,
+               typename RhsXprType::Scalar)>::type Scalar;
+  typedef traits<LhsXprType> XprTraits;
+  typedef typename promote_storage_type<
+      typename traits<LhsXprType>::StorageKind,
+      typename traits<RhsXprType>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<
+      typename traits<LhsXprType>::Index,
+      typename traits<RhsXprType>::Index>::type Index;
+  typedef typename LhsXprType::Nested LhsNested;
+  typedef typename RhsXprType::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename TypeConversion<Scalar,
+                                  typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                                                      typename traits<LhsXprType>::PointerType,
+                                                      typename traits<RhsXprType>::PointerType>::type
+                                  >::type 
+                                  PointerType;
+  enum {
+    Flags = 0
+  };
+};
+
+template<typename BinaryOp, typename LhsXprType, typename RhsXprType>
+struct eval<TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType>, Eigen::Dense>
+{
+  typedef const TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType>& type;
+};
+
+template<typename BinaryOp, typename LhsXprType, typename RhsXprType>
+struct nested<TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType>, 1, typename eval<TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType> >::type>
+{
+  typedef TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename BinaryOp, typename LhsXprType, typename RhsXprType>
+class TensorCwiseBinaryOp : public TensorBase<TensorCwiseBinaryOp<BinaryOp, LhsXprType, RhsXprType>, ReadOnlyAccessors>
+{
+  public:
+    // TODO(phli): Add Lhs/RhsScalar, Lhs/RhsPacket.  Check references to
+    // current Scalar/Packet to see if the intent is Inputs or Output.
+    typedef typename Eigen::internal::traits<TensorCwiseBinaryOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef Scalar CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorCwiseBinaryOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorCwiseBinaryOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorCwiseBinaryOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseBinaryOp(const LhsXprType& lhs, const RhsXprType& rhs, const BinaryOp& func = BinaryOp())
+        : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_functor(func) {}
+
+    EIGEN_DEVICE_FUNC
+    const BinaryOp& functor() const { return m_functor; }
+
+    /** \returns the nested expressions */
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename LhsXprType::Nested>::type&
+    lhsExpression() const { return m_lhs_xpr; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename RhsXprType::Nested>::type&
+    rhsExpression() const { return m_rhs_xpr; }
+
+  protected:
+    typename LhsXprType::Nested m_lhs_xpr;
+    typename RhsXprType::Nested m_rhs_xpr;
+    const BinaryOp m_functor;
+};
+
+
+namespace internal {
+template<typename TernaryOp, typename Arg1XprType, typename Arg2XprType, typename Arg3XprType>
+struct traits<TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprType> >
+{
+  // Type promotion to handle the case where the types of the args are different.
+  typedef typename result_of<
+      TernaryOp(typename Arg1XprType::Scalar,
+                typename Arg2XprType::Scalar,
+                typename Arg3XprType::Scalar)>::type Scalar;
+  typedef traits<Arg1XprType> XprTraits;
+  typedef typename traits<Arg1XprType>::StorageKind StorageKind;
+  typedef typename traits<Arg1XprType>::Index Index;
+  typedef typename Arg1XprType::Nested Arg1Nested;
+  typedef typename Arg2XprType::Nested Arg2Nested;
+  typedef typename Arg3XprType::Nested Arg3Nested;
+  typedef typename remove_reference<Arg1Nested>::type _Arg1Nested;
+  typedef typename remove_reference<Arg2Nested>::type _Arg2Nested;
+  typedef typename remove_reference<Arg3Nested>::type _Arg3Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename TypeConversion<Scalar,
+                                  typename conditional<Pointer_type_promotion<typename Arg2XprType::Scalar, Scalar>::val,
+                                                      typename traits<Arg2XprType>::PointerType,
+                                                      typename traits<Arg3XprType>::PointerType>::type
+                                  >::type 
+                                  PointerType;
+  enum {
+    Flags = 0
+  };
+};
+
+template<typename TernaryOp, typename Arg1XprType, typename Arg2XprType, typename Arg3XprType>
+struct eval<TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprType>, Eigen::Dense>
+{
+  typedef const TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprType>& type;
+};
+
+template<typename TernaryOp, typename Arg1XprType, typename Arg2XprType, typename Arg3XprType>
+struct nested<TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprType>, 1, typename eval<TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprType> >::type>
+{
+  typedef TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename TernaryOp, typename Arg1XprType, typename Arg2XprType, typename Arg3XprType>
+class TensorCwiseTernaryOp : public TensorBase<TensorCwiseTernaryOp<TernaryOp, Arg1XprType, Arg2XprType, Arg3XprType>, ReadOnlyAccessors>
+{
+  public:
+    typedef typename Eigen::internal::traits<TensorCwiseTernaryOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef Scalar CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorCwiseTernaryOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorCwiseTernaryOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorCwiseTernaryOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseTernaryOp(const Arg1XprType& arg1, const Arg2XprType& arg2, const Arg3XprType& arg3, const TernaryOp& func = TernaryOp())
+        : m_arg1_xpr(arg1), m_arg2_xpr(arg2), m_arg3_xpr(arg3), m_functor(func) {}
+
+    EIGEN_DEVICE_FUNC
+    const TernaryOp& functor() const { return m_functor; }
+
+    /** \returns the nested expressions */
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename Arg1XprType::Nested>::type&
+    arg1Expression() const { return m_arg1_xpr; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename Arg2XprType::Nested>::type&
+    arg2Expression() const { return m_arg2_xpr; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename Arg3XprType::Nested>::type&
+    arg3Expression() const { return m_arg3_xpr; }
+
+  protected:
+    typename Arg1XprType::Nested m_arg1_xpr;
+    typename Arg2XprType::Nested m_arg2_xpr;
+    typename Arg3XprType::Nested m_arg3_xpr;
+    const TernaryOp m_functor;
+};
+
+
+namespace internal {
+template<typename IfXprType, typename ThenXprType, typename ElseXprType>
+struct traits<TensorSelectOp<IfXprType, ThenXprType, ElseXprType> >
+    : traits<ThenXprType>
+{
+  typedef typename traits<ThenXprType>::Scalar Scalar;
+  typedef traits<ThenXprType> XprTraits;
+  typedef typename promote_storage_type<typename traits<ThenXprType>::StorageKind,
+                                        typename traits<ElseXprType>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<ElseXprType>::Index,
+                                      typename traits<ThenXprType>::Index>::type Index;
+  typedef typename IfXprType::Nested IfNested;
+  typedef typename ThenXprType::Nested ThenNested;
+  typedef typename ElseXprType::Nested ElseNested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename conditional<Pointer_type_promotion<typename ThenXprType::Scalar, Scalar>::val,
+                               typename traits<ThenXprType>::PointerType,
+                               typename traits<ElseXprType>::PointerType>::type PointerType;
+};
+
+template<typename IfXprType, typename ThenXprType, typename ElseXprType>
+struct eval<TensorSelectOp<IfXprType, ThenXprType, ElseXprType>, Eigen::Dense>
+{
+  typedef const TensorSelectOp<IfXprType, ThenXprType, ElseXprType>& type;
+};
+
+template<typename IfXprType, typename ThenXprType, typename ElseXprType>
+struct nested<TensorSelectOp<IfXprType, ThenXprType, ElseXprType>, 1, typename eval<TensorSelectOp<IfXprType, ThenXprType, ElseXprType> >::type>
+{
+  typedef TensorSelectOp<IfXprType, ThenXprType, ElseXprType> type;
+};
+
+}  // end namespace internal
+
+
+template<typename IfXprType, typename ThenXprType, typename ElseXprType>
+class TensorSelectOp : public TensorBase<TensorSelectOp<IfXprType, ThenXprType, ElseXprType>, ReadOnlyAccessors>
+{
+  public:
+    typedef typename Eigen::internal::traits<TensorSelectOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename internal::promote_storage_type<typename ThenXprType::CoeffReturnType,
+                                                    typename ElseXprType::CoeffReturnType>::ret CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorSelectOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorSelectOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorSelectOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC
+    TensorSelectOp(const IfXprType& a_condition,
+                   const ThenXprType& a_then,
+                   const ElseXprType& a_else)
+      : m_condition(a_condition), m_then(a_then), m_else(a_else)
+    { }
+
+    EIGEN_DEVICE_FUNC
+    const IfXprType& ifExpression() const { return m_condition; }
+
+    EIGEN_DEVICE_FUNC
+    const ThenXprType& thenExpression() const { return m_then; }
+
+    EIGEN_DEVICE_FUNC
+    const ElseXprType& elseExpression() const { return m_else; }
+
+  protected:
+    typename IfXprType::Nested m_condition;
+    typename ThenXprType::Nested m_then;
+    typename ElseXprType::Nested m_else;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_EXPR_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorFFT.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorFFT.h
new file mode 100644
index 0000000..4a1a068
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorFFT.h
@@ -0,0 +1,669 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Jianwei Cui <thucjw@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_FFT_H
+#define EIGEN_CXX11_TENSOR_TENSOR_FFT_H
+
+namespace Eigen {
+
+/** \class TensorFFT
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor FFT class.
+  *
+  * TODO:
+  * Vectorize the Cooley Tukey and the Bluestein algorithm
+  * Add support for multithreaded evaluation
+  * Improve the performance on GPU
+  */
+
+template <bool NeedUprade> struct MakeComplex {
+  template <typename T>
+  EIGEN_DEVICE_FUNC
+  T operator() (const T& val) const { return val; }
+};
+
+template <> struct MakeComplex<true> {
+  template <typename T>
+  EIGEN_DEVICE_FUNC
+  std::complex<T> operator() (const T& val) const { return std::complex<T>(val, 0); }
+};
+
+template <> struct MakeComplex<false> {
+  template <typename T>
+  EIGEN_DEVICE_FUNC
+  std::complex<T> operator() (const std::complex<T>& val) const { return val; }
+};
+
+template <int ResultType> struct PartOf {
+  template <typename T> T operator() (const T& val) const { return val; }
+};
+
+template <> struct PartOf<RealPart> {
+  template <typename T> T operator() (const std::complex<T>& val) const { return val.real(); }
+};
+
+template <> struct PartOf<ImagPart> {
+  template <typename T> T operator() (const std::complex<T>& val) const { return val.imag(); }
+};
+
+namespace internal {
+template <typename FFT, typename XprType, int FFTResultType, int FFTDir>
+struct traits<TensorFFTOp<FFT, XprType, FFTResultType, FFTDir> > : public traits<XprType> {
+  typedef traits<XprType> XprTraits;
+  typedef typename NumTraits<typename XprTraits::Scalar>::Real RealScalar;
+  typedef typename std::complex<RealScalar> ComplexScalar;
+  typedef typename XprTraits::Scalar InputScalar;
+  typedef typename conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename traits<XprType>::PointerType PointerType;
+};
+
+template <typename FFT, typename XprType, int FFTResultType, int FFTDirection>
+struct eval<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>, Eigen::Dense> {
+  typedef const TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>& type;
+};
+
+template <typename FFT, typename XprType, int FFTResultType, int FFTDirection>
+struct nested<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>, 1, typename eval<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection> >::type> {
+  typedef TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection> type;
+};
+
+}  // end namespace internal
+
+template <typename FFT, typename XprType, int FFTResultType, int FFTDir>
+class TensorFFTOp : public TensorBase<TensorFFTOp<FFT, XprType, FFTResultType, FFTDir>, ReadOnlyAccessors> {
+ public:
+  typedef typename Eigen::internal::traits<TensorFFTOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename std::complex<RealScalar> ComplexScalar;
+  typedef typename internal::conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
+  typedef OutputScalar CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorFFTOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorFFTOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorFFTOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorFFTOp(const XprType& expr, const FFT& fft)
+      : m_xpr(expr), m_fft(fft) {}
+
+  EIGEN_DEVICE_FUNC
+  const FFT& fft() const { return m_fft; }
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename XprType::Nested>::type& expression() const {
+    return m_xpr;
+  }
+
+ protected:
+  typename XprType::Nested m_xpr;
+  const FFT m_fft;
+};
+
+// Eval as rvalue
+template <typename FFT, typename ArgType, typename Device, int FFTResultType, int FFTDir>
+struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, Device> {
+  typedef TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename std::complex<RealScalar> ComplexScalar;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions;
+  typedef internal::traits<XprType> XprTraits;
+  typedef typename XprTraits::Scalar InputScalar;
+  typedef typename internal::conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
+  typedef OutputScalar CoeffReturnType;
+  typedef typename PacketType<OutputScalar, Device>::type PacketReturnType;
+  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+    typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = true,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_fft(op.fft()), m_impl(op.expression(), device), m_data(NULL), m_device(device) {
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    for (int i = 0; i < NumDims; ++i) {
+      eigen_assert(input_dims[i] > 0);
+      m_dimensions[i] = input_dims[i];
+    }
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_strides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_strides[i] = m_strides[i - 1] * m_dimensions[i - 1];
+      }
+    } else {
+      m_strides[NumDims - 1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_strides[i] = m_strides[i + 1] * m_dimensions[i + 1];
+      }
+    }
+    m_size = m_dimensions.TotalSize();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
+    return m_dimensions;
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    if (data) {
+      evalToBuf(data);
+      return false;
+    } else {
+      m_data = (EvaluatorPointerType)m_device.get((CoeffReturnType*)(m_device.allocate_temp(sizeof(CoeffReturnType) * m_size)));
+      evalToBuf(m_data);
+      return true;
+    }
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    if (m_data) {
+      m_device.deallocate(m_data);
+      m_data = NULL;
+    }
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffReturnType coeff(Index index) const {
+    return m_data[index];
+  }
+
+  template <int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketReturnType
+  packet(Index index) const {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_data + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_data; }
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_data.bind(cgh);
+  }
+#endif
+
+ private:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalToBuf(EvaluatorPointerType data) {
+    const bool write_to_out = internal::is_same<OutputScalar, ComplexScalar>::value;
+    ComplexScalar* buf = write_to_out ? (ComplexScalar*)data : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * m_size);
+
+    for (Index i = 0; i < m_size; ++i) {
+      buf[i] = MakeComplex<internal::is_same<InputScalar, RealScalar>::value>()(m_impl.coeff(i));
+    }
+
+    for (size_t i = 0; i < m_fft.size(); ++i) {
+      Index dim = m_fft[i];
+      eigen_assert(dim >= 0 && dim < NumDims);
+      Index line_len = m_dimensions[dim];
+      eigen_assert(line_len >= 1);
+      ComplexScalar* line_buf = (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * line_len);
+      const bool is_power_of_two = isPowerOfTwo(line_len);
+      const Index good_composite = is_power_of_two ? 0 : findGoodComposite(line_len);
+      const Index log_len = is_power_of_two ? getLog2(line_len) : getLog2(good_composite);
+
+      ComplexScalar* a = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * good_composite);
+      ComplexScalar* b = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * good_composite);
+      ComplexScalar* pos_j_base_powered = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * (line_len + 1));
+      if (!is_power_of_two) {
+        // Compute twiddle factors
+        //   t_n = exp(sqrt(-1) * pi * n^2 / line_len)
+        // for n = 0, 1,..., line_len-1.
+        // For n > 2 we use the recurrence t_n = t_{n-1}^2 / t_{n-2} * t_1^2
+
+        // The recurrence is correct in exact arithmetic, but causes
+        // numerical issues for large transforms, especially in
+        // single-precision floating point.
+        //
+        // pos_j_base_powered[0] = ComplexScalar(1, 0);
+        // if (line_len > 1) {
+        //   const ComplexScalar pos_j_base = ComplexScalar(
+        //       numext::cos(M_PI / line_len), numext::sin(M_PI / line_len));
+        //   pos_j_base_powered[1] = pos_j_base;
+        //   if (line_len > 2) {
+        //     const ComplexScalar pos_j_base_sq = pos_j_base * pos_j_base;
+        //     for (int i = 2; i < line_len + 1; ++i) {
+        //       pos_j_base_powered[i] = pos_j_base_powered[i - 1] *
+        //           pos_j_base_powered[i - 1] /
+        //           pos_j_base_powered[i - 2] *
+        //           pos_j_base_sq;
+        //     }
+        //   }
+        // }
+        // TODO(rmlarsen): Find a way to use Eigen's vectorized sin
+        // and cosine functions here.
+        for (int j = 0; j < line_len + 1; ++j) {
+          double arg = ((EIGEN_PI * j) * j) / line_len;
+          std::complex<double> tmp(numext::cos(arg), numext::sin(arg));
+          pos_j_base_powered[j] = static_cast<ComplexScalar>(tmp);
+        }
+      }
+
+      for (Index partial_index = 0; partial_index < m_size / line_len; ++partial_index) {
+        const Index base_offset = getBaseOffsetFromIndex(partial_index, dim);
+
+        // get data into line_buf
+        const Index stride = m_strides[dim];
+        if (stride == 1) {
+          m_device.memcpy(line_buf, &buf[base_offset], line_len*sizeof(ComplexScalar));
+        } else {
+          Index offset = base_offset;
+          for (int j = 0; j < line_len; ++j, offset += stride) {
+            line_buf[j] = buf[offset];
+          }
+        }
+
+        // process the line
+        if (is_power_of_two) {
+          processDataLineCooleyTukey(line_buf, line_len, log_len);
+        }
+        else {
+          processDataLineBluestein(line_buf, line_len, good_composite, log_len, a, b, pos_j_base_powered);
+        }
+
+        // write back
+        if (FFTDir == FFT_FORWARD && stride == 1) {
+          m_device.memcpy(&buf[base_offset], line_buf, line_len*sizeof(ComplexScalar));
+        } else {
+          Index offset = base_offset;
+          const ComplexScalar div_factor =  ComplexScalar(1.0 / line_len, 0);
+          for (int j = 0; j < line_len; ++j, offset += stride) {
+             buf[offset] = (FFTDir == FFT_FORWARD) ? line_buf[j] : line_buf[j] * div_factor;
+          }
+        }
+      }
+      m_device.deallocate(line_buf);
+      if (!is_power_of_two) {
+        m_device.deallocate(a);
+        m_device.deallocate(b);
+        m_device.deallocate(pos_j_base_powered);
+      }
+    }
+
+    if(!write_to_out) {
+      for (Index i = 0; i < m_size; ++i) {
+        data[i] = PartOf<FFTResultType>()(buf[i]);
+      }
+      m_device.deallocate(buf);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static bool isPowerOfTwo(Index x) {
+    eigen_assert(x > 0);
+    return !(x & (x - 1));
+  }
+
+  // The composite number for padding, used in Bluestein's FFT algorithm
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static Index findGoodComposite(Index n) {
+    Index i = 2;
+    while (i < 2 * n - 1) i *= 2;
+    return i;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static Index getLog2(Index m) {
+    Index log2m = 0;
+    while (m >>= 1) log2m++;
+    return log2m;
+  }
+
+  // Call Cooley Tukey algorithm directly, data length must be power of 2
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void processDataLineCooleyTukey(ComplexScalar* line_buf, Index line_len, Index log_len) {
+    eigen_assert(isPowerOfTwo(line_len));
+    scramble_FFT(line_buf, line_len);
+    compute_1D_Butterfly<FFTDir>(line_buf, line_len, log_len);
+  }
+
+  // Call Bluestein's FFT algorithm, m is a good composite number greater than (2 * n - 1), used as the padding length
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void processDataLineBluestein(ComplexScalar* line_buf, Index line_len, Index good_composite, Index log_len, ComplexScalar* a, ComplexScalar* b, const ComplexScalar* pos_j_base_powered) {
+    Index n = line_len;
+    Index m = good_composite;
+    ComplexScalar* data = line_buf;
+
+    for (Index i = 0; i < n; ++i) {
+      if(FFTDir == FFT_FORWARD) {
+        a[i] = data[i] * numext::conj(pos_j_base_powered[i]);
+      }
+      else {
+        a[i] = data[i] * pos_j_base_powered[i];
+      }
+    }
+    for (Index i = n; i < m; ++i) {
+      a[i] = ComplexScalar(0, 0);
+    }
+
+    for (Index i = 0; i < n; ++i) {
+      if(FFTDir == FFT_FORWARD) {
+        b[i] = pos_j_base_powered[i];
+      }
+      else {
+        b[i] = numext::conj(pos_j_base_powered[i]);
+      }
+    }
+    for (Index i = n; i < m - n; ++i) {
+      b[i] = ComplexScalar(0, 0);
+    }
+    for (Index i = m - n; i < m; ++i) {
+      if(FFTDir == FFT_FORWARD) {
+        b[i] = pos_j_base_powered[m-i];
+      }
+      else {
+        b[i] = numext::conj(pos_j_base_powered[m-i]);
+      }
+    }
+
+    scramble_FFT(a, m);
+    compute_1D_Butterfly<FFT_FORWARD>(a, m, log_len);
+
+    scramble_FFT(b, m);
+    compute_1D_Butterfly<FFT_FORWARD>(b, m, log_len);
+
+    for (Index i = 0; i < m; ++i) {
+      a[i] *= b[i];
+    }
+
+    scramble_FFT(a, m);
+    compute_1D_Butterfly<FFT_REVERSE>(a, m, log_len);
+
+    //Do the scaling after ifft
+    for (Index i = 0; i < m; ++i) {
+      a[i] /= m;
+    }
+
+    for (Index i = 0; i < n; ++i) {
+      if(FFTDir == FFT_FORWARD) {
+        data[i] = a[i] * numext::conj(pos_j_base_powered[i]);
+      }
+      else {
+        data[i] = a[i] * pos_j_base_powered[i];
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static void scramble_FFT(ComplexScalar* data, Index n) {
+    eigen_assert(isPowerOfTwo(n));
+    Index j = 1;
+    for (Index i = 1; i < n; ++i){
+      if (j > i) {
+        std::swap(data[j-1], data[i-1]);
+      }
+      Index m = n >> 1;
+      while (m >= 2 && j > m) {
+        j -= m;
+        m >>= 1;
+      }
+      j += m;
+    }
+  }
+
+  template <int Dir>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_2(ComplexScalar* data) {
+    ComplexScalar tmp = data[1];
+    data[1] = data[0] - data[1];
+    data[0] += tmp;
+  }
+
+  template <int Dir>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_4(ComplexScalar* data) {
+    ComplexScalar tmp[4];
+    tmp[0] = data[0] + data[1];
+    tmp[1] = data[0] - data[1];
+    tmp[2] = data[2] + data[3];
+    if (Dir == FFT_FORWARD) {
+      tmp[3] = ComplexScalar(0.0, -1.0) * (data[2] - data[3]);
+    } else {
+      tmp[3] = ComplexScalar(0.0, 1.0) * (data[2] - data[3]);
+    }
+    data[0] = tmp[0] + tmp[2];
+    data[1] = tmp[1] + tmp[3];
+    data[2] = tmp[0] - tmp[2];
+    data[3] = tmp[1] - tmp[3];
+  }
+
+  template <int Dir>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_8(ComplexScalar* data) {
+    ComplexScalar tmp_1[8];
+    ComplexScalar tmp_2[8];
+
+    tmp_1[0] = data[0] + data[1];
+    tmp_1[1] = data[0] - data[1];
+    tmp_1[2] = data[2] + data[3];
+    if (Dir == FFT_FORWARD) {
+      tmp_1[3] = (data[2] - data[3]) * ComplexScalar(0, -1);
+    } else {
+      tmp_1[3] = (data[2] - data[3]) * ComplexScalar(0, 1);
+    }
+    tmp_1[4] = data[4] + data[5];
+    tmp_1[5] = data[4] - data[5];
+    tmp_1[6] = data[6] + data[7];
+    if (Dir == FFT_FORWARD) {
+      tmp_1[7] = (data[6] - data[7]) * ComplexScalar(0, -1);
+    } else {
+      tmp_1[7] = (data[6] - data[7]) * ComplexScalar(0, 1);
+    }
+    tmp_2[0] = tmp_1[0] + tmp_1[2];
+    tmp_2[1] = tmp_1[1] + tmp_1[3];
+    tmp_2[2] = tmp_1[0] - tmp_1[2];
+    tmp_2[3] = tmp_1[1] - tmp_1[3];
+    tmp_2[4] = tmp_1[4] + tmp_1[6];
+// SQRT2DIV2 = sqrt(2)/2
+#define SQRT2DIV2 0.7071067811865476
+    if (Dir == FFT_FORWARD) {
+      tmp_2[5] = (tmp_1[5] + tmp_1[7]) * ComplexScalar(SQRT2DIV2, -SQRT2DIV2);
+      tmp_2[6] = (tmp_1[4] - tmp_1[6]) * ComplexScalar(0, -1);
+      tmp_2[7] = (tmp_1[5] - tmp_1[7]) * ComplexScalar(-SQRT2DIV2, -SQRT2DIV2);
+    } else {
+      tmp_2[5] = (tmp_1[5] + tmp_1[7]) * ComplexScalar(SQRT2DIV2, SQRT2DIV2);
+      tmp_2[6] = (tmp_1[4] - tmp_1[6]) * ComplexScalar(0, 1);
+      tmp_2[7] = (tmp_1[5] - tmp_1[7]) * ComplexScalar(-SQRT2DIV2, SQRT2DIV2);
+    }
+    data[0] = tmp_2[0] + tmp_2[4];
+    data[1] = tmp_2[1] + tmp_2[5];
+    data[2] = tmp_2[2] + tmp_2[6];
+    data[3] = tmp_2[3] + tmp_2[7];
+    data[4] = tmp_2[0] - tmp_2[4];
+    data[5] = tmp_2[1] - tmp_2[5];
+    data[6] = tmp_2[2] - tmp_2[6];
+    data[7] = tmp_2[3] - tmp_2[7];
+  }
+
+  template <int Dir>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_1D_merge(
+      ComplexScalar* data, Index n, Index n_power_of_2) {
+    // Original code:
+    // RealScalar wtemp = std::sin(M_PI/n);
+    // RealScalar wpi =  -std::sin(2 * M_PI/n);
+    const RealScalar wtemp = m_sin_PI_div_n_LUT[n_power_of_2];
+    const RealScalar wpi = (Dir == FFT_FORWARD)
+                               ? m_minus_sin_2_PI_div_n_LUT[n_power_of_2]
+                               : -m_minus_sin_2_PI_div_n_LUT[n_power_of_2];
+
+    const ComplexScalar wp(wtemp, wpi);
+    const ComplexScalar wp_one = wp + ComplexScalar(1, 0);
+    const ComplexScalar wp_one_2 = wp_one * wp_one;
+    const ComplexScalar wp_one_3 = wp_one_2 * wp_one;
+    const ComplexScalar wp_one_4 = wp_one_3 * wp_one;
+    const Index n2 = n / 2;
+    ComplexScalar w(1.0, 0.0);
+    for (Index i = 0; i < n2; i += 4) {
+       ComplexScalar temp0(data[i + n2] * w);
+       ComplexScalar temp1(data[i + 1 + n2] * w * wp_one);
+       ComplexScalar temp2(data[i + 2 + n2] * w * wp_one_2);
+       ComplexScalar temp3(data[i + 3 + n2] * w * wp_one_3);
+       w = w * wp_one_4;
+
+       data[i + n2] = data[i] - temp0;
+       data[i] += temp0;
+
+       data[i + 1 + n2] = data[i + 1] - temp1;
+       data[i + 1] += temp1;
+
+       data[i + 2 + n2] = data[i + 2] - temp2;
+       data[i + 2] += temp2;
+
+       data[i + 3 + n2] = data[i + 3] - temp3;
+       data[i + 3] += temp3;
+    }
+  }
+
+ template <int Dir>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void compute_1D_Butterfly(
+      ComplexScalar* data, Index n, Index n_power_of_2) {
+    eigen_assert(isPowerOfTwo(n));
+    if (n > 8) {
+      compute_1D_Butterfly<Dir>(data, n / 2, n_power_of_2 - 1);
+      compute_1D_Butterfly<Dir>(data + n / 2, n / 2, n_power_of_2 - 1);
+      butterfly_1D_merge<Dir>(data, n, n_power_of_2);
+    } else if (n == 8) {
+      butterfly_8<Dir>(data);
+    } else if (n == 4) {
+      butterfly_4<Dir>(data);
+    } else if (n == 2) {
+      butterfly_2<Dir>(data);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index getBaseOffsetFromIndex(Index index, Index omitted_dim) const {
+    Index result = 0;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > omitted_dim; --i) {
+        const Index partial_m_stride = m_strides[i] / m_dimensions[omitted_dim];
+        const Index idx = index / partial_m_stride;
+        index -= idx * partial_m_stride;
+        result += idx * m_strides[i];
+      }
+      result += index;
+    }
+    else {
+      for (Index i = 0; i < omitted_dim; ++i) {
+        const Index partial_m_stride = m_strides[i] / m_dimensions[omitted_dim];
+        const Index idx = index / partial_m_stride;
+        index -= idx * partial_m_stride;
+        result += idx * m_strides[i];
+      }
+      result += index;
+    }
+    // Value of index_coords[omitted_dim] is not determined to this step
+    return result;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index getIndexFromOffset(Index base, Index omitted_dim, Index offset) const {
+    Index result = base + offset * m_strides[omitted_dim] ;
+    return result;
+  }
+
+ protected:
+  Index m_size;
+  const FFT EIGEN_DEVICE_REF m_fft;
+  Dimensions m_dimensions;
+  array<Index, NumDims> m_strides;
+  TensorEvaluator<ArgType, Device> m_impl;
+  EvaluatorPointerType m_data;
+  const Device EIGEN_DEVICE_REF m_device;
+
+  // This will support a maximum FFT size of 2^32 for each dimension
+  // m_sin_PI_div_n_LUT[i] = (-2) * std::sin(M_PI / std::pow(2,i)) ^ 2;
+  const RealScalar m_sin_PI_div_n_LUT[32] = {
+    RealScalar(0.0),
+    RealScalar(-2),
+    RealScalar(-0.999999999999999),
+    RealScalar(-0.292893218813453),
+    RealScalar(-0.0761204674887130),
+    RealScalar(-0.0192147195967696),
+    RealScalar(-0.00481527332780311),
+    RealScalar(-0.00120454379482761),
+    RealScalar(-3.01181303795779e-04),
+    RealScalar(-7.52981608554592e-05),
+    RealScalar(-1.88247173988574e-05),
+    RealScalar(-4.70619042382852e-06),
+    RealScalar(-1.17654829809007e-06),
+    RealScalar(-2.94137117780840e-07),
+    RealScalar(-7.35342821488550e-08),
+    RealScalar(-1.83835707061916e-08),
+    RealScalar(-4.59589268710903e-09),
+    RealScalar(-1.14897317243732e-09),
+    RealScalar(-2.87243293150586e-10),
+    RealScalar( -7.18108232902250e-11),
+    RealScalar(-1.79527058227174e-11),
+    RealScalar(-4.48817645568941e-12),
+    RealScalar(-1.12204411392298e-12),
+    RealScalar(-2.80511028480785e-13),
+    RealScalar(-7.01277571201985e-14),
+    RealScalar(-1.75319392800498e-14),
+    RealScalar(-4.38298482001247e-15),
+    RealScalar(-1.09574620500312e-15),
+    RealScalar(-2.73936551250781e-16),
+    RealScalar(-6.84841378126949e-17),
+    RealScalar(-1.71210344531737e-17),
+    RealScalar(-4.28025861329343e-18)
+  };
+
+  // m_minus_sin_2_PI_div_n_LUT[i] = -std::sin(2 * M_PI / std::pow(2,i));
+  const RealScalar m_minus_sin_2_PI_div_n_LUT[32] = {
+    RealScalar(0.0),
+    RealScalar(0.0),
+    RealScalar(-1.00000000000000e+00),
+    RealScalar(-7.07106781186547e-01),
+    RealScalar(-3.82683432365090e-01),
+    RealScalar(-1.95090322016128e-01),
+    RealScalar(-9.80171403295606e-02),
+    RealScalar(-4.90676743274180e-02),
+    RealScalar(-2.45412285229123e-02),
+    RealScalar(-1.22715382857199e-02),
+    RealScalar(-6.13588464915448e-03),
+    RealScalar(-3.06795676296598e-03),
+    RealScalar(-1.53398018628477e-03),
+    RealScalar(-7.66990318742704e-04),
+    RealScalar(-3.83495187571396e-04),
+    RealScalar(-1.91747597310703e-04),
+    RealScalar(-9.58737990959773e-05),
+    RealScalar(-4.79368996030669e-05),
+    RealScalar(-2.39684498084182e-05),
+    RealScalar(-1.19842249050697e-05),
+    RealScalar(-5.99211245264243e-06),
+    RealScalar(-2.99605622633466e-06),
+    RealScalar(-1.49802811316901e-06),
+    RealScalar(-7.49014056584716e-07),
+    RealScalar(-3.74507028292384e-07),
+    RealScalar(-1.87253514146195e-07),
+    RealScalar(-9.36267570730981e-08),
+    RealScalar(-4.68133785365491e-08),
+    RealScalar(-2.34066892682746e-08),
+    RealScalar(-1.17033446341373e-08),
+    RealScalar(-5.85167231706864e-09),
+    RealScalar(-2.92583615853432e-09)
+  };
+};
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_FFT_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorFixedSize.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorFixedSize.h
new file mode 100644
index 0000000..ca39bb8
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorFixedSize.h
@@ -0,0 +1,379 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_FIXED_SIZE_H
+#define EIGEN_CXX11_TENSOR_TENSOR_FIXED_SIZE_H
+
+namespace Eigen {
+
+/** \class TensorFixedSize
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief The fixed sized version of the tensor class.
+  *
+  * The fixed sized equivalent of
+  * Eigen::Tensor<float, 3> t(3, 5, 7);
+  * is
+  * Eigen::TensorFixedSize<float, Sizes<3,5,7>> t;
+  */
+
+template<typename Scalar_, typename Dimensions_, int Options_, typename IndexType>
+class TensorFixedSize : public TensorBase<TensorFixedSize<Scalar_, Dimensions_, Options_, IndexType> >
+{
+  public:
+    typedef TensorFixedSize<Scalar_, Dimensions_, Options_, IndexType> Self;
+    typedef TensorBase<TensorFixedSize<Scalar_, Dimensions_, Options_, IndexType> > Base;
+    typedef typename Eigen::internal::nested<Self>::type Nested;
+    typedef typename internal::traits<Self>::StorageKind StorageKind;
+    typedef typename internal::traits<Self>::Index Index;
+    typedef Scalar_ Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef typename Base::CoeffReturnType CoeffReturnType;
+
+    static const int Options = Options_;
+
+    enum {
+      IsAligned = bool(EIGEN_MAX_ALIGN_BYTES>0),
+      PacketAccess = (internal::packet_traits<Scalar>::size > 1),
+      BlockAccess = false,
+      PreferBlockAccess = false,
+      Layout = Options_ & RowMajor ? RowMajor : ColMajor,
+      CoordAccess = true,
+      RawAccess = true
+    };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  typedef Dimensions_ Dimensions;
+  static const std::size_t NumIndices = Dimensions::count;
+
+  protected:
+  TensorStorage<Scalar, Dimensions, Options> m_storage;
+
+  public:
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index                    rank()                   const { return NumIndices; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index                    dimension(std::size_t n) const { return m_storage.dimensions()[n]; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions&        dimensions()             const { return m_storage.dimensions(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index                    size()                   const { return m_storage.size(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar                   *data()                        { return m_storage.data(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar             *data()                  const { return m_storage.data(); }
+
+    // This makes EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    // work, because that uses base().coeffRef() - and we don't yet
+    // implement a similar class hierarchy
+    inline Self& base()             { return *this; }
+    inline const Self& base() const { return *this; }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(Index firstIndex, IndexTypes... otherIndices) const
+    {
+      // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return coeff(array<Index, NumIndices>{{firstIndex, otherIndices...}});
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& coeff(const array<Index, NumIndices>& indices) const
+    {
+      eigen_internal_assert(checkIndexRange(indices));
+      return m_storage.data()[linearizedIndex(indices)];
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const
+    {
+      eigen_internal_assert(index >= 0 && index < size());
+      return m_storage.data()[index];
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& coeff() const
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return m_storage.data()[0];
+    }
+
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index firstIndex, IndexTypes... otherIndices)
+    {
+      // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return coeffRef(array<Index, NumIndices>{{firstIndex, otherIndices...}});
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(const array<Index, NumIndices>& indices)
+    {
+      eigen_internal_assert(checkIndexRange(indices));
+      return m_storage.data()[linearizedIndex(indices)];
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
+    {
+      eigen_internal_assert(index >= 0 && index < size());
+      return m_storage.data()[index];
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef()
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return m_storage.data()[0];
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()(Index firstIndex, IndexTypes... otherIndices) const
+    {
+      // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return this->operator()(array<Index, NumIndices>{{firstIndex, otherIndices...}});
+    }
+#else
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1) const
+    {
+      if (Options&RowMajor) {
+        const Index index = i1 + i0 * m_storage.dimensions()[1];
+        return m_storage.data()[index];
+      } else {
+        const Index index = i0 + i1 * m_storage.dimensions()[0];
+        return m_storage.data()[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2) const
+    {
+      if (Options&RowMajor) {
+         const Index index = i2 + m_storage.dimensions()[2] * (i1 + m_storage.dimensions()[1] * i0);
+         return m_storage.data()[index];
+      } else {
+         const Index index = i0 + m_storage.dimensions()[0] * (i1 + m_storage.dimensions()[1] * i2);
+        return m_storage.data()[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2, Index i3) const
+    {
+      if (Options&RowMajor) {
+        const Index index = i3 + m_storage.dimensions()[3] * (i2 + m_storage.dimensions()[2] * (i1 + m_storage.dimensions()[1] * i0));
+        return m_storage.data()[index];
+      } else {
+        const Index index = i0 + m_storage.dimensions()[0] * (i1 + m_storage.dimensions()[1] * (i2 + m_storage.dimensions()[2] * i3));
+        return m_storage.data()[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2, Index i3, Index i4) const
+    {
+      if (Options&RowMajor) {
+        const Index index = i4 + m_storage.dimensions()[4] * (i3 + m_storage.dimensions()[3] * (i2 + m_storage.dimensions()[2] * (i1 + m_storage.dimensions()[1] * i0)));
+        return m_storage.data()[index];
+      } else {
+        const Index index = i0 + m_storage.dimensions()[0] * (i1 + m_storage.dimensions()[1] * (i2 + m_storage.dimensions()[2] * (i3 + m_storage.dimensions()[3] * i4)));
+        return m_storage.data()[index];
+      }
+    }
+#endif
+
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(const array<Index, NumIndices>& indices) const
+    {
+      eigen_assert(checkIndexRange(indices));
+      return coeff(indices);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()(Index index) const
+    {
+      eigen_internal_assert(index >= 0 && index < size());
+      return coeff(index);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator()() const
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return coeff();
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& operator[](Index index) const
+    {
+      // The bracket operator is only for vectors, use the parenthesis operator instead.
+      EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return coeff(index);
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()(Index firstIndex, IndexTypes... otherIndices)
+    {
+      // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return operator()(array<Index, NumIndices>{{firstIndex, otherIndices...}});
+    }
+#else
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1)
+    {
+       if (Options&RowMajor) {
+         const Index index = i1 + i0 * m_storage.dimensions()[1];
+        return m_storage.data()[index];
+      } else {
+        const Index index = i0 + i1 * m_storage.dimensions()[0];
+        return m_storage.data()[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2)
+    {
+       if (Options&RowMajor) {
+         const Index index = i2 + m_storage.dimensions()[2] * (i1 + m_storage.dimensions()[1] * i0);
+        return m_storage.data()[index];
+      } else {
+         const Index index = i0 + m_storage.dimensions()[0] * (i1 + m_storage.dimensions()[1] * i2);
+        return m_storage.data()[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2, Index i3)
+    {
+      if (Options&RowMajor) {
+        const Index index = i3 + m_storage.dimensions()[3] * (i2 + m_storage.dimensions()[2] * (i1 + m_storage.dimensions()[1] * i0));
+        return m_storage.data()[index];
+      } else {
+        const Index index = i0 + m_storage.dimensions()[0] * (i1 + m_storage.dimensions()[1] * (i2 + m_storage.dimensions()[2] * i3));
+        return m_storage.data()[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2, Index i3, Index i4)
+    {
+      if (Options&RowMajor) {
+        const Index index = i4 + m_storage.dimensions()[4] * (i3 + m_storage.dimensions()[3] * (i2 + m_storage.dimensions()[2] * (i1 + m_storage.dimensions()[1] * i0)));
+        return m_storage.data()[index];
+      } else {
+        const Index index = i0 + m_storage.dimensions()[0] * (i1 + m_storage.dimensions()[1] * (i2 + m_storage.dimensions()[2] * (i3 + m_storage.dimensions()[3] * i4)));
+        return m_storage.data()[index];
+      }
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(const array<Index, NumIndices>& indices)
+    {
+      eigen_assert(checkIndexRange(indices));
+      return coeffRef(indices);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index index)
+    {
+      eigen_assert(index >= 0 && index < size());
+      return coeffRef(index);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()()
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return coeffRef();
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator[](Index index)
+    {
+      // The bracket operator is only for vectors, use the parenthesis operator instead
+      EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return coeffRef(index);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorFixedSize()
+      : m_storage()
+    {
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorFixedSize(const Self& other)
+      : m_storage(other.m_storage)
+    {
+    }
+
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorFixedSize(Self&& other)
+      : m_storage(other.m_storage)
+    {
+    }
+#endif
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorFixedSize(const TensorBase<OtherDerived, ReadOnlyAccessors>& other)
+    {
+      typedef TensorAssignOp<TensorFixedSize, const OtherDerived> Assign;
+      Assign assign(*this, other.derived());
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
+    }
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorFixedSize(const TensorBase<OtherDerived, WriteAccessors>& other)
+    {
+      typedef TensorAssignOp<TensorFixedSize, const OtherDerived> Assign;
+      Assign assign(*this, other.derived());
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
+    }
+
+    // FIXME: check that the dimensions of other match the dimensions of *this.
+    // Unfortunately this isn't possible yet when the rhs is an expression.
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(TensorFixedSize)
+
+
+  protected:
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE bool checkIndexRange(const array<Index, NumIndices>& /*indices*/) const
+    {
+      using internal::array_apply_and_reduce;
+      using internal::array_zip_and_reduce;
+      using internal::greater_equal_zero_op;
+      using internal::logical_and_op;
+      using internal::lesser_op;
+
+      return true;
+        // check whether the indices are all >= 0
+          /*       array_apply_and_reduce<logical_and_op, greater_equal_zero_op>(indices) &&
+        // check whether the indices fit in the dimensions
+        array_zip_and_reduce<logical_and_op, lesser_op>(indices, m_storage.dimensions());*/
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index linearizedIndex(const array<Index, NumIndices>& indices) const
+    {
+      if (Options&RowMajor) {
+        return m_storage.dimensions().IndexOfRowMajor(indices);
+      } else {
+        return m_storage.dimensions().IndexOfColMajor(indices);
+      }
+    }
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_FIXED_SIZE_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorForcedEval.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorForcedEval.h
new file mode 100644
index 0000000..e800ded
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorForcedEval.h
@@ -0,0 +1,237 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H
+#define EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H
+
+namespace Eigen {
+
+/** \class TensorForcedEval
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor reshaping class.
+  *
+  *
+  */
+namespace internal {
+template<typename XprType>
+struct traits<TensorForcedEvalOp<XprType> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs are different.
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename traits<XprType>::StorageKind StorageKind;
+  typedef typename traits<XprType>::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+
+  enum {
+    Flags = 0
+  };
+};
+
+template<typename XprType>
+struct eval<TensorForcedEvalOp<XprType>, Eigen::Dense>
+{
+  typedef const TensorForcedEvalOp<XprType>& type;
+};
+
+template<typename XprType>
+struct nested<TensorForcedEvalOp<XprType>, 1, typename eval<TensorForcedEvalOp<XprType> >::type>
+{
+  typedef TensorForcedEvalOp<XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename XprType>
+class TensorForcedEvalOp : public TensorBase<TensorForcedEvalOp<XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorForcedEvalOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorForcedEvalOp(const XprType& expr)
+      : m_xpr(expr) {}
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+};
+
+namespace internal {
+template <typename Device, typename CoeffReturnType>
+struct non_integral_type_placement_new{
+  template <typename StorageType>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index numValues, StorageType m_buffer) {
+   // Initialize non-trivially constructible types.
+    if (!internal::is_arithmetic<CoeffReturnType>::value) {
+      for (Index i = 0; i < numValues; ++i) new (m_buffer + i) CoeffReturnType();
+    }
+}
+};
+
+// SYCL does not support non-integral types 
+// having new (m_buffer + i) CoeffReturnType() causes the following compiler error for SYCL Devices 
+// no matching function for call to 'operator new'
+template <typename CoeffReturnType>
+struct non_integral_type_placement_new<Eigen::SyclDevice, CoeffReturnType> {
+  template <typename StorageType>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index, StorageType) {
+}
+};
+} // end namespace internal
+
+template<typename ArgType_, typename Device>
+struct TensorEvaluator<const TensorForcedEvalOp<ArgType_>, Device>
+{
+  typedef const typename internal::remove_all<ArgType_>::type ArgType;
+  typedef TensorForcedEvalOp<ArgType> XprType;
+  typedef typename ArgType::Scalar Scalar;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename Eigen::internal::traits<XprType>::PointerType TensorPointerType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned         = true,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = internal::is_arithmetic<CoeffReturnType>::value,
+    PreferBlockAccess = false,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess         = true
+  };
+
+  static const int NumDims = internal::traits<ArgType>::NumDimensions;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename internal::TensorMaterializedBlock<CoeffReturnType, NumDims,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_op(op.expression()),
+      m_device(device), m_buffer(NULL)
+  { }
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    const Index numValues =  internal::array_prod(m_impl.dimensions());
+    m_buffer = m_device.get((CoeffReturnType*)m_device.allocate_temp(numValues * sizeof(CoeffReturnType)));
+
+   internal::non_integral_type_placement_new<Device, CoeffReturnType>()(numValues, m_buffer);
+
+    typedef TensorEvalToOp< const typename internal::remove_const<ArgType>::type > EvalTo;
+    EvalTo evalToTmp(m_device.get(m_buffer), m_op);
+
+    internal::TensorExecutor<
+        const EvalTo, typename internal::remove_const<Device>::type,
+        /*Vectorizable=*/internal::IsVectorizable<Device, const ArgType>::value,
+        /*Tiling=*/internal::IsTileable<Device, const ArgType>::value>::
+        run(evalToTmp, m_device);
+
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    const Index numValues = internal::array_prod(m_impl.dimensions());
+    m_buffer = m_device.get((CoeffReturnType*)m_device.allocate_temp(
+        numValues * sizeof(CoeffReturnType)));
+    typedef TensorEvalToOp<const typename internal::remove_const<ArgType>::type>
+        EvalTo;
+    EvalTo evalToTmp(m_device.get(m_buffer), m_op);
+
+    auto on_done = std::bind([](EvalSubExprsCallback done_) { done_(true); },
+                             std::move(done));
+    internal::TensorAsyncExecutor<
+        const EvalTo, typename internal::remove_const<Device>::type,
+        decltype(on_done),
+        /*Vectorizable=*/internal::IsVectorizable<Device, const ArgType>::value,
+        /*Tiling=*/internal::IsTileable<Device, const ArgType>::value>::
+        runAsync(evalToTmp, m_device, std::move(on_done));
+  }
+#endif
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_device.deallocate_temp(m_buffer);
+    m_buffer = NULL;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return m_buffer[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_buffer + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    return internal::TensorBlockResourceRequirements::any();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    assert(m_buffer != NULL);
+    return TensorBlock::materialize(m_buffer, m_impl.dimensions(), desc, scratch);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  EvaluatorPointerType data() const { return m_buffer; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_buffer.bind(cgh);
+    m_impl.bind(cgh);
+  }
+#endif
+ private:
+  TensorEvaluator<ArgType, Device> m_impl;
+  const ArgType m_op;
+  const Device EIGEN_DEVICE_REF m_device;
+  EvaluatorPointerType m_buffer;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorForwardDeclarations.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorForwardDeclarations.h
new file mode 100644
index 0000000..246ebe4
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -0,0 +1,191 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_FORWARD_DECLARATIONS_H
+#define EIGEN_CXX11_TENSOR_TENSOR_FORWARD_DECLARATIONS_H
+
+namespace Eigen {
+
+// MakePointer class is used as a container of the address space of the pointer
+// on the host and on the device. From the host side it generates the T* pointer
+// and when EIGEN_USE_SYCL is used it construct a buffer with a map_allocator to
+// T* m_data on the host. It is always called on the device.
+// Specialisation of MakePointer class for creating the sycl buffer with
+// map_allocator.
+template<typename T> struct MakePointer {
+  typedef T* Type;
+  typedef const T* ConstType;
+};
+
+template <typename T>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T* constCast(const T* data) {
+  return const_cast<T*>(data);
+}
+
+// The StorageMemory class is a container of the device specific pointer
+// used for refering to a Pointer on TensorEvaluator class. While the TensorExpression
+// is a device-agnostic type and need MakePointer class for type conversion,
+// the TensorEvaluator class can be specialized for a device, hence it is possible
+// to construct different types of temproray storage memory in TensorEvaluator
+// for different devices by specializing the following StorageMemory class.
+template<typename T, typename device> struct StorageMemory: MakePointer <T> {};
+
+namespace internal{
+template<typename A, typename B> struct Pointer_type_promotion {
+  static const bool val=false;
+};
+template<typename A> struct Pointer_type_promotion<A, A> {
+  static const bool val = true;
+};
+template<typename A, typename B> struct TypeConversion {
+  typedef A* type;
+};
+}
+
+
+template<typename PlainObjectType, int Options_ = Unaligned, template <class> class MakePointer_ = MakePointer> class TensorMap;
+template<typename Scalar_, int NumIndices_, int Options_ = 0, typename IndexType = DenseIndex> class Tensor;
+template<typename Scalar_, typename Dimensions, int Options_ = 0, typename IndexType = DenseIndex> class TensorFixedSize;
+template<typename PlainObjectType> class TensorRef;
+template<typename Derived, int AccessLevel> class TensorBase;
+
+template<typename NullaryOp, typename PlainObjectType> class TensorCwiseNullaryOp;
+template<typename UnaryOp, typename XprType> class TensorCwiseUnaryOp;
+template<typename BinaryOp, typename LeftXprType, typename RightXprType> class TensorCwiseBinaryOp;
+template<typename TernaryOp, typename Arg1XprType, typename Arg2XprType, typename Arg3XprType> class TensorCwiseTernaryOp;
+template<typename IfXprType, typename ThenXprType, typename ElseXprType> class TensorSelectOp;
+template<typename Op, typename Dims, typename XprType, template <class> class MakePointer_ = MakePointer > class TensorReductionOp;
+template<typename XprType> class TensorIndexTupleOp;
+template<typename ReduceOp, typename Dims, typename XprType> class TensorTupleReducerOp;
+template<typename Axis, typename LeftXprType, typename RightXprType> class TensorConcatenationOp;
+template<typename Dimensions, typename LeftXprType, typename RightXprType, typename OutputKernelType> class TensorContractionOp;
+template<typename TargetType, typename XprType> class TensorConversionOp;
+template<typename Dimensions, typename InputXprType, typename KernelXprType> class TensorConvolutionOp;
+template<typename FFT, typename XprType, int FFTDataType, int FFTDirection> class TensorFFTOp;
+template<typename PatchDim, typename XprType> class TensorPatchOp;
+template<DenseIndex Rows, DenseIndex Cols, typename XprType> class TensorImagePatchOp;
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType> class TensorVolumePatchOp;
+template<typename Broadcast, typename XprType> class TensorBroadcastingOp;
+template<DenseIndex DimId, typename XprType> class TensorChippingOp;
+template<typename NewDimensions, typename XprType> class TensorReshapingOp;
+template<typename XprType> class TensorLayoutSwapOp;
+template<typename StartIndices, typename Sizes, typename XprType> class TensorSlicingOp;
+template<typename ReverseDimensions, typename XprType> class TensorReverseOp;
+template<typename PaddingDimensions, typename XprType> class TensorPaddingOp;
+template<typename Shuffle, typename XprType> class TensorShufflingOp;
+template<typename Strides, typename XprType> class TensorStridingOp;
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType> class TensorStridingSlicingOp;
+template<typename Strides, typename XprType> class TensorInflationOp;
+template<typename Generator, typename XprType> class TensorGeneratorOp;
+template<typename LeftXprType, typename RightXprType> class TensorAssignOp;
+template<typename Op, typename XprType> class TensorScanOp;
+template<typename Dims, typename XprType> class TensorTraceOp;
+
+template<typename CustomUnaryFunc, typename XprType> class TensorCustomUnaryOp;
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType> class TensorCustomBinaryOp;
+
+template<typename XprType, template <class> class MakePointer_ = MakePointer> class TensorEvalToOp;
+template<typename XprType> class TensorForcedEvalOp;
+
+template<typename ExpressionType, typename DeviceType> class TensorDevice;
+template<typename ExpressionType, typename DeviceType, typename DoneCallback> class TensorAsyncDevice;
+template<typename Derived, typename Device> struct TensorEvaluator;
+
+struct NoOpOutputKernel;
+
+struct DefaultDevice;
+struct ThreadPoolDevice;
+struct GpuDevice;
+struct SyclDevice;
+
+#ifdef EIGEN_USE_SYCL
+
+template <typename T> struct MakeSYCLPointer {
+  typedef Eigen::TensorSycl::internal::RangeAccess<cl::sycl::access::mode::read_write, T> Type;
+};
+
+template <typename T>
+EIGEN_STRONG_INLINE const Eigen::TensorSycl::internal::RangeAccess<cl::sycl::access::mode::read_write, T>&
+constCast(const Eigen::TensorSycl::internal::RangeAccess<cl::sycl::access::mode::read_write, T>& data) {
+  return data;
+}
+
+template <typename T>
+struct StorageMemory<T, SyclDevice> : MakeSYCLPointer<T> {};
+template <typename T>
+struct StorageMemory<T, const SyclDevice> : StorageMemory<T, SyclDevice> {};
+
+namespace TensorSycl {
+namespace internal{
+template <typename Evaluator, typename Op> class GenericNondeterministicReducer;
+}
+}
+#endif
+
+
+enum FFTResultType {
+  RealPart = 0,
+  ImagPart = 1,
+  BothParts = 2
+};
+
+enum FFTDirection {
+    FFT_FORWARD = 0,
+    FFT_REVERSE = 1
+};
+
+
+namespace internal {
+
+template <typename Device, typename Expression>
+struct IsVectorizable {
+  static const bool value = TensorEvaluator<Expression, Device>::PacketAccess;
+};
+
+template <typename Expression>
+struct IsVectorizable<GpuDevice, Expression> {
+  static const bool value = TensorEvaluator<Expression, GpuDevice>::PacketAccess &&
+                            TensorEvaluator<Expression, GpuDevice>::IsAligned;
+};
+
+// Tiled evaluation strategy.
+enum TiledEvaluation {
+  Off = 0,    // tiled evaluation is not supported
+  On = 1,     // still work in progress (see TensorBlock.h)
+};
+
+template <typename Device, typename Expression>
+struct IsTileable {
+  // Check that block evaluation is supported and it's a preferred option (at
+  // least one sub-expression has much faster block evaluation, e.g.
+  // broadcasting).
+  static const bool BlockAccess =
+      TensorEvaluator<Expression, Device>::BlockAccess &&
+      TensorEvaluator<Expression, Device>::PreferBlockAccess;
+
+  static const TiledEvaluation value =
+      BlockAccess ? TiledEvaluation::On : TiledEvaluation::Off;
+};
+
+template <typename Expression, typename Device,
+          bool Vectorizable      = IsVectorizable<Device, Expression>::value,
+          TiledEvaluation Tiling = IsTileable<Device, Expression>::value>
+class TensorExecutor;
+
+template <typename Expression, typename Device, typename DoneCallback,
+          bool Vectorizable = IsVectorizable<Device, Expression>::value,
+          TiledEvaluation Tiling = IsTileable<Device, Expression>::value>
+class TensorAsyncExecutor;
+
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_FORWARD_DECLARATIONS_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorFunctors.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorFunctors.h
new file mode 100644
index 0000000..d963032
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorFunctors.h
@@ -0,0 +1,488 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_FUNCTORS_H
+#define EIGEN_CXX11_TENSOR_TENSOR_FUNCTORS_H
+
+namespace Eigen {
+namespace internal {
+
+
+/** \internal
+ * \brief Template functor to compute the modulo between an array and a scalar.
+ */
+template <typename Scalar>
+struct scalar_mod_op {
+  EIGEN_DEVICE_FUNC scalar_mod_op(const Scalar& divisor) : m_divisor(divisor) {}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a % m_divisor; }
+  const Scalar m_divisor;
+};
+template <typename Scalar>
+struct functor_traits<scalar_mod_op<Scalar> >
+{ enum { Cost = scalar_div_cost<Scalar,false>::value, PacketAccess = false }; };
+
+
+/** \internal
+ * \brief Template functor to compute the modulo between 2 arrays.
+ */
+template <typename Scalar>
+struct scalar_mod2_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_mod2_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a, const Scalar& b) const { return a % b; }
+};
+template <typename Scalar>
+struct functor_traits<scalar_mod2_op<Scalar> >
+{ enum { Cost = scalar_div_cost<Scalar,false>::value, PacketAccess = false }; };
+
+template <typename Scalar>
+struct scalar_fmod_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_fmod_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar
+  operator()(const Scalar& a, const Scalar& b) const {
+    return numext::fmod(a, b);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_fmod_op<Scalar> > {
+  enum { Cost = 13,  // Reciprocal throughput of FPREM on Haswell.
+         PacketAccess = false };
+};
+
+template<typename Reducer, typename Device>
+struct reducer_traits {
+  enum {
+    Cost = 1,
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
+  };
+};
+
+// Standard reduction functors
+template <typename T> struct SumReducer
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
+    internal::scalar_sum_op<T> sum_op;
+    *accum = sum_op(*accum, t);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) const {
+    (*accum) = padd<Packet>(*accum, p);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const {
+    internal::scalar_cast_op<int, T> conv;
+    return conv(0);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet initializePacket() const {
+    return pset1<Packet>(initialize());
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T accum) const {
+    return accum;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const {
+    return vaccum;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
+    internal::scalar_sum_op<T> sum_op;
+    return sum_op(saccum, predux(vaccum));
+  }
+};
+
+template <typename T, typename Device>
+struct reducer_traits<SumReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = PacketType<T, Device>::HasAdd,
+    IsStateful = false,
+    IsExactlyAssociative = NumTraits<T>::IsInteger
+  };
+};
+
+template <typename T> struct MeanReducer
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  MeanReducer() : scalarCount_(0), packetCount_(0) { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) {
+    internal::scalar_sum_op<T> sum_op;
+    *accum = sum_op(*accum, t);
+    scalarCount_++;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) {
+    (*accum) = padd<Packet>(*accum, p);
+    packetCount_++;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const {
+    internal::scalar_cast_op<int, T> conv;
+    return conv(0);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet initializePacket() const {
+    return pset1<Packet>(initialize());
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T accum) const {
+    internal::scalar_quotient_op<T> quotient_op;
+    return quotient_op(accum, T(scalarCount_));
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const {
+    return pdiv(vaccum, pset1<Packet>(T(packetCount_)));
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
+    internal::scalar_sum_op<T> sum_op;
+    internal::scalar_quotient_op<T> quotient_op;
+    return quotient_op(
+        sum_op(saccum, predux(vaccum)),
+        T(scalarCount_ + packetCount_ * unpacket_traits<Packet>::size));
+  }
+
+  protected:
+    DenseIndex scalarCount_;
+    DenseIndex packetCount_;
+};
+
+template <typename T, typename Device>
+struct reducer_traits<MeanReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = PacketType<T, Device>::HasAdd &&
+                   PacketType<T, Device>::HasDiv && !NumTraits<T>::IsInteger,
+    IsStateful = true,
+    IsExactlyAssociative = NumTraits<T>::IsInteger
+  };
+};
+
+
+template <typename T, bool IsMax = true, bool IsInteger = true>
+struct MinMaxBottomValue {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T bottom_value() {
+    return Eigen::NumTraits<T>::lowest();
+  }
+};
+template <typename T>
+struct MinMaxBottomValue<T, true, false> {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T bottom_value() {
+    return -Eigen::NumTraits<T>::infinity();
+  }
+};
+template <typename T>
+struct MinMaxBottomValue<T, false, true> {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T bottom_value() {
+    return Eigen::NumTraits<T>::highest();
+  }
+};
+template <typename T>
+struct MinMaxBottomValue<T, false, false> {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T bottom_value() {
+    return Eigen::NumTraits<T>::infinity();
+  }
+};
+
+
+template <typename T, int NaNPropagation=PropagateFast> struct MaxReducer
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
+    scalar_max_op<T, T, NaNPropagation> op;
+    *accum = op(t, *accum);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) const {
+    scalar_max_op<T, T, NaNPropagation> op;
+    (*accum) = op.packetOp(*accum, p);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const {
+    return MinMaxBottomValue<T, /*IsMax=*/true, Eigen::NumTraits<T>::IsInteger>::bottom_value();
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet initializePacket() const {
+    return pset1<Packet>(initialize());
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T accum) const {
+    return accum;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const {
+    return vaccum;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
+    scalar_max_op<T, T, NaNPropagation> op;
+    return op(saccum, op.predux(vaccum));
+  }
+};
+
+template <typename T, typename Device, int NaNPropagation>
+    struct reducer_traits<MaxReducer<T, NaNPropagation>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = PacketType<T, Device>::HasMax,
+    IsStateful = false,
+    IsExactlyAssociative = (NaNPropagation!=PropagateFast)
+  };
+};
+
+template <typename T, int NaNPropagation=PropagateFast> struct MinReducer
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
+    scalar_min_op<T, T, NaNPropagation> op;
+    *accum = op(t, *accum);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) const {
+    scalar_min_op<T, T, NaNPropagation> op;
+    (*accum) = op.packetOp(*accum, p);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const {
+    return MinMaxBottomValue<T, /*IsMax=*/false, Eigen::NumTraits<T>::IsInteger>::bottom_value();
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet initializePacket() const {
+    return pset1<Packet>(initialize());
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T accum) const {
+    return accum;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const {
+    return vaccum;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
+    scalar_min_op<T, T, NaNPropagation> op;
+    return op(saccum, op.predux(vaccum));
+  }
+};
+
+template <typename T, typename Device, int NaNPropagation>
+    struct reducer_traits<MinReducer<T, NaNPropagation>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = PacketType<T, Device>::HasMin,
+    IsStateful = false,
+    IsExactlyAssociative = (NaNPropagation!=PropagateFast)
+  };
+};
+
+template <typename T> struct ProdReducer
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
+    internal::scalar_product_op<T> prod_op;
+    (*accum) = prod_op(*accum, t);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) const {
+    (*accum) = pmul<Packet>(*accum, p);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const {
+    internal::scalar_cast_op<int, T> conv;
+    return conv(1);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet initializePacket() const {
+    return pset1<Packet>(initialize());
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T accum) const {
+    return accum;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const {
+    return vaccum;
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
+    internal::scalar_product_op<T> prod_op;
+    return prod_op(saccum, predux_mul(vaccum));
+  }
+};
+
+template <typename T, typename Device>
+struct reducer_traits<ProdReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::MulCost,
+    PacketAccess = PacketType<T, Device>::HasMul,
+    IsStateful = false,
+    IsExactlyAssociative = true
+  };
+};
+
+
+struct AndReducer
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(bool t, bool* accum) const {
+    *accum = *accum && t;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool initialize() const {
+    return true;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool finalize(bool accum) const {
+    return accum;
+  }
+};
+
+template <typename Device>
+struct reducer_traits<AndReducer, Device> {
+  enum {
+    Cost = 1,
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
+  };
+};
+
+
+struct OrReducer {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(bool t, bool* accum) const {
+    *accum = *accum || t;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool initialize() const {
+    return false;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool finalize(bool accum) const {
+    return accum;
+  }
+};
+
+template <typename Device>
+struct reducer_traits<OrReducer, Device> {
+  enum {
+    Cost = 1,
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
+  };
+};
+
+// Argmin/Argmax reducers.  Returns the first occurrence if multiple locations
+// contain the same min/max value.
+template <typename T> struct ArgMaxTupleReducer
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
+    if (t.second < accum->second) {
+      return;
+    } else if (t.second > accum->second || accum->first > t.first ) {
+      *accum = t;
+    }
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const {
+    return T(0, NumTraits<typename T::second_type>::lowest());
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T& accum) const {
+    return accum;
+  }
+};
+
+template <typename T, typename Device>
+struct reducer_traits<ArgMaxTupleReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
+  };
+};
+
+
+template <typename T> struct ArgMinTupleReducer
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T& t, T* accum) const {
+    if (t.second > accum->second) {
+      return;
+    } else if (t.second < accum->second || accum->first > t.first) {
+      *accum = t;
+    }
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const {
+    return T(0, NumTraits<typename T::second_type>::highest());
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T& accum) const {
+    return accum;
+  }
+};
+
+template <typename T, typename Device>
+struct reducer_traits<ArgMinTupleReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = false,
+    IsStateful = false,
+    IsExactlyAssociative = true
+  };
+};
+
+
+template <typename T, typename Index, size_t NumDims>
+class GaussianGenerator {
+ public:
+  static const bool PacketAccess = false;
+
+  EIGEN_DEVICE_FUNC GaussianGenerator(const array<T, NumDims>& means,
+                                      const array<T, NumDims>& std_devs)
+      : m_means(means)
+  {
+    EIGEN_UNROLL_LOOP
+    for (size_t i = 0; i < NumDims; ++i) {
+      m_two_sigmas[i] = std_devs[i] * std_devs[i] * 2;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC T operator()(const array<Index, NumDims>& coordinates) const {
+    T tmp = T(0);
+    EIGEN_UNROLL_LOOP
+    for (size_t i = 0; i < NumDims; ++i) {
+      T offset = coordinates[i] - m_means[i];
+      tmp += offset * offset / m_two_sigmas[i];
+    }
+    return numext::exp(-tmp);
+  }
+
+ private:
+  array<T, NumDims> m_means;
+  array<T, NumDims> m_two_sigmas;
+};
+
+template <typename T, typename Index, size_t NumDims>
+struct functor_traits<GaussianGenerator<T, Index, NumDims> > {
+  enum {
+    Cost = NumDims * (2 * NumTraits<T>::AddCost + NumTraits<T>::MulCost +
+                      functor_traits<scalar_quotient_op<T, T> >::Cost) +
+           functor_traits<scalar_exp_op<T> >::Cost,
+    PacketAccess = GaussianGenerator<T, Index, NumDims>::PacketAccess
+  };
+};
+
+template <typename Scalar>
+struct scalar_clamp_op {
+  EIGEN_DEVICE_FUNC inline scalar_clamp_op(const Scalar& _min, const Scalar& _max) : m_min(_min), m_max(_max) {}
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar
+  operator()(const Scalar& x) const {
+    return numext::mini(numext::maxi(x, m_min), m_max);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet
+  packetOp(const Packet& x) const {
+    return internal::pmin(internal::pmax(x, pset1<Packet>(m_min)), pset1<Packet>(m_max));
+  }
+  const Scalar m_min;
+  const Scalar m_max;
+};
+template<typename Scalar>
+struct functor_traits<scalar_clamp_op<Scalar> >
+{ enum { Cost = 2 * NumTraits<Scalar>::AddCost, PacketAccess = (packet_traits<Scalar>::HasMin && packet_traits<Scalar>::HasMax)}; };
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_FUNCTORS_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorGenerator.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorGenerator.h
new file mode 100644
index 0000000..174bf06
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorGenerator.h
@@ -0,0 +1,302 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_GENERATOR_H
+#define EIGEN_CXX11_TENSOR_TENSOR_GENERATOR_H
+
+namespace Eigen {
+
+/** \class TensorGeneratorOp
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor generator class.
+  *
+  *
+  */
+namespace internal {
+template<typename Generator, typename XprType>
+struct traits<TensorGeneratorOp<Generator, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename Generator, typename XprType>
+struct eval<TensorGeneratorOp<Generator, XprType>, Eigen::Dense>
+{
+  typedef const TensorGeneratorOp<Generator, XprType>& type;
+};
+
+template<typename Generator, typename XprType>
+struct nested<TensorGeneratorOp<Generator, XprType>, 1, typename eval<TensorGeneratorOp<Generator, XprType> >::type>
+{
+  typedef TensorGeneratorOp<Generator, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename Generator, typename XprType>
+class TensorGeneratorOp : public TensorBase<TensorGeneratorOp<Generator, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorGeneratorOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorGeneratorOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorGeneratorOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorGeneratorOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorGeneratorOp(const XprType& expr, const Generator& generator)
+      : m_xpr(expr), m_generator(generator) {}
+
+    EIGEN_DEVICE_FUNC
+    const Generator& generator() const { return m_generator; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Generator m_generator;
+};
+
+
+// Eval as rvalue
+template<typename Generator, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorGeneratorOp<Generator, ArgType>, Device>
+{
+  typedef TensorGeneratorOp<Generator, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
+  static const int NumDims = internal::array_size<Dimensions>::value;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  enum {
+    IsAligned         = false,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = true,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
+  };
+
+  typedef internal::TensorIntDivisor<Index> IndexDivisor;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename internal::TensorMaterializedBlock<CoeffReturnType, NumDims,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      :  m_device(device), m_generator(op.generator())
+  {
+    TensorEvaluator<ArgType, Device> argImpl(op.expression(), device);
+    m_dimensions = argImpl.dimensions();
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_strides[0] = 1;
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < NumDims; ++i) {
+        m_strides[i] = m_strides[i - 1] * m_dimensions[i - 1];
+        if (m_strides[i] != 0) m_fast_strides[i] = IndexDivisor(m_strides[i]);
+      }
+    } else {
+      m_strides[NumDims - 1] = 1;
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_strides[i] = m_strides[i + 1] * m_dimensions[i + 1];
+        if (m_strides[i] != 0) m_fast_strides[i] = IndexDivisor(m_strides[i]);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    return true;
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    array<Index, NumDims> coords;
+    extract_coordinates(index, coords);
+    return m_generator(coords);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
+    EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+packetSize-1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[packetSize];
+    for (int i = 0; i < packetSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    const size_t target_size = m_device.firstLevelCacheSize();
+    // TODO(ezhulenev): Generator should have a cost.
+    return internal::TensorBlockResourceRequirements::skewed<Scalar>(
+        target_size);
+  }
+
+  struct BlockIteratorState {
+    Index stride;
+    Index span;
+    Index size;
+    Index count;
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    static const bool is_col_major =
+        static_cast<int>(Layout) == static_cast<int>(ColMajor);
+
+    // Compute spatial coordinates for the first block element.
+    array<Index, NumDims> coords;
+    extract_coordinates(desc.offset(), coords);
+    array<Index, NumDims> initial_coords = coords;
+
+    // Offset in the output block buffer.
+    Index offset = 0;
+
+    // Initialize output block iterator state. Dimension in this array are
+    // always in inner_most -> outer_most order (col major layout).
+    array<BlockIteratorState, NumDims> it;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = is_col_major ? i : NumDims - 1 - i;
+      it[i].size = desc.dimension(dim);
+      it[i].stride = i == 0 ? 1 : (it[i - 1].size * it[i - 1].stride);
+      it[i].span = it[i].stride * (it[i].size - 1);
+      it[i].count = 0;
+    }
+    eigen_assert(it[0].stride == 1);
+
+    // Prepare storage for the materialized generator result.
+    const typename TensorBlock::Storage block_storage =
+        TensorBlock::prepareStorage(desc, scratch);
+
+    CoeffReturnType* block_buffer = block_storage.data();
+
+    static const int packet_size = PacketType<CoeffReturnType, Device>::size;
+
+    static const int inner_dim = is_col_major ? 0 : NumDims - 1;
+    const Index inner_dim_size = it[0].size;
+    const Index inner_dim_vectorized = inner_dim_size - packet_size;
+
+    while (it[NumDims - 1].count < it[NumDims - 1].size) {
+      Index i = 0;
+      // Generate data for the vectorized part of the inner-most dimension.
+      for (; i <= inner_dim_vectorized; i += packet_size) {
+        for (Index j = 0; j < packet_size; ++j) {
+          array<Index, NumDims> j_coords = coords;  // Break loop dependence.
+          j_coords[inner_dim] += j;
+          *(block_buffer + offset + i + j) = m_generator(j_coords);
+        }
+        coords[inner_dim] += packet_size;
+      }
+      // Finalize non-vectorized part of the inner-most dimension.
+      for (; i < inner_dim_size; ++i) {
+        *(block_buffer + offset + i) = m_generator(coords);
+        coords[inner_dim]++;
+      }
+      coords[inner_dim] = initial_coords[inner_dim];
+
+      // For the 1d tensor we need to generate only one inner-most dimension.
+      if (NumDims == 1) break;
+
+      // Update offset.
+      for (i = 1; i < NumDims; ++i) {
+        if (++it[i].count < it[i].size) {
+          offset += it[i].stride;
+          coords[is_col_major ? i : NumDims - 1 - i]++;
+          break;
+        }
+        if (i != NumDims - 1) it[i].count = 0;
+        coords[is_col_major ? i : NumDims - 1 - i] =
+            initial_coords[is_col_major ? i : NumDims - 1 - i];
+        offset -= it[i].span;
+      }
+    }
+
+    return block_storage.AsTensorMaterializedBlock();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool) const {
+    // TODO(rmlarsen): This is just a placeholder. Define interface to make
+    // generators return their cost.
+    return TensorOpCost(0, 0, TensorOpCost::AddCost<Scalar>() +
+                                  TensorOpCost::MulCost<Scalar>());
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType  data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler&) const {}
+#endif
+
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void extract_coordinates(Index index, array<Index, NumDims>& coords) const {
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_fast_strides[i];
+        index -= idx * m_strides[i];
+        coords[i] = idx;
+      }
+      coords[0] = index;
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_fast_strides[i];
+        index -= idx * m_strides[i];
+        coords[i] = idx;
+      }
+      coords[NumDims-1] = index;
+    }
+  }
+
+  const Device EIGEN_DEVICE_REF m_device;
+  Dimensions m_dimensions;
+  array<Index, NumDims> m_strides;
+  array<IndexDivisor, NumDims> m_fast_strides;
+  Generator m_generator;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_GENERATOR_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorGlobalFunctions.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorGlobalFunctions.h
new file mode 100644
index 0000000..665b861
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorGlobalFunctions.h
@@ -0,0 +1,33 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_GLOBAL_FUNCTIONS_H
+#define EIGEN_CXX11_TENSOR_TENSOR_GLOBAL_FUNCTIONS_H
+
+namespace Eigen {
+
+/** \cpp11 \returns an expression of the coefficient-wise betainc(\a x, \a a, \a b) to the given tensors.
+ *
+ * This function computes the regularized incomplete beta function (integral).
+ *
+ */
+template <typename ADerived, typename BDerived, typename XDerived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const
+    TensorCwiseTernaryOp<internal::scalar_betainc_op<typename XDerived::Scalar>,
+                         const ADerived, const BDerived, const XDerived>
+    betainc(const ADerived& a, const BDerived& b, const XDerived& x) {
+  return TensorCwiseTernaryOp<
+      internal::scalar_betainc_op<typename XDerived::Scalar>, const ADerived,
+      const BDerived, const XDerived>(
+      a, b, x, internal::scalar_betainc_op<typename XDerived::Scalar>());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_GLOBAL_FUNCTIONS_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorGpuHipCudaDefines.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
new file mode 100644
index 0000000..cb53ce2
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
@@ -0,0 +1,99 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2018 Deven Desai <deven.desai.amd@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#if defined(EIGEN_USE_GPU) && !defined(EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H)
+#define EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
+
+// Note that we are using EIGEN_USE_HIP here instead of EIGEN_HIPCC...this is by design
+// There is code in the Tensorflow codebase that will define EIGEN_USE_GPU,  but
+// for some reason gets sent to the gcc/host compiler instead of the gpu/nvcc/hipcc compiler
+// When compiling such files, gcc will end up trying to pick up the CUDA headers by 
+// default (see the code within "unsupported/Eigen/CXX11/Tensor" that is guarded by EIGEN_USE_GPU)
+// This will obviously not work when trying to compile tensorflow on a system with no CUDA
+// To work around this issue for HIP systems (and leave the default behaviour intact), the
+// HIP tensorflow build defines EIGEN_USE_HIP when compiling all source files, and 
+// "unsupported/Eigen/CXX11/Tensor" has been updated to use HIP header when EIGEN_USE_HIP is
+// defined. In continuation of that requirement, the guard here needs to be EIGEN_USE_HIP as well
+
+#if defined(EIGEN_USE_HIP)
+
+#define gpuStream_t hipStream_t
+#define gpuDeviceProp_t hipDeviceProp_t
+#define gpuError_t hipError_t
+#define gpuSuccess hipSuccess
+#define gpuErrorNotReady hipErrorNotReady
+#define gpuGetDeviceCount hipGetDeviceCount
+#define gpuGetLastError hipGetLastError
+#define gpuPeekAtLastError hipPeekAtLastError
+#define gpuGetErrorName hipGetErrorName
+#define gpuGetErrorString hipGetErrorString
+#define gpuGetDeviceProperties hipGetDeviceProperties
+#define gpuStreamDefault hipStreamDefault
+#define gpuGetDevice hipGetDevice
+#define gpuSetDevice hipSetDevice
+#define gpuMalloc hipMalloc
+#define gpuFree hipFree
+#define gpuMemsetAsync hipMemsetAsync
+#define gpuMemcpyAsync hipMemcpyAsync
+#define gpuMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define gpuMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define gpuMemcpyHostToDevice hipMemcpyHostToDevice
+#define gpuStreamQuery hipStreamQuery
+#define gpuSharedMemConfig hipSharedMemConfig
+#define gpuDeviceSetSharedMemConfig hipDeviceSetSharedMemConfig
+#define gpuStreamSynchronize hipStreamSynchronize
+#define gpuDeviceSynchronize hipDeviceSynchronize
+#define gpuMemcpy hipMemcpy
+
+#else
+
+#define gpuStream_t cudaStream_t
+#define gpuDeviceProp_t cudaDeviceProp
+#define gpuError_t cudaError_t
+#define gpuSuccess cudaSuccess
+#define gpuErrorNotReady cudaErrorNotReady
+#define gpuGetDeviceCount cudaGetDeviceCount
+#define gpuGetLastError cudaGetLastError
+#define gpuPeekAtLastError cudaPeekAtLastError
+#define gpuGetErrorName cudaGetErrorName
+#define gpuGetErrorString cudaGetErrorString
+#define gpuGetDeviceProperties cudaGetDeviceProperties
+#define gpuStreamDefault cudaStreamDefault
+#define gpuGetDevice cudaGetDevice
+#define gpuSetDevice cudaSetDevice
+#define gpuMalloc cudaMalloc
+#define gpuFree cudaFree
+#define gpuMemsetAsync cudaMemsetAsync
+#define gpuMemcpyAsync cudaMemcpyAsync
+#define gpuMemcpyDeviceToDevice cudaMemcpyDeviceToDevice
+#define gpuMemcpyDeviceToHost cudaMemcpyDeviceToHost
+#define gpuMemcpyHostToDevice cudaMemcpyHostToDevice
+#define gpuStreamQuery cudaStreamQuery
+#define gpuSharedMemConfig cudaSharedMemConfig
+#define gpuDeviceSetSharedMemConfig cudaDeviceSetSharedMemConfig
+#define gpuStreamSynchronize cudaStreamSynchronize
+#define gpuDeviceSynchronize cudaDeviceSynchronize
+#define gpuMemcpy cudaMemcpy
+
+#endif
+
+// gpu_assert can be overridden
+#ifndef gpu_assert
+
+#if defined(EIGEN_HIP_DEVICE_COMPILE)
+// HIPCC do not support the use of assert on the GPU side.
+#define gpu_assert(COND)
+#else
+#define gpu_assert(COND) assert(COND)
+#endif
+
+#endif // gpu_assert
+
+#endif  // EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h
new file mode 100644
index 0000000..1d142f2
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorGpuHipCudaUndefines.h
@@ -0,0 +1,44 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2018 Deven Desai <deven.desai.amd@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#if defined(EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H)
+
+#ifndef EIGEN_PERMANENTLY_ENABLE_GPU_HIP_CUDA_DEFINES
+
+#undef gpuStream_t
+#undef gpuDeviceProp_t 
+#undef gpuError_t
+#undef gpuSuccess
+#undef gpuErrorNotReady
+#undef gpuGetDeviceCount
+#undef gpuGetErrorString
+#undef gpuGetDeviceProperties
+#undef gpuStreamDefault
+#undef gpuGetDevice
+#undef gpuSetDevice
+#undef gpuMalloc
+#undef gpuFree
+#undef gpuMemsetAsync
+#undef gpuMemcpyAsync
+#undef gpuMemcpyDeviceToDevice
+#undef gpuMemcpyDeviceToHost
+#undef gpuMemcpyHostToDevice
+#undef gpuStreamQuery
+#undef gpuSharedMemConfig
+#undef gpuDeviceSetSharedMemConfig
+#undef gpuStreamSynchronize
+#undef gpuDeviceSynchronize
+#undef gpuMemcpy
+
+#endif // EIGEN_PERMANENTLY_ENABLE_GPU_HIP_CUDA_DEFINES
+
+#undef EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
+
+#endif // EIGEN_CXX11_TENSOR_GPU_HIP_CUDA_DEFINES_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorIO.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorIO.h
new file mode 100644
index 0000000..a901c5d
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorIO.h
@@ -0,0 +1,79 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_IO_H
+#define EIGEN_CXX11_TENSOR_TENSOR_IO_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Print the tensor as a 2d matrix
+template <typename Tensor, int Rank>
+struct TensorPrinter {
+  static void run (std::ostream& os, const Tensor& tensor) {
+    typedef typename internal::remove_const<typename Tensor::Scalar>::type Scalar;
+    typedef typename Tensor::Index Index;
+    const Index total_size = internal::array_prod(tensor.dimensions());
+    if (total_size > 0) {
+      const Index first_dim = Eigen::internal::array_get<0>(tensor.dimensions());
+      static const int layout = Tensor::Layout;
+      Map<const Array<Scalar, Dynamic, Dynamic, layout> > matrix(const_cast<Scalar*>(tensor.data()), first_dim, total_size/first_dim);
+      os << matrix;
+    }
+  }
+};
+
+
+// Print the tensor as a vector
+template <typename Tensor>
+struct TensorPrinter<Tensor, 1> {
+  static void run (std::ostream& os, const Tensor& tensor) {
+    typedef typename internal::remove_const<typename Tensor::Scalar>::type Scalar;
+    typedef typename Tensor::Index Index;
+    const Index total_size = internal::array_prod(tensor.dimensions());
+    if (total_size > 0) {
+      Map<const Array<Scalar, Dynamic, 1> > array(const_cast<Scalar*>(tensor.data()), total_size);
+      os << array;
+    }
+  }
+};
+
+
+// Print the tensor as a scalar
+template <typename Tensor>
+struct TensorPrinter<Tensor, 0> {
+  static void run (std::ostream& os, const Tensor& tensor) {
+    os << tensor.coeff(0);
+  }
+};
+}
+
+template <typename T>
+std::ostream& operator << (std::ostream& os, const TensorBase<T, ReadOnlyAccessors>& expr) {
+  typedef TensorEvaluator<const TensorForcedEvalOp<const T>, DefaultDevice> Evaluator;
+  typedef typename Evaluator::Dimensions Dimensions;
+
+  // Evaluate the expression if needed
+  TensorForcedEvalOp<const T> eval = expr.eval();
+  Evaluator tensor(eval, DefaultDevice());
+  tensor.evalSubExprsIfNeeded(NULL);
+
+  // Print the result
+  static const int rank = internal::array_size<Dimensions>::value;
+  internal::TensorPrinter<Evaluator, rank>::run(os, tensor);
+
+  // Cleanup.
+  tensor.cleanup();
+  return os;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_IO_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorImagePatch.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorImagePatch.h
new file mode 100644
index 0000000..dd51850
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorImagePatch.h
@@ -0,0 +1,603 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
+#define EIGEN_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
+
+namespace Eigen {
+
+/** \class TensorImagePatch
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Patch extraction specialized for image processing.
+  * This assumes that the input has a least 3 dimensions ordered as follow:
+  *  1st dimension: channels (of size d)
+  *  2nd dimension: rows (of size r)
+  *  3rd dimension: columns (of size c)
+  *  There can be additional dimensions such as time (for video) or batch (for
+  * bulk processing after the first 3.
+  * Calling the image patch code with patch_rows and patch_cols is equivalent
+  * to calling the regular patch extraction code with parameters d, patch_rows,
+  * patch_cols, and 1 for all the additional dimensions.
+  */
+namespace internal {
+
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct traits<TensorImagePatchOp<Rows, Cols, XprType> > : public traits<XprType>
+{
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions + 1;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct eval<TensorImagePatchOp<Rows, Cols, XprType>, Eigen::Dense>
+{
+  typedef const TensorImagePatchOp<Rows, Cols, XprType>& type;
+};
+
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct nested<TensorImagePatchOp<Rows, Cols, XprType>, 1, typename eval<TensorImagePatchOp<Rows, Cols, XprType> >::type>
+{
+  typedef TensorImagePatchOp<Rows, Cols, XprType> type;
+};
+
+template <typename Self, bool Vectorizable>
+struct ImagePatchCopyOp {
+  typedef typename Self::Index Index;
+  typedef typename Self::Scalar Scalar;
+  typedef typename Self::Impl Impl;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Self& self, const Index num_coeff_to_copy, const Index dst_index,
+      Scalar* dst_data, const Index src_index) {
+    const Impl& impl = self.impl();
+    for (Index i = 0; i < num_coeff_to_copy; ++i) {
+      dst_data[dst_index + i] = impl.coeff(src_index + i);
+    }
+  }
+};
+
+template <typename Self>
+struct ImagePatchCopyOp<Self, true> {
+  typedef typename Self::Index Index;
+  typedef typename Self::Scalar Scalar;
+  typedef typename Self::Impl Impl;
+  typedef typename packet_traits<Scalar>::type Packet;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Self& self, const Index num_coeff_to_copy, const Index dst_index,
+      Scalar* dst_data, const Index src_index) {
+    const Impl& impl = self.impl();
+    const Index packet_size = internal::unpacket_traits<Packet>::size;
+    const Index vectorized_size =
+        (num_coeff_to_copy / packet_size) * packet_size;
+    for (Index i = 0; i < vectorized_size; i += packet_size) {
+      Packet p = impl.template packet<Unaligned>(src_index + i);
+      internal::pstoret<Scalar, Packet, Unaligned>(dst_data + dst_index + i, p);
+    }
+    for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
+      dst_data[dst_index + i] = impl.coeff(src_index + i);
+    }
+  }
+};
+
+template <typename Self>
+struct ImagePatchPaddingOp {
+  typedef typename Self::Index Index;
+  typedef typename Self::Scalar Scalar;
+  typedef typename packet_traits<Scalar>::type Packet;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
+      const Index num_coeff_to_pad, const Scalar padding_value,
+      const Index dst_index, Scalar* dst_data) {
+    const Index packet_size = internal::unpacket_traits<Packet>::size;
+    const Packet padded_packet = internal::pset1<Packet>(padding_value);
+    const Index vectorized_size =
+        (num_coeff_to_pad / packet_size) * packet_size;
+    for (Index i = 0; i < vectorized_size; i += packet_size) {
+      internal::pstoret<Scalar, Packet, Unaligned>(dst_data + dst_index + i,
+                                                   padded_packet);
+    }
+    for (Index i = vectorized_size; i < num_coeff_to_pad; ++i) {
+      dst_data[dst_index + i] = padding_value;
+    }
+  }
+};
+
+}  // end namespace internal
+
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>
+class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorImagePatchOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorImagePatchOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorImagePatchOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorImagePatchOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
+                                                           DenseIndex row_strides, DenseIndex col_strides,
+                                                           DenseIndex in_row_strides, DenseIndex in_col_strides,
+                                                           DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
+                                                           PaddingType padding_type, Scalar padding_value)
+                                                           : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                           m_row_strides(row_strides), m_col_strides(col_strides),
+                                                           m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                           m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                           m_padding_explicit(false), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
+                                                           m_padding_type(padding_type), m_padding_value(padding_value) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
+                                                           DenseIndex row_strides, DenseIndex col_strides,
+                                                           DenseIndex in_row_strides, DenseIndex in_col_strides,
+                                                           DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
+                                                           DenseIndex padding_top, DenseIndex padding_bottom,
+                                                           DenseIndex padding_left, DenseIndex padding_right,
+                                                           Scalar padding_value)
+                                                           : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                           m_row_strides(row_strides), m_col_strides(col_strides),
+                                                           m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                           m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                           m_padding_explicit(true), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+                                                           m_padding_left(padding_left), m_padding_right(padding_right),
+                                                           m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
+
+
+    EIGEN_DEVICE_FUNC
+    DenseIndex patch_rows() const { return m_patch_rows; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex patch_cols() const { return m_patch_cols; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex row_strides() const { return m_row_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex col_strides() const { return m_col_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex in_row_strides() const { return m_in_row_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex in_col_strides() const { return m_in_col_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex row_inflate_strides() const { return m_row_inflate_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex col_inflate_strides() const { return m_col_inflate_strides; }
+    EIGEN_DEVICE_FUNC
+    bool padding_explicit() const { return m_padding_explicit; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_top() const { return m_padding_top; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_bottom() const { return m_padding_bottom; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_left() const { return m_padding_left; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_right() const { return m_padding_right; }
+    EIGEN_DEVICE_FUNC
+    PaddingType padding_type() const { return m_padding_type; }
+    EIGEN_DEVICE_FUNC
+    Scalar padding_value() const { return m_padding_value; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const DenseIndex m_patch_rows;
+    const DenseIndex m_patch_cols;
+    const DenseIndex m_row_strides;
+    const DenseIndex m_col_strides;
+    const DenseIndex m_in_row_strides;
+    const DenseIndex m_in_col_strides;
+    const DenseIndex m_row_inflate_strides;
+    const DenseIndex m_col_inflate_strides;
+    const bool m_padding_explicit;
+    const DenseIndex m_padding_top;
+    const DenseIndex m_padding_bottom;
+    const DenseIndex m_padding_left;
+    const DenseIndex m_padding_right;
+    const PaddingType m_padding_type;
+    const Scalar m_padding_value;
+};
+
+// Eval as rvalue
+template<DenseIndex Rows, DenseIndex Cols, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
+{
+  typedef TensorImagePatchOp<Rows, Cols, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  static const int NumDims = NumInputDims + 1;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>,
+                          Device> Self;
+  typedef TensorEvaluator<ArgType, Device> Impl;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = false,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,
+    RawAccess         = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator( const XprType& op, const Device& device)
+      : m_device(device), m_impl(op.expression(), device)
+  {
+    EIGEN_STATIC_ASSERT((NumDims >= 4), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    m_paddingValue = op.padding_value();
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+
+    // Caches a few variables.
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputDepth = input_dims[0];
+      m_inputRows = input_dims[1];
+      m_inputCols = input_dims[2];
+    } else {
+      m_inputDepth = input_dims[NumInputDims-1];
+      m_inputRows = input_dims[NumInputDims-2];
+      m_inputCols = input_dims[NumInputDims-3];
+    }
+
+    m_row_strides = op.row_strides();
+    m_col_strides = op.col_strides();
+
+    // Input strides and effective input/patch size
+    m_in_row_strides = op.in_row_strides();
+    m_in_col_strides = op.in_col_strides();
+    m_row_inflate_strides = op.row_inflate_strides();
+    m_col_inflate_strides = op.col_inflate_strides();
+    // The "effective" input rows and input cols are the input rows and cols
+    // after inflating them with zeros.
+    // For examples, a 2x3 matrix with row_inflate_strides and
+    // col_inflate_strides of 2 comes from:
+    //   A B C
+    //   D E F
+    //
+    // to a matrix is 3 x 5:
+    //
+    //   A . B . C
+    //   . . . . .
+    //   D . E . F
+
+    m_input_rows_eff = (m_inputRows - 1) * m_row_inflate_strides + 1;
+    m_input_cols_eff = (m_inputCols - 1) * m_col_inflate_strides + 1;
+    m_patch_rows_eff = op.patch_rows() + (op.patch_rows() - 1) * (m_in_row_strides - 1);
+    m_patch_cols_eff = op.patch_cols() + (op.patch_cols() - 1) * (m_in_col_strides - 1);
+
+    if (op.padding_explicit()) {
+      m_outputRows = numext::ceil((m_input_rows_eff + op.padding_top() + op.padding_bottom() - m_patch_rows_eff + 1.f) / static_cast<float>(m_row_strides));
+      m_outputCols = numext::ceil((m_input_cols_eff + op.padding_left() + op.padding_right() - m_patch_cols_eff + 1.f) / static_cast<float>(m_col_strides));
+      m_rowPaddingTop = op.padding_top();
+      m_colPaddingLeft = op.padding_left();
+    } else {
+      // Computing padding from the type
+      switch (op.padding_type()) {
+        case PADDING_VALID:
+          m_outputRows = numext::ceil((m_input_rows_eff - m_patch_rows_eff + 1.f) / static_cast<float>(m_row_strides));
+          m_outputCols = numext::ceil((m_input_cols_eff - m_patch_cols_eff + 1.f) / static_cast<float>(m_col_strides));
+          // Calculate the padding
+          m_rowPaddingTop = numext::maxi<Index>(0, ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2);
+          m_colPaddingLeft = numext::maxi<Index>(0, ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2);
+          break;
+        case PADDING_SAME:
+          m_outputRows = numext::ceil(m_input_rows_eff / static_cast<float>(m_row_strides));
+          m_outputCols = numext::ceil(m_input_cols_eff / static_cast<float>(m_col_strides));
+          // Calculate the padding
+          m_rowPaddingTop = ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2;
+          m_colPaddingLeft = ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2;
+          // The padding size calculation for PADDING_SAME has been updated to
+          // be consistent with how TensorFlow extracts its paddings.
+          m_rowPaddingTop = numext::maxi<Index>(0, m_rowPaddingTop);
+          m_colPaddingLeft = numext::maxi<Index>(0, m_colPaddingLeft);
+          break;
+        default:
+          eigen_assert(false && "unexpected padding");
+          m_outputCols=0; // silence the uninitialised warning;
+          m_outputRows=0; //// silence the uninitialised warning;
+      }
+    }
+    eigen_assert(m_outputRows > 0);
+    eigen_assert(m_outputCols > 0);
+
+    // Dimensions for result of extraction.
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      // ColMajor
+      // 0: depth
+      // 1: patch_rows
+      // 2: patch_cols
+      // 3: number of patches
+      // 4 and beyond: anything else (such as batch).
+      m_dimensions[0] = input_dims[0];
+      m_dimensions[1] = op.patch_rows();
+      m_dimensions[2] = op.patch_cols();
+      m_dimensions[3] = m_outputRows * m_outputCols;
+      for (int i = 4; i < NumDims; ++i) {
+        m_dimensions[i] = input_dims[i-1];
+      }
+    } else {
+      // RowMajor
+      // NumDims-1: depth
+      // NumDims-2: patch_rows
+      // NumDims-3: patch_cols
+      // NumDims-4: number of patches
+      // NumDims-5 and beyond: anything else (such as batch).
+      m_dimensions[NumDims-1] = input_dims[NumInputDims-1];
+      m_dimensions[NumDims-2] = op.patch_rows();
+      m_dimensions[NumDims-3] = op.patch_cols();
+      m_dimensions[NumDims-4] = m_outputRows * m_outputCols;
+      for (int i = NumDims-5; i >= 0; --i) {
+        m_dimensions[i] = input_dims[i];
+      }
+    }
+
+    // Strides for moving the patch in various dimensions.
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_colStride = m_dimensions[1];
+      m_patchStride = m_colStride * m_dimensions[2] * m_dimensions[0];
+      m_otherStride = m_patchStride * m_dimensions[3];
+    } else {
+      m_colStride = m_dimensions[NumDims-2];
+      m_patchStride = m_colStride * m_dimensions[NumDims-3] * m_dimensions[NumDims-1];
+      m_otherStride = m_patchStride * m_dimensions[NumDims-4];
+    }
+
+    // Strides for navigating through the input tensor.
+    m_rowInputStride = m_inputDepth;
+    m_colInputStride = m_inputDepth * m_inputRows;
+    m_patchInputStride = m_inputDepth * m_inputRows * m_inputCols;
+
+    // Fast representations of different variables.
+    m_fastOtherStride = internal::TensorIntDivisor<Index>(m_otherStride);
+    m_fastPatchStride = internal::TensorIntDivisor<Index>(m_patchStride);
+    m_fastColStride = internal::TensorIntDivisor<Index>(m_colStride);
+    m_fastInflateRowStride = internal::TensorIntDivisor<Index>(m_row_inflate_strides);
+    m_fastInflateColStride = internal::TensorIntDivisor<Index>(m_col_inflate_strides);
+    m_fastInputColsEff = internal::TensorIntDivisor<Index>(m_input_cols_eff);
+
+    // Number of patches in the width dimension.
+    m_fastOutputRows = internal::TensorIntDivisor<Index>(m_outputRows);
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[0]);
+    } else {
+      m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[NumDims-1]);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    // Patch index corresponding to the passed in index.
+    const Index patchIndex = index / m_fastPatchStride;
+    // Find the offset of the element wrt the location of the first element.
+    const Index patchOffset = (index - patchIndex * m_patchStride) / m_fastOutputDepth;
+
+    // Other ways to index this element.
+    const Index otherIndex = (NumDims == 4) ? 0 : index / m_fastOtherStride;
+    const Index patch2DIndex = (NumDims == 4) ? patchIndex : (index - otherIndex * m_otherStride) / m_fastPatchStride;
+
+    // Calculate col index in the input original tensor.
+    const Index colIndex = patch2DIndex / m_fastOutputRows;
+    const Index colOffset = patchOffset / m_fastColStride;
+    const Index inputCol = colIndex * m_col_strides + colOffset * m_in_col_strides - m_colPaddingLeft;
+    const Index origInputCol = (m_col_inflate_strides == 1) ? inputCol : ((inputCol >= 0) ? (inputCol / m_fastInflateColStride) : 0);
+    if (inputCol < 0 || inputCol >= m_input_cols_eff ||
+        ((m_col_inflate_strides != 1) && (inputCol != origInputCol * m_col_inflate_strides))) {
+      return Scalar(m_paddingValue);
+    }
+
+    // Calculate row index in the original input tensor.
+    const Index rowIndex = patch2DIndex - colIndex * m_outputRows;
+    const Index rowOffset = patchOffset - colOffset * m_colStride;
+    const Index inputRow = rowIndex * m_row_strides + rowOffset * m_in_row_strides - m_rowPaddingTop;
+    const Index origInputRow = (m_row_inflate_strides == 1) ? inputRow : ((inputRow >= 0) ? (inputRow / m_fastInflateRowStride) : 0);
+    if (inputRow < 0 || inputRow >= m_input_rows_eff ||
+        ((m_row_inflate_strides != 1) && (inputRow != origInputRow * m_row_inflate_strides))) {
+      return Scalar(m_paddingValue);
+    }
+
+    const int depth_index = static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
+    const Index depth = index - (index / m_fastOutputDepth) * m_dimensions[depth_index];
+
+    const Index inputIndex = depth + origInputRow * m_rowInputStride + origInputCol * m_colInputStride + otherIndex * m_patchInputStride;
+    return m_impl.coeff(inputIndex);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    if (m_in_row_strides != 1 || m_in_col_strides != 1 || m_row_inflate_strides != 1 || m_col_inflate_strides != 1) {
+      return packetWithPossibleZero(index);
+    }
+
+    const Index indices[2] = {index, index + PacketSize - 1};
+    const Index patchIndex = indices[0] / m_fastPatchStride;
+    if (patchIndex != indices[1] / m_fastPatchStride) {
+      return packetWithPossibleZero(index);
+    }
+    const Index otherIndex = (NumDims == 4) ? 0 : indices[0] / m_fastOtherStride;
+    eigen_assert(otherIndex == indices[1] / m_fastOtherStride);
+
+    // Find the offset of the element wrt the location of the first element.
+    const Index patchOffsets[2] = {(indices[0] - patchIndex * m_patchStride) / m_fastOutputDepth,
+                                   (indices[1] - patchIndex * m_patchStride) / m_fastOutputDepth};
+
+    const Index patch2DIndex = (NumDims == 4) ? patchIndex : (indices[0] - otherIndex * m_otherStride) / m_fastPatchStride;
+    eigen_assert(patch2DIndex == (indices[1] - otherIndex * m_otherStride) / m_fastPatchStride);
+
+    const Index colIndex = patch2DIndex / m_fastOutputRows;
+    const Index colOffsets[2] = {patchOffsets[0] / m_fastColStride, patchOffsets[1] / m_fastColStride};
+
+    // Calculate col indices in the original input tensor.
+    const Index inputCols[2] = {colIndex * m_col_strides + colOffsets[0] -
+      m_colPaddingLeft, colIndex * m_col_strides + colOffsets[1] - m_colPaddingLeft};
+    if (inputCols[1] < 0 || inputCols[0] >= m_inputCols) {
+      return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
+    }
+
+    if (inputCols[0] == inputCols[1]) {
+      const Index rowIndex = patch2DIndex - colIndex * m_outputRows;
+      const Index rowOffsets[2] = {patchOffsets[0] - colOffsets[0]*m_colStride, patchOffsets[1] - colOffsets[1]*m_colStride};
+      eigen_assert(rowOffsets[0] <= rowOffsets[1]);
+      // Calculate col indices in the original input tensor.
+      const Index inputRows[2] = {rowIndex * m_row_strides + rowOffsets[0] -
+        m_rowPaddingTop, rowIndex * m_row_strides + rowOffsets[1] - m_rowPaddingTop};
+
+      if (inputRows[1] < 0 || inputRows[0] >= m_inputRows) {
+        return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
+      }
+
+      if (inputRows[0] >= 0 && inputRows[1] < m_inputRows) {
+        // no padding
+        const int depth_index = static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
+        const Index depth = index - (index / m_fastOutputDepth) * m_dimensions[depth_index];
+        const Index inputIndex = depth + inputRows[0] * m_rowInputStride + inputCols[0] * m_colInputStride + otherIndex * m_patchInputStride;
+        return m_impl.template packet<Unaligned>(inputIndex);
+      }
+    }
+
+    return packetWithPossibleZero(index);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowPaddingTop() const { return m_rowPaddingTop; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colPaddingLeft() const { return m_colPaddingLeft; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputRows() const { return m_outputRows; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputCols() const { return m_outputCols; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userRowStride() const { return m_row_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userColStride() const { return m_col_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInRowStride() const { return m_in_row_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInColStride() const { return m_in_col_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowInflateStride() const { return m_row_inflate_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colInflateStride() const { return m_col_inflate_strides; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    // We conservatively estimate the cost for the code path where the computed
+    // index is inside the original image and
+    // TensorEvaluator<ArgType, Device>::CoordAccess is false.
+    const double compute_cost = 3 * TensorOpCost::DivCost<Index>() +
+                                6 * TensorOpCost::MulCost<Index>() +
+                                8 * TensorOpCost::MulCost<Index>();
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
+  }
+
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const
+  {
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  Dimensions m_dimensions;
+
+  Index m_otherStride;
+  Index m_patchStride;
+  Index m_colStride;
+  Index m_row_strides;
+  Index m_col_strides;
+
+  Index m_in_row_strides;
+  Index m_in_col_strides;
+  Index m_row_inflate_strides;
+  Index m_col_inflate_strides;
+
+  Index m_input_rows_eff;
+  Index m_input_cols_eff;
+  Index m_patch_rows_eff;
+  Index m_patch_cols_eff;
+
+  internal::TensorIntDivisor<Index> m_fastOtherStride;
+  internal::TensorIntDivisor<Index> m_fastPatchStride;
+  internal::TensorIntDivisor<Index> m_fastColStride;
+  internal::TensorIntDivisor<Index> m_fastInflateRowStride;
+  internal::TensorIntDivisor<Index> m_fastInflateColStride;
+  internal::TensorIntDivisor<Index> m_fastInputColsEff;
+
+  Index m_rowInputStride;
+  Index m_colInputStride;
+  Index m_patchInputStride;
+
+  Index m_inputDepth;
+  Index m_inputRows;
+  Index m_inputCols;
+
+  Index m_outputRows;
+  Index m_outputCols;
+
+  Index m_rowPaddingTop;
+  Index m_colPaddingLeft;
+
+  internal::TensorIntDivisor<Index> m_fastOutputRows;
+  internal::TensorIntDivisor<Index> m_fastOutputDepth;
+
+  Scalar m_paddingValue;
+
+  const Device EIGEN_DEVICE_REF m_device;
+  TensorEvaluator<ArgType, Device> m_impl;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorIndexList.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorIndexList.h
new file mode 100644
index 0000000..2d8c7b9
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorIndexList.h
@@ -0,0 +1,738 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_INDEX_LIST_H
+#define EIGEN_CXX11_TENSOR_TENSOR_INDEX_LIST_H
+
+
+#if EIGEN_HAS_CONSTEXPR && EIGEN_HAS_VARIADIC_TEMPLATES
+
+#define EIGEN_HAS_INDEX_LIST
+
+namespace Eigen {
+
+/** \internal
+  *
+  * \class TensorIndexList
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Set of classes used to encode a set of Tensor dimensions/indices.
+  *
+  * The indices in the list can be known at compile time or at runtime. A mix
+  * of static and dynamic indices can also be provided if needed. The tensor
+  * code will attempt to take advantage of the indices that are known at
+  * compile time to optimize the code it generates.
+  *
+  * This functionality requires a c++11 compliant compiler. If your compiler
+  * is older you need to use arrays of indices instead.
+  *
+  * Several examples are provided in the cxx11_tensor_index_list.cpp file.
+  *
+  * \sa Tensor
+  */
+
+template <Index n>
+struct type2index {
+  static const Index value = n;
+  EIGEN_DEVICE_FUNC constexpr operator Index() const { return n; }
+  EIGEN_DEVICE_FUNC void set(Index val) {
+    eigen_assert(val == n);
+  }
+};
+
+// This can be used with IndexPairList to get compile-time constant pairs,
+// such as IndexPairList<type2indexpair<1,2>, type2indexpair<3,4>>().
+template <Index f, Index s>
+struct type2indexpair {
+  static const Index first = f;
+  static const Index second = s;
+
+  constexpr EIGEN_DEVICE_FUNC operator IndexPair<Index>() const {
+    return IndexPair<Index>(f, s);
+  }
+
+  EIGEN_DEVICE_FUNC void set(const IndexPair<Index>& val) {
+    eigen_assert(val.first == f);
+    eigen_assert(val.second == s);
+  }
+};
+
+
+template<Index n> struct NumTraits<type2index<n> >
+{
+  typedef Index Real;
+  enum {
+    IsComplex = 0,
+    RequireInitialization = false,
+    ReadCost = 1,
+    AddCost = 1,
+    MulCost = 1
+  };
+
+  EIGEN_DEVICE_FUNC static EIGEN_CONSTEXPR EIGEN_STRONG_INLINE Real epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC static EIGEN_CONSTEXPR EIGEN_STRONG_INLINE Real dummy_precision() { return 0; }
+  EIGEN_DEVICE_FUNC static EIGEN_CONSTEXPR EIGEN_STRONG_INLINE Real highest() { return n; }
+  EIGEN_DEVICE_FUNC static EIGEN_CONSTEXPR EIGEN_STRONG_INLINE Real lowest() { return n; }
+};
+
+namespace internal {
+template <typename T>
+EIGEN_DEVICE_FUNC void update_value(T& val, Index new_val) {
+  val = internal::convert_index<T>(new_val);
+}
+template <Index n>
+EIGEN_DEVICE_FUNC void update_value(type2index<n>& val, Index new_val) {
+  val.set(new_val);
+}
+
+template <typename T>
+EIGEN_DEVICE_FUNC void update_value(T& val, IndexPair<Index> new_val) {
+  val = new_val;
+}
+template <Index f, Index s>
+EIGEN_DEVICE_FUNC void update_value(type2indexpair<f, s>& val, IndexPair<Index> new_val) {
+  val.set(new_val);
+}
+
+
+template <typename T>
+struct is_compile_time_constant {
+  static constexpr bool value = false;
+};
+
+template <Index idx>
+struct is_compile_time_constant<type2index<idx> > {
+  static constexpr bool value = true;
+};
+template <Index idx>
+struct is_compile_time_constant<const type2index<idx> > {
+  static constexpr bool value = true;
+};
+template <Index idx>
+struct is_compile_time_constant<type2index<idx>& > {
+  static constexpr bool value = true;
+};
+template <Index idx>
+struct is_compile_time_constant<const type2index<idx>& > {
+  static constexpr bool value = true;
+};
+
+template <Index f, Index s>
+struct is_compile_time_constant<type2indexpair<f, s> > {
+  static constexpr bool value = true;
+};
+template <Index f, Index s>
+struct is_compile_time_constant<const type2indexpair<f, s> > {
+  static constexpr bool value = true;
+};
+template <Index f, Index s>
+struct is_compile_time_constant<type2indexpair<f, s>& > {
+  static constexpr bool value = true;
+};
+template <Index f, Index s>
+struct is_compile_time_constant<const type2indexpair<f, s>& > {
+  static constexpr bool value = true;
+};
+
+
+template<typename... T>
+struct IndexTuple;
+
+template<typename T, typename... O>
+struct IndexTuple<T, O...> {
+  EIGEN_DEVICE_FUNC constexpr IndexTuple() : head(), others() { }
+  EIGEN_DEVICE_FUNC constexpr IndexTuple(const T& v, const O... o) : head(v), others(o...) { }
+
+  constexpr static int count = 1 + sizeof...(O);
+  T head;
+  IndexTuple<O...> others;
+  typedef T Head;
+  typedef IndexTuple<O...> Other;
+};
+
+template<typename T>
+  struct IndexTuple<T> {
+  EIGEN_DEVICE_FUNC constexpr IndexTuple() : head() { }
+  EIGEN_DEVICE_FUNC constexpr IndexTuple(const T& v) : head(v) { }
+
+  constexpr static int count = 1;
+  T head;
+  typedef T Head;
+};
+
+
+template<int N, typename... T>
+struct IndexTupleExtractor;
+
+template<int N, typename T, typename... O>
+struct IndexTupleExtractor<N, T, O...> {
+
+  typedef typename IndexTupleExtractor<N-1, O...>::ValType ValType;
+
+  EIGEN_DEVICE_FUNC static constexpr ValType& get_val(IndexTuple<T, O...>& val) {
+    return IndexTupleExtractor<N-1, O...>::get_val(val.others);
+  }
+
+  EIGEN_DEVICE_FUNC static constexpr const ValType& get_val(const IndexTuple<T, O...>& val) {
+    return IndexTupleExtractor<N-1, O...>::get_val(val.others);
+  }
+  template <typename V>
+  EIGEN_DEVICE_FUNC static void set_val(IndexTuple<T, O...>& val, V& new_val) {
+    IndexTupleExtractor<N-1, O...>::set_val(val.others, new_val);
+  }
+
+};
+
+template<typename T, typename... O>
+  struct IndexTupleExtractor<0, T, O...> {
+
+  typedef T ValType;
+
+  EIGEN_DEVICE_FUNC static constexpr ValType& get_val(IndexTuple<T, O...>& val) {
+    return val.head;
+  }
+  EIGEN_DEVICE_FUNC static constexpr const ValType& get_val(const IndexTuple<T, O...>& val) {
+    return val.head;
+  }
+  template <typename V>
+  EIGEN_DEVICE_FUNC static void set_val(IndexTuple<T, O...>& val, V& new_val) {
+    val.head = new_val;
+  }
+};
+
+
+
+template <int N, typename T, typename... O>
+EIGEN_DEVICE_FUNC constexpr typename IndexTupleExtractor<N, T, O...>::ValType& array_get(IndexTuple<T, O...>& tuple) {
+  return IndexTupleExtractor<N, T, O...>::get_val(tuple);
+}
+template <int N, typename T, typename... O>
+EIGEN_DEVICE_FUNC constexpr const typename IndexTupleExtractor<N, T, O...>::ValType& array_get(const IndexTuple<T, O...>& tuple) {
+  return IndexTupleExtractor<N, T, O...>::get_val(tuple);
+}
+template <typename T, typename... O>
+  struct array_size<IndexTuple<T, O...> > {
+  static const size_t value = IndexTuple<T, O...>::count;
+};
+template <typename T, typename... O>
+  struct array_size<const IndexTuple<T, O...> > {
+  static const size_t value = IndexTuple<T, O...>::count;
+};
+
+
+
+
+template <Index Idx, typename ValueT>
+struct tuple_coeff {
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static constexpr ValueT get(const Index i, const IndexTuple<T...>& t) {
+    //    return array_get<Idx>(t) * (i == Idx) + tuple_coeff<Idx-1>::get(i, t) * (i != Idx);
+    return (i == Idx ? array_get<Idx>(t) : tuple_coeff<Idx-1, ValueT>::get(i, t));
+  }
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static void set(const Index i, IndexTuple<T...>& t, const ValueT& value) {
+    if (i == Idx) {
+      update_value(array_get<Idx>(t), value);
+    } else {
+      tuple_coeff<Idx-1, ValueT>::set(i, t, value);
+    }
+  }
+
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static constexpr bool value_known_statically(const Index i, const IndexTuple<T...>& t) {
+    return ((i == Idx) & is_compile_time_constant<typename IndexTupleExtractor<Idx, T...>::ValType>::value) ||
+        tuple_coeff<Idx-1, ValueT>::value_known_statically(i, t);
+  }
+
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static constexpr bool values_up_to_known_statically(const IndexTuple<T...>& t) {
+    return is_compile_time_constant<typename IndexTupleExtractor<Idx, T...>::ValType>::value &&
+        tuple_coeff<Idx-1, ValueT>::values_up_to_known_statically(t);
+  }
+
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static constexpr bool values_up_to_statically_known_to_increase(const IndexTuple<T...>& t) {
+    return is_compile_time_constant<typename IndexTupleExtractor<Idx, T...>::ValType>::value &&
+           is_compile_time_constant<typename IndexTupleExtractor<Idx, T...>::ValType>::value &&
+           array_get<Idx>(t) > array_get<Idx-1>(t) &&
+           tuple_coeff<Idx-1, ValueT>::values_up_to_statically_known_to_increase(t);
+  }
+};
+
+template <typename ValueT>
+struct tuple_coeff<0, ValueT> {
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static constexpr ValueT get(const Index /*i*/, const IndexTuple<T...>& t) {
+    //  eigen_assert (i == 0);  // gcc fails to compile assertions in constexpr
+    return array_get<0>(t)/* * (i == 0)*/;
+  }
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static void set(const Index i, IndexTuple<T...>& t, const ValueT value) {
+    eigen_assert (i == 0);
+    update_value(array_get<0>(t), value);
+  }
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static constexpr bool value_known_statically(const Index i, const IndexTuple<T...>&) {
+    return is_compile_time_constant<typename IndexTupleExtractor<0, T...>::ValType>::value && (i == 0);
+  }
+
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static constexpr bool values_up_to_known_statically(const IndexTuple<T...>&) {
+    return is_compile_time_constant<typename IndexTupleExtractor<0, T...>::ValType>::value;
+  }
+
+  template <typename... T>
+  EIGEN_DEVICE_FUNC static constexpr bool values_up_to_statically_known_to_increase(const IndexTuple<T...>&) {
+    return true;
+  }
+};
+}  // namespace internal
+
+
+
+template<typename FirstType, typename... OtherTypes>
+struct IndexList : internal::IndexTuple<FirstType, OtherTypes...> {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr Index operator[] (const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::get(i, *this);
+  }
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr Index get(const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::get(i, *this);
+  }
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void set(const Index i, const Index value) {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::set(i, *this, value);
+  }
+
+  EIGEN_DEVICE_FUNC constexpr IndexList(const internal::IndexTuple<FirstType, OtherTypes...>& other) : internal::IndexTuple<FirstType, OtherTypes...>(other) { }
+  EIGEN_DEVICE_FUNC constexpr IndexList(FirstType& first, OtherTypes... other) : internal::IndexTuple<FirstType, OtherTypes...>(first, other...) { }
+  EIGEN_DEVICE_FUNC constexpr IndexList() : internal::IndexTuple<FirstType, OtherTypes...>() { }
+
+  EIGEN_DEVICE_FUNC constexpr bool value_known_statically(const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::value_known_statically(i, *this);
+  }
+  EIGEN_DEVICE_FUNC constexpr bool all_values_known_statically() const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::values_up_to_known_statically(*this);
+  }
+
+  EIGEN_DEVICE_FUNC constexpr bool values_statically_known_to_increase() const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::values_up_to_statically_known_to_increase(*this);
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+std::ostream& operator<<(std::ostream& os,
+                         const IndexList<FirstType, OtherTypes...>& dims) {
+  os << "[";
+  for (size_t i = 0; i < 1 + sizeof...(OtherTypes); ++i) {
+    if (i > 0) os << ", ";
+    os << dims[i];
+  }
+  os << "]";
+  return os;
+}
+
+template<typename FirstType, typename... OtherTypes>
+constexpr IndexList<FirstType, OtherTypes...> make_index_list(FirstType val1, OtherTypes... other_vals) {
+  return IndexList<FirstType, OtherTypes...>(val1, other_vals...);
+}
+
+
+template<typename FirstType, typename... OtherTypes>
+struct IndexPairList : internal::IndexTuple<FirstType, OtherTypes...> {
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC constexpr IndexPair<Index> operator[] (const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, IndexPair<Index>>::get(i, *this);
+  }
+  EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void set(const Index i, const IndexPair<Index> value) {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...>>::value-1, IndexPair<Index> >::set(i, *this, value);
+  }
+
+  EIGEN_DEVICE_FUNC  constexpr IndexPairList(const internal::IndexTuple<FirstType, OtherTypes...>& other) : internal::IndexTuple<FirstType, OtherTypes...>(other) { }
+  EIGEN_DEVICE_FUNC  constexpr IndexPairList() : internal::IndexTuple<FirstType, OtherTypes...>() { }
+
+  EIGEN_DEVICE_FUNC constexpr bool value_known_statically(const Index i) const {
+    return internal::tuple_coeff<internal::array_size<internal::IndexTuple<FirstType, OtherTypes...> >::value-1, Index>::value_known_statically(i, *this);
+  }
+};
+
+namespace internal {
+
+template<typename FirstType, typename... OtherTypes>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index array_prod(const IndexList<FirstType, OtherTypes...>& sizes) {
+  Index result = 1;
+  EIGEN_UNROLL_LOOP
+  for (size_t i = 0; i < array_size<IndexList<FirstType, OtherTypes...> >::value; ++i) {
+    result *= sizes[i];
+  }
+  return result;
+}
+
+template<typename FirstType, typename... OtherTypes> struct array_size<IndexList<FirstType, OtherTypes...> > {
+  static const size_t value = array_size<IndexTuple<FirstType, OtherTypes...> >::value;
+};
+template<typename FirstType, typename... OtherTypes> struct array_size<const IndexList<FirstType, OtherTypes...> > {
+  static const size_t value = array_size<IndexTuple<FirstType, OtherTypes...> >::value;
+};
+
+template<typename FirstType, typename... OtherTypes> struct array_size<IndexPairList<FirstType, OtherTypes...> > {
+  static const size_t value = std::tuple_size<std::tuple<FirstType, OtherTypes...> >::value;
+};
+template<typename FirstType, typename... OtherTypes> struct array_size<const IndexPairList<FirstType, OtherTypes...> > {
+  static const size_t value = std::tuple_size<std::tuple<FirstType, OtherTypes...> >::value;
+};
+
+template<Index N, typename FirstType, typename... OtherTypes> EIGEN_DEVICE_FUNC constexpr Index array_get(IndexList<FirstType, OtherTypes...>& a) {
+  return IndexTupleExtractor<N, FirstType, OtherTypes...>::get_val(a);
+}
+template<Index N, typename FirstType, typename... OtherTypes> EIGEN_DEVICE_FUNC constexpr Index array_get(const IndexList<FirstType, OtherTypes...>& a) {
+  return IndexTupleExtractor<N, FirstType, OtherTypes...>::get_val(a);
+}
+
+template <typename T>
+struct index_known_statically_impl {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index) {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_known_statically_impl<IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i);
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_known_statically_impl<const IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i);
+  }
+};
+
+
+template <typename T>
+struct all_indices_known_statically_impl {
+  static constexpr bool run() {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct all_indices_known_statically_impl<IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return IndexList<FirstType, OtherTypes...>().all_values_known_statically();
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct all_indices_known_statically_impl<const IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return IndexList<FirstType, OtherTypes...>().all_values_known_statically();
+  }
+};
+
+
+template <typename T>
+struct indices_statically_known_to_increase_impl {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+  struct indices_statically_known_to_increase_impl<IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return Eigen::IndexList<FirstType, OtherTypes...>().values_statically_known_to_increase();
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+  struct indices_statically_known_to_increase_impl<const IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run() {
+    return Eigen::IndexList<FirstType, OtherTypes...>().values_statically_known_to_increase();
+  }
+};
+
+
+template <typename Tx>
+struct index_statically_eq_impl {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_statically_eq_impl<IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexList<FirstType, OtherTypes...>().get(i) == value);
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_statically_eq_impl<const IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexList<FirstType, OtherTypes...>().get(i) == value);
+  }
+};
+
+
+template <typename T>
+struct index_statically_ne_impl {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_statically_ne_impl<IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexList<FirstType, OtherTypes...>().get(i) != value);
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_statically_ne_impl<const IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexList<FirstType, OtherTypes...>().get(i) != value);
+  }
+};
+
+
+template <typename T>
+struct index_statically_gt_impl {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_statically_gt_impl<IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexList<FirstType, OtherTypes...>().get(i) > value);
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_statically_gt_impl<const IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexList<FirstType, OtherTypes...>().get(i) > value);
+  }
+};
+
+
+
+template <typename T>
+struct index_statically_lt_impl {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_statically_lt_impl<IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexList<FirstType, OtherTypes...>().get(i) < value);
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_statically_lt_impl<const IndexList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexList<FirstType, OtherTypes...>().get(i) < value);
+  }
+};
+
+
+
+template <typename Tx>
+struct index_pair_first_statically_eq_impl {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_pair_first_statically_eq_impl<IndexPairList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexPairList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexPairList<FirstType, OtherTypes...>().operator[](i).first == value);
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_pair_first_statically_eq_impl<const IndexPairList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexPairList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexPairList<FirstType, OtherTypes...>().operator[](i).first == value);
+  }
+};
+
+
+
+template <typename Tx>
+struct index_pair_second_statically_eq_impl {
+  EIGEN_DEVICE_FUNC static constexpr bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_pair_second_statically_eq_impl<IndexPairList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexPairList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexPairList<FirstType, OtherTypes...>().operator[](i).second == value);
+  }
+};
+
+template <typename FirstType, typename... OtherTypes>
+struct index_pair_second_statically_eq_impl<const IndexPairList<FirstType, OtherTypes...> > {
+  EIGEN_DEVICE_FUNC static constexpr bool run(const Index i, const Index value) {
+    return IndexPairList<FirstType, OtherTypes...>().value_known_statically(i) &
+        (IndexPairList<FirstType, OtherTypes...>().operator[](i).second == value);
+  }
+};
+
+
+}  // end namespace internal
+}  // end namespace Eigen
+
+#else
+
+namespace Eigen {
+namespace internal {
+
+template <typename T>
+struct index_known_statically_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(const Index) {
+    return false;
+  }
+};
+
+template <typename T>
+struct all_indices_known_statically_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run() {
+    return false;
+  }
+};
+
+template <typename T>
+struct indices_statically_known_to_increase_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run() {
+    return false;
+  }
+};
+
+template <typename T>
+struct index_statically_eq_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename T>
+struct index_statically_ne_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename T>
+struct index_statically_gt_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename T>
+struct index_statically_lt_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename Tx>
+struct index_pair_first_statically_eq_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
+    return false;
+  }
+};
+
+template <typename Tx>
+struct index_pair_second_statically_eq_impl {
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool run(Index, Index) {
+    return false;
+  }
+};
+
+
+
+}  // end namespace internal
+}  // end namespace Eigen
+
+#endif
+
+
+namespace Eigen {
+namespace internal {
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_known_statically(Index i) {
+  return index_known_statically_impl<T>::run(i);
+}
+
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool all_indices_known_statically() {
+  return all_indices_known_statically_impl<T>::run();
+}
+
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool indices_statically_known_to_increase() {
+  return indices_statically_known_to_increase_impl<T>::run();
+}
+
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_eq(Index i, Index value) {
+  return index_statically_eq_impl<T>::run(i, value);
+}
+
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_ne(Index i, Index value) {
+  return index_statically_ne_impl<T>::run(i, value);
+}
+
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_gt(Index i, Index value) {
+  return index_statically_gt_impl<T>::run(i, value);
+}
+
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_statically_lt(Index i, Index value) {
+  return index_statically_lt_impl<T>::run(i, value);
+}
+
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_pair_first_statically_eq(Index i, Index value) {
+  return index_pair_first_statically_eq_impl<T>::run(i, value);
+}
+
+template <typename T>
+static EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bool index_pair_second_statically_eq(Index i, Index value) {
+  return index_pair_second_statically_eq_impl<T>::run(i, value);
+}
+
+}  // end namespace internal
+}  // end namespace Eigen
+
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_INDEX_LIST_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorInflation.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorInflation.h
new file mode 100644
index 0000000..c5cb61a
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorInflation.h
@@ -0,0 +1,247 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Ke Yang <yangke@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_INFLATION_H
+#define EIGEN_CXX11_TENSOR_TENSOR_INFLATION_H
+
+namespace Eigen {
+
+/** \class TensorInflation
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor inflation class.
+  *
+  *
+  */
+namespace internal {
+template<typename Strides, typename XprType>
+struct traits<TensorInflationOp<Strides, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename Strides, typename XprType>
+struct eval<TensorInflationOp<Strides, XprType>, Eigen::Dense>
+{
+  typedef const TensorInflationOp<Strides, XprType>& type;
+};
+
+template<typename Strides, typename XprType>
+struct nested<TensorInflationOp<Strides, XprType>, 1, typename eval<TensorInflationOp<Strides, XprType> >::type>
+{
+  typedef TensorInflationOp<Strides, XprType> type;
+};
+
+}  // end namespace internal
+
+template<typename Strides, typename XprType>
+class TensorInflationOp : public TensorBase<TensorInflationOp<Strides, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorInflationOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorInflationOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorInflationOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorInflationOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorInflationOp(const XprType& expr, const Strides& strides)
+      : m_xpr(expr), m_strides(strides) {}
+
+    EIGEN_DEVICE_FUNC
+    const Strides& strides() const { return m_strides; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Strides m_strides;
+};
+
+// Eval as rvalue
+template<typename Strides, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorInflationOp<Strides, ArgType>, Device>
+{
+  typedef TensorInflationOp<Strides, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_strides(op.strides())
+  {
+    m_dimensions = m_impl.dimensions();
+    // Expand each dimension to the inflated dimension.
+    for (int i = 0; i < NumDims; ++i) {
+      m_dimensions[i] = (m_dimensions[i] - 1) * op.strides()[i] + 1;
+    }
+
+    // Remember the strides for fast division.
+    for (int i = 0; i < NumDims; ++i) {
+      m_fastStrides[i] = internal::TensorIntDivisor<Index>(m_strides[i]);
+    }
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_outputStrides[0] = 1;
+      m_inputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
+        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+      }
+    } else {  // RowMajor
+      m_outputStrides[NumDims-1] = 1;
+      m_inputStrides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
+        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  // Computes the input index given the output index. Returns true if the output
+  // index doesn't fall into a hole.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool getInputIndex(Index index, Index* inputIndex) const
+  {
+    eigen_assert(index < dimensions().TotalSize());
+    *inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_outputStrides[i];
+        if (idx != idx / m_fastStrides[i] * m_strides[i]) {
+          return false;
+        }
+        *inputIndex += idx / m_strides[i] * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      if (index != index / m_fastStrides[0] * m_strides[0]) {
+        return false;
+      }
+      *inputIndex += index / m_strides[0];
+      return true;
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_outputStrides[i];
+        if (idx != idx / m_fastStrides[i] * m_strides[i]) {
+          return false;
+        }
+        *inputIndex += idx / m_strides[i] * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      if (index != index / m_fastStrides[NumDims-1] * m_strides[NumDims-1]) {
+        return false;
+      }
+      *inputIndex += index / m_strides[NumDims - 1];
+    }
+    return true;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    Index inputIndex = 0;
+    if (getInputIndex(index, &inputIndex)) {
+     return m_impl.coeff(inputIndex);
+    } else {
+     return Scalar(0);
+    }
+  }
+
+  // TODO(yangke): optimize this function so that we can detect and produce
+  // all-zero packets
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    const double compute_cost = NumDims * (3 * TensorOpCost::DivCost<Index>() +
+                                           3 * TensorOpCost::MulCost<Index>() +
+                                           2 * TensorOpCost::AddCost<Index>());
+    const double input_size = m_impl.dimensions().TotalSize();
+    const double output_size = m_dimensions.TotalSize();
+    if (output_size == 0)
+      return TensorOpCost();
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(sizeof(CoeffReturnType) * input_size / output_size, 0,
+                        compute_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ protected:
+  Dimensions m_dimensions;
+  array<Index, NumDims> m_outputStrides;
+  array<Index, NumDims> m_inputStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+  const Strides m_strides;
+  array<internal::TensorIntDivisor<Index>, NumDims> m_fastStrides;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_INFLATION_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorInitializer.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorInitializer.h
new file mode 100644
index 0000000..26a3818
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorInitializer.h
@@ -0,0 +1,82 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_INITIALIZER_H
+#define EIGEN_CXX11_TENSOR_TENSOR_INITIALIZER_H
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+
+#include <initializer_list>
+
+namespace Eigen {
+
+/** \class TensorInitializer
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Helper template to initialize Tensors from std::initializer_lists.
+  */
+namespace internal {
+
+template <typename Derived, int N>
+struct Initializer {
+  typedef std::initializer_list<
+    typename Initializer<Derived, N - 1>::InitList> InitList;
+
+  static void run(TensorEvaluator<Derived, DefaultDevice>& tensor,
+                  Eigen::array<typename traits<Derived>::Index, traits<Derived>::NumDimensions>* indices,
+                  const InitList& vals) {
+    int i = 0;
+    for (const auto& v : vals) {
+      (*indices)[traits<Derived>::NumDimensions - N] = i++;
+      Initializer<Derived, N - 1>::run(tensor, indices, v);
+    }
+  }
+};
+
+template <typename Derived>
+struct Initializer<Derived, 1> {
+  typedef std::initializer_list<typename traits<Derived>::Scalar> InitList;
+
+  static void run(TensorEvaluator<Derived, DefaultDevice>& tensor,
+                  Eigen::array<typename traits<Derived>::Index, traits<Derived>::NumDimensions>* indices,
+                  const InitList& vals) {
+    int i = 0;
+    // There is likely a faster way to do that than iterating.
+    for (const auto& v : vals) {
+      (*indices)[traits<Derived>::NumDimensions - 1] = i++;
+      tensor.coeffRef(*indices) = v;
+    }
+  }
+};
+
+template <typename Derived>
+struct Initializer<Derived, 0> {
+  typedef typename traits<Derived>::Scalar InitList;
+
+  static void run(TensorEvaluator<Derived, DefaultDevice>& tensor,
+                  Eigen::array<typename traits<Derived>::Index, traits<Derived>::NumDimensions>*,
+                  const InitList& v) {
+    tensor.coeffRef(0) = v;
+  }
+};
+
+
+template <typename Derived, int N>
+void initialize_tensor(TensorEvaluator<Derived, DefaultDevice>& tensor,
+                       const typename Initializer<Derived, traits<Derived>::NumDimensions>::InitList& vals) {
+  Eigen::array<typename traits<Derived>::Index, traits<Derived>::NumDimensions> indices;
+  Initializer<Derived, traits<Derived>::NumDimensions>::run(tensor, &indices, vals);
+}
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_HAS_VARIADIC_TEMPLATES
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_INITIALIZER_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorIntDiv.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorIntDiv.h
new file mode 100644
index 0000000..6d5cce4
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorIntDiv.h
@@ -0,0 +1,263 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H
+#define EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H
+
+
+namespace Eigen {
+
+/** \internal
+  *
+  * \class TensorIntDiv
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Fast integer division by a constant.
+  *
+  * See the paper from Granlund and Montgomery for explanation.
+  *   (at https://doi.org/10.1145/773473.178249)
+  *
+  * \sa Tensor
+  */
+
+namespace internal {
+
+namespace {
+
+  // Note: result is undefined if val == 0
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  typename internal::enable_if<sizeof(T)==4,int>::type count_leading_zeros(const T val)
+  {
+#ifdef EIGEN_GPU_COMPILE_PHASE
+    return __clz(val);
+#elif defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::clz(val);
+#elif EIGEN_COMP_MSVC
+    unsigned long index;
+    _BitScanReverse(&index, val);
+    return 31 - index;
+#else
+    EIGEN_STATIC_ASSERT(sizeof(unsigned long long) == 8, YOU_MADE_A_PROGRAMMING_MISTAKE);
+    return __builtin_clz(static_cast<uint32_t>(val));
+#endif
+  }
+
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  typename internal::enable_if<sizeof(T)==8,int>::type count_leading_zeros(const T val)
+  {
+#ifdef EIGEN_GPU_COMPILE_PHASE
+    return __clzll(val);
+#elif defined(SYCL_DEVICE_ONLY)
+    return static_cast<int>(cl::sycl::clz(val));
+#elif EIGEN_COMP_MSVC && EIGEN_ARCH_x86_64
+    unsigned long index;
+    _BitScanReverse64(&index, val);
+    return 63 - index;
+#elif EIGEN_COMP_MSVC
+    // MSVC's _BitScanReverse64 is not available for 32bits builds.
+    unsigned int lo = (unsigned int)(val&0xffffffff);
+    unsigned int hi = (unsigned int)((val>>32)&0xffffffff);
+    int n;
+    if(hi==0)
+      n = 32 + count_leading_zeros<unsigned int>(lo);
+    else
+      n = count_leading_zeros<unsigned int>(hi);
+    return n;
+#else
+    EIGEN_STATIC_ASSERT(sizeof(unsigned long long) == 8, YOU_MADE_A_PROGRAMMING_MISTAKE);
+    return __builtin_clzll(static_cast<uint64_t>(val));
+#endif
+  }
+
+  template <typename T>
+  struct UnsignedTraits {
+    typedef typename conditional<sizeof(T) == 8, uint64_t, uint32_t>::type type;
+  };
+
+  template <typename T>
+  struct DividerTraits {
+    typedef typename UnsignedTraits<T>::type type;
+    static const int N = sizeof(T) * 8;
+  };
+
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t muluh(const uint32_t a, const T b) {
+#if defined(EIGEN_GPU_COMPILE_PHASE)
+    return __umulhi(a, b);
+#elif defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::mul_hi(a, static_cast<uint32_t>(b));
+#else
+    return (static_cast<uint64_t>(a) * b) >> 32;
+#endif
+  }
+
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
+#if defined(EIGEN_GPU_COMPILE_PHASE)
+    return __umul64hi(a, b);
+#elif defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::mul_hi(a, static_cast<uint64_t>(b));
+#elif EIGEN_HAS_BUILTIN_INT128
+    __uint128_t v = static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b);
+    return static_cast<uint64_t>(v >> 64);
+#else
+    return (TensorUInt128<static_val<0>, uint64_t>(a) * TensorUInt128<static_val<0>, uint64_t>(b)).upper();
+#endif
+  }
+
+  template <int N, typename T>
+  struct DividerHelper {
+    static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t computeMultiplier(const int log_div, const T divider) {
+      EIGEN_STATIC_ASSERT(N == 32, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return static_cast<uint32_t>((static_cast<uint64_t>(1) << (N+log_div)) / divider - (static_cast<uint64_t>(1) << N) + 1);
+    }
+  };
+
+  template <typename T>
+  struct DividerHelper<64, T> {
+    static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
+#if EIGEN_HAS_BUILTIN_INT128 && !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY)
+      return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1);
+#else
+      const uint64_t shift = 1ULL << log_div;
+      TensorUInt128<uint64_t, uint64_t> result = TensorUInt128<uint64_t, static_val<0> >(shift, 0) / TensorUInt128<static_val<0>, uint64_t>(divider)
+                                               - TensorUInt128<static_val<1>, static_val<0> >(1, 0)
+                                               + TensorUInt128<static_val<0>, static_val<1> >(1);
+      return static_cast<uint64_t>(result);
+#endif
+    }
+  };
+}
+
+
+template <typename T, bool div_gt_one = false>
+struct TensorIntDivisor {
+ public:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() {
+    multiplier = 0;
+    shift1 = 0;
+    shift2 = 0;
+  }
+
+  // Must have 0 < divider < 2^31. This is relaxed to
+  // 0 < divider < 2^63 when using 64-bit indices on platforms that support
+  // the __uint128_t type.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor(const T divider) {
+    const int N = DividerTraits<T>::N;
+    eigen_assert(static_cast<typename UnsignedTraits<T>::type>(divider) < NumTraits<UnsignedType>::highest()/2);
+    eigen_assert(divider > 0);
+
+    // fast ln2
+    const int leading_zeros = count_leading_zeros(static_cast<UnsignedType>(divider));
+    int log_div = N - leading_zeros;
+    // if divider is a power of two then log_div is 1 more than it should be.
+    if ((static_cast<typename UnsignedTraits<T>::type>(1) << (log_div-1)) == static_cast<typename UnsignedTraits<T>::type>(divider))
+      log_div--;
+
+    multiplier = DividerHelper<N, T>::computeMultiplier(log_div, divider);
+    shift1 = log_div > 1 ? 1 : log_div;
+    shift2 = log_div > 1 ? log_div-1 : 0;
+  }
+
+  // Must have 0 <= numerator. On platforms that don't support the __uint128_t
+  // type numerator should also be less than 2^32-1.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T divide(const T numerator) const {
+    eigen_assert(static_cast<typename UnsignedTraits<T>::type>(numerator) < NumTraits<UnsignedType>::highest()/2);
+    //eigen_assert(numerator >= 0); // this is implicitly asserted by the line above
+
+    UnsignedType t1 = muluh(multiplier, numerator);
+    UnsignedType t = (static_cast<UnsignedType>(numerator) - t1) >> shift1;
+    return (t1 + t) >> shift2;
+  }
+
+ private:
+  typedef typename DividerTraits<T>::type UnsignedType;
+  UnsignedType multiplier;
+  int32_t shift1;
+  int32_t shift2;
+};
+
+
+// Optimized version for signed 32 bit integers.
+// Derived from Hacker's Delight.
+// Only works for divisors strictly greater than one
+template <>
+class TensorIntDivisor<int32_t, true> {
+ public:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() {
+    magic = 0;
+    shift = 0;
+  }
+  // Must have 2 <= divider
+  EIGEN_DEVICE_FUNC TensorIntDivisor(int32_t divider)  {
+    eigen_assert(divider >= 2);
+    calcMagic(divider);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE int divide(const int32_t n) const {
+#ifdef EIGEN_GPU_COMPILE_PHASE
+    return (__umulhi(magic, n) >> shift);
+#elif defined(SYCL_DEVICE_ONLY)
+    return (cl::sycl::mul_hi(magic, static_cast<uint32_t>(n)) >> shift);
+#else
+    uint64_t v = static_cast<uint64_t>(magic) * static_cast<uint64_t>(n);
+    return (static_cast<uint32_t>(v >> 32) >> shift);
+#endif
+  }
+
+private:
+  // Compute the magic numbers. See Hacker's Delight section 10 for an in
+  // depth explanation.
+  EIGEN_DEVICE_FUNC void calcMagic(int32_t d) {
+   const unsigned two31 = 0x80000000;     // 2**31.
+   unsigned ad = d;
+   unsigned t = two31 + (ad >> 31);
+   unsigned anc = t - 1 - t%ad;     // Absolute value of nc.
+   int p = 31;                      // Init. p.
+   unsigned q1 = two31/anc;         // Init. q1 = 2**p/|nc|.
+   unsigned r1 = two31 - q1*anc;    // Init. r1 = rem(2**p, |nc|).
+   unsigned q2 = two31/ad;          // Init. q2 = 2**p/|d|.
+   unsigned r2 = two31 - q2*ad;     // Init. r2 = rem(2**p, |d|).
+   unsigned delta = 0;
+   do {
+      p = p + 1;
+      q1 = 2*q1;           // Update q1 = 2**p/|nc|.
+      r1 = 2*r1;           // Update r1 = rem(2**p, |nc|).
+      if (r1 >= anc) {     // (Must be an unsigned
+         q1 = q1 + 1;      // comparison here).
+         r1 = r1 - anc;}
+      q2 = 2*q2;           // Update q2 = 2**p/|d|.
+      r2 = 2*r2;           // Update r2 = rem(2**p, |d|).
+      if (r2 >= ad) {      // (Must be an unsigned
+         q2 = q2 + 1;      // comparison here).
+         r2 = r2 - ad;}
+      delta = ad - r2;
+   } while (q1 < delta || (q1 == delta && r1 == 0));
+
+   magic = (unsigned)(q2 + 1);
+   shift = p - 32;
+  }
+
+  uint32_t magic;
+  int32_t shift;
+};
+
+
+template <typename T, bool div_gt_one>
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator / (const T& numerator, const TensorIntDivisor<T, div_gt_one>& divisor) {
+  return divisor.divide(numerator);
+}
+
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorLayoutSwap.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorLayoutSwap.h
new file mode 100644
index 0000000..80106c1
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorLayoutSwap.h
@@ -0,0 +1,216 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_LAYOUT_SWAP_H
+#define EIGEN_CXX11_TENSOR_TENSOR_LAYOUT_SWAP_H
+
+namespace Eigen {
+
+/** \class TensorLayoutSwap
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Swap the layout from col-major to row-major, or row-major
+  * to col-major, and invert the order of the dimensions.
+  *
+  * Beware: the dimensions are reversed by this operation. If you want to
+  * preserve the ordering of the dimensions, you need to combine this
+  * operation with a shuffle.
+  *
+  * \example:
+  * Tensor<float, 2, ColMajor> input(2, 4);
+  * Tensor<float, 2, RowMajor> output = input.swap_layout();
+  * eigen_assert(output.dimension(0) == 4);
+  * eigen_assert(output.dimension(1) == 2);
+  *
+  * array<int, 2> shuffle(1, 0);
+  * output = input.swap_layout().shuffle(shuffle);
+  * eigen_assert(output.dimension(0) == 2);
+  * eigen_assert(output.dimension(1) == 4);
+  *
+  */
+namespace internal {
+template<typename XprType>
+struct traits<TensorLayoutSwapOp<XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = traits<XprType>::NumDimensions;
+  static const int Layout = (traits<XprType>::Layout == ColMajor) ? RowMajor : ColMajor;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename XprType>
+struct eval<TensorLayoutSwapOp<XprType>, Eigen::Dense>
+{
+  typedef const TensorLayoutSwapOp<XprType>& type;
+};
+
+template<typename XprType>
+struct nested<TensorLayoutSwapOp<XprType>, 1, typename eval<TensorLayoutSwapOp<XprType> >::type>
+{
+  typedef TensorLayoutSwapOp<XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename XprType>
+class TensorLayoutSwapOp : public TensorBase<TensorLayoutSwapOp<XprType>, WriteAccessors>
+{
+  public:
+    typedef TensorBase<TensorLayoutSwapOp<XprType>, WriteAccessors> Base;
+    typedef typename Eigen::internal::traits<TensorLayoutSwapOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorLayoutSwapOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorLayoutSwapOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorLayoutSwapOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorLayoutSwapOp(const XprType& expr)
+        : m_xpr(expr) {}
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorLayoutSwapOp)
+  protected:
+    typename XprType::Nested m_xpr;
+};
+
+
+// Eval as rvalue
+template<typename ArgType, typename Device>
+struct TensorEvaluator<const TensorLayoutSwapOp<ArgType>, Device>
+{
+  typedef TensorLayoutSwapOp<ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  enum {
+    IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = (static_cast<int>(TensorEvaluator<ArgType, Device>::Layout) == static_cast<int>(ColMajor)) ? RowMajor : ColMajor,
+    CoordAccess = false,  // to be implemented
+    RawAccess = TensorEvaluator<ArgType, Device>::RawAccess
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device)
+  {
+    for(int i = 0; i < NumDims; ++i) {
+      m_dimensions[i] = m_impl.dimensions()[NumDims-1-i];
+    }
+  }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    return m_impl.evalSubExprsIfNeeded(data);
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return m_impl.coeff(index);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return m_impl.template packet<LoadMode>(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    return m_impl.costPerCoeff(vectorized);
+  }
+
+  EIGEN_DEVICE_FUNC typename Storage::Type data() const {
+    return constCast(m_impl.data());
+  }
+
+  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+
+ protected:
+  TensorEvaluator<ArgType, Device> m_impl;
+  Dimensions m_dimensions;
+};
+
+
+// Eval as lvalue
+template<typename ArgType, typename Device>
+  struct TensorEvaluator<TensorLayoutSwapOp<ArgType>, Device>
+  : public TensorEvaluator<const TensorLayoutSwapOp<ArgType>, Device>
+{
+  typedef TensorEvaluator<const TensorLayoutSwapOp<ArgType>, Device> Base;
+  typedef TensorLayoutSwapOp<ArgType> XprType;
+
+  enum {
+    IsAligned = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = (static_cast<int>(TensorEvaluator<ArgType, Device>::Layout) == static_cast<int>(ColMajor)) ? RowMajor : ColMajor,
+    CoordAccess = false  // to be implemented
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : Base(op, device)
+  { }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
+  {
+    return this->m_impl.coeffRef(index);
+  }
+  template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x)
+  {
+    this->m_impl.template writePacket<StoreMode>(index, x);
+  }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_LAYOUT_SWAP_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorMacros.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorMacros.h
new file mode 100644
index 0000000..73ff3d2
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorMacros.h
@@ -0,0 +1,98 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_META_MACROS_H
+#define EIGEN_CXX11_TENSOR_TENSOR_META_MACROS_H
+
+
+/** use this macro in sfinae selection in templated functions
+ *
+ *   template<typename T,
+ *            typename std::enable_if< isBanana<T>::value , int >::type = 0
+ *   >
+ *   void foo(){}
+ *
+ *   becomes =>
+ *
+ *   template<typename TopoType,
+ *           SFINAE_ENABLE_IF( isBanana<T>::value )
+ *   >
+ *   void foo(){}
+ */
+
+// SFINAE requires variadic templates
+#if !defined(EIGEN_GPUCC)
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+  // SFINAE doesn't work for gcc <= 4.7
+  #ifdef EIGEN_COMP_GNUC
+    #if EIGEN_GNUC_AT_LEAST(4,8)
+      #define EIGEN_HAS_SFINAE
+    #endif
+  #else
+    #define EIGEN_HAS_SFINAE
+  #endif
+#endif
+#endif
+
+#define EIGEN_SFINAE_ENABLE_IF( __condition__ ) \
+    typename internal::enable_if< ( __condition__ ) , int >::type = 0
+
+// Define a macro to use a reference on the host but a value on the device
+#if defined(SYCL_DEVICE_ONLY)
+  #define EIGEN_DEVICE_REF
+#else
+  #define EIGEN_DEVICE_REF &
+#endif
+
+// Define a macro for catching SYCL exceptions if exceptions are enabled
+#define EIGEN_SYCL_TRY_CATCH(X) \
+  do { \
+    EIGEN_TRY {X;} \
+    EIGEN_CATCH(const cl::sycl::exception& e) { \
+      EIGEN_THROW_X(std::runtime_error("SYCL exception at " + \
+                                       std::string(__FILE__) + ":" + \
+                                       std::to_string(__LINE__) + "\n" + \
+                                       e.what())); \
+    } \
+  } while (false)
+
+// Define a macro if local memory flags are unset or one of them is set
+// Setting both flags is the same as unsetting them
+#if (!defined(EIGEN_SYCL_LOCAL_MEM) && !defined(EIGEN_SYCL_NO_LOCAL_MEM)) || \
+     (defined(EIGEN_SYCL_LOCAL_MEM) &&  defined(EIGEN_SYCL_NO_LOCAL_MEM))
+  #define EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON 1
+  #define EIGEN_SYCL_LOCAL_MEM_UNSET_OR_OFF 1
+#elif defined(EIGEN_SYCL_LOCAL_MEM) && !defined(EIGEN_SYCL_NO_LOCAL_MEM)
+  #define EIGEN_SYCL_LOCAL_MEM_UNSET_OR_ON 1
+#elif !defined(EIGEN_SYCL_LOCAL_MEM) && defined(EIGEN_SYCL_NO_LOCAL_MEM)
+  #define EIGEN_SYCL_LOCAL_MEM_UNSET_OR_OFF 1
+#endif
+
+#if EIGEN_COMP_CLANG // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653)
+  #define EIGEN_TENSOR_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    using Base::operator =; \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) { Base::operator=(other); return *this; } \
+    template <typename OtherDerived> \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const OtherDerived& other) { Base::operator=(other); return *this; }
+#else
+  #define EIGEN_TENSOR_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived)
+#endif
+
+/** \internal
+ * \brief Macro to manually inherit assignment operators.
+ * This is necessary, because the implicitly defined assignment operator gets deleted when a custom operator= is defined.
+ * This also inherits template<OtherDerived> operator=(const OtherDerived&) assignments.
+ * With C++11 or later this also default-implements the copy-constructor
+ */
+#define EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(Derived)  \
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    EIGEN_DEFAULT_COPY_CONSTRUCTOR(Derived)
+
+#endif
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorMap.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorMap.h
new file mode 100644
index 0000000..6834c97
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorMap.h
@@ -0,0 +1,327 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_MAP_H
+#define EIGEN_CXX11_TENSOR_TENSOR_MAP_H
+
+namespace Eigen {
+
+// FIXME use proper doxygen documentation (e.g. \tparam MakePointer_)
+
+/** \class TensorMap
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief A tensor expression mapping an existing array of data.
+  *
+  */
+/// `template <class> class MakePointer_` is added to convert the host pointer to the device pointer.
+/// It is added due to the fact that for our device compiler `T*` is not allowed.
+/// If we wanted to use the same Evaluator functions we have to convert that type to our pointer `T`.
+/// This is done through our `MakePointer_` class. By default the Type in the `MakePointer_<T>` is `T*` .
+/// Therefore, by adding the default value, we managed to convert the type and it does not break any
+/// existing code as its default value is `T*`.
+template<typename PlainObjectType, int Options_, template <class> class MakePointer_> class TensorMap : public TensorBase<TensorMap<PlainObjectType, Options_, MakePointer_> >
+{
+  public:
+    typedef TensorMap<PlainObjectType, Options_, MakePointer_> Self;
+    typedef TensorBase<TensorMap<PlainObjectType, Options_, MakePointer_> > Base;
+  #ifdef EIGEN_USE_SYCL
+    typedef  typename Eigen::internal::remove_reference<typename Eigen::internal::nested<Self>::type>::type Nested;
+  #else
+     typedef typename Eigen::internal::nested<Self>::type Nested;
+  #endif
+   typedef typename internal::traits<PlainObjectType>::StorageKind StorageKind;
+    typedef typename internal::traits<PlainObjectType>::Index Index;
+    typedef typename internal::traits<PlainObjectType>::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef typename PlainObjectType::Base::CoeffReturnType CoeffReturnType;
+
+    typedef typename MakePointer_<Scalar>::Type PointerType;
+    typedef typename MakePointer_<Scalar>::ConstType PointerConstType;
+
+    // WARN: PointerType still can be a pointer to const (const Scalar*), for
+    // example in TensorMap<Tensor<const Scalar, ...>> expression. This type of
+    // expression should be illegal, but adding this restriction is not possible
+    // in practice (see https://bitbucket.org/eigen/eigen/pull-requests/488).
+    typedef typename internal::conditional<
+        bool(internal::is_lvalue<PlainObjectType>::value),
+        PointerType,      // use simple pointer in lvalue expressions
+        PointerConstType  // use const pointer in rvalue expressions
+        >::type StoragePointerType;
+
+    // If TensorMap was constructed over rvalue expression (e.g. const Tensor),
+    // we should return a reference to const from operator() (and others), even
+    // if TensorMap itself is not const.
+    typedef typename internal::conditional<
+        bool(internal::is_lvalue<PlainObjectType>::value),
+        Scalar&,
+        const Scalar&
+        >::type StorageRefType;
+
+    static const int Options = Options_;
+
+    static const Index NumIndices = PlainObjectType::NumIndices;
+    typedef typename PlainObjectType::Dimensions Dimensions;
+
+    enum {
+      IsAligned = ((int(Options_)&Aligned)==Aligned),
+      Layout = PlainObjectType::Layout,
+      CoordAccess = true,
+      RawAccess = true
+    };
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr) : m_data(dataPtr), m_dimensions() {
+      // The number of dimensions used to construct a tensor must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT((0 == NumIndices || NumIndices == Dynamic), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions(firstDimension, otherDimensions...) {
+      // The number of dimensions used to construct a tensor must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT((sizeof...(otherDimensions) + 1 == NumIndices || NumIndices == Dynamic), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+#else
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr, Index firstDimension) : m_data(dataPtr), m_dimensions(firstDimension) {
+      // The number of dimensions used to construct a tensor must be equal to the rank of the tensor.
+      EIGEN_STATIC_ASSERT((1 == NumIndices || NumIndices == Dynamic), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr, Index dim1, Index dim2) : m_data(dataPtr), m_dimensions(dim1, dim2) {
+      EIGEN_STATIC_ASSERT(2 == NumIndices || NumIndices == Dynamic, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr, Index dim1, Index dim2, Index dim3) : m_data(dataPtr), m_dimensions(dim1, dim2, dim3) {
+      EIGEN_STATIC_ASSERT(3 == NumIndices || NumIndices == Dynamic, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr, Index dim1, Index dim2, Index dim3, Index dim4) : m_data(dataPtr), m_dimensions(dim1, dim2, dim3, dim4) {
+      EIGEN_STATIC_ASSERT(4 == NumIndices || NumIndices == Dynamic, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr, Index dim1, Index dim2, Index dim3, Index dim4, Index dim5) : m_data(dataPtr), m_dimensions(dim1, dim2, dim3, dim4, dim5) {
+      EIGEN_STATIC_ASSERT(5 == NumIndices || NumIndices == Dynamic, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    }
+#endif
+
+   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr, const array<Index, NumIndices>& dimensions)
+      : m_data(dataPtr), m_dimensions(dimensions)
+    { }
+
+    template <typename Dimensions>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorMap(StoragePointerType dataPtr, const Dimensions& dimensions)
+      : m_data(dataPtr), m_dimensions(dimensions)
+    { }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorMap(PlainObjectType& tensor)
+      : m_data(tensor.data()), m_dimensions(tensor.dimensions())
+    { }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index rank() const { return m_dimensions.rank(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index dimension(Index n) const { return m_dimensions[n]; }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index size() const { return m_dimensions.TotalSize(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StoragePointerType data() { return m_data; }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StoragePointerType data() const { return m_data; }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(const array<Index, NumIndices>& indices) const
+    {
+      //      eigen_assert(checkIndexRange(indices));
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = m_dimensions.IndexOfRowMajor(indices);
+        return m_data[index];
+      } else {
+        const Index index = m_dimensions.IndexOfColMajor(indices);
+        return m_data[index];
+      }
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()() const
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return m_data[0];
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index index) const
+    {
+      eigen_internal_assert(index >= 0 && index < size());
+      return m_data[index];
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const
+    {
+      EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      eigen_assert(internal::all((Eigen::NumTraits<Index>::highest() >= otherIndices)...));
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = m_dimensions.IndexOfRowMajor(array<Index, NumIndices>{{firstIndex, secondIndex, otherIndices...}});
+        return m_data[index];
+      } else {
+        const Index index = m_dimensions.IndexOfColMajor(array<Index, NumIndices>{{firstIndex, secondIndex, otherIndices...}});
+        return m_data[index];
+      }
+    }
+#else
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index i0, Index i1) const
+    {
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = i1 + i0 * m_dimensions[1];
+        return m_data[index];
+      } else {
+        const Index index = i0 + i1 * m_dimensions[0];
+        return m_data[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index i0, Index i1, Index i2) const
+    {
+      if (PlainObjectType::Options&RowMajor) {
+         const Index index = i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0);
+         return m_data[index];
+      } else {
+         const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * i2);
+        return m_data[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index i0, Index i1, Index i2, Index i3) const
+    {
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = i3 + m_dimensions[3] * (i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0));
+        return m_data[index];
+      } else {
+        const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * (i2 + m_dimensions[2] * i3));
+        return m_data[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index i0, Index i1, Index i2, Index i3, Index i4) const
+    {
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = i4 + m_dimensions[4] * (i3 + m_dimensions[3] * (i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0)));
+        return m_data[index];
+      } else {
+        const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * (i2 + m_dimensions[2] * (i3 + m_dimensions[3] * i4)));
+        return m_data[index];
+      }
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(const array<Index, NumIndices>& indices)
+    {
+      //      eigen_assert(checkIndexRange(indices));
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = m_dimensions.IndexOfRowMajor(indices);
+        return m_data[index];
+      } else {
+        const Index index = m_dimensions.IndexOfColMajor(indices);
+        return m_data[index];
+      }
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()()
+    {
+      EIGEN_STATIC_ASSERT(NumIndices == 0, YOU_MADE_A_PROGRAMMING_MISTAKE)
+      return m_data[0];
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index index)
+    {
+      eigen_internal_assert(index >= 0 && index < size());
+      return m_data[index];
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices)
+    {
+      static_assert(sizeof...(otherIndices) + 2 == NumIndices || NumIndices == Dynamic, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor.");
+       eigen_assert(internal::all((Eigen::NumTraits<Index>::highest() >= otherIndices)...));
+      const std::size_t NumDims = sizeof...(otherIndices) + 2;
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = m_dimensions.IndexOfRowMajor(array<Index, NumDims>{{firstIndex, secondIndex, otherIndices...}});
+        return m_data[index];
+      } else {
+        const Index index = m_dimensions.IndexOfColMajor(array<Index, NumDims>{{firstIndex, secondIndex, otherIndices...}});
+        return m_data[index];
+      }
+    }
+#else
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index i0, Index i1)
+    {
+       if (PlainObjectType::Options&RowMajor) {
+         const Index index = i1 + i0 * m_dimensions[1];
+        return m_data[index];
+      } else {
+        const Index index = i0 + i1 * m_dimensions[0];
+        return m_data[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index i0, Index i1, Index i2)
+    {
+       if (PlainObjectType::Options&RowMajor) {
+         const Index index = i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0);
+        return m_data[index];
+      } else {
+         const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * i2);
+        return m_data[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index i0, Index i1, Index i2, Index i3)
+    {
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = i3 + m_dimensions[3] * (i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0));
+        return m_data[index];
+      } else {
+        const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * (i2 + m_dimensions[2] * i3));
+        return m_data[index];
+      }
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE StorageRefType operator()(Index i0, Index i1, Index i2, Index i3, Index i4)
+    {
+      if (PlainObjectType::Options&RowMajor) {
+        const Index index = i4 + m_dimensions[4] * (i3 + m_dimensions[3] * (i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0)));
+        return m_data[index];
+      } else {
+        const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * (i2 + m_dimensions[2] * (i3 + m_dimensions[3] * i4)));
+        return m_data[index];
+      }
+    }
+#endif
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorMap)
+
+  private:
+    StoragePointerType m_data;
+    Dimensions m_dimensions;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_MAP_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorMeta.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorMeta.h
new file mode 100644
index 0000000..a6181d3
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorMeta.h
@@ -0,0 +1,311 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_META_H
+#define EIGEN_CXX11_TENSOR_TENSOR_META_H
+
+namespace Eigen {
+
+template<bool cond> struct Cond {};
+
+template<typename T1, typename T2> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+const T1& choose(Cond<true>, const T1& first, const T2&) {
+  return first;
+}
+
+template<typename T1, typename T2> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+const T2& choose(Cond<false>, const T1&, const T2& second) {
+  return second;
+}
+
+
+template <typename T, typename X, typename Y>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T divup(const X x, const Y y) {
+  return static_cast<T>((x + y - 1) / y);
+}
+
+template <typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T divup(const T x, const T y) {
+  return static_cast<T>((x + y - 1) / y);
+}
+
+template <size_t n> struct max_n_1 {
+  static const size_t size = n;
+};
+template <> struct max_n_1<0> {
+  static const size_t size = 1;
+};
+
+
+// Default packet types
+template <typename Scalar, typename Device>
+struct PacketType : internal::packet_traits<Scalar> {
+  typedef typename internal::packet_traits<Scalar>::type type;
+};
+
+// For CUDA packet types when using a GpuDevice
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_HAS_GPU_FP16)
+
+typedef ulonglong2 Packet4h2;
+template<>
+struct PacketType<half, GpuDevice> {
+  typedef Packet4h2 type;
+  static const int size = 8;
+  enum {
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasNegate = 1,
+    HasAbs    = 1,
+    HasArg    = 0,
+    HasAbs2   = 0,
+    HasMin    = 1,
+    HasMax    = 1,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasBlend  = 0,
+
+    HasDiv    = 1,
+    HasSqrt   = 1,
+    HasRsqrt  = 1,
+    HasExp    = 1,
+    HasExpm1  = 0,
+    HasLog    = 1,
+    HasLog1p  = 0,
+    HasLog10  = 0,
+    HasPow    = 1,
+  };
+};
+#endif
+
+#if defined(EIGEN_USE_SYCL)
+
+namespace TensorSycl {
+namespace internal {
+
+template <typename Index, Index A, Index B> struct PlusOp {
+  static constexpr Index Value = A + B;
+};
+
+template <typename Index, Index A, Index B> struct DivOp {
+  static constexpr Index Value = A / B;
+};
+
+template <typename Index, Index start, Index end, Index step,
+          template <class Indx, Indx...> class StepOp>
+struct static_for {
+  template <typename UnaryOperator>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void loop(UnaryOperator op) {
+    op(start);
+    static_for<Index, StepOp<Index, start, step>::Value, end, step,
+               StepOp>::loop(op);
+  }
+};
+template <typename Index, Index end, Index step,
+          template <class Indx, Indx...> class StepOp>
+struct static_for<Index, end, end, step, StepOp> {
+  template <typename UnaryOperator>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void loop(UnaryOperator) {}
+};
+
+template <typename OutScalar, typename Device, bool Vectorizable>
+struct Vectorise {
+  static const int PacketSize = 1;
+  typedef OutScalar PacketReturnType;
+};
+
+template <typename OutScalar, typename Device>
+struct Vectorise<OutScalar, Device, true> {
+  static const int PacketSize = Eigen::PacketType<OutScalar, Device>::size;
+  typedef typename Eigen::PacketType<OutScalar, Device>::type PacketReturnType;
+};
+
+static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Index roundUp(Index x, Index y) {
+  return ((((x) + (y)-1) / (y)) * (y));
+}
+
+} // namespace internal
+} // namespace TensorSycl
+
+template <>
+  struct PacketType<half, SyclDevice> {
+  typedef half type;
+  static const int size = 1;
+  enum {
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasArg    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasBlend  = 0
+  };
+};
+template <typename Scalar>
+struct PacketType<Scalar, SyclDevice> : internal::default_packet_traits {
+  typedef Scalar type;
+  typedef Scalar half;
+  enum {
+    Vectorizable = 0,
+    size = 1,
+    AlignedOnScalar = 0,
+    HasHalfPacket = 0
+  };
+  enum {
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0
+  };
+
+};
+
+template <typename Scalar>
+struct PacketType<Scalar, const SyclDevice> : PacketType<Scalar, SyclDevice>{};
+
+#ifndef EIGEN_DONT_VECTORIZE_SYCL
+#define PACKET_TYPE(CVQual, Type, val, lengths, DEV)\
+template<> struct PacketType<CVQual Type, DEV> : internal::sycl_packet_traits<val, lengths> \
+{\
+  typedef typename internal::packet_traits<Type>::type type;\
+  typedef typename internal::packet_traits<Type>::half half;\
+};
+
+
+PACKET_TYPE(const, float, 1, 4, SyclDevice)
+PACKET_TYPE(, float, 1, 4, SyclDevice)
+PACKET_TYPE(const, float, 1, 4, const SyclDevice)
+PACKET_TYPE(, float, 1, 4, const SyclDevice)
+
+PACKET_TYPE(const, double, 0, 2, SyclDevice)
+PACKET_TYPE(, double, 0, 2, SyclDevice)
+PACKET_TYPE(const, double, 0, 2, const SyclDevice)
+PACKET_TYPE(, double, 0, 2, const SyclDevice)
+#undef PACKET_TYPE
+
+template<> struct PacketType<half, const SyclDevice>: PacketType<half, SyclDevice>{};
+template<> struct PacketType<const half, const SyclDevice>: PacketType<half, SyclDevice>{};
+#endif
+#endif
+
+// Tuple mimics std::pair but works on e.g. nvcc.
+template <typename U, typename V> struct Tuple {
+ public:
+  U first;
+  V second;
+
+  typedef U first_type;
+  typedef V second_type;
+
+  EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Tuple() : first(), second() {}
+
+  EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Tuple(const U& f, const V& s) : first(f), second(s) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void swap(Tuple& rhs) {
+    using numext::swap;
+    swap(first, rhs.first);
+    swap(second, rhs.second);
+  }
+};
+
+template <typename U, typename V>
+EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+bool operator==(const Tuple<U, V>& x, const Tuple<U, V>& y) {
+  return (x.first == y.first && x.second == y.second);
+}
+
+template <typename U, typename V>
+EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+bool operator!=(const Tuple<U, V>& x, const Tuple<U, V>& y) {
+  return !(x == y);
+}
+
+
+// Can't use std::pairs on cuda devices
+template <typename Idx> struct IndexPair {
+  EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexPair() : first(0), second(0) {}
+  EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexPair(Idx f, Idx s) : first(f), second(s) {}
+
+  EIGEN_DEVICE_FUNC void set(IndexPair<Idx> val) {
+    first = val.first;
+    second = val.second;
+  }
+
+  Idx first;
+  Idx second;
+};
+
+
+#ifdef EIGEN_HAS_SFINAE
+namespace internal {
+
+  template<typename IndexType, typename Index, Index... Is>
+  EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  array<Index, sizeof...(Is)> customIndices2Array(IndexType& idx, numeric_list<Index, Is...>) {
+    return { idx[Is]... };
+  }
+  template<typename IndexType, typename Index>
+  EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  array<Index, 0> customIndices2Array(IndexType&, numeric_list<Index>) {
+    return array<Index, 0>();
+  }
+
+  /** Make an array (for index/dimensions) out of a custom index */
+  template<typename Index, std::size_t NumIndices, typename IndexType>
+  EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  array<Index, NumIndices> customIndices2Array(IndexType& idx) {
+    return customIndices2Array(idx, typename gen_numeric_list<Index, NumIndices>::type{});
+  }
+
+
+  template <typename B, typename D>
+  struct is_base_of
+  {
+
+    typedef char (&yes)[1];
+    typedef char (&no)[2];
+
+    template <typename BB, typename DD>
+    struct Host
+    {
+      operator BB*() const;
+      operator DD*();
+    };
+
+    template<typename T>
+    static yes check(D*, T);
+    static no check(B*, int);
+
+    static const bool value = sizeof(check(Host<B,D>(), int())) == sizeof(yes);
+  };
+
+}
+#endif
+
+
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_META_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorMorphing.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorMorphing.h
new file mode 100644
index 0000000..b3f00f7
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorMorphing.h
@@ -0,0 +1,1102 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_MORPHING_H
+#define EIGEN_CXX11_TENSOR_TENSOR_MORPHING_H
+
+namespace Eigen {
+
+/** \class TensorReshaping
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor reshaping class.
+  *
+  *
+  */
+namespace internal {
+template<typename NewDimensions, typename XprType>
+struct traits<TensorReshapingOp<NewDimensions, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = array_size<NewDimensions>::value;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename NewDimensions, typename XprType>
+struct eval<TensorReshapingOp<NewDimensions, XprType>, Eigen::Dense>
+{
+  typedef const TensorReshapingOp<NewDimensions, XprType>EIGEN_DEVICE_REF type;
+};
+
+template<typename NewDimensions, typename XprType>
+struct nested<TensorReshapingOp<NewDimensions, XprType>, 1, typename eval<TensorReshapingOp<NewDimensions, XprType> >::type>
+{
+  typedef TensorReshapingOp<NewDimensions, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename NewDimensions, typename XprType>
+class TensorReshapingOp : public TensorBase<TensorReshapingOp<NewDimensions, XprType>, WriteAccessors>
+{
+  public:
+  typedef TensorBase<TensorReshapingOp<NewDimensions, XprType>, WriteAccessors> Base;
+  typedef typename Eigen::internal::traits<TensorReshapingOp>::Scalar Scalar;
+  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorReshapingOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorReshapingOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorReshapingOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorReshapingOp(const XprType& expr, const NewDimensions& dims)
+      : m_xpr(expr), m_dims(dims) {}
+
+    EIGEN_DEVICE_FUNC
+    const NewDimensions& dimensions() const { return m_dims; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorReshapingOp)
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const NewDimensions m_dims;
+};
+
+
+// Eval as rvalue
+template<typename NewDimensions, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
+{
+  typedef TensorReshapingOp<NewDimensions, ArgType> XprType;
+  typedef NewDimensions Dimensions;
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  typedef StorageMemory<typename internal::remove_const<CoeffReturnType>::type, Device> ConstCastStorage;
+
+  static const int NumOutputDims = internal::array_size<Dimensions>::value;
+  static const int NumInputDims  = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+
+  enum ReshapingKind {
+    // We do not use layout information to determine reshaping kind.
+    // Depending on the layout `N` can be inner or outer dimension.
+    OneByN = 0,  // expr.reshape(1, N)
+    NByOne = 1,  // expr.reshape(N, 1)
+    Runtime = 2  // Reshape dimensions are dynamic (specified at runtime).
+  };
+
+  // clang-format off
+  static const ReshapingKind kind =
+#if defined(EIGEN_HAS_INDEX_LIST)
+        (NumOutputDims == 2 && internal::index_statically_eq<NewDimensions>(/*index=*/0, /*value=*/1)) ? OneByN
+      : (NumOutputDims == 2 && internal::index_statically_eq<NewDimensions>(/*index=*/1, /*value=*/1)) ? NByOne
+      : Runtime;
+#else
+        Runtime;
+#endif
+  // clang-format on
+
+  enum {
+    IsAligned         = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    // For trivial reshapes with raw access to underlying data we will provide
+    // zero overhead block access.
+    // TODO(ezhulenev): Consider adding block access without raw access?
+    BlockAccess       = TensorEvaluator<ArgType, Device>::RawAccess &&
+                        NumInputDims > 0 && NumOutputDims > 0,
+    PreferBlockAccess = false,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = TensorEvaluator<ArgType, Device>::RawAccess
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumOutputDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef
+      typename internal::TensorMaterializedBlock<ScalarNoConst, NumOutputDims,
+                                                 Layout, Index>
+          TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_dimensions(op.dimensions())
+  {
+    // The total size of the reshaped tensor must be equal to the total size
+    // of the input tensor.
+    eigen_assert(internal::array_prod(m_impl.dimensions()) == internal::array_prod(op.dimensions()));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType data, EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(data, std::move(done));
+  }
+#endif
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    return m_impl.evalSubExprsIfNeeded(data);
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return m_impl.coeff(index);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return m_impl.template packet<LoadMode>(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    return m_impl.costPerCoeff(vectorized);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    return internal::TensorBlockResourceRequirements::any();
+  }
+
+  // required in block(OutputTensorBlock* output_block) const
+  // For C++03 compatibility this must be defined outside the method
+  struct BlockIteratorState {
+    Index stride;
+    Index span;
+    Index size;
+    Index count;
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    eigen_assert(m_impl.data() != NULL);
+    eigen_assert((kind == Runtime) ||
+                 (kind == OneByN && desc.dimensions()[0] == 1) ||
+                 (kind == NByOne && desc.dimensions()[1] == 1));
+
+    if (kind == OneByN || kind == NByOne) {
+      // We can guarantee at compile time that block is just a contiguous slice
+      // of the underlying expression memory buffer.
+      return TensorBlock(internal::TensorBlockKind::kView,
+                           m_impl.data() + desc.offset(), desc.dimensions());
+    } else {
+      // This will do additional runtime checks, and in the end it might be also
+      // a view, or it might be a block materialized in the temporary buffer.
+      return TensorBlock::materialize(m_impl.data(), m_dimensions, desc,
+                                        scratch);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC typename Storage::Type data() const {
+    return constCast(m_impl.data());
+  }
+
+  EIGEN_DEVICE_FUNC const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+
+  #ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+  #endif
+ protected:
+  TensorEvaluator<ArgType, Device> m_impl;
+  NewDimensions m_dimensions;
+};
+
+
+// Eval as lvalue
+template<typename NewDimensions, typename ArgType, typename Device>
+  struct TensorEvaluator<TensorReshapingOp<NewDimensions, ArgType>, Device>
+  : public TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device>
+
+{
+  typedef TensorEvaluator<const TensorReshapingOp<NewDimensions, ArgType>, Device> Base;
+  typedef TensorReshapingOp<NewDimensions, ArgType> XprType;
+  typedef NewDimensions Dimensions;
+
+  enum {
+    IsAligned         = TensorEvaluator<ArgType, Device>::IsAligned,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::RawAccess,
+    PreferBlockAccess = false,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = TensorEvaluator<ArgType, Device>::RawAccess
+  };
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : Base(op, device)
+  { }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<TensorEvaluator::NumOutputDims, Index>
+      TensorBlockDesc;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
+  {
+    return this->m_impl.coeffRef(index);
+  }
+
+  template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x)
+  {
+    this->m_impl.template writePacket<StoreMode>(index, x);
+  }
+
+  template <typename TensorBlock>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const TensorBlockDesc& desc, const TensorBlock& block) {
+    assert(this->m_impl.data() != NULL);
+
+    typedef typename TensorBlock::XprType TensorBlockExpr;
+    typedef internal::TensorBlockAssignment<
+        Scalar, TensorEvaluator::NumOutputDims, TensorBlockExpr, Index>
+        TensorBlockAssign;
+
+    TensorBlockAssign::Run(
+        TensorBlockAssign::target(desc.dimensions(),
+                                  internal::strides<Layout>(this->dimensions()),
+                                  this->m_impl.data(), desc.offset()),
+        block.expr());
+  }
+};
+
+
+/** \class TensorSlicing
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor slicing class.
+  *
+  *
+  */
+namespace internal {
+template<typename StartIndices, typename Sizes, typename XprType>
+struct traits<TensorSlicingOp<StartIndices, Sizes, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = array_size<StartIndices>::value;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename StartIndices, typename Sizes, typename XprType>
+struct eval<TensorSlicingOp<StartIndices, Sizes, XprType>, Eigen::Dense>
+{
+  typedef const TensorSlicingOp<StartIndices, Sizes, XprType>EIGEN_DEVICE_REF type;
+};
+
+template<typename StartIndices, typename Sizes, typename XprType>
+struct nested<TensorSlicingOp<StartIndices, Sizes, XprType>, 1, typename eval<TensorSlicingOp<StartIndices, Sizes, XprType> >::type>
+{
+  typedef TensorSlicingOp<StartIndices, Sizes, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename StartIndices, typename Sizes, typename XprType>
+class TensorSlicingOp : public TensorBase<TensorSlicingOp<StartIndices, Sizes, XprType> >
+{
+  public:
+  typedef TensorBase<TensorSlicingOp<StartIndices, Sizes, XprType> > Base;
+  typedef typename Eigen::internal::traits<TensorSlicingOp>::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorSlicingOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorSlicingOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorSlicingOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorSlicingOp(const XprType& expr, const StartIndices& indices, const Sizes& sizes)
+      : m_xpr(expr), m_indices(indices), m_sizes(sizes) {}
+
+    EIGEN_DEVICE_FUNC
+    const StartIndices& startIndices() const { return m_indices; }
+    EIGEN_DEVICE_FUNC
+    const Sizes& sizes() const { return m_sizes; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorSlicingOp)
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const StartIndices m_indices;
+    const Sizes m_sizes;
+};
+
+
+// Fixme: figure out the exact threshold
+namespace {
+template <typename Index, typename Device, bool BlockAccess> struct MemcpyTriggerForSlicing {
+  EIGEN_DEVICE_FUNC MemcpyTriggerForSlicing(const Device& device) : threshold_(2 * device.numThreads()) { }
+  EIGEN_DEVICE_FUNC bool operator ()(Index total, Index contiguous) const {
+    const bool prefer_block_evaluation = BlockAccess && total > 32*1024;
+    return !prefer_block_evaluation && contiguous > threshold_;
+  }
+
+ private:
+  Index threshold_;
+};
+
+// It is very expensive to start the memcpy kernel on GPU: we therefore only
+// use it for large copies.
+#ifdef EIGEN_USE_GPU
+template <typename Index, bool BlockAccess> struct MemcpyTriggerForSlicing<Index, GpuDevice, BlockAccess>  {
+  EIGEN_DEVICE_FUNC MemcpyTriggerForSlicing(const GpuDevice&) { }
+  EIGEN_DEVICE_FUNC bool operator ()(Index, Index contiguous) const { return contiguous > 4*1024*1024; }
+};
+#endif
+
+// It is very expensive to start the memcpy kernel on GPU: we therefore only
+// use it for large copies.
+#ifdef EIGEN_USE_SYCL
+template <typename Index, bool BlockAccess> struct MemcpyTriggerForSlicing<Index, Eigen::SyclDevice, BlockAccess>  {
+  EIGEN_DEVICE_FUNC MemcpyTriggerForSlicing(const SyclDevice&) { }
+  EIGEN_DEVICE_FUNC bool operator ()(Index, Index contiguous) const { return contiguous > 4*1024*1024; }
+};
+#endif
+
+}
+
+// Eval as rvalue
+template<typename StartIndices, typename Sizes, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
+{
+  typedef TensorSlicingOp<StartIndices, Sizes, ArgType> XprType;
+  static const int NumDims = internal::array_size<Sizes>::value;
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef Sizes Dimensions;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef StorageMemory<typename internal::remove_const<CoeffReturnType>::type, Device> ConstCastStorage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    // Alignment can't be guaranteed at compile time since it depends on the
+    // slice offsets and sizes.
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess &&
+                        // FIXME: Temporary workaround for bug in slicing of bool tensors.
+                        !internal::is_same<typename internal::remove_const<Scalar>::type, bool>::value,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,
+    RawAccess         = false
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  // Tensor slicing does not change the block type.
+  typedef typename TensorEvaluator<const ArgType, Device>::TensorBlock
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_device(device), m_dimensions(op.sizes()), m_offsets(op.startIndices())
+  {
+    m_is_identity = true;
+    for (int i = 0; i < internal::array_size<Dimensions>::value; ++i) {
+      eigen_assert(m_impl.dimensions()[i] >=
+                   op.sizes()[i] + op.startIndices()[i]);
+      if (m_impl.dimensions()[i] != op.sizes()[i] ||
+          op.startIndices()[i] != 0) {
+        m_is_identity = false;
+      }
+    }
+
+    // No strides for scalars.
+    if (NumDims == 0) return;
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    const Sizes& output_dims = op.sizes();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+      }
+
+     // Don't initialize m_fastOutputStrides[0] since it won't ever be accessed.
+      m_outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1];
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i] > 0 ? m_outputStrides[i] : 1);
+      }
+    } else {
+      m_inputStrides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
+      }
+
+     // Don't initialize m_fastOutputStrides[NumDims-1] since it won't ever be accessed.
+      m_outputStrides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_outputStrides[i] = m_outputStrides[i+1] * output_dims[i+1];
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i] > 0 ? m_outputStrides[i] : 1);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    if (!NumTraits<typename internal::remove_const<Scalar>::type>::RequireInitialization
+        && data && m_impl.data()) {
+      Index contiguous_values = 1;
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        for (int i = 0; i < NumDims; ++i) {
+          contiguous_values *= dimensions()[i];
+          if (dimensions()[i] != m_impl.dimensions()[i]) {
+            break;
+          }
+        }
+      } else {
+        for (int i = NumDims-1; i >= 0; --i) {
+          contiguous_values *= dimensions()[i];
+          if (dimensions()[i] != m_impl.dimensions()[i]) {
+            break;
+          }
+        }
+      }
+      // Use memcpy if it's going to be faster than using the regular evaluation.
+      const MemcpyTriggerForSlicing<Index, Device, BlockAccess> trigger(m_device);
+      if (trigger(internal::array_prod(dimensions()), contiguous_values)) {
+        EvaluatorPointerType src = (EvaluatorPointerType)m_impl.data();
+        for (Index i = 0; i < internal::array_prod(dimensions()); i += contiguous_values) {
+          Index offset = srcCoeff(i);
+          m_device.memcpy((void*)(m_device.get(data + i)), m_device.get(src+offset), contiguous_values * sizeof(Scalar));
+        }
+        return false;
+      }
+    }
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType /*data*/, EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    if (m_is_identity) {
+      return m_impl.coeff(index);
+    } else {
+      return m_impl.coeff(srcCoeff(index));
+    }
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
+    EIGEN_STATIC_ASSERT((packetSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+packetSize-1 < internal::array_prod(dimensions()));
+
+    if (m_is_identity) {
+      return m_impl.template packet<LoadMode>(index);
+    }
+
+    Index inputIndices[] = {0, 0};
+    Index indices[] = {index, index + packetSize - 1};
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx0 = indices[0] / m_fastOutputStrides[i];
+        const Index idx1 = indices[1] / m_fastOutputStrides[i];
+        inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i];
+        inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i];
+        indices[0] -= idx0 * m_outputStrides[i];
+        indices[1] -= idx1 * m_outputStrides[i];
+      }
+      inputIndices[0] += (indices[0] + m_offsets[0]);
+      inputIndices[1] += (indices[1] + m_offsets[0]);
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx0 = indices[0] / m_fastOutputStrides[i];
+        const Index idx1 = indices[1] / m_fastOutputStrides[i];
+        inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i];
+        inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i];
+        indices[0] -= idx0 * m_outputStrides[i];
+        indices[1] -= idx1 * m_outputStrides[i];
+      }
+      inputIndices[0] += (indices[0] + m_offsets[NumDims-1]);
+      inputIndices[1] += (indices[1] + m_offsets[NumDims-1]);
+    }
+    if (inputIndices[1] - inputIndices[0] == packetSize - 1) {
+      PacketReturnType rslt = m_impl.template packet<Unaligned>(inputIndices[0]);
+      return rslt;
+    }
+    else {
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[packetSize];
+      values[0] = m_impl.coeff(inputIndices[0]);
+      values[packetSize-1] = m_impl.coeff(inputIndices[1]);
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < packetSize-1; ++i) {
+        values[i] = coeff(index+i);
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, m_is_identity ? 1 : NumDims);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    const size_t target_size = m_device.lastLevelCacheSize();
+    return internal::TensorBlockResourceRequirements::merge(
+        internal::TensorBlockResourceRequirements::skewed<Scalar>(target_size),
+        m_impl.getResourceRequirements());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    TensorBlockDesc arg_desc = desc.WithOffset(srcCoeff(desc.offset()));
+    TensorBlock block = m_impl.block(arg_desc, scratch);
+    if (!arg_desc.HasDestinationBuffer()) desc.DropDestinationBuffer();
+    return block;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Storage::Type data() const {
+    typename Storage::Type result = constCast(m_impl.data());
+    if (result) {
+      Index offset = 0;
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        for (int i = 0; i < NumDims; ++i) {
+          if (m_dimensions[i] != m_impl.dimensions()[i]) {
+            offset += m_offsets[i] * m_inputStrides[i];
+            for (int j = i+1; j < NumDims; ++j) {
+              if (m_dimensions[j] > 1) {
+                return NULL;
+              }
+              offset += m_offsets[j] * m_inputStrides[j];
+            }
+            break;
+          }
+        }
+      } else {
+        for (int i = NumDims - 1; i >= 0; --i) {
+          if (m_dimensions[i] != m_impl.dimensions()[i]) {
+            offset += m_offsets[i] * m_inputStrides[i];
+            for (int j = i-1; j >= 0; --j) {
+              if (m_dimensions[j] > 1) {
+                return NULL;
+              }
+              offset += m_offsets[j] * m_inputStrides[j];
+            }
+            break;
+          }
+        }
+      }
+      return result + offset;
+    }
+    return NULL;
+  }
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const
+  {
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        inputIndex += (idx + m_offsets[i]) * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      inputIndex += (index + m_offsets[0]);
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        inputIndex += (idx + m_offsets[i]) * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      inputIndex += (index + m_offsets[NumDims-1]);
+    }
+    return inputIndex;
+  }
+
+  array<Index, NumDims> m_outputStrides;
+  array<internal::TensorIntDivisor<Index>, NumDims> m_fastOutputStrides;
+  array<Index, NumDims> m_inputStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+  const Device EIGEN_DEVICE_REF m_device;
+  Dimensions m_dimensions;
+  bool m_is_identity;
+  const StartIndices m_offsets;
+};
+
+
+// Eval as lvalue
+template<typename StartIndices, typename Sizes, typename ArgType, typename Device>
+struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
+  : public TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
+{
+  typedef TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Device> Base;
+  typedef TensorSlicingOp<StartIndices, Sizes, ArgType> XprType;
+  static const int NumDims = internal::array_size<Sizes>::value;
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef Sizes Dimensions;
+
+  enum {
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,
+    RawAccess         = (NumDims == 1) & TensorEvaluator<ArgType, Device>::RawAccess
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : Base(op, device)
+    { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
+  {
+    if (this->m_is_identity) {
+      return this->m_impl.coeffRef(index);
+    } else {
+      return this->m_impl.coeffRef(this->srcCoeff(index));
+    }
+  }
+
+  template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x)
+  {
+    if (this->m_is_identity) {
+      this->m_impl.template writePacket<StoreMode>(index, x);
+      return;
+    }
+
+    const int packetSize = PacketType<CoeffReturnType, Device>::size;
+    Index inputIndices[] = {0, 0};
+    Index indices[] = {index, index + packetSize - 1};
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx0 = indices[0] / this->m_fastOutputStrides[i];
+        const Index idx1 = indices[1] / this->m_fastOutputStrides[i];
+        inputIndices[0] += (idx0 + this->m_offsets[i]) * this->m_inputStrides[i];
+        inputIndices[1] += (idx1 + this->m_offsets[i]) * this->m_inputStrides[i];
+        indices[0] -= idx0 * this->m_outputStrides[i];
+        indices[1] -= idx1 * this->m_outputStrides[i];
+      }
+      inputIndices[0] += (indices[0] + this->m_offsets[0]);
+      inputIndices[1] += (indices[1] + this->m_offsets[0]);
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx0 = indices[0] / this->m_fastOutputStrides[i];
+        const Index idx1 = indices[1] / this->m_fastOutputStrides[i];
+        inputIndices[0] += (idx0 + this->m_offsets[i]) * this->m_inputStrides[i];
+        inputIndices[1] += (idx1 + this->m_offsets[i]) * this->m_inputStrides[i];
+        indices[0] -= idx0 * this->m_outputStrides[i];
+        indices[1] -= idx1 * this->m_outputStrides[i];
+      }
+      inputIndices[0] += (indices[0] + this->m_offsets[NumDims-1]);
+      inputIndices[1] += (indices[1] + this->m_offsets[NumDims-1]);
+    }
+    if (inputIndices[1] - inputIndices[0] == packetSize - 1) {
+      this->m_impl.template writePacket<StoreMode>(inputIndices[0], x);
+    }
+    else {
+      EIGEN_ALIGN_MAX CoeffReturnType values[packetSize];
+      internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
+      this->m_impl.coeffRef(inputIndices[0]) = values[0];
+      this->m_impl.coeffRef(inputIndices[1]) = values[packetSize-1];
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < packetSize-1; ++i) {
+        this->coeffRef(index+i) = values[i];
+      }
+    }
+  }
+
+  template<typename TensorBlock>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const TensorBlockDesc& desc, const TensorBlock& block) {
+    TensorBlockDesc arg_desc = desc.WithOffset(this->srcCoeff(desc.offset()));
+    this->m_impl.writeBlock(arg_desc, block);
+  }
+};
+
+namespace internal {
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
+struct traits<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = array_size<StartIndices>::value;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
+struct eval<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>, Eigen::Dense>
+{
+  typedef const TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>EIGEN_DEVICE_REF type;
+};
+
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
+struct nested<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>, 1, typename eval<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> >::type>
+{
+  typedef TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> type;
+};
+
+}  // end namespace internal
+
+
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
+class TensorStridingSlicingOp : public TensorBase<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> >
+{
+  public:
+  typedef TensorBase<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> > Base;
+  typedef typename internal::traits<TensorStridingSlicingOp>::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename internal::nested<TensorStridingSlicingOp>::type Nested;
+  typedef typename internal::traits<TensorStridingSlicingOp>::StorageKind StorageKind;
+  typedef typename internal::traits<TensorStridingSlicingOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorStridingSlicingOp(
+    const XprType& expr, const StartIndices& startIndices,
+    const StopIndices& stopIndices, const Strides& strides)
+      : m_xpr(expr), m_startIndices(startIndices), m_stopIndices(stopIndices),
+        m_strides(strides) {}
+
+    EIGEN_DEVICE_FUNC
+    const StartIndices& startIndices() const { return m_startIndices; }
+    EIGEN_DEVICE_FUNC
+    const StartIndices& stopIndices() const { return m_stopIndices; }
+    EIGEN_DEVICE_FUNC
+    const StartIndices& strides() const { return m_strides; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorStridingSlicingOp)
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const StartIndices m_startIndices;
+    const StopIndices m_stopIndices;
+    const Strides m_strides;
+};
+
+// Eval as rvalue
+template<typename StartIndices, typename StopIndices, typename Strides, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType>, Device>
+{
+  typedef TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType> XprType;
+  static const int NumDims = internal::array_size<Strides>::value;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  typedef Strides Dimensions;
+
+  enum {
+    // Alignment can't be guaranteed at compile time since it depends on the
+    // slice offsets and sizes.
+    IsAligned = false,
+    PacketAccess = false,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device),
+        m_device(device),
+        m_strides(op.strides())
+  {
+    // Handle degenerate intervals by gracefully clamping and allowing m_dimensions to be zero
+    DSizes<Index, NumDims> startIndicesClamped, stopIndicesClamped;
+    for (ptrdiff_t i = 0; i < internal::array_size<Dimensions>::value; ++i) {
+      eigen_assert(m_strides[i] != 0 && "0 stride is invalid");
+      if (m_strides[i] > 0) {
+        startIndicesClamped[i] =
+            clamp(op.startIndices()[i], 0, m_impl.dimensions()[i]);
+        stopIndicesClamped[i] =
+            clamp(op.stopIndices()[i], 0, m_impl.dimensions()[i]);
+      } else {
+        /* implies m_strides[i] < 0 by assert */
+        startIndicesClamped[i] =
+            clamp(op.startIndices()[i], -1, m_impl.dimensions()[i] - 1);
+        stopIndicesClamped[i] =
+            clamp(op.stopIndices()[i], -1, m_impl.dimensions()[i] - 1);
+      }
+      m_startIndices[i] = startIndicesClamped[i];
+    }
+
+    typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions;
+    const InputDimensions& input_dims = m_impl.dimensions();
+
+    // compute output tensor shape
+    m_is_identity = true;
+    for (int i = 0; i < NumDims; i++) {
+      Index interval = stopIndicesClamped[i] - startIndicesClamped[i];
+      if (interval == 0 || ((interval < 0) != (m_strides[i] < 0))) {
+        m_dimensions[i] = 0;
+      } else {
+        m_dimensions[i] =
+            (interval / m_strides[i]) + (interval % m_strides[i] != 0 ? 1 : 0);
+        eigen_assert(m_dimensions[i] >= 0);
+      }
+      if (m_strides[i] != 1 || interval != m_impl.dimensions()[i]) {
+        m_is_identity = false;
+      }
+    }
+
+    Strides output_dims = m_dimensions;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputStrides[0] = m_strides[0];
+      m_offsets[0] = startIndicesClamped[0];
+      Index previousDimProduct = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        previousDimProduct *= input_dims[i-1];
+        m_inputStrides[i] = previousDimProduct * m_strides[i];
+        m_offsets[i] = startIndicesClamped[i] * previousDimProduct;
+      }
+
+      // Don't initialize m_fastOutputStrides[0] since it won't ever be accessed.
+      m_outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1];
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i] > 0 ? m_outputStrides[i] : 1);
+      }
+    } else {
+      m_inputStrides[NumDims-1] = m_strides[NumDims-1];
+      m_offsets[NumDims-1] = startIndicesClamped[NumDims-1];
+      Index previousDimProduct = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        previousDimProduct *= input_dims[i+1];
+        m_inputStrides[i] = previousDimProduct * m_strides[i];
+        m_offsets[i] = startIndicesClamped[i] * previousDimProduct;
+      }
+
+      m_outputStrides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_outputStrides[i] = m_outputStrides[i+1] * output_dims[i+1];
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i] > 0 ? m_outputStrides[i] : 1);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    if (m_is_identity) {
+      return m_impl.coeff(index);
+    } else {
+      return m_impl.coeff(srcCoeff(index));
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, m_is_identity ? 1 : NumDims);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Storage::Type data() const {
+    return NULL;
+  }
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const
+  {
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i >= 0; --i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        inputIndex += idx * m_inputStrides[i] + m_offsets[i];
+        index -= idx * m_outputStrides[i];
+      }
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims; ++i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        inputIndex += idx * m_inputStrides[i] + m_offsets[i];
+        index -= idx * m_outputStrides[i];
+      }
+    }
+    return inputIndex;
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index clamp(Index value, Index min, Index max) {
+#ifndef SYCL_DEVICE_ONLY
+    return numext::maxi(min, numext::mini(max,value));
+#else
+    return cl::sycl::clamp(value, min, max);
+#endif
+  }
+
+  array<Index, NumDims> m_outputStrides;
+  array<internal::TensorIntDivisor<Index>, NumDims> m_fastOutputStrides;
+  array<Index, NumDims> m_inputStrides;
+  bool m_is_identity;
+  TensorEvaluator<ArgType, Device> m_impl;
+  const Device EIGEN_DEVICE_REF m_device;
+  DSizes<Index, NumDims> m_startIndices; // clamped startIndices
+  DSizes<Index, NumDims> m_dimensions;
+  DSizes<Index, NumDims> m_offsets; // offset in a flattened shape
+  const Strides m_strides;
+};
+
+// Eval as lvalue
+template<typename StartIndices, typename StopIndices, typename Strides, typename ArgType, typename Device>
+struct TensorEvaluator<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType>, Device>
+  : public TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType>, Device>
+{
+  typedef TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType>, Device> Base;
+  typedef TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType> XprType;
+  static const int NumDims = internal::array_size<Strides>::value;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = false,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = TensorEvaluator<ArgType, Device>::CoordAccess,
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : Base(op, device)
+    { }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef Strides Dimensions;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
+  {
+    if (this->m_is_identity) {
+      return this->m_impl.coeffRef(index);
+    } else {
+      return this->m_impl.coeffRef(this->srcCoeff(index));
+    }
+  }
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_MORPHING_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorPadding.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorPadding.h
new file mode 100644
index 0000000..ee44382
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorPadding.h
@@ -0,0 +1,708 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_PADDING_H
+#define EIGEN_CXX11_TENSOR_TENSOR_PADDING_H
+
+namespace Eigen {
+
+/** \class TensorPadding
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor padding class.
+  * At the moment only padding with a constant value is supported.
+  *
+  */
+namespace internal {
+template<typename PaddingDimensions, typename XprType>
+struct traits<TensorPaddingOp<PaddingDimensions, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename PaddingDimensions, typename XprType>
+struct eval<TensorPaddingOp<PaddingDimensions, XprType>, Eigen::Dense>
+{
+  typedef const TensorPaddingOp<PaddingDimensions, XprType>& type;
+};
+
+template<typename PaddingDimensions, typename XprType>
+struct nested<TensorPaddingOp<PaddingDimensions, XprType>, 1, typename eval<TensorPaddingOp<PaddingDimensions, XprType> >::type>
+{
+  typedef TensorPaddingOp<PaddingDimensions, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename PaddingDimensions, typename XprType>
+class TensorPaddingOp : public TensorBase<TensorPaddingOp<PaddingDimensions, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorPaddingOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorPaddingOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorPaddingOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorPaddingOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorPaddingOp(const XprType& expr, const PaddingDimensions& padding_dims, const Scalar padding_value)
+      : m_xpr(expr), m_padding_dims(padding_dims), m_padding_value(padding_value) {}
+
+    EIGEN_DEVICE_FUNC
+    const PaddingDimensions& padding() const { return m_padding_dims; }
+    EIGEN_DEVICE_FUNC
+    Scalar padding_value() const { return m_padding_value; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const PaddingDimensions m_padding_dims;
+    const Scalar m_padding_value;
+};
+
+
+// Eval as rvalue
+template<typename PaddingDimensions, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorPaddingOp<PaddingDimensions, ArgType>, Device>
+{
+  typedef TensorPaddingOp<PaddingDimensions, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<PaddingDimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned         = true,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = TensorEvaluator<ArgType, Device>::RawAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = true,
+    RawAccess         = false
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename internal::TensorMaterializedBlock<ScalarNoConst, NumDims,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_padding(op.padding()), m_paddingValue(op.padding_value()), m_device(device)
+  {
+    // The padding op doesn't change the rank of the tensor. Directly padding a scalar would lead
+    // to a vector, which doesn't make sense. Instead one should reshape the scalar into a vector
+    // of 1 element first and then pad.
+    EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    // Compute dimensions
+    m_dimensions = m_impl.dimensions();
+    for (int i = 0; i < NumDims; ++i) {
+      m_dimensions[i] += m_padding[i].first + m_padding[i].second;
+    }
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputStrides[0] = 1;
+      m_outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+        m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
+      }
+      m_outputStrides[NumDims] = m_outputStrides[NumDims-1] * m_dimensions[NumDims-1];
+    } else {
+      m_inputStrides[NumDims - 1] = 1;
+      m_outputStrides[NumDims] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
+        m_outputStrides[i+1] = m_outputStrides[i+2] * m_dimensions[i+1];
+      }
+      m_outputStrides[0] = m_outputStrides[1] * m_dimensions[0];
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    eigen_assert(index < dimensions().TotalSize());
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_outputStrides[i];
+        if (isPaddingAtIndexForDim(idx, i)) {
+          return m_paddingValue;
+        }
+        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      if (isPaddingAtIndexForDim(index, 0)) {
+        return m_paddingValue;
+      }
+      inputIndex += (index - m_padding[0].first);
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_outputStrides[i+1];
+        if (isPaddingAtIndexForDim(idx, i)) {
+          return m_paddingValue;
+        }
+        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
+        index -= idx * m_outputStrides[i+1];
+      }
+      if (isPaddingAtIndexForDim(index, NumDims-1)) {
+        return m_paddingValue;
+      }
+      inputIndex += (index - m_padding[NumDims-1].first);
+    }
+    return m_impl.coeff(inputIndex);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      return packetColMajor(index);
+    }
+    return packetRowMajor(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    TensorOpCost cost = m_impl.costPerCoeff(vectorized);
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims; ++i)
+        updateCostPerDimension(cost, i, i == 0);
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i >= 0; --i)
+        updateCostPerDimension(cost, i, i == NumDims - 1);
+    }
+    return cost;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    const size_t target_size = m_device.lastLevelCacheSize();
+    return internal::TensorBlockResourceRequirements::merge(
+        internal::TensorBlockResourceRequirements::skewed<Scalar>(target_size),
+        m_impl.getResourceRequirements());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    // If one of the dimensions is zero, return empty block view.
+    if (desc.size() == 0) {
+      return TensorBlock(internal::TensorBlockKind::kView, NULL,
+                           desc.dimensions());
+    }
+
+    static const bool IsColMajor = Layout == static_cast<int>(ColMajor);
+    const int inner_dim_idx = IsColMajor ? 0 : NumDims - 1;
+
+    Index offset = desc.offset();
+
+    // Compute offsets in the output tensor corresponding to the desc.offset().
+    DSizes<Index, NumDims> output_offsets;
+    for (int i = NumDims - 1; i > 0; --i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+      const int stride_dim = IsColMajor ? dim : dim + 1;
+      output_offsets[dim] = offset / m_outputStrides[stride_dim];
+      offset -= output_offsets[dim] * m_outputStrides[stride_dim];
+    }
+    output_offsets[inner_dim_idx] = offset;
+
+    // Offsets in the input corresponding to output offsets.
+    DSizes<Index, NumDims> input_offsets = output_offsets;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+      input_offsets[dim] = input_offsets[dim] - m_padding[dim].first;
+    }
+
+    // Compute offset in the input buffer (at this point it might be illegal and
+    // point outside of the input buffer, because we don't check for negative
+    // offsets, it will be autocorrected in the block iteration loop below).
+    Index input_offset = 0;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+      input_offset += input_offsets[dim] * m_inputStrides[dim];
+    }
+
+    // Destination buffer and scratch buffer both indexed from 0 and have the
+    // same dimensions as the requested block (for destination buffer this
+    // property is guaranteed by `desc.destination()`).
+    Index output_offset = 0;
+    const DSizes<Index, NumDims> output_strides =
+        internal::strides<Layout>(desc.dimensions());
+
+    // NOTE(ezhulenev): We initialize bock iteration state for `NumDims - 1`
+    // dimensions, skipping innermost dimension. In theory it should be possible
+    // to squeeze matching innermost dimensions, however in practice that did
+    // not show any improvements in benchmarks. Also in practice first outer
+    // dimension usually has padding, and will prevent squeezing.
+
+    // Initialize output block iterator state. Dimension in this array are
+    // always in inner_most -> outer_most order (col major layout).
+    array<BlockIteratorState, NumDims - 1> it;
+    for (int i = 0; i < NumDims - 1; ++i) {
+      const int dim = IsColMajor ? i + 1 : NumDims - i - 2;
+      it[i].count = 0;
+      it[i].size = desc.dimension(dim);
+
+      it[i].input_stride = m_inputStrides[dim];
+      it[i].input_span = it[i].input_stride * (it[i].size - 1);
+
+      it[i].output_stride = output_strides[dim];
+      it[i].output_span = it[i].output_stride * (it[i].size - 1);
+    }
+
+    const Index input_inner_dim_size =
+        static_cast<Index>(m_impl.dimensions()[inner_dim_idx]);
+
+    // Total output size.
+    const Index output_size = desc.size();
+
+    // We will fill inner dimension of this size in the output. It might be
+    // larger than the inner dimension in the input, so we might have to pad
+    // before/after we copy values from the input inner dimension.
+    const Index output_inner_dim_size = desc.dimension(inner_dim_idx);
+
+    // How many values to fill with padding BEFORE reading from the input inner
+    // dimension.
+    const Index output_inner_pad_before_size =
+        input_offsets[inner_dim_idx] < 0
+            ? numext::mini(numext::abs(input_offsets[inner_dim_idx]),
+                           output_inner_dim_size)
+            : 0;
+
+    // How many values we can actually copy from the input inner dimension.
+    const Index output_inner_copy_size = numext::mini(
+        // Want to copy from input.
+        (output_inner_dim_size - output_inner_pad_before_size),
+        // Can copy from input.
+        numext::maxi(input_inner_dim_size - (input_offsets[inner_dim_idx] +
+                                             output_inner_pad_before_size),
+                     Index(0)));
+
+    eigen_assert(output_inner_copy_size >= 0);
+
+    // How many values to fill with padding AFTER reading from the input inner
+    // dimension.
+    const Index output_inner_pad_after_size =
+        (output_inner_dim_size - output_inner_copy_size -
+         output_inner_pad_before_size);
+
+    // Sanity check, sum of all sizes must be equal to the output size.
+    eigen_assert(output_inner_dim_size ==
+                 (output_inner_pad_before_size + output_inner_copy_size +
+                  output_inner_pad_after_size));
+
+    // Keep track of current coordinates and padding in the output.
+    DSizes<Index, NumDims> output_coord = output_offsets;
+    DSizes<Index, NumDims> output_padded;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = IsColMajor ? i : NumDims - i - 1;
+      output_padded[dim] = isPaddingAtIndexForDim(output_coord[dim], dim);
+    }
+
+    typedef internal::StridedLinearBufferCopy<ScalarNoConst, Index> LinCopy;
+
+    // Prepare storage for the materialized padding result.
+    const typename TensorBlock::Storage block_storage =
+        TensorBlock::prepareStorage(desc, scratch);
+
+    // TODO(ezhulenev): Squeeze multiple non-padded inner dimensions into a
+    // single logical inner dimension.
+
+    // When possible we squeeze writes for the innermost (only if non-padded)
+    // dimension with the first padded dimension. This allows to reduce the
+    // number of calls to LinCopy and better utilize vector instructions.
+    const bool squeeze_writes =
+        NumDims > 1 &&
+        // inner dimension is not padded
+        (input_inner_dim_size == m_dimensions[inner_dim_idx]) &&
+        // and equal to the block inner dimension
+        (input_inner_dim_size == output_inner_dim_size);
+
+    const int squeeze_dim = IsColMajor ? inner_dim_idx + 1 : inner_dim_idx - 1;
+
+    // Maximum coordinate on a squeeze dimension that we can write to.
+    const Index squeeze_max_coord =
+        squeeze_writes ? numext::mini(
+                             // max non-padded element in the input
+                             static_cast<Index>(m_dimensions[squeeze_dim] -
+                                                m_padding[squeeze_dim].second),
+                             // max element in the output buffer
+                             static_cast<Index>(output_offsets[squeeze_dim] +
+                                                desc.dimension(squeeze_dim)))
+                       : static_cast<Index>(0);
+
+    // Iterate copying data from `m_impl.data()` to the output buffer.
+    for (Index size = 0; size < output_size;) {
+      // Detect if we are in the padded region (exclude innermost dimension).
+      bool is_padded = false;
+      for (int j = 1; j < NumDims; ++j) {
+        const int dim = IsColMajor ? j : NumDims - j - 1;
+        is_padded = output_padded[dim];
+        if (is_padded) break;
+      }
+
+      if (is_padded) {
+        // Fill single innermost dimension with padding value.
+        size += output_inner_dim_size;
+
+        LinCopy::template Run<LinCopy::Kind::FillLinear>(
+            typename LinCopy::Dst(output_offset, 1, block_storage.data()),
+            typename LinCopy::Src(0, 0, &m_paddingValue),
+            output_inner_dim_size);
+
+
+      } else if (squeeze_writes) {
+        // Squeeze multiple reads from innermost dimensions.
+        const Index squeeze_num = squeeze_max_coord - output_coord[squeeze_dim];
+        size += output_inner_dim_size * squeeze_num;
+
+        // Copy `squeeze_num` inner dimensions from input to output.
+        LinCopy::template Run<LinCopy::Kind::Linear>(
+            typename LinCopy::Dst(output_offset, 1, block_storage.data()),
+            typename LinCopy::Src(input_offset, 1, m_impl.data()),
+            output_inner_dim_size * squeeze_num);
+
+        // Update iteration state for only `squeeze_num - 1` processed inner
+        // dimensions, because we have another iteration state update at the end
+        // of the loop that will update iteration state for the last inner
+        // processed dimension.
+        it[0].count += (squeeze_num - 1);
+        input_offset += it[0].input_stride * (squeeze_num - 1);
+        output_offset += it[0].output_stride * (squeeze_num - 1);
+        output_coord[squeeze_dim] += (squeeze_num - 1);
+
+      } else {
+        // Single read from innermost dimension.
+        size += output_inner_dim_size;
+
+        {  // Fill with padding before copying from input inner dimension.
+          const Index out = output_offset;
+
+          LinCopy::template Run<LinCopy::Kind::FillLinear>(
+              typename LinCopy::Dst(out, 1, block_storage.data()),
+              typename LinCopy::Src(0, 0, &m_paddingValue),
+              output_inner_pad_before_size);
+        }
+
+        {  // Copy data from input inner dimension.
+          const Index out = output_offset + output_inner_pad_before_size;
+          const Index in = input_offset + output_inner_pad_before_size;
+
+          eigen_assert(output_inner_copy_size == 0 || m_impl.data() != NULL);
+
+          LinCopy::template Run<LinCopy::Kind::Linear>(
+              typename LinCopy::Dst(out, 1, block_storage.data()),
+              typename LinCopy::Src(in, 1, m_impl.data()),
+              output_inner_copy_size);
+        }
+
+        {  // Fill with padding after copying from input inner dimension.
+          const Index out = output_offset + output_inner_pad_before_size +
+                            output_inner_copy_size;
+
+          LinCopy::template Run<LinCopy::Kind::FillLinear>(
+              typename LinCopy::Dst(out, 1, block_storage.data()),
+              typename LinCopy::Src(0, 0, &m_paddingValue),
+              output_inner_pad_after_size);
+        }
+      }
+
+      for (int j = 0; j < NumDims - 1; ++j) {
+        const int dim = IsColMajor ? j + 1 : NumDims - j - 2;
+
+        if (++it[j].count < it[j].size) {
+          input_offset += it[j].input_stride;
+          output_offset += it[j].output_stride;
+          output_coord[dim] += 1;
+          output_padded[dim] = isPaddingAtIndexForDim(output_coord[dim], dim);
+          break;
+        }
+        it[j].count = 0;
+        input_offset -= it[j].input_span;
+        output_offset -= it[j].output_span;
+        output_coord[dim] -= it[j].size - 1;
+        output_padded[dim] = isPaddingAtIndexForDim(output_coord[dim], dim);
+      }
+    }
+
+    return block_storage.AsTensorMaterializedBlock();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ private:
+  struct BlockIteratorState {
+    BlockIteratorState()
+        : count(0),
+          size(0),
+          input_stride(0),
+          input_span(0),
+          output_stride(0),
+          output_span(0) {}
+
+    Index count;
+    Index size;
+    Index input_stride;
+    Index input_span;
+    Index output_stride;
+    Index output_span;
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool isPaddingAtIndexForDim(
+      Index index, int dim_index) const {
+#if defined(EIGEN_HAS_INDEX_LIST)
+    return (!internal::index_pair_first_statically_eq<PaddingDimensions>(dim_index, 0) &&
+            index < m_padding[dim_index].first) ||
+        (!internal::index_pair_second_statically_eq<PaddingDimensions>(dim_index, 0) &&
+         index >= m_dimensions[dim_index] - m_padding[dim_index].second);
+#else
+    return (index < m_padding[dim_index].first) ||
+           (index >= m_dimensions[dim_index] - m_padding[dim_index].second);
+#endif
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool isLeftPaddingCompileTimeZero(
+      int dim_index) const {
+#if defined(EIGEN_HAS_INDEX_LIST)
+    return internal::index_pair_first_statically_eq<PaddingDimensions>(dim_index, 0);
+#else
+    EIGEN_UNUSED_VARIABLE(dim_index);
+    return false;
+#endif
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool isRightPaddingCompileTimeZero(
+      int dim_index) const {
+#if defined(EIGEN_HAS_INDEX_LIST)
+    return internal::index_pair_second_statically_eq<PaddingDimensions>(dim_index, 0);
+#else
+    EIGEN_UNUSED_VARIABLE(dim_index);
+    return false;
+#endif
+  }
+
+
+  void updateCostPerDimension(TensorOpCost& cost, int i, bool first) const {
+    const double in = static_cast<double>(m_impl.dimensions()[i]);
+    const double out = in + m_padding[i].first + m_padding[i].second;
+    if (out == 0)
+      return;
+    const double reduction = in / out;
+    cost *= reduction;
+    if (first) {
+      cost += TensorOpCost(0, 0, 2 * TensorOpCost::AddCost<Index>() +
+                    reduction * (1 * TensorOpCost::AddCost<Index>()));
+    } else {
+      cost += TensorOpCost(0, 0, 2 * TensorOpCost::AddCost<Index>() +
+                                 2 * TensorOpCost::MulCost<Index>() +
+                    reduction * (2 * TensorOpCost::MulCost<Index>() +
+                                 1 * TensorOpCost::DivCost<Index>()));
+    }
+  }
+
+ protected:
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetColMajor(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    const Index initialIndex = index;
+    Index inputIndex = 0;
+    EIGEN_UNROLL_LOOP
+    for (int i = NumDims - 1; i > 0; --i) {
+      const Index firstIdx = index;
+      const Index lastIdx = index + PacketSize - 1;
+      const Index lastPaddedLeft = m_padding[i].first * m_outputStrides[i];
+      const Index firstPaddedRight = (m_dimensions[i] - m_padding[i].second) * m_outputStrides[i];
+      const Index lastPaddedRight = m_outputStrides[i+1];
+
+      if (!isLeftPaddingCompileTimeZero(i) && lastIdx < lastPaddedLeft) {
+        // all the coefficient are in the padding zone.
+        return internal::pset1<PacketReturnType>(m_paddingValue);
+      }
+      else if (!isRightPaddingCompileTimeZero(i) && firstIdx >= firstPaddedRight && lastIdx < lastPaddedRight) {
+        // all the coefficient are in the padding zone.
+        return internal::pset1<PacketReturnType>(m_paddingValue);
+      }
+      else if ((isLeftPaddingCompileTimeZero(i) && isRightPaddingCompileTimeZero(i)) || (firstIdx >= lastPaddedLeft && lastIdx < firstPaddedRight)) {
+        // all the coefficient are between the 2 padding zones.
+        const Index idx = index / m_outputStrides[i];
+        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      else {
+        // Every other case
+        return packetWithPossibleZero(initialIndex);
+      }
+    }
+
+    const Index lastIdx = index + PacketSize - 1;
+    const Index firstIdx = index;
+    const Index lastPaddedLeft = m_padding[0].first;
+    const Index firstPaddedRight = (m_dimensions[0] - m_padding[0].second);
+    const Index lastPaddedRight = m_outputStrides[1];
+
+    if (!isLeftPaddingCompileTimeZero(0) && lastIdx < lastPaddedLeft) {
+      // all the coefficient are in the padding zone.
+      return internal::pset1<PacketReturnType>(m_paddingValue);
+    }
+    else if (!isRightPaddingCompileTimeZero(0) && firstIdx >= firstPaddedRight && lastIdx < lastPaddedRight) {
+      // all the coefficient are in the padding zone.
+      return internal::pset1<PacketReturnType>(m_paddingValue);
+    }
+    else if ((isLeftPaddingCompileTimeZero(0) && isRightPaddingCompileTimeZero(0)) || (firstIdx >= lastPaddedLeft && lastIdx < firstPaddedRight)) {
+      // all the coefficient are between the 2 padding zones.
+      inputIndex += (index - m_padding[0].first);
+      return m_impl.template packet<Unaligned>(inputIndex);
+    }
+    // Every other case
+    return packetWithPossibleZero(initialIndex);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetRowMajor(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    const Index initialIndex = index;
+    Index inputIndex = 0;
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < NumDims - 1; ++i) {
+      const Index firstIdx = index;
+      const Index lastIdx = index + PacketSize - 1;
+      const Index lastPaddedLeft = m_padding[i].first * m_outputStrides[i+1];
+      const Index firstPaddedRight = (m_dimensions[i] - m_padding[i].second) * m_outputStrides[i+1];
+      const Index lastPaddedRight = m_outputStrides[i];
+
+      if (!isLeftPaddingCompileTimeZero(i) && lastIdx < lastPaddedLeft) {
+        // all the coefficient are in the padding zone.
+        return internal::pset1<PacketReturnType>(m_paddingValue);
+      }
+      else if (!isRightPaddingCompileTimeZero(i) && firstIdx >= firstPaddedRight && lastIdx < lastPaddedRight) {
+        // all the coefficient are in the padding zone.
+        return internal::pset1<PacketReturnType>(m_paddingValue);
+      }
+      else if ((isLeftPaddingCompileTimeZero(i) && isRightPaddingCompileTimeZero(i)) || (firstIdx >= lastPaddedLeft && lastIdx < firstPaddedRight)) {
+        // all the coefficient are between the 2 padding zones.
+        const Index idx = index / m_outputStrides[i+1];
+        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
+        index -= idx * m_outputStrides[i+1];
+      }
+      else {
+        // Every other case
+        return packetWithPossibleZero(initialIndex);
+      }
+    }
+
+    const Index lastIdx = index + PacketSize - 1;
+    const Index firstIdx = index;
+    const Index lastPaddedLeft = m_padding[NumDims-1].first;
+    const Index firstPaddedRight = (m_dimensions[NumDims-1] - m_padding[NumDims-1].second);
+    const Index lastPaddedRight = m_outputStrides[NumDims-1];
+
+    if (!isLeftPaddingCompileTimeZero(NumDims-1) && lastIdx < lastPaddedLeft) {
+      // all the coefficient are in the padding zone.
+      return internal::pset1<PacketReturnType>(m_paddingValue);
+    }
+    else if (!isRightPaddingCompileTimeZero(NumDims-1) && firstIdx >= firstPaddedRight && lastIdx < lastPaddedRight) {
+      // all the coefficient are in the padding zone.
+      return internal::pset1<PacketReturnType>(m_paddingValue);
+    }
+    else if ((isLeftPaddingCompileTimeZero(NumDims-1) && isRightPaddingCompileTimeZero(NumDims-1)) || (firstIdx >= lastPaddedLeft && lastIdx < firstPaddedRight)) {
+      // all the coefficient are between the 2 padding zones.
+      inputIndex += (index - m_padding[NumDims-1].first);
+      return m_impl.template packet<Unaligned>(inputIndex);
+    }
+    // Every other case
+    return packetWithPossibleZero(initialIndex);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const
+  {
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  Dimensions m_dimensions;
+  array<Index, NumDims+1> m_outputStrides;
+  array<Index, NumDims> m_inputStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+  PaddingDimensions m_padding;
+
+  Scalar m_paddingValue;
+
+  const Device EIGEN_DEVICE_REF m_device;
+};
+
+
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_PADDING_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorPatch.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorPatch.h
new file mode 100644
index 0000000..413d25d
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorPatch.h
@@ -0,0 +1,291 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_PATCH_H
+#define EIGEN_CXX11_TENSOR_TENSOR_PATCH_H
+
+namespace Eigen {
+
+/** \class TensorPatch
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor patch class.
+  *
+  *
+  */
+namespace internal {
+template<typename PatchDim, typename XprType>
+struct traits<TensorPatchOp<PatchDim, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions + 1;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename PatchDim, typename XprType>
+struct eval<TensorPatchOp<PatchDim, XprType>, Eigen::Dense>
+{
+  typedef const TensorPatchOp<PatchDim, XprType>& type;
+};
+
+template<typename PatchDim, typename XprType>
+struct nested<TensorPatchOp<PatchDim, XprType>, 1, typename eval<TensorPatchOp<PatchDim, XprType> >::type>
+{
+  typedef TensorPatchOp<PatchDim, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename PatchDim, typename XprType>
+class TensorPatchOp : public TensorBase<TensorPatchOp<PatchDim, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorPatchOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorPatchOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorPatchOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorPatchOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorPatchOp(const XprType& expr, const PatchDim& patch_dims)
+      : m_xpr(expr), m_patch_dims(patch_dims) {}
+
+    EIGEN_DEVICE_FUNC
+    const PatchDim& patch_dims() const { return m_patch_dims; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const PatchDim m_patch_dims;
+};
+
+
+// Eval as rvalue
+template<typename PatchDim, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
+{
+  typedef TensorPatchOp<PatchDim, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value + 1;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+
+  enum {
+    IsAligned = false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,
+    RawAccess = false
+ };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device)
+  {
+    Index num_patches = 1;
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    const PatchDim& patch_dims = op.patch_dims();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < NumDims-1; ++i) {
+        m_dimensions[i] = patch_dims[i];
+        num_patches *= (input_dims[i] - patch_dims[i] + 1);
+      }
+      m_dimensions[NumDims-1] = num_patches;
+
+      m_inputStrides[0] = 1;
+      m_patchStrides[0] = 1;
+      for (int i = 1; i < NumDims-1; ++i) {
+        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+        m_patchStrides[i] = m_patchStrides[i-1] * (input_dims[i-1] - patch_dims[i-1] + 1);
+      }
+      m_outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
+      }
+    } else {
+      for (int i = 0; i < NumDims-1; ++i) {
+        m_dimensions[i+1] = patch_dims[i];
+        num_patches *= (input_dims[i] - patch_dims[i] + 1);
+      }
+      m_dimensions[0] = num_patches;
+
+      m_inputStrides[NumDims-2] = 1;
+      m_patchStrides[NumDims-2] = 1;
+      for (int i = NumDims-3; i >= 0; --i) {
+        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
+        m_patchStrides[i] = m_patchStrides[i+1] * (input_dims[i+1] - patch_dims[i+1] + 1);
+      }
+      m_outputStrides[NumDims-1] = 1;
+      for (int i = NumDims-2; i >= 0; --i) {
+        m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    Index output_stride_index = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? NumDims - 1 : 0;
+    // Find the location of the first element of the patch.
+    Index patchIndex = index / m_outputStrides[output_stride_index];
+    // Find the offset of the element wrt the location of the first element.
+    Index patchOffset = index - patchIndex * m_outputStrides[output_stride_index];
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 2; i > 0; --i) {
+        const Index patchIdx = patchIndex / m_patchStrides[i];
+        patchIndex -= patchIdx * m_patchStrides[i];
+        const Index offsetIdx = patchOffset / m_outputStrides[i];
+        patchOffset -= offsetIdx * m_outputStrides[i];
+        inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i];
+      }
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 2; ++i) {
+        const Index patchIdx = patchIndex / m_patchStrides[i];
+        patchIndex -= patchIdx * m_patchStrides[i];
+        const Index offsetIdx = patchOffset / m_outputStrides[i+1];
+        patchOffset -= offsetIdx * m_outputStrides[i+1];
+        inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i];
+      }
+    }
+    inputIndex += (patchIndex + patchOffset);
+    return m_impl.coeff(inputIndex);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    Index output_stride_index = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? NumDims - 1 : 0;
+    Index indices[2] = {index, index + PacketSize - 1};
+    Index patchIndices[2] = {indices[0] / m_outputStrides[output_stride_index],
+                             indices[1] / m_outputStrides[output_stride_index]};
+    Index patchOffsets[2] = {indices[0] - patchIndices[0] * m_outputStrides[output_stride_index],
+                             indices[1] - patchIndices[1] * m_outputStrides[output_stride_index]};
+
+    Index inputIndices[2] = {0, 0};
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 2; i > 0; --i) {
+        const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i],
+                                   patchIndices[1] / m_patchStrides[i]};
+        patchIndices[0] -= patchIdx[0] * m_patchStrides[i];
+        patchIndices[1] -= patchIdx[1] * m_patchStrides[i];
+
+        const Index offsetIdx[2] = {patchOffsets[0] / m_outputStrides[i],
+                                    patchOffsets[1] / m_outputStrides[i]};
+        patchOffsets[0] -= offsetIdx[0] * m_outputStrides[i];
+        patchOffsets[1] -= offsetIdx[1] * m_outputStrides[i];
+
+        inputIndices[0] += (patchIdx[0] + offsetIdx[0]) * m_inputStrides[i];
+        inputIndices[1] += (patchIdx[1] + offsetIdx[1]) * m_inputStrides[i];
+      }
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 2; ++i) {
+        const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i],
+                                   patchIndices[1] / m_patchStrides[i]};
+        patchIndices[0] -= patchIdx[0] * m_patchStrides[i];
+        patchIndices[1] -= patchIdx[1] * m_patchStrides[i];
+
+        const Index offsetIdx[2] = {patchOffsets[0] / m_outputStrides[i+1],
+                                    patchOffsets[1] / m_outputStrides[i+1]};
+        patchOffsets[0] -= offsetIdx[0] * m_outputStrides[i+1];
+        patchOffsets[1] -= offsetIdx[1] * m_outputStrides[i+1];
+
+        inputIndices[0] += (patchIdx[0] + offsetIdx[0]) * m_inputStrides[i];
+        inputIndices[1] += (patchIdx[1] + offsetIdx[1]) * m_inputStrides[i];
+      }
+    }
+    inputIndices[0] += (patchIndices[0] + patchOffsets[0]);
+    inputIndices[1] += (patchIndices[1] + patchOffsets[1]);
+
+    if (inputIndices[1] - inputIndices[0] == PacketSize - 1) {
+      PacketReturnType rslt = m_impl.template packet<Unaligned>(inputIndices[0]);
+      return rslt;
+    }
+    else {
+      EIGEN_ALIGN_MAX CoeffReturnType values[PacketSize];
+      values[0] = m_impl.coeff(inputIndices[0]);
+      values[PacketSize-1] = m_impl.coeff(inputIndices[1]);
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < PacketSize-1; ++i) {
+        values[i] = coeff(index+i);
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    const double compute_cost = NumDims * (TensorOpCost::DivCost<Index>() +
+                                           TensorOpCost::MulCost<Index>() +
+                                           2 * TensorOpCost::AddCost<Index>());
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh); 
+  }
+#endif
+
+ protected:
+  Dimensions m_dimensions;
+  array<Index, NumDims> m_outputStrides;
+  array<Index, NumDims-1> m_inputStrides;
+  array<Index, NumDims-1> m_patchStrides;
+
+  TensorEvaluator<ArgType, Device> m_impl;
+
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_PATCH_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorRandom.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorRandom.h
new file mode 100644
index 0000000..37c1d1c
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorRandom.h
@@ -0,0 +1,322 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2018 Mehdi Goli <eigen@codeplay.com> Codeplay Software Ltd.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_RANDOM_H
+#define EIGEN_CXX11_TENSOR_TENSOR_RANDOM_H
+
+namespace Eigen {
+namespace internal {
+
+namespace {
+
+EIGEN_DEVICE_FUNC uint64_t get_random_seed() {
+#if defined(EIGEN_GPU_COMPILE_PHASE)
+  // We don't support 3d kernels since we currently only use 1 and
+  // 2d kernels.
+  gpu_assert(threadIdx.z == 0);
+  return blockIdx.x * blockDim.x + threadIdx.x 
+         + gridDim.x * blockDim.x * (blockIdx.y * blockDim.y + threadIdx.y);
+#else
+  // Rely on Eigen's random implementation.
+  return random<uint64_t>();
+#endif
+}
+
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE unsigned PCG_XSH_RS_generator(uint64_t* state, uint64_t stream) {
+  // TODO: Unify with the implementation in the non blocking thread pool.
+  uint64_t current = *state;
+  // Update the internal state
+  *state = current * 6364136223846793005ULL + (stream << 1 | 1);
+  // Generate the random output (using the PCG-XSH-RS scheme)
+  return static_cast<unsigned>((current ^ (current >> 22)) >> (22 + (current >> 61)));
+}
+
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE uint64_t PCG_XSH_RS_state(uint64_t seed) {
+  seed = seed ? seed : get_random_seed();
+  return seed * 6364136223846793005ULL + 0xda3e39cb94b95bdbULL;
+}
+
+}  // namespace
+
+
+template <typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+T RandomToTypeUniform(uint64_t* state, uint64_t stream) {
+  unsigned rnd = PCG_XSH_RS_generator(state, stream);
+  return static_cast<T>(rnd);
+}
+
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+Eigen::half RandomToTypeUniform<Eigen::half>(uint64_t* state, uint64_t stream) {
+  // Generate 10 random bits for the mantissa, merge with exponent.
+  unsigned rnd = PCG_XSH_RS_generator(state, stream);
+  const uint16_t half_bits = static_cast<uint16_t>(rnd & 0x3ffu) | (static_cast<uint16_t>(15) << 10);
+  Eigen::half result = Eigen::numext::bit_cast<Eigen::half>(half_bits);
+  // Return the final result
+  return result - Eigen::half(1.0f);
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+Eigen::bfloat16 RandomToTypeUniform<Eigen::bfloat16>(uint64_t* state, uint64_t stream) {
+
+  // Generate 7 random bits for the mantissa, merge with exponent.
+  unsigned rnd = PCG_XSH_RS_generator(state, stream);
+  const uint16_t half_bits = static_cast<uint16_t>(rnd & 0x7fu) | (static_cast<uint16_t>(127) << 7);
+  Eigen::bfloat16 result = Eigen::numext::bit_cast<Eigen::bfloat16>(half_bits);
+  // Return the final result
+  return result - Eigen::bfloat16(1.0f);
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float RandomToTypeUniform<float>(uint64_t* state, uint64_t stream) {
+  typedef union {
+    uint32_t raw;
+    float fp;
+  } internal;
+  internal result;
+  // Generate 23 random bits for the mantissa mantissa
+  const unsigned rnd = PCG_XSH_RS_generator(state, stream);
+  result.raw = rnd & 0x7fffffu;
+  // Set the exponent
+  result.raw |= (static_cast<uint32_t>(127) << 23);
+  // Return the final result
+  return result.fp - 1.0f;
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double RandomToTypeUniform<double>(uint64_t* state, uint64_t stream) {
+  typedef union {
+    uint64_t raw;
+    double dp;
+  } internal;
+  internal result;
+  result.raw = 0;
+  // Generate 52 random bits for the mantissa
+  // First generate the upper 20 bits
+  unsigned rnd1 = PCG_XSH_RS_generator(state, stream) & 0xfffffu;
+  // The generate the lower 32 bits
+  unsigned rnd2 = PCG_XSH_RS_generator(state, stream);
+  result.raw = (static_cast<uint64_t>(rnd1) << 32) | rnd2;
+  // Set the exponent
+  result.raw |= (static_cast<uint64_t>(1023) << 52);
+  // Return the final result
+  return result.dp - 1.0;
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+std::complex<float> RandomToTypeUniform<std::complex<float> >(uint64_t* state, uint64_t stream) {
+  return std::complex<float>(RandomToTypeUniform<float>(state, stream),
+                             RandomToTypeUniform<float>(state, stream));
+}
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+std::complex<double> RandomToTypeUniform<std::complex<double> >(uint64_t* state, uint64_t stream) {
+  return std::complex<double>(RandomToTypeUniform<double>(state, stream),
+                              RandomToTypeUniform<double>(state, stream));
+}
+
+template <typename T> class UniformRandomGenerator {
+ public:
+  static const bool PacketAccess = true;
+
+  // Uses the given "seed" if non-zero, otherwise uses a random seed.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE UniformRandomGenerator(
+      uint64_t seed = 0) {
+    m_state = PCG_XSH_RS_state(seed);
+    #ifdef EIGEN_USE_SYCL
+    // In SYCL it is not possible to build PCG_XSH_RS_state in one step.
+    // Therefor, we need two step to initializate the m_state.
+    // IN SYCL, the constructor of the functor is s called on the CPU
+    // and we get the clock seed here from the CPU. However, This seed is
+    //the same for all the thread. As unlike CUDA, the thread.ID, BlockID, etc is not a global function.
+    // and only  available on the Operator() function (which is called on the GPU).
+    // Thus for CUDA (((CLOCK  + global_thread_id)* 6364136223846793005ULL) + 0xda3e39cb94b95bdbULL) is passed to each thread
+    // but for SYCL ((CLOCK * 6364136223846793005ULL) + 0xda3e39cb94b95bdbULL) is passed to each thread and each thread adds
+    // the  (global_thread_id* 6364136223846793005ULL) for itself only once, in order to complete the construction
+    // similar to CUDA Therefore, the thread Id injection is not available at this stage.
+    //However when the operator() is called the thread ID will be avilable. So inside the opeator,
+    // we add the thrreadID, BlockId,... (which is equivalent of i)
+    //to the seed and construct the unique m_state per thead similar to cuda.
+    m_exec_once =false;
+   #endif
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE UniformRandomGenerator(
+      const UniformRandomGenerator& other) {
+    m_state = other.m_state;
+    #ifdef EIGEN_USE_SYCL
+     m_exec_once =other.m_exec_once;
+    #endif
+  }
+
+  template<typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T operator()(Index i) const {
+    #ifdef EIGEN_USE_SYCL
+      if(!m_exec_once) {
+      // This is the second stage of adding thread Id to the CPU clock seed and build unique seed per thread
+      // The (i * 6364136223846793005ULL) is the remaining part of the PCG_XSH_RS_state on the GPU side
+       m_state += (i * 6364136223846793005ULL);
+       m_exec_once =true;
+      }
+    #endif
+    T result = RandomToTypeUniform<T>(&m_state, i);
+    return result;
+  }
+
+  template<typename Packet, typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Packet packetOp(Index i) const {
+    const int packetSize = internal::unpacket_traits<Packet>::size;
+    EIGEN_ALIGN_MAX T values[packetSize];
+      #ifdef EIGEN_USE_SYCL
+      if(!m_exec_once) {
+      // This is the second stage of adding thread Id to the CPU clock seed and build unique seed per thread
+       m_state += (i * 6364136223846793005ULL);
+       m_exec_once =true;
+      }
+    #endif
+    EIGEN_UNROLL_LOOP
+    for (int j = 0; j < packetSize; ++j) {
+      values[j] = RandomToTypeUniform<T>(&m_state, i);
+    }
+    return internal::pload<Packet>(values);
+  }
+
+ private:
+  mutable uint64_t m_state;
+  #ifdef EIGEN_USE_SYCL
+  mutable bool m_exec_once;
+  #endif
+};
+
+template <typename Scalar>
+struct functor_traits<UniformRandomGenerator<Scalar> > {
+  enum {
+    // Rough estimate for floating point, multiplied by ceil(sizeof(T) / sizeof(float)).
+    Cost = 12 * NumTraits<Scalar>::AddCost *
+           ((sizeof(Scalar) + sizeof(float) - 1) / sizeof(float)),
+    PacketAccess = UniformRandomGenerator<Scalar>::PacketAccess
+  };
+};
+
+
+
+template <typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+T RandomToTypeNormal(uint64_t* state, uint64_t stream) {
+  // Use the ratio of uniform method to generate numbers following a normal
+  // distribution. See for example Numerical Recipes chapter 7.3.9 for the
+  // details.
+  T u, v, q;
+  do {
+    u = RandomToTypeUniform<T>(state, stream);
+    v = T(1.7156) * (RandomToTypeUniform<T>(state, stream) - T(0.5));
+    const T x = u - T(0.449871);
+    const T y = numext::abs(v) + T(0.386595);
+    q = x*x + y * (T(0.196)*y - T(0.25472)*x);
+  } while (q > T(0.27597) &&
+           (q > T(0.27846) || v*v > T(-4) * numext::log(u) * u*u));
+
+  return v/u;
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+std::complex<float> RandomToTypeNormal<std::complex<float> >(uint64_t* state, uint64_t stream) {
+  return std::complex<float>(RandomToTypeNormal<float>(state, stream),
+                             RandomToTypeNormal<float>(state, stream));
+}
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+std::complex<double> RandomToTypeNormal<std::complex<double> >(uint64_t* state, uint64_t stream) {
+  return std::complex<double>(RandomToTypeNormal<double>(state, stream),
+                              RandomToTypeNormal<double>(state, stream));
+}
+
+
+template <typename T> class NormalRandomGenerator {
+ public:
+  static const bool PacketAccess = true;
+
+  // Uses the given "seed" if non-zero, otherwise uses a random seed.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE NormalRandomGenerator(uint64_t seed = 0) {
+    m_state = PCG_XSH_RS_state(seed);
+    #ifdef EIGEN_USE_SYCL
+    // In SYCL it is not possible to build PCG_XSH_RS_state in one step.
+    // Therefor, we need two steps to initializate the m_state.
+    // IN SYCL, the constructor of the functor is s called on the CPU
+    // and we get the clock seed here from the CPU. However, This seed is
+    //the same for all the thread. As unlike CUDA, the thread.ID, BlockID, etc is not a global function.
+    // and only  available on the Operator() function (which is called on the GPU).
+    // Therefore, the thread Id injection is not available at this stage. However when the operator()
+    //is called the thread ID will be avilable. So inside the opeator,
+    // we add the thrreadID, BlockId,... (which is equivalent of i)
+    //to the seed and construct the unique m_state per thead similar to cuda.
+    m_exec_once =false;
+   #endif
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE NormalRandomGenerator(
+      const NormalRandomGenerator& other) {
+    m_state = other.m_state;
+#ifdef EIGEN_USE_SYCL
+    m_exec_once=other.m_exec_once;
+#endif
+  }
+
+ template<typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T operator()(Index i) const {
+    #ifdef EIGEN_USE_SYCL
+    if(!m_exec_once) {
+      // This is the second stage of adding thread Id to the CPU clock seed and build unique seed per thread
+      m_state += (i * 6364136223846793005ULL);
+      m_exec_once =true;
+    }
+    #endif
+    T result = RandomToTypeNormal<T>(&m_state, i);
+    return result;
+  }
+
+  template<typename Packet, typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Packet packetOp(Index i) const {
+    const int packetSize = internal::unpacket_traits<Packet>::size;
+    EIGEN_ALIGN_MAX T values[packetSize];
+    #ifdef EIGEN_USE_SYCL
+    if(!m_exec_once) {
+      // This is the second stage of adding thread Id to the CPU clock seed and build unique seed per thread
+      m_state += (i * 6364136223846793005ULL);
+      m_exec_once =true;
+    }
+    #endif
+    EIGEN_UNROLL_LOOP
+    for (int j = 0; j < packetSize; ++j) {
+      values[j] = RandomToTypeNormal<T>(&m_state, i);
+    }
+    return internal::pload<Packet>(values);
+  }
+
+ private:
+  mutable uint64_t m_state;
+   #ifdef EIGEN_USE_SYCL
+  mutable bool m_exec_once;
+  #endif
+};
+
+
+template <typename Scalar>
+struct functor_traits<NormalRandomGenerator<Scalar> > {
+  enum {
+    // On average, we need to generate about 3 random numbers
+    // 15 mul, 8 add, 1.5 logs
+    Cost = 3 * functor_traits<UniformRandomGenerator<Scalar> >::Cost +
+           15 * NumTraits<Scalar>::AddCost + 8 * NumTraits<Scalar>::AddCost +
+           3 * functor_traits<scalar_log_op<Scalar> >::Cost / 2,
+    PacketAccess = NormalRandomGenerator<Scalar>::PacketAccess
+  };
+};
+
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_RANDOM_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorReduction.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorReduction.h
new file mode 100644
index 0000000..583f462
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorReduction.h
@@ -0,0 +1,998 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2016 Mehdi Goli, Codeplay Software Ltd <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_H
+#define EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_H
+
+// clang is incompatible with the CUDA syntax wrt making a kernel a class friend,
+// so we'll use a macro to make clang happy.
+#ifndef KERNEL_FRIEND
+#if defined(__clang__) && (defined(__CUDA__) || defined(__HIP__))
+#define KERNEL_FRIEND friend __global__ EIGEN_HIP_LAUNCH_BOUNDS_1024
+#else
+#define KERNEL_FRIEND friend
+#endif
+#endif
+
+
+namespace Eigen {
+
+
+/** \class TensorReduction
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor reduction class.
+  *
+  */
+
+namespace internal {
+  template<typename Op, typename Dims, typename XprType,template <class> class MakePointer_ >
+  struct traits<TensorReductionOp<Op, Dims, XprType, MakePointer_> >
+ : traits<XprType>
+{
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::Scalar Scalar;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  static const int NumDimensions = XprTraits::NumDimensions - array_size<Dims>::value;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+
+  template <class T> struct MakePointer {
+    // Intermediate typedef to workaround MSVC issue.
+    typedef MakePointer_<T> MakePointerT;
+    typedef typename MakePointerT::Type Type;
+  };
+};
+
+template<typename Op, typename Dims, typename XprType, template <class> class MakePointer_>
+struct eval<TensorReductionOp<Op, Dims, XprType, MakePointer_>, Eigen::Dense>
+{
+  typedef const TensorReductionOp<Op, Dims, XprType, MakePointer_>& type;
+};
+
+template<typename Op, typename Dims, typename XprType, template <class> class MakePointer_>
+struct nested<TensorReductionOp<Op, Dims, XprType, MakePointer_>, 1, typename eval<TensorReductionOp<Op, Dims, XprType, MakePointer_> >::type>
+{
+  typedef TensorReductionOp<Op, Dims, XprType, MakePointer_> type;
+};
+
+
+template <typename OutputDims> struct DimInitializer {
+  template <typename InputDims, typename ReducedDims> EIGEN_DEVICE_FUNC
+  static void run(const InputDims& input_dims,
+                  const array<bool, internal::array_size<InputDims>::value>& reduced,
+                  OutputDims* output_dims, ReducedDims* reduced_dims) {
+    const int NumInputDims = internal::array_size<InputDims>::value;
+    int outputIndex = 0;
+    int reduceIndex = 0;
+    for (int i = 0; i < NumInputDims; ++i) {
+      if (reduced[i]) {
+        (*reduced_dims)[reduceIndex] = input_dims[i];
+        ++reduceIndex;
+      } else {
+        (*output_dims)[outputIndex] = input_dims[i];
+        ++outputIndex;
+      }
+    }
+  }
+};
+
+template <> struct DimInitializer<Sizes<> > {
+  template <typename InputDims, typename Index, size_t Rank> EIGEN_DEVICE_FUNC
+  static void run(const InputDims& input_dims, const array<bool, Rank>&,
+                  Sizes<>*, array<Index, Rank>* reduced_dims) {
+    const int NumInputDims = internal::array_size<InputDims>::value;
+    for (int i = 0; i < NumInputDims; ++i) {
+      (*reduced_dims)[i] = input_dims[i];
+    }
+  }
+};
+
+
+template <typename ReducedDims, int NumTensorDims, int Layout>
+struct are_inner_most_dims {
+  static const bool value = false;
+};
+template <typename ReducedDims, int NumTensorDims, int Layout>
+struct preserve_inner_most_dims {
+  static const bool value = false;
+};
+
+#if EIGEN_HAS_CONSTEXPR && EIGEN_HAS_VARIADIC_TEMPLATES
+template <typename ReducedDims, int NumTensorDims>
+struct are_inner_most_dims<ReducedDims, NumTensorDims, ColMajor>{
+  static const bool tmp1 = indices_statically_known_to_increase<ReducedDims>();
+  static const bool tmp2 = index_statically_eq<ReducedDims>(0, 0);
+  static const bool tmp3 = index_statically_eq<ReducedDims>(array_size<ReducedDims>::value-1, array_size<ReducedDims>::value-1);
+  static const bool value = tmp1 & tmp2 & tmp3;
+};
+template <typename ReducedDims, int NumTensorDims>
+struct are_inner_most_dims<ReducedDims, NumTensorDims, RowMajor>{
+  static const bool tmp1 = indices_statically_known_to_increase<ReducedDims>();
+  static const bool tmp2 = index_statically_eq<ReducedDims>(0, NumTensorDims - array_size<ReducedDims>::value);
+  static const bool tmp3 = index_statically_eq<ReducedDims>(array_size<ReducedDims>::value - 1, NumTensorDims - 1);
+  static const bool value = tmp1 & tmp2 & tmp3;
+
+};
+template <typename ReducedDims, int NumTensorDims>
+struct preserve_inner_most_dims<ReducedDims, NumTensorDims, ColMajor>{
+  static const bool tmp1 = indices_statically_known_to_increase<ReducedDims>();
+  static const bool tmp2 = index_statically_gt<ReducedDims>(0, 0);
+  static const bool value = tmp1 & tmp2;
+
+};
+template <typename ReducedDims, int NumTensorDims>
+struct preserve_inner_most_dims<ReducedDims, NumTensorDims, RowMajor>{
+  static const bool tmp1 = indices_statically_known_to_increase<ReducedDims>();
+  static const bool tmp2 = index_statically_lt<ReducedDims>(array_size<ReducedDims>::value - 1, NumTensorDims - 1);
+  static const bool value = tmp1 & tmp2;
+};
+#endif
+
+
+template <int DimIndex, typename Self, typename Op>
+struct GenericDimReducer {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index firstIndex, Op& reducer, typename Self::CoeffReturnType* accum) {
+    EIGEN_STATIC_ASSERT((DimIndex > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    for (int j = 0; j < self.m_reducedDims[DimIndex]; ++j) {
+      const typename Self::Index input = firstIndex + j * self.m_reducedStrides[DimIndex];
+      GenericDimReducer<DimIndex-1, Self, Op>::reduce(self, input, reducer, accum);
+    }
+  }
+};
+template <typename Self, typename Op>
+struct GenericDimReducer<0, Self, Op> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index firstIndex, Op& reducer, typename Self::CoeffReturnType* accum) {
+    for (int j = 0; j < self.m_reducedDims[0]; ++j) {
+      const typename Self::Index input = firstIndex + j * self.m_reducedStrides[0];
+      reducer.reduce(self.m_impl.coeff(input), accum);
+    }
+  }
+};
+template <typename Self, typename Op>
+struct GenericDimReducer<-1, Self, Op> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index index, Op& reducer, typename Self::CoeffReturnType* accum) {
+    reducer.reduce(self.m_impl.coeff(index), accum);
+  }
+};
+
+template <typename Self, typename Op, bool Vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess),
+          bool UseTreeReduction = (!Self::ReducerTraits::IsStateful &&
+                                   !Self::ReducerTraits::IsExactlyAssociative)>
+struct InnerMostDimReducer {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType reduce(const Self& self, typename Self::Index firstIndex, typename Self::Index numValuesToReduce, Op& reducer) {
+    typename Self::CoeffReturnType accum = reducer.initialize();
+    for (typename Self::Index j = 0; j < numValuesToReduce; ++j) {
+      reducer.reduce(self.m_impl.coeff(firstIndex + j), &accum);
+    }
+    return reducer.finalize(accum);
+  }
+};
+
+template <typename Self, typename Op>
+struct InnerMostDimReducer<Self, Op, true, false> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType reduce(const Self& self, typename Self::Index firstIndex, typename Self::Index numValuesToReduce, Op& reducer) {
+    const typename Self::Index packetSize = internal::unpacket_traits<typename Self::PacketReturnType>::size;
+    const typename Self::Index VectorizedSize = (numValuesToReduce / packetSize) * packetSize;
+    typename Self::PacketReturnType paccum = reducer.template initializePacket<typename Self::PacketReturnType>();
+    for (typename Self::Index j = 0; j < VectorizedSize; j += packetSize) {
+      reducer.reducePacket(self.m_impl.template packet<Unaligned>(firstIndex + j), &paccum);
+    }
+    typename Self::CoeffReturnType accum = reducer.initialize();
+    for (typename Self::Index j = VectorizedSize; j < numValuesToReduce; ++j) {
+      reducer.reduce(self.m_impl.coeff(firstIndex + j), &accum);
+    }
+    return reducer.finalizeBoth(accum, paccum);
+  }
+};
+
+#if !defined(EIGEN_HIPCC) 
+static const int kLeafSize = 1024;
+
+template <typename Self, typename Op>
+struct InnerMostDimReducer<Self, Op, false, true> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType
+  reduce(const Self& self, typename Self::Index firstIndex,
+         typename Self::Index numValuesToReduce, Op& reducer) {
+    typename Self::CoeffReturnType accum = reducer.initialize();
+    if (numValuesToReduce > kLeafSize) {
+      const typename Self::Index half = numValuesToReduce / 2;
+      reducer.reduce(reduce(self, firstIndex, half, reducer), &accum);
+      reducer.reduce(
+          reduce(self, firstIndex + half, numValuesToReduce - half, reducer),
+          &accum);
+    } else {
+      for (typename Self::Index j = 0; j < numValuesToReduce; ++j) {
+        reducer.reduce(self.m_impl.coeff(firstIndex + j), &accum);
+      }
+    }
+    return reducer.finalize(accum);
+  }
+};
+
+template <typename Self, typename Op>
+struct InnerMostDimReducer<Self, Op, true, true> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType
+  reduce(const Self& self, typename Self::Index firstIndex,
+         typename Self::Index numValuesToReduce, Op& reducer) {
+    const typename Self::Index packetSize =
+        internal::unpacket_traits<typename Self::PacketReturnType>::size;
+    typename Self::CoeffReturnType accum = reducer.initialize();
+    if (numValuesToReduce > packetSize * kLeafSize) {
+      // Make sure the split point is aligned on a packet boundary.
+      const typename Self::Index split =
+          packetSize *
+          divup(firstIndex + divup(numValuesToReduce, typename Self::Index(2)),
+                packetSize);
+      const typename Self::Index num_left =
+          numext::mini(split - firstIndex, numValuesToReduce);
+      reducer.reduce(reduce(self, firstIndex, num_left, reducer), &accum);
+      if (num_left < numValuesToReduce) {
+        reducer.reduce(
+            reduce(self, split, numValuesToReduce - num_left, reducer), &accum);
+      }
+      return reducer.finalize(accum);
+    } else {
+      const typename Self::Index UnrollSize =
+          (numValuesToReduce / (2*packetSize)) * 2*packetSize;
+      const typename Self::Index VectorizedSize =
+          (numValuesToReduce / packetSize) * packetSize;
+      typename Self::PacketReturnType paccum =
+          reducer.template initializePacket<typename Self::PacketReturnType>();
+      typename Self::PacketReturnType paccum2 =
+          reducer.template initializePacket<typename Self::PacketReturnType>();
+      for (typename Self::Index j = 0; j < UnrollSize; j += packetSize * 2) {
+        reducer.reducePacket(
+            self.m_impl.template packet<Unaligned>(firstIndex + j), &paccum);
+        reducer.reducePacket(
+            self.m_impl.template packet<Unaligned>(firstIndex + j + packetSize),
+            &paccum2);
+      }
+      for (typename Self::Index j = UnrollSize; j < VectorizedSize; j+= packetSize) {
+        reducer.reducePacket(self.m_impl.template packet<Unaligned>(
+                                 firstIndex + j), &paccum);
+      }
+      reducer.reducePacket(paccum2, &paccum);
+      for (typename Self::Index j = VectorizedSize; j < numValuesToReduce;
+           ++j) {
+        reducer.reduce(self.m_impl.coeff(firstIndex + j), &accum);
+      }
+      return reducer.finalizeBoth(accum, paccum);
+    }
+  }
+};
+#endif
+ 
+template <int DimIndex, typename Self, typename Op, bool vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess)>
+struct InnerMostDimPreserver {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self&, typename Self::Index, Op&, typename Self::PacketReturnType*) {
+    eigen_assert(false && "should never be called");
+  }
+};
+
+template <int DimIndex, typename Self, typename Op>
+struct InnerMostDimPreserver<DimIndex, Self, Op, true> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index firstIndex, Op& reducer, typename Self::PacketReturnType* accum) {
+    EIGEN_STATIC_ASSERT((DimIndex > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    for (typename Self::Index j = 0; j < self.m_reducedDims[DimIndex]; ++j) {
+      const typename Self::Index input = firstIndex + j * self.m_reducedStrides[DimIndex];
+      InnerMostDimPreserver<DimIndex-1, Self, Op>::reduce(self, input, reducer, accum);
+    }
+  }
+};
+
+template <typename Self, typename Op>
+struct InnerMostDimPreserver<0, Self, Op, true> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index firstIndex, Op& reducer, typename Self::PacketReturnType* accum) {
+    for (typename Self::Index j = 0; j < self.m_reducedDims[0]; ++j) {
+      const typename Self::Index input = firstIndex + j * self.m_reducedStrides[0];
+      reducer.reducePacket(self.m_impl.template packet<Unaligned>(input), accum);
+    }
+  }
+};
+template <typename Self, typename Op>
+struct InnerMostDimPreserver<-1, Self, Op, true> {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self&, typename Self::Index, Op&, typename Self::PacketReturnType*) {
+    eigen_assert(false && "should never be called");
+  }
+};
+
+// Default full reducer
+template <typename Self, typename Op, typename Device, bool Vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess)>
+struct FullReducer {
+  static const bool HasOptimizedImplementation = false;
+
+  static EIGEN_DEVICE_FUNC void run(const Self& self, Op& reducer, const Device&, typename Self::EvaluatorPointerType output) {
+    const typename Self::Index num_coeffs = array_prod(self.m_impl.dimensions());
+    *output = InnerMostDimReducer<Self, Op, Vectorizable>::reduce(self, 0, num_coeffs, reducer);
+  }
+};
+
+
+#ifdef EIGEN_USE_THREADS
+// Multithreaded full reducers
+template <typename Self, typename Op,
+          bool Vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess)>
+struct FullReducerShard {
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(const Self& self, typename Self::Index firstIndex,
+                  typename Self::Index numValuesToReduce, Op& reducer,
+                  typename Self::CoeffReturnType* output) {
+    *output = InnerMostDimReducer<Self, Op, Vectorizable>::reduce(
+        self, firstIndex, numValuesToReduce, reducer);
+  }
+};
+
+// Multithreaded full reducer
+template <typename Self, typename Op, bool Vectorizable>
+struct FullReducer<Self, Op, ThreadPoolDevice, Vectorizable> {
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful;
+  static const Index PacketSize =
+      unpacket_traits<typename Self::PacketReturnType>::size;
+
+  // launch one reducer per thread and accumulate the result.
+  static void run(const Self& self, Op& reducer, const ThreadPoolDevice& device,
+                  typename Self::CoeffReturnType* output) {
+    typedef typename Self::Index Index;
+    const Index num_coeffs = array_prod(self.m_impl.dimensions());
+    if (num_coeffs == 0) {
+      *output = reducer.finalize(reducer.initialize());
+      return;
+    }
+    const TensorOpCost cost =
+        self.m_impl.costPerCoeff(Vectorizable) +
+        TensorOpCost(0, 0, internal::functor_traits<Op>::Cost, Vectorizable,
+                     PacketSize);
+    const int num_threads = TensorCostModel<ThreadPoolDevice>::numThreads(
+        num_coeffs, cost, device.numThreads());
+    if (num_threads == 1) {
+      *output =
+          InnerMostDimReducer<Self, Op, Vectorizable>::reduce(self, 0, num_coeffs, reducer);
+      return;
+    }
+    const Index blocksize =
+        std::floor<Index>(static_cast<float>(num_coeffs) / num_threads);
+    const Index numblocks = blocksize > 0 ? num_coeffs / blocksize : 0;
+    eigen_assert(num_coeffs >= numblocks * blocksize);
+
+    Barrier barrier(internal::convert_index<unsigned int>(numblocks));
+    MaxSizeVector<typename Self::CoeffReturnType> shards(numblocks, reducer.initialize());
+    for (Index i = 0; i < numblocks; ++i) {
+      device.enqueue_with_barrier(&barrier, &FullReducerShard<Self, Op, Vectorizable>::run,
+                                  self, i * blocksize, blocksize, reducer,
+                                  &shards[i]);
+    }
+    typename Self::CoeffReturnType finalShard;
+    if (numblocks * blocksize < num_coeffs) {
+      finalShard = InnerMostDimReducer<Self, Op, Vectorizable>::reduce(
+          self, numblocks * blocksize, num_coeffs - numblocks * blocksize,
+          reducer);
+    } else {
+      finalShard = reducer.initialize();
+    }
+    barrier.Wait();
+
+    for (Index i = 0; i < numblocks; ++i) {
+      reducer.reduce(shards[i], &finalShard);
+    }
+    *output = reducer.finalize(finalShard);
+  }
+};
+
+#endif
+
+
+// Default inner reducer
+template <typename Self, typename Op, typename Device>
+struct InnerReducer {
+  static const bool HasOptimizedImplementation = false;
+
+  EIGEN_DEVICE_FUNC static bool run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
+    eigen_assert(false && "Not implemented");
+    return true;
+  }
+};
+
+// Default outer reducer
+template <typename Self, typename Op, typename Device>
+struct OuterReducer {
+  static const bool HasOptimizedImplementation = false;
+
+  EIGEN_DEVICE_FUNC static bool run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
+    eigen_assert(false && "Not implemented");
+    return true;
+  }
+};
+
+#ifdef EIGEN_USE_SYCL
+// Default Generic reducer
+template <typename Self, typename Op, typename Device>
+struct GenericReducer {
+  static const bool HasOptimizedImplementation = false;
+
+  EIGEN_DEVICE_FUNC static bool run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
+    eigen_assert(false && "Not implemented");
+    return true;
+  }
+};
+#endif
+
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
+template <int B, int N, typename S, typename R, typename I_>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void FullReductionKernel(R, const S, I_, typename S::CoeffReturnType*, unsigned int*);
+
+
+#if defined(EIGEN_HAS_GPU_FP16)
+template <typename S, typename R, typename I_>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void ReductionInitFullReduxKernelHalfFloat(R, const S, I_, internal::packet_traits<half>::type*);
+template <int B, int N, typename S, typename R, typename I_>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void FullReductionKernelHalfFloat(R, const S, I_, half*, internal::packet_traits<half>::type*);
+template <int NPT, typename S, typename R, typename I_>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void InnerReductionKernelHalfFloat(R, const S, I_, I_, half*);
+
+#endif
+
+template <int NPT, typename S, typename R, typename I_>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void InnerReductionKernel(R, const S, I_, I_, typename S::CoeffReturnType*);
+
+template <int NPT, typename S, typename R, typename I_>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void OuterReductionKernel(R, const S, I_, I_, typename S::CoeffReturnType*);
+#endif
+
+/**
+ * For SYCL, the return type of the reduction is deduced from the initialize method of the given Op.
+ * This allows the reduction to have a different type for the accumulator than the input data type.
+ * If this is the case, the functor needs to have two reduce method: one for reducing an element of the input
+ * with the accumulator and the other for reducing two accumulators.
+ * Such a reducer can be useful for instance when the accumulator is a boolean or a bitset that checks for
+ * some properties of the input.
+ */
+template <typename Op, typename CoeffReturnType>
+struct ReductionReturnType {
+#if defined(EIGEN_USE_SYCL)
+  typedef typename remove_const<decltype(std::declval<Op>().initialize())>::type type;
+#else
+  typedef typename remove_const<CoeffReturnType>::type type;
+#endif
+};
+
+}  // end namespace internal
+
+
+template <typename Op, typename Dims, typename XprType,  template <class> class MakePointer_>
+class TensorReductionOp : public TensorBase<TensorReductionOp<Op, Dims, XprType, MakePointer_>, ReadOnlyAccessors> {
+  public:
+    typedef typename Eigen::internal::traits<TensorReductionOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorReductionOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorReductionOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorReductionOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorReductionOp(const XprType& expr, const Dims& dims) : m_expr(expr), m_dims(dims)
+    { }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    TensorReductionOp(const XprType& expr, const Dims& dims, const Op& reducer) : m_expr(expr), m_dims(dims), m_reducer(reducer)
+    { }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const XprType& expression() const { return m_expr; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const Dims& dims() const { return m_dims; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const Op& reducer() const { return m_reducer; }
+
+  protected:
+    typename XprType::Nested m_expr;
+    const Dims m_dims;
+    const Op m_reducer;
+};
+
+template<typename ArgType, typename Device>
+struct TensorReductionEvaluatorBase;
+
+// Eval as rvalue
+template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device>
+struct TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>
+{
+  typedef internal::reducer_traits<Op, Device> ReducerTraits;
+  typedef Dims ReducedDims;
+  typedef TensorReductionOp<Op, Dims, ArgType, MakePointer_> XprType;
+  typedef typename XprType::Index Index;
+  typedef ArgType ChildType;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions;
+  static const int NumInputDims = internal::array_size<InputDimensions>::value;
+  static const int NumReducedDims = internal::array_size<Dims>::value;
+  static const int NumOutputDims = NumInputDims - NumReducedDims;
+  typedef typename internal::conditional<NumOutputDims==0, Sizes<>, DSizes<Index, NumOutputDims> >::type Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device> Self;
+  static const bool InputPacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess;
+  typedef typename internal::ReductionReturnType<Op, typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const Index PacketSize = PacketType<CoeffReturnType, Device>::size;
+
+  typedef typename Eigen::internal::traits<XprType>::PointerType TensorPointerType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+    // Subset of strides of the input tensor for the non-reduced dimensions.
+  // Indexed by output dimensions.
+  static const int NumPreservedStrides = max_n_1<NumOutputDims>::size;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = Self::InputPacketAccess && ReducerTraits::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = true,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  static const bool ReducingInnerMostDims = internal::are_inner_most_dims<Dims, NumInputDims, Layout>::value;
+  static const bool PreservingInnerMostDims = internal::preserve_inner_most_dims<Dims, NumInputDims, Layout>::value;
+  static const bool RunningFullReduction = (NumOutputDims==0);
+
+  EIGEN_STRONG_INLINE TensorReductionEvaluatorBase(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_reducer(op.reducer()), m_result(NULL), m_device(device)
+  {
+    EIGEN_STATIC_ASSERT((NumInputDims >= NumReducedDims), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT((!ReducingInnerMostDims | !PreservingInnerMostDims | (NumReducedDims == NumInputDims)),
+                        YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    // Build the bitmap indicating if an input dimension is reduced or not.
+    for (int i = 0; i < NumInputDims; ++i) {
+      m_reduced[i] = false;
+    }
+    for (int i = 0; i < NumReducedDims; ++i) {
+      eigen_assert(op.dims()[i] >= 0);
+      eigen_assert(op.dims()[i] < NumInputDims);
+      m_reduced[op.dims()[i]] = true;
+    }
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    internal::DimInitializer<Dimensions>::run(input_dims, m_reduced, &m_dimensions, &m_reducedDims);
+
+    // Precompute output strides.
+    if (NumOutputDims > 0) {
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        m_outputStrides[0] = 1;
+        for (int i = 1; i < NumOutputDims; ++i) {
+          m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
+          m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
+        }
+      } else {
+        m_outputStrides[NumOutputDims - 1] = 1;
+        for (int i = NumOutputDims - 2; i >= 0; --i) {
+          m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
+          m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
+        }
+      }
+    }
+
+    // Precompute input strides.
+    if (NumInputDims > 0) {
+      array<Index, NumInputDims> input_strides;
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        input_strides[0] = 1;
+        for (int i = 1; i < NumInputDims; ++i) {
+          input_strides[i] = input_strides[i-1] * input_dims[i-1];
+        }
+      } else {
+        input_strides.back() = 1;
+        for (int i = NumInputDims - 2; i >= 0; --i) {
+          input_strides[i] = input_strides[i + 1] * input_dims[i + 1];
+        }
+      }
+
+      int outputIndex = 0;
+      int reduceIndex = 0;
+      for (int i = 0; i < NumInputDims; ++i) {
+        if (m_reduced[i]) {
+          m_reducedStrides[reduceIndex] = input_strides[i];
+          ++reduceIndex;
+        } else {
+          m_preservedStrides[outputIndex] = input_strides[i];
+          m_output_to_input_dim_map[outputIndex] = i;
+          ++outputIndex;
+        }
+      }
+    }
+
+    // Special case for full reductions
+    if (NumOutputDims == 0) {
+      m_preservedStrides[0] = internal::array_prod(input_dims);
+    }
+
+    m_numValuesToReduce =
+        NumOutputDims == 0
+            ? internal::array_prod(input_dims)
+            : (static_cast<int>(Layout) == static_cast<int>(ColMajor))
+                  ? m_preservedStrides[0]
+                  : m_preservedStrides[NumOutputDims - 1];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE
+  bool evalSubExprsIfNeededCommon(EvaluatorPointerType data) {
+    // Use the FullReducer if possible.
+    if ((RunningFullReduction && RunningOnSycl) ||(RunningFullReduction &&
+        internal::FullReducer<Self, Op, Device>::HasOptimizedImplementation &&
+        ((RunningOnGPU && (m_device.majorDeviceVersion() >= 3)) ||
+         !RunningOnGPU))) {
+      bool need_assign = false;
+      if (!data) {
+        m_result = static_cast<EvaluatorPointerType>(m_device.get((CoeffReturnType*)m_device.allocate_temp(sizeof(CoeffReturnType))));
+        data = m_result;
+        need_assign = true;
+      }
+      Op reducer(m_reducer);
+      internal::FullReducer<Self, Op, Device>::run(*this, reducer, m_device, data);
+      return need_assign;
+    }
+
+    // Attempt to use an optimized reduction.
+    else if ((RunningOnGPU && (m_device.majorDeviceVersion() >= 3)) || (RunningOnSycl)) {
+      bool reducing_inner_dims = true;
+      for (int i = 0; i < NumReducedDims; ++i) {
+        if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+          reducing_inner_dims &= m_reduced[i];
+        } else {
+          reducing_inner_dims &= m_reduced[NumInputDims - 1 - i];
+        }
+      }
+      if (internal::InnerReducer<Self, Op, Device>::HasOptimizedImplementation &&
+          (reducing_inner_dims || ReducingInnerMostDims)) {
+        const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
+        const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
+        if (!data) {
+          if ((num_coeffs_to_preserve < 1024 && num_values_to_reduce > num_coeffs_to_preserve && num_values_to_reduce > 128) || (RunningOnSycl)) {
+            data = static_cast<EvaluatorPointerType>(m_device.get((CoeffReturnType*)m_device.allocate_temp(sizeof(CoeffReturnType) * num_coeffs_to_preserve)));
+            m_result = data;
+          }
+          else {
+            return true;
+          }
+        }
+        Op reducer(m_reducer);
+        // For SYCL this if always return false
+        if (internal::InnerReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve)) {
+          if (m_result) {
+            m_device.deallocate_temp(m_result);
+            m_result = NULL;
+          }
+          return true;
+        } else {
+          return (m_result != NULL);
+        }
+      }
+
+      bool preserving_inner_dims = true;
+      for (int i = 0; i < NumReducedDims; ++i) {
+        if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+          preserving_inner_dims &= m_reduced[NumInputDims - 1 - i];
+        } else {
+          preserving_inner_dims &= m_reduced[i];
+        }
+      }
+      if (internal::OuterReducer<Self, Op, Device>::HasOptimizedImplementation &&
+          preserving_inner_dims) {
+        const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
+        const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
+        if (!data) {
+          if ((num_coeffs_to_preserve < 1024 && num_values_to_reduce > num_coeffs_to_preserve && num_values_to_reduce > 32) || (RunningOnSycl)) {
+            data = static_cast<EvaluatorPointerType>(m_device.get((CoeffReturnType*)m_device.allocate_temp(sizeof(CoeffReturnType) * num_coeffs_to_preserve)));
+            m_result = data;
+          }
+          else {
+            return true;
+          }
+        }
+        Op reducer(m_reducer);
+        // For SYCL this if always return false
+        if (internal::OuterReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve)) {
+          if (m_result) {
+            m_device.deallocate_temp(m_result);
+            m_result = NULL;
+          }
+          return true;
+        } else {
+          return (m_result != NULL);
+        }
+      }
+      #if defined(EIGEN_USE_SYCL)
+      // If there is no Optimised version for SYCL, the reduction expression 
+      // must break into two subexpression and use the SYCL generic Reducer on the device.
+      if(RunningOnSycl) {
+         const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
+         const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
+         if (!data) {
+           data = static_cast<EvaluatorPointerType>(m_device.get((CoeffReturnType*)m_device.allocate_temp(sizeof(CoeffReturnType) * num_coeffs_to_preserve)));
+           m_result = data;
+         }
+         Op reducer(m_reducer);
+         internal::GenericReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
+         return (m_result != NULL);
+       }
+      #endif
+    }
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE
+      void
+      evalSubExprsIfNeededAsync(EvaluatorPointerType data,
+                                EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(NULL, [this, data, done](bool) {
+      done(evalSubExprsIfNeededCommon(data));
+    });
+  }
+#endif
+
+  EIGEN_STRONG_INLINE
+  bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return evalSubExprsIfNeededCommon(data);
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+    if (m_result) {
+      m_device.deallocate_temp(m_result);
+      m_result = NULL;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    if (( RunningFullReduction || RunningOnGPU) && m_result ) {
+      return *(m_result + index);
+    }
+    Op reducer(m_reducer);
+    if (ReducingInnerMostDims || RunningFullReduction) {
+      const Index num_values_to_reduce =
+        (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? m_preservedStrides[0] : m_preservedStrides[NumPreservedStrides - 1];
+      return internal::InnerMostDimReducer<Self, Op>::reduce(*this, firstInput(index),
+                                                             num_values_to_reduce, reducer);
+    } else {
+      typename Self::CoeffReturnType accum = reducer.initialize();
+      internal::GenericDimReducer<NumReducedDims-1, Self, Op>::reduce(*this, firstInput(index), reducer, &accum);
+      return reducer.finalize(accum);
+    }
+  }
+
+  // TODO(bsteiner): provide a more efficient implementation.
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index + PacketSize - 1 < Index(internal::array_prod(dimensions())));
+
+    if (RunningOnGPU && m_result) {
+      return internal::pload<PacketReturnType>(m_result + index);
+    }
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    if (ReducingInnerMostDims) {
+      const Index num_values_to_reduce =
+        (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? m_preservedStrides[0] : m_preservedStrides[NumPreservedStrides - 1];
+      const Index firstIndex = firstInput(index);
+      for (Index i = 0; i < PacketSize; ++i) {
+        Op reducer(m_reducer);
+        values[i] = internal::InnerMostDimReducer<Self, Op>::reduce(*this, firstIndex + i * num_values_to_reduce,
+                                                                    num_values_to_reduce, reducer);
+      }
+    } else if (PreservingInnerMostDims) {
+      const Index firstIndex = firstInput(index);
+      const int innermost_dim = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? 0 : NumOutputDims - 1;
+      // TBD: extend this the the n innermost dimensions that we preserve.
+      if (((firstIndex % m_dimensions[innermost_dim]) + PacketSize - 1) < m_dimensions[innermost_dim]) {
+        Op reducer(m_reducer);
+        typename Self::PacketReturnType accum = reducer.template initializePacket<typename Self::PacketReturnType>();
+        internal::InnerMostDimPreserver<NumReducedDims-1, Self, Op>::reduce(*this, firstIndex, reducer, &accum);
+        return reducer.finalizePacket(accum);
+      } else {
+        for (int i = 0; i < PacketSize; ++i) {
+          values[i] = coeff(index + i);
+        }
+      }
+    } else {
+      for (int i = 0; i < PacketSize; ++i) {
+        values[i] = coeff(index + i);
+      }
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  // Must be called after evalSubExprsIfNeeded().
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    if (RunningFullReduction && m_result) {
+      return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
+    } else {
+      const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
+      const double compute_cost = num_values_to_reduce * internal::functor_traits<Op>::Cost;
+      return m_impl.costPerCoeff(vectorized) * num_values_to_reduce +
+          TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_result; }
+  EIGEN_DEVICE_FUNC const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+  EIGEN_DEVICE_FUNC const Device& device() const { return m_device; }
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+    m_result.bind(cgh);
+  }
+#endif
+
+  private:
+  template <int, typename, typename> friend struct internal::GenericDimReducer;
+  template <typename, typename, bool, bool> friend struct internal::InnerMostDimReducer;
+  template <int, typename, typename, bool> friend struct internal::InnerMostDimPreserver;
+  template <typename S, typename O, typename D, bool V> friend struct internal::FullReducer;
+#ifdef EIGEN_USE_THREADS
+  template <typename S, typename O, bool V> friend struct internal::FullReducerShard;
+#endif
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
+  template <int B, int N, typename S, typename R, typename I_> KERNEL_FRIEND void internal::FullReductionKernel(R, const S, I_, typename S::CoeffReturnType*, unsigned int*);
+#if defined(EIGEN_HAS_GPU_FP16)
+  template <typename S, typename R, typename I_> KERNEL_FRIEND void internal::ReductionInitFullReduxKernelHalfFloat(R, const S, I_, internal::packet_traits<Eigen::half>::type*);
+  template <int B, int N, typename S, typename R, typename I_> KERNEL_FRIEND void internal::FullReductionKernelHalfFloat(R, const S, I_, half*, internal::packet_traits<Eigen::half>::type*);
+  template <int NPT, typename S, typename R, typename I_> KERNEL_FRIEND void internal::InnerReductionKernelHalfFloat(R, const S, I_, I_, half*);
+#endif
+  template <int NPT, typename S, typename R, typename I_> KERNEL_FRIEND void internal::InnerReductionKernel(R, const S, I_, I_, typename S::CoeffReturnType*);
+
+  template <int NPT, typename S, typename R, typename I_> KERNEL_FRIEND void internal::OuterReductionKernel(R, const S, I_, I_, typename S::CoeffReturnType*);
+#endif
+
+#if defined(EIGEN_USE_SYCL)
+ template < typename Evaluator_, typename Op__> friend class TensorSycl::internal::GenericNondeterministicReducer;
+ // SYCL need the Generic reducer for the case the recution algorithm is neither inner, outer, and full reducer
+ template <typename, typename, typename> friend struct internal::GenericReducer;
+#endif
+
+
+  template <typename S, typename O, typename D> friend struct internal::InnerReducer;
+
+  struct BlockIteratorState {
+    Index input_dim;
+    Index output_size;
+    Index output_count;
+  };
+
+  // Returns the Index in the input tensor of the first value that needs to be
+  // used to compute the reduction at output index "index".
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const {
+    if (ReducingInnerMostDims) {
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        return index * m_preservedStrides[0];
+      } else {
+        return index * m_preservedStrides[NumPreservedStrides - 1];
+      }
+    }
+    // TBD: optimize the case where we preserve the innermost dimensions.
+    Index startInput = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumOutputDims - 1; i > 0; --i) {
+        // This is index_i in the output tensor.
+        const Index idx = index / m_outputStrides[i];
+        startInput += idx * m_preservedStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      if (PreservingInnerMostDims) {
+        eigen_assert(m_preservedStrides[0] == 1);
+        startInput += index;
+      } else {
+        startInput += index * m_preservedStrides[0];
+      }
+    } else {
+      for (int i = 0; i < NumOutputDims - 1; ++i) {
+        // This is index_i in the output tensor.
+        const Index idx = index / m_outputStrides[i];
+        startInput += idx * m_preservedStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      if (PreservingInnerMostDims) {
+        eigen_assert(m_preservedStrides[NumPreservedStrides - 1] == 1);
+        startInput += index;
+      } else {
+        startInput += index * m_preservedStrides[NumPreservedStrides - 1];
+      }
+    }
+    return startInput;
+  }
+
+  // Bitmap indicating if an input dimension is reduced or not.
+  array<bool, NumInputDims> m_reduced;
+  // Dimensions of the output of the operation.
+  Dimensions m_dimensions;
+  // Precomputed strides for the output tensor.
+  array<Index, NumOutputDims> m_outputStrides;
+  array<internal::TensorIntDivisor<Index>, NumOutputDims> m_fastOutputStrides;
+  array<Index, NumPreservedStrides> m_preservedStrides;
+  // Map from output to input dimension index.
+  array<Index, NumOutputDims> m_output_to_input_dim_map;
+  // How many values go into each reduction
+  Index m_numValuesToReduce;
+
+  // Subset of strides of the input tensor for the reduced dimensions.
+  // Indexed by reduced dimensions.
+  array<Index, NumReducedDims> m_reducedStrides;
+  // Size of the input dimensions that are reduced.
+  // Indexed by reduced dimensions.
+  array<Index, NumReducedDims> m_reducedDims;
+
+  // Evaluator for the input expression.
+  TensorEvaluator<ArgType, Device> m_impl;
+
+  // Operation to apply for computing the reduction.
+  Op m_reducer;
+
+  // For full reductions
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
+  static const bool RunningOnGPU = internal::is_same<Device, Eigen::GpuDevice>::value;
+  static const bool RunningOnSycl = false;
+#elif defined(EIGEN_USE_SYCL)
+static const bool RunningOnSycl = internal::is_same<typename internal::remove_all<Device>::type, Eigen::SyclDevice>::value;
+static const bool RunningOnGPU = false;
+#else
+  static const bool RunningOnGPU = false;
+  static const bool RunningOnSycl = false;
+#endif
+  EvaluatorPointerType m_result;
+
+  const Device EIGEN_DEVICE_REF m_device;
+};
+
+template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device>
+struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>
+: public TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device> {
+  typedef TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device> Base;
+  EIGEN_STRONG_INLINE TensorEvaluator(const typename Base::XprType& op, const Device& device) : Base(op, device){}
+};
+
+
+template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_>
+struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Eigen::SyclDevice>
+: public TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Eigen::SyclDevice> {
+
+  typedef TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Eigen::SyclDevice> Base;
+  EIGEN_STRONG_INLINE TensorEvaluator(const typename Base::XprType& op, const Eigen::SyclDevice& device) : Base(op, device){}
+  // The coeff function in the base the recursive method which is not an standard layout and cannot be used in the SYCL kernel
+  //Therefore the coeff function should be overridden by for SYCL kernel
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Base::CoeffReturnType coeff(typename Base::Index index) const {
+    return *(this->data() + index);
+  }
+  // The packet function in the base the recursive method which is not an standard layout and cannot be used in the SYCL kernel
+  //Therefore the packet function should be overridden by for SYCL kernel
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Base::PacketReturnType packet(typename Base::Index index) const {
+    return internal::pload<typename Base::PacketReturnType>(this->data() + index);
+  }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionCuda.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionCuda.h
new file mode 100644
index 0000000..68780cd
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionCuda.h
@@ -0,0 +1,6 @@
+
+#if defined(__clang__) || defined(__GNUC__)
+#warning "Deprecated header file, please either include the main Eigen/CXX11/Tensor header or the respective TensorReductionGpu.h file"
+#endif
+
+#include "TensorReductionGpu.h"
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionGpu.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionGpu.h
new file mode 100644
index 0000000..db4e8d8
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionGpu.h
@@ -0,0 +1,966 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_GPU_H
+#define EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_GPU_H
+
+namespace Eigen {
+namespace internal {
+
+
+#if defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
+// Full reducers for GPU, don't vectorize for now
+
+// Reducer function that enables multiple gpu thread to safely accumulate at the same
+// output address. It basically reads the current value of the output variable, and
+// attempts to update it with the new value. If in the meantime another gpu thread
+// updated the content of the output address it will try again.
+template <typename T, typename R>
+__device__ EIGEN_ALWAYS_INLINE void atomicReduce(T* output, T accum, R& reducer) {
+#if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
+  if (sizeof(T) == 4)
+  {
+    unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
+    unsigned int newval = oldval;
+    reducer.reduce(accum, reinterpret_cast<T*>(&newval));
+    if (newval == oldval) {
+      return;
+    }
+    unsigned int readback;
+    while ((readback = atomicCAS((unsigned int*)output, oldval, newval)) != oldval) {
+      oldval = readback;
+      newval = oldval;
+      reducer.reduce(accum, reinterpret_cast<T*>(&newval));
+      if (newval == oldval) {
+        return;
+      }
+    }
+  }
+  else if (sizeof(T) == 8) {
+    unsigned long long oldval = *reinterpret_cast<unsigned long long*>(output);
+    unsigned long long newval = oldval;
+    reducer.reduce(accum, reinterpret_cast<T*>(&newval));
+    if (newval == oldval) {
+      return;
+    }
+    unsigned long long readback;
+    while ((readback = atomicCAS((unsigned long long*)output, oldval, newval)) != oldval) {
+      oldval = readback;
+      newval = oldval;
+      reducer.reduce(accum, reinterpret_cast<T*>(&newval));
+      if (newval == oldval) {
+        return;
+      }
+    }
+  }
+  else {
+    gpu_assert(0 && "Wordsize not supported");
+  }
+#else // EIGEN_CUDA_ARCH >= 300
+  gpu_assert(0 && "Shouldn't be called on unsupported device");
+#endif // EIGEN_CUDA_ARCH >= 300
+}
+
+// We extend atomicExch to support extra data types
+template <typename Type>
+__device__ inline Type atomicExchCustom(Type* address, Type val) {
+  return atomicExch(address, val);
+}
+
+template <>
+__device__ inline double atomicExchCustom(double* address, double val) {
+  unsigned long long int* address_as_ull = reinterpret_cast<unsigned long long int*>(address);
+  return __longlong_as_double(atomicExch(address_as_ull, __double_as_longlong(val)));
+}
+
+#ifdef EIGEN_HAS_GPU_FP16
+template <typename R>
+__device__ inline void atomicReduce(half2* output, half2 accum, R& reducer) {
+  unsigned int oldval = *reinterpret_cast<unsigned int*>(output);
+  unsigned int newval = oldval;
+  reducer.reducePacket(accum, reinterpret_cast<half2*>(&newval));
+  if (newval == oldval) {
+    return;
+  }
+  unsigned int readback;
+  while ((readback = atomicCAS((unsigned int*)output, oldval, newval)) != oldval) {
+    oldval = readback;
+    newval = oldval;
+    reducer.reducePacket(accum, reinterpret_cast<half2*>(&newval));
+    if (newval == oldval) {
+      return;
+    }
+  }
+}
+// reduction should be associative since reduction is not atomic in wide vector but atomic in half2 operations
+template <typename R>
+__device__ inline void atomicReduce(Packet4h2* output, Packet4h2 accum, R& reducer) {
+  half2* houtput=reinterpret_cast<half2*>(output);
+  half2* haccum=reinterpret_cast<half2*>(&accum);
+  for(int i=0;i<4;++i){
+    atomicReduce(houtput+i,*(haccum+i),reducer);
+  }
+}
+#endif  // EIGEN_HAS_GPU_FP16
+
+template <>
+__device__ inline void atomicReduce(float* output, float accum, SumReducer<float>&) {
+#if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
+  atomicAdd(output, accum);
+#else // EIGEN_CUDA_ARCH >= 300
+  gpu_assert(0 && "Shouldn't be called on unsupported device");
+#endif // EIGEN_CUDA_ARCH >= 300
+}
+
+
+template <typename CoeffType, typename Index>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void ReductionInitKernel(const CoeffType val, Index num_preserved_coeffs, CoeffType* output) {
+  const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+  const Index num_threads = blockDim.x * gridDim.x;
+  for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
+    output[i] = val;
+  }
+}
+
+
+template <int BlockSize, int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void FullReductionKernel(Reducer reducer, const Self input, Index num_coeffs,
+                                    typename Self::CoeffReturnType* output, unsigned int* semaphore) {
+#if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
+  // Initialize the output value
+  const Index first_index = blockIdx.x * BlockSize * NumPerThread + threadIdx.x;
+  if (gridDim.x == 1) {
+    if (first_index == 0) {
+      *output = reducer.initialize();
+    }
+  }
+  else {
+    if (threadIdx.x == 0) {
+      unsigned int block = atomicCAS(semaphore, 0u, 1u);
+      if (block == 0) {
+        // We're the first block to run, initialize the output value
+        atomicExchCustom(output, reducer.initialize());
+        __threadfence();
+        atomicExch(semaphore, 2u);
+      }
+      else {
+        // Wait for the first block to initialize the output value.
+        // Use atomicCAS here to ensure that the reads aren't cached
+        unsigned int val;
+        do {
+          val = atomicCAS(semaphore, 2u, 2u);
+        }
+        while (val < 2u);
+      }
+    }
+  }
+
+  __syncthreads();
+
+  eigen_assert(gridDim.x == 1 || *semaphore >= 2u);
+
+  typename Self::CoeffReturnType accum = reducer.initialize();
+  Index max_iter = numext::mini<Index>(num_coeffs - first_index, NumPerThread*BlockSize);
+  for (Index i = 0; i < max_iter; i+=BlockSize) {
+    const Index index = first_index + i;
+    eigen_assert(index < num_coeffs);
+    typename Self::CoeffReturnType val = input.m_impl.coeff(index);
+    reducer.reduce(val, &accum);
+  }
+
+#pragma unroll
+  for (int offset = warpSize/2; offset > 0; offset /= 2) {
+  #if defined(EIGEN_HIPCC)
+    // use std::is_floating_point to determine the type of reduced_val 
+    // This is needed because when Type == double, hipcc will give a "call to __shfl_down is ambguous" error 
+    // and list the float and int versions of __shfl_down as the candidate functions. 
+    if (std::is_floating_point<typename Self::CoeffReturnType>::value) {
+      reducer.reduce(__shfl_down(static_cast<float>(accum), offset, warpSize), &accum);
+    } else {
+      reducer.reduce(__shfl_down(static_cast<int>(accum), offset, warpSize), &accum);
+    }
+  #elif defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
+    reducer.reduce(__shfl_down(accum, offset, warpSize), &accum);
+  #else
+    reducer.reduce(__shfl_down_sync(0xFFFFFFFF, accum, offset, warpSize), &accum);
+  #endif
+  }
+
+  if ((threadIdx.x & (warpSize - 1)) == 0) {
+    atomicReduce(output, accum, reducer);
+  }
+
+  if (gridDim.x > 1 && threadIdx.x == 0) {
+    // Let the last block reset the semaphore
+    atomicInc(semaphore, gridDim.x + 1);
+#if defined(EIGEN_HIPCC)
+    __threadfence_system();
+#endif
+  }
+#else // EIGEN_CUDA_ARCH >= 300
+  gpu_assert(0 && "Shouldn't be called on unsupported device");
+#endif // EIGEN_CUDA_ARCH >= 300
+}
+
+
+#ifdef EIGEN_HAS_GPU_FP16
+template <typename Self,
+          typename Reducer, typename Index>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void ReductionInitFullReduxKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs,
+                                                      packet_traits<Eigen::half>::type* scratch) {
+  eigen_assert(blockDim.x == 1);
+  eigen_assert(gridDim.x == 1);
+  typedef packet_traits<Eigen::half>::type packet_type;
+  Index packet_remainder =
+      num_coeffs % Index(unpacket_traits<packet_type>::size);
+  if (packet_remainder != 0) {
+    half2* h2scratch = reinterpret_cast<half2*>(scratch);
+    for (Index i = num_coeffs - packet_remainder; i + 2 <= num_coeffs; i += 2) {
+      *h2scratch =
+          __halves2half2(input.m_impl.coeff(i), input.m_impl.coeff(i + 1));
+      h2scratch++;
+    }
+    if ((num_coeffs & 1) != 0) {
+      half lastCoeff = input.m_impl.coeff(num_coeffs - 1);
+      *h2scratch = __halves2half2(lastCoeff, reducer.initialize());
+    }
+  } else {
+    *scratch = reducer.template initializePacket<packet_type>();
+  }
+}
+
+template <typename Self,
+          typename Reducer, typename Index>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void ReductionInitKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half* output) {
+  const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+  const Index num_threads = blockDim.x * gridDim.x;
+  typedef typename packet_traits<Eigen::half>::type PacketType;
+
+  const Index num_packets =
+      num_coeffs / Index(unpacket_traits<PacketType>::size);
+  PacketType* p_output = reinterpret_cast<PacketType*>(output);
+  for (Index i = thread_id; i < num_packets; i += num_threads) {
+    p_output[i] = reducer.template initializePacket<PacketType>();
+  }
+  Index packet_remainder =
+      num_coeffs % Index(unpacket_traits<PacketType>::size);
+  if (thread_id < packet_remainder) {
+    output[num_coeffs - packet_remainder + thread_id] = reducer.initialize();
+  }
+}
+
+template <int BlockSize, int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void FullReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs,
+                                    half* output, packet_traits<Eigen::half>::type* scratch) {
+  typedef typename packet_traits<Eigen::half>::type PacketType;
+  const int packet_width = unpacket_traits<PacketType>::size;
+  eigen_assert(NumPerThread % packet_width == 0);
+  const Index first_index =
+      blockIdx.x * BlockSize * NumPerThread + packet_width * threadIdx.x;
+
+  // Initialize the output value if it wasn't initialized by the ReductionInitKernel
+
+  if (gridDim.x == 1) {
+    if (first_index == 0) {
+      int rem = num_coeffs % packet_width;
+      if (rem != 0) {
+        half2* p_scratch = reinterpret_cast<half2*>(scratch);
+        *scratch = reducer.template initializePacket<PacketType>();
+        for (int i = 0; i < rem / 2; i++) {
+          *p_scratch = __halves2half2(
+              input.m_impl.coeff(num_coeffs - packet_width + 2 * i),
+              input.m_impl.coeff(num_coeffs - packet_width + 2 * i + 1));
+          p_scratch++;
+        }
+        if ((num_coeffs & 1) != 0) {
+          half last = input.m_impl.coeff(num_coeffs - 1);
+          *p_scratch = __halves2half2(last, reducer.initialize());
+        }
+      } else {
+        *scratch = reducer.template initializePacket<PacketType>();
+      }
+    }
+    __syncthreads();
+  }
+
+  PacketType accum = reducer.template initializePacket<PacketType>();
+  const Index max_iter =
+      numext::mini<Index>((num_coeffs - first_index) / packet_width,
+                          NumPerThread * BlockSize / packet_width);
+  for (Index i = 0; i < max_iter; i += BlockSize) {
+    const Index index = first_index + packet_width * i;
+    eigen_assert(index + packet_width < num_coeffs);
+    PacketType val = input.m_impl.template packet<Unaligned>(index);
+    reducer.reducePacket(val, &accum);
+  }
+
+#pragma unroll
+  for (int offset = warpSize/2; offset > 0; offset /= 2) {
+  #if defined(EIGEN_HIPCC)
+    PacketType r1;
+    half2* hr = reinterpret_cast<half2*>(&r1);
+    half2* hacc = reinterpret_cast<half2*>(&accum);
+    for (int i = 0; i < packet_width / 2; i++) {
+      // FIXME : remove this workaround once we have native half/half2 support for __shfl_down
+      union { int i; half2 h; } wka_in, wka_out;
+      wka_in.h = hacc[i];
+      wka_out.i = __shfl_down(wka_in.i, offset, warpSize);
+      hr[i] = wka_out.h;
+    }
+    reducer.reducePacket(r1, &accum);
+  #elif defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
+    PacketType r1;
+    half2* hr = reinterpret_cast<half2*>(&r1);
+    half2* hacc = reinterpret_cast<half2*>(&accum);
+    for (int i = 0; i < packet_width / 2; i++) {
+      hr[i] = __shfl_down(hacc[i], offset, warpSize);
+    }
+    reducer.reducePacket(r1, &accum);
+  #else
+    PacketType r1;
+    half2* hr = reinterpret_cast<half2*>(&r1);
+    half2* hacc = reinterpret_cast<half2*>(&accum);
+    for (int i = 0; i < packet_width / 2; i++) {
+      hr[i] = __shfl_down_sync(0xFFFFFFFF, hacc[i], (unsigned)offset, warpSize);
+    }
+    reducer.reducePacket(r1, &accum);
+
+  #endif
+  }
+
+  if ((threadIdx.x & (warpSize - 1)) == 0) {
+    atomicReduce(scratch, accum, reducer);
+  }
+
+  __syncthreads();
+  half2* rv1 = reinterpret_cast<half2*>(scratch);
+  if (packet_width > 2) {
+    reducer.reducePacket(rv1[2], rv1);
+    reducer.reducePacket(rv1[3], rv1 + 1);
+    reducer.reducePacket(rv1[1], rv1);
+  }
+  if (gridDim.x == 1) {
+    if (first_index == 0) {
+      half tmp = __low2half(*rv1);
+      reducer.reduce(__high2half(*rv1), &tmp);
+      *output = tmp;
+    }
+  }
+}
+
+template <typename Op>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void ReductionCleanupKernelHalfFloat(Op reducer, half* output, packet_traits<Eigen::half>::type* scratch) {
+  eigen_assert(threadIdx.x == 1);
+  half2* pscratch = reinterpret_cast<half2*>(scratch);
+  half tmp = __float2half(0.f);
+  typedef packet_traits<Eigen::half>::type packet_type;
+  for (int i = 0; i < unpacket_traits<packet_type>::size; i += 2) {
+    reducer.reduce(__low2half(*pscratch), &tmp);
+    reducer.reduce(__high2half(*pscratch), &tmp);
+    pscratch++;
+  }
+  *output = tmp;
+}
+
+#endif // EIGEN_HAS_GPU_FP16
+
+template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
+struct FullReductionLauncher {
+  static void run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index) {
+    gpu_assert(false && "Should only be called on doubles, floats and half floats");
+  }
+};
+
+// Specialization for float and double
+template <typename Self, typename Op, typename OutputType, bool PacketAccess>
+struct FullReductionLauncher<
+    Self, Op, OutputType, PacketAccess,
+    typename internal::enable_if<
+      internal::is_same<float, OutputType>::value ||
+      internal::is_same<double, OutputType>::value,
+    void>::type> {
+  static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs) {
+
+    typedef typename Self::Index Index;
+    const int block_size = 256;
+    const int num_per_thread = 128;
+    const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+
+    unsigned int* semaphore = NULL;
+    if (num_blocks > 1) {
+      semaphore = device.semaphore();
+    }
+
+    LAUNCH_GPU_KERNEL((FullReductionKernel<block_size, num_per_thread, Self, Op, Index>),
+                       num_blocks, block_size, 0, device, reducer, self, num_coeffs, output, semaphore);
+  }
+};
+
+#ifdef EIGEN_HAS_GPU_FP16
+template <typename Self, typename Op>
+struct FullReductionLauncher<Self, Op, Eigen::half, false> {
+  static void run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index) {
+    gpu_assert(false && "Should not be called since there is no packet accessor");
+  }
+};
+
+template <typename Self, typename Op>
+struct FullReductionLauncher<Self, Op, Eigen::half, true> {
+  static void run(const Self& self, Op& reducer, const GpuDevice& device, half* output, typename Self::Index num_coeffs) {
+    typedef typename Self::Index Index;
+    typedef typename packet_traits<Eigen::half>::type PacketType;
+
+    const int block_size = 256;
+    const int num_per_thread = 128;
+    const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+    PacketType* scratch = static_cast<PacketType*>(device.scratchpad());
+    // half2* scratch = static_cast<half2*>(device.scratchpad());
+
+    if (num_blocks > 1) {
+      // We initialize the output and the scrathpad outside the reduction kernel when we can't be sure that there
+      // won't be a race conditions between multiple thread blocks.
+      LAUNCH_GPU_KERNEL((ReductionInitFullReduxKernelHalfFloat<Self, Op, Index>),
+                         1, 1, 0, device, reducer, self, num_coeffs, scratch);
+    }
+
+    LAUNCH_GPU_KERNEL((FullReductionKernelHalfFloat<block_size, num_per_thread, Self, Op, Index>),
+                       num_blocks, block_size, 0, device, reducer, self, num_coeffs, output, scratch);
+
+    if (num_blocks > 1) {
+      LAUNCH_GPU_KERNEL((ReductionCleanupKernelHalfFloat<Op>),
+                         1, 1, 0, device, reducer, output, scratch);
+    }
+  }
+};
+#endif // EIGEN_HAS_GPU_FP16
+
+
+template <typename Self, typename Op, bool Vectorizable>
+struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
+  // Unfortunately nvidia doesn't support well exotic types such as complex,
+  // so reduce the scope of the optimized version of the code to the simple cases
+  // of doubles, floats and half floats
+#ifdef EIGEN_HAS_GPU_FP16
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
+      (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+       internal::is_same<typename Self::CoeffReturnType, double>::value ||
+       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
+#else // EIGEN_HAS_GPU_FP16
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
+                                                (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+                                                 internal::is_same<typename Self::CoeffReturnType, double>::value);
+#endif // EIGEN_HAS_GPU_FP16
+
+  template <typename OutputType>
+  static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) {
+    gpu_assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
+    const Index num_coeffs = array_prod(self.m_impl.dimensions());
+    // Don't crash when we're called with an input tensor of size 0.
+    if (num_coeffs == 0) {
+      return;
+    }
+
+    FullReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs);
+  }
+};
+
+
+template <int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void InnerReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
+                                         typename Self::CoeffReturnType* output) {
+#if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300)
+  typedef typename Self::CoeffReturnType Type;
+  eigen_assert(blockDim.y == 1);
+  eigen_assert(blockDim.z == 1);
+  eigen_assert(gridDim.y == 1);
+  eigen_assert(gridDim.z == 1);
+
+  const int unroll_times = 16;
+  eigen_assert(NumPerThread % unroll_times == 0);
+
+  const Index input_col_blocks = divup<Index>(num_coeffs_to_reduce, blockDim.x * NumPerThread);
+  const Index num_input_blocks = input_col_blocks * num_preserved_coeffs;
+
+  const Index num_threads = blockDim.x * gridDim.x;
+  const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+
+  // Initialize the output values if they weren't initialized by the ReductionInitKernel
+  if (gridDim.x == 1) {
+    for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
+      output[i] = reducer.initialize();
+    }
+    __syncthreads();
+  }
+
+  for (Index i = blockIdx.x; i < num_input_blocks; i += gridDim.x) {
+    const Index row = i / input_col_blocks;
+
+    if (row < num_preserved_coeffs) {
+      const Index col_block = i % input_col_blocks;
+      const Index col_begin = col_block * blockDim.x * NumPerThread + threadIdx.x;
+
+      Type reduced_val = reducer.initialize();
+
+      for (Index j = 0; j < NumPerThread; j += unroll_times) {
+        const Index last_col = col_begin + blockDim.x * (j + unroll_times - 1);
+        if (last_col >= num_coeffs_to_reduce) {
+          for (Index col = col_begin + blockDim.x * j; col < num_coeffs_to_reduce; col += blockDim.x) {
+            const Type val = input.m_impl.coeff(row * num_coeffs_to_reduce + col);
+            reducer.reduce(val, &reduced_val);
+          }
+          break;
+        } else {
+          // Faster version of the loop with no branches after unrolling.
+#pragma unroll
+          for (int k = 0; k < unroll_times; ++k) {
+            const Index col = col_begin + blockDim.x * (j + k);
+            reducer.reduce(input.m_impl.coeff(row * num_coeffs_to_reduce + col), &reduced_val);
+          }
+        }
+      }
+
+#pragma unroll
+      for (int offset = warpSize/2; offset > 0; offset /= 2) {
+      #if defined(EIGEN_HIPCC)
+        // use std::is_floating_point to determine the type of reduced_val 
+       // This is needed because when Type == double, hipcc will give a "call to __shfl_down is ambguous" error 
+       // and list the float and int versions of __shfl_down as the candidate functions. 
+        if (std::is_floating_point<Type>::value) {
+          reducer.reduce(__shfl_down(static_cast<float>(reduced_val), offset), &reduced_val);
+        } else {
+          reducer.reduce(__shfl_down(static_cast<int>(reduced_val), offset), &reduced_val);
+        }
+      #elif defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
+        reducer.reduce(__shfl_down(reduced_val, offset), &reduced_val);
+      #else
+        reducer.reduce(__shfl_down_sync(0xFFFFFFFF, reduced_val, offset), &reduced_val);
+      #endif
+      }
+
+      if ((threadIdx.x & (warpSize - 1)) == 0) {
+        atomicReduce(&(output[row]), reduced_val, reducer);
+      }
+    }
+  }
+#else // EIGEN_CUDA_ARCH >= 300
+  gpu_assert(0 && "Shouldn't be called on unsupported device");
+#endif // EIGEN_CUDA_ARCH >= 300
+}
+
+#ifdef EIGEN_HAS_GPU_FP16
+
+template <int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void InnerReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
+                                              half* output) {
+  eigen_assert(blockDim.y == 1);
+  eigen_assert(blockDim.z == 1);
+  eigen_assert(gridDim.y == 1);
+  eigen_assert(gridDim.z == 1);
+
+  typedef typename packet_traits<Eigen::half>::type PacketType;
+  const int packet_width = unpacket_traits<PacketType>::size;
+  const int unroll_times = 16 / packet_width;
+  eigen_assert(NumPerThread % unroll_times == 0);
+  eigen_assert(unroll_times % 2 == 0);
+
+  const Index input_col_blocks = divup<Index>(num_coeffs_to_reduce, blockDim.x * NumPerThread * 2);
+  const Index num_input_blocks = divup<Index>(input_col_blocks * num_preserved_coeffs, 2);
+
+  const Index num_threads = blockDim.x * gridDim.x;
+  const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+
+  // Initialize the output values if they weren't initialized by the ReductionInitKernel
+  if (gridDim.x == 1) {
+    Index i = packet_width * thread_id;
+    for (; i + packet_width <= num_preserved_coeffs;
+         i += packet_width * num_threads) {
+      PacketType* poutput = reinterpret_cast<PacketType*>(output + i);
+      *poutput = reducer.template initializePacket<PacketType>();
+    }
+    if (i < num_preserved_coeffs) {
+      output[i] = reducer.initialize();
+    }
+    __syncthreads();
+  }
+
+  for (Index i = blockIdx.x; i < num_input_blocks; i += gridDim.x) {
+    const Index row = 2 * (i / input_col_blocks);  // everybody takes 2 rows
+
+    if (row + 1 < num_preserved_coeffs) {
+      const Index col_block = i % input_col_blocks;
+      const Index col_begin =
+          packet_width * (col_block * blockDim.x * NumPerThread + threadIdx.x);
+
+      PacketType reduced_val1 = reducer.template initializePacket<PacketType>();
+      PacketType reduced_val2 = reducer.template initializePacket<PacketType>();
+
+      for (Index j = 0; j < NumPerThread; j += unroll_times) {
+        const Index last_col =
+            col_begin + blockDim.x * (j + unroll_times - 1) * packet_width;
+        if (last_col >= num_coeffs_to_reduce) {
+          Index col = col_begin + blockDim.x * j;
+          for (; col + packet_width <= num_coeffs_to_reduce;
+               col += blockDim.x) {
+            const PacketType val1 = input.m_impl.template packet<Unaligned>(
+                row * num_coeffs_to_reduce + col);
+            reducer.reducePacket(val1, &reduced_val1);
+            const PacketType val2 = input.m_impl.template packet<Unaligned>(
+                (row + 1) * num_coeffs_to_reduce + col);
+            reducer.reducePacket(val2, &reduced_val2);
+          }
+          if (col < num_coeffs_to_reduce) {
+            PacketType r1 = reducer.template initializePacket<PacketType>();
+            PacketType r2 = reducer.template initializePacket<PacketType>();
+            half2* hr1 = reinterpret_cast<half2*>(&r1);
+            half2* hr2 = reinterpret_cast<half2*>(&r2);
+            while (col + 1 < num_coeffs_to_reduce) {
+              *hr1 = __halves2half2(
+                  input.m_impl.coeff(row * num_coeffs_to_reduce + col),
+                  input.m_impl.coeff(row * num_coeffs_to_reduce + col + 1));
+              *hr2 = __halves2half2(
+                  input.m_impl.coeff((row + 1) * num_coeffs_to_reduce + col),
+                  input.m_impl.coeff((row + 1) * num_coeffs_to_reduce + col +
+                                     1));
+              hr1++;
+              hr2++;
+              col += 2;
+            }
+            if (col < num_coeffs_to_reduce) {
+              // Peel;
+              const half last1 =
+                  input.m_impl.coeff(row * num_coeffs_to_reduce + col);
+              *hr1 = __halves2half2(last1, reducer.initialize());
+              const half last2 =
+                  input.m_impl.coeff((row + 1) * num_coeffs_to_reduce + col);
+              *hr2 = __halves2half2(last2, reducer.initialize());
+            }
+            reducer.reducePacket(r1, &reduced_val1);
+            reducer.reducePacket(r2, &reduced_val2);
+          }
+          break;
+        } else {
+          // Faster version of the loop with no branches after unrolling.
+#pragma unroll
+          for (int k = 0; k < unroll_times; ++k) {
+            const Index col = col_begin + blockDim.x * (j + k) * packet_width;
+            reducer.reducePacket(input.m_impl.template packet<Unaligned>(
+                                     row * num_coeffs_to_reduce + col),
+                                 &reduced_val1);
+            reducer.reducePacket(input.m_impl.template packet<Unaligned>(
+                                     (row + 1) * num_coeffs_to_reduce + col),
+                                 &reduced_val2);
+          }
+        }
+      }
+
+#pragma unroll
+      for (int offset = warpSize/2; offset > 0; offset /= 2) {
+      #if defined(EIGEN_HIPCC)
+        PacketType r1;
+        PacketType r2;
+        half2* hr1 = reinterpret_cast<half2*>(&r1);
+        half2* hr2 = reinterpret_cast<half2*>(&r2);
+        half2* rv1 = reinterpret_cast<half2*>(&reduced_val1);
+        half2* rv2 = reinterpret_cast<half2*>(&reduced_val2);
+        for (int i = 0; i < packet_width / 2; i++) {
+	  // FIXME : remove this workaround once we have native half/half2 support for __shfl_down
+	  union { int i; half2 h; } wka_in1, wka_out1;
+	  wka_in1.h = rv1[i];
+	  wka_out1.i = __shfl_down(wka_in1.i, offset, warpSize);
+	  hr1[i] = wka_out1.h;
+
+	  union { int i; half2 h; } wka_in2, wka_out2;
+	  wka_in2.h = rv2[i];
+	  wka_out2.i = __shfl_down(wka_in2.i, offset, warpSize);
+	  hr2[i] = wka_out2.h;
+        }
+        reducer.reducePacket(r1, &reduced_val1);
+        reducer.reducePacket(r2, &reduced_val2);
+      #elif defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
+        PacketType r1;
+        PacketType r2;
+        half2* hr1 = reinterpret_cast<half2*>(&r1);
+        half2* hr2 = reinterpret_cast<half2*>(&r2);
+        half2* rv1 = reinterpret_cast<half2*>(&reduced_val1);
+        half2* rv2 = reinterpret_cast<half2*>(&reduced_val2);
+        for (int i = 0; i < packet_width / 2; i++) {
+          hr1[i] = __shfl_down(rv1[i], offset, warpSize);
+          hr2[i] = __shfl_down(rv2[i], offset, warpSize);
+        }
+        reducer.reducePacket(r1, &reduced_val1);
+        reducer.reducePacket(r2, &reduced_val2);
+      #else
+        PacketType r1;
+        PacketType r2;
+        half2* hr1 = reinterpret_cast<half2*>(&r1);
+        half2* hr2 = reinterpret_cast<half2*>(&r2);
+        half2* rr1 = reinterpret_cast<half2*>(&reduced_val1);
+        half2* rr2 = reinterpret_cast<half2*>(&reduced_val2);
+        for (int i = 0; i < packet_width / 2; i++) {
+          hr1[i] =
+              __shfl_down_sync(0xFFFFFFFF, rr1[i], (unsigned)offset, warpSize);
+          hr2[i] =
+              __shfl_down_sync(0xFFFFFFFF, rr2[i], (unsigned)offset, warpSize);
+        }
+        reducer.reducePacket(r1, &reduced_val1);
+        reducer.reducePacket(r2, &reduced_val2);
+
+      #endif
+      }
+      half2* rv1 = reinterpret_cast<half2*>(&reduced_val1);
+      half2* rv2 = reinterpret_cast<half2*>(&reduced_val2);
+      half2 val;
+      if (packet_width > 2) {
+        reducer.reducePacket(rv1[2], rv1);
+        reducer.reducePacket(rv1[3], rv1 + 1);
+        reducer.reducePacket(rv1[1], rv1);
+        reducer.reducePacket(rv2[2], rv2);
+        reducer.reducePacket(rv2[3], rv2 + 1);
+        reducer.reducePacket(rv2[1], rv2);
+      }
+      half val1 = __low2half(*rv1);
+      reducer.reduce(__high2half(*rv1), &val1);
+      half val2 = __low2half(*rv2);
+      reducer.reduce(__high2half(*rv2), &val2);
+      val = __halves2half2(val1, val2);
+      if ((threadIdx.x & (warpSize - 1)) == 0) {
+        half* loc = output + row;
+        atomicReduce((half2*)loc, val, reducer);
+      }
+    }
+  }
+}
+
+#endif // EIGEN_HAS_GPU_FP16
+
+template <typename Self, typename Op, typename OutputType, bool PacketAccess, typename Enabled = void>
+struct InnerReductionLauncher {
+  static EIGEN_DEVICE_FUNC bool run(const Self&, Op&, const GpuDevice&, OutputType*, typename Self::Index, typename Self::Index) {
+    gpu_assert(false && "Should only be called to reduce doubles, floats and half floats on a gpu device");
+    return true;
+  }
+};
+
+// Specialization for float and double
+template <typename Self, typename Op, typename OutputType, bool PacketAccess>
+struct InnerReductionLauncher<
+  Self, Op, OutputType, PacketAccess,
+  typename internal::enable_if<
+    internal::is_same<float, OutputType>::value ||
+    internal::is_same<double, OutputType>::value,
+  void>::type> {
+  static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    typedef typename Self::Index Index;
+
+    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
+    const int block_size = 256;
+    const int num_per_thread = 128;
+    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+    const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / block_size;
+    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+
+    if (num_blocks > 1) {
+      // We initialize the outputs outside the reduction kernel when we can't be sure that there
+      // won't be a race conditions between multiple thread blocks.
+      const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
+      const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / 1024;
+      const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+      LAUNCH_GPU_KERNEL((ReductionInitKernel<OutputType, Index>),
+                         num_blocks, 1024, 0, device, reducer.initialize(),
+                         num_preserved_vals, output);
+    }
+
+    LAUNCH_GPU_KERNEL((InnerReductionKernel<num_per_thread, Self, Op, Index>),
+                       num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
+
+    return false;
+  }
+};
+
+#ifdef EIGEN_HAS_GPU_FP16
+template <typename Self, typename Op>
+struct InnerReductionLauncher<Self, Op, Eigen::half, false> {
+  static bool run(const Self&, Op&, const GpuDevice&, half*, typename Self::Index, typename Self::Index) {
+    gpu_assert(false && "Should not be called since there is no packet accessor");
+    return true;
+  }
+};
+
+template <typename Self, typename Op>
+struct InnerReductionLauncher<Self, Op, Eigen::half, true> {
+  static bool run(const Self& self, Op& reducer, const GpuDevice& device, half* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    typedef typename Self::Index Index;
+
+    if (num_preserved_vals % 2 != 0) {
+      // Not supported yet, revert to the slower code path
+      return true;
+    }
+
+    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
+    const int block_size = /*256*/128;
+    const int num_per_thread = /*128*/64;
+    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+    const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / block_size;
+    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+
+    if (num_blocks > 1) {
+      // We initialize the outputs outside the reduction kernel when we can't be sure that there
+      // won't be a race conditions between multiple thread blocks.
+      LAUNCH_GPU_KERNEL((ReductionInitKernelHalfFloat<Self, Op, Index>),
+                         1, 1, 0, device, reducer, self, num_preserved_vals, output);
+    }
+
+    LAUNCH_GPU_KERNEL((InnerReductionKernelHalfFloat<num_per_thread, Self, Op, Index>),
+                       num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
+
+    return false;
+  }
+};
+#endif // EIGEN_HAS_GPU_FP16
+
+
+template <typename Self, typename Op>
+struct InnerReducer<Self, Op, GpuDevice> {
+  // Unfortunately nvidia doesn't support well exotic types such as complex,
+  // so reduce the scope of the optimized version of the code to the simple case
+  // of floats and half floats.
+#ifdef EIGEN_HAS_GPU_FP16
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
+      (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+       internal::is_same<typename Self::CoeffReturnType, double>::value ||
+       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
+#else // EIGEN_HAS_GPU_FP16
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
+                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
+#endif // EIGEN_HAS_GPU_FP16
+
+  template <typename OutputType>
+  static bool run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    gpu_assert(HasOptimizedImplementation && "Should only be called on doubles, floats or half floats");
+    const Index num_coeffs = array_prod(self.m_impl.dimensions());
+    // Don't crash when we're called with an input tensor of size 0.
+    if (num_coeffs == 0) {
+      return true;
+    }
+    // It's faster to use the usual code.
+    if (num_coeffs_to_reduce <= 128) {
+      return true;
+    }
+
+    return InnerReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs_to_reduce, num_preserved_vals);
+  }
+};
+
+template <int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void OuterReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
+                                     typename Self::CoeffReturnType* output) {
+  const Index num_threads = blockDim.x * gridDim.x;
+  const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+  // Initialize the output values if they weren't initialized by the ReductionInitKernel
+  if (gridDim.x == 1) {
+    for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
+      output[i] = reducer.initialize();
+    }
+    __syncthreads();
+  }
+
+  // Do the reduction.
+  const Index max_iter = num_preserved_coeffs * divup<Index>(num_coeffs_to_reduce, NumPerThread);
+  for (Index i = thread_id; i < max_iter; i += num_threads) {
+    const Index input_col = i % num_preserved_coeffs;
+    const Index input_row = (i / num_preserved_coeffs) * NumPerThread;
+    typename Self::CoeffReturnType reduced_val = reducer.initialize();
+    const Index max_row = numext::mini(input_row + NumPerThread, num_coeffs_to_reduce);
+    for (Index j = input_row; j < max_row; j++) {
+      typename Self::CoeffReturnType val = input.m_impl.coeff(j * num_preserved_coeffs + input_col);
+      reducer.reduce(val, &reduced_val);
+    }
+    atomicReduce(&(output[input_col]), reduced_val, reducer);
+  }
+}
+
+
+template <typename Self, typename Op>
+struct OuterReducer<Self, Op, GpuDevice> {
+  // Unfortunately nvidia doesn't support well exotic types such as complex,
+  // so reduce the scope of the optimized version of the code to the simple case
+  // of floats.
+  static const bool HasOptimizedImplementation = !Self::ReducerTraits::IsStateful &&
+                                                 (internal::is_same<typename Self::CoeffReturnType, float>::value ||
+                                                  internal::is_same<typename Self::CoeffReturnType, double>::value);
+  template <typename Device, typename OutputType>
+  static
+    #if !defined(EIGEN_HIPCC)
+    // FIXME :  leaving this EIGEN_DEVICE_FUNC in, results in the following runtime error
+    //          (in the cxx11_tensor_reduction_gpu test)
+    //
+    // terminate called after throwing an instance of 'std::runtime_error'
+    //   what():  No device code available for function: _ZN5Eigen8internal20OuterReductionKernelIL...
+    //
+    // don't know why this happens (and why is it a runtime error instead of a compile time error)
+    //
+    // this will be fixed by HIP PR#457
+    EIGEN_DEVICE_FUNC
+    #endif
+    bool run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) {
+    gpu_assert(false && "Should only be called to reduce doubles or floats on a gpu device");
+    return true;
+  }
+
+  static bool run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    typedef typename Self::Index Index;
+
+    // It's faster to use the usual code.
+    if (num_coeffs_to_reduce <= 32) {
+      return true;
+    }
+
+    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
+    const int block_size = 256;
+    const int num_per_thread = 16;
+    const int dyn_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
+    const int max_blocks = device.getNumGpuMultiProcessors() *
+                           device.maxGpuThreadsPerMultiProcessor() / block_size;
+    const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+
+    if (num_blocks > 1) {
+      // We initialize the outputs in the reduction kernel itself when we don't have to worry
+      // about race conditions between multiple thread blocks.
+      const int dyn_blocks = divup<int>(num_preserved_vals, 1024);
+      const int max_blocks = device.getNumGpuMultiProcessors() *
+                             device.maxGpuThreadsPerMultiProcessor() / 1024;
+      const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
+      LAUNCH_GPU_KERNEL((ReductionInitKernel<float, Index>),
+                         num_blocks, 1024, 0, device, reducer.initialize(),
+                         num_preserved_vals, output);
+    }
+
+    LAUNCH_GPU_KERNEL((OuterReductionKernel<num_per_thread, Self, Op, Index>),
+                       num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
+
+    return false;
+  }
+};
+
+#endif // defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
+
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_GPU_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionSycl.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionSycl.h
new file mode 100644
index 0000000..474eba0
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorReductionSycl.h
@@ -0,0 +1,582 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/*****************************************************************
+ * TensorReductionSycl.h
+ *
+ * \brief:
+ *  This is the specialization of the reduction operation. Two phase reduction approach 
+ * is used since the GPU does not have Global Synchronization for global memory among 
+ * different work-group/thread block. To solve the problem, we need to create two kernels 
+ * to reduce the data, where the first kernel reduce the data locally and each local 
+ * workgroup/thread-block save the input data into global memory. In the second phase (global reduction)
+ * one work-group uses one work-group/thread-block to reduces the intermediate data into one single element. 
+ * Here is an NVIDIA presentation explaining the optimized two phase reduction algorithm on GPU:
+ * https://developer.download.nvidia.com/assets/cuda/files/reduction.pdf
+ *
+ *****************************************************************/
+
+#ifndef UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP
+#define UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP
+namespace Eigen {
+namespace TensorSycl {
+namespace internal {
+
+template <typename Op, typename CoeffReturnType, typename Index, bool Vectorizable>
+struct OpDefiner {
+  typedef typename Vectorise<CoeffReturnType, Eigen::SyclDevice, Vectorizable>::PacketReturnType PacketReturnType;
+  typedef Op type;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE type get_op(Op &op) { return op; }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType finalise_op(const PacketReturnType &accumulator,
+                                                                            const Index &) {
+    return accumulator;
+  }
+};
+
+template <typename CoeffReturnType, typename Index>
+struct OpDefiner<Eigen::internal::MeanReducer<CoeffReturnType>, CoeffReturnType, Index, false> {
+  typedef Eigen::internal::SumReducer<CoeffReturnType> type;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE type get_op(Eigen::internal::MeanReducer<CoeffReturnType> &) {
+    return type();
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType finalise_op(const CoeffReturnType &accumulator,
+                                                                           const Index &scale) {
+    ::Eigen::internal::scalar_quotient_op<CoeffReturnType> quotient_op;
+    return quotient_op(accumulator, CoeffReturnType(scale));
+  }
+};
+
+template <typename CoeffReturnType, typename Index>
+struct OpDefiner<Eigen::internal::MeanReducer<CoeffReturnType>, CoeffReturnType, Index, true> {
+  typedef typename Vectorise<CoeffReturnType, Eigen::SyclDevice, true>::PacketReturnType PacketReturnType;
+  typedef Eigen::internal::SumReducer<CoeffReturnType> type;
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE type get_op(Eigen::internal::MeanReducer<CoeffReturnType> &) {
+    return type();
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType finalise_op(const PacketReturnType &accumulator,
+                                                                            const Index &scale) {
+    return ::Eigen::internal::pdiv(accumulator, ::Eigen::internal::pset1<PacketReturnType>(CoeffReturnType(scale)));
+  }
+};
+
+template <typename CoeffReturnType, typename OpType, typename InputAccessor, typename OutputAccessor, typename Index,
+          Index local_range>
+struct SecondStepFullReducer {
+  typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      LocalAccessor;
+  typedef OpDefiner<OpType, CoeffReturnType, Index, true> OpDef;
+  typedef typename OpDef::type Op;
+  LocalAccessor scratch;
+  InputAccessor aI;
+  OutputAccessor outAcc;
+  Op op;
+  SecondStepFullReducer(LocalAccessor scratch_, InputAccessor aI_, OutputAccessor outAcc_, OpType op_)
+      : scratch(scratch_), aI(aI_), outAcc(outAcc_), op(OpDef::get_op(op_)) {}
+
+  void operator()(cl::sycl::nd_item<1> itemID) {
+    // Our empirical research shows that the best performance will be achieved
+    // when there is only one element per thread to reduce in the second step.
+    // in this step the second step reduction time is almost negligible.
+    // Hence, in the second step of reduction the input size is fixed to the
+    // local size, thus, there is only one element read per thread. The
+    // algorithm must be changed if the number of reduce per thread in the
+    // second step is greater than 1. Otherwise, the result will be wrong.
+    const Index localid = itemID.get_local_id(0);
+    auto aInPtr = aI.get_pointer() + localid;
+    auto aOutPtr = outAcc.get_pointer();
+    CoeffReturnType *scratchptr = scratch.get_pointer();
+    CoeffReturnType accumulator = *aInPtr;
+
+    scratchptr[localid] = op.finalize(accumulator);
+    for (Index offset = itemID.get_local_range(0) / 2; offset > 0; offset /= 2) {
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+      if (localid < offset) {
+        op.reduce(scratchptr[localid + offset], &accumulator);
+        scratchptr[localid] = op.finalize(accumulator);
+      }
+    }
+    if (localid == 0) *aOutPtr = op.finalize(accumulator);
+  }
+};
+
+// Full reduction first phase. In this version the vectorization is true and the reduction accept 
+// any generic reducerOp  e.g( max, min, sum, mean, iamax, iamin, etc ). 
+template <typename Evaluator, typename OpType, typename Evaluator::Index local_range>
+class FullReductionKernelFunctor {
+ public:
+  typedef typename Evaluator::CoeffReturnType CoeffReturnType;
+  typedef typename Evaluator::Index Index;
+  typedef OpDefiner<OpType, typename Evaluator::CoeffReturnType, Index,
+                    (Evaluator::ReducerTraits::PacketAccess & Evaluator::InputPacketAccess)>
+      OpDef;
+
+  typedef typename OpDef::type Op;
+  typedef typename Evaluator::EvaluatorPointerType EvaluatorPointerType;
+  typedef typename Evaluator::PacketReturnType PacketReturnType;
+  typedef
+      typename ::Eigen::internal::conditional<(Evaluator::ReducerTraits::PacketAccess & Evaluator::InputPacketAccess),
+                                              PacketReturnType, CoeffReturnType>::type OutType;
+  typedef cl::sycl::accessor<OutType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      LocalAccessor;
+  LocalAccessor scratch;
+  Evaluator evaluator;
+  EvaluatorPointerType final_output;
+  Index rng;
+  Op op;
+
+  FullReductionKernelFunctor(LocalAccessor scratch_, Evaluator evaluator_, EvaluatorPointerType final_output_,
+                             Index rng_, OpType op_)
+      : scratch(scratch_), evaluator(evaluator_), final_output(final_output_), rng(rng_), op(OpDef::get_op(op_)) {}
+
+  void operator()(cl::sycl::nd_item<1> itemID) { compute_reduction(itemID); }
+
+  template <bool Vect = (Evaluator::ReducerTraits::PacketAccess & Evaluator::InputPacketAccess)>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<Vect>::type compute_reduction(
+      const cl::sycl::nd_item<1> &itemID) {
+    auto output_ptr = final_output.get_pointer();
+    Index VectorizedRange = (rng / Evaluator::PacketSize) * Evaluator::PacketSize;
+    Index globalid = itemID.get_global_id(0);
+    Index localid = itemID.get_local_id(0);
+    Index step = Evaluator::PacketSize * itemID.get_global_range(0);
+    Index start = Evaluator::PacketSize * globalid;
+    // vectorizable parts
+    PacketReturnType packetAccumulator = op.template initializePacket<PacketReturnType>();
+    for (Index i = start; i < VectorizedRange; i += step) {
+      op.template reducePacket<PacketReturnType>(evaluator.impl().template packet<Unaligned>(i), &packetAccumulator);
+    }
+    globalid += VectorizedRange;
+    // non vectorizable parts
+    for (Index i = globalid; i < rng; i += itemID.get_global_range(0)) {
+      op.template reducePacket<PacketReturnType>(
+          ::Eigen::TensorSycl::internal::PacketWrapper<PacketReturnType, Evaluator::PacketSize>::convert_to_packet_type(
+              evaluator.impl().coeff(i), op.initialize()),
+          &packetAccumulator);
+    }
+    scratch[localid] = packetAccumulator =
+        OpDef::finalise_op(op.template finalizePacket<PacketReturnType>(packetAccumulator), rng);
+    // reduction parts // Local size is always power of 2
+    EIGEN_UNROLL_LOOP
+    for (Index offset = local_range / 2; offset > 0; offset /= 2) {
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+      if (localid < offset) {
+        op.template reducePacket<PacketReturnType>(scratch[localid + offset], &packetAccumulator);
+        scratch[localid] = op.template finalizePacket<PacketReturnType>(packetAccumulator);
+      }
+    }
+    if (localid == 0) {
+      output_ptr[itemID.get_group(0)] =
+          op.finalizeBoth(op.initialize(), op.template finalizePacket<PacketReturnType>(packetAccumulator));
+    }
+  }
+
+  template <bool Vect = (Evaluator::ReducerTraits::PacketAccess & Evaluator::InputPacketAccess)>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<!Vect>::type compute_reduction(
+      const cl::sycl::nd_item<1> &itemID) {
+    auto output_ptr = final_output.get_pointer();
+    Index globalid = itemID.get_global_id(0);
+    Index localid = itemID.get_local_id(0);
+    // vectorizable parts
+    CoeffReturnType accumulator = op.initialize();
+    // non vectorizable parts
+    for (Index i = globalid; i < rng; i += itemID.get_global_range(0)) {
+      op.reduce(evaluator.impl().coeff(i), &accumulator);
+    }
+    scratch[localid] = accumulator = OpDef::finalise_op(op.finalize(accumulator), rng);
+
+    // reduction parts. the local size is always power of 2
+    EIGEN_UNROLL_LOOP
+    for (Index offset = local_range / 2; offset > 0; offset /= 2) {
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+      if (localid < offset) {
+        op.reduce(scratch[localid + offset], &accumulator);
+        scratch[localid] = op.finalize(accumulator);
+      }
+    }
+    if (localid == 0) {
+      output_ptr[itemID.get_group(0)] = op.finalize(accumulator);
+    }
+  }
+};
+
+template <typename Evaluator, typename OpType>
+class GenericNondeterministicReducer {
+ public:
+  typedef typename Evaluator::CoeffReturnType CoeffReturnType;
+  typedef typename Evaluator::EvaluatorPointerType EvaluatorPointerType;
+  typedef typename Evaluator::Index Index;
+  typedef OpDefiner<OpType, CoeffReturnType, Index, false> OpDef;
+  typedef typename OpDef::type Op;
+  template <typename Scratch>
+  GenericNondeterministicReducer(Scratch, Evaluator evaluator_, EvaluatorPointerType output_accessor_, OpType functor_,
+                       Index range_, Index num_values_to_reduce_)
+      : evaluator(evaluator_),
+        output_accessor(output_accessor_),
+        functor(OpDef::get_op(functor_)),
+        range(range_),
+        num_values_to_reduce(num_values_to_reduce_) {}
+
+  void operator()(cl::sycl::nd_item<1> itemID) {
+    auto output_accessor_ptr = output_accessor.get_pointer();
+    /// const cast added as a naive solution to solve the qualifier drop error
+    Index globalid = static_cast<Index>(itemID.get_global_linear_id());
+    if (globalid < range) {
+      CoeffReturnType accum = functor.initialize();
+      Eigen::internal::GenericDimReducer<Evaluator::NumReducedDims - 1, Evaluator, Op>::reduce(
+          evaluator, evaluator.firstInput(globalid), functor, &accum);
+      output_accessor_ptr[globalid] = OpDef::finalise_op(functor.finalize(accum), num_values_to_reduce);
+    }
+  }
+
+ private:
+  Evaluator evaluator;
+  EvaluatorPointerType output_accessor;
+  Op functor;
+  Index range;
+  Index num_values_to_reduce;
+};
+
+enum class reduction_dim { inner_most, outer_most };
+// default is preserver
+template <typename Evaluator, typename OpType, typename PannelParameters, reduction_dim rt>
+struct PartialReductionKernel {
+  typedef typename Evaluator::CoeffReturnType CoeffReturnType;
+  typedef typename Evaluator::EvaluatorPointerType EvaluatorPointerType;
+  typedef typename Evaluator::Index Index;
+  typedef OpDefiner<OpType, CoeffReturnType, Index, false> OpDef;
+  typedef typename OpDef::type Op;
+  typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      ScratchAcc;
+  ScratchAcc scratch;
+  Evaluator evaluator;
+  EvaluatorPointerType output_accessor;
+  Op op;
+  const Index preserve_elements_num_groups;
+  const Index reduce_elements_num_groups;
+  const Index num_coeffs_to_preserve;
+  const Index num_coeffs_to_reduce;
+
+  PartialReductionKernel(ScratchAcc scratch_, Evaluator evaluator_, EvaluatorPointerType output_accessor_, OpType op_,
+                         const Index preserve_elements_num_groups_, const Index reduce_elements_num_groups_,
+                         const Index num_coeffs_to_preserve_, const Index num_coeffs_to_reduce_)
+      : scratch(scratch_),
+        evaluator(evaluator_),
+        output_accessor(output_accessor_),
+        op(OpDef::get_op(op_)),
+        preserve_elements_num_groups(preserve_elements_num_groups_),
+        reduce_elements_num_groups(reduce_elements_num_groups_),
+        num_coeffs_to_preserve(num_coeffs_to_preserve_),
+        num_coeffs_to_reduce(num_coeffs_to_reduce_) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void element_wise_reduce(Index globalRId, Index globalPId,
+                                                                 CoeffReturnType &accumulator) {
+    if (globalPId >= num_coeffs_to_preserve) {
+      return;
+    }
+    Index global_offset = rt == reduction_dim::outer_most ? globalPId + (globalRId * num_coeffs_to_preserve)
+                                                          : globalRId + (globalPId * num_coeffs_to_reduce);
+    Index localOffset = globalRId;
+
+    const Index per_thread_local_stride = PannelParameters::LocalThreadSizeR * reduce_elements_num_groups;
+    const Index per_thread_global_stride =
+        rt == reduction_dim::outer_most ? num_coeffs_to_preserve * per_thread_local_stride : per_thread_local_stride;
+    for (Index i = globalRId; i < num_coeffs_to_reduce; i += per_thread_local_stride) {
+      op.reduce(evaluator.impl().coeff(global_offset), &accumulator);
+      localOffset += per_thread_local_stride;
+      global_offset += per_thread_global_stride;
+    }
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(cl::sycl::nd_item<1> itemID) {
+    const Index linearLocalThreadId = itemID.get_local_id(0);
+    Index pLocalThreadId = rt == reduction_dim::outer_most ? linearLocalThreadId % PannelParameters::LocalThreadSizeP
+                                                           : linearLocalThreadId / PannelParameters::LocalThreadSizeR;
+    Index rLocalThreadId = rt == reduction_dim::outer_most ? linearLocalThreadId / PannelParameters::LocalThreadSizeP
+                                                           : linearLocalThreadId % PannelParameters::LocalThreadSizeR;
+    const Index pGroupId = rt == reduction_dim::outer_most ? itemID.get_group(0) % preserve_elements_num_groups
+                                                           : itemID.get_group(0) / reduce_elements_num_groups;
+    const Index rGroupId = rt == reduction_dim::outer_most ? itemID.get_group(0) / preserve_elements_num_groups
+                                                           : itemID.get_group(0) % reduce_elements_num_groups;
+
+    Index globalPId = pGroupId * PannelParameters::LocalThreadSizeP + pLocalThreadId;
+    const Index globalRId = rGroupId * PannelParameters::LocalThreadSizeR + rLocalThreadId;
+    auto scratchPtr = scratch.get_pointer().get();
+    auto outPtr =
+        output_accessor.get_pointer() + (reduce_elements_num_groups > 1 ? rGroupId * num_coeffs_to_preserve : 0);
+    CoeffReturnType accumulator = op.initialize();
+
+    element_wise_reduce(globalRId, globalPId, accumulator);
+
+    accumulator = OpDef::finalise_op(op.finalize(accumulator), num_coeffs_to_reduce);
+    scratchPtr[pLocalThreadId + rLocalThreadId * (PannelParameters::LocalThreadSizeP + PannelParameters::BC)] =
+        accumulator;
+    if (rt == reduction_dim::inner_most) {
+      pLocalThreadId = linearLocalThreadId % PannelParameters::LocalThreadSizeP;
+      rLocalThreadId = linearLocalThreadId / PannelParameters::LocalThreadSizeP;
+      globalPId = pGroupId * PannelParameters::LocalThreadSizeP + pLocalThreadId;
+    }
+
+    /* Apply the reduction operation between the current local
+     * id and the one on the other half of the vector. */
+    auto out_scratch_ptr =
+        scratchPtr + (pLocalThreadId + (rLocalThreadId * (PannelParameters::LocalThreadSizeP + PannelParameters::BC)));
+    itemID.barrier(cl::sycl::access::fence_space::local_space);
+    if (rt == reduction_dim::inner_most) {
+      accumulator = *out_scratch_ptr;
+    }
+    // The Local LocalThreadSizeR is always power of 2
+    EIGEN_UNROLL_LOOP
+    for (Index offset = PannelParameters::LocalThreadSizeR >> 1; offset > 0; offset >>= 1) {
+      if (rLocalThreadId < offset) {
+        op.reduce(out_scratch_ptr[(PannelParameters::LocalThreadSizeP + PannelParameters::BC) * offset], &accumulator);
+        // The result has already been divided for mean reducer in the
+        // previous reduction so no need to divide furthermore
+        *out_scratch_ptr = op.finalize(accumulator);
+      }
+      /* All threads collectively read from global memory into local.
+       * The barrier ensures all threads' IO is resolved before
+       * execution continues (strictly speaking, all threads within
+       * a single work-group - there is no co-ordination between
+       * work-groups, only work-items). */
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+    }
+
+    if (rLocalThreadId == 0 && (globalPId < num_coeffs_to_preserve)) {
+      outPtr[globalPId] = op.finalize(accumulator);
+    }
+  }
+};
+
+template <typename OutScalar, typename Index, typename InputAccessor, typename OutputAccessor, typename OpType>
+struct SecondStepPartialReduction {
+  typedef OpDefiner<OpType, OutScalar, Index, false> OpDef;
+  typedef typename OpDef::type Op;
+  typedef cl::sycl::accessor<OutScalar, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      ScratchAccessor;
+  InputAccessor input_accessor;
+  OutputAccessor output_accessor;
+  Op op;
+  const Index num_coeffs_to_preserve;
+  const Index num_coeffs_to_reduce;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE SecondStepPartialReduction(ScratchAccessor, InputAccessor input_accessor_,
+                                                                   OutputAccessor output_accessor_, OpType op_,
+                                                                   const Index num_coeffs_to_preserve_,
+                                                                   const Index num_coeffs_to_reduce_)
+      : input_accessor(input_accessor_),
+        output_accessor(output_accessor_),
+        op(OpDef::get_op(op_)),
+        num_coeffs_to_preserve(num_coeffs_to_preserve_),
+        num_coeffs_to_reduce(num_coeffs_to_reduce_) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(cl::sycl::nd_item<1> itemID) {
+    const Index globalId = itemID.get_global_id(0);
+
+    if (globalId >= num_coeffs_to_preserve) return;
+
+    auto in_ptr = input_accessor.get_pointer() + globalId;
+
+    OutScalar accumulator = op.initialize();
+// num_coeffs_to_reduce is not bigger that 256
+    for (Index i = 0; i < num_coeffs_to_reduce; i++) {
+      op.reduce(*in_ptr, &accumulator);
+      in_ptr += num_coeffs_to_preserve;
+    }
+    output_accessor.get_pointer()[globalId] = op.finalize(accumulator);
+  }
+};  // namespace internal
+
+template <typename Index, Index LTP, Index LTR, bool BC_>
+struct ReductionPannel {
+  static EIGEN_CONSTEXPR Index LocalThreadSizeP = LTP;
+  static EIGEN_CONSTEXPR Index LocalThreadSizeR = LTR;
+  static EIGEN_CONSTEXPR bool BC = BC_;
+};
+
+template <typename Self, typename Op, TensorSycl::internal::reduction_dim rt>
+struct PartialReducerLauncher {
+  typedef typename Self::EvaluatorPointerType EvaluatorPointerType;
+  typedef typename Self::CoeffReturnType CoeffReturnType;
+  typedef typename Self::Storage Storage;
+  typedef typename Self::Index Index;
+  typedef ReductionPannel<typename Self::Index, EIGEN_SYCL_LOCAL_THREAD_DIM0, EIGEN_SYCL_LOCAL_THREAD_DIM1, true>
+      PannelParameters;
+
+  typedef PartialReductionKernel<Self, Op, PannelParameters, rt> SyclReducerKerneType;
+
+  static bool run(const Self &self, const Op &reducer, const Eigen::SyclDevice &dev, EvaluatorPointerType output,
+                  Index num_coeffs_to_reduce, Index num_coeffs_to_preserve) {
+    Index roundUpP = roundUp(num_coeffs_to_preserve, PannelParameters::LocalThreadSizeP);
+
+    // getPowerOfTwo makes sure local range is power of 2 and <=
+    // maxSyclThreadPerBlock this will help us to avoid extra check on the
+    // kernel
+    static_assert(!((PannelParameters::LocalThreadSizeP * PannelParameters::LocalThreadSizeR) &
+                    (PannelParameters::LocalThreadSizeP * PannelParameters::LocalThreadSizeR - 1)),
+                  "The Local thread size must be a power of 2 for the reduction "
+                  "operation");
+
+    EIGEN_CONSTEXPR Index localRange = PannelParameters::LocalThreadSizeP * PannelParameters::LocalThreadSizeR;
+    // In this step, we force the code not to be more than 2-step reduction:
+    // Our empirical research shows that if each thread reduces at least 64
+    // elemnts individually, we get better performance. However, this can change
+    // on different platforms. In this step we force the code not to be
+    // morthan step reduction: Our empirical research shows that for inner_most
+    // dim reducer, it is better to have 8 group in a reduce dimension for sizes
+    // > 1024 to achieve the best performance.
+    const Index reductionPerThread = 64;
+    Index cu = dev.getPowerOfTwo(dev.getNumSyclMultiProcessors(), true);
+    const Index pNumGroups = roundUpP / PannelParameters::LocalThreadSizeP;
+    Index rGroups = (cu + pNumGroups - 1) / pNumGroups;
+    const Index rNumGroups = num_coeffs_to_reduce > reductionPerThread * localRange ? std::min(rGroups, localRange) : 1;
+    const Index globalRange = pNumGroups * rNumGroups * localRange;
+
+    EIGEN_CONSTEXPR Index scratchSize =
+        PannelParameters::LocalThreadSizeR * (PannelParameters::LocalThreadSizeP + PannelParameters::BC);
+    auto thread_range = cl::sycl::nd_range<1>(cl::sycl::range<1>(globalRange), cl::sycl::range<1>(localRange));
+    if (rNumGroups > 1) {
+      CoeffReturnType *temp_pointer = static_cast<CoeffReturnType *>(
+          dev.allocate_temp(num_coeffs_to_preserve * rNumGroups * sizeof(CoeffReturnType)));
+      EvaluatorPointerType temp_accessor = dev.get(temp_pointer);
+      dev.template unary_kernel_launcher<CoeffReturnType, SyclReducerKerneType>(
+          self, temp_accessor, thread_range, scratchSize, reducer, pNumGroups, rNumGroups, num_coeffs_to_preserve,
+          num_coeffs_to_reduce);
+
+      typedef SecondStepPartialReduction<CoeffReturnType, Index, EvaluatorPointerType, EvaluatorPointerType, Op>
+          SecondStepPartialReductionKernel;
+
+      dev.template unary_kernel_launcher<CoeffReturnType, SecondStepPartialReductionKernel>(
+          temp_accessor, output,
+          cl::sycl::nd_range<1>(cl::sycl::range<1>(pNumGroups * localRange), cl::sycl::range<1>(localRange)), Index(1),
+          reducer, num_coeffs_to_preserve, rNumGroups);
+
+      self.device().deallocate_temp(temp_pointer);
+    } else {
+      dev.template unary_kernel_launcher<CoeffReturnType, SyclReducerKerneType>(
+          self, output, thread_range, scratchSize, reducer, pNumGroups, rNumGroups, num_coeffs_to_preserve,
+          num_coeffs_to_reduce);
+    }
+    return false;
+  }
+};
+}  // namespace internal
+}  // namespace TensorSycl
+
+namespace internal {
+
+template <typename Self, typename Op, bool Vectorizable>
+struct FullReducer<Self, Op, Eigen::SyclDevice, Vectorizable> {
+  typedef typename Self::CoeffReturnType CoeffReturnType;
+  typedef typename Self::EvaluatorPointerType EvaluatorPointerType;
+  static EIGEN_CONSTEXPR bool HasOptimizedImplementation = true;
+  static EIGEN_CONSTEXPR int PacketSize = Self::PacketAccess ? Self::PacketSize : 1;
+  static void run(const Self &self, Op &reducer, const Eigen::SyclDevice &dev, EvaluatorPointerType data) {
+    typedef typename conditional<Self::PacketAccess, typename Self::PacketReturnType, CoeffReturnType>::type OutType;
+    static_assert(!((EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1) &
+                    (EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1 - 1)),
+                  "The Local thread size must be a power of 2 for the reduction "
+                  "operation");
+    EIGEN_CONSTEXPR Index local_range = EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1;
+
+    typename Self::Index inputSize = self.impl().dimensions().TotalSize();
+    // In this step we force the code not to be more than 2-step reduction:
+    // Our empirical research shows that if each thread reduces at least 512
+    // elemnts individually, we get better performance.
+    const Index reductionPerThread = 2048;
+    // const Index num_work_group =
+    Index reductionGroup = dev.getPowerOfTwo(
+        (inputSize + (reductionPerThread * local_range - 1)) / (reductionPerThread * local_range), true);
+    const Index num_work_group = std::min(reductionGroup, local_range);
+    // 1
+    // ? local_range
+    // : 1);
+    const Index global_range = num_work_group * local_range;
+
+    auto thread_range = cl::sycl::nd_range<1>(cl::sycl::range<1>(global_range), cl::sycl::range<1>(local_range));
+    typedef TensorSycl::internal::FullReductionKernelFunctor<Self, Op, local_range> reduction_kernel_t;
+    if (num_work_group > 1) {
+      CoeffReturnType *temp_pointer =
+          static_cast<CoeffReturnType *>(dev.allocate_temp(num_work_group * sizeof(CoeffReturnType)));
+      typename Self::EvaluatorPointerType tmp_global_accessor = dev.get(temp_pointer);
+      dev.template unary_kernel_launcher<OutType, reduction_kernel_t>(self, tmp_global_accessor, thread_range,
+                                                                      local_range, inputSize, reducer);
+
+      typedef TensorSycl::internal::SecondStepFullReducer<CoeffReturnType, Op, EvaluatorPointerType,
+                                                          EvaluatorPointerType, Index, local_range>
+          GenericRKernel;
+      dev.template unary_kernel_launcher<CoeffReturnType, GenericRKernel>(
+          tmp_global_accessor, data,
+          cl::sycl::nd_range<1>(cl::sycl::range<1>(num_work_group), cl::sycl::range<1>(num_work_group)), num_work_group,
+          reducer);
+
+      dev.deallocate_temp(temp_pointer);
+    } else {
+      dev.template unary_kernel_launcher<OutType, reduction_kernel_t>(self, data, thread_range, local_range, inputSize,
+                                                                      reducer);
+    }
+  }
+};
+// vectorizable inner_most most dim preserver
+// col reduction
+template <typename Self, typename Op>
+struct OuterReducer<Self, Op, Eigen::SyclDevice> {
+  static EIGEN_CONSTEXPR bool HasOptimizedImplementation = true;
+
+  static bool run(const Self &self, const Op &reducer, const Eigen::SyclDevice &dev,
+                  typename Self::EvaluatorPointerType output, typename Self::Index num_coeffs_to_reduce,
+                  typename Self::Index num_coeffs_to_preserve) {
+    return ::Eigen::TensorSycl::internal::PartialReducerLauncher<
+        Self, Op, ::Eigen::TensorSycl::internal::reduction_dim::outer_most>::run(self, reducer, dev, output,
+                                                                                 num_coeffs_to_reduce,
+                                                                                 num_coeffs_to_preserve);
+  }
+};
+// row reduction
+template <typename Self, typename Op>
+struct InnerReducer<Self, Op, Eigen::SyclDevice> {
+  static EIGEN_CONSTEXPR bool HasOptimizedImplementation = true;
+
+  static bool run(const Self &self, const Op &reducer, const Eigen::SyclDevice &dev,
+                  typename Self::EvaluatorPointerType output, typename Self::Index num_coeffs_to_reduce,
+                  typename Self::Index num_coeffs_to_preserve) {
+    return ::Eigen::TensorSycl::internal::PartialReducerLauncher<
+        Self, Op, ::Eigen::TensorSycl::internal::reduction_dim::inner_most>::run(self, reducer, dev, output,
+                                                                                 num_coeffs_to_reduce,
+                                                                                 num_coeffs_to_preserve);
+  }
+};
+
+// ArmgMax uses this kernel for partial reduction//
+// TODO(@mehdi.goli) come up with a better kernel
+// generic partial reduction
+template <typename Self, typename Op>
+struct GenericReducer<Self, Op, Eigen::SyclDevice> {
+  static EIGEN_CONSTEXPR bool HasOptimizedImplementation = false;
+  static bool run(const Self &self, const Op &reducer, const Eigen::SyclDevice &dev,
+                  typename Self::EvaluatorPointerType output, typename Self::Index num_values_to_reduce,
+                  typename Self::Index num_coeffs_to_preserve) {
+    typename Self::Index range, GRange, tileSize;
+    dev.parallel_for_setup(num_coeffs_to_preserve, tileSize, range, GRange);
+
+    dev.template unary_kernel_launcher<typename Self::CoeffReturnType,
+                                       TensorSycl::internal::GenericNondeterministicReducer<Self, Op>>(
+        self, output, cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), Index(1),
+        reducer, range, (num_values_to_reduce != 0) ? num_values_to_reduce : static_cast<Index>(1));
+    return false;
+  }
+};
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorRef.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorRef.h
new file mode 100644
index 0000000..a27d364
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorRef.h
@@ -0,0 +1,454 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_REF_H
+#define EIGEN_CXX11_TENSOR_TENSOR_REF_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <typename Dimensions, typename Scalar>
+class TensorLazyBaseEvaluator {
+ public:
+  TensorLazyBaseEvaluator() : m_refcount(0) { }
+  virtual ~TensorLazyBaseEvaluator() { }
+
+  EIGEN_DEVICE_FUNC virtual const Dimensions& dimensions() const = 0;
+  EIGEN_DEVICE_FUNC virtual const Scalar* data() const = 0;
+
+  EIGEN_DEVICE_FUNC virtual const Scalar coeff(DenseIndex index) const = 0;
+  EIGEN_DEVICE_FUNC virtual Scalar& coeffRef(DenseIndex index) = 0;
+
+  void incrRefCount() { ++m_refcount; }
+  void decrRefCount() { --m_refcount; }
+  int refCount() const { return m_refcount; }
+
+ private:
+  // No copy, no assignment;
+  TensorLazyBaseEvaluator(const TensorLazyBaseEvaluator& other);
+  TensorLazyBaseEvaluator& operator = (const TensorLazyBaseEvaluator& other);
+
+  int m_refcount;
+};
+
+
+template <typename Dimensions, typename Expr, typename Device>
+class TensorLazyEvaluatorReadOnly : public TensorLazyBaseEvaluator<Dimensions, typename TensorEvaluator<Expr, Device>::Scalar> {
+ public:
+  //  typedef typename TensorEvaluator<Expr, Device>::Dimensions Dimensions;
+  typedef typename TensorEvaluator<Expr, Device>::Scalar Scalar;
+  typedef StorageMemory<Scalar, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+  typedef  TensorEvaluator<Expr, Device> EvalType;
+
+  TensorLazyEvaluatorReadOnly(const Expr& expr, const Device& device) : m_impl(expr, device), m_dummy(Scalar(0)) {
+    m_dims = m_impl.dimensions();
+    m_impl.evalSubExprsIfNeeded(NULL);
+  }
+  virtual ~TensorLazyEvaluatorReadOnly() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC virtual const Dimensions& dimensions() const {
+    return m_dims;
+  }
+  EIGEN_DEVICE_FUNC virtual const Scalar* data() const {
+    return m_impl.data();
+  }
+
+  EIGEN_DEVICE_FUNC virtual const Scalar coeff(DenseIndex index) const {
+    return m_impl.coeff(index);
+  }
+  EIGEN_DEVICE_FUNC virtual Scalar& coeffRef(DenseIndex /*index*/) {
+    eigen_assert(false && "can't reference the coefficient of a rvalue");
+    return m_dummy;
+  };
+
+ protected:
+  TensorEvaluator<Expr, Device> m_impl;
+  Dimensions m_dims;
+  Scalar m_dummy;
+};
+
+template <typename Dimensions, typename Expr, typename Device>
+class TensorLazyEvaluatorWritable : public TensorLazyEvaluatorReadOnly<Dimensions, Expr, Device> {
+ public:
+  typedef TensorLazyEvaluatorReadOnly<Dimensions, Expr, Device> Base;
+  typedef typename Base::Scalar Scalar;
+  typedef StorageMemory<Scalar, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  TensorLazyEvaluatorWritable(const Expr& expr, const Device& device) : Base(expr, device) {
+  }
+  virtual ~TensorLazyEvaluatorWritable() {
+  }
+
+  EIGEN_DEVICE_FUNC virtual Scalar& coeffRef(DenseIndex index) {
+    return this->m_impl.coeffRef(index);
+  }
+};
+
+template <typename Dimensions, typename Expr, typename Device>
+class TensorLazyEvaluator : public internal::conditional<bool(internal::is_lvalue<Expr>::value),
+                            TensorLazyEvaluatorWritable<Dimensions, Expr, Device>,
+                            TensorLazyEvaluatorReadOnly<Dimensions, const Expr, Device> >::type {
+ public:
+  typedef typename internal::conditional<bool(internal::is_lvalue<Expr>::value),
+                                         TensorLazyEvaluatorWritable<Dimensions, Expr, Device>,
+                                         TensorLazyEvaluatorReadOnly<Dimensions, const Expr, Device> >::type Base;
+  typedef typename Base::Scalar Scalar;
+
+  TensorLazyEvaluator(const Expr& expr, const Device& device) : Base(expr, device) {
+  }
+  virtual ~TensorLazyEvaluator() {
+  }
+};
+
+}  // namespace internal
+
+
+/** \class TensorRef
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief A reference to a tensor expression
+  * The expression will be evaluated lazily (as much as possible).
+  *
+  */
+template<typename PlainObjectType> class TensorRef : public TensorBase<TensorRef<PlainObjectType> >
+{
+  public:
+    typedef TensorRef<PlainObjectType> Self;
+    typedef typename PlainObjectType::Base Base;
+    typedef typename Eigen::internal::nested<Self>::type Nested;
+    typedef typename internal::traits<PlainObjectType>::StorageKind StorageKind;
+    typedef typename internal::traits<PlainObjectType>::Index Index;
+    typedef typename internal::traits<PlainObjectType>::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef typename Base::CoeffReturnType CoeffReturnType;
+    typedef Scalar* PointerType;
+    typedef PointerType PointerArgType;
+
+    static const Index NumIndices = PlainObjectType::NumIndices;
+    typedef typename PlainObjectType::Dimensions Dimensions;
+
+    enum {
+      IsAligned = false,
+      PacketAccess = false,
+      BlockAccess = false,
+      PreferBlockAccess = false,
+      Layout = PlainObjectType::Layout,
+      CoordAccess = false,  // to be implemented
+      RawAccess = false
+    };
+
+    //===- Tensor block evaluation strategy (see TensorBlock.h) -----------===//
+    typedef internal::TensorBlockNotImplemented TensorBlock;
+    //===------------------------------------------------------------------===//
+
+    EIGEN_STRONG_INLINE TensorRef() : m_evaluator(NULL) {
+    }
+
+    template <typename Expression>
+    EIGEN_STRONG_INLINE TensorRef(const Expression& expr) : m_evaluator(new internal::TensorLazyEvaluator<Dimensions, Expression, DefaultDevice>(expr, DefaultDevice())) {
+      m_evaluator->incrRefCount();
+    }
+
+    template <typename Expression>
+    EIGEN_STRONG_INLINE TensorRef& operator = (const Expression& expr) {
+      unrefEvaluator();
+      m_evaluator = new internal::TensorLazyEvaluator<Dimensions, Expression, DefaultDevice>(expr, DefaultDevice());
+      m_evaluator->incrRefCount();
+      return *this;
+    }
+
+    ~TensorRef() {
+      unrefEvaluator();
+    }
+
+    TensorRef(const TensorRef& other) : m_evaluator(other.m_evaluator) {
+      eigen_assert(m_evaluator->refCount() > 0);
+      m_evaluator->incrRefCount();
+    }
+
+    TensorRef& operator = (const TensorRef& other) {
+      if (this != &other) {
+        unrefEvaluator();
+        m_evaluator = other.m_evaluator;
+        eigen_assert(m_evaluator->refCount() > 0);
+        m_evaluator->incrRefCount();
+      }
+      return *this;
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index rank() const { return m_evaluator->dimensions().size(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index dimension(Index n) const { return m_evaluator->dimensions()[n]; }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_evaluator->dimensions(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index size() const { return m_evaluator->dimensions().TotalSize(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar* data() const { return m_evaluator->data(); }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar operator()(Index index) const
+    {
+      return m_evaluator->coeff(index);
+    }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template<typename... IndexTypes> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar operator()(Index firstIndex, IndexTypes... otherIndices) const
+    {
+      const std::size_t num_indices = (sizeof...(otherIndices) + 1);
+      const array<Index, num_indices> indices{{firstIndex, otherIndices...}};
+      return coeff(indices);
+    }
+    template<typename... IndexTypes> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index firstIndex, IndexTypes... otherIndices)
+    {
+      const std::size_t num_indices = (sizeof...(otherIndices) + 1);
+      const array<Index, num_indices> indices{{firstIndex, otherIndices...}};
+      return coeffRef(indices);
+    }
+#else
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar operator()(Index i0, Index i1) const
+    {
+      array<Index, 2> indices;
+      indices[0] = i0;
+      indices[1] = i1;
+      return coeff(indices);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar operator()(Index i0, Index i1, Index i2) const
+    {
+      array<Index, 3> indices;
+      indices[0] = i0;
+      indices[1] = i1;
+      indices[2] = i2;
+      return coeff(indices);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar operator()(Index i0, Index i1, Index i2, Index i3) const
+    {
+      array<Index, 4> indices;
+      indices[0] = i0;
+      indices[1] = i1;
+      indices[2] = i2;
+      indices[3] = i3;
+      return coeff(indices);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar operator()(Index i0, Index i1, Index i2, Index i3, Index i4) const
+    {
+      array<Index, 5> indices;
+      indices[0] = i0;
+      indices[1] = i1;
+      indices[2] = i2;
+      indices[3] = i3;
+      indices[4] = i4;
+      return coeff(indices);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index i0, Index i1)
+    {
+      array<Index, 2> indices;
+      indices[0] = i0;
+      indices[1] = i1;
+      return coeffRef(indices);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index i0, Index i1, Index i2)
+    {
+      array<Index, 3> indices;
+      indices[0] = i0;
+      indices[1] = i1;
+      indices[2] = i2;
+      return coeffRef(indices);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2, Index i3)
+    {
+      array<Index, 4> indices;
+      indices[0] = i0;
+      indices[1] = i1;
+      indices[2] = i2;
+      indices[3] = i3;
+      return coeffRef(indices);
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index i0, Index i1, Index i2, Index i3, Index i4)
+    {
+      array<Index, 5> indices;
+      indices[0] = i0;
+      indices[1] = i1;
+      indices[2] = i2;
+      indices[3] = i3;
+      indices[4] = i4;
+      return coeffRef(indices);
+    }
+#endif
+
+    template <std::size_t NumIndices> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar coeff(const array<Index, NumIndices>& indices) const
+    {
+      const Dimensions& dims = this->dimensions();
+      Index index = 0;
+      if (PlainObjectType::Options & RowMajor) {
+        index += indices[0];
+        for (size_t i = 1; i < NumIndices; ++i) {
+          index = index * dims[i] + indices[i];
+        }
+      } else {
+        index += indices[NumIndices-1];
+        for (int i = NumIndices-2; i >= 0; --i) {
+          index = index * dims[i] + indices[i];
+        }
+      }
+      return m_evaluator->coeff(index);
+    }
+    template <std::size_t NumIndices> EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(const array<Index, NumIndices>& indices)
+    {
+      const Dimensions& dims = this->dimensions();
+      Index index = 0;
+      if (PlainObjectType::Options & RowMajor) {
+        index += indices[0];
+        for (size_t i = 1; i < NumIndices; ++i) {
+          index = index * dims[i] + indices[i];
+        }
+      } else {
+        index += indices[NumIndices-1];
+        for (int i = NumIndices-2; i >= 0; --i) {
+          index = index * dims[i] + indices[i];
+        }
+      }
+      return m_evaluator->coeffRef(index);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
+    {
+      return m_evaluator->coeff(index);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
+    {
+      return m_evaluator->coeffRef(index);
+    }
+
+  private:
+    EIGEN_STRONG_INLINE void unrefEvaluator() {
+      if (m_evaluator) {
+        m_evaluator->decrRefCount();
+        if (m_evaluator->refCount() == 0) {
+          delete m_evaluator;
+        }
+      }
+    }
+
+  internal::TensorLazyBaseEvaluator<Dimensions, Scalar>* m_evaluator;
+};
+
+
+// evaluator for rvalues
+template<typename Derived, typename Device>
+struct TensorEvaluator<const TensorRef<Derived>, Device>
+{
+  typedef typename Derived::Index Index;
+  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef typename Derived::Dimensions Dimensions;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = false,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorRef<Derived>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const TensorRef<Derived>& m, const Device&)
+      : m_ref(m)
+  { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_ref.dimensions(); }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    return true;
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    return m_ref.coeff(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) {
+    return m_ref.coeffRef(index);
+  }
+
+  EIGEN_DEVICE_FUNC const Scalar* data() const { return m_ref.data(); }
+
+ protected:
+  TensorRef<Derived> m_ref;
+};
+
+
+// evaluator for lvalues
+template<typename Derived, typename Device>
+struct TensorEvaluator<TensorRef<Derived>, Device> : public TensorEvaluator<const TensorRef<Derived>, Device>
+{
+  typedef typename Derived::Index Index;
+  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::Scalar CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef typename Derived::Dimensions Dimensions;
+
+  typedef TensorEvaluator<const TensorRef<Derived>, Device> Base;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = false,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(TensorRef<Derived>& m, const Device& d) : Base(m, d)
+  { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) {
+    return this->m_ref.coeffRef(index);
+  }
+};
+
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_REF_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorReverse.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorReverse.h
new file mode 100644
index 0000000..586ce68
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorReverse.h
@@ -0,0 +1,465 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Navdeep Jaitly <ndjaitly@google.com>
+//                    Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_REVERSE_H
+#define EIGEN_CXX11_TENSOR_TENSOR_REVERSE_H
+namespace Eigen {
+
+/** \class TensorReverse
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor reverse elements class.
+  *
+  */
+namespace internal {
+template<typename ReverseDimensions, typename XprType>
+struct traits<TensorReverseOp<ReverseDimensions,
+                              XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename ReverseDimensions, typename XprType>
+struct eval<TensorReverseOp<ReverseDimensions, XprType>, Eigen::Dense>
+{
+  typedef const TensorReverseOp<ReverseDimensions, XprType>& type;
+};
+
+template<typename ReverseDimensions, typename XprType>
+struct nested<TensorReverseOp<ReverseDimensions, XprType>, 1,
+            typename eval<TensorReverseOp<ReverseDimensions, XprType> >::type>
+{
+  typedef TensorReverseOp<ReverseDimensions, XprType> type;
+};
+
+}  // end namespace internal
+
+template<typename ReverseDimensions, typename XprType>
+class TensorReverseOp : public TensorBase<TensorReverseOp<ReverseDimensions,
+                                          XprType>, WriteAccessors>
+{
+  public:
+    typedef TensorBase<TensorReverseOp<ReverseDimensions, XprType>, WriteAccessors>Base;
+    typedef typename Eigen::internal::traits<TensorReverseOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename XprType::CoeffReturnType CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorReverseOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorReverseOp>::StorageKind
+                                                                      StorageKind;
+    typedef typename Eigen::internal::traits<TensorReverseOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorReverseOp(
+      const XprType& expr, const ReverseDimensions& reverse_dims)
+      : m_xpr(expr), m_reverse_dims(reverse_dims) { }
+
+    EIGEN_DEVICE_FUNC
+    const ReverseDimensions& reverse() const { return m_reverse_dims; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorReverseOp)
+
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const ReverseDimensions m_reverse_dims;
+};
+
+// Eval as rvalue
+template<typename ReverseDimensions, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorReverseOp<ReverseDimensions, ArgType>, Device>
+{
+  typedef TensorReverseOp<ReverseDimensions, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<ReverseDimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned         = false,
+    PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess       = NumDims > 0,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
+  };
+
+  typedef internal::TensorIntDivisor<Index> IndexDivisor;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename TensorEvaluator<const ArgType, Device>::TensorBlock
+      ArgTensorBlock;
+
+  typedef typename internal::TensorMaterializedBlock<CoeffReturnType, NumDims,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device),
+        m_reverse(op.reverse()),
+        m_device(device)
+  {
+    // Reversing a scalar isn't supported yet. It would be a no-op anyway.
+    EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    // Compute strides
+    m_dimensions = m_impl.dimensions();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_strides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_strides[i] = m_strides[i-1] * m_dimensions[i-1];
+        if (m_strides[i] > 0) m_fastStrides[i] = IndexDivisor(m_strides[i]);
+      }
+    } else {
+      m_strides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_strides[i] = m_strides[i+1] * m_dimensions[i+1];
+        if (m_strides[i] > 0) m_fastStrides[i] = IndexDivisor(m_strides[i]);
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index reverseIndex(
+      Index index) const {
+    eigen_assert(index < dimensions().TotalSize());
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        Index idx = index / m_fastStrides[i];
+        index -= idx * m_strides[i];
+        if (m_reverse[i]) {
+          idx = m_dimensions[i] - idx - 1;
+        }
+        inputIndex += idx * m_strides[i] ;
+      }
+      if (m_reverse[0]) {
+        inputIndex += (m_dimensions[0] - index - 1);
+      } else {
+        inputIndex += index;
+      }
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        Index idx = index / m_fastStrides[i];
+        index -= idx * m_strides[i];
+        if (m_reverse[i]) {
+          idx = m_dimensions[i] - idx - 1;
+        }
+        inputIndex += idx * m_strides[i] ;
+      }
+      if (m_reverse[NumDims-1]) {
+        inputIndex += (m_dimensions[NumDims-1] - index - 1);
+      } else {
+        inputIndex += index;
+      }
+    }
+    return inputIndex;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(
+      Index index) const  {
+    return m_impl.coeff(reverseIndex(index));
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    // TODO(ndjaitly): write a better packing routine that uses
+    // local structure.
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type
+                                                            values[PacketSize];
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    const size_t target_size = m_device.lastLevelCacheSize();
+    // Block evaluation reads underlying memory in reverse order, and default
+    // cost model does not properly catch this in bytes stored/loaded.
+    return internal::TensorBlockResourceRequirements::skewed<Scalar>(
+               target_size)
+        .addCostPerCoeff({0, 0, 24});
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool /*root_of_expr_ast*/ = false) const {
+    // TODO(ezhulenev): If underlying tensor expression supports and prefers
+    // block evaluation we must use it. Currently we use coeff and packet
+    // access into the underlying tensor expression.
+    // static const bool useBlockAccessForArgType =
+    //     TensorEvaluator<ArgType, Device>::BlockAccess &&
+    //     TensorEvaluator<ArgType, Device>::PreferBlockAccess;
+
+    static const bool isColMajor =
+        static_cast<int>(Layout) == static_cast<int>(ColMajor);
+
+    static const Index inner_dim_idx = isColMajor ? 0 : NumDims - 1;
+    const bool inner_dim_reversed = m_reverse[inner_dim_idx];
+
+    // Offset in the output block.
+    Index block_offset = 0;
+
+    // Offset in the input Tensor.
+    Index input_offset = reverseIndex(desc.offset());
+
+    // Initialize output block iterator state. Dimension in this array are
+    // always in inner_most -> outer_most order (col major layout).
+    array<BlockIteratorState, NumDims> it;
+    for (int i = 0; i < NumDims; ++i) {
+      const int dim = isColMajor ? i : NumDims - 1 - i;
+      it[i].size = desc.dimension(dim);
+      it[i].count = 0;
+      it[i].reverse = m_reverse[dim];
+
+      it[i].block_stride =
+          i == 0 ? 1 : (it[i - 1].size * it[i - 1].block_stride);
+      it[i].block_span = it[i].block_stride * (it[i].size - 1);
+
+      it[i].input_stride = m_strides[dim];
+      it[i].input_span = it[i].input_stride * (it[i].size - 1);
+
+      if (it[i].reverse) {
+        it[i].input_stride = -1 * it[i].input_stride;
+        it[i].input_span = -1 * it[i].input_span;
+      }
+    }
+
+    // If multiple inner dimensions have the same reverse flag, check if we can
+    // merge them into a single virtual inner dimension.
+    int effective_inner_dim = 0;
+    for (int i = 1; i < NumDims; ++i) {
+      if (it[i].reverse != it[effective_inner_dim].reverse) break;
+      if (it[i].block_stride != it[effective_inner_dim].size) break;
+      if (it[i].block_stride != numext::abs(it[i].input_stride)) break;
+
+      it[i].size = it[effective_inner_dim].size * it[i].size;
+
+      it[i].block_stride = 1;
+      it[i].input_stride = (inner_dim_reversed ? -1 : 1);
+
+      it[i].block_span = it[i].block_stride * (it[i].size - 1);
+      it[i].input_span = it[i].input_stride * (it[i].size - 1);
+
+      effective_inner_dim = i;
+    }
+
+    eigen_assert(it[effective_inner_dim].block_stride == 1);
+    eigen_assert(it[effective_inner_dim].input_stride ==
+                 (inner_dim_reversed ? -1 : 1));
+
+    const Index inner_dim_size = it[effective_inner_dim].size;
+
+    // Prepare storage for the materialized reverse result.
+    const typename TensorBlock::Storage block_storage =
+        TensorBlock::prepareStorage(desc, scratch);
+    CoeffReturnType* block_buffer = block_storage.data();
+
+    while (it[NumDims - 1].count < it[NumDims - 1].size) {
+      // Copy inner-most dimension data from reversed location in input.
+      Index dst = block_offset;
+      Index src = input_offset;
+
+      // NOTE(ezhulenev): Adding vectorized path with internal::preverse showed
+      // worse results in benchmarks than a simple coefficient loop.
+      if (inner_dim_reversed) {
+        for (Index i = 0; i < inner_dim_size; ++i) {
+          block_buffer[dst] = m_impl.coeff(src);
+          ++dst;
+          --src;
+        }
+      } else {
+        for (Index i = 0; i < inner_dim_size; ++i) {
+          block_buffer[dst] = m_impl.coeff(src);
+          ++dst;
+          ++src;
+        }
+      }
+
+      // For the 1d tensor we need to generate only one inner-most dimension.
+      if ((NumDims - effective_inner_dim) == 1) break;
+
+      // Update offset.
+      for (Index i = effective_inner_dim + 1; i < NumDims; ++i) {
+        if (++it[i].count < it[i].size) {
+          block_offset += it[i].block_stride;
+          input_offset += it[i].input_stride;
+          break;
+        }
+        if (i != NumDims - 1) it[i].count = 0;
+        block_offset -= it[i].block_span;
+        input_offset -= it[i].input_span;
+      }
+    }
+
+    return block_storage.AsTensorMaterializedBlock();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    double compute_cost = NumDims * (2 * TensorOpCost::AddCost<Index>() +
+                                     2 * TensorOpCost::MulCost<Index>() +
+                                     TensorOpCost::DivCost<Index>());
+    for (int i = 0; i < NumDims; ++i) {
+      if (m_reverse[i]) {
+        compute_cost += 2 * TensorOpCost::AddCost<Index>();
+      }
+    }
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, compute_cost, false /* vectorized */, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC typename Storage::Type data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ protected:
+  Dimensions m_dimensions;
+  array<Index, NumDims> m_strides;
+  array<IndexDivisor, NumDims> m_fastStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+  ReverseDimensions m_reverse;
+  const Device EIGEN_DEVICE_REF m_device;
+
+ private:
+  struct BlockIteratorState {
+    BlockIteratorState()
+        : size(0),
+          count(0),
+          reverse(false),
+          block_stride(0),
+          block_span(0),
+          input_stride(0),
+          input_span(0) {}
+
+    Index size;
+    Index count;
+    bool reverse;
+    Index block_stride;
+    Index block_span;
+    Index input_stride;
+    Index input_span;
+  };
+};
+
+// Eval as lvalue
+
+template <typename ReverseDimensions, typename ArgType, typename Device>
+struct TensorEvaluator<TensorReverseOp<ReverseDimensions, ArgType>, Device>
+    : public TensorEvaluator<const TensorReverseOp<ReverseDimensions, ArgType>,
+                             Device> {
+  typedef TensorEvaluator<const TensorReverseOp<ReverseDimensions, ArgType>,
+                          Device> Base;
+  typedef TensorReverseOp<ReverseDimensions, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<ReverseDimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : Base(op, device) {}
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Dimensions& dimensions() const { return this->m_dimensions; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) {
+    return this->m_impl.coeffRef(this->reverseIndex(index));
+  }
+
+  template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x) {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    // This code is pilfered from TensorMorphing.h
+    EIGEN_ALIGN_MAX CoeffReturnType values[PacketSize];
+    internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      this->coeffRef(index+i) = values[i];
+    }
+  }
+};
+
+
+}  // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_REVERSE_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorScan.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorScan.h
new file mode 100644
index 0000000..beae854
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorScan.h
@@ -0,0 +1,528 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Igor Babuschkin <igor@babuschk.in>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_SCAN_H
+#define EIGEN_CXX11_TENSOR_TENSOR_SCAN_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <typename Op, typename XprType>
+struct traits<TensorScanOp<Op, XprType> >
+    : public traits<XprType> {
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename Op, typename XprType>
+struct eval<TensorScanOp<Op, XprType>, Eigen::Dense>
+{
+  typedef const TensorScanOp<Op, XprType>& type;
+};
+
+template<typename Op, typename XprType>
+struct nested<TensorScanOp<Op, XprType>, 1,
+            typename eval<TensorScanOp<Op, XprType> >::type>
+{
+  typedef TensorScanOp<Op, XprType> type;
+};
+} // end namespace internal
+
+/** \class TensorScan
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor scan class.
+  */
+template <typename Op, typename XprType>
+class TensorScanOp
+    : public TensorBase<TensorScanOp<Op, XprType>, ReadOnlyAccessors> {
+public:
+  typedef typename Eigen::internal::traits<TensorScanOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorScanOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorScanOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorScanOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorScanOp(
+      const XprType& expr, const Index& axis, bool exclusive = false, const Op& op = Op())
+      : m_expr(expr), m_axis(axis), m_accumulator(op), m_exclusive(exclusive) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Index axis() const { return m_axis; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const XprType& expression() const { return m_expr; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Op accumulator() const { return m_accumulator; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  bool exclusive() const { return m_exclusive; }
+
+protected:
+  typename XprType::Nested m_expr;
+  const Index m_axis;
+  const Op m_accumulator;
+  const bool m_exclusive;
+};
+
+
+namespace internal {
+
+template <typename Self>
+EIGEN_STRONG_INLINE void ReduceScalar(Self& self, Index offset,
+                                      typename Self::CoeffReturnType* data) {
+  // Compute the scan along the axis, starting at the given offset
+  typename Self::CoeffReturnType accum = self.accumulator().initialize();
+  if (self.stride() == 1) {
+    if (self.exclusive()) {
+      for (Index curr = offset; curr < offset + self.size(); ++curr) {
+        data[curr] = self.accumulator().finalize(accum);
+        self.accumulator().reduce(self.inner().coeff(curr), &accum);
+      }
+    } else {
+      for (Index curr = offset; curr < offset + self.size(); ++curr) {
+        self.accumulator().reduce(self.inner().coeff(curr), &accum);
+        data[curr] = self.accumulator().finalize(accum);
+      }
+    }
+  } else {
+    if (self.exclusive()) {
+      for (Index idx3 = 0; idx3 < self.size(); idx3++) {
+        Index curr = offset + idx3 * self.stride();
+        data[curr] = self.accumulator().finalize(accum);
+        self.accumulator().reduce(self.inner().coeff(curr), &accum);
+      }
+    } else {
+      for (Index idx3 = 0; idx3 < self.size(); idx3++) {
+        Index curr = offset + idx3 * self.stride();
+        self.accumulator().reduce(self.inner().coeff(curr), &accum);
+        data[curr] = self.accumulator().finalize(accum);
+      }
+    }
+  }
+}
+
+template <typename Self>
+EIGEN_STRONG_INLINE void ReducePacket(Self& self, Index offset,
+                                      typename Self::CoeffReturnType* data) {
+  using Scalar = typename Self::CoeffReturnType;
+  using Packet = typename Self::PacketReturnType;
+  // Compute the scan along the axis, starting at the calculated offset
+  Packet accum = self.accumulator().template initializePacket<Packet>();
+  if (self.stride() == 1) {
+    if (self.exclusive()) {
+      for (Index curr = offset; curr < offset + self.size(); ++curr) {
+        internal::pstoreu<Scalar, Packet>(data + curr, self.accumulator().finalizePacket(accum));
+        self.accumulator().reducePacket(self.inner().template packet<Unaligned>(curr), &accum);
+      }
+    } else {
+      for (Index curr = offset; curr < offset + self.size(); ++curr) {
+        self.accumulator().reducePacket(self.inner().template packet<Unaligned>(curr), &accum);
+        internal::pstoreu<Scalar, Packet>(data + curr, self.accumulator().finalizePacket(accum));
+      }
+    }
+  } else {
+    if (self.exclusive()) {
+      for (Index idx3 = 0; idx3 < self.size(); idx3++) {
+        const Index curr = offset + idx3 * self.stride();
+        internal::pstoreu<Scalar, Packet>(data + curr, self.accumulator().finalizePacket(accum));
+        self.accumulator().reducePacket(self.inner().template packet<Unaligned>(curr), &accum);
+      }
+    } else {
+      for (Index idx3 = 0; idx3 < self.size(); idx3++) {
+        const Index curr = offset + idx3 * self.stride();
+        self.accumulator().reducePacket(self.inner().template packet<Unaligned>(curr), &accum);
+        internal::pstoreu<Scalar, Packet>(data + curr, self.accumulator().finalizePacket(accum));
+      }
+    }
+  }
+}
+
+template <typename Self, bool Vectorize, bool Parallel>
+struct ReduceBlock {
+  EIGEN_STRONG_INLINE void operator()(Self& self, Index idx1,
+                                      typename Self::CoeffReturnType* data) {
+    for (Index idx2 = 0; idx2 < self.stride(); idx2++) {
+      // Calculate the starting offset for the scan
+      Index offset = idx1 + idx2;
+      ReduceScalar(self, offset, data);
+    }
+  }
+};
+
+// Specialization for vectorized reduction.
+template <typename Self>
+struct ReduceBlock<Self, /*Vectorize=*/true, /*Parallel=*/false> {
+  EIGEN_STRONG_INLINE void operator()(Self& self, Index idx1,
+                                      typename Self::CoeffReturnType* data) {
+    using Packet = typename Self::PacketReturnType;
+    const int PacketSize = internal::unpacket_traits<Packet>::size;
+    Index idx2 = 0;
+    for (; idx2 + PacketSize <= self.stride(); idx2 += PacketSize) {
+      // Calculate the starting offset for the packet scan
+      Index offset = idx1 + idx2;
+      ReducePacket(self, offset, data);
+    }
+    for (; idx2 < self.stride(); idx2++) {
+      // Calculate the starting offset for the scan
+      Index offset = idx1 + idx2;
+      ReduceScalar(self, offset, data);
+    }
+  }
+};
+
+// Single-threaded CPU implementation of scan
+template <typename Self, typename Reducer, typename Device,
+          bool Vectorize =
+              (TensorEvaluator<typename Self::ChildTypeNoConst, Device>::PacketAccess &&
+               internal::reducer_traits<Reducer, Device>::PacketAccess)>
+struct ScanLauncher {
+  void operator()(Self& self, typename Self::CoeffReturnType* data) {
+    Index total_size = internal::array_prod(self.dimensions());
+
+    // We fix the index along the scan axis to 0 and perform a
+    // scan per remaining entry. The iteration is split into two nested
+    // loops to avoid an integer division by keeping track of each idx1 and
+    // idx2.
+    for (Index idx1 = 0; idx1 < total_size; idx1 += self.stride() * self.size()) {
+      ReduceBlock<Self, Vectorize, /*Parallel=*/false> block_reducer;
+      block_reducer(self, idx1, data);
+    }
+  }
+};
+
+#ifdef EIGEN_USE_THREADS
+
+// Adjust block_size to avoid false sharing of cachelines among
+// threads. Currently set to twice the cache line size on Intel and ARM
+// processors.
+EIGEN_STRONG_INLINE Index AdjustBlockSize(Index item_size, Index block_size) {
+  EIGEN_CONSTEXPR Index kBlockAlignment = 128;
+  const Index items_per_cacheline =
+      numext::maxi<Index>(1, kBlockAlignment / item_size);
+  return items_per_cacheline * divup(block_size, items_per_cacheline);
+}
+
+template <typename Self>
+struct ReduceBlock<Self, /*Vectorize=*/true, /*Parallel=*/true> {
+  EIGEN_STRONG_INLINE void operator()(Self& self, Index idx1,
+                                      typename Self::CoeffReturnType* data) {
+    using Scalar = typename Self::CoeffReturnType;
+    using Packet = typename Self::PacketReturnType;
+    const int PacketSize = internal::unpacket_traits<Packet>::size;
+    Index num_scalars = self.stride();
+    Index num_packets = 0;
+    if (self.stride() >= PacketSize) {
+      num_packets = self.stride() / PacketSize;
+      self.device().parallelFor(
+          num_packets,
+        TensorOpCost(PacketSize * self.size(), PacketSize * self.size(),
+                     16 * PacketSize * self.size(), true, PacketSize),
+        // Make the shard size large enough that two neighboring threads
+        // won't write to the same cacheline of `data`.
+        [=](Index blk_size) {
+          return AdjustBlockSize(PacketSize * sizeof(Scalar), blk_size);
+        },
+        [&](Index first, Index last) {
+          for (Index packet = first; packet < last; ++packet) {
+            const Index idx2 = packet * PacketSize;
+            ReducePacket(self, idx1 + idx2, data);
+          }
+        });
+      num_scalars -= num_packets * PacketSize;
+    }
+    self.device().parallelFor(
+        num_scalars, TensorOpCost(self.size(), self.size(), 16 * self.size()),
+        // Make the shard size large enough that two neighboring threads
+        // won't write to the same cacheline of `data`.
+        [=](Index blk_size) {
+          return AdjustBlockSize(sizeof(Scalar), blk_size);
+        },
+        [&](Index first, Index last) {
+          for (Index scalar = first; scalar < last; ++scalar) {
+            const Index idx2 = num_packets * PacketSize + scalar;
+            ReduceScalar(self, idx1 + idx2, data);
+          }
+        });
+  }
+};
+
+template <typename Self>
+struct ReduceBlock<Self, /*Vectorize=*/false, /*Parallel=*/true> {
+  EIGEN_STRONG_INLINE void operator()(Self& self, Index idx1,
+                                      typename Self::CoeffReturnType* data) {
+    using Scalar = typename Self::CoeffReturnType;
+    self.device().parallelFor(
+        self.stride(), TensorOpCost(self.size(), self.size(), 16 * self.size()),
+        // Make the shard size large enough that two neighboring threads
+        // won't write to the same cacheline of `data`.
+        [=](Index blk_size) {
+          return AdjustBlockSize(sizeof(Scalar), blk_size);
+        },
+        [&](Index first, Index last) {
+          for (Index idx2 = first; idx2 < last; ++idx2) {
+            ReduceScalar(self, idx1 + idx2, data);
+          }
+        });
+  }
+};
+
+// Specialization for multi-threaded execution.
+template <typename Self, typename Reducer, bool Vectorize>
+struct ScanLauncher<Self, Reducer, ThreadPoolDevice, Vectorize> {
+  void operator()(Self& self, typename Self::CoeffReturnType* data) {
+    using Scalar = typename Self::CoeffReturnType;
+    using Packet = typename Self::PacketReturnType;
+    const int PacketSize = internal::unpacket_traits<Packet>::size;
+    const Index total_size = internal::array_prod(self.dimensions());
+    const Index inner_block_size = self.stride() * self.size();
+    bool parallelize_by_outer_blocks = (total_size >= (self.stride() * inner_block_size));
+
+    if ((parallelize_by_outer_blocks && total_size <= 4096) ||
+        (!parallelize_by_outer_blocks && self.stride() < PacketSize)) {
+      ScanLauncher<Self, Reducer, DefaultDevice, Vectorize> launcher;
+      launcher(self, data);
+      return;
+    }
+
+    if (parallelize_by_outer_blocks) {
+      // Parallelize over outer blocks.
+      const Index num_outer_blocks = total_size / inner_block_size;
+      self.device().parallelFor(
+          num_outer_blocks,
+          TensorOpCost(inner_block_size, inner_block_size,
+                       16 * PacketSize * inner_block_size, Vectorize,
+                       PacketSize),
+          [=](Index blk_size) {
+            return AdjustBlockSize(inner_block_size * sizeof(Scalar), blk_size);
+          },
+          [&](Index first, Index last) {
+            for (Index idx1 = first; idx1 < last; ++idx1) {
+              ReduceBlock<Self, Vectorize, /*Parallelize=*/false> block_reducer;
+              block_reducer(self, idx1 * inner_block_size, data);
+            }
+          });
+    } else {
+      // Parallelize over inner packets/scalars dimensions when the reduction
+      // axis is not an inner dimension.
+      ReduceBlock<Self, Vectorize, /*Parallelize=*/true> block_reducer;
+      for (Index idx1 = 0; idx1 < total_size;
+           idx1 += self.stride() * self.size()) {
+        block_reducer(self, idx1, data);
+      }
+    }
+  }
+};
+#endif  // EIGEN_USE_THREADS
+
+#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
+
+// GPU implementation of scan
+// TODO(ibab) This placeholder implementation performs multiple scans in
+// parallel, but it would be better to use a parallel scan algorithm and
+// optimize memory access.
+template <typename Self, typename Reducer>
+__global__ EIGEN_HIP_LAUNCH_BOUNDS_1024 void ScanKernel(Self self, Index total_size, typename Self::CoeffReturnType* data) {
+  // Compute offset as in the CPU version
+  Index val = threadIdx.x + blockIdx.x * blockDim.x;
+  Index offset = (val / self.stride()) * self.stride() * self.size() + val % self.stride();
+
+  if (offset + (self.size() - 1) * self.stride() < total_size) {
+    // Compute the scan along the axis, starting at the calculated offset
+    typename Self::CoeffReturnType accum = self.accumulator().initialize();
+    for (Index idx = 0; idx < self.size(); idx++) {
+      Index curr = offset + idx * self.stride();
+      if (self.exclusive()) {
+        data[curr] = self.accumulator().finalize(accum);
+        self.accumulator().reduce(self.inner().coeff(curr), &accum);
+      } else {
+        self.accumulator().reduce(self.inner().coeff(curr), &accum);
+        data[curr] = self.accumulator().finalize(accum);
+      }
+    }
+  }
+  __syncthreads();
+
+}
+
+template <typename Self, typename Reducer, bool Vectorize>
+struct ScanLauncher<Self, Reducer, GpuDevice, Vectorize> {
+  void operator()(const Self& self, typename Self::CoeffReturnType* data) {
+     Index total_size = internal::array_prod(self.dimensions());
+     Index num_blocks = (total_size / self.size() + 63) / 64;
+     Index block_size = 64;
+
+     LAUNCH_GPU_KERNEL((ScanKernel<Self, Reducer>), num_blocks, block_size, 0, self.device(), self, total_size, data);
+  }
+};
+#endif  // EIGEN_USE_GPU && (EIGEN_GPUCC)
+
+}  // namespace internal
+
+// Eval as rvalue
+template <typename Op, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorScanOp<Op, ArgType>, Device> {
+
+  typedef TensorScanOp<Op, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  typedef const ArgType ChildTypeNoConst;
+  typedef const ArgType ChildType;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef TensorEvaluator<const TensorScanOp<Op, ArgType>, Device> Self;
+  typedef StorageMemory<Scalar, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,
+    RawAccess = true
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device),
+        m_device(device),
+        m_exclusive(op.exclusive()),
+        m_accumulator(op.accumulator()),
+        m_size(m_impl.dimensions()[op.axis()]),
+        m_stride(1), m_consume_dim(op.axis()),
+        m_output(NULL) {
+
+    // Accumulating a scalar isn't supported.
+    EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    eigen_assert(op.axis() >= 0 && op.axis() < NumDims);
+
+    // Compute stride of scan axis
+    const Dimensions& dims = m_impl.dimensions();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < op.axis(); ++i) {
+        m_stride = m_stride * dims[i];
+      }
+    } else {
+      // dims can only be indexed through unsigned integers,
+      // so let's use an unsigned type to let the compiler knows.
+      // This prevents stupid warnings: ""'*((void*)(& evaluator)+64)[18446744073709551615]' may be used uninitialized in this function"
+      unsigned int axis = internal::convert_index<unsigned int>(op.axis());
+      for (unsigned int i = NumDims - 1; i > axis; --i) {
+        m_stride = m_stride * dims[i];
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
+    return m_impl.dimensions();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& stride() const {
+    return m_stride;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& consume_dim() const {
+    return m_consume_dim;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& size() const {
+    return m_size;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Op& accumulator() const {
+    return m_accumulator;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool exclusive() const {
+    return m_exclusive;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& inner() const {
+    return m_impl;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const {
+    return m_device;
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    internal::ScanLauncher<Self, Op, Device> launcher;
+    if (data) {
+      launcher(*this, data);
+      return false;
+    }
+
+    const Index total_size = internal::array_prod(dimensions());
+    m_output = static_cast<EvaluatorPointerType>(m_device.get((Scalar*) m_device.allocate_temp(total_size * sizeof(Scalar))));
+    launcher(*this, m_output);
+    return true;
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_output + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE EvaluatorPointerType data() const
+  {
+    return m_output;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return m_output[index];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool) const {
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0);
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    if (m_output) {
+      m_device.deallocate_temp(m_output);
+      m_output = NULL;
+    }
+    m_impl.cleanup();
+  }
+
+#ifdef EIGEN_USE_SYCL
+ // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+    m_output.bind(cgh);
+  }
+#endif
+protected:
+  TensorEvaluator<ArgType, Device> m_impl;
+  const Device EIGEN_DEVICE_REF m_device;
+  const bool m_exclusive;
+  Op m_accumulator;
+  const Index m_size;
+  Index m_stride;
+  Index m_consume_dim;
+  EvaluatorPointerType m_output;
+};
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_SCAN_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorScanSycl.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorScanSycl.h
new file mode 100644
index 0000000..7f68ecb
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorScanSycl.h
@@ -0,0 +1,513 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Mehdi Goli    Codeplay Software Ltd.
+// Ralph Potter  Codeplay Software Ltd.
+// Luke Iwanski  Codeplay Software Ltd.
+// Contact: <eigen@codeplay.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/*****************************************************************
+ * TensorScanSycl.h
+ *
+ * \brief:
+ *  Tensor Scan Sycl implement the extend  version of
+ * "Efficient parallel scan algorithms for GPUs." .for Tensor operations.
+ * The algorithm requires up to 3 stage (consequently 3 kernels) depending on
+ * the size of the tensor. In the first kernel (ScanKernelFunctor), each
+ * threads within the work-group individually reduces the allocated elements per
+ * thread in order to reduces the total number of blocks. In the next step all
+ * thread within the work-group will reduce the associated blocks into the
+ * temporary buffers. In the next kernel(ScanBlockKernelFunctor), the temporary
+ * buffer is given as an input and all the threads within a work-group scan and
+ * reduces the boundaries between the blocks (generated from the previous
+ * kernel). and write the data on the temporary buffer. If the second kernel is
+ * required, the third and final kerenl (ScanAdjustmentKernelFunctor) will
+ * adjust the final result into the output buffer.
+ * The original algorithm for the parallel prefix sum can be found here:
+ *
+ * Sengupta, Shubhabrata, Mark Harris, and Michael Garland. "Efficient parallel
+ * scan algorithms for GPUs." NVIDIA, Santa Clara, CA, Tech. Rep. NVR-2008-003
+ *1, no. 1 (2008): 1-17.
+ *****************************************************************/
+
+#ifndef UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_SYCL_SYCL_HPP
+#define UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_SYCL_SYCL_HPP
+
+namespace Eigen {
+namespace TensorSycl {
+namespace internal {
+
+#ifndef EIGEN_SYCL_MAX_GLOBAL_RANGE
+#define EIGEN_SYCL_MAX_GLOBAL_RANGE (EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1 * 4)
+#endif
+
+template <typename index_t>
+struct ScanParameters {
+  // must be power of 2
+  static EIGEN_CONSTEXPR index_t ScanPerThread = 8;
+  const index_t total_size;
+  const index_t non_scan_size;
+  const index_t scan_size;
+  const index_t non_scan_stride;
+  const index_t scan_stride;
+  const index_t panel_threads;
+  const index_t group_threads;
+  const index_t block_threads;
+  const index_t elements_per_group;
+  const index_t elements_per_block;
+  const index_t loop_range;
+
+  ScanParameters(index_t total_size_, index_t non_scan_size_, index_t scan_size_, index_t non_scan_stride_,
+                 index_t scan_stride_, index_t panel_threads_, index_t group_threads_, index_t block_threads_,
+                 index_t elements_per_group_, index_t elements_per_block_, index_t loop_range_)
+      : total_size(total_size_),
+        non_scan_size(non_scan_size_),
+        scan_size(scan_size_),
+        non_scan_stride(non_scan_stride_),
+        scan_stride(scan_stride_),
+        panel_threads(panel_threads_),
+        group_threads(group_threads_),
+        block_threads(block_threads_),
+        elements_per_group(elements_per_group_),
+        elements_per_block(elements_per_block_),
+        loop_range(loop_range_) {}
+};
+
+enum class scan_step { first, second };
+template <typename Evaluator, typename CoeffReturnType, typename OutAccessor, typename Op, typename Index,
+          scan_step stp>
+struct ScanKernelFunctor {
+  typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      LocalAccessor;
+  static EIGEN_CONSTEXPR int PacketSize = ScanParameters<Index>::ScanPerThread / 2;
+
+  LocalAccessor scratch;
+  Evaluator dev_eval;
+  OutAccessor out_accessor;
+  OutAccessor temp_accessor;
+  const ScanParameters<Index> scanParameters;
+  Op accumulator;
+  const bool inclusive;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ScanKernelFunctor(LocalAccessor scratch_, const Evaluator dev_eval_,
+                                                          OutAccessor out_accessor_, OutAccessor temp_accessor_,
+                                                          const ScanParameters<Index> scanParameters_, Op accumulator_,
+                                                          const bool inclusive_)
+      : scratch(scratch_),
+        dev_eval(dev_eval_),
+        out_accessor(out_accessor_),
+        temp_accessor(temp_accessor_),
+        scanParameters(scanParameters_),
+        accumulator(accumulator_),
+        inclusive(inclusive_) {}
+
+  template <scan_step sst = stp, typename Input>
+  typename ::Eigen::internal::enable_if<sst == scan_step::first, CoeffReturnType>::type EIGEN_DEVICE_FUNC
+      EIGEN_STRONG_INLINE
+      read(const Input &inpt, Index global_id) {
+    return inpt.coeff(global_id);
+  }
+
+  template <scan_step sst = stp, typename Input>
+  typename ::Eigen::internal::enable_if<sst != scan_step::first, CoeffReturnType>::type EIGEN_DEVICE_FUNC
+      EIGEN_STRONG_INLINE
+      read(const Input &inpt, Index global_id) {
+    return inpt[global_id];
+  }
+
+  template <scan_step sst = stp, typename InclusiveOp>
+  typename ::Eigen::internal::enable_if<sst == scan_step::first>::type EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  first_step_inclusive_Operation(InclusiveOp inclusive_op) {
+    inclusive_op();
+  }
+
+  template <scan_step sst = stp, typename InclusiveOp>
+  typename ::Eigen::internal::enable_if<sst != scan_step::first>::type EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  first_step_inclusive_Operation(InclusiveOp) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(cl::sycl::nd_item<1> itemID) {
+    auto out_ptr = out_accessor.get_pointer();
+    auto tmp_ptr = temp_accessor.get_pointer();
+    auto scratch_ptr = scratch.get_pointer().get();
+
+    for (Index loop_offset = 0; loop_offset < scanParameters.loop_range; loop_offset++) {
+      Index data_offset = (itemID.get_global_id(0) + (itemID.get_global_range(0) * loop_offset));
+      Index tmp = data_offset % scanParameters.panel_threads;
+      const Index panel_id = data_offset / scanParameters.panel_threads;
+      const Index group_id = tmp / scanParameters.group_threads;
+      tmp = tmp % scanParameters.group_threads;
+      const Index block_id = tmp / scanParameters.block_threads;
+      const Index local_id = tmp % scanParameters.block_threads;
+      // we put one element per packet in scratch_mem
+      const Index scratch_stride = scanParameters.elements_per_block / PacketSize;
+      const Index scratch_offset = (itemID.get_local_id(0) / scanParameters.block_threads) * scratch_stride;
+      CoeffReturnType private_scan[ScanParameters<Index>::ScanPerThread];
+      CoeffReturnType inclusive_scan;
+      // the actual panel size is scan_size * non_scan_size.
+      // elements_per_panel is roundup to power of 2 for binary tree
+      const Index panel_offset = panel_id * scanParameters.scan_size * scanParameters.non_scan_size;
+      const Index group_offset = group_id * scanParameters.non_scan_stride;
+      // This will be effective when the size is bigger than elements_per_block
+      const Index block_offset = block_id * scanParameters.elements_per_block * scanParameters.scan_stride;
+      const Index thread_offset = (ScanParameters<Index>::ScanPerThread * local_id * scanParameters.scan_stride);
+      const Index global_offset = panel_offset + group_offset + block_offset + thread_offset;
+      Index next_elements = 0;
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < ScanParameters<Index>::ScanPerThread; i++) {
+        Index global_id = global_offset + next_elements;
+        private_scan[i] = ((((block_id * scanParameters.elements_per_block) +
+                             (ScanParameters<Index>::ScanPerThread * local_id) + i) < scanParameters.scan_size) &&
+                           (global_id < scanParameters.total_size))
+                              ? read(dev_eval, global_id)
+                              : accumulator.initialize();
+        next_elements += scanParameters.scan_stride;
+      }
+      first_step_inclusive_Operation([&]() EIGEN_DEVICE_FUNC {
+        if (inclusive) {
+          inclusive_scan = private_scan[ScanParameters<Index>::ScanPerThread - 1];
+        }
+      });
+      // This for loop must be 2
+      EIGEN_UNROLL_LOOP
+      for (int packetIndex = 0; packetIndex < ScanParameters<Index>::ScanPerThread; packetIndex += PacketSize) {
+        Index private_offset = 1;
+        // build sum in place up the tree
+        EIGEN_UNROLL_LOOP
+        for (Index d = PacketSize >> 1; d > 0; d >>= 1) {
+          EIGEN_UNROLL_LOOP
+          for (Index l = 0; l < d; l++) {
+            Index ai = private_offset * (2 * l + 1) - 1 + packetIndex;
+            Index bi = private_offset * (2 * l + 2) - 1 + packetIndex;
+            CoeffReturnType accum = accumulator.initialize();
+            accumulator.reduce(private_scan[ai], &accum);
+            accumulator.reduce(private_scan[bi], &accum);
+            private_scan[bi] = accumulator.finalize(accum);
+          }
+          private_offset *= 2;
+        }
+        scratch_ptr[2 * local_id + (packetIndex / PacketSize) + scratch_offset] =
+            private_scan[PacketSize - 1 + packetIndex];
+        private_scan[PacketSize - 1 + packetIndex] = accumulator.initialize();
+        // traverse down tree & build scan
+        EIGEN_UNROLL_LOOP
+        for (Index d = 1; d < PacketSize; d *= 2) {
+          private_offset >>= 1;
+          EIGEN_UNROLL_LOOP
+          for (Index l = 0; l < d; l++) {
+            Index ai = private_offset * (2 * l + 1) - 1 + packetIndex;
+            Index bi = private_offset * (2 * l + 2) - 1 + packetIndex;
+            CoeffReturnType accum = accumulator.initialize();
+            accumulator.reduce(private_scan[ai], &accum);
+            accumulator.reduce(private_scan[bi], &accum);
+            private_scan[ai] = private_scan[bi];
+            private_scan[bi] = accumulator.finalize(accum);
+          }
+        }
+      }
+
+      Index offset = 1;
+      // build sum in place up the tree
+      for (Index d = scratch_stride >> 1; d > 0; d >>= 1) {
+        // Synchronise
+        itemID.barrier(cl::sycl::access::fence_space::local_space);
+        if (local_id < d) {
+          Index ai = offset * (2 * local_id + 1) - 1 + scratch_offset;
+          Index bi = offset * (2 * local_id + 2) - 1 + scratch_offset;
+          CoeffReturnType accum = accumulator.initialize();
+          accumulator.reduce(scratch_ptr[ai], &accum);
+          accumulator.reduce(scratch_ptr[bi], &accum);
+          scratch_ptr[bi] = accumulator.finalize(accum);
+        }
+        offset *= 2;
+      }
+      // Synchronise
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+      // next step optimisation
+      if (local_id == 0) {
+        if (((scanParameters.elements_per_group / scanParameters.elements_per_block) > 1)) {
+          const Index temp_id = panel_id * (scanParameters.elements_per_group / scanParameters.elements_per_block) *
+                                    scanParameters.non_scan_size +
+                                group_id * (scanParameters.elements_per_group / scanParameters.elements_per_block) +
+                                block_id;
+          tmp_ptr[temp_id] = scratch_ptr[scratch_stride - 1 + scratch_offset];
+        }
+        // clear the last element
+        scratch_ptr[scratch_stride - 1 + scratch_offset] = accumulator.initialize();
+      }
+      // traverse down tree & build scan
+      for (Index d = 1; d < scratch_stride; d *= 2) {
+        offset >>= 1;
+        // Synchronise
+        itemID.barrier(cl::sycl::access::fence_space::local_space);
+        if (local_id < d) {
+          Index ai = offset * (2 * local_id + 1) - 1 + scratch_offset;
+          Index bi = offset * (2 * local_id + 2) - 1 + scratch_offset;
+          CoeffReturnType accum = accumulator.initialize();
+          accumulator.reduce(scratch_ptr[ai], &accum);
+          accumulator.reduce(scratch_ptr[bi], &accum);
+          scratch_ptr[ai] = scratch_ptr[bi];
+          scratch_ptr[bi] = accumulator.finalize(accum);
+        }
+      }
+      // Synchronise
+      itemID.barrier(cl::sycl::access::fence_space::local_space);
+      // This for loop must be 2
+      EIGEN_UNROLL_LOOP
+      for (int packetIndex = 0; packetIndex < ScanParameters<Index>::ScanPerThread; packetIndex += PacketSize) {
+        EIGEN_UNROLL_LOOP
+        for (Index i = 0; i < PacketSize; i++) {
+          CoeffReturnType accum = private_scan[packetIndex + i];
+          accumulator.reduce(scratch_ptr[2 * local_id + (packetIndex / PacketSize) + scratch_offset], &accum);
+          private_scan[packetIndex + i] = accumulator.finalize(accum);
+        }
+      }
+      first_step_inclusive_Operation([&]() EIGEN_DEVICE_FUNC {
+        if (inclusive) {
+          accumulator.reduce(private_scan[ScanParameters<Index>::ScanPerThread - 1], &inclusive_scan);
+          private_scan[0] = accumulator.finalize(inclusive_scan);
+        }
+      });
+      next_elements = 0;
+      // right the first set of private param
+      EIGEN_UNROLL_LOOP
+      for (Index i = 0; i < ScanParameters<Index>::ScanPerThread; i++) {
+        Index global_id = global_offset + next_elements;
+        if ((((block_id * scanParameters.elements_per_block) + (ScanParameters<Index>::ScanPerThread * local_id) + i) <
+             scanParameters.scan_size) &&
+            (global_id < scanParameters.total_size)) {
+          Index private_id = (i * !inclusive) + (((i + 1) % ScanParameters<Index>::ScanPerThread) * (inclusive));
+          out_ptr[global_id] = private_scan[private_id];
+        }
+        next_elements += scanParameters.scan_stride;
+      }
+    }  // end for loop
+  }
+};
+
+template <typename CoeffReturnType, typename InAccessor, typename OutAccessor, typename Op, typename Index>
+struct ScanAdjustmentKernelFunctor {
+  typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
+      LocalAccessor;
+  static EIGEN_CONSTEXPR int PacketSize = ScanParameters<Index>::ScanPerThread / 2;
+  InAccessor in_accessor;
+  OutAccessor out_accessor;
+  const ScanParameters<Index> scanParameters;
+  Op accumulator;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ScanAdjustmentKernelFunctor(LocalAccessor, InAccessor in_accessor_,
+                                                                    OutAccessor out_accessor_,
+                                                                    const ScanParameters<Index> scanParameters_,
+                                                                    Op accumulator_)
+      : in_accessor(in_accessor_),
+        out_accessor(out_accessor_),
+        scanParameters(scanParameters_),
+        accumulator(accumulator_) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(cl::sycl::nd_item<1> itemID) {
+    auto in_ptr = in_accessor.get_pointer();
+    auto out_ptr = out_accessor.get_pointer();
+
+    for (Index loop_offset = 0; loop_offset < scanParameters.loop_range; loop_offset++) {
+      Index data_offset = (itemID.get_global_id(0) + (itemID.get_global_range(0) * loop_offset));
+      Index tmp = data_offset % scanParameters.panel_threads;
+      const Index panel_id = data_offset / scanParameters.panel_threads;
+      const Index group_id = tmp / scanParameters.group_threads;
+      tmp = tmp % scanParameters.group_threads;
+      const Index block_id = tmp / scanParameters.block_threads;
+      const Index local_id = tmp % scanParameters.block_threads;
+
+      // the actual panel size is scan_size * non_scan_size.
+      // elements_per_panel is roundup to power of 2 for binary tree
+      const Index panel_offset = panel_id * scanParameters.scan_size * scanParameters.non_scan_size;
+      const Index group_offset = group_id * scanParameters.non_scan_stride;
+      // This will be effective when the size is bigger than elements_per_block
+      const Index block_offset = block_id * scanParameters.elements_per_block * scanParameters.scan_stride;
+      const Index thread_offset = ScanParameters<Index>::ScanPerThread * local_id * scanParameters.scan_stride;
+
+      const Index global_offset = panel_offset + group_offset + block_offset + thread_offset;
+      const Index block_size = scanParameters.elements_per_group / scanParameters.elements_per_block;
+      const Index in_id = (panel_id * block_size * scanParameters.non_scan_size) + (group_id * block_size) + block_id;
+      CoeffReturnType adjust_val = in_ptr[in_id];
+
+      Index next_elements = 0;
+      EIGEN_UNROLL_LOOP
+      for (Index i = 0; i < ScanParameters<Index>::ScanPerThread; i++) {
+        Index global_id = global_offset + next_elements;
+        if ((((block_id * scanParameters.elements_per_block) + (ScanParameters<Index>::ScanPerThread * local_id) + i) <
+             scanParameters.scan_size) &&
+            (global_id < scanParameters.total_size)) {
+          CoeffReturnType accum = adjust_val;
+          accumulator.reduce(out_ptr[global_id], &accum);
+          out_ptr[global_id] = accumulator.finalize(accum);
+        }
+        next_elements += scanParameters.scan_stride;
+      }
+    }
+  }
+};
+
+template <typename Index>
+struct ScanInfo {
+  const Index &total_size;
+  const Index &scan_size;
+  const Index &panel_size;
+  const Index &non_scan_size;
+  const Index &scan_stride;
+  const Index &non_scan_stride;
+
+  Index max_elements_per_block;
+  Index block_size;
+  Index panel_threads;
+  Index group_threads;
+  Index block_threads;
+  Index elements_per_group;
+  Index elements_per_block;
+  Index loop_range;
+  Index global_range;
+  Index local_range;
+  const Eigen::SyclDevice &dev;
+  EIGEN_STRONG_INLINE ScanInfo(const Index &total_size_, const Index &scan_size_, const Index &panel_size_,
+                               const Index &non_scan_size_, const Index &scan_stride_, const Index &non_scan_stride_,
+                               const Eigen::SyclDevice &dev_)
+      : total_size(total_size_),
+        scan_size(scan_size_),
+        panel_size(panel_size_),
+        non_scan_size(non_scan_size_),
+        scan_stride(scan_stride_),
+        non_scan_stride(non_scan_stride_),
+        dev(dev_) {
+    // must be power of 2
+    local_range = std::min(Index(dev.getNearestPowerOfTwoWorkGroupSize()),
+                           Index(EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1));
+
+    max_elements_per_block = local_range * ScanParameters<Index>::ScanPerThread;
+
+    elements_per_group =
+        dev.getPowerOfTwo(Index(roundUp(Index(scan_size), ScanParameters<Index>::ScanPerThread)), true);
+    const Index elements_per_panel = elements_per_group * non_scan_size;
+    elements_per_block = std::min(Index(elements_per_group), Index(max_elements_per_block));
+    panel_threads = elements_per_panel / ScanParameters<Index>::ScanPerThread;
+    group_threads = elements_per_group / ScanParameters<Index>::ScanPerThread;
+    block_threads = elements_per_block / ScanParameters<Index>::ScanPerThread;
+    block_size = elements_per_group / elements_per_block;
+#ifdef EIGEN_SYCL_MAX_GLOBAL_RANGE
+    const Index max_threads = std::min(Index(panel_threads * panel_size), Index(EIGEN_SYCL_MAX_GLOBAL_RANGE));
+#else
+    const Index max_threads = panel_threads * panel_size;
+#endif
+    global_range = roundUp(max_threads, local_range);
+    loop_range = Index(
+        std::ceil(double(elements_per_panel * panel_size) / (global_range * ScanParameters<Index>::ScanPerThread)));
+  }
+  inline ScanParameters<Index> get_scan_parameter() {
+    return ScanParameters<Index>(total_size, non_scan_size, scan_size, non_scan_stride, scan_stride, panel_threads,
+                                 group_threads, block_threads, elements_per_group, elements_per_block, loop_range);
+  }
+  inline cl::sycl::nd_range<1> get_thread_range() {
+    return cl::sycl::nd_range<1>(cl::sycl::range<1>(global_range), cl::sycl::range<1>(local_range));
+  }
+};
+
+template <typename EvaluatorPointerType, typename CoeffReturnType, typename Reducer, typename Index>
+struct SYCLAdjustBlockOffset {
+  EIGEN_STRONG_INLINE static void adjust_scan_block_offset(EvaluatorPointerType in_ptr, EvaluatorPointerType out_ptr,
+                                                           Reducer &accumulator, const Index total_size,
+                                                           const Index scan_size, const Index panel_size,
+                                                           const Index non_scan_size, const Index scan_stride,
+                                                           const Index non_scan_stride, const Eigen::SyclDevice &dev) {
+    auto scan_info =
+        ScanInfo<Index>(total_size, scan_size, panel_size, non_scan_size, scan_stride, non_scan_stride, dev);
+
+    typedef ScanAdjustmentKernelFunctor<CoeffReturnType, EvaluatorPointerType, EvaluatorPointerType, Reducer, Index>
+        AdjustFuctor;
+    dev.template unary_kernel_launcher<CoeffReturnType, AdjustFuctor>(in_ptr, out_ptr, scan_info.get_thread_range(),
+                                                                      scan_info.max_elements_per_block,
+                                                                      scan_info.get_scan_parameter(), accumulator);
+  }
+};
+
+template <typename CoeffReturnType, scan_step stp>
+struct ScanLauncher_impl {
+  template <typename Input, typename EvaluatorPointerType, typename Reducer, typename Index>
+  EIGEN_STRONG_INLINE static void scan_block(Input in_ptr, EvaluatorPointerType out_ptr, Reducer &accumulator,
+                                             const Index total_size, const Index scan_size, const Index panel_size,
+                                             const Index non_scan_size, const Index scan_stride,
+                                             const Index non_scan_stride, const bool inclusive,
+                                             const Eigen::SyclDevice &dev) {
+    auto scan_info =
+        ScanInfo<Index>(total_size, scan_size, panel_size, non_scan_size, scan_stride, non_scan_stride, dev);
+    const Index temp_pointer_size = scan_info.block_size * non_scan_size * panel_size;
+    const Index scratch_size = scan_info.max_elements_per_block / (ScanParameters<Index>::ScanPerThread / 2);
+    CoeffReturnType *temp_pointer =
+        static_cast<CoeffReturnType *>(dev.allocate_temp(temp_pointer_size * sizeof(CoeffReturnType)));
+    EvaluatorPointerType tmp_global_accessor = dev.get(temp_pointer);
+
+    typedef ScanKernelFunctor<Input, CoeffReturnType, EvaluatorPointerType, Reducer, Index, stp> ScanFunctor;
+    dev.template binary_kernel_launcher<CoeffReturnType, ScanFunctor>(
+        in_ptr, out_ptr, tmp_global_accessor, scan_info.get_thread_range(), scratch_size,
+        scan_info.get_scan_parameter(), accumulator, inclusive);
+
+    if (scan_info.block_size > 1) {
+      ScanLauncher_impl<CoeffReturnType, scan_step::second>::scan_block(
+          tmp_global_accessor, tmp_global_accessor, accumulator, temp_pointer_size, scan_info.block_size, panel_size,
+          non_scan_size, Index(1), scan_info.block_size, false, dev);
+
+      SYCLAdjustBlockOffset<EvaluatorPointerType, CoeffReturnType, Reducer, Index>::adjust_scan_block_offset(
+          tmp_global_accessor, out_ptr, accumulator, total_size, scan_size, panel_size, non_scan_size, scan_stride,
+          non_scan_stride, dev);
+    }
+    dev.deallocate_temp(temp_pointer);
+  }
+};
+
+}  // namespace internal
+}  // namespace TensorSycl
+namespace internal {
+template <typename Self, typename Reducer, bool vectorize>
+struct ScanLauncher<Self, Reducer, Eigen::SyclDevice, vectorize> {
+  typedef typename Self::Index Index;
+  typedef typename Self::CoeffReturnType CoeffReturnType;
+  typedef typename Self::Storage Storage;
+  typedef typename Self::EvaluatorPointerType EvaluatorPointerType;
+  void operator()(Self &self, EvaluatorPointerType data) {
+    const Index total_size = internal::array_prod(self.dimensions());
+    const Index scan_size = self.size();
+    const Index scan_stride = self.stride();
+    // this is the scan op (can be sum or ...)
+    auto accumulator = self.accumulator();
+    auto inclusive = !self.exclusive();
+    auto consume_dim = self.consume_dim();
+    auto dev = self.device();
+
+    auto dims = self.inner().dimensions();
+
+    Index non_scan_size = 1;
+    Index panel_size = 1;
+    if (static_cast<int>(Self::Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < consume_dim; i++) {
+        non_scan_size *= dims[i];
+      }
+      for (int i = consume_dim + 1; i < Self::NumDims; i++) {
+        panel_size *= dims[i];
+      }
+    } else {
+      for (int i = Self::NumDims - 1; i > consume_dim; i--) {
+        non_scan_size *= dims[i];
+      }
+      for (int i = consume_dim - 1; i >= 0; i--) {
+        panel_size *= dims[i];
+      }
+    }
+    const Index non_scan_stride = (scan_stride > 1) ? 1 : scan_size;
+    auto eval_impl = self.inner();
+    TensorSycl::internal::ScanLauncher_impl<CoeffReturnType, TensorSycl::internal::scan_step::first>::scan_block(
+        eval_impl, data, accumulator, total_size, scan_size, panel_size, non_scan_size, scan_stride, non_scan_stride,
+        inclusive, dev);
+  }
+};
+} // namespace internal
+}  // namespace Eigen
+
+#endif  // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_SYCL_SYCL_HPP
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorShuffling.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorShuffling.h
new file mode 100644
index 0000000..e5e5efd
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorShuffling.h
@@ -0,0 +1,471 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H
+#define EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H
+
+namespace Eigen {
+
+/** \class TensorShuffling
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor shuffling class.
+  *
+  *
+  */
+namespace internal {
+template<typename Shuffle, typename XprType>
+struct traits<TensorShufflingOp<Shuffle, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename Shuffle, typename XprType>
+struct eval<TensorShufflingOp<Shuffle, XprType>, Eigen::Dense>
+{
+  typedef const TensorShufflingOp<Shuffle, XprType>& type;
+};
+
+template<typename Shuffle, typename XprType>
+struct nested<TensorShufflingOp<Shuffle, XprType>, 1, typename eval<TensorShufflingOp<Shuffle, XprType> >::type>
+{
+  typedef TensorShufflingOp<Shuffle, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename Shuffle, typename XprType>
+class TensorShufflingOp : public TensorBase<TensorShufflingOp<Shuffle, XprType> >
+{
+  public:
+    typedef TensorBase<TensorShufflingOp<Shuffle, XprType> > Base;
+    typedef typename Eigen::internal::traits<TensorShufflingOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename XprType::CoeffReturnType CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorShufflingOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorShufflingOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorShufflingOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp(const XprType& expr, const Shuffle& shfl)
+      : m_xpr(expr), m_shuffle(shfl) {}
+
+    EIGEN_DEVICE_FUNC
+    const Shuffle& shufflePermutation() const { return m_shuffle; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorShufflingOp)
+
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Shuffle m_shuffle;
+};
+
+
+// Eval as rvalue
+template<typename Shuffle, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
+{
+  typedef TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device> Self;
+  typedef TensorShufflingOp<Shuffle, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned         = false,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = TensorEvaluator<ArgType, Device>::RawAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess       = false,  // to be implemented
+    RawAccess         = false
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+
+  typedef typename internal::TensorMaterializedBlock<ScalarNoConst, NumDims,
+                                                     Layout, Index>
+      TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_device(device),
+        m_impl(op.expression(), device)
+  {
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    const Shuffle& shuffle = op.shufflePermutation();
+    m_is_identity = true;
+    for (int i = 0; i < NumDims; ++i) {
+      m_shuffle[i] = static_cast<int>(shuffle[i]);
+      m_dimensions[i] = input_dims[shuffle[i]];
+      m_inverseShuffle[shuffle[i]] = i;
+      if (m_is_identity && shuffle[i] != i) {
+        m_is_identity = false;
+      }
+    }
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_unshuffledInputStrides[0] = 1;
+      m_outputStrides[0] = 1;
+
+      for (int i = 1; i < NumDims; ++i) {
+        m_unshuffledInputStrides[i] =
+            m_unshuffledInputStrides[i - 1] * input_dims[i - 1];
+        m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(
+                  m_outputStrides[i] > 0 ? m_outputStrides[i] : Index(1));
+      }
+    } else {
+      m_unshuffledInputStrides[NumDims - 1] = 1;
+      m_outputStrides[NumDims - 1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_unshuffledInputStrides[i] =
+            m_unshuffledInputStrides[i + 1] * input_dims[i + 1];
+        m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(
+                  m_outputStrides[i] > 0 ? m_outputStrides[i] : Index(1));
+      }
+    }
+
+    for (int i = 0; i < NumDims; ++i) {
+      m_inputStrides[i] = m_unshuffledInputStrides[shuffle[i]];
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType, EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
+  }
+#endif  // EIGEN_USE_THREADS
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    if (m_is_identity) {
+      return m_impl.coeff(index);
+    } else {
+      return m_impl.coeff(srcCoeff(index));
+    }
+  }
+
+  template <int LoadMode, typename Self, bool ImplPacketAccess>
+  struct PacketLoader {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    static PacketReturnType Run(const Self& self, Index index) {
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < PacketSize; ++i) {
+        values[i] = self.coeff(index + i);
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    }
+  };
+
+  template<int LoadMode, typename Self>
+  struct PacketLoader<LoadMode, Self, true> {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    static PacketReturnType Run(const Self& self, Index index) {
+      if (self.m_is_identity) {
+        return self.m_impl.template packet<LoadMode>(index);
+      } else {
+        EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+        EIGEN_UNROLL_LOOP
+        for (int i = 0; i < PacketSize; ++i) {
+          values[i] = self.coeff(index + i);
+        }
+        PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+        return rslt;
+      }
+    }
+  };
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+        eigen_assert(index + PacketSize - 1 < dimensions().TotalSize());
+    return PacketLoader<LoadMode, Self, TensorEvaluator<ArgType, Device>::PacketAccess>::Run(*this, index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  internal::TensorBlockResourceRequirements getResourceRequirements() const {
+    static const int inner_dim =
+        Layout == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
+
+    const size_t target_size = m_device.firstLevelCacheSize();
+    const bool inner_dim_shuffled = m_shuffle[inner_dim] != inner_dim;
+
+    // Shuffled inner dimensions leads to a random memory access, which is not
+    // captured by default cost model bytes loaded/stored. We add this cost
+    // explicitly. The number of cycles picked based on the benchmarks.
+    // TODO(ezhulenev): This number was picked based on a very questionable
+    // benchmarks, add benchmarks that are representative of real workloads.
+    using BlockRequirements = internal::TensorBlockResourceRequirements;
+    if (inner_dim_shuffled) {
+      return BlockRequirements::uniform<Scalar>(target_size)
+          .addCostPerCoeff({0, 0, NumDims * 28});
+    } else {
+      return BlockRequirements::skewed<Scalar>(target_size);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
+  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
+          bool root_of_expr_ast = false) const {
+    assert(m_impl.data() != NULL);
+
+    typedef internal::TensorBlockIO<ScalarNoConst, Index, NumDims, Layout>
+        TensorBlockIO;
+    typedef typename TensorBlockIO::Dst TensorBlockIODst;
+    typedef typename TensorBlockIO::Src TensorBlockIOSrc;
+
+    const typename TensorBlock::Storage block_storage =
+        TensorBlock::prepareStorage(
+            desc, scratch, /*allow_strided_storage=*/root_of_expr_ast);
+
+    typename TensorBlockIO::Dimensions input_strides(m_unshuffledInputStrides);
+    TensorBlockIOSrc src(input_strides, m_impl.data(), srcCoeff(desc.offset()));
+
+    TensorBlockIODst dst(block_storage.dimensions(), block_storage.strides(),
+                         block_storage.data());
+
+    typename TensorBlockIO::DimensionsMap dst_to_src_dim_map(m_shuffle);
+    TensorBlockIO::Copy(dst, src, dst_to_src_dim_map);
+
+    return block_storage.AsTensorMaterializedBlock();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    const double compute_cost = m_is_identity ? TensorOpCost::AddCost<Index>() :
+                                NumDims * (2 * TensorOpCost::AddCost<Index>() +
+                                           2 * TensorOpCost::MulCost<Index>() +
+                                           TensorOpCost::DivCost<Index>());
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 0, compute_cost, m_is_identity /* vectorized */, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC typename Storage::Type data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+   // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index GetBlockOutputIndex(
+      Index input_index,
+      const DSizes<Index, NumDims>& input_block_strides,
+      const DSizes<Index, NumDims>& output_block_strides,
+      const DSizes<internal::TensorIntDivisor<Index>, NumDims>& fast_input_block_strides) const {
+    Index output_index = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = input_index / fast_input_block_strides[i];
+        output_index += idx * output_block_strides[m_inverseShuffle[i]];
+        input_index -= idx * input_block_strides[i];
+      }
+      return output_index + input_index *
+          output_block_strides[m_inverseShuffle[0]];
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = input_index / fast_input_block_strides[i];
+        output_index += idx * output_block_strides[m_inverseShuffle[i]];
+        input_index -= idx * input_block_strides[i];
+      }
+      return output_index + input_index *
+          output_block_strides[m_inverseShuffle[NumDims - 1]];
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const {
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        inputIndex += idx * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      return inputIndex + index * m_inputStrides[0];
+    } else {
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        inputIndex += idx * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      return inputIndex + index * m_inputStrides[NumDims - 1];
+    }
+  }
+
+  Dimensions m_dimensions;
+  bool m_is_identity;
+  array<int, NumDims> m_shuffle;
+  array<Index, NumDims> m_inverseShuffle;  // TODO(ezhulenev): Make it int type.
+  array<Index, NumDims> m_outputStrides;
+  array<internal::TensorIntDivisor<Index>, NumDims> m_fastOutputStrides;
+  array<Index, NumDims> m_inputStrides;
+  array<Index, NumDims> m_unshuffledInputStrides;
+
+  const Device EIGEN_DEVICE_REF m_device;
+  TensorEvaluator<ArgType, Device> m_impl;
+};
+
+
+// Eval as lvalue
+template<typename Shuffle, typename ArgType, typename Device>
+struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
+    : public TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
+{
+  typedef TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device> Base;
+
+  typedef TensorShufflingOp<Shuffle, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+
+  enum {
+    IsAligned         = false,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = TensorEvaluator<ArgType, Device>::RawAccess,
+    PreferBlockAccess = true,
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess         = false
+  };
+
+  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : Base(op, device)
+  { }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
+  {
+    return this->m_impl.coeffRef(this->srcCoeff(index));
+  }
+
+  template <int StoreMode> EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x)
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      this->coeffRef(index+i) = values[i];
+    }
+  }
+
+  template <typename TensorBlock>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
+      const TensorBlockDesc& desc, const TensorBlock& block) {
+    eigen_assert(this->m_impl.data() != NULL);
+
+    typedef internal::TensorBlockIO<ScalarNoConst, Index, NumDims, Layout>
+        TensorBlockIO;
+    typedef typename TensorBlockIO::Dst TensorBlockIODst;
+    typedef typename TensorBlockIO::Src TensorBlockIOSrc;
+
+    const Scalar* block_buffer = block.data();
+
+    // TODO(ezhulenev): TensorBlockIO should be able to read from any Eigen
+    // expression with coefficient and packet access as `src`.
+    void* mem = NULL;
+    if (block_buffer == NULL) {
+      mem = this->m_device.allocate(desc.size() * sizeof(Scalar));
+      ScalarNoConst* buf = static_cast<ScalarNoConst*>(mem);
+
+      typedef internal::TensorBlockAssignment<
+          ScalarNoConst, NumDims, typename TensorBlock::XprType, Index>
+          TensorBlockAssignment;
+
+      TensorBlockAssignment::Run(
+          TensorBlockAssignment::target(
+              desc.dimensions(), internal::strides<Layout>(desc.dimensions()),
+              buf),
+          block.expr());
+
+      block_buffer = buf;
+    }
+
+    // Read from block.
+    TensorBlockIOSrc src(internal::strides<Layout>(desc.dimensions()),
+                         block_buffer);
+
+    // Write to the output buffer.
+    typename TensorBlockIO::Dimensions output_strides(
+        this->m_unshuffledInputStrides);
+    typename TensorBlockIO::Dimensions output_dimensions;
+    for (int i = 0; i < NumDims; ++i) {
+      output_dimensions[this->m_shuffle[i]] = desc.dimension(i);
+    }
+    TensorBlockIODst dst(output_dimensions, output_strides, this->m_impl.data(),
+                         this->srcCoeff(desc.offset()));
+
+    // Reorder dimensions according to the shuffle.
+    typename TensorBlockIO::DimensionsMap dst_to_src_dim_map;
+    for (int i = 0; i < NumDims; ++i) {
+      dst_to_src_dim_map[i] = static_cast<int>(this->m_inverseShuffle[i]);
+    }
+    TensorBlockIO::Copy(dst, src, dst_to_src_dim_map);
+
+    // Deallocate temporary buffer used for the block materialization.
+    if (mem != NULL) this->m_device.deallocate(mem);
+  }
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorStorage.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorStorage.h
new file mode 100644
index 0000000..5ff0880
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorStorage.h
@@ -0,0 +1,161 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+// Copyright (C) 2014-2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSORSTORAGE_H
+#define EIGEN_CXX11_TENSOR_TENSORSTORAGE_H
+
+#ifdef EIGEN_TENSOR_STORAGE_CTOR_PLUGIN
+  #define EIGEN_INTERNAL_TENSOR_STORAGE_CTOR_PLUGIN EIGEN_TENSOR_STORAGE_CTOR_PLUGIN;
+#else
+  #define EIGEN_INTERNAL_TENSOR_STORAGE_CTOR_PLUGIN
+#endif
+
+namespace Eigen {
+
+/** \internal
+  *
+  * \class TensorStorage
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Stores the data of a tensor
+  *
+  * This class stores the data of fixed-size, dynamic-size or mixed tensors
+  * in a way as compact as possible.
+  *
+  * \sa Tensor
+  */
+template<typename T, typename Dimensions, int Options> class TensorStorage;
+
+
+// Pure fixed-size storage
+template<typename T, typename FixedDimensions, int Options_>
+class TensorStorage
+{
+ private:
+  static const std::size_t Size = FixedDimensions::total_size;
+
+  // Allocate an array of size at least one to prevent compiler warnings.
+  static const std::size_t MinSize = max_n_1<Size>::size;
+  EIGEN_ALIGN_MAX T m_data[MinSize];
+
+ public:
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE TensorStorage() {
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T *data() { return m_data; }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T *data() const { return m_data; }
+
+  static EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const FixedDimensions& dimensions()
+  {
+    static const FixedDimensions* singleton_dimensions = new FixedDimensions();
+    return *singleton_dimensions;
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE DenseIndex size() const { return Size; }
+};
+
+// pure dynamic
+template<typename T, typename IndexType, int NumIndices_, int Options_>
+class TensorStorage<T, DSizes<IndexType, NumIndices_>, Options_>
+{
+  public:
+    typedef IndexType Index;
+    typedef DSizes<IndexType, NumIndices_> Dimensions;
+    typedef TensorStorage<T, DSizes<IndexType, NumIndices_>, Options_> Self;
+
+    EIGEN_DEVICE_FUNC TensorStorage() : m_data(0), m_dimensions() {
+      if (NumIndices_ == 0) {
+	m_data = internal::conditional_aligned_new_auto<T,(Options_&DontAlign)==0>(1);
+      }
+    }
+    EIGEN_DEVICE_FUNC TensorStorage(internal::constructor_without_unaligned_array_assert)
+      : m_data(0), m_dimensions(internal::template repeat<NumIndices_, Index>(0)) {}
+    EIGEN_DEVICE_FUNC TensorStorage(Index size, const array<Index, NumIndices_>& dimensions)
+        : m_data(internal::conditional_aligned_new_auto<T,(Options_&DontAlign)==0>(size)), m_dimensions(dimensions)
+      { EIGEN_INTERNAL_TENSOR_STORAGE_CTOR_PLUGIN }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    template <typename... DenseIndex>
+    EIGEN_DEVICE_FUNC TensorStorage(DenseIndex... indices) : m_dimensions(indices...) {
+      m_data = internal::conditional_aligned_new_auto<T,(Options_&DontAlign)==0>(internal::array_prod(m_dimensions));
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC TensorStorage(const Self& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(Options_&DontAlign)==0>(internal::array_prod(other.m_dimensions)))
+      , m_dimensions(other.m_dimensions)
+    {
+      internal::smart_copy(other.m_data, other.m_data+internal::array_prod(other.m_dimensions), m_data);
+    }
+    EIGEN_DEVICE_FUNC Self& operator=(const Self& other)
+    {
+      if (this != &other) {
+        Self tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }
+
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC TensorStorage(Self&& other) : TensorStorage()
+    {
+      *this = std::move(other);
+    }
+    
+    EIGEN_DEVICE_FUNC Self& operator=(Self&& other)
+    {
+      numext::swap(m_data, other.m_data);
+      numext::swap(m_dimensions, other.m_dimensions);
+      return *this;
+    }
+#endif
+
+    EIGEN_DEVICE_FUNC  ~TensorStorage() { internal::conditional_aligned_delete_auto<T,(Options_&DontAlign)==0>(m_data, internal::array_prod(m_dimensions)); }
+    EIGEN_DEVICE_FUNC  void swap(Self& other)
+    { numext::swap(m_data,other.m_data); numext::swap(m_dimensions,other.m_dimensions); }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {return m_dimensions;}
+
+    EIGEN_DEVICE_FUNC void resize(Index size, const array<Index, NumIndices_>& nbDimensions)
+    {
+      const Index currentSz = internal::array_prod(m_dimensions);
+      if(size != currentSz)
+      {
+        internal::conditional_aligned_delete_auto<T,(Options_&DontAlign)==0>(m_data, currentSz);
+        if (size)
+          m_data = internal::conditional_aligned_new_auto<T,(Options_&DontAlign)==0>(size);
+        else if (NumIndices_ == 0) {
+	  m_data = internal::conditional_aligned_new_auto<T,(Options_&DontAlign)==0>(1);
+	}
+	else 
+          m_data = 0;
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      }
+      m_dimensions = nbDimensions;
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T *data() { return m_data; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T *data() const { return m_data; }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index size() const { return m_dimensions.TotalSize(); }
+
+ private:
+  T *m_data;
+  Dimensions m_dimensions;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSORSTORAGE_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorStriding.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorStriding.h
new file mode 100644
index 0000000..2f62a66
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorStriding.h
@@ -0,0 +1,346 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H
+#define EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H
+
+namespace Eigen {
+
+/** \class TensorStriding
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor striding class.
+  *
+  *
+  */
+namespace internal {
+template<typename Strides, typename XprType>
+struct traits<TensorStridingOp<Strides, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename Strides, typename XprType>
+struct eval<TensorStridingOp<Strides, XprType>, Eigen::Dense>
+{
+  typedef const TensorStridingOp<Strides, XprType>EIGEN_DEVICE_REF type;
+};
+
+template<typename Strides, typename XprType>
+struct nested<TensorStridingOp<Strides, XprType>, 1, typename eval<TensorStridingOp<Strides, XprType> >::type>
+{
+  typedef TensorStridingOp<Strides, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename Strides, typename XprType>
+class TensorStridingOp : public TensorBase<TensorStridingOp<Strides, XprType> >
+{
+  public:
+    typedef TensorBase<TensorStridingOp<Strides, XprType> > Base;
+    typedef typename Eigen::internal::traits<TensorStridingOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename XprType::CoeffReturnType CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorStridingOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorStridingOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorStridingOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorStridingOp(const XprType& expr, const Strides& dims)
+      : m_xpr(expr), m_dims(dims) {}
+
+    EIGEN_DEVICE_FUNC
+    const Strides& strides() const { return m_dims; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorStridingOp)
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Strides m_dims;
+};
+
+
+// Eval as rvalue
+template<typename Strides, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
+{
+  typedef TensorStridingOp<Strides, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device)
+  {
+    m_dimensions = m_impl.dimensions();
+    for (int i = 0; i < NumDims; ++i) {
+      m_dimensions[i] =Eigen::numext::ceil(static_cast<float>(m_dimensions[i]) / op.strides()[i]);
+    }
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_outputStrides[0] = 1;
+      m_inputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
+        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+        m_inputStrides[i-1] *= op.strides()[i-1];
+      }
+      m_inputStrides[NumDims-1] *= op.strides()[NumDims-1];
+    } else {  // RowMajor
+      m_outputStrides[NumDims-1] = 1;
+      m_inputStrides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
+        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
+        m_inputStrides[i+1] *= op.strides()[i+1];
+      }
+      m_inputStrides[0] *= op.strides()[0];
+    }
+  }
+
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType/*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return m_impl.coeff(srcCoeff(index));
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    Index inputIndices[] = {0, 0};
+    Index indices[] = {index, index + PacketSize - 1};
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx0 = indices[0] / m_outputStrides[i];
+        const Index idx1 = indices[1] / m_outputStrides[i];
+        inputIndices[0] += idx0 * m_inputStrides[i];
+        inputIndices[1] += idx1 * m_inputStrides[i];
+        indices[0] -= idx0 * m_outputStrides[i];
+        indices[1] -= idx1 * m_outputStrides[i];
+      }
+      inputIndices[0] += indices[0] * m_inputStrides[0];
+      inputIndices[1] += indices[1] * m_inputStrides[0];
+    } else {  // RowMajor
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx0 = indices[0] / m_outputStrides[i];
+        const Index idx1 = indices[1] / m_outputStrides[i];
+        inputIndices[0] += idx0 * m_inputStrides[i];
+        inputIndices[1] += idx1 * m_inputStrides[i];
+        indices[0] -= idx0 * m_outputStrides[i];
+        indices[1] -= idx1 * m_outputStrides[i];
+      }
+      inputIndices[0] += indices[0] * m_inputStrides[NumDims-1];
+      inputIndices[1] += indices[1] * m_inputStrides[NumDims-1];
+    }
+    if (inputIndices[1] - inputIndices[0] == PacketSize - 1) {
+      PacketReturnType rslt = m_impl.template packet<Unaligned>(inputIndices[0]);
+      return rslt;
+    }
+    else {
+      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+      values[0] = m_impl.coeff(inputIndices[0]);
+      values[PacketSize-1] = m_impl.coeff(inputIndices[1]);
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < PacketSize-1; ++i) {
+        values[i] = coeff(index+i);
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    double compute_cost = (NumDims - 1) * (TensorOpCost::AddCost<Index>() +
+                                           TensorOpCost::MulCost<Index>() +
+                                           TensorOpCost::DivCost<Index>()) +
+        TensorOpCost::MulCost<Index>();
+    if (vectorized) {
+      compute_cost *= 2;  // packet() computes two indices
+    }
+    const int innerDim = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? 0 : (NumDims - 1);
+    return m_impl.costPerCoeff(vectorized && m_inputStrides[innerDim] == 1) +
+        // Computation is not vectorized per se, but it is done once per packet.
+        TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC typename Storage::Type data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const
+  {
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_outputStrides[i];
+        inputIndex += idx * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      inputIndex += index * m_inputStrides[0];
+    } else {  // RowMajor
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_outputStrides[i];
+        inputIndex += idx * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      inputIndex += index * m_inputStrides[NumDims-1];
+    }
+    return inputIndex;
+  }
+
+  Dimensions m_dimensions;
+  array<Index, NumDims> m_outputStrides;
+  array<Index, NumDims> m_inputStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+};
+
+// Eval as lvalue
+template<typename Strides, typename ArgType, typename Device>
+struct TensorEvaluator<TensorStridingOp<Strides, ArgType>, Device>
+    : public TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
+{
+  typedef TensorStridingOp<Strides, ArgType> XprType;
+  typedef TensorEvaluator<const XprType, Device> Base;
+  //  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  //  typedef DSizes<Index, NumDims> Dimensions;
+
+  enum {
+    IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : Base(op, device) { }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
+  {
+    return this->m_impl.coeffRef(this->srcCoeff(index));
+  }
+
+  template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketReturnType& x)
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < this->dimensions().TotalSize());
+
+    Index inputIndices[] = {0, 0};
+    Index indices[] = {index, index + PacketSize - 1};
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx0 = indices[0] / this->m_outputStrides[i];
+        const Index idx1 = indices[1] / this->m_outputStrides[i];
+        inputIndices[0] += idx0 * this->m_inputStrides[i];
+        inputIndices[1] += idx1 * this->m_inputStrides[i];
+        indices[0] -= idx0 * this->m_outputStrides[i];
+        indices[1] -= idx1 * this->m_outputStrides[i];
+      }
+      inputIndices[0] += indices[0] * this->m_inputStrides[0];
+      inputIndices[1] += indices[1] * this->m_inputStrides[0];
+    } else {  // RowMajor
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx0 = indices[0] / this->m_outputStrides[i];
+        const Index idx1 = indices[1] / this->m_outputStrides[i];
+        inputIndices[0] += idx0 * this->m_inputStrides[i];
+        inputIndices[1] += idx1 * this->m_inputStrides[i];
+        indices[0] -= idx0 * this->m_outputStrides[i];
+        indices[1] -= idx1 * this->m_outputStrides[i];
+      }
+      inputIndices[0] += indices[0] * this->m_inputStrides[NumDims-1];
+      inputIndices[1] += indices[1] * this->m_inputStrides[NumDims-1];
+    }
+    if (inputIndices[1] - inputIndices[0] == PacketSize - 1) {
+      this->m_impl.template writePacket<Unaligned>(inputIndices[0], x);
+    }
+    else {
+      EIGEN_ALIGN_MAX Scalar values[PacketSize];
+      internal::pstore<Scalar, PacketReturnType>(values, x);
+      this->m_impl.coeffRef(inputIndices[0]) = values[0];
+      this->m_impl.coeffRef(inputIndices[1]) = values[PacketSize-1];
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < PacketSize-1; ++i) {
+        this->coeffRef(index+i) = values[i];
+      }
+    }
+  }
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorTrace.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorTrace.h
new file mode 100644
index 0000000..926ecdd
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorTrace.h
@@ -0,0 +1,303 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2017 Gagan Goel <gagan.nith@gmail.com>
+// Copyright (C) 2017 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_TRACE_H
+#define EIGEN_CXX11_TENSOR_TENSOR_TRACE_H
+
+namespace Eigen {
+
+/** \class TensorTrace
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor Trace class.
+  *
+  *
+  */
+
+namespace internal {
+template<typename Dims, typename XprType>
+struct traits<TensorTraceOp<Dims, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions - array_size<Dims>::value;
+  static const int Layout = XprTraits::Layout;
+};
+
+template<typename Dims, typename XprType>
+struct eval<TensorTraceOp<Dims, XprType>, Eigen::Dense>
+{
+  typedef const TensorTraceOp<Dims, XprType>& type;
+};
+
+template<typename Dims, typename XprType>
+struct nested<TensorTraceOp<Dims, XprType>, 1, typename eval<TensorTraceOp<Dims, XprType> >::type>
+{
+  typedef TensorTraceOp<Dims, XprType> type;
+};
+
+} // end namespace internal
+
+
+template<typename Dims, typename XprType>
+class TensorTraceOp : public TensorBase<TensorTraceOp<Dims, XprType> >
+{
+  public:
+    typedef typename Eigen::internal::traits<TensorTraceOp>::Scalar Scalar;
+    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+    typedef typename XprType::CoeffReturnType CoeffReturnType;
+    typedef typename Eigen::internal::nested<TensorTraceOp>::type Nested;
+    typedef typename Eigen::internal::traits<TensorTraceOp>::StorageKind StorageKind;
+    typedef typename Eigen::internal::traits<TensorTraceOp>::Index Index;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorTraceOp(const XprType& expr, const Dims& dims)
+      : m_xpr(expr), m_dims(dims) {
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const Dims& dims() const { return m_dims; }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const typename internal::remove_all<typename XprType::Nested>::type& expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Dims m_dims;
+};
+
+
+// Eval as rvalue
+template<typename Dims, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
+{
+  typedef TensorTraceOp<Dims, ArgType> XprType;
+  static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  static const int NumReducedDims = internal::array_size<Dims>::value;
+  static const int NumOutputDims = NumInputDims - NumReducedDims;
+  typedef typename XprType::Index Index;
+  typedef DSizes<Index, NumOutputDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : m_impl(op.expression(), device), m_traceDim(1), m_device(device)
+  {
+
+    EIGEN_STATIC_ASSERT((NumOutputDims >= 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT((NumReducedDims >= 2) || ((NumReducedDims == 0) && (NumInputDims == 0)), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    for (int i = 0; i < NumInputDims; ++i) {
+      m_reduced[i] = false;
+    }
+
+    const Dims& op_dims = op.dims();
+    for (int i = 0; i < NumReducedDims; ++i) {
+      eigen_assert(op_dims[i] >= 0);
+      eigen_assert(op_dims[i] < NumInputDims);
+      m_reduced[op_dims[i]] = true;
+    }
+
+    // All the dimensions should be distinct to compute the trace
+    int num_distinct_reduce_dims = 0;
+    for (int i = 0; i < NumInputDims; ++i) {
+      if (m_reduced[i]) {
+        ++num_distinct_reduce_dims;
+      }
+    }
+
+    eigen_assert(num_distinct_reduce_dims == NumReducedDims);
+
+    // Compute the dimensions of the result.
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+
+    int output_index = 0;
+    int reduced_index = 0;
+    for (int i = 0; i < NumInputDims; ++i) {
+      if (m_reduced[i]) {
+        m_reducedDims[reduced_index] = input_dims[i];
+        if (reduced_index > 0) {
+          // All the trace dimensions must have the same size
+          eigen_assert(m_reducedDims[0] == m_reducedDims[reduced_index]);
+        }
+        ++reduced_index;
+      }
+      else {
+        m_dimensions[output_index] = input_dims[i];
+        ++output_index;
+      }
+    }
+
+    if (NumReducedDims != 0) {
+      m_traceDim = m_reducedDims[0];
+    }
+
+    // Compute the output strides
+    if (NumOutputDims > 0) {
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        m_outputStrides[0] = 1;
+        for (int i = 1; i < NumOutputDims; ++i) {
+          m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
+        }
+      }
+      else {
+        m_outputStrides.back() = 1;
+        for (int i = NumOutputDims - 2; i >= 0; --i) {
+          m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
+        }
+      }
+    }
+
+    // Compute the input strides
+    if (NumInputDims > 0) {
+      array<Index, NumInputDims> input_strides;
+      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+        input_strides[0] = 1;
+        for (int i = 1; i < NumInputDims; ++i) {
+          input_strides[i] = input_strides[i - 1] * input_dims[i - 1];
+        }
+      }
+      else {
+        input_strides.back() = 1;
+        for (int i = NumInputDims - 2; i >= 0; --i) {
+          input_strides[i] = input_strides[i + 1] * input_dims[i + 1];
+        }
+      }
+
+      output_index = 0;
+      reduced_index = 0;
+      for (int i = 0; i < NumInputDims; ++i) {
+        if(m_reduced[i]) {
+          m_reducedStrides[reduced_index] = input_strides[i];
+          ++reduced_index;
+        }
+        else {
+          m_preservedStrides[output_index] = input_strides[i];
+          ++output_index;
+        }
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
+    return m_dimensions;
+  }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    // Initialize the result
+    CoeffReturnType result = internal::cast<int, CoeffReturnType>(0);
+    Index index_stride = 0;
+    for (int i = 0; i < NumReducedDims; ++i) {
+      index_stride += m_reducedStrides[i];
+    }
+
+    // If trace is requested along all dimensions, starting index would be 0
+    Index cur_index = 0;
+    if (NumOutputDims != 0)
+      cur_index = firstInput(index);
+    for (Index i = 0; i < m_traceDim; ++i) {
+        result += m_impl.coeff(cur_index);
+        cur_index += index_stride;
+    }
+
+    return result;
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const {
+
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    eigen_assert(index + PacketSize - 1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    for (int i = 0; i < PacketSize; ++i) {
+        values[i] = coeff(index + i);
+    }
+    PacketReturnType result = internal::ploadt<PacketReturnType, LoadMode>(values);
+    return result;
+  }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ protected:
+  // Given the output index, finds the first index in the input tensor used to compute the trace
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const {
+    Index startInput = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumOutputDims - 1; i > 0; --i) {
+        const Index idx = index / m_outputStrides[i];
+        startInput += idx * m_preservedStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      startInput += index * m_preservedStrides[0];
+    }
+    else {
+      for (int i = 0; i < NumOutputDims - 1; ++i) {
+        const Index idx = index / m_outputStrides[i];
+        startInput += idx * m_preservedStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      startInput += index * m_preservedStrides[NumOutputDims - 1];
+    }
+    return startInput;
+  }
+
+  Dimensions m_dimensions;
+  TensorEvaluator<ArgType, Device> m_impl;
+  // Initialize the size of the trace dimension
+  Index m_traceDim;
+  const Device EIGEN_DEVICE_REF m_device;
+  array<bool, NumInputDims> m_reduced;
+  array<Index, NumReducedDims> m_reducedDims;
+  array<Index, NumOutputDims> m_outputStrides;
+  array<Index, NumReducedDims> m_reducedStrides;
+  array<Index, NumOutputDims> m_preservedStrides;
+};
+
+
+} // End namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_TRACE_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorTraits.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorTraits.h
new file mode 100644
index 0000000..4f7fd34
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorTraits.h
@@ -0,0 +1,264 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_TRAITS_H
+#define EIGEN_CXX11_TENSOR_TENSOR_TRAITS_H
+
+namespace Eigen {
+namespace internal {
+
+
+template<typename Scalar, int Options>
+class compute_tensor_flags
+{
+  enum {
+    is_dynamic_size_storage = 1,
+
+    is_aligned =
+    (
+        ((Options&DontAlign)==0) && (
+#if EIGEN_MAX_STATIC_ALIGN_BYTES>0
+            (!is_dynamic_size_storage)
+#else
+            0
+#endif
+            |
+#if EIGEN_MAX_ALIGN_BYTES>0
+            is_dynamic_size_storage
+#else
+            0
+#endif
+      )
+     ),
+    packet_access_bit = packet_traits<Scalar>::Vectorizable && is_aligned ? PacketAccessBit : 0
+  };
+
+  public:
+    enum { ret = packet_access_bit };
+};
+
+
+template<typename Scalar_, int NumIndices_, int Options_, typename IndexType_>
+struct traits<Tensor<Scalar_, NumIndices_, Options_, IndexType_> >
+{
+  typedef Scalar_ Scalar;
+  typedef Dense StorageKind;
+  typedef IndexType_ Index;
+  static const int NumDimensions = NumIndices_;
+  static const int Layout = Options_ & RowMajor ? RowMajor : ColMajor;
+  enum {
+    Options = Options_,
+    Flags = compute_tensor_flags<Scalar_, Options_>::ret | (is_const<Scalar_>::value ? 0 : LvalueBit)
+  };
+  template <typename T> struct MakePointer {
+    typedef T* Type;
+  };
+  typedef typename MakePointer<Scalar>::Type PointerType;
+};
+
+
+template<typename Scalar_, typename Dimensions, int Options_, typename IndexType_>
+struct traits<TensorFixedSize<Scalar_, Dimensions, Options_, IndexType_> >
+{
+  typedef Scalar_ Scalar;
+  typedef Dense StorageKind;
+  typedef IndexType_ Index;
+  static const int NumDimensions = array_size<Dimensions>::value;
+  static const int Layout = Options_ & RowMajor ? RowMajor : ColMajor;
+  enum {
+    Options = Options_,
+    Flags = compute_tensor_flags<Scalar_, Options_>::ret | (is_const<Scalar_>::value ? 0: LvalueBit)
+  };
+  template <typename T> struct MakePointer {
+    typedef T* Type;
+  };
+  typedef typename MakePointer<Scalar>::Type PointerType;
+};
+
+
+template<typename PlainObjectType, int Options_, template <class> class MakePointer_>
+struct traits<TensorMap<PlainObjectType, Options_, MakePointer_> >
+  : public traits<PlainObjectType>
+{
+  typedef traits<PlainObjectType> BaseTraits;
+  typedef typename BaseTraits::Scalar Scalar;
+  typedef typename BaseTraits::StorageKind StorageKind;
+  typedef typename BaseTraits::Index Index;
+  static const int NumDimensions = BaseTraits::NumDimensions;
+  static const int Layout = BaseTraits::Layout;
+  enum {
+    Options = Options_,
+    Flags = BaseTraits::Flags
+  };
+  template <class T> struct MakePointer {
+    // Intermediate typedef to workaround MSVC issue.
+    typedef MakePointer_<T> MakePointerT;
+    typedef typename MakePointerT::Type Type;
+  };
+  typedef typename MakePointer<Scalar>::Type PointerType;
+};
+
+template<typename PlainObjectType>
+struct traits<TensorRef<PlainObjectType> >
+  : public traits<PlainObjectType>
+{
+  typedef traits<PlainObjectType> BaseTraits;
+  typedef typename BaseTraits::Scalar Scalar;
+  typedef typename BaseTraits::StorageKind StorageKind;
+  typedef typename BaseTraits::Index Index;
+  static const int NumDimensions = BaseTraits::NumDimensions;
+  static const int Layout = BaseTraits::Layout;
+  enum {
+    Options = BaseTraits::Options,
+    Flags = BaseTraits::Flags
+  };
+  typedef typename BaseTraits::PointerType PointerType;
+};
+
+
+template<typename _Scalar, int NumIndices_, int Options, typename IndexType_>
+struct eval<Tensor<_Scalar, NumIndices_, Options, IndexType_>, Eigen::Dense>
+{
+  typedef const Tensor<_Scalar, NumIndices_, Options, IndexType_>EIGEN_DEVICE_REF type;
+};
+
+template<typename _Scalar, int NumIndices_, int Options, typename IndexType_>
+struct eval<const Tensor<_Scalar, NumIndices_, Options, IndexType_>, Eigen::Dense>
+{
+  typedef const Tensor<_Scalar, NumIndices_, Options, IndexType_>EIGEN_DEVICE_REF type;
+};
+
+template<typename Scalar_, typename Dimensions, int Options, typename IndexType_>
+struct eval<TensorFixedSize<Scalar_, Dimensions, Options, IndexType_>, Eigen::Dense>
+{
+  typedef const TensorFixedSize<Scalar_, Dimensions, Options, IndexType_>EIGEN_DEVICE_REF type;
+};
+
+template<typename Scalar_, typename Dimensions, int Options, typename IndexType_>
+struct eval<const TensorFixedSize<Scalar_, Dimensions, Options, IndexType_>, Eigen::Dense>
+{
+  typedef const TensorFixedSize<Scalar_, Dimensions, Options, IndexType_>EIGEN_DEVICE_REF type;
+};
+
+template<typename PlainObjectType, int Options, template <class> class MakePointer>
+struct eval<TensorMap<PlainObjectType, Options, MakePointer>, Eigen::Dense>
+{
+  typedef const TensorMap<PlainObjectType, Options, MakePointer>EIGEN_DEVICE_REF type;
+};
+
+template<typename PlainObjectType, int Options, template <class> class MakePointer>
+struct eval<const TensorMap<PlainObjectType, Options, MakePointer>, Eigen::Dense>
+{
+  typedef const TensorMap<PlainObjectType, Options, MakePointer>EIGEN_DEVICE_REF type;
+};
+
+template<typename PlainObjectType>
+struct eval<TensorRef<PlainObjectType>, Eigen::Dense>
+{
+  typedef const TensorRef<PlainObjectType>EIGEN_DEVICE_REF type;
+};
+
+template<typename PlainObjectType>
+struct eval<const TensorRef<PlainObjectType>, Eigen::Dense>
+{
+  typedef const TensorRef<PlainObjectType>EIGEN_DEVICE_REF type;
+};
+
+// TODO nested<> does not exist anymore in Eigen/Core, and it thus has to be removed in favor of ref_selector.
+template<typename T, int n=1, typename PlainObject = void> struct nested
+{
+  typedef typename ref_selector<T>::type type;
+};
+
+template <typename Scalar_, int NumIndices_, int Options_, typename IndexType_>
+struct nested<Tensor<Scalar_, NumIndices_, Options_, IndexType_> >
+{
+  typedef const Tensor<Scalar_, NumIndices_, Options_, IndexType_>EIGEN_DEVICE_REF type;
+};
+
+template <typename Scalar_, int NumIndices_, int Options_, typename IndexType_>
+struct nested<const Tensor<Scalar_, NumIndices_, Options_, IndexType_> >
+{
+  typedef const Tensor<Scalar_, NumIndices_, Options_, IndexType_>EIGEN_DEVICE_REF type;
+};
+
+template <typename Scalar_, typename Dimensions, int Options, typename IndexType_>
+struct nested<TensorFixedSize<Scalar_, Dimensions, Options, IndexType_> >
+{
+  typedef const TensorFixedSize<Scalar_, Dimensions, Options, IndexType_>EIGEN_DEVICE_REF type;
+};
+
+template <typename Scalar_, typename Dimensions, int Options, typename IndexType_>
+struct nested<const TensorFixedSize<Scalar_, Dimensions, Options, IndexType_> >
+{
+  typedef const TensorFixedSize<Scalar_, Dimensions, Options, IndexType_>EIGEN_DEVICE_REF type;
+};
+
+
+template <typename PlainObjectType>
+struct nested<TensorRef<PlainObjectType> >
+{
+  typedef const TensorRef<PlainObjectType>EIGEN_DEVICE_REF type;
+};
+
+template <typename PlainObjectType>
+struct nested<const TensorRef<PlainObjectType> >
+{
+  typedef const TensorRef<PlainObjectType>EIGEN_DEVICE_REF type;
+};
+
+}  // end namespace internal
+
+// Convolutional layers take in an input tensor of shape (D, R, C, B), or (D, C,
+// R, B), and convolve it with a set of filters, which can also be presented as
+// a tensor (D, K, K, M), where M is the number of filters, K is the filter
+// size, and each 3-dimensional tensor of size (D, K, K) is a filter. For
+// simplicity we assume that we always use square filters (which is usually the
+// case in images), hence the two Ks in the tensor dimension.  It also takes in
+// a few additional parameters:
+// Stride (S): The convolution stride is the offset between locations where we
+//             apply the filters.  A larger stride means that the output will be
+//             spatially smaller.
+// Padding (P): The padding we apply to the input tensor along the R and C
+//              dimensions.  This is usually used to make sure that the spatial
+//              dimensions of the output matches our intention.
+//
+// Two types of padding are often used:
+//   SAME: The pad value is computed so that the output will have size
+//         R/S and C/S.
+//   VALID: no padding is carried out.
+// When we do padding, the padded values at the padded locations are usually
+// zero.
+//
+// The output dimensions for convolution, when given all the parameters above,
+// are as follows:
+// When Padding = SAME: the output size is (B, R', C', M), where
+//   R' = ceil(float(R) / float(S))
+//   C' = ceil(float(C) / float(S))
+// where ceil is the ceiling function.  The input tensor is padded with 0 as
+// needed.  The number of padded rows and columns are computed as:
+//   Pr = ((R' - 1) * S + K - R) / 2
+//   Pc = ((C' - 1) * S + K - C) / 2
+// when the stride is 1, we have the simplified case R'=R, C'=C, Pr=Pc=(K-1)/2.
+// This is where SAME comes from - the output has the same size as the input has.
+// When Padding = VALID: the output size is computed as
+//   R' = ceil(float(R - K + 1) / float(S))
+//   C' = ceil(float(C - K + 1) / float(S))
+// and the number of padded rows and columns are computed in the same way as in
+// the SAME case.
+// When the stride is 1, we have the simplified case R'=R-K+1, C'=C-K+1, Pr=0,
+// Pc=0.
+typedef enum {
+  PADDING_VALID = 1,
+  PADDING_SAME = 2
+} PaddingType;
+
+}  // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_TRAITS_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorUInt128.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorUInt128.h
new file mode 100644
index 0000000..d23f2e4
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorUInt128.h
@@ -0,0 +1,249 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_UINT128_H
+#define EIGEN_CXX11_TENSOR_TENSOR_UINT128_H
+
+namespace Eigen {
+namespace internal {
+
+
+template <uint64_t n>
+struct static_val {
+  static const uint64_t value = n;
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE operator uint64_t() const { return n; }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE static_val() { }
+
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE static_val(const T& v) {
+    EIGEN_UNUSED_VARIABLE(v);
+    eigen_assert(v == n);
+  }
+};
+
+
+template <typename HIGH = uint64_t, typename LOW = uint64_t>
+struct TensorUInt128
+{
+  HIGH high;
+  LOW low;
+
+  template<typename OTHER_HIGH, typename OTHER_LOW>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  TensorUInt128(const TensorUInt128<OTHER_HIGH, OTHER_LOW>& other) : high(other.high), low(other.low) {
+    EIGEN_STATIC_ASSERT(sizeof(OTHER_HIGH) <= sizeof(HIGH), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT(sizeof(OTHER_LOW) <= sizeof(LOW), YOU_MADE_A_PROGRAMMING_MISTAKE);
+  }
+
+  template<typename OTHER_HIGH, typename OTHER_LOW>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  TensorUInt128& operator = (const TensorUInt128<OTHER_HIGH, OTHER_LOW>& other) {
+    EIGEN_STATIC_ASSERT(sizeof(OTHER_HIGH) <= sizeof(HIGH), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    EIGEN_STATIC_ASSERT(sizeof(OTHER_LOW) <= sizeof(LOW), YOU_MADE_A_PROGRAMMING_MISTAKE);
+    high = other.high;
+    low = other.low;
+    return *this;
+  }
+
+  template<typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  explicit TensorUInt128(const T& x) : high(0), low(x) {
+    eigen_assert((static_cast<typename conditional<sizeof(T) == 8, uint64_t, uint32_t>::type>(x) <= NumTraits<uint64_t>::highest()));
+    eigen_assert(x >= 0);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  TensorUInt128(HIGH y, LOW x) : high(y), low(x) { }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE operator LOW() const {
+    return low;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE LOW lower() const {
+    return low;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HIGH upper() const {
+    return high;
+  }
+};
+
+
+template <typename HL, typename LL, typename HR, typename LR>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool operator == (const TensorUInt128<HL, LL>& lhs, const TensorUInt128<HR, LR>& rhs)
+{
+  return (lhs.high == rhs.high) & (lhs.low == rhs.low);
+}
+
+template <typename HL, typename LL, typename HR, typename LR>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool operator != (const TensorUInt128<HL, LL>& lhs, const TensorUInt128<HR, LR>& rhs)
+{
+  return (lhs.high != rhs.high) | (lhs.low != rhs.low);
+}
+
+template <typename HL, typename LL, typename HR, typename LR>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool operator >= (const TensorUInt128<HL, LL>& lhs, const TensorUInt128<HR, LR>& rhs)
+{
+  if (lhs.high != rhs.high) {
+    return lhs.high > rhs.high;
+  }
+  return lhs.low >= rhs.low;
+}
+
+template <typename HL, typename LL, typename HR, typename LR>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool operator < (const TensorUInt128<HL, LL>& lhs, const TensorUInt128<HR, LR>& rhs)
+{
+  if (lhs.high != rhs.high) {
+    return lhs.high < rhs.high;
+  }
+  return lhs.low < rhs.low;
+}
+
+template <typename HL, typename LL, typename HR, typename LR>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+TensorUInt128<uint64_t, uint64_t> operator + (const TensorUInt128<HL, LL>& lhs, const TensorUInt128<HR, LR>& rhs)
+{
+  TensorUInt128<uint64_t, uint64_t> result(lhs.high + rhs.high, lhs.low + rhs.low);
+  if (result.low < rhs.low) {
+    result.high += 1;
+  }
+  return result;
+}
+
+template <typename HL, typename LL, typename HR, typename LR>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+TensorUInt128<uint64_t, uint64_t> operator - (const TensorUInt128<HL, LL>& lhs, const TensorUInt128<HR, LR>& rhs)
+{
+  TensorUInt128<uint64_t, uint64_t> result(lhs.high - rhs.high, lhs.low - rhs.low);
+  if (result.low > lhs.low) {
+    result.high -= 1;
+  }
+  return result;
+}
+
+
+template <typename HL, typename LL, typename HR, typename LR>
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+TensorUInt128<uint64_t, uint64_t> operator * (const TensorUInt128<HL, LL>& lhs, const TensorUInt128<HR, LR>& rhs)
+{
+  // Split each 128-bit integer into 4 32-bit integers, and then do the
+  // multiplications by hand as follow:
+  //   lhs      a  b  c  d
+  //   rhs      e  f  g  h
+  //           -----------
+  //           ah bh ch dh
+  //           bg cg dg
+  //           cf df
+  //           de
+  // The result is stored in 2 64bit integers, high and low.
+
+  const uint64_t LOW = 0x00000000FFFFFFFFLL;
+  const uint64_t HIGH = 0xFFFFFFFF00000000LL;
+
+  uint64_t d = lhs.low & LOW;
+  uint64_t c = (lhs.low & HIGH) >> 32LL;
+  uint64_t b = lhs.high & LOW;
+  uint64_t a = (lhs.high & HIGH) >> 32LL;
+
+  uint64_t h = rhs.low & LOW;
+  uint64_t g = (rhs.low & HIGH) >> 32LL;
+  uint64_t f = rhs.high & LOW;
+  uint64_t e = (rhs.high & HIGH) >> 32LL;
+
+  // Compute the low 32 bits of low
+  uint64_t acc = d * h;
+  uint64_t low = acc & LOW;
+  //  Compute the high 32 bits of low. Add a carry every time we wrap around
+  acc >>= 32LL;
+  uint64_t carry = 0;
+  uint64_t acc2 = acc + c * h;
+  if (acc2 < acc) {
+    carry++;
+  }
+  acc = acc2 + d * g;
+  if (acc < acc2) {
+    carry++;
+  }
+  low |= (acc << 32LL);
+
+  // Carry forward the high bits of acc to initiate the computation of the
+  // low 32 bits of high
+  acc2 = (acc >> 32LL) | (carry << 32LL);
+  carry = 0;
+
+  acc = acc2 + b * h;
+  if (acc < acc2) {
+    carry++;
+  }
+  acc2 = acc + c * g;
+  if (acc2 < acc) {
+    carry++;
+  }
+  acc = acc2 + d * f;
+  if (acc < acc2) {
+    carry++;
+  }
+  uint64_t high = acc & LOW;
+
+  // Start to compute the high 32 bits of high.
+  acc2 = (acc >> 32LL) | (carry << 32LL);
+
+  acc = acc2 + a * h;
+  acc2 = acc + b * g;
+  acc = acc2 + c * f;
+  acc2 = acc + d * e;
+  high |= (acc2 << 32LL);
+
+  return TensorUInt128<uint64_t, uint64_t>(high, low);
+}
+
+template <typename HL, typename LL, typename HR, typename LR>
+static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+TensorUInt128<uint64_t, uint64_t> operator / (const TensorUInt128<HL, LL>& lhs, const TensorUInt128<HR, LR>& rhs)
+{
+  if (rhs == TensorUInt128<static_val<0>, static_val<1> >(1)) {
+    return TensorUInt128<uint64_t, uint64_t>(lhs.high, lhs.low);
+  } else if (lhs < rhs) {
+    return TensorUInt128<uint64_t, uint64_t>(0);
+  } else {
+    // calculate the biggest power of 2 times rhs that's less than or equal to lhs
+    TensorUInt128<uint64_t, uint64_t> power2(1);
+    TensorUInt128<uint64_t, uint64_t> d(rhs);
+    TensorUInt128<uint64_t, uint64_t> tmp(lhs - d);
+    while (lhs >= d) {
+      tmp = tmp - d;
+      d = d + d;
+      power2 = power2 + power2;
+    }
+
+    tmp = TensorUInt128<uint64_t, uint64_t>(lhs.high, lhs.low);
+    TensorUInt128<uint64_t, uint64_t> result(0);
+    while (power2 != TensorUInt128<static_val<0>, static_val<0> >(0)) {
+      if (tmp >= d) {
+        tmp = tmp - d;
+        result = result + power2;
+      }
+      // Shift right
+      power2 = TensorUInt128<uint64_t, uint64_t>(power2.high >> 1, (power2.low >> 1) | (power2.high << 63));
+      d = TensorUInt128<uint64_t, uint64_t>(d.high >> 1, (d.low >> 1) | (d.high << 63));
+    }
+
+    return result;
+  }
+}
+
+
+}  // namespace internal
+}  // namespace Eigen
+
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_UINT128_H
diff --git a/src/EigenUnsupported/CXX11/src/Tensor/TensorVolumePatch.h b/src/EigenUnsupported/CXX11/src/Tensor/TensorVolumePatch.h
new file mode 100644
index 0000000..0beb9ff
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/Tensor/TensorVolumePatch.h
@@ -0,0 +1,629 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_VOLUME_PATCH_H
+#define EIGEN_CXX11_TENSOR_TENSOR_VOLUME_PATCH_H
+
+namespace Eigen {
+
+/** \class TensorVolumePatch
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Patch extraction specialized for processing of volumetric data.
+  * This assumes that the input has a least 4 dimensions ordered as follows:
+  *  - channels
+  *  - planes
+  *  - rows
+  *  - columns
+  *  - (optional) additional dimensions such as time or batch size.
+  * Calling the volume patch code with patch_planes, patch_rows, and patch_cols
+  * is equivalent to calling the regular patch extraction code with parameters
+  * d, patch_planes, patch_rows, patch_cols, and 1 for all the additional
+  * dimensions.
+  */
+namespace internal {
+
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct traits<TensorVolumePatchOp<Planes, Rows, Cols, XprType> > : public traits<XprType>
+{
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions + 1;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+
+};
+
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct eval<TensorVolumePatchOp<Planes, Rows, Cols, XprType>, Eigen::Dense>
+{
+  typedef const TensorVolumePatchOp<Planes, Rows, Cols, XprType>& type;
+};
+
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct nested<TensorVolumePatchOp<Planes, Rows, Cols, XprType>, 1, typename eval<TensorVolumePatchOp<Planes, Rows, Cols, XprType> >::type>
+{
+  typedef TensorVolumePatchOp<Planes, Rows, Cols, XprType> type;
+};
+
+}  // end namespace internal
+
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
+class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows, Cols, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorVolumePatchOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorVolumePatchOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorVolumePatchOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorVolumePatchOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
+                                                            DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
+                                                            DenseIndex in_plane_strides, DenseIndex in_row_strides, DenseIndex in_col_strides,
+                                                            DenseIndex plane_inflate_strides, DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
+                                                            PaddingType padding_type, Scalar padding_value)
+                                                            : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                            m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
+                                                            m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                            m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                            m_padding_explicit(false), m_padding_top_z(0), m_padding_bottom_z(0), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
+                                                            m_padding_type(padding_type), m_padding_value(padding_value) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
+                                                           DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
+                                                           DenseIndex in_plane_strides, DenseIndex in_row_strides, DenseIndex in_col_strides,
+                                                           DenseIndex plane_inflate_strides, DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
+                                                           DenseIndex padding_top_z, DenseIndex padding_bottom_z,
+                                                           DenseIndex padding_top, DenseIndex padding_bottom,
+                                                           DenseIndex padding_left, DenseIndex padding_right,
+                                                           Scalar padding_value)
+                                                           : m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+                                                           m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
+                                                           m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
+                                                           m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
+                                                           m_padding_explicit(true), m_padding_top_z(padding_top_z), m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
+                                                           m_padding_left(padding_left), m_padding_right(padding_right),
+                                                           m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
+
+    EIGEN_DEVICE_FUNC
+    DenseIndex patch_planes() const { return m_patch_planes; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex patch_rows() const { return m_patch_rows; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex patch_cols() const { return m_patch_cols; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex plane_strides() const { return m_plane_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex row_strides() const { return m_row_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex col_strides() const { return m_col_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex in_plane_strides() const { return m_in_plane_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex in_row_strides() const { return m_in_row_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex in_col_strides() const { return m_in_col_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex plane_inflate_strides() const { return m_plane_inflate_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex row_inflate_strides() const { return m_row_inflate_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex col_inflate_strides() const { return m_col_inflate_strides; }
+    EIGEN_DEVICE_FUNC
+    bool padding_explicit() const { return m_padding_explicit; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_top_z() const { return m_padding_top_z; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_bottom_z() const { return m_padding_bottom_z; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_top() const { return m_padding_top; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_bottom() const { return m_padding_bottom; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_left() const { return m_padding_left; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex padding_right() const { return m_padding_right; }
+    EIGEN_DEVICE_FUNC
+    PaddingType padding_type() const { return m_padding_type; }
+    EIGEN_DEVICE_FUNC
+    Scalar padding_value() const { return m_padding_value; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const DenseIndex m_patch_planes;
+    const DenseIndex m_patch_rows;
+    const DenseIndex m_patch_cols;
+    const DenseIndex m_plane_strides;
+    const DenseIndex m_row_strides;
+    const DenseIndex m_col_strides;
+    const DenseIndex m_in_plane_strides;
+    const DenseIndex m_in_row_strides;
+    const DenseIndex m_in_col_strides;
+    const DenseIndex m_plane_inflate_strides;
+    const DenseIndex m_row_inflate_strides;
+    const DenseIndex m_col_inflate_strides;
+    const bool m_padding_explicit;
+    const DenseIndex m_padding_top_z;
+    const DenseIndex m_padding_bottom_z;
+    const DenseIndex m_padding_top;
+    const DenseIndex m_padding_bottom;
+    const DenseIndex m_padding_left;
+    const DenseIndex m_padding_right;
+    const PaddingType m_padding_type;
+    const Scalar m_padding_value;
+};
+
+
+// Eval as rvalue
+template<DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, Device>
+{
+  typedef TensorVolumePatchOp<Planes, Rows, Cols, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  static const int NumDims = NumInputDims + 1;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) :
+ m_impl(op.expression(), device)
+  {
+    EIGEN_STATIC_ASSERT((NumDims >= 5), YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    m_paddingValue = op.padding_value();
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+
+    // Cache a few variables.
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputDepth = input_dims[0];
+      m_inputPlanes = input_dims[1];
+      m_inputRows = input_dims[2];
+      m_inputCols = input_dims[3];
+    } else {
+      m_inputDepth = input_dims[NumInputDims-1];
+      m_inputPlanes = input_dims[NumInputDims-2];
+      m_inputRows = input_dims[NumInputDims-3];
+      m_inputCols = input_dims[NumInputDims-4];
+    }
+
+    m_plane_strides = op.plane_strides();
+    m_row_strides = op.row_strides();
+    m_col_strides = op.col_strides();
+
+    // Input strides and effective input/patch size
+    m_in_plane_strides = op.in_plane_strides();
+    m_in_row_strides = op.in_row_strides();
+    m_in_col_strides = op.in_col_strides();
+    m_plane_inflate_strides = op.plane_inflate_strides();
+    m_row_inflate_strides = op.row_inflate_strides();
+    m_col_inflate_strides = op.col_inflate_strides();
+
+    // The "effective" spatial size after inflating data with zeros.
+    m_input_planes_eff = (m_inputPlanes - 1) * m_plane_inflate_strides + 1;
+    m_input_rows_eff = (m_inputRows - 1) * m_row_inflate_strides + 1;
+    m_input_cols_eff = (m_inputCols - 1) * m_col_inflate_strides + 1;
+    m_patch_planes_eff = op.patch_planes() + (op.patch_planes() - 1) * (m_in_plane_strides - 1);
+    m_patch_rows_eff = op.patch_rows() + (op.patch_rows() - 1) * (m_in_row_strides - 1);
+    m_patch_cols_eff = op.patch_cols() + (op.patch_cols() - 1) * (m_in_col_strides - 1);
+
+    if (op.padding_explicit()) {
+      m_outputPlanes = numext::ceil((m_input_planes_eff + op.padding_top_z() + op.padding_bottom_z() - m_patch_planes_eff + 1.f) / static_cast<float>(m_plane_strides));
+      m_outputRows = numext::ceil((m_input_rows_eff + op.padding_top() + op.padding_bottom() - m_patch_rows_eff + 1.f) / static_cast<float>(m_row_strides));
+      m_outputCols = numext::ceil((m_input_cols_eff + op.padding_left() + op.padding_right() - m_patch_cols_eff + 1.f) / static_cast<float>(m_col_strides));
+      m_planePaddingTop = op.padding_top_z();
+      m_rowPaddingTop = op.padding_top();
+      m_colPaddingLeft = op.padding_left();
+    } else {
+      // Computing padding from the type
+      switch (op.padding_type()) {
+        case PADDING_VALID:
+          m_outputPlanes = numext::ceil((m_input_planes_eff - m_patch_planes_eff + 1.f) / static_cast<float>(m_plane_strides));
+          m_outputRows = numext::ceil((m_input_rows_eff - m_patch_rows_eff + 1.f) / static_cast<float>(m_row_strides));
+          m_outputCols = numext::ceil((m_input_cols_eff - m_patch_cols_eff + 1.f) / static_cast<float>(m_col_strides));
+          m_planePaddingTop = 0;
+          m_rowPaddingTop = 0;
+          m_colPaddingLeft = 0;
+          break;
+        case PADDING_SAME: {
+          m_outputPlanes = numext::ceil(m_input_planes_eff / static_cast<float>(m_plane_strides));
+          m_outputRows = numext::ceil(m_input_rows_eff / static_cast<float>(m_row_strides));
+          m_outputCols = numext::ceil(m_input_cols_eff / static_cast<float>(m_col_strides));
+          const Index dz = (m_outputPlanes - 1) * m_plane_strides + m_patch_planes_eff - m_input_planes_eff;
+          const Index dy = (m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff;
+          const Index dx = (m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff;
+          m_planePaddingTop = dz / 2;
+          m_rowPaddingTop = dy / 2;
+          m_colPaddingLeft = dx / 2;
+          break;
+        }
+        default:
+          eigen_assert(false && "unexpected padding");
+      }
+    }
+    eigen_assert(m_outputRows > 0);
+    eigen_assert(m_outputCols > 0);
+    eigen_assert(m_outputPlanes > 0);
+
+    // Dimensions for result of extraction.
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      // ColMajor
+      // 0: depth
+      // 1: patch_planes
+      // 2: patch_rows
+      // 3: patch_cols
+      // 4: number of patches
+      // 5 and beyond: anything else (such as batch).
+      m_dimensions[0] = input_dims[0];
+      m_dimensions[1] = op.patch_planes();
+      m_dimensions[2] = op.patch_rows();
+      m_dimensions[3] = op.patch_cols();
+      m_dimensions[4] = m_outputPlanes * m_outputRows * m_outputCols;
+      for (int i = 5; i < NumDims; ++i) {
+        m_dimensions[i] = input_dims[i-1];
+      }
+    } else {
+      // RowMajor
+      // NumDims-1: depth
+      // NumDims-2: patch_planes
+      // NumDims-3: patch_rows
+      // NumDims-4: patch_cols
+      // NumDims-5: number of patches
+      // NumDims-6 and beyond: anything else (such as batch).
+      m_dimensions[NumDims-1] = input_dims[NumInputDims-1];
+      m_dimensions[NumDims-2] = op.patch_planes();
+      m_dimensions[NumDims-3] = op.patch_rows();
+      m_dimensions[NumDims-4] = op.patch_cols();
+      m_dimensions[NumDims-5] = m_outputPlanes * m_outputRows * m_outputCols;
+      for (int i = NumDims-6; i >= 0; --i) {
+        m_dimensions[i] = input_dims[i];
+      }
+    }
+
+    // Strides for the output tensor.
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_rowStride = m_dimensions[1];
+      m_colStride = m_dimensions[2] * m_rowStride;
+      m_patchStride = m_colStride * m_dimensions[3] * m_dimensions[0];
+      m_otherStride = m_patchStride * m_dimensions[4];
+    } else {
+      m_rowStride = m_dimensions[NumDims-2];
+      m_colStride = m_dimensions[NumDims-3] * m_rowStride;
+      m_patchStride = m_colStride * m_dimensions[NumDims-4] * m_dimensions[NumDims-1];
+      m_otherStride = m_patchStride * m_dimensions[NumDims-5];
+    }
+
+    // Strides for navigating through the input tensor.
+    m_planeInputStride = m_inputDepth;
+    m_rowInputStride = m_inputDepth * m_inputPlanes;
+    m_colInputStride = m_inputDepth * m_inputRows * m_inputPlanes;
+    m_otherInputStride = m_inputDepth * m_inputRows * m_inputCols * m_inputPlanes;
+
+    m_outputPlanesRows = m_outputPlanes * m_outputRows;
+
+    // Fast representations of different variables.
+    m_fastOtherStride = internal::TensorIntDivisor<Index>(m_otherStride);
+
+    m_fastPatchStride = internal::TensorIntDivisor<Index>(m_patchStride);
+    m_fastColStride = internal::TensorIntDivisor<Index>(m_colStride);
+    m_fastRowStride = internal::TensorIntDivisor<Index>(m_rowStride);
+    m_fastInputRowStride = internal::TensorIntDivisor<Index>(m_row_inflate_strides);
+    m_fastInputColStride = internal::TensorIntDivisor<Index>(m_col_inflate_strides);
+    m_fastInputPlaneStride = internal::TensorIntDivisor<Index>(m_plane_inflate_strides);
+    m_fastInputColsEff = internal::TensorIntDivisor<Index>(m_input_cols_eff);
+    m_fastOutputPlanes = internal::TensorIntDivisor<Index>(m_outputPlanes);
+    m_fastOutputPlanesRows = internal::TensorIntDivisor<Index>(m_outputPlanesRows);
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[0]);
+    } else {
+      m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[NumDims-1]);
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    // Patch index corresponding to the passed in index.
+    const Index patchIndex = index / m_fastPatchStride;
+
+    // Spatial offset within the patch. This has to be translated into 3D
+    // coordinates within the patch.
+    const Index patchOffset = (index - patchIndex * m_patchStride) / m_fastOutputDepth;
+
+    // Batch, etc.
+    const Index otherIndex = (NumDims == 5) ? 0 : index / m_fastOtherStride;
+    const Index patch3DIndex = (NumDims == 5) ? patchIndex : (index - otherIndex * m_otherStride) / m_fastPatchStride;
+
+    // Calculate column index in the input original tensor.
+    const Index colIndex = patch3DIndex / m_fastOutputPlanesRows;
+    const Index colOffset = patchOffset / m_fastColStride;
+    const Index inputCol = colIndex * m_col_strides + colOffset * m_in_col_strides - m_colPaddingLeft;
+    const Index origInputCol = (m_col_inflate_strides == 1) ? inputCol : ((inputCol >= 0) ? (inputCol / m_fastInputColStride) : 0);
+    if (inputCol < 0 || inputCol >= m_input_cols_eff ||
+        ((m_col_inflate_strides != 1) && (inputCol != origInputCol * m_col_inflate_strides))) {
+      return Scalar(m_paddingValue);
+    }
+
+    // Calculate row index in the original input tensor.
+    const Index rowIndex = (patch3DIndex - colIndex * m_outputPlanesRows) / m_fastOutputPlanes;
+    const Index rowOffset = (patchOffset - colOffset * m_colStride) / m_fastRowStride;
+    const Index inputRow = rowIndex * m_row_strides + rowOffset * m_in_row_strides - m_rowPaddingTop;
+    const Index origInputRow = (m_row_inflate_strides == 1) ? inputRow : ((inputRow >= 0) ? (inputRow / m_fastInputRowStride) : 0);
+    if (inputRow < 0 || inputRow >= m_input_rows_eff ||
+        ((m_row_inflate_strides != 1) && (inputRow != origInputRow * m_row_inflate_strides))) {
+      return Scalar(m_paddingValue);
+    }
+
+    // Calculate plane index in the original input tensor.
+    const Index planeIndex = (patch3DIndex - m_outputPlanes * (colIndex * m_outputRows + rowIndex));
+    const Index planeOffset = patchOffset - colOffset * m_colStride - rowOffset * m_rowStride;
+    const Index inputPlane = planeIndex * m_plane_strides + planeOffset * m_in_plane_strides - m_planePaddingTop;
+    const Index origInputPlane = (m_plane_inflate_strides == 1) ? inputPlane : ((inputPlane >= 0) ? (inputPlane / m_fastInputPlaneStride) : 0);
+    if (inputPlane < 0 || inputPlane >= m_input_planes_eff ||
+        ((m_plane_inflate_strides != 1) && (inputPlane != origInputPlane * m_plane_inflate_strides))) {
+      return Scalar(m_paddingValue);
+    }
+
+    const int depth_index = static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
+    const Index depth = index - (index / m_fastOutputDepth) * m_dimensions[depth_index];
+
+    const Index inputIndex = depth +
+        origInputRow * m_rowInputStride +
+        origInputCol * m_colInputStride +
+        origInputPlane * m_planeInputStride +
+        otherIndex * m_otherInputStride;
+
+    return m_impl.coeff(inputIndex);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    if (m_in_row_strides != 1 || m_in_col_strides != 1 || m_row_inflate_strides != 1 || m_col_inflate_strides != 1 ||
+        m_in_plane_strides != 1 || m_plane_inflate_strides != 1) {
+      return packetWithPossibleZero(index);
+    }
+
+    const Index indices[2] = {index, index + PacketSize - 1};
+    const Index patchIndex = indices[0] / m_fastPatchStride;
+    if (patchIndex != indices[1] / m_fastPatchStride) {
+      return packetWithPossibleZero(index);
+    }
+    const Index otherIndex = (NumDims == 5) ? 0 : indices[0] / m_fastOtherStride;
+    eigen_assert(otherIndex == indices[1] / m_fastOtherStride);
+
+    // Find the offset of the element wrt the location of the first element.
+    const Index patchOffsets[2] = {(indices[0] - patchIndex * m_patchStride) / m_fastOutputDepth,
+                                   (indices[1] - patchIndex * m_patchStride) / m_fastOutputDepth};
+
+    const Index patch3DIndex = (NumDims == 5) ? patchIndex : (indices[0] - otherIndex * m_otherStride) / m_fastPatchStride;
+    eigen_assert(patch3DIndex == (indices[1] - otherIndex * m_otherStride) / m_fastPatchStride);
+
+    const Index colIndex = patch3DIndex / m_fastOutputPlanesRows;
+    const Index colOffsets[2] = {
+      patchOffsets[0] / m_fastColStride,
+      patchOffsets[1] / m_fastColStride};
+
+    // Calculate col indices in the original input tensor.
+    const Index inputCols[2] = {
+      colIndex * m_col_strides + colOffsets[0] - m_colPaddingLeft,
+      colIndex * m_col_strides + colOffsets[1] - m_colPaddingLeft};
+    if (inputCols[1] < 0 || inputCols[0] >= m_inputCols) {
+      return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
+    }
+
+    if (inputCols[0] != inputCols[1]) {
+      return packetWithPossibleZero(index);
+    }
+
+    const Index rowIndex = (patch3DIndex - colIndex * m_outputPlanesRows) / m_fastOutputPlanes;
+    const Index rowOffsets[2] = {
+      (patchOffsets[0] - colOffsets[0] * m_colStride) / m_fastRowStride,
+      (patchOffsets[1] - colOffsets[1] * m_colStride) / m_fastRowStride};
+    eigen_assert(rowOffsets[0] <= rowOffsets[1]);
+    // Calculate col indices in the original input tensor.
+    const Index inputRows[2] = {
+      rowIndex * m_row_strides + rowOffsets[0] - m_rowPaddingTop,
+      rowIndex * m_row_strides + rowOffsets[1] - m_rowPaddingTop};
+
+    if (inputRows[1] < 0 || inputRows[0] >= m_inputRows) {
+      return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
+    }
+
+    if (inputRows[0] != inputRows[1]) {
+      return packetWithPossibleZero(index);
+    }
+
+    const Index planeIndex = (patch3DIndex - m_outputPlanes * (colIndex * m_outputRows + rowIndex));
+    const Index planeOffsets[2] = {
+      patchOffsets[0] - colOffsets[0] * m_colStride - rowOffsets[0] * m_rowStride,
+      patchOffsets[1] - colOffsets[1] * m_colStride - rowOffsets[1] * m_rowStride};
+    eigen_assert(planeOffsets[0] <= planeOffsets[1]);
+    const Index inputPlanes[2] = {
+      planeIndex * m_plane_strides + planeOffsets[0] - m_planePaddingTop,
+      planeIndex * m_plane_strides + planeOffsets[1] - m_planePaddingTop};
+
+    if (inputPlanes[1] < 0 || inputPlanes[0] >= m_inputPlanes) {
+      return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
+    }
+
+    if (inputPlanes[0] >= 0 && inputPlanes[1] < m_inputPlanes) {
+      // no padding
+      const int depth_index = static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
+      const Index depth = index - (index / m_fastOutputDepth) * m_dimensions[depth_index];
+      const Index inputIndex = depth +
+          inputRows[0] * m_rowInputStride +
+          inputCols[0] * m_colInputStride +
+          m_planeInputStride * inputPlanes[0] +
+          otherIndex * m_otherInputStride;
+      return m_impl.template packet<Unaligned>(inputIndex);
+    }
+
+    return packetWithPossibleZero(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
+  costPerCoeff(bool vectorized) const {
+    const double compute_cost =
+        10 * TensorOpCost::DivCost<Index>() + 21 * TensorOpCost::MulCost<Index>() +
+        8 * TensorOpCost::AddCost<Index>();
+    return TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
+
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index planePaddingTop() const { return m_planePaddingTop; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowPaddingTop() const { return m_rowPaddingTop; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colPaddingLeft() const { return m_colPaddingLeft; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputPlanes() const { return m_outputPlanes; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputRows() const { return m_outputRows; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputCols() const { return m_outputCols; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userPlaneStride() const { return m_plane_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userRowStride() const { return m_row_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userColStride() const { return m_col_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInPlaneStride() const { return m_in_plane_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInRowStride() const { return m_in_row_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInColStride() const { return m_in_col_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index planeInflateStride() const { return m_plane_inflate_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowInflateStride() const { return m_row_inflate_strides; }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colInflateStride() const { return m_col_inflate_strides; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const
+  {
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  Dimensions m_dimensions;
+
+  // Parameters passed to the constructor.
+  Index m_plane_strides;
+  Index m_row_strides;
+  Index m_col_strides;
+
+  Index m_outputPlanes;
+  Index m_outputRows;
+  Index m_outputCols;
+
+  Index m_planePaddingTop;
+  Index m_rowPaddingTop;
+  Index m_colPaddingLeft;
+
+  Index m_in_plane_strides;
+  Index m_in_row_strides;
+  Index m_in_col_strides;
+
+  Index m_plane_inflate_strides;
+  Index m_row_inflate_strides;
+  Index m_col_inflate_strides;
+
+  // Cached input size.
+  Index m_inputDepth;
+  Index m_inputPlanes;
+  Index m_inputRows;
+  Index m_inputCols;
+
+  // Other cached variables.
+  Index m_outputPlanesRows;
+
+  // Effective input/patch post-inflation size.
+  Index m_input_planes_eff;
+  Index m_input_rows_eff;
+  Index m_input_cols_eff;
+  Index m_patch_planes_eff;
+  Index m_patch_rows_eff;
+  Index m_patch_cols_eff;
+
+  // Strides for the output tensor.
+  Index m_otherStride;
+  Index m_patchStride;
+  Index m_rowStride;
+  Index m_colStride;
+
+  // Strides for the input tensor.
+  Index m_planeInputStride;
+  Index m_rowInputStride;
+  Index m_colInputStride;
+  Index m_otherInputStride;
+
+  internal::TensorIntDivisor<Index> m_fastOtherStride;
+  internal::TensorIntDivisor<Index> m_fastPatchStride;
+  internal::TensorIntDivisor<Index> m_fastColStride;
+  internal::TensorIntDivisor<Index> m_fastRowStride;
+  internal::TensorIntDivisor<Index> m_fastInputPlaneStride;
+  internal::TensorIntDivisor<Index> m_fastInputRowStride;
+  internal::TensorIntDivisor<Index> m_fastInputColStride;
+  internal::TensorIntDivisor<Index> m_fastInputColsEff;
+  internal::TensorIntDivisor<Index> m_fastOutputPlanesRows;
+  internal::TensorIntDivisor<Index> m_fastOutputPlanes;
+  internal::TensorIntDivisor<Index> m_fastOutputDepth;
+
+  Scalar m_paddingValue;
+
+  TensorEvaluator<ArgType, Device> m_impl;
+
+
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_VOLUME_PATCH_H
diff --git a/src/EigenUnsupported/CXX11/src/TensorSymmetry/DynamicSymmetry.h b/src/EigenUnsupported/CXX11/src/TensorSymmetry/DynamicSymmetry.h
new file mode 100644
index 0000000..bc4f202
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/TensorSymmetry/DynamicSymmetry.h
@@ -0,0 +1,293 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSORSYMMETRY_DYNAMICSYMMETRY_H
+#define EIGEN_CXX11_TENSORSYMMETRY_DYNAMICSYMMETRY_H
+
+namespace Eigen {
+
+class DynamicSGroup
+{
+  public:
+    inline explicit DynamicSGroup() : m_numIndices(1), m_elements(), m_generators(), m_globalFlags(0) { m_elements.push_back(ge(Generator(0, 0, 0))); }
+    inline DynamicSGroup(const DynamicSGroup& o) : m_numIndices(o.m_numIndices), m_elements(o.m_elements), m_generators(o.m_generators), m_globalFlags(o.m_globalFlags) { }
+    inline DynamicSGroup(DynamicSGroup&& o) : m_numIndices(o.m_numIndices), m_elements(), m_generators(o.m_generators), m_globalFlags(o.m_globalFlags) { std::swap(m_elements, o.m_elements); }
+    inline DynamicSGroup& operator=(const DynamicSGroup& o) { m_numIndices = o.m_numIndices; m_elements = o.m_elements; m_generators = o.m_generators; m_globalFlags = o.m_globalFlags; return *this; }
+    inline DynamicSGroup& operator=(DynamicSGroup&& o) { m_numIndices = o.m_numIndices; std::swap(m_elements, o.m_elements); m_generators = o.m_generators; m_globalFlags = o.m_globalFlags; return *this; }
+
+    void add(int one, int two, int flags = 0);
+
+    template<typename Gen_>
+    inline void add(Gen_) { add(Gen_::One, Gen_::Two, Gen_::Flags); }
+    inline void addSymmetry(int one, int two) { add(one, two, 0); }
+    inline void addAntiSymmetry(int one, int two) { add(one, two, NegationFlag); }
+    inline void addHermiticity(int one, int two) { add(one, two, ConjugationFlag); }
+    inline void addAntiHermiticity(int one, int two) { add(one, two, NegationFlag | ConjugationFlag); }
+
+    template<typename Op, typename RV, typename Index, std::size_t N, typename... Args>
+    inline RV apply(const std::array<Index, N>& idx, RV initial, Args&&... args) const
+    {
+      eigen_assert(N >= m_numIndices && "Can only apply symmetry group to objects that have at least the required amount of indices.");
+      for (std::size_t i = 0; i < size(); i++)
+        initial = Op::run(h_permute(i, idx, typename internal::gen_numeric_list<int, N>::type()), m_elements[i].flags, initial, std::forward<Args>(args)...);
+      return initial;
+    }
+
+    template<typename Op, typename RV, typename Index, typename... Args>
+    inline RV apply(const std::vector<Index>& idx, RV initial, Args&&... args) const
+    {
+      eigen_assert(idx.size() >= m_numIndices && "Can only apply symmetry group to objects that have at least the required amount of indices.");
+      for (std::size_t i = 0; i < size(); i++)
+        initial = Op::run(h_permute(i, idx), m_elements[i].flags, initial, std::forward<Args>(args)...);
+      return initial;
+    }
+
+    inline int globalFlags() const { return m_globalFlags; }
+    inline std::size_t size() const { return m_elements.size(); }
+
+    template<typename Tensor_, typename... IndexTypes>
+    inline internal::tensor_symmetry_value_setter<Tensor_, DynamicSGroup> operator()(Tensor_& tensor, typename Tensor_::Index firstIndex, IndexTypes... otherIndices) const
+    {
+      static_assert(sizeof...(otherIndices) + 1 == Tensor_::NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor.");
+      return operator()(tensor, std::array<typename Tensor_::Index, Tensor_::NumIndices>{{firstIndex, otherIndices...}});
+    }
+
+    template<typename Tensor_>
+    inline internal::tensor_symmetry_value_setter<Tensor_, DynamicSGroup> operator()(Tensor_& tensor, std::array<typename Tensor_::Index, Tensor_::NumIndices> const& indices) const
+    {
+      return internal::tensor_symmetry_value_setter<Tensor_, DynamicSGroup>(tensor, *this, indices);
+    }
+  private:
+    struct GroupElement {
+      std::vector<int> representation;
+      int flags;
+      bool isId() const
+      {
+        for (std::size_t i = 0; i < representation.size(); i++)
+          if (i != (size_t)representation[i])
+            return false;
+        return true;
+      }
+    };
+    struct Generator {
+      int one;
+      int two;
+      int flags;
+      constexpr inline Generator(int one_, int two_, int flags_) : one(one_), two(two_), flags(flags_) {}
+    };
+
+    std::size_t m_numIndices;
+    std::vector<GroupElement> m_elements;
+    std::vector<Generator> m_generators;
+    int m_globalFlags;
+
+    template<typename Index, std::size_t N, int... n>
+    inline std::array<Index, N> h_permute(std::size_t which, const std::array<Index, N>& idx, internal::numeric_list<int, n...>) const
+    {
+      return std::array<Index, N>{{ idx[n >= m_numIndices ? n : m_elements[which].representation[n]]... }};
+    }
+
+    template<typename Index>
+    inline std::vector<Index> h_permute(std::size_t which, std::vector<Index> idx) const
+    {
+      std::vector<Index> result;
+      result.reserve(idx.size());
+      for (auto k : m_elements[which].representation)
+        result.push_back(idx[k]);
+      for (std::size_t i = m_numIndices; i < idx.size(); i++)
+        result.push_back(idx[i]);
+      return result;
+    }
+
+    inline GroupElement ge(Generator const& g) const
+    {
+      GroupElement result;
+      result.representation.reserve(m_numIndices);
+      result.flags = g.flags;
+      for (std::size_t k = 0; k < m_numIndices; k++) {
+        if (k == (std::size_t)g.one)
+          result.representation.push_back(g.two);
+        else if (k == (std::size_t)g.two)
+          result.representation.push_back(g.one);
+        else
+          result.representation.push_back(int(k));
+      }
+      return result;
+    }
+
+    GroupElement mul(GroupElement, GroupElement) const;
+    inline GroupElement mul(Generator g1, GroupElement g2) const
+    {
+      return mul(ge(g1), g2);
+    }
+
+    inline GroupElement mul(GroupElement g1, Generator g2) const
+    {
+      return mul(g1, ge(g2));
+    }
+
+    inline GroupElement mul(Generator g1, Generator g2) const
+    {
+      return mul(ge(g1), ge(g2));
+    }
+
+    inline int findElement(GroupElement e) const
+    {
+      for (auto ee : m_elements) {
+        if (ee.representation == e.representation)
+          return ee.flags ^ e.flags;
+      }
+      return -1;
+    }
+
+    void updateGlobalFlags(int flagDiffOfSameGenerator);
+};
+
+// dynamic symmetry group that auto-adds the template parameters in the constructor
+template<typename... Gen>
+class DynamicSGroupFromTemplateArgs : public DynamicSGroup
+{
+  public:
+    inline DynamicSGroupFromTemplateArgs() : DynamicSGroup()
+    {
+      add_all(internal::type_list<Gen...>());
+    }
+    inline DynamicSGroupFromTemplateArgs(DynamicSGroupFromTemplateArgs const& other) : DynamicSGroup(other) { }
+    inline DynamicSGroupFromTemplateArgs(DynamicSGroupFromTemplateArgs&& other) : DynamicSGroup(other) { }
+    inline DynamicSGroupFromTemplateArgs<Gen...>& operator=(const DynamicSGroupFromTemplateArgs<Gen...>& o) { DynamicSGroup::operator=(o); return *this; }
+    inline DynamicSGroupFromTemplateArgs<Gen...>& operator=(DynamicSGroupFromTemplateArgs<Gen...>&& o) { DynamicSGroup::operator=(o); return *this; }
+  
+  private:
+    template<typename Gen1, typename... GenNext>
+    inline void add_all(internal::type_list<Gen1, GenNext...>)
+    {
+      add(Gen1());
+      add_all(internal::type_list<GenNext...>());
+    }
+
+    inline void add_all(internal::type_list<>)
+    {
+    }
+};
+
+inline DynamicSGroup::GroupElement DynamicSGroup::mul(GroupElement g1, GroupElement g2) const
+{
+  eigen_internal_assert(g1.representation.size() == m_numIndices);
+  eigen_internal_assert(g2.representation.size() == m_numIndices);
+
+  GroupElement result;
+  result.representation.reserve(m_numIndices);
+  for (std::size_t i = 0; i < m_numIndices; i++) {
+    int v = g2.representation[g1.representation[i]];
+    eigen_assert(v >= 0);
+    result.representation.push_back(v);
+  }
+  result.flags = g1.flags ^ g2.flags;
+  return result;
+}
+
+inline void DynamicSGroup::add(int one, int two, int flags)
+{
+  eigen_assert(one >= 0);
+  eigen_assert(two >= 0);
+  eigen_assert(one != two);
+
+  if ((std::size_t)one >= m_numIndices || (std::size_t)two >= m_numIndices) {
+    std::size_t newNumIndices = (one > two) ? one : two + 1;
+    for (auto& gelem : m_elements) {
+      gelem.representation.reserve(newNumIndices);
+      for (std::size_t i = m_numIndices; i < newNumIndices; i++)
+        gelem.representation.push_back(i);
+    }
+    m_numIndices = newNumIndices;
+  }
+
+  Generator g{one, two, flags};
+  GroupElement e = ge(g);
+
+  /* special case for first generator */
+  if (m_elements.size() == 1) {
+    while (!e.isId()) {
+      m_elements.push_back(e);
+      e = mul(e, g);
+    }
+
+    if (e.flags > 0)
+      updateGlobalFlags(e.flags);
+
+    // only add in case we didn't have identity
+    if (m_elements.size() > 1)
+      m_generators.push_back(g);
+    return;
+  }
+
+  int p = findElement(e);
+  if (p >= 0) {
+    updateGlobalFlags(p);
+    return;
+  }
+
+  std::size_t coset_order = m_elements.size();
+  m_elements.push_back(e);
+  for (std::size_t i = 1; i < coset_order; i++)
+    m_elements.push_back(mul(m_elements[i], e));
+  m_generators.push_back(g);
+
+  std::size_t coset_rep = coset_order;
+  do {
+    for (auto g : m_generators) {
+      e = mul(m_elements[coset_rep], g);
+      p = findElement(e);
+      if (p < 0) {
+        // element not yet in group
+        m_elements.push_back(e);
+        for (std::size_t i = 1; i < coset_order; i++)
+          m_elements.push_back(mul(m_elements[i], e));
+      } else if (p > 0) {
+        updateGlobalFlags(p);
+      }
+    }
+    coset_rep += coset_order;
+  } while (coset_rep < m_elements.size());
+}
+
+inline void DynamicSGroup::updateGlobalFlags(int flagDiffOfSameGenerator)
+{
+    switch (flagDiffOfSameGenerator) {
+      case 0:
+      default:
+        // nothing happened
+        break;
+      case NegationFlag:
+        // every element is it's own negative => whole tensor is zero
+        m_globalFlags |= GlobalZeroFlag;
+        break;
+      case ConjugationFlag:
+        // every element is it's own conjugate => whole tensor is real
+        m_globalFlags |= GlobalRealFlag;
+        break;
+      case (NegationFlag | ConjugationFlag):
+        // every element is it's own negative conjugate => whole tensor is imaginary
+        m_globalFlags |= GlobalImagFlag;
+        break;
+      /* NOTE:
+       *   since GlobalZeroFlag == GlobalRealFlag | GlobalImagFlag, if one generator
+       *   causes the tensor to be real and the next one to be imaginary, this will
+       *   trivially give the correct result
+       */
+    }
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSORSYMMETRY_DYNAMICSYMMETRY_H
+
+/*
+ * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle;
+ */
diff --git a/src/EigenUnsupported/CXX11/src/TensorSymmetry/StaticSymmetry.h b/src/EigenUnsupported/CXX11/src/TensorSymmetry/StaticSymmetry.h
new file mode 100644
index 0000000..942293b
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/TensorSymmetry/StaticSymmetry.h
@@ -0,0 +1,236 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSORSYMMETRY_STATICSYMMETRY_H
+#define EIGEN_CXX11_TENSORSYMMETRY_STATICSYMMETRY_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename list> struct tensor_static_symgroup_permutate;
+
+template<int... nn>
+struct tensor_static_symgroup_permutate<numeric_list<int, nn...>>
+{
+  constexpr static std::size_t N = sizeof...(nn);
+
+  template<typename T>
+  constexpr static inline std::array<T, N> run(const std::array<T, N>& indices)
+  {
+    return {{indices[nn]...}};
+  }
+};
+
+template<typename indices_, int flags_>
+struct tensor_static_symgroup_element
+{
+  typedef indices_ indices;
+  constexpr static int flags = flags_;
+};
+
+template<typename Gen, int N>
+struct tensor_static_symgroup_element_ctor
+{
+  typedef tensor_static_symgroup_element<
+    typename gen_numeric_list_swapped_pair<int, N, Gen::One, Gen::Two>::type,
+    Gen::Flags
+  > type;
+};
+
+template<int N>
+struct tensor_static_symgroup_identity_ctor
+{
+  typedef tensor_static_symgroup_element<
+    typename gen_numeric_list<int, N>::type,
+    0
+  > type;
+};
+
+template<typename iib>
+struct tensor_static_symgroup_multiply_helper
+{
+  template<int... iia>
+  constexpr static inline numeric_list<int, get<iia, iib>::value...> helper(numeric_list<int, iia...>) {
+    return numeric_list<int, get<iia, iib>::value...>();
+  }
+};
+
+template<typename A, typename B>
+struct tensor_static_symgroup_multiply
+{
+  private:
+    typedef typename A::indices iia;
+    typedef typename B::indices iib;
+    constexpr static int ffa = A::flags;
+    constexpr static int ffb = B::flags;
+  
+  public:
+    static_assert(iia::count == iib::count, "Cannot multiply symmetry elements with different number of indices.");
+
+    typedef tensor_static_symgroup_element<
+      decltype(tensor_static_symgroup_multiply_helper<iib>::helper(iia())),
+      ffa ^ ffb
+    > type;
+};
+
+template<typename A, typename B>
+struct tensor_static_symgroup_equality
+{
+    typedef typename A::indices iia;
+    typedef typename B::indices iib;
+    constexpr static int ffa = A::flags;
+    constexpr static int ffb = B::flags;
+    static_assert(iia::count == iib::count, "Cannot compare symmetry elements with different number of indices.");
+
+    constexpr static bool value = is_same<iia, iib>::value;
+
+  private:
+    /* this should be zero if they are identical, or else the tensor
+     * will be forced to be pure real, pure imaginary or even pure zero
+     */
+    constexpr static int flags_cmp_ = ffa ^ ffb;
+
+    /* either they are not equal, then we don't care whether the flags
+     * match, or they are equal, and then we have to check
+     */
+    constexpr static bool is_zero      = value && flags_cmp_ == NegationFlag;
+    constexpr static bool is_real      = value && flags_cmp_ == ConjugationFlag;
+    constexpr static bool is_imag      = value && flags_cmp_ == (NegationFlag | ConjugationFlag);
+
+  public:
+    constexpr static int global_flags = 
+      (is_real ? GlobalRealFlag : 0) |
+      (is_imag ? GlobalImagFlag : 0) |
+      (is_zero ? GlobalZeroFlag : 0);
+};
+
+template<std::size_t NumIndices, typename... Gen>
+struct tensor_static_symgroup
+{
+  typedef StaticSGroup<Gen...> type;
+  constexpr static std::size_t size = type::static_size;
+};
+
+template<typename Index, std::size_t N, int... ii, int... jj>
+constexpr static inline std::array<Index, N> tensor_static_symgroup_index_permute(std::array<Index, N> idx, internal::numeric_list<int, ii...>, internal::numeric_list<int, jj...>)
+{
+  return {{ idx[ii]..., idx[jj]... }};
+}
+
+template<typename Index, int... ii>
+static inline std::vector<Index> tensor_static_symgroup_index_permute(std::vector<Index> idx, internal::numeric_list<int, ii...>)
+{
+  std::vector<Index> result{{ idx[ii]... }};
+  std::size_t target_size = idx.size();
+  for (std::size_t i = result.size(); i < target_size; i++)
+    result.push_back(idx[i]);
+  return result;
+}
+
+template<typename T> struct tensor_static_symgroup_do_apply;
+
+template<typename first, typename... next>
+struct tensor_static_symgroup_do_apply<internal::type_list<first, next...>>
+{
+  template<typename Op, typename RV, std::size_t SGNumIndices, typename Index, std::size_t NumIndices, typename... Args>
+  static inline RV run(const std::array<Index, NumIndices>& idx, RV initial, Args&&... args)
+  {
+    static_assert(NumIndices >= SGNumIndices, "Can only apply symmetry group to objects that have at least the required amount of indices.");
+    typedef typename internal::gen_numeric_list<int, NumIndices - SGNumIndices, SGNumIndices>::type remaining_indices;
+    initial = Op::run(tensor_static_symgroup_index_permute(idx, typename first::indices(), remaining_indices()), first::flags, initial, std::forward<Args>(args)...);
+    return tensor_static_symgroup_do_apply<internal::type_list<next...>>::template run<Op, RV, SGNumIndices>(idx, initial, args...);
+  }
+
+  template<typename Op, typename RV, std::size_t SGNumIndices, typename Index, typename... Args>
+  static inline RV run(const std::vector<Index>& idx, RV initial, Args&&... args)
+  {
+    eigen_assert(idx.size() >= SGNumIndices && "Can only apply symmetry group to objects that have at least the required amount of indices.");
+    initial = Op::run(tensor_static_symgroup_index_permute(idx, typename first::indices()), first::flags, initial, std::forward<Args>(args)...);
+    return tensor_static_symgroup_do_apply<internal::type_list<next...>>::template run<Op, RV, SGNumIndices>(idx, initial, args...);
+  }
+};
+
+template<EIGEN_TPL_PP_SPEC_HACK_DEF(typename, empty)>
+struct tensor_static_symgroup_do_apply<internal::type_list<EIGEN_TPL_PP_SPEC_HACK_USE(empty)>>
+{
+  template<typename Op, typename RV, std::size_t SGNumIndices, typename Index, std::size_t NumIndices, typename... Args>
+  static inline RV run(const std::array<Index, NumIndices>&, RV initial, Args&&...)
+  {
+    // do nothing
+    return initial;
+  }
+
+  template<typename Op, typename RV, std::size_t SGNumIndices, typename Index, typename... Args>
+  static inline RV run(const std::vector<Index>&, RV initial, Args&&...)
+  {
+    // do nothing
+    return initial;
+  }
+};
+
+} // end namespace internal
+
+template<typename... Gen>
+class StaticSGroup
+{
+    constexpr static std::size_t NumIndices = internal::tensor_symmetry_num_indices<Gen...>::value;
+    typedef internal::group_theory::enumerate_group_elements<
+      internal::tensor_static_symgroup_multiply,
+      internal::tensor_static_symgroup_equality,
+      typename internal::tensor_static_symgroup_identity_ctor<NumIndices>::type,
+      internal::type_list<typename internal::tensor_static_symgroup_element_ctor<Gen, NumIndices>::type...>
+    > group_elements;
+    typedef typename group_elements::type ge;
+  public:
+    constexpr inline StaticSGroup() {}
+    constexpr inline StaticSGroup(const StaticSGroup<Gen...>&) {}
+    constexpr inline StaticSGroup(StaticSGroup<Gen...>&&) {}
+
+    template<typename Op, typename RV, typename Index, std::size_t N, typename... Args>
+    static inline RV apply(const std::array<Index, N>& idx, RV initial, Args&&... args)
+    {
+      return internal::tensor_static_symgroup_do_apply<ge>::template run<Op, RV, NumIndices>(idx, initial, args...);
+    }
+
+    template<typename Op, typename RV, typename Index, typename... Args>
+    static inline RV apply(const std::vector<Index>& idx, RV initial, Args&&... args)
+    {
+      eigen_assert(idx.size() == NumIndices);
+      return internal::tensor_static_symgroup_do_apply<ge>::template run<Op, RV, NumIndices>(idx, initial, args...);
+    }
+
+    constexpr static std::size_t static_size = ge::count;
+
+    constexpr static inline std::size_t size() {
+      return ge::count;
+    }
+    constexpr static inline int globalFlags() { return group_elements::global_flags; }
+
+    template<typename Tensor_, typename... IndexTypes>
+    inline internal::tensor_symmetry_value_setter<Tensor_, StaticSGroup<Gen...>> operator()(Tensor_& tensor, typename Tensor_::Index firstIndex, IndexTypes... otherIndices) const
+    {
+      static_assert(sizeof...(otherIndices) + 1 == Tensor_::NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor.");
+      return operator()(tensor, std::array<typename Tensor_::Index, Tensor_::NumIndices>{{firstIndex, otherIndices...}});
+    }
+
+    template<typename Tensor_>
+    inline internal::tensor_symmetry_value_setter<Tensor_, StaticSGroup<Gen...>> operator()(Tensor_& tensor, std::array<typename Tensor_::Index, Tensor_::NumIndices> const& indices) const
+    {
+      return internal::tensor_symmetry_value_setter<Tensor_, StaticSGroup<Gen...>>(tensor, *this, indices);
+    }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSORSYMMETRY_STATICSYMMETRY_H
+
+/*
+ * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle;
+ */
diff --git a/src/EigenUnsupported/CXX11/src/TensorSymmetry/Symmetry.h b/src/EigenUnsupported/CXX11/src/TensorSymmetry/Symmetry.h
new file mode 100644
index 0000000..879d6cd
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/TensorSymmetry/Symmetry.h
@@ -0,0 +1,338 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSORSYMMETRY_SYMMETRY_H
+#define EIGEN_CXX11_TENSORSYMMETRY_SYMMETRY_H
+
+namespace Eigen {
+
+enum {
+  NegationFlag           = 0x01,
+  ConjugationFlag        = 0x02
+};
+
+enum {
+  GlobalRealFlag         = 0x01,
+  GlobalImagFlag         = 0x02,
+  GlobalZeroFlag         = 0x03
+};
+
+namespace internal {
+
+template<std::size_t NumIndices, typename... Sym>                   struct tensor_symmetry_pre_analysis;
+template<std::size_t NumIndices, typename... Sym>                   struct tensor_static_symgroup;
+template<bool instantiate, std::size_t NumIndices, typename... Sym> struct tensor_static_symgroup_if;
+template<typename Tensor_> struct tensor_symmetry_calculate_flags;
+template<typename Tensor_> struct tensor_symmetry_assign_value;
+template<typename... Sym> struct tensor_symmetry_num_indices;
+
+} // end namespace internal
+
+template<int One_, int Two_>
+struct Symmetry
+{
+  static_assert(One_ != Two_, "Symmetries must cover distinct indices.");
+  constexpr static int One = One_;
+  constexpr static int Two = Two_;
+  constexpr static int Flags = 0;
+};
+
+template<int One_, int Two_>
+struct AntiSymmetry
+{
+  static_assert(One_ != Two_, "Symmetries must cover distinct indices.");
+  constexpr static int One = One_;
+  constexpr static int Two = Two_;
+  constexpr static int Flags = NegationFlag;
+};
+
+template<int One_, int Two_>
+struct Hermiticity
+{
+  static_assert(One_ != Two_, "Symmetries must cover distinct indices.");
+  constexpr static int One = One_;
+  constexpr static int Two = Two_;
+  constexpr static int Flags = ConjugationFlag;
+};
+
+template<int One_, int Two_>
+struct AntiHermiticity
+{
+  static_assert(One_ != Two_, "Symmetries must cover distinct indices.");
+  constexpr static int One = One_;
+  constexpr static int Two = Two_;
+  constexpr static int Flags = ConjugationFlag | NegationFlag;
+};
+
+/** \class DynamicSGroup
+  * \ingroup TensorSymmetry_Module
+  *
+  * \brief Dynamic symmetry group
+  *
+  * The %DynamicSGroup class represents a symmetry group that need not be known at
+  * compile time. It is useful if one wants to support arbitrary run-time defineable
+  * symmetries for tensors, but it is also instantiated if a symmetry group is defined
+  * at compile time that would be either too large for the compiler to reasonably
+  * generate (using templates to calculate this at compile time is very inefficient)
+  * or that the compiler could generate the group but that it wouldn't make sense to
+  * unroll the loop for setting coefficients anymore.
+  */
+class DynamicSGroup;
+
+/** \internal
+  *
+  * \class DynamicSGroupFromTemplateArgs
+  * \ingroup TensorSymmetry_Module
+  *
+  * \brief Dynamic symmetry group, initialized from template arguments
+  *
+  * This class is a child class of DynamicSGroup. It uses the template arguments
+  * specified to initialize itself.
+  */
+template<typename... Gen>
+class DynamicSGroupFromTemplateArgs;
+
+/** \class StaticSGroup
+  * \ingroup TensorSymmetry_Module
+  *
+  * \brief Static symmetry group
+  *
+  * This class represents a symmetry group that is known and resolved completely
+  * at compile time. Ideally, no run-time penalty is incurred compared to the
+  * manual unrolling of the symmetry.
+  *
+  * <b><i>CAUTION:</i></b>
+  *
+  * Do not use this class directly for large symmetry groups. The compiler
+  * may run into a limit, or segfault or in the very least will take a very,
+  * very, very long time to compile the code. Use the SGroup class instead
+  * if you want a static group. That class contains logic that will
+  * automatically select the DynamicSGroup class instead if the symmetry
+  * group becomes too large. (In that case, unrolling may not even be
+  * beneficial.)
+  */
+template<typename... Gen>
+class StaticSGroup;
+
+/** \class SGroup
+  * \ingroup TensorSymmetry_Module
+  *
+  * \brief Symmetry group, initialized from template arguments
+  *
+  * This class represents a symmetry group whose generators are already
+  * known at compile time. It may or may not be resolved at compile time,
+  * depending on the estimated size of the group.
+  *
+  * \sa StaticSGroup
+  * \sa DynamicSGroup
+  */
+template<typename... Gen>
+class SGroup : public internal::tensor_symmetry_pre_analysis<internal::tensor_symmetry_num_indices<Gen...>::value, Gen...>::root_type
+{
+  public:
+    constexpr static std::size_t NumIndices = internal::tensor_symmetry_num_indices<Gen...>::value;
+    typedef typename internal::tensor_symmetry_pre_analysis<NumIndices, Gen...>::root_type Base;
+
+    // make standard constructors + assignment operators public
+    inline SGroup() : Base() { }
+    inline SGroup(const SGroup<Gen...>& other) : Base(other) { }
+    inline SGroup(SGroup<Gen...>&& other) : Base(other) { }
+    inline SGroup<Gen...>& operator=(const SGroup<Gen...>& other) { Base::operator=(other); return *this; }
+    inline SGroup<Gen...>& operator=(SGroup<Gen...>&& other) { Base::operator=(other); return *this; }
+
+    // all else is defined in the base class
+};
+
+namespace internal {
+
+template<typename... Sym> struct tensor_symmetry_num_indices
+{
+  constexpr static std::size_t value = 1;
+};
+
+template<int One_, int Two_, typename... Sym> struct tensor_symmetry_num_indices<Symmetry<One_, Two_>, Sym...>
+{
+private:
+  constexpr static std::size_t One = static_cast<std::size_t>(One_);
+  constexpr static std::size_t Two = static_cast<std::size_t>(Two_);
+  constexpr static std::size_t Three = tensor_symmetry_num_indices<Sym...>::value;
+
+  // don't use std::max, since it's not constexpr until C++14...
+  constexpr static std::size_t maxOneTwoPlusOne = ((One > Two) ? One : Two) + 1;
+public:
+  constexpr static std::size_t value = (maxOneTwoPlusOne > Three) ? maxOneTwoPlusOne : Three;
+};
+
+template<int One_, int Two_, typename... Sym> struct tensor_symmetry_num_indices<AntiSymmetry<One_, Two_>, Sym...>
+  : public tensor_symmetry_num_indices<Symmetry<One_, Two_>, Sym...> {};
+template<int One_, int Two_, typename... Sym> struct tensor_symmetry_num_indices<Hermiticity<One_, Two_>, Sym...>
+  : public tensor_symmetry_num_indices<Symmetry<One_, Two_>, Sym...> {};
+template<int One_, int Two_, typename... Sym> struct tensor_symmetry_num_indices<AntiHermiticity<One_, Two_>, Sym...>
+  : public tensor_symmetry_num_indices<Symmetry<One_, Two_>, Sym...> {};
+
+/** \internal
+  *
+  * \class tensor_symmetry_pre_analysis
+  * \ingroup TensorSymmetry_Module
+  *
+  * \brief Pre-select whether to use a static or dynamic symmetry group
+  *
+  * When a symmetry group could in principle be determined at compile time,
+  * this template implements the logic whether to actually do that or whether
+  * to rather defer that to runtime.
+  *
+  * The logic is as follows:
+  * <dl>
+  * <dt><b>No generators (trivial symmetry):</b></dt>
+  * <dd>Use a trivial static group. Ideally, this has no performance impact
+  *     compared to not using symmetry at all. In practice, this might not
+  *     be the case.</dd>
+  * <dt><b>More than 4 generators:</b></dt>
+  * <dd>Calculate the group at run time, it is likely far too large for the
+  *     compiler to be able to properly generate it in a realistic time.</dd>
+  * <dt><b>Up to and including 4 generators:</b></dt>
+  * <dd>Actually enumerate all group elements, but then check how many there
+  *     are. If there are more than 16, it is unlikely that unrolling the
+  *     loop (as is done in the static compile-time case) is sensible, so
+  *     use a dynamic group instead. If there are at most 16 elements, actually
+  *     use that static group. Note that the largest group with 4 generators
+  *     still compiles with reasonable resources.</dd>
+  * </dl>
+  *
+  * Note: Example compile time performance with g++-4.6 on an Intenl Core i5-3470
+  *       with 16 GiB RAM (all generators non-redundant and the subgroups don't
+  *       factorize):
+  *
+  *          # Generators          -O0 -ggdb               -O2
+  *          -------------------------------------------------------------------
+  *          1                 0.5 s  /   250 MiB     0.45s /   230 MiB
+  *          2                 0.5 s  /   260 MiB     0.5 s /   250 MiB
+  *          3                 0.65s  /   310 MiB     0.62s /   310 MiB
+  *          4                 2.2 s  /   860 MiB     1.7 s /   770 MiB
+  *          5               130   s  / 13000 MiB   120   s / 11000 MiB
+  *
+  * It is clear that everything is still very efficient up to 4 generators, then
+  * the memory and CPU requirements become unreasonable. Thus we only instantiate
+  * the template group theory logic if the number of generators supplied is 4 or
+  * lower, otherwise this will be forced to be done during runtime, where the
+  * algorithm is reasonably fast.
+  */
+template<std::size_t NumIndices>
+struct tensor_symmetry_pre_analysis<NumIndices>
+{
+  typedef StaticSGroup<> root_type;
+};
+
+template<std::size_t NumIndices, typename Gen_, typename... Gens_>
+struct tensor_symmetry_pre_analysis<NumIndices, Gen_, Gens_...>
+{
+  constexpr static std::size_t max_static_generators = 4;
+  constexpr static std::size_t max_static_elements = 16;
+  typedef tensor_static_symgroup_if<(sizeof...(Gens_) + 1 <= max_static_generators), NumIndices, Gen_, Gens_...> helper;
+  constexpr static std::size_t possible_size = helper::size;
+
+  typedef typename conditional<
+    possible_size == 0 || possible_size >= max_static_elements,
+    DynamicSGroupFromTemplateArgs<Gen_, Gens_...>,
+    typename helper::type
+  >::type root_type;
+};
+
+template<bool instantiate, std::size_t NumIndices, typename... Gens>
+struct tensor_static_symgroup_if
+{
+  constexpr static std::size_t size = 0;
+  typedef void type;
+};
+
+template<std::size_t NumIndices, typename... Gens>
+struct tensor_static_symgroup_if<true, NumIndices, Gens...> : tensor_static_symgroup<NumIndices, Gens...> {};
+
+template<typename Tensor_>
+struct tensor_symmetry_assign_value
+{
+  typedef typename Tensor_::Index Index;
+  typedef typename Tensor_::Scalar Scalar;
+  constexpr static std::size_t NumIndices = Tensor_::NumIndices;
+
+  static inline int run(const std::array<Index, NumIndices>& transformed_indices, int transformation_flags, int dummy, Tensor_& tensor, const Scalar& value_)
+  {
+    Scalar value(value_);
+    if (transformation_flags & ConjugationFlag)
+      value = numext::conj(value);
+    if (transformation_flags & NegationFlag)
+      value = -value;
+    tensor.coeffRef(transformed_indices) = value;
+    return dummy;
+  }
+};
+
+template<typename Tensor_>
+struct tensor_symmetry_calculate_flags
+{
+  typedef typename Tensor_::Index Index;
+  constexpr static std::size_t NumIndices = Tensor_::NumIndices;
+
+  static inline int run(const std::array<Index, NumIndices>& transformed_indices, int transform_flags, int current_flags, const std::array<Index, NumIndices>& orig_indices)
+  {
+    if (transformed_indices == orig_indices) {
+      if (transform_flags & (ConjugationFlag | NegationFlag))
+        return current_flags | GlobalImagFlag; // anti-hermitian diagonal
+      else if (transform_flags & ConjugationFlag)
+        return current_flags | GlobalRealFlag; // hermitian diagonal
+      else if (transform_flags & NegationFlag)
+        return current_flags | GlobalZeroFlag; // anti-symmetric diagonal
+    }
+    return current_flags;
+  }
+};
+
+template<typename Tensor_, typename Symmetry_, int Flags = 0>
+class tensor_symmetry_value_setter
+{
+  public:
+    typedef typename Tensor_::Index Index;
+    typedef typename Tensor_::Scalar Scalar;
+    constexpr static std::size_t NumIndices = Tensor_::NumIndices;
+
+    inline tensor_symmetry_value_setter(Tensor_& tensor, Symmetry_ const& symmetry, std::array<Index, NumIndices> const& indices)
+      : m_tensor(tensor), m_symmetry(symmetry), m_indices(indices) { }
+
+    inline tensor_symmetry_value_setter<Tensor_, Symmetry_, Flags>& operator=(Scalar const& value)
+    {
+      doAssign(value);
+      return *this;
+    }
+  private:
+    Tensor_& m_tensor;
+    Symmetry_ m_symmetry;
+    std::array<Index, NumIndices> m_indices;
+
+    inline void doAssign(Scalar const& value)
+    {
+      #ifdef EIGEN_TENSOR_SYMMETRY_CHECK_VALUES
+        int value_flags = m_symmetry.template apply<internal::tensor_symmetry_calculate_flags<Tensor_>, int>(m_indices, m_symmetry.globalFlags(), m_indices);
+        if (value_flags & GlobalRealFlag)
+          eigen_assert(numext::imag(value) == 0);
+        if (value_flags & GlobalImagFlag)
+          eigen_assert(numext::real(value) == 0);
+      #endif
+      m_symmetry.template apply<internal::tensor_symmetry_assign_value<Tensor_>, int>(m_indices, 0, m_tensor, value);
+    }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSORSYMMETRY_SYMMETRY_H
+
+/*
+ * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle;
+ */
diff --git a/src/EigenUnsupported/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h b/src/EigenUnsupported/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h
new file mode 100644
index 0000000..54bf9db
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h
@@ -0,0 +1,669 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSORSYMMETRY_TEMPLATEGROUPTHEORY_H
+#define EIGEN_CXX11_TENSORSYMMETRY_TEMPLATEGROUPTHEORY_H
+
+namespace Eigen {
+
+namespace internal {
+
+namespace group_theory {
+
+/** \internal
+  * \file CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h
+  * This file contains C++ templates that implement group theory algorithms.
+  *
+  * The algorithms allow for a compile-time analysis of finite groups.
+  *
+  * Currently only Dimino's algorithm is implemented, which returns a list
+  * of all elements in a group given a set of (possibly redundant) generators.
+  * (One could also do that with the so-called orbital algorithm, but that
+  * is much more expensive and usually has no advantages.)
+  */
+
+/**********************************************************************
+ *                "Ok kid, here is where it gets complicated."
+ *                         - Amelia Pond in the "Doctor Who" episode
+ *                           "The Big Bang"
+ *
+ * Dimino's algorithm
+ * ==================
+ *
+ * The following is Dimino's algorithm in sequential form:
+ *
+ * Input: identity element, list of generators, equality check,
+ *        multiplication operation
+ * Output: list of group elements
+ *
+ * 1. add identity element
+ * 2. remove identities from list of generators
+ * 3. add all powers of first generator that aren't the
+ *    identity element
+ * 4. go through all remaining generators:
+ *        a. if generator is already in the list of elements
+ *                -> do nothing
+ *        b. otherwise
+ *                i.   remember current # of elements
+ *                     (i.e. the size of the current subgroup)
+ *                ii.  add all current elements (which includes
+ *                     the identity) each multiplied from right
+ *                     with the current generator to the group
+ *                iii. add all remaining cosets that are generated
+ *                     by products of the new generator with itself
+ *                     and all other generators seen so far
+ *
+ * In functional form, this is implemented as a long set of recursive
+ * templates that have a complicated relationship.
+ *
+ * The main interface for Dimino's algorithm is the template
+ * enumerate_group_elements. All lists are implemented as variadic
+ * type_list<typename...> and numeric_list<typename = int, int...>
+ * templates.
+ *
+ * 'Calling' templates is usually done via typedefs.
+ *
+ * This algorithm is an extended version of the basic version. The
+ * extension consists in the fact that each group element has a set
+ * of flags associated with it. Multiplication of two group elements
+ * with each other results in a group element whose flags are the
+ * XOR of the flags of the previous elements. Each time the algorithm
+ * notices that a group element it just calculated is already in the
+ * list of current elements, the flags of both will be compared and
+ * added to the so-called 'global flags' of the group.
+ *
+ * The rationale behind this extension is that this allows not only
+ * for the description of symmetries between tensor indices, but
+ * also allows for the description of hermiticity, antisymmetry and
+ * antihermiticity. Negation and conjugation each are specific bit
+ * in the flags value and if two different ways to reach a group
+ * element lead to two different flags, this poses a constraint on
+ * the allowed values of the resulting tensor. For example, if a
+ * group element is reach both with and without the conjugation
+ * flags, it is clear that the resulting tensor has to be real.
+ *
+ * Note that this flag mechanism is quite generic and may have other
+ * uses beyond tensor properties.
+ *
+ * IMPORTANT: 
+ *     This algorithm assumes the group to be finite. If you try to
+ *     run it with a group that's infinite, the algorithm will only
+ *     terminate once you hit a compiler limit (max template depth).
+ *     Also note that trying to use this implementation to create a
+ *     very large group will probably either make you hit the same
+ *     limit, cause the compiler to segfault or at the very least
+ *     take a *really* long time (hours, days, weeks - sic!) to
+ *     compile. It is not recommended to plug in more than 4
+ *     generators, unless they are independent of each other.
+ */
+
+/** \internal
+  *
+  * \class strip_identities
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Cleanse a list of group elements of the identity element
+  *
+  * This template is used to make a first pass through all initial
+  * generators of Dimino's algorithm and remove the identity
+  * elements.
+  *
+  * \sa enumerate_group_elements
+  */
+template<template<typename, typename> class Equality, typename id, typename L> struct strip_identities;
+
+template<
+  template<typename, typename> class Equality,
+  typename id,
+  typename t,
+  typename... ts
+>
+struct strip_identities<Equality, id, type_list<t, ts...>>
+{
+  typedef typename conditional<
+    Equality<id, t>::value,
+    typename strip_identities<Equality, id, type_list<ts...>>::type,
+    typename concat<type_list<t>, typename strip_identities<Equality, id, type_list<ts...>>::type>::type
+  >::type type;
+  constexpr static int global_flags = Equality<id, t>::global_flags | strip_identities<Equality, id, type_list<ts...>>::global_flags;
+};
+
+template<
+  template<typename, typename> class Equality,
+  typename id
+  EIGEN_TPL_PP_SPEC_HACK_DEFC(typename, ts)
+>
+struct strip_identities<Equality, id, type_list<EIGEN_TPL_PP_SPEC_HACK_USE(ts)>>
+{
+  typedef type_list<> type;
+  constexpr static int global_flags = 0;
+};
+
+/** \internal
+  *
+  * \class dimino_first_step_elements_helper 
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Recursive template that adds powers of the first generator to the list of group elements
+  *
+  * This template calls itself recursively to add powers of the first
+  * generator to the list of group elements. It stops if it reaches
+  * the identity element again.
+  *
+  * \sa enumerate_group_elements, dimino_first_step_elements
+  */
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename g,
+  typename current_element,
+  typename elements,
+  bool dont_add_current_element   // = false
+>
+struct dimino_first_step_elements_helper
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+  : // recursive inheritance is too difficult for Doxygen
+  public dimino_first_step_elements_helper<
+    Multiply,
+    Equality,
+    id,
+    g,
+    typename Multiply<current_element, g>::type,
+    typename concat<elements, type_list<current_element>>::type,
+    Equality<typename Multiply<current_element, g>::type, id>::value
+  > {};
+
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename g,
+  typename current_element,
+  typename elements
+>
+struct dimino_first_step_elements_helper<Multiply, Equality, id, g, current_element, elements, true>
+#endif // EIGEN_PARSED_BY_DOXYGEN
+{
+  typedef elements type;
+  constexpr static int global_flags = Equality<current_element, id>::global_flags;
+};
+
+/** \internal
+  *
+  * \class dimino_first_step_elements
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Add all powers of the first generator to the list of group elements
+  *
+  * This template takes the first non-identity generator and generates the initial
+  * list of elements which consists of all powers of that generator. For a group
+  * with just one generated, it would be enumerated after this.
+  *
+  * \sa enumerate_group_elements
+  */
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename generators
+>
+struct dimino_first_step_elements
+{
+  typedef typename get<0, generators>::type first_generator;
+  typedef typename skip<1, generators>::type next_generators;
+  typedef type_list<first_generator> generators_done;
+
+  typedef dimino_first_step_elements_helper<
+    Multiply,
+    Equality,
+    id,
+    first_generator,
+    first_generator,
+    type_list<id>,
+    false
+  > helper;
+  typedef typename helper::type type;
+  constexpr static int global_flags = helper::global_flags;
+};
+
+/** \internal
+  *
+  * \class dimino_get_coset_elements
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Generate all elements of a specific coset
+  *
+  * This template generates all the elements of a specific coset by
+  * multiplying all elements in the given subgroup with the new
+  * coset representative. Note that the first element of the
+  * subgroup is always the identity element, so the first element of
+  * the result of this template is going to be the coset
+  * representative itself.
+  *
+  * Note that this template accepts an additional boolean parameter
+  * that specifies whether to actually generate the coset (true) or
+  * just return an empty list (false).
+  *
+  * \sa enumerate_group_elements, dimino_add_cosets_for_rep
+  */
+template<
+  template<typename, typename> class Multiply,
+  typename sub_group_elements,
+  typename new_coset_rep,
+  bool generate_coset      // = true
+>
+struct dimino_get_coset_elements
+{
+  typedef typename apply_op_from_right<Multiply, new_coset_rep, sub_group_elements>::type type;
+};
+
+template<
+  template<typename, typename> class Multiply,
+  typename sub_group_elements,
+  typename new_coset_rep
+>
+struct dimino_get_coset_elements<Multiply, sub_group_elements, new_coset_rep, false>
+{
+  typedef type_list<> type;
+};
+
+/** \internal
+  *
+  * \class dimino_add_cosets_for_rep
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Recursive template for adding coset spaces
+  *
+  * This template multiplies the coset representative with a generator
+  * from the list of previous generators. If the new element is not in
+  * the group already, it adds the corresponding coset. Finally it
+  * proceeds to call itself with the next generator from the list.
+  *
+  * \sa enumerate_group_elements, dimino_add_all_coset_spaces
+  */
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename sub_group_elements,
+  typename elements,
+  typename generators,
+  typename rep_element,
+  int sub_group_size
+>
+struct dimino_add_cosets_for_rep;
+
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename sub_group_elements,
+  typename elements,
+  typename g,
+  typename... gs,
+  typename rep_element,
+  int sub_group_size
+>
+struct dimino_add_cosets_for_rep<Multiply, Equality, id, sub_group_elements, elements, type_list<g, gs...>, rep_element, sub_group_size>
+{
+  typedef typename Multiply<rep_element, g>::type new_coset_rep;
+  typedef contained_in_list_gf<Equality, new_coset_rep, elements> _cil;
+  constexpr static bool add_coset = !_cil::value;
+
+  typedef typename dimino_get_coset_elements<
+    Multiply,
+    sub_group_elements,
+    new_coset_rep,
+    add_coset
+  >::type coset_elements;
+
+  typedef dimino_add_cosets_for_rep<
+    Multiply,
+    Equality,
+    id,
+    sub_group_elements,
+    typename concat<elements, coset_elements>::type,
+    type_list<gs...>,
+    rep_element,
+    sub_group_size
+  > _helper;
+
+  typedef typename _helper::type type;
+  constexpr static int global_flags = _cil::global_flags | _helper::global_flags;
+
+  /* Note that we don't have to update global flags here, since
+   * we will only add these elements if they are not part of
+   * the group already. But that only happens if the coset rep
+   * is not already in the group, so the check for the coset rep
+   * will catch this.
+   */
+};
+
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename sub_group_elements,
+  typename elements
+  EIGEN_TPL_PP_SPEC_HACK_DEFC(typename, empty),
+  typename rep_element,
+  int sub_group_size
+>
+struct dimino_add_cosets_for_rep<Multiply, Equality, id, sub_group_elements, elements, type_list<EIGEN_TPL_PP_SPEC_HACK_USE(empty)>, rep_element, sub_group_size>
+{
+  typedef elements type;
+  constexpr static int global_flags = 0;
+};
+
+/** \internal
+  *
+  * \class dimino_add_all_coset_spaces
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Recursive template for adding all coset spaces for a new generator
+  *
+  * This template tries to go through the list of generators (with
+  * the help of the dimino_add_cosets_for_rep template) as long as
+  * it still finds elements that are not part of the group and add
+  * the corresponding cosets.
+  *
+  * \sa enumerate_group_elements, dimino_add_cosets_for_rep
+  */
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename sub_group_elements,
+  typename elements,
+  typename generators,
+  int sub_group_size,
+  int rep_pos,
+  bool stop_condition        // = false
+>
+struct dimino_add_all_coset_spaces
+{
+  typedef typename get<rep_pos, elements>::type rep_element;
+  typedef dimino_add_cosets_for_rep<
+    Multiply,
+    Equality,
+    id,
+    sub_group_elements,
+    elements,
+    generators,
+    rep_element,
+    sub_group_elements::count
+  > _ac4r;
+  typedef typename _ac4r::type new_elements;
+  
+  constexpr static int new_rep_pos = rep_pos + sub_group_elements::count;
+  constexpr static bool new_stop_condition = new_rep_pos >= new_elements::count;
+
+  typedef dimino_add_all_coset_spaces<
+    Multiply,
+    Equality,
+    id,
+    sub_group_elements,
+    new_elements,
+    generators,
+    sub_group_size,
+    new_rep_pos,
+    new_stop_condition
+  > _helper;
+
+  typedef typename _helper::type type;
+  constexpr static int global_flags = _helper::global_flags | _ac4r::global_flags;
+};
+
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename sub_group_elements,
+  typename elements,
+  typename generators,
+  int sub_group_size,
+  int rep_pos
+>
+struct dimino_add_all_coset_spaces<Multiply, Equality, id, sub_group_elements, elements, generators, sub_group_size, rep_pos, true>
+{
+  typedef elements type;
+  constexpr static int global_flags = 0;
+};
+
+/** \internal
+  *
+  * \class dimino_add_generator
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Enlarge the group by adding a new generator.
+  *
+  * It accepts a boolean parameter that determines if the generator is redundant,
+  * i.e. was already seen in the group. In that case, it reduces to a no-op.
+  *
+  * \sa enumerate_group_elements, dimino_add_all_coset_spaces
+  */
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename elements,
+  typename generators_done,
+  typename current_generator,
+  bool redundant          // = false
+>
+struct dimino_add_generator
+{
+  /* this template is only called if the generator is not redundant
+   * => all elements of the group multiplied with the new generator
+   *    are going to be new elements of the most trivial coset space
+   */
+  typedef typename apply_op_from_right<Multiply, current_generator, elements>::type multiplied_elements;
+  typedef typename concat<elements, multiplied_elements>::type new_elements;
+
+  constexpr static int rep_pos = elements::count;
+
+  typedef dimino_add_all_coset_spaces<
+    Multiply,
+    Equality,
+    id,
+    elements, // elements of previous subgroup
+    new_elements,
+    typename concat<generators_done, type_list<current_generator>>::type,
+    elements::count, // size of previous subgroup
+    rep_pos,
+    false // don't stop (because rep_pos >= new_elements::count is always false at this point)
+  > _helper;
+  typedef typename _helper::type type;
+  constexpr static int global_flags = _helper::global_flags;
+};
+
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename elements,
+  typename generators_done,
+  typename current_generator
+>
+struct dimino_add_generator<Multiply, Equality, id, elements, generators_done, current_generator, true>
+{
+  // redundant case
+  typedef elements type;
+  constexpr static int global_flags = 0;
+};
+
+/** \internal
+  *
+  * \class dimino_add_remaining_generators
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Recursive template that adds all remaining generators to a group
+  *
+  * Loop through the list of generators that remain and successively
+  * add them to the group.
+  *
+  * \sa enumerate_group_elements, dimino_add_generator
+  */
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename generators_done,
+  typename remaining_generators,
+  typename elements
+>
+struct dimino_add_remaining_generators
+{
+  typedef typename get<0, remaining_generators>::type first_generator;
+  typedef typename skip<1, remaining_generators>::type next_generators;
+
+  typedef contained_in_list_gf<Equality, first_generator, elements> _cil;
+
+  typedef dimino_add_generator<
+    Multiply,
+    Equality,
+    id,
+    elements,
+    generators_done,
+    first_generator,
+    _cil::value
+  > _helper;
+
+  typedef typename _helper::type new_elements;
+
+  typedef dimino_add_remaining_generators<
+    Multiply,
+    Equality,
+    id,
+    typename concat<generators_done, type_list<first_generator>>::type,
+    next_generators,
+    new_elements
+  > _next_iter;
+
+  typedef typename _next_iter::type type;
+  constexpr static int global_flags =
+    _cil::global_flags |
+    _helper::global_flags |
+    _next_iter::global_flags;
+};
+
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename generators_done,
+  typename elements
+>
+struct dimino_add_remaining_generators<Multiply, Equality, id, generators_done, type_list<>, elements>
+{
+  typedef elements type;
+  constexpr static int global_flags = 0;
+};
+
+/** \internal
+  *
+  * \class enumerate_group_elements_noid
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Helper template that implements group element enumeration
+  *
+  * This is a helper template that implements the actual enumeration
+  * of group elements. This has been split so that the list of
+  * generators can be cleansed of the identity element before
+  * performing the actual operation.
+  *
+  * \sa enumerate_group_elements
+  */
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename generators,
+  int initial_global_flags = 0
+>
+struct enumerate_group_elements_noid
+{
+  typedef dimino_first_step_elements<Multiply, Equality, id, generators> first_step;
+  typedef typename first_step::type first_step_elements;
+
+  typedef dimino_add_remaining_generators<
+    Multiply,
+    Equality,
+    id,
+    typename first_step::generators_done,
+    typename first_step::next_generators, // remaining_generators
+    typename first_step::type // first_step elements
+  > _helper;
+
+  typedef typename _helper::type type;
+  constexpr static int global_flags =
+    initial_global_flags |
+    first_step::global_flags |
+    _helper::global_flags;
+};
+
+// in case when no generators are specified
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  int initial_global_flags
+>
+struct enumerate_group_elements_noid<Multiply, Equality, id, type_list<>, initial_global_flags>
+{
+  typedef type_list<id> type;
+  constexpr static int global_flags = initial_global_flags;
+};
+
+/** \internal
+  *
+  * \class enumerate_group_elements
+  * \ingroup CXX11_TensorSymmetry_Module
+  *
+  * \brief Enumerate all elements in a finite group
+  *
+  * This template enumerates all elements in a finite group. It accepts
+  * the following template parameters:
+  *
+  * \tparam Multiply      The multiplication operation that multiplies two group elements
+  *                       with each other.
+  * \tparam Equality      The equality check operation that checks if two group elements
+  *                       are equal to another.
+  * \tparam id            The identity element
+  * \tparam _generators   A list of (possibly redundant) generators of the group
+  */
+template<
+  template<typename, typename> class Multiply,
+  template<typename, typename> class Equality,
+  typename id,
+  typename _generators
+>
+struct enumerate_group_elements
+  : public enumerate_group_elements_noid<
+      Multiply,
+      Equality,
+      id,
+      typename strip_identities<Equality, id, _generators>::type,
+      strip_identities<Equality, id, _generators>::global_flags
+    >
+{
+};
+
+} // end namespace group_theory
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSORSYMMETRY_TEMPLATEGROUPTHEORY_H
+
+/*
+ * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle;
+ */
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/Barrier.h b/src/EigenUnsupported/CXX11/src/ThreadPool/Barrier.h
new file mode 100644
index 0000000..e4c59dc
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/Barrier.h
@@ -0,0 +1,67 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Rasmus Munk Larsen <rmlarsen@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// Barrier is an object that allows one or more threads to wait until
+// Notify has been called a specified number of times.
+
+#ifndef EIGEN_CXX11_THREADPOOL_BARRIER_H
+#define EIGEN_CXX11_THREADPOOL_BARRIER_H
+
+namespace Eigen {
+
+class Barrier {
+ public:
+  Barrier(unsigned int count) : state_(count << 1), notified_(false) {
+    eigen_plain_assert(((count << 1) >> 1) == count);
+  }
+  ~Barrier() { eigen_plain_assert((state_ >> 1) == 0); }
+
+  void Notify() {
+    unsigned int v = state_.fetch_sub(2, std::memory_order_acq_rel) - 2;
+    if (v != 1) {
+      // Clear the lowest bit (waiter flag) and check that the original state
+      // value was not zero. If it was zero, it means that notify was called
+      // more times than the original count.
+      eigen_plain_assert(((v + 2) & ~1) != 0);
+      return;  // either count has not dropped to 0, or waiter is not waiting
+    }
+    std::unique_lock<std::mutex> l(mu_);
+    eigen_plain_assert(!notified_);
+    notified_ = true;
+    cv_.notify_all();
+  }
+
+  void Wait() {
+    unsigned int v = state_.fetch_or(1, std::memory_order_acq_rel);
+    if ((v >> 1) == 0) return;
+    std::unique_lock<std::mutex> l(mu_);
+    while (!notified_) {
+      cv_.wait(l);
+    }
+  }
+
+ private:
+  std::mutex mu_;
+  std::condition_variable cv_;
+  std::atomic<unsigned int> state_;  // low bit is waiter flag
+  bool notified_;
+};
+
+// Notification is an object that allows a user to to wait for another
+// thread to signal a notification that an event has occurred.
+//
+// Multiple threads can wait on the same Notification object,
+// but only one caller must call Notify() on the object.
+struct Notification : Barrier {
+  Notification() : Barrier(1){};
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_BARRIER_H
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/EventCount.h b/src/EigenUnsupported/CXX11/src/ThreadPool/EventCount.h
new file mode 100644
index 0000000..4549aa0
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/EventCount.h
@@ -0,0 +1,249 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Dmitry Vyukov <dvyukov@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_EVENTCOUNT_H_
+#define EIGEN_CXX11_THREADPOOL_EVENTCOUNT_H_
+
+namespace Eigen {
+
+// EventCount allows to wait for arbitrary predicates in non-blocking
+// algorithms. Think of condition variable, but wait predicate does not need to
+// be protected by a mutex. Usage:
+// Waiting thread does:
+//
+//   if (predicate)
+//     return act();
+//   EventCount::Waiter& w = waiters[my_index];
+//   ec.Prewait(&w);
+//   if (predicate) {
+//     ec.CancelWait(&w);
+//     return act();
+//   }
+//   ec.CommitWait(&w);
+//
+// Notifying thread does:
+//
+//   predicate = true;
+//   ec.Notify(true);
+//
+// Notify is cheap if there are no waiting threads. Prewait/CommitWait are not
+// cheap, but they are executed only if the preceding predicate check has
+// failed.
+//
+// Algorithm outline:
+// There are two main variables: predicate (managed by user) and state_.
+// Operation closely resembles Dekker mutual algorithm:
+// https://en.wikipedia.org/wiki/Dekker%27s_algorithm
+// Waiting thread sets state_ then checks predicate, Notifying thread sets
+// predicate then checks state_. Due to seq_cst fences in between these
+// operations it is guaranteed than either waiter will see predicate change
+// and won't block, or notifying thread will see state_ change and will unblock
+// the waiter, or both. But it can't happen that both threads don't see each
+// other changes, which would lead to deadlock.
+class EventCount {
+ public:
+  class Waiter;
+
+  EventCount(MaxSizeVector<Waiter>& waiters)
+      : state_(kStackMask), waiters_(waiters) {
+    eigen_plain_assert(waiters.size() < (1 << kWaiterBits) - 1);
+  }
+
+  ~EventCount() {
+    // Ensure there are no waiters.
+    eigen_plain_assert(state_.load() == kStackMask);
+  }
+
+  // Prewait prepares for waiting.
+  // After calling Prewait, the thread must re-check the wait predicate
+  // and then call either CancelWait or CommitWait.
+  void Prewait() {
+    uint64_t state = state_.load(std::memory_order_relaxed);
+    for (;;) {
+      CheckState(state);
+      uint64_t newstate = state + kWaiterInc;
+      CheckState(newstate);
+      if (state_.compare_exchange_weak(state, newstate,
+                                       std::memory_order_seq_cst))
+        return;
+    }
+  }
+
+  // CommitWait commits waiting after Prewait.
+  void CommitWait(Waiter* w) {
+    eigen_plain_assert((w->epoch & ~kEpochMask) == 0);
+    w->state = Waiter::kNotSignaled;
+    const uint64_t me = (w - &waiters_[0]) | w->epoch;
+    uint64_t state = state_.load(std::memory_order_seq_cst);
+    for (;;) {
+      CheckState(state, true);
+      uint64_t newstate;
+      if ((state & kSignalMask) != 0) {
+        // Consume the signal and return immidiately.
+        newstate = state - kWaiterInc - kSignalInc;
+      } else {
+        // Remove this thread from pre-wait counter and add to the waiter stack.
+        newstate = ((state & kWaiterMask) - kWaiterInc) | me;
+        w->next.store(state & (kStackMask | kEpochMask),
+                      std::memory_order_relaxed);
+      }
+      CheckState(newstate);
+      if (state_.compare_exchange_weak(state, newstate,
+                                       std::memory_order_acq_rel)) {
+        if ((state & kSignalMask) == 0) {
+          w->epoch += kEpochInc;
+          Park(w);
+        }
+        return;
+      }
+    }
+  }
+
+  // CancelWait cancels effects of the previous Prewait call.
+  void CancelWait() {
+    uint64_t state = state_.load(std::memory_order_relaxed);
+    for (;;) {
+      CheckState(state, true);
+      uint64_t newstate = state - kWaiterInc;
+      // We don't know if the thread was also notified or not,
+      // so we should not consume a signal unconditionaly.
+      // Only if number of waiters is equal to number of signals,
+      // we know that the thread was notified and we must take away the signal.
+      if (((state & kWaiterMask) >> kWaiterShift) ==
+          ((state & kSignalMask) >> kSignalShift))
+        newstate -= kSignalInc;
+      CheckState(newstate);
+      if (state_.compare_exchange_weak(state, newstate,
+                                       std::memory_order_acq_rel))
+        return;
+    }
+  }
+
+  // Notify wakes one or all waiting threads.
+  // Must be called after changing the associated wait predicate.
+  void Notify(bool notifyAll) {
+    std::atomic_thread_fence(std::memory_order_seq_cst);
+    uint64_t state = state_.load(std::memory_order_acquire);
+    for (;;) {
+      CheckState(state);
+      const uint64_t waiters = (state & kWaiterMask) >> kWaiterShift;
+      const uint64_t signals = (state & kSignalMask) >> kSignalShift;
+      // Easy case: no waiters.
+      if ((state & kStackMask) == kStackMask && waiters == signals) return;
+      uint64_t newstate;
+      if (notifyAll) {
+        // Empty wait stack and set signal to number of pre-wait threads.
+        newstate =
+            (state & kWaiterMask) | (waiters << kSignalShift) | kStackMask;
+      } else if (signals < waiters) {
+        // There is a thread in pre-wait state, unblock it.
+        newstate = state + kSignalInc;
+      } else {
+        // Pop a waiter from list and unpark it.
+        Waiter* w = &waiters_[state & kStackMask];
+        uint64_t next = w->next.load(std::memory_order_relaxed);
+        newstate = (state & (kWaiterMask | kSignalMask)) | next;
+      }
+      CheckState(newstate);
+      if (state_.compare_exchange_weak(state, newstate,
+                                       std::memory_order_acq_rel)) {
+        if (!notifyAll && (signals < waiters))
+          return;  // unblocked pre-wait thread
+        if ((state & kStackMask) == kStackMask) return;
+        Waiter* w = &waiters_[state & kStackMask];
+        if (!notifyAll) w->next.store(kStackMask, std::memory_order_relaxed);
+        Unpark(w);
+        return;
+      }
+    }
+  }
+
+  class Waiter {
+    friend class EventCount;
+    // Align to 128 byte boundary to prevent false sharing with other Waiter
+    // objects in the same vector.
+    EIGEN_ALIGN_TO_BOUNDARY(128) std::atomic<uint64_t> next;
+    std::mutex mu;
+    std::condition_variable cv;
+    uint64_t epoch = 0;
+    unsigned state = kNotSignaled;
+    enum {
+      kNotSignaled,
+      kWaiting,
+      kSignaled,
+    };
+  };
+
+ private:
+  // State_ layout:
+  // - low kWaiterBits is a stack of waiters committed wait
+  //   (indexes in waiters_ array are used as stack elements,
+  //   kStackMask means empty stack).
+  // - next kWaiterBits is count of waiters in prewait state.
+  // - next kWaiterBits is count of pending signals.
+  // - remaining bits are ABA counter for the stack.
+  //   (stored in Waiter node and incremented on push).
+  static const uint64_t kWaiterBits = 14;
+  static const uint64_t kStackMask = (1ull << kWaiterBits) - 1;
+  static const uint64_t kWaiterShift = kWaiterBits;
+  static const uint64_t kWaiterMask = ((1ull << kWaiterBits) - 1)
+                                      << kWaiterShift;
+  static const uint64_t kWaiterInc = 1ull << kWaiterShift;
+  static const uint64_t kSignalShift = 2 * kWaiterBits;
+  static const uint64_t kSignalMask = ((1ull << kWaiterBits) - 1)
+                                      << kSignalShift;
+  static const uint64_t kSignalInc = 1ull << kSignalShift;
+  static const uint64_t kEpochShift = 3 * kWaiterBits;
+  static const uint64_t kEpochBits = 64 - kEpochShift;
+  static const uint64_t kEpochMask = ((1ull << kEpochBits) - 1) << kEpochShift;
+  static const uint64_t kEpochInc = 1ull << kEpochShift;
+  std::atomic<uint64_t> state_;
+  MaxSizeVector<Waiter>& waiters_;
+
+  static void CheckState(uint64_t state, bool waiter = false) {
+    static_assert(kEpochBits >= 20, "not enough bits to prevent ABA problem");
+    const uint64_t waiters = (state & kWaiterMask) >> kWaiterShift;
+    const uint64_t signals = (state & kSignalMask) >> kSignalShift;
+    eigen_plain_assert(waiters >= signals);
+    eigen_plain_assert(waiters < (1 << kWaiterBits) - 1);
+    eigen_plain_assert(!waiter || waiters > 0);
+    (void)waiters;
+    (void)signals;
+  }
+
+  void Park(Waiter* w) {
+    std::unique_lock<std::mutex> lock(w->mu);
+    while (w->state != Waiter::kSignaled) {
+      w->state = Waiter::kWaiting;
+      w->cv.wait(lock);
+    }
+  }
+
+  void Unpark(Waiter* w) {
+    for (Waiter* next; w; w = next) {
+      uint64_t wnext = w->next.load(std::memory_order_relaxed) & kStackMask;
+      next = wnext == kStackMask ? nullptr : &waiters_[wnext];
+      unsigned state;
+      {
+        std::unique_lock<std::mutex> lock(w->mu);
+        state = w->state;
+        w->state = Waiter::kSignaled;
+      }
+      // Avoid notifying if it wasn't waiting.
+      if (state == Waiter::kWaiting) w->cv.notify_one();
+    }
+  }
+
+  EventCount(const EventCount&) = delete;
+  void operator=(const EventCount&) = delete;
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_EVENTCOUNT_H_
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/NonBlockingThreadPool.h b/src/EigenUnsupported/CXX11/src/ThreadPool/NonBlockingThreadPool.h
new file mode 100644
index 0000000..23a2b54
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/NonBlockingThreadPool.h
@@ -0,0 +1,486 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Dmitry Vyukov <dvyukov@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_NONBLOCKING_THREAD_POOL_H
+#define EIGEN_CXX11_THREADPOOL_NONBLOCKING_THREAD_POOL_H
+
+namespace Eigen {
+
+template <typename Environment>
+class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
+ public:
+  typedef typename Environment::Task Task;
+  typedef RunQueue<Task, 1024> Queue;
+
+  ThreadPoolTempl(int num_threads, Environment env = Environment())
+      : ThreadPoolTempl(num_threads, true, env) {}
+
+  ThreadPoolTempl(int num_threads, bool allow_spinning,
+                  Environment env = Environment())
+      : env_(env),
+        num_threads_(num_threads),
+        allow_spinning_(allow_spinning),
+        thread_data_(num_threads),
+        all_coprimes_(num_threads),
+        waiters_(num_threads),
+        global_steal_partition_(EncodePartition(0, num_threads_)),
+        blocked_(0),
+        spinning_(0),
+        done_(false),
+        cancelled_(false),
+        ec_(waiters_) {
+    waiters_.resize(num_threads_);
+    // Calculate coprimes of all numbers [1, num_threads].
+    // Coprimes are used for random walks over all threads in Steal
+    // and NonEmptyQueueIndex. Iteration is based on the fact that if we take
+    // a random starting thread index t and calculate num_threads - 1 subsequent
+    // indices as (t + coprime) % num_threads, we will cover all threads without
+    // repetitions (effectively getting a presudo-random permutation of thread
+    // indices).
+    eigen_plain_assert(num_threads_ < kMaxThreads);
+    for (int i = 1; i <= num_threads_; ++i) {
+      all_coprimes_.emplace_back(i);
+      ComputeCoprimes(i, &all_coprimes_.back());
+    }
+#ifndef EIGEN_THREAD_LOCAL
+    init_barrier_.reset(new Barrier(num_threads_));
+#endif
+    thread_data_.resize(num_threads_);
+    for (int i = 0; i < num_threads_; i++) {
+      SetStealPartition(i, EncodePartition(0, num_threads_));
+      thread_data_[i].thread.reset(
+          env_.CreateThread([this, i]() { WorkerLoop(i); }));
+    }
+#ifndef EIGEN_THREAD_LOCAL
+    // Wait for workers to initialize per_thread_map_. Otherwise we might race
+    // with them in Schedule or CurrentThreadId.
+    init_barrier_->Wait();
+#endif
+  }
+
+  ~ThreadPoolTempl() {
+    done_ = true;
+
+    // Now if all threads block without work, they will start exiting.
+    // But note that threads can continue to work arbitrary long,
+    // block, submit new work, unblock and otherwise live full life.
+    if (!cancelled_) {
+      ec_.Notify(true);
+    } else {
+      // Since we were cancelled, there might be entries in the queues.
+      // Empty them to prevent their destructor from asserting.
+      for (size_t i = 0; i < thread_data_.size(); i++) {
+        thread_data_[i].queue.Flush();
+      }
+    }
+    // Join threads explicitly (by destroying) to avoid destruction order within
+    // this class.
+    for (size_t i = 0; i < thread_data_.size(); ++i)
+      thread_data_[i].thread.reset();
+  }
+
+  void SetStealPartitions(const std::vector<std::pair<unsigned, unsigned>>& partitions) {
+    eigen_plain_assert(partitions.size() == static_cast<std::size_t>(num_threads_));
+
+    // Pass this information to each thread queue.
+    for (int i = 0; i < num_threads_; i++) {
+      const auto& pair = partitions[i];
+      unsigned start = pair.first, end = pair.second;
+      AssertBounds(start, end);
+      unsigned val = EncodePartition(start, end);
+      SetStealPartition(i, val);
+    }
+  }
+
+  void Schedule(std::function<void()> fn) EIGEN_OVERRIDE {
+    ScheduleWithHint(std::move(fn), 0, num_threads_);
+  }
+
+  void ScheduleWithHint(std::function<void()> fn, int start,
+                        int limit) override {
+    Task t = env_.CreateTask(std::move(fn));
+    PerThread* pt = GetPerThread();
+    if (pt->pool == this) {
+      // Worker thread of this pool, push onto the thread's queue.
+      Queue& q = thread_data_[pt->thread_id].queue;
+      t = q.PushFront(std::move(t));
+    } else {
+      // A free-standing thread (or worker of another pool), push onto a random
+      // queue.
+      eigen_plain_assert(start < limit);
+      eigen_plain_assert(limit <= num_threads_);
+      int num_queues = limit - start;
+      int rnd = Rand(&pt->rand) % num_queues;
+      eigen_plain_assert(start + rnd < limit);
+      Queue& q = thread_data_[start + rnd].queue;
+      t = q.PushBack(std::move(t));
+    }
+    // Note: below we touch this after making w available to worker threads.
+    // Strictly speaking, this can lead to a racy-use-after-free. Consider that
+    // Schedule is called from a thread that is neither main thread nor a worker
+    // thread of this pool. Then, execution of w directly or indirectly
+    // completes overall computations, which in turn leads to destruction of
+    // this. We expect that such scenario is prevented by program, that is,
+    // this is kept alive while any threads can potentially be in Schedule.
+    if (!t.f) {
+      ec_.Notify(false);
+    } else {
+      env_.ExecuteTask(t);  // Push failed, execute directly.
+    }
+  }
+
+  void Cancel() EIGEN_OVERRIDE {
+    cancelled_ = true;
+    done_ = true;
+
+    // Let each thread know it's been cancelled.
+#ifdef EIGEN_THREAD_ENV_SUPPORTS_CANCELLATION
+    for (size_t i = 0; i < thread_data_.size(); i++) {
+      thread_data_[i].thread->OnCancel();
+    }
+#endif
+
+    // Wake up the threads without work to let them exit on their own.
+    ec_.Notify(true);
+  }
+
+  int NumThreads() const EIGEN_FINAL { return num_threads_; }
+
+  int CurrentThreadId() const EIGEN_FINAL {
+    const PerThread* pt = const_cast<ThreadPoolTempl*>(this)->GetPerThread();
+    if (pt->pool == this) {
+      return pt->thread_id;
+    } else {
+      return -1;
+    }
+  }
+
+ private:
+  // Create a single atomic<int> that encodes start and limit information for
+  // each thread.
+  // We expect num_threads_ < 65536, so we can store them in a single
+  // std::atomic<unsigned>.
+  // Exposed publicly as static functions so that external callers can reuse
+  // this encode/decode logic for maintaining their own thread-safe copies of
+  // scheduling and steal domain(s).
+  static const int kMaxPartitionBits = 16;
+  static const int kMaxThreads = 1 << kMaxPartitionBits;
+
+  inline unsigned EncodePartition(unsigned start, unsigned limit) {
+    return (start << kMaxPartitionBits) | limit;
+  }
+
+  inline void DecodePartition(unsigned val, unsigned* start, unsigned* limit) {
+    *limit = val & (kMaxThreads - 1);
+    val >>= kMaxPartitionBits;
+    *start = val;
+  }
+
+  void AssertBounds(int start, int end) {
+    eigen_plain_assert(start >= 0);
+    eigen_plain_assert(start < end);  // non-zero sized partition
+    eigen_plain_assert(end <= num_threads_);
+  }
+
+  inline void SetStealPartition(size_t i, unsigned val) {
+    thread_data_[i].steal_partition.store(val, std::memory_order_relaxed);
+  }
+
+  inline unsigned GetStealPartition(int i) {
+    return thread_data_[i].steal_partition.load(std::memory_order_relaxed);
+  }
+
+  void ComputeCoprimes(int N, MaxSizeVector<unsigned>* coprimes) {
+    for (int i = 1; i <= N; i++) {
+      unsigned a = i;
+      unsigned b = N;
+      // If GCD(a, b) == 1, then a and b are coprimes.
+      while (b != 0) {
+        unsigned tmp = a;
+        a = b;
+        b = tmp % b;
+      }
+      if (a == 1) {
+        coprimes->push_back(i);
+      }
+    }
+  }
+
+  typedef typename Environment::EnvThread Thread;
+
+  struct PerThread {
+    constexpr PerThread() : pool(NULL), rand(0), thread_id(-1) {}
+    ThreadPoolTempl* pool;  // Parent pool, or null for normal threads.
+    uint64_t rand;          // Random generator state.
+    int thread_id;          // Worker thread index in pool.
+#ifndef EIGEN_THREAD_LOCAL
+    // Prevent false sharing.
+    char pad_[128];
+#endif
+  };
+
+  struct ThreadData {
+    constexpr ThreadData() : thread(), steal_partition(0), queue() {}
+    std::unique_ptr<Thread> thread;
+    std::atomic<unsigned> steal_partition;
+    Queue queue;
+  };
+
+  Environment env_;
+  const int num_threads_;
+  const bool allow_spinning_;
+  MaxSizeVector<ThreadData> thread_data_;
+  MaxSizeVector<MaxSizeVector<unsigned>> all_coprimes_;
+  MaxSizeVector<EventCount::Waiter> waiters_;
+  unsigned global_steal_partition_;
+  std::atomic<unsigned> blocked_;
+  std::atomic<bool> spinning_;
+  std::atomic<bool> done_;
+  std::atomic<bool> cancelled_;
+  EventCount ec_;
+#ifndef EIGEN_THREAD_LOCAL
+  std::unique_ptr<Barrier> init_barrier_;
+  std::mutex per_thread_map_mutex_;  // Protects per_thread_map_.
+  std::unordered_map<uint64_t, std::unique_ptr<PerThread>> per_thread_map_;
+#endif
+
+  // Main worker thread loop.
+  void WorkerLoop(int thread_id) {
+#ifndef EIGEN_THREAD_LOCAL
+    std::unique_ptr<PerThread> new_pt(new PerThread());
+    per_thread_map_mutex_.lock();
+    bool insertOK = per_thread_map_.emplace(GlobalThreadIdHash(), std::move(new_pt)).second;
+    eigen_plain_assert(insertOK);
+    EIGEN_UNUSED_VARIABLE(insertOK);
+    per_thread_map_mutex_.unlock();
+    init_barrier_->Notify();
+    init_barrier_->Wait();
+#endif
+    PerThread* pt = GetPerThread();
+    pt->pool = this;
+    pt->rand = GlobalThreadIdHash();
+    pt->thread_id = thread_id;
+    Queue& q = thread_data_[thread_id].queue;
+    EventCount::Waiter* waiter = &waiters_[thread_id];
+    // TODO(dvyukov,rmlarsen): The time spent in NonEmptyQueueIndex() is
+    // proportional to num_threads_ and we assume that new work is scheduled at
+    // a constant rate, so we set spin_count to 5000 / num_threads_. The
+    // constant was picked based on a fair dice roll, tune it.
+    const int spin_count =
+        allow_spinning_ && num_threads_ > 0 ? 5000 / num_threads_ : 0;
+    if (num_threads_ == 1) {
+      // For num_threads_ == 1 there is no point in going through the expensive
+      // steal loop. Moreover, since NonEmptyQueueIndex() calls PopBack() on the
+      // victim queues it might reverse the order in which ops are executed
+      // compared to the order in which they are scheduled, which tends to be
+      // counter-productive for the types of I/O workloads the single thread
+      // pools tend to be used for.
+      while (!cancelled_) {
+        Task t = q.PopFront();
+        for (int i = 0; i < spin_count && !t.f; i++) {
+          if (!cancelled_.load(std::memory_order_relaxed)) {
+            t = q.PopFront();
+          }
+        }
+        if (!t.f) {
+          if (!WaitForWork(waiter, &t)) {
+            return;
+          }
+        }
+        if (t.f) {
+          env_.ExecuteTask(t);
+        }
+      }
+    } else {
+      while (!cancelled_) {
+        Task t = q.PopFront();
+        if (!t.f) {
+          t = LocalSteal();
+          if (!t.f) {
+            t = GlobalSteal();
+            if (!t.f) {
+              // Leave one thread spinning. This reduces latency.
+              if (allow_spinning_ && !spinning_ && !spinning_.exchange(true)) {
+                for (int i = 0; i < spin_count && !t.f; i++) {
+                  if (!cancelled_.load(std::memory_order_relaxed)) {
+                    t = GlobalSteal();
+                  } else {
+                    return;
+                  }
+                }
+                spinning_ = false;
+              }
+              if (!t.f) {
+                if (!WaitForWork(waiter, &t)) {
+                  return;
+                }
+              }
+            }
+          }
+        }
+        if (t.f) {
+          env_.ExecuteTask(t);
+        }
+      }
+    }
+  }
+
+  // Steal tries to steal work from other worker threads in the range [start,
+  // limit) in best-effort manner.
+  Task Steal(unsigned start, unsigned limit) {
+    PerThread* pt = GetPerThread();
+    const size_t size = limit - start;
+    unsigned r = Rand(&pt->rand);
+    // Reduce r into [0, size) range, this utilizes trick from
+    // https://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction/
+    eigen_plain_assert(all_coprimes_[size - 1].size() < (1<<30));
+    unsigned victim = ((uint64_t)r * (uint64_t)size) >> 32;
+    unsigned index = ((uint64_t) all_coprimes_[size - 1].size() * (uint64_t)r) >> 32;
+    unsigned inc = all_coprimes_[size - 1][index];
+
+    for (unsigned i = 0; i < size; i++) {
+      eigen_plain_assert(start + victim < limit);
+      Task t = thread_data_[start + victim].queue.PopBack();
+      if (t.f) {
+        return t;
+      }
+      victim += inc;
+      if (victim >= size) {
+        victim -= size;
+      }
+    }
+    return Task();
+  }
+
+  // Steals work within threads belonging to the partition.
+  Task LocalSteal() {
+    PerThread* pt = GetPerThread();
+    unsigned partition = GetStealPartition(pt->thread_id);
+    // If thread steal partition is the same as global partition, there is no
+    // need to go through the steal loop twice.
+    if (global_steal_partition_ == partition) return Task();
+    unsigned start, limit;
+    DecodePartition(partition, &start, &limit);
+    AssertBounds(start, limit);
+
+    return Steal(start, limit);
+  }
+
+  // Steals work from any other thread in the pool.
+  Task GlobalSteal() {
+    return Steal(0, num_threads_);
+  }
+
+
+  // WaitForWork blocks until new work is available (returns true), or if it is
+  // time to exit (returns false). Can optionally return a task to execute in t
+  // (in such case t.f != nullptr on return).
+  bool WaitForWork(EventCount::Waiter* waiter, Task* t) {
+    eigen_plain_assert(!t->f);
+    // We already did best-effort emptiness check in Steal, so prepare for
+    // blocking.
+    ec_.Prewait();
+    // Now do a reliable emptiness check.
+    int victim = NonEmptyQueueIndex();
+    if (victim != -1) {
+      ec_.CancelWait();
+      if (cancelled_) {
+        return false;
+      } else {
+        *t = thread_data_[victim].queue.PopBack();
+        return true;
+      }
+    }
+    // Number of blocked threads is used as termination condition.
+    // If we are shutting down and all worker threads blocked without work,
+    // that's we are done.
+    blocked_++;
+    // TODO is blocked_ required to be unsigned?
+    if (done_ && blocked_ == static_cast<unsigned>(num_threads_)) {
+      ec_.CancelWait();
+      // Almost done, but need to re-check queues.
+      // Consider that all queues are empty and all worker threads are preempted
+      // right after incrementing blocked_ above. Now a free-standing thread
+      // submits work and calls destructor (which sets done_). If we don't
+      // re-check queues, we will exit leaving the work unexecuted.
+      if (NonEmptyQueueIndex() != -1) {
+        // Note: we must not pop from queues before we decrement blocked_,
+        // otherwise the following scenario is possible. Consider that instead
+        // of checking for emptiness we popped the only element from queues.
+        // Now other worker threads can start exiting, which is bad if the
+        // work item submits other work. So we just check emptiness here,
+        // which ensures that all worker threads exit at the same time.
+        blocked_--;
+        return true;
+      }
+      // Reached stable termination state.
+      ec_.Notify(true);
+      return false;
+    }
+    ec_.CommitWait(waiter);
+    blocked_--;
+    return true;
+  }
+
+  int NonEmptyQueueIndex() {
+    PerThread* pt = GetPerThread();
+    // We intentionally design NonEmptyQueueIndex to steal work from
+    // anywhere in the queue so threads don't block in WaitForWork() forever
+    // when all threads in their partition go to sleep. Steal is still local.
+    const size_t size = thread_data_.size();
+    unsigned r = Rand(&pt->rand);
+    unsigned inc = all_coprimes_[size - 1][r % all_coprimes_[size - 1].size()];
+    unsigned victim = r % size;
+    for (unsigned i = 0; i < size; i++) {
+      if (!thread_data_[victim].queue.Empty()) {
+        return victim;
+      }
+      victim += inc;
+      if (victim >= size) {
+        victim -= size;
+      }
+    }
+    return -1;
+  }
+
+  static EIGEN_STRONG_INLINE uint64_t GlobalThreadIdHash() {
+    return std::hash<std::thread::id>()(std::this_thread::get_id());
+  }
+
+  EIGEN_STRONG_INLINE PerThread* GetPerThread() {
+#ifndef EIGEN_THREAD_LOCAL
+    static PerThread dummy;
+    auto it = per_thread_map_.find(GlobalThreadIdHash());
+    if (it == per_thread_map_.end()) {
+      return &dummy;
+    } else {
+      return it->second.get();
+    }
+#else
+    EIGEN_THREAD_LOCAL PerThread per_thread_;
+    PerThread* pt = &per_thread_;
+    return pt;
+#endif
+  }
+
+  static EIGEN_STRONG_INLINE unsigned Rand(uint64_t* state) {
+    uint64_t current = *state;
+    // Update the internal state
+    *state = current * 6364136223846793005ULL + 0xda3e39cb94b95bdbULL;
+    // Generate the random output (using the PCG-XSH-RS scheme)
+    return static_cast<unsigned>((current ^ (current >> 22)) >>
+                                 (22 + (current >> 61)));
+  }
+};
+
+typedef ThreadPoolTempl<StlThreadEnvironment> ThreadPool;
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_NONBLOCKING_THREAD_POOL_H
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/RunQueue.h b/src/EigenUnsupported/CXX11/src/ThreadPool/RunQueue.h
new file mode 100644
index 0000000..b572ebc
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/RunQueue.h
@@ -0,0 +1,236 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Dmitry Vyukov <dvyukov@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_RUNQUEUE_H_
+#define EIGEN_CXX11_THREADPOOL_RUNQUEUE_H_
+
+namespace Eigen {
+
+// RunQueue is a fixed-size, partially non-blocking deque or Work items.
+// Operations on front of the queue must be done by a single thread (owner),
+// operations on back of the queue can be done by multiple threads concurrently.
+//
+// Algorithm outline:
+// All remote threads operating on the queue back are serialized by a mutex.
+// This ensures that at most two threads access state: owner and one remote
+// thread (Size aside). The algorithm ensures that the occupied region of the
+// underlying array is logically continuous (can wraparound, but no stray
+// occupied elements). Owner operates on one end of this region, remote thread
+// operates on the other end. Synchronization between these threads
+// (potential consumption of the last element and take up of the last empty
+// element) happens by means of state variable in each element. States are:
+// empty, busy (in process of insertion of removal) and ready. Threads claim
+// elements (empty->busy and ready->busy transitions) by means of a CAS
+// operation. The finishing transition (busy->empty and busy->ready) are done
+// with plain store as the element is exclusively owned by the current thread.
+//
+// Note: we could permit only pointers as elements, then we would not need
+// separate state variable as null/non-null pointer value would serve as state,
+// but that would require malloc/free per operation for large, complex values
+// (and this is designed to store std::function<()>).
+template <typename Work, unsigned kSize>
+class RunQueue {
+ public:
+  RunQueue() : front_(0), back_(0) {
+    // require power-of-two for fast masking
+    eigen_plain_assert((kSize & (kSize - 1)) == 0);
+    eigen_plain_assert(kSize > 2);            // why would you do this?
+    eigen_plain_assert(kSize <= (64 << 10));  // leave enough space for counter
+    for (unsigned i = 0; i < kSize; i++)
+      array_[i].state.store(kEmpty, std::memory_order_relaxed);
+  }
+
+  ~RunQueue() { eigen_plain_assert(Size() == 0); }
+
+  // PushFront inserts w at the beginning of the queue.
+  // If queue is full returns w, otherwise returns default-constructed Work.
+  Work PushFront(Work w) {
+    unsigned front = front_.load(std::memory_order_relaxed);
+    Elem* e = &array_[front & kMask];
+    uint8_t s = e->state.load(std::memory_order_relaxed);
+    if (s != kEmpty ||
+        !e->state.compare_exchange_strong(s, kBusy, std::memory_order_acquire))
+      return w;
+    front_.store(front + 1 + (kSize << 1), std::memory_order_relaxed);
+    e->w = std::move(w);
+    e->state.store(kReady, std::memory_order_release);
+    return Work();
+  }
+
+  // PopFront removes and returns the first element in the queue.
+  // If the queue was empty returns default-constructed Work.
+  Work PopFront() {
+    unsigned front = front_.load(std::memory_order_relaxed);
+    Elem* e = &array_[(front - 1) & kMask];
+    uint8_t s = e->state.load(std::memory_order_relaxed);
+    if (s != kReady ||
+        !e->state.compare_exchange_strong(s, kBusy, std::memory_order_acquire))
+      return Work();
+    Work w = std::move(e->w);
+    e->state.store(kEmpty, std::memory_order_release);
+    front = ((front - 1) & kMask2) | (front & ~kMask2);
+    front_.store(front, std::memory_order_relaxed);
+    return w;
+  }
+
+  // PushBack adds w at the end of the queue.
+  // If queue is full returns w, otherwise returns default-constructed Work.
+  Work PushBack(Work w) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    unsigned back = back_.load(std::memory_order_relaxed);
+    Elem* e = &array_[(back - 1) & kMask];
+    uint8_t s = e->state.load(std::memory_order_relaxed);
+    if (s != kEmpty ||
+        !e->state.compare_exchange_strong(s, kBusy, std::memory_order_acquire))
+      return w;
+    back = ((back - 1) & kMask2) | (back & ~kMask2);
+    back_.store(back, std::memory_order_relaxed);
+    e->w = std::move(w);
+    e->state.store(kReady, std::memory_order_release);
+    return Work();
+  }
+
+  // PopBack removes and returns the last elements in the queue.
+  Work PopBack() {
+    if (Empty()) return Work();
+    std::unique_lock<std::mutex> lock(mutex_);
+    unsigned back = back_.load(std::memory_order_relaxed);
+    Elem* e = &array_[back & kMask];
+    uint8_t s = e->state.load(std::memory_order_relaxed);
+    if (s != kReady ||
+        !e->state.compare_exchange_strong(s, kBusy, std::memory_order_acquire))
+      return Work();
+    Work w = std::move(e->w);
+    e->state.store(kEmpty, std::memory_order_release);
+    back_.store(back + 1 + (kSize << 1), std::memory_order_relaxed);
+    return w;
+  }
+
+  // PopBackHalf removes and returns half last elements in the queue.
+  // Returns number of elements removed.
+  unsigned PopBackHalf(std::vector<Work>* result) {
+    if (Empty()) return 0;
+    std::unique_lock<std::mutex> lock(mutex_);
+    unsigned back = back_.load(std::memory_order_relaxed);
+    unsigned size = Size();
+    unsigned mid = back;
+    if (size > 1) mid = back + (size - 1) / 2;
+    unsigned n = 0;
+    unsigned start = 0;
+    for (; static_cast<int>(mid - back) >= 0; mid--) {
+      Elem* e = &array_[mid & kMask];
+      uint8_t s = e->state.load(std::memory_order_relaxed);
+      if (n == 0) {
+        if (s != kReady || !e->state.compare_exchange_strong(
+                               s, kBusy, std::memory_order_acquire))
+          continue;
+        start = mid;
+      } else {
+        // Note: no need to store temporal kBusy, we exclusively own these
+        // elements.
+        eigen_plain_assert(s == kReady);
+      }
+      result->push_back(std::move(e->w));
+      e->state.store(kEmpty, std::memory_order_release);
+      n++;
+    }
+    if (n != 0)
+      back_.store(start + 1 + (kSize << 1), std::memory_order_relaxed);
+    return n;
+  }
+
+  // Size returns current queue size.
+  // Can be called by any thread at any time.
+  unsigned Size() const { return SizeOrNotEmpty<true>(); }
+
+  // Empty tests whether container is empty.
+  // Can be called by any thread at any time.
+  bool Empty() const { return SizeOrNotEmpty<false>() == 0; }
+
+  // Delete all the elements from the queue.
+  void Flush() {
+    while (!Empty()) {
+      PopFront();
+    }
+  }
+
+ private:
+  static const unsigned kMask = kSize - 1;
+  static const unsigned kMask2 = (kSize << 1) - 1;
+  struct Elem {
+    std::atomic<uint8_t> state;
+    Work w;
+  };
+  enum {
+    kEmpty,
+    kBusy,
+    kReady,
+  };
+  std::mutex mutex_;
+  // Low log(kSize) + 1 bits in front_ and back_ contain rolling index of
+  // front/back, respectively. The remaining bits contain modification counters
+  // that are incremented on Push operations. This allows us to (1) distinguish
+  // between empty and full conditions (if we would use log(kSize) bits for
+  // position, these conditions would be indistinguishable); (2) obtain
+  // consistent snapshot of front_/back_ for Size operation using the
+  // modification counters.
+  std::atomic<unsigned> front_;
+  std::atomic<unsigned> back_;
+  Elem array_[kSize];
+
+  // SizeOrNotEmpty returns current queue size; if NeedSizeEstimate is false,
+  // only whether the size is 0 is guaranteed to be correct.
+  // Can be called by any thread at any time.
+  template<bool NeedSizeEstimate>
+  unsigned SizeOrNotEmpty() const {
+    // Emptiness plays critical role in thread pool blocking. So we go to great
+    // effort to not produce false positives (claim non-empty queue as empty).
+    unsigned front = front_.load(std::memory_order_acquire);
+    for (;;) {
+      // Capture a consistent snapshot of front/tail.
+      unsigned back = back_.load(std::memory_order_acquire);
+      unsigned front1 = front_.load(std::memory_order_relaxed);
+      if (front != front1) {
+        front = front1;
+        std::atomic_thread_fence(std::memory_order_acquire);
+        continue;
+      }
+      if (NeedSizeEstimate) {
+        return CalculateSize(front, back);
+      } else {
+        // This value will be 0 if the queue is empty, and undefined otherwise.
+        unsigned maybe_zero = ((front ^ back) & kMask2);
+        // Queue size estimate must agree with maybe zero check on the queue
+        // empty/non-empty state.
+        eigen_assert((CalculateSize(front, back) == 0) == (maybe_zero == 0));
+        return maybe_zero;
+      }
+    }
+  }
+
+  EIGEN_ALWAYS_INLINE
+  unsigned CalculateSize(unsigned front, unsigned back) const {
+    int size = (front & kMask2) - (back & kMask2);
+    // Fix overflow.
+    if (size < 0) size += 2 * kSize;
+    // Order of modification in push/pop is crafted to make the queue look
+    // larger than it is during concurrent modifications. E.g. push can
+    // increment size before the corresponding pop has decremented it.
+    // So the computed size can be up to kSize + 1, fix it.
+    if (size > static_cast<int>(kSize)) size = kSize;
+    return static_cast<unsigned>(size);
+  }
+
+  RunQueue(const RunQueue&) = delete;
+  void operator=(const RunQueue&) = delete;
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_RUNQUEUE_H_
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadCancel.h b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadCancel.h
new file mode 100644
index 0000000..a05685f
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadCancel.h
@@ -0,0 +1,23 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_THREAD_CANCEL_H
+#define EIGEN_CXX11_THREADPOOL_THREAD_CANCEL_H
+
+// Try to come up with a portable way to cancel a thread
+#if EIGEN_OS_GNULINUX
+  #define EIGEN_THREAD_CANCEL(t)                  \
+    pthread_cancel(t.native_handle());
+  #define EIGEN_SUPPORTS_THREAD_CANCELLATION 1
+#else
+#define EIGEN_THREAD_CANCEL(t)
+#endif
+
+
+#endif  // EIGEN_CXX11_THREADPOOL_THREAD_CANCEL_H
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadEnvironment.h b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadEnvironment.h
new file mode 100644
index 0000000..d94a064
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadEnvironment.h
@@ -0,0 +1,40 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_THREAD_ENVIRONMENT_H
+#define EIGEN_CXX11_THREADPOOL_THREAD_ENVIRONMENT_H
+
+namespace Eigen {
+
+struct StlThreadEnvironment {
+  struct Task {
+    std::function<void()> f;
+  };
+
+  // EnvThread constructor must start the thread,
+  // destructor must join the thread.
+  class EnvThread {
+   public:
+    EnvThread(std::function<void()> f) : thr_(std::move(f)) {}
+    ~EnvThread() { thr_.join(); }
+    // This function is called when the threadpool is cancelled.
+    void OnCancel() { }
+
+   private:
+    std::thread thr_;
+  };
+
+  EnvThread* CreateThread(std::function<void()> f) { return new EnvThread(std::move(f)); }
+  Task CreateTask(std::function<void()> f) { return Task{std::move(f)}; }
+  void ExecuteTask(const Task& t) { t.f(); }
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_THREAD_ENVIRONMENT_H
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadLocal.h b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadLocal.h
new file mode 100644
index 0000000..4e68474
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadLocal.h
@@ -0,0 +1,301 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_THREAD_LOCAL_H
+#define EIGEN_CXX11_THREADPOOL_THREAD_LOCAL_H
+
+#ifdef EIGEN_AVOID_THREAD_LOCAL
+
+#ifdef EIGEN_THREAD_LOCAL
+#undef EIGEN_THREAD_LOCAL
+#endif
+
+#else
+
+#if EIGEN_MAX_CPP_VER >= 11 &&                         \
+    ((EIGEN_COMP_GNUC && EIGEN_GNUC_AT_LEAST(4, 8)) || \
+     __has_feature(cxx_thread_local)                || \
+     (EIGEN_COMP_MSVC >= 1900) )
+#define EIGEN_THREAD_LOCAL static thread_local
+#endif
+
+// Disable TLS for Apple and Android builds with older toolchains.
+#if defined(__APPLE__)
+// Included for TARGET_OS_IPHONE, __IPHONE_OS_VERSION_MIN_REQUIRED,
+// __IPHONE_8_0.
+#include <Availability.h>
+#include <TargetConditionals.h>
+#endif
+// Checks whether C++11's `thread_local` storage duration specifier is
+// supported.
+#if defined(__apple_build_version__) &&     \
+    ((__apple_build_version__ < 8000042) || \
+     (TARGET_OS_IPHONE && __IPHONE_OS_VERSION_MIN_REQUIRED < __IPHONE_9_0))
+// Notes: Xcode's clang did not support `thread_local` until version
+// 8, and even then not for all iOS < 9.0.
+#undef EIGEN_THREAD_LOCAL
+
+#elif defined(__ANDROID__) && EIGEN_COMP_CLANG
+// There are platforms for which TLS should not be used even though the compiler
+// makes it seem like it's supported (Android NDK < r12b for example).
+// This is primarily because of linker problems and toolchain misconfiguration:
+// TLS isn't supported until NDK r12b per
+// https://developer.android.com/ndk/downloads/revision_history.html
+// Since NDK r16, `__NDK_MAJOR__` and `__NDK_MINOR__` are defined in
+// <android/ndk-version.h>. For NDK < r16, users should define these macros,
+// e.g. `-D__NDK_MAJOR__=11 -D__NKD_MINOR__=0` for NDK r11.
+#if __has_include(<android/ndk-version.h>)
+#include <android/ndk-version.h>
+#endif  // __has_include(<android/ndk-version.h>)
+#if defined(__ANDROID__) && defined(__clang__) && defined(__NDK_MAJOR__) && \
+    defined(__NDK_MINOR__) &&                                               \
+    ((__NDK_MAJOR__ < 12) || ((__NDK_MAJOR__ == 12) && (__NDK_MINOR__ < 1)))
+#undef EIGEN_THREAD_LOCAL
+#endif
+#endif  // defined(__ANDROID__) && defined(__clang__)
+
+#endif  // EIGEN_AVOID_THREAD_LOCAL
+
+namespace Eigen {
+
+namespace internal {
+template <typename T>
+struct ThreadLocalNoOpInitialize {
+  void operator()(T&) const {}
+};
+
+template <typename T>
+struct ThreadLocalNoOpRelease {
+  void operator()(T&) const {}
+};
+
+}  // namespace internal
+
+// Thread local container for elements of type T, that does not use thread local
+// storage. As long as the number of unique threads accessing this storage
+// is smaller than `capacity_`, it is lock-free and wait-free. Otherwise it will
+// use a mutex for synchronization.
+//
+// Type `T` has to be default constructible, and by default each thread will get
+// a default constructed value. It is possible to specify custom `initialize`
+// callable, that will be called lazily from each thread accessing this object,
+// and will be passed a default initialized object of type `T`. Also it's
+// possible to pass a custom `release` callable, that will be invoked before
+// calling ~T().
+//
+// Example:
+//
+//   struct Counter {
+//     int value = 0;
+//   }
+//
+//   Eigen::ThreadLocal<Counter> counter(10);
+//
+//   // Each thread will have access to it's own counter object.
+//   Counter& cnt = counter.local();
+//   cnt++;
+//
+// WARNING: Eigen::ThreadLocal uses the OS-specific value returned by
+// std::this_thread::get_id() to identify threads. This value is not guaranteed
+// to be unique except for the life of the thread. A newly created thread may
+// get an OS-specific ID equal to that of an already destroyed thread.
+//
+// Somewhat similar to TBB thread local storage, with similar restrictions:
+// https://www.threadingbuildingblocks.org/docs/help/reference/thread_local_storage/enumerable_thread_specific_cls.html
+//
+template <typename T,
+          typename Initialize = internal::ThreadLocalNoOpInitialize<T>,
+          typename Release = internal::ThreadLocalNoOpRelease<T>>
+class ThreadLocal {
+  // We preallocate default constructed elements in MaxSizedVector.
+  static_assert(std::is_default_constructible<T>::value,
+                "ThreadLocal data type must be default constructible");
+
+ public:
+  explicit ThreadLocal(int capacity)
+      : ThreadLocal(capacity, internal::ThreadLocalNoOpInitialize<T>(),
+                    internal::ThreadLocalNoOpRelease<T>()) {}
+
+  ThreadLocal(int capacity, Initialize initialize)
+      : ThreadLocal(capacity, std::move(initialize),
+                    internal::ThreadLocalNoOpRelease<T>()) {}
+
+  ThreadLocal(int capacity, Initialize initialize, Release release)
+      : initialize_(std::move(initialize)),
+        release_(std::move(release)),
+        capacity_(capacity),
+        data_(capacity_),
+        ptr_(capacity_),
+        filled_records_(0) {
+    eigen_assert(capacity_ >= 0);
+    data_.resize(capacity_);
+    for (int i = 0; i < capacity_; ++i) {
+      ptr_.emplace_back(nullptr);
+    }
+  }
+
+  T& local() {
+    std::thread::id this_thread = std::this_thread::get_id();
+    if (capacity_ == 0) return SpilledLocal(this_thread);
+
+    std::size_t h = std::hash<std::thread::id>()(this_thread);
+    const int start_idx = h % capacity_;
+
+    // NOTE: From the definition of `std::this_thread::get_id()` it is
+    // guaranteed that we never can have concurrent insertions with the same key
+    // to our hash-map like data structure. If we didn't find an element during
+    // the initial traversal, it's guaranteed that no one else could have
+    // inserted it while we are in this function. This allows to massively
+    // simplify out lock-free insert-only hash map.
+
+    // Check if we already have an element for `this_thread`.
+    int idx = start_idx;
+    while (ptr_[idx].load() != nullptr) {
+      ThreadIdAndValue& record = *(ptr_[idx].load());
+      if (record.thread_id == this_thread) return record.value;
+
+      idx += 1;
+      if (idx >= capacity_) idx -= capacity_;
+      if (idx == start_idx) break;
+    }
+
+    // If we are here, it means that we found an insertion point in lookup
+    // table at `idx`, or we did a full traversal and table is full.
+
+    // If lock-free storage is full, fallback on mutex.
+    if (filled_records_.load() >= capacity_) return SpilledLocal(this_thread);
+
+    // We double check that we still have space to insert an element into a lock
+    // free storage. If old value in `filled_records_` is larger than the
+    // records capacity, it means that some other thread added an element while
+    // we were traversing lookup table.
+    int insertion_index =
+        filled_records_.fetch_add(1, std::memory_order_relaxed);
+    if (insertion_index >= capacity_) return SpilledLocal(this_thread);
+
+    // At this point it's guaranteed that we can access to
+    // data_[insertion_index_] without a data race.
+    data_[insertion_index].thread_id = this_thread;
+    initialize_(data_[insertion_index].value);
+
+    // That's the pointer we'll put into the lookup table.
+    ThreadIdAndValue* inserted = &data_[insertion_index];
+
+    // We'll use nullptr pointer to ThreadIdAndValue in a compare-and-swap loop.
+    ThreadIdAndValue* empty = nullptr;
+
+    // Now we have to find an insertion point into the lookup table. We start
+    // from the `idx` that was identified as an insertion point above, it's
+    // guaranteed that we will have an empty record somewhere in a lookup table
+    // (because we created a record in the `data_`).
+    const int insertion_idx = idx;
+
+    do {
+      // Always start search from the original insertion candidate.
+      idx = insertion_idx;
+      while (ptr_[idx].load() != nullptr) {
+        idx += 1;
+        if (idx >= capacity_) idx -= capacity_;
+        // If we did a full loop, it means that we don't have any free entries
+        // in the lookup table, and this means that something is terribly wrong.
+        eigen_assert(idx != insertion_idx);
+      }
+      // Atomic CAS of the pointer guarantees that any other thread, that will
+      // follow this pointer will see all the mutations in the `data_`.
+    } while (!ptr_[idx].compare_exchange_weak(empty, inserted));
+
+    return inserted->value;
+  }
+
+  // WARN: It's not thread safe to call it concurrently with `local()`.
+  void ForEach(std::function<void(std::thread::id, T&)> f) {
+    // Reading directly from `data_` is unsafe, because only CAS to the
+    // record in `ptr_` makes all changes visible to other threads.
+    for (auto& ptr : ptr_) {
+      ThreadIdAndValue* record = ptr.load();
+      if (record == nullptr) continue;
+      f(record->thread_id, record->value);
+    }
+
+    // We did not spill into the map based storage.
+    if (filled_records_.load(std::memory_order_relaxed) < capacity_) return;
+
+    // Adds a happens before edge from the last call to SpilledLocal().
+    std::unique_lock<std::mutex> lock(mu_);
+    for (auto& kv : per_thread_map_) {
+      f(kv.first, kv.second);
+    }
+  }
+
+  // WARN: It's not thread safe to call it concurrently with `local()`.
+  ~ThreadLocal() {
+    // Reading directly from `data_` is unsafe, because only CAS to the record
+    // in `ptr_` makes all changes visible to other threads.
+    for (auto& ptr : ptr_) {
+      ThreadIdAndValue* record = ptr.load();
+      if (record == nullptr) continue;
+      release_(record->value);
+    }
+
+    // We did not spill into the map based storage.
+    if (filled_records_.load(std::memory_order_relaxed) < capacity_) return;
+
+    // Adds a happens before edge from the last call to SpilledLocal().
+    std::unique_lock<std::mutex> lock(mu_);
+    for (auto& kv : per_thread_map_) {
+      release_(kv.second);
+    }
+  }
+
+ private:
+  struct ThreadIdAndValue {
+    std::thread::id thread_id;
+    T value;
+  };
+
+  // Use unordered map guarded by a mutex when lock free storage is full.
+  T& SpilledLocal(std::thread::id this_thread) {
+    std::unique_lock<std::mutex> lock(mu_);
+
+    auto it = per_thread_map_.find(this_thread);
+    if (it == per_thread_map_.end()) {
+      auto result = per_thread_map_.emplace(this_thread, T());
+      eigen_assert(result.second);
+      initialize_((*result.first).second);
+      return (*result.first).second;
+    } else {
+      return it->second;
+    }
+  }
+
+  Initialize initialize_;
+  Release release_;
+  const int capacity_;
+
+  // Storage that backs lock-free lookup table `ptr_`. Records stored in this
+  // storage contiguously starting from index 0.
+  MaxSizeVector<ThreadIdAndValue> data_;
+
+  // Atomic pointers to the data stored in `data_`. Used as a lookup table for
+  // linear probing hash map (https://en.wikipedia.org/wiki/Linear_probing).
+  MaxSizeVector<std::atomic<ThreadIdAndValue*>> ptr_;
+
+  // Number of records stored in the `data_`.
+  std::atomic<int> filled_records_;
+
+  // We fallback on per thread map if lock-free storage is full. In practice
+  // this should never happen, if `capacity_` is a reasonable estimate of the
+  // number of threads running in a system.
+  std::mutex mu_;  // Protects per_thread_map_.
+  std::unordered_map<std::thread::id, T> per_thread_map_;
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_THREAD_LOCAL_H
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadPoolInterface.h b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadPoolInterface.h
new file mode 100644
index 0000000..25030dc
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadPoolInterface.h
@@ -0,0 +1,48 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_THREAD_POOL_INTERFACE_H
+#define EIGEN_CXX11_THREADPOOL_THREAD_POOL_INTERFACE_H
+
+namespace Eigen {
+
+// This defines an interface that ThreadPoolDevice can take to use
+// custom thread pools underneath.
+class ThreadPoolInterface {
+ public:
+  // Submits a closure to be run by a thread in the pool.
+  virtual void Schedule(std::function<void()> fn) = 0;
+
+  // Submits a closure to be run by threads in the range [start, end) in the
+  // pool.
+  virtual void ScheduleWithHint(std::function<void()> fn, int /*start*/,
+                                int /*end*/) {
+    // Just defer to Schedule in case sub-classes aren't interested in
+    // overriding this functionality.
+    Schedule(fn);
+  }
+
+  // If implemented, stop processing the closures that have been enqueued.
+  // Currently running closures may still be processed.
+  // If not implemented, does nothing.
+  virtual void Cancel() {}
+
+  // Returns the number of threads in the pool.
+  virtual int NumThreads() const = 0;
+
+  // Returns a logical thread index between 0 and NumThreads() - 1 if called
+  // from one of the threads in the pool. Returns -1 otherwise.
+  virtual int CurrentThreadId() const = 0;
+
+  virtual ~ThreadPoolInterface() {}
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_THREAD_POOL_INTERFACE_H
diff --git a/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadYield.h b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadYield.h
new file mode 100644
index 0000000..a859c7b
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/ThreadPool/ThreadYield.h
@@ -0,0 +1,20 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_THREADPOOL_THREAD_YIELD_H
+#define EIGEN_CXX11_THREADPOOL_THREAD_YIELD_H
+
+// Try to come up with a portable way to yield
+#if EIGEN_COMP_GNUC && EIGEN_GNUC_AT_MOST(4, 7)
+#define EIGEN_THREAD_YIELD() sched_yield()
+#else
+#define EIGEN_THREAD_YIELD() std::this_thread::yield()
+#endif
+
+#endif  // EIGEN_CXX11_THREADPOOL_THREAD_YIELD_H
diff --git a/src/EigenUnsupported/CXX11/src/util/CXX11Meta.h b/src/EigenUnsupported/CXX11/src/util/CXX11Meta.h
new file mode 100644
index 0000000..149ceaf
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/util/CXX11Meta.h
@@ -0,0 +1,537 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11META_H
+#define EIGEN_CXX11META_H
+
+#include <vector>
+#include "EmulateArray.h"
+
+#include "CXX11Workarounds.h"
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal
+  * \file CXX11/util/CXX11Meta.h
+  * This file contains generic metaprogramming classes which are not specifically related to Eigen.
+  * This file expands upon Core/util/Meta.h and adds support for C++11 specific features.
+  */
+
+template<typename... tt>
+struct type_list { constexpr static int count = sizeof...(tt); };
+
+template<typename t, typename... tt>
+struct type_list<t, tt...> { constexpr static int count = sizeof...(tt) + 1; typedef t first_type; };
+
+template<typename T, T... nn>
+struct numeric_list { constexpr static std::size_t count = sizeof...(nn); };
+
+template<typename T, T n, T... nn>
+struct numeric_list<T, n, nn...> { static const std::size_t count = sizeof...(nn) + 1; const static T first_value = n; };
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+/* numeric list constructors
+ *
+ * equivalencies:
+ *     constructor                                              result
+ *     typename gen_numeric_list<int, 5>::type                  numeric_list<int, 0,1,2,3,4>
+ *     typename gen_numeric_list_reversed<int, 5>::type         numeric_list<int, 4,3,2,1,0>
+ *     typename gen_numeric_list_swapped_pair<int, 5,1,2>::type numeric_list<int, 0,2,1,3,4>
+ *     typename gen_numeric_list_repeated<int, 0, 5>::type      numeric_list<int, 0,0,0,0,0>
+ */
+
+template<typename T, std::size_t n, T start = 0, T... ii> struct gen_numeric_list                     : gen_numeric_list<T, n-1, start, start + n-1, ii...> {};
+template<typename T, T start, T... ii>                    struct gen_numeric_list<T, 0, start, ii...> { typedef numeric_list<T, ii...> type; };
+
+template<typename T, std::size_t n, T start = 0, T... ii> struct gen_numeric_list_reversed                     : gen_numeric_list_reversed<T, n-1, start, ii..., start + n-1> {};
+template<typename T, T start, T... ii>                    struct gen_numeric_list_reversed<T, 0, start, ii...> { typedef numeric_list<T, ii...> type; };
+
+template<typename T, std::size_t n, T a, T b, T start = 0, T... ii> struct gen_numeric_list_swapped_pair                           : gen_numeric_list_swapped_pair<T, n-1, a, b, start, (start + n-1) == a ? b : ((start + n-1) == b ? a : (start + n-1)), ii...> {};
+template<typename T, T a, T b, T start, T... ii>                    struct gen_numeric_list_swapped_pair<T, 0, a, b, start, ii...> { typedef numeric_list<T, ii...> type; };
+
+template<typename T, std::size_t n, T V, T... nn> struct gen_numeric_list_repeated                 : gen_numeric_list_repeated<T, n-1, V, V, nn...> {};
+template<typename T, T V, T... nn>                struct gen_numeric_list_repeated<T, 0, V, nn...> { typedef numeric_list<T, nn...> type; };
+
+/* list manipulation: concatenate */
+
+template<class a, class b> struct concat;
+
+template<typename... as, typename... bs> struct concat<type_list<as...>,       type_list<bs...>>        { typedef type_list<as..., bs...> type; };
+template<typename T, T... as, T... bs>   struct concat<numeric_list<T, as...>, numeric_list<T, bs...> > { typedef numeric_list<T, as..., bs...> type; };
+
+template<typename... p> struct mconcat;
+template<typename a>                             struct mconcat<a>           { typedef a type; };
+template<typename a, typename b>                 struct mconcat<a, b>        : concat<a, b> {};
+template<typename a, typename b, typename... cs> struct mconcat<a, b, cs...> : concat<a, typename mconcat<b, cs...>::type> {};
+
+/* list manipulation: extract slices */
+
+template<int n, typename x> struct take;
+template<int n, typename a, typename... as> struct take<n, type_list<a, as...>> : concat<type_list<a>, typename take<n-1, type_list<as...>>::type> {};
+template<int n>                             struct take<n, type_list<>>         { typedef type_list<> type; };
+template<typename a, typename... as>        struct take<0, type_list<a, as...>> { typedef type_list<> type; };
+template<>                                  struct take<0, type_list<>>         { typedef type_list<> type; };
+
+template<typename T, int n, T a, T... as> struct take<n, numeric_list<T, a, as...>> : concat<numeric_list<T, a>, typename take<n-1, numeric_list<T, as...>>::type> {};
+template<typename T, int n>               struct take<n, numeric_list<T>>           { typedef numeric_list<T> type; };
+template<typename T, T a, T... as>        struct take<0, numeric_list<T, a, as...>> { typedef numeric_list<T> type; };
+template<typename T>                      struct take<0, numeric_list<T>>           { typedef numeric_list<T> type; };
+
+template<typename T, int n, T... ii>      struct h_skip_helper_numeric;
+template<typename T, int n, T i, T... ii> struct h_skip_helper_numeric<T, n, i, ii...> : h_skip_helper_numeric<T, n-1, ii...> {};
+template<typename T, T i, T... ii>        struct h_skip_helper_numeric<T, 0, i, ii...> { typedef numeric_list<T, i, ii...> type; };
+template<typename T, int n>               struct h_skip_helper_numeric<T, n>           { typedef numeric_list<T> type; };
+template<typename T>                      struct h_skip_helper_numeric<T, 0>           { typedef numeric_list<T> type; };
+
+template<int n, typename... tt>             struct h_skip_helper_type;
+template<int n, typename t, typename... tt> struct h_skip_helper_type<n, t, tt...> : h_skip_helper_type<n-1, tt...> {};
+template<typename t, typename... tt>        struct h_skip_helper_type<0, t, tt...> { typedef type_list<t, tt...> type; };
+template<int n>                             struct h_skip_helper_type<n>           { typedef type_list<> type; };
+template<>                                  struct h_skip_helper_type<0>           { typedef type_list<> type; };
+#endif //not EIGEN_PARSED_BY_DOXYGEN
+
+template<int n>
+struct h_skip {
+  template<typename T, T... ii>
+  constexpr static EIGEN_STRONG_INLINE typename h_skip_helper_numeric<T, n, ii...>::type helper(numeric_list<T, ii...>) { return typename h_skip_helper_numeric<T, n, ii...>::type(); }
+  template<typename... tt>
+  constexpr static EIGEN_STRONG_INLINE typename h_skip_helper_type<n, tt...>::type helper(type_list<tt...>) { return typename h_skip_helper_type<n, tt...>::type(); }
+};
+
+template<int n, typename a> struct skip { typedef decltype(h_skip<n>::helper(a())) type; };
+
+template<int start, int count, typename a> struct slice : take<count, typename skip<start, a>::type> {};
+
+/* list manipulation: retrieve single element from list */
+
+template<int n, typename x> struct get;
+
+template<int n, typename a, typename... as>               struct get<n, type_list<a, as...>>   : get<n-1, type_list<as...>> {};
+template<typename a, typename... as>                      struct get<0, type_list<a, as...>>   { typedef a type; };
+
+template<typename T, int n, T a, T... as>                        struct get<n, numeric_list<T, a, as...>>   : get<n-1, numeric_list<T, as...>> {};
+template<typename T, T a, T... as>                               struct get<0, numeric_list<T, a, as...>>   { constexpr static T value = a; };
+
+template<std::size_t n, typename T, T a, T... as> constexpr T       array_get(const numeric_list<T, a, as...>&) {
+   return get<(int)n, numeric_list<T, a, as...>>::value;
+}
+
+/* always get type, regardless of dummy; good for parameter pack expansion */
+
+template<typename T, T dummy, typename t> struct id_numeric  { typedef t type; };
+template<typename dummy, typename t>      struct id_type     { typedef t type; };
+
+/* equality checking, flagged version */
+
+template<typename a, typename b> struct is_same_gf : is_same<a, b> { constexpr static int global_flags = 0; };
+
+/* apply_op to list */
+
+template<
+  bool from_left, // false
+  template<typename, typename> class op,
+  typename additional_param,
+  typename... values
+>
+struct h_apply_op_helper                                        { typedef type_list<typename op<values, additional_param>::type...> type; };
+template<
+  template<typename, typename> class op,
+  typename additional_param,
+  typename... values
+>
+struct h_apply_op_helper<true, op, additional_param, values...> { typedef type_list<typename op<additional_param, values>::type...> type; };
+
+template<
+  bool from_left,
+  template<typename, typename> class op,
+  typename additional_param
+>
+struct h_apply_op
+{
+  template<typename... values>
+  constexpr static typename h_apply_op_helper<from_left, op, additional_param, values...>::type helper(type_list<values...>)
+  { return typename h_apply_op_helper<from_left, op, additional_param, values...>::type(); }
+};
+
+template<
+  template<typename, typename> class op,
+  typename additional_param,
+  typename a
+>
+struct apply_op_from_left { typedef decltype(h_apply_op<true, op, additional_param>::helper(a())) type; };
+
+template<
+  template<typename, typename> class op,
+  typename additional_param,
+  typename a
+>
+struct apply_op_from_right { typedef decltype(h_apply_op<false, op, additional_param>::helper(a())) type; };
+
+/* see if an element is in a list */
+
+template<
+  template<typename, typename> class test,
+  typename check_against,
+  typename h_list,
+  bool last_check_positive = false
+>
+struct contained_in_list;
+
+template<
+  template<typename, typename> class test,
+  typename check_against,
+  typename h_list
+>
+struct contained_in_list<test, check_against, h_list, true>
+{
+  constexpr static bool value = true;
+};
+
+template<
+  template<typename, typename> class test,
+  typename check_against,
+  typename a,
+  typename... as
+>
+struct contained_in_list<test, check_against, type_list<a, as...>, false> : contained_in_list<test, check_against, type_list<as...>, test<check_against, a>::value> {};
+
+template<
+  template<typename, typename> class test,
+  typename check_against
+  EIGEN_TPL_PP_SPEC_HACK_DEFC(typename, empty)
+>
+struct contained_in_list<test, check_against, type_list<EIGEN_TPL_PP_SPEC_HACK_USE(empty)>, false> { constexpr static bool value = false; };
+
+/* see if an element is in a list and check for global flags */
+
+template<
+  template<typename, typename> class test,
+  typename check_against,
+  typename h_list,
+  int default_flags = 0,
+  bool last_check_positive = false,
+  int last_check_flags = default_flags
+>
+struct contained_in_list_gf;
+
+template<
+  template<typename, typename> class test,
+  typename check_against,
+  typename h_list,
+  int default_flags,
+  int last_check_flags
+>
+struct contained_in_list_gf<test, check_against, h_list, default_flags, true, last_check_flags>
+{
+  constexpr static bool value = true;
+  constexpr static int global_flags = last_check_flags;
+};
+
+template<
+  template<typename, typename> class test,
+  typename check_against,
+  typename a,
+  typename... as,
+  int default_flags,
+  int last_check_flags
+>
+struct contained_in_list_gf<test, check_against, type_list<a, as...>, default_flags, false, last_check_flags> : contained_in_list_gf<test, check_against, type_list<as...>, default_flags, test<check_against, a>::value, test<check_against, a>::global_flags> {};
+
+template<
+  template<typename, typename> class test,
+  typename check_against
+  EIGEN_TPL_PP_SPEC_HACK_DEFC(typename, empty),
+  int default_flags,
+  int last_check_flags
+>
+struct contained_in_list_gf<test, check_against, type_list<EIGEN_TPL_PP_SPEC_HACK_USE(empty)>, default_flags, false, last_check_flags> { constexpr static bool value = false; constexpr static int global_flags = default_flags; };
+
+/* generic reductions */
+
+template<
+  typename Reducer,
+  typename... Ts
+> struct reduce;
+
+template<
+  typename Reducer
+> struct reduce<Reducer>
+{
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE int run() { return Reducer::Identity; }
+};
+
+template<
+  typename Reducer,
+  typename A
+> struct reduce<Reducer, A>
+{
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE A run(A a) { return a; }
+};
+
+template<
+  typename Reducer,
+  typename A,
+  typename... Ts
+> struct reduce<Reducer, A, Ts...>
+{
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE auto run(A a, Ts... ts) -> decltype(Reducer::run(a, reduce<Reducer, Ts...>::run(ts...))) {
+    return Reducer::run(a, reduce<Reducer, Ts...>::run(ts...));
+  }
+};
+
+/* generic binary operations */
+
+struct sum_op           {
+  template<typename A, typename B> EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a + b)   { return a + b;   }
+  static constexpr int Identity = 0;
+};
+struct product_op       {
+  template<typename A, typename B> EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a * b)   { return a * b;   }
+  static constexpr int Identity = 1;
+};
+
+struct logical_and_op   { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a && b)  { return a && b;  } };
+struct logical_or_op    { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a || b)  { return a || b;  } };
+
+struct equal_op         { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a == b)  { return a == b;  } };
+struct not_equal_op     { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a != b)  { return a != b;  } };
+struct lesser_op        { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a < b)   { return a < b;   } };
+struct lesser_equal_op  { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a <= b)  { return a <= b;  } };
+struct greater_op       { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a > b)   { return a > b;   } };
+struct greater_equal_op { template<typename A, typename B> constexpr static EIGEN_STRONG_INLINE auto run(A a, B b) -> decltype(a >= b)  { return a >= b;  } };
+
+/* generic unary operations */
+
+struct not_op                { template<typename A> constexpr static EIGEN_STRONG_INLINE auto run(A a) -> decltype(!a)      { return !a;      } };
+struct negation_op           { template<typename A> constexpr static EIGEN_STRONG_INLINE auto run(A a) -> decltype(-a)      { return -a;      } };
+struct greater_equal_zero_op { template<typename A> constexpr static EIGEN_STRONG_INLINE auto run(A a) -> decltype(a >= 0)  { return a >= 0;  } };
+
+
+/* reductions for lists */
+
+// using auto -> return value spec makes ICC 13.0 and 13.1 crash here, so we have to hack it
+// together in front... (13.0 doesn't work with array_prod/array_reduce/... anyway, but 13.1
+// does...
+template<typename... Ts>
+EIGEN_DEVICE_FUNC constexpr EIGEN_STRONG_INLINE decltype(reduce<product_op, Ts...>::run((*((Ts*)0))...)) arg_prod(Ts... ts)
+{
+  return reduce<product_op, Ts...>::run(ts...);
+}
+
+template<typename... Ts>
+constexpr EIGEN_STRONG_INLINE decltype(reduce<sum_op, Ts...>::run((*((Ts*)0))...)) arg_sum(Ts... ts)
+{
+  return reduce<sum_op, Ts...>::run(ts...);
+}
+
+/* reverse arrays */
+
+template<typename Array, int... n>
+constexpr EIGEN_STRONG_INLINE Array h_array_reverse(Array arr, numeric_list<int, n...>)
+{
+  return {{array_get<sizeof...(n) - n - 1>(arr)...}};
+}
+
+template<typename T, std::size_t N>
+constexpr EIGEN_STRONG_INLINE array<T, N> array_reverse(array<T, N> arr)
+{
+  return h_array_reverse(arr, typename gen_numeric_list<int, N>::type());
+}
+
+
+/* generic array reductions */
+
+// can't reuse standard reduce() interface above because Intel's Compiler
+// *really* doesn't like it, so we just reimplement the stuff
+// (start from N - 1 and work down to 0 because specialization for
+// n == N - 1 also doesn't work in Intel's compiler, so it goes into
+// an infinite loop)
+template<typename Reducer, typename T, std::size_t N, std::size_t n = N - 1>
+struct h_array_reduce {
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE auto run(array<T, N> arr, T identity) -> decltype(Reducer::run(h_array_reduce<Reducer, T, N, n - 1>::run(arr, identity), array_get<n>(arr)))
+  {
+    return Reducer::run(h_array_reduce<Reducer, T, N, n - 1>::run(arr, identity), array_get<n>(arr));
+  }
+};
+
+template<typename Reducer, typename T, std::size_t N>
+struct h_array_reduce<Reducer, T, N, 0>
+{
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE T run(const array<T, N>& arr, T)
+  {
+    return array_get<0>(arr);
+  }
+};
+
+template<typename Reducer, typename T>
+struct h_array_reduce<Reducer, T, 0>
+{
+  EIGEN_DEVICE_FUNC constexpr static EIGEN_STRONG_INLINE T run(const array<T, 0>&, T identity)
+  {
+    return identity;
+  }
+};
+
+template<typename Reducer, typename T, std::size_t N>
+EIGEN_DEVICE_FUNC constexpr EIGEN_STRONG_INLINE auto array_reduce(const array<T, N>& arr, T identity) -> decltype(h_array_reduce<Reducer, T, N>::run(arr, identity))
+{
+  return h_array_reduce<Reducer, T, N>::run(arr, identity);
+}
+
+/* standard array reductions */
+
+template<typename T, std::size_t N>
+EIGEN_DEVICE_FUNC constexpr EIGEN_STRONG_INLINE auto array_sum(const array<T, N>& arr) -> decltype(array_reduce<sum_op, T, N>(arr, static_cast<T>(0)))
+{
+  return array_reduce<sum_op, T, N>(arr, static_cast<T>(0));
+}
+
+template<typename T, std::size_t N>
+EIGEN_DEVICE_FUNC constexpr EIGEN_STRONG_INLINE auto array_prod(const array<T, N>& arr) -> decltype(array_reduce<product_op, T, N>(arr, static_cast<T>(1)))
+{
+  return array_reduce<product_op, T, N>(arr, static_cast<T>(1));
+}
+
+template<typename t>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const std::vector<t>& a) {
+  eigen_assert(a.size() > 0);
+  t prod = 1;
+  for (size_t i = 0; i < a.size(); ++i) { prod *= a[i]; }
+  return prod;
+}
+
+/* zip an array */
+
+template<typename Op, typename A, typename B, std::size_t N, int... n>
+constexpr EIGEN_STRONG_INLINE array<decltype(Op::run(A(), B())),N> h_array_zip(array<A, N> a, array<B, N> b, numeric_list<int, n...>)
+{
+  return array<decltype(Op::run(A(), B())),N>{{ Op::run(array_get<n>(a), array_get<n>(b))... }};
+}
+
+template<typename Op, typename A, typename B, std::size_t N>
+constexpr EIGEN_STRONG_INLINE array<decltype(Op::run(A(), B())),N> array_zip(array<A, N> a, array<B, N> b)
+{
+  return h_array_zip<Op>(a, b, typename gen_numeric_list<int, N>::type());
+}
+
+/* zip an array and reduce the result */
+
+template<typename Reducer, typename Op, typename A, typename B, std::size_t N, int... n>
+constexpr EIGEN_STRONG_INLINE auto h_array_zip_and_reduce(array<A, N> a, array<B, N> b, numeric_list<int, n...>) -> decltype(reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A(), B()))>::type...>::run(Op::run(array_get<n>(a), array_get<n>(b))...))
+{
+  return reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A(), B()))>::type...>::run(Op::run(array_get<n>(a), array_get<n>(b))...);
+}
+
+template<typename Reducer, typename Op, typename A, typename B, std::size_t N>
+constexpr EIGEN_STRONG_INLINE auto array_zip_and_reduce(array<A, N> a, array<B, N> b) -> decltype(h_array_zip_and_reduce<Reducer, Op, A, B, N>(a, b, typename gen_numeric_list<int, N>::type()))
+{
+  return h_array_zip_and_reduce<Reducer, Op, A, B, N>(a, b, typename gen_numeric_list<int, N>::type());
+}
+
+/* apply stuff to an array */
+
+template<typename Op, typename A, std::size_t N, int... n>
+constexpr EIGEN_STRONG_INLINE array<decltype(Op::run(A())),N> h_array_apply(array<A, N> a, numeric_list<int, n...>)
+{
+  return array<decltype(Op::run(A())),N>{{ Op::run(array_get<n>(a))... }};
+}
+
+template<typename Op, typename A, std::size_t N>
+constexpr EIGEN_STRONG_INLINE array<decltype(Op::run(A())),N> array_apply(array<A, N> a)
+{
+  return h_array_apply<Op>(a, typename gen_numeric_list<int, N>::type());
+}
+
+/* apply stuff to an array and reduce */
+
+template<typename Reducer, typename Op, typename A, std::size_t N, int... n>
+constexpr EIGEN_STRONG_INLINE auto h_array_apply_and_reduce(array<A, N> arr, numeric_list<int, n...>) -> decltype(reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A()))>::type...>::run(Op::run(array_get<n>(arr))...))
+{
+  return reduce<Reducer, typename id_numeric<int,n,decltype(Op::run(A()))>::type...>::run(Op::run(array_get<n>(arr))...);
+}
+
+template<typename Reducer, typename Op, typename A, std::size_t N>
+constexpr EIGEN_STRONG_INLINE auto array_apply_and_reduce(array<A, N> a) -> decltype(h_array_apply_and_reduce<Reducer, Op, A, N>(a, typename gen_numeric_list<int, N>::type()))
+{
+  return h_array_apply_and_reduce<Reducer, Op, A, N>(a, typename gen_numeric_list<int, N>::type());
+}
+
+/* repeat a value n times (and make an array out of it
+ * usage:
+ *   array<int, 16> = repeat<16>(42);
+ */
+
+template<int n>
+struct h_repeat
+{
+  template<typename t, int... ii>
+  constexpr static EIGEN_STRONG_INLINE array<t, n> run(t v, numeric_list<int, ii...>)
+  {
+    return {{ typename id_numeric<int, ii, t>::type(v)... }};
+  }
+};
+
+template<int n, typename t>
+constexpr array<t, n> repeat(t v) { return h_repeat<n>::run(v, typename gen_numeric_list<int, n>::type()); }
+
+/* instantiate a class by a C-style array */
+template<class InstType, typename ArrType, std::size_t N, bool Reverse, typename... Ps>
+struct h_instantiate_by_c_array;
+
+template<class InstType, typename ArrType, std::size_t N, typename... Ps>
+struct h_instantiate_by_c_array<InstType, ArrType, N, false, Ps...>
+{
+  static InstType run(ArrType* arr, Ps... args)
+  {
+    return h_instantiate_by_c_array<InstType, ArrType, N - 1, false, Ps..., ArrType>::run(arr + 1, args..., arr[0]);
+  }
+};
+
+template<class InstType, typename ArrType, std::size_t N, typename... Ps>
+struct h_instantiate_by_c_array<InstType, ArrType, N, true, Ps...>
+{
+  static InstType run(ArrType* arr, Ps... args)
+  {
+    return h_instantiate_by_c_array<InstType, ArrType, N - 1, false, ArrType, Ps...>::run(arr + 1, arr[0], args...);
+  }
+};
+
+template<class InstType, typename ArrType, typename... Ps>
+struct h_instantiate_by_c_array<InstType, ArrType, 0, false, Ps...>
+{
+  static InstType run(ArrType* arr, Ps... args)
+  {
+    (void)arr;
+    return InstType(args...);
+  }
+};
+
+template<class InstType, typename ArrType, typename... Ps>
+struct h_instantiate_by_c_array<InstType, ArrType, 0, true, Ps...>
+{
+  static InstType run(ArrType* arr, Ps... args)
+  {
+    (void)arr;
+    return InstType(args...);
+  }
+};
+
+template<class InstType, typename ArrType, std::size_t N, bool Reverse = false>
+InstType instantiate_by_c_array(ArrType* arr)
+{
+  return h_instantiate_by_c_array<InstType, ArrType, N, Reverse>::run(arr);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11META_H
diff --git a/src/EigenUnsupported/CXX11/src/util/CXX11Workarounds.h b/src/EigenUnsupported/CXX11/src/util/CXX11Workarounds.h
new file mode 100644
index 0000000..056736c
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/util/CXX11Workarounds.h
@@ -0,0 +1,88 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11WORKAROUNDS_H
+#define EIGEN_CXX11WORKAROUNDS_H
+
+/* COMPATIBILITY CHECKS
+ * (so users of compilers that are too old get some realistic error messages)
+ */
+#if defined(__INTEL_COMPILER) && (__INTEL_COMPILER < 1310)
+#error Intel Compiler only supports required C++ features since version 13.1.
+// note that most stuff in principle works with 13.0 but when combining
+// some features, at some point 13.0 will just fail with an internal assertion
+#elif defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && (__GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 6))
+// G++ < 4.6 by default will continue processing the source files - even if we use #error to make
+// it error out. For this reason, we use the pragma to make sure G++ aborts at the first error
+// it sees. Unfortunately, that is still not our #error directive, but at least the output is
+// short enough the user has a chance to see that the compiler version is not sufficient for
+// the funky template mojo we use.
+#pragma GCC diagnostic error "-Wfatal-errors"
+#error GNU C++ Compiler (g++) only supports required C++ features since version 4.6.
+#endif
+
+/* Check that the compiler at least claims to support C++11. It might not be sufficient
+ * because the compiler may not implement it correctly, but at least we'll know.
+ * On the other hand, visual studio still doesn't claim to support C++11 although it's
+ * compliant enugh for our purpose.
+ */
+#if (EIGEN_COMP_CXXVER < 11)
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER)
+#pragma GCC diagnostic error "-Wfatal-errors"
+#endif
+#error This library needs at least a C++11 compliant compiler. If you use g++/clang, please enable the -std=c++11 compiler flag. (-std=c++0x on older versions.)
+#endif
+
+namespace Eigen {
+
+namespace internal {
+
+/* std::get is only constexpr in C++14, not yet in C++11
+ */
+
+
+template<std::size_t I_, class T> constexpr inline T&       array_get(std::vector<T>&       a) { return a[I_]; }
+template<std::size_t I_, class T> constexpr inline T&&      array_get(std::vector<T>&&      a) { return a[I_]; }
+template<std::size_t I_, class T> constexpr inline T const& array_get(std::vector<T> const& a) { return a[I_]; }
+
+/* Suppose you have a template of the form
+ * template<typename T> struct X;
+ * And you want to specialize it in such a way:
+ *    template<typename S1, typename... SN> struct X<Foo<S1, SN...>> { ::: };
+ *    template<>                            struct X<Foo<>>          { ::: };
+ * This will work in Intel's compiler 13.0, but only to some extent in g++ 4.6, since
+ * g++ can only match templates called with parameter packs if the number of template
+ * arguments is not a fixed size (so inside the first specialization, referencing
+ * X<Foo<Sn...>> will fail in g++). On the other hand, g++ will accept the following:
+ *    template<typename S...> struct X<Foo<S...>> { ::: }:
+ * as an additional (!) specialization, which will then only match the empty case.
+ * But Intel's compiler 13.0 won't accept that, it will only accept the empty syntax,
+ * so we have to create a workaround for this.
+ */
+#if defined(__GNUC__) && !defined(__INTEL_COMPILER)
+#define EIGEN_TPL_PP_SPEC_HACK_DEF(mt, n)    mt... n
+#define EIGEN_TPL_PP_SPEC_HACK_DEFC(mt, n)   , EIGEN_TPL_PP_SPEC_HACK_DEF(mt, n)
+#define EIGEN_TPL_PP_SPEC_HACK_USE(n)        n...
+#define EIGEN_TPL_PP_SPEC_HACK_USEC(n)       , n...
+#else
+#define EIGEN_TPL_PP_SPEC_HACK_DEF(mt, n)
+#define EIGEN_TPL_PP_SPEC_HACK_DEFC(mt, n)
+#define EIGEN_TPL_PP_SPEC_HACK_USE(n)
+#define EIGEN_TPL_PP_SPEC_HACK_USEC(n)
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11WORKAROUNDS_H
+
+/*
+ * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle;
+ */
diff --git a/src/EigenUnsupported/CXX11/src/util/EmulateArray.h b/src/EigenUnsupported/CXX11/src/util/EmulateArray.h
new file mode 100644
index 0000000..834b20b
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/util/EmulateArray.h
@@ -0,0 +1,261 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_EMULATE_ARRAY_H
+#define EIGEN_EMULATE_ARRAY_H
+
+
+
+// The array class is only available starting with cxx11. Emulate our own here
+// if needed. Beware, msvc still doesn't advertise itself as a c++11 compiler!
+// Moreover, CUDA doesn't support the STL containers, so we use our own instead.
+#if (__cplusplus <= 199711L && EIGEN_COMP_MSVC < 1900) || defined(EIGEN_GPUCC) || defined(EIGEN_AVOID_STL_ARRAY)
+
+namespace Eigen {
+template <typename T, size_t n> class array {
+ public:
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T& operator[] (size_t index) { eigen_internal_assert(index < size()); return values[index]; }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T& operator[] (size_t index) const { eigen_internal_assert(index < size()); return values[index]; }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T& at(size_t index) { eigen_assert(index < size()); return values[index]; }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T& at(size_t index) const { eigen_assert(index < size()); return values[index]; }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T& front() { return values[0]; }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T& front() const { return values[0]; }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T& back() { return values[n-1]; }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T& back() const { return values[n-1]; }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  static std::size_t size() { return n; }
+
+  T values[n];
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array() { }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(const T& v) {
+    EIGEN_STATIC_ASSERT(n==1, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    values[0] = v;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(const T& v1, const T& v2) {
+    EIGEN_STATIC_ASSERT(n==2, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    values[0] = v1;
+    values[1] = v2;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3) {
+    EIGEN_STATIC_ASSERT(n==3, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    values[0] = v1;
+    values[1] = v2;
+    values[2] = v3;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3,
+                            const T& v4) {
+    EIGEN_STATIC_ASSERT(n==4, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    values[0] = v1;
+    values[1] = v2;
+    values[2] = v3;
+    values[3] = v4;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4,
+                            const T& v5) {
+    EIGEN_STATIC_ASSERT(n==5, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    values[0] = v1;
+    values[1] = v2;
+    values[2] = v3;
+    values[3] = v4;
+    values[4] = v5;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4,
+                            const T& v5, const T& v6) {
+    EIGEN_STATIC_ASSERT(n==6, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    values[0] = v1;
+    values[1] = v2;
+    values[2] = v3;
+    values[3] = v4;
+    values[4] = v5;
+    values[5] = v6;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4,
+                            const T& v5, const T& v6, const T& v7) {
+    EIGEN_STATIC_ASSERT(n==7, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    values[0] = v1;
+    values[1] = v2;
+    values[2] = v3;
+    values[3] = v4;
+    values[4] = v5;
+    values[5] = v6;
+    values[6] = v7;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(
+      const T& v1, const T& v2, const T& v3, const T& v4,
+      const T& v5, const T& v6, const T& v7, const T& v8) {
+    EIGEN_STATIC_ASSERT(n==8, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    values[0] = v1;
+    values[1] = v2;
+    values[2] = v3;
+    values[3] = v4;
+    values[4] = v5;
+    values[5] = v6;
+    values[6] = v7;
+    values[7] = v8;
+  }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array(std::initializer_list<T> l) {
+    eigen_assert(l.size() == n);
+    internal::smart_copy(l.begin(), l.end(), values);
+  }
+#endif
+};
+
+
+// Specialize array for zero size
+template <typename T> class array<T, 0> {
+ public:
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T& operator[] (size_t) {
+    eigen_assert(false && "Can't index a zero size array");
+    return dummy;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T& operator[] (size_t) const {
+    eigen_assert(false && "Can't index a zero size array");
+    return dummy;
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T& front() {
+    eigen_assert(false && "Can't index a zero size array");
+    return dummy;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T& front() const {
+    eigen_assert(false && "Can't index a zero size array");
+    return dummy;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE T& back() {
+    eigen_assert(false && "Can't index a zero size array");
+    return dummy;
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const T& back() const {
+    eigen_assert(false && "Can't index a zero size array");
+    return dummy;
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE std::size_t size() { return 0; }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE array() : dummy() { }
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+  EIGEN_DEVICE_FUNC array(std::initializer_list<T> l) : dummy() {
+    EIGEN_UNUSED_VARIABLE(l);
+    eigen_assert(l.size() == 0);
+  }
+#endif
+
+ private:
+  T dummy;
+};
+
+// Comparison operator
+// Todo: implement !=, <, <=, >,  and >=
+template<class T, std::size_t N>
+EIGEN_DEVICE_FUNC bool operator==(const array<T,N>& lhs, const array<T,N>& rhs) {
+  for (std::size_t i = 0; i < N; ++i) {
+    if (lhs[i] != rhs[i]) {
+      return false;
+    }
+  }
+  return true;
+}
+
+
+namespace internal {
+template<std::size_t I_, class T, std::size_t N>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T& array_get(array<T,N>& a) {
+  return a[I_];
+}
+template<std::size_t I_, class T, std::size_t N>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T& array_get(const array<T,N>& a) {
+  return a[I_];
+}
+
+template<class T, std::size_t N> struct array_size<array<T,N> > {
+  enum { value = N };
+};
+template<class T, std::size_t N> struct array_size<array<T,N>& > {
+  enum { value = N };
+};
+template<class T, std::size_t N> struct array_size<const array<T,N> > {
+  enum { value = N };
+};
+template<class T, std::size_t N> struct array_size<const array<T,N>& > {
+  enum { value = N };
+};
+
+}  // end namespace internal
+}  // end namespace Eigen
+
+#else
+
+// The compiler supports c++11, and we're not targeting cuda: use std::array as Eigen::array
+#include <array>
+namespace Eigen {
+
+template <typename T, std::size_t N> using array = std::array<T, N>;
+
+namespace internal {
+/* std::get is only constexpr in C++14, not yet in C++11
+ *     - libstdc++ from version 4.7 onwards has it nevertheless,
+ *                                          so use that
+ *     - libstdc++ older versions: use _M_instance directly
+ *     - libc++ all versions so far: use __elems_ directly
+ *     - all other libs: use std::get to be portable, but
+ *                       this may not be constexpr
+ */
+#if defined(__GLIBCXX__) && __GLIBCXX__ < 20120322
+#define STD_GET_ARR_HACK             a._M_instance[I_]
+#elif defined(_LIBCPP_VERSION)
+#define STD_GET_ARR_HACK             a.__elems_[I_]
+#else
+#define STD_GET_ARR_HACK             std::template get<I_, T, N>(a)
+#endif
+
+template<std::size_t I_, class T, std::size_t N> constexpr inline T&       array_get(std::array<T,N>&       a) { return (T&)       STD_GET_ARR_HACK; }
+template<std::size_t I_, class T, std::size_t N> constexpr inline T&&      array_get(std::array<T,N>&&      a) { return (T&&)      STD_GET_ARR_HACK; }
+template<std::size_t I_, class T, std::size_t N> constexpr inline T const& array_get(std::array<T,N> const& a) { return (T const&) STD_GET_ARR_HACK; }
+
+#undef STD_GET_ARR_HACK
+
+}  // end namespace internal
+}  // end namespace Eigen
+
+#endif
+
+#endif  // EIGEN_EMULATE_ARRAY_H
diff --git a/src/EigenUnsupported/CXX11/src/util/MaxSizeVector.h b/src/EigenUnsupported/CXX11/src/util/MaxSizeVector.h
new file mode 100644
index 0000000..277ab14
--- /dev/null
+++ b/src/EigenUnsupported/CXX11/src/util/MaxSizeVector.h
@@ -0,0 +1,158 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_FIXEDSIZEVECTOR_H
+#define EIGEN_FIXEDSIZEVECTOR_H
+
+namespace Eigen {
+
+/** \class MaxSizeVector
+  * \ingroup Core
+  *
+  * \brief The MaxSizeVector class.
+  *
+  * The %MaxSizeVector provides a subset of std::vector functionality.
+  *
+  * The goal is to provide basic std::vector operations when using
+  * std::vector is not an option (e.g. on GPU or when compiling using
+  * FMA/AVX, as this can cause either compilation failures or illegal
+  * instruction failures).
+  *
+  * Beware: The constructors are not API compatible with these of
+  * std::vector.
+  */
+template <typename T>
+class MaxSizeVector {
+  static const size_t alignment = EIGEN_PLAIN_ENUM_MAX(EIGEN_ALIGNOF(T), sizeof(void*));
+ public:
+  // Construct a new MaxSizeVector, reserve n elements.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit MaxSizeVector(size_t n)
+      : reserve_(n), size_(0),
+        data_(static_cast<T*>(internal::handmade_aligned_malloc(n * sizeof(T), alignment))) {
+  }
+
+  // Construct a new MaxSizeVector, reserve and resize to n.
+  // Copy the init value to all elements.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  MaxSizeVector(size_t n, const T& init)
+      : reserve_(n), size_(n),
+        data_(static_cast<T*>(internal::handmade_aligned_malloc(n * sizeof(T), alignment))) {
+    size_t i = 0;
+    EIGEN_TRY
+    {
+      for(; i < size_; ++i) { new (&data_[i]) T(init); }
+    }
+    EIGEN_CATCH(...)
+    {
+      // Construction failed, destruct in reverse order:
+      for(; (i+1) > 0; --i) { data_[i-1].~T(); }
+      internal::handmade_aligned_free(data_);
+      EIGEN_THROW;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  ~MaxSizeVector() {
+    for (size_t i = size_; i > 0; --i) {
+      data_[i-1].~T();
+    }
+    internal::handmade_aligned_free(data_);
+  }
+
+  void resize(size_t n) {
+    eigen_assert(n <= reserve_);
+    for (; size_ < n; ++size_) {
+      new (&data_[size_]) T;
+    }
+    for (; size_ > n; --size_) {
+      data_[size_-1].~T();
+    }
+    eigen_assert(size_ == n);
+  }
+
+  // Append new elements (up to reserved size).
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void push_back(const T& t) {
+    eigen_assert(size_ < reserve_);
+    new (&data_[size_++]) T(t);
+  }
+
+  // For C++03 compatibility this only takes one argument
+  template<class X>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void emplace_back(const X& x) {
+    eigen_assert(size_ < reserve_);
+    new (&data_[size_++]) T(x);
+  }
+
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T& operator[] (size_t i) const {
+    eigen_assert(i < size_);
+    return data_[i];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T& operator[] (size_t i) {
+    eigen_assert(i < size_);
+    return data_[i];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T& back() {
+    eigen_assert(size_ > 0);
+    return data_[size_ - 1];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T& back() const {
+    eigen_assert(size_ > 0);
+    return data_[size_ - 1];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void pop_back() {
+    eigen_assert(size_ > 0);
+    data_[--size_].~T();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  size_t size() const { return size_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  bool empty() const { return size_ == 0; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T* data() { return data_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T* data() const { return data_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T* begin() { return data_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T* end() { return data_ + size_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T* begin() const { return data_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T* end() const { return data_ + size_; }
+
+ private:
+  size_t reserve_;
+  size_t size_;
+  T* data_;
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_FIXEDSIZEVECTOR_H
diff --git a/src/EigenUnsupported/EulerAngles b/src/EigenUnsupported/EulerAngles
new file mode 100644
index 0000000..f8f1c5d
--- /dev/null
+++ b/src/EigenUnsupported/EulerAngles
@@ -0,0 +1,43 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Tal Hadad <tal_hd@hotmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_EULERANGLES_MODULE_H
+#define EIGEN_EULERANGLES_MODULE_H
+
+
+#include "../../Eigen/Core"
+#include "../../Eigen/Geometry"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+namespace Eigen {
+
+/**
+  * \defgroup EulerAngles_Module EulerAngles module
+  * \brief This module provides generic euler angles rotation.
+  *
+  * Euler angles are a way to represent 3D rotation.
+  *
+  * In order to use this module in your code, include this header:
+  * \code
+  * #include <unsupported/Eigen/EulerAngles>
+  * \endcode
+  *
+  * See \ref EulerAngles for more information.
+  *
+  */
+
+}
+
+#include "src/EulerAngles/EulerSystem.h"
+#include "src/EulerAngles/EulerAngles.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_EULERANGLES_MODULE_H
diff --git a/src/EigenUnsupported/FFT b/src/EigenUnsupported/FFT
new file mode 100644
index 0000000..c8c311a
--- /dev/null
+++ b/src/EigenUnsupported/FFT
@@ -0,0 +1,419 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra. 
+//
+// Copyright (C) 2009 Mark Borgerding mark a borgerding net
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_FFT_H
+#define EIGEN_FFT_H
+
+#include <complex>
+#include <vector>
+#include <map>
+#include "../../Eigen/Core"
+
+
+/**
+  * \defgroup FFT_Module Fast Fourier Transform module
+  *
+  * \code
+  * #include <unsupported/Eigen/FFT>
+  * \endcode
+  *
+  * This module provides Fast Fourier transformation, with a configurable backend
+  * implementation.
+  *
+  * The default implementation is based on kissfft. It is a small, free, and
+  * reasonably efficient default.
+  *
+  * There are currently two implementation backend:
+  *
+  * - fftw (http://www.fftw.org) : faster, GPL -- incompatible with Eigen in LGPL form, bigger code size.
+  * - MKL (http://en.wikipedia.org/wiki/Math_Kernel_Library) : fastest, commercial -- may be incompatible with Eigen in GPL form.
+  *
+  * \section FFTDesign Design
+  *
+  * The following design decisions were made concerning scaling and
+  * half-spectrum for real FFT.
+  *
+  * The intent is to facilitate generic programming and ease migrating code
+  * from  Matlab/octave.
+  * We think the default behavior of Eigen/FFT should favor correctness and
+  * generality over speed. Of course, the caller should be able to "opt-out" from this
+  * behavior and get the speed increase if they want it.
+  *
+  * 1) %Scaling:
+  * Other libraries (FFTW,IMKL,KISSFFT)  do not perform scaling, so there
+  * is a constant gain incurred after the forward&inverse transforms , so 
+  * IFFT(FFT(x)) = Kx;  this is done to avoid a vector-by-value multiply.  
+  * The downside is that algorithms that worked correctly in Matlab/octave 
+  * don't behave the same way once implemented in C++.
+  *
+  * How Eigen/FFT differs: invertible scaling is performed so IFFT( FFT(x) ) = x. 
+  *
+  * 2) Real FFT half-spectrum
+  * Other libraries use only half the frequency spectrum (plus one extra 
+  * sample for the Nyquist bin) for a real FFT, the other half is the 
+  * conjugate-symmetric of the first half.  This saves them a copy and some 
+  * memory.  The downside is the caller needs to have special logic for the 
+  * number of bins in complex vs real.
+  *
+  * How Eigen/FFT differs: The full spectrum is returned from the forward 
+  * transform.  This facilitates generic template programming by obviating 
+  * separate specializations for real vs complex.  On the inverse
+  * transform, only half the spectrum is actually used if the output type is real.
+  */
+ 
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#ifdef EIGEN_FFTW_DEFAULT
+// FFTW: faster, GPL -- incompatible with Eigen in LGPL form, bigger code size
+#  include <fftw3.h>
+#  include "src/FFT/ei_fftw_impl.h"
+   namespace Eigen {
+     //template <typename T> typedef struct internal::fftw_impl  default_fft_impl; this does not work
+     template <typename T> struct default_fft_impl : public internal::fftw_impl<T> {};
+   }
+#elif defined EIGEN_MKL_DEFAULT
+// TODO 
+// intel Math Kernel Library: fastest, commercial -- may be incompatible with Eigen in GPL form
+#  include "src/FFT/ei_imklfft_impl.h"
+   namespace Eigen {
+     template <typename T> struct default_fft_impl : public internal::imklfft_impl {};
+   }
+#else
+// internal::kissfft_impl:  small, free, reasonably efficient default, derived from kissfft
+//
+# include "src/FFT/ei_kissfft_impl.h"
+  namespace Eigen {
+     template <typename T> 
+       struct default_fft_impl : public internal::kissfft_impl<T> {};
+  }
+#endif
+
+namespace Eigen {
+
+ 
+// 
+template<typename T_SrcMat,typename T_FftIfc> struct fft_fwd_proxy;
+template<typename T_SrcMat,typename T_FftIfc> struct fft_inv_proxy;
+
+namespace internal {
+template<typename T_SrcMat,typename T_FftIfc>
+struct traits< fft_fwd_proxy<T_SrcMat,T_FftIfc> >
+{
+  typedef typename T_SrcMat::PlainObject ReturnType;
+};
+template<typename T_SrcMat,typename T_FftIfc>
+struct traits< fft_inv_proxy<T_SrcMat,T_FftIfc> >
+{
+  typedef typename T_SrcMat::PlainObject ReturnType;
+};
+}
+
+template<typename T_SrcMat,typename T_FftIfc> 
+struct fft_fwd_proxy
+ : public ReturnByValue<fft_fwd_proxy<T_SrcMat,T_FftIfc> >
+{
+  typedef DenseIndex Index;
+
+  fft_fwd_proxy(const T_SrcMat& src,T_FftIfc & fft, Index nfft) : m_src(src),m_ifc(fft), m_nfft(nfft) {}
+
+  template<typename T_DestMat> void evalTo(T_DestMat& dst) const;
+
+  Index rows() const { return m_src.rows(); }
+  Index cols() const { return m_src.cols(); }
+protected:
+  const T_SrcMat & m_src;
+  T_FftIfc & m_ifc;
+  Index m_nfft;
+};
+
+template<typename T_SrcMat,typename T_FftIfc> 
+struct fft_inv_proxy
+ : public ReturnByValue<fft_inv_proxy<T_SrcMat,T_FftIfc> >
+{
+  typedef DenseIndex Index;
+
+  fft_inv_proxy(const T_SrcMat& src,T_FftIfc & fft, Index nfft) : m_src(src),m_ifc(fft), m_nfft(nfft) {}
+
+  template<typename T_DestMat> void evalTo(T_DestMat& dst) const;
+
+  Index rows() const { return m_src.rows(); }
+  Index cols() const { return m_src.cols(); }
+protected:
+  const T_SrcMat & m_src;
+  T_FftIfc & m_ifc;
+  Index m_nfft;
+};
+
+
+template <typename T_Scalar,
+         typename T_Impl=default_fft_impl<T_Scalar> >
+class FFT
+{
+  public:
+    typedef T_Impl impl_type;
+    typedef DenseIndex Index;
+    typedef typename impl_type::Scalar Scalar;
+    typedef typename impl_type::Complex Complex;
+
+    enum Flag {
+      Default=0, // goof proof
+      Unscaled=1,
+      HalfSpectrum=2,
+      // SomeOtherSpeedOptimization=4
+      Speedy=32767
+    };
+
+    FFT( const impl_type & impl=impl_type() , Flag flags=Default ) :m_impl(impl),m_flag(flags) { }
+
+    inline
+    bool HasFlag(Flag f) const { return (m_flag & (int)f) == f;}
+
+    inline
+    void SetFlag(Flag f) { m_flag |= (int)f;}
+
+    inline
+    void ClearFlag(Flag f) { m_flag &= (~(int)f);}
+
+    inline
+    void fwd( Complex * dst, const Scalar * src, Index nfft)
+    {
+        m_impl.fwd(dst,src,static_cast<int>(nfft));
+        if ( HasFlag(HalfSpectrum) == false)
+          ReflectSpectrum(dst,nfft);
+    }
+
+    inline
+    void fwd( Complex * dst, const Complex * src, Index nfft)
+    {
+        m_impl.fwd(dst,src,static_cast<int>(nfft));
+    }
+
+    /*
+    inline 
+    void fwd2(Complex * dst, const Complex * src, int n0,int n1)
+    {
+      m_impl.fwd2(dst,src,n0,n1);
+    }
+    */
+
+    template <typename _Input>
+    inline
+    void fwd( std::vector<Complex> & dst, const std::vector<_Input> & src) 
+    {
+      if ( NumTraits<_Input>::IsComplex == 0 && HasFlag(HalfSpectrum) )
+        dst.resize( (src.size()>>1)+1); // half the bins + Nyquist bin
+      else
+        dst.resize(src.size());
+      fwd(&dst[0],&src[0],src.size());
+    }
+
+    template<typename InputDerived, typename ComplexDerived>
+    inline
+    void fwd( MatrixBase<ComplexDerived> & dst, const MatrixBase<InputDerived> & src, Index nfft=-1)
+    {
+      typedef typename ComplexDerived::Scalar dst_type;
+      typedef typename InputDerived::Scalar src_type;
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(InputDerived)
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(ComplexDerived)
+      EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(ComplexDerived,InputDerived) // size at compile-time
+      EIGEN_STATIC_ASSERT((internal::is_same<dst_type, Complex>::value),
+            YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+      EIGEN_STATIC_ASSERT(int(InputDerived::Flags)&int(ComplexDerived::Flags)&DirectAccessBit,
+            THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_WITH_DIRECT_MEMORY_ACCESS_SUCH_AS_MAP_OR_PLAIN_MATRICES)
+
+      if (nfft<1)
+        nfft = src.size();
+
+      if ( NumTraits< src_type >::IsComplex == 0 && HasFlag(HalfSpectrum) )
+        dst.derived().resize( (nfft>>1)+1);
+      else
+        dst.derived().resize(nfft);
+
+      if ( src.innerStride() != 1 || src.size() < nfft ) {
+        Matrix<src_type,1,Dynamic> tmp;
+        if (src.size()<nfft) {
+          tmp.setZero(nfft);
+          tmp.block(0,0,src.size(),1 ) = src;
+        }else{
+          tmp = src;
+        }
+        fwd( &dst[0],&tmp[0],nfft );
+      }else{
+        fwd( &dst[0],&src[0],nfft );
+      }
+    }
+ 
+    template<typename InputDerived>
+    inline
+    fft_fwd_proxy< MatrixBase<InputDerived>, FFT<T_Scalar,T_Impl> >
+    fwd( const MatrixBase<InputDerived> & src, Index nfft=-1)
+    {
+      return fft_fwd_proxy< MatrixBase<InputDerived> ,FFT<T_Scalar,T_Impl> >( src, *this,nfft );
+    }
+
+    template<typename InputDerived>
+    inline
+    fft_inv_proxy< MatrixBase<InputDerived>, FFT<T_Scalar,T_Impl> >
+    inv( const MatrixBase<InputDerived> & src, Index nfft=-1)
+    {
+      return  fft_inv_proxy< MatrixBase<InputDerived> ,FFT<T_Scalar,T_Impl> >( src, *this,nfft );
+    }
+
+    inline
+    void inv( Complex * dst, const Complex * src, Index nfft)
+    {
+      m_impl.inv( dst,src,static_cast<int>(nfft) );
+      if ( HasFlag( Unscaled ) == false)
+        scale(dst,Scalar(1./nfft),nfft); // scale the time series
+    }
+
+    inline
+    void inv( Scalar * dst, const Complex * src, Index nfft)
+    {
+      m_impl.inv( dst,src,static_cast<int>(nfft) );
+      if ( HasFlag( Unscaled ) == false)
+        scale(dst,Scalar(1./nfft),nfft); // scale the time series
+    }
+
+    template<typename OutputDerived, typename ComplexDerived>
+    inline
+    void inv( MatrixBase<OutputDerived> & dst, const MatrixBase<ComplexDerived> & src, Index nfft=-1)
+    {
+      typedef typename ComplexDerived::Scalar src_type;
+      typedef typename ComplexDerived::RealScalar real_type;
+      typedef typename OutputDerived::Scalar dst_type;
+      const bool realfft= (NumTraits<dst_type>::IsComplex == 0);
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OutputDerived)
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(ComplexDerived)
+      EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(ComplexDerived,OutputDerived) // size at compile-time
+      EIGEN_STATIC_ASSERT((internal::is_same<src_type, Complex>::value),
+            YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+      EIGEN_STATIC_ASSERT(int(OutputDerived::Flags)&int(ComplexDerived::Flags)&DirectAccessBit,
+            THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_WITH_DIRECT_MEMORY_ACCESS_SUCH_AS_MAP_OR_PLAIN_MATRICES)
+
+      if (nfft<1) { //automatic FFT size determination
+        if ( realfft && HasFlag(HalfSpectrum) ) 
+          nfft = 2*(src.size()-1); //assume even fft size
+        else
+          nfft = src.size();
+      }
+      dst.derived().resize( nfft );
+
+      // check for nfft that does not fit the input data size
+      Index resize_input= ( realfft && HasFlag(HalfSpectrum) )
+        ? ( (nfft/2+1) - src.size() )
+        : ( nfft - src.size() );
+
+      if ( src.innerStride() != 1 || resize_input ) {
+        // if the vector is strided, then we need to copy it to a packed temporary
+        Matrix<src_type,1,Dynamic> tmp;
+        if ( resize_input ) {
+          size_t ncopy = (std::min)(src.size(),src.size() + resize_input);
+          tmp.setZero(src.size() + resize_input);
+          if ( realfft && HasFlag(HalfSpectrum) ) {
+            // pad at the Nyquist bin
+            tmp.head(ncopy) = src.head(ncopy);
+            tmp(ncopy-1) = real(tmp(ncopy-1)); // enforce real-only Nyquist bin
+          }else{
+            size_t nhead,ntail;
+            nhead = 1+ncopy/2-1; // range  [0:pi)
+            ntail = ncopy/2-1;   // range (-pi:0)
+            tmp.head(nhead) = src.head(nhead);
+            tmp.tail(ntail) = src.tail(ntail);
+            if (resize_input<0) { //shrinking -- create the Nyquist bin as the average of the two bins that fold into it
+              tmp(nhead) = ( src(nfft/2) + src( src.size() - nfft/2 ) )*real_type(.5);
+            }else{ // expanding -- split the old Nyquist bin into two halves
+              tmp(nhead) = src(nhead) * real_type(.5);
+              tmp(tmp.size()-nhead) = tmp(nhead);
+            }
+          }
+        }else{
+          tmp = src;
+        }
+        inv( &dst[0],&tmp[0], nfft);
+      }else{
+        inv( &dst[0],&src[0], nfft);
+      }
+    }
+
+    template <typename _Output>
+    inline
+    void inv( std::vector<_Output> & dst, const std::vector<Complex> & src,Index nfft=-1)
+    {
+      if (nfft<1)
+        nfft = ( NumTraits<_Output>::IsComplex == 0 && HasFlag(HalfSpectrum) ) ? 2*(src.size()-1) : src.size();
+      dst.resize( nfft );
+      inv( &dst[0],&src[0],nfft);
+    }
+
+
+    /*
+    // TODO: multi-dimensional FFTs
+    inline 
+    void inv2(Complex * dst, const Complex * src, int n0,int n1)
+    {
+      m_impl.inv2(dst,src,n0,n1);
+      if ( HasFlag( Unscaled ) == false)
+          scale(dst,1./(n0*n1),n0*n1);
+    }
+  */
+
+    inline
+    impl_type & impl() {return m_impl;}
+  private:
+
+    template <typename T_Data>
+    inline
+    void scale(T_Data * x,Scalar s,Index nx)
+    {
+#if 1
+      for (int k=0;k<nx;++k)
+        *x++ *= s;
+#else
+      if ( ((ptrdiff_t)x) & 15 )
+        Matrix<T_Data, Dynamic, 1>::Map(x,nx) *= s;
+      else
+        Matrix<T_Data, Dynamic, 1>::MapAligned(x,nx) *= s;
+         //Matrix<T_Data, Dynamic, Dynamic>::Map(x,nx) * s;
+#endif  
+    }
+
+    inline
+    void ReflectSpectrum(Complex * freq, Index nfft)
+    {
+      // create the implicit right-half spectrum (conjugate-mirror of the left-half)
+      Index nhbins=(nfft>>1)+1;
+      for (Index k=nhbins;k < nfft; ++k )
+        freq[k] = conj(freq[nfft-k]);
+    }
+
+    impl_type m_impl;
+    int m_flag;
+};
+
+template<typename T_SrcMat,typename T_FftIfc> 
+template<typename T_DestMat> inline 
+void fft_fwd_proxy<T_SrcMat,T_FftIfc>::evalTo(T_DestMat& dst) const
+{
+    m_ifc.fwd( dst, m_src, m_nfft);
+}
+
+template<typename T_SrcMat,typename T_FftIfc> 
+template<typename T_DestMat> inline 
+void fft_inv_proxy<T_SrcMat,T_FftIfc>::evalTo(T_DestMat& dst) const
+{
+    m_ifc.inv( dst, m_src, m_nfft);
+}
+
+}
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif
diff --git a/src/EigenUnsupported/IterativeSolvers b/src/EigenUnsupported/IterativeSolvers
new file mode 100644
index 0000000..a3f58d6
--- /dev/null
+++ b/src/EigenUnsupported/IterativeSolvers
@@ -0,0 +1,51 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <g.gael@free.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ITERATIVE_SOLVERS_MODULE_H
+#define EIGEN_ITERATIVE_SOLVERS_MODULE_H
+
+#include "../../Eigen/Sparse"
+#include "../../Eigen/Jacobi"
+#include "../../Eigen/Householder"
+
+
+/**
+  * \defgroup IterativeLinearSolvers_Module Iterative solvers module
+  * This module aims to provide various iterative linear and non linear solver algorithms.
+  * It currently provides:
+  *  - a constrained conjugate gradient
+  *  - a Householder GMRES implementation
+  *  - an IDR(s) implementation
+  *  - a DGMRES implementation
+  *  - a MINRES implementation
+  *
+  * \code
+  * #include <unsupported/Eigen/IterativeSolvers>
+  * \endcode
+  */
+
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#ifndef EIGEN_MPL2_ONLY
+#include "src/IterativeSolvers/IterationController.h"
+#include "src/IterativeSolvers/ConstrainedConjGrad.h"
+#endif
+
+#include "src/IterativeSolvers/IncompleteLU.h"
+#include "src/IterativeSolvers/GMRES.h"
+#include "src/IterativeSolvers/DGMRES.h"
+//#include "src/IterativeSolvers/SSORPreconditioner.h"
+#include "src/IterativeSolvers/MINRES.h"
+#include "src/IterativeSolvers/IDRS.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+
+#endif // EIGEN_ITERATIVE_SOLVERS_MODULE_H
diff --git a/src/EigenUnsupported/KroneckerProduct b/src/EigenUnsupported/KroneckerProduct
new file mode 100644
index 0000000..5f5afb8
--- /dev/null
+++ b/src/EigenUnsupported/KroneckerProduct
@@ -0,0 +1,36 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_KRONECKER_PRODUCT_MODULE_H
+#define EIGEN_KRONECKER_PRODUCT_MODULE_H
+
+#include "../../Eigen/Core"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#include "../../Eigen/src/SparseCore/SparseUtil.h"
+
+namespace Eigen {
+
+/**
+  * \defgroup KroneckerProduct_Module KroneckerProduct module
+  *
+  * This module contains an experimental Kronecker product implementation.
+  *
+  * \code
+  * #include <Eigen/KroneckerProduct>
+  * \endcode
+  */
+
+} // namespace Eigen
+
+#include "src/KroneckerProduct/KroneckerTensorProduct.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_KRONECKER_PRODUCT_MODULE_H
diff --git a/src/EigenUnsupported/LevenbergMarquardt b/src/EigenUnsupported/LevenbergMarquardt
new file mode 100644
index 0000000..1090505
--- /dev/null
+++ b/src/EigenUnsupported/LevenbergMarquardt
@@ -0,0 +1,49 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_LEVENBERGMARQUARDT_MODULE
+#define EIGEN_LEVENBERGMARQUARDT_MODULE
+
+// #include <vector>
+
+#include "../../Eigen/Core"
+#include "../../Eigen/Jacobi"
+#include "../../Eigen/QR"
+#include "NumericalDiff"
+
+#include "../../Eigen/SparseQR"
+
+/**
+  * \defgroup LevenbergMarquardt_Module Levenberg-Marquardt module
+  *
+  * \code
+  * #include </Eigen/LevenbergMarquardt>
+  * \endcode
+  *
+  * 
+  */
+
+#include "../../Eigen/SparseCore"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+
+#include "src/LevenbergMarquardt/LMqrsolv.h"
+#include "src/LevenbergMarquardt/LMcovar.h"
+#include "src/LevenbergMarquardt/LMpar.h"
+
+#endif
+
+#include "src/LevenbergMarquardt/LevenbergMarquardt.h"
+#include "src/LevenbergMarquardt/LMonestep.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_LEVENBERGMARQUARDT_MODULE
diff --git a/src/EigenUnsupported/MPRealSupport b/src/EigenUnsupported/MPRealSupport
new file mode 100644
index 0000000..c4ea4ec
--- /dev/null
+++ b/src/EigenUnsupported/MPRealSupport
@@ -0,0 +1,213 @@
+// This file is part of a joint effort between Eigen, a lightweight C++ template library
+// for linear algebra, and MPFR C++, a C++ interface to MPFR library (http://www.holoborodko.com/pavel/)
+//
+// Copyright (C) 2010-2012 Pavel Holoborodko <pavel@holoborodko.com>
+// Copyright (C) 2010 Konstantin Holoborodko <konstantin@holoborodko.com>
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MPREALSUPPORT_MODULE_H
+#define EIGEN_MPREALSUPPORT_MODULE_H
+
+#include "../../Eigen/Core"
+#include <mpreal.h>
+
+namespace Eigen {
+  
+/**
+  * \defgroup MPRealSupport_Module MPFRC++ Support module
+  * \code
+  * #include <Eigen/MPRealSupport>
+  * \endcode
+  *
+  * This module provides support for multi precision floating point numbers
+  * via the <a href="http://www.holoborodko.com/pavel/mpfr">MPFR C++</a>
+  * library which itself is built upon <a href="http://www.mpfr.org/">MPFR</a>/<a href="http://gmplib.org/">GMP</a>.
+  *
+  * \warning MPFR C++ is licensed under the GPL.
+  *
+  * You can find a copy of MPFR C++ that is known to be compatible in the unsupported/test/mpreal folder.
+  *
+  * Here is an example:
+  *
+\code
+#include <iostream>
+#include <Eigen/MPRealSupport>
+#include <Eigen/LU>
+using namespace mpfr;
+using namespace Eigen;
+int main()
+{
+  // set precision to 256 bits (double has only 53 bits)
+  mpreal::set_default_prec(256);
+  // Declare matrix and vector types with multi-precision scalar type
+  typedef Matrix<mpreal,Dynamic,Dynamic>  MatrixXmp;
+  typedef Matrix<mpreal,Dynamic,1>        VectorXmp;
+
+  MatrixXmp A = MatrixXmp::Random(100,100);
+  VectorXmp b = VectorXmp::Random(100);
+
+  // Solve Ax=b using LU
+  VectorXmp x = A.lu().solve(b);
+  std::cout << "relative error: " << (A*x - b).norm() / b.norm() << std::endl;
+  return 0;
+}
+\endcode
+  *
+  */
+	
+  template<> struct NumTraits<mpfr::mpreal>
+    : GenericNumTraits<mpfr::mpreal>
+  {
+    enum {
+      IsInteger = 0,
+      IsSigned = 1,
+      IsComplex = 0,
+      RequireInitialization = 1,
+      ReadCost = HugeCost,
+      AddCost  = HugeCost,
+      MulCost  = HugeCost
+    };
+
+    typedef mpfr::mpreal Real;
+    typedef mpfr::mpreal NonInteger;
+    
+    static inline Real highest  (long Precision = mpfr::mpreal::get_default_prec()) { return  mpfr::maxval(Precision); }
+    static inline Real lowest   (long Precision = mpfr::mpreal::get_default_prec()) { return -mpfr::maxval(Precision); }
+
+    // Constants
+    static inline Real Pi      (long Precision = mpfr::mpreal::get_default_prec())  { return mpfr::const_pi(Precision);        }
+    static inline Real Euler   (long Precision = mpfr::mpreal::get_default_prec())  { return mpfr::const_euler(Precision);     }
+    static inline Real Log2    (long Precision = mpfr::mpreal::get_default_prec())  { return mpfr::const_log2(Precision);      }
+    static inline Real Catalan (long Precision = mpfr::mpreal::get_default_prec())  { return mpfr::const_catalan(Precision);   }
+
+    static inline Real epsilon (long Precision = mpfr::mpreal::get_default_prec())  { return mpfr::machine_epsilon(Precision); }
+    static inline Real epsilon (const Real& x)                                      { return mpfr::machine_epsilon(x); }
+
+#ifdef MPREAL_HAVE_DYNAMIC_STD_NUMERIC_LIMITS
+    static inline int digits10 (long Precision = mpfr::mpreal::get_default_prec())  { return std::numeric_limits<Real>::digits10(Precision); }
+    static inline int digits10 (const Real& x)                                      { return std::numeric_limits<Real>::digits10(x); }
+    
+    static inline int digits ()               { return std::numeric_limits<Real>::digits(); }
+    static inline int digits (const Real& x)  { return std::numeric_limits<Real>::digits(x); }
+#endif
+
+    static inline Real dummy_precision()
+    {
+      mpfr_prec_t weak_prec = ((mpfr::mpreal::get_default_prec()-1) * 90) / 100;
+      return mpfr::machine_epsilon(weak_prec);
+    }
+  };
+
+  namespace internal {
+
+  template<> inline mpfr::mpreal random<mpfr::mpreal>()
+  {
+    return mpfr::random();
+  }
+
+  template<> inline mpfr::mpreal random<mpfr::mpreal>(const mpfr::mpreal& a, const mpfr::mpreal& b)
+  {
+    return a + (b-a) * random<mpfr::mpreal>();
+  }
+
+  inline bool isMuchSmallerThan(const mpfr::mpreal& a, const mpfr::mpreal& b, const mpfr::mpreal& eps)
+  {
+    return mpfr::abs(a) <= mpfr::abs(b) * eps;
+  }
+
+  inline bool isApprox(const mpfr::mpreal& a, const mpfr::mpreal& b, const mpfr::mpreal& eps)
+  {
+    return mpfr::isEqualFuzzy(a,b,eps);
+  }
+
+  inline bool isApproxOrLessThan(const mpfr::mpreal& a, const mpfr::mpreal& b, const mpfr::mpreal& eps)
+  {
+    return a <= b || mpfr::isEqualFuzzy(a,b,eps);
+  }
+
+  template<> inline long double cast<mpfr::mpreal,long double>(const mpfr::mpreal& x)
+  { return x.toLDouble(); }
+
+  template<> inline double cast<mpfr::mpreal,double>(const mpfr::mpreal& x)
+  { return x.toDouble(); }
+
+  template<> inline long cast<mpfr::mpreal,long>(const mpfr::mpreal& x)
+  { return x.toLong(); }
+
+  template<> inline int cast<mpfr::mpreal,int>(const mpfr::mpreal& x)
+  { return int(x.toLong()); }
+
+  // Specialize GEBP kernel and traits for mpreal (no need for peeling, nor complicated stuff)
+  // This also permits to directly call mpfr's routines and avoid many temporaries produced by mpreal
+    template<>
+    class gebp_traits<mpfr::mpreal, mpfr::mpreal, false, false>
+    {
+    public:
+      typedef mpfr::mpreal ResScalar;
+      enum {
+        Vectorizable = false,
+        LhsPacketSize = 1,
+        RhsPacketSize = 1,
+        ResPacketSize = 1,
+        NumberOfRegisters = 1,
+        nr = 1,
+        mr = 1,
+        LhsProgress = 1,
+        RhsProgress = 1
+      };
+      typedef ResScalar LhsPacket;
+      typedef ResScalar RhsPacket;
+      typedef ResScalar ResPacket;
+      typedef LhsPacket LhsPacket4Packing;
+      
+    };
+
+
+
+    template<typename Index, typename DataMapper, bool ConjugateLhs, bool ConjugateRhs>
+    struct gebp_kernel<mpfr::mpreal,mpfr::mpreal,Index,DataMapper,1,1,ConjugateLhs,ConjugateRhs>
+    {
+      typedef mpfr::mpreal mpreal;
+
+      EIGEN_DONT_INLINE
+      void operator()(const DataMapper& res, const mpreal* blockA, const mpreal* blockB, 
+                      Index rows, Index depth, Index cols, const mpreal& alpha,
+                      Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0)
+      {
+        if(rows==0 || cols==0 || depth==0)
+          return;
+
+        mpreal  acc1(0,mpfr_get_prec(blockA[0].mpfr_srcptr())),
+                tmp (0,mpfr_get_prec(blockA[0].mpfr_srcptr()));
+
+        if(strideA==-1) strideA = depth;
+        if(strideB==-1) strideB = depth;
+
+        for(Index i=0; i<rows; ++i)
+        {
+          for(Index j=0; j<cols; ++j)
+          {
+            const mpreal *A = blockA + i*strideA + offsetA;
+            const mpreal *B = blockB + j*strideB + offsetB;
+            
+            acc1 = 0;
+            for(Index k=0; k<depth; k++)
+            {
+              mpfr_mul(tmp.mpfr_ptr(), A[k].mpfr_srcptr(), B[k].mpfr_srcptr(), mpreal::get_default_rnd());
+              mpfr_add(acc1.mpfr_ptr(), acc1.mpfr_ptr(), tmp.mpfr_ptr(),  mpreal::get_default_rnd());
+            }
+            
+            mpfr_mul(acc1.mpfr_ptr(), acc1.mpfr_srcptr(), alpha.mpfr_srcptr(), mpreal::get_default_rnd());
+            mpfr_add(res(i,j).mpfr_ptr(), res(i,j).mpfr_srcptr(), acc1.mpfr_srcptr(),  mpreal::get_default_rnd());
+          }
+        }
+      }
+    };
+  } // end namespace internal
+}
+
+#endif // EIGEN_MPREALSUPPORT_MODULE_H
diff --git a/src/EigenUnsupported/MatrixFunctions b/src/EigenUnsupported/MatrixFunctions
new file mode 100644
index 0000000..20c23d1
--- /dev/null
+++ b/src/EigenUnsupported/MatrixFunctions
@@ -0,0 +1,504 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Jitse Niesen <jitse@maths.leeds.ac.uk>
+// Copyright (C) 2012 Chen-Pang He <jdh8@ms63.hinet.net>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIX_FUNCTIONS
+#define EIGEN_MATRIX_FUNCTIONS
+
+#include <cfloat>
+#include <list>
+
+#include "../../Eigen/Core"
+#include "../../Eigen/LU"
+#include "../../Eigen/Eigenvalues"
+
+/**
+  * \defgroup MatrixFunctions_Module Matrix functions module
+  * \brief This module aims to provide various methods for the computation of
+  * matrix functions. 
+  *
+  * To use this module, add 
+  * \code
+  * #include <unsupported/Eigen/MatrixFunctions>
+  * \endcode
+  * at the start of your source file.
+  *
+  * This module defines the following MatrixBase methods.
+  *  - \ref matrixbase_cos "MatrixBase::cos()", for computing the matrix cosine
+  *  - \ref matrixbase_cosh "MatrixBase::cosh()", for computing the matrix hyperbolic cosine
+  *  - \ref matrixbase_exp "MatrixBase::exp()", for computing the matrix exponential
+  *  - \ref matrixbase_log "MatrixBase::log()", for computing the matrix logarithm
+  *  - \ref matrixbase_pow "MatrixBase::pow()", for computing the matrix power
+  *  - \ref matrixbase_matrixfunction "MatrixBase::matrixFunction()", for computing general matrix functions
+  *  - \ref matrixbase_sin "MatrixBase::sin()", for computing the matrix sine
+  *  - \ref matrixbase_sinh "MatrixBase::sinh()", for computing the matrix hyperbolic sine
+  *  - \ref matrixbase_sqrt "MatrixBase::sqrt()", for computing the matrix square root
+  *
+  * These methods are the main entry points to this module. 
+  *
+  * %Matrix functions are defined as follows.  Suppose that \f$ f \f$
+  * is an entire function (that is, a function on the complex plane
+  * that is everywhere complex differentiable).  Then its Taylor
+  * series
+  * \f[ f(0) + f'(0) x + \frac{f''(0)}{2} x^2 + \frac{f'''(0)}{3!} x^3 + \cdots \f]
+  * converges to \f$ f(x) \f$. In this case, we can define the matrix
+  * function by the same series:
+  * \f[ f(M) = f(0) + f'(0) M + \frac{f''(0)}{2} M^2 + \frac{f'''(0)}{3!} M^3 + \cdots \f]
+  *
+  */
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#include "src/MatrixFunctions/MatrixExponential.h"
+#include "src/MatrixFunctions/MatrixFunction.h"
+#include "src/MatrixFunctions/MatrixSquareRoot.h"
+#include "src/MatrixFunctions/MatrixLogarithm.h"
+#include "src/MatrixFunctions/MatrixPower.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+
+/** 
+\page matrixbaseextra_page
+\ingroup MatrixFunctions_Module
+
+\section matrixbaseextra MatrixBase methods defined in the MatrixFunctions module
+
+The remainder of the page documents the following MatrixBase methods
+which are defined in the MatrixFunctions module.
+
+
+
+\subsection matrixbase_cos MatrixBase::cos()
+
+Compute the matrix cosine.
+
+\code
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::cos() const
+\endcode
+
+\param[in]  M  a square matrix.
+\returns  expression representing \f$ \cos(M) \f$.
+
+This function computes the matrix cosine. Use ArrayBase::cos() for computing the entry-wise cosine.
+
+The implementation calls \ref matrixbase_matrixfunction "matrixFunction()" with StdStemFunctions::cos().
+
+\sa \ref matrixbase_sin "sin()" for an example.
+
+
+
+\subsection matrixbase_cosh MatrixBase::cosh()
+
+Compute the matrix hyberbolic cosine.
+
+\code
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::cosh() const
+\endcode
+
+\param[in]  M  a square matrix.
+\returns  expression representing \f$ \cosh(M) \f$
+
+This function calls \ref matrixbase_matrixfunction "matrixFunction()" with StdStemFunctions::cosh().
+
+\sa \ref matrixbase_sinh "sinh()" for an example.
+
+
+
+\subsection matrixbase_exp MatrixBase::exp()
+
+Compute the matrix exponential.
+
+\code
+const MatrixExponentialReturnValue<Derived> MatrixBase<Derived>::exp() const
+\endcode
+
+\param[in]  M  matrix whose exponential is to be computed.
+\returns    expression representing the matrix exponential of \p M.
+
+The matrix exponential of \f$ M \f$ is defined by
+\f[ \exp(M) = \sum_{k=0}^\infty \frac{M^k}{k!}. \f]
+The matrix exponential can be used to solve linear ordinary
+differential equations: the solution of \f$ y' = My \f$ with the
+initial condition \f$ y(0) = y_0 \f$ is given by
+\f$ y(t) = \exp(M) y_0 \f$.
+
+The matrix exponential is different from applying the exp function to all the entries in the matrix.
+Use ArrayBase::exp() if you want to do the latter.
+
+The cost of the computation is approximately \f$ 20 n^3 \f$ for
+matrices of size \f$ n \f$. The number 20 depends weakly on the
+norm of the matrix.
+
+The matrix exponential is computed using the scaling-and-squaring
+method combined with Pad&eacute; approximation. The matrix is first
+rescaled, then the exponential of the reduced matrix is computed
+approximant, and then the rescaling is undone by repeated
+squaring. The degree of the Pad&eacute; approximant is chosen such
+that the approximation error is less than the round-off
+error. However, errors may accumulate during the squaring phase.
+
+Details of the algorithm can be found in: Nicholas J. Higham, "The
+scaling and squaring method for the matrix exponential revisited,"
+<em>SIAM J. %Matrix Anal. Applic.</em>, <b>26</b>:1179&ndash;1193,
+2005.
+
+Example: The following program checks that
+\f[ \exp \left[ \begin{array}{ccc}
+      0 & \frac14\pi & 0 \\
+      -\frac14\pi & 0 & 0 \\
+      0 & 0 & 0
+    \end{array} \right] = \left[ \begin{array}{ccc}
+      \frac12\sqrt2 & -\frac12\sqrt2 & 0 \\
+      \frac12\sqrt2 & \frac12\sqrt2 & 0 \\
+      0 & 0 & 1
+    \end{array} \right]. \f]
+This corresponds to a rotation of \f$ \frac14\pi \f$ radians around
+the z-axis.
+
+\include MatrixExponential.cpp
+Output: \verbinclude MatrixExponential.out
+
+\note \p M has to be a matrix of \c float, \c double, `long double`
+\c complex<float>, \c complex<double>, or `complex<long double>` .
+
+
+\subsection matrixbase_log MatrixBase::log()
+
+Compute the matrix logarithm.
+
+\code
+const MatrixLogarithmReturnValue<Derived> MatrixBase<Derived>::log() const
+\endcode
+
+\param[in]  M  invertible matrix whose logarithm is to be computed.
+\returns    expression representing the matrix logarithm root of \p M.
+
+The matrix logarithm of \f$ M \f$ is a matrix \f$ X \f$ such that 
+\f$ \exp(X) = M \f$ where exp denotes the matrix exponential. As for
+the scalar logarithm, the equation \f$ \exp(X) = M \f$ may have
+multiple solutions; this function returns a matrix whose eigenvalues
+have imaginary part in the interval \f$ (-\pi,\pi] \f$.
+
+The matrix logarithm is different from applying the log function to all the entries in the matrix.
+Use ArrayBase::log() if you want to do the latter.
+
+In the real case, the matrix \f$ M \f$ should be invertible and
+it should have no eigenvalues which are real and negative (pairs of
+complex conjugate eigenvalues are allowed). In the complex case, it
+only needs to be invertible.
+
+This function computes the matrix logarithm using the Schur-Parlett
+algorithm as implemented by MatrixBase::matrixFunction(). The
+logarithm of an atomic block is computed by MatrixLogarithmAtomic,
+which uses direct computation for 1-by-1 and 2-by-2 blocks and an
+inverse scaling-and-squaring algorithm for bigger blocks, with the
+square roots computed by MatrixBase::sqrt().
+
+Details of the algorithm can be found in Section 11.6.2 of:
+Nicholas J. Higham,
+<em>Functions of Matrices: Theory and Computation</em>,
+SIAM 2008. ISBN 978-0-898716-46-7.
+
+Example: The following program checks that
+\f[ \log \left[ \begin{array}{ccc} 
+      \frac12\sqrt2 & -\frac12\sqrt2 & 0 \\
+      \frac12\sqrt2 & \frac12\sqrt2 & 0 \\
+      0 & 0 & 1
+    \end{array} \right] = \left[ \begin{array}{ccc}
+      0 & \frac14\pi & 0 \\ 
+      -\frac14\pi & 0 & 0 \\
+      0 & 0 & 0 
+    \end{array} \right]. \f]
+This corresponds to a rotation of \f$ \frac14\pi \f$ radians around
+the z-axis. This is the inverse of the example used in the
+documentation of \ref matrixbase_exp "exp()".
+
+\include MatrixLogarithm.cpp
+Output: \verbinclude MatrixLogarithm.out
+
+\note \p M has to be a matrix of \c float, \c double, `long
+double`, \c complex<float>, \c complex<double>, or `complex<long double>`.
+
+\sa MatrixBase::exp(), MatrixBase::matrixFunction(), 
+    class MatrixLogarithmAtomic, MatrixBase::sqrt().
+
+
+\subsection matrixbase_pow MatrixBase::pow()
+
+Compute the matrix raised to arbitrary real power.
+
+\code
+const MatrixPowerReturnValue<Derived> MatrixBase<Derived>::pow(RealScalar p) const
+\endcode
+
+\param[in]  M  base of the matrix power, should be a square matrix.
+\param[in]  p  exponent of the matrix power.
+
+The matrix power \f$ M^p \f$ is defined as \f$ \exp(p \log(M)) \f$,
+where exp denotes the matrix exponential, and log denotes the matrix
+logarithm. This is different from raising all the entries in the matrix
+to the p-th power. Use ArrayBase::pow() if you want to do the latter.
+
+If \p p is complex, the scalar type of \p M should be the type of \p
+p . \f$ M^p \f$ simply evaluates into \f$ \exp(p \log(M)) \f$.
+Therefore, the matrix \f$ M \f$ should meet the conditions to be an
+argument of matrix logarithm.
+
+If \p p is real, it is casted into the real scalar type of \p M. Then
+this function computes the matrix power using the Schur-Pad&eacute;
+algorithm as implemented by class MatrixPower. The exponent is split
+into integral part and fractional part, where the fractional part is
+in the interval \f$ (-1, 1) \f$. The main diagonal and the first
+super-diagonal is directly computed.
+
+If \p M is singular with a semisimple zero eigenvalue and \p p is
+positive, the Schur factor \f$ T \f$ is reordered with Givens
+rotations, i.e.
+
+\f[ T = \left[ \begin{array}{cc}
+      T_1 & T_2 \\
+      0   & 0
+    \end{array} \right] \f]
+
+where \f$ T_1 \f$ is invertible. Then \f$ T^p \f$ is given by
+
+\f[ T^p = \left[ \begin{array}{cc}
+      T_1^p & T_1^{-1} T_1^p T_2 \\
+      0     & 0
+    \end{array}. \right] \f]
+
+\warning Fractional power of a matrix with a non-semisimple zero
+eigenvalue is not well-defined. We introduce an assertion failure
+against inaccurate result, e.g. \code
+#include <unsupported/Eigen/MatrixFunctions>
+#include <iostream>
+
+int main()
+{
+  Eigen::Matrix4d A;
+  A << 0, 0, 2, 3,
+       0, 0, 4, 5,
+       0, 0, 6, 7,
+       0, 0, 8, 9;
+  std::cout << A.pow(0.37) << std::endl;
+  
+  // The 1 makes eigenvalue 0 non-semisimple.
+  A.coeffRef(0, 1) = 1;
+
+  // This fails if EIGEN_NO_DEBUG is undefined.
+  std::cout << A.pow(0.37) << std::endl;
+
+  return 0;
+}
+\endcode
+
+Details of the algorithm can be found in: Nicholas J. Higham and
+Lijing Lin, "A Schur-Pad&eacute; algorithm for fractional powers of a
+matrix," <em>SIAM J. %Matrix Anal. Applic.</em>,
+<b>32(3)</b>:1056&ndash;1078, 2011.
+
+Example: The following program checks that
+\f[ \left[ \begin{array}{ccc}
+      \cos1 & -\sin1 & 0 \\
+      \sin1 & \cos1 & 0 \\
+      0 & 0 & 1
+    \end{array} \right]^{\frac14\pi} = \left[ \begin{array}{ccc}
+      \frac12\sqrt2 & -\frac12\sqrt2 & 0 \\
+      \frac12\sqrt2 & \frac12\sqrt2 & 0 \\
+      0 & 0 & 1
+    \end{array} \right]. \f]
+This corresponds to \f$ \frac14\pi \f$ rotations of 1 radian around
+the z-axis.
+
+\include MatrixPower.cpp
+Output: \verbinclude MatrixPower.out
+
+MatrixBase::pow() is user-friendly. However, there are some
+circumstances under which you should use class MatrixPower directly.
+MatrixPower can save the result of Schur decomposition, so it's
+better for computing various powers for the same matrix.
+
+Example:
+\include MatrixPower_optimal.cpp
+Output: \verbinclude MatrixPower_optimal.out
+
+\note \p M has to be a matrix of \c float, \c double, `long
+double`, \c complex<float>, \c complex<double>, or
+\c complex<long double> .
+
+\sa MatrixBase::exp(), MatrixBase::log(), class MatrixPower.
+
+
+\subsection matrixbase_matrixfunction MatrixBase::matrixFunction()
+
+Compute a matrix function.
+
+\code
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::matrixFunction(typename internal::stem_function<typename internal::traits<Derived>::Scalar>::type f) const
+\endcode
+
+\param[in]  M  argument of matrix function, should be a square matrix.
+\param[in]  f  an entire function; \c f(x,n) should compute the n-th
+derivative of f at x.
+\returns  expression representing \p f applied to \p M.
+
+Suppose that \p M is a matrix whose entries have type \c Scalar. 
+Then, the second argument, \p f, should be a function with prototype
+\code 
+ComplexScalar f(ComplexScalar, int) 
+\endcode
+where \c ComplexScalar = \c std::complex<Scalar> if \c Scalar is
+real (e.g., \c float or \c double) and \c ComplexScalar =
+\c Scalar if \c Scalar is complex. The return value of \c f(x,n)
+should be \f$ f^{(n)}(x) \f$, the n-th derivative of f at x.
+
+This routine uses the algorithm described in:
+Philip Davies and Nicholas J. Higham, 
+"A Schur-Parlett algorithm for computing matrix functions", 
+<em>SIAM J. %Matrix Anal. Applic.</em>, <b>25</b>:464&ndash;485, 2003.
+
+The actual work is done by the MatrixFunction class.
+
+Example: The following program checks that
+\f[ \exp \left[ \begin{array}{ccc} 
+      0 & \frac14\pi & 0 \\ 
+      -\frac14\pi & 0 & 0 \\
+      0 & 0 & 0 
+    \end{array} \right] = \left[ \begin{array}{ccc}
+      \frac12\sqrt2 & -\frac12\sqrt2 & 0 \\
+      \frac12\sqrt2 & \frac12\sqrt2 & 0 \\
+      0 & 0 & 1
+    \end{array} \right]. \f]
+This corresponds to a rotation of \f$ \frac14\pi \f$ radians around
+the z-axis. This is the same example as used in the documentation
+of \ref matrixbase_exp "exp()".
+
+\include MatrixFunction.cpp
+Output: \verbinclude MatrixFunction.out
+
+Note that the function \c expfn is defined for complex numbers 
+\c x, even though the matrix \c A is over the reals. Instead of
+\c expfn, we could also have used StdStemFunctions::exp:
+\code
+A.matrixFunction(StdStemFunctions<std::complex<double> >::exp, &B);
+\endcode
+
+
+
+\subsection matrixbase_sin MatrixBase::sin()
+
+Compute the matrix sine.
+
+\code
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::sin() const
+\endcode
+
+\param[in]  M  a square matrix.
+\returns  expression representing \f$ \sin(M) \f$.
+
+This function computes the matrix sine. Use ArrayBase::sin() for computing the entry-wise sine.
+
+The implementation calls \ref matrixbase_matrixfunction "matrixFunction()" with StdStemFunctions::sin().
+
+Example: \include MatrixSine.cpp
+Output: \verbinclude MatrixSine.out
+
+
+
+\subsection matrixbase_sinh MatrixBase::sinh()
+
+Compute the matrix hyperbolic sine.
+
+\code
+MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::sinh() const
+\endcode
+
+\param[in]  M  a square matrix.
+\returns  expression representing \f$ \sinh(M) \f$
+
+This function calls \ref matrixbase_matrixfunction "matrixFunction()" with StdStemFunctions::sinh().
+
+Example: \include MatrixSinh.cpp
+Output: \verbinclude MatrixSinh.out
+
+
+\subsection matrixbase_sqrt MatrixBase::sqrt()
+
+Compute the matrix square root.
+
+\code
+const MatrixSquareRootReturnValue<Derived> MatrixBase<Derived>::sqrt() const
+\endcode
+
+\param[in]  M  invertible matrix whose square root is to be computed.
+\returns    expression representing the matrix square root of \p M.
+
+The matrix square root of \f$ M \f$ is the matrix \f$ M^{1/2} \f$
+whose square is the original matrix; so if \f$ S = M^{1/2} \f$ then
+\f$ S^2 = M \f$. This is different from taking the square root of all
+the entries in the matrix; use ArrayBase::sqrt() if you want to do the
+latter.
+
+In the <b>real case</b>, the matrix \f$ M \f$ should be invertible and
+it should have no eigenvalues which are real and negative (pairs of
+complex conjugate eigenvalues are allowed). In that case, the matrix
+has a square root which is also real, and this is the square root
+computed by this function. 
+
+The matrix square root is computed by first reducing the matrix to
+quasi-triangular form with the real Schur decomposition. The square
+root of the quasi-triangular matrix can then be computed directly. The
+cost is approximately \f$ 25 n^3 \f$ real flops for the real Schur
+decomposition and \f$ 3\frac13 n^3 \f$ real flops for the remainder
+(though the computation time in practice is likely more than this
+indicates).
+
+Details of the algorithm can be found in: Nicholas J. Highan,
+"Computing real square roots of a real matrix", <em>Linear Algebra
+Appl.</em>, 88/89:405&ndash;430, 1987.
+
+If the matrix is <b>positive-definite symmetric</b>, then the square
+root is also positive-definite symmetric. In this case, it is best to
+use SelfAdjointEigenSolver::operatorSqrt() to compute it.
+
+In the <b>complex case</b>, the matrix \f$ M \f$ should be invertible;
+this is a restriction of the algorithm. The square root computed by
+this algorithm is the one whose eigenvalues have an argument in the
+interval \f$ (-\frac12\pi, \frac12\pi] \f$. This is the usual branch
+cut.
+
+The computation is the same as in the real case, except that the
+complex Schur decomposition is used to reduce the matrix to a
+triangular matrix. The theoretical cost is the same. Details are in:
+&Aring;ke Bj&ouml;rck and Sven Hammarling, "A Schur method for the
+square root of a matrix", <em>Linear Algebra Appl.</em>,
+52/53:127&ndash;140, 1983.
+
+Example: The following program checks that the square root of
+\f[ \left[ \begin{array}{cc} 
+              \cos(\frac13\pi) & -\sin(\frac13\pi) \\
+              \sin(\frac13\pi) & \cos(\frac13\pi)
+    \end{array} \right], \f]
+corresponding to a rotation over 60 degrees, is a rotation over 30 degrees:
+\f[ \left[ \begin{array}{cc} 
+              \cos(\frac16\pi) & -\sin(\frac16\pi) \\
+              \sin(\frac16\pi) & \cos(\frac16\pi)
+    \end{array} \right]. \f]
+
+\include MatrixSquareRoot.cpp
+Output: \verbinclude MatrixSquareRoot.out
+
+\sa class RealSchur, class ComplexSchur, class MatrixSquareRoot,
+    SelfAdjointEigenSolver::operatorSqrt().
+
+*/
+
+#endif // EIGEN_MATRIX_FUNCTIONS
+
diff --git a/src/EigenUnsupported/MoreVectorization b/src/EigenUnsupported/MoreVectorization
new file mode 100644
index 0000000..7662b47
--- /dev/null
+++ b/src/EigenUnsupported/MoreVectorization
@@ -0,0 +1,24 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MOREVECTORIZATION_MODULE_H
+#define EIGEN_MOREVECTORIZATION_MODULE_H
+
+#include "../../Eigen/Core"
+
+namespace Eigen {
+
+/**
+  * \defgroup MoreVectorization More vectorization module
+  */
+
+}
+
+#include "src/MoreVectorization/MathFunctions.h"
+
+#endif // EIGEN_MOREVECTORIZATION_MODULE_H
diff --git a/src/EigenUnsupported/NonLinearOptimization b/src/EigenUnsupported/NonLinearOptimization
new file mode 100644
index 0000000..961f192
--- /dev/null
+++ b/src/EigenUnsupported/NonLinearOptimization
@@ -0,0 +1,140 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_NONLINEAROPTIMIZATION_MODULE
+#define EIGEN_NONLINEAROPTIMIZATION_MODULE
+
+#include <vector>
+
+#include "../../Eigen/Core"
+#include "../../Eigen/Jacobi"
+#include "../../Eigen/QR"
+#include "NumericalDiff"
+
+/**
+  * \defgroup NonLinearOptimization_Module Non linear optimization module
+  *
+  * \code
+  * #include <unsupported/Eigen/NonLinearOptimization>
+  * \endcode
+  *
+  * This module provides implementation of two important algorithms in non linear
+  * optimization. In both cases, we consider a system of non linear functions. Of
+  * course, this should work, and even work very well if those functions are
+  * actually linear. But if this is so, you should probably better use other
+  * methods more fitted to this special case.
+  *
+  * One algorithm allows to find a least-squares solution of such a system
+  * (Levenberg-Marquardt algorithm) and the second one is used to find 
+  * a zero for the system (Powell hybrid "dogleg" method).
+  *
+  * This code is a port of minpack (http://en.wikipedia.org/wiki/MINPACK).
+  * Minpack is a very famous, old, robust and well renowned package, written in
+  * fortran. Those implementations have been carefully tuned, tested, and used
+  * for several decades.
+  *
+  * The original fortran code was automatically translated using f2c (http://en.wikipedia.org/wiki/F2c) in C,
+  * then c++, and then cleaned by several different authors.
+  * The last one of those cleanings being our starting point : 
+  * http://devernay.free.fr/hacks/cminpack.html
+  * 
+  * Finally, we ported this code to Eigen, creating classes and API
+  * coherent with Eigen. When possible, we switched to Eigen
+  * implementation, such as most linear algebra (vectors, matrices, stable norms).
+  *
+  * Doing so, we were very careful to check the tests we setup at the very
+  * beginning, which ensure that the same results are found.
+  *
+  * \section Tests Tests
+  * 
+  * The tests are placed in the file unsupported/test/NonLinear.cpp.
+  * 
+  * There are two kinds of tests : those that come from examples bundled with cminpack.
+  * They guaranty we get the same results as the original algorithms (value for 'x',
+  * for the number of evaluations of the function, and for the number of evaluations
+  * of the Jacobian if ever).
+  * 
+  * Other tests were added by myself at the very beginning of the 
+  * process and check the results for Levenberg-Marquardt using the reference data 
+  * on http://www.itl.nist.gov/div898/strd/nls/nls_main.shtml. Since then i've 
+  * carefully checked that the same results were obtained when modifying the
+  * code. Please note that we do not always get the exact same decimals as they do,
+  * but this is ok : they use 128bits float, and we do the tests using the C type 'double',
+  * which is 64 bits on most platforms (x86 and amd64, at least).
+  * I've performed those tests on several other implementations of Levenberg-Marquardt, and
+  * (c)minpack performs VERY well compared to those, both in accuracy and speed.
+  * 
+  * The documentation for running the tests is on the wiki
+  * http://eigen.tuxfamily.org/index.php?title=Tests
+  * 
+  * \section API API: overview of methods
+  * 
+  * Both algorithms needs a functor computing the Jacobian. It can be computed by
+  * hand, using auto-differentiation (see \ref AutoDiff_Module), or using numerical
+  * differences (see \ref NumericalDiff_Module). For instance:
+  *\code
+  * MyFunc func;
+  * NumericalDiff<MyFunc> func_with_num_diff(func);
+  * LevenbergMarquardt<NumericalDiff<MyFunc> > lm(func_with_num_diff);
+  * \endcode
+  * For HybridNonLinearSolver, the method solveNumericalDiff() does the above wrapping for
+  * you.
+  * 
+  * The methods LevenbergMarquardt.lmder1()/lmdif1()/lmstr1() and 
+  * HybridNonLinearSolver.hybrj1()/hybrd1() are specific methods from the original 
+  * minpack package that you probably should NOT use until you are porting a code that
+  * was previously using minpack. They just define a 'simple' API with default values 
+  * for some parameters.
+  * 
+  * All algorithms are provided using two APIs :
+  *     - one where the user inits the algorithm, and uses '*OneStep()' as much as he wants : 
+  * this way the caller have control over the steps
+  *     - one where the user just calls a method (optimize() or solve()) which will 
+  * handle the loop: init + loop until a stop condition is met. Those are provided for
+  *  convenience.
+  * 
+  * As an example, the method LevenbergMarquardt::minimize() is 
+  * implemented as follow: 
+  * \code
+  * Status LevenbergMarquardt<FunctorType,Scalar>::minimize(FVectorType  &x, const int mode)
+  * {
+  *     Status status = minimizeInit(x, mode);
+  *     do {
+  *         status = minimizeOneStep(x, mode);
+  *     } while (status==Running);
+  *     return status;
+  * }
+  * \endcode
+  * 
+  * \section examples Examples
+  * 
+  * The easiest way to understand how to use this module is by looking at the many examples in the file
+  * unsupported/test/NonLinearOptimization.cpp.
+  */
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+
+#include "src/NonLinearOptimization/qrsolv.h"
+#include "src/NonLinearOptimization/r1updt.h"
+#include "src/NonLinearOptimization/r1mpyq.h"
+#include "src/NonLinearOptimization/rwupdt.h"
+#include "src/NonLinearOptimization/fdjac1.h"
+#include "src/NonLinearOptimization/lmpar.h"
+#include "src/NonLinearOptimization/dogleg.h"
+#include "src/NonLinearOptimization/covar.h"
+
+#include "src/NonLinearOptimization/chkder.h"
+
+#endif
+
+#include "src/NonLinearOptimization/HybridNonLinearSolver.h"
+#include "src/NonLinearOptimization/LevenbergMarquardt.h"
+
+
+#endif // EIGEN_NONLINEAROPTIMIZATION_MODULE
diff --git a/src/EigenUnsupported/NumericalDiff b/src/EigenUnsupported/NumericalDiff
new file mode 100644
index 0000000..0668f96
--- /dev/null
+++ b/src/EigenUnsupported/NumericalDiff
@@ -0,0 +1,56 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_NUMERICALDIFF_MODULE
+#define EIGEN_NUMERICALDIFF_MODULE
+
+#include "../../Eigen/Core"
+
+namespace Eigen {
+
+/**
+  * \defgroup NumericalDiff_Module Numerical differentiation module
+  *
+  * \code
+  * #include <unsupported/Eigen/NumericalDiff>
+  * \endcode
+  *
+  * See http://en.wikipedia.org/wiki/Numerical_differentiation
+  *
+  * Warning : this should NOT be confused with automatic differentiation, which
+  * is a different method and has its own module in Eigen : \ref
+  * AutoDiff_Module.
+  *
+  * Currently only "Forward" and "Central" schemes are implemented. Those
+  * are basic methods, and there exist some more elaborated way of
+  * computing such approximates. They are implemented using both
+  * proprietary and free software, and usually requires linking to an
+  * external library. It is very easy for you to write a functor
+  * using such software, and the purpose is quite orthogonal to what we
+  * want to achieve with Eigen.
+  *
+  * This is why we will not provide wrappers for every great numerical
+  * differentiation software that exist, but should rather stick with those
+  * basic ones, that still are useful for testing.
+  *
+  * Also, the \ref NonLinearOptimization_Module needs this in order to
+  * provide full features compatibility with the original (c)minpack
+  * package.
+  *
+  */
+}
+
+//@{
+
+#include "src/NumericalDiff/NumericalDiff.h"
+
+//@}
+
+
+#endif // EIGEN_NUMERICALDIFF_MODULE
diff --git a/src/EigenUnsupported/OpenGLSupport b/src/EigenUnsupported/OpenGLSupport
new file mode 100644
index 0000000..f8c2130
--- /dev/null
+++ b/src/EigenUnsupported/OpenGLSupport
@@ -0,0 +1,322 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_OPENGL_MODULE
+#define EIGEN_OPENGL_MODULE
+
+#include "../../Eigen/Geometry"
+
+#if defined(__APPLE_CC__)
+  #include <OpenGL/gl.h>
+#else
+  #include <GL/gl.h>
+#endif
+
+namespace Eigen {
+
+/**
+  * \defgroup OpenGLSUpport_Module OpenGL Support module
+  *
+  * This module provides wrapper functions for a couple of OpenGL functions
+  * which simplify the way to pass Eigen's object to openGL.
+  * Here is an example:
+  * 
+  * \code
+  * // You need to add path_to_eigen/unsupported to your include path.
+  * #include <Eigen/OpenGLSupport>
+  * // ...
+  * Vector3f x, y;
+  * Matrix3f rot;
+  * 
+  * glVertex(y + x * rot);
+  * 
+  * Quaternion q;
+  * glRotate(q);
+  * 
+  * // ...
+  * \endcode
+  *
+  */
+//@{
+
+#define EIGEN_GL_FUNC_DECLARATION(FUNC)                                                                             \
+namespace internal {                                                                                                \
+  template< typename XprType,                                                                                       \
+            typename Scalar = typename XprType::Scalar,                                                             \
+            int Rows = XprType::RowsAtCompileTime,                                                                  \
+            int Cols = XprType::ColsAtCompileTime,                                                                  \
+            bool IsGLCompatible = bool(internal::evaluator<XprType>::Flags&LinearAccessBit)                         \
+                              && bool(XprType::Flags&DirectAccessBit)                                               \
+                              && (XprType::IsVectorAtCompileTime || (XprType::Flags&RowMajorBit)==0)>               \
+  struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl);                                                                      \
+                                                                                                                    \
+  template<typename XprType, typename Scalar, int Rows, int Cols>                                                   \
+  struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<XprType,Scalar,Rows,Cols,false> {                                     \
+    inline static void run(const XprType& p) {                                                                      \
+      EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<typename plain_matrix_type_column_major<XprType>::type>::run(p); }       \
+  };                                                                                                                \
+}                                                                                                                   \
+                                                                                                                    \
+template<typename Derived> inline void FUNC(const Eigen::DenseBase<Derived>& p) {                                   \
+  EIGEN_CAT(EIGEN_CAT(internal::gl_,FUNC),_impl)<Derived>::run(p.derived());                                        \
+}
+
+
+#define EIGEN_GL_FUNC_SPECIALIZATION_MAT(FUNC,SCALAR,ROWS,COLS,SUFFIX)                                              \
+namespace internal {                                                                                                \
+  template< typename XprType> struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<XprType, SCALAR, ROWS, COLS, true> {      \
+    inline static void run(const XprType& p) { FUNC##SUFFIX(p.data()); }                                            \
+  };                                                                                                                \
+}
+
+  
+#define EIGEN_GL_FUNC_SPECIALIZATION_VEC(FUNC,SCALAR,SIZE,SUFFIX)                                                   \
+namespace internal {                                                                                                \
+  template< typename XprType> struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<XprType, SCALAR, SIZE, 1, true> {         \
+    inline static void run(const XprType& p) { FUNC##SUFFIX(p.data()); }                                            \
+  };                                                                                                                \
+  template< typename XprType> struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<XprType, SCALAR, 1, SIZE, true> {         \
+    inline static void run(const XprType& p) { FUNC##SUFFIX(p.data()); }                                            \
+  };                                                                                                                \
+}
+
+  
+EIGEN_GL_FUNC_DECLARATION       (glVertex)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,int,    2,2iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,short,  2,2sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,float,  2,2fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,double, 2,2dv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,int,    3,3iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,short,  3,3sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,float,  3,3fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,double, 3,3dv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,int,    4,4iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,short,  4,4sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,float,  4,4fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glVertex,double, 4,4dv)
+
+EIGEN_GL_FUNC_DECLARATION       (glTexCoord)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,int,    2,2iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,short,  2,2sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,float,  2,2fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,double, 2,2dv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,int,    3,3iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,short,  3,3sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,float,  3,3fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,double, 3,3dv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,int,    4,4iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,short,  4,4sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,float,  4,4fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTexCoord,double, 4,4dv)
+
+EIGEN_GL_FUNC_DECLARATION       (glColor)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,int,    2,2iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,short,  2,2sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,float,  2,2fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,double, 2,2dv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,int,    3,3iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,short,  3,3sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,float,  3,3fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,double, 3,3dv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,int,    4,4iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,short,  4,4sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,float,  4,4fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glColor,double, 4,4dv)
+
+EIGEN_GL_FUNC_DECLARATION       (glNormal)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glNormal,int,    3,3iv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glNormal,short,  3,3sv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glNormal,float,  3,3fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glNormal,double, 3,3dv)
+
+inline void glScale2fv(const float*  v) { glScalef(v[0], v[1], 1.f);  }
+inline void glScale2dv(const double* v) { glScaled(v[0], v[1], 1.0);  }
+inline void glScale3fv(const float*  v) { glScalef(v[0], v[1], v[2]); }
+inline void glScale3dv(const double* v) { glScaled(v[0], v[1], v[2]); }
+
+EIGEN_GL_FUNC_DECLARATION       (glScale)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glScale,float,  2,2fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glScale,double, 2,2dv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glScale,float,  3,3fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glScale,double, 3,3dv)
+
+template<typename Scalar> void glScale(const UniformScaling<Scalar>& s)  { glScale(Matrix<Scalar,3,1>::Constant(s.factor())); }
+
+inline void glTranslate2fv(const float*  v) { glTranslatef(v[0], v[1], 0.f);  }
+inline void glTranslate2dv(const double* v) { glTranslated(v[0], v[1], 0.0);  }
+inline void glTranslate3fv(const float*  v) { glTranslatef(v[0], v[1], v[2]); }
+inline void glTranslate3dv(const double* v) { glTranslated(v[0], v[1], v[2]); }
+
+EIGEN_GL_FUNC_DECLARATION       (glTranslate)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTranslate,float,  2,2fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTranslate,double, 2,2dv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTranslate,float,  3,3fv)
+EIGEN_GL_FUNC_SPECIALIZATION_VEC(glTranslate,double, 3,3dv)
+
+template<typename Scalar> void glTranslate(const Translation<Scalar,2>& t)  { glTranslate(t.vector()); }
+template<typename Scalar> void glTranslate(const Translation<Scalar,3>& t)  { glTranslate(t.vector()); }
+
+EIGEN_GL_FUNC_DECLARATION       (glMultMatrix)
+EIGEN_GL_FUNC_SPECIALIZATION_MAT(glMultMatrix,float,  4,4,f)
+EIGEN_GL_FUNC_SPECIALIZATION_MAT(glMultMatrix,double, 4,4,d)
+
+template<typename Scalar> void glMultMatrix(const Transform<Scalar,3,Affine>& t)        { glMultMatrix(t.matrix()); }
+template<typename Scalar> void glMultMatrix(const Transform<Scalar,3,Projective>& t)    { glMultMatrix(t.matrix()); }
+template<typename Scalar> void glMultMatrix(const Transform<Scalar,3,AffineCompact>& t) { glMultMatrix(Transform<Scalar,3,Affine>(t).matrix()); }
+
+EIGEN_GL_FUNC_DECLARATION       (glLoadMatrix)
+EIGEN_GL_FUNC_SPECIALIZATION_MAT(glLoadMatrix,float,  4,4,f)
+EIGEN_GL_FUNC_SPECIALIZATION_MAT(glLoadMatrix,double, 4,4,d)
+
+template<typename Scalar> void glLoadMatrix(const Transform<Scalar,3,Affine>& t)        { glLoadMatrix(t.matrix()); }
+template<typename Scalar> void glLoadMatrix(const Transform<Scalar,3,Projective>& t)    { glLoadMatrix(t.matrix()); }
+template<typename Scalar> void glLoadMatrix(const Transform<Scalar,3,AffineCompact>& t) { glLoadMatrix(Transform<Scalar,3,Affine>(t).matrix()); }
+
+inline void glRotate(const Rotation2D<float>& rot)
+{
+  glRotatef(rot.angle()*180.f/float(EIGEN_PI), 0.f, 0.f, 1.f);
+}
+inline void glRotate(const Rotation2D<double>& rot)
+{
+  glRotated(rot.angle()*180.0/double(EIGEN_PI), 0.0, 0.0, 1.0);
+}
+
+template<typename Derived> void glRotate(const RotationBase<Derived,3>& rot)
+{  
+  Transform<typename Derived::Scalar,3,Projective> tr(rot);
+  glMultMatrix(tr.matrix());
+}
+
+#define EIGEN_GL_MAKE_CONST_const const
+#define EIGEN_GL_MAKE_CONST__ 
+#define EIGEN_GL_EVAL(X) X
+
+#define EIGEN_GL_FUNC1_DECLARATION(FUNC,ARG1,CONST)                                                                             \
+namespace internal {                                                                                                            \
+  template< typename XprType,                                                                                                   \
+            typename Scalar = typename XprType::Scalar,                                                                         \
+            int Rows = XprType::RowsAtCompileTime,                                                                              \
+            int Cols = XprType::ColsAtCompileTime,                                                                              \
+            bool IsGLCompatible = bool(internal::evaluator<XprType>::Flags&LinearAccessBit)                                     \
+                              && bool(XprType::Flags&DirectAccessBit)                                                           \
+                              && (XprType::IsVectorAtCompileTime || (XprType::Flags&RowMajorBit)==0)>                           \
+  struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl);                                                                                  \
+                                                                                                                                \
+  template<typename XprType, typename Scalar, int Rows, int Cols>                                                               \
+  struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<XprType,Scalar,Rows,Cols,false> {                                                 \
+    inline static void run(ARG1 a,EIGEN_GL_EVAL(EIGEN_GL_MAKE_CONST_##CONST) XprType& p) {                                      \
+      EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<typename plain_matrix_type_column_major<XprType>::type>::run(a,p); }                 \
+  };                                                                                                                            \
+}                                                                                                                               \
+                                                                                                                                \
+template<typename Derived> inline void FUNC(ARG1 a,EIGEN_GL_EVAL(EIGEN_GL_MAKE_CONST_##CONST) Eigen::DenseBase<Derived>& p) {   \
+  EIGEN_CAT(EIGEN_CAT(internal::gl_,FUNC),_impl)<Derived>::run(a,p.derived());                                                  \
+}
+
+
+#define EIGEN_GL_FUNC1_SPECIALIZATION_MAT(FUNC,ARG1,CONST,SCALAR,ROWS,COLS,SUFFIX)                                              \
+namespace internal {                                                                                                            \
+  template< typename XprType> struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<XprType, SCALAR, ROWS, COLS, true> {                  \
+    inline static void run(ARG1 a, EIGEN_GL_EVAL(EIGEN_GL_MAKE_CONST_##CONST) XprType& p) { FUNC##SUFFIX(a,p.data()); }         \
+  }; \
+}
+
+  
+#define EIGEN_GL_FUNC1_SPECIALIZATION_VEC(FUNC,ARG1,CONST,SCALAR,SIZE,SUFFIX)                                                   \
+namespace internal {                                                                                                            \
+  template< typename XprType> struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<XprType, SCALAR, SIZE, 1, true> {                     \
+    inline static void run(ARG1 a, EIGEN_GL_EVAL(EIGEN_GL_MAKE_CONST_##CONST) XprType& p) { FUNC##SUFFIX(a,p.data()); }         \
+  };                                                                                                                            \
+  template< typename XprType> struct EIGEN_CAT(EIGEN_CAT(gl_,FUNC),_impl)<XprType, SCALAR, 1, SIZE, true> {                     \
+    inline static void run(ARG1 a, EIGEN_GL_EVAL(EIGEN_GL_MAKE_CONST_##CONST) XprType& p) { FUNC##SUFFIX(a,p.data()); }         \
+  };                                                                                                                            \
+}
+
+EIGEN_GL_FUNC1_DECLARATION       (glGet,GLenum,_)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glGet,GLenum,_,float,  4,4,Floatv)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glGet,GLenum,_,double, 4,4,Doublev)
+
+// glUniform API
+
+#ifdef GL_VERSION_2_0
+
+inline void glUniform2fv_ei  (GLint loc, const float* v)         { glUniform2fv(loc,1,v); }
+inline void glUniform2iv_ei  (GLint loc, const int* v)           { glUniform2iv(loc,1,v); }
+
+inline void glUniform3fv_ei  (GLint loc, const float* v)         { glUniform3fv(loc,1,v); }
+inline void glUniform3iv_ei  (GLint loc, const int* v)           { glUniform3iv(loc,1,v); }
+
+inline void glUniform4fv_ei  (GLint loc, const float* v)         { glUniform4fv(loc,1,v); }
+inline void glUniform4iv_ei  (GLint loc, const int* v)           { glUniform4iv(loc,1,v); }
+
+inline void glUniformMatrix2fv_ei  (GLint loc, const float* v)         { glUniformMatrix2fv(loc,1,false,v); }
+inline void glUniformMatrix3fv_ei  (GLint loc, const float* v)         { glUniformMatrix3fv(loc,1,false,v); }
+inline void glUniformMatrix4fv_ei  (GLint loc, const float* v)         { glUniformMatrix4fv(loc,1,false,v); }
+
+
+EIGEN_GL_FUNC1_DECLARATION       (glUniform,GLint,const)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,float,        2,2fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,int,          2,2iv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,float,        3,3fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,int,          3,3iv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,float,        4,4fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,int,          4,4iv_ei)
+
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        2,2,Matrix2fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        3,3,Matrix3fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        4,4,Matrix4fv_ei)
+
+#endif
+
+#ifdef GL_VERSION_2_1
+
+inline void glUniformMatrix2x3fv_ei(GLint loc, const float* v)         { glUniformMatrix2x3fv(loc,1,false,v); }
+inline void glUniformMatrix3x2fv_ei(GLint loc, const float* v)         { glUniformMatrix3x2fv(loc,1,false,v); }
+inline void glUniformMatrix2x4fv_ei(GLint loc, const float* v)         { glUniformMatrix2x4fv(loc,1,false,v); }
+inline void glUniformMatrix4x2fv_ei(GLint loc, const float* v)         { glUniformMatrix4x2fv(loc,1,false,v); }
+inline void glUniformMatrix3x4fv_ei(GLint loc, const float* v)         { glUniformMatrix3x4fv(loc,1,false,v); }
+inline void glUniformMatrix4x3fv_ei(GLint loc, const float* v)         { glUniformMatrix4x3fv(loc,1,false,v); }
+
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        2,3,Matrix2x3fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        3,2,Matrix3x2fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        2,4,Matrix2x4fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        4,2,Matrix4x2fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        3,4,Matrix3x4fv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_MAT(glUniform,GLint,const,float,        4,3,Matrix4x3fv_ei)
+
+#endif
+
+#ifdef GL_VERSION_3_0
+
+inline void glUniform2uiv_ei (GLint loc, const unsigned int* v)  { glUniform2uiv(loc,1,v); }
+inline void glUniform3uiv_ei (GLint loc, const unsigned int* v)  { glUniform3uiv(loc,1,v); }
+inline void glUniform4uiv_ei (GLint loc, const unsigned int* v)  { glUniform4uiv(loc,1,v); }
+
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,unsigned int, 2,2uiv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,unsigned int, 3,3uiv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,unsigned int, 4,4uiv_ei)
+
+#endif
+
+#ifdef GL_ARB_gpu_shader_fp64
+inline void glUniform2dv_ei  (GLint loc, const double* v)        { glUniform2dv(loc,1,v); }
+inline void glUniform3dv_ei  (GLint loc, const double* v)        { glUniform3dv(loc,1,v); }
+inline void glUniform4dv_ei  (GLint loc, const double* v)        { glUniform4dv(loc,1,v); }
+
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,double,       2,2dv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,double,       3,3dv_ei)
+EIGEN_GL_FUNC1_SPECIALIZATION_VEC(glUniform,GLint,const,double,       4,4dv_ei)
+#endif
+
+
+//@}
+
+}
+
+#endif // EIGEN_OPENGL_MODULE
diff --git a/src/EigenUnsupported/Polynomials b/src/EigenUnsupported/Polynomials
new file mode 100644
index 0000000..32ce2a2
--- /dev/null
+++ b/src/EigenUnsupported/Polynomials
@@ -0,0 +1,137 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_POLYNOMIALS_MODULE_H
+#define EIGEN_POLYNOMIALS_MODULE_H
+
+#include "../../Eigen/Core"
+
+#include "../../Eigen/Eigenvalues"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+// Note that EIGEN_HIDE_HEAVY_CODE has to be defined per module
+#if (defined EIGEN_EXTERN_INSTANTIATIONS) && (EIGEN_EXTERN_INSTANTIATIONS>=2)
+  #ifndef EIGEN_HIDE_HEAVY_CODE
+  #define EIGEN_HIDE_HEAVY_CODE
+  #endif
+#elif defined EIGEN_HIDE_HEAVY_CODE
+  #undef EIGEN_HIDE_HEAVY_CODE
+#endif
+
+/**
+  * \defgroup Polynomials_Module Polynomials module
+  * \brief This module provides a QR based polynomial solver.
+	*
+  * To use this module, add
+  * \code
+  * #include <unsupported/Eigen/Polynomials>
+  * \endcode
+	* at the start of your source file.
+  */
+
+#include "src/Polynomials/PolynomialUtils.h"
+#include "src/Polynomials/Companion.h"
+#include "src/Polynomials/PolynomialSolver.h"
+
+/**
+	\page polynomials Polynomials defines functions for dealing with polynomials
+	and a QR based polynomial solver.
+	\ingroup Polynomials_Module
+
+	The remainder of the page documents first the functions for evaluating, computing
+	polynomials, computing estimates about polynomials and next the QR based polynomial
+	solver.
+
+	\section polynomialUtils convenient functions to deal with polynomials
+	\subsection roots_to_monicPolynomial
+	The function
+	\code
+	void roots_to_monicPolynomial( const RootVector& rv, Polynomial& poly )
+	\endcode
+	computes the coefficients \f$ a_i \f$ of
+
+	\f$ p(x) = a_0 + a_{1}x + ... + a_{n-1}x^{n-1} + x^n \f$
+
+	where \f$ p \f$ is known through its roots i.e. \f$ p(x) = (x-r_1)(x-r_2)...(x-r_n) \f$.
+
+	\subsection poly_eval
+	The function
+	\code
+	T poly_eval( const Polynomials& poly, const T& x )
+	\endcode
+	evaluates a polynomial at a given point using stabilized H&ouml;rner method.
+
+	The following code: first computes the coefficients in the monomial basis of the monic polynomial that has the provided roots;
+	then, it evaluates the computed polynomial, using a stabilized H&ouml;rner method.
+
+	\include PolynomialUtils1.cpp
+  Output: \verbinclude PolynomialUtils1.out
+
+	\subsection Cauchy bounds
+	The function
+	\code
+	Real cauchy_max_bound( const Polynomial& poly )
+	\endcode
+	provides a maximum bound (the Cauchy one: \f$C(p)\f$) for the absolute value of a root of the given polynomial i.e.
+	\f$ \forall r_i \f$ root of \f$ p(x) = \sum_{k=0}^d a_k x^k \f$,
+	\f$ |r_i| \le C(p) = \sum_{k=0}^{d} \left | \frac{a_k}{a_d} \right | \f$
+	The leading coefficient \f$ p \f$: should be non zero \f$a_d \neq 0\f$.
+
+
+	The function
+	\code
+	Real cauchy_min_bound( const Polynomial& poly )
+	\endcode
+	provides a minimum bound (the Cauchy one: \f$c(p)\f$) for the absolute value of a non zero root of the given polynomial i.e.
+	\f$ \forall r_i \neq 0 \f$ root of \f$ p(x) = \sum_{k=0}^d a_k x^k \f$,
+	\f$ |r_i| \ge c(p) = \left( \sum_{k=0}^{d} \left | \frac{a_k}{a_0} \right | \right)^{-1} \f$
+
+
+
+
+	\section QR polynomial solver class
+	Computes the complex roots of a polynomial by computing the eigenvalues of the associated companion matrix with the QR algorithm.
+	
+	The roots of \f$ p(x) = a_0 + a_1 x + a_2 x^2 + a_{3} x^3 + x^4 \f$ are the eigenvalues of
+	\f$
+	\left [
+	\begin{array}{cccc}
+	0 & 0 &  0 & a_0 \\
+	1 & 0 &  0 & a_1 \\
+	0 & 1 &  0 & a_2 \\
+	0 & 0 &  1 & a_3
+	\end{array} \right ]
+	\f$
+
+	However, the QR algorithm is not guaranteed to converge when there are several eigenvalues with same modulus.
+
+	Therefore the current polynomial solver is guaranteed to provide a correct result only when the complex roots \f$r_1,r_2,...,r_d\f$ have distinct moduli i.e.
+	
+	\f$ \forall i,j \in [1;d],~ \| r_i \| \neq \| r_j \| \f$.
+
+	With 32bit (float) floating types this problem shows up frequently.
+  However, almost always, correct accuracy is reached even in these cases for 64bit
+  (double) floating types and small polynomial degree (<20).
+
+	\include PolynomialSolver1.cpp
+	
+	In the above example:
+	
+	-# a simple use of the polynomial solver is shown;
+	-# the accuracy problem with the QR algorithm is presented: a polynomial with almost conjugate roots is provided to the solver.
+	Those roots have almost same module therefore the QR algorithm failed to converge: the accuracy
+	of the last root is bad;
+	-# a simple way to circumvent the problem is shown: use doubles instead of floats.
+
+  Output: \verbinclude PolynomialSolver1.out
+*/
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_POLYNOMIALS_MODULE_H
diff --git a/src/EigenUnsupported/Skyline b/src/EigenUnsupported/Skyline
new file mode 100644
index 0000000..ebdf143
--- /dev/null
+++ b/src/EigenUnsupported/Skyline
@@ -0,0 +1,39 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SKYLINE_MODULE_H
+#define EIGEN_SKYLINE_MODULE_H
+
+
+#include "../../Eigen/Core"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#include <map>
+#include <cstdlib>
+#include <cstring>
+#include <algorithm>
+
+/**
+ *  \defgroup Skyline_Module Skyline module
+ *
+ *
+ *
+ *
+ */
+
+#include "src/Skyline/SkylineUtil.h"
+#include "src/Skyline/SkylineMatrixBase.h"
+#include "src/Skyline/SkylineStorage.h"
+#include "src/Skyline/SkylineMatrix.h"
+#include "src/Skyline/SkylineInplaceLU.h"
+#include "src/Skyline/SkylineProduct.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SKYLINE_MODULE_H
diff --git a/src/EigenUnsupported/SparseExtra b/src/EigenUnsupported/SparseExtra
new file mode 100644
index 0000000..ba5cbd6
--- /dev/null
+++ b/src/EigenUnsupported/SparseExtra
@@ -0,0 +1,54 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <g.gael@free.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_EXTRA_MODULE_H
+#define EIGEN_SPARSE_EXTRA_MODULE_H
+
+#include "../../Eigen/Sparse"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#include <vector>
+#include <map>
+#include <cstdlib>
+#include <cstring>
+#include <algorithm>
+#include <fstream>
+#include <sstream>
+
+#ifdef EIGEN_GOOGLEHASH_SUPPORT
+  #include <google/dense_hash_map>
+  #include <google/sparse_hash_map>
+#endif
+
+/**
+  * \defgroup SparseExtra_Module SparseExtra module
+  *
+  * This module contains some experimental features extending the sparse module.
+  *
+  * \code
+  * #include <Eigen/SparseExtra>
+  * \endcode
+  */
+
+
+#include "src/SparseExtra/DynamicSparseMatrix.h"
+#include "src/SparseExtra/BlockOfDynamicSparseMatrix.h"
+#include "src/SparseExtra/RandomSetter.h"
+
+#include "src/SparseExtra/MarketIO.h"
+
+#if !defined(_WIN32)
+#include <dirent.h>
+#include "src/SparseExtra/MatrixMarketIterator.h"
+#endif
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SPARSE_EXTRA_MODULE_H
diff --git a/src/EigenUnsupported/SpecialFunctions b/src/EigenUnsupported/SpecialFunctions
new file mode 100644
index 0000000..f6a2460
--- /dev/null
+++ b/src/EigenUnsupported/SpecialFunctions
@@ -0,0 +1,103 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Gael Guennebaud <g.gael@free.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPECIALFUNCTIONS_MODULE
+#define EIGEN_SPECIALFUNCTIONS_MODULE
+
+#include <math.h>
+
+#include "../../Eigen/Core"
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+namespace Eigen {
+
+/**
+  * \defgroup SpecialFunctions_Module Special math functions module
+  *
+  * This module features additional coefficient-wise math functions available
+  * within the numext:: namespace for the scalar version, and as method and/or free
+  * functions of Array. Those include:
+  *
+  * - erf
+  * - erfc
+  * - lgamma
+  * - igamma
+  * - igamma_der_a
+  * - gamma_sample_der_alpha
+  * - igammac
+  * - digamma
+  * - ndtri
+  * - polygamma
+  * - zeta
+  * - betainc
+  *
+  * Bessel Functions
+  * - bessel_i0
+  * - bessel_i0e
+  * - bessel_i1
+  * - bessel_i1e
+  * - bessel_j0
+  * - bessel_j1
+  * - bessel_k0
+  * - bessel_k0e
+  * - bessel_k1
+  * - bessel_k1e
+  * - bessel_y0
+  * - bessel_y1
+  *
+  * \code
+  * #include <unsupported/Eigen/SpecialFunctions>
+  * \endcode
+  */
+//@{
+
+}
+
+#include "src/SpecialFunctions/BesselFunctionsImpl.h"
+#include "src/SpecialFunctions/BesselFunctionsBFloat16.h"
+#include "src/SpecialFunctions/BesselFunctionsHalf.h"
+#include "src/SpecialFunctions/BesselFunctionsPacketMath.h"
+#include "src/SpecialFunctions/BesselFunctionsFunctors.h"
+#include "src/SpecialFunctions/BesselFunctionsArrayAPI.h"
+#include "src/SpecialFunctions/SpecialFunctionsImpl.h"
+#if defined(EIGEN_HIPCC)
+#include "src/SpecialFunctions/HipVectorCompatibility.h"
+#endif
+#include "src/SpecialFunctions/SpecialFunctionsBFloat16.h"
+#include "src/SpecialFunctions/SpecialFunctionsHalf.h"
+#include "src/SpecialFunctions/SpecialFunctionsPacketMath.h"
+#include "src/SpecialFunctions/SpecialFunctionsFunctors.h"
+#include "src/SpecialFunctions/SpecialFunctionsArrayAPI.h"
+
+#if defined EIGEN_VECTORIZE_AVX512
+  #include "src/SpecialFunctions/arch/AVX/BesselFunctions.h"
+  #include "src/SpecialFunctions/arch/AVX/SpecialFunctions.h"
+  #include "src/SpecialFunctions/arch/AVX512/BesselFunctions.h"
+  #include "src/SpecialFunctions/arch/AVX512/SpecialFunctions.h"
+#elif defined EIGEN_VECTORIZE_AVX
+  #include "src/SpecialFunctions/arch/AVX/BesselFunctions.h"
+  #include "src/SpecialFunctions/arch/AVX/SpecialFunctions.h"
+#elif defined EIGEN_VECTORIZE_NEON
+  #include "src/SpecialFunctions/arch/NEON/BesselFunctions.h"
+  #include "src/SpecialFunctions/arch/NEON/SpecialFunctions.h"
+#endif
+
+#if defined EIGEN_VECTORIZE_GPU
+  #include "src/SpecialFunctions/arch/GPU/SpecialFunctions.h"
+#endif
+
+namespace Eigen {
+//@}
+}
+
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SPECIALFUNCTIONS_MODULE
diff --git a/src/EigenUnsupported/Splines b/src/EigenUnsupported/Splines
new file mode 100644
index 0000000..2ca5813
--- /dev/null
+++ b/src/EigenUnsupported/Splines
@@ -0,0 +1,35 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 20010-2011 Hauke Heibel <hauke.heibel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPLINES_MODULE_H
+#define EIGEN_SPLINES_MODULE_H
+
+namespace Eigen 
+{
+/**
+  * \defgroup Splines_Module Spline and spline fitting module
+  *
+  * This module provides a simple multi-dimensional spline class while
+  * offering most basic functionality to fit a spline to point sets.
+  *
+  * \code
+  * #include <unsupported/Eigen/Splines>
+  * \endcode
+  */
+}
+
+#include "../../Eigen/src/Core/util/DisableStupidWarnings.h"
+
+#include "src/Splines/SplineFwd.h"
+#include "src/Splines/Spline.h"
+#include "src/Splines/SplineFitting.h"
+
+#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SPLINES_MODULE_H
diff --git a/src/EigenUnsupported/src/AutoDiff/AutoDiffJacobian.h b/src/EigenUnsupported/src/AutoDiff/AutoDiffJacobian.h
new file mode 100644
index 0000000..33b6c39
--- /dev/null
+++ b/src/EigenUnsupported/src/AutoDiff/AutoDiffJacobian.h
@@ -0,0 +1,108 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_AUTODIFF_JACOBIAN_H
+#define EIGEN_AUTODIFF_JACOBIAN_H
+
+namespace Eigen
+{
+
+template<typename Functor> class AutoDiffJacobian : public Functor
+{
+public:
+  AutoDiffJacobian() : Functor() {}
+  AutoDiffJacobian(const Functor& f) : Functor(f) {}
+
+  // forward constructors
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+  template<typename... T>
+  AutoDiffJacobian(const T& ...Values) : Functor(Values...) {}
+#else
+  template<typename T0>
+  AutoDiffJacobian(const T0& a0) : Functor(a0) {}
+  template<typename T0, typename T1>
+  AutoDiffJacobian(const T0& a0, const T1& a1) : Functor(a0, a1) {}
+  template<typename T0, typename T1, typename T2>
+  AutoDiffJacobian(const T0& a0, const T1& a1, const T2& a2) : Functor(a0, a1, a2) {}
+#endif
+
+  typedef typename Functor::InputType InputType;
+  typedef typename Functor::ValueType ValueType;
+  typedef typename ValueType::Scalar Scalar;
+
+  enum {
+    InputsAtCompileTime = InputType::RowsAtCompileTime,
+    ValuesAtCompileTime = ValueType::RowsAtCompileTime
+  };
+
+  typedef Matrix<Scalar, ValuesAtCompileTime, InputsAtCompileTime> JacobianType;
+  typedef typename JacobianType::Index Index;
+
+  typedef Matrix<Scalar, InputsAtCompileTime, 1> DerivativeType;
+  typedef AutoDiffScalar<DerivativeType> ActiveScalar;
+
+  typedef Matrix<ActiveScalar, InputsAtCompileTime, 1> ActiveInput;
+  typedef Matrix<ActiveScalar, ValuesAtCompileTime, 1> ActiveValue;
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+  // Some compilers don't accept variadic parameters after a default parameter,
+  // i.e., we can't just write _jac=0 but we need to overload operator():
+  EIGEN_STRONG_INLINE
+  void operator() (const InputType& x, ValueType* v) const
+  {
+      this->operator()(x, v, 0);
+  }
+  template<typename... ParamsType>
+  void operator() (const InputType& x, ValueType* v, JacobianType* _jac,
+                   const ParamsType&... Params) const
+#else
+  void operator() (const InputType& x, ValueType* v, JacobianType* _jac=0) const
+#endif
+  {
+    eigen_assert(v!=0);
+
+    if (!_jac)
+    {
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+      Functor::operator()(x, v, Params...);
+#else
+      Functor::operator()(x, v);
+#endif
+      return;
+    }
+
+    JacobianType& jac = *_jac;
+
+    ActiveInput ax = x.template cast<ActiveScalar>();
+    ActiveValue av(jac.rows());
+
+    if(InputsAtCompileTime==Dynamic)
+      for (Index j=0; j<jac.rows(); j++)
+        av[j].derivatives().resize(x.rows());
+
+    for (Index i=0; i<jac.cols(); i++)
+      ax[i].derivatives() = DerivativeType::Unit(x.rows(),i);
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+    Functor::operator()(ax, &av, Params...);
+#else
+    Functor::operator()(ax, &av);
+#endif
+
+    for (Index i=0; i<jac.rows(); i++)
+    {
+      (*v)[i] = av[i].value();
+      jac.row(i) = av[i].derivatives();
+    }
+  }
+};
+
+}
+
+#endif // EIGEN_AUTODIFF_JACOBIAN_H
diff --git a/src/EigenUnsupported/src/AutoDiff/AutoDiffScalar.h b/src/EigenUnsupported/src/AutoDiff/AutoDiffScalar.h
new file mode 100755
index 0000000..0f166e3
--- /dev/null
+++ b/src/EigenUnsupported/src/AutoDiff/AutoDiffScalar.h
@@ -0,0 +1,730 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_AUTODIFF_SCALAR_H
+#define EIGEN_AUTODIFF_SCALAR_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename A, typename B>
+struct make_coherent_impl {
+  static void run(A&, B&) {}
+};
+
+// resize a to match b is a.size()==0, and conversely.
+template<typename A, typename B>
+void make_coherent(const A& a, const B&b)
+{
+  make_coherent_impl<A,B>::run(a.const_cast_derived(), b.const_cast_derived());
+}
+
+template<typename DerivativeType, bool Enable> struct auto_diff_special_op;
+
+} // end namespace internal
+
+template<typename DerivativeType> class AutoDiffScalar;
+
+template<typename NewDerType>
+inline AutoDiffScalar<NewDerType> MakeAutoDiffScalar(const typename NewDerType::Scalar& value, const NewDerType &der) {
+  return AutoDiffScalar<NewDerType>(value,der);
+}
+
+/** \class AutoDiffScalar
+  * \brief A scalar type replacement with automatic differentiation capability
+  *
+  * \param DerivativeType the vector type used to store/represent the derivatives. The base scalar type
+  *                 as well as the number of derivatives to compute are determined from this type.
+  *                 Typical choices include, e.g., \c Vector4f for 4 derivatives, or \c VectorXf
+  *                 if the number of derivatives is not known at compile time, and/or, the number
+  *                 of derivatives is large.
+  *                 Note that DerivativeType can also be a reference (e.g., \c VectorXf&) to wrap a
+  *                 existing vector into an AutoDiffScalar.
+  *                 Finally, DerivativeType can also be any Eigen compatible expression.
+  *
+  * This class represents a scalar value while tracking its respective derivatives using Eigen's expression
+  * template mechanism.
+  *
+  * It supports the following list of global math function:
+  *  - std::abs, std::sqrt, std::pow, std::exp, std::log, std::sin, std::cos,
+  *  - internal::abs, internal::sqrt, numext::pow, internal::exp, internal::log, internal::sin, internal::cos,
+  *  - internal::conj, internal::real, internal::imag, numext::abs2.
+  *
+  * AutoDiffScalar can be used as the scalar type of an Eigen::Matrix object. However,
+  * in that case, the expression template mechanism only occurs at the top Matrix level,
+  * while derivatives are computed right away.
+  *
+  */
+
+template<typename DerivativeType>
+class AutoDiffScalar
+  : public internal::auto_diff_special_op
+            <DerivativeType, !internal::is_same<typename internal::traits<typename internal::remove_all<DerivativeType>::type>::Scalar,
+                                          typename NumTraits<typename internal::traits<typename internal::remove_all<DerivativeType>::type>::Scalar>::Real>::value>
+{
+  public:
+    typedef internal::auto_diff_special_op
+            <DerivativeType, !internal::is_same<typename internal::traits<typename internal::remove_all<DerivativeType>::type>::Scalar,
+                       typename NumTraits<typename internal::traits<typename internal::remove_all<DerivativeType>::type>::Scalar>::Real>::value> Base;
+    typedef typename internal::remove_all<DerivativeType>::type DerType;
+    typedef typename internal::traits<DerType>::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real Real;
+
+    using Base::operator+;
+    using Base::operator*;
+
+    /** Default constructor without any initialization. */
+    AutoDiffScalar() {}
+
+    /** Constructs an active scalar from its \a value,
+        and initializes the \a nbDer derivatives such that it corresponds to the \a derNumber -th variable */
+    AutoDiffScalar(const Scalar& value, int nbDer, int derNumber)
+      : m_value(value), m_derivatives(DerType::Zero(nbDer))
+    {
+      m_derivatives.coeffRef(derNumber) = Scalar(1);
+    }
+
+    /** Conversion from a scalar constant to an active scalar.
+      * The derivatives are set to zero. */
+    /*explicit*/ AutoDiffScalar(const Real& value)
+      : m_value(value)
+    {
+      if(m_derivatives.size()>0)
+        m_derivatives.setZero();
+    }
+
+    /** Constructs an active scalar from its \a value and derivatives \a der */
+    AutoDiffScalar(const Scalar& value, const DerType& der)
+      : m_value(value), m_derivatives(der)
+    {}
+
+    template<typename OtherDerType>
+    AutoDiffScalar(const AutoDiffScalar<OtherDerType>& other
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    , typename internal::enable_if<
+            internal::is_same<Scalar, typename internal::traits<typename internal::remove_all<OtherDerType>::type>::Scalar>::value
+        &&  internal::is_convertible<OtherDerType,DerType>::value , void*>::type = 0
+#endif
+    )
+      : m_value(other.value()), m_derivatives(other.derivatives())
+    {}
+
+    friend  std::ostream & operator << (std::ostream & s, const AutoDiffScalar& a)
+    {
+      return s << a.value();
+    }
+
+    AutoDiffScalar(const AutoDiffScalar& other)
+      : m_value(other.value()), m_derivatives(other.derivatives())
+    {}
+
+    template<typename OtherDerType>
+    inline AutoDiffScalar& operator=(const AutoDiffScalar<OtherDerType>& other)
+    {
+      m_value = other.value();
+      m_derivatives = other.derivatives();
+      return *this;
+    }
+
+    inline AutoDiffScalar& operator=(const AutoDiffScalar& other)
+    {
+      m_value = other.value();
+      m_derivatives = other.derivatives();
+      return *this;
+    }
+
+    inline AutoDiffScalar& operator=(const Scalar& other)
+    {
+      m_value = other;
+      if(m_derivatives.size()>0)
+        m_derivatives.setZero();
+      return *this;
+    }
+
+//     inline operator const Scalar& () const { return m_value; }
+//     inline operator Scalar& () { return m_value; }
+
+    inline const Scalar& value() const { return m_value; }
+    inline Scalar& value() { return m_value; }
+
+    inline const DerType& derivatives() const { return m_derivatives; }
+    inline DerType& derivatives() { return m_derivatives; }
+
+    inline bool operator< (const Scalar& other) const  { return m_value <  other; }
+    inline bool operator<=(const Scalar& other) const  { return m_value <= other; }
+    inline bool operator> (const Scalar& other) const  { return m_value >  other; }
+    inline bool operator>=(const Scalar& other) const  { return m_value >= other; }
+    inline bool operator==(const Scalar& other) const  { return m_value == other; }
+    inline bool operator!=(const Scalar& other) const  { return m_value != other; }
+
+    friend inline bool operator< (const Scalar& a, const AutoDiffScalar& b) { return a <  b.value(); }
+    friend inline bool operator<=(const Scalar& a, const AutoDiffScalar& b) { return a <= b.value(); }
+    friend inline bool operator> (const Scalar& a, const AutoDiffScalar& b) { return a >  b.value(); }
+    friend inline bool operator>=(const Scalar& a, const AutoDiffScalar& b) { return a >= b.value(); }
+    friend inline bool operator==(const Scalar& a, const AutoDiffScalar& b) { return a == b.value(); }
+    friend inline bool operator!=(const Scalar& a, const AutoDiffScalar& b) { return a != b.value(); }
+
+    template<typename OtherDerType> inline bool operator< (const AutoDiffScalar<OtherDerType>& b) const  { return m_value <  b.value(); }
+    template<typename OtherDerType> inline bool operator<=(const AutoDiffScalar<OtherDerType>& b) const  { return m_value <= b.value(); }
+    template<typename OtherDerType> inline bool operator> (const AutoDiffScalar<OtherDerType>& b) const  { return m_value >  b.value(); }
+    template<typename OtherDerType> inline bool operator>=(const AutoDiffScalar<OtherDerType>& b) const  { return m_value >= b.value(); }
+    template<typename OtherDerType> inline bool operator==(const AutoDiffScalar<OtherDerType>& b) const  { return m_value == b.value(); }
+    template<typename OtherDerType> inline bool operator!=(const AutoDiffScalar<OtherDerType>& b) const  { return m_value != b.value(); }
+
+    inline const AutoDiffScalar<DerType&> operator+(const Scalar& other) const
+    {
+      return AutoDiffScalar<DerType&>(m_value + other, m_derivatives);
+    }
+
+    friend inline const AutoDiffScalar<DerType&> operator+(const Scalar& a, const AutoDiffScalar& b)
+    {
+      return AutoDiffScalar<DerType&>(a + b.value(), b.derivatives());
+    }
+
+//     inline const AutoDiffScalar<DerType&> operator+(const Real& other) const
+//     {
+//       return AutoDiffScalar<DerType&>(m_value + other, m_derivatives);
+//     }
+
+//     friend inline const AutoDiffScalar<DerType&> operator+(const Real& a, const AutoDiffScalar& b)
+//     {
+//       return AutoDiffScalar<DerType&>(a + b.value(), b.derivatives());
+//     }
+
+    inline AutoDiffScalar& operator+=(const Scalar& other)
+    {
+      value() += other;
+      return *this;
+    }
+
+    template<typename OtherDerType>
+    inline const AutoDiffScalar<CwiseBinaryOp<internal::scalar_sum_op<Scalar>,const DerType,const typename internal::remove_all<OtherDerType>::type> >
+    operator+(const AutoDiffScalar<OtherDerType>& other) const
+    {
+      internal::make_coherent(m_derivatives, other.derivatives());
+      return AutoDiffScalar<CwiseBinaryOp<internal::scalar_sum_op<Scalar>,const DerType,const typename internal::remove_all<OtherDerType>::type> >(
+        m_value + other.value(),
+        m_derivatives + other.derivatives());
+    }
+
+    template<typename OtherDerType>
+    inline AutoDiffScalar&
+    operator+=(const AutoDiffScalar<OtherDerType>& other)
+    {
+      (*this) = (*this) + other;
+      return *this;
+    }
+
+    inline const AutoDiffScalar<DerType&> operator-(const Scalar& b) const
+    {
+      return AutoDiffScalar<DerType&>(m_value - b, m_derivatives);
+    }
+
+    friend inline const AutoDiffScalar<CwiseUnaryOp<internal::scalar_opposite_op<Scalar>, const DerType> >
+    operator-(const Scalar& a, const AutoDiffScalar& b)
+    {
+      return AutoDiffScalar<CwiseUnaryOp<internal::scalar_opposite_op<Scalar>, const DerType> >
+            (a - b.value(), -b.derivatives());
+    }
+
+    inline AutoDiffScalar& operator-=(const Scalar& other)
+    {
+      value() -= other;
+      return *this;
+    }
+
+    template<typename OtherDerType>
+    inline const AutoDiffScalar<CwiseBinaryOp<internal::scalar_difference_op<Scalar>, const DerType,const typename internal::remove_all<OtherDerType>::type> >
+    operator-(const AutoDiffScalar<OtherDerType>& other) const
+    {
+      internal::make_coherent(m_derivatives, other.derivatives());
+      return AutoDiffScalar<CwiseBinaryOp<internal::scalar_difference_op<Scalar>, const DerType,const typename internal::remove_all<OtherDerType>::type> >(
+        m_value - other.value(),
+        m_derivatives - other.derivatives());
+    }
+
+    template<typename OtherDerType>
+    inline AutoDiffScalar&
+    operator-=(const AutoDiffScalar<OtherDerType>& other)
+    {
+      *this = *this - other;
+      return *this;
+    }
+
+    inline const AutoDiffScalar<CwiseUnaryOp<internal::scalar_opposite_op<Scalar>, const DerType> >
+    operator-() const
+    {
+      return AutoDiffScalar<CwiseUnaryOp<internal::scalar_opposite_op<Scalar>, const DerType> >(
+        -m_value,
+        -m_derivatives);
+    }
+
+    inline const AutoDiffScalar<EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DerType,Scalar,product) >
+    operator*(const Scalar& other) const
+    {
+      return MakeAutoDiffScalar(m_value * other, m_derivatives * other);
+    }
+
+    friend inline const AutoDiffScalar<EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DerType,Scalar,product) >
+    operator*(const Scalar& other, const AutoDiffScalar& a)
+    {
+      return MakeAutoDiffScalar(a.value() * other, a.derivatives() * other);
+    }
+
+//     inline const AutoDiffScalar<typename CwiseUnaryOp<internal::scalar_multiple_op<Real>, DerType>::Type >
+//     operator*(const Real& other) const
+//     {
+//       return AutoDiffScalar<typename CwiseUnaryOp<internal::scalar_multiple_op<Real>, DerType>::Type >(
+//         m_value * other,
+//         (m_derivatives * other));
+//     }
+//
+//     friend inline const AutoDiffScalar<typename CwiseUnaryOp<internal::scalar_multiple_op<Real>, DerType>::Type >
+//     operator*(const Real& other, const AutoDiffScalar& a)
+//     {
+//       return AutoDiffScalar<typename CwiseUnaryOp<internal::scalar_multiple_op<Real>, DerType>::Type >(
+//         a.value() * other,
+//         a.derivatives() * other);
+//     }
+
+    inline const AutoDiffScalar<EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DerType,Scalar,product) >
+    operator/(const Scalar& other) const
+    {
+      return MakeAutoDiffScalar(m_value / other, (m_derivatives * (Scalar(1)/other)));
+    }
+
+    friend inline const AutoDiffScalar<EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DerType,Scalar,product) >
+    operator/(const Scalar& other, const AutoDiffScalar& a)
+    {
+      return MakeAutoDiffScalar(other / a.value(), a.derivatives() * (Scalar(-other) / (a.value()*a.value())));
+    }
+
+//     inline const AutoDiffScalar<typename CwiseUnaryOp<internal::scalar_multiple_op<Real>, DerType>::Type >
+//     operator/(const Real& other) const
+//     {
+//       return AutoDiffScalar<typename CwiseUnaryOp<internal::scalar_multiple_op<Real>, DerType>::Type >(
+//         m_value / other,
+//         (m_derivatives * (Real(1)/other)));
+//     }
+//
+//     friend inline const AutoDiffScalar<typename CwiseUnaryOp<internal::scalar_multiple_op<Real>, DerType>::Type >
+//     operator/(const Real& other, const AutoDiffScalar& a)
+//     {
+//       return AutoDiffScalar<typename CwiseUnaryOp<internal::scalar_multiple_op<Real>, DerType>::Type >(
+//         other / a.value(),
+//         a.derivatives() * (-Real(1)/other));
+//     }
+
+    template<typename OtherDerType>
+    inline const AutoDiffScalar<EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(
+        CwiseBinaryOp<internal::scalar_difference_op<Scalar> EIGEN_COMMA
+          const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DerType,Scalar,product) EIGEN_COMMA
+          const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(typename internal::remove_all<OtherDerType>::type,Scalar,product) >,Scalar,product) >
+    operator/(const AutoDiffScalar<OtherDerType>& other) const
+    {
+      internal::make_coherent(m_derivatives, other.derivatives());
+      return MakeAutoDiffScalar(
+        m_value / other.value(),
+          ((m_derivatives * other.value()) - (other.derivatives() * m_value))
+        * (Scalar(1)/(other.value()*other.value())));
+    }
+
+    template<typename OtherDerType>
+    inline const AutoDiffScalar<CwiseBinaryOp<internal::scalar_sum_op<Scalar>,
+        const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DerType,Scalar,product),
+        const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(typename internal::remove_all<OtherDerType>::type,Scalar,product) > >
+    operator*(const AutoDiffScalar<OtherDerType>& other) const
+    {
+      internal::make_coherent(m_derivatives, other.derivatives());
+      return MakeAutoDiffScalar(
+        m_value * other.value(),
+        (m_derivatives * other.value()) + (other.derivatives() * m_value));
+    }
+
+    inline AutoDiffScalar& operator*=(const Scalar& other)
+    {
+      *this = *this * other;
+      return *this;
+    }
+
+    template<typename OtherDerType>
+    inline AutoDiffScalar& operator*=(const AutoDiffScalar<OtherDerType>& other)
+    {
+      *this = *this * other;
+      return *this;
+    }
+
+    inline AutoDiffScalar& operator/=(const Scalar& other)
+    {
+      *this = *this / other;
+      return *this;
+    }
+
+    template<typename OtherDerType>
+    inline AutoDiffScalar& operator/=(const AutoDiffScalar<OtherDerType>& other)
+    {
+      *this = *this / other;
+      return *this;
+    }
+
+  protected:
+    Scalar m_value;
+    DerType m_derivatives;
+
+};
+
+namespace internal {
+
+template<typename DerivativeType>
+struct auto_diff_special_op<DerivativeType, true>
+//   : auto_diff_scalar_op<DerivativeType, typename NumTraits<Scalar>::Real,
+//                            is_same<Scalar,typename NumTraits<Scalar>::Real>::value>
+{
+  typedef typename remove_all<DerivativeType>::type DerType;
+  typedef typename traits<DerType>::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real Real;
+
+//   typedef auto_diff_scalar_op<DerivativeType, typename NumTraits<Scalar>::Real,
+//                            is_same<Scalar,typename NumTraits<Scalar>::Real>::value> Base;
+
+//   using Base::operator+;
+//   using Base::operator+=;
+//   using Base::operator-;
+//   using Base::operator-=;
+//   using Base::operator*;
+//   using Base::operator*=;
+
+  const AutoDiffScalar<DerivativeType>& derived() const { return *static_cast<const AutoDiffScalar<DerivativeType>*>(this); }
+  AutoDiffScalar<DerivativeType>& derived() { return *static_cast<AutoDiffScalar<DerivativeType>*>(this); }
+
+
+  inline const AutoDiffScalar<DerType&> operator+(const Real& other) const
+  {
+    return AutoDiffScalar<DerType&>(derived().value() + other, derived().derivatives());
+  }
+
+  friend inline const AutoDiffScalar<DerType&> operator+(const Real& a, const AutoDiffScalar<DerivativeType>& b)
+  {
+    return AutoDiffScalar<DerType&>(a + b.value(), b.derivatives());
+  }
+
+  inline AutoDiffScalar<DerivativeType>& operator+=(const Real& other)
+  {
+    derived().value() += other;
+    return derived();
+  }
+
+
+  inline const AutoDiffScalar<typename CwiseUnaryOp<bind2nd_op<scalar_product_op<Scalar,Real> >, DerType>::Type >
+  operator*(const Real& other) const
+  {
+    return AutoDiffScalar<typename CwiseUnaryOp<bind2nd_op<scalar_product_op<Scalar,Real> >, DerType>::Type >(
+      derived().value() * other,
+      derived().derivatives() * other);
+  }
+
+  friend inline const AutoDiffScalar<typename CwiseUnaryOp<bind1st_op<scalar_product_op<Real,Scalar> >, DerType>::Type >
+  operator*(const Real& other, const AutoDiffScalar<DerivativeType>& a)
+  {
+    return AutoDiffScalar<typename CwiseUnaryOp<bind1st_op<scalar_product_op<Real,Scalar> >, DerType>::Type >(
+      a.value() * other,
+      a.derivatives() * other);
+  }
+
+  inline AutoDiffScalar<DerivativeType>& operator*=(const Scalar& other)
+  {
+    *this = *this * other;
+    return derived();
+  }
+};
+
+template<typename DerivativeType>
+struct auto_diff_special_op<DerivativeType, false>
+{
+  void operator*() const;
+  void operator-() const;
+  void operator+() const;
+};
+
+template<typename BinOp, typename A, typename B, typename RefType>
+void make_coherent_expression(CwiseBinaryOp<BinOp,A,B> xpr, const RefType &ref)
+{
+  make_coherent(xpr.const_cast_derived().lhs(), ref);
+  make_coherent(xpr.const_cast_derived().rhs(), ref);
+}
+
+template<typename UnaryOp, typename A, typename RefType>
+void make_coherent_expression(const CwiseUnaryOp<UnaryOp,A> &xpr, const RefType &ref)
+{
+  make_coherent(xpr.nestedExpression().const_cast_derived(), ref);
+}
+
+// needed for compilation only
+template<typename UnaryOp, typename A, typename RefType>
+void make_coherent_expression(const CwiseNullaryOp<UnaryOp,A> &, const RefType &)
+{}
+
+template<typename A_Scalar, int A_Rows, int A_Cols, int A_Options, int A_MaxRows, int A_MaxCols, typename B>
+struct make_coherent_impl<Matrix<A_Scalar, A_Rows, A_Cols, A_Options, A_MaxRows, A_MaxCols>, B> {
+  typedef Matrix<A_Scalar, A_Rows, A_Cols, A_Options, A_MaxRows, A_MaxCols> A;
+  static void run(A& a, B& b) {
+    if((A_Rows==Dynamic || A_Cols==Dynamic) && (a.size()==0))
+    {
+      a.resize(b.size());
+      a.setZero();
+    }
+    else if (B::SizeAtCompileTime==Dynamic && a.size()!=0 && b.size()==0)
+    {
+      make_coherent_expression(b,a);
+    }
+  }
+};
+
+template<typename A, typename B_Scalar, int B_Rows, int B_Cols, int B_Options, int B_MaxRows, int B_MaxCols>
+struct make_coherent_impl<A, Matrix<B_Scalar, B_Rows, B_Cols, B_Options, B_MaxRows, B_MaxCols> > {
+  typedef Matrix<B_Scalar, B_Rows, B_Cols, B_Options, B_MaxRows, B_MaxCols> B;
+  static void run(A& a, B& b) {
+    if((B_Rows==Dynamic || B_Cols==Dynamic) && (b.size()==0))
+    {
+      b.resize(a.size());
+      b.setZero();
+    }
+    else if (A::SizeAtCompileTime==Dynamic && b.size()!=0 && a.size()==0)
+    {
+      make_coherent_expression(a,b);
+    }
+  }
+};
+
+template<typename A_Scalar, int A_Rows, int A_Cols, int A_Options, int A_MaxRows, int A_MaxCols,
+         typename B_Scalar, int B_Rows, int B_Cols, int B_Options, int B_MaxRows, int B_MaxCols>
+struct make_coherent_impl<Matrix<A_Scalar, A_Rows, A_Cols, A_Options, A_MaxRows, A_MaxCols>,
+                          Matrix<B_Scalar, B_Rows, B_Cols, B_Options, B_MaxRows, B_MaxCols> > {
+  typedef Matrix<A_Scalar, A_Rows, A_Cols, A_Options, A_MaxRows, A_MaxCols> A;
+  typedef Matrix<B_Scalar, B_Rows, B_Cols, B_Options, B_MaxRows, B_MaxCols> B;
+  static void run(A& a, B& b) {
+    if((A_Rows==Dynamic || A_Cols==Dynamic) && (a.size()==0))
+    {
+      a.resize(b.size());
+      a.setZero();
+    }
+    else if((B_Rows==Dynamic || B_Cols==Dynamic) && (b.size()==0))
+    {
+      b.resize(a.size());
+      b.setZero();
+    }
+  }
+};
+
+} // end namespace internal
+
+template<typename DerType, typename BinOp>
+struct ScalarBinaryOpTraits<AutoDiffScalar<DerType>,typename DerType::Scalar,BinOp>
+{
+  typedef AutoDiffScalar<DerType> ReturnType;
+};
+
+template<typename DerType, typename BinOp>
+struct ScalarBinaryOpTraits<typename DerType::Scalar,AutoDiffScalar<DerType>, BinOp>
+{
+  typedef AutoDiffScalar<DerType> ReturnType;
+};
+
+
+// The following is an attempt to let Eigen's known about expression template, but that's more tricky!
+
+// template<typename DerType, typename BinOp>
+// struct ScalarBinaryOpTraits<AutoDiffScalar<DerType>,AutoDiffScalar<DerType>, BinOp>
+// {
+//   enum { Defined = 1 };
+//   typedef AutoDiffScalar<typename DerType::PlainObject> ReturnType;
+// };
+//
+// template<typename DerType1,typename DerType2, typename BinOp>
+// struct ScalarBinaryOpTraits<AutoDiffScalar<DerType1>,AutoDiffScalar<DerType2>, BinOp>
+// {
+//   enum { Defined = 1 };//internal::is_same<typename DerType1::Scalar,typename DerType2::Scalar>::value };
+//   typedef AutoDiffScalar<typename DerType1::PlainObject> ReturnType;
+// };
+
+#define EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(FUNC,CODE) \
+  template<typename DerType> \
+  inline const Eigen::AutoDiffScalar< \
+  EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(typename Eigen::internal::remove_all<DerType>::type, typename Eigen::internal::traits<typename Eigen::internal::remove_all<DerType>::type>::Scalar, product) > \
+  FUNC(const Eigen::AutoDiffScalar<DerType>& x) { \
+    using namespace Eigen; \
+    typedef typename Eigen::internal::traits<typename Eigen::internal::remove_all<DerType>::type>::Scalar Scalar; \
+    EIGEN_UNUSED_VARIABLE(sizeof(Scalar)); \
+    CODE; \
+  }
+
+template<typename DerType>
+struct CleanedUpDerType {
+  typedef AutoDiffScalar<typename Eigen::internal::remove_all<DerType>::type::PlainObject> type;
+};
+
+template<typename DerType>
+inline const AutoDiffScalar<DerType>& conj(const AutoDiffScalar<DerType>& x)  { return x; }
+template<typename DerType>
+inline const AutoDiffScalar<DerType>& real(const AutoDiffScalar<DerType>& x)  { return x; }
+template<typename DerType>
+inline typename DerType::Scalar imag(const AutoDiffScalar<DerType>&)    { return 0.; }
+template<typename DerType, typename T>
+inline typename CleanedUpDerType<DerType>::type (min)(const AutoDiffScalar<DerType>& x, const T& y) {
+  typedef typename CleanedUpDerType<DerType>::type ADS;
+  return (x <= y ? ADS(x) : ADS(y));
+}
+template<typename DerType, typename T>
+inline typename CleanedUpDerType<DerType>::type (max)(const AutoDiffScalar<DerType>& x, const T& y) {
+  typedef typename CleanedUpDerType<DerType>::type ADS;
+  return (x >= y ? ADS(x) : ADS(y));
+}
+template<typename DerType, typename T>
+inline typename CleanedUpDerType<DerType>::type (min)(const T& x, const AutoDiffScalar<DerType>& y) {
+  typedef typename CleanedUpDerType<DerType>::type ADS;
+  return (x < y ? ADS(x) : ADS(y));
+}
+template<typename DerType, typename T>
+inline typename CleanedUpDerType<DerType>::type (max)(const T& x, const AutoDiffScalar<DerType>& y) {
+  typedef typename CleanedUpDerType<DerType>::type ADS;
+  return (x > y ? ADS(x) : ADS(y));
+}
+template<typename DerType>
+inline typename CleanedUpDerType<DerType>::type (min)(const AutoDiffScalar<DerType>& x, const AutoDiffScalar<DerType>& y) {
+  return (x.value() < y.value() ? x : y);
+}
+template<typename DerType>
+inline typename CleanedUpDerType<DerType>::type (max)(const AutoDiffScalar<DerType>& x, const AutoDiffScalar<DerType>& y) {
+  return (x.value() >= y.value() ? x : y);
+}
+
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(abs,
+  using std::abs;
+  return Eigen::MakeAutoDiffScalar(abs(x.value()), x.derivatives() * (x.value()<0 ? -1 : 1) );)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(abs2,
+  using numext::abs2;
+  return Eigen::MakeAutoDiffScalar(abs2(x.value()), x.derivatives() * (Scalar(2)*x.value()));)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(sqrt,
+  using std::sqrt;
+  Scalar sqrtx = sqrt(x.value());
+  return Eigen::MakeAutoDiffScalar(sqrtx,x.derivatives() * (Scalar(0.5) / sqrtx));)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(cos,
+  using std::cos;
+  using std::sin;
+  return Eigen::MakeAutoDiffScalar(cos(x.value()), x.derivatives() * (-sin(x.value())));)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(sin,
+  using std::sin;
+  using std::cos;
+  return Eigen::MakeAutoDiffScalar(sin(x.value()),x.derivatives() * cos(x.value()));)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(exp,
+  using std::exp;
+  Scalar expx = exp(x.value());
+  return Eigen::MakeAutoDiffScalar(expx,x.derivatives() * expx);)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(log,
+  using std::log;
+  return Eigen::MakeAutoDiffScalar(log(x.value()),x.derivatives() * (Scalar(1)/x.value()));)
+
+template<typename DerType>
+inline const Eigen::AutoDiffScalar<
+EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(typename internal::remove_all<DerType>::type,typename internal::traits<typename internal::remove_all<DerType>::type>::Scalar,product) >
+pow(const Eigen::AutoDiffScalar<DerType> &x, const typename internal::traits<typename internal::remove_all<DerType>::type>::Scalar &y)
+{
+  using namespace Eigen;
+  using std::pow;
+  return Eigen::MakeAutoDiffScalar(pow(x.value(),y), x.derivatives() * (y * pow(x.value(),y-1)));
+}
+
+
+template<typename DerTypeA,typename DerTypeB>
+inline const AutoDiffScalar<Matrix<typename internal::traits<typename internal::remove_all<DerTypeA>::type>::Scalar,Dynamic,1> >
+atan2(const AutoDiffScalar<DerTypeA>& a, const AutoDiffScalar<DerTypeB>& b)
+{
+  using std::atan2;
+  typedef typename internal::traits<typename internal::remove_all<DerTypeA>::type>::Scalar Scalar;
+  typedef AutoDiffScalar<Matrix<Scalar,Dynamic,1> > PlainADS;
+  PlainADS ret;
+  ret.value() = atan2(a.value(), b.value());
+  
+  Scalar squared_hypot = a.value() * a.value() + b.value() * b.value();
+  
+  // if (squared_hypot==0) the derivation is undefined and the following results in a NaN:
+  ret.derivatives() = (a.derivatives() * b.value() - a.value() * b.derivatives()) / squared_hypot;
+
+  return ret;
+}
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(tan,
+  using std::tan;
+  using std::cos;
+  return Eigen::MakeAutoDiffScalar(tan(x.value()),x.derivatives() * (Scalar(1)/numext::abs2(cos(x.value()))));)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(asin,
+  using std::sqrt;
+  using std::asin;
+  return Eigen::MakeAutoDiffScalar(asin(x.value()),x.derivatives() * (Scalar(1)/sqrt(1-numext::abs2(x.value()))));)
+  
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(acos,
+  using std::sqrt;
+  using std::acos;
+  return Eigen::MakeAutoDiffScalar(acos(x.value()),x.derivatives() * (Scalar(-1)/sqrt(1-numext::abs2(x.value()))));)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(tanh,
+  using std::cosh;
+  using std::tanh;
+  return Eigen::MakeAutoDiffScalar(tanh(x.value()),x.derivatives() * (Scalar(1)/numext::abs2(cosh(x.value()))));)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(sinh,
+  using std::sinh;
+  using std::cosh;
+  return Eigen::MakeAutoDiffScalar(sinh(x.value()),x.derivatives() * cosh(x.value()));)
+
+EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY(cosh,
+  using std::sinh;
+  using std::cosh;
+  return Eigen::MakeAutoDiffScalar(cosh(x.value()),x.derivatives() * sinh(x.value()));)
+
+#undef EIGEN_AUTODIFF_DECLARE_GLOBAL_UNARY
+
+template<typename DerType> struct NumTraits<AutoDiffScalar<DerType> >
+  : NumTraits< typename NumTraits<typename internal::remove_all<DerType>::type::Scalar>::Real >
+{
+  typedef typename internal::remove_all<DerType>::type DerTypeCleaned;
+  typedef AutoDiffScalar<Matrix<typename NumTraits<typename DerTypeCleaned::Scalar>::Real,DerTypeCleaned::RowsAtCompileTime,DerTypeCleaned::ColsAtCompileTime,
+                                0, DerTypeCleaned::MaxRowsAtCompileTime, DerTypeCleaned::MaxColsAtCompileTime> > Real;
+  typedef AutoDiffScalar<DerType> NonInteger;
+  typedef AutoDiffScalar<DerType> Nested;
+  typedef typename NumTraits<typename DerTypeCleaned::Scalar>::Literal Literal;
+  enum{
+    RequireInitialization = 1
+  };
+};
+
+}
+
+namespace std {
+
+template <typename T>
+class numeric_limits<Eigen::AutoDiffScalar<T> >
+  : public numeric_limits<typename T::Scalar> {};
+
+template <typename T>
+class numeric_limits<Eigen::AutoDiffScalar<T&> >
+  : public numeric_limits<typename T::Scalar> {};
+
+}  // namespace std
+
+#endif // EIGEN_AUTODIFF_SCALAR_H
diff --git a/src/EigenUnsupported/src/AutoDiff/AutoDiffVector.h b/src/EigenUnsupported/src/AutoDiff/AutoDiffVector.h
new file mode 100644
index 0000000..8c2d048
--- /dev/null
+++ b/src/EigenUnsupported/src/AutoDiff/AutoDiffVector.h
@@ -0,0 +1,220 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_AUTODIFF_VECTOR_H
+#define EIGEN_AUTODIFF_VECTOR_H
+
+namespace Eigen {
+
+/* \class AutoDiffScalar
+  * \brief A scalar type replacement with automatic differentation capability
+  *
+  * \param DerType the vector type used to store/represent the derivatives (e.g. Vector3f)
+  *
+  * This class represents a scalar value while tracking its respective derivatives.
+  *
+  * It supports the following list of global math function:
+  *  - std::abs, std::sqrt, std::pow, std::exp, std::log, std::sin, std::cos,
+  *  - internal::abs, internal::sqrt, numext::pow, internal::exp, internal::log, internal::sin, internal::cos,
+  *  - internal::conj, internal::real, internal::imag, numext::abs2.
+  *
+  * AutoDiffScalar can be used as the scalar type of an Eigen::Matrix object. However,
+  * in that case, the expression template mechanism only occurs at the top Matrix level,
+  * while derivatives are computed right away.
+  *
+  */
+template<typename ValueType, typename JacobianType>
+class AutoDiffVector
+{
+  public:
+    //typedef typename internal::traits<ValueType>::Scalar Scalar;
+    typedef typename internal::traits<ValueType>::Scalar BaseScalar;
+    typedef AutoDiffScalar<Matrix<BaseScalar,JacobianType::RowsAtCompileTime,1> > ActiveScalar;
+    typedef ActiveScalar Scalar;
+    typedef AutoDiffScalar<typename JacobianType::ColXpr> CoeffType;
+    typedef typename JacobianType::Index Index;
+
+    inline AutoDiffVector() {}
+
+    inline AutoDiffVector(const ValueType& values)
+      : m_values(values)
+    {
+      m_jacobian.setZero();
+    }
+
+
+    CoeffType operator[] (Index i) { return CoeffType(m_values[i], m_jacobian.col(i)); }
+    const CoeffType operator[] (Index i) const { return CoeffType(m_values[i], m_jacobian.col(i)); }
+
+    CoeffType operator() (Index i) { return CoeffType(m_values[i], m_jacobian.col(i)); }
+    const CoeffType operator() (Index i) const { return CoeffType(m_values[i], m_jacobian.col(i)); }
+
+    CoeffType coeffRef(Index i) { return CoeffType(m_values[i], m_jacobian.col(i)); }
+    const CoeffType coeffRef(Index i) const { return CoeffType(m_values[i], m_jacobian.col(i)); }
+
+    Index size() const { return m_values.size(); }
+
+    // FIXME here we could return an expression of the sum
+    Scalar sum() const { /*std::cerr << "sum \n\n";*/ /*std::cerr << m_jacobian.rowwise().sum() << "\n\n";*/ return Scalar(m_values.sum(), m_jacobian.rowwise().sum()); }
+
+
+    inline AutoDiffVector(const ValueType& values, const JacobianType& jac)
+      : m_values(values), m_jacobian(jac)
+    {}
+
+    template<typename OtherValueType, typename OtherJacobianType>
+    inline AutoDiffVector(const AutoDiffVector<OtherValueType, OtherJacobianType>& other)
+      : m_values(other.values()), m_jacobian(other.jacobian())
+    {}
+
+    inline AutoDiffVector(const AutoDiffVector& other)
+      : m_values(other.values()), m_jacobian(other.jacobian())
+    {}
+
+    template<typename OtherValueType, typename OtherJacobianType>
+    inline AutoDiffVector& operator=(const AutoDiffVector<OtherValueType, OtherJacobianType>& other)
+    {
+      m_values = other.values();
+      m_jacobian = other.jacobian();
+      return *this;
+    }
+
+    inline AutoDiffVector& operator=(const AutoDiffVector& other)
+    {
+      m_values = other.values();
+      m_jacobian = other.jacobian();
+      return *this;
+    }
+
+    inline const ValueType& values() const { return m_values; }
+    inline ValueType& values() { return m_values; }
+
+    inline const JacobianType& jacobian() const { return m_jacobian; }
+    inline JacobianType& jacobian() { return m_jacobian; }
+
+    template<typename OtherValueType,typename OtherJacobianType>
+    inline const AutoDiffVector<
+      typename MakeCwiseBinaryOp<internal::scalar_sum_op<BaseScalar>,ValueType,OtherValueType>::Type,
+      typename MakeCwiseBinaryOp<internal::scalar_sum_op<BaseScalar>,JacobianType,OtherJacobianType>::Type >
+    operator+(const AutoDiffVector<OtherValueType,OtherJacobianType>& other) const
+    {
+      return AutoDiffVector<
+      typename MakeCwiseBinaryOp<internal::scalar_sum_op<BaseScalar>,ValueType,OtherValueType>::Type,
+      typename MakeCwiseBinaryOp<internal::scalar_sum_op<BaseScalar>,JacobianType,OtherJacobianType>::Type >(
+        m_values + other.values(),
+        m_jacobian + other.jacobian());
+    }
+
+    template<typename OtherValueType, typename OtherJacobianType>
+    inline AutoDiffVector&
+    operator+=(const AutoDiffVector<OtherValueType,OtherJacobianType>& other)
+    {
+      m_values += other.values();
+      m_jacobian += other.jacobian();
+      return *this;
+    }
+
+    template<typename OtherValueType,typename OtherJacobianType>
+    inline const AutoDiffVector<
+      typename MakeCwiseBinaryOp<internal::scalar_difference_op<Scalar>,ValueType,OtherValueType>::Type,
+      typename MakeCwiseBinaryOp<internal::scalar_difference_op<Scalar>,JacobianType,OtherJacobianType>::Type >
+    operator-(const AutoDiffVector<OtherValueType,OtherJacobianType>& other) const
+    {
+      return AutoDiffVector<
+        typename MakeCwiseBinaryOp<internal::scalar_difference_op<Scalar>,ValueType,OtherValueType>::Type,
+        typename MakeCwiseBinaryOp<internal::scalar_difference_op<Scalar>,JacobianType,OtherJacobianType>::Type >(
+          m_values - other.values(),
+          m_jacobian - other.jacobian());
+    }
+
+    template<typename OtherValueType, typename OtherJacobianType>
+    inline AutoDiffVector&
+    operator-=(const AutoDiffVector<OtherValueType,OtherJacobianType>& other)
+    {
+      m_values -= other.values();
+      m_jacobian -= other.jacobian();
+      return *this;
+    }
+
+    inline const AutoDiffVector<
+      typename MakeCwiseUnaryOp<internal::scalar_opposite_op<Scalar>, ValueType>::Type,
+      typename MakeCwiseUnaryOp<internal::scalar_opposite_op<Scalar>, JacobianType>::Type >
+    operator-() const
+    {
+      return AutoDiffVector<
+        typename MakeCwiseUnaryOp<internal::scalar_opposite_op<Scalar>, ValueType>::Type,
+        typename MakeCwiseUnaryOp<internal::scalar_opposite_op<Scalar>, JacobianType>::Type >(
+          -m_values,
+          -m_jacobian);
+    }
+
+    inline const AutoDiffVector<
+      typename MakeCwiseUnaryOp<internal::scalar_multiple_op<Scalar>, ValueType>::Type,
+      typename MakeCwiseUnaryOp<internal::scalar_multiple_op<Scalar>, JacobianType>::Type>
+    operator*(const BaseScalar& other) const
+    {
+      return AutoDiffVector<
+        typename MakeCwiseUnaryOp<internal::scalar_multiple_op<Scalar>, ValueType>::Type,
+        typename MakeCwiseUnaryOp<internal::scalar_multiple_op<Scalar>, JacobianType>::Type >(
+          m_values * other,
+          m_jacobian * other);
+    }
+
+    friend inline const AutoDiffVector<
+      typename MakeCwiseUnaryOp<internal::scalar_multiple_op<Scalar>, ValueType>::Type,
+      typename MakeCwiseUnaryOp<internal::scalar_multiple_op<Scalar>, JacobianType>::Type >
+    operator*(const Scalar& other, const AutoDiffVector& v)
+    {
+      return AutoDiffVector<
+        typename MakeCwiseUnaryOp<internal::scalar_multiple_op<Scalar>, ValueType>::Type,
+        typename MakeCwiseUnaryOp<internal::scalar_multiple_op<Scalar>, JacobianType>::Type >(
+          v.values() * other,
+          v.jacobian() * other);
+    }
+
+//     template<typename OtherValueType,typename OtherJacobianType>
+//     inline const AutoDiffVector<
+//       CwiseBinaryOp<internal::scalar_multiple_op<Scalar>, ValueType, OtherValueType>
+//       CwiseBinaryOp<internal::scalar_sum_op<Scalar>,
+//         CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, JacobianType>,
+//         CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, OtherJacobianType> > >
+//     operator*(const AutoDiffVector<OtherValueType,OtherJacobianType>& other) const
+//     {
+//       return AutoDiffVector<
+//         CwiseBinaryOp<internal::scalar_multiple_op<Scalar>, ValueType, OtherValueType>
+//         CwiseBinaryOp<internal::scalar_sum_op<Scalar>,
+//           CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, JacobianType>,
+//           CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, OtherJacobianType> > >(
+//             m_values.cwise() * other.values(),
+//             (m_jacobian * other.values()) + (m_values * other.jacobian()));
+//     }
+
+    inline AutoDiffVector& operator*=(const Scalar& other)
+    {
+      m_values *= other;
+      m_jacobian *= other;
+      return *this;
+    }
+
+    template<typename OtherValueType,typename OtherJacobianType>
+    inline AutoDiffVector& operator*=(const AutoDiffVector<OtherValueType,OtherJacobianType>& other)
+    {
+      *this = *this * other;
+      return *this;
+    }
+
+  protected:
+    ValueType m_values;
+    JacobianType m_jacobian;
+
+};
+
+}
+
+#endif // EIGEN_AUTODIFF_VECTOR_H
diff --git a/src/EigenUnsupported/src/BVH/BVAlgorithms.h b/src/EigenUnsupported/src/BVH/BVAlgorithms.h
new file mode 100644
index 0000000..994c8af
--- /dev/null
+++ b/src/EigenUnsupported/src/BVH/BVAlgorithms.h
@@ -0,0 +1,293 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Ilya Baran <ibaran@mit.edu>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BVALGORITHMS_H
+#define EIGEN_BVALGORITHMS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename BVH, typename Intersector>
+bool intersect_helper(const BVH &tree, Intersector &intersector, typename BVH::Index root)
+{
+  typedef typename BVH::Index Index;
+  typedef typename BVH::VolumeIterator VolIter;
+  typedef typename BVH::ObjectIterator ObjIter;
+
+  VolIter vBegin = VolIter(), vEnd = VolIter();
+  ObjIter oBegin = ObjIter(), oEnd = ObjIter();
+
+  std::vector<Index> todo(1, root);
+
+  while(!todo.empty()) {
+    tree.getChildren(todo.back(), vBegin, vEnd, oBegin, oEnd);
+    todo.pop_back();
+
+    for(; vBegin != vEnd; ++vBegin) //go through child volumes
+      if(intersector.intersectVolume(tree.getVolume(*vBegin)))
+        todo.push_back(*vBegin);
+
+    for(; oBegin != oEnd; ++oBegin) //go through child objects
+      if(intersector.intersectObject(*oBegin))
+        return true; //intersector said to stop query
+  }
+  return false;
+}
+#endif //not EIGEN_PARSED_BY_DOXYGEN
+
+template<typename Volume1, typename Object1, typename Object2, typename Intersector>
+struct intersector_helper1
+{
+  intersector_helper1(const Object2 &inStored, Intersector &in) : stored(inStored), intersector(in) {}
+  bool intersectVolume(const Volume1 &vol) { return intersector.intersectVolumeObject(vol, stored); }
+  bool intersectObject(const Object1 &obj) { return intersector.intersectObjectObject(obj, stored); }
+  Object2 stored;
+  Intersector &intersector;
+private:
+  intersector_helper1& operator=(const intersector_helper1&);
+};
+
+template<typename Volume2, typename Object2, typename Object1, typename Intersector>
+struct intersector_helper2
+{
+  intersector_helper2(const Object1 &inStored, Intersector &in) : stored(inStored), intersector(in) {}
+  bool intersectVolume(const Volume2 &vol) { return intersector.intersectObjectVolume(stored, vol); }
+  bool intersectObject(const Object2 &obj) { return intersector.intersectObjectObject(stored, obj); }
+  Object1 stored;
+  Intersector &intersector;
+private:
+  intersector_helper2& operator=(const intersector_helper2&);
+};
+
+} // end namespace internal
+
+/**  Given a BVH, runs the query encapsulated by \a intersector.
+  *  The Intersector type must provide the following members: \code
+     bool intersectVolume(const BVH::Volume &volume) //returns true if volume intersects the query
+     bool intersectObject(const BVH::Object &object) //returns true if the search should terminate immediately
+  \endcode
+  */
+template<typename BVH, typename Intersector>
+void BVIntersect(const BVH &tree, Intersector &intersector)
+{
+  internal::intersect_helper(tree, intersector, tree.getRootIndex());
+}
+
+/**  Given two BVH's, runs the query on their Cartesian product encapsulated by \a intersector.
+  *  The Intersector type must provide the following members: \code
+     bool intersectVolumeVolume(const BVH1::Volume &v1, const BVH2::Volume &v2) //returns true if product of volumes intersects the query
+     bool intersectVolumeObject(const BVH1::Volume &v1, const BVH2::Object &o2) //returns true if the volume-object product intersects the query
+     bool intersectObjectVolume(const BVH1::Object &o1, const BVH2::Volume &v2) //returns true if the volume-object product intersects the query
+     bool intersectObjectObject(const BVH1::Object &o1, const BVH2::Object &o2) //returns true if the search should terminate immediately
+  \endcode
+  */
+template<typename BVH1, typename BVH2, typename Intersector>
+void BVIntersect(const BVH1 &tree1, const BVH2 &tree2, Intersector &intersector) //TODO: tandem descent when it makes sense
+{
+  typedef typename BVH1::Index Index1;
+  typedef typename BVH2::Index Index2;
+  typedef internal::intersector_helper1<typename BVH1::Volume, typename BVH1::Object, typename BVH2::Object, Intersector> Helper1;
+  typedef internal::intersector_helper2<typename BVH2::Volume, typename BVH2::Object, typename BVH1::Object, Intersector> Helper2;
+  typedef typename BVH1::VolumeIterator VolIter1;
+  typedef typename BVH1::ObjectIterator ObjIter1;
+  typedef typename BVH2::VolumeIterator VolIter2;
+  typedef typename BVH2::ObjectIterator ObjIter2;
+
+  VolIter1 vBegin1 = VolIter1(), vEnd1 = VolIter1();
+  ObjIter1 oBegin1 = ObjIter1(), oEnd1 = ObjIter1();
+  VolIter2 vBegin2 = VolIter2(), vEnd2 = VolIter2(), vCur2 = VolIter2();
+  ObjIter2 oBegin2 = ObjIter2(), oEnd2 = ObjIter2(), oCur2 = ObjIter2();
+
+  std::vector<std::pair<Index1, Index2> > todo(1, std::make_pair(tree1.getRootIndex(), tree2.getRootIndex()));
+
+  while(!todo.empty()) {
+    tree1.getChildren(todo.back().first, vBegin1, vEnd1, oBegin1, oEnd1);
+    tree2.getChildren(todo.back().second, vBegin2, vEnd2, oBegin2, oEnd2);
+    todo.pop_back();
+
+    for(; vBegin1 != vEnd1; ++vBegin1) { //go through child volumes of first tree
+      const typename BVH1::Volume &vol1 = tree1.getVolume(*vBegin1);
+      for(vCur2 = vBegin2; vCur2 != vEnd2; ++vCur2) { //go through child volumes of second tree
+        if(intersector.intersectVolumeVolume(vol1, tree2.getVolume(*vCur2)))
+          todo.push_back(std::make_pair(*vBegin1, *vCur2));
+      }
+
+      for(oCur2 = oBegin2; oCur2 != oEnd2; ++oCur2) {//go through child objects of second tree
+        Helper1 helper(*oCur2, intersector);
+        if(internal::intersect_helper(tree1, helper, *vBegin1))
+          return; //intersector said to stop query
+      }
+    }
+
+    for(; oBegin1 != oEnd1; ++oBegin1) { //go through child objects of first tree
+      for(vCur2 = vBegin2; vCur2 != vEnd2; ++vCur2) { //go through child volumes of second tree
+        Helper2 helper(*oBegin1, intersector);
+        if(internal::intersect_helper(tree2, helper, *vCur2))
+          return; //intersector said to stop query
+      }
+
+      for(oCur2 = oBegin2; oCur2 != oEnd2; ++oCur2) {//go through child objects of second tree
+        if(intersector.intersectObjectObject(*oBegin1, *oCur2))
+          return; //intersector said to stop query
+      }
+    }
+  }
+}
+
+namespace internal {
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename BVH, typename Minimizer>
+typename Minimizer::Scalar minimize_helper(const BVH &tree, Minimizer &minimizer, typename BVH::Index root, typename Minimizer::Scalar minimum)
+{
+  typedef typename Minimizer::Scalar Scalar;
+  typedef typename BVH::Index Index;
+  typedef std::pair<Scalar, Index> QueueElement; //first element is priority
+  typedef typename BVH::VolumeIterator VolIter;
+  typedef typename BVH::ObjectIterator ObjIter;
+
+  VolIter vBegin = VolIter(), vEnd = VolIter();
+  ObjIter oBegin = ObjIter(), oEnd = ObjIter();
+  std::priority_queue<QueueElement, std::vector<QueueElement>, std::greater<QueueElement> > todo; //smallest is at the top
+
+  todo.push(std::make_pair(Scalar(), root));
+
+  while(!todo.empty()) {
+    tree.getChildren(todo.top().second, vBegin, vEnd, oBegin, oEnd);
+    todo.pop();
+
+    for(; oBegin != oEnd; ++oBegin) //go through child objects
+      minimum = (std::min)(minimum, minimizer.minimumOnObject(*oBegin));
+
+    for(; vBegin != vEnd; ++vBegin) { //go through child volumes
+      Scalar val = minimizer.minimumOnVolume(tree.getVolume(*vBegin));
+      if(val < minimum)
+        todo.push(std::make_pair(val, *vBegin));
+    }
+  }
+
+  return minimum;
+}
+#endif //not EIGEN_PARSED_BY_DOXYGEN
+
+
+template<typename Volume1, typename Object1, typename Object2, typename Minimizer>
+struct minimizer_helper1
+{
+  typedef typename Minimizer::Scalar Scalar;
+  minimizer_helper1(const Object2 &inStored, Minimizer &m) : stored(inStored), minimizer(m) {}
+  Scalar minimumOnVolume(const Volume1 &vol) { return minimizer.minimumOnVolumeObject(vol, stored); }
+  Scalar minimumOnObject(const Object1 &obj) { return minimizer.minimumOnObjectObject(obj, stored); }
+  Object2 stored;
+  Minimizer &minimizer;
+private:
+  minimizer_helper1& operator=(const minimizer_helper1&);
+};
+
+template<typename Volume2, typename Object2, typename Object1, typename Minimizer>
+struct minimizer_helper2
+{
+  typedef typename Minimizer::Scalar Scalar;
+  minimizer_helper2(const Object1 &inStored, Minimizer &m) : stored(inStored), minimizer(m) {}
+  Scalar minimumOnVolume(const Volume2 &vol) { return minimizer.minimumOnObjectVolume(stored, vol); }
+  Scalar minimumOnObject(const Object2 &obj) { return minimizer.minimumOnObjectObject(stored, obj); }
+  Object1 stored;
+  Minimizer &minimizer;
+private:
+  minimizer_helper2& operator=(const minimizer_helper2&);
+};
+
+} // end namespace internal
+
+/**  Given a BVH, runs the query encapsulated by \a minimizer.
+  *  \returns the minimum value.
+  *  The Minimizer type must provide the following members: \code
+     typedef Scalar //the numeric type of what is being minimized--not necessarily the Scalar type of the BVH (if it has one)
+     Scalar minimumOnVolume(const BVH::Volume &volume)
+     Scalar minimumOnObject(const BVH::Object &object)
+  \endcode
+  */
+template<typename BVH, typename Minimizer>
+typename Minimizer::Scalar BVMinimize(const BVH &tree, Minimizer &minimizer)
+{
+  return internal::minimize_helper(tree, minimizer, tree.getRootIndex(), (std::numeric_limits<typename Minimizer::Scalar>::max)());
+}
+
+/**  Given two BVH's, runs the query on their cartesian product encapsulated by \a minimizer.
+  *  \returns the minimum value.
+  *  The Minimizer type must provide the following members: \code
+     typedef Scalar //the numeric type of what is being minimized--not necessarily the Scalar type of the BVH (if it has one)
+     Scalar minimumOnVolumeVolume(const BVH1::Volume &v1, const BVH2::Volume &v2)
+     Scalar minimumOnVolumeObject(const BVH1::Volume &v1, const BVH2::Object &o2)
+     Scalar minimumOnObjectVolume(const BVH1::Object &o1, const BVH2::Volume &v2)
+     Scalar minimumOnObjectObject(const BVH1::Object &o1, const BVH2::Object &o2)
+  \endcode
+  */
+template<typename BVH1, typename BVH2, typename Minimizer>
+typename Minimizer::Scalar BVMinimize(const BVH1 &tree1, const BVH2 &tree2, Minimizer &minimizer)
+{
+  typedef typename Minimizer::Scalar Scalar;
+  typedef typename BVH1::Index Index1;
+  typedef typename BVH2::Index Index2;
+  typedef internal::minimizer_helper1<typename BVH1::Volume, typename BVH1::Object, typename BVH2::Object, Minimizer> Helper1;
+  typedef internal::minimizer_helper2<typename BVH2::Volume, typename BVH2::Object, typename BVH1::Object, Minimizer> Helper2;
+  typedef std::pair<Scalar, std::pair<Index1, Index2> > QueueElement; //first element is priority
+  typedef typename BVH1::VolumeIterator VolIter1;
+  typedef typename BVH1::ObjectIterator ObjIter1;
+  typedef typename BVH2::VolumeIterator VolIter2;
+  typedef typename BVH2::ObjectIterator ObjIter2;
+
+  VolIter1 vBegin1 = VolIter1(), vEnd1 = VolIter1();
+  ObjIter1 oBegin1 = ObjIter1(), oEnd1 = ObjIter1();
+  VolIter2 vBegin2 = VolIter2(), vEnd2 = VolIter2(), vCur2 = VolIter2();
+  ObjIter2 oBegin2 = ObjIter2(), oEnd2 = ObjIter2(), oCur2 = ObjIter2();
+  std::priority_queue<QueueElement, std::vector<QueueElement>, std::greater<QueueElement> > todo; //smallest is at the top
+
+  Scalar minimum = (std::numeric_limits<Scalar>::max)();
+  todo.push(std::make_pair(Scalar(), std::make_pair(tree1.getRootIndex(), tree2.getRootIndex())));
+
+  while(!todo.empty()) {
+    tree1.getChildren(todo.top().second.first, vBegin1, vEnd1, oBegin1, oEnd1);
+    tree2.getChildren(todo.top().second.second, vBegin2, vEnd2, oBegin2, oEnd2);
+    todo.pop();
+
+    for(; oBegin1 != oEnd1; ++oBegin1) { //go through child objects of first tree
+      for(oCur2 = oBegin2; oCur2 != oEnd2; ++oCur2) {//go through child objects of second tree
+        minimum = (std::min)(minimum, minimizer.minimumOnObjectObject(*oBegin1, *oCur2));
+      }
+
+      for(vCur2 = vBegin2; vCur2 != vEnd2; ++vCur2) { //go through child volumes of second tree
+        Helper2 helper(*oBegin1, minimizer);
+        minimum = (std::min)(minimum, internal::minimize_helper(tree2, helper, *vCur2, minimum));
+      }
+    }
+
+    for(; vBegin1 != vEnd1; ++vBegin1) { //go through child volumes of first tree
+      const typename BVH1::Volume &vol1 = tree1.getVolume(*vBegin1);
+
+      for(oCur2 = oBegin2; oCur2 != oEnd2; ++oCur2) {//go through child objects of second tree
+        Helper1 helper(*oCur2, minimizer);
+        minimum = (std::min)(minimum, internal::minimize_helper(tree1, helper, *vBegin1, minimum));
+      }
+
+      for(vCur2 = vBegin2; vCur2 != vEnd2; ++vCur2) { //go through child volumes of second tree
+        Scalar val = minimizer.minimumOnVolumeVolume(vol1, tree2.getVolume(*vCur2));
+        if(val < minimum)
+          todo.push(std::make_pair(val, std::make_pair(*vBegin1, *vCur2)));
+      }
+    }
+  }
+  return minimum;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_BVALGORITHMS_H
diff --git a/src/EigenUnsupported/src/BVH/KdBVH.h b/src/EigenUnsupported/src/BVH/KdBVH.h
new file mode 100644
index 0000000..2d5b76a
--- /dev/null
+++ b/src/EigenUnsupported/src/BVH/KdBVH.h
@@ -0,0 +1,223 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Ilya Baran <ibaran@mit.edu>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef KDBVH_H_INCLUDED
+#define KDBVH_H_INCLUDED
+
+namespace Eigen { 
+
+namespace internal {
+
+//internal pair class for the BVH--used instead of std::pair because of alignment
+template<typename Scalar, int Dim>
+struct vector_int_pair
+{
+EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar, Dim)
+  typedef Matrix<Scalar, Dim, 1> VectorType;
+
+  vector_int_pair(const VectorType &v, int i) : first(v), second(i) {}
+
+  VectorType first;
+  int second;
+};
+
+//these templates help the tree initializer get the bounding boxes either from a provided
+//iterator range or using bounding_box in a unified way
+template<typename ObjectList, typename VolumeList, typename BoxIter>
+struct get_boxes_helper {
+  void operator()(const ObjectList &objects, BoxIter boxBegin, BoxIter boxEnd, VolumeList &outBoxes)
+  {
+    outBoxes.insert(outBoxes.end(), boxBegin, boxEnd);
+    eigen_assert(outBoxes.size() == objects.size());
+    EIGEN_ONLY_USED_FOR_DEBUG(objects);
+  }
+};
+
+template<typename ObjectList, typename VolumeList>
+struct get_boxes_helper<ObjectList, VolumeList, int> {
+  void operator()(const ObjectList &objects, int, int, VolumeList &outBoxes)
+  {
+    outBoxes.reserve(objects.size());
+    for(int i = 0; i < (int)objects.size(); ++i)
+      outBoxes.push_back(bounding_box(objects[i]));
+  }
+};
+
+} // end namespace internal
+
+
+/** \class KdBVH
+ *  \brief A simple bounding volume hierarchy based on AlignedBox
+ *
+ *  \param _Scalar The underlying scalar type of the bounding boxes
+ *  \param _Dim The dimension of the space in which the hierarchy lives
+ *  \param _Object The object type that lives in the hierarchy.  It must have value semantics.  Either bounding_box(_Object) must
+ *                 be defined and return an AlignedBox<_Scalar, _Dim> or bounding boxes must be provided to the tree initializer.
+ *
+ *  This class provides a simple (as opposed to optimized) implementation of a bounding volume hierarchy analogous to a Kd-tree.
+ *  Given a sequence of objects, it computes their bounding boxes, constructs a Kd-tree of their centers
+ *  and builds a BVH with the structure of that Kd-tree.  When the elements of the tree are too expensive to be copied around,
+ *  it is useful for _Object to be a pointer.
+ */
+template<typename _Scalar, int _Dim, typename _Object> class KdBVH
+{
+public:
+  enum { Dim = _Dim };
+  typedef _Object Object;
+  typedef std::vector<Object, aligned_allocator<Object> > ObjectList;
+  typedef _Scalar Scalar;
+  typedef AlignedBox<Scalar, Dim> Volume;
+  typedef std::vector<Volume, aligned_allocator<Volume> > VolumeList;
+  typedef int Index;
+  typedef const int *VolumeIterator; //the iterators are just pointers into the tree's vectors
+  typedef const Object *ObjectIterator;
+
+  KdBVH() {}
+
+  /** Given an iterator range over \a Object references, constructs the BVH.  Requires that bounding_box(Object) return a Volume. */
+  template<typename Iter> KdBVH(Iter begin, Iter end) { init(begin, end, 0, 0); } //int is recognized by init as not being an iterator type
+
+  /** Given an iterator range over \a Object references and an iterator range over their bounding boxes, constructs the BVH */
+  template<typename OIter, typename BIter> KdBVH(OIter begin, OIter end, BIter boxBegin, BIter boxEnd) { init(begin, end, boxBegin, boxEnd); }
+
+  /** Given an iterator range over \a Object references, constructs the BVH, overwriting whatever is in there currently.
+    * Requires that bounding_box(Object) return a Volume. */
+  template<typename Iter> void init(Iter begin, Iter end) { init(begin, end, 0, 0); }
+
+  /** Given an iterator range over \a Object references and an iterator range over their bounding boxes,
+    * constructs the BVH, overwriting whatever is in there currently. */
+  template<typename OIter, typename BIter> void init(OIter begin, OIter end, BIter boxBegin, BIter boxEnd)
+  {
+    objects.clear();
+    boxes.clear();
+    children.clear();
+
+    objects.insert(objects.end(), begin, end);
+    int n = static_cast<int>(objects.size());
+
+    if(n < 2)
+      return; //if we have at most one object, we don't need any internal nodes
+
+    VolumeList objBoxes;
+    VIPairList objCenters;
+
+    //compute the bounding boxes depending on BIter type
+    internal::get_boxes_helper<ObjectList, VolumeList, BIter>()(objects, boxBegin, boxEnd, objBoxes);
+
+    objCenters.reserve(n);
+    boxes.reserve(n - 1);
+    children.reserve(2 * n - 2);
+
+    for(int i = 0; i < n; ++i)
+      objCenters.push_back(VIPair(objBoxes[i].center(), i));
+
+    build(objCenters, 0, n, objBoxes, 0); //the recursive part of the algorithm
+
+    ObjectList tmp(n);
+    tmp.swap(objects);
+    for(int i = 0; i < n; ++i)
+      objects[i] = tmp[objCenters[i].second];
+  }
+
+  /** \returns the index of the root of the hierarchy */
+  inline Index getRootIndex() const { return (int)boxes.size() - 1; }
+
+  /** Given an \a index of a node, on exit, \a outVBegin and \a outVEnd range over the indices of the volume children of the node
+    * and \a outOBegin and \a outOEnd range over the object children of the node */
+  EIGEN_STRONG_INLINE void getChildren(Index index, VolumeIterator &outVBegin, VolumeIterator &outVEnd,
+                                       ObjectIterator &outOBegin, ObjectIterator &outOEnd) const
+  { //inlining this function should open lots of optimization opportunities to the compiler
+    if(index < 0) {
+      outVBegin = outVEnd;
+      if(!objects.empty())
+        outOBegin = &(objects[0]);
+      outOEnd = outOBegin + objects.size(); //output all objects--necessary when the tree has only one object
+      return;
+    }
+
+    int numBoxes = static_cast<int>(boxes.size());
+
+    int idx = index * 2;
+    if(children[idx + 1] < numBoxes) { //second index is always bigger
+      outVBegin = &(children[idx]);
+      outVEnd = outVBegin + 2;
+      outOBegin = outOEnd;
+    }
+    else if(children[idx] >= numBoxes) { //if both children are objects
+      outVBegin = outVEnd;
+      outOBegin = &(objects[children[idx] - numBoxes]);
+      outOEnd = outOBegin + 2;
+    } else { //if the first child is a volume and the second is an object
+      outVBegin = &(children[idx]);
+      outVEnd = outVBegin + 1;
+      outOBegin = &(objects[children[idx + 1] - numBoxes]);
+      outOEnd = outOBegin + 1;
+    }
+  }
+
+  /** \returns the bounding box of the node at \a index */
+  inline const Volume &getVolume(Index index) const
+  {
+    return boxes[index];
+  }
+
+private:
+  typedef internal::vector_int_pair<Scalar, Dim> VIPair;
+  typedef std::vector<VIPair, aligned_allocator<VIPair> > VIPairList;
+  typedef Matrix<Scalar, Dim, 1> VectorType;
+  struct VectorComparator //compares vectors, or more specifically, VIPairs along a particular dimension
+  {
+    VectorComparator(int inDim) : dim(inDim) {}
+    inline bool operator()(const VIPair &v1, const VIPair &v2) const { return v1.first[dim] < v2.first[dim]; }
+    int dim;
+  };
+
+  //Build the part of the tree between objects[from] and objects[to] (not including objects[to]).
+  //This routine partitions the objCenters in [from, to) along the dimension dim, recursively constructs
+  //the two halves, and adds their parent node.  TODO: a cache-friendlier layout
+  void build(VIPairList &objCenters, int from, int to, const VolumeList &objBoxes, int dim)
+  {
+    eigen_assert(to - from > 1);
+    if(to - from == 2) {
+      boxes.push_back(objBoxes[objCenters[from].second].merged(objBoxes[objCenters[from + 1].second]));
+      children.push_back(from + (int)objects.size() - 1); //there are objects.size() - 1 tree nodes
+      children.push_back(from + (int)objects.size());
+    }
+    else if(to - from == 3) {
+      int mid = from + 2;
+      std::nth_element(objCenters.begin() + from, objCenters.begin() + mid,
+                        objCenters.begin() + to, VectorComparator(dim)); //partition
+      build(objCenters, from, mid, objBoxes, (dim + 1) % Dim);
+      int idx1 = (int)boxes.size() - 1;
+      boxes.push_back(boxes[idx1].merged(objBoxes[objCenters[mid].second]));
+      children.push_back(idx1);
+      children.push_back(mid + (int)objects.size() - 1);
+    }
+    else {
+      int mid = from + (to - from) / 2;
+      nth_element(objCenters.begin() + from, objCenters.begin() + mid,
+                  objCenters.begin() + to, VectorComparator(dim)); //partition
+      build(objCenters, from, mid, objBoxes, (dim + 1) % Dim);
+      int idx1 = (int)boxes.size() - 1;
+      build(objCenters, mid, to, objBoxes, (dim + 1) % Dim);
+      int idx2 = (int)boxes.size() - 1;
+      boxes.push_back(boxes[idx1].merged(boxes[idx2]));
+      children.push_back(idx1);
+      children.push_back(idx2);
+    }
+  }
+
+  std::vector<int> children; //children of x are children[2x] and children[2x+1], indices bigger than boxes.size() index into objects.
+  VolumeList boxes;
+  ObjectList objects;
+};
+
+} // end namespace Eigen
+
+#endif //KDBVH_H_INCLUDED
diff --git a/src/EigenUnsupported/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h b/src/EigenUnsupported/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h
new file mode 100644
index 0000000..0fbd847
--- /dev/null
+++ b/src/EigenUnsupported/src/Eigenvalues/ArpackSelfAdjointEigenSolver.h
@@ -0,0 +1,790 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 David Harmon <dharmon@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ARPACKGENERALIZEDSELFADJOINTEIGENSOLVER_H
+#define EIGEN_ARPACKGENERALIZEDSELFADJOINTEIGENSOLVER_H
+
+#include "../../../../Eigen/Dense"
+
+namespace Eigen { 
+
+namespace internal {
+  template<typename Scalar, typename RealScalar> struct arpack_wrapper;
+  template<typename MatrixSolver, typename MatrixType, typename Scalar, bool BisSPD> struct OP;
+}
+
+
+
+template<typename MatrixType, typename MatrixSolver=SimplicialLLT<MatrixType>, bool BisSPD=false>
+class ArpackGeneralizedSelfAdjointEigenSolver
+{
+public:
+  //typedef typename MatrixSolver::MatrixType MatrixType;
+
+  /** \brief Scalar type for matrices of type \p MatrixType. */
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::Index Index;
+
+  /** \brief Real scalar type for \p MatrixType.
+   *
+   * This is just \c Scalar if #Scalar is real (e.g., \c float or
+   * \c Scalar), and the type of the real part of \c Scalar if #Scalar is
+   * complex.
+   */
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+
+  /** \brief Type for vector of eigenvalues as returned by eigenvalues().
+   *
+   * This is a column vector with entries of type #RealScalar.
+   * The length of the vector is the size of \p nbrEigenvalues.
+   */
+  typedef typename internal::plain_col_type<MatrixType, RealScalar>::type RealVectorType;
+
+  /** \brief Default constructor.
+   *
+   * The default constructor is for cases in which the user intends to
+   * perform decompositions via compute().
+   *
+   */
+  ArpackGeneralizedSelfAdjointEigenSolver()
+   : m_eivec(),
+     m_eivalues(),
+     m_isInitialized(false),
+     m_eigenvectorsOk(false),
+     m_nbrConverged(0),
+     m_nbrIterations(0)
+  { }
+
+  /** \brief Constructor; computes generalized eigenvalues of given matrix with respect to another matrix.
+   *
+   * \param[in] A Self-adjoint matrix whose eigenvalues / eigenvectors will
+   *    computed. By default, the upper triangular part is used, but can be changed
+   *    through the template parameter.
+   * \param[in] B Self-adjoint matrix for the generalized eigenvalue problem.
+   * \param[in] nbrEigenvalues The number of eigenvalues / eigenvectors to compute.
+   *    Must be less than the size of the input matrix, or an error is returned.
+   * \param[in] eigs_sigma String containing either "LM", "SM", "LA", or "SA", with
+   *    respective meanings to find the largest magnitude , smallest magnitude,
+   *    largest algebraic, or smallest algebraic eigenvalues. Alternatively, this
+   *    value can contain floating point value in string form, in which case the
+   *    eigenvalues closest to this value will be found.
+   * \param[in]  options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly.
+   * \param[in] tol What tolerance to find the eigenvalues to. Default is 0, which
+   *    means machine precision.
+   *
+   * This constructor calls compute(const MatrixType&, const MatrixType&, Index, string, int, RealScalar)
+   * to compute the eigenvalues of the matrix \p A with respect to \p B. The eigenvectors are computed if
+   * \p options equals #ComputeEigenvectors.
+   *
+   */
+  ArpackGeneralizedSelfAdjointEigenSolver(const MatrixType& A, const MatrixType& B,
+                                          Index nbrEigenvalues, std::string eigs_sigma="LM",
+                               int options=ComputeEigenvectors, RealScalar tol=0.0)
+    : m_eivec(),
+      m_eivalues(),
+      m_isInitialized(false),
+      m_eigenvectorsOk(false),
+      m_nbrConverged(0),
+      m_nbrIterations(0)
+  {
+    compute(A, B, nbrEigenvalues, eigs_sigma, options, tol);
+  }
+
+  /** \brief Constructor; computes eigenvalues of given matrix.
+   *
+   * \param[in] A Self-adjoint matrix whose eigenvalues / eigenvectors will
+   *    computed. By default, the upper triangular part is used, but can be changed
+   *    through the template parameter.
+   * \param[in] nbrEigenvalues The number of eigenvalues / eigenvectors to compute.
+   *    Must be less than the size of the input matrix, or an error is returned.
+   * \param[in] eigs_sigma String containing either "LM", "SM", "LA", or "SA", with
+   *    respective meanings to find the largest magnitude , smallest magnitude,
+   *    largest algebraic, or smallest algebraic eigenvalues. Alternatively, this
+   *    value can contain floating point value in string form, in which case the
+   *    eigenvalues closest to this value will be found.
+   * \param[in]  options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly.
+   * \param[in] tol What tolerance to find the eigenvalues to. Default is 0, which
+   *    means machine precision.
+   *
+   * This constructor calls compute(const MatrixType&, Index, string, int, RealScalar)
+   * to compute the eigenvalues of the matrix \p A. The eigenvectors are computed if
+   * \p options equals #ComputeEigenvectors.
+   *
+   */
+
+  ArpackGeneralizedSelfAdjointEigenSolver(const MatrixType& A,
+                                          Index nbrEigenvalues, std::string eigs_sigma="LM",
+                               int options=ComputeEigenvectors, RealScalar tol=0.0)
+    : m_eivec(),
+      m_eivalues(),
+      m_isInitialized(false),
+      m_eigenvectorsOk(false),
+      m_nbrConverged(0),
+      m_nbrIterations(0)
+  {
+    compute(A, nbrEigenvalues, eigs_sigma, options, tol);
+  }
+
+
+  /** \brief Computes generalized eigenvalues / eigenvectors of given matrix using the external ARPACK library.
+   *
+   * \param[in]  A  Selfadjoint matrix whose eigendecomposition is to be computed.
+   * \param[in]  B  Selfadjoint matrix for generalized eigenvalues.
+   * \param[in] nbrEigenvalues The number of eigenvalues / eigenvectors to compute.
+   *    Must be less than the size of the input matrix, or an error is returned.
+   * \param[in] eigs_sigma String containing either "LM", "SM", "LA", or "SA", with
+   *    respective meanings to find the largest magnitude , smallest magnitude,
+   *    largest algebraic, or smallest algebraic eigenvalues. Alternatively, this
+   *    value can contain floating point value in string form, in which case the
+   *    eigenvalues closest to this value will be found.
+   * \param[in]  options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly.
+   * \param[in] tol What tolerance to find the eigenvalues to. Default is 0, which
+   *    means machine precision.
+   *
+   * \returns    Reference to \c *this
+   *
+   * This function computes the generalized eigenvalues of \p A with respect to \p B using ARPACK.  The eigenvalues()
+   * function can be used to retrieve them.  If \p options equals #ComputeEigenvectors,
+   * then the eigenvectors are also computed and can be retrieved by
+   * calling eigenvectors().
+   *
+   */
+  ArpackGeneralizedSelfAdjointEigenSolver& compute(const MatrixType& A, const MatrixType& B,
+                                                   Index nbrEigenvalues, std::string eigs_sigma="LM",
+                                        int options=ComputeEigenvectors, RealScalar tol=0.0);
+  
+  /** \brief Computes eigenvalues / eigenvectors of given matrix using the external ARPACK library.
+   *
+   * \param[in]  A  Selfadjoint matrix whose eigendecomposition is to be computed.
+   * \param[in] nbrEigenvalues The number of eigenvalues / eigenvectors to compute.
+   *    Must be less than the size of the input matrix, or an error is returned.
+   * \param[in] eigs_sigma String containing either "LM", "SM", "LA", or "SA", with
+   *    respective meanings to find the largest magnitude , smallest magnitude,
+   *    largest algebraic, or smallest algebraic eigenvalues. Alternatively, this
+   *    value can contain floating point value in string form, in which case the
+   *    eigenvalues closest to this value will be found.
+   * \param[in]  options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly.
+   * \param[in] tol What tolerance to find the eigenvalues to. Default is 0, which
+   *    means machine precision.
+   *
+   * \returns    Reference to \c *this
+   *
+   * This function computes the eigenvalues of \p A using ARPACK.  The eigenvalues()
+   * function can be used to retrieve them.  If \p options equals #ComputeEigenvectors,
+   * then the eigenvectors are also computed and can be retrieved by
+   * calling eigenvectors().
+   *
+   */
+  ArpackGeneralizedSelfAdjointEigenSolver& compute(const MatrixType& A,
+                                                   Index nbrEigenvalues, std::string eigs_sigma="LM",
+                                        int options=ComputeEigenvectors, RealScalar tol=0.0);
+
+
+  /** \brief Returns the eigenvectors of given matrix.
+   *
+   * \returns  A const reference to the matrix whose columns are the eigenvectors.
+   *
+   * \pre The eigenvectors have been computed before.
+   *
+   * Column \f$ k \f$ of the returned matrix is an eigenvector corresponding
+   * to eigenvalue number \f$ k \f$ as returned by eigenvalues().  The
+   * eigenvectors are normalized to have (Euclidean) norm equal to one. If
+   * this object was used to solve the eigenproblem for the selfadjoint
+   * matrix \f$ A \f$, then the matrix returned by this function is the
+   * matrix \f$ V \f$ in the eigendecomposition \f$ A V = D V \f$.
+   * For the generalized eigenproblem, the matrix returned is the solution \f$ A V = D B V \f$
+   *
+   * Example: \include SelfAdjointEigenSolver_eigenvectors.cpp
+   * Output: \verbinclude SelfAdjointEigenSolver_eigenvectors.out
+   *
+   * \sa eigenvalues()
+   */
+  const Matrix<Scalar, Dynamic, Dynamic>& eigenvectors() const
+  {
+    eigen_assert(m_isInitialized && "ArpackGeneralizedSelfAdjointEigenSolver is not initialized.");
+    eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+    return m_eivec;
+  }
+
+  /** \brief Returns the eigenvalues of given matrix.
+   *
+   * \returns A const reference to the column vector containing the eigenvalues.
+   *
+   * \pre The eigenvalues have been computed before.
+   *
+   * The eigenvalues are repeated according to their algebraic multiplicity,
+   * so there are as many eigenvalues as rows in the matrix. The eigenvalues
+   * are sorted in increasing order.
+   *
+   * Example: \include SelfAdjointEigenSolver_eigenvalues.cpp
+   * Output: \verbinclude SelfAdjointEigenSolver_eigenvalues.out
+   *
+   * \sa eigenvectors(), MatrixBase::eigenvalues()
+   */
+  const Matrix<Scalar, Dynamic, 1>& eigenvalues() const
+  {
+    eigen_assert(m_isInitialized && "ArpackGeneralizedSelfAdjointEigenSolver is not initialized.");
+    return m_eivalues;
+  }
+
+  /** \brief Computes the positive-definite square root of the matrix.
+   *
+   * \returns the positive-definite square root of the matrix
+   *
+   * \pre The eigenvalues and eigenvectors of a positive-definite matrix
+   * have been computed before.
+   *
+   * The square root of a positive-definite matrix \f$ A \f$ is the
+   * positive-definite matrix whose square equals \f$ A \f$. This function
+   * uses the eigendecomposition \f$ A = V D V^{-1} \f$ to compute the
+   * square root as \f$ A^{1/2} = V D^{1/2} V^{-1} \f$.
+   *
+   * Example: \include SelfAdjointEigenSolver_operatorSqrt.cpp
+   * Output: \verbinclude SelfAdjointEigenSolver_operatorSqrt.out
+   *
+   * \sa operatorInverseSqrt(),
+   *     \ref MatrixFunctions_Module "MatrixFunctions Module"
+   */
+  Matrix<Scalar, Dynamic, Dynamic> operatorSqrt() const
+  {
+    eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
+    eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+    return m_eivec * m_eivalues.cwiseSqrt().asDiagonal() * m_eivec.adjoint();
+  }
+
+  /** \brief Computes the inverse square root of the matrix.
+   *
+   * \returns the inverse positive-definite square root of the matrix
+   *
+   * \pre The eigenvalues and eigenvectors of a positive-definite matrix
+   * have been computed before.
+   *
+   * This function uses the eigendecomposition \f$ A = V D V^{-1} \f$ to
+   * compute the inverse square root as \f$ V D^{-1/2} V^{-1} \f$. This is
+   * cheaper than first computing the square root with operatorSqrt() and
+   * then its inverse with MatrixBase::inverse().
+   *
+   * Example: \include SelfAdjointEigenSolver_operatorInverseSqrt.cpp
+   * Output: \verbinclude SelfAdjointEigenSolver_operatorInverseSqrt.out
+   *
+   * \sa operatorSqrt(), MatrixBase::inverse(),
+   *     \ref MatrixFunctions_Module "MatrixFunctions Module"
+   */
+  Matrix<Scalar, Dynamic, Dynamic> operatorInverseSqrt() const
+  {
+    eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
+    eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+    return m_eivec * m_eivalues.cwiseInverse().cwiseSqrt().asDiagonal() * m_eivec.adjoint();
+  }
+
+  /** \brief Reports whether previous computation was successful.
+   *
+   * \returns \c Success if computation was successful, \c NoConvergence otherwise.
+   */
+  ComputationInfo info() const
+  {
+    eigen_assert(m_isInitialized && "ArpackGeneralizedSelfAdjointEigenSolver is not initialized.");
+    return m_info;
+  }
+
+  size_t getNbrConvergedEigenValues() const
+  { return m_nbrConverged; }
+
+  size_t getNbrIterations() const
+  { return m_nbrIterations; }
+
+protected:
+  Matrix<Scalar, Dynamic, Dynamic> m_eivec;
+  Matrix<Scalar, Dynamic, 1> m_eivalues;
+  ComputationInfo m_info;
+  bool m_isInitialized;
+  bool m_eigenvectorsOk;
+
+  size_t m_nbrConverged;
+  size_t m_nbrIterations;
+};
+
+
+
+
+
+template<typename MatrixType, typename MatrixSolver, bool BisSPD>
+ArpackGeneralizedSelfAdjointEigenSolver<MatrixType, MatrixSolver, BisSPD>&
+    ArpackGeneralizedSelfAdjointEigenSolver<MatrixType, MatrixSolver, BisSPD>
+::compute(const MatrixType& A, Index nbrEigenvalues,
+          std::string eigs_sigma, int options, RealScalar tol)
+{
+    MatrixType B(0,0);
+    compute(A, B, nbrEigenvalues, eigs_sigma, options, tol);
+    
+    return *this;
+}
+
+
+template<typename MatrixType, typename MatrixSolver, bool BisSPD>
+ArpackGeneralizedSelfAdjointEigenSolver<MatrixType, MatrixSolver, BisSPD>&
+    ArpackGeneralizedSelfAdjointEigenSolver<MatrixType, MatrixSolver, BisSPD>
+::compute(const MatrixType& A, const MatrixType& B, Index nbrEigenvalues,
+          std::string eigs_sigma, int options, RealScalar tol)
+{
+  eigen_assert(A.cols() == A.rows());
+  eigen_assert(B.cols() == B.rows());
+  eigen_assert(B.rows() == 0 || A.cols() == B.rows());
+  eigen_assert((options &~ (EigVecMask | GenEigMask)) == 0
+            && (options & EigVecMask) != EigVecMask
+            && "invalid option parameter");
+
+  bool isBempty = (B.rows() == 0) || (B.cols() == 0);
+
+  // For clarity, all parameters match their ARPACK name
+  //
+  // Always 0 on the first call
+  //
+  int ido = 0;
+
+  int n = (int)A.cols();
+
+  // User options: "LA", "SA", "SM", "LM", "BE"
+  //
+  char whch[3] = "LM";
+    
+  // Specifies the shift if iparam[6] = { 3, 4, 5 }, not used if iparam[6] = { 1, 2 }
+  //
+  RealScalar sigma = 0.0;
+
+  if (eigs_sigma.length() >= 2 && isalpha(eigs_sigma[0]) && isalpha(eigs_sigma[1]))
+  {
+      eigs_sigma[0] = toupper(eigs_sigma[0]);
+      eigs_sigma[1] = toupper(eigs_sigma[1]);
+
+      // In the following special case we're going to invert the problem, since solving
+      // for larger magnitude is much much faster
+      // i.e., if 'SM' is specified, we're going to really use 'LM', the default
+      //
+      if (eigs_sigma.substr(0,2) != "SM")
+      {
+          whch[0] = eigs_sigma[0];
+          whch[1] = eigs_sigma[1];
+      }
+  }
+  else
+  {
+      eigen_assert(false && "Specifying clustered eigenvalues is not yet supported!");
+
+      // If it's not scalar values, then the user may be explicitly
+      // specifying the sigma value to cluster the evs around
+      //
+      sigma = atof(eigs_sigma.c_str());
+
+      // If atof fails, it returns 0.0, which is a fine default
+      //
+  }
+
+  // "I" means normal eigenvalue problem, "G" means generalized
+  //
+  char bmat[2] = "I";
+  if (eigs_sigma.substr(0,2) == "SM" || !(isalpha(eigs_sigma[0]) && isalpha(eigs_sigma[1])) || (!isBempty && !BisSPD))
+      bmat[0] = 'G';
+
+  // Now we determine the mode to use
+  //
+  int mode = (bmat[0] == 'G') + 1;
+  if (eigs_sigma.substr(0,2) == "SM" || !(isalpha(eigs_sigma[0]) && isalpha(eigs_sigma[1])))
+  {
+      // We're going to use shift-and-invert mode, and basically find
+      // the largest eigenvalues of the inverse operator
+      //
+      mode = 3;
+  }
+
+  // The user-specified number of eigenvalues/vectors to compute
+  //
+  int nev = (int)nbrEigenvalues;
+
+  // Allocate space for ARPACK to store the residual
+  //
+  Scalar *resid = new Scalar[n];
+
+  // Number of Lanczos vectors, must satisfy nev < ncv <= n
+  // Note that this indicates that nev != n, and we cannot compute
+  // all eigenvalues of a mtrix
+  //
+  int ncv = std::min(std::max(2*nev, 20), n);
+
+  // The working n x ncv matrix, also store the final eigenvectors (if computed)
+  //
+  Scalar *v = new Scalar[n*ncv];
+  int ldv = n;
+
+  // Working space
+  //
+  Scalar *workd = new Scalar[3*n];
+  int lworkl = ncv*ncv+8*ncv; // Must be at least this length
+  Scalar *workl = new Scalar[lworkl];
+
+  int *iparam= new int[11];
+  iparam[0] = 1; // 1 means we let ARPACK perform the shifts, 0 means we'd have to do it
+  iparam[2] = std::max(300, (int)std::ceil(2*n/std::max(ncv,1)));
+  iparam[6] = mode; // The mode, 1 is standard ev problem, 2 for generalized ev, 3 for shift-and-invert
+
+  // Used during reverse communicate to notify where arrays start
+  //
+  int *ipntr = new int[11]; 
+
+  // Error codes are returned in here, initial value of 0 indicates a random initial
+  // residual vector is used, any other values means resid contains the initial residual
+  // vector, possibly from a previous run
+  //
+  int info = 0;
+
+  Scalar scale = 1.0;
+  //if (!isBempty)
+  //{
+  //Scalar scale = B.norm() / std::sqrt(n);
+  //scale = std::pow(2, std::floor(std::log(scale+1)));
+  ////M /= scale;
+  //for (size_t i=0; i<(size_t)B.outerSize(); i++)
+  //    for (typename MatrixType::InnerIterator it(B, i); it; ++it)
+  //        it.valueRef() /= scale;
+  //}
+
+  MatrixSolver OP;
+  if (mode == 1 || mode == 2)
+  {
+      if (!isBempty)
+          OP.compute(B);
+  }
+  else if (mode == 3)
+  {
+      if (sigma == 0.0)
+      {
+          OP.compute(A);
+      }
+      else
+      {
+          // Note: We will never enter here because sigma must be 0.0
+          //
+          if (isBempty)
+          {
+            MatrixType AminusSigmaB(A);
+            for (Index i=0; i<A.rows(); ++i)
+                AminusSigmaB.coeffRef(i,i) -= sigma;
+            
+            OP.compute(AminusSigmaB);
+          }
+          else
+          {
+              MatrixType AminusSigmaB = A - sigma * B;
+              OP.compute(AminusSigmaB);
+          }
+      }
+  }
+ 
+  if (!(mode == 1 && isBempty) && !(mode == 2 && isBempty) && OP.info() != Success)
+      std::cout << "Error factoring matrix" << std::endl;
+
+  do
+  {
+    internal::arpack_wrapper<Scalar, RealScalar>::saupd(&ido, bmat, &n, whch, &nev, &tol, resid, 
+                                                        &ncv, v, &ldv, iparam, ipntr, workd, workl,
+                                                        &lworkl, &info);
+
+    if (ido == -1 || ido == 1)
+    {
+      Scalar *in  = workd + ipntr[0] - 1;
+      Scalar *out = workd + ipntr[1] - 1;
+
+      if (ido == 1 && mode != 2)
+      {
+          Scalar *out2 = workd + ipntr[2] - 1;
+          if (isBempty || mode == 1)
+            Matrix<Scalar, Dynamic, 1>::Map(out2, n) = Matrix<Scalar, Dynamic, 1>::Map(in, n);
+          else
+            Matrix<Scalar, Dynamic, 1>::Map(out2, n) = B * Matrix<Scalar, Dynamic, 1>::Map(in, n);
+          
+          in = workd + ipntr[2] - 1;
+      }
+
+      if (mode == 1)
+      {
+        if (isBempty)
+        {
+          // OP = A
+          //
+          Matrix<Scalar, Dynamic, 1>::Map(out, n) = A * Matrix<Scalar, Dynamic, 1>::Map(in, n);
+        }
+        else
+        {
+          // OP = L^{-1}AL^{-T}
+          //
+          internal::OP<MatrixSolver, MatrixType, Scalar, BisSPD>::applyOP(OP, A, n, in, out);
+        }
+      }
+      else if (mode == 2)
+      {
+        if (ido == 1)
+          Matrix<Scalar, Dynamic, 1>::Map(in, n)  = A * Matrix<Scalar, Dynamic, 1>::Map(in, n);
+        
+        // OP = B^{-1} A
+        //
+        Matrix<Scalar, Dynamic, 1>::Map(out, n) = OP.solve(Matrix<Scalar, Dynamic, 1>::Map(in, n));
+      }
+      else if (mode == 3)
+      {
+        // OP = (A-\sigmaB)B (\sigma could be 0, and B could be I)
+        // The B * in is already computed and stored at in if ido == 1
+        //
+        if (ido == 1 || isBempty)
+          Matrix<Scalar, Dynamic, 1>::Map(out, n) = OP.solve(Matrix<Scalar, Dynamic, 1>::Map(in, n));
+        else
+          Matrix<Scalar, Dynamic, 1>::Map(out, n) = OP.solve(B * Matrix<Scalar, Dynamic, 1>::Map(in, n));
+      }
+    }
+    else if (ido == 2)
+    {
+      Scalar *in  = workd + ipntr[0] - 1;
+      Scalar *out = workd + ipntr[1] - 1;
+
+      if (isBempty || mode == 1)
+        Matrix<Scalar, Dynamic, 1>::Map(out, n) = Matrix<Scalar, Dynamic, 1>::Map(in, n);
+      else
+        Matrix<Scalar, Dynamic, 1>::Map(out, n) = B * Matrix<Scalar, Dynamic, 1>::Map(in, n);
+    }
+  } while (ido != 99);
+
+  if (info == 1)
+    m_info = NoConvergence;
+  else if (info == 3)
+    m_info = NumericalIssue;
+  else if (info < 0)
+    m_info = InvalidInput;
+  else if (info != 0)
+    eigen_assert(false && "Unknown ARPACK return value!");
+  else
+  {
+    // Do we compute eigenvectors or not?
+    //
+    int rvec = (options & ComputeEigenvectors) == ComputeEigenvectors;
+
+    // "A" means "All", use "S" to choose specific eigenvalues (not yet supported in ARPACK))
+    //
+    char howmny[2] = "A"; 
+
+    // if howmny == "S", specifies the eigenvalues to compute (not implemented in ARPACK)
+    //
+    int *select = new int[ncv];
+
+    // Final eigenvalues
+    //
+    m_eivalues.resize(nev, 1);
+
+    internal::arpack_wrapper<Scalar, RealScalar>::seupd(&rvec, howmny, select, m_eivalues.data(), v, &ldv,
+                                                        &sigma, bmat, &n, whch, &nev, &tol, resid, &ncv,
+                                                        v, &ldv, iparam, ipntr, workd, workl, &lworkl, &info);
+
+    if (info == -14)
+      m_info = NoConvergence;
+    else if (info != 0)
+      m_info = InvalidInput;
+    else
+    {
+      if (rvec)
+      {
+        m_eivec.resize(A.rows(), nev);
+        for (int i=0; i<nev; i++)
+          for (int j=0; j<n; j++)
+            m_eivec(j,i) = v[i*n+j] / scale;
+      
+        if (mode == 1 && !isBempty && BisSPD)
+          internal::OP<MatrixSolver, MatrixType, Scalar, BisSPD>::project(OP, n, nev, m_eivec.data());
+
+        m_eigenvectorsOk = true;
+      }
+
+      m_nbrIterations = iparam[2];
+      m_nbrConverged  = iparam[4];
+
+      m_info = Success;
+    }
+
+    delete[] select;
+  }
+
+  delete[] v;
+  delete[] iparam;
+  delete[] ipntr;
+  delete[] workd;
+  delete[] workl;
+  delete[] resid;
+
+  m_isInitialized = true;
+
+  return *this;
+}
+
+
+// Single precision
+//
+extern "C" void ssaupd_(int *ido, char *bmat, int *n, char *which,
+    int *nev, float *tol, float *resid, int *ncv,
+    float *v, int *ldv, int *iparam, int *ipntr,
+    float *workd, float *workl, int *lworkl,
+    int *info);
+
+extern "C" void sseupd_(int *rvec, char *All, int *select, float *d,
+    float *z, int *ldz, float *sigma, 
+    char *bmat, int *n, char *which, int *nev,
+    float *tol, float *resid, int *ncv, float *v,
+    int *ldv, int *iparam, int *ipntr, float *workd,
+    float *workl, int *lworkl, int *ierr);
+
+// Double precision
+//
+extern "C" void dsaupd_(int *ido, char *bmat, int *n, char *which,
+    int *nev, double *tol, double *resid, int *ncv,
+    double *v, int *ldv, int *iparam, int *ipntr,
+    double *workd, double *workl, int *lworkl,
+    int *info);
+
+extern "C" void dseupd_(int *rvec, char *All, int *select, double *d,
+    double *z, int *ldz, double *sigma, 
+    char *bmat, int *n, char *which, int *nev,
+    double *tol, double *resid, int *ncv, double *v,
+    int *ldv, int *iparam, int *ipntr, double *workd,
+    double *workl, int *lworkl, int *ierr);
+
+
+namespace internal {
+
+template<typename Scalar, typename RealScalar> struct arpack_wrapper
+{
+  static inline void saupd(int *ido, char *bmat, int *n, char *which,
+      int *nev, RealScalar *tol, Scalar *resid, int *ncv,
+      Scalar *v, int *ldv, int *iparam, int *ipntr,
+      Scalar *workd, Scalar *workl, int *lworkl, int *info)
+  { 
+    EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL)
+  }
+
+  static inline void seupd(int *rvec, char *All, int *select, Scalar *d,
+      Scalar *z, int *ldz, RealScalar *sigma,
+      char *bmat, int *n, char *which, int *nev,
+      RealScalar *tol, Scalar *resid, int *ncv, Scalar *v,
+      int *ldv, int *iparam, int *ipntr, Scalar *workd,
+      Scalar *workl, int *lworkl, int *ierr)
+  {
+    EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL)
+  }
+};
+
+template <> struct arpack_wrapper<float, float>
+{
+  static inline void saupd(int *ido, char *bmat, int *n, char *which,
+      int *nev, float *tol, float *resid, int *ncv,
+      float *v, int *ldv, int *iparam, int *ipntr,
+      float *workd, float *workl, int *lworkl, int *info)
+  {
+    ssaupd_(ido, bmat, n, which, nev, tol, resid, ncv, v, ldv, iparam, ipntr, workd, workl, lworkl, info);
+  }
+
+  static inline void seupd(int *rvec, char *All, int *select, float *d,
+      float *z, int *ldz, float *sigma,
+      char *bmat, int *n, char *which, int *nev,
+      float *tol, float *resid, int *ncv, float *v,
+      int *ldv, int *iparam, int *ipntr, float *workd,
+      float *workl, int *lworkl, int *ierr)
+  {
+    sseupd_(rvec, All, select, d, z, ldz, sigma, bmat, n, which, nev, tol, resid, ncv, v, ldv, iparam, ipntr,
+        workd, workl, lworkl, ierr);
+  }
+};
+
+template <> struct arpack_wrapper<double, double>
+{
+  static inline void saupd(int *ido, char *bmat, int *n, char *which,
+      int *nev, double *tol, double *resid, int *ncv,
+      double *v, int *ldv, int *iparam, int *ipntr,
+      double *workd, double *workl, int *lworkl, int *info)
+  {
+    dsaupd_(ido, bmat, n, which, nev, tol, resid, ncv, v, ldv, iparam, ipntr, workd, workl, lworkl, info);
+  }
+
+  static inline void seupd(int *rvec, char *All, int *select, double *d,
+      double *z, int *ldz, double *sigma,
+      char *bmat, int *n, char *which, int *nev,
+      double *tol, double *resid, int *ncv, double *v,
+      int *ldv, int *iparam, int *ipntr, double *workd,
+      double *workl, int *lworkl, int *ierr)
+  {
+    dseupd_(rvec, All, select, d, v, ldv, sigma, bmat, n, which, nev, tol, resid, ncv, v, ldv, iparam, ipntr,
+        workd, workl, lworkl, ierr);
+  }
+};
+
+
+template<typename MatrixSolver, typename MatrixType, typename Scalar, bool BisSPD>
+struct OP
+{
+    static inline void applyOP(MatrixSolver &OP, const MatrixType &A, int n, Scalar *in, Scalar *out);
+    static inline void project(MatrixSolver &OP, int n, int k, Scalar *vecs);
+};
+
+template<typename MatrixSolver, typename MatrixType, typename Scalar>
+struct OP<MatrixSolver, MatrixType, Scalar, true>
+{
+  static inline void applyOP(MatrixSolver &OP, const MatrixType &A, int n, Scalar *in, Scalar *out)
+{
+    // OP = L^{-1} A L^{-T}  (B = LL^T)
+    //
+    // First solve L^T out = in
+    //
+    Matrix<Scalar, Dynamic, 1>::Map(out, n) = OP.matrixU().solve(Matrix<Scalar, Dynamic, 1>::Map(in, n));
+    Matrix<Scalar, Dynamic, 1>::Map(out, n) = OP.permutationPinv() * Matrix<Scalar, Dynamic, 1>::Map(out, n);
+
+    // Then compute out = A out
+    //
+    Matrix<Scalar, Dynamic, 1>::Map(out, n) = A * Matrix<Scalar, Dynamic, 1>::Map(out, n);
+
+    // Then solve L out = out
+    //
+    Matrix<Scalar, Dynamic, 1>::Map(out, n) = OP.permutationP() * Matrix<Scalar, Dynamic, 1>::Map(out, n);
+    Matrix<Scalar, Dynamic, 1>::Map(out, n) = OP.matrixL().solve(Matrix<Scalar, Dynamic, 1>::Map(out, n));
+}
+
+  static inline void project(MatrixSolver &OP, int n, int k, Scalar *vecs)
+{
+    // Solve L^T out = in
+    //
+    Matrix<Scalar, Dynamic, Dynamic>::Map(vecs, n, k) = OP.matrixU().solve(Matrix<Scalar, Dynamic, Dynamic>::Map(vecs, n, k));
+    Matrix<Scalar, Dynamic, Dynamic>::Map(vecs, n, k) = OP.permutationPinv() * Matrix<Scalar, Dynamic, Dynamic>::Map(vecs, n, k);
+}
+
+};
+
+template<typename MatrixSolver, typename MatrixType, typename Scalar>
+struct OP<MatrixSolver, MatrixType, Scalar, false>
+{
+  static inline void applyOP(MatrixSolver &OP, const MatrixType &A, int n, Scalar *in, Scalar *out)
+{
+    eigen_assert(false && "Should never be in here...");
+}
+
+  static inline void project(MatrixSolver &OP, int n, int k, Scalar *vecs)
+{
+    eigen_assert(false && "Should never be in here...");
+}
+
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_ARPACKSELFADJOINTEIGENSOLVER_H
+
diff --git a/src/EigenUnsupported/src/EulerAngles/CMakeLists.txt b/src/EigenUnsupported/src/EulerAngles/CMakeLists.txt
new file mode 100644
index 0000000..22088eb
--- /dev/null
+++ b/src/EigenUnsupported/src/EulerAngles/CMakeLists.txt
@@ -0,0 +1,6 @@
+file(GLOB Eigen_EulerAngles_SRCS "*.h")
+
+install(FILES
+  ${Eigen_EulerAngles_SRCS}
+  DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/EulerAngles COMPONENT Devel
+  )
diff --git a/src/EigenUnsupported/src/EulerAngles/EulerAngles.h b/src/EigenUnsupported/src/EulerAngles/EulerAngles.h
new file mode 100644
index 0000000..e43cdb7
--- /dev/null
+++ b/src/EigenUnsupported/src/EulerAngles/EulerAngles.h
@@ -0,0 +1,355 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Tal Hadad <tal_hd@hotmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_EULERANGLESCLASS_H// TODO: Fix previous "EIGEN_EULERANGLES_H" definition?
+#define EIGEN_EULERANGLESCLASS_H
+
+namespace Eigen
+{
+  /** \class EulerAngles
+    *
+    * \ingroup EulerAngles_Module
+    *
+    * \brief Represents a rotation in a 3 dimensional space as three Euler angles.
+    *
+    * Euler rotation is a set of three rotation of three angles over three fixed axes, defined by the EulerSystem given as a template parameter.
+    * 
+    * Here is how intrinsic Euler angles works:
+    *  - first, rotate the axes system over the alpha axis in angle alpha
+    *  - then, rotate the axes system over the beta axis(which was rotated in the first stage) in angle beta
+    *  - then, rotate the axes system over the gamma axis(which was rotated in the two stages above) in angle gamma
+    *
+    * \note This class support only intrinsic Euler angles for simplicity,
+    *  see EulerSystem how to easily overcome this for extrinsic systems.
+    *
+    * ### Rotation representation and conversions ###
+    *
+    * It has been proved(see Wikipedia link below) that every rotation can be represented
+    *  by Euler angles, but there is no single representation (e.g. unlike rotation matrices).
+    * Therefore, you can convert from Eigen rotation and to them
+    *  (including rotation matrices, which is not called "rotations" by Eigen design).
+    *
+    * Euler angles usually used for:
+    *  - convenient human representation of rotation, especially in interactive GUI.
+    *  - gimbal systems and robotics
+    *  - efficient encoding(i.e. 3 floats only) of rotation for network protocols.
+    *
+    * However, Euler angles are slow comparing to quaternion or matrices,
+    *  because their unnatural math definition, although it's simple for human.
+    * To overcome this, this class provide easy movement from the math friendly representation
+    *  to the human friendly representation, and vise-versa.
+    *
+    * All the user need to do is a safe simple C++ type conversion,
+    *  and this class take care for the math.
+    * Additionally, some axes related computation is done in compile time.
+    *
+    * #### Euler angles ranges in conversions ####
+    * Rotations representation as EulerAngles are not single (unlike matrices),
+    *  and even have infinite EulerAngles representations.<BR>
+    * For example, add or subtract 2*PI from either angle of EulerAngles
+    *  and you'll get the same rotation.
+    * This is the general reason for infinite representation,
+    *  but it's not the only general reason for not having a single representation.
+    *
+    * When converting rotation to EulerAngles, this class convert it to specific ranges
+    * When converting some rotation to EulerAngles, the rules for ranges are as follow:
+    * - If the rotation we converting from is an EulerAngles
+    *  (even when it represented as RotationBase explicitly), angles ranges are __undefined__.
+    * - otherwise, alpha and gamma angles will be in the range [-PI, PI].<BR>
+    *   As for Beta angle:
+    *    - If the system is Tait-Bryan, the beta angle will be in the range [-PI/2, PI/2].
+    *    - otherwise:
+    *      - If the beta axis is positive, the beta angle will be in the range [0, PI]
+    *      - If the beta axis is negative, the beta angle will be in the range [-PI, 0]
+    *
+    * \sa EulerAngles(const MatrixBase<Derived>&)
+    * \sa EulerAngles(const RotationBase<Derived, 3>&)
+    *
+    * ### Convenient user typedefs ###
+    *
+    * Convenient typedefs for EulerAngles exist for float and double scalar,
+    *  in a form of EulerAngles{A}{B}{C}{scalar},
+    *  e.g. \ref EulerAnglesXYZd, \ref EulerAnglesZYZf.
+    *
+    * Only for positive axes{+x,+y,+z} Euler systems are have convenient typedef.
+    * If you need negative axes{-x,-y,-z}, it is recommended to create you own typedef with
+    *  a word that represent what you need.
+    *
+    * ### Example ###
+    *
+    * \include EulerAngles.cpp
+    * Output: \verbinclude EulerAngles.out
+    *
+    * ### Additional reading ###
+    *
+    * If you're want to get more idea about how Euler system work in Eigen see EulerSystem.
+    *
+    * More information about Euler angles: https://en.wikipedia.org/wiki/Euler_angles
+    *
+    * \tparam _Scalar the scalar type, i.e. the type of the angles.
+    *
+    * \tparam _System the EulerSystem to use, which represents the axes of rotation.
+    */
+  template <typename _Scalar, class _System>
+  class EulerAngles : public RotationBase<EulerAngles<_Scalar, _System>, 3>
+  {
+    public:
+      typedef RotationBase<EulerAngles<_Scalar, _System>, 3> Base;
+      
+      /** the scalar type of the angles */
+      typedef _Scalar Scalar;
+      typedef typename NumTraits<Scalar>::Real RealScalar;
+      
+      /** the EulerSystem to use, which represents the axes of rotation. */
+      typedef _System System;
+    
+      typedef Matrix<Scalar,3,3> Matrix3; /*!< the equivalent rotation matrix type */
+      typedef Matrix<Scalar,3,1> Vector3; /*!< the equivalent 3 dimension vector type */
+      typedef Quaternion<Scalar> QuaternionType; /*!< the equivalent quaternion type */
+      typedef AngleAxis<Scalar> AngleAxisType; /*!< the equivalent angle-axis type */
+      
+      /** \returns the axis vector of the first (alpha) rotation */
+      static Vector3 AlphaAxisVector() {
+        const Vector3& u = Vector3::Unit(System::AlphaAxisAbs - 1);
+        return System::IsAlphaOpposite ? -u : u;
+      }
+      
+      /** \returns the axis vector of the second (beta) rotation */
+      static Vector3 BetaAxisVector() {
+        const Vector3& u = Vector3::Unit(System::BetaAxisAbs - 1);
+        return System::IsBetaOpposite ? -u : u;
+      }
+      
+      /** \returns the axis vector of the third (gamma) rotation */
+      static Vector3 GammaAxisVector() {
+        const Vector3& u = Vector3::Unit(System::GammaAxisAbs - 1);
+        return System::IsGammaOpposite ? -u : u;
+      }
+
+    private:
+      Vector3 m_angles;
+
+    public:
+      /** Default constructor without initialization. */
+      EulerAngles() {}
+      /** Constructs and initialize an EulerAngles (\p alpha, \p beta, \p gamma). */
+      EulerAngles(const Scalar& alpha, const Scalar& beta, const Scalar& gamma) :
+        m_angles(alpha, beta, gamma) {}
+      
+      // TODO: Test this constructor
+      /** Constructs and initialize an EulerAngles from the array data {alpha, beta, gamma} */
+      explicit EulerAngles(const Scalar* data) : m_angles(data) {}
+      
+      /** Constructs and initializes an EulerAngles from either:
+        *  - a 3x3 rotation matrix expression(i.e. pure orthogonal matrix with determinant of +1),
+        *  - a 3D vector expression representing Euler angles.
+        *
+        * \note If \p other is a 3x3 rotation matrix, the angles range rules will be as follow:<BR>
+        *  Alpha and gamma angles will be in the range [-PI, PI].<BR>
+        *  As for Beta angle:
+        *   - If the system is Tait-Bryan, the beta angle will be in the range [-PI/2, PI/2].
+        *   - otherwise:
+        *     - If the beta axis is positive, the beta angle will be in the range [0, PI]
+        *     - If the beta axis is negative, the beta angle will be in the range [-PI, 0]
+       */
+      template<typename Derived>
+      explicit EulerAngles(const MatrixBase<Derived>& other) { *this = other; }
+      
+      /** Constructs and initialize Euler angles from a rotation \p rot.
+        *
+        * \note If \p rot is an EulerAngles (even when it represented as RotationBase explicitly),
+        *  angles ranges are __undefined__.
+        *  Otherwise, alpha and gamma angles will be in the range [-PI, PI].<BR>
+        *  As for Beta angle:
+        *   - If the system is Tait-Bryan, the beta angle will be in the range [-PI/2, PI/2].
+        *   - otherwise:
+        *     - If the beta axis is positive, the beta angle will be in the range [0, PI]
+        *     - If the beta axis is negative, the beta angle will be in the range [-PI, 0]
+      */
+      template<typename Derived>
+      EulerAngles(const RotationBase<Derived, 3>& rot) { System::CalcEulerAngles(*this, rot.toRotationMatrix()); }
+      
+      /*EulerAngles(const QuaternionType& q)
+      {
+        // TODO: Implement it in a faster way for quaternions
+        // According to http://www.euclideanspace.com/maths/geometry/rotations/conversions/quaternionToEuler/
+        //  we can compute only the needed matrix cells and then convert to euler angles. (see ZYX example below)
+        // Currently we compute all matrix cells from quaternion.
+
+        // Special case only for ZYX
+        //Scalar y2 = q.y() * q.y();
+        //m_angles[0] = std::atan2(2*(q.w()*q.z() + q.x()*q.y()), (1 - 2*(y2 + q.z()*q.z())));
+        //m_angles[1] = std::asin( 2*(q.w()*q.y() - q.z()*q.x()));
+        //m_angles[2] = std::atan2(2*(q.w()*q.x() + q.y()*q.z()), (1 - 2*(q.x()*q.x() + y2)));
+      }*/
+
+      /** \returns The angle values stored in a vector (alpha, beta, gamma). */
+      const Vector3& angles() const { return m_angles; }
+      /** \returns A read-write reference to the angle values stored in a vector (alpha, beta, gamma). */
+      Vector3& angles() { return m_angles; }
+
+      /** \returns The value of the first angle. */
+      Scalar alpha() const { return m_angles[0]; }
+      /** \returns A read-write reference to the angle of the first angle. */
+      Scalar& alpha() { return m_angles[0]; }
+
+      /** \returns The value of the second angle. */
+      Scalar beta() const { return m_angles[1]; }
+      /** \returns A read-write reference to the angle of the second angle. */
+      Scalar& beta() { return m_angles[1]; }
+
+      /** \returns The value of the third angle. */
+      Scalar gamma() const { return m_angles[2]; }
+      /** \returns A read-write reference to the angle of the third angle. */
+      Scalar& gamma() { return m_angles[2]; }
+
+      /** \returns The Euler angles rotation inverse (which is as same as the negative),
+        *  (-alpha, -beta, -gamma).
+      */
+      EulerAngles inverse() const
+      {
+        EulerAngles res;
+        res.m_angles = -m_angles;
+        return res;
+      }
+
+      /** \returns The Euler angles rotation negative (which is as same as the inverse),
+        *  (-alpha, -beta, -gamma).
+      */
+      EulerAngles operator -() const
+      {
+        return inverse();
+      }
+      
+      /** Set \c *this from either:
+        *  - a 3x3 rotation matrix expression(i.e. pure orthogonal matrix with determinant of +1),
+        *  - a 3D vector expression representing Euler angles.
+        *
+        * See EulerAngles(const MatrixBase<Derived, 3>&) for more information about
+        *  angles ranges output.
+      */
+      template<class Derived>
+      EulerAngles& operator=(const MatrixBase<Derived>& other)
+      {
+        EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename Derived::Scalar>::value),
+         YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+        
+        internal::eulerangles_assign_impl<System, Derived>::run(*this, other.derived());
+        return *this;
+      }
+
+      // TODO: Assign and construct from another EulerAngles (with different system)
+      
+      /** Set \c *this from a rotation.
+        *
+        * See EulerAngles(const RotationBase<Derived, 3>&) for more information about
+        *  angles ranges output.
+      */
+      template<typename Derived>
+      EulerAngles& operator=(const RotationBase<Derived, 3>& rot) {
+        System::CalcEulerAngles(*this, rot.toRotationMatrix());
+        return *this;
+      }
+      
+      /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+        * determined by \a prec.
+        *
+        * \sa MatrixBase::isApprox() */
+      bool isApprox(const EulerAngles& other,
+        const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const
+      { return angles().isApprox(other.angles(), prec); }
+
+      /** \returns an equivalent 3x3 rotation matrix. */
+      Matrix3 toRotationMatrix() const
+      {
+        // TODO: Calc it faster
+        return static_cast<QuaternionType>(*this).toRotationMatrix();
+      }
+
+      /** Convert the Euler angles to quaternion. */
+      operator QuaternionType() const
+      {
+        return
+          AngleAxisType(alpha(), AlphaAxisVector()) *
+          AngleAxisType(beta(), BetaAxisVector())   *
+          AngleAxisType(gamma(), GammaAxisVector());
+      }
+      
+      friend std::ostream& operator<<(std::ostream& s, const EulerAngles<Scalar, System>& eulerAngles)
+      {
+        s << eulerAngles.angles().transpose();
+        return s;
+      }
+      
+      /** \returns \c *this with scalar type casted to \a NewScalarType */
+      template <typename NewScalarType>
+      EulerAngles<NewScalarType, System> cast() const
+      {
+        EulerAngles<NewScalarType, System> e;
+        e.angles() = angles().template cast<NewScalarType>();
+        return e;
+      }
+  };
+
+#define EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(AXES, SCALAR_TYPE, SCALAR_POSTFIX) \
+  /** \ingroup EulerAngles_Module */ \
+  typedef EulerAngles<SCALAR_TYPE, EulerSystem##AXES> EulerAngles##AXES##SCALAR_POSTFIX;
+
+#define EIGEN_EULER_ANGLES_TYPEDEFS(SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(XYZ, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(XYX, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(XZY, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(XZX, SCALAR_TYPE, SCALAR_POSTFIX) \
+ \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(YZX, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(YZY, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(YXZ, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(YXY, SCALAR_TYPE, SCALAR_POSTFIX) \
+ \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(ZXY, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(ZXZ, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(ZYX, SCALAR_TYPE, SCALAR_POSTFIX) \
+  EIGEN_EULER_ANGLES_SINGLE_TYPEDEF(ZYZ, SCALAR_TYPE, SCALAR_POSTFIX)
+
+EIGEN_EULER_ANGLES_TYPEDEFS(float, f)
+EIGEN_EULER_ANGLES_TYPEDEFS(double, d)
+
+  namespace internal
+  {
+    template<typename _Scalar, class _System>
+    struct traits<EulerAngles<_Scalar, _System> >
+    {
+      typedef _Scalar Scalar;
+    };
+    
+    // set from a rotation matrix
+    template<class System, class Other>
+    struct eulerangles_assign_impl<System,Other,3,3>
+    {
+      typedef typename Other::Scalar Scalar;
+      static void run(EulerAngles<Scalar, System>& e, const Other& m)
+      {
+        System::CalcEulerAngles(e, m);
+      }
+    };
+    
+    // set from a vector of Euler angles
+    template<class System, class Other>
+    struct eulerangles_assign_impl<System,Other,3,1>
+    {
+      typedef typename Other::Scalar Scalar;
+      static void run(EulerAngles<Scalar, System>& e, const Other& vec)
+      {
+        e.angles() = vec;
+      }
+    };
+  }
+}
+
+#endif // EIGEN_EULERANGLESCLASS_H
diff --git a/src/EigenUnsupported/src/EulerAngles/EulerSystem.h b/src/EigenUnsupported/src/EulerAngles/EulerSystem.h
new file mode 100644
index 0000000..2a833b0
--- /dev/null
+++ b/src/EigenUnsupported/src/EulerAngles/EulerSystem.h
@@ -0,0 +1,305 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Tal Hadad <tal_hd@hotmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_EULERSYSTEM_H
+#define EIGEN_EULERSYSTEM_H
+
+namespace Eigen
+{
+  // Forward declarations
+  template <typename _Scalar, class _System>
+  class EulerAngles;
+  
+  namespace internal
+  {
+    // TODO: Add this trait to the Eigen internal API?
+    template <int Num, bool IsPositive = (Num > 0)>
+    struct Abs
+    {
+      enum { value = Num };
+    };
+  
+    template <int Num>
+    struct Abs<Num, false>
+    {
+      enum { value = -Num };
+    };
+
+    template <int Axis>
+    struct IsValidAxis
+    {
+      enum { value = Axis != 0 && Abs<Axis>::value <= 3 };
+    };
+    
+    template<typename System,
+            typename Other,
+            int OtherRows=Other::RowsAtCompileTime,
+            int OtherCols=Other::ColsAtCompileTime>
+    struct eulerangles_assign_impl;
+  }
+  
+  #define EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT(COND,MSG) typedef char static_assertion_##MSG[(COND)?1:-1]
+  
+  /** \brief Representation of a fixed signed rotation axis for EulerSystem.
+    *
+    * \ingroup EulerAngles_Module
+    *
+    * Values here represent:
+    *  - The axis of the rotation: X, Y or Z.
+    *  - The sign (i.e. direction of the rotation along the axis): positive(+) or negative(-)
+    *
+    * Therefore, this could express all the axes {+X,+Y,+Z,-X,-Y,-Z}
+    *
+    * For positive axis, use +EULER_{axis}, and for negative axis use -EULER_{axis}.
+    */
+  enum EulerAxis
+  {
+    EULER_X = 1, /*!< the X axis */
+    EULER_Y = 2, /*!< the Y axis */
+    EULER_Z = 3  /*!< the Z axis */
+  };
+  
+  /** \class EulerSystem
+    *
+    * \ingroup EulerAngles_Module
+    *
+    * \brief Represents a fixed Euler rotation system.
+    *
+    * This meta-class goal is to represent the Euler system in compilation time, for EulerAngles.
+    *
+    * You can use this class to get two things:
+    *  - Build an Euler system, and then pass it as a template parameter to EulerAngles.
+    *  - Query some compile time data about an Euler system. (e.g. Whether it's Tait-Bryan)
+    *
+    * Euler rotation is a set of three rotation on fixed axes. (see \ref EulerAngles)
+    * This meta-class store constantly those signed axes. (see \ref EulerAxis)
+    *
+    * ### Types of Euler systems ###
+    *
+    * All and only valid 3 dimension Euler rotation over standard
+    *  signed axes{+X,+Y,+Z,-X,-Y,-Z} are supported:
+    *  - all axes X, Y, Z in each valid order (see below what order is valid)
+    *  - rotation over the axis is supported both over the positive and negative directions.
+    *  - both Tait-Bryan and proper/classic Euler angles (i.e. the opposite).
+    *
+    * Since EulerSystem support both positive and negative directions,
+    *  you may call this rotation distinction in other names:
+    *  - _right handed_ or _left handed_
+    *  - _counterclockwise_ or _clockwise_
+    *
+    * Notice all axed combination are valid, and would trigger a static assertion.
+    * Same unsigned axes can't be neighbors, e.g. {X,X,Y} is invalid.
+    * This yield two and only two classes:
+    *  - _Tait-Bryan_ - all unsigned axes are distinct, e.g. {X,Y,Z}
+    *  - _proper/classic Euler angles_ - The first and the third unsigned axes is equal,
+    *     and the second is different, e.g. {X,Y,X}
+    *
+    * ### Intrinsic vs extrinsic Euler systems ###
+    *
+    * Only intrinsic Euler systems are supported for simplicity.
+    *  If you want to use extrinsic Euler systems,
+    *   just use the equal intrinsic opposite order for axes and angles.
+    *  I.e axes (A,B,C) becomes (C,B,A), and angles (a,b,c) becomes (c,b,a).
+    *
+    * ### Convenient user typedefs ###
+    *
+    * Convenient typedefs for EulerSystem exist (only for positive axes Euler systems),
+    *  in a form of EulerSystem{A}{B}{C}, e.g. \ref EulerSystemXYZ.
+    *
+    * ### Additional reading ###
+    *
+    * More information about Euler angles: https://en.wikipedia.org/wiki/Euler_angles
+    *
+    * \tparam _AlphaAxis the first fixed EulerAxis
+    *
+    * \tparam _BetaAxis the second fixed EulerAxis
+    *
+    * \tparam _GammaAxis the third fixed EulerAxis
+    */
+  template <int _AlphaAxis, int _BetaAxis, int _GammaAxis>
+  class EulerSystem
+  {
+    public:
+    // It's defined this way and not as enum, because I think
+    //  that enum is not guerantee to support negative numbers
+    
+    /** The first rotation axis */
+    static const int AlphaAxis = _AlphaAxis;
+    
+    /** The second rotation axis */
+    static const int BetaAxis = _BetaAxis;
+    
+    /** The third rotation axis */
+    static const int GammaAxis = _GammaAxis;
+
+    enum
+    {
+      AlphaAxisAbs = internal::Abs<AlphaAxis>::value, /*!< the first rotation axis unsigned */
+      BetaAxisAbs = internal::Abs<BetaAxis>::value, /*!< the second rotation axis unsigned */
+      GammaAxisAbs = internal::Abs<GammaAxis>::value, /*!< the third rotation axis unsigned */
+      
+      IsAlphaOpposite = (AlphaAxis < 0) ? 1 : 0, /*!< whether alpha axis is negative */
+      IsBetaOpposite = (BetaAxis < 0) ? 1 : 0, /*!< whether beta axis is negative */
+      IsGammaOpposite = (GammaAxis < 0) ? 1 : 0, /*!< whether gamma axis is negative */
+
+      // Parity is even if alpha axis X is followed by beta axis Y, or Y is followed
+      // by Z, or Z is followed by X; otherwise it is odd.
+      IsOdd = ((AlphaAxisAbs)%3 == (BetaAxisAbs - 1)%3) ? 0 : 1, /*!< whether the Euler system is odd */
+      IsEven = IsOdd ? 0 : 1, /*!< whether the Euler system is even */
+
+      IsTaitBryan = ((unsigned)AlphaAxisAbs != (unsigned)GammaAxisAbs) ? 1 : 0 /*!< whether the Euler system is Tait-Bryan */
+    };
+    
+    private:
+    
+    EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT(internal::IsValidAxis<AlphaAxis>::value,
+      ALPHA_AXIS_IS_INVALID);
+      
+    EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT(internal::IsValidAxis<BetaAxis>::value,
+      BETA_AXIS_IS_INVALID);
+      
+    EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT(internal::IsValidAxis<GammaAxis>::value,
+      GAMMA_AXIS_IS_INVALID);
+      
+    EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT((unsigned)AlphaAxisAbs != (unsigned)BetaAxisAbs,
+      ALPHA_AXIS_CANT_BE_EQUAL_TO_BETA_AXIS);
+      
+    EIGEN_EULER_ANGLES_CLASS_STATIC_ASSERT((unsigned)BetaAxisAbs != (unsigned)GammaAxisAbs,
+      BETA_AXIS_CANT_BE_EQUAL_TO_GAMMA_AXIS);
+
+    static const int
+      // I, J, K are the pivot indexes permutation for the rotation matrix, that match this Euler system. 
+      // They are used in this class converters.
+      // They are always different from each other, and their possible values are: 0, 1, or 2.
+      I_ = AlphaAxisAbs - 1,
+      J_ = (AlphaAxisAbs - 1 + 1 + IsOdd)%3,
+      K_ = (AlphaAxisAbs - 1 + 2 - IsOdd)%3
+    ;
+    
+    // TODO: Get @mat parameter in form that avoids double evaluation.
+    template <typename Derived>
+    static void CalcEulerAngles_imp(Matrix<typename MatrixBase<Derived>::Scalar, 3, 1>& res, const MatrixBase<Derived>& mat, internal::true_type /*isTaitBryan*/)
+    {
+      using std::atan2;
+      using std::sqrt;
+      
+      typedef typename Derived::Scalar Scalar;
+
+      const Scalar plusMinus = IsEven? 1 : -1;
+      const Scalar minusPlus = IsOdd?  1 : -1;
+
+      const Scalar Rsum = sqrt((mat(I_,I_) * mat(I_,I_) + mat(I_,J_) * mat(I_,J_) + mat(J_,K_) * mat(J_,K_) + mat(K_,K_) * mat(K_,K_))/2);
+      res[1] = atan2(plusMinus * mat(I_,K_), Rsum);
+
+      // There is a singularity when cos(beta) == 0
+      if(Rsum > 4 * NumTraits<Scalar>::epsilon()) {// cos(beta) != 0
+        res[0] = atan2(minusPlus * mat(J_, K_), mat(K_, K_));
+        res[2] = atan2(minusPlus * mat(I_, J_), mat(I_, I_));
+      }
+      else if(plusMinus * mat(I_, K_) > 0) {// cos(beta) == 0 and sin(beta) == 1
+        Scalar spos = mat(J_, I_) + plusMinus * mat(K_, J_); // 2*sin(alpha + plusMinus * gamma
+        Scalar cpos = mat(J_, J_) + minusPlus * mat(K_, I_); // 2*cos(alpha + plusMinus * gamma)
+        Scalar alphaPlusMinusGamma = atan2(spos, cpos);
+        res[0] = alphaPlusMinusGamma;
+        res[2] = 0;
+      }
+      else {// cos(beta) == 0 and sin(beta) == -1
+        Scalar sneg = plusMinus * (mat(K_, J_) + minusPlus * mat(J_, I_)); // 2*sin(alpha + minusPlus*gamma)
+        Scalar cneg = mat(J_, J_) + plusMinus * mat(K_, I_);               // 2*cos(alpha + minusPlus*gamma)
+        Scalar alphaMinusPlusBeta = atan2(sneg, cneg);
+        res[0] = alphaMinusPlusBeta;
+        res[2] = 0;
+      }
+    }
+
+    template <typename Derived>
+    static void CalcEulerAngles_imp(Matrix<typename MatrixBase<Derived>::Scalar,3,1>& res,
+                                    const MatrixBase<Derived>& mat, internal::false_type /*isTaitBryan*/)
+    {
+      using std::atan2;
+      using std::sqrt;
+
+      typedef typename Derived::Scalar Scalar;
+
+      const Scalar plusMinus = IsEven? 1 : -1;
+      const Scalar minusPlus = IsOdd?  1 : -1;
+
+      const Scalar Rsum = sqrt((mat(I_, J_) * mat(I_, J_) + mat(I_, K_) * mat(I_, K_) + mat(J_, I_) * mat(J_, I_) + mat(K_, I_) * mat(K_, I_)) / 2);
+
+      res[1] = atan2(Rsum, mat(I_, I_));
+
+      // There is a singularity when sin(beta) == 0
+      if(Rsum > 4 * NumTraits<Scalar>::epsilon()) {// sin(beta) != 0
+        res[0] = atan2(mat(J_, I_), minusPlus * mat(K_, I_));
+        res[2] = atan2(mat(I_, J_), plusMinus * mat(I_, K_));
+      }
+      else if(mat(I_, I_) > 0) {// sin(beta) == 0 and cos(beta) == 1
+        Scalar spos = plusMinus * mat(K_, J_) + minusPlus * mat(J_, K_); // 2*sin(alpha + gamma)
+        Scalar cpos = mat(J_, J_) + mat(K_, K_);                         // 2*cos(alpha + gamma)
+        res[0] = atan2(spos, cpos);
+        res[2] = 0;
+      }
+      else {// sin(beta) == 0 and cos(beta) == -1
+        Scalar sneg = plusMinus * mat(K_, J_) + plusMinus * mat(J_, K_); // 2*sin(alpha - gamma)
+        Scalar cneg = mat(J_, J_) - mat(K_, K_);                         // 2*cos(alpha - gamma)
+        res[0] = atan2(sneg, cneg);
+        res[2] = 0;
+      }
+    }
+    
+    template<typename Scalar>
+    static void CalcEulerAngles(
+      EulerAngles<Scalar, EulerSystem>& res,
+      const typename EulerAngles<Scalar, EulerSystem>::Matrix3& mat)
+    {
+      CalcEulerAngles_imp(
+        res.angles(), mat,
+        typename internal::conditional<IsTaitBryan, internal::true_type, internal::false_type>::type());
+
+      if (IsAlphaOpposite)
+        res.alpha() = -res.alpha();
+        
+      if (IsBetaOpposite)
+        res.beta() = -res.beta();
+        
+      if (IsGammaOpposite)
+        res.gamma() = -res.gamma();
+    }
+    
+    template <typename _Scalar, class _System>
+    friend class Eigen::EulerAngles;
+    
+    template<typename System,
+            typename Other,
+            int OtherRows,
+            int OtherCols>
+    friend struct internal::eulerangles_assign_impl;
+  };
+
+#define EIGEN_EULER_SYSTEM_TYPEDEF(A, B, C) \
+  /** \ingroup EulerAngles_Module */ \
+  typedef EulerSystem<EULER_##A, EULER_##B, EULER_##C> EulerSystem##A##B##C;
+  
+  EIGEN_EULER_SYSTEM_TYPEDEF(X,Y,Z)
+  EIGEN_EULER_SYSTEM_TYPEDEF(X,Y,X)
+  EIGEN_EULER_SYSTEM_TYPEDEF(X,Z,Y)
+  EIGEN_EULER_SYSTEM_TYPEDEF(X,Z,X)
+  
+  EIGEN_EULER_SYSTEM_TYPEDEF(Y,Z,X)
+  EIGEN_EULER_SYSTEM_TYPEDEF(Y,Z,Y)
+  EIGEN_EULER_SYSTEM_TYPEDEF(Y,X,Z)
+  EIGEN_EULER_SYSTEM_TYPEDEF(Y,X,Y)
+  
+  EIGEN_EULER_SYSTEM_TYPEDEF(Z,X,Y)
+  EIGEN_EULER_SYSTEM_TYPEDEF(Z,X,Z)
+  EIGEN_EULER_SYSTEM_TYPEDEF(Z,Y,X)
+  EIGEN_EULER_SYSTEM_TYPEDEF(Z,Y,Z)
+}
+
+#endif // EIGEN_EULERSYSTEM_H
diff --git a/src/EigenUnsupported/src/FFT/ei_fftw_impl.h b/src/EigenUnsupported/src/FFT/ei_fftw_impl.h
new file mode 100644
index 0000000..1c2cd24
--- /dev/null
+++ b/src/EigenUnsupported/src/FFT/ei_fftw_impl.h
@@ -0,0 +1,261 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra. 
+//
+// Copyright (C) 2009 Mark Borgerding mark a borgerding net
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+namespace Eigen { 
+
+namespace internal {
+
+  // FFTW uses non-const arguments
+  // so we must use ugly const_cast calls for all the args it uses
+  //
+  // This should be safe as long as 
+  // 1. we use FFTW_ESTIMATE for all our planning
+  //       see the FFTW docs section 4.3.2 "Planner Flags"
+  // 2. fftw_complex is compatible with std::complex
+  //    This assumes std::complex<T> layout is array of size 2 with real,imag
+  template <typename T> 
+  inline 
+  T * fftw_cast(const T* p)
+  { 
+      return const_cast<T*>( p); 
+  }
+
+  inline 
+  fftw_complex * fftw_cast( const std::complex<double> * p)
+  {
+      return const_cast<fftw_complex*>( reinterpret_cast<const fftw_complex*>(p) ); 
+  }
+
+  inline 
+  fftwf_complex * fftw_cast( const std::complex<float> * p)
+  { 
+      return const_cast<fftwf_complex*>( reinterpret_cast<const fftwf_complex*>(p) ); 
+  }
+
+  inline 
+  fftwl_complex * fftw_cast( const std::complex<long double> * p)
+  { 
+      return const_cast<fftwl_complex*>( reinterpret_cast<const fftwl_complex*>(p) ); 
+  }
+
+  template <typename T> 
+  struct fftw_plan {};
+
+  template <> 
+  struct fftw_plan<float>
+  {
+      typedef float scalar_type;
+      typedef fftwf_complex complex_type;
+      fftwf_plan m_plan;
+      fftw_plan() :m_plan(NULL) {}
+      ~fftw_plan() {if (m_plan) fftwf_destroy_plan(m_plan);}
+
+      inline
+      void fwd(complex_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftwf_plan_dft_1d(nfft,src,dst, FFTW_FORWARD, FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwf_execute_dft( m_plan, src,dst);
+      }
+      inline
+      void inv(complex_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftwf_plan_dft_1d(nfft,src,dst, FFTW_BACKWARD , FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwf_execute_dft( m_plan, src,dst);
+      }
+      inline
+      void fwd(complex_type * dst,scalar_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftwf_plan_dft_r2c_1d(nfft,src,dst,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwf_execute_dft_r2c( m_plan,src,dst);
+      }
+      inline
+      void inv(scalar_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL)
+              m_plan = fftwf_plan_dft_c2r_1d(nfft,src,dst,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwf_execute_dft_c2r( m_plan, src,dst);
+      }
+
+      inline 
+      void fwd2( complex_type * dst,complex_type * src,int n0,int n1) {
+          if (m_plan==NULL) m_plan = fftwf_plan_dft_2d(n0,n1,src,dst,FFTW_FORWARD,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwf_execute_dft( m_plan, src,dst);
+      }
+      inline 
+      void inv2( complex_type * dst,complex_type * src,int n0,int n1) {
+          if (m_plan==NULL) m_plan = fftwf_plan_dft_2d(n0,n1,src,dst,FFTW_BACKWARD,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwf_execute_dft( m_plan, src,dst);
+      }
+
+  };
+  template <> 
+  struct fftw_plan<double>
+  {
+      typedef double scalar_type;
+      typedef fftw_complex complex_type;
+      ::fftw_plan m_plan;
+      fftw_plan() :m_plan(NULL) {}
+      ~fftw_plan() {if (m_plan) fftw_destroy_plan(m_plan);}
+
+      inline
+      void fwd(complex_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftw_plan_dft_1d(nfft,src,dst, FFTW_FORWARD, FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftw_execute_dft( m_plan, src,dst);
+      }
+      inline
+      void inv(complex_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftw_plan_dft_1d(nfft,src,dst, FFTW_BACKWARD , FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftw_execute_dft( m_plan, src,dst);
+      }
+      inline
+      void fwd(complex_type * dst,scalar_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftw_plan_dft_r2c_1d(nfft,src,dst,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftw_execute_dft_r2c( m_plan,src,dst);
+      }
+      inline
+      void inv(scalar_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL)
+              m_plan = fftw_plan_dft_c2r_1d(nfft,src,dst,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftw_execute_dft_c2r( m_plan, src,dst);
+      }
+      inline 
+      void fwd2( complex_type * dst,complex_type * src,int n0,int n1) {
+          if (m_plan==NULL) m_plan = fftw_plan_dft_2d(n0,n1,src,dst,FFTW_FORWARD,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftw_execute_dft( m_plan, src,dst);
+      }
+      inline 
+      void inv2( complex_type * dst,complex_type * src,int n0,int n1) {
+          if (m_plan==NULL) m_plan = fftw_plan_dft_2d(n0,n1,src,dst,FFTW_BACKWARD,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftw_execute_dft( m_plan, src,dst);
+      }
+  };
+  template <> 
+  struct fftw_plan<long double>
+  {
+      typedef long double scalar_type;
+      typedef fftwl_complex complex_type;
+      fftwl_plan m_plan;
+      fftw_plan() :m_plan(NULL) {}
+      ~fftw_plan() {if (m_plan) fftwl_destroy_plan(m_plan);}
+
+      inline
+      void fwd(complex_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftwl_plan_dft_1d(nfft,src,dst, FFTW_FORWARD, FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwl_execute_dft( m_plan, src,dst);
+      }
+      inline
+      void inv(complex_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftwl_plan_dft_1d(nfft,src,dst, FFTW_BACKWARD , FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwl_execute_dft( m_plan, src,dst);
+      }
+      inline
+      void fwd(complex_type * dst,scalar_type * src,int nfft) {
+          if (m_plan==NULL) m_plan = fftwl_plan_dft_r2c_1d(nfft,src,dst,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwl_execute_dft_r2c( m_plan,src,dst);
+      }
+      inline
+      void inv(scalar_type * dst,complex_type * src,int nfft) {
+          if (m_plan==NULL)
+              m_plan = fftwl_plan_dft_c2r_1d(nfft,src,dst,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwl_execute_dft_c2r( m_plan, src,dst);
+      }
+      inline 
+      void fwd2( complex_type * dst,complex_type * src,int n0,int n1) {
+          if (m_plan==NULL) m_plan = fftwl_plan_dft_2d(n0,n1,src,dst,FFTW_FORWARD,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwl_execute_dft( m_plan, src,dst);
+      }
+      inline 
+      void inv2( complex_type * dst,complex_type * src,int n0,int n1) {
+          if (m_plan==NULL) m_plan = fftwl_plan_dft_2d(n0,n1,src,dst,FFTW_BACKWARD,FFTW_ESTIMATE|FFTW_PRESERVE_INPUT);
+          fftwl_execute_dft( m_plan, src,dst);
+      }
+  };
+
+  template <typename _Scalar>
+  struct fftw_impl
+  {
+      typedef _Scalar Scalar;
+      typedef std::complex<Scalar> Complex;
+
+      inline
+      void clear() 
+      {
+        m_plans.clear();
+      }
+
+      // complex-to-complex forward FFT
+      inline
+      void fwd( Complex * dst,const Complex *src,int nfft)
+      {
+        get_plan(nfft,false,dst,src).fwd(fftw_cast(dst), fftw_cast(src),nfft );
+      }
+
+      // real-to-complex forward FFT
+      inline
+      void fwd( Complex * dst,const Scalar * src,int nfft) 
+      {
+          get_plan(nfft,false,dst,src).fwd(fftw_cast(dst), fftw_cast(src) ,nfft);
+      }
+
+      // 2-d complex-to-complex
+      inline
+      void fwd2(Complex * dst, const Complex * src, int n0,int n1)
+      {
+          get_plan(n0,n1,false,dst,src).fwd2(fftw_cast(dst), fftw_cast(src) ,n0,n1);
+      }
+
+      // inverse complex-to-complex
+      inline
+      void inv(Complex * dst,const Complex  *src,int nfft)
+      {
+        get_plan(nfft,true,dst,src).inv(fftw_cast(dst), fftw_cast(src),nfft );
+      }
+
+      // half-complex to scalar
+      inline
+      void inv( Scalar * dst,const Complex * src,int nfft) 
+      {
+        get_plan(nfft,true,dst,src).inv(fftw_cast(dst), fftw_cast(src),nfft );
+      }
+
+      // 2-d complex-to-complex
+      inline
+      void inv2(Complex * dst, const Complex * src, int n0,int n1)
+      {
+        get_plan(n0,n1,true,dst,src).inv2(fftw_cast(dst), fftw_cast(src) ,n0,n1);
+      }
+
+
+  protected:
+      typedef fftw_plan<Scalar> PlanData;
+
+      typedef Eigen::numext::int64_t int64_t;
+
+      typedef std::map<int64_t,PlanData> PlanMap;
+
+      PlanMap m_plans;
+
+      inline
+      PlanData & get_plan(int nfft,bool inverse,void * dst,const void * src)
+      {
+          bool inplace = (dst==src);
+          bool aligned = ( (reinterpret_cast<size_t>(src)&15) | (reinterpret_cast<size_t>(dst)&15) ) == 0;
+          int64_t key = ( (nfft<<3 ) | (inverse<<2) | (inplace<<1) | aligned ) << 1;
+          return m_plans[key];
+      }
+
+      inline
+      PlanData & get_plan(int n0,int n1,bool inverse,void * dst,const void * src)
+      {
+          bool inplace = (dst==src);
+          bool aligned = ( (reinterpret_cast<size_t>(src)&15) | (reinterpret_cast<size_t>(dst)&15) ) == 0;
+          int64_t key = ( ( (((int64_t)n0) << 30)|(n1<<3 ) | (inverse<<2) | (inplace<<1) | aligned ) << 1 ) + 1;
+          return m_plans[key];
+      }
+  };
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/FFT/ei_kissfft_impl.h b/src/EigenUnsupported/src/FFT/ei_kissfft_impl.h
new file mode 100644
index 0000000..430953a
--- /dev/null
+++ b/src/EigenUnsupported/src/FFT/ei_kissfft_impl.h
@@ -0,0 +1,449 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Mark Borgerding mark a borgerding net
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+namespace Eigen { 
+
+namespace internal {
+
+  // This FFT implementation was derived from kissfft http:sourceforge.net/projects/kissfft
+  // Copyright 2003-2009 Mark Borgerding
+
+template <typename _Scalar>
+struct kiss_cpx_fft
+{
+  typedef _Scalar Scalar;
+  typedef std::complex<Scalar> Complex;
+  std::vector<Complex> m_twiddles;
+  std::vector<int> m_stageRadix;
+  std::vector<int> m_stageRemainder;
+  std::vector<Complex> m_scratchBuf;
+  bool m_inverse;
+
+  inline void make_twiddles(int nfft, bool inverse)
+  {
+    using numext::sin;
+    using numext::cos;
+    m_inverse = inverse;
+    m_twiddles.resize(nfft);
+    double phinc =  0.25 * double(EIGEN_PI) / nfft;
+    Scalar flip = inverse ? Scalar(1) : Scalar(-1);
+    m_twiddles[0] = Complex(Scalar(1), Scalar(0));
+    if ((nfft&1)==0)
+      m_twiddles[nfft/2] = Complex(Scalar(-1), Scalar(0));
+    int i=1;
+    for (;i*8<nfft;++i)
+    {
+      Scalar c = Scalar(cos(i*8*phinc));
+      Scalar s = Scalar(sin(i*8*phinc));
+      m_twiddles[i] = Complex(c, s*flip);
+      m_twiddles[nfft-i] = Complex(c, -s*flip);
+    }
+    for (;i*4<nfft;++i)
+    {
+      Scalar c = Scalar(cos((2*nfft-8*i)*phinc));
+      Scalar s = Scalar(sin((2*nfft-8*i)*phinc));
+      m_twiddles[i] = Complex(s, c*flip);
+      m_twiddles[nfft-i] = Complex(s, -c*flip);
+    }
+    for (;i*8<3*nfft;++i)
+    {
+      Scalar c = Scalar(cos((8*i-2*nfft)*phinc));
+      Scalar s = Scalar(sin((8*i-2*nfft)*phinc));
+      m_twiddles[i] = Complex(-s, c*flip);
+      m_twiddles[nfft-i] = Complex(-s, -c*flip);
+    }
+    for (;i*2<nfft;++i)
+    {
+      Scalar c = Scalar(cos((4*nfft-8*i)*phinc));
+      Scalar s = Scalar(sin((4*nfft-8*i)*phinc));
+      m_twiddles[i] = Complex(-c, s*flip);
+      m_twiddles[nfft-i] = Complex(-c, -s*flip);
+    }
+  }
+
+  void factorize(int nfft)
+  {
+    //start factoring out 4's, then 2's, then 3,5,7,9,...
+    int n= nfft;
+    int p=4;
+    do {
+      while (n % p) {
+        switch (p) {
+          case 4: p = 2; break;
+          case 2: p = 3; break;
+          default: p += 2; break;
+        }
+        if (p*p>n)
+          p=n;// impossible to have a factor > sqrt(n)
+      }
+      n /= p;
+      m_stageRadix.push_back(p);
+      m_stageRemainder.push_back(n);
+      if ( p > 5 )
+        m_scratchBuf.resize(p); // scratchbuf will be needed in bfly_generic
+    }while(n>1);
+  }
+
+  template <typename _Src>
+    inline
+    void work( int stage,Complex * xout, const _Src * xin, size_t fstride,size_t in_stride)
+    {
+      int p = m_stageRadix[stage];
+      int m = m_stageRemainder[stage];
+      Complex * Fout_beg = xout;
+      Complex * Fout_end = xout + p*m;
+
+      if (m>1) {
+        do{
+          // recursive call:
+          // DFT of size m*p performed by doing
+          // p instances of smaller DFTs of size m, 
+          // each one takes a decimated version of the input
+          work(stage+1, xout , xin, fstride*p,in_stride);
+          xin += fstride*in_stride;
+        }while( (xout += m) != Fout_end );
+      }else{
+        do{
+          *xout = *xin;
+          xin += fstride*in_stride;
+        }while(++xout != Fout_end );
+      }
+      xout=Fout_beg;
+
+      // recombine the p smaller DFTs 
+      switch (p) {
+        case 2: bfly2(xout,fstride,m); break;
+        case 3: bfly3(xout,fstride,m); break;
+        case 4: bfly4(xout,fstride,m); break;
+        case 5: bfly5(xout,fstride,m); break;
+        default: bfly_generic(xout,fstride,m,p); break;
+      }
+    }
+
+  inline
+    void bfly2( Complex * Fout, const size_t fstride, int m)
+    {
+      for (int k=0;k<m;++k) {
+        Complex t = Fout[m+k] * m_twiddles[k*fstride];
+        Fout[m+k] = Fout[k] - t;
+        Fout[k] += t;
+      }
+    }
+
+  inline
+    void bfly4( Complex * Fout, const size_t fstride, const size_t m)
+    {
+      Complex scratch[6];
+      int negative_if_inverse = m_inverse * -2 +1;
+      for (size_t k=0;k<m;++k) {
+        scratch[0] = Fout[k+m] * m_twiddles[k*fstride];
+        scratch[1] = Fout[k+2*m] * m_twiddles[k*fstride*2];
+        scratch[2] = Fout[k+3*m] * m_twiddles[k*fstride*3];
+        scratch[5] = Fout[k] - scratch[1];
+
+        Fout[k] += scratch[1];
+        scratch[3] = scratch[0] + scratch[2];
+        scratch[4] = scratch[0] - scratch[2];
+        scratch[4] = Complex( scratch[4].imag()*negative_if_inverse , -scratch[4].real()* negative_if_inverse );
+
+        Fout[k+2*m]  = Fout[k] - scratch[3];
+        Fout[k] += scratch[3];
+        Fout[k+m] = scratch[5] + scratch[4];
+        Fout[k+3*m] = scratch[5] - scratch[4];
+      }
+    }
+
+  inline
+    void bfly3( Complex * Fout, const size_t fstride, const size_t m)
+    {
+      size_t k=m;
+      const size_t m2 = 2*m;
+      Complex *tw1,*tw2;
+      Complex scratch[5];
+      Complex epi3;
+      epi3 = m_twiddles[fstride*m];
+
+      tw1=tw2=&m_twiddles[0];
+
+      do{
+        scratch[1]=Fout[m] * *tw1;
+        scratch[2]=Fout[m2] * *tw2;
+
+        scratch[3]=scratch[1]+scratch[2];
+        scratch[0]=scratch[1]-scratch[2];
+        tw1 += fstride;
+        tw2 += fstride*2;
+        Fout[m] = Complex( Fout->real() - Scalar(.5)*scratch[3].real() , Fout->imag() - Scalar(.5)*scratch[3].imag() );
+        scratch[0] *= epi3.imag();
+        *Fout += scratch[3];
+        Fout[m2] = Complex(  Fout[m].real() + scratch[0].imag() , Fout[m].imag() - scratch[0].real() );
+        Fout[m] += Complex( -scratch[0].imag(),scratch[0].real() );
+        ++Fout;
+      }while(--k);
+    }
+
+  inline
+    void bfly5( Complex * Fout, const size_t fstride, const size_t m)
+    {
+      Complex *Fout0,*Fout1,*Fout2,*Fout3,*Fout4;
+      size_t u;
+      Complex scratch[13];
+      Complex * twiddles = &m_twiddles[0];
+      Complex *tw;
+      Complex ya,yb;
+      ya = twiddles[fstride*m];
+      yb = twiddles[fstride*2*m];
+
+      Fout0=Fout;
+      Fout1=Fout0+m;
+      Fout2=Fout0+2*m;
+      Fout3=Fout0+3*m;
+      Fout4=Fout0+4*m;
+
+      tw=twiddles;
+      for ( u=0; u<m; ++u ) {
+        scratch[0] = *Fout0;
+
+        scratch[1]  = *Fout1 * tw[u*fstride];
+        scratch[2]  = *Fout2 * tw[2*u*fstride];
+        scratch[3]  = *Fout3 * tw[3*u*fstride];
+        scratch[4]  = *Fout4 * tw[4*u*fstride];
+
+        scratch[7] = scratch[1] + scratch[4];
+        scratch[10] = scratch[1] - scratch[4];
+        scratch[8] = scratch[2] + scratch[3];
+        scratch[9] = scratch[2] - scratch[3];
+
+        *Fout0 +=  scratch[7];
+        *Fout0 +=  scratch[8];
+
+        scratch[5] = scratch[0] + Complex(
+            (scratch[7].real()*ya.real() ) + (scratch[8].real() *yb.real() ),
+            (scratch[7].imag()*ya.real()) + (scratch[8].imag()*yb.real())
+            );
+
+        scratch[6] = Complex(
+            (scratch[10].imag()*ya.imag()) + (scratch[9].imag()*yb.imag()),
+            -(scratch[10].real()*ya.imag()) - (scratch[9].real()*yb.imag())
+            );
+
+        *Fout1 = scratch[5] - scratch[6];
+        *Fout4 = scratch[5] + scratch[6];
+
+        scratch[11] = scratch[0] +
+          Complex(
+              (scratch[7].real()*yb.real()) + (scratch[8].real()*ya.real()),
+              (scratch[7].imag()*yb.real()) + (scratch[8].imag()*ya.real())
+              );
+
+        scratch[12] = Complex(
+            -(scratch[10].imag()*yb.imag()) + (scratch[9].imag()*ya.imag()),
+            (scratch[10].real()*yb.imag()) - (scratch[9].real()*ya.imag())
+            );
+
+        *Fout2=scratch[11]+scratch[12];
+        *Fout3=scratch[11]-scratch[12];
+
+        ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4;
+      }
+    }
+
+  /* perform the butterfly for one stage of a mixed radix FFT */
+  inline
+    void bfly_generic(
+        Complex * Fout,
+        const size_t fstride,
+        int m,
+        int p
+        )
+    {
+      int u,k,q1,q;
+      Complex * twiddles = &m_twiddles[0];
+      Complex t;
+      int Norig = static_cast<int>(m_twiddles.size());
+      Complex * scratchbuf = &m_scratchBuf[0];
+
+      for ( u=0; u<m; ++u ) {
+        k=u;
+        for ( q1=0 ; q1<p ; ++q1 ) {
+          scratchbuf[q1] = Fout[ k  ];
+          k += m;
+        }
+
+        k=u;
+        for ( q1=0 ; q1<p ; ++q1 ) {
+          int twidx=0;
+          Fout[ k ] = scratchbuf[0];
+          for (q=1;q<p;++q ) {
+            twidx += static_cast<int>(fstride) * k;
+            if (twidx>=Norig) twidx-=Norig;
+            t=scratchbuf[q] * twiddles[twidx];
+            Fout[ k ] += t;
+          }
+          k += m;
+        }
+      }
+    }
+};
+
+template <typename _Scalar>
+struct kissfft_impl
+{
+  typedef _Scalar Scalar;
+  typedef std::complex<Scalar> Complex;
+
+  void clear() 
+  {
+    m_plans.clear();
+    m_realTwiddles.clear();
+  }
+
+  inline
+    void fwd( Complex * dst,const Complex *src,int nfft)
+    {
+      get_plan(nfft,false).work(0, dst, src, 1,1);
+    }
+
+  inline
+    void fwd2( Complex * dst,const Complex *src,int n0,int n1)
+    {
+        EIGEN_UNUSED_VARIABLE(dst);
+        EIGEN_UNUSED_VARIABLE(src);
+        EIGEN_UNUSED_VARIABLE(n0);
+        EIGEN_UNUSED_VARIABLE(n1);
+    }
+
+  inline
+    void inv2( Complex * dst,const Complex *src,int n0,int n1)
+    {
+        EIGEN_UNUSED_VARIABLE(dst);
+        EIGEN_UNUSED_VARIABLE(src);
+        EIGEN_UNUSED_VARIABLE(n0);
+        EIGEN_UNUSED_VARIABLE(n1);
+    }
+
+  // real-to-complex forward FFT
+  // perform two FFTs of src even and src odd
+  // then twiddle to recombine them into the half-spectrum format
+  // then fill in the conjugate symmetric half
+  inline
+    void fwd( Complex * dst,const Scalar * src,int nfft) 
+    {
+      if ( nfft&3  ) {
+        // use generic mode for odd
+        m_tmpBuf1.resize(nfft);
+        get_plan(nfft,false).work(0, &m_tmpBuf1[0], src, 1,1);
+        std::copy(m_tmpBuf1.begin(),m_tmpBuf1.begin()+(nfft>>1)+1,dst );
+      }else{
+        int ncfft = nfft>>1;
+        int ncfft2 = nfft>>2;
+        Complex * rtw = real_twiddles(ncfft2);
+
+        // use optimized mode for even real
+        fwd( dst, reinterpret_cast<const Complex*> (src), ncfft);
+        Complex dc(dst[0].real() +  dst[0].imag());
+        Complex nyquist(dst[0].real() -  dst[0].imag());
+        int k;
+        for ( k=1;k <= ncfft2 ; ++k ) {
+          Complex fpk = dst[k];
+          Complex fpnk = conj(dst[ncfft-k]);
+          Complex f1k = fpk + fpnk;
+          Complex f2k = fpk - fpnk;
+          Complex tw= f2k * rtw[k-1];
+          dst[k] =  (f1k + tw) * Scalar(.5);
+          dst[ncfft-k] =  conj(f1k -tw)*Scalar(.5);
+        }
+        dst[0] = dc;
+        dst[ncfft] = nyquist;
+      }
+    }
+
+  // inverse complex-to-complex
+  inline
+    void inv(Complex * dst,const Complex  *src,int nfft)
+    {
+      get_plan(nfft,true).work(0, dst, src, 1,1);
+    }
+
+  // half-complex to scalar
+  inline
+    void inv( Scalar * dst,const Complex * src,int nfft) 
+    {
+      if (nfft&3) {
+        m_tmpBuf1.resize(nfft);
+        m_tmpBuf2.resize(nfft);
+        std::copy(src,src+(nfft>>1)+1,m_tmpBuf1.begin() );
+        for (int k=1;k<(nfft>>1)+1;++k)
+          m_tmpBuf1[nfft-k] = conj(m_tmpBuf1[k]);
+        inv(&m_tmpBuf2[0],&m_tmpBuf1[0],nfft);
+        for (int k=0;k<nfft;++k)
+          dst[k] = m_tmpBuf2[k].real();
+      }else{
+        // optimized version for multiple of 4
+        int ncfft = nfft>>1;
+        int ncfft2 = nfft>>2;
+        Complex * rtw = real_twiddles(ncfft2);
+        m_tmpBuf1.resize(ncfft);
+        m_tmpBuf1[0] = Complex( src[0].real() + src[ncfft].real(), src[0].real() - src[ncfft].real() );
+        for (int k = 1; k <= ncfft / 2; ++k) {
+          Complex fk = src[k];
+          Complex fnkc = conj(src[ncfft-k]);
+          Complex fek = fk + fnkc;
+          Complex tmp = fk - fnkc;
+          Complex fok = tmp * conj(rtw[k-1]);
+          m_tmpBuf1[k] = fek + fok;
+          m_tmpBuf1[ncfft-k] = conj(fek - fok);
+        }
+        get_plan(ncfft,true).work(0, reinterpret_cast<Complex*>(dst), &m_tmpBuf1[0], 1,1);
+      }
+    }
+
+  protected:
+  typedef kiss_cpx_fft<Scalar> PlanData;
+  typedef std::map<int,PlanData> PlanMap;
+
+  PlanMap m_plans;
+  std::map<int, std::vector<Complex> > m_realTwiddles;
+  std::vector<Complex> m_tmpBuf1;
+  std::vector<Complex> m_tmpBuf2;
+
+  inline
+    int PlanKey(int nfft, bool isinverse) const { return (nfft<<1) | int(isinverse); }
+
+  inline
+    PlanData & get_plan(int nfft, bool inverse)
+    {
+      // TODO look for PlanKey(nfft, ! inverse) and conjugate the twiddles
+      PlanData & pd = m_plans[ PlanKey(nfft,inverse) ];
+      if ( pd.m_twiddles.size() == 0 ) {
+        pd.make_twiddles(nfft,inverse);
+        pd.factorize(nfft);
+      }
+      return pd;
+    }
+
+  inline
+    Complex * real_twiddles(int ncfft2)
+    {
+      using std::acos;
+      std::vector<Complex> & twidref = m_realTwiddles[ncfft2];// creates new if not there
+      if ( (int)twidref.size() != ncfft2 ) {
+        twidref.resize(ncfft2);
+        int ncfft= ncfft2<<1;
+        Scalar pi =  acos( Scalar(-1) );
+        for (int k=1;k<=ncfft2;++k) 
+          twidref[k-1] = exp( Complex(0,-pi * (Scalar(k) / ncfft + Scalar(.5)) ) );
+      }
+      return &twidref[0];
+    }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/IterativeSolvers/ConstrainedConjGrad.h b/src/EigenUnsupported/src/IterativeSolvers/ConstrainedConjGrad.h
new file mode 100644
index 0000000..e7d70f3
--- /dev/null
+++ b/src/EigenUnsupported/src/IterativeSolvers/ConstrainedConjGrad.h
@@ -0,0 +1,187 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+
+/* NOTE The functions of this file have been adapted from the GMM++ library */
+
+//========================================================================
+//
+// Copyright (C) 2002-2007 Yves Renard
+//
+// This file is a part of GETFEM++
+//
+// Getfem++ is free software; you can redistribute it and/or modify
+// it under the terms of the GNU Lesser General Public License as
+// published by the Free Software Foundation; version 2.1 of the License.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU Lesser General Public License for more details.
+// You should have received a copy of the GNU Lesser General Public
+// License along with this program; if not, write to the Free Software
+// Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301,
+// USA.
+//
+//========================================================================
+
+#include "../../../../Eigen/src/Core/util/NonMPL2.h"
+
+#ifndef EIGEN_CONSTRAINEDCG_H
+#define EIGEN_CONSTRAINEDCG_H
+
+#include "../../../../Eigen/Core"
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \ingroup IterativeLinearSolvers_Module
+  * Compute the pseudo inverse of the non-square matrix C such that
+  * \f$ CINV = (C * C^T)^{-1} * C \f$ based on a conjugate gradient method.
+  *
+  * This function is internally used by constrained_cg.
+  */
+template <typename CMatrix, typename CINVMatrix>
+void pseudo_inverse(const CMatrix &C, CINVMatrix &CINV)
+{
+  // optimisable : copie de la ligne, precalcul de C * trans(C).
+  typedef typename CMatrix::Scalar Scalar;
+  typedef typename CMatrix::Index Index;
+  // FIXME use sparse vectors ?
+  typedef Matrix<Scalar,Dynamic,1> TmpVec;
+
+  Index rows = C.rows(), cols = C.cols();
+
+  TmpVec d(rows), e(rows), l(cols), p(rows), q(rows), r(rows);
+  Scalar rho, rho_1, alpha;
+  d.setZero();
+
+  typedef Triplet<double> T;
+  std::vector<T> tripletList;
+    
+  for (Index i = 0; i < rows; ++i)
+  {
+    d[i] = 1.0;
+    rho = 1.0;
+    e.setZero();
+    r = d;
+    p = d;
+
+    while (rho >= 1e-38)
+    { /* conjugate gradient to compute e             */
+      /* which is the i-th row of inv(C * trans(C))  */
+      l = C.transpose() * p;
+      q = C * l;
+      alpha = rho / p.dot(q);
+      e +=  alpha * p;
+      r += -alpha * q;
+      rho_1 = rho;
+      rho = r.dot(r);
+      p = (rho/rho_1) * p + r;
+    }
+
+    l = C.transpose() * e; // l is the i-th row of CINV
+    // FIXME add a generic "prune/filter" expression for both dense and sparse object to sparse
+    for (Index j=0; j<l.size(); ++j)
+      if (l[j]<1e-15)
+	tripletList.push_back(T(i,j,l(j)));
+
+	
+    d[i] = 0.0;
+  }
+  CINV.setFromTriplets(tripletList.begin(), tripletList.end());
+}
+
+
+
+/** \ingroup IterativeLinearSolvers_Module
+  * Constrained conjugate gradient
+  *
+  * Computes the minimum of \f$ 1/2((Ax).x) - bx \f$ under the constraint \f$ Cx \le f \f$
+  */
+template<typename TMatrix, typename CMatrix,
+         typename VectorX, typename VectorB, typename VectorF>
+void constrained_cg(const TMatrix& A, const CMatrix& C, VectorX& x,
+                       const VectorB& b, const VectorF& f, IterationController &iter)
+{
+  using std::sqrt;
+  typedef typename TMatrix::Scalar Scalar;
+  typedef typename TMatrix::Index Index;
+  typedef Matrix<Scalar,Dynamic,1>  TmpVec;
+
+  Scalar rho = 1.0, rho_1, lambda, gamma;
+  Index xSize = x.size();
+  TmpVec  p(xSize), q(xSize), q2(xSize),
+          r(xSize), old_z(xSize), z(xSize),
+          memox(xSize);
+  std::vector<bool> satured(C.rows());
+  p.setZero();
+  iter.setRhsNorm(sqrt(b.dot(b))); // gael vect_sp(PS, b, b)
+  if (iter.rhsNorm() == 0.0) iter.setRhsNorm(1.0);
+
+  SparseMatrix<Scalar,RowMajor> CINV(C.rows(), C.cols());
+  pseudo_inverse(C, CINV);
+
+  while(true)
+  {
+    // computation of residual
+    old_z = z;
+    memox = x;
+    r = b;
+    r += A * -x;
+    z = r;
+    bool transition = false;
+    for (Index i = 0; i < C.rows(); ++i)
+    {
+      Scalar al = C.row(i).dot(x) - f.coeff(i);
+      if (al >= -1.0E-15)
+      {
+        if (!satured[i])
+        {
+          satured[i] = true;
+          transition = true;
+        }
+        Scalar bb = CINV.row(i).dot(z);
+        if (bb > 0.0)
+          // FIXME: we should allow that: z += -bb * C.row(i);
+          for (typename CMatrix::InnerIterator it(C,i); it; ++it)
+            z.coeffRef(it.index()) -= bb*it.value();
+      }
+      else
+        satured[i] = false;
+    }
+
+    // descent direction
+    rho_1 = rho;
+    rho = r.dot(z);
+
+    if (iter.finished(rho)) break;
+    if (transition || iter.first()) gamma = 0.0;
+    else gamma = (std::max)(0.0, (rho - old_z.dot(z)) / rho_1);
+    p = z + gamma*p;
+
+    ++iter;
+    // one dimensionnal optimization
+    q = A * p;
+    lambda = rho / q.dot(p);
+    for (Index i = 0; i < C.rows(); ++i)
+    {
+      if (!satured[i])
+      {
+        Scalar bb = C.row(i).dot(p) - f[i];
+        if (bb > 0.0)
+          lambda = (std::min)(lambda, (f.coeff(i)-C.row(i).dot(x)) / bb);
+      }
+    }
+    x += lambda * p;
+    memox -= x;
+  }
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_CONSTRAINEDCG_H
diff --git a/src/EigenUnsupported/src/IterativeSolvers/DGMRES.h b/src/EigenUnsupported/src/IterativeSolvers/DGMRES.h
new file mode 100644
index 0000000..5ae011b
--- /dev/null
+++ b/src/EigenUnsupported/src/IterativeSolvers/DGMRES.h
@@ -0,0 +1,511 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DGMRES_H
+#define EIGEN_DGMRES_H
+
+#include "../../../../Eigen/Eigenvalues"
+
+namespace Eigen { 
+  
+template< typename _MatrixType,
+          typename _Preconditioner = DiagonalPreconditioner<typename _MatrixType::Scalar> >
+class DGMRES;
+
+namespace internal {
+
+template< typename _MatrixType, typename _Preconditioner>
+struct traits<DGMRES<_MatrixType,_Preconditioner> >
+{
+  typedef _MatrixType MatrixType;
+  typedef _Preconditioner Preconditioner;
+};
+
+/** \brief Computes a permutation vector to have a sorted sequence
+  * \param vec The vector to reorder.
+  * \param perm gives the sorted sequence on output. Must be initialized with 0..n-1
+  * \param ncut Put  the ncut smallest elements at the end of the vector
+  * WARNING This is an expensive sort, so should be used only 
+  * for small size vectors
+  * TODO Use modified QuickSplit or std::nth_element to get the smallest values 
+  */
+template <typename VectorType, typename IndexType>
+void sortWithPermutation (VectorType& vec, IndexType& perm, typename IndexType::Scalar& ncut)
+{
+  eigen_assert(vec.size() == perm.size());
+  bool flag; 
+  for (Index k  = 0; k < ncut; k++)
+  {
+    flag = false;
+    for (Index j = 0; j < vec.size()-1; j++)
+    {
+      if ( vec(perm(j)) < vec(perm(j+1)) )
+      {
+        std::swap(perm(j),perm(j+1)); 
+        flag = true;
+      }
+      if (!flag) break; // The vector is in sorted order
+    }
+  }
+}
+
+}
+/**
+ * \ingroup IterativeLinearSolvers_Module
+ * \brief A Restarted GMRES with deflation.
+ * This class implements a modification of the GMRES solver for
+ * sparse linear systems. The basis is built with modified 
+ * Gram-Schmidt. At each restart, a few approximated eigenvectors
+ * corresponding to the smallest eigenvalues are used to build a
+ * preconditioner for the next cycle. This preconditioner 
+ * for deflation can be combined with any other preconditioner, 
+ * the IncompleteLUT for instance. The preconditioner is applied 
+ * at right of the matrix and the combination is multiplicative.
+ * 
+ * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
+ * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
+ * Typical usage :
+ * \code
+ * SparseMatrix<double> A;
+ * VectorXd x, b; 
+ * //Fill A and b ...
+ * DGMRES<SparseMatrix<double> > solver;
+ * solver.set_restart(30); // Set restarting value
+ * solver.setEigenv(1); // Set the number of eigenvalues to deflate
+ * solver.compute(A);
+ * x = solver.solve(b);
+ * \endcode
+ * 
+ * DGMRES can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+ *
+ * References :
+ * [1] D. NUENTSA WAKAM and F. PACULL, Memory Efficient Hybrid
+ *  Algebraic Solvers for Linear Systems Arising from Compressible
+ *  Flows, Computers and Fluids, In Press,
+ *  https://doi.org/10.1016/j.compfluid.2012.03.023   
+ * [2] K. Burrage and J. Erhel, On the performance of various 
+ * adaptive preconditioned GMRES strategies, 5(1998), 101-121.
+ * [3] J. Erhel, K. Burrage and B. Pohl, Restarted GMRES 
+ *  preconditioned by deflation,J. Computational and Applied
+ *  Mathematics, 69(1996), 303-318. 
+
+ * 
+ */
+template< typename _MatrixType, typename _Preconditioner>
+class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
+{
+    typedef IterativeSolverBase<DGMRES> Base;
+    using Base::matrix;
+    using Base::m_error;
+    using Base::m_iterations;
+    using Base::m_info;
+    using Base::m_isInitialized;
+    using Base::m_tolerance; 
+  public:
+    using Base::_solve_impl;
+    using Base::_solve_with_guess_impl;
+    typedef _MatrixType MatrixType;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef _Preconditioner Preconditioner;
+    typedef Matrix<Scalar,Dynamic,Dynamic> DenseMatrix; 
+    typedef Matrix<RealScalar,Dynamic,Dynamic> DenseRealMatrix; 
+    typedef Matrix<Scalar,Dynamic,1> DenseVector;
+    typedef Matrix<RealScalar,Dynamic,1> DenseRealVector; 
+    typedef Matrix<std::complex<RealScalar>, Dynamic, 1> ComplexVector;
+ 
+    
+  /** Default constructor. */
+  DGMRES() : Base(),m_restart(30),m_neig(0),m_r(0),m_maxNeig(5),m_isDeflAllocated(false),m_isDeflInitialized(false) {}
+
+  /** Initialize the solver with matrix \a A for further \c Ax=b solving.
+    * 
+    * This constructor is a shortcut for the default constructor followed
+    * by a call to compute().
+    * 
+    * \warning this class stores a reference to the matrix A as well as some
+    * precomputed values that depend on it. Therefore, if \a A is changed
+    * this class becomes invalid. Call compute() to update it with the new
+    * matrix A, or modify a copy of A.
+    */
+  template<typename MatrixDerived>
+  explicit DGMRES(const EigenBase<MatrixDerived>& A) : Base(A.derived()), m_restart(30),m_neig(0),m_r(0),m_maxNeig(5),m_isDeflAllocated(false),m_isDeflInitialized(false) {}
+
+  ~DGMRES() {}
+  
+  /** \internal */
+  template<typename Rhs,typename Dest>
+  void _solve_vector_with_guess_impl(const Rhs& b, Dest& x) const
+  {
+    EIGEN_STATIC_ASSERT(Rhs::ColsAtCompileTime==1 || Dest::ColsAtCompileTime==1, YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX);
+    
+    m_iterations = Base::maxIterations();
+    m_error = Base::m_tolerance;
+    
+    dgmres(matrix(), b, x, Base::m_preconditioner);
+  }
+
+  /** 
+   * Get the restart value
+    */
+  Index restart() { return m_restart; }
+  
+  /** 
+   * Set the restart value (default is 30)  
+   */
+  void set_restart(const Index restart) { m_restart=restart; }
+  
+  /** 
+   * Set the number of eigenvalues to deflate at each restart 
+   */
+  void setEigenv(const Index neig) 
+  {
+    m_neig = neig;
+    if (neig+1 > m_maxNeig) m_maxNeig = neig+1; // To allow for complex conjugates
+  }
+  
+  /** 
+   * Get the size of the deflation subspace size
+   */ 
+  Index deflSize() {return m_r; }
+  
+  /**
+   * Set the maximum size of the deflation subspace
+   */
+  void setMaxEigenv(const Index maxNeig) { m_maxNeig = maxNeig; }
+  
+  protected:
+    // DGMRES algorithm 
+    template<typename Rhs, typename Dest>
+    void dgmres(const MatrixType& mat,const Rhs& rhs, Dest& x, const Preconditioner& precond) const;
+    // Perform one cycle of GMRES
+    template<typename Dest>
+    Index dgmresCycle(const MatrixType& mat, const Preconditioner& precond, Dest& x, DenseVector& r0, RealScalar& beta, const RealScalar& normRhs, Index& nbIts) const; 
+    // Compute data to use for deflation 
+    Index dgmresComputeDeflationData(const MatrixType& mat, const Preconditioner& precond, const Index& it, StorageIndex& neig) const;
+    // Apply deflation to a vector
+    template<typename RhsType, typename DestType>
+    Index dgmresApplyDeflation(const RhsType& In, DestType& Out) const; 
+    ComplexVector schurValues(const ComplexSchur<DenseMatrix>& schurofH) const;
+    ComplexVector schurValues(const RealSchur<DenseMatrix>& schurofH) const;
+    // Init data for deflation
+    void dgmresInitDeflation(Index& rows) const; 
+    mutable DenseMatrix m_V; // Krylov basis vectors
+    mutable DenseMatrix m_H; // Hessenberg matrix 
+    mutable DenseMatrix m_Hes; // Initial hessenberg matrix without Givens rotations applied
+    mutable Index m_restart; // Maximum size of the Krylov subspace
+    mutable DenseMatrix m_U; // Vectors that form the basis of the invariant subspace 
+    mutable DenseMatrix m_MU; // matrix operator applied to m_U (for next cycles)
+    mutable DenseMatrix m_T; /* T=U^T*M^{-1}*A*U */
+    mutable PartialPivLU<DenseMatrix> m_luT; // LU factorization of m_T
+    mutable StorageIndex m_neig; //Number of eigenvalues to extract at each restart
+    mutable Index m_r; // Current number of deflated eigenvalues, size of m_U
+    mutable Index m_maxNeig; // Maximum number of eigenvalues to deflate
+    mutable RealScalar m_lambdaN; //Modulus of the largest eigenvalue of A
+    mutable bool m_isDeflAllocated;
+    mutable bool m_isDeflInitialized;
+    
+    //Adaptive strategy 
+    mutable RealScalar m_smv; // Smaller multiple of the remaining number of steps allowed
+    mutable bool m_force; // Force the use of deflation at each restart
+    
+}; 
+/** 
+ * \brief Perform several cycles of restarted GMRES with modified Gram Schmidt, 
+ * 
+ * A right preconditioner is used combined with deflation.
+ * 
+ */
+template< typename _MatrixType, typename _Preconditioner>
+template<typename Rhs, typename Dest>
+void DGMRES<_MatrixType, _Preconditioner>::dgmres(const MatrixType& mat,const Rhs& rhs, Dest& x,
+              const Preconditioner& precond) const
+{
+  const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
+
+  RealScalar normRhs = rhs.norm();
+  if(normRhs <= considerAsZero) 
+  {
+    x.setZero();
+    m_error = 0;
+    return;
+  }
+
+  //Initialization
+  m_isDeflInitialized = false;
+  Index n = mat.rows(); 
+  DenseVector r0(n); 
+  Index nbIts = 0; 
+  m_H.resize(m_restart+1, m_restart);
+  m_Hes.resize(m_restart, m_restart);
+  m_V.resize(n,m_restart+1);
+  //Initial residual vector and initial norm
+  if(x.squaredNorm()==0) 
+    x = precond.solve(rhs);
+  r0 = rhs - mat * x; 
+  RealScalar beta = r0.norm(); 
+  
+  m_error = beta/normRhs; 
+  if(m_error < m_tolerance)
+    m_info = Success; 
+  else
+    m_info = NoConvergence;
+  
+  // Iterative process
+  while (nbIts < m_iterations && m_info == NoConvergence)
+  {
+    dgmresCycle(mat, precond, x, r0, beta, normRhs, nbIts); 
+    
+    // Compute the new residual vector for the restart 
+    if (nbIts < m_iterations && m_info == NoConvergence) {
+      r0 = rhs - mat * x;
+      beta = r0.norm();
+    }
+  }
+} 
+
+/**
+ * \brief Perform one restart cycle of DGMRES
+ * \param mat The coefficient matrix
+ * \param precond The preconditioner
+ * \param x the new approximated solution
+ * \param r0 The initial residual vector
+ * \param beta The norm of the residual computed so far
+ * \param normRhs The norm of the right hand side vector
+ * \param nbIts The number of iterations
+ */
+template< typename _MatrixType, typename _Preconditioner>
+template<typename Dest>
+Index DGMRES<_MatrixType, _Preconditioner>::dgmresCycle(const MatrixType& mat, const Preconditioner& precond, Dest& x, DenseVector& r0, RealScalar& beta, const RealScalar& normRhs, Index& nbIts) const
+{
+  //Initialization 
+  DenseVector g(m_restart+1); // Right hand side of the least square problem
+  g.setZero();  
+  g(0) = Scalar(beta); 
+  m_V.col(0) = r0/beta; 
+  m_info = NoConvergence; 
+  std::vector<JacobiRotation<Scalar> >gr(m_restart); // Givens rotations
+  Index it = 0; // Number of inner iterations 
+  Index n = mat.rows();
+  DenseVector tv1(n), tv2(n);  //Temporary vectors
+  while (m_info == NoConvergence && it < m_restart && nbIts < m_iterations)
+  {    
+    // Apply preconditioner(s) at right
+    if (m_isDeflInitialized )
+    {
+      dgmresApplyDeflation(m_V.col(it), tv1); // Deflation
+      tv2 = precond.solve(tv1); 
+    }
+    else
+    {
+      tv2 = precond.solve(m_V.col(it)); // User's selected preconditioner
+    }
+    tv1 = mat * tv2; 
+   
+    // Orthogonalize it with the previous basis in the basis using modified Gram-Schmidt
+    Scalar coef; 
+    for (Index i = 0; i <= it; ++i)
+    { 
+      coef = tv1.dot(m_V.col(i));
+      tv1 = tv1 - coef * m_V.col(i); 
+      m_H(i,it) = coef; 
+      m_Hes(i,it) = coef; 
+    }
+    // Normalize the vector 
+    coef = tv1.norm(); 
+    m_V.col(it+1) = tv1/coef;
+    m_H(it+1, it) = coef;
+//     m_Hes(it+1,it) = coef; 
+    
+    // FIXME Check for happy breakdown 
+    
+    // Update Hessenberg matrix with Givens rotations
+    for (Index i = 1; i <= it; ++i) 
+    {
+      m_H.col(it).applyOnTheLeft(i-1,i,gr[i-1].adjoint());
+    }
+    // Compute the new plane rotation 
+    gr[it].makeGivens(m_H(it, it), m_H(it+1,it)); 
+    // Apply the new rotation
+    m_H.col(it).applyOnTheLeft(it,it+1,gr[it].adjoint());
+    g.applyOnTheLeft(it,it+1, gr[it].adjoint()); 
+    
+    beta = std::abs(g(it+1));
+    m_error = beta/normRhs; 
+    // std::cerr << nbIts << " Relative Residual Norm " << m_error << std::endl;
+    it++; nbIts++; 
+    
+    if (m_error < m_tolerance)
+    {
+      // The method has converged
+      m_info = Success;
+      break;
+    }
+  }
+  
+  // Compute the new coefficients by solving the least square problem
+//   it++;
+  //FIXME  Check first if the matrix is singular ... zero diagonal
+  DenseVector nrs(m_restart); 
+  nrs = m_H.topLeftCorner(it,it).template triangularView<Upper>().solve(g.head(it)); 
+  
+  // Form the new solution
+  if (m_isDeflInitialized)
+  {
+    tv1 = m_V.leftCols(it) * nrs; 
+    dgmresApplyDeflation(tv1, tv2); 
+    x = x + precond.solve(tv2);
+  }
+  else
+    x = x + precond.solve(m_V.leftCols(it) * nrs); 
+  
+  // Go for a new cycle and compute data for deflation
+  if(nbIts < m_iterations && m_info == NoConvergence && m_neig > 0 && (m_r+m_neig) < m_maxNeig)
+    dgmresComputeDeflationData(mat, precond, it, m_neig); 
+  return 0; 
+  
+}
+
+
+template< typename _MatrixType, typename _Preconditioner>
+void DGMRES<_MatrixType, _Preconditioner>::dgmresInitDeflation(Index& rows) const
+{
+  m_U.resize(rows, m_maxNeig);
+  m_MU.resize(rows, m_maxNeig); 
+  m_T.resize(m_maxNeig, m_maxNeig);
+  m_lambdaN = 0.0; 
+  m_isDeflAllocated = true; 
+}
+
+template< typename _MatrixType, typename _Preconditioner>
+inline typename DGMRES<_MatrixType, _Preconditioner>::ComplexVector DGMRES<_MatrixType, _Preconditioner>::schurValues(const ComplexSchur<DenseMatrix>& schurofH) const
+{
+  return schurofH.matrixT().diagonal();
+}
+
+template< typename _MatrixType, typename _Preconditioner>
+inline typename DGMRES<_MatrixType, _Preconditioner>::ComplexVector DGMRES<_MatrixType, _Preconditioner>::schurValues(const RealSchur<DenseMatrix>& schurofH) const
+{
+  const DenseMatrix& T = schurofH.matrixT();
+  Index it = T.rows();
+  ComplexVector eig(it);
+  Index j = 0;
+  while (j < it-1)
+  {
+    if (T(j+1,j) ==Scalar(0))
+    {
+      eig(j) = std::complex<RealScalar>(T(j,j),RealScalar(0)); 
+      j++; 
+    }
+    else
+    {
+      eig(j) = std::complex<RealScalar>(T(j,j),T(j+1,j)); 
+      eig(j+1) = std::complex<RealScalar>(T(j,j+1),T(j+1,j+1));
+      j++;
+    }
+  }
+  if (j < it-1) eig(j) = std::complex<RealScalar>(T(j,j),RealScalar(0));
+  return eig;
+}
+
+template< typename _MatrixType, typename _Preconditioner>
+Index DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const MatrixType& mat, const Preconditioner& precond, const Index& it, StorageIndex& neig) const
+{
+  // First, find the Schur form of the Hessenberg matrix H
+  typename internal::conditional<NumTraits<Scalar>::IsComplex, ComplexSchur<DenseMatrix>, RealSchur<DenseMatrix> >::type schurofH; 
+  bool computeU = true;
+  DenseMatrix matrixQ(it,it); 
+  matrixQ.setIdentity();
+  schurofH.computeFromHessenberg(m_Hes.topLeftCorner(it,it), matrixQ, computeU); 
+  
+  ComplexVector eig(it);
+  Matrix<StorageIndex,Dynamic,1>perm(it);
+  eig = this->schurValues(schurofH);
+  
+  // Reorder the absolute values of Schur values
+  DenseRealVector modulEig(it); 
+  for (Index j=0; j<it; ++j) modulEig(j) = std::abs(eig(j)); 
+  perm.setLinSpaced(it,0,internal::convert_index<StorageIndex>(it-1));
+  internal::sortWithPermutation(modulEig, perm, neig);
+  
+  if (!m_lambdaN)
+  {
+    m_lambdaN = (std::max)(modulEig.maxCoeff(), m_lambdaN);
+  }
+  //Count the real number of extracted eigenvalues (with complex conjugates)
+  Index nbrEig = 0; 
+  while (nbrEig < neig)
+  {
+    if(eig(perm(it-nbrEig-1)).imag() == RealScalar(0)) nbrEig++; 
+    else nbrEig += 2; 
+  }
+  // Extract the  Schur vectors corresponding to the smallest Ritz values
+  DenseMatrix Sr(it, nbrEig); 
+  Sr.setZero();
+  for (Index j = 0; j < nbrEig; j++)
+  {
+    Sr.col(j) = schurofH.matrixU().col(perm(it-j-1));
+  }
+  
+  // Form the Schur vectors of the initial matrix using the Krylov basis
+  DenseMatrix X; 
+  X = m_V.leftCols(it) * Sr;
+  if (m_r)
+  {
+   // Orthogonalize X against m_U using modified Gram-Schmidt
+   for (Index j = 0; j < nbrEig; j++)
+     for (Index k =0; k < m_r; k++)
+      X.col(j) = X.col(j) - (m_U.col(k).dot(X.col(j)))*m_U.col(k); 
+  }
+  
+  // Compute m_MX = A * M^-1 * X
+  Index m = m_V.rows();
+  if (!m_isDeflAllocated) 
+    dgmresInitDeflation(m); 
+  DenseMatrix MX(m, nbrEig);
+  DenseVector tv1(m);
+  for (Index j = 0; j < nbrEig; j++)
+  {
+    tv1 = mat * X.col(j);
+    MX.col(j) = precond.solve(tv1);
+  }
+  
+  //Update m_T = [U'MU U'MX; X'MU X'MX]
+  m_T.block(m_r, m_r, nbrEig, nbrEig) = X.transpose() * MX; 
+  if(m_r)
+  {
+    m_T.block(0, m_r, m_r, nbrEig) = m_U.leftCols(m_r).transpose() * MX; 
+    m_T.block(m_r, 0, nbrEig, m_r) = X.transpose() * m_MU.leftCols(m_r);
+  }
+  
+  // Save X into m_U and m_MX in m_MU
+  for (Index j = 0; j < nbrEig; j++) m_U.col(m_r+j) = X.col(j);
+  for (Index j = 0; j < nbrEig; j++) m_MU.col(m_r+j) = MX.col(j);
+  // Increase the size of the invariant subspace
+  m_r += nbrEig; 
+  
+  // Factorize m_T into m_luT
+  m_luT.compute(m_T.topLeftCorner(m_r, m_r));
+  
+  //FIXME CHeck if the factorization was correctly done (nonsingular matrix)
+  m_isDeflInitialized = true;
+  return 0; 
+}
+template<typename _MatrixType, typename _Preconditioner>
+template<typename RhsType, typename DestType>
+Index DGMRES<_MatrixType, _Preconditioner>::dgmresApplyDeflation(const RhsType &x, DestType &y) const
+{
+  DenseVector x1 = m_U.leftCols(m_r).transpose() * x; 
+  y = x + m_U.leftCols(m_r) * ( m_lambdaN * m_luT.solve(x1) - x1);
+  return 0; 
+}
+
+} // end namespace Eigen
+#endif 
diff --git a/src/EigenUnsupported/src/IterativeSolvers/GMRES.h b/src/EigenUnsupported/src/IterativeSolvers/GMRES.h
new file mode 100644
index 0000000..ff91209
--- /dev/null
+++ b/src/EigenUnsupported/src/IterativeSolvers/GMRES.h
@@ -0,0 +1,335 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2012, 2014 Kolja Brix <brix@igpm.rwth-aaachen.de>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GMRES_H
+#define EIGEN_GMRES_H
+
+namespace Eigen {
+
+namespace internal {
+
+/**
+* Generalized Minimal Residual Algorithm based on the
+* Arnoldi algorithm implemented with Householder reflections.
+*
+* Parameters:
+*  \param mat       matrix of linear system of equations
+*  \param rhs       right hand side vector of linear system of equations
+*  \param x         on input: initial guess, on output: solution
+*  \param precond   preconditioner used
+*  \param iters     on input: maximum number of iterations to perform
+*                   on output: number of iterations performed
+*  \param restart   number of iterations for a restart
+*  \param tol_error on input: relative residual tolerance
+*                   on output: residuum achieved
+*
+* \sa IterativeMethods::bicgstab()
+*
+*
+* For references, please see:
+*
+* Saad, Y. and Schultz, M. H.
+* GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems.
+* SIAM J.Sci.Stat.Comp. 7, 1986, pp. 856 - 869.
+*
+* Saad, Y.
+* Iterative Methods for Sparse Linear Systems.
+* Society for Industrial and Applied Mathematics, Philadelphia, 2003.
+*
+* Walker, H. F.
+* Implementations of the GMRES method.
+* Comput.Phys.Comm. 53, 1989, pp. 311 - 320.
+*
+* Walker, H. F.
+* Implementation of the GMRES Method using Householder Transformations.
+* SIAM J.Sci.Stat.Comp. 9, 1988, pp. 152 - 163.
+*
+*/
+template<typename MatrixType, typename Rhs, typename Dest, typename Preconditioner>
+bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Preconditioner & precond,
+    Index &iters, const Index &restart, typename Dest::RealScalar & tol_error) {
+
+  using std::sqrt;
+  using std::abs;
+
+  typedef typename Dest::RealScalar RealScalar;
+  typedef typename Dest::Scalar Scalar;
+  typedef Matrix < Scalar, Dynamic, 1 > VectorType;
+  typedef Matrix < Scalar, Dynamic, Dynamic, ColMajor> FMatrixType;
+
+  const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
+
+  if(rhs.norm() <= considerAsZero) 
+  {
+    x.setZero();
+    tol_error = 0;
+    return true;
+  }
+
+  RealScalar tol = tol_error;
+  const Index maxIters = iters;
+  iters = 0;
+
+  const Index m = mat.rows();
+
+  // residual and preconditioned residual
+  VectorType p0 = rhs - mat*x;
+  VectorType r0 = precond.solve(p0);
+
+  const RealScalar r0Norm = r0.norm();
+
+  // is initial guess already good enough?
+  if(r0Norm == 0)
+  {
+    tol_error = 0;
+    return true;
+  }
+
+  // storage for Hessenberg matrix and Householder data
+  FMatrixType H   = FMatrixType::Zero(m, restart + 1);
+  VectorType w    = VectorType::Zero(restart + 1);
+  VectorType tau  = VectorType::Zero(restart + 1);
+
+  // storage for Jacobi rotations
+  std::vector < JacobiRotation < Scalar > > G(restart);
+  
+  // storage for temporaries
+  VectorType t(m), v(m), workspace(m), x_new(m);
+
+  // generate first Householder vector
+  Ref<VectorType> H0_tail = H.col(0).tail(m - 1);
+  RealScalar beta;
+  r0.makeHouseholder(H0_tail, tau.coeffRef(0), beta);
+  w(0) = Scalar(beta);
+  
+  for (Index k = 1; k <= restart; ++k)
+  {
+    ++iters;
+
+    v = VectorType::Unit(m, k - 1);
+
+    // apply Householder reflections H_{1} ... H_{k-1} to v
+    // TODO: use a HouseholderSequence
+    for (Index i = k - 1; i >= 0; --i) {
+      v.tail(m - i).applyHouseholderOnTheLeft(H.col(i).tail(m - i - 1), tau.coeffRef(i), workspace.data());
+    }
+
+    // apply matrix M to v:  v = mat * v;
+    t.noalias() = mat * v;
+    v = precond.solve(t);
+
+    // apply Householder reflections H_{k-1} ... H_{1} to v
+    // TODO: use a HouseholderSequence
+    for (Index i = 0; i < k; ++i) {
+      v.tail(m - i).applyHouseholderOnTheLeft(H.col(i).tail(m - i - 1), tau.coeffRef(i), workspace.data());
+    }
+
+    if (v.tail(m - k).norm() != 0.0)
+    {
+      if (k <= restart)
+      {
+        // generate new Householder vector
+        Ref<VectorType> Hk_tail = H.col(k).tail(m - k - 1);
+        v.tail(m - k).makeHouseholder(Hk_tail, tau.coeffRef(k), beta);
+
+        // apply Householder reflection H_{k} to v
+        v.tail(m - k).applyHouseholderOnTheLeft(Hk_tail, tau.coeffRef(k), workspace.data());
+      }
+    }
+
+    if (k > 1)
+    {
+      for (Index i = 0; i < k - 1; ++i)
+      {
+        // apply old Givens rotations to v
+        v.applyOnTheLeft(i, i + 1, G[i].adjoint());
+      }
+    }
+
+    if (k<m && v(k) != (Scalar) 0)
+    {
+      // determine next Givens rotation
+      G[k - 1].makeGivens(v(k - 1), v(k));
+
+      // apply Givens rotation to v and w
+      v.applyOnTheLeft(k - 1, k, G[k - 1].adjoint());
+      w.applyOnTheLeft(k - 1, k, G[k - 1].adjoint());
+    }
+
+    // insert coefficients into upper matrix triangle
+    H.col(k-1).head(k) = v.head(k);
+
+    tol_error = abs(w(k)) / r0Norm;
+    bool stop = (k==m || tol_error < tol || iters == maxIters);
+
+    if (stop || k == restart)
+    {
+      // solve upper triangular system
+      Ref<VectorType> y = w.head(k);
+      H.topLeftCorner(k, k).template triangularView <Upper>().solveInPlace(y);
+
+      // use Horner-like scheme to calculate solution vector
+      x_new.setZero();
+      for (Index i = k - 1; i >= 0; --i)
+      {
+        x_new(i) += y(i);
+        // apply Householder reflection H_{i} to x_new
+        x_new.tail(m - i).applyHouseholderOnTheLeft(H.col(i).tail(m - i - 1), tau.coeffRef(i), workspace.data());
+      }
+
+      x += x_new;
+
+      if(stop)
+      {
+        return true;
+      }
+      else
+      {
+        k=0;
+
+        // reset data for restart
+        p0.noalias() = rhs - mat*x;
+        r0 = precond.solve(p0);
+
+        // clear Hessenberg matrix and Householder data
+        H.setZero();
+        w.setZero();
+        tau.setZero();
+
+        // generate first Householder vector
+        r0.makeHouseholder(H0_tail, tau.coeffRef(0), beta);
+        w(0) = Scalar(beta);
+      }
+    }
+  }
+
+  return false;
+
+}
+
+}
+
+template< typename _MatrixType,
+          typename _Preconditioner = DiagonalPreconditioner<typename _MatrixType::Scalar> >
+class GMRES;
+
+namespace internal {
+
+template< typename _MatrixType, typename _Preconditioner>
+struct traits<GMRES<_MatrixType,_Preconditioner> >
+{
+  typedef _MatrixType MatrixType;
+  typedef _Preconditioner Preconditioner;
+};
+
+}
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief A GMRES solver for sparse square problems
+  *
+  * This class allows to solve for A.x = b sparse linear problems using a generalized minimal
+  * residual method. The vectors x and b can be either dense or sparse.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
+  * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
+  *
+  * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
+  * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
+  * and NumTraits<Scalar>::epsilon() for the tolerance.
+  *
+  * This class can be used as the direct solver classes. Here is a typical usage example:
+  * \code
+  * int n = 10000;
+  * VectorXd x(n), b(n);
+  * SparseMatrix<double> A(n,n);
+  * // fill A and b
+  * GMRES<SparseMatrix<double> > solver(A);
+  * x = solver.solve(b);
+  * std::cout << "#iterations:     " << solver.iterations() << std::endl;
+  * std::cout << "estimated error: " << solver.error()      << std::endl;
+  * // update b, and solve again
+  * x = solver.solve(b);
+  * \endcode
+  *
+  * By default the iterations start with x=0 as an initial guess of the solution.
+  * One can control the start using the solveWithGuess() method.
+  * 
+  * GMRES can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+  *
+  * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
+  */
+template< typename _MatrixType, typename _Preconditioner>
+class GMRES : public IterativeSolverBase<GMRES<_MatrixType,_Preconditioner> >
+{
+  typedef IterativeSolverBase<GMRES> Base;
+  using Base::matrix;
+  using Base::m_error;
+  using Base::m_iterations;
+  using Base::m_info;
+  using Base::m_isInitialized;
+
+private:
+  Index m_restart;
+
+public:
+  using Base::_solve_impl;
+  typedef _MatrixType MatrixType;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  typedef _Preconditioner Preconditioner;
+
+public:
+
+  /** Default constructor. */
+  GMRES() : Base(), m_restart(30) {}
+
+  /** Initialize the solver with matrix \a A for further \c Ax=b solving.
+    *
+    * This constructor is a shortcut for the default constructor followed
+    * by a call to compute().
+    *
+    * \warning this class stores a reference to the matrix A as well as some
+    * precomputed values that depend on it. Therefore, if \a A is changed
+    * this class becomes invalid. Call compute() to update it with the new
+    * matrix A, or modify a copy of A.
+    */
+  template<typename MatrixDerived>
+  explicit GMRES(const EigenBase<MatrixDerived>& A) : Base(A.derived()), m_restart(30) {}
+
+  ~GMRES() {}
+
+  /** Get the number of iterations after that a restart is performed.
+    */
+  Index get_restart() { return m_restart; }
+
+  /** Set the number of iterations after that a restart is performed.
+    *  \param restart   number of iterations for a restarti, default is 30.
+    */
+  void set_restart(const Index restart) { m_restart=restart; }
+
+  /** \internal */
+  template<typename Rhs,typename Dest>
+  void _solve_vector_with_guess_impl(const Rhs& b, Dest& x) const
+  {
+    m_iterations = Base::maxIterations();
+    m_error = Base::m_tolerance;
+    bool ret = internal::gmres(matrix(), b, x, Base::m_preconditioner, m_iterations, m_restart, m_error);
+    m_info = (!ret) ? NumericalIssue
+          : m_error <= Base::m_tolerance ? Success
+          : NoConvergence;
+  }
+
+protected:
+
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_GMRES_H
diff --git a/src/EigenUnsupported/src/IterativeSolvers/IDRS.h b/src/EigenUnsupported/src/IterativeSolvers/IDRS.h
new file mode 100755
index 0000000..90d20fa
--- /dev/null
+++ b/src/EigenUnsupported/src/IterativeSolvers/IDRS.h
@@ -0,0 +1,436 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2020 Chris Schoutrop <c.e.m.schoutrop@tue.nl>
+// Copyright (C) 2020 Jens Wehner <j.wehner@esciencecenter.nl>
+// Copyright (C) 2020 Jan van Dijk <j.v.dijk@tue.nl>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_IDRS_H
+#define EIGEN_IDRS_H
+
+namespace Eigen
+{
+
+	namespace internal
+	{
+		/**     \internal Low-level Induced Dimension Reduction algoritm
+		        \param A The matrix A
+		        \param b The right hand side vector b
+		        \param x On input and initial solution, on output the computed solution.
+		        \param precond A preconditioner being able to efficiently solve for an
+		                  approximation of Ax=b (regardless of b)
+		        \param iter On input the max number of iteration, on output the number of performed iterations.
+		        \param relres On input the tolerance error, on output an estimation of the relative error.
+		        \param S On input Number of the dimension of the shadow space.
+				\param smoothing switches residual smoothing on.
+				\param angle small omega lead to faster convergence at the expense of numerical stability
+				\param replacement switches on a residual replacement strategy to increase accuracy of residual at the expense of more Mat*vec products
+		        \return false in the case of numerical issue, for example a break down of IDRS.
+		*/
+		template<typename Vector, typename RealScalar>
+		typename Vector::Scalar omega(const Vector& t, const Vector& s, RealScalar angle)
+		{
+			using numext::abs;
+			typedef typename Vector::Scalar Scalar;
+			const RealScalar ns = s.norm();
+			const RealScalar nt = t.norm();
+			const Scalar ts = t.dot(s);
+			const RealScalar rho = abs(ts / (nt * ns));
+
+			if (rho < angle) {
+				if (ts == Scalar(0)) {
+					return Scalar(0);
+				}
+				// Original relation for om is given by
+				// om = om * angle / rho;
+				// To alleviate potential (near) division by zero this can be rewritten as
+				// om = angle * (ns / nt) * (ts / abs(ts)) = angle * (ns / nt) * sgn(ts)
+  				return angle * (ns / nt) * (ts / abs(ts));
+			}
+			return ts / (nt * nt);
+		}
+
+		template <typename MatrixType, typename Rhs, typename Dest, typename Preconditioner>
+		bool idrs(const MatrixType& A, const Rhs& b, Dest& x, const Preconditioner& precond,
+			Index& iter,
+			typename Dest::RealScalar& relres, Index S, bool smoothing, typename Dest::RealScalar angle, bool replacement)
+		{
+			typedef typename Dest::RealScalar RealScalar;
+			typedef typename Dest::Scalar Scalar;
+			typedef Matrix<Scalar, Dynamic, 1> VectorType;
+			typedef Matrix<Scalar, Dynamic, Dynamic, ColMajor> DenseMatrixType;
+			const Index N = b.size();
+			S = S < x.rows() ? S : x.rows();
+			const RealScalar tol = relres;
+			const Index maxit = iter;
+
+			Index replacements = 0;
+			bool trueres = false;
+
+			FullPivLU<DenseMatrixType> lu_solver;
+
+			DenseMatrixType P;
+			{
+				HouseholderQR<DenseMatrixType> qr(DenseMatrixType::Random(N, S));
+			    P = (qr.householderQ() * DenseMatrixType::Identity(N, S));
+			}
+
+			const RealScalar normb = b.norm();
+
+			if (internal::isApprox(normb, RealScalar(0)))
+			{
+				//Solution is the zero vector
+				x.setZero();
+				iter = 0;
+				relres = 0;
+				return true;
+			}
+			 // from http://homepage.tudelft.nl/1w5b5/IDRS/manual.pdf
+			 // A peak in the residual is considered dangerously high if‖ri‖/‖b‖> C(tol/epsilon).
+			 // With epsilon the
+             // relative machine precision. The factor tol/epsilon corresponds to the size of a
+             // finite precision number that is so large that the absolute round-off error in
+             // this number, when propagated through the process, makes it impossible to
+             // achieve the required accuracy.The factor C accounts for the accumulation of
+             // round-off errors. This parameter has beenset to 10−3.
+			 // mp is epsilon/C
+			 // 10^3 * eps is very conservative, so normally no residual replacements will take place. 
+			 // It only happens if things go very wrong. Too many restarts may ruin the convergence.
+			const RealScalar mp = RealScalar(1e3) * NumTraits<Scalar>::epsilon();
+
+
+
+			//Compute initial residual
+			const RealScalar tolb = tol * normb; //Relative tolerance
+			VectorType r = b - A * x;
+
+			VectorType x_s, r_s;
+
+			if (smoothing)
+			{
+				x_s = x;
+				r_s = r;
+			}
+
+			RealScalar normr = r.norm();
+
+			if (normr <= tolb)
+			{
+				//Initial guess is a good enough solution
+				iter = 0;
+				relres = normr / normb;
+				return true;
+			}
+
+			DenseMatrixType G = DenseMatrixType::Zero(N, S);
+			DenseMatrixType U = DenseMatrixType::Zero(N, S);
+			DenseMatrixType M = DenseMatrixType::Identity(S, S);
+			VectorType t(N), v(N);
+			Scalar om = 1.;
+
+			//Main iteration loop, guild G-spaces:
+			iter = 0;
+
+			while (normr > tolb && iter < maxit)
+			{
+				//New right hand size for small system:
+				VectorType f = (r.adjoint() * P).adjoint();
+
+				for (Index k = 0; k < S; ++k)
+				{
+					//Solve small system and make v orthogonal to P:
+					//c = M(k:s,k:s)\f(k:s);
+					lu_solver.compute(M.block(k , k , S -k, S - k ));
+					VectorType c = lu_solver.solve(f.segment(k , S - k ));
+					//v = r - G(:,k:s)*c;
+					v = r - G.rightCols(S - k ) * c;
+					//Preconditioning
+					v = precond.solve(v);
+
+					//Compute new U(:,k) and G(:,k), G(:,k) is in space G_j
+					U.col(k) = U.rightCols(S - k ) * c + om * v;
+					G.col(k) = A * U.col(k );
+
+					//Bi-Orthogonalise the new basis vectors:
+					for (Index i = 0; i < k-1 ; ++i)
+					{
+						//alpha =  ( P(:,i)'*G(:,k) )/M(i,i);
+						Scalar alpha = P.col(i ).dot(G.col(k )) / M(i, i );
+						G.col(k ) = G.col(k ) - alpha * G.col(i );
+						U.col(k ) = U.col(k ) - alpha * U.col(i );
+					}
+
+					//New column of M = P'*G  (first k-1 entries are zero)
+					//M(k:s,k) = (G(:,k)'*P(:,k:s))';
+					M.block(k , k , S - k , 1) = (G.col(k ).adjoint() * P.rightCols(S - k )).adjoint();
+
+					if (internal::isApprox(M(k,k), Scalar(0)))
+					{
+						return false;
+					}
+
+					//Make r orthogonal to q_i, i = 0..k-1
+					Scalar beta = f(k ) / M(k , k );
+					r = r - beta * G.col(k );
+					x = x + beta * U.col(k );
+					normr = r.norm();
+
+					if (replacement && normr > tolb / mp)
+					{
+						trueres = true;
+					}
+
+					//Smoothing:
+					if (smoothing)
+					{
+						t = r_s - r;
+						//gamma is a Scalar, but the conversion is not allowed
+						Scalar gamma = t.dot(r_s) / t.norm();
+						r_s = r_s - gamma * t;
+						x_s = x_s - gamma * (x_s - x);
+						normr = r_s.norm();
+					}
+
+					if (normr < tolb || iter == maxit)
+					{
+						break;
+					}
+
+					//New f = P'*r (first k  components are zero)
+					if (k < S-1)
+					{
+						f.segment(k + 1, S - (k + 1) ) = f.segment(k + 1 , S - (k + 1)) - beta * M.block(k + 1 , k , S - (k + 1), 1);
+					}
+				}//end for
+
+				if (normr < tolb || iter == maxit)
+				{
+					break;
+				}
+
+				//Now we have sufficient vectors in G_j to compute residual in G_j+1
+				//Note: r is already perpendicular to P so v = r
+				//Preconditioning
+				v = r;
+				v = precond.solve(v);
+
+				//Matrix-vector multiplication:
+				t = A * v;
+
+				//Computation of a new omega
+				om = internal::omega(t, r, angle);
+
+				if (om == RealScalar(0.0))
+				{
+					return false;
+				}
+
+				r = r - om * t;
+				x = x + om * v;
+				normr = r.norm();
+
+				if (replacement && normr > tolb / mp)
+				{
+					trueres = true;
+				}
+
+				//Residual replacement?
+				if (trueres && normr < normb)
+				{
+					r = b - A * x;
+					trueres = false;
+					replacements++;
+				}
+
+				//Smoothing:
+				if (smoothing)
+				{
+					t = r_s - r;
+					Scalar gamma = t.dot(r_s) /t.norm();
+					r_s = r_s - gamma * t;
+					x_s = x_s - gamma * (x_s - x);
+					normr = r_s.norm();
+				}
+
+				iter++;
+
+			}//end while
+
+			if (smoothing)
+			{
+				x = x_s;
+			}
+			relres=normr/normb;
+			return true;
+		}
+
+	}  // namespace internal
+
+	template <typename _MatrixType, typename _Preconditioner = DiagonalPreconditioner<typename _MatrixType::Scalar> >
+	class IDRS;
+
+	namespace internal
+	{
+
+		template <typename _MatrixType, typename _Preconditioner>
+		struct traits<Eigen::IDRS<_MatrixType, _Preconditioner> >
+		{
+			typedef _MatrixType MatrixType;
+			typedef _Preconditioner Preconditioner;
+		};
+
+	}  // namespace internal
+
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief The Induced Dimension Reduction method (IDR(s)) is a short-recurrences Krylov method for sparse square problems.
+  *
+  * This class allows to solve for A.x = b sparse linear problems. The vectors x and b can be either dense or sparse.
+  * he Induced Dimension Reduction method, IDR(), is a robust and efficient short-recurrence Krylov subspace method for
+  * solving large nonsymmetric systems of linear equations.
+  *
+  * For indefinite systems IDR(S) outperforms both BiCGStab and BiCGStab(L). Additionally, IDR(S) can handle matrices
+  * with complex eigenvalues more efficiently than BiCGStab.
+  *
+  * Many problems that do not converge for BiCGSTAB converge for IDR(s) (for larger values of s). And if both methods 
+  * converge the convergence for IDR(s) is typically much faster for difficult systems (for example indefinite problems). 
+  *
+  * IDR(s) is a limited memory finite termination method. In exact arithmetic it converges in at most N+N/s iterations,
+  * with N the system size.  It uses a fixed number of 4+3s vector. In comparison, BiCGSTAB terminates in 2N iterations 
+  * and uses 7 vectors. GMRES terminates in at most N iterations, and uses I+3 vectors, with I the number of iterations. 
+  * Restarting GMRES limits the memory consumption, but destroys the finite termination property.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
+  * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
+  *
+  * \implsparsesolverconcept
+  *
+  * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
+  * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
+  * and NumTraits<Scalar>::epsilon() for the tolerance.
+  *
+  * The tolerance corresponds to the relative residual error: |Ax-b|/|b|
+  *
+  * \b Performance: when using sparse matrices, best performance is achied for a row-major sparse matrix format.
+  * Moreover, in this case multi-threading can be exploited if the user code is compiled with OpenMP enabled.
+  * See \ref TopicMultiThreading for details.
+  *
+  * By default the iterations start with x=0 as an initial guess of the solution.
+  * One can control the start using the solveWithGuess() method.
+  *
+  * IDR(s) can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+  *
+  * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
+  */
+	template <typename _MatrixType, typename _Preconditioner>
+	class IDRS : public IterativeSolverBase<IDRS<_MatrixType, _Preconditioner> >
+	{
+
+		public:
+			typedef _MatrixType MatrixType;
+			typedef typename MatrixType::Scalar Scalar;
+			typedef typename MatrixType::RealScalar RealScalar;
+			typedef _Preconditioner Preconditioner;
+
+		private:
+			typedef IterativeSolverBase<IDRS> Base;
+			using Base::m_error;
+			using Base::m_info;
+			using Base::m_isInitialized;
+			using Base::m_iterations;
+			using Base::matrix;
+			Index m_S;
+			bool m_smoothing;
+			RealScalar m_angle;
+			bool m_residual;
+
+		public:
+			/** Default constructor. */
+			IDRS(): m_S(4), m_smoothing(false), m_angle(RealScalar(0.7)), m_residual(false) {}
+
+			/**     Initialize the solver with matrix \a A for further \c Ax=b solving.
+
+			        This constructor is a shortcut for the default constructor followed
+			        by a call to compute().
+
+			        \warning this class stores a reference to the matrix A as well as some
+			        precomputed values that depend on it. Therefore, if \a A is changed
+			        this class becomes invalid. Call compute() to update it with the new
+			        matrix A, or modify a copy of A.
+			*/
+			template <typename MatrixDerived>
+			explicit IDRS(const EigenBase<MatrixDerived>& A) : Base(A.derived()), m_S(4), m_smoothing(false),
+															   m_angle(RealScalar(0.7)), m_residual(false) {}
+
+
+			/** \internal */
+			/**     Loops over the number of columns of b and does the following:
+			                1. sets the tolerence and maxIterations
+			                2. Calls the function that has the core solver routine
+			*/
+			template <typename Rhs, typename Dest>
+			void _solve_vector_with_guess_impl(const Rhs& b, Dest& x) const
+			{
+				m_iterations = Base::maxIterations();
+				m_error = Base::m_tolerance;
+
+				bool ret = internal::idrs(matrix(), b, x, Base::m_preconditioner, m_iterations, m_error, m_S,m_smoothing,m_angle,m_residual);
+
+				m_info = (!ret) ? NumericalIssue : m_error <= Base::m_tolerance ? Success : NoConvergence;
+			}
+
+			/** Sets the parameter S, indicating the dimension of the shadow space. Default is 4*/
+			void setS(Index S)
+			{
+				if (S < 1)
+				{
+					S = 4;
+				}
+
+				m_S = S;
+			}
+
+			/** Switches off and on smoothing.
+			Residual smoothing results in monotonically decreasing residual norms at
+			the expense of two extra vectors of storage and a few extra vector
+			operations. Although monotonic decrease of the residual norms is a
+			desirable property, the rate of convergence of the unsmoothed process and
+			the smoothed process is basically the same. Default is off */
+			void setSmoothing(bool smoothing)
+			{
+				m_smoothing=smoothing;
+			}
+
+			/** The angle must be a real scalar. In IDR(s), a value for the
+			iteration parameter omega must be chosen in every s+1th step. The most
+			natural choice is to select a value to minimize the norm of the next residual.
+			This corresponds to the parameter omega = 0. In practice, this may lead to
+			values of omega that are so small that the other iteration parameters
+			cannot be computed with sufficient accuracy. In such cases it is better to
+			increase the value of omega sufficiently such that a compromise is reached
+			between accurate computations and reduction of the residual norm. The
+			parameter angle =0.7 (”maintaining the convergence strategy”)
+			results in such a compromise. */
+			void setAngle(RealScalar angle)
+			{
+				m_angle=angle;
+			}
+
+			/** The parameter replace is a logical that determines whether a
+			residual replacement strategy is employed to increase the accuracy of the
+			solution. */
+			void setResidualUpdate(bool update)
+			{
+				m_residual=update;
+			}
+
+	};
+
+}  // namespace Eigen
+
+#endif /* EIGEN_IDRS_H */
diff --git a/src/EigenUnsupported/src/IterativeSolvers/IncompleteLU.h b/src/EigenUnsupported/src/IterativeSolvers/IncompleteLU.h
new file mode 100644
index 0000000..7d08c35
--- /dev/null
+++ b/src/EigenUnsupported/src/IterativeSolvers/IncompleteLU.h
@@ -0,0 +1,90 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_INCOMPLETE_LU_H
+#define EIGEN_INCOMPLETE_LU_H
+
+namespace Eigen { 
+
+template <typename _Scalar>
+class IncompleteLU : public SparseSolverBase<IncompleteLU<_Scalar> >
+{
+  protected:
+    typedef SparseSolverBase<IncompleteLU<_Scalar> > Base;
+    using Base::m_isInitialized;
+    
+    typedef _Scalar Scalar;
+    typedef Matrix<Scalar,Dynamic,1> Vector;
+    typedef typename Vector::Index Index;
+    typedef SparseMatrix<Scalar,RowMajor> FactorType;
+
+  public:
+    typedef Matrix<Scalar,Dynamic,Dynamic> MatrixType;
+
+    IncompleteLU() {}
+
+    template<typename MatrixType>
+    IncompleteLU(const MatrixType& mat)
+    {
+      compute(mat);
+    }
+
+    Index rows() const { return m_lu.rows(); }
+    Index cols() const { return m_lu.cols(); }
+
+    template<typename MatrixType>
+    IncompleteLU& compute(const MatrixType& mat)
+    {
+      m_lu = mat;
+      int size = mat.cols();
+      Vector diag(size);
+      for(int i=0; i<size; ++i)
+      {
+        typename FactorType::InnerIterator k_it(m_lu,i);
+        for(; k_it && k_it.index()<i; ++k_it)
+        {
+          int k = k_it.index();
+          k_it.valueRef() /= diag(k);
+
+          typename FactorType::InnerIterator j_it(k_it);
+          typename FactorType::InnerIterator kj_it(m_lu, k);
+          while(kj_it && kj_it.index()<=k) ++kj_it;
+          for(++j_it; j_it; )
+          {
+            if(kj_it.index()==j_it.index())
+            {
+              j_it.valueRef() -= k_it.value() * kj_it.value();
+              ++j_it;
+              ++kj_it;
+            }
+            else if(kj_it.index()<j_it.index()) ++kj_it;
+            else                                ++j_it;
+          }
+        }
+        if(k_it && k_it.index()==i) diag(i) = k_it.value();
+        else                        diag(i) = 1;
+      }
+      m_isInitialized = true;
+      return *this;
+    }
+
+    template<typename Rhs, typename Dest>
+    void _solve_impl(const Rhs& b, Dest& x) const
+    {
+      x = m_lu.template triangularView<UnitLower>().solve(b);
+      x = m_lu.template triangularView<Upper>().solve(x);
+    }
+
+  protected:
+    FactorType m_lu;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_INCOMPLETE_LU_H
diff --git a/src/EigenUnsupported/src/IterativeSolvers/IterationController.h b/src/EigenUnsupported/src/IterativeSolvers/IterationController.h
new file mode 100644
index 0000000..a116e09
--- /dev/null
+++ b/src/EigenUnsupported/src/IterativeSolvers/IterationController.h
@@ -0,0 +1,154 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+
+/* NOTE The class IterationController has been adapted from the iteration
+ *      class of the GMM++ and ITL libraries.
+ */
+
+//=======================================================================
+// Copyright (C) 1997-2001
+// Authors: Andrew Lumsdaine <lums@osl.iu.edu> 
+//          Lie-Quan Lee     <llee@osl.iu.edu>
+//
+// This file is part of the Iterative Template Library
+//
+// You should have received a copy of the License Agreement for the
+// Iterative Template Library along with the software;  see the
+// file LICENSE.  
+//
+// Permission to modify the code and to distribute modified code is
+// granted, provided the text of this NOTICE is retained, a notice that
+// the code was modified is included with the above COPYRIGHT NOTICE and
+// with the COPYRIGHT NOTICE in the LICENSE file, and that the LICENSE
+// file is distributed with the modified code.
+//
+// LICENSOR MAKES NO REPRESENTATIONS OR WARRANTIES, EXPRESS OR IMPLIED.
+// By way of example, but not limitation, Licensor MAKES NO
+// REPRESENTATIONS OR WARRANTIES OF MERCHANTABILITY OR FITNESS FOR ANY
+// PARTICULAR PURPOSE OR THAT THE USE OF THE LICENSED SOFTWARE COMPONENTS
+// OR DOCUMENTATION WILL NOT INFRINGE ANY PATENTS, COPYRIGHTS, TRADEMARKS
+// OR OTHER RIGHTS.
+//=======================================================================
+
+//========================================================================
+//
+// Copyright (C) 2002-2007 Yves Renard
+//
+// This file is a part of GETFEM++
+//
+// Getfem++ is free software; you can redistribute it and/or modify
+// it under the terms of the GNU Lesser General Public License as
+// published by the Free Software Foundation; version 2.1 of the License.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU Lesser General Public License for more details.
+// You should have received a copy of the GNU Lesser General Public
+// License along with this program; if not, write to the Free Software
+// Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301,
+// USA.
+//
+//========================================================================
+
+#include "../../../../Eigen/src/Core/util/NonMPL2.h"
+
+#ifndef EIGEN_ITERATION_CONTROLLER_H
+#define EIGEN_ITERATION_CONTROLLER_H
+
+namespace Eigen { 
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \class IterationController
+  *
+  * \brief Controls the iterations of the iterative solvers
+  *
+  * This class has been adapted from the iteration class of GMM++ and ITL libraries.
+  *
+  */
+class IterationController
+{
+  protected :
+    double m_rhsn;        ///< Right hand side norm
+    size_t m_maxiter;     ///< Max. number of iterations
+    int m_noise;          ///< if noise > 0 iterations are printed
+    double m_resmax;      ///< maximum residual
+    double m_resminreach, m_resadd;
+    size_t m_nit;         ///< iteration number
+    double m_res;         ///< last computed residual
+    bool m_written;
+    void (*m_callback)(const IterationController&);
+  public :
+
+    void init()
+    {
+      m_nit = 0; m_res = 0.0; m_written = false;
+      m_resminreach = 1E50; m_resadd = 0.0;
+      m_callback = 0;
+    }
+
+    IterationController(double r = 1.0E-8, int noi = 0, size_t mit = size_t(-1))
+      : m_rhsn(1.0), m_maxiter(mit), m_noise(noi), m_resmax(r) { init(); }
+
+    void operator ++(int) { m_nit++; m_written = false; m_resadd += m_res; }
+    void operator ++() { (*this)++; }
+
+    bool first() { return m_nit == 0; }
+
+    /* get/set the "noisyness" (verbosity) of the solvers */
+    int noiseLevel() const { return m_noise; }
+    void setNoiseLevel(int n) { m_noise = n; }
+    void reduceNoiseLevel() { if (m_noise > 0) m_noise--; }
+
+    double maxResidual() const { return m_resmax; }
+    void setMaxResidual(double r) { m_resmax = r; }
+
+    double residual() const { return m_res; }
+
+    /* change the user-definable callback, called after each iteration */
+    void setCallback(void (*t)(const IterationController&))
+    {
+      m_callback = t;
+    }
+
+    size_t iteration() const { return m_nit; }
+    void setIteration(size_t i) { m_nit = i; }
+
+    size_t maxIterarions() const { return m_maxiter; }
+    void setMaxIterations(size_t i) { m_maxiter = i; }
+
+    double rhsNorm() const { return m_rhsn; }
+    void setRhsNorm(double r) { m_rhsn = r; }
+
+    bool converged() const { return m_res <= m_rhsn * m_resmax; }
+    bool converged(double nr)
+    {
+      using std::abs;
+      m_res = abs(nr); 
+      m_resminreach = (std::min)(m_resminreach, m_res);
+      return converged();
+    }
+    template<typename VectorType> bool converged(const VectorType &v)
+    { return converged(v.squaredNorm()); }
+
+    bool finished(double nr)
+    {
+      if (m_callback) m_callback(*this);
+      if (m_noise > 0 && !m_written)
+      {
+        converged(nr);
+        m_written = true;
+      }
+      return (m_nit >= m_maxiter || converged(nr));
+    }
+    template <typename VectorType>
+    bool finished(const MatrixBase<VectorType> &v)
+    { return finished(double(v.squaredNorm())); }
+
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_ITERATION_CONTROLLER_H
diff --git a/src/EigenUnsupported/src/IterativeSolvers/MINRES.h b/src/EigenUnsupported/src/IterativeSolvers/MINRES.h
new file mode 100644
index 0000000..5db454d
--- /dev/null
+++ b/src/EigenUnsupported/src/IterativeSolvers/MINRES.h
@@ -0,0 +1,267 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Giacomo Po <gpo@ucla.edu>
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2018 David Hyde <dabh@stanford.edu>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_MINRES_H_
+#define EIGEN_MINRES_H_
+
+
+namespace Eigen {
+    
+    namespace internal {
+        
+        /** \internal Low-level MINRES algorithm
+         * \param mat The matrix A
+         * \param rhs The right hand side vector b
+         * \param x On input and initial solution, on output the computed solution.
+         * \param precond A right preconditioner being able to efficiently solve for an
+         *                approximation of Ax=b (regardless of b)
+         * \param iters On input the max number of iteration, on output the number of performed iterations.
+         * \param tol_error On input the tolerance error, on output an estimation of the relative error.
+         */
+        template<typename MatrixType, typename Rhs, typename Dest, typename Preconditioner>
+        EIGEN_DONT_INLINE
+        void minres(const MatrixType& mat, const Rhs& rhs, Dest& x,
+                    const Preconditioner& precond, Index& iters,
+                    typename Dest::RealScalar& tol_error)
+        {
+            using std::sqrt;
+            typedef typename Dest::RealScalar RealScalar;
+            typedef typename Dest::Scalar Scalar;
+            typedef Matrix<Scalar,Dynamic,1> VectorType;
+
+            // Check for zero rhs
+            const RealScalar rhsNorm2(rhs.squaredNorm());
+            if(rhsNorm2 == 0)
+            {
+                x.setZero();
+                iters = 0;
+                tol_error = 0;
+                return;
+            }
+            
+            // initialize
+            const Index maxIters(iters);  // initialize maxIters to iters
+            const Index N(mat.cols());    // the size of the matrix
+            const RealScalar threshold2(tol_error*tol_error*rhsNorm2); // convergence threshold (compared to residualNorm2)
+            
+            // Initialize preconditioned Lanczos
+            VectorType v_old(N); // will be initialized inside loop
+            VectorType v( VectorType::Zero(N) ); //initialize v
+            VectorType v_new(rhs-mat*x); //initialize v_new
+            RealScalar residualNorm2(v_new.squaredNorm());
+            VectorType w(N); // will be initialized inside loop
+            VectorType w_new(precond.solve(v_new)); // initialize w_new
+//            RealScalar beta; // will be initialized inside loop
+            RealScalar beta_new2(v_new.dot(w_new));
+            eigen_assert(beta_new2 >= 0.0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE");
+            RealScalar beta_new(sqrt(beta_new2));
+            const RealScalar beta_one(beta_new);
+            // Initialize other variables
+            RealScalar c(1.0); // the cosine of the Givens rotation
+            RealScalar c_old(1.0);
+            RealScalar s(0.0); // the sine of the Givens rotation
+            RealScalar s_old(0.0); // the sine of the Givens rotation
+            VectorType p_oold(N); // will be initialized in loop
+            VectorType p_old(VectorType::Zero(N)); // initialize p_old=0
+            VectorType p(p_old); // initialize p=0
+            RealScalar eta(1.0);
+                        
+            iters = 0; // reset iters
+            while ( iters < maxIters )
+            {
+                // Preconditioned Lanczos
+                /* Note that there are 4 variants on the Lanczos algorithm. These are
+                 * described in Paige, C. C. (1972). Computational variants of
+                 * the Lanczos method for the eigenproblem. IMA Journal of Applied
+                 * Mathematics, 10(3), 373-381. The current implementation corresponds 
+                 * to the case A(2,7) in the paper. It also corresponds to 
+                 * algorithm 6.14 in Y. Saad, Iterative Methods for Sparse Linear
+                 * Systems, 2003 p.173. For the preconditioned version see 
+                 * A. Greenbaum, Iterative Methods for Solving Linear Systems, SIAM (1987).
+                 */
+                const RealScalar beta(beta_new);
+                v_old = v; // update: at first time step, this makes v_old = 0 so value of beta doesn't matter
+                v_new /= beta_new; // overwrite v_new for next iteration
+                w_new /= beta_new; // overwrite w_new for next iteration
+                v = v_new; // update
+                w = w_new; // update
+                v_new.noalias() = mat*w - beta*v_old; // compute v_new
+                const RealScalar alpha = v_new.dot(w);
+                v_new -= alpha*v; // overwrite v_new
+                w_new = precond.solve(v_new); // overwrite w_new
+                beta_new2 = v_new.dot(w_new); // compute beta_new
+                eigen_assert(beta_new2 >= 0.0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE");
+                beta_new = sqrt(beta_new2); // compute beta_new
+                
+                // Givens rotation
+                const RealScalar r2 =s*alpha+c*c_old*beta; // s, s_old, c and c_old are still from previous iteration
+                const RealScalar r3 =s_old*beta; // s, s_old, c and c_old are still from previous iteration
+                const RealScalar r1_hat=c*alpha-c_old*s*beta;
+                const RealScalar r1 =sqrt( std::pow(r1_hat,2) + std::pow(beta_new,2) );
+                c_old = c; // store for next iteration
+                s_old = s; // store for next iteration
+                c=r1_hat/r1; // new cosine
+                s=beta_new/r1; // new sine
+                
+                // Update solution
+                p_oold = p_old;
+                p_old = p;
+                p.noalias()=(w-r2*p_old-r3*p_oold) /r1; // IS NOALIAS REQUIRED?
+                x += beta_one*c*eta*p;
+                
+                /* Update the squared residual. Note that this is the estimated residual.
+                The real residual |Ax-b|^2 may be slightly larger */
+                residualNorm2 *= s*s;
+                
+                if ( residualNorm2 < threshold2)
+                {
+                    break;
+                }
+                
+                eta=-s*eta; // update eta
+                iters++; // increment iteration number (for output purposes)
+            }
+            
+            /* Compute error. Note that this is the estimated error. The real 
+             error |Ax-b|/|b| may be slightly larger */
+            tol_error = std::sqrt(residualNorm2 / rhsNorm2);
+        }
+        
+    }
+    
+    template< typename _MatrixType, int _UpLo=Lower,
+    typename _Preconditioner = IdentityPreconditioner>
+    class MINRES;
+    
+    namespace internal {
+        
+        template< typename _MatrixType, int _UpLo, typename _Preconditioner>
+        struct traits<MINRES<_MatrixType,_UpLo,_Preconditioner> >
+        {
+            typedef _MatrixType MatrixType;
+            typedef _Preconditioner Preconditioner;
+        };
+        
+    }
+    
+    /** \ingroup IterativeLinearSolvers_Module
+     * \brief A minimal residual solver for sparse symmetric problems
+     *
+     * This class allows to solve for A.x = b sparse linear problems using the MINRES algorithm
+     * of Paige and Saunders (1975). The sparse matrix A must be symmetric (possibly indefinite).
+     * The vectors x and b can be either dense or sparse.
+     *
+     * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
+     * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower,
+     *               Upper, or Lower|Upper in which the full matrix entries will be considered. Default is Lower.
+     * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
+     *
+     * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
+     * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
+     * and NumTraits<Scalar>::epsilon() for the tolerance.
+     *
+     * This class can be used as the direct solver classes. Here is a typical usage example:
+     * \code
+     * int n = 10000;
+     * VectorXd x(n), b(n);
+     * SparseMatrix<double> A(n,n);
+     * // fill A and b
+     * MINRES<SparseMatrix<double> > mr;
+     * mr.compute(A);
+     * x = mr.solve(b);
+     * std::cout << "#iterations:     " << mr.iterations() << std::endl;
+     * std::cout << "estimated error: " << mr.error()      << std::endl;
+     * // update b, and solve again
+     * x = mr.solve(b);
+     * \endcode
+     *
+     * By default the iterations start with x=0 as an initial guess of the solution.
+     * One can control the start using the solveWithGuess() method.
+     *
+     * MINRES can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+     *
+     * \sa class ConjugateGradient, BiCGSTAB, SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
+     */
+    template< typename _MatrixType, int _UpLo, typename _Preconditioner>
+    class MINRES : public IterativeSolverBase<MINRES<_MatrixType,_UpLo,_Preconditioner> >
+    {
+        
+        typedef IterativeSolverBase<MINRES> Base;
+        using Base::matrix;
+        using Base::m_error;
+        using Base::m_iterations;
+        using Base::m_info;
+        using Base::m_isInitialized;
+    public:
+        using Base::_solve_impl;
+        typedef _MatrixType MatrixType;
+        typedef typename MatrixType::Scalar Scalar;
+        typedef typename MatrixType::RealScalar RealScalar;
+        typedef _Preconditioner Preconditioner;
+        
+        enum {UpLo = _UpLo};
+        
+    public:
+        
+        /** Default constructor. */
+        MINRES() : Base() {}
+        
+        /** Initialize the solver with matrix \a A for further \c Ax=b solving.
+         *
+         * This constructor is a shortcut for the default constructor followed
+         * by a call to compute().
+         *
+         * \warning this class stores a reference to the matrix A as well as some
+         * precomputed values that depend on it. Therefore, if \a A is changed
+         * this class becomes invalid. Call compute() to update it with the new
+         * matrix A, or modify a copy of A.
+         */
+        template<typename MatrixDerived>
+        explicit MINRES(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
+        
+        /** Destructor. */
+        ~MINRES(){}
+
+        /** \internal */
+        template<typename Rhs,typename Dest>
+        void _solve_vector_with_guess_impl(const Rhs& b, Dest& x) const
+        {
+            typedef typename Base::MatrixWrapper MatrixWrapper;
+            typedef typename Base::ActualMatrixType ActualMatrixType;
+            enum {
+              TransposeInput  =   (!MatrixWrapper::MatrixFree)
+                              &&  (UpLo==(Lower|Upper))
+                              &&  (!MatrixType::IsRowMajor)
+                              &&  (!NumTraits<Scalar>::IsComplex)
+            };
+            typedef typename internal::conditional<TransposeInput,Transpose<const ActualMatrixType>, ActualMatrixType const&>::type RowMajorWrapper;
+            EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(MatrixWrapper::MatrixFree,UpLo==(Lower|Upper)),MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY);
+            typedef typename internal::conditional<UpLo==(Lower|Upper),
+                                                  RowMajorWrapper,
+                                                  typename MatrixWrapper::template ConstSelfAdjointViewReturnType<UpLo>::Type
+                                            >::type SelfAdjointWrapper;
+
+            m_iterations = Base::maxIterations();
+            m_error = Base::m_tolerance;
+            RowMajorWrapper row_mat(matrix());
+            internal::minres(SelfAdjointWrapper(row_mat), b, x,
+                             Base::m_preconditioner, m_iterations, m_error);
+            m_info = m_error <= Base::m_tolerance ? Success : NoConvergence;
+        }
+        
+    protected:
+        
+    };
+
+} // end namespace Eigen
+
+#endif // EIGEN_MINRES_H
diff --git a/src/EigenUnsupported/src/IterativeSolvers/Scaling.h b/src/EigenUnsupported/src/IterativeSolvers/Scaling.h
new file mode 100644
index 0000000..9b3eb53
--- /dev/null
+++ b/src/EigenUnsupported/src/IterativeSolvers/Scaling.h
@@ -0,0 +1,193 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Desire NUENTSA WAKAM <desire.nuentsa_wakam@inria.fr
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ITERSCALING_H
+#define EIGEN_ITERSCALING_H
+
+namespace Eigen {
+
+/**
+  * \ingroup IterativeSolvers_Module
+  * \brief iterative scaling algorithm to equilibrate rows and column norms in matrices
+  * 
+  * This class can be used as a preprocessing tool to accelerate the convergence of iterative methods 
+  * 
+  * This feature is  useful to limit the pivoting amount during LU/ILU factorization
+  * The  scaling strategy as presented here preserves the symmetry of the problem
+  * NOTE It is assumed that the matrix does not have empty row or column, 
+  * 
+  * Example with key steps 
+  * \code
+  * VectorXd x(n), b(n);
+  * SparseMatrix<double> A;
+  * // fill A and b;
+  * IterScaling<SparseMatrix<double> > scal; 
+  * // Compute the left and right scaling vectors. The matrix is equilibrated at output
+  * scal.computeRef(A); 
+  * // Scale the right hand side
+  * b = scal.LeftScaling().cwiseProduct(b); 
+  * // Now, solve the equilibrated linear system with any available solver
+  * 
+  * // Scale back the computed solution
+  * x = scal.RightScaling().cwiseProduct(x); 
+  * \endcode
+  * 
+  * \tparam _MatrixType the type of the matrix. It should be a real square sparsematrix
+  * 
+  * References : D. Ruiz and B. Ucar, A Symmetry Preserving Algorithm for Matrix Scaling, INRIA Research report RR-7552
+  * 
+  * \sa \ref IncompleteLUT 
+  */
+template<typename _MatrixType>
+class IterScaling
+{
+  public:
+    typedef _MatrixType MatrixType; 
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::Index Index;
+    
+  public:
+    IterScaling() { init(); }
+    
+    IterScaling(const MatrixType& matrix)
+    {
+      init();
+      compute(matrix);
+    }
+    
+    ~IterScaling() { }
+    
+    /** 
+     * Compute the left and right diagonal matrices to scale the input matrix @p mat
+     * 
+     * FIXME This algorithm will be modified such that the diagonal elements are permuted on the diagonal. 
+     * 
+     * \sa LeftScaling() RightScaling()
+     */
+    void compute (const MatrixType& mat)
+    {
+      using std::abs;
+      int m = mat.rows(); 
+      int n = mat.cols();
+      eigen_assert((m>0 && m == n) && "Please give a non - empty matrix");
+      m_left.resize(m); 
+      m_right.resize(n);
+      m_left.setOnes();
+      m_right.setOnes();
+      m_matrix = mat;
+      VectorXd Dr, Dc, DrRes, DcRes; // Temporary Left and right scaling vectors
+      Dr.resize(m); Dc.resize(n);
+      DrRes.resize(m); DcRes.resize(n);
+      double EpsRow = 1.0, EpsCol = 1.0;
+      int its = 0; 
+      do
+      { // Iterate until the infinite norm of each row and column is approximately 1
+        // Get the maximum value in each row and column
+        Dr.setZero(); Dc.setZero();
+        for (int k=0; k<m_matrix.outerSize(); ++k)
+        {
+          for (typename MatrixType::InnerIterator it(m_matrix, k); it; ++it)
+          {
+            if ( Dr(it.row()) < abs(it.value()) )
+              Dr(it.row()) = abs(it.value());
+            
+            if ( Dc(it.col()) < abs(it.value()) )
+              Dc(it.col()) = abs(it.value());
+          }
+        }
+        for (int i = 0; i < m; ++i) 
+        {
+          Dr(i) = std::sqrt(Dr(i));
+        }
+        for (int i = 0; i < n; ++i) 
+        {
+          Dc(i) = std::sqrt(Dc(i));
+        }
+        // Save the scaling factors 
+        for (int i = 0; i < m; ++i) 
+        {
+          m_left(i) /= Dr(i);
+        }
+        for (int i = 0; i < n; ++i) 
+        {
+          m_right(i) /= Dc(i);
+        }
+        // Scale the rows and the columns of the matrix
+        DrRes.setZero(); DcRes.setZero(); 
+        for (int k=0; k<m_matrix.outerSize(); ++k)
+        {
+          for (typename MatrixType::InnerIterator it(m_matrix, k); it; ++it)
+          {
+            it.valueRef() = it.value()/( Dr(it.row()) * Dc(it.col()) );
+            // Accumulate the norms of the row and column vectors   
+            if ( DrRes(it.row()) < abs(it.value()) )
+              DrRes(it.row()) = abs(it.value());
+            
+            if ( DcRes(it.col()) < abs(it.value()) )
+              DcRes(it.col()) = abs(it.value());
+          }
+        }  
+        DrRes.array() = (1-DrRes.array()).abs();
+        EpsRow = DrRes.maxCoeff();
+        DcRes.array() = (1-DcRes.array()).abs();
+        EpsCol = DcRes.maxCoeff();
+        its++;
+      }while ( (EpsRow >m_tol || EpsCol > m_tol) && (its < m_maxits) );
+      m_isInitialized = true;
+    }
+    /** Compute the left and right vectors to scale the vectors
+     * the input matrix is scaled with the computed vectors at output
+     * 
+     * \sa compute()
+     */
+    void computeRef (MatrixType& mat)
+    {
+      compute (mat);
+      mat = m_matrix;
+    }
+    /** Get the vector to scale the rows of the matrix 
+     */
+    VectorXd& LeftScaling()
+    {
+      return m_left;
+    }
+    
+    /** Get the vector to scale the columns of the matrix 
+     */
+    VectorXd& RightScaling()
+    {
+      return m_right;
+    }
+    
+    /** Set the tolerance for the convergence of the iterative scaling algorithm
+     */
+    void setTolerance(double tol)
+    {
+      m_tol = tol; 
+    }
+      
+  protected:
+    
+    void init()
+    {
+      m_tol = 1e-10;
+      m_maxits = 5;
+      m_isInitialized = false;
+    }
+    
+    MatrixType m_matrix;
+    mutable ComputationInfo m_info; 
+    bool m_isInitialized; 
+    VectorXd m_left; // Left scaling vector
+    VectorXd m_right; // m_right scaling vector
+    double m_tol; 
+    int m_maxits; // Maximum number of iterations allowed
+};
+}
+#endif
diff --git a/src/EigenUnsupported/src/KroneckerProduct/KroneckerTensorProduct.h b/src/EigenUnsupported/src/KroneckerProduct/KroneckerTensorProduct.h
new file mode 100644
index 0000000..6a9b0be
--- /dev/null
+++ b/src/EigenUnsupported/src/KroneckerProduct/KroneckerTensorProduct.h
@@ -0,0 +1,305 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Kolja Brix <brix@igpm.rwth-aachen.de>
+// Copyright (C) 2011 Andreas Platen <andiplaten@gmx.de>
+// Copyright (C) 2012 Chen-Pang He <jdh8@ms63.hinet.net>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef KRONECKER_TENSOR_PRODUCT_H
+#define KRONECKER_TENSOR_PRODUCT_H
+
+namespace Eigen {
+
+/*!
+ * \ingroup KroneckerProduct_Module
+ *
+ * \brief The base class of dense and sparse Kronecker product.
+ *
+ * \tparam Derived is the derived type.
+ */
+template<typename Derived>
+class KroneckerProductBase : public ReturnByValue<Derived>
+{
+  private:
+    typedef typename internal::traits<Derived> Traits;
+    typedef typename Traits::Scalar Scalar;
+
+  protected:
+    typedef typename Traits::Lhs Lhs;
+    typedef typename Traits::Rhs Rhs;
+
+  public:
+    /*! \brief Constructor. */
+    KroneckerProductBase(const Lhs& A, const Rhs& B)
+      : m_A(A), m_B(B)
+    {}
+
+    inline Index rows() const { return m_A.rows() * m_B.rows(); }
+    inline Index cols() const { return m_A.cols() * m_B.cols(); }
+
+    /*!
+     * This overrides ReturnByValue::coeff because this function is
+     * efficient enough.
+     */
+    Scalar coeff(Index row, Index col) const
+    {
+      return m_A.coeff(row / m_B.rows(), col / m_B.cols()) *
+             m_B.coeff(row % m_B.rows(), col % m_B.cols());
+    }
+
+    /*!
+     * This overrides ReturnByValue::coeff because this function is
+     * efficient enough.
+     */
+    Scalar coeff(Index i) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+      return m_A.coeff(i / m_A.size()) * m_B.coeff(i % m_A.size());
+    }
+
+  protected:
+    typename Lhs::Nested m_A;
+    typename Rhs::Nested m_B;
+};
+
+/*!
+ * \ingroup KroneckerProduct_Module
+ *
+ * \brief Kronecker tensor product helper class for dense matrices
+ *
+ * This class is the return value of kroneckerProduct(MatrixBase,
+ * MatrixBase). Use the function rather than construct this class
+ * directly to avoid specifying template prarameters.
+ *
+ * \tparam Lhs  Type of the left-hand side, a matrix expression.
+ * \tparam Rhs  Type of the rignt-hand side, a matrix expression.
+ */
+template<typename Lhs, typename Rhs>
+class KroneckerProduct : public KroneckerProductBase<KroneckerProduct<Lhs,Rhs> >
+{
+  private:
+    typedef KroneckerProductBase<KroneckerProduct> Base;
+    using Base::m_A;
+    using Base::m_B;
+
+  public:
+    /*! \brief Constructor. */
+    KroneckerProduct(const Lhs& A, const Rhs& B)
+      : Base(A, B)
+    {}
+
+    /*! \brief Evaluate the Kronecker tensor product. */
+    template<typename Dest> void evalTo(Dest& dst) const;
+};
+
+/*!
+ * \ingroup KroneckerProduct_Module
+ *
+ * \brief Kronecker tensor product helper class for sparse matrices
+ *
+ * If at least one of the operands is a sparse matrix expression,
+ * then this class is returned and evaluates into a sparse matrix.
+ *
+ * This class is the return value of kroneckerProduct(EigenBase,
+ * EigenBase). Use the function rather than construct this class
+ * directly to avoid specifying template prarameters.
+ *
+ * \tparam Lhs  Type of the left-hand side, a matrix expression.
+ * \tparam Rhs  Type of the rignt-hand side, a matrix expression.
+ */
+template<typename Lhs, typename Rhs>
+class KroneckerProductSparse : public KroneckerProductBase<KroneckerProductSparse<Lhs,Rhs> >
+{
+  private:
+    typedef KroneckerProductBase<KroneckerProductSparse> Base;
+    using Base::m_A;
+    using Base::m_B;
+
+  public:
+    /*! \brief Constructor. */
+    KroneckerProductSparse(const Lhs& A, const Rhs& B)
+      : Base(A, B)
+    {}
+
+    /*! \brief Evaluate the Kronecker tensor product. */
+    template<typename Dest> void evalTo(Dest& dst) const;
+};
+
+template<typename Lhs, typename Rhs>
+template<typename Dest>
+void KroneckerProduct<Lhs,Rhs>::evalTo(Dest& dst) const
+{
+  const int BlockRows = Rhs::RowsAtCompileTime,
+            BlockCols = Rhs::ColsAtCompileTime;
+  const Index Br = m_B.rows(),
+              Bc = m_B.cols();
+  for (Index i=0; i < m_A.rows(); ++i)
+    for (Index j=0; j < m_A.cols(); ++j)
+      Block<Dest,BlockRows,BlockCols>(dst,i*Br,j*Bc,Br,Bc) = m_A.coeff(i,j) * m_B;
+}
+
+template<typename Lhs, typename Rhs>
+template<typename Dest>
+void KroneckerProductSparse<Lhs,Rhs>::evalTo(Dest& dst) const
+{
+  Index Br = m_B.rows(), Bc = m_B.cols();
+  dst.resize(this->rows(), this->cols());
+  dst.resizeNonZeros(0);
+  
+  // 1 - evaluate the operands if needed:
+  typedef typename internal::nested_eval<Lhs,Dynamic>::type Lhs1;
+  typedef typename internal::remove_all<Lhs1>::type Lhs1Cleaned;
+  const Lhs1 lhs1(m_A);
+  typedef typename internal::nested_eval<Rhs,Dynamic>::type Rhs1;
+  typedef typename internal::remove_all<Rhs1>::type Rhs1Cleaned;
+  const Rhs1 rhs1(m_B);
+    
+  // 2 - construct respective iterators
+  typedef Eigen::InnerIterator<Lhs1Cleaned> LhsInnerIterator;
+  typedef Eigen::InnerIterator<Rhs1Cleaned> RhsInnerIterator;
+  
+  // compute number of non-zeros per innervectors of dst
+  {
+    // TODO VectorXi is not necessarily big enough!
+    VectorXi nnzA = VectorXi::Zero(Dest::IsRowMajor ? m_A.rows() : m_A.cols());
+    for (Index kA=0; kA < m_A.outerSize(); ++kA)
+      for (LhsInnerIterator itA(lhs1,kA); itA; ++itA)
+        nnzA(Dest::IsRowMajor ? itA.row() : itA.col())++;
+      
+    VectorXi nnzB = VectorXi::Zero(Dest::IsRowMajor ? m_B.rows() : m_B.cols());
+    for (Index kB=0; kB < m_B.outerSize(); ++kB)
+      for (RhsInnerIterator itB(rhs1,kB); itB; ++itB)
+        nnzB(Dest::IsRowMajor ? itB.row() : itB.col())++;
+    
+    Matrix<int,Dynamic,Dynamic,ColMajor> nnzAB = nnzB * nnzA.transpose();
+    dst.reserve(VectorXi::Map(nnzAB.data(), nnzAB.size()));
+  }
+
+  for (Index kA=0; kA < m_A.outerSize(); ++kA)
+  {
+    for (Index kB=0; kB < m_B.outerSize(); ++kB)
+    {
+      for (LhsInnerIterator itA(lhs1,kA); itA; ++itA)
+      {
+        for (RhsInnerIterator itB(rhs1,kB); itB; ++itB)
+        {
+          Index i = itA.row() * Br + itB.row(),
+                j = itA.col() * Bc + itB.col();
+          dst.insert(i,j) = itA.value() * itB.value();
+        }
+      }
+    }
+  }
+}
+
+namespace internal {
+
+template<typename _Lhs, typename _Rhs>
+struct traits<KroneckerProduct<_Lhs,_Rhs> >
+{
+  typedef typename remove_all<_Lhs>::type Lhs;
+  typedef typename remove_all<_Rhs>::type Rhs;
+  typedef typename ScalarBinaryOpTraits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType Scalar;
+  typedef typename promote_index_type<typename Lhs::StorageIndex, typename Rhs::StorageIndex>::type StorageIndex;
+
+  enum {
+    Rows = size_at_compile_time<traits<Lhs>::RowsAtCompileTime, traits<Rhs>::RowsAtCompileTime>::ret,
+    Cols = size_at_compile_time<traits<Lhs>::ColsAtCompileTime, traits<Rhs>::ColsAtCompileTime>::ret,
+    MaxRows = size_at_compile_time<traits<Lhs>::MaxRowsAtCompileTime, traits<Rhs>::MaxRowsAtCompileTime>::ret,
+    MaxCols = size_at_compile_time<traits<Lhs>::MaxColsAtCompileTime, traits<Rhs>::MaxColsAtCompileTime>::ret
+  };
+
+  typedef Matrix<Scalar,Rows,Cols> ReturnType;
+};
+
+template<typename _Lhs, typename _Rhs>
+struct traits<KroneckerProductSparse<_Lhs,_Rhs> >
+{
+  typedef MatrixXpr XprKind;
+  typedef typename remove_all<_Lhs>::type Lhs;
+  typedef typename remove_all<_Rhs>::type Rhs;
+  typedef typename ScalarBinaryOpTraits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType Scalar;
+  typedef typename cwise_promote_storage_type<typename traits<Lhs>::StorageKind, typename traits<Rhs>::StorageKind, scalar_product_op<typename Lhs::Scalar, typename Rhs::Scalar> >::ret StorageKind;
+  typedef typename promote_index_type<typename Lhs::StorageIndex, typename Rhs::StorageIndex>::type StorageIndex;
+
+  enum {
+    LhsFlags = Lhs::Flags,
+    RhsFlags = Rhs::Flags,
+
+    RowsAtCompileTime = size_at_compile_time<traits<Lhs>::RowsAtCompileTime, traits<Rhs>::RowsAtCompileTime>::ret,
+    ColsAtCompileTime = size_at_compile_time<traits<Lhs>::ColsAtCompileTime, traits<Rhs>::ColsAtCompileTime>::ret,
+    MaxRowsAtCompileTime = size_at_compile_time<traits<Lhs>::MaxRowsAtCompileTime, traits<Rhs>::MaxRowsAtCompileTime>::ret,
+    MaxColsAtCompileTime = size_at_compile_time<traits<Lhs>::MaxColsAtCompileTime, traits<Rhs>::MaxColsAtCompileTime>::ret,
+
+    EvalToRowMajor = (int(LhsFlags) & int(RhsFlags) & RowMajorBit),
+    RemovedBits = ~(EvalToRowMajor ? 0 : RowMajorBit),
+
+    Flags = ((int(LhsFlags) | int(RhsFlags)) & HereditaryBits & RemovedBits)
+          | EvalBeforeNestingBit,
+    CoeffReadCost = HugeCost
+  };
+
+  typedef SparseMatrix<Scalar, 0, StorageIndex> ReturnType;
+};
+
+} // end namespace internal
+
+/*!
+ * \ingroup KroneckerProduct_Module
+ *
+ * Computes Kronecker tensor product of two dense matrices
+ *
+ * \warning If you want to replace a matrix by its Kronecker product
+ *          with some matrix, do \b NOT do this:
+ * \code
+ * A = kroneckerProduct(A,B); // bug!!! caused by aliasing effect
+ * \endcode
+ * instead, use eval() to work around this:
+ * \code
+ * A = kroneckerProduct(A,B).eval();
+ * \endcode
+ *
+ * \param a  Dense matrix a
+ * \param b  Dense matrix b
+ * \return   Kronecker tensor product of a and b
+ */
+template<typename A, typename B>
+KroneckerProduct<A,B> kroneckerProduct(const MatrixBase<A>& a, const MatrixBase<B>& b)
+{
+  return KroneckerProduct<A, B>(a.derived(), b.derived());
+}
+
+/*!
+ * \ingroup KroneckerProduct_Module
+ *
+ * Computes Kronecker tensor product of two matrices, at least one of
+ * which is sparse
+ *
+ * \warning If you want to replace a matrix by its Kronecker product
+ *          with some matrix, do \b NOT do this:
+ * \code
+ * A = kroneckerProduct(A,B); // bug!!! caused by aliasing effect
+ * \endcode
+ * instead, use eval() to work around this:
+ * \code
+ * A = kroneckerProduct(A,B).eval();
+ * \endcode
+ *
+ * \param a  Dense/sparse matrix a
+ * \param b  Dense/sparse matrix b
+ * \return   Kronecker tensor product of a and b, stored in a sparse
+ *           matrix
+ */
+template<typename A, typename B>
+KroneckerProductSparse<A,B> kroneckerProduct(const EigenBase<A>& a, const EigenBase<B>& b)
+{
+  return KroneckerProductSparse<A,B>(a.derived(), b.derived());
+}
+
+} // end namespace Eigen
+
+#endif // KRONECKER_TENSOR_PRODUCT_H
diff --git a/src/EigenUnsupported/src/LevenbergMarquardt/CopyrightMINPACK.txt b/src/EigenUnsupported/src/LevenbergMarquardt/CopyrightMINPACK.txt
new file mode 100644
index 0000000..ae7984d
--- /dev/null
+++ b/src/EigenUnsupported/src/LevenbergMarquardt/CopyrightMINPACK.txt
@@ -0,0 +1,52 @@
+Minpack Copyright Notice (1999) University of Chicago.  All rights reserved
+
+Redistribution and use in source and binary forms, with or
+without modification, are permitted provided that the
+following conditions are met:
+
+1. Redistributions of source code must retain the above
+copyright notice, this list of conditions and the following
+disclaimer.
+
+2. Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following
+disclaimer in the documentation and/or other materials
+provided with the distribution.
+
+3. The end-user documentation included with the
+redistribution, if any, must include the following
+acknowledgment:
+
+   "This product includes software developed by the
+   University of Chicago, as Operator of Argonne National
+   Laboratory.
+
+Alternately, this acknowledgment may appear in the software
+itself, if and wherever such third-party acknowledgments
+normally appear.
+
+4. WARRANTY DISCLAIMER. THE SOFTWARE IS SUPPLIED "AS IS"
+WITHOUT WARRANTY OF ANY KIND. THE COPYRIGHT HOLDER, THE
+UNITED STATES, THE UNITED STATES DEPARTMENT OF ENERGY, AND
+THEIR EMPLOYEES: (1) DISCLAIM ANY WARRANTIES, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO ANY IMPLIED WARRANTIES
+OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE
+OR NON-INFRINGEMENT, (2) DO NOT ASSUME ANY LEGAL LIABILITY
+OR RESPONSIBILITY FOR THE ACCURACY, COMPLETENESS, OR
+USEFULNESS OF THE SOFTWARE, (3) DO NOT REPRESENT THAT USE OF
+THE SOFTWARE WOULD NOT INFRINGE PRIVATELY OWNED RIGHTS, (4)
+DO NOT WARRANT THAT THE SOFTWARE WILL FUNCTION
+UNINTERRUPTED, THAT IT IS ERROR-FREE OR THAT ANY ERRORS WILL
+BE CORRECTED.
+
+5. LIMITATION OF LIABILITY. IN NO EVENT WILL THE COPYRIGHT
+HOLDER, THE UNITED STATES, THE UNITED STATES DEPARTMENT OF
+ENERGY, OR THEIR EMPLOYEES: BE LIABLE FOR ANY INDIRECT,
+INCIDENTAL, CONSEQUENTIAL, SPECIAL OR PUNITIVE DAMAGES OF
+ANY KIND OR NATURE, INCLUDING BUT NOT LIMITED TO LOSS OF
+PROFITS OR LOSS OF DATA, FOR ANY REASON WHATSOEVER, WHETHER
+SUCH LIABILITY IS ASSERTED ON THE BASIS OF CONTRACT, TORT
+(INCLUDING NEGLIGENCE OR STRICT LIABILITY), OR OTHERWISE,
+EVEN IF ANY OF SAID PARTIES HAS BEEN WARNED OF THE
+POSSIBILITY OF SUCH LOSS OR DAMAGES.
+
diff --git a/src/EigenUnsupported/src/LevenbergMarquardt/LMcovar.h b/src/EigenUnsupported/src/LevenbergMarquardt/LMcovar.h
new file mode 100644
index 0000000..b75bea2
--- /dev/null
+++ b/src/EigenUnsupported/src/LevenbergMarquardt/LMcovar.h
@@ -0,0 +1,84 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This code initially comes from MINPACK whose original authors are:
+// Copyright Jorge More - Argonne National Laboratory
+// Copyright Burt Garbow - Argonne National Laboratory
+// Copyright Ken Hillstrom - Argonne National Laboratory
+//
+// This Source Code Form is subject to the terms of the Minpack license
+// (a BSD-like license) described in the campaigned CopyrightMINPACK.txt file.
+
+#ifndef EIGEN_LMCOVAR_H
+#define EIGEN_LMCOVAR_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template <typename Scalar>
+void covar(
+        Matrix< Scalar, Dynamic, Dynamic > &r,
+        const VectorXi& ipvt,
+        Scalar tol = std::sqrt(NumTraits<Scalar>::epsilon()) )
+{
+    using std::abs;
+    /* Local variables */
+    Index i, j, k, l, ii, jj;
+    bool sing;
+    Scalar temp;
+
+    /* Function Body */
+    const Index n = r.cols();
+    const Scalar tolr = tol * abs(r(0,0));
+    Matrix< Scalar, Dynamic, 1 > wa(n);
+    eigen_assert(ipvt.size()==n);
+
+    /* form the inverse of r in the full upper triangle of r. */
+    l = -1;
+    for (k = 0; k < n; ++k)
+        if (abs(r(k,k)) > tolr) {
+            r(k,k) = 1. / r(k,k);
+            for (j = 0; j <= k-1; ++j) {
+                temp = r(k,k) * r(j,k);
+                r(j,k) = 0.;
+                r.col(k).head(j+1) -= r.col(j).head(j+1) * temp;
+            }
+            l = k;
+        }
+
+    /* form the full upper triangle of the inverse of (r transpose)*r */
+    /* in the full upper triangle of r. */
+    for (k = 0; k <= l; ++k) {
+        for (j = 0; j <= k-1; ++j)
+            r.col(j).head(j+1) += r.col(k).head(j+1) * r(j,k);
+        r.col(k).head(k+1) *= r(k,k);
+    }
+
+    /* form the full lower triangle of the covariance matrix */
+    /* in the strict lower triangle of r and in wa. */
+    for (j = 0; j < n; ++j) {
+        jj = ipvt[j];
+        sing = j > l;
+        for (i = 0; i <= j; ++i) {
+            if (sing)
+                r(i,j) = 0.;
+            ii = ipvt[i];
+            if (ii > jj)
+                r(ii,jj) = r(i,j);
+            if (ii < jj)
+                r(jj,ii) = r(i,j);
+        }
+        wa[jj] = r(j,j);
+    }
+
+    /* symmetrize the covariance matrix in r. */
+    r.topLeftCorner(n,n).template triangularView<StrictlyUpper>() = r.topLeftCorner(n,n).transpose();
+    r.diagonal() = wa;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_LMCOVAR_H
diff --git a/src/EigenUnsupported/src/LevenbergMarquardt/LMonestep.h b/src/EigenUnsupported/src/LevenbergMarquardt/LMonestep.h
new file mode 100644
index 0000000..25b32ec
--- /dev/null
+++ b/src/EigenUnsupported/src/LevenbergMarquardt/LMonestep.h
@@ -0,0 +1,202 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+//
+// This code initially comes from MINPACK whose original authors are:
+// Copyright Jorge More - Argonne National Laboratory
+// Copyright Burt Garbow - Argonne National Laboratory
+// Copyright Ken Hillstrom - Argonne National Laboratory
+//
+// This Source Code Form is subject to the terms of the Minpack license
+// (a BSD-like license) described in the campaigned CopyrightMINPACK.txt file.
+
+#ifndef EIGEN_LMONESTEP_H
+#define EIGEN_LMONESTEP_H
+
+namespace Eigen {
+
+template<typename FunctorType>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType>::minimizeOneStep(FVectorType  &x)
+{
+  using std::abs;
+  using std::sqrt;
+  RealScalar temp, temp1,temp2; 
+  RealScalar ratio; 
+  RealScalar pnorm, xnorm, fnorm1, actred, dirder, prered;
+  eigen_assert(x.size()==n); // check the caller is not cheating us
+
+  temp = 0.0; xnorm = 0.0;
+  /* calculate the jacobian matrix. */
+  Index df_ret = m_functor.df(x, m_fjac);
+  if (df_ret<0)
+      return LevenbergMarquardtSpace::UserAsked;
+  if (df_ret>0)
+      // numerical diff, we evaluated the function df_ret times
+      m_nfev += df_ret;
+  else m_njev++;
+
+  /* compute the qr factorization of the jacobian. */
+  for (int j = 0; j < x.size(); ++j)
+    m_wa2(j) = m_fjac.col(j).blueNorm();
+  QRSolver qrfac(m_fjac);
+  if(qrfac.info() != Success) {
+    m_info = NumericalIssue;
+    return LevenbergMarquardtSpace::ImproperInputParameters;
+  }
+  // Make a copy of the first factor with the associated permutation
+  m_rfactor = qrfac.matrixR();
+  m_permutation = (qrfac.colsPermutation());
+
+  /* on the first iteration and if external scaling is not used, scale according */
+  /* to the norms of the columns of the initial jacobian. */
+  if (m_iter == 1) {
+      if (!m_useExternalScaling)
+          for (Index j = 0; j < n; ++j)
+              m_diag[j] = (m_wa2[j]==0.)? 1. : m_wa2[j];
+
+      /* on the first iteration, calculate the norm of the scaled x */
+      /* and initialize the step bound m_delta. */
+      xnorm = m_diag.cwiseProduct(x).stableNorm();
+      m_delta = m_factor * xnorm;
+      if (m_delta == 0.)
+          m_delta = m_factor;
+  }
+
+  /* form (q transpose)*m_fvec and store the first n components in */
+  /* m_qtf. */
+  m_wa4 = m_fvec;
+  m_wa4 = qrfac.matrixQ().adjoint() * m_fvec; 
+  m_qtf = m_wa4.head(n);
+
+  /* compute the norm of the scaled gradient. */
+  m_gnorm = 0.;
+  if (m_fnorm != 0.)
+      for (Index j = 0; j < n; ++j)
+          if (m_wa2[m_permutation.indices()[j]] != 0.)
+              m_gnorm = (std::max)(m_gnorm, abs( m_rfactor.col(j).head(j+1).dot(m_qtf.head(j+1)/m_fnorm) / m_wa2[m_permutation.indices()[j]]));
+
+  /* test for convergence of the gradient norm. */
+  if (m_gnorm <= m_gtol) {
+    m_info = Success;
+    return LevenbergMarquardtSpace::CosinusTooSmall;
+  }
+
+  /* rescale if necessary. */
+  if (!m_useExternalScaling)
+      m_diag = m_diag.cwiseMax(m_wa2);
+
+  do {
+    /* determine the levenberg-marquardt parameter. */
+    internal::lmpar2(qrfac, m_diag, m_qtf, m_delta, m_par, m_wa1);
+
+    /* store the direction p and x + p. calculate the norm of p. */
+    m_wa1 = -m_wa1;
+    m_wa2 = x + m_wa1;
+    pnorm = m_diag.cwiseProduct(m_wa1).stableNorm();
+
+    /* on the first iteration, adjust the initial step bound. */
+    if (m_iter == 1)
+        m_delta = (std::min)(m_delta,pnorm);
+
+    /* evaluate the function at x + p and calculate its norm. */
+    if ( m_functor(m_wa2, m_wa4) < 0)
+        return LevenbergMarquardtSpace::UserAsked;
+    ++m_nfev;
+    fnorm1 = m_wa4.stableNorm();
+
+    /* compute the scaled actual reduction. */
+    actred = -1.;
+    if (Scalar(.1) * fnorm1 < m_fnorm)
+        actred = 1. - numext::abs2(fnorm1 / m_fnorm);
+
+    /* compute the scaled predicted reduction and */
+    /* the scaled directional derivative. */
+    m_wa3 = m_rfactor.template triangularView<Upper>() * (m_permutation.inverse() *m_wa1);
+    temp1 = numext::abs2(m_wa3.stableNorm() / m_fnorm);
+    temp2 = numext::abs2(sqrt(m_par) * pnorm / m_fnorm);
+    prered = temp1 + temp2 / Scalar(.5);
+    dirder = -(temp1 + temp2);
+
+    /* compute the ratio of the actual to the predicted */
+    /* reduction. */
+    ratio = 0.;
+    if (prered != 0.)
+        ratio = actred / prered;
+
+    /* update the step bound. */
+    if (ratio <= Scalar(.25)) {
+        if (actred >= 0.)
+            temp = RealScalar(.5);
+        if (actred < 0.)
+            temp = RealScalar(.5) * dirder / (dirder + RealScalar(.5) * actred);
+        if (RealScalar(.1) * fnorm1 >= m_fnorm || temp < RealScalar(.1))
+            temp = Scalar(.1);
+        /* Computing MIN */
+        m_delta = temp * (std::min)(m_delta, pnorm / RealScalar(.1));
+        m_par /= temp;
+    } else if (!(m_par != 0. && ratio < RealScalar(.75))) {
+        m_delta = pnorm / RealScalar(.5);
+        m_par = RealScalar(.5) * m_par;
+    }
+
+    /* test for successful iteration. */
+    if (ratio >= RealScalar(1e-4)) {
+        /* successful iteration. update x, m_fvec, and their norms. */
+        x = m_wa2;
+        m_wa2 = m_diag.cwiseProduct(x);
+        m_fvec = m_wa4;
+        xnorm = m_wa2.stableNorm();
+        m_fnorm = fnorm1;
+        ++m_iter;
+    }
+
+    /* tests for convergence. */
+    if (abs(actred) <= m_ftol && prered <= m_ftol && Scalar(.5) * ratio <= 1. && m_delta <= m_xtol * xnorm)
+    {
+       m_info = Success;
+      return LevenbergMarquardtSpace::RelativeErrorAndReductionTooSmall;
+    }
+    if (abs(actred) <= m_ftol && prered <= m_ftol && Scalar(.5) * ratio <= 1.) 
+    {
+      m_info = Success;
+      return LevenbergMarquardtSpace::RelativeReductionTooSmall;
+    }
+    if (m_delta <= m_xtol * xnorm)
+    {
+      m_info = Success;
+      return LevenbergMarquardtSpace::RelativeErrorTooSmall;
+    }
+
+    /* tests for termination and stringent tolerances. */
+    if (m_nfev >= m_maxfev) 
+    {
+      m_info = NoConvergence;
+      return LevenbergMarquardtSpace::TooManyFunctionEvaluation;
+    }
+    if (abs(actred) <= NumTraits<Scalar>::epsilon() && prered <= NumTraits<Scalar>::epsilon() && Scalar(.5) * ratio <= 1.)
+    {
+      m_info = Success;
+      return LevenbergMarquardtSpace::FtolTooSmall;
+    }
+    if (m_delta <= NumTraits<Scalar>::epsilon() * xnorm) 
+    {
+      m_info = Success;
+      return LevenbergMarquardtSpace::XtolTooSmall;
+    }
+    if (m_gnorm <= NumTraits<Scalar>::epsilon())
+    {
+      m_info = Success;
+      return LevenbergMarquardtSpace::GtolTooSmall;
+    }
+
+  } while (ratio < Scalar(1e-4));
+
+  return LevenbergMarquardtSpace::Running;
+}
+
+  
+} // end namespace Eigen
+
+#endif // EIGEN_LMONESTEP_H
diff --git a/src/EigenUnsupported/src/LevenbergMarquardt/LMpar.h b/src/EigenUnsupported/src/LevenbergMarquardt/LMpar.h
new file mode 100644
index 0000000..9a48365
--- /dev/null
+++ b/src/EigenUnsupported/src/LevenbergMarquardt/LMpar.h
@@ -0,0 +1,160 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This code initially comes from MINPACK whose original authors are:
+// Copyright Jorge More - Argonne National Laboratory
+// Copyright Burt Garbow - Argonne National Laboratory
+// Copyright Ken Hillstrom - Argonne National Laboratory
+//
+// This Source Code Form is subject to the terms of the Minpack license
+// (a BSD-like license) described in the campaigned CopyrightMINPACK.txt file.
+
+#ifndef EIGEN_LMPAR_H
+#define EIGEN_LMPAR_H
+
+namespace Eigen {
+
+namespace internal {
+  
+  template <typename QRSolver, typename VectorType>
+    void lmpar2(
+    const QRSolver &qr,
+    const VectorType  &diag,
+    const VectorType  &qtb,
+    typename VectorType::Scalar m_delta,
+    typename VectorType::Scalar &par,
+    VectorType  &x)
+
+  {
+    using std::sqrt;
+    using std::abs;
+    typedef typename QRSolver::MatrixType MatrixType;
+    typedef typename QRSolver::Scalar Scalar;
+//    typedef typename QRSolver::StorageIndex StorageIndex;
+
+    /* Local variables */
+    Index j;
+    Scalar fp;
+    Scalar parc, parl;
+    Index iter;
+    Scalar temp, paru;
+    Scalar gnorm;
+    Scalar dxnorm;
+    
+    // Make a copy of the triangular factor. 
+    // This copy is modified during call the qrsolv
+    MatrixType s;
+    s = qr.matrixR();
+
+    /* Function Body */
+    const Scalar dwarf = (std::numeric_limits<Scalar>::min)();
+    const Index n = qr.matrixR().cols();
+    eigen_assert(n==diag.size());
+    eigen_assert(n==qtb.size());
+
+    VectorType  wa1, wa2;
+
+    /* compute and store in x the gauss-newton direction. if the */
+    /* jacobian is rank-deficient, obtain a least squares solution. */
+
+    //    const Index rank = qr.nonzeroPivots(); // exactly double(0.)
+    const Index rank = qr.rank(); // use a threshold
+    wa1 = qtb;
+    wa1.tail(n-rank).setZero();
+    //FIXME There is no solve in place for sparse triangularView
+    wa1.head(rank) = s.topLeftCorner(rank,rank).template triangularView<Upper>().solve(qtb.head(rank));
+
+    x = qr.colsPermutation()*wa1;
+
+    /* initialize the iteration counter. */
+    /* evaluate the function at the origin, and test */
+    /* for acceptance of the gauss-newton direction. */
+    iter = 0;
+    wa2 = diag.cwiseProduct(x);
+    dxnorm = wa2.blueNorm();
+    fp = dxnorm - m_delta;
+    if (fp <= Scalar(0.1) * m_delta) {
+      par = 0;
+      return;
+    }
+
+    /* if the jacobian is not rank deficient, the newton */
+    /* step provides a lower bound, parl, for the zero of */
+    /* the function. otherwise set this bound to zero. */
+    parl = 0.;
+    if (rank==n) {
+      wa1 = qr.colsPermutation().inverse() *  diag.cwiseProduct(wa2)/dxnorm;
+      s.topLeftCorner(n,n).transpose().template triangularView<Lower>().solveInPlace(wa1);
+      temp = wa1.blueNorm();
+      parl = fp / m_delta / temp / temp;
+    }
+
+    /* calculate an upper bound, paru, for the zero of the function. */
+    for (j = 0; j < n; ++j)
+      wa1[j] = s.col(j).head(j+1).dot(qtb.head(j+1)) / diag[qr.colsPermutation().indices()(j)];
+
+    gnorm = wa1.stableNorm();
+    paru = gnorm / m_delta;
+    if (paru == 0.)
+      paru = dwarf / (std::min)(m_delta,Scalar(0.1));
+
+    /* if the input par lies outside of the interval (parl,paru), */
+    /* set par to the closer endpoint. */
+    par = (std::max)(par,parl);
+    par = (std::min)(par,paru);
+    if (par == 0.)
+      par = gnorm / dxnorm;
+
+    /* beginning of an iteration. */
+    while (true) {
+      ++iter;
+
+      /* evaluate the function at the current value of par. */
+      if (par == 0.)
+        par = (std::max)(dwarf,Scalar(.001) * paru); /* Computing MAX */
+      wa1 = sqrt(par)* diag;
+
+      VectorType sdiag(n);
+      lmqrsolv(s, qr.colsPermutation(), wa1, qtb, x, sdiag);
+
+      wa2 = diag.cwiseProduct(x);
+      dxnorm = wa2.blueNorm();
+      temp = fp;
+      fp = dxnorm - m_delta;
+
+      /* if the function is small enough, accept the current value */
+      /* of par. also test for the exceptional cases where parl */
+      /* is zero or the number of iterations has reached 10. */
+      if (abs(fp) <= Scalar(0.1) * m_delta || (parl == 0. && fp <= temp && temp < 0.) || iter == 10)
+        break;
+
+      /* compute the newton correction. */
+      wa1 = qr.colsPermutation().inverse() * diag.cwiseProduct(wa2/dxnorm);
+      // we could almost use this here, but the diagonal is outside qr, in sdiag[]
+      for (j = 0; j < n; ++j) {
+        wa1[j] /= sdiag[j];
+        temp = wa1[j];
+        for (Index i = j+1; i < n; ++i)
+          wa1[i] -= s.coeff(i,j) * temp;
+      }
+      temp = wa1.blueNorm();
+      parc = fp / m_delta / temp / temp;
+
+      /* depending on the sign of the function, update parl or paru. */
+      if (fp > 0.)
+        parl = (std::max)(parl,par);
+      if (fp < 0.)
+        paru = (std::min)(paru,par);
+
+      /* compute an improved estimate for par. */
+      par = (std::max)(parl,par+parc);
+    }
+    if (iter == 0)
+      par = 0.;
+    return;
+  }
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_LMPAR_H
diff --git a/src/EigenUnsupported/src/LevenbergMarquardt/LMqrsolv.h b/src/EigenUnsupported/src/LevenbergMarquardt/LMqrsolv.h
new file mode 100644
index 0000000..1234858
--- /dev/null
+++ b/src/EigenUnsupported/src/LevenbergMarquardt/LMqrsolv.h
@@ -0,0 +1,188 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+// Copyright (C) 2012 Desire Nuentsa <desire.nuentsa_wakam@inria.fr>
+//
+// This code initially comes from MINPACK whose original authors are:
+// Copyright Jorge More - Argonne National Laboratory
+// Copyright Burt Garbow - Argonne National Laboratory
+// Copyright Ken Hillstrom - Argonne National Laboratory
+//
+// This Source Code Form is subject to the terms of the Minpack license
+// (a BSD-like license) described in the campaigned CopyrightMINPACK.txt file.
+
+#ifndef EIGEN_LMQRSOLV_H
+#define EIGEN_LMQRSOLV_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template <typename Scalar,int Rows, int Cols, typename PermIndex>
+void lmqrsolv(
+  Matrix<Scalar,Rows,Cols> &s,
+  const PermutationMatrix<Dynamic,Dynamic,PermIndex> &iPerm,
+  const Matrix<Scalar,Dynamic,1> &diag,
+  const Matrix<Scalar,Dynamic,1> &qtb,
+  Matrix<Scalar,Dynamic,1> &x,
+  Matrix<Scalar,Dynamic,1> &sdiag)
+{
+    /* Local variables */
+    Index i, j, k;
+    Scalar temp;
+    Index n = s.cols();
+    Matrix<Scalar,Dynamic,1>  wa(n);
+    JacobiRotation<Scalar> givens;
+
+    /* Function Body */
+    // the following will only change the lower triangular part of s, including
+    // the diagonal, though the diagonal is restored afterward
+
+    /*     copy r and (q transpose)*b to preserve input and initialize s. */
+    /*     in particular, save the diagonal elements of r in x. */
+    x = s.diagonal();
+    wa = qtb;
+    
+   
+    s.topLeftCorner(n,n).template triangularView<StrictlyLower>() = s.topLeftCorner(n,n).transpose();
+    /*     eliminate the diagonal matrix d using a givens rotation. */
+    for (j = 0; j < n; ++j) {
+
+        /*        prepare the row of d to be eliminated, locating the */
+        /*        diagonal element using p from the qr factorization. */
+        const PermIndex l = iPerm.indices()(j);
+        if (diag[l] == 0.)
+            break;
+        sdiag.tail(n-j).setZero();
+        sdiag[j] = diag[l];
+
+        /*        the transformations to eliminate the row of d */
+        /*        modify only a single element of (q transpose)*b */
+        /*        beyond the first n, which is initially zero. */
+        Scalar qtbpj = 0.;
+        for (k = j; k < n; ++k) {
+            /*           determine a givens rotation which eliminates the */
+            /*           appropriate element in the current row of d. */
+            givens.makeGivens(-s(k,k), sdiag[k]);
+
+            /*           compute the modified diagonal element of r and */
+            /*           the modified element of ((q transpose)*b,0). */
+            s(k,k) = givens.c() * s(k,k) + givens.s() * sdiag[k];
+            temp = givens.c() * wa[k] + givens.s() * qtbpj;
+            qtbpj = -givens.s() * wa[k] + givens.c() * qtbpj;
+            wa[k] = temp;
+
+            /*           accumulate the transformation in the row of s. */
+            for (i = k+1; i<n; ++i) {
+                temp = givens.c() * s(i,k) + givens.s() * sdiag[i];
+                sdiag[i] = -givens.s() * s(i,k) + givens.c() * sdiag[i];
+                s(i,k) = temp;
+            }
+        }
+    }
+  
+    /*     solve the triangular system for z. if the system is */
+    /*     singular, then obtain a least squares solution. */
+    Index nsing;
+    for(nsing=0; nsing<n && sdiag[nsing]!=0; nsing++) {}
+
+    wa.tail(n-nsing).setZero();
+    s.topLeftCorner(nsing, nsing).transpose().template triangularView<Upper>().solveInPlace(wa.head(nsing));
+  
+    // restore
+    sdiag = s.diagonal();
+    s.diagonal() = x;
+
+    /* permute the components of z back to components of x. */
+    x = iPerm * wa; 
+}
+
+template <typename Scalar, int _Options, typename Index>
+void lmqrsolv(
+  SparseMatrix<Scalar,_Options,Index> &s,
+  const PermutationMatrix<Dynamic,Dynamic> &iPerm,
+  const Matrix<Scalar,Dynamic,1> &diag,
+  const Matrix<Scalar,Dynamic,1> &qtb,
+  Matrix<Scalar,Dynamic,1> &x,
+  Matrix<Scalar,Dynamic,1> &sdiag)
+{
+  /* Local variables */
+  typedef SparseMatrix<Scalar,RowMajor,Index> FactorType;
+    Index i, j, k, l;
+    Scalar temp;
+    Index n = s.cols();
+    Matrix<Scalar,Dynamic,1>  wa(n);
+    JacobiRotation<Scalar> givens;
+
+    /* Function Body */
+    // the following will only change the lower triangular part of s, including
+    // the diagonal, though the diagonal is restored afterward
+
+    /*     copy r and (q transpose)*b to preserve input and initialize R. */
+    wa = qtb;
+    FactorType R(s);
+    // Eliminate the diagonal matrix d using a givens rotation
+    for (j = 0; j < n; ++j)
+    {
+      // Prepare the row of d to be eliminated, locating the 
+      // diagonal element using p from the qr factorization
+      l = iPerm.indices()(j);
+      if (diag(l) == Scalar(0)) 
+        break; 
+      sdiag.tail(n-j).setZero();
+      sdiag[j] = diag[l];
+      // the transformations to eliminate the row of d
+      // modify only a single element of (q transpose)*b
+      // beyond the first n, which is initially zero. 
+      
+      Scalar qtbpj = 0; 
+      // Browse the nonzero elements of row j of the upper triangular s
+      for (k = j; k < n; ++k)
+      {
+        typename FactorType::InnerIterator itk(R,k);
+        for (; itk; ++itk){
+          if (itk.index() < k) continue;
+          else break;
+        }
+        //At this point, we have the diagonal element R(k,k)
+        // Determine a givens rotation which eliminates 
+        // the appropriate element in the current row of d
+        givens.makeGivens(-itk.value(), sdiag(k));
+        
+        // Compute the modified diagonal element of r and 
+        // the modified element of ((q transpose)*b,0).
+        itk.valueRef() = givens.c() * itk.value() + givens.s() * sdiag(k);
+        temp = givens.c() * wa(k) + givens.s() * qtbpj; 
+        qtbpj = -givens.s() * wa(k) + givens.c() * qtbpj;
+        wa(k) = temp;
+        
+        // Accumulate the transformation in the remaining k row/column of R
+        for (++itk; itk; ++itk)
+        {
+          i = itk.index();
+          temp = givens.c() *  itk.value() + givens.s() * sdiag(i);
+          sdiag(i) = -givens.s() * itk.value() + givens.c() * sdiag(i);
+          itk.valueRef() = temp;
+        }
+      }
+    }
+    
+    // Solve the triangular system for z. If the system is 
+    // singular, then obtain a least squares solution
+    Index nsing;
+    for(nsing = 0; nsing<n && sdiag(nsing) !=0; nsing++) {}
+    
+    wa.tail(n-nsing).setZero();
+//     x = wa; 
+    wa.head(nsing) = R.topLeftCorner(nsing,nsing).template triangularView<Upper>().solve/*InPlace*/(wa.head(nsing));
+    
+    sdiag = R.diagonal();
+    // Permute the components of z back to components of x
+    x = iPerm * wa; 
+}
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_LMQRSOLV_H
diff --git a/src/EigenUnsupported/src/LevenbergMarquardt/LevenbergMarquardt.h b/src/EigenUnsupported/src/LevenbergMarquardt/LevenbergMarquardt.h
new file mode 100644
index 0000000..62561da
--- /dev/null
+++ b/src/EigenUnsupported/src/LevenbergMarquardt/LevenbergMarquardt.h
@@ -0,0 +1,396 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+// Copyright (C) 2012 Desire Nuentsa <desire.nuentsa_wakam@inria.fr>
+//
+// The algorithm of this class initially comes from MINPACK whose original authors are:
+// Copyright Jorge More - Argonne National Laboratory
+// Copyright Burt Garbow - Argonne National Laboratory
+// Copyright Ken Hillstrom - Argonne National Laboratory
+//
+// This Source Code Form is subject to the terms of the Minpack license
+// (a BSD-like license) described in the campaigned CopyrightMINPACK.txt file.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_LEVENBERGMARQUARDT_H
+#define EIGEN_LEVENBERGMARQUARDT_H
+
+
+namespace Eigen {
+namespace LevenbergMarquardtSpace {
+    enum Status {
+        NotStarted = -2,
+        Running = -1,
+        ImproperInputParameters = 0,
+        RelativeReductionTooSmall = 1,
+        RelativeErrorTooSmall = 2,
+        RelativeErrorAndReductionTooSmall = 3,
+        CosinusTooSmall = 4,
+        TooManyFunctionEvaluation = 5,
+        FtolTooSmall = 6,
+        XtolTooSmall = 7,
+        GtolTooSmall = 8,
+        UserAsked = 9
+    };
+}
+
+template <typename _Scalar, int NX=Dynamic, int NY=Dynamic>
+struct DenseFunctor
+{
+  typedef _Scalar Scalar;
+  enum {
+    InputsAtCompileTime = NX,
+    ValuesAtCompileTime = NY
+  };
+  typedef Matrix<Scalar,InputsAtCompileTime,1> InputType;
+  typedef Matrix<Scalar,ValuesAtCompileTime,1> ValueType;
+  typedef Matrix<Scalar,ValuesAtCompileTime,InputsAtCompileTime> JacobianType;
+  typedef ColPivHouseholderQR<JacobianType> QRSolver;
+  const int m_inputs, m_values;
+
+  DenseFunctor() : m_inputs(InputsAtCompileTime), m_values(ValuesAtCompileTime) {}
+  DenseFunctor(int inputs, int values) : m_inputs(inputs), m_values(values) {}
+
+  int inputs() const { return m_inputs; }
+  int values() const { return m_values; }
+
+  //int operator()(const InputType &x, ValueType& fvec) { }
+  // should be defined in derived classes
+  
+  //int df(const InputType &x, JacobianType& fjac) { }
+  // should be defined in derived classes
+};
+
+template <typename _Scalar, typename _Index>
+struct SparseFunctor
+{
+  typedef _Scalar Scalar;
+  typedef _Index Index;
+  typedef Matrix<Scalar,Dynamic,1> InputType;
+  typedef Matrix<Scalar,Dynamic,1> ValueType;
+  typedef SparseMatrix<Scalar, ColMajor, Index> JacobianType;
+  typedef SparseQR<JacobianType, COLAMDOrdering<int> > QRSolver;
+  enum {
+    InputsAtCompileTime = Dynamic,
+    ValuesAtCompileTime = Dynamic
+  };
+  
+  SparseFunctor(int inputs, int values) : m_inputs(inputs), m_values(values) {}
+
+  int inputs() const { return m_inputs; }
+  int values() const { return m_values; }
+  
+  const int m_inputs, m_values;
+  //int operator()(const InputType &x, ValueType& fvec) { }
+  // to be defined in the functor
+  
+  //int df(const InputType &x, JacobianType& fjac) { }
+  // to be defined in the functor if no automatic differentiation
+  
+};
+namespace internal {
+template <typename QRSolver, typename VectorType>
+void lmpar2(const QRSolver &qr, const VectorType  &diag, const VectorType  &qtb,
+	    typename VectorType::Scalar m_delta, typename VectorType::Scalar &par,
+	    VectorType  &x);
+    }
+/**
+  * \ingroup NonLinearOptimization_Module
+  * \brief Performs non linear optimization over a non-linear function,
+  * using a variant of the Levenberg Marquardt algorithm.
+  *
+  * Check wikipedia for more information.
+  * http://en.wikipedia.org/wiki/Levenberg%E2%80%93Marquardt_algorithm
+  */
+template<typename _FunctorType>
+class LevenbergMarquardt : internal::no_assignment_operator
+{
+  public:
+    typedef _FunctorType FunctorType;
+    typedef typename FunctorType::QRSolver QRSolver;
+    typedef typename FunctorType::JacobianType JacobianType;
+    typedef typename JacobianType::Scalar Scalar;
+    typedef typename JacobianType::RealScalar RealScalar; 
+    typedef typename QRSolver::StorageIndex PermIndex;
+    typedef Matrix<Scalar,Dynamic,1> FVectorType;
+    typedef PermutationMatrix<Dynamic,Dynamic,int> PermutationType;
+  public:
+    LevenbergMarquardt(FunctorType& functor) 
+    : m_functor(functor),m_nfev(0),m_njev(0),m_fnorm(0.0),m_gnorm(0),
+      m_isInitialized(false),m_info(InvalidInput)
+    {
+      resetParameters();
+      m_useExternalScaling=false; 
+    }
+    
+    LevenbergMarquardtSpace::Status minimize(FVectorType &x);
+    LevenbergMarquardtSpace::Status minimizeInit(FVectorType &x);
+    LevenbergMarquardtSpace::Status minimizeOneStep(FVectorType &x);
+    LevenbergMarquardtSpace::Status lmder1(
+      FVectorType  &x, 
+      const Scalar tol = std::sqrt(NumTraits<Scalar>::epsilon())
+    );
+    static LevenbergMarquardtSpace::Status lmdif1(
+            FunctorType &functor,
+            FVectorType  &x,
+            Index *nfev,
+            const Scalar tol = std::sqrt(NumTraits<Scalar>::epsilon())
+            );
+    
+    /** Sets the default parameters */
+    void resetParameters() 
+    {
+      using std::sqrt;        
+
+      m_factor = 100.; 
+      m_maxfev = 400; 
+      m_ftol = sqrt(NumTraits<RealScalar>::epsilon());
+      m_xtol = sqrt(NumTraits<RealScalar>::epsilon());
+      m_gtol = 0. ; 
+      m_epsfcn = 0. ;
+    }
+    
+    /** Sets the tolerance for the norm of the solution vector*/
+    void setXtol(RealScalar xtol) { m_xtol = xtol; }
+    
+    /** Sets the tolerance for the norm of the vector function*/
+    void setFtol(RealScalar ftol) { m_ftol = ftol; }
+    
+    /** Sets the tolerance for the norm of the gradient of the error vector*/
+    void setGtol(RealScalar gtol) { m_gtol = gtol; }
+    
+    /** Sets the step bound for the diagonal shift */
+    void setFactor(RealScalar factor) { m_factor = factor; }    
+    
+    /** Sets the error precision  */
+    void setEpsilon (RealScalar epsfcn) { m_epsfcn = epsfcn; }
+    
+    /** Sets the maximum number of function evaluation */
+    void setMaxfev(Index maxfev) {m_maxfev = maxfev; }
+    
+    /** Use an external Scaling. If set to true, pass a nonzero diagonal to diag() */
+    void setExternalScaling(bool value) {m_useExternalScaling  = value; }
+    
+    /** \returns the tolerance for the norm of the solution vector */
+    RealScalar xtol() const {return m_xtol; }
+    
+    /** \returns the tolerance for the norm of the vector function */
+    RealScalar ftol() const {return m_ftol; }
+    
+    /** \returns the tolerance for the norm of the gradient of the error vector */
+    RealScalar gtol() const {return m_gtol; }
+    
+    /** \returns the step bound for the diagonal shift */
+    RealScalar factor() const {return m_factor; }
+    
+    /** \returns the error precision */
+    RealScalar epsilon() const {return m_epsfcn; }
+    
+    /** \returns the maximum number of function evaluation */
+    Index maxfev() const {return m_maxfev; }
+    
+    /** \returns a reference to the diagonal of the jacobian */
+    FVectorType& diag() {return m_diag; }
+    
+    /** \returns the number of iterations performed */
+    Index iterations() { return m_iter; }
+    
+    /** \returns the number of functions evaluation */
+    Index nfev() { return m_nfev; }
+    
+    /** \returns the number of jacobian evaluation */
+    Index njev() { return m_njev; }
+    
+    /** \returns the norm of current vector function */
+    RealScalar fnorm() {return m_fnorm; }
+    
+    /** \returns the norm of the gradient of the error */
+    RealScalar gnorm() {return m_gnorm; }
+    
+    /** \returns the LevenbergMarquardt parameter */
+    RealScalar lm_param(void) { return m_par; }
+    
+    /** \returns a reference to the  current vector function 
+     */
+    FVectorType& fvec() {return m_fvec; }
+    
+    /** \returns a reference to the matrix where the current Jacobian matrix is stored
+     */
+    JacobianType& jacobian() {return m_fjac; }
+    
+    /** \returns a reference to the triangular matrix R from the QR of the jacobian matrix.
+     * \sa jacobian()
+     */
+    JacobianType& matrixR() {return m_rfactor; }
+    
+    /** the permutation used in the QR factorization
+     */
+    PermutationType permutation() {return m_permutation; }
+    
+    /** 
+     * \brief Reports whether the minimization was successful
+     * \returns \c Success if the minimization was successful,
+     *         \c NumericalIssue if a numerical problem arises during the 
+     *          minimization process, for example during the QR factorization
+     *         \c NoConvergence if the minimization did not converge after 
+     *          the maximum number of function evaluation allowed
+     *          \c InvalidInput if the input matrix is invalid
+     */
+    ComputationInfo info() const
+    {
+      
+      return m_info;
+    }
+  private:
+    JacobianType m_fjac; 
+    JacobianType m_rfactor; // The triangular matrix R from the QR of the jacobian matrix m_fjac
+    FunctorType &m_functor;
+    FVectorType m_fvec, m_qtf, m_diag; 
+    Index n;
+    Index m; 
+    Index m_nfev;
+    Index m_njev; 
+    RealScalar m_fnorm; // Norm of the current vector function
+    RealScalar m_gnorm; //Norm of the gradient of the error 
+    RealScalar m_factor; //
+    Index m_maxfev; // Maximum number of function evaluation
+    RealScalar m_ftol; //Tolerance in the norm of the vector function
+    RealScalar m_xtol; // 
+    RealScalar m_gtol; //tolerance of the norm of the error gradient
+    RealScalar m_epsfcn; //
+    Index m_iter; // Number of iterations performed
+    RealScalar m_delta;
+    bool m_useExternalScaling;
+    PermutationType m_permutation;
+    FVectorType m_wa1, m_wa2, m_wa3, m_wa4; //Temporary vectors
+    RealScalar m_par;
+    bool m_isInitialized; // Check whether the minimization step has been called
+    ComputationInfo m_info; 
+};
+
+template<typename FunctorType>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType>::minimize(FVectorType  &x)
+{
+    LevenbergMarquardtSpace::Status status = minimizeInit(x);
+    if (status==LevenbergMarquardtSpace::ImproperInputParameters) {
+      m_isInitialized = true;
+      return status;
+    }
+    do {
+//       std::cout << " uv " << x.transpose() << "\n";
+        status = minimizeOneStep(x);
+    } while (status==LevenbergMarquardtSpace::Running);
+     m_isInitialized = true;
+     return status;
+}
+
+template<typename FunctorType>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType>::minimizeInit(FVectorType  &x)
+{
+    n = x.size();
+    m = m_functor.values();
+
+    m_wa1.resize(n); m_wa2.resize(n); m_wa3.resize(n);
+    m_wa4.resize(m);
+    m_fvec.resize(m);
+    //FIXME Sparse Case : Allocate space for the jacobian
+    m_fjac.resize(m, n);
+//     m_fjac.reserve(VectorXi::Constant(n,5)); // FIXME Find a better alternative
+    if (!m_useExternalScaling)
+        m_diag.resize(n);
+    eigen_assert( (!m_useExternalScaling || m_diag.size()==n) && "When m_useExternalScaling is set, the caller must provide a valid 'm_diag'");
+    m_qtf.resize(n);
+
+    /* Function Body */
+    m_nfev = 0;
+    m_njev = 0;
+
+    /*     check the input parameters for errors. */
+    if (n <= 0 || m < n || m_ftol < 0. || m_xtol < 0. || m_gtol < 0. || m_maxfev <= 0 || m_factor <= 0.){
+      m_info = InvalidInput;
+      return LevenbergMarquardtSpace::ImproperInputParameters;
+    }
+
+    if (m_useExternalScaling)
+        for (Index j = 0; j < n; ++j)
+            if (m_diag[j] <= 0.) 
+            {
+              m_info = InvalidInput;
+              return LevenbergMarquardtSpace::ImproperInputParameters;
+            }
+
+    /*     evaluate the function at the starting point */
+    /*     and calculate its norm. */
+    m_nfev = 1;
+    if ( m_functor(x, m_fvec) < 0)
+        return LevenbergMarquardtSpace::UserAsked;
+    m_fnorm = m_fvec.stableNorm();
+
+    /*     initialize levenberg-marquardt parameter and iteration counter. */
+    m_par = 0.;
+    m_iter = 1;
+
+    return LevenbergMarquardtSpace::NotStarted;
+}
+
+template<typename FunctorType>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType>::lmder1(
+        FVectorType  &x,
+        const Scalar tol
+        )
+{
+    n = x.size();
+    m = m_functor.values();
+
+    /* check the input parameters for errors. */
+    if (n <= 0 || m < n || tol < 0.)
+        return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    resetParameters();
+    m_ftol = tol;
+    m_xtol = tol;
+    m_maxfev = 100*(n+1);
+
+    return minimize(x);
+}
+
+
+template<typename FunctorType>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType>::lmdif1(
+        FunctorType &functor,
+        FVectorType  &x,
+        Index *nfev,
+        const Scalar tol
+        )
+{
+    Index n = x.size();
+    Index m = functor.values();
+
+    /* check the input parameters for errors. */
+    if (n <= 0 || m < n || tol < 0.)
+        return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    NumericalDiff<FunctorType> numDiff(functor);
+    // embedded LevenbergMarquardt
+    LevenbergMarquardt<NumericalDiff<FunctorType> > lm(numDiff);
+    lm.setFtol(tol);
+    lm.setXtol(tol);
+    lm.setMaxfev(200*(n+1));
+
+    LevenbergMarquardtSpace::Status info = LevenbergMarquardtSpace::Status(lm.minimize(x));
+    if (nfev)
+        * nfev = lm.nfev();
+    return info;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_LEVENBERGMARQUARDT_H
diff --git a/src/EigenUnsupported/src/MatrixFunctions/MatrixExponential.h b/src/EigenUnsupported/src/MatrixFunctions/MatrixExponential.h
new file mode 100644
index 0000000..02284b0
--- /dev/null
+++ b/src/EigenUnsupported/src/MatrixFunctions/MatrixExponential.h
@@ -0,0 +1,441 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009, 2010, 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
+// Copyright (C) 2011, 2013 Chen-Pang He <jdh8@ms63.hinet.net>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIX_EXPONENTIAL
+#define EIGEN_MATRIX_EXPONENTIAL
+
+#include "StemFunction.h"
+
+namespace Eigen {
+namespace internal {
+
+/** \brief Scaling operator.
+ *
+ * This struct is used by CwiseUnaryOp to scale a matrix by \f$ 2^{-s} \f$.
+ */
+template <typename RealScalar>
+struct MatrixExponentialScalingOp
+{
+  /** \brief Constructor.
+   *
+   * \param[in] squarings  The integer \f$ s \f$ in this document.
+   */
+  MatrixExponentialScalingOp(int squarings) : m_squarings(squarings) { }
+
+
+  /** \brief Scale a matrix coefficient.
+   *
+   * \param[in,out] x  The scalar to be scaled, becoming \f$ 2^{-s} x \f$.
+   */
+  inline const RealScalar operator() (const RealScalar& x) const
+  {
+    using std::ldexp;
+    return ldexp(x, -m_squarings);
+  }
+
+  typedef std::complex<RealScalar> ComplexScalar;
+
+  /** \brief Scale a matrix coefficient.
+   *
+   * \param[in,out] x  The scalar to be scaled, becoming \f$ 2^{-s} x \f$.
+   */
+  inline const ComplexScalar operator() (const ComplexScalar& x) const
+  {
+    using std::ldexp;
+    return ComplexScalar(ldexp(x.real(), -m_squarings), ldexp(x.imag(), -m_squarings));
+  }
+
+  private:
+    int m_squarings;
+};
+
+/** \brief Compute the (3,3)-Pad&eacute; approximant to the exponential.
+ *
+ *  After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Pad&eacute;
+ *  approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$.
+ */
+template <typename MatA, typename MatU, typename MatV>
+void matrix_exp_pade3(const MatA& A, MatU& U, MatV& V)
+{
+  typedef typename MatA::PlainObject MatrixType;
+  typedef typename NumTraits<typename traits<MatA>::Scalar>::Real RealScalar;
+  const RealScalar b[] = {120.L, 60.L, 12.L, 1.L};
+  const MatrixType A2 = A * A;
+  const MatrixType tmp = b[3] * A2 + b[1] * MatrixType::Identity(A.rows(), A.cols());
+  U.noalias() = A * tmp;
+  V = b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols());
+}
+
+/** \brief Compute the (5,5)-Pad&eacute; approximant to the exponential.
+ *
+ *  After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Pad&eacute;
+ *  approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$.
+ */
+template <typename MatA, typename MatU, typename MatV>
+void matrix_exp_pade5(const MatA& A, MatU& U, MatV& V)
+{
+  typedef typename MatA::PlainObject MatrixType;
+  typedef typename NumTraits<typename traits<MatrixType>::Scalar>::Real RealScalar;
+  const RealScalar b[] = {30240.L, 15120.L, 3360.L, 420.L, 30.L, 1.L};
+  const MatrixType A2 = A * A;
+  const MatrixType A4 = A2 * A2;
+  const MatrixType tmp = b[5] * A4 + b[3] * A2 + b[1] * MatrixType::Identity(A.rows(), A.cols());
+  U.noalias() = A * tmp;
+  V = b[4] * A4 + b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols());
+}
+
+/** \brief Compute the (7,7)-Pad&eacute; approximant to the exponential.
+ *
+ *  After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Pad&eacute;
+ *  approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$.
+ */
+template <typename MatA, typename MatU, typename MatV>
+void matrix_exp_pade7(const MatA& A, MatU& U, MatV& V)
+{
+  typedef typename MatA::PlainObject MatrixType;
+  typedef typename NumTraits<typename traits<MatrixType>::Scalar>::Real RealScalar;
+  const RealScalar b[] = {17297280.L, 8648640.L, 1995840.L, 277200.L, 25200.L, 1512.L, 56.L, 1.L};
+  const MatrixType A2 = A * A;
+  const MatrixType A4 = A2 * A2;
+  const MatrixType A6 = A4 * A2;
+  const MatrixType tmp = b[7] * A6 + b[5] * A4 + b[3] * A2 
+    + b[1] * MatrixType::Identity(A.rows(), A.cols());
+  U.noalias() = A * tmp;
+  V = b[6] * A6 + b[4] * A4 + b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols());
+
+}
+
+/** \brief Compute the (9,9)-Pad&eacute; approximant to the exponential.
+ *
+ *  After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Pad&eacute;
+ *  approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$.
+ */
+template <typename MatA, typename MatU, typename MatV>
+void matrix_exp_pade9(const MatA& A, MatU& U, MatV& V)
+{
+  typedef typename MatA::PlainObject MatrixType;
+  typedef typename NumTraits<typename traits<MatrixType>::Scalar>::Real RealScalar;
+  const RealScalar b[] = {17643225600.L, 8821612800.L, 2075673600.L, 302702400.L, 30270240.L,
+                          2162160.L, 110880.L, 3960.L, 90.L, 1.L};
+  const MatrixType A2 = A * A;
+  const MatrixType A4 = A2 * A2;
+  const MatrixType A6 = A4 * A2;
+  const MatrixType A8 = A6 * A2;
+  const MatrixType tmp = b[9] * A8 + b[7] * A6 + b[5] * A4 + b[3] * A2 
+    + b[1] * MatrixType::Identity(A.rows(), A.cols());
+  U.noalias() = A * tmp;
+  V = b[8] * A8 + b[6] * A6 + b[4] * A4 + b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols());
+}
+
+/** \brief Compute the (13,13)-Pad&eacute; approximant to the exponential.
+ *
+ *  After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Pad&eacute;
+ *  approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$.
+ */
+template <typename MatA, typename MatU, typename MatV>
+void matrix_exp_pade13(const MatA& A, MatU& U, MatV& V)
+{
+  typedef typename MatA::PlainObject MatrixType;
+  typedef typename NumTraits<typename traits<MatrixType>::Scalar>::Real RealScalar;
+  const RealScalar b[] = {64764752532480000.L, 32382376266240000.L, 7771770303897600.L,
+                          1187353796428800.L, 129060195264000.L, 10559470521600.L, 670442572800.L,
+                          33522128640.L, 1323241920.L, 40840800.L, 960960.L, 16380.L, 182.L, 1.L};
+  const MatrixType A2 = A * A;
+  const MatrixType A4 = A2 * A2;
+  const MatrixType A6 = A4 * A2;
+  V = b[13] * A6 + b[11] * A4 + b[9] * A2; // used for temporary storage
+  MatrixType tmp = A6 * V;
+  tmp += b[7] * A6 + b[5] * A4 + b[3] * A2 + b[1] * MatrixType::Identity(A.rows(), A.cols());
+  U.noalias() = A * tmp;
+  tmp = b[12] * A6 + b[10] * A4 + b[8] * A2;
+  V.noalias() = A6 * tmp;
+  V += b[6] * A6 + b[4] * A4 + b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols());
+}
+
+/** \brief Compute the (17,17)-Pad&eacute; approximant to the exponential.
+ *
+ *  After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Pad&eacute;
+ *  approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$.
+ *
+ *  This function activates only if your long double is double-double or quadruple.
+ */
+#if LDBL_MANT_DIG > 64
+template <typename MatA, typename MatU, typename MatV>
+void matrix_exp_pade17(const MatA& A, MatU& U, MatV& V)
+{
+  typedef typename MatA::PlainObject MatrixType;
+  typedef typename NumTraits<typename traits<MatrixType>::Scalar>::Real RealScalar;
+  const RealScalar b[] = {830034394580628357120000.L, 415017197290314178560000.L,
+                          100610229646136770560000.L, 15720348382208870400000.L,
+                          1774878043152614400000.L, 153822763739893248000.L, 10608466464820224000.L,
+                          595373117923584000.L, 27563570274240000.L, 1060137318240000.L,
+                          33924394183680.L, 899510451840.L, 19554575040.L, 341863200.L, 4651200.L,
+                          46512.L, 306.L, 1.L};
+  const MatrixType A2 = A * A;
+  const MatrixType A4 = A2 * A2;
+  const MatrixType A6 = A4 * A2;
+  const MatrixType A8 = A4 * A4;
+  V = b[17] * A8 + b[15] * A6 + b[13] * A4 + b[11] * A2; // used for temporary storage
+  MatrixType tmp = A8 * V;
+  tmp += b[9] * A8 + b[7] * A6 + b[5] * A4 + b[3] * A2 
+    + b[1] * MatrixType::Identity(A.rows(), A.cols());
+  U.noalias() = A * tmp;
+  tmp = b[16] * A8 + b[14] * A6 + b[12] * A4 + b[10] * A2;
+  V.noalias() = tmp * A8;
+  V += b[8] * A8 + b[6] * A6 + b[4] * A4 + b[2] * A2 
+    + b[0] * MatrixType::Identity(A.rows(), A.cols());
+}
+#endif
+
+template <typename MatrixType, typename RealScalar = typename NumTraits<typename traits<MatrixType>::Scalar>::Real>
+struct matrix_exp_computeUV
+{
+  /** \brief Compute Pad&eacute; approximant to the exponential.
+    *
+    * Computes \c U, \c V and \c squarings such that \f$ (V+U)(V-U)^{-1} \f$ is a Pad&eacute;
+    * approximant of \f$ \exp(2^{-\mbox{squarings}}M) \f$ around \f$ M = 0 \f$, where \f$ M \f$
+    * denotes the matrix \c arg. The degree of the Pad&eacute; approximant and the value of squarings
+    * are chosen such that the approximation error is no more than the round-off error.
+    */
+  static void run(const MatrixType& arg, MatrixType& U, MatrixType& V, int& squarings);
+};
+
+template <typename MatrixType>
+struct matrix_exp_computeUV<MatrixType, float>
+{
+  template <typename ArgType>
+  static void run(const ArgType& arg, MatrixType& U, MatrixType& V, int& squarings)
+  {
+    using std::frexp;
+    using std::pow;
+    const float l1norm = arg.cwiseAbs().colwise().sum().maxCoeff();
+    squarings = 0;
+    if (l1norm < 4.258730016922831e-001f) {
+      matrix_exp_pade3(arg, U, V);
+    } else if (l1norm < 1.880152677804762e+000f) {
+      matrix_exp_pade5(arg, U, V);
+    } else {
+      const float maxnorm = 3.925724783138660f;
+      frexp(l1norm / maxnorm, &squarings);
+      if (squarings < 0) squarings = 0;
+      MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp<float>(squarings));
+      matrix_exp_pade7(A, U, V);
+    }
+  }
+};
+
+template <typename MatrixType>
+struct matrix_exp_computeUV<MatrixType, double>
+{
+  typedef typename NumTraits<typename traits<MatrixType>::Scalar>::Real RealScalar;
+  template <typename ArgType>
+  static void run(const ArgType& arg, MatrixType& U, MatrixType& V, int& squarings)
+  {
+    using std::frexp;
+    using std::pow;
+    const RealScalar l1norm = arg.cwiseAbs().colwise().sum().maxCoeff();
+    squarings = 0;
+    if (l1norm < 1.495585217958292e-002) {
+      matrix_exp_pade3(arg, U, V);
+    } else if (l1norm < 2.539398330063230e-001) {
+      matrix_exp_pade5(arg, U, V);
+    } else if (l1norm < 9.504178996162932e-001) {
+      matrix_exp_pade7(arg, U, V);
+    } else if (l1norm < 2.097847961257068e+000) {
+      matrix_exp_pade9(arg, U, V);
+    } else {
+      const RealScalar maxnorm = 5.371920351148152;
+      frexp(l1norm / maxnorm, &squarings);
+      if (squarings < 0) squarings = 0;
+      MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp<RealScalar>(squarings));
+      matrix_exp_pade13(A, U, V);
+    }
+  }
+};
+  
+template <typename MatrixType>
+struct matrix_exp_computeUV<MatrixType, long double>
+{
+  template <typename ArgType>
+  static void run(const ArgType& arg, MatrixType& U, MatrixType& V, int& squarings)
+  {
+#if   LDBL_MANT_DIG == 53   // double precision
+    matrix_exp_computeUV<MatrixType, double>::run(arg, U, V, squarings);
+  
+#else
+  
+    using std::frexp;
+    using std::pow;
+    const long double l1norm = arg.cwiseAbs().colwise().sum().maxCoeff();
+    squarings = 0;
+  
+#if LDBL_MANT_DIG <= 64   // extended precision
+  
+    if (l1norm < 4.1968497232266989671e-003L) {
+      matrix_exp_pade3(arg, U, V);
+    } else if (l1norm < 1.1848116734693823091e-001L) {
+      matrix_exp_pade5(arg, U, V);
+    } else if (l1norm < 5.5170388480686700274e-001L) {
+      matrix_exp_pade7(arg, U, V);
+    } else if (l1norm < 1.3759868875587845383e+000L) {
+      matrix_exp_pade9(arg, U, V);
+    } else {
+      const long double maxnorm = 4.0246098906697353063L;
+      frexp(l1norm / maxnorm, &squarings);
+      if (squarings < 0) squarings = 0;
+      MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp<long double>(squarings));
+      matrix_exp_pade13(A, U, V);
+    }
+  
+#elif LDBL_MANT_DIG <= 106  // double-double
+  
+    if (l1norm < 3.2787892205607026992947488108213e-005L) {
+      matrix_exp_pade3(arg, U, V);
+    } else if (l1norm < 6.4467025060072760084130906076332e-003L) {
+      matrix_exp_pade5(arg, U, V);
+    } else if (l1norm < 6.8988028496595374751374122881143e-002L) {
+      matrix_exp_pade7(arg, U, V);
+    } else if (l1norm < 2.7339737518502231741495857201670e-001L) {
+      matrix_exp_pade9(arg, U, V);
+    } else if (l1norm < 1.3203382096514474905666448850278e+000L) {
+      matrix_exp_pade13(arg, U, V);
+    } else {
+      const long double maxnorm = 3.2579440895405400856599663723517L;
+      frexp(l1norm / maxnorm, &squarings);
+      if (squarings < 0) squarings = 0;
+      MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp<long double>(squarings));
+      matrix_exp_pade17(A, U, V);
+    }
+  
+#elif LDBL_MANT_DIG <= 113  // quadruple precision
+  
+    if (l1norm < 1.639394610288918690547467954466970e-005L) {
+      matrix_exp_pade3(arg, U, V);
+    } else if (l1norm < 4.253237712165275566025884344433009e-003L) {
+      matrix_exp_pade5(arg, U, V);
+    } else if (l1norm < 5.125804063165764409885122032933142e-002L) {
+      matrix_exp_pade7(arg, U, V);
+    } else if (l1norm < 2.170000765161155195453205651889853e-001L) {
+      matrix_exp_pade9(arg, U, V);
+    } else if (l1norm < 1.125358383453143065081397882891878e+000L) {
+      matrix_exp_pade13(arg, U, V);
+    } else {
+      const long double maxnorm = 2.884233277829519311757165057717815L;
+      frexp(l1norm / maxnorm, &squarings);
+      if (squarings < 0) squarings = 0;
+      MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp<long double>(squarings));
+      matrix_exp_pade17(A, U, V);
+    }
+  
+#else
+  
+    // this case should be handled in compute()
+    eigen_assert(false && "Bug in MatrixExponential"); 
+  
+#endif
+#endif  // LDBL_MANT_DIG
+  }
+};
+
+template<typename T> struct is_exp_known_type : false_type {};
+template<> struct is_exp_known_type<float> : true_type {};
+template<> struct is_exp_known_type<double> : true_type {};
+#if LDBL_MANT_DIG <= 113
+template<> struct is_exp_known_type<long double> : true_type {};
+#endif
+
+template <typename ArgType, typename ResultType>
+void matrix_exp_compute(const ArgType& arg, ResultType &result, true_type) // natively supported scalar type
+{
+  typedef typename ArgType::PlainObject MatrixType;
+  MatrixType U, V;
+  int squarings;
+  matrix_exp_computeUV<MatrixType>::run(arg, U, V, squarings); // Pade approximant is (U+V) / (-U+V)
+  MatrixType numer = U + V;
+  MatrixType denom = -U + V;
+  result = denom.partialPivLu().solve(numer);
+  for (int i=0; i<squarings; i++)
+    result *= result;   // undo scaling by repeated squaring
+}
+
+
+/* Computes the matrix exponential
+ *
+ * \param arg    argument of matrix exponential (should be plain object)
+ * \param result variable in which result will be stored
+ */
+template <typename ArgType, typename ResultType>
+void matrix_exp_compute(const ArgType& arg, ResultType &result, false_type) // default
+{
+  typedef typename ArgType::PlainObject MatrixType;
+  typedef typename traits<MatrixType>::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  typedef typename std::complex<RealScalar> ComplexScalar;
+  result = arg.matrixFunction(internal::stem_function_exp<ComplexScalar>);
+}
+
+} // end namespace Eigen::internal
+
+/** \ingroup MatrixFunctions_Module
+  *
+  * \brief Proxy for the matrix exponential of some matrix (expression).
+  *
+  * \tparam Derived  Type of the argument to the matrix exponential.
+  *
+  * This class holds the argument to the matrix exponential until it is assigned or evaluated for
+  * some other reason (so the argument should not be changed in the meantime). It is the return type
+  * of MatrixBase::exp() and most of the time this is the only way it is used.
+  */
+template<typename Derived> struct MatrixExponentialReturnValue
+: public ReturnByValue<MatrixExponentialReturnValue<Derived> >
+{
+  public:
+    /** \brief Constructor.
+      *
+      * \param src %Matrix (expression) forming the argument of the matrix exponential.
+      */
+    MatrixExponentialReturnValue(const Derived& src) : m_src(src) { }
+
+    /** \brief Compute the matrix exponential.
+      *
+      * \param result the matrix exponential of \p src in the constructor.
+      */
+    template <typename ResultType>
+    inline void evalTo(ResultType& result) const
+    {
+      const typename internal::nested_eval<Derived, 10>::type tmp(m_src);
+      internal::matrix_exp_compute(tmp, result, internal::is_exp_known_type<typename Derived::RealScalar>());
+    }
+
+    Index rows() const { return m_src.rows(); }
+    Index cols() const { return m_src.cols(); }
+
+  protected:
+    const typename internal::ref_selector<Derived>::type m_src;
+};
+
+namespace internal {
+template<typename Derived>
+struct traits<MatrixExponentialReturnValue<Derived> >
+{
+  typedef typename Derived::PlainObject ReturnType;
+};
+}
+
+template <typename Derived>
+const MatrixExponentialReturnValue<Derived> MatrixBase<Derived>::exp() const
+{
+  eigen_assert(rows() == cols());
+  return MatrixExponentialReturnValue<Derived>(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATRIX_EXPONENTIAL
diff --git a/src/EigenUnsupported/src/MatrixFunctions/MatrixFunction.h b/src/EigenUnsupported/src/MatrixFunctions/MatrixFunction.h
new file mode 100644
index 0000000..cc12ab6
--- /dev/null
+++ b/src/EigenUnsupported/src/MatrixFunctions/MatrixFunction.h
@@ -0,0 +1,569 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2011, 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIX_FUNCTION_H
+#define EIGEN_MATRIX_FUNCTION_H
+
+#include "StemFunction.h"
+
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \brief Maximum distance allowed between eigenvalues to be considered "close". */
+static const float matrix_function_separation = 0.1f;
+
+/** \ingroup MatrixFunctions_Module
+  * \class MatrixFunctionAtomic
+  * \brief Helper class for computing matrix functions of atomic matrices.
+  *
+  * Here, an atomic matrix is a triangular matrix whose diagonal entries are close to each other.
+  */
+template <typename MatrixType>
+class MatrixFunctionAtomic 
+{
+  public:
+
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename stem_function<Scalar>::type StemFunction;
+
+    /** \brief Constructor
+      * \param[in]  f  matrix function to compute.
+      */
+    MatrixFunctionAtomic(StemFunction f) : m_f(f) { }
+
+    /** \brief Compute matrix function of atomic matrix
+      * \param[in]  A  argument of matrix function, should be upper triangular and atomic
+      * \returns  f(A), the matrix function evaluated at the given matrix
+      */
+    MatrixType compute(const MatrixType& A);
+
+  private:
+    StemFunction* m_f;
+};
+
+template <typename MatrixType>
+typename NumTraits<typename MatrixType::Scalar>::Real matrix_function_compute_mu(const MatrixType& A)
+{
+  typedef typename plain_col_type<MatrixType>::type VectorType;
+  Index rows = A.rows();
+  const MatrixType N = MatrixType::Identity(rows, rows) - A;
+  VectorType e = VectorType::Ones(rows);
+  N.template triangularView<Upper>().solveInPlace(e);
+  return e.cwiseAbs().maxCoeff();
+}
+
+template <typename MatrixType>
+MatrixType MatrixFunctionAtomic<MatrixType>::compute(const MatrixType& A)
+{
+  // TODO: Use that A is upper triangular
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  Index rows = A.rows();
+  Scalar avgEival = A.trace() / Scalar(RealScalar(rows));
+  MatrixType Ashifted = A - avgEival * MatrixType::Identity(rows, rows);
+  RealScalar mu = matrix_function_compute_mu(Ashifted);
+  MatrixType F = m_f(avgEival, 0) * MatrixType::Identity(rows, rows);
+  MatrixType P = Ashifted;
+  MatrixType Fincr;
+  for (Index s = 1; double(s) < 1.1 * double(rows) + 10.0; s++) { // upper limit is fairly arbitrary
+    Fincr = m_f(avgEival, static_cast<int>(s)) * P;
+    F += Fincr;
+    P = Scalar(RealScalar(1)/RealScalar(s + 1)) * P * Ashifted;
+
+    // test whether Taylor series converged
+    const RealScalar F_norm = F.cwiseAbs().rowwise().sum().maxCoeff();
+    const RealScalar Fincr_norm = Fincr.cwiseAbs().rowwise().sum().maxCoeff();
+    if (Fincr_norm < NumTraits<Scalar>::epsilon() * F_norm) {
+      RealScalar delta = 0;
+      RealScalar rfactorial = 1;
+      for (Index r = 0; r < rows; r++) {
+        RealScalar mx = 0;
+        for (Index i = 0; i < rows; i++)
+          mx = (std::max)(mx, std::abs(m_f(Ashifted(i, i) + avgEival, static_cast<int>(s+r))));
+        if (r != 0)
+          rfactorial *= RealScalar(r);
+        delta = (std::max)(delta, mx / rfactorial);
+      }
+      const RealScalar P_norm = P.cwiseAbs().rowwise().sum().maxCoeff();
+      if (mu * delta * P_norm < NumTraits<Scalar>::epsilon() * F_norm) // series converged
+        break;
+    }
+  }
+  return F;
+}
+
+/** \brief Find cluster in \p clusters containing some value 
+  * \param[in] key Value to find
+  * \returns Iterator to cluster containing \p key, or \c clusters.end() if no cluster in \p m_clusters
+  * contains \p key.
+  */
+template <typename Index, typename ListOfClusters>
+typename ListOfClusters::iterator matrix_function_find_cluster(Index key, ListOfClusters& clusters)
+{
+  typename std::list<Index>::iterator j;
+  for (typename ListOfClusters::iterator i = clusters.begin(); i != clusters.end(); ++i) {
+    j = std::find(i->begin(), i->end(), key);
+    if (j != i->end())
+      return i;
+  }
+  return clusters.end();
+}
+
+/** \brief Partition eigenvalues in clusters of ei'vals close to each other
+  * 
+  * \param[in]  eivals    Eigenvalues
+  * \param[out] clusters  Resulting partition of eigenvalues
+  *
+  * The partition satisfies the following two properties:
+  * # Any eigenvalue in a certain cluster is at most matrix_function_separation() away from another eigenvalue
+  *   in the same cluster.
+  * # The distance between two eigenvalues in different clusters is more than matrix_function_separation().  
+  * The implementation follows Algorithm 4.1 in the paper of Davies and Higham.
+  */
+template <typename EivalsType, typename Cluster>
+void matrix_function_partition_eigenvalues(const EivalsType& eivals, std::list<Cluster>& clusters)
+{
+  typedef typename EivalsType::RealScalar RealScalar;
+  for (Index i=0; i<eivals.rows(); ++i) {
+    // Find cluster containing i-th ei'val, adding a new cluster if necessary
+    typename std::list<Cluster>::iterator qi = matrix_function_find_cluster(i, clusters);
+    if (qi == clusters.end()) {
+      Cluster l;
+      l.push_back(i);
+      clusters.push_back(l);
+      qi = clusters.end();
+      --qi;
+    }
+
+    // Look for other element to add to the set
+    for (Index j=i+1; j<eivals.rows(); ++j) {
+      if (abs(eivals(j) - eivals(i)) <= RealScalar(matrix_function_separation)
+          && std::find(qi->begin(), qi->end(), j) == qi->end()) {
+        typename std::list<Cluster>::iterator qj = matrix_function_find_cluster(j, clusters);
+        if (qj == clusters.end()) {
+          qi->push_back(j);
+        } else {
+          qi->insert(qi->end(), qj->begin(), qj->end());
+          clusters.erase(qj);
+        }
+      }
+    }
+  }
+}
+
+/** \brief Compute size of each cluster given a partitioning */
+template <typename ListOfClusters, typename Index>
+void matrix_function_compute_cluster_size(const ListOfClusters& clusters, Matrix<Index, Dynamic, 1>& clusterSize)
+{
+  const Index numClusters = static_cast<Index>(clusters.size());
+  clusterSize.setZero(numClusters);
+  Index clusterIndex = 0;
+  for (typename ListOfClusters::const_iterator cluster = clusters.begin(); cluster != clusters.end(); ++cluster) {
+    clusterSize[clusterIndex] = cluster->size();
+    ++clusterIndex;
+  }
+}
+
+/** \brief Compute start of each block using clusterSize */
+template <typename VectorType>
+void matrix_function_compute_block_start(const VectorType& clusterSize, VectorType& blockStart)
+{
+  blockStart.resize(clusterSize.rows());
+  blockStart(0) = 0;
+  for (Index i = 1; i < clusterSize.rows(); i++) {
+    blockStart(i) = blockStart(i-1) + clusterSize(i-1);
+  }
+}
+
+/** \brief Compute mapping of eigenvalue indices to cluster indices */
+template <typename EivalsType, typename ListOfClusters, typename VectorType>
+void matrix_function_compute_map(const EivalsType& eivals, const ListOfClusters& clusters, VectorType& eivalToCluster)
+{
+  eivalToCluster.resize(eivals.rows());
+  Index clusterIndex = 0;
+  for (typename ListOfClusters::const_iterator cluster = clusters.begin(); cluster != clusters.end(); ++cluster) {
+    for (Index i = 0; i < eivals.rows(); ++i) {
+      if (std::find(cluster->begin(), cluster->end(), i) != cluster->end()) {
+        eivalToCluster[i] = clusterIndex;
+      }
+    }
+    ++clusterIndex;
+  }
+}
+
+/** \brief Compute permutation which groups ei'vals in same cluster together */
+template <typename DynVectorType, typename VectorType>
+void matrix_function_compute_permutation(const DynVectorType& blockStart, const DynVectorType& eivalToCluster, VectorType& permutation)
+{
+  DynVectorType indexNextEntry = blockStart;
+  permutation.resize(eivalToCluster.rows());
+  for (Index i = 0; i < eivalToCluster.rows(); i++) {
+    Index cluster = eivalToCluster[i];
+    permutation[i] = indexNextEntry[cluster];
+    ++indexNextEntry[cluster];
+  }
+}  
+
+/** \brief Permute Schur decomposition in U and T according to permutation */
+template <typename VectorType, typename MatrixType>
+void matrix_function_permute_schur(VectorType& permutation, MatrixType& U, MatrixType& T)
+{
+  for (Index i = 0; i < permutation.rows() - 1; i++) {
+    Index j;
+    for (j = i; j < permutation.rows(); j++) {
+      if (permutation(j) == i) break;
+    }
+    eigen_assert(permutation(j) == i);
+    for (Index k = j-1; k >= i; k--) {
+      JacobiRotation<typename MatrixType::Scalar> rotation;
+      rotation.makeGivens(T(k, k+1), T(k+1, k+1) - T(k, k));
+      T.applyOnTheLeft(k, k+1, rotation.adjoint());
+      T.applyOnTheRight(k, k+1, rotation);
+      U.applyOnTheRight(k, k+1, rotation);
+      std::swap(permutation.coeffRef(k), permutation.coeffRef(k+1));
+    }
+  }
+}
+
+/** \brief Compute block diagonal part of matrix function.
+  *
+  * This routine computes the matrix function applied to the block diagonal part of \p T (which should be
+  * upper triangular), with the blocking given by \p blockStart and \p clusterSize. The matrix function of
+  * each diagonal block is computed by \p atomic. The off-diagonal parts of \p fT are set to zero.
+  */
+template <typename MatrixType, typename AtomicType, typename VectorType>
+void matrix_function_compute_block_atomic(const MatrixType& T, AtomicType& atomic, const VectorType& blockStart, const VectorType& clusterSize, MatrixType& fT)
+{ 
+  fT.setZero(T.rows(), T.cols());
+  for (Index i = 0; i < clusterSize.rows(); ++i) {
+    fT.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i))
+      = atomic.compute(T.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i)));
+  }
+}
+
+/** \brief Solve a triangular Sylvester equation AX + XB = C 
+  *
+  * \param[in]  A  the matrix A; should be square and upper triangular
+  * \param[in]  B  the matrix B; should be square and upper triangular
+  * \param[in]  C  the matrix C; should have correct size.
+  *
+  * \returns the solution X.
+  *
+  * If A is m-by-m and B is n-by-n, then both C and X are m-by-n.  The (i,j)-th component of the Sylvester
+  * equation is
+  * \f[ 
+  *     \sum_{k=i}^m A_{ik} X_{kj} + \sum_{k=1}^j X_{ik} B_{kj} = C_{ij}. 
+  * \f]
+  * This can be re-arranged to yield:
+  * \f[ 
+  *     X_{ij} = \frac{1}{A_{ii} + B_{jj}} \Bigl( C_{ij}
+  *     - \sum_{k=i+1}^m A_{ik} X_{kj} - \sum_{k=1}^{j-1} X_{ik} B_{kj} \Bigr).
+  * \f]
+  * It is assumed that A and B are such that the numerator is never zero (otherwise the Sylvester equation
+  * does not have a unique solution). In that case, these equations can be evaluated in the order 
+  * \f$ i=m,\ldots,1 \f$ and \f$ j=1,\ldots,n \f$.
+  */
+template <typename MatrixType>
+MatrixType matrix_function_solve_triangular_sylvester(const MatrixType& A, const MatrixType& B, const MatrixType& C)
+{
+  eigen_assert(A.rows() == A.cols());
+  eigen_assert(A.isUpperTriangular());
+  eigen_assert(B.rows() == B.cols());
+  eigen_assert(B.isUpperTriangular());
+  eigen_assert(C.rows() == A.rows());
+  eigen_assert(C.cols() == B.rows());
+
+  typedef typename MatrixType::Scalar Scalar;
+
+  Index m = A.rows();
+  Index n = B.rows();
+  MatrixType X(m, n);
+
+  for (Index i = m - 1; i >= 0; --i) {
+    for (Index j = 0; j < n; ++j) {
+
+      // Compute AX = \sum_{k=i+1}^m A_{ik} X_{kj}
+      Scalar AX;
+      if (i == m - 1) {
+	AX = 0; 
+      } else {
+	Matrix<Scalar,1,1> AXmatrix = A.row(i).tail(m-1-i) * X.col(j).tail(m-1-i);
+	AX = AXmatrix(0,0);
+      }
+
+      // Compute XB = \sum_{k=1}^{j-1} X_{ik} B_{kj}
+      Scalar XB;
+      if (j == 0) {
+	XB = 0; 
+      } else {
+	Matrix<Scalar,1,1> XBmatrix = X.row(i).head(j) * B.col(j).head(j);
+	XB = XBmatrix(0,0);
+      }
+
+      X(i,j) = (C(i,j) - AX - XB) / (A(i,i) + B(j,j));
+    }
+  }
+  return X;
+}
+
+/** \brief Compute part of matrix function above block diagonal.
+  *
+  * This routine completes the computation of \p fT, denoting a matrix function applied to the triangular
+  * matrix \p T. It assumes that the block diagonal part of \p fT has already been computed. The part below
+  * the diagonal is zero, because \p T is upper triangular.
+  */
+template <typename MatrixType, typename VectorType>
+void matrix_function_compute_above_diagonal(const MatrixType& T, const VectorType& blockStart, const VectorType& clusterSize, MatrixType& fT)
+{ 
+  typedef internal::traits<MatrixType> Traits;
+  typedef typename MatrixType::Scalar Scalar;
+  static const int Options = MatrixType::Options;
+  typedef Matrix<Scalar, Dynamic, Dynamic, Options, Traits::RowsAtCompileTime, Traits::ColsAtCompileTime> DynMatrixType;
+
+  for (Index k = 1; k < clusterSize.rows(); k++) {
+    for (Index i = 0; i < clusterSize.rows() - k; i++) {
+      // compute (i, i+k) block
+      DynMatrixType A = T.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i));
+      DynMatrixType B = -T.block(blockStart(i+k), blockStart(i+k), clusterSize(i+k), clusterSize(i+k));
+      DynMatrixType C = fT.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i))
+        * T.block(blockStart(i), blockStart(i+k), clusterSize(i), clusterSize(i+k));
+      C -= T.block(blockStart(i), blockStart(i+k), clusterSize(i), clusterSize(i+k))
+        * fT.block(blockStart(i+k), blockStart(i+k), clusterSize(i+k), clusterSize(i+k));
+      for (Index m = i + 1; m < i + k; m++) {
+        C += fT.block(blockStart(i), blockStart(m), clusterSize(i), clusterSize(m))
+          * T.block(blockStart(m), blockStart(i+k), clusterSize(m), clusterSize(i+k));
+        C -= T.block(blockStart(i), blockStart(m), clusterSize(i), clusterSize(m))
+          * fT.block(blockStart(m), blockStart(i+k), clusterSize(m), clusterSize(i+k));
+      }
+      fT.block(blockStart(i), blockStart(i+k), clusterSize(i), clusterSize(i+k))
+        = matrix_function_solve_triangular_sylvester(A, B, C);
+    }
+  }
+}
+
+/** \ingroup MatrixFunctions_Module
+  * \brief Class for computing matrix functions.
+  * \tparam  MatrixType  type of the argument of the matrix function,
+  *                      expected to be an instantiation of the Matrix class template.
+  * \tparam  AtomicType  type for computing matrix function of atomic blocks.
+  * \tparam  IsComplex   used internally to select correct specialization.
+  *
+  * This class implements the Schur-Parlett algorithm for computing matrix functions. The spectrum of the
+  * matrix is divided in clustered of eigenvalues that lies close together. This class delegates the
+  * computation of the matrix function on every block corresponding to these clusters to an object of type
+  * \p AtomicType and uses these results to compute the matrix function of the whole matrix. The class
+  * \p AtomicType should have a \p compute() member function for computing the matrix function of a block.
+  *
+  * \sa class MatrixFunctionAtomic, class MatrixLogarithmAtomic
+  */
+template <typename MatrixType, int IsComplex = NumTraits<typename internal::traits<MatrixType>::Scalar>::IsComplex>
+struct matrix_function_compute
+{  
+    /** \brief Compute the matrix function.
+      *
+      * \param[in]  A       argument of matrix function, should be a square matrix.
+      * \param[in]  atomic  class for computing matrix function of atomic blocks.
+      * \param[out] result  the function \p f applied to \p A, as
+      * specified in the constructor.
+      *
+      * See MatrixBase::matrixFunction() for details on how this computation
+      * is implemented.
+      */
+    template <typename AtomicType, typename ResultType> 
+    static void run(const MatrixType& A, AtomicType& atomic, ResultType &result);    
+};
+
+/** \internal \ingroup MatrixFunctions_Module 
+  * \brief Partial specialization of MatrixFunction for real matrices
+  *
+  * This converts the real matrix to a complex matrix, compute the matrix function of that matrix, and then
+  * converts the result back to a real matrix.
+  */
+template <typename MatrixType>
+struct matrix_function_compute<MatrixType, 0>
+{  
+  template <typename MatA, typename AtomicType, typename ResultType>
+  static void run(const MatA& A, AtomicType& atomic, ResultType &result)
+  {
+    typedef internal::traits<MatrixType> Traits;
+    typedef typename Traits::Scalar Scalar;
+    static const int Rows = Traits::RowsAtCompileTime, Cols = Traits::ColsAtCompileTime;
+    static const int MaxRows = Traits::MaxRowsAtCompileTime, MaxCols = Traits::MaxColsAtCompileTime;
+
+    typedef std::complex<Scalar> ComplexScalar;
+    typedef Matrix<ComplexScalar, Rows, Cols, 0, MaxRows, MaxCols> ComplexMatrix;
+
+    ComplexMatrix CA = A.template cast<ComplexScalar>();
+    ComplexMatrix Cresult;
+    matrix_function_compute<ComplexMatrix>::run(CA, atomic, Cresult);
+    result = Cresult.real();
+  }
+};
+
+/** \internal \ingroup MatrixFunctions_Module 
+  * \brief Partial specialization of MatrixFunction for complex matrices
+  */
+template <typename MatrixType>
+struct matrix_function_compute<MatrixType, 1>
+{
+  template <typename MatA, typename AtomicType, typename ResultType>
+  static void run(const MatA& A, AtomicType& atomic, ResultType &result)
+  {
+    typedef internal::traits<MatrixType> Traits;
+    
+    // compute Schur decomposition of A
+    const ComplexSchur<MatrixType> schurOfA(A);
+    eigen_assert(schurOfA.info()==Success);
+    MatrixType T = schurOfA.matrixT();
+    MatrixType U = schurOfA.matrixU();
+
+    // partition eigenvalues into clusters of ei'vals "close" to each other
+    std::list<std::list<Index> > clusters; 
+    matrix_function_partition_eigenvalues(T.diagonal(), clusters);
+
+    // compute size of each cluster
+    Matrix<Index, Dynamic, 1> clusterSize;
+    matrix_function_compute_cluster_size(clusters, clusterSize);
+
+    // blockStart[i] is row index at which block corresponding to i-th cluster starts 
+    Matrix<Index, Dynamic, 1> blockStart; 
+    matrix_function_compute_block_start(clusterSize, blockStart);
+
+    // compute map so that eivalToCluster[i] = j means that i-th ei'val is in j-th cluster 
+    Matrix<Index, Dynamic, 1> eivalToCluster;
+    matrix_function_compute_map(T.diagonal(), clusters, eivalToCluster);
+
+    // compute permutation which groups ei'vals in same cluster together 
+    Matrix<Index, Traits::RowsAtCompileTime, 1> permutation;
+    matrix_function_compute_permutation(blockStart, eivalToCluster, permutation);
+
+    // permute Schur decomposition
+    matrix_function_permute_schur(permutation, U, T);
+
+    // compute result
+    MatrixType fT; // matrix function applied to T
+    matrix_function_compute_block_atomic(T, atomic, blockStart, clusterSize, fT);
+    matrix_function_compute_above_diagonal(T, blockStart, clusterSize, fT);
+    result = U * (fT.template triangularView<Upper>() * U.adjoint());
+  }
+};
+
+} // end of namespace internal
+
+/** \ingroup MatrixFunctions_Module
+  *
+  * \brief Proxy for the matrix function of some matrix (expression).
+  *
+  * \tparam Derived  Type of the argument to the matrix function.
+  *
+  * This class holds the argument to the matrix function until it is assigned or evaluated for some other
+  * reason (so the argument should not be changed in the meantime). It is the return type of
+  * matrixBase::matrixFunction() and related functions and most of the time this is the only way it is used.
+  */
+template<typename Derived> class MatrixFunctionReturnValue
+: public ReturnByValue<MatrixFunctionReturnValue<Derived> >
+{
+  public:
+    typedef typename Derived::Scalar Scalar;
+    typedef typename internal::stem_function<Scalar>::type StemFunction;
+
+  protected:
+    typedef typename internal::ref_selector<Derived>::type DerivedNested;
+
+  public:
+
+    /** \brief Constructor.
+      *
+      * \param[in] A  %Matrix (expression) forming the argument of the matrix function.
+      * \param[in] f  Stem function for matrix function under consideration.
+      */
+    MatrixFunctionReturnValue(const Derived& A, StemFunction f) : m_A(A), m_f(f) { }
+
+    /** \brief Compute the matrix function.
+      *
+      * \param[out] result \p f applied to \p A, where \p f and \p A are as in the constructor.
+      */
+    template <typename ResultType>
+    inline void evalTo(ResultType& result) const
+    {
+      typedef typename internal::nested_eval<Derived, 10>::type NestedEvalType;
+      typedef typename internal::remove_all<NestedEvalType>::type NestedEvalTypeClean;
+      typedef internal::traits<NestedEvalTypeClean> Traits;
+      typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
+      typedef Matrix<ComplexScalar, Dynamic, Dynamic, 0, Traits::RowsAtCompileTime, Traits::ColsAtCompileTime> DynMatrixType;
+
+      typedef internal::MatrixFunctionAtomic<DynMatrixType> AtomicType;
+      AtomicType atomic(m_f);
+
+      internal::matrix_function_compute<typename NestedEvalTypeClean::PlainObject>::run(m_A, atomic, result);
+    }
+
+    Index rows() const { return m_A.rows(); }
+    Index cols() const { return m_A.cols(); }
+
+  private:
+    const DerivedNested m_A;
+    StemFunction *m_f;
+};
+
+namespace internal {
+template<typename Derived>
+struct traits<MatrixFunctionReturnValue<Derived> >
+{
+  typedef typename Derived::PlainObject ReturnType;
+};
+}
+
+
+/********** MatrixBase methods **********/
+
+
+template <typename Derived>
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::matrixFunction(typename internal::stem_function<typename internal::traits<Derived>::Scalar>::type f) const
+{
+  eigen_assert(rows() == cols());
+  return MatrixFunctionReturnValue<Derived>(derived(), f);
+}
+
+template <typename Derived>
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::sin() const
+{
+  eigen_assert(rows() == cols());
+  typedef typename internal::stem_function<Scalar>::ComplexScalar ComplexScalar;
+  return MatrixFunctionReturnValue<Derived>(derived(), internal::stem_function_sin<ComplexScalar>);
+}
+
+template <typename Derived>
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::cos() const
+{
+  eigen_assert(rows() == cols());
+  typedef typename internal::stem_function<Scalar>::ComplexScalar ComplexScalar;
+  return MatrixFunctionReturnValue<Derived>(derived(), internal::stem_function_cos<ComplexScalar>);
+}
+
+template <typename Derived>
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::sinh() const
+{
+  eigen_assert(rows() == cols());
+  typedef typename internal::stem_function<Scalar>::ComplexScalar ComplexScalar;
+  return MatrixFunctionReturnValue<Derived>(derived(), internal::stem_function_sinh<ComplexScalar>);
+}
+
+template <typename Derived>
+const MatrixFunctionReturnValue<Derived> MatrixBase<Derived>::cosh() const
+{
+  eigen_assert(rows() == cols());
+  typedef typename internal::stem_function<Scalar>::ComplexScalar ComplexScalar;
+  return MatrixFunctionReturnValue<Derived>(derived(), internal::stem_function_cosh<ComplexScalar>);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATRIX_FUNCTION_H
diff --git a/src/EigenUnsupported/src/MatrixFunctions/MatrixLogarithm.h b/src/EigenUnsupported/src/MatrixFunctions/MatrixLogarithm.h
new file mode 100644
index 0000000..e917013
--- /dev/null
+++ b/src/EigenUnsupported/src/MatrixFunctions/MatrixLogarithm.h
@@ -0,0 +1,373 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011, 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
+// Copyright (C) 2011 Chen-Pang He <jdh8@ms63.hinet.net>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIX_LOGARITHM
+#define EIGEN_MATRIX_LOGARITHM
+
+namespace Eigen { 
+
+namespace internal { 
+
+template <typename Scalar>
+struct matrix_log_min_pade_degree 
+{
+  static const int value = 3;
+};
+
+template <typename Scalar>
+struct matrix_log_max_pade_degree 
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  static const int value = std::numeric_limits<RealScalar>::digits<= 24?  5:  // single precision
+                           std::numeric_limits<RealScalar>::digits<= 53?  7:  // double precision
+                           std::numeric_limits<RealScalar>::digits<= 64?  8:  // extended precision
+                           std::numeric_limits<RealScalar>::digits<=106? 10:  // double-double
+                                                                         11;  // quadruple precision
+};
+
+/** \brief Compute logarithm of 2x2 triangular matrix. */
+template <typename MatrixType>
+void matrix_log_compute_2x2(const MatrixType& A, MatrixType& result)
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  using std::abs;
+  using std::ceil;
+  using std::imag;
+  using std::log;
+
+  Scalar logA00 = log(A(0,0));
+  Scalar logA11 = log(A(1,1));
+
+  result(0,0) = logA00;
+  result(1,0) = Scalar(0);
+  result(1,1) = logA11;
+
+  Scalar y = A(1,1) - A(0,0);
+  if (y==Scalar(0))
+  {
+    result(0,1) = A(0,1) / A(0,0);
+  }
+  else if ((abs(A(0,0)) < RealScalar(0.5)*abs(A(1,1))) || (abs(A(0,0)) > 2*abs(A(1,1))))
+  {
+    result(0,1) = A(0,1) * (logA11 - logA00) / y;
+  }
+  else
+  {
+    // computation in previous branch is inaccurate if A(1,1) \approx A(0,0)
+    RealScalar unwindingNumber = ceil((imag(logA11 - logA00) - RealScalar(EIGEN_PI)) / RealScalar(2*EIGEN_PI));
+    result(0,1) = A(0,1) * (numext::log1p(y/A(0,0)) + Scalar(0,RealScalar(2*EIGEN_PI)*unwindingNumber)) / y;
+  }
+}
+
+/* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = float) */
+inline int matrix_log_get_pade_degree(float normTminusI)
+{
+  const float maxNormForPade[] = { 2.5111573934555054e-1 /* degree = 3 */ , 4.0535837411880493e-1,
+            5.3149729967117310e-1 };
+  const int minPadeDegree = matrix_log_min_pade_degree<float>::value;
+  const int maxPadeDegree = matrix_log_max_pade_degree<float>::value;
+  int degree = minPadeDegree;
+  for (; degree <= maxPadeDegree; ++degree) 
+    if (normTminusI <= maxNormForPade[degree - minPadeDegree])
+      break;
+  return degree;
+}
+
+/* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = double) */
+inline int matrix_log_get_pade_degree(double normTminusI)
+{
+  const double maxNormForPade[] = { 1.6206284795015624e-2 /* degree = 3 */ , 5.3873532631381171e-2,
+            1.1352802267628681e-1, 1.8662860613541288e-1, 2.642960831111435e-1 };
+  const int minPadeDegree = matrix_log_min_pade_degree<double>::value;
+  const int maxPadeDegree = matrix_log_max_pade_degree<double>::value;
+  int degree = minPadeDegree;
+  for (; degree <= maxPadeDegree; ++degree)
+    if (normTminusI <= maxNormForPade[degree - minPadeDegree])
+      break;
+  return degree;
+}
+
+/* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = long double) */
+inline int matrix_log_get_pade_degree(long double normTminusI)
+{
+#if   LDBL_MANT_DIG == 53         // double precision
+  const long double maxNormForPade[] = { 1.6206284795015624e-2L /* degree = 3 */ , 5.3873532631381171e-2L,
+            1.1352802267628681e-1L, 1.8662860613541288e-1L, 2.642960831111435e-1L };
+#elif LDBL_MANT_DIG <= 64         // extended precision
+  const long double maxNormForPade[] = { 5.48256690357782863103e-3L /* degree = 3 */, 2.34559162387971167321e-2L,
+            5.84603923897347449857e-2L, 1.08486423756725170223e-1L, 1.68385767881294446649e-1L,
+            2.32777776523703892094e-1L };
+#elif LDBL_MANT_DIG <= 106        // double-double
+  const long double maxNormForPade[] = { 8.58970550342939562202529664318890e-5L /* degree = 3 */,
+            9.34074328446359654039446552677759e-4L, 4.26117194647672175773064114582860e-3L,
+            1.21546224740281848743149666560464e-2L, 2.61100544998339436713088248557444e-2L,
+            4.66170074627052749243018566390567e-2L, 7.32585144444135027565872014932387e-2L,
+            1.05026503471351080481093652651105e-1L };
+#else                             // quadruple precision
+  const long double maxNormForPade[] = { 4.7419931187193005048501568167858103e-5L /* degree = 3 */,
+            5.8853168473544560470387769480192666e-4L, 2.9216120366601315391789493628113520e-3L,
+            8.8415758124319434347116734705174308e-3L, 1.9850836029449446668518049562565291e-2L,
+            3.6688019729653446926585242192447447e-2L, 5.9290962294020186998954055264528393e-2L,
+            8.6998436081634343903250580992127677e-2L, 1.1880960220216759245467951592883642e-1L };
+#endif
+  const int minPadeDegree = matrix_log_min_pade_degree<long double>::value;
+  const int maxPadeDegree = matrix_log_max_pade_degree<long double>::value;
+  int degree = minPadeDegree;
+  for (; degree <= maxPadeDegree; ++degree)
+    if (normTminusI <= maxNormForPade[degree - minPadeDegree])
+      break;
+  return degree;
+}
+
+/* \brief Compute Pade approximation to matrix logarithm */
+template <typename MatrixType>
+void matrix_log_compute_pade(MatrixType& result, const MatrixType& T, int degree)
+{
+  typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
+  const int minPadeDegree = 3;
+  const int maxPadeDegree = 11;
+  assert(degree >= minPadeDegree && degree <= maxPadeDegree);
+  // FIXME this creates float-conversion-warnings if these are enabled.
+  // Either manually convert each value, or disable the warning locally
+  const RealScalar nodes[][maxPadeDegree] = { 
+    { 0.1127016653792583114820734600217600L, 0.5000000000000000000000000000000000L,  // degree 3
+      0.8872983346207416885179265399782400L }, 
+    { 0.0694318442029737123880267555535953L, 0.3300094782075718675986671204483777L,  // degree 4
+      0.6699905217924281324013328795516223L, 0.9305681557970262876119732444464048L },
+    { 0.0469100770306680036011865608503035L, 0.2307653449471584544818427896498956L,  // degree 5
+      0.5000000000000000000000000000000000L, 0.7692346550528415455181572103501044L,
+      0.9530899229693319963988134391496965L },
+    { 0.0337652428984239860938492227530027L, 0.1693953067668677431693002024900473L,  // degree 6
+      0.3806904069584015456847491391596440L, 0.6193095930415984543152508608403560L,
+      0.8306046932331322568306997975099527L, 0.9662347571015760139061507772469973L },
+    { 0.0254460438286207377369051579760744L, 0.1292344072003027800680676133596058L,  // degree 7
+      0.2970774243113014165466967939615193L, 0.5000000000000000000000000000000000L,
+      0.7029225756886985834533032060384807L, 0.8707655927996972199319323866403942L,
+      0.9745539561713792622630948420239256L },
+    { 0.0198550717512318841582195657152635L, 0.1016667612931866302042230317620848L,  // degree 8
+      0.2372337950418355070911304754053768L, 0.4082826787521750975302619288199080L,
+      0.5917173212478249024697380711800920L, 0.7627662049581644929088695245946232L,
+      0.8983332387068133697957769682379152L, 0.9801449282487681158417804342847365L },
+    { 0.0159198802461869550822118985481636L, 0.0819844463366821028502851059651326L,  // degree 9
+      0.1933142836497048013456489803292629L, 0.3378732882980955354807309926783317L,
+      0.5000000000000000000000000000000000L, 0.6621267117019044645192690073216683L,
+      0.8066857163502951986543510196707371L, 0.9180155536633178971497148940348674L,
+      0.9840801197538130449177881014518364L },
+    { 0.0130467357414141399610179939577740L, 0.0674683166555077446339516557882535L,  // degree 10
+      0.1602952158504877968828363174425632L, 0.2833023029353764046003670284171079L,
+      0.4255628305091843945575869994351400L, 0.5744371694908156054424130005648600L,
+      0.7166976970646235953996329715828921L, 0.8397047841495122031171636825574368L,
+      0.9325316833444922553660483442117465L, 0.9869532642585858600389820060422260L },
+    { 0.0108856709269715035980309994385713L, 0.0564687001159523504624211153480364L,  // degree 11
+      0.1349239972129753379532918739844233L, 0.2404519353965940920371371652706952L,
+      0.3652284220238275138342340072995692L, 0.5000000000000000000000000000000000L,
+      0.6347715779761724861657659927004308L, 0.7595480646034059079628628347293048L,
+      0.8650760027870246620467081260155767L, 0.9435312998840476495375788846519636L,
+      0.9891143290730284964019690005614287L } };
+
+  const RealScalar weights[][maxPadeDegree] = { 
+    { 0.2777777777777777777777777777777778L, 0.4444444444444444444444444444444444L,  // degree 3
+      0.2777777777777777777777777777777778L },
+    { 0.1739274225687269286865319746109997L, 0.3260725774312730713134680253890003L,  // degree 4
+      0.3260725774312730713134680253890003L, 0.1739274225687269286865319746109997L },
+    { 0.1184634425280945437571320203599587L, 0.2393143352496832340206457574178191L,  // degree 5
+      0.2844444444444444444444444444444444L, 0.2393143352496832340206457574178191L,
+      0.1184634425280945437571320203599587L },
+    { 0.0856622461895851725201480710863665L, 0.1803807865240693037849167569188581L,  // degree 6
+      0.2339569672863455236949351719947755L, 0.2339569672863455236949351719947755L,
+      0.1803807865240693037849167569188581L, 0.0856622461895851725201480710863665L },
+    { 0.0647424830844348466353057163395410L, 0.1398526957446383339507338857118898L,  // degree 7
+      0.1909150252525594724751848877444876L, 0.2089795918367346938775510204081633L,
+      0.1909150252525594724751848877444876L, 0.1398526957446383339507338857118898L,
+      0.0647424830844348466353057163395410L },
+    { 0.0506142681451881295762656771549811L, 0.1111905172266872352721779972131204L,  // degree 8
+      0.1568533229389436436689811009933007L, 0.1813418916891809914825752246385978L,
+      0.1813418916891809914825752246385978L, 0.1568533229389436436689811009933007L,
+      0.1111905172266872352721779972131204L, 0.0506142681451881295762656771549811L },
+    { 0.0406371941807872059859460790552618L, 0.0903240803474287020292360156214564L,  // degree 9
+      0.1303053482014677311593714347093164L, 0.1561735385200014200343152032922218L,
+      0.1651196775006298815822625346434870L, 0.1561735385200014200343152032922218L,
+      0.1303053482014677311593714347093164L, 0.0903240803474287020292360156214564L,
+      0.0406371941807872059859460790552618L },
+    { 0.0333356721543440687967844049466659L, 0.0747256745752902965728881698288487L,  // degree 10
+      0.1095431812579910219977674671140816L, 0.1346333596549981775456134607847347L,
+      0.1477621123573764350869464973256692L, 0.1477621123573764350869464973256692L,
+      0.1346333596549981775456134607847347L, 0.1095431812579910219977674671140816L,
+      0.0747256745752902965728881698288487L, 0.0333356721543440687967844049466659L },
+    { 0.0278342835580868332413768602212743L, 0.0627901847324523123173471496119701L,  // degree 11
+      0.0931451054638671257130488207158280L, 0.1165968822959952399592618524215876L,
+      0.1314022722551233310903444349452546L, 0.1364625433889503153572417641681711L,
+      0.1314022722551233310903444349452546L, 0.1165968822959952399592618524215876L,
+      0.0931451054638671257130488207158280L, 0.0627901847324523123173471496119701L,
+      0.0278342835580868332413768602212743L } };
+
+  MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows());
+  result.setZero(T.rows(), T.rows());
+  for (int k = 0; k < degree; ++k) {
+    RealScalar weight = weights[degree-minPadeDegree][k];
+    RealScalar node = nodes[degree-minPadeDegree][k];
+    result += weight * (MatrixType::Identity(T.rows(), T.rows()) + node * TminusI)
+                       .template triangularView<Upper>().solve(TminusI);
+  }
+} 
+
+/** \brief Compute logarithm of triangular matrices with size > 2. 
+  * \details This uses a inverse scale-and-square algorithm. */
+template <typename MatrixType>
+void matrix_log_compute_big(const MatrixType& A, MatrixType& result)
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  using std::pow;
+
+  int numberOfSquareRoots = 0;
+  int numberOfExtraSquareRoots = 0;
+  int degree;
+  MatrixType T = A, sqrtT;
+
+  const int maxPadeDegree = matrix_log_max_pade_degree<Scalar>::value;
+  const RealScalar maxNormForPade = RealScalar(
+                                    maxPadeDegree<= 5? 5.3149729967117310e-1L:                    // single precision
+                                    maxPadeDegree<= 7? 2.6429608311114350e-1L:                    // double precision
+                                    maxPadeDegree<= 8? 2.32777776523703892094e-1L:                // extended precision
+                                    maxPadeDegree<=10? 1.05026503471351080481093652651105e-1L:    // double-double
+                                                       1.1880960220216759245467951592883642e-1L); // quadruple precision
+
+  while (true) {
+    RealScalar normTminusI = (T - MatrixType::Identity(T.rows(), T.rows())).cwiseAbs().colwise().sum().maxCoeff();
+    if (normTminusI < maxNormForPade) {
+      degree = matrix_log_get_pade_degree(normTminusI);
+      int degree2 = matrix_log_get_pade_degree(normTminusI / RealScalar(2));
+      if ((degree - degree2 <= 1) || (numberOfExtraSquareRoots == 1)) 
+        break;
+      ++numberOfExtraSquareRoots;
+    }
+    matrix_sqrt_triangular(T, sqrtT);
+    T = sqrtT.template triangularView<Upper>();
+    ++numberOfSquareRoots;
+  }
+
+  matrix_log_compute_pade(result, T, degree);
+  result *= pow(RealScalar(2), RealScalar(numberOfSquareRoots)); // TODO replace by bitshift if possible
+}
+
+/** \ingroup MatrixFunctions_Module
+  * \class MatrixLogarithmAtomic
+  * \brief Helper class for computing matrix logarithm of atomic matrices.
+  *
+  * Here, an atomic matrix is a triangular matrix whose diagonal entries are close to each other.
+  *
+  * \sa class MatrixFunctionAtomic, MatrixBase::log()
+  */
+template <typename MatrixType>
+class MatrixLogarithmAtomic
+{
+public:
+  /** \brief Compute matrix logarithm of atomic matrix
+    * \param[in]  A  argument of matrix logarithm, should be upper triangular and atomic
+    * \returns  The logarithm of \p A.
+    */
+  MatrixType compute(const MatrixType& A);
+};
+
+template <typename MatrixType>
+MatrixType MatrixLogarithmAtomic<MatrixType>::compute(const MatrixType& A)
+{
+  using std::log;
+  MatrixType result(A.rows(), A.rows());
+  if (A.rows() == 1)
+    result(0,0) = log(A(0,0));
+  else if (A.rows() == 2)
+    matrix_log_compute_2x2(A, result);
+  else
+    matrix_log_compute_big(A, result);
+  return result;
+}
+
+} // end of namespace internal
+
+/** \ingroup MatrixFunctions_Module
+  *
+  * \brief Proxy for the matrix logarithm of some matrix (expression).
+  *
+  * \tparam Derived  Type of the argument to the matrix function.
+  *
+  * This class holds the argument to the matrix function until it is
+  * assigned or evaluated for some other reason (so the argument
+  * should not be changed in the meantime). It is the return type of
+  * MatrixBase::log() and most of the time this is the only way it
+  * is used.
+  */
+template<typename Derived> class MatrixLogarithmReturnValue
+: public ReturnByValue<MatrixLogarithmReturnValue<Derived> >
+{
+public:
+  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::Index Index;
+
+protected:
+  typedef typename internal::ref_selector<Derived>::type DerivedNested;
+
+public:
+
+  /** \brief Constructor.
+    *
+    * \param[in]  A  %Matrix (expression) forming the argument of the matrix logarithm.
+    */
+  explicit MatrixLogarithmReturnValue(const Derived& A) : m_A(A) { }
+  
+  /** \brief Compute the matrix logarithm.
+    *
+    * \param[out]  result  Logarithm of \c A, where \c A is as specified in the constructor.
+    */
+  template <typename ResultType>
+  inline void evalTo(ResultType& result) const
+  {
+    typedef typename internal::nested_eval<Derived, 10>::type DerivedEvalType;
+    typedef typename internal::remove_all<DerivedEvalType>::type DerivedEvalTypeClean;
+    typedef internal::traits<DerivedEvalTypeClean> Traits;
+    typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
+    typedef Matrix<ComplexScalar, Dynamic, Dynamic, 0, Traits::RowsAtCompileTime, Traits::ColsAtCompileTime> DynMatrixType;
+    typedef internal::MatrixLogarithmAtomic<DynMatrixType> AtomicType;
+    AtomicType atomic;
+    
+    internal::matrix_function_compute<typename DerivedEvalTypeClean::PlainObject>::run(m_A, atomic, result);
+  }
+
+  Index rows() const { return m_A.rows(); }
+  Index cols() const { return m_A.cols(); }
+  
+private:
+  const DerivedNested m_A;
+};
+
+namespace internal {
+  template<typename Derived>
+  struct traits<MatrixLogarithmReturnValue<Derived> >
+  {
+    typedef typename Derived::PlainObject ReturnType;
+  };
+}
+
+
+/********** MatrixBase method **********/
+
+
+template <typename Derived>
+const MatrixLogarithmReturnValue<Derived> MatrixBase<Derived>::log() const
+{
+  eigen_assert(rows() == cols());
+  return MatrixLogarithmReturnValue<Derived>(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATRIX_LOGARITHM
diff --git a/src/EigenUnsupported/src/MatrixFunctions/MatrixPower.h b/src/EigenUnsupported/src/MatrixFunctions/MatrixPower.h
new file mode 100644
index 0000000..d7672d7
--- /dev/null
+++ b/src/EigenUnsupported/src/MatrixFunctions/MatrixPower.h
@@ -0,0 +1,705 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012, 2013 Chen-Pang He <jdh8@ms63.hinet.net>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIX_POWER
+#define EIGEN_MATRIX_POWER
+
+namespace Eigen {
+
+template<typename MatrixType> class MatrixPower;
+
+/**
+ * \ingroup MatrixFunctions_Module
+ *
+ * \brief Proxy for the matrix power of some matrix.
+ *
+ * \tparam MatrixType  type of the base, a matrix.
+ *
+ * This class holds the arguments to the matrix power until it is
+ * assigned or evaluated for some other reason (so the argument
+ * should not be changed in the meantime). It is the return type of
+ * MatrixPower::operator() and related functions and most of the
+ * time this is the only way it is used.
+ */
+/* TODO This class is only used by MatrixPower, so it should be nested
+ * into MatrixPower, like MatrixPower::ReturnValue. However, my
+ * compiler complained about unused template parameter in the
+ * following declaration in namespace internal.
+ *
+ * template<typename MatrixType>
+ * struct traits<MatrixPower<MatrixType>::ReturnValue>;
+ */
+template<typename MatrixType>
+class MatrixPowerParenthesesReturnValue : public ReturnByValue< MatrixPowerParenthesesReturnValue<MatrixType> >
+{
+  public:
+    typedef typename MatrixType::RealScalar RealScalar;
+
+    /**
+     * \brief Constructor.
+     *
+     * \param[in] pow  %MatrixPower storing the base.
+     * \param[in] p    scalar, the exponent of the matrix power.
+     */
+    MatrixPowerParenthesesReturnValue(MatrixPower<MatrixType>& pow, RealScalar p) : m_pow(pow), m_p(p)
+    { }
+
+    /**
+     * \brief Compute the matrix power.
+     *
+     * \param[out] result
+     */
+    template<typename ResultType>
+    inline void evalTo(ResultType& result) const
+    { m_pow.compute(result, m_p); }
+
+    Index rows() const { return m_pow.rows(); }
+    Index cols() const { return m_pow.cols(); }
+
+  private:
+    MatrixPower<MatrixType>& m_pow;
+    const RealScalar m_p;
+};
+
+/**
+ * \ingroup MatrixFunctions_Module
+ *
+ * \brief Class for computing matrix powers.
+ *
+ * \tparam MatrixType  type of the base, expected to be an instantiation
+ * of the Matrix class template.
+ *
+ * This class is capable of computing triangular real/complex matrices
+ * raised to a power in the interval \f$ (-1, 1) \f$.
+ *
+ * \note Currently this class is only used by MatrixPower. One may
+ * insist that this be nested into MatrixPower. This class is here to
+ * facilitate future development of triangular matrix functions.
+ */
+template<typename MatrixType>
+class MatrixPowerAtomic : internal::noncopyable
+{
+  private:
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime
+    };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef std::complex<RealScalar> ComplexScalar;
+    typedef Block<MatrixType,Dynamic,Dynamic> ResultType;
+
+    const MatrixType& m_A;
+    RealScalar m_p;
+
+    void computePade(int degree, const MatrixType& IminusT, ResultType& res) const;
+    void compute2x2(ResultType& res, RealScalar p) const;
+    void computeBig(ResultType& res) const;
+    static int getPadeDegree(float normIminusT);
+    static int getPadeDegree(double normIminusT);
+    static int getPadeDegree(long double normIminusT);
+    static ComplexScalar computeSuperDiag(const ComplexScalar&, const ComplexScalar&, RealScalar p);
+    static RealScalar computeSuperDiag(RealScalar, RealScalar, RealScalar p);
+
+  public:
+    /**
+     * \brief Constructor.
+     *
+     * \param[in] T  the base of the matrix power.
+     * \param[in] p  the exponent of the matrix power, should be in
+     * \f$ (-1, 1) \f$.
+     *
+     * The class stores a reference to T, so it should not be changed
+     * (or destroyed) before evaluation. Only the upper triangular
+     * part of T is read.
+     */
+    MatrixPowerAtomic(const MatrixType& T, RealScalar p);
+    
+    /**
+     * \brief Compute the matrix power.
+     *
+     * \param[out] res  \f$ A^p \f$ where A and p are specified in the
+     * constructor.
+     */
+    void compute(ResultType& res) const;
+};
+
+template<typename MatrixType>
+MatrixPowerAtomic<MatrixType>::MatrixPowerAtomic(const MatrixType& T, RealScalar p) :
+  m_A(T), m_p(p)
+{
+  eigen_assert(T.rows() == T.cols());
+  eigen_assert(p > -1 && p < 1);
+}
+
+template<typename MatrixType>
+void MatrixPowerAtomic<MatrixType>::compute(ResultType& res) const
+{
+  using std::pow;
+  switch (m_A.rows()) {
+    case 0:
+      break;
+    case 1:
+      res(0,0) = pow(m_A(0,0), m_p);
+      break;
+    case 2:
+      compute2x2(res, m_p);
+      break;
+    default:
+      computeBig(res);
+  }
+}
+
+template<typename MatrixType>
+void MatrixPowerAtomic<MatrixType>::computePade(int degree, const MatrixType& IminusT, ResultType& res) const
+{
+  int i = 2*degree;
+  res = (m_p-RealScalar(degree)) / RealScalar(2*i-2) * IminusT;
+
+  for (--i; i; --i) {
+    res = (MatrixType::Identity(IminusT.rows(), IminusT.cols()) + res).template triangularView<Upper>()
+	.solve((i==1 ? -m_p : i&1 ? (-m_p-RealScalar(i/2))/RealScalar(2*i) : (m_p-RealScalar(i/2))/RealScalar(2*i-2)) * IminusT).eval();
+  }
+  res += MatrixType::Identity(IminusT.rows(), IminusT.cols());
+}
+
+// This function assumes that res has the correct size (see bug 614)
+template<typename MatrixType>
+void MatrixPowerAtomic<MatrixType>::compute2x2(ResultType& res, RealScalar p) const
+{
+  using std::abs;
+  using std::pow;
+  res.coeffRef(0,0) = pow(m_A.coeff(0,0), p);
+
+  for (Index i=1; i < m_A.cols(); ++i) {
+    res.coeffRef(i,i) = pow(m_A.coeff(i,i), p);
+    if (m_A.coeff(i-1,i-1) == m_A.coeff(i,i))
+      res.coeffRef(i-1,i) = p * pow(m_A.coeff(i,i), p-1);
+    else if (2*abs(m_A.coeff(i-1,i-1)) < abs(m_A.coeff(i,i)) || 2*abs(m_A.coeff(i,i)) < abs(m_A.coeff(i-1,i-1)))
+      res.coeffRef(i-1,i) = (res.coeff(i,i)-res.coeff(i-1,i-1)) / (m_A.coeff(i,i)-m_A.coeff(i-1,i-1));
+    else
+      res.coeffRef(i-1,i) = computeSuperDiag(m_A.coeff(i,i), m_A.coeff(i-1,i-1), p);
+    res.coeffRef(i-1,i) *= m_A.coeff(i-1,i);
+  }
+}
+
+template<typename MatrixType>
+void MatrixPowerAtomic<MatrixType>::computeBig(ResultType& res) const
+{
+  using std::ldexp;
+  const int digits = std::numeric_limits<RealScalar>::digits;
+  const RealScalar maxNormForPade = RealScalar(
+                                    digits <=  24? 4.3386528e-1L                            // single precision
+                                  : digits <=  53? 2.789358995219730e-1L                    // double precision
+                                  : digits <=  64? 2.4471944416607995472e-1L                // extended precision
+                                  : digits <= 106? 1.1016843812851143391275867258512e-1L    // double-double
+                                  :                9.134603732914548552537150753385375e-2L); // quadruple precision
+  MatrixType IminusT, sqrtT, T = m_A.template triangularView<Upper>();
+  RealScalar normIminusT;
+  int degree, degree2, numberOfSquareRoots = 0;
+  bool hasExtraSquareRoot = false;
+
+  for (Index i=0; i < m_A.cols(); ++i)
+    eigen_assert(m_A(i,i) != RealScalar(0));
+
+  while (true) {
+    IminusT = MatrixType::Identity(m_A.rows(), m_A.cols()) - T;
+    normIminusT = IminusT.cwiseAbs().colwise().sum().maxCoeff();
+    if (normIminusT < maxNormForPade) {
+      degree = getPadeDegree(normIminusT);
+      degree2 = getPadeDegree(normIminusT/2);
+      if (degree - degree2 <= 1 || hasExtraSquareRoot)
+	break;
+      hasExtraSquareRoot = true;
+    }
+    matrix_sqrt_triangular(T, sqrtT);
+    T = sqrtT.template triangularView<Upper>();
+    ++numberOfSquareRoots;
+  }
+  computePade(degree, IminusT, res);
+
+  for (; numberOfSquareRoots; --numberOfSquareRoots) {
+    compute2x2(res, ldexp(m_p, -numberOfSquareRoots));
+    res = res.template triangularView<Upper>() * res;
+  }
+  compute2x2(res, m_p);
+}
+  
+template<typename MatrixType>
+inline int MatrixPowerAtomic<MatrixType>::getPadeDegree(float normIminusT)
+{
+  const float maxNormForPade[] = { 2.8064004e-1f /* degree = 3 */ , 4.3386528e-1f };
+  int degree = 3;
+  for (; degree <= 4; ++degree)
+    if (normIminusT <= maxNormForPade[degree - 3])
+      break;
+  return degree;
+}
+
+template<typename MatrixType>
+inline int MatrixPowerAtomic<MatrixType>::getPadeDegree(double normIminusT)
+{
+  const double maxNormForPade[] = { 1.884160592658218e-2 /* degree = 3 */ , 6.038881904059573e-2, 1.239917516308172e-1,
+      1.999045567181744e-1, 2.789358995219730e-1 };
+  int degree = 3;
+  for (; degree <= 7; ++degree)
+    if (normIminusT <= maxNormForPade[degree - 3])
+      break;
+  return degree;
+}
+
+template<typename MatrixType>
+inline int MatrixPowerAtomic<MatrixType>::getPadeDegree(long double normIminusT)
+{
+#if   LDBL_MANT_DIG == 53
+  const int maxPadeDegree = 7;
+  const double maxNormForPade[] = { 1.884160592658218e-2L /* degree = 3 */ , 6.038881904059573e-2L, 1.239917516308172e-1L,
+      1.999045567181744e-1L, 2.789358995219730e-1L };
+#elif LDBL_MANT_DIG <= 64
+  const int maxPadeDegree = 8;
+  const long double maxNormForPade[] = { 6.3854693117491799460e-3L /* degree = 3 */ , 2.6394893435456973676e-2L,
+      6.4216043030404063729e-2L, 1.1701165502926694307e-1L, 1.7904284231268670284e-1L, 2.4471944416607995472e-1L };
+#elif LDBL_MANT_DIG <= 106
+  const int maxPadeDegree = 10;
+  const double maxNormForPade[] = { 1.0007161601787493236741409687186e-4L /* degree = 3 */ ,
+      1.0007161601787493236741409687186e-3L, 4.7069769360887572939882574746264e-3L, 1.3220386624169159689406653101695e-2L,
+      2.8063482381631737920612944054906e-2L, 4.9625993951953473052385361085058e-2L, 7.7367040706027886224557538328171e-2L,
+      1.1016843812851143391275867258512e-1L };
+#else
+  const int maxPadeDegree = 10;
+  const double maxNormForPade[] = { 5.524506147036624377378713555116378e-5L /* degree = 3 */ ,
+      6.640600568157479679823602193345995e-4L, 3.227716520106894279249709728084626e-3L,
+      9.619593944683432960546978734646284e-3L, 2.134595382433742403911124458161147e-2L,
+      3.908166513900489428442993794761185e-2L, 6.266780814639442865832535460550138e-2L,
+      9.134603732914548552537150753385375e-2L };
+#endif
+  int degree = 3;
+  for (; degree <= maxPadeDegree; ++degree)
+    if (normIminusT <= maxNormForPade[degree - 3])
+      break;
+  return degree;
+}
+
+template<typename MatrixType>
+inline typename MatrixPowerAtomic<MatrixType>::ComplexScalar
+MatrixPowerAtomic<MatrixType>::computeSuperDiag(const ComplexScalar& curr, const ComplexScalar& prev, RealScalar p)
+{
+  using std::ceil;
+  using std::exp;
+  using std::log;
+  using std::sinh;
+
+  ComplexScalar logCurr = log(curr);
+  ComplexScalar logPrev = log(prev);
+  RealScalar unwindingNumber = ceil((numext::imag(logCurr - logPrev) - RealScalar(EIGEN_PI)) / RealScalar(2*EIGEN_PI));
+  ComplexScalar w = numext::log1p((curr-prev)/prev)/RealScalar(2) + ComplexScalar(0, RealScalar(EIGEN_PI)*unwindingNumber);
+  return RealScalar(2) * exp(RealScalar(0.5) * p * (logCurr + logPrev)) * sinh(p * w) / (curr - prev);
+}
+
+template<typename MatrixType>
+inline typename MatrixPowerAtomic<MatrixType>::RealScalar
+MatrixPowerAtomic<MatrixType>::computeSuperDiag(RealScalar curr, RealScalar prev, RealScalar p)
+{
+  using std::exp;
+  using std::log;
+  using std::sinh;
+
+  RealScalar w = numext::log1p((curr-prev)/prev)/RealScalar(2);
+  return 2 * exp(p * (log(curr) + log(prev)) / 2) * sinh(p * w) / (curr - prev);
+}
+
+/**
+ * \ingroup MatrixFunctions_Module
+ *
+ * \brief Class for computing matrix powers.
+ *
+ * \tparam MatrixType  type of the base, expected to be an instantiation
+ * of the Matrix class template.
+ *
+ * This class is capable of computing real/complex matrices raised to
+ * an arbitrary real power. Meanwhile, it saves the result of Schur
+ * decomposition if an non-integral power has even been calculated.
+ * Therefore, if you want to compute multiple (>= 2) matrix powers
+ * for the same matrix, using the class directly is more efficient than
+ * calling MatrixBase::pow().
+ *
+ * Example:
+ * \include MatrixPower_optimal.cpp
+ * Output: \verbinclude MatrixPower_optimal.out
+ */
+template<typename MatrixType>
+class MatrixPower : internal::noncopyable
+{
+  private:
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+
+  public:
+    /**
+     * \brief Constructor.
+     *
+     * \param[in] A  the base of the matrix power.
+     *
+     * The class stores a reference to A, so it should not be changed
+     * (or destroyed) before evaluation.
+     */
+    explicit MatrixPower(const MatrixType& A) :
+      m_A(A),
+      m_conditionNumber(0),
+      m_rank(A.cols()),
+      m_nulls(0)
+    { eigen_assert(A.rows() == A.cols()); }
+
+    /**
+     * \brief Returns the matrix power.
+     *
+     * \param[in] p  exponent, a real scalar.
+     * \return The expression \f$ A^p \f$, where A is specified in the
+     * constructor.
+     */
+    const MatrixPowerParenthesesReturnValue<MatrixType> operator()(RealScalar p)
+    { return MatrixPowerParenthesesReturnValue<MatrixType>(*this, p); }
+
+    /**
+     * \brief Compute the matrix power.
+     *
+     * \param[in]  p    exponent, a real scalar.
+     * \param[out] res  \f$ A^p \f$ where A is specified in the
+     * constructor.
+     */
+    template<typename ResultType>
+    void compute(ResultType& res, RealScalar p);
+    
+    Index rows() const { return m_A.rows(); }
+    Index cols() const { return m_A.cols(); }
+
+  private:
+    typedef std::complex<RealScalar> ComplexScalar;
+    typedef Matrix<ComplexScalar, Dynamic, Dynamic, 0,
+              MatrixType::RowsAtCompileTime, MatrixType::ColsAtCompileTime> ComplexMatrix;
+
+    /** \brief Reference to the base of matrix power. */
+    typename MatrixType::Nested m_A;
+
+    /** \brief Temporary storage. */
+    MatrixType m_tmp;
+
+    /** \brief Store the result of Schur decomposition. */
+    ComplexMatrix m_T, m_U;
+    
+    /** \brief Store fractional power of m_T. */
+    ComplexMatrix m_fT;
+
+    /**
+     * \brief Condition number of m_A.
+     *
+     * It is initialized as 0 to avoid performing unnecessary Schur
+     * decomposition, which is the bottleneck.
+     */
+    RealScalar m_conditionNumber;
+
+    /** \brief Rank of m_A. */
+    Index m_rank;
+    
+    /** \brief Rank deficiency of m_A. */
+    Index m_nulls;
+
+    /**
+     * \brief Split p into integral part and fractional part.
+     *
+     * \param[in]  p        The exponent.
+     * \param[out] p        The fractional part ranging in \f$ (-1, 1) \f$.
+     * \param[out] intpart  The integral part.
+     *
+     * Only if the fractional part is nonzero, it calls initialize().
+     */
+    void split(RealScalar& p, RealScalar& intpart);
+
+    /** \brief Perform Schur decomposition for fractional power. */
+    void initialize();
+
+    template<typename ResultType>
+    void computeIntPower(ResultType& res, RealScalar p);
+
+    template<typename ResultType>
+    void computeFracPower(ResultType& res, RealScalar p);
+
+    template<int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+    static void revertSchur(
+        Matrix<ComplexScalar, Rows, Cols, Options, MaxRows, MaxCols>& res,
+        const ComplexMatrix& T,
+        const ComplexMatrix& U);
+
+    template<int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+    static void revertSchur(
+        Matrix<RealScalar, Rows, Cols, Options, MaxRows, MaxCols>& res,
+        const ComplexMatrix& T,
+        const ComplexMatrix& U);
+};
+
+template<typename MatrixType>
+template<typename ResultType>
+void MatrixPower<MatrixType>::compute(ResultType& res, RealScalar p)
+{
+  using std::pow;
+  switch (cols()) {
+    case 0:
+      break;
+    case 1:
+      res(0,0) = pow(m_A.coeff(0,0), p);
+      break;
+    default:
+      RealScalar intpart;
+      split(p, intpart);
+
+      res = MatrixType::Identity(rows(), cols());
+      computeIntPower(res, intpart);
+      if (p) computeFracPower(res, p);
+  }
+}
+
+template<typename MatrixType>
+void MatrixPower<MatrixType>::split(RealScalar& p, RealScalar& intpart)
+{
+  using std::floor;
+  using std::pow;
+
+  intpart = floor(p);
+  p -= intpart;
+
+  // Perform Schur decomposition if it is not yet performed and the power is
+  // not an integer.
+  if (!m_conditionNumber && p)
+    initialize();
+
+  // Choose the more stable of intpart = floor(p) and intpart = ceil(p).
+  if (p > RealScalar(0.5) && p > (1-p) * pow(m_conditionNumber, p)) {
+    --p;
+    ++intpart;
+  }
+}
+
+template<typename MatrixType>
+void MatrixPower<MatrixType>::initialize()
+{
+  const ComplexSchur<MatrixType> schurOfA(m_A);
+  JacobiRotation<ComplexScalar> rot;
+  ComplexScalar eigenvalue;
+
+  m_fT.resizeLike(m_A);
+  m_T = schurOfA.matrixT();
+  m_U = schurOfA.matrixU();
+  m_conditionNumber = m_T.diagonal().array().abs().maxCoeff() / m_T.diagonal().array().abs().minCoeff();
+
+  // Move zero eigenvalues to the bottom right corner.
+  for (Index i = cols()-1; i>=0; --i) {
+    if (m_rank <= 2)
+      return;
+    if (m_T.coeff(i,i) == RealScalar(0)) {
+      for (Index j=i+1; j < m_rank; ++j) {
+        eigenvalue = m_T.coeff(j,j);
+        rot.makeGivens(m_T.coeff(j-1,j), eigenvalue);
+        m_T.applyOnTheRight(j-1, j, rot);
+        m_T.applyOnTheLeft(j-1, j, rot.adjoint());
+        m_T.coeffRef(j-1,j-1) = eigenvalue;
+        m_T.coeffRef(j,j) = RealScalar(0);
+        m_U.applyOnTheRight(j-1, j, rot);
+      }
+      --m_rank;
+    }
+  }
+
+  m_nulls = rows() - m_rank;
+  if (m_nulls) {
+    eigen_assert(m_T.bottomRightCorner(m_nulls, m_nulls).isZero()
+        && "Base of matrix power should be invertible or with a semisimple zero eigenvalue.");
+    m_fT.bottomRows(m_nulls).fill(RealScalar(0));
+  }
+}
+
+template<typename MatrixType>
+template<typename ResultType>
+void MatrixPower<MatrixType>::computeIntPower(ResultType& res, RealScalar p)
+{
+  using std::abs;
+  using std::fmod;
+  RealScalar pp = abs(p);
+
+  if (p<0) 
+    m_tmp = m_A.inverse();
+  else     
+    m_tmp = m_A;
+
+  while (true) {
+    if (fmod(pp, 2) >= 1)
+      res = m_tmp * res;
+    pp /= 2;
+    if (pp < 1)
+      break;
+    m_tmp *= m_tmp;
+  }
+}
+
+template<typename MatrixType>
+template<typename ResultType>
+void MatrixPower<MatrixType>::computeFracPower(ResultType& res, RealScalar p)
+{
+  Block<ComplexMatrix,Dynamic,Dynamic> blockTp(m_fT, 0, 0, m_rank, m_rank);
+  eigen_assert(m_conditionNumber);
+  eigen_assert(m_rank + m_nulls == rows());
+
+  MatrixPowerAtomic<ComplexMatrix>(m_T.topLeftCorner(m_rank, m_rank), p).compute(blockTp);
+  if (m_nulls) {
+    m_fT.topRightCorner(m_rank, m_nulls) = m_T.topLeftCorner(m_rank, m_rank).template triangularView<Upper>()
+        .solve(blockTp * m_T.topRightCorner(m_rank, m_nulls));
+  }
+  revertSchur(m_tmp, m_fT, m_U);
+  res = m_tmp * res;
+}
+
+template<typename MatrixType>
+template<int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+inline void MatrixPower<MatrixType>::revertSchur(
+    Matrix<ComplexScalar, Rows, Cols, Options, MaxRows, MaxCols>& res,
+    const ComplexMatrix& T,
+    const ComplexMatrix& U)
+{ res.noalias() = U * (T.template triangularView<Upper>() * U.adjoint()); }
+
+template<typename MatrixType>
+template<int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+inline void MatrixPower<MatrixType>::revertSchur(
+    Matrix<RealScalar, Rows, Cols, Options, MaxRows, MaxCols>& res,
+    const ComplexMatrix& T,
+    const ComplexMatrix& U)
+{ res.noalias() = (U * (T.template triangularView<Upper>() * U.adjoint())).real(); }
+
+/**
+ * \ingroup MatrixFunctions_Module
+ *
+ * \brief Proxy for the matrix power of some matrix (expression).
+ *
+ * \tparam Derived  type of the base, a matrix (expression).
+ *
+ * This class holds the arguments to the matrix power until it is
+ * assigned or evaluated for some other reason (so the argument
+ * should not be changed in the meantime). It is the return type of
+ * MatrixBase::pow() and related functions and most of the
+ * time this is the only way it is used.
+ */
+template<typename Derived>
+class MatrixPowerReturnValue : public ReturnByValue< MatrixPowerReturnValue<Derived> >
+{
+  public:
+    typedef typename Derived::PlainObject PlainObject;
+    typedef typename Derived::RealScalar RealScalar;
+
+    /**
+     * \brief Constructor.
+     *
+     * \param[in] A  %Matrix (expression), the base of the matrix power.
+     * \param[in] p  real scalar, the exponent of the matrix power.
+     */
+    MatrixPowerReturnValue(const Derived& A, RealScalar p) : m_A(A), m_p(p)
+    { }
+
+    /**
+     * \brief Compute the matrix power.
+     *
+     * \param[out] result  \f$ A^p \f$ where \p A and \p p are as in the
+     * constructor.
+     */
+    template<typename ResultType>
+    inline void evalTo(ResultType& result) const
+    { MatrixPower<PlainObject>(m_A.eval()).compute(result, m_p); }
+
+    Index rows() const { return m_A.rows(); }
+    Index cols() const { return m_A.cols(); }
+
+  private:
+    const Derived& m_A;
+    const RealScalar m_p;
+};
+
+/**
+ * \ingroup MatrixFunctions_Module
+ *
+ * \brief Proxy for the matrix power of some matrix (expression).
+ *
+ * \tparam Derived  type of the base, a matrix (expression).
+ *
+ * This class holds the arguments to the matrix power until it is
+ * assigned or evaluated for some other reason (so the argument
+ * should not be changed in the meantime). It is the return type of
+ * MatrixBase::pow() and related functions and most of the
+ * time this is the only way it is used.
+ */
+template<typename Derived>
+class MatrixComplexPowerReturnValue : public ReturnByValue< MatrixComplexPowerReturnValue<Derived> >
+{
+  public:
+    typedef typename Derived::PlainObject PlainObject;
+    typedef typename std::complex<typename Derived::RealScalar> ComplexScalar;
+
+    /**
+     * \brief Constructor.
+     *
+     * \param[in] A  %Matrix (expression), the base of the matrix power.
+     * \param[in] p  complex scalar, the exponent of the matrix power.
+     */
+    MatrixComplexPowerReturnValue(const Derived& A, const ComplexScalar& p) : m_A(A), m_p(p)
+    { }
+
+    /**
+     * \brief Compute the matrix power.
+     *
+     * Because \p p is complex, \f$ A^p \f$ is simply evaluated as \f$
+     * \exp(p \log(A)) \f$.
+     *
+     * \param[out] result  \f$ A^p \f$ where \p A and \p p are as in the
+     * constructor.
+     */
+    template<typename ResultType>
+    inline void evalTo(ResultType& result) const
+    { result = (m_p * m_A.log()).exp(); }
+
+    Index rows() const { return m_A.rows(); }
+    Index cols() const { return m_A.cols(); }
+
+  private:
+    const Derived& m_A;
+    const ComplexScalar m_p;
+};
+
+namespace internal {
+
+template<typename MatrixPowerType>
+struct traits< MatrixPowerParenthesesReturnValue<MatrixPowerType> >
+{ typedef typename MatrixPowerType::PlainObject ReturnType; };
+
+template<typename Derived>
+struct traits< MatrixPowerReturnValue<Derived> >
+{ typedef typename Derived::PlainObject ReturnType; };
+
+template<typename Derived>
+struct traits< MatrixComplexPowerReturnValue<Derived> >
+{ typedef typename Derived::PlainObject ReturnType; };
+
+}
+
+template<typename Derived>
+const MatrixPowerReturnValue<Derived> MatrixBase<Derived>::pow(const RealScalar& p) const
+{ return MatrixPowerReturnValue<Derived>(derived(), p); }
+
+template<typename Derived>
+const MatrixComplexPowerReturnValue<Derived> MatrixBase<Derived>::pow(const std::complex<RealScalar>& p) const
+{ return MatrixComplexPowerReturnValue<Derived>(derived(), p); }
+
+} // namespace Eigen
+
+#endif // EIGEN_MATRIX_POWER
diff --git a/src/EigenUnsupported/src/MatrixFunctions/MatrixSquareRoot.h b/src/EigenUnsupported/src/MatrixFunctions/MatrixSquareRoot.h
new file mode 100644
index 0000000..e363e77
--- /dev/null
+++ b/src/EigenUnsupported/src/MatrixFunctions/MatrixSquareRoot.h
@@ -0,0 +1,368 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011, 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIX_SQUARE_ROOT
+#define EIGEN_MATRIX_SQUARE_ROOT
+
+namespace Eigen { 
+
+namespace internal {
+
+// pre:  T.block(i,i,2,2) has complex conjugate eigenvalues
+// post: sqrtT.block(i,i,2,2) is square root of T.block(i,i,2,2)
+template <typename MatrixType, typename ResultType>
+void matrix_sqrt_quasi_triangular_2x2_diagonal_block(const MatrixType& T, Index i, ResultType& sqrtT)
+{
+  // TODO: This case (2-by-2 blocks with complex conjugate eigenvalues) is probably hidden somewhere
+  //       in EigenSolver. If we expose it, we could call it directly from here.
+  typedef typename traits<MatrixType>::Scalar Scalar;
+  Matrix<Scalar,2,2> block = T.template block<2,2>(i,i);
+  EigenSolver<Matrix<Scalar,2,2> > es(block);
+  sqrtT.template block<2,2>(i,i)
+    = (es.eigenvectors() * es.eigenvalues().cwiseSqrt().asDiagonal() * es.eigenvectors().inverse()).real();
+}
+
+// pre:  block structure of T is such that (i,j) is a 1x1 block,
+//       all blocks of sqrtT to left of and below (i,j) are correct
+// post: sqrtT(i,j) has the correct value
+template <typename MatrixType, typename ResultType>
+void matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
+{
+  typedef typename traits<MatrixType>::Scalar Scalar;
+  Scalar tmp = (sqrtT.row(i).segment(i+1,j-i-1) * sqrtT.col(j).segment(i+1,j-i-1)).value();
+  sqrtT.coeffRef(i,j) = (T.coeff(i,j) - tmp) / (sqrtT.coeff(i,i) + sqrtT.coeff(j,j));
+}
+
+// similar to compute1x1offDiagonalBlock()
+template <typename MatrixType, typename ResultType>
+void matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
+{
+  typedef typename traits<MatrixType>::Scalar Scalar;
+  Matrix<Scalar,1,2> rhs = T.template block<1,2>(i,j);
+  if (j-i > 1)
+    rhs -= sqrtT.block(i, i+1, 1, j-i-1) * sqrtT.block(i+1, j, j-i-1, 2);
+  Matrix<Scalar,2,2> A = sqrtT.coeff(i,i) * Matrix<Scalar,2,2>::Identity();
+  A += sqrtT.template block<2,2>(j,j).transpose();
+  sqrtT.template block<1,2>(i,j).transpose() = A.fullPivLu().solve(rhs.transpose());
+}
+
+// similar to compute1x1offDiagonalBlock()
+template <typename MatrixType, typename ResultType>
+void matrix_sqrt_quasi_triangular_2x1_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
+{
+  typedef typename traits<MatrixType>::Scalar Scalar;
+  Matrix<Scalar,2,1> rhs = T.template block<2,1>(i,j);
+  if (j-i > 2)
+    rhs -= sqrtT.block(i, i+2, 2, j-i-2) * sqrtT.block(i+2, j, j-i-2, 1);
+  Matrix<Scalar,2,2> A = sqrtT.coeff(j,j) * Matrix<Scalar,2,2>::Identity();
+  A += sqrtT.template block<2,2>(i,i);
+  sqrtT.template block<2,1>(i,j) = A.fullPivLu().solve(rhs);
+}
+
+// solves the equation A X + X B = C where all matrices are 2-by-2
+template <typename MatrixType>
+void matrix_sqrt_quasi_triangular_solve_auxiliary_equation(MatrixType& X, const MatrixType& A, const MatrixType& B, const MatrixType& C)
+{
+  typedef typename traits<MatrixType>::Scalar Scalar;
+  Matrix<Scalar,4,4> coeffMatrix = Matrix<Scalar,4,4>::Zero();
+  coeffMatrix.coeffRef(0,0) = A.coeff(0,0) + B.coeff(0,0);
+  coeffMatrix.coeffRef(1,1) = A.coeff(0,0) + B.coeff(1,1);
+  coeffMatrix.coeffRef(2,2) = A.coeff(1,1) + B.coeff(0,0);
+  coeffMatrix.coeffRef(3,3) = A.coeff(1,1) + B.coeff(1,1);
+  coeffMatrix.coeffRef(0,1) = B.coeff(1,0);
+  coeffMatrix.coeffRef(0,2) = A.coeff(0,1);
+  coeffMatrix.coeffRef(1,0) = B.coeff(0,1);
+  coeffMatrix.coeffRef(1,3) = A.coeff(0,1);
+  coeffMatrix.coeffRef(2,0) = A.coeff(1,0);
+  coeffMatrix.coeffRef(2,3) = B.coeff(1,0);
+  coeffMatrix.coeffRef(3,1) = A.coeff(1,0);
+  coeffMatrix.coeffRef(3,2) = B.coeff(0,1);
+
+  Matrix<Scalar,4,1> rhs;
+  rhs.coeffRef(0) = C.coeff(0,0);
+  rhs.coeffRef(1) = C.coeff(0,1);
+  rhs.coeffRef(2) = C.coeff(1,0);
+  rhs.coeffRef(3) = C.coeff(1,1);
+
+  Matrix<Scalar,4,1> result;
+  result = coeffMatrix.fullPivLu().solve(rhs);
+
+  X.coeffRef(0,0) = result.coeff(0);
+  X.coeffRef(0,1) = result.coeff(1);
+  X.coeffRef(1,0) = result.coeff(2);
+  X.coeffRef(1,1) = result.coeff(3);
+}
+
+// similar to compute1x1offDiagonalBlock()
+template <typename MatrixType, typename ResultType>
+void matrix_sqrt_quasi_triangular_2x2_off_diagonal_block(const MatrixType& T, Index i, Index j, ResultType& sqrtT)
+{
+  typedef typename traits<MatrixType>::Scalar Scalar;
+  Matrix<Scalar,2,2> A = sqrtT.template block<2,2>(i,i);
+  Matrix<Scalar,2,2> B = sqrtT.template block<2,2>(j,j);
+  Matrix<Scalar,2,2> C = T.template block<2,2>(i,j);
+  if (j-i > 2)
+    C -= sqrtT.block(i, i+2, 2, j-i-2) * sqrtT.block(i+2, j, j-i-2, 2);
+  Matrix<Scalar,2,2> X;
+  matrix_sqrt_quasi_triangular_solve_auxiliary_equation(X, A, B, C);
+  sqrtT.template block<2,2>(i,j) = X;
+}
+
+// pre:  T is quasi-upper-triangular and sqrtT is a zero matrix of the same size
+// post: the diagonal blocks of sqrtT are the square roots of the diagonal blocks of T
+template <typename MatrixType, typename ResultType>
+void matrix_sqrt_quasi_triangular_diagonal(const MatrixType& T, ResultType& sqrtT)
+{
+  using std::sqrt;
+  const Index size = T.rows();
+  for (Index i = 0; i < size; i++) {
+    if (i == size - 1 || T.coeff(i+1, i) == 0) {
+      eigen_assert(T(i,i) >= 0);
+      sqrtT.coeffRef(i,i) = sqrt(T.coeff(i,i));
+    }
+    else {
+      matrix_sqrt_quasi_triangular_2x2_diagonal_block(T, i, sqrtT);
+      ++i;
+    }
+  }
+}
+
+// pre:  T is quasi-upper-triangular and diagonal blocks of sqrtT are square root of diagonal blocks of T.
+// post: sqrtT is the square root of T.
+template <typename MatrixType, typename ResultType>
+void matrix_sqrt_quasi_triangular_off_diagonal(const MatrixType& T, ResultType& sqrtT)
+{
+  const Index size = T.rows();
+  for (Index j = 1; j < size; j++) {
+      if (T.coeff(j, j-1) != 0)  // if T(j-1:j, j-1:j) is a 2-by-2 block
+	continue;
+    for (Index i = j-1; i >= 0; i--) {
+      if (i > 0 && T.coeff(i, i-1) != 0)  // if T(i-1:i, i-1:i) is a 2-by-2 block
+	continue;
+      bool iBlockIs2x2 = (i < size - 1) && (T.coeff(i+1, i) != 0);
+      bool jBlockIs2x2 = (j < size - 1) && (T.coeff(j+1, j) != 0);
+      if (iBlockIs2x2 && jBlockIs2x2) 
+        matrix_sqrt_quasi_triangular_2x2_off_diagonal_block(T, i, j, sqrtT);
+      else if (iBlockIs2x2 && !jBlockIs2x2) 
+        matrix_sqrt_quasi_triangular_2x1_off_diagonal_block(T, i, j, sqrtT);
+      else if (!iBlockIs2x2 && jBlockIs2x2) 
+        matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(T, i, j, sqrtT);
+      else if (!iBlockIs2x2 && !jBlockIs2x2) 
+        matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(T, i, j, sqrtT);
+    }
+  }
+}
+
+} // end of namespace internal
+
+/** \ingroup MatrixFunctions_Module
+  * \brief Compute matrix square root of quasi-triangular matrix.
+  *
+  * \tparam  MatrixType  type of \p arg, the argument of matrix square root,
+  *                      expected to be an instantiation of the Matrix class template.
+  * \tparam  ResultType  type of \p result, where result is to be stored.
+  * \param[in]  arg      argument of matrix square root.
+  * \param[out] result   matrix square root of upper Hessenberg part of \p arg.
+  *
+  * This function computes the square root of the upper quasi-triangular matrix stored in the upper
+  * Hessenberg part of \p arg.  Only the upper Hessenberg part of \p result is updated, the rest is
+  * not touched.  See MatrixBase::sqrt() for details on how this computation is implemented.
+  *
+  * \sa MatrixSquareRoot, MatrixSquareRootQuasiTriangular
+  */
+template <typename MatrixType, typename ResultType> 
+void matrix_sqrt_quasi_triangular(const MatrixType &arg, ResultType &result)
+{
+  eigen_assert(arg.rows() == arg.cols());
+  result.resize(arg.rows(), arg.cols());
+  internal::matrix_sqrt_quasi_triangular_diagonal(arg, result);
+  internal::matrix_sqrt_quasi_triangular_off_diagonal(arg, result);
+}
+
+
+/** \ingroup MatrixFunctions_Module
+  * \brief Compute matrix square root of triangular matrix.
+  *
+  * \tparam  MatrixType  type of \p arg, the argument of matrix square root,
+  *                      expected to be an instantiation of the Matrix class template.
+  * \tparam  ResultType  type of \p result, where result is to be stored.
+  * \param[in]  arg      argument of matrix square root.
+  * \param[out] result   matrix square root of upper triangular part of \p arg.
+  *
+  * Only the upper triangular part (including the diagonal) of \p result is updated, the rest is not
+  * touched.  See MatrixBase::sqrt() for details on how this computation is implemented.
+  *
+  * \sa MatrixSquareRoot, MatrixSquareRootQuasiTriangular
+  */
+template <typename MatrixType, typename ResultType> 
+void matrix_sqrt_triangular(const MatrixType &arg, ResultType &result)
+{
+  using std::sqrt;
+  typedef typename MatrixType::Scalar Scalar;
+
+  eigen_assert(arg.rows() == arg.cols());
+
+  // Compute square root of arg and store it in upper triangular part of result
+  // This uses that the square root of triangular matrices can be computed directly.
+  result.resize(arg.rows(), arg.cols());
+  for (Index i = 0; i < arg.rows(); i++) {
+    result.coeffRef(i,i) = sqrt(arg.coeff(i,i));
+  }
+  for (Index j = 1; j < arg.cols(); j++) {
+    for (Index i = j-1; i >= 0; i--) {
+      // if i = j-1, then segment has length 0 so tmp = 0
+      Scalar tmp = (result.row(i).segment(i+1,j-i-1) * result.col(j).segment(i+1,j-i-1)).value();
+      // denominator may be zero if original matrix is singular
+      result.coeffRef(i,j) = (arg.coeff(i,j) - tmp) / (result.coeff(i,i) + result.coeff(j,j));
+    }
+  }
+}
+
+
+namespace internal {
+
+/** \ingroup MatrixFunctions_Module
+  * \brief Helper struct for computing matrix square roots of general matrices.
+  * \tparam  MatrixType  type of the argument of the matrix square root,
+  *                      expected to be an instantiation of the Matrix class template.
+  *
+  * \sa MatrixSquareRootTriangular, MatrixSquareRootQuasiTriangular, MatrixBase::sqrt()
+  */
+template <typename MatrixType, int IsComplex = NumTraits<typename internal::traits<MatrixType>::Scalar>::IsComplex>
+struct matrix_sqrt_compute
+{
+  /** \brief Compute the matrix square root
+    *
+    * \param[in]  arg     matrix whose square root is to be computed.
+    * \param[out] result  square root of \p arg.
+    *
+    * See MatrixBase::sqrt() for details on how this computation is implemented.
+    */
+  template <typename ResultType> static void run(const MatrixType &arg, ResultType &result);    
+};
+
+
+// ********** Partial specialization for real matrices **********
+
+template <typename MatrixType>
+struct matrix_sqrt_compute<MatrixType, 0>
+{
+  typedef typename MatrixType::PlainObject PlainType;
+  template <typename ResultType>
+  static void run(const MatrixType &arg, ResultType &result)
+  {
+    eigen_assert(arg.rows() == arg.cols());
+
+    // Compute Schur decomposition of arg
+    const RealSchur<PlainType> schurOfA(arg);
+    const PlainType& T = schurOfA.matrixT();
+    const PlainType& U = schurOfA.matrixU();
+    
+    // Compute square root of T
+    PlainType sqrtT = PlainType::Zero(arg.rows(), arg.cols());
+    matrix_sqrt_quasi_triangular(T, sqrtT);
+    
+    // Compute square root of arg
+    result = U * sqrtT * U.adjoint();
+  }
+};
+
+
+// ********** Partial specialization for complex matrices **********
+
+template <typename MatrixType>
+struct matrix_sqrt_compute<MatrixType, 1>
+{
+  typedef typename MatrixType::PlainObject PlainType;
+  template <typename ResultType>
+  static void run(const MatrixType &arg, ResultType &result)
+  {
+    eigen_assert(arg.rows() == arg.cols());
+
+    // Compute Schur decomposition of arg
+    const ComplexSchur<PlainType> schurOfA(arg);
+    const PlainType& T = schurOfA.matrixT();
+    const PlainType& U = schurOfA.matrixU();
+    
+    // Compute square root of T
+    PlainType sqrtT;
+    matrix_sqrt_triangular(T, sqrtT);
+    
+    // Compute square root of arg
+    result = U * (sqrtT.template triangularView<Upper>() * U.adjoint());
+  }
+};
+
+} // end namespace internal
+
+/** \ingroup MatrixFunctions_Module
+  *
+  * \brief Proxy for the matrix square root of some matrix (expression).
+  *
+  * \tparam Derived  Type of the argument to the matrix square root.
+  *
+  * This class holds the argument to the matrix square root until it
+  * is assigned or evaluated for some other reason (so the argument
+  * should not be changed in the meantime). It is the return type of
+  * MatrixBase::sqrt() and most of the time this is the only way it is
+  * used.
+  */
+template<typename Derived> class MatrixSquareRootReturnValue
+: public ReturnByValue<MatrixSquareRootReturnValue<Derived> >
+{
+  protected:
+    typedef typename internal::ref_selector<Derived>::type DerivedNested;
+
+  public:
+    /** \brief Constructor.
+      *
+      * \param[in]  src  %Matrix (expression) forming the argument of the
+      * matrix square root.
+      */
+    explicit MatrixSquareRootReturnValue(const Derived& src) : m_src(src) { }
+
+    /** \brief Compute the matrix square root.
+      *
+      * \param[out]  result  the matrix square root of \p src in the
+      * constructor.
+      */
+    template <typename ResultType>
+    inline void evalTo(ResultType& result) const
+    {
+      typedef typename internal::nested_eval<Derived, 10>::type DerivedEvalType;
+      typedef typename internal::remove_all<DerivedEvalType>::type DerivedEvalTypeClean;
+      DerivedEvalType tmp(m_src);
+      internal::matrix_sqrt_compute<DerivedEvalTypeClean>::run(tmp, result);
+    }
+
+    Index rows() const { return m_src.rows(); }
+    Index cols() const { return m_src.cols(); }
+
+  protected:
+    const DerivedNested m_src;
+};
+
+namespace internal {
+template<typename Derived>
+struct traits<MatrixSquareRootReturnValue<Derived> >
+{
+  typedef typename Derived::PlainObject ReturnType;
+};
+}
+
+template <typename Derived>
+const MatrixSquareRootReturnValue<Derived> MatrixBase<Derived>::sqrt() const
+{
+  eigen_assert(rows() == cols());
+  return MatrixSquareRootReturnValue<Derived>(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATRIX_FUNCTION
diff --git a/src/EigenUnsupported/src/MatrixFunctions/StemFunction.h b/src/EigenUnsupported/src/MatrixFunctions/StemFunction.h
new file mode 100644
index 0000000..7604df9
--- /dev/null
+++ b/src/EigenUnsupported/src/MatrixFunctions/StemFunction.h
@@ -0,0 +1,117 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010, 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STEM_FUNCTION
+#define EIGEN_STEM_FUNCTION
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \brief The exponential function (and its derivatives). */
+template <typename Scalar>
+Scalar stem_function_exp(Scalar x, int)
+{
+  using std::exp;
+  return exp(x);
+}
+
+/** \brief Cosine (and its derivatives). */
+template <typename Scalar>
+Scalar stem_function_cos(Scalar x, int n)
+{
+  using std::cos;
+  using std::sin;
+  Scalar res;
+
+  switch (n % 4) {
+  case 0: 
+    res = std::cos(x);
+    break;
+  case 1:
+    res = -std::sin(x);
+    break;
+  case 2:
+    res = -std::cos(x);
+    break;
+  case 3:
+    res = std::sin(x);
+    break;
+  }
+  return res;
+}
+
+/** \brief Sine (and its derivatives). */
+template <typename Scalar>
+Scalar stem_function_sin(Scalar x, int n)
+{
+  using std::cos;
+  using std::sin;
+  Scalar res;
+
+  switch (n % 4) {
+  case 0:
+    res = std::sin(x);
+    break;
+  case 1:
+    res = std::cos(x);
+    break;
+  case 2:
+    res = -std::sin(x);
+    break;
+  case 3:
+    res = -std::cos(x);
+    break;
+  }
+  return res;
+}
+
+/** \brief Hyperbolic cosine (and its derivatives). */
+template <typename Scalar>
+Scalar stem_function_cosh(Scalar x, int n)
+{
+  using std::cosh;
+  using std::sinh;
+  Scalar res;
+  
+  switch (n % 2) {
+  case 0:
+    res = std::cosh(x);
+    break;
+  case 1:
+    res = std::sinh(x);
+    break;
+  }
+  return res;
+}
+	
+/** \brief Hyperbolic sine (and its derivatives). */
+template <typename Scalar>
+Scalar stem_function_sinh(Scalar x, int n)
+{
+  using std::cosh;
+  using std::sinh;
+  Scalar res;
+  
+  switch (n % 2) {
+  case 0:
+    res = std::sinh(x);
+    break;
+  case 1:
+    res = std::cosh(x);
+    break;
+  }
+  return res;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_STEM_FUNCTION
diff --git a/src/EigenUnsupported/src/MoreVectorization/MathFunctions.h b/src/EigenUnsupported/src/MoreVectorization/MathFunctions.h
new file mode 100644
index 0000000..63cb28d
--- /dev/null
+++ b/src/EigenUnsupported/src/MoreVectorization/MathFunctions.h
@@ -0,0 +1,95 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Rohit Garg <rpg.314@gmail.com>
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H
+#define EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \internal \returns the arcsin of \a a (coeff-wise) */
+template<typename Packet> inline static Packet pasin(Packet a) { return std::asin(a); }
+
+#ifdef EIGEN_VECTORIZE_SSE
+
+template<> EIGEN_DONT_INLINE Packet4f pasin(Packet4f x)
+{
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5);
+  _EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5);
+  _EIGEN_DECLARE_CONST_Packet4f(3half, 1.5);
+
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(sign_mask, 0x80000000);
+
+  _EIGEN_DECLARE_CONST_Packet4f(pi, 3.141592654);
+  _EIGEN_DECLARE_CONST_Packet4f(pi_over_2, 3.141592654*0.5);
+
+  _EIGEN_DECLARE_CONST_Packet4f(asin1, 4.2163199048E-2);
+  _EIGEN_DECLARE_CONST_Packet4f(asin2, 2.4181311049E-2);
+  _EIGEN_DECLARE_CONST_Packet4f(asin3, 4.5470025998E-2);
+  _EIGEN_DECLARE_CONST_Packet4f(asin4, 7.4953002686E-2);
+  _EIGEN_DECLARE_CONST_Packet4f(asin5, 1.6666752422E-1);
+
+  Packet4f a = pabs(x);//got the absolute value
+
+  Packet4f sign_bit= _mm_and_ps(x, p4f_sign_mask);//extracted the sign bit
+
+  Packet4f z1,z2;//will need them during computation    
+
+
+//will compute the two branches for asin
+//so first compare with half
+
+  Packet4f branch_mask= _mm_cmpgt_ps(a, p4f_half);//this is to select which branch to take
+//both will be taken, and finally results will be merged
+//the branch for values >0.5
+
+    {
+//the core series expansion 
+    z1=pmadd(p4f_minus_half,a,p4f_half);
+    Packet4f x1=psqrt(z1);
+    Packet4f s1=pmadd(p4f_asin1, z1, p4f_asin2);
+    Packet4f s2=pmadd(s1, z1, p4f_asin3);
+    Packet4f s3=pmadd(s2,z1, p4f_asin4);
+    Packet4f s4=pmadd(s3,z1, p4f_asin5);
+    Packet4f temp=pmul(s4,z1);//not really a madd but a mul by z so that the next term can be a madd
+    z1=pmadd(temp,x1,x1);
+    z1=padd(z1,z1);
+    z1=psub(p4f_pi_over_2,z1);
+    }
+
+    {
+//the core series expansion 
+    Packet4f x2=a;
+    z2=pmul(x2,x2);
+    Packet4f s1=pmadd(p4f_asin1, z2, p4f_asin2);
+    Packet4f s2=pmadd(s1, z2, p4f_asin3);
+    Packet4f s3=pmadd(s2,z2, p4f_asin4);
+    Packet4f s4=pmadd(s3,z2, p4f_asin5);
+    Packet4f temp=pmul(s4,z2);//not really a madd but a mul by z so that the next term can be a madd
+    z2=pmadd(temp,x2,x2);
+    }
+
+/* select the correct result from the two branch evaluations */
+  z1  = _mm_and_ps(branch_mask, z1);
+  z2  = _mm_andnot_ps(branch_mask, z2);
+  Packet4f z  = _mm_or_ps(z1,z2);
+
+/* update the sign */
+  return _mm_xor_ps(z, sign_bit);
+}
+
+#endif // EIGEN_VECTORIZE_SSE
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/HybridNonLinearSolver.h b/src/EigenUnsupported/src/NonLinearOptimization/HybridNonLinearSolver.h
new file mode 100644
index 0000000..07c5ef0
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/HybridNonLinearSolver.h
@@ -0,0 +1,601 @@
+// -*- coding: utf-8
+// vim: set fileencoding=utf-8
+
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_HYBRIDNONLINEARSOLVER_H
+#define EIGEN_HYBRIDNONLINEARSOLVER_H
+
+namespace Eigen { 
+
+namespace HybridNonLinearSolverSpace { 
+    enum Status {
+        Running = -1,
+        ImproperInputParameters = 0,
+        RelativeErrorTooSmall = 1,
+        TooManyFunctionEvaluation = 2,
+        TolTooSmall = 3,
+        NotMakingProgressJacobian = 4,
+        NotMakingProgressIterations = 5,
+        UserAsked = 6
+    };
+}
+
+/**
+  * \ingroup NonLinearOptimization_Module
+  * \brief Finds a zero of a system of n
+  * nonlinear functions in n variables by a modification of the Powell
+  * hybrid method ("dogleg").
+  *
+  * The user must provide a subroutine which calculates the
+  * functions. The Jacobian is either provided by the user, or approximated
+  * using a forward-difference method.
+  *
+  */
+template<typename FunctorType, typename Scalar=double>
+class HybridNonLinearSolver
+{
+public:
+    typedef DenseIndex Index;
+
+    HybridNonLinearSolver(FunctorType &_functor)
+        : functor(_functor) { nfev=njev=iter = 0;  fnorm= 0.; useExternalScaling=false;}
+
+    struct Parameters {
+        Parameters()
+            : factor(Scalar(100.))
+            , maxfev(1000)
+            , xtol(numext::sqrt(NumTraits<Scalar>::epsilon()))
+            , nb_of_subdiagonals(-1)
+            , nb_of_superdiagonals(-1)
+            , epsfcn(Scalar(0.)) {}
+        Scalar factor;
+        Index maxfev;   // maximum number of function evaluation
+        Scalar xtol;
+        Index nb_of_subdiagonals;
+        Index nb_of_superdiagonals;
+        Scalar epsfcn;
+    };
+    typedef Matrix< Scalar, Dynamic, 1 > FVectorType;
+    typedef Matrix< Scalar, Dynamic, Dynamic > JacobianType;
+    /* TODO: if eigen provides a triangular storage, use it here */
+    typedef Matrix< Scalar, Dynamic, Dynamic > UpperTriangularType;
+
+    HybridNonLinearSolverSpace::Status hybrj1(
+            FVectorType  &x,
+            const Scalar tol = numext::sqrt(NumTraits<Scalar>::epsilon())
+            );
+
+    HybridNonLinearSolverSpace::Status solveInit(FVectorType  &x);
+    HybridNonLinearSolverSpace::Status solveOneStep(FVectorType  &x);
+    HybridNonLinearSolverSpace::Status solve(FVectorType  &x);
+
+    HybridNonLinearSolverSpace::Status hybrd1(
+            FVectorType  &x,
+            const Scalar tol = numext::sqrt(NumTraits<Scalar>::epsilon())
+            );
+
+    HybridNonLinearSolverSpace::Status solveNumericalDiffInit(FVectorType  &x);
+    HybridNonLinearSolverSpace::Status solveNumericalDiffOneStep(FVectorType  &x);
+    HybridNonLinearSolverSpace::Status solveNumericalDiff(FVectorType  &x);
+
+    void resetParameters(void) { parameters = Parameters(); }
+    Parameters parameters;
+    FVectorType  fvec, qtf, diag;
+    JacobianType fjac;
+    UpperTriangularType R;
+    Index nfev;
+    Index njev;
+    Index iter;
+    Scalar fnorm;
+    bool useExternalScaling; 
+private:
+    FunctorType &functor;
+    Index n;
+    Scalar sum;
+    bool sing;
+    Scalar temp;
+    Scalar delta;
+    bool jeval;
+    Index ncsuc;
+    Scalar ratio;
+    Scalar pnorm, xnorm, fnorm1;
+    Index nslow1, nslow2;
+    Index ncfail;
+    Scalar actred, prered;
+    FVectorType wa1, wa2, wa3, wa4;
+
+    HybridNonLinearSolver& operator=(const HybridNonLinearSolver&);
+};
+
+
+
+template<typename FunctorType, typename Scalar>
+HybridNonLinearSolverSpace::Status
+HybridNonLinearSolver<FunctorType,Scalar>::hybrj1(
+        FVectorType  &x,
+        const Scalar tol
+        )
+{
+    n = x.size();
+
+    /* check the input parameters for errors. */
+    if (n <= 0 || tol < 0.)
+        return HybridNonLinearSolverSpace::ImproperInputParameters;
+
+    resetParameters();
+    parameters.maxfev = 100*(n+1);
+    parameters.xtol = tol;
+    diag.setConstant(n, 1.);
+    useExternalScaling = true;
+    return solve(x);
+}
+
+template<typename FunctorType, typename Scalar>
+HybridNonLinearSolverSpace::Status
+HybridNonLinearSolver<FunctorType,Scalar>::solveInit(FVectorType  &x)
+{
+    n = x.size();
+
+    wa1.resize(n); wa2.resize(n); wa3.resize(n); wa4.resize(n);
+    fvec.resize(n);
+    qtf.resize(n);
+    fjac.resize(n, n);
+    if (!useExternalScaling)
+        diag.resize(n);
+    eigen_assert( (!useExternalScaling || diag.size()==n) && "When useExternalScaling is set, the caller must provide a valid 'diag'");
+
+    /* Function Body */
+    nfev = 0;
+    njev = 0;
+
+    /*     check the input parameters for errors. */
+    if (n <= 0 || parameters.xtol < 0. || parameters.maxfev <= 0 || parameters.factor <= 0. )
+        return HybridNonLinearSolverSpace::ImproperInputParameters;
+    if (useExternalScaling)
+        for (Index j = 0; j < n; ++j)
+            if (diag[j] <= 0.)
+                return HybridNonLinearSolverSpace::ImproperInputParameters;
+
+    /*     evaluate the function at the starting point */
+    /*     and calculate its norm. */
+    nfev = 1;
+    if ( functor(x, fvec) < 0)
+        return HybridNonLinearSolverSpace::UserAsked;
+    fnorm = fvec.stableNorm();
+
+    /*     initialize iteration counter and monitors. */
+    iter = 1;
+    ncsuc = 0;
+    ncfail = 0;
+    nslow1 = 0;
+    nslow2 = 0;
+
+    return HybridNonLinearSolverSpace::Running;
+}
+
+template<typename FunctorType, typename Scalar>
+HybridNonLinearSolverSpace::Status
+HybridNonLinearSolver<FunctorType,Scalar>::solveOneStep(FVectorType  &x)
+{
+    using std::abs;
+    
+    eigen_assert(x.size()==n); // check the caller is not cheating us
+
+    Index j;
+    std::vector<JacobiRotation<Scalar> > v_givens(n), w_givens(n);
+
+    jeval = true;
+
+    /* calculate the jacobian matrix. */
+    if ( functor.df(x, fjac) < 0)
+        return HybridNonLinearSolverSpace::UserAsked;
+    ++njev;
+
+    wa2 = fjac.colwise().blueNorm();
+
+    /* on the first iteration and if external scaling is not used, scale according */
+    /* to the norms of the columns of the initial jacobian. */
+    if (iter == 1) {
+        if (!useExternalScaling)
+            for (j = 0; j < n; ++j)
+                diag[j] = (wa2[j]==0.) ? 1. : wa2[j];
+
+        /* on the first iteration, calculate the norm of the scaled x */
+        /* and initialize the step bound delta. */
+        xnorm = diag.cwiseProduct(x).stableNorm();
+        delta = parameters.factor * xnorm;
+        if (delta == 0.)
+            delta = parameters.factor;
+    }
+
+    /* compute the qr factorization of the jacobian. */
+    HouseholderQR<JacobianType> qrfac(fjac); // no pivoting:
+
+    /* copy the triangular factor of the qr factorization into r. */
+    R = qrfac.matrixQR();
+
+    /* accumulate the orthogonal factor in fjac. */
+    fjac = qrfac.householderQ();
+
+    /* form (q transpose)*fvec and store in qtf. */
+    qtf = fjac.transpose() * fvec;
+
+    /* rescale if necessary. */
+    if (!useExternalScaling)
+        diag = diag.cwiseMax(wa2);
+
+    while (true) {
+        /* determine the direction p. */
+        internal::dogleg<Scalar>(R, diag, qtf, delta, wa1);
+
+        /* store the direction p and x + p. calculate the norm of p. */
+        wa1 = -wa1;
+        wa2 = x + wa1;
+        pnorm = diag.cwiseProduct(wa1).stableNorm();
+
+        /* on the first iteration, adjust the initial step bound. */
+        if (iter == 1)
+            delta = (std::min)(delta,pnorm);
+
+        /* evaluate the function at x + p and calculate its norm. */
+        if ( functor(wa2, wa4) < 0)
+            return HybridNonLinearSolverSpace::UserAsked;
+        ++nfev;
+        fnorm1 = wa4.stableNorm();
+
+        /* compute the scaled actual reduction. */
+        actred = -1.;
+        if (fnorm1 < fnorm) /* Computing 2nd power */
+            actred = 1. - numext::abs2(fnorm1 / fnorm);
+
+        /* compute the scaled predicted reduction. */
+        wa3 = R.template triangularView<Upper>()*wa1 + qtf;
+        temp = wa3.stableNorm();
+        prered = 0.;
+        if (temp < fnorm) /* Computing 2nd power */
+            prered = 1. - numext::abs2(temp / fnorm);
+
+        /* compute the ratio of the actual to the predicted reduction. */
+        ratio = 0.;
+        if (prered > 0.)
+            ratio = actred / prered;
+
+        /* update the step bound. */
+        if (ratio < Scalar(.1)) {
+            ncsuc = 0;
+            ++ncfail;
+            delta = Scalar(.5) * delta;
+        } else {
+            ncfail = 0;
+            ++ncsuc;
+            if (ratio >= Scalar(.5) || ncsuc > 1)
+                delta = (std::max)(delta, pnorm / Scalar(.5));
+            if (abs(ratio - 1.) <= Scalar(.1)) {
+                delta = pnorm / Scalar(.5);
+            }
+        }
+
+        /* test for successful iteration. */
+        if (ratio >= Scalar(1e-4)) {
+            /* successful iteration. update x, fvec, and their norms. */
+            x = wa2;
+            wa2 = diag.cwiseProduct(x);
+            fvec = wa4;
+            xnorm = wa2.stableNorm();
+            fnorm = fnorm1;
+            ++iter;
+        }
+
+        /* determine the progress of the iteration. */
+        ++nslow1;
+        if (actred >= Scalar(.001))
+            nslow1 = 0;
+        if (jeval)
+            ++nslow2;
+        if (actred >= Scalar(.1))
+            nslow2 = 0;
+
+        /* test for convergence. */
+        if (delta <= parameters.xtol * xnorm || fnorm == 0.)
+            return HybridNonLinearSolverSpace::RelativeErrorTooSmall;
+
+        /* tests for termination and stringent tolerances. */
+        if (nfev >= parameters.maxfev)
+            return HybridNonLinearSolverSpace::TooManyFunctionEvaluation;
+        if (Scalar(.1) * (std::max)(Scalar(.1) * delta, pnorm) <= NumTraits<Scalar>::epsilon() * xnorm)
+            return HybridNonLinearSolverSpace::TolTooSmall;
+        if (nslow2 == 5)
+            return HybridNonLinearSolverSpace::NotMakingProgressJacobian;
+        if (nslow1 == 10)
+            return HybridNonLinearSolverSpace::NotMakingProgressIterations;
+
+        /* criterion for recalculating jacobian. */
+        if (ncfail == 2)
+            break; // leave inner loop and go for the next outer loop iteration
+
+        /* calculate the rank one modification to the jacobian */
+        /* and update qtf if necessary. */
+        wa1 = diag.cwiseProduct( diag.cwiseProduct(wa1)/pnorm );
+        wa2 = fjac.transpose() * wa4;
+        if (ratio >= Scalar(1e-4))
+            qtf = wa2;
+        wa2 = (wa2-wa3)/pnorm;
+
+        /* compute the qr factorization of the updated jacobian. */
+        internal::r1updt<Scalar>(R, wa1, v_givens, w_givens, wa2, wa3, &sing);
+        internal::r1mpyq<Scalar>(n, n, fjac.data(), v_givens, w_givens);
+        internal::r1mpyq<Scalar>(1, n, qtf.data(), v_givens, w_givens);
+
+        jeval = false;
+    }
+    return HybridNonLinearSolverSpace::Running;
+}
+
+template<typename FunctorType, typename Scalar>
+HybridNonLinearSolverSpace::Status
+HybridNonLinearSolver<FunctorType,Scalar>::solve(FVectorType  &x)
+{
+    HybridNonLinearSolverSpace::Status status = solveInit(x);
+    if (status==HybridNonLinearSolverSpace::ImproperInputParameters)
+        return status;
+    while (status==HybridNonLinearSolverSpace::Running)
+        status = solveOneStep(x);
+    return status;
+}
+
+
+
+template<typename FunctorType, typename Scalar>
+HybridNonLinearSolverSpace::Status
+HybridNonLinearSolver<FunctorType,Scalar>::hybrd1(
+        FVectorType  &x,
+        const Scalar tol
+        )
+{
+    n = x.size();
+
+    /* check the input parameters for errors. */
+    if (n <= 0 || tol < 0.)
+        return HybridNonLinearSolverSpace::ImproperInputParameters;
+
+    resetParameters();
+    parameters.maxfev = 200*(n+1);
+    parameters.xtol = tol;
+
+    diag.setConstant(n, 1.);
+    useExternalScaling = true;
+    return solveNumericalDiff(x);
+}
+
+template<typename FunctorType, typename Scalar>
+HybridNonLinearSolverSpace::Status
+HybridNonLinearSolver<FunctorType,Scalar>::solveNumericalDiffInit(FVectorType  &x)
+{
+    n = x.size();
+
+    if (parameters.nb_of_subdiagonals<0) parameters.nb_of_subdiagonals= n-1;
+    if (parameters.nb_of_superdiagonals<0) parameters.nb_of_superdiagonals= n-1;
+
+    wa1.resize(n); wa2.resize(n); wa3.resize(n); wa4.resize(n);
+    qtf.resize(n);
+    fjac.resize(n, n);
+    fvec.resize(n);
+    if (!useExternalScaling)
+        diag.resize(n);
+    eigen_assert( (!useExternalScaling || diag.size()==n) && "When useExternalScaling is set, the caller must provide a valid 'diag'");
+
+    /* Function Body */
+    nfev = 0;
+    njev = 0;
+
+    /*     check the input parameters for errors. */
+    if (n <= 0 || parameters.xtol < 0. || parameters.maxfev <= 0 || parameters.nb_of_subdiagonals< 0 || parameters.nb_of_superdiagonals< 0 || parameters.factor <= 0. )
+        return HybridNonLinearSolverSpace::ImproperInputParameters;
+    if (useExternalScaling)
+        for (Index j = 0; j < n; ++j)
+            if (diag[j] <= 0.)
+                return HybridNonLinearSolverSpace::ImproperInputParameters;
+
+    /*     evaluate the function at the starting point */
+    /*     and calculate its norm. */
+    nfev = 1;
+    if ( functor(x, fvec) < 0)
+        return HybridNonLinearSolverSpace::UserAsked;
+    fnorm = fvec.stableNorm();
+
+    /*     initialize iteration counter and monitors. */
+    iter = 1;
+    ncsuc = 0;
+    ncfail = 0;
+    nslow1 = 0;
+    nslow2 = 0;
+
+    return HybridNonLinearSolverSpace::Running;
+}
+
+template<typename FunctorType, typename Scalar>
+HybridNonLinearSolverSpace::Status
+HybridNonLinearSolver<FunctorType,Scalar>::solveNumericalDiffOneStep(FVectorType  &x)
+{
+    using std::sqrt;
+    using std::abs;
+    
+    assert(x.size()==n); // check the caller is not cheating us
+
+    Index j;
+    std::vector<JacobiRotation<Scalar> > v_givens(n), w_givens(n);
+
+    jeval = true;
+    if (parameters.nb_of_subdiagonals<0) parameters.nb_of_subdiagonals= n-1;
+    if (parameters.nb_of_superdiagonals<0) parameters.nb_of_superdiagonals= n-1;
+
+    /* calculate the jacobian matrix. */
+    if (internal::fdjac1(functor, x, fvec, fjac, parameters.nb_of_subdiagonals, parameters.nb_of_superdiagonals, parameters.epsfcn) <0)
+        return HybridNonLinearSolverSpace::UserAsked;
+    nfev += (std::min)(parameters.nb_of_subdiagonals+parameters.nb_of_superdiagonals+ 1, n);
+
+    wa2 = fjac.colwise().blueNorm();
+
+    /* on the first iteration and if external scaling is not used, scale according */
+    /* to the norms of the columns of the initial jacobian. */
+    if (iter == 1) {
+        if (!useExternalScaling)
+            for (j = 0; j < n; ++j)
+                diag[j] = (wa2[j]==0.) ? 1. : wa2[j];
+
+        /* on the first iteration, calculate the norm of the scaled x */
+        /* and initialize the step bound delta. */
+        xnorm = diag.cwiseProduct(x).stableNorm();
+        delta = parameters.factor * xnorm;
+        if (delta == 0.)
+            delta = parameters.factor;
+    }
+
+    /* compute the qr factorization of the jacobian. */
+    HouseholderQR<JacobianType> qrfac(fjac); // no pivoting:
+
+    /* copy the triangular factor of the qr factorization into r. */
+    R = qrfac.matrixQR();
+
+    /* accumulate the orthogonal factor in fjac. */
+    fjac = qrfac.householderQ();
+
+    /* form (q transpose)*fvec and store in qtf. */
+    qtf = fjac.transpose() * fvec;
+
+    /* rescale if necessary. */
+    if (!useExternalScaling)
+        diag = diag.cwiseMax(wa2);
+
+    while (true) {
+        /* determine the direction p. */
+        internal::dogleg<Scalar>(R, diag, qtf, delta, wa1);
+
+        /* store the direction p and x + p. calculate the norm of p. */
+        wa1 = -wa1;
+        wa2 = x + wa1;
+        pnorm = diag.cwiseProduct(wa1).stableNorm();
+
+        /* on the first iteration, adjust the initial step bound. */
+        if (iter == 1)
+            delta = (std::min)(delta,pnorm);
+
+        /* evaluate the function at x + p and calculate its norm. */
+        if ( functor(wa2, wa4) < 0)
+            return HybridNonLinearSolverSpace::UserAsked;
+        ++nfev;
+        fnorm1 = wa4.stableNorm();
+
+        /* compute the scaled actual reduction. */
+        actred = -1.;
+        if (fnorm1 < fnorm) /* Computing 2nd power */
+            actred = 1. - numext::abs2(fnorm1 / fnorm);
+
+        /* compute the scaled predicted reduction. */
+        wa3 = R.template triangularView<Upper>()*wa1 + qtf;
+        temp = wa3.stableNorm();
+        prered = 0.;
+        if (temp < fnorm) /* Computing 2nd power */
+            prered = 1. - numext::abs2(temp / fnorm);
+
+        /* compute the ratio of the actual to the predicted reduction. */
+        ratio = 0.;
+        if (prered > 0.)
+            ratio = actred / prered;
+
+        /* update the step bound. */
+        if (ratio < Scalar(.1)) {
+            ncsuc = 0;
+            ++ncfail;
+            delta = Scalar(.5) * delta;
+        } else {
+            ncfail = 0;
+            ++ncsuc;
+            if (ratio >= Scalar(.5) || ncsuc > 1)
+                delta = (std::max)(delta, pnorm / Scalar(.5));
+            if (abs(ratio - 1.) <= Scalar(.1)) {
+                delta = pnorm / Scalar(.5);
+            }
+        }
+
+        /* test for successful iteration. */
+        if (ratio >= Scalar(1e-4)) {
+            /* successful iteration. update x, fvec, and their norms. */
+            x = wa2;
+            wa2 = diag.cwiseProduct(x);
+            fvec = wa4;
+            xnorm = wa2.stableNorm();
+            fnorm = fnorm1;
+            ++iter;
+        }
+
+        /* determine the progress of the iteration. */
+        ++nslow1;
+        if (actred >= Scalar(.001))
+            nslow1 = 0;
+        if (jeval)
+            ++nslow2;
+        if (actred >= Scalar(.1))
+            nslow2 = 0;
+
+        /* test for convergence. */
+        if (delta <= parameters.xtol * xnorm || fnorm == 0.)
+            return HybridNonLinearSolverSpace::RelativeErrorTooSmall;
+
+        /* tests for termination and stringent tolerances. */
+        if (nfev >= parameters.maxfev)
+            return HybridNonLinearSolverSpace::TooManyFunctionEvaluation;
+        if (Scalar(.1) * (std::max)(Scalar(.1) * delta, pnorm) <= NumTraits<Scalar>::epsilon() * xnorm)
+            return HybridNonLinearSolverSpace::TolTooSmall;
+        if (nslow2 == 5)
+            return HybridNonLinearSolverSpace::NotMakingProgressJacobian;
+        if (nslow1 == 10)
+            return HybridNonLinearSolverSpace::NotMakingProgressIterations;
+
+        /* criterion for recalculating jacobian. */
+        if (ncfail == 2)
+            break; // leave inner loop and go for the next outer loop iteration
+
+        /* calculate the rank one modification to the jacobian */
+        /* and update qtf if necessary. */
+        wa1 = diag.cwiseProduct( diag.cwiseProduct(wa1)/pnorm );
+        wa2 = fjac.transpose() * wa4;
+        if (ratio >= Scalar(1e-4))
+            qtf = wa2;
+        wa2 = (wa2-wa3)/pnorm;
+
+        /* compute the qr factorization of the updated jacobian. */
+        internal::r1updt<Scalar>(R, wa1, v_givens, w_givens, wa2, wa3, &sing);
+        internal::r1mpyq<Scalar>(n, n, fjac.data(), v_givens, w_givens);
+        internal::r1mpyq<Scalar>(1, n, qtf.data(), v_givens, w_givens);
+
+        jeval = false;
+    }
+    return HybridNonLinearSolverSpace::Running;
+}
+
+template<typename FunctorType, typename Scalar>
+HybridNonLinearSolverSpace::Status
+HybridNonLinearSolver<FunctorType,Scalar>::solveNumericalDiff(FVectorType  &x)
+{
+    HybridNonLinearSolverSpace::Status status = solveNumericalDiffInit(x);
+    if (status==HybridNonLinearSolverSpace::ImproperInputParameters)
+        return status;
+    while (status==HybridNonLinearSolverSpace::Running)
+        status = solveNumericalDiffOneStep(x);
+    return status;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_HYBRIDNONLINEARSOLVER_H
+
+//vim: ai ts=4 sts=4 et sw=4
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/LevenbergMarquardt.h b/src/EigenUnsupported/src/NonLinearOptimization/LevenbergMarquardt.h
new file mode 100644
index 0000000..fe3b79c
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/LevenbergMarquardt.h
@@ -0,0 +1,657 @@
+// -*- coding: utf-8
+// vim: set fileencoding=utf-8
+
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_LEVENBERGMARQUARDT__H
+#define EIGEN_LEVENBERGMARQUARDT__H
+
+namespace Eigen { 
+
+namespace LevenbergMarquardtSpace {
+    enum Status {
+        NotStarted = -2,
+        Running = -1,
+        ImproperInputParameters = 0,
+        RelativeReductionTooSmall = 1,
+        RelativeErrorTooSmall = 2,
+        RelativeErrorAndReductionTooSmall = 3,
+        CosinusTooSmall = 4,
+        TooManyFunctionEvaluation = 5,
+        FtolTooSmall = 6,
+        XtolTooSmall = 7,
+        GtolTooSmall = 8,
+        UserAsked = 9
+    };
+}
+
+
+
+/**
+  * \ingroup NonLinearOptimization_Module
+  * \brief Performs non linear optimization over a non-linear function,
+  * using a variant of the Levenberg Marquardt algorithm.
+  *
+  * Check wikipedia for more information.
+  * http://en.wikipedia.org/wiki/Levenberg%E2%80%93Marquardt_algorithm
+  */
+template<typename FunctorType, typename Scalar=double>
+class LevenbergMarquardt
+{
+    static Scalar sqrt_epsilon()
+    {
+      using std::sqrt;
+      return sqrt(NumTraits<Scalar>::epsilon());
+    }
+    
+public:
+    LevenbergMarquardt(FunctorType &_functor)
+        : functor(_functor) { nfev = njev = iter = 0;  fnorm = gnorm = 0.; useExternalScaling=false; }
+
+    typedef DenseIndex Index;
+    
+    struct Parameters {
+        Parameters()
+            : factor(Scalar(100.))
+            , maxfev(400)
+            , ftol(sqrt_epsilon())
+            , xtol(sqrt_epsilon())
+            , gtol(Scalar(0.))
+            , epsfcn(Scalar(0.)) {}
+        Scalar factor;
+        Index maxfev;   // maximum number of function evaluation
+        Scalar ftol;
+        Scalar xtol;
+        Scalar gtol;
+        Scalar epsfcn;
+    };
+
+    typedef Matrix< Scalar, Dynamic, 1 > FVectorType;
+    typedef Matrix< Scalar, Dynamic, Dynamic > JacobianType;
+
+    LevenbergMarquardtSpace::Status lmder1(
+            FVectorType &x,
+            const Scalar tol = sqrt_epsilon()
+            );
+
+    LevenbergMarquardtSpace::Status minimize(FVectorType &x);
+    LevenbergMarquardtSpace::Status minimizeInit(FVectorType &x);
+    LevenbergMarquardtSpace::Status minimizeOneStep(FVectorType &x);
+
+    static LevenbergMarquardtSpace::Status lmdif1(
+            FunctorType &functor,
+            FVectorType &x,
+            Index *nfev,
+            const Scalar tol = sqrt_epsilon()
+            );
+
+    LevenbergMarquardtSpace::Status lmstr1(
+            FVectorType  &x,
+            const Scalar tol = sqrt_epsilon()
+            );
+
+    LevenbergMarquardtSpace::Status minimizeOptimumStorage(FVectorType  &x);
+    LevenbergMarquardtSpace::Status minimizeOptimumStorageInit(FVectorType  &x);
+    LevenbergMarquardtSpace::Status minimizeOptimumStorageOneStep(FVectorType  &x);
+
+    void resetParameters(void) { parameters = Parameters(); }
+
+    Parameters parameters;
+    FVectorType  fvec, qtf, diag;
+    JacobianType fjac;
+    PermutationMatrix<Dynamic,Dynamic> permutation;
+    Index nfev;
+    Index njev;
+    Index iter;
+    Scalar fnorm, gnorm;
+    bool useExternalScaling; 
+
+    Scalar lm_param(void) { return par; }
+private:
+    
+    FunctorType &functor;
+    Index n;
+    Index m;
+    FVectorType wa1, wa2, wa3, wa4;
+
+    Scalar par, sum;
+    Scalar temp, temp1, temp2;
+    Scalar delta;
+    Scalar ratio;
+    Scalar pnorm, xnorm, fnorm1, actred, dirder, prered;
+
+    LevenbergMarquardt& operator=(const LevenbergMarquardt&);
+};
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::lmder1(
+        FVectorType  &x,
+        const Scalar tol
+        )
+{
+    n = x.size();
+    m = functor.values();
+
+    /* check the input parameters for errors. */
+    if (n <= 0 || m < n || tol < 0.)
+        return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    resetParameters();
+    parameters.ftol = tol;
+    parameters.xtol = tol;
+    parameters.maxfev = 100*(n+1);
+
+    return minimize(x);
+}
+
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::minimize(FVectorType  &x)
+{
+    LevenbergMarquardtSpace::Status status = minimizeInit(x);
+    if (status==LevenbergMarquardtSpace::ImproperInputParameters)
+        return status;
+    do {
+        status = minimizeOneStep(x);
+    } while (status==LevenbergMarquardtSpace::Running);
+    return status;
+}
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::minimizeInit(FVectorType  &x)
+{
+    n = x.size();
+    m = functor.values();
+
+    wa1.resize(n); wa2.resize(n); wa3.resize(n);
+    wa4.resize(m);
+    fvec.resize(m);
+    fjac.resize(m, n);
+    if (!useExternalScaling)
+        diag.resize(n);
+    eigen_assert( (!useExternalScaling || diag.size()==n) && "When useExternalScaling is set, the caller must provide a valid 'diag'");
+    qtf.resize(n);
+
+    /* Function Body */
+    nfev = 0;
+    njev = 0;
+
+    /*     check the input parameters for errors. */
+    if (n <= 0 || m < n || parameters.ftol < 0. || parameters.xtol < 0. || parameters.gtol < 0. || parameters.maxfev <= 0 || parameters.factor <= 0.)
+        return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    if (useExternalScaling)
+        for (Index j = 0; j < n; ++j)
+            if (diag[j] <= 0.)
+                return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    /*     evaluate the function at the starting point */
+    /*     and calculate its norm. */
+    nfev = 1;
+    if ( functor(x, fvec) < 0)
+        return LevenbergMarquardtSpace::UserAsked;
+    fnorm = fvec.stableNorm();
+
+    /*     initialize levenberg-marquardt parameter and iteration counter. */
+    par = 0.;
+    iter = 1;
+
+    return LevenbergMarquardtSpace::NotStarted;
+}
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::minimizeOneStep(FVectorType  &x)
+{
+    using std::abs;
+    using std::sqrt;
+
+    eigen_assert(x.size()==n); // check the caller is not cheating us
+
+    /* calculate the jacobian matrix. */
+    Index df_ret = functor.df(x, fjac);
+    if (df_ret<0)
+        return LevenbergMarquardtSpace::UserAsked;
+    if (df_ret>0)
+        // numerical diff, we evaluated the function df_ret times
+        nfev += df_ret;
+    else njev++;
+
+    /* compute the qr factorization of the jacobian. */
+    wa2 = fjac.colwise().blueNorm();
+    ColPivHouseholderQR<JacobianType> qrfac(fjac);
+    fjac = qrfac.matrixQR();
+    permutation = qrfac.colsPermutation();
+
+    /* on the first iteration and if external scaling is not used, scale according */
+    /* to the norms of the columns of the initial jacobian. */
+    if (iter == 1) {
+        if (!useExternalScaling)
+            for (Index j = 0; j < n; ++j)
+                diag[j] = (wa2[j]==0.)? 1. : wa2[j];
+
+        /* on the first iteration, calculate the norm of the scaled x */
+        /* and initialize the step bound delta. */
+        xnorm = diag.cwiseProduct(x).stableNorm();
+        delta = parameters.factor * xnorm;
+        if (delta == 0.)
+            delta = parameters.factor;
+    }
+
+    /* form (q transpose)*fvec and store the first n components in */
+    /* qtf. */
+    wa4 = fvec;
+    wa4.applyOnTheLeft(qrfac.householderQ().adjoint());
+    qtf = wa4.head(n);
+
+    /* compute the norm of the scaled gradient. */
+    gnorm = 0.;
+    if (fnorm != 0.)
+        for (Index j = 0; j < n; ++j)
+            if (wa2[permutation.indices()[j]] != 0.)
+                gnorm = (std::max)(gnorm, abs( fjac.col(j).head(j+1).dot(qtf.head(j+1)/fnorm) / wa2[permutation.indices()[j]]));
+
+    /* test for convergence of the gradient norm. */
+    if (gnorm <= parameters.gtol)
+        return LevenbergMarquardtSpace::CosinusTooSmall;
+
+    /* rescale if necessary. */
+    if (!useExternalScaling)
+        diag = diag.cwiseMax(wa2);
+
+    do {
+
+        /* determine the levenberg-marquardt parameter. */
+        internal::lmpar2<Scalar>(qrfac, diag, qtf, delta, par, wa1);
+
+        /* store the direction p and x + p. calculate the norm of p. */
+        wa1 = -wa1;
+        wa2 = x + wa1;
+        pnorm = diag.cwiseProduct(wa1).stableNorm();
+
+        /* on the first iteration, adjust the initial step bound. */
+        if (iter == 1)
+            delta = (std::min)(delta,pnorm);
+
+        /* evaluate the function at x + p and calculate its norm. */
+        if ( functor(wa2, wa4) < 0)
+            return LevenbergMarquardtSpace::UserAsked;
+        ++nfev;
+        fnorm1 = wa4.stableNorm();
+
+        /* compute the scaled actual reduction. */
+        actred = -1.;
+        if (Scalar(.1) * fnorm1 < fnorm)
+            actred = 1. - numext::abs2(fnorm1 / fnorm);
+
+        /* compute the scaled predicted reduction and */
+        /* the scaled directional derivative. */
+        wa3 = fjac.template triangularView<Upper>() * (qrfac.colsPermutation().inverse() *wa1);
+        temp1 = numext::abs2(wa3.stableNorm() / fnorm);
+        temp2 = numext::abs2(sqrt(par) * pnorm / fnorm);
+        prered = temp1 + temp2 / Scalar(.5);
+        dirder = -(temp1 + temp2);
+
+        /* compute the ratio of the actual to the predicted */
+        /* reduction. */
+        ratio = 0.;
+        if (prered != 0.)
+            ratio = actred / prered;
+
+        /* update the step bound. */
+        if (ratio <= Scalar(.25)) {
+            if (actred >= 0.)
+                temp = Scalar(.5);
+            if (actred < 0.)
+                temp = Scalar(.5) * dirder / (dirder + Scalar(.5) * actred);
+            if (Scalar(.1) * fnorm1 >= fnorm || temp < Scalar(.1))
+                temp = Scalar(.1);
+            /* Computing MIN */
+            delta = temp * (std::min)(delta, pnorm / Scalar(.1));
+            par /= temp;
+        } else if (!(par != 0. && ratio < Scalar(.75))) {
+            delta = pnorm / Scalar(.5);
+            par = Scalar(.5) * par;
+        }
+
+        /* test for successful iteration. */
+        if (ratio >= Scalar(1e-4)) {
+            /* successful iteration. update x, fvec, and their norms. */
+            x = wa2;
+            wa2 = diag.cwiseProduct(x);
+            fvec = wa4;
+            xnorm = wa2.stableNorm();
+            fnorm = fnorm1;
+            ++iter;
+        }
+
+        /* tests for convergence. */
+        if (abs(actred) <= parameters.ftol && prered <= parameters.ftol && Scalar(.5) * ratio <= 1. && delta <= parameters.xtol * xnorm)
+            return LevenbergMarquardtSpace::RelativeErrorAndReductionTooSmall;
+        if (abs(actred) <= parameters.ftol && prered <= parameters.ftol && Scalar(.5) * ratio <= 1.)
+            return LevenbergMarquardtSpace::RelativeReductionTooSmall;
+        if (delta <= parameters.xtol * xnorm)
+            return LevenbergMarquardtSpace::RelativeErrorTooSmall;
+
+        /* tests for termination and stringent tolerances. */
+        if (nfev >= parameters.maxfev)
+            return LevenbergMarquardtSpace::TooManyFunctionEvaluation;
+        if (abs(actred) <= NumTraits<Scalar>::epsilon() && prered <= NumTraits<Scalar>::epsilon() && Scalar(.5) * ratio <= 1.)
+            return LevenbergMarquardtSpace::FtolTooSmall;
+        if (delta <= NumTraits<Scalar>::epsilon() * xnorm)
+            return LevenbergMarquardtSpace::XtolTooSmall;
+        if (gnorm <= NumTraits<Scalar>::epsilon())
+            return LevenbergMarquardtSpace::GtolTooSmall;
+
+    } while (ratio < Scalar(1e-4));
+
+    return LevenbergMarquardtSpace::Running;
+}
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::lmstr1(
+        FVectorType  &x,
+        const Scalar tol
+        )
+{
+    n = x.size();
+    m = functor.values();
+
+    /* check the input parameters for errors. */
+    if (n <= 0 || m < n || tol < 0.)
+        return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    resetParameters();
+    parameters.ftol = tol;
+    parameters.xtol = tol;
+    parameters.maxfev = 100*(n+1);
+
+    return minimizeOptimumStorage(x);
+}
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::minimizeOptimumStorageInit(FVectorType  &x)
+{
+    n = x.size();
+    m = functor.values();
+
+    wa1.resize(n); wa2.resize(n); wa3.resize(n);
+    wa4.resize(m);
+    fvec.resize(m);
+    // Only R is stored in fjac. Q is only used to compute 'qtf', which is
+    // Q.transpose()*rhs. qtf will be updated using givens rotation,
+    // instead of storing them in Q.
+    // The purpose it to only use a nxn matrix, instead of mxn here, so
+    // that we can handle cases where m>>n :
+    fjac.resize(n, n);
+    if (!useExternalScaling)
+        diag.resize(n);
+    eigen_assert( (!useExternalScaling || diag.size()==n) && "When useExternalScaling is set, the caller must provide a valid 'diag'");
+    qtf.resize(n);
+
+    /* Function Body */
+    nfev = 0;
+    njev = 0;
+
+    /*     check the input parameters for errors. */
+    if (n <= 0 || m < n || parameters.ftol < 0. || parameters.xtol < 0. || parameters.gtol < 0. || parameters.maxfev <= 0 || parameters.factor <= 0.)
+        return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    if (useExternalScaling)
+        for (Index j = 0; j < n; ++j)
+            if (diag[j] <= 0.)
+                return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    /*     evaluate the function at the starting point */
+    /*     and calculate its norm. */
+    nfev = 1;
+    if ( functor(x, fvec) < 0)
+        return LevenbergMarquardtSpace::UserAsked;
+    fnorm = fvec.stableNorm();
+
+    /*     initialize levenberg-marquardt parameter and iteration counter. */
+    par = 0.;
+    iter = 1;
+
+    return LevenbergMarquardtSpace::NotStarted;
+}
+
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::minimizeOptimumStorageOneStep(FVectorType  &x)
+{
+    using std::abs;
+    using std::sqrt;
+    
+    eigen_assert(x.size()==n); // check the caller is not cheating us
+
+    Index i, j;
+    bool sing;
+
+    /* compute the qr factorization of the jacobian matrix */
+    /* calculated one row at a time, while simultaneously */
+    /* forming (q transpose)*fvec and storing the first */
+    /* n components in qtf. */
+    qtf.fill(0.);
+    fjac.fill(0.);
+    Index rownb = 2;
+    for (i = 0; i < m; ++i) {
+        if (functor.df(x, wa3, rownb) < 0) return LevenbergMarquardtSpace::UserAsked;
+        internal::rwupdt<Scalar>(fjac, wa3, qtf, fvec[i]);
+        ++rownb;
+    }
+    ++njev;
+
+    /* if the jacobian is rank deficient, call qrfac to */
+    /* reorder its columns and update the components of qtf. */
+    sing = false;
+    for (j = 0; j < n; ++j) {
+        if (fjac(j,j) == 0.)
+            sing = true;
+        wa2[j] = fjac.col(j).head(j).stableNorm();
+    }
+    permutation.setIdentity(n);
+    if (sing) {
+        wa2 = fjac.colwise().blueNorm();
+        // TODO We have no unit test covering this code path, do not modify
+        // until it is carefully tested
+        ColPivHouseholderQR<JacobianType> qrfac(fjac);
+        fjac = qrfac.matrixQR();
+        wa1 = fjac.diagonal();
+        fjac.diagonal() = qrfac.hCoeffs();
+        permutation = qrfac.colsPermutation();
+        // TODO : avoid this:
+        for(Index ii=0; ii< fjac.cols(); ii++) fjac.col(ii).segment(ii+1, fjac.rows()-ii-1) *= fjac(ii,ii); // rescale vectors
+
+        for (j = 0; j < n; ++j) {
+            if (fjac(j,j) != 0.) {
+                sum = 0.;
+                for (i = j; i < n; ++i)
+                    sum += fjac(i,j) * qtf[i];
+                temp = -sum / fjac(j,j);
+                for (i = j; i < n; ++i)
+                    qtf[i] += fjac(i,j) * temp;
+            }
+            fjac(j,j) = wa1[j];
+        }
+    }
+
+    /* on the first iteration and if external scaling is not used, scale according */
+    /* to the norms of the columns of the initial jacobian. */
+    if (iter == 1) {
+        if (!useExternalScaling)
+            for (j = 0; j < n; ++j)
+                diag[j] = (wa2[j]==0.)? 1. : wa2[j];
+
+        /* on the first iteration, calculate the norm of the scaled x */
+        /* and initialize the step bound delta. */
+        xnorm = diag.cwiseProduct(x).stableNorm();
+        delta = parameters.factor * xnorm;
+        if (delta == 0.)
+            delta = parameters.factor;
+    }
+
+    /* compute the norm of the scaled gradient. */
+    gnorm = 0.;
+    if (fnorm != 0.)
+        for (j = 0; j < n; ++j)
+            if (wa2[permutation.indices()[j]] != 0.)
+                gnorm = (std::max)(gnorm, abs( fjac.col(j).head(j+1).dot(qtf.head(j+1)/fnorm) / wa2[permutation.indices()[j]]));
+
+    /* test for convergence of the gradient norm. */
+    if (gnorm <= parameters.gtol)
+        return LevenbergMarquardtSpace::CosinusTooSmall;
+
+    /* rescale if necessary. */
+    if (!useExternalScaling)
+        diag = diag.cwiseMax(wa2);
+
+    do {
+
+        /* determine the levenberg-marquardt parameter. */
+        internal::lmpar<Scalar>(fjac, permutation.indices(), diag, qtf, delta, par, wa1);
+
+        /* store the direction p and x + p. calculate the norm of p. */
+        wa1 = -wa1;
+        wa2 = x + wa1;
+        pnorm = diag.cwiseProduct(wa1).stableNorm();
+
+        /* on the first iteration, adjust the initial step bound. */
+        if (iter == 1)
+            delta = (std::min)(delta,pnorm);
+
+        /* evaluate the function at x + p and calculate its norm. */
+        if ( functor(wa2, wa4) < 0)
+            return LevenbergMarquardtSpace::UserAsked;
+        ++nfev;
+        fnorm1 = wa4.stableNorm();
+
+        /* compute the scaled actual reduction. */
+        actred = -1.;
+        if (Scalar(.1) * fnorm1 < fnorm)
+            actred = 1. - numext::abs2(fnorm1 / fnorm);
+
+        /* compute the scaled predicted reduction and */
+        /* the scaled directional derivative. */
+        wa3 = fjac.topLeftCorner(n,n).template triangularView<Upper>() * (permutation.inverse() * wa1);
+        temp1 = numext::abs2(wa3.stableNorm() / fnorm);
+        temp2 = numext::abs2(sqrt(par) * pnorm / fnorm);
+        prered = temp1 + temp2 / Scalar(.5);
+        dirder = -(temp1 + temp2);
+
+        /* compute the ratio of the actual to the predicted */
+        /* reduction. */
+        ratio = 0.;
+        if (prered != 0.)
+            ratio = actred / prered;
+
+        /* update the step bound. */
+        if (ratio <= Scalar(.25)) {
+            if (actred >= 0.)
+                temp = Scalar(.5);
+            if (actred < 0.)
+                temp = Scalar(.5) * dirder / (dirder + Scalar(.5) * actred);
+            if (Scalar(.1) * fnorm1 >= fnorm || temp < Scalar(.1))
+                temp = Scalar(.1);
+            /* Computing MIN */
+            delta = temp * (std::min)(delta, pnorm / Scalar(.1));
+            par /= temp;
+        } else if (!(par != 0. && ratio < Scalar(.75))) {
+            delta = pnorm / Scalar(.5);
+            par = Scalar(.5) * par;
+        }
+
+        /* test for successful iteration. */
+        if (ratio >= Scalar(1e-4)) {
+            /* successful iteration. update x, fvec, and their norms. */
+            x = wa2;
+            wa2 = diag.cwiseProduct(x);
+            fvec = wa4;
+            xnorm = wa2.stableNorm();
+            fnorm = fnorm1;
+            ++iter;
+        }
+
+        /* tests for convergence. */
+        if (abs(actred) <= parameters.ftol && prered <= parameters.ftol && Scalar(.5) * ratio <= 1. && delta <= parameters.xtol * xnorm)
+            return LevenbergMarquardtSpace::RelativeErrorAndReductionTooSmall;
+        if (abs(actred) <= parameters.ftol && prered <= parameters.ftol && Scalar(.5) * ratio <= 1.)
+            return LevenbergMarquardtSpace::RelativeReductionTooSmall;
+        if (delta <= parameters.xtol * xnorm)
+            return LevenbergMarquardtSpace::RelativeErrorTooSmall;
+
+        /* tests for termination and stringent tolerances. */
+        if (nfev >= parameters.maxfev)
+            return LevenbergMarquardtSpace::TooManyFunctionEvaluation;
+        if (abs(actred) <= NumTraits<Scalar>::epsilon() && prered <= NumTraits<Scalar>::epsilon() && Scalar(.5) * ratio <= 1.)
+            return LevenbergMarquardtSpace::FtolTooSmall;
+        if (delta <= NumTraits<Scalar>::epsilon() * xnorm)
+            return LevenbergMarquardtSpace::XtolTooSmall;
+        if (gnorm <= NumTraits<Scalar>::epsilon())
+            return LevenbergMarquardtSpace::GtolTooSmall;
+
+    } while (ratio < Scalar(1e-4));
+
+    return LevenbergMarquardtSpace::Running;
+}
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::minimizeOptimumStorage(FVectorType  &x)
+{
+    LevenbergMarquardtSpace::Status status = minimizeOptimumStorageInit(x);
+    if (status==LevenbergMarquardtSpace::ImproperInputParameters)
+        return status;
+    do {
+        status = minimizeOptimumStorageOneStep(x);
+    } while (status==LevenbergMarquardtSpace::Running);
+    return status;
+}
+
+template<typename FunctorType, typename Scalar>
+LevenbergMarquardtSpace::Status
+LevenbergMarquardt<FunctorType,Scalar>::lmdif1(
+        FunctorType &functor,
+        FVectorType  &x,
+        Index *nfev,
+        const Scalar tol
+        )
+{
+    Index n = x.size();
+    Index m = functor.values();
+
+    /* check the input parameters for errors. */
+    if (n <= 0 || m < n || tol < 0.)
+        return LevenbergMarquardtSpace::ImproperInputParameters;
+
+    NumericalDiff<FunctorType> numDiff(functor);
+    // embedded LevenbergMarquardt
+    LevenbergMarquardt<NumericalDiff<FunctorType>, Scalar > lm(numDiff);
+    lm.parameters.ftol = tol;
+    lm.parameters.xtol = tol;
+    lm.parameters.maxfev = 200*(n+1);
+
+    LevenbergMarquardtSpace::Status info = LevenbergMarquardtSpace::Status(lm.minimize(x));
+    if (nfev)
+        * nfev = lm.nfev;
+    return info;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_LEVENBERGMARQUARDT__H
+
+//vim: ai ts=4 sts=4 et sw=4
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/chkder.h b/src/EigenUnsupported/src/NonLinearOptimization/chkder.h
new file mode 100644
index 0000000..db8ff7d
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/chkder.h
@@ -0,0 +1,66 @@
+#define chkder_log10e 0.43429448190325182765
+#define chkder_factor 100.
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Scalar>
+void chkder(
+        const Matrix< Scalar, Dynamic, 1 >  &x,
+        const Matrix< Scalar, Dynamic, 1 >  &fvec,
+        const Matrix< Scalar, Dynamic, Dynamic > &fjac,
+        Matrix< Scalar, Dynamic, 1 >  &xp,
+        const Matrix< Scalar, Dynamic, 1 >  &fvecp,
+        int mode,
+        Matrix< Scalar, Dynamic, 1 >  &err
+        )
+{
+    using std::sqrt;
+    using std::abs;
+    using std::log;
+    
+    typedef DenseIndex Index;
+
+    const Scalar eps = sqrt(NumTraits<Scalar>::epsilon());
+    const Scalar epsf = chkder_factor * NumTraits<Scalar>::epsilon();
+    const Scalar epslog = chkder_log10e * log(eps);
+    Scalar temp;
+
+    const Index m = fvec.size(), n = x.size();
+
+    if (mode != 2) {
+        /* mode = 1. */
+        xp.resize(n);
+        for (Index j = 0; j < n; ++j) {
+            temp = eps * abs(x[j]);
+            if (temp == 0.)
+                temp = eps;
+            xp[j] = x[j] + temp;
+        }
+    }
+    else {
+        /* mode = 2. */
+        err.setZero(m); 
+        for (Index j = 0; j < n; ++j) {
+            temp = abs(x[j]);
+            if (temp == 0.)
+                temp = 1.;
+            err += temp * fjac.col(j);
+        }
+        for (Index i = 0; i < m; ++i) {
+            temp = 1.;
+            if (fvec[i] != 0. && fvecp[i] != 0. && abs(fvecp[i] - fvec[i]) >= epsf * abs(fvec[i]))
+                temp = eps * abs((fvecp[i] - fvec[i]) / eps - err[i]) / (abs(fvec[i]) + abs(fvecp[i]));
+            err[i] = 1.;
+            if (temp > NumTraits<Scalar>::epsilon() && temp < eps)
+                err[i] = (chkder_log10e * log(temp) - epslog) / epslog;
+            if (temp >= eps)
+                err[i] = 0.;
+        }
+    }
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/covar.h b/src/EigenUnsupported/src/NonLinearOptimization/covar.h
new file mode 100644
index 0000000..68260d1
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/covar.h
@@ -0,0 +1,70 @@
+namespace Eigen { 
+
+namespace internal {
+
+template <typename Scalar>
+void covar(
+        Matrix< Scalar, Dynamic, Dynamic > &r,
+        const VectorXi &ipvt,
+        Scalar tol = std::sqrt(NumTraits<Scalar>::epsilon()) )
+{
+    using std::abs;
+    typedef DenseIndex Index;
+
+    /* Local variables */
+    Index i, j, k, l, ii, jj;
+    bool sing;
+    Scalar temp;
+
+    /* Function Body */
+    const Index n = r.cols();
+    const Scalar tolr = tol * abs(r(0,0));
+    Matrix< Scalar, Dynamic, 1 > wa(n);
+    eigen_assert(ipvt.size()==n);
+
+    /* form the inverse of r in the full upper triangle of r. */
+    l = -1;
+    for (k = 0; k < n; ++k)
+        if (abs(r(k,k)) > tolr) {
+            r(k,k) = 1. / r(k,k);
+            for (j = 0; j <= k-1; ++j) {
+                temp = r(k,k) * r(j,k);
+                r(j,k) = 0.;
+                r.col(k).head(j+1) -= r.col(j).head(j+1) * temp;
+            }
+            l = k;
+        }
+
+    /* form the full upper triangle of the inverse of (r transpose)*r */
+    /* in the full upper triangle of r. */
+    for (k = 0; k <= l; ++k) {
+        for (j = 0; j <= k-1; ++j)
+            r.col(j).head(j+1) += r.col(k).head(j+1) * r(j,k);
+        r.col(k).head(k+1) *= r(k,k);
+    }
+
+    /* form the full lower triangle of the covariance matrix */
+    /* in the strict lower triangle of r and in wa. */
+    for (j = 0; j < n; ++j) {
+        jj = ipvt[j];
+        sing = j > l;
+        for (i = 0; i <= j; ++i) {
+            if (sing)
+                r(i,j) = 0.;
+            ii = ipvt[i];
+            if (ii > jj)
+                r(ii,jj) = r(i,j);
+            if (ii < jj)
+                r(jj,ii) = r(i,j);
+        }
+        wa[jj] = r(j,j);
+    }
+
+    /* symmetrize the covariance matrix in r. */
+    r.topLeftCorner(n,n).template triangularView<StrictlyUpper>() = r.topLeftCorner(n,n).transpose();
+    r.diagonal() = wa;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/dogleg.h b/src/EigenUnsupported/src/NonLinearOptimization/dogleg.h
new file mode 100644
index 0000000..80c5d27
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/dogleg.h
@@ -0,0 +1,107 @@
+namespace Eigen { 
+
+namespace internal {
+
+template <typename Scalar>
+void dogleg(
+        const Matrix< Scalar, Dynamic, Dynamic >  &qrfac,
+        const Matrix< Scalar, Dynamic, 1 >  &diag,
+        const Matrix< Scalar, Dynamic, 1 >  &qtb,
+        Scalar delta,
+        Matrix< Scalar, Dynamic, 1 >  &x)
+{
+    using std::abs;
+    using std::sqrt;
+    
+    typedef DenseIndex Index;
+
+    /* Local variables */
+    Index i, j;
+    Scalar sum, temp, alpha, bnorm;
+    Scalar gnorm, qnorm;
+    Scalar sgnorm;
+
+    /* Function Body */
+    const Scalar epsmch = NumTraits<Scalar>::epsilon();
+    const Index n = qrfac.cols();
+    eigen_assert(n==qtb.size());
+    eigen_assert(n==x.size());
+    eigen_assert(n==diag.size());
+    Matrix< Scalar, Dynamic, 1 >  wa1(n), wa2(n);
+
+    /* first, calculate the gauss-newton direction. */
+    for (j = n-1; j >=0; --j) {
+        temp = qrfac(j,j);
+        if (temp == 0.) {
+            temp = epsmch * qrfac.col(j).head(j+1).maxCoeff();
+            if (temp == 0.)
+                temp = epsmch;
+        }
+        if (j==n-1)
+            x[j] = qtb[j] / temp;
+        else
+            x[j] = (qtb[j] - qrfac.row(j).tail(n-j-1).dot(x.tail(n-j-1))) / temp;
+    }
+
+    /* test whether the gauss-newton direction is acceptable. */
+    qnorm = diag.cwiseProduct(x).stableNorm();
+    if (qnorm <= delta)
+        return;
+
+    // TODO : this path is not tested by Eigen unit tests
+
+    /* the gauss-newton direction is not acceptable. */
+    /* next, calculate the scaled gradient direction. */
+
+    wa1.fill(0.);
+    for (j = 0; j < n; ++j) {
+        wa1.tail(n-j) += qrfac.row(j).tail(n-j) * qtb[j];
+        wa1[j] /= diag[j];
+    }
+
+    /* calculate the norm of the scaled gradient and test for */
+    /* the special case in which the scaled gradient is zero. */
+    gnorm = wa1.stableNorm();
+    sgnorm = 0.;
+    alpha = delta / qnorm;
+    if (gnorm == 0.)
+        goto algo_end;
+
+    /* calculate the point along the scaled gradient */
+    /* at which the quadratic is minimized. */
+    wa1.array() /= (diag*gnorm).array();
+    // TODO : once unit tests cover this part,:
+    // wa2 = qrfac.template triangularView<Upper>() * wa1;
+    for (j = 0; j < n; ++j) {
+        sum = 0.;
+        for (i = j; i < n; ++i) {
+            sum += qrfac(j,i) * wa1[i];
+        }
+        wa2[j] = sum;
+    }
+    temp = wa2.stableNorm();
+    sgnorm = gnorm / temp / temp;
+
+    /* test whether the scaled gradient direction is acceptable. */
+    alpha = 0.;
+    if (sgnorm >= delta)
+        goto algo_end;
+
+    /* the scaled gradient direction is not acceptable. */
+    /* finally, calculate the point along the dogleg */
+    /* at which the quadratic is minimized. */
+    bnorm = qtb.stableNorm();
+    temp = bnorm / gnorm * (bnorm / qnorm) * (sgnorm / delta);
+    temp = temp - delta / qnorm * numext::abs2(sgnorm / delta) + sqrt(numext::abs2(temp - delta / qnorm) + (1.-numext::abs2(delta / qnorm)) * (1.-numext::abs2(sgnorm / delta)));
+    alpha = delta / qnorm * (1. - numext::abs2(sgnorm / delta)) / temp;
+algo_end:
+
+    /* form appropriate convex combination of the gauss-newton */
+    /* direction and the scaled gradient direction. */
+    temp = (1.-alpha) * (std::min)(sgnorm,delta);
+    x = temp * wa1 + alpha * x;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/fdjac1.h b/src/EigenUnsupported/src/NonLinearOptimization/fdjac1.h
new file mode 100644
index 0000000..bb7cf26
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/fdjac1.h
@@ -0,0 +1,79 @@
+namespace Eigen { 
+
+namespace internal {
+
+template<typename FunctorType, typename Scalar>
+DenseIndex fdjac1(
+        const FunctorType &Functor,
+        Matrix< Scalar, Dynamic, 1 >  &x,
+        Matrix< Scalar, Dynamic, 1 >  &fvec,
+        Matrix< Scalar, Dynamic, Dynamic > &fjac,
+        DenseIndex ml, DenseIndex mu,
+        Scalar epsfcn)
+{
+    using std::sqrt;
+    using std::abs;
+    
+    typedef DenseIndex Index;
+
+    /* Local variables */
+    Scalar h;
+    Index j, k;
+    Scalar eps, temp;
+    Index msum;
+    int iflag;
+    Index start, length;
+
+    /* Function Body */
+    const Scalar epsmch = NumTraits<Scalar>::epsilon();
+    const Index n = x.size();
+    eigen_assert(fvec.size()==n);
+    Matrix< Scalar, Dynamic, 1 >  wa1(n);
+    Matrix< Scalar, Dynamic, 1 >  wa2(n);
+
+    eps = sqrt((std::max)(epsfcn,epsmch));
+    msum = ml + mu + 1;
+    if (msum >= n) {
+        /* computation of dense approximate jacobian. */
+        for (j = 0; j < n; ++j) {
+            temp = x[j];
+            h = eps * abs(temp);
+            if (h == 0.)
+                h = eps;
+            x[j] = temp + h;
+            iflag = Functor(x, wa1);
+            if (iflag < 0)
+                return iflag;
+            x[j] = temp;
+            fjac.col(j) = (wa1-fvec)/h;
+        }
+
+    }else {
+        /* computation of banded approximate jacobian. */
+        for (k = 0; k < msum; ++k) {
+            for (j = k; (msum<0) ? (j>n): (j<n); j += msum) {
+                wa2[j] = x[j];
+                h = eps * abs(wa2[j]);
+                if (h == 0.) h = eps;
+                x[j] = wa2[j] + h;
+            }
+            iflag = Functor(x, wa1);
+            if (iflag < 0)
+                return iflag;
+            for (j = k; (msum<0) ? (j>n): (j<n); j += msum) {
+                x[j] = wa2[j];
+                h = eps * abs(wa2[j]);
+                if (h == 0.) h = eps;
+                fjac.col(j).setZero();
+                start = std::max<Index>(0,j-mu);
+                length = (std::min)(n-1, j+ml) - start + 1;
+                fjac.col(j).segment(start, length) = ( wa1.segment(start, length)-fvec.segment(start, length))/h;
+            }
+        }
+    }
+    return 0;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/lmpar.h b/src/EigenUnsupported/src/NonLinearOptimization/lmpar.h
new file mode 100644
index 0000000..4c17d4c
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/lmpar.h
@@ -0,0 +1,298 @@
+namespace Eigen { 
+
+namespace internal {
+
+template <typename Scalar>
+void lmpar(
+        Matrix< Scalar, Dynamic, Dynamic > &r,
+        const VectorXi &ipvt,
+        const Matrix< Scalar, Dynamic, 1 >  &diag,
+        const Matrix< Scalar, Dynamic, 1 >  &qtb,
+        Scalar delta,
+        Scalar &par,
+        Matrix< Scalar, Dynamic, 1 >  &x)
+{
+    using std::abs;
+    using std::sqrt;
+    typedef DenseIndex Index;
+
+    /* Local variables */
+    Index i, j, l;
+    Scalar fp;
+    Scalar parc, parl;
+    Index iter;
+    Scalar temp, paru;
+    Scalar gnorm;
+    Scalar dxnorm;
+
+
+    /* Function Body */
+    const Scalar dwarf = (std::numeric_limits<Scalar>::min)();
+    const Index n = r.cols();
+    eigen_assert(n==diag.size());
+    eigen_assert(n==qtb.size());
+    eigen_assert(n==x.size());
+
+    Matrix< Scalar, Dynamic, 1 >  wa1, wa2;
+
+    /* compute and store in x the gauss-newton direction. if the */
+    /* jacobian is rank-deficient, obtain a least squares solution. */
+    Index nsing = n-1;
+    wa1 = qtb;
+    for (j = 0; j < n; ++j) {
+        if (r(j,j) == 0. && nsing == n-1)
+            nsing = j - 1;
+        if (nsing < n-1)
+            wa1[j] = 0.;
+    }
+    for (j = nsing; j>=0; --j) {
+        wa1[j] /= r(j,j);
+        temp = wa1[j];
+        for (i = 0; i < j ; ++i)
+            wa1[i] -= r(i,j) * temp;
+    }
+
+    for (j = 0; j < n; ++j)
+        x[ipvt[j]] = wa1[j];
+
+    /* initialize the iteration counter. */
+    /* evaluate the function at the origin, and test */
+    /* for acceptance of the gauss-newton direction. */
+    iter = 0;
+    wa2 = diag.cwiseProduct(x);
+    dxnorm = wa2.blueNorm();
+    fp = dxnorm - delta;
+    if (fp <= Scalar(0.1) * delta) {
+        par = 0;
+        return;
+    }
+
+    /* if the jacobian is not rank deficient, the newton */
+    /* step provides a lower bound, parl, for the zero of */
+    /* the function. otherwise set this bound to zero. */
+    parl = 0.;
+    if (nsing >= n-1) {
+        for (j = 0; j < n; ++j) {
+            l = ipvt[j];
+            wa1[j] = diag[l] * (wa2[l] / dxnorm);
+        }
+        // it's actually a triangularView.solveInplace(), though in a weird
+        // way:
+        for (j = 0; j < n; ++j) {
+            Scalar sum = 0.;
+            for (i = 0; i < j; ++i)
+                sum += r(i,j) * wa1[i];
+            wa1[j] = (wa1[j] - sum) / r(j,j);
+        }
+        temp = wa1.blueNorm();
+        parl = fp / delta / temp / temp;
+    }
+
+    /* calculate an upper bound, paru, for the zero of the function. */
+    for (j = 0; j < n; ++j)
+        wa1[j] = r.col(j).head(j+1).dot(qtb.head(j+1)) / diag[ipvt[j]];
+
+    gnorm = wa1.stableNorm();
+    paru = gnorm / delta;
+    if (paru == 0.)
+        paru = dwarf / (std::min)(delta,Scalar(0.1));
+
+    /* if the input par lies outside of the interval (parl,paru), */
+    /* set par to the closer endpoint. */
+    par = (std::max)(par,parl);
+    par = (std::min)(par,paru);
+    if (par == 0.)
+        par = gnorm / dxnorm;
+
+    /* beginning of an iteration. */
+    while (true) {
+        ++iter;
+
+        /* evaluate the function at the current value of par. */
+        if (par == 0.)
+            par = (std::max)(dwarf,Scalar(.001) * paru); /* Computing MAX */
+        wa1 = sqrt(par)* diag;
+
+        Matrix< Scalar, Dynamic, 1 > sdiag(n);
+        qrsolv<Scalar>(r, ipvt, wa1, qtb, x, sdiag);
+
+        wa2 = diag.cwiseProduct(x);
+        dxnorm = wa2.blueNorm();
+        temp = fp;
+        fp = dxnorm - delta;
+
+        /* if the function is small enough, accept the current value */
+        /* of par. also test for the exceptional cases where parl */
+        /* is zero or the number of iterations has reached 10. */
+        if (abs(fp) <= Scalar(0.1) * delta || (parl == 0. && fp <= temp && temp < 0.) || iter == 10)
+            break;
+
+        /* compute the newton correction. */
+        for (j = 0; j < n; ++j) {
+            l = ipvt[j];
+            wa1[j] = diag[l] * (wa2[l] / dxnorm);
+        }
+        for (j = 0; j < n; ++j) {
+            wa1[j] /= sdiag[j];
+            temp = wa1[j];
+            for (i = j+1; i < n; ++i)
+                wa1[i] -= r(i,j) * temp;
+        }
+        temp = wa1.blueNorm();
+        parc = fp / delta / temp / temp;
+
+        /* depending on the sign of the function, update parl or paru. */
+        if (fp > 0.)
+            parl = (std::max)(parl,par);
+        if (fp < 0.)
+            paru = (std::min)(paru,par);
+
+        /* compute an improved estimate for par. */
+        /* Computing MAX */
+        par = (std::max)(parl,par+parc);
+
+        /* end of an iteration. */
+    }
+
+    /* termination. */
+    if (iter == 0)
+        par = 0.;
+    return;
+}
+
+template <typename Scalar>
+void lmpar2(
+        const ColPivHouseholderQR<Matrix< Scalar, Dynamic, Dynamic> > &qr,
+        const Matrix< Scalar, Dynamic, 1 >  &diag,
+        const Matrix< Scalar, Dynamic, 1 >  &qtb,
+        Scalar delta,
+        Scalar &par,
+        Matrix< Scalar, Dynamic, 1 >  &x)
+
+{
+    using std::sqrt;
+    using std::abs;
+    typedef DenseIndex Index;
+
+    /* Local variables */
+    Index j;
+    Scalar fp;
+    Scalar parc, parl;
+    Index iter;
+    Scalar temp, paru;
+    Scalar gnorm;
+    Scalar dxnorm;
+
+
+    /* Function Body */
+    const Scalar dwarf = (std::numeric_limits<Scalar>::min)();
+    const Index n = qr.matrixQR().cols();
+    eigen_assert(n==diag.size());
+    eigen_assert(n==qtb.size());
+
+    Matrix< Scalar, Dynamic, 1 >  wa1, wa2;
+
+    /* compute and store in x the gauss-newton direction. if the */
+    /* jacobian is rank-deficient, obtain a least squares solution. */
+
+//    const Index rank = qr.nonzeroPivots(); // exactly double(0.)
+    const Index rank = qr.rank(); // use a threshold
+    wa1 = qtb;
+    wa1.tail(n-rank).setZero();
+    qr.matrixQR().topLeftCorner(rank, rank).template triangularView<Upper>().solveInPlace(wa1.head(rank));
+
+    x = qr.colsPermutation()*wa1;
+
+    /* initialize the iteration counter. */
+    /* evaluate the function at the origin, and test */
+    /* for acceptance of the gauss-newton direction. */
+    iter = 0;
+    wa2 = diag.cwiseProduct(x);
+    dxnorm = wa2.blueNorm();
+    fp = dxnorm - delta;
+    if (fp <= Scalar(0.1) * delta) {
+        par = 0;
+        return;
+    }
+
+    /* if the jacobian is not rank deficient, the newton */
+    /* step provides a lower bound, parl, for the zero of */
+    /* the function. otherwise set this bound to zero. */
+    parl = 0.;
+    if (rank==n) {
+        wa1 = qr.colsPermutation().inverse() *  diag.cwiseProduct(wa2)/dxnorm;
+        qr.matrixQR().topLeftCorner(n, n).transpose().template triangularView<Lower>().solveInPlace(wa1);
+        temp = wa1.blueNorm();
+        parl = fp / delta / temp / temp;
+    }
+
+    /* calculate an upper bound, paru, for the zero of the function. */
+    for (j = 0; j < n; ++j)
+        wa1[j] = qr.matrixQR().col(j).head(j+1).dot(qtb.head(j+1)) / diag[qr.colsPermutation().indices()(j)];
+
+    gnorm = wa1.stableNorm();
+    paru = gnorm / delta;
+    if (paru == 0.)
+        paru = dwarf / (std::min)(delta,Scalar(0.1));
+
+    /* if the input par lies outside of the interval (parl,paru), */
+    /* set par to the closer endpoint. */
+    par = (std::max)(par,parl);
+    par = (std::min)(par,paru);
+    if (par == 0.)
+        par = gnorm / dxnorm;
+
+    /* beginning of an iteration. */
+    Matrix< Scalar, Dynamic, Dynamic > s = qr.matrixQR();
+    while (true) {
+        ++iter;
+
+        /* evaluate the function at the current value of par. */
+        if (par == 0.)
+            par = (std::max)(dwarf,Scalar(.001) * paru); /* Computing MAX */
+        wa1 = sqrt(par)* diag;
+
+        Matrix< Scalar, Dynamic, 1 > sdiag(n);
+        qrsolv<Scalar>(s, qr.colsPermutation().indices(), wa1, qtb, x, sdiag);
+
+        wa2 = diag.cwiseProduct(x);
+        dxnorm = wa2.blueNorm();
+        temp = fp;
+        fp = dxnorm - delta;
+
+        /* if the function is small enough, accept the current value */
+        /* of par. also test for the exceptional cases where parl */
+        /* is zero or the number of iterations has reached 10. */
+        if (abs(fp) <= Scalar(0.1) * delta || (parl == 0. && fp <= temp && temp < 0.) || iter == 10)
+            break;
+
+        /* compute the newton correction. */
+        wa1 = qr.colsPermutation().inverse() * diag.cwiseProduct(wa2/dxnorm);
+        // we could almost use this here, but the diagonal is outside qr, in sdiag[]
+        // qr.matrixQR().topLeftCorner(n, n).transpose().template triangularView<Lower>().solveInPlace(wa1);
+        for (j = 0; j < n; ++j) {
+            wa1[j] /= sdiag[j];
+            temp = wa1[j];
+            for (Index i = j+1; i < n; ++i)
+                wa1[i] -= s(i,j) * temp;
+        }
+        temp = wa1.blueNorm();
+        parc = fp / delta / temp / temp;
+
+        /* depending on the sign of the function, update parl or paru. */
+        if (fp > 0.)
+            parl = (std::max)(parl,par);
+        if (fp < 0.)
+            paru = (std::min)(paru,par);
+
+        /* compute an improved estimate for par. */
+        par = (std::max)(parl,par+parc);
+    }
+    if (iter == 0)
+        par = 0.;
+    return;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/qrsolv.h b/src/EigenUnsupported/src/NonLinearOptimization/qrsolv.h
new file mode 100644
index 0000000..4f2f560
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/qrsolv.h
@@ -0,0 +1,91 @@
+namespace Eigen { 
+
+namespace internal {
+
+// TODO : once qrsolv2 is removed, use ColPivHouseholderQR or PermutationMatrix instead of ipvt
+template <typename Scalar>
+void qrsolv(
+        Matrix< Scalar, Dynamic, Dynamic > &s,
+        // TODO : use a PermutationMatrix once lmpar is no more:
+        const VectorXi &ipvt,
+        const Matrix< Scalar, Dynamic, 1 >  &diag,
+        const Matrix< Scalar, Dynamic, 1 >  &qtb,
+        Matrix< Scalar, Dynamic, 1 >  &x,
+        Matrix< Scalar, Dynamic, 1 >  &sdiag)
+
+{
+    typedef DenseIndex Index;
+
+    /* Local variables */
+    Index i, j, k, l;
+    Scalar temp;
+    Index n = s.cols();
+    Matrix< Scalar, Dynamic, 1 >  wa(n);
+    JacobiRotation<Scalar> givens;
+
+    /* Function Body */
+    // the following will only change the lower triangular part of s, including
+    // the diagonal, though the diagonal is restored afterward
+
+    /*     copy r and (q transpose)*b to preserve input and initialize s. */
+    /*     in particular, save the diagonal elements of r in x. */
+    x = s.diagonal();
+    wa = qtb;
+
+    s.topLeftCorner(n,n).template triangularView<StrictlyLower>() = s.topLeftCorner(n,n).transpose();
+
+    /*     eliminate the diagonal matrix d using a givens rotation. */
+    for (j = 0; j < n; ++j) {
+
+        /*        prepare the row of d to be eliminated, locating the */
+        /*        diagonal element using p from the qr factorization. */
+        l = ipvt[j];
+        if (diag[l] == 0.)
+            break;
+        sdiag.tail(n-j).setZero();
+        sdiag[j] = diag[l];
+
+        /*        the transformations to eliminate the row of d */
+        /*        modify only a single element of (q transpose)*b */
+        /*        beyond the first n, which is initially zero. */
+        Scalar qtbpj = 0.;
+        for (k = j; k < n; ++k) {
+            /*           determine a givens rotation which eliminates the */
+            /*           appropriate element in the current row of d. */
+            givens.makeGivens(-s(k,k), sdiag[k]);
+
+            /*           compute the modified diagonal element of r and */
+            /*           the modified element of ((q transpose)*b,0). */
+            s(k,k) = givens.c() * s(k,k) + givens.s() * sdiag[k];
+            temp = givens.c() * wa[k] + givens.s() * qtbpj;
+            qtbpj = -givens.s() * wa[k] + givens.c() * qtbpj;
+            wa[k] = temp;
+
+            /*           accumulate the transformation in the row of s. */
+            for (i = k+1; i<n; ++i) {
+                temp = givens.c() * s(i,k) + givens.s() * sdiag[i];
+                sdiag[i] = -givens.s() * s(i,k) + givens.c() * sdiag[i];
+                s(i,k) = temp;
+            }
+        }
+    }
+
+    /*     solve the triangular system for z. if the system is */
+    /*     singular, then obtain a least squares solution. */
+    Index nsing;
+    for(nsing=0; nsing<n && sdiag[nsing]!=0; nsing++) {}
+
+    wa.tail(n-nsing).setZero();
+    s.topLeftCorner(nsing, nsing).transpose().template triangularView<Upper>().solveInPlace(wa.head(nsing));
+
+    // restore
+    sdiag = s.diagonal();
+    s.diagonal() = x;
+
+    /*     permute the components of z back to components of x. */
+    for (j = 0; j < n; ++j) x[ipvt[j]] = wa[j];
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/r1mpyq.h b/src/EigenUnsupported/src/NonLinearOptimization/r1mpyq.h
new file mode 100644
index 0000000..36ff700
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/r1mpyq.h
@@ -0,0 +1,30 @@
+namespace Eigen { 
+
+namespace internal {
+
+// TODO : move this to GivensQR once there's such a thing in Eigen
+
+template <typename Scalar>
+void r1mpyq(DenseIndex m, DenseIndex n, Scalar *a, const std::vector<JacobiRotation<Scalar> > &v_givens, const std::vector<JacobiRotation<Scalar> > &w_givens)
+{
+    typedef DenseIndex Index;
+
+    /*     apply the first set of givens rotations to a. */
+    for (Index j = n-2; j>=0; --j)
+        for (Index i = 0; i<m; ++i) {
+            Scalar temp = v_givens[j].c() * a[i+m*j] - v_givens[j].s() * a[i+m*(n-1)];
+            a[i+m*(n-1)] = v_givens[j].s() * a[i+m*j] + v_givens[j].c() * a[i+m*(n-1)];
+            a[i+m*j] = temp;
+        }
+    /*     apply the second set of givens rotations to a. */
+    for (Index j = 0; j<n-1; ++j)
+        for (Index i = 0; i<m; ++i) {
+            Scalar temp = w_givens[j].c() * a[i+m*j] + w_givens[j].s() * a[i+m*(n-1)];
+            a[i+m*(n-1)] = -w_givens[j].s() * a[i+m*j] + w_givens[j].c() * a[i+m*(n-1)];
+            a[i+m*j] = temp;
+        }
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/r1updt.h b/src/EigenUnsupported/src/NonLinearOptimization/r1updt.h
new file mode 100644
index 0000000..09fc652
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/r1updt.h
@@ -0,0 +1,99 @@
+namespace Eigen { 
+
+namespace internal {
+
+template <typename Scalar>
+void r1updt(
+        Matrix< Scalar, Dynamic, Dynamic > &s,
+        const Matrix< Scalar, Dynamic, 1> &u,
+        std::vector<JacobiRotation<Scalar> > &v_givens,
+        std::vector<JacobiRotation<Scalar> > &w_givens,
+        Matrix< Scalar, Dynamic, 1> &v,
+        Matrix< Scalar, Dynamic, 1> &w,
+        bool *sing)
+{
+    typedef DenseIndex Index;
+    const JacobiRotation<Scalar> IdentityRotation = JacobiRotation<Scalar>(1,0);
+
+    /* Local variables */
+    const Index m = s.rows();
+    const Index n = s.cols();
+    Index i, j=1;
+    Scalar temp;
+    JacobiRotation<Scalar> givens;
+
+    // r1updt had a broader usecase, but we don't use it here. And, more
+    // importantly, we can not test it.
+    eigen_assert(m==n);
+    eigen_assert(u.size()==m);
+    eigen_assert(v.size()==n);
+    eigen_assert(w.size()==n);
+
+    /* move the nontrivial part of the last column of s into w. */
+    w[n-1] = s(n-1,n-1);
+
+    /* rotate the vector v into a multiple of the n-th unit vector */
+    /* in such a way that a spike is introduced into w. */
+    for (j=n-2; j>=0; --j) {
+        w[j] = 0.;
+        if (v[j] != 0.) {
+            /* determine a givens rotation which eliminates the */
+            /* j-th element of v. */
+            givens.makeGivens(-v[n-1], v[j]);
+
+            /* apply the transformation to v and store the information */
+            /* necessary to recover the givens rotation. */
+            v[n-1] = givens.s() * v[j] + givens.c() * v[n-1];
+            v_givens[j] = givens;
+
+            /* apply the transformation to s and extend the spike in w. */
+            for (i = j; i < m; ++i) {
+                temp = givens.c() * s(j,i) - givens.s() * w[i];
+                w[i] = givens.s() * s(j,i) + givens.c() * w[i];
+                s(j,i) = temp;
+            }
+        } else
+            v_givens[j] = IdentityRotation;
+    }
+
+    /* add the spike from the rank 1 update to w. */
+    w += v[n-1] * u;
+
+    /* eliminate the spike. */
+    *sing = false;
+    for (j = 0; j < n-1; ++j) {
+        if (w[j] != 0.) {
+            /* determine a givens rotation which eliminates the */
+            /* j-th element of the spike. */
+            givens.makeGivens(-s(j,j), w[j]);
+
+            /* apply the transformation to s and reduce the spike in w. */
+            for (i = j; i < m; ++i) {
+                temp = givens.c() * s(j,i) + givens.s() * w[i];
+                w[i] = -givens.s() * s(j,i) + givens.c() * w[i];
+                s(j,i) = temp;
+            }
+
+            /* store the information necessary to recover the */
+            /* givens rotation. */
+            w_givens[j] = givens;
+        } else
+            v_givens[j] = IdentityRotation;
+
+        /* test for zero diagonal elements in the output s. */
+        if (s(j,j) == 0.) {
+            *sing = true;
+        }
+    }
+    /* move w back into the last column of the output s. */
+    s(n-1,n-1) = w[n-1];
+
+    if (s(j,j) == 0.) {
+        *sing = true;
+    }
+    return;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NonLinearOptimization/rwupdt.h b/src/EigenUnsupported/src/NonLinearOptimization/rwupdt.h
new file mode 100644
index 0000000..6ebf856
--- /dev/null
+++ b/src/EigenUnsupported/src/NonLinearOptimization/rwupdt.h
@@ -0,0 +1,49 @@
+namespace Eigen { 
+
+namespace internal {
+
+template <typename Scalar>
+void rwupdt(
+        Matrix< Scalar, Dynamic, Dynamic >  &r,
+        const Matrix< Scalar, Dynamic, 1>  &w,
+        Matrix< Scalar, Dynamic, 1>  &b,
+        Scalar alpha)
+{
+    typedef DenseIndex Index;
+
+    const Index n = r.cols();
+    eigen_assert(r.rows()>=n);
+    std::vector<JacobiRotation<Scalar> > givens(n);
+
+    /* Local variables */
+    Scalar temp, rowj;
+
+    /* Function Body */
+    for (Index j = 0; j < n; ++j) {
+        rowj = w[j];
+
+        /* apply the previous transformations to */
+        /* r(i,j), i=0,1,...,j-1, and to w(j). */
+        for (Index i = 0; i < j; ++i) {
+            temp = givens[i].c() * r(i,j) + givens[i].s() * rowj;
+            rowj = -givens[i].s() * r(i,j) + givens[i].c() * rowj;
+            r(i,j) = temp;
+        }
+
+        /* determine a givens rotation which eliminates w(j). */
+        givens[j].makeGivens(-r(j,j), rowj);
+
+        if (rowj == 0.)
+            continue; // givens[j] is identity
+
+        /* apply the current transformation to r(j,j), b(j), and alpha. */
+        r(j,j) = givens[j].c() * r(j,j) + givens[j].s() * rowj;
+        temp = givens[j].c() * b[j] + givens[j].s() * alpha;
+        alpha = -givens[j].s() * b[j] + givens[j].c() * alpha;
+        b[j] = temp;
+    }
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
diff --git a/src/EigenUnsupported/src/NumericalDiff/NumericalDiff.h b/src/EigenUnsupported/src/NumericalDiff/NumericalDiff.h
new file mode 100644
index 0000000..ea5d8bc
--- /dev/null
+++ b/src/EigenUnsupported/src/NumericalDiff/NumericalDiff.h
@@ -0,0 +1,130 @@
+// -*- coding: utf-8
+// vim: set fileencoding=utf-8
+
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Thomas Capricelli <orzel@freehackers.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_NUMERICAL_DIFF_H
+#define EIGEN_NUMERICAL_DIFF_H
+
+namespace Eigen { 
+
+enum NumericalDiffMode {
+    Forward,
+    Central
+};
+
+
+/**
+  * This class allows you to add a method df() to your functor, which will 
+  * use numerical differentiation to compute an approximate of the
+  * derivative for the functor. Of course, if you have an analytical form
+  * for the derivative, you should rather implement df() by yourself.
+  *
+  * More information on
+  * http://en.wikipedia.org/wiki/Numerical_differentiation
+  *
+  * Currently only "Forward" and "Central" scheme are implemented.
+  */
+template<typename _Functor, NumericalDiffMode mode=Forward>
+class NumericalDiff : public _Functor
+{
+public:
+    typedef _Functor Functor;
+    typedef typename Functor::Scalar Scalar;
+    typedef typename Functor::InputType InputType;
+    typedef typename Functor::ValueType ValueType;
+    typedef typename Functor::JacobianType JacobianType;
+
+    NumericalDiff(Scalar _epsfcn=0.) : Functor(), epsfcn(_epsfcn) {}
+    NumericalDiff(const Functor& f, Scalar _epsfcn=0.) : Functor(f), epsfcn(_epsfcn) {}
+
+    // forward constructors
+    template<typename T0>
+        NumericalDiff(const T0& a0) : Functor(a0), epsfcn(0) {}
+    template<typename T0, typename T1>
+        NumericalDiff(const T0& a0, const T1& a1) : Functor(a0, a1), epsfcn(0) {}
+    template<typename T0, typename T1, typename T2>
+        NumericalDiff(const T0& a0, const T1& a1, const T2& a2) : Functor(a0, a1, a2), epsfcn(0) {}
+
+    enum {
+        InputsAtCompileTime = Functor::InputsAtCompileTime,
+        ValuesAtCompileTime = Functor::ValuesAtCompileTime
+    };
+
+    /**
+      * return the number of evaluation of functor
+     */
+    int df(const InputType& _x, JacobianType &jac) const
+    {
+        using std::sqrt;
+        using std::abs;
+        /* Local variables */
+        Scalar h;
+        int nfev=0;
+        const typename InputType::Index n = _x.size();
+        const Scalar eps = sqrt(((std::max)(epsfcn,NumTraits<Scalar>::epsilon() )));
+        ValueType val1, val2;
+        InputType x = _x;
+        // TODO : we should do this only if the size is not already known
+        val1.resize(Functor::values());
+        val2.resize(Functor::values());
+
+        // initialization
+        switch(mode) {
+            case Forward:
+                // compute f(x)
+                Functor::operator()(x, val1); nfev++;
+                break;
+            case Central:
+                // do nothing
+                break;
+            default:
+                eigen_assert(false);
+        };
+
+        // Function Body
+        for (int j = 0; j < n; ++j) {
+            h = eps * abs(x[j]);
+            if (h == 0.) {
+                h = eps;
+            }
+            switch(mode) {
+                case Forward:
+                    x[j] += h;
+                    Functor::operator()(x, val2);
+                    nfev++;
+                    x[j] = _x[j];
+                    jac.col(j) = (val2-val1)/h;
+                    break;
+                case Central:
+                    x[j] += h;
+                    Functor::operator()(x, val2); nfev++;
+                    x[j] -= 2*h;
+                    Functor::operator()(x, val1); nfev++;
+                    x[j] = _x[j];
+                    jac.col(j) = (val2-val1)/(2*h);
+                    break;
+                default:
+                    eigen_assert(false);
+            };
+        }
+        return nfev;
+    }
+private:
+    Scalar epsfcn;
+
+    NumericalDiff& operator=(const NumericalDiff&);
+};
+
+} // end namespace Eigen
+
+//vim: ai ts=4 sts=4 et sw=4
+#endif // EIGEN_NUMERICAL_DIFF_H
+
diff --git a/src/EigenUnsupported/src/Polynomials/Companion.h b/src/EigenUnsupported/src/Polynomials/Companion.h
new file mode 100644
index 0000000..59a15b0
--- /dev/null
+++ b/src/EigenUnsupported/src/Polynomials/Companion.h
@@ -0,0 +1,280 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Manuel Yguel <manuel.yguel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPANION_H
+#define EIGEN_COMPANION_H
+
+// This file requires the user to include
+// * Eigen/Core
+// * Eigen/src/PolynomialSolver.h
+
+namespace Eigen { 
+
+namespace internal {
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+
+template<int Size>
+struct decrement_if_fixed_size
+{
+  enum {
+    ret = (Size == Dynamic) ? Dynamic : Size-1 };
+};
+
+#endif
+
+template< typename _Scalar, int _Deg >
+class companion
+{
+  public:
+    EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_Deg==Dynamic ? Dynamic : _Deg)
+
+    enum {
+      Deg = _Deg,
+      Deg_1=decrement_if_fixed_size<Deg>::ret
+    };
+
+    typedef _Scalar                                Scalar;
+    typedef typename NumTraits<Scalar>::Real       RealScalar;
+    typedef Matrix<Scalar, Deg, 1>                 RightColumn;
+    //typedef DiagonalMatrix< Scalar, Deg_1, Deg_1 > BottomLeftDiagonal;
+    typedef Matrix<Scalar, Deg_1, 1>               BottomLeftDiagonal;
+
+    typedef Matrix<Scalar, Deg, Deg>               DenseCompanionMatrixType;
+    typedef Matrix< Scalar, _Deg, Deg_1 >          LeftBlock;
+    typedef Matrix< Scalar, Deg_1, Deg_1 >         BottomLeftBlock;
+    typedef Matrix< Scalar, 1, Deg_1 >             LeftBlockFirstRow;
+
+    typedef DenseIndex Index;
+
+  public:
+    EIGEN_STRONG_INLINE const _Scalar operator()(Index row, Index col ) const
+    {
+      if( m_bl_diag.rows() > col )
+      {
+        if( 0 < row ){ return m_bl_diag[col]; }
+        else{ return 0; }
+      }
+      else{ return m_monic[row]; }
+    }
+
+  public:
+    template<typename VectorType>
+    void setPolynomial( const VectorType& poly )
+    {
+      const Index deg = poly.size()-1;
+      m_monic = -poly.head(deg)/poly[deg];
+      m_bl_diag.setOnes(deg-1);
+    }
+
+    template<typename VectorType>
+    companion( const VectorType& poly ){
+      setPolynomial( poly ); }
+
+  public:
+    DenseCompanionMatrixType denseMatrix() const
+    {
+      const Index deg   = m_monic.size();
+      const Index deg_1 = deg-1;
+      DenseCompanionMatrixType companMat(deg,deg);
+      companMat <<
+        ( LeftBlock(deg,deg_1)
+          << LeftBlockFirstRow::Zero(1,deg_1),
+          BottomLeftBlock::Identity(deg-1,deg-1)*m_bl_diag.asDiagonal() ).finished()
+        , m_monic;
+      return companMat;
+    }
+
+
+
+  protected:
+    /** Helper function for the balancing algorithm.
+     * \returns true if the row and the column, having colNorm and rowNorm
+     * as norms, are balanced, false otherwise.
+     * colB and rowB are respectively the multipliers for
+     * the column and the row in order to balance them.
+     * */
+    bool balanced( RealScalar colNorm, RealScalar rowNorm,
+        bool& isBalanced, RealScalar& colB, RealScalar& rowB );
+
+    /** Helper function for the balancing algorithm.
+     * \returns true if the row and the column, having colNorm and rowNorm
+     * as norms, are balanced, false otherwise.
+     * colB and rowB are respectively the multipliers for
+     * the column and the row in order to balance them.
+     * */
+    bool balancedR( RealScalar colNorm, RealScalar rowNorm,
+        bool& isBalanced, RealScalar& colB, RealScalar& rowB );
+
+  public:
+    /**
+     * Balancing algorithm from B. N. PARLETT and C. REINSCH (1969)
+     * "Balancing a matrix for calculation of eigenvalues and eigenvectors"
+     * adapted to the case of companion matrices.
+     * A matrix with non zero row and non zero column is balanced
+     * for a certain norm if the i-th row and the i-th column
+     * have same norm for all i.
+     */
+    void balance();
+
+  protected:
+      RightColumn                m_monic;
+      BottomLeftDiagonal         m_bl_diag;
+};
+
+
+
+template< typename _Scalar, int _Deg >
+inline
+bool companion<_Scalar,_Deg>::balanced( RealScalar colNorm, RealScalar rowNorm,
+    bool& isBalanced, RealScalar& colB, RealScalar& rowB )
+{
+  if( RealScalar(0) == colNorm || RealScalar(0) == rowNorm 
+      || !(numext::isfinite)(colNorm) || !(numext::isfinite)(rowNorm)){
+    return true;
+  }
+  else
+  {
+    //To find the balancing coefficients, if the radix is 2,
+    //one finds \f$ \sigma \f$ such that
+    // \f$ 2^{2\sigma-1} < rowNorm / colNorm \le 2^{2\sigma+1} \f$
+    // then the balancing coefficient for the row is \f$ 1/2^{\sigma} \f$
+    // and the balancing coefficient for the column is \f$ 2^{\sigma} \f$
+    const RealScalar radix = RealScalar(2);
+    const RealScalar radix2 = RealScalar(4);
+    
+    rowB = rowNorm / radix;
+    colB = RealScalar(1);
+    const RealScalar s = colNorm + rowNorm;
+
+    // Find sigma s.t. rowNorm / 2 <= 2^(2*sigma) * colNorm
+    RealScalar scout = colNorm;
+    while (scout < rowB)
+    {
+      colB *= radix;
+      scout *= radix2;
+    }
+    
+    // We now have an upper-bound for sigma, try to lower it.
+    // Find sigma s.t. 2^(2*sigma) * colNorm / 2 < rowNorm
+    scout = colNorm * (colB / radix) * colB;  // Avoid overflow.
+    while (scout >= rowNorm)
+    {
+      colB /= radix;
+      scout /= radix2;
+    }
+
+    // This line is used to avoid insubstantial balancing.
+    if ((rowNorm + radix * scout) < RealScalar(0.95) * s * colB)
+    {
+      isBalanced = false;
+      rowB = RealScalar(1) / colB;
+      return false;
+    }
+    else
+    {
+      return true;
+    }
+  }
+}
+
+template< typename _Scalar, int _Deg >
+inline
+bool companion<_Scalar,_Deg>::balancedR( RealScalar colNorm, RealScalar rowNorm,
+    bool& isBalanced, RealScalar& colB, RealScalar& rowB )
+{
+  if( RealScalar(0) == colNorm || RealScalar(0) == rowNorm ){ return true; }
+  else
+  {
+    /**
+     * Set the norm of the column and the row to the geometric mean
+     * of the row and column norm
+     */
+    const RealScalar q = colNorm/rowNorm;
+    if( !isApprox( q, _Scalar(1) ) )
+    {
+      rowB = sqrt( colNorm/rowNorm );
+      colB = RealScalar(1)/rowB;
+
+      isBalanced = false;
+      return false;
+    }
+    else{
+      return true; }
+  }
+}
+
+
+template< typename _Scalar, int _Deg >
+void companion<_Scalar,_Deg>::balance()
+{
+  using std::abs;
+  EIGEN_STATIC_ASSERT( Deg == Dynamic || 1 < Deg, YOU_MADE_A_PROGRAMMING_MISTAKE );
+  const Index deg   = m_monic.size();
+  const Index deg_1 = deg-1;
+
+  bool hasConverged=false;
+  while( !hasConverged )
+  {
+    hasConverged = true;
+    RealScalar colNorm,rowNorm;
+    RealScalar colB,rowB;
+
+    //First row, first column excluding the diagonal
+    //==============================================
+    colNorm = abs(m_bl_diag[0]);
+    rowNorm = abs(m_monic[0]);
+
+    //Compute balancing of the row and the column
+    if( !balanced( colNorm, rowNorm, hasConverged, colB, rowB ) )
+    {
+      m_bl_diag[0] *= colB;
+      m_monic[0] *= rowB;
+    }
+
+    //Middle rows and columns excluding the diagonal
+    //==============================================
+    for( Index i=1; i<deg_1; ++i )
+    {
+      // column norm, excluding the diagonal
+      colNorm = abs(m_bl_diag[i]);
+
+      // row norm, excluding the diagonal
+      rowNorm = abs(m_bl_diag[i-1]) + abs(m_monic[i]);
+
+      //Compute balancing of the row and the column
+      if( !balanced( colNorm, rowNorm, hasConverged, colB, rowB ) )
+      {
+        m_bl_diag[i]   *= colB;
+        m_bl_diag[i-1] *= rowB;
+        m_monic[i]     *= rowB;
+      }
+    }
+
+    //Last row, last column excluding the diagonal
+    //============================================
+    const Index ebl = m_bl_diag.size()-1;
+    VectorBlock<RightColumn,Deg_1> headMonic( m_monic, 0, deg_1 );
+    colNorm = headMonic.array().abs().sum();
+    rowNorm = abs( m_bl_diag[ebl] );
+
+    //Compute balancing of the row and the column
+    if( !balanced( colNorm, rowNorm, hasConverged, colB, rowB ) )
+    {
+      headMonic      *= colB;
+      m_bl_diag[ebl] *= rowB;
+    }
+  }
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPANION_H
diff --git a/src/EigenUnsupported/src/Polynomials/PolynomialSolver.h b/src/EigenUnsupported/src/Polynomials/PolynomialSolver.h
new file mode 100644
index 0000000..5e0ecbb
--- /dev/null
+++ b/src/EigenUnsupported/src/Polynomials/PolynomialSolver.h
@@ -0,0 +1,428 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Manuel Yguel <manuel.yguel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_POLYNOMIAL_SOLVER_H
+#define EIGEN_POLYNOMIAL_SOLVER_H
+
+namespace Eigen { 
+
+/** \ingroup Polynomials_Module
+ *  \class PolynomialSolverBase.
+ *
+ * \brief Defined to be inherited by polynomial solvers: it provides
+ * convenient methods such as
+ *  - real roots,
+ *  - greatest, smallest complex roots,
+ *  - real roots with greatest, smallest absolute real value,
+ *  - greatest, smallest real roots.
+ *
+ * It stores the set of roots as a vector of complexes.
+ *
+ */
+template< typename _Scalar, int _Deg >
+class PolynomialSolverBase
+{
+  public:
+    EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_Deg==Dynamic ? Dynamic : _Deg)
+
+    typedef _Scalar                             Scalar;
+    typedef typename NumTraits<Scalar>::Real    RealScalar;
+    typedef std::complex<RealScalar>            RootType;
+    typedef Matrix<RootType,_Deg,1>             RootsType;
+
+    typedef DenseIndex Index;
+
+  protected:
+    template< typename OtherPolynomial >
+    inline void setPolynomial( const OtherPolynomial& poly ){
+      m_roots.resize(poly.size()-1); }
+
+  public:
+    template< typename OtherPolynomial >
+    inline PolynomialSolverBase( const OtherPolynomial& poly ){
+      setPolynomial( poly() ); }
+
+    inline PolynomialSolverBase(){}
+
+  public:
+    /** \returns the complex roots of the polynomial */
+    inline const RootsType& roots() const { return m_roots; }
+
+  public:
+    /** Clear and fills the back insertion sequence with the real roots of the polynomial
+     * i.e. the real part of the complex roots that have an imaginary part which
+     * absolute value is smaller than absImaginaryThreshold.
+     * absImaginaryThreshold takes the dummy_precision associated
+     * with the _Scalar template parameter of the PolynomialSolver class as the default value.
+     *
+     * \param[out] bi_seq : the back insertion sequence (stl concept)
+     * \param[in]  absImaginaryThreshold : the maximum bound of the imaginary part of a complex
+     *  number that is considered as real.
+     * */
+    template<typename Stl_back_insertion_sequence>
+    inline void realRoots( Stl_back_insertion_sequence& bi_seq,
+        const RealScalar& absImaginaryThreshold = NumTraits<Scalar>::dummy_precision() ) const
+    {
+      using std::abs;
+      bi_seq.clear();
+      for(Index i=0; i<m_roots.size(); ++i )
+      {
+        if( abs( m_roots[i].imag() ) < absImaginaryThreshold ){
+          bi_seq.push_back( m_roots[i].real() ); }
+      }
+    }
+
+  protected:
+    template<typename squaredNormBinaryPredicate>
+    inline const RootType& selectComplexRoot_withRespectToNorm( squaredNormBinaryPredicate& pred ) const
+    {
+      Index res=0;
+      RealScalar norm2 = numext::abs2( m_roots[0] );
+      for( Index i=1; i<m_roots.size(); ++i )
+      {
+        const RealScalar currNorm2 = numext::abs2( m_roots[i] );
+        if( pred( currNorm2, norm2 ) ){
+          res=i; norm2=currNorm2; }
+      }
+      return m_roots[res];
+    }
+
+  public:
+    /**
+     * \returns the complex root with greatest norm.
+     */
+    inline const RootType& greatestRoot() const
+    {
+      std::greater<RealScalar> greater;
+      return selectComplexRoot_withRespectToNorm( greater );
+    }
+
+    /**
+     * \returns the complex root with smallest norm.
+     */
+    inline const RootType& smallestRoot() const
+    {
+      std::less<RealScalar> less;
+      return selectComplexRoot_withRespectToNorm( less );
+    }
+
+  protected:
+    template<typename squaredRealPartBinaryPredicate>
+    inline const RealScalar& selectRealRoot_withRespectToAbsRealPart(
+        squaredRealPartBinaryPredicate& pred,
+        bool& hasArealRoot,
+        const RealScalar& absImaginaryThreshold = NumTraits<Scalar>::dummy_precision() ) const
+    {
+      using std::abs;
+      hasArealRoot = false;
+      Index res=0;
+      RealScalar abs2(0);
+
+      for( Index i=0; i<m_roots.size(); ++i )
+      {
+        if( abs( m_roots[i].imag() ) <= absImaginaryThreshold )
+        {
+          if( !hasArealRoot )
+          {
+            hasArealRoot = true;
+            res = i;
+            abs2 = m_roots[i].real() * m_roots[i].real();
+          }
+          else
+          {
+            const RealScalar currAbs2 = m_roots[i].real() * m_roots[i].real();
+            if( pred( currAbs2, abs2 ) )
+            {
+              abs2 = currAbs2;
+              res = i;
+            }
+          }
+        }
+        else if(!hasArealRoot)
+        {
+          if( abs( m_roots[i].imag() ) < abs( m_roots[res].imag() ) ){
+            res = i;}
+        }
+      }
+      return numext::real_ref(m_roots[res]);
+    }
+
+
+    template<typename RealPartBinaryPredicate>
+    inline const RealScalar& selectRealRoot_withRespectToRealPart(
+        RealPartBinaryPredicate& pred,
+        bool& hasArealRoot,
+        const RealScalar& absImaginaryThreshold = NumTraits<Scalar>::dummy_precision() ) const
+    {
+      using std::abs;
+      hasArealRoot = false;
+      Index res=0;
+      RealScalar val(0);
+
+      for( Index i=0; i<m_roots.size(); ++i )
+      {
+        if( abs( m_roots[i].imag() ) <= absImaginaryThreshold )
+        {
+          if( !hasArealRoot )
+          {
+            hasArealRoot = true;
+            res = i;
+            val = m_roots[i].real();
+          }
+          else
+          {
+            const RealScalar curr = m_roots[i].real();
+            if( pred( curr, val ) )
+            {
+              val = curr;
+              res = i;
+            }
+          }
+        }
+        else
+        {
+          if( abs( m_roots[i].imag() ) < abs( m_roots[res].imag() ) ){
+            res = i; }
+        }
+      }
+      return numext::real_ref(m_roots[res]);
+    }
+
+  public:
+    /**
+     * \returns a real root with greatest absolute magnitude.
+     * A real root is defined as the real part of a complex root with absolute imaginary
+     * part smallest than absImaginaryThreshold.
+     * absImaginaryThreshold takes the dummy_precision associated
+     * with the _Scalar template parameter of the PolynomialSolver class as the default value.
+     * If no real root is found the boolean hasArealRoot is set to false and the real part of
+     * the root with smallest absolute imaginary part is returned instead.
+     *
+     * \param[out] hasArealRoot : boolean true if a real root is found according to the
+     *  absImaginaryThreshold criterion, false otherwise.
+     * \param[in] absImaginaryThreshold : threshold on the absolute imaginary part to decide
+     *  whether or not a root is real.
+     */
+    inline const RealScalar& absGreatestRealRoot(
+        bool& hasArealRoot,
+        const RealScalar& absImaginaryThreshold = NumTraits<Scalar>::dummy_precision() ) const
+    {
+      std::greater<RealScalar> greater;
+      return selectRealRoot_withRespectToAbsRealPart( greater, hasArealRoot, absImaginaryThreshold );
+    }
+
+
+    /**
+     * \returns a real root with smallest absolute magnitude.
+     * A real root is defined as the real part of a complex root with absolute imaginary
+     * part smallest than absImaginaryThreshold.
+     * absImaginaryThreshold takes the dummy_precision associated
+     * with the _Scalar template parameter of the PolynomialSolver class as the default value.
+     * If no real root is found the boolean hasArealRoot is set to false and the real part of
+     * the root with smallest absolute imaginary part is returned instead.
+     *
+     * \param[out] hasArealRoot : boolean true if a real root is found according to the
+     *  absImaginaryThreshold criterion, false otherwise.
+     * \param[in] absImaginaryThreshold : threshold on the absolute imaginary part to decide
+     *  whether or not a root is real.
+     */
+    inline const RealScalar& absSmallestRealRoot(
+        bool& hasArealRoot,
+        const RealScalar& absImaginaryThreshold = NumTraits<Scalar>::dummy_precision() ) const
+    {
+      std::less<RealScalar> less;
+      return selectRealRoot_withRespectToAbsRealPart( less, hasArealRoot, absImaginaryThreshold );
+    }
+
+
+    /**
+     * \returns the real root with greatest value.
+     * A real root is defined as the real part of a complex root with absolute imaginary
+     * part smallest than absImaginaryThreshold.
+     * absImaginaryThreshold takes the dummy_precision associated
+     * with the _Scalar template parameter of the PolynomialSolver class as the default value.
+     * If no real root is found the boolean hasArealRoot is set to false and the real part of
+     * the root with smallest absolute imaginary part is returned instead.
+     *
+     * \param[out] hasArealRoot : boolean true if a real root is found according to the
+     *  absImaginaryThreshold criterion, false otherwise.
+     * \param[in] absImaginaryThreshold : threshold on the absolute imaginary part to decide
+     *  whether or not a root is real.
+     */
+    inline const RealScalar& greatestRealRoot(
+        bool& hasArealRoot,
+        const RealScalar& absImaginaryThreshold = NumTraits<Scalar>::dummy_precision() ) const
+    {
+      std::greater<RealScalar> greater;
+      return selectRealRoot_withRespectToRealPart( greater, hasArealRoot, absImaginaryThreshold );
+    }
+
+
+    /**
+     * \returns the real root with smallest value.
+     * A real root is defined as the real part of a complex root with absolute imaginary
+     * part smallest than absImaginaryThreshold.
+     * absImaginaryThreshold takes the dummy_precision associated
+     * with the _Scalar template parameter of the PolynomialSolver class as the default value.
+     * If no real root is found the boolean hasArealRoot is set to false and the real part of
+     * the root with smallest absolute imaginary part is returned instead.
+     *
+     * \param[out] hasArealRoot : boolean true if a real root is found according to the
+     *  absImaginaryThreshold criterion, false otherwise.
+     * \param[in] absImaginaryThreshold : threshold on the absolute imaginary part to decide
+     *  whether or not a root is real.
+     */
+    inline const RealScalar& smallestRealRoot(
+        bool& hasArealRoot,
+        const RealScalar& absImaginaryThreshold = NumTraits<Scalar>::dummy_precision() ) const
+    {
+      std::less<RealScalar> less;
+      return selectRealRoot_withRespectToRealPart( less, hasArealRoot, absImaginaryThreshold );
+    }
+
+  protected:
+    RootsType               m_roots;
+};
+
+#define EIGEN_POLYNOMIAL_SOLVER_BASE_INHERITED_TYPES( BASE )  \
+  typedef typename BASE::Scalar                 Scalar;       \
+  typedef typename BASE::RealScalar             RealScalar;   \
+  typedef typename BASE::RootType               RootType;     \
+  typedef typename BASE::RootsType              RootsType;
+
+
+
+/** \ingroup Polynomials_Module
+  *
+  * \class PolynomialSolver
+  *
+  * \brief A polynomial solver
+  *
+  * Computes the complex roots of a real polynomial.
+  *
+  * \param _Scalar the scalar type, i.e., the type of the polynomial coefficients
+  * \param _Deg the degree of the polynomial, can be a compile time value or Dynamic.
+  *             Notice that the number of polynomial coefficients is _Deg+1.
+  *
+  * This class implements a polynomial solver and provides convenient methods such as
+  * - real roots,
+  * - greatest, smallest complex roots,
+  * - real roots with greatest, smallest absolute real value.
+  * - greatest, smallest real roots.
+  *
+  * WARNING: this polynomial solver is experimental, part of the unsupported Eigen modules.
+  *
+  *
+  * Currently a QR algorithm is used to compute the eigenvalues of the companion matrix of
+  * the polynomial to compute its roots.
+  * This supposes that the complex moduli of the roots are all distinct: e.g. there should
+  * be no multiple roots or conjugate roots for instance.
+  * With 32bit (float) floating types this problem shows up frequently.
+  * However, almost always, correct accuracy is reached even in these cases for 64bit
+  * (double) floating types and small polynomial degree (<20).
+  */
+template<typename _Scalar, int _Deg>
+class PolynomialSolver : public PolynomialSolverBase<_Scalar,_Deg>
+{
+  public:
+    EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_Deg==Dynamic ? Dynamic : _Deg)
+
+    typedef PolynomialSolverBase<_Scalar,_Deg>    PS_Base;
+    EIGEN_POLYNOMIAL_SOLVER_BASE_INHERITED_TYPES( PS_Base )
+
+    typedef Matrix<Scalar,_Deg,_Deg>                 CompanionMatrixType;
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                                          ComplexEigenSolver<CompanionMatrixType>,
+                                          EigenSolver<CompanionMatrixType> >::type EigenSolverType;
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex, Scalar, std::complex<Scalar> >::type ComplexScalar;
+
+  public:
+    /** Computes the complex roots of a new polynomial. */
+    template< typename OtherPolynomial >
+    void compute( const OtherPolynomial& poly )
+    {
+      eigen_assert( Scalar(0) != poly[poly.size()-1] );
+      eigen_assert( poly.size() > 1 );
+      if(poly.size() >  2 )
+      {
+        internal::companion<Scalar,_Deg> companion( poly );
+        companion.balance();
+        m_eigenSolver.compute( companion.denseMatrix() );
+        m_roots = m_eigenSolver.eigenvalues();
+        // cleanup noise in imaginary part of real roots:
+        // if the imaginary part is rather small compared to the real part
+        // and that cancelling the imaginary part yield a smaller evaluation,
+        // then it's safe to keep the real part only.
+        RealScalar coarse_prec = RealScalar(std::pow(4,poly.size()+1))*NumTraits<RealScalar>::epsilon();
+        for(Index i = 0; i<m_roots.size(); ++i)
+        {
+          if( internal::isMuchSmallerThan(numext::abs(numext::imag(m_roots[i])),
+                                          numext::abs(numext::real(m_roots[i])),
+                                          coarse_prec) )
+          {
+            ComplexScalar as_real_root = ComplexScalar(numext::real(m_roots[i]));
+            if(    numext::abs(poly_eval(poly, as_real_root))
+                <= numext::abs(poly_eval(poly, m_roots[i])))
+            {
+              m_roots[i] = as_real_root;
+            }
+          }
+        }
+      }
+      else if(poly.size () == 2)
+      {
+        m_roots.resize(1);
+        m_roots[0] = -poly[0]/poly[1];
+      }
+    }
+
+  public:
+    template< typename OtherPolynomial >
+    inline PolynomialSolver( const OtherPolynomial& poly ){
+      compute( poly ); }
+
+    inline PolynomialSolver(){}
+
+  protected:
+    using                   PS_Base::m_roots;
+    EigenSolverType         m_eigenSolver;
+};
+
+
+template< typename _Scalar >
+class PolynomialSolver<_Scalar,1> : public PolynomialSolverBase<_Scalar,1>
+{
+  public:
+    typedef PolynomialSolverBase<_Scalar,1>    PS_Base;
+    EIGEN_POLYNOMIAL_SOLVER_BASE_INHERITED_TYPES( PS_Base )
+
+  public:
+    /** Computes the complex roots of a new polynomial. */
+    template< typename OtherPolynomial >
+    void compute( const OtherPolynomial& poly )
+    {
+      eigen_assert( poly.size() == 2 );
+      eigen_assert( Scalar(0) != poly[1] );
+      m_roots[0] = -poly[0]/poly[1];
+    }
+
+  public:
+    template< typename OtherPolynomial >
+    inline PolynomialSolver( const OtherPolynomial& poly ){
+      compute( poly ); }
+
+    inline PolynomialSolver(){}
+
+  protected:
+    using                   PS_Base::m_roots;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_POLYNOMIAL_SOLVER_H
diff --git a/src/EigenUnsupported/src/Polynomials/PolynomialUtils.h b/src/EigenUnsupported/src/Polynomials/PolynomialUtils.h
new file mode 100644
index 0000000..394e857
--- /dev/null
+++ b/src/EigenUnsupported/src/Polynomials/PolynomialUtils.h
@@ -0,0 +1,143 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Manuel Yguel <manuel.yguel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_POLYNOMIAL_UTILS_H
+#define EIGEN_POLYNOMIAL_UTILS_H
+
+namespace Eigen { 
+
+/** \ingroup Polynomials_Module
+ * \returns the evaluation of the polynomial at x using Horner algorithm.
+ *
+ * \param[in] poly : the vector of coefficients of the polynomial ordered
+ *  by degrees i.e. poly[i] is the coefficient of degree i of the polynomial
+ *  e.g. \f$ 1 + 3x^2 \f$ is stored as a vector \f$ [ 1, 0, 3 ] \f$.
+ * \param[in] x : the value to evaluate the polynomial at.
+ *
+ * \note for stability:
+ *   \f$ |x| \le 1 \f$
+ */
+template <typename Polynomials, typename T>
+inline
+T poly_eval_horner( const Polynomials& poly, const T& x )
+{
+  T val=poly[poly.size()-1];
+  for(DenseIndex i=poly.size()-2; i>=0; --i ){
+    val = val*x + poly[i]; }
+  return val;
+}
+
+/** \ingroup Polynomials_Module
+ * \returns the evaluation of the polynomial at x using stabilized Horner algorithm.
+ *
+ * \param[in] poly : the vector of coefficients of the polynomial ordered
+ *  by degrees i.e. poly[i] is the coefficient of degree i of the polynomial
+ *  e.g. \f$ 1 + 3x^2 \f$ is stored as a vector \f$ [ 1, 0, 3 ] \f$.
+ * \param[in] x : the value to evaluate the polynomial at.
+ */
+template <typename Polynomials, typename T>
+inline
+T poly_eval( const Polynomials& poly, const T& x )
+{
+  typedef typename NumTraits<T>::Real Real;
+
+  if( numext::abs2( x ) <= Real(1) ){
+    return poly_eval_horner( poly, x ); }
+  else
+  {
+    T val=poly[0];
+    T inv_x = T(1)/x;
+    for( DenseIndex i=1; i<poly.size(); ++i ){
+      val = val*inv_x + poly[i]; }
+
+    return numext::pow(x,(T)(poly.size()-1)) * val;
+  }
+}
+
+/** \ingroup Polynomials_Module
+ * \returns a maximum bound for the absolute value of any root of the polynomial.
+ *
+ * \param[in] poly : the vector of coefficients of the polynomial ordered
+ *  by degrees i.e. poly[i] is the coefficient of degree i of the polynomial
+ *  e.g. \f$ 1 + 3x^2 \f$ is stored as a vector \f$ [ 1, 0, 3 ] \f$.
+ *
+ *  \pre
+ *   the leading coefficient of the input polynomial poly must be non zero
+ */
+template <typename Polynomial>
+inline
+typename NumTraits<typename Polynomial::Scalar>::Real cauchy_max_bound( const Polynomial& poly )
+{
+  using std::abs;
+  typedef typename Polynomial::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real Real;
+
+  eigen_assert( Scalar(0) != poly[poly.size()-1] );
+  const Scalar inv_leading_coeff = Scalar(1)/poly[poly.size()-1];
+  Real cb(0);
+
+  for( DenseIndex i=0; i<poly.size()-1; ++i ){
+    cb += abs(poly[i]*inv_leading_coeff); }
+  return cb + Real(1);
+}
+
+/** \ingroup Polynomials_Module
+ * \returns a minimum bound for the absolute value of any non zero root of the polynomial.
+ * \param[in] poly : the vector of coefficients of the polynomial ordered
+ *  by degrees i.e. poly[i] is the coefficient of degree i of the polynomial
+ *  e.g. \f$ 1 + 3x^2 \f$ is stored as a vector \f$ [ 1, 0, 3 ] \f$.
+ */
+template <typename Polynomial>
+inline
+typename NumTraits<typename Polynomial::Scalar>::Real cauchy_min_bound( const Polynomial& poly )
+{
+  using std::abs;
+  typedef typename Polynomial::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real Real;
+
+  DenseIndex i=0;
+  while( i<poly.size()-1 && Scalar(0) == poly(i) ){ ++i; }
+  if( poly.size()-1 == i ){
+    return Real(1); }
+
+  const Scalar inv_min_coeff = Scalar(1)/poly[i];
+  Real cb(1);
+  for( DenseIndex j=i+1; j<poly.size(); ++j ){
+    cb += abs(poly[j]*inv_min_coeff); }
+  return Real(1)/cb;
+}
+
+/** \ingroup Polynomials_Module
+ * Given the roots of a polynomial compute the coefficients in the
+ * monomial basis of the monic polynomial with same roots and minimal degree.
+ * If RootVector is a vector of complexes, Polynomial should also be a vector
+ * of complexes.
+ * \param[in] rv : a vector containing the roots of a polynomial.
+ * \param[out] poly : the vector of coefficients of the polynomial ordered
+ *  by degrees i.e. poly[i] is the coefficient of degree i of the polynomial
+ *  e.g. \f$ 3 + x^2 \f$ is stored as a vector \f$ [ 3, 0, 1 ] \f$.
+ */
+template <typename RootVector, typename Polynomial>
+void roots_to_monicPolynomial( const RootVector& rv, Polynomial& poly )
+{
+
+  typedef typename Polynomial::Scalar Scalar;
+
+  poly.setZero( rv.size()+1 );
+  poly[0] = -rv[0]; poly[1] = Scalar(1);
+  for( DenseIndex i=1; i< rv.size(); ++i )
+  {
+    for( DenseIndex j=i+1; j>0; --j ){ poly[j] = poly[j-1] - rv[i]*poly[j]; }
+    poly[0] = -rv[i]*poly[0];
+  }
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_POLYNOMIAL_UTILS_H
diff --git a/src/EigenUnsupported/src/Skyline/SkylineInplaceLU.h b/src/EigenUnsupported/src/Skyline/SkylineInplaceLU.h
new file mode 100644
index 0000000..6d0370d
--- /dev/null
+++ b/src/EigenUnsupported/src/Skyline/SkylineInplaceLU.h
@@ -0,0 +1,352 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Guillaume Saupin <guillaume.saupin@cea.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SKYLINEINPLACELU_H
+#define EIGEN_SKYLINEINPLACELU_H
+
+namespace Eigen { 
+
+/** \ingroup Skyline_Module
+ *
+ * \class SkylineInplaceLU
+ *
+ * \brief Inplace LU decomposition of a skyline matrix and associated features
+ *
+ * \param MatrixType the type of the matrix of which we are computing the LU factorization
+ *
+ */
+template<typename MatrixType>
+class SkylineInplaceLU {
+protected:
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::Index Index;
+    
+    typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
+
+public:
+
+    /** Creates a LU object and compute the respective factorization of \a matrix using
+     * flags \a flags. */
+    SkylineInplaceLU(MatrixType& matrix, int flags = 0)
+    : /*m_matrix(matrix.rows(), matrix.cols()),*/ m_flags(flags), m_status(0), m_lu(matrix) {
+        m_precision = RealScalar(0.1) * Eigen::dummy_precision<RealScalar > ();
+        m_lu.IsRowMajor ? computeRowMajor() : compute();
+    }
+
+    /** Sets the relative threshold value used to prune zero coefficients during the decomposition.
+     *
+     * Setting a value greater than zero speeds up computation, and yields to an incomplete
+     * factorization with fewer non zero coefficients. Such approximate factors are especially
+     * useful to initialize an iterative solver.
+     *
+     * Note that the exact meaning of this parameter might depends on the actual
+     * backend. Moreover, not all backends support this feature.
+     *
+     * \sa precision() */
+    void setPrecision(RealScalar v) {
+        m_precision = v;
+    }
+
+    /** \returns the current precision.
+     *
+     * \sa setPrecision() */
+    RealScalar precision() const {
+        return m_precision;
+    }
+
+    /** Sets the flags. Possible values are:
+     *  - CompleteFactorization
+     *  - IncompleteFactorization
+     *  - MemoryEfficient
+     *  - one of the ordering methods
+     *  - etc...
+     *
+     * \sa flags() */
+    void setFlags(int f) {
+        m_flags = f;
+    }
+
+    /** \returns the current flags */
+    int flags() const {
+        return m_flags;
+    }
+
+    void setOrderingMethod(int m) {
+        m_flags = m;
+    }
+
+    int orderingMethod() const {
+        return m_flags;
+    }
+
+    /** Computes/re-computes the LU factorization */
+    void compute();
+    void computeRowMajor();
+
+    /** \returns the lower triangular matrix L */
+    //inline const MatrixType& matrixL() const { return m_matrixL; }
+
+    /** \returns the upper triangular matrix U */
+    //inline const MatrixType& matrixU() const { return m_matrixU; }
+
+    template<typename BDerived, typename XDerived>
+    bool solve(const MatrixBase<BDerived> &b, MatrixBase<XDerived>* x,
+            const int transposed = 0) const;
+
+    /** \returns true if the factorization succeeded */
+    inline bool succeeded(void) const {
+        return m_succeeded;
+    }
+
+protected:
+    RealScalar m_precision;
+    int m_flags;
+    mutable int m_status;
+    bool m_succeeded;
+    MatrixType& m_lu;
+};
+
+/** Computes / recomputes the in place LU decomposition of the SkylineInplaceLU.
+ * using the default algorithm.
+ */
+template<typename MatrixType>
+//template<typename _Scalar>
+void SkylineInplaceLU<MatrixType>::compute() {
+    const size_t rows = m_lu.rows();
+    const size_t cols = m_lu.cols();
+
+    eigen_assert(rows == cols && "We do not (yet) support rectangular LU.");
+    eigen_assert(!m_lu.IsRowMajor && "LU decomposition does not work with rowMajor Storage");
+
+    for (Index row = 0; row < rows; row++) {
+        const double pivot = m_lu.coeffDiag(row);
+
+        //Lower matrix Columns update
+        const Index& col = row;
+        for (typename MatrixType::InnerLowerIterator lIt(m_lu, col); lIt; ++lIt) {
+            lIt.valueRef() /= pivot;
+        }
+
+        //Upper matrix update -> contiguous memory access
+        typename MatrixType::InnerLowerIterator lIt(m_lu, col);
+        for (Index rrow = row + 1; rrow < m_lu.rows(); rrow++) {
+            typename MatrixType::InnerUpperIterator uItPivot(m_lu, row);
+            typename MatrixType::InnerUpperIterator uIt(m_lu, rrow);
+            const double coef = lIt.value();
+
+            uItPivot += (rrow - row - 1);
+
+            //update upper part  -> contiguous memory access
+            for (++uItPivot; uIt && uItPivot;) {
+                uIt.valueRef() -= uItPivot.value() * coef;
+
+                ++uIt;
+                ++uItPivot;
+            }
+            ++lIt;
+        }
+
+        //Upper matrix update -> non contiguous memory access
+        typename MatrixType::InnerLowerIterator lIt3(m_lu, col);
+        for (Index rrow = row + 1; rrow < m_lu.rows(); rrow++) {
+            typename MatrixType::InnerUpperIterator uItPivot(m_lu, row);
+            const double coef = lIt3.value();
+
+            //update lower part ->  non contiguous memory access
+            for (Index i = 0; i < rrow - row - 1; i++) {
+                m_lu.coeffRefLower(rrow, row + i + 1) -= uItPivot.value() * coef;
+                ++uItPivot;
+            }
+            ++lIt3;
+        }
+        //update diag -> contiguous
+        typename MatrixType::InnerLowerIterator lIt2(m_lu, col);
+        for (Index rrow = row + 1; rrow < m_lu.rows(); rrow++) {
+
+            typename MatrixType::InnerUpperIterator uItPivot(m_lu, row);
+            typename MatrixType::InnerUpperIterator uIt(m_lu, rrow);
+            const double coef = lIt2.value();
+
+            uItPivot += (rrow - row - 1);
+            m_lu.coeffRefDiag(rrow) -= uItPivot.value() * coef;
+            ++lIt2;
+        }
+    }
+}
+
+template<typename MatrixType>
+void SkylineInplaceLU<MatrixType>::computeRowMajor() {
+    const size_t rows = m_lu.rows();
+    const size_t cols = m_lu.cols();
+
+    eigen_assert(rows == cols && "We do not (yet) support rectangular LU.");
+    eigen_assert(m_lu.IsRowMajor && "You're trying to apply rowMajor decomposition on a ColMajor matrix !");
+
+    for (Index row = 0; row < rows; row++) {
+        typename MatrixType::InnerLowerIterator llIt(m_lu, row);
+
+
+        for (Index col = llIt.col(); col < row; col++) {
+            if (m_lu.coeffExistLower(row, col)) {
+                const double diag = m_lu.coeffDiag(col);
+
+                typename MatrixType::InnerLowerIterator lIt(m_lu, row);
+                typename MatrixType::InnerUpperIterator uIt(m_lu, col);
+
+
+                const Index offset = lIt.col() - uIt.row();
+
+
+                Index stop = offset > 0 ? col - lIt.col() : col - uIt.row();
+
+                //#define VECTORIZE
+#ifdef VECTORIZE
+                Map<VectorXd > rowVal(lIt.valuePtr() + (offset > 0 ? 0 : -offset), stop);
+                Map<VectorXd > colVal(uIt.valuePtr() + (offset > 0 ? offset : 0), stop);
+
+
+                Scalar newCoeff = m_lu.coeffLower(row, col) - rowVal.dot(colVal);
+#else
+                if (offset > 0) //Skip zero value of lIt
+                    uIt += offset;
+                else //Skip zero values of uIt
+                    lIt += -offset;
+                Scalar newCoeff = m_lu.coeffLower(row, col);
+
+                for (Index k = 0; k < stop; ++k) {
+                    const Scalar tmp = newCoeff;
+                    newCoeff = tmp - lIt.value() * uIt.value();
+                    ++lIt;
+                    ++uIt;
+                }
+#endif
+
+                m_lu.coeffRefLower(row, col) = newCoeff / diag;
+            }
+        }
+
+        //Upper matrix update
+        const Index col = row;
+        typename MatrixType::InnerUpperIterator uuIt(m_lu, col);
+        for (Index rrow = uuIt.row(); rrow < col; rrow++) {
+
+            typename MatrixType::InnerLowerIterator lIt(m_lu, rrow);
+            typename MatrixType::InnerUpperIterator uIt(m_lu, col);
+            const Index offset = lIt.col() - uIt.row();
+
+            Index stop = offset > 0 ? rrow - lIt.col() : rrow - uIt.row();
+
+#ifdef VECTORIZE
+            Map<VectorXd > rowVal(lIt.valuePtr() + (offset > 0 ? 0 : -offset), stop);
+            Map<VectorXd > colVal(uIt.valuePtr() + (offset > 0 ? offset : 0), stop);
+
+            Scalar newCoeff = m_lu.coeffUpper(rrow, col) - rowVal.dot(colVal);
+#else
+            if (offset > 0) //Skip zero value of lIt
+                uIt += offset;
+            else //Skip zero values of uIt
+                lIt += -offset;
+            Scalar newCoeff = m_lu.coeffUpper(rrow, col);
+            for (Index k = 0; k < stop; ++k) {
+                const Scalar tmp = newCoeff;
+                newCoeff = tmp - lIt.value() * uIt.value();
+
+                ++lIt;
+                ++uIt;
+            }
+#endif
+            m_lu.coeffRefUpper(rrow, col) = newCoeff;
+        }
+
+
+        //Diag matrix update
+        typename MatrixType::InnerLowerIterator lIt(m_lu, row);
+        typename MatrixType::InnerUpperIterator uIt(m_lu, row);
+
+        const Index offset = lIt.col() - uIt.row();
+
+
+        Index stop = offset > 0 ? lIt.size() : uIt.size();
+#ifdef VECTORIZE
+        Map<VectorXd > rowVal(lIt.valuePtr() + (offset > 0 ? 0 : -offset), stop);
+        Map<VectorXd > colVal(uIt.valuePtr() + (offset > 0 ? offset : 0), stop);
+        Scalar newCoeff = m_lu.coeffDiag(row) - rowVal.dot(colVal);
+#else
+        if (offset > 0) //Skip zero value of lIt
+            uIt += offset;
+        else //Skip zero values of uIt
+            lIt += -offset;
+        Scalar newCoeff = m_lu.coeffDiag(row);
+        for (Index k = 0; k < stop; ++k) {
+            const Scalar tmp = newCoeff;
+            newCoeff = tmp - lIt.value() * uIt.value();
+            ++lIt;
+            ++uIt;
+        }
+#endif
+        m_lu.coeffRefDiag(row) = newCoeff;
+    }
+}
+
+/** Computes *x = U^-1 L^-1 b
+ *
+ * If \a transpose is set to SvTranspose or SvAdjoint, the solution
+ * of the transposed/adjoint system is computed instead.
+ *
+ * Not all backends implement the solution of the transposed or
+ * adjoint system.
+ */
+template<typename MatrixType>
+template<typename BDerived, typename XDerived>
+bool SkylineInplaceLU<MatrixType>::solve(const MatrixBase<BDerived> &b, MatrixBase<XDerived>* x, const int transposed) const {
+    const size_t rows = m_lu.rows();
+    const size_t cols = m_lu.cols();
+
+
+    for (Index row = 0; row < rows; row++) {
+        x->coeffRef(row) = b.coeff(row);
+        Scalar newVal = x->coeff(row);
+        typename MatrixType::InnerLowerIterator lIt(m_lu, row);
+
+        Index col = lIt.col();
+        while (lIt.col() < row) {
+
+            newVal -= x->coeff(col++) * lIt.value();
+            ++lIt;
+        }
+
+        x->coeffRef(row) = newVal;
+    }
+
+
+    for (Index col = rows - 1; col > 0; col--) {
+        x->coeffRef(col) = x->coeff(col) / m_lu.coeffDiag(col);
+
+        const Scalar x_col = x->coeff(col);
+
+        typename MatrixType::InnerUpperIterator uIt(m_lu, col);
+        uIt += uIt.size()-1;
+
+
+        while (uIt) {
+            x->coeffRef(uIt.row()) -= x_col * uIt.value();
+            //TODO : introduce --operator
+            uIt += -1;
+        }
+
+
+    }
+    x->coeffRef(0) = x->coeff(0) / m_lu.coeffDiag(0);
+
+    return true;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SKYLINEINPLACELU_H
diff --git a/src/EigenUnsupported/src/Skyline/SkylineMatrix.h b/src/EigenUnsupported/src/Skyline/SkylineMatrix.h
new file mode 100644
index 0000000..7c7eace
--- /dev/null
+++ b/src/EigenUnsupported/src/Skyline/SkylineMatrix.h
@@ -0,0 +1,862 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Guillaume Saupin <guillaume.saupin@cea.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SKYLINEMATRIX_H
+#define EIGEN_SKYLINEMATRIX_H
+
+#include "SkylineStorage.h"
+#include "SkylineMatrixBase.h"
+
+namespace Eigen { 
+
+/** \ingroup Skyline_Module
+ *
+ * \class SkylineMatrix
+ *
+ * \brief The main skyline matrix class
+ *
+ * This class implements a skyline matrix using the very uncommon storage
+ * scheme.
+ *
+ * \param _Scalar the scalar type, i.e. the type of the coefficients
+ * \param _Options Union of bit flags controlling the storage scheme. Currently the only possibility
+ *                 is RowMajor. The default is 0 which means column-major.
+ *
+ *
+ */
+namespace internal {
+template<typename _Scalar, int _Options>
+struct traits<SkylineMatrix<_Scalar, _Options> > {
+    typedef _Scalar Scalar;
+    typedef Sparse StorageKind;
+
+    enum {
+        RowsAtCompileTime = Dynamic,
+        ColsAtCompileTime = Dynamic,
+        MaxRowsAtCompileTime = Dynamic,
+        MaxColsAtCompileTime = Dynamic,
+        Flags = SkylineBit | _Options,
+        CoeffReadCost = NumTraits<Scalar>::ReadCost,
+    };
+};
+}
+
+template<typename _Scalar, int _Options>
+class SkylineMatrix
+: public SkylineMatrixBase<SkylineMatrix<_Scalar, _Options> > {
+public:
+    EIGEN_SKYLINE_GENERIC_PUBLIC_INTERFACE(SkylineMatrix)
+    EIGEN_SKYLINE_INHERIT_ASSIGNMENT_OPERATOR(SkylineMatrix, +=)
+    EIGEN_SKYLINE_INHERIT_ASSIGNMENT_OPERATOR(SkylineMatrix, -=)
+
+    using Base::IsRowMajor;
+
+protected:
+
+    typedef SkylineMatrix<Scalar, (Flags&~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0) > TransposedSkylineMatrix;
+
+    Index m_outerSize;
+    Index m_innerSize;
+
+public:
+    Index* m_colStartIndex;
+    Index* m_rowStartIndex;
+    SkylineStorage<Scalar> m_data;
+
+public:
+
+    inline Index rows() const {
+        return IsRowMajor ? m_outerSize : m_innerSize;
+    }
+
+    inline Index cols() const {
+        return IsRowMajor ? m_innerSize : m_outerSize;
+    }
+
+    inline Index innerSize() const {
+        return m_innerSize;
+    }
+
+    inline Index outerSize() const {
+        return m_outerSize;
+    }
+
+    inline Index upperNonZeros() const {
+        return m_data.upperSize();
+    }
+
+    inline Index lowerNonZeros() const {
+        return m_data.lowerSize();
+    }
+
+    inline Index upperNonZeros(Index j) const {
+        return m_colStartIndex[j + 1] - m_colStartIndex[j];
+    }
+
+    inline Index lowerNonZeros(Index j) const {
+        return m_rowStartIndex[j + 1] - m_rowStartIndex[j];
+    }
+
+    inline const Scalar* _diagPtr() const {
+        return &m_data.diag(0);
+    }
+
+    inline Scalar* _diagPtr() {
+        return &m_data.diag(0);
+    }
+
+    inline const Scalar* _upperPtr() const {
+        return &m_data.upper(0);
+    }
+
+    inline Scalar* _upperPtr() {
+        return &m_data.upper(0);
+    }
+
+    inline const Scalar* _lowerPtr() const {
+        return &m_data.lower(0);
+    }
+
+    inline Scalar* _lowerPtr() {
+        return &m_data.lower(0);
+    }
+
+    inline const Index* _upperProfilePtr() const {
+        return &m_data.upperProfile(0);
+    }
+
+    inline Index* _upperProfilePtr() {
+        return &m_data.upperProfile(0);
+    }
+
+    inline const Index* _lowerProfilePtr() const {
+        return &m_data.lowerProfile(0);
+    }
+
+    inline Index* _lowerProfilePtr() {
+        return &m_data.lowerProfile(0);
+    }
+
+    inline Scalar coeff(Index row, Index col) const {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+
+        if (outer == inner)
+            return this->m_data.diag(outer);
+
+        if (IsRowMajor) {
+            if (inner > outer) //upper matrix
+            {
+                const Index minOuterIndex = inner - m_data.upperProfile(inner);
+                if (outer >= minOuterIndex)
+                    return this->m_data.upper(m_colStartIndex[inner] + outer - (inner - m_data.upperProfile(inner)));
+                else
+                    return Scalar(0);
+            }
+            if (inner < outer) //lower matrix
+            {
+                const Index minInnerIndex = outer - m_data.lowerProfile(outer);
+                if (inner >= minInnerIndex)
+                    return this->m_data.lower(m_rowStartIndex[outer] + inner - (outer - m_data.lowerProfile(outer)));
+                else
+                    return Scalar(0);
+            }
+            return m_data.upper(m_colStartIndex[inner] + outer - inner);
+        } else {
+            if (outer > inner) //upper matrix
+            {
+                const Index maxOuterIndex = inner + m_data.upperProfile(inner);
+                if (outer <= maxOuterIndex)
+                    return this->m_data.upper(m_colStartIndex[inner] + (outer - inner));
+                else
+                    return Scalar(0);
+            }
+            if (outer < inner) //lower matrix
+            {
+                const Index maxInnerIndex = outer + m_data.lowerProfile(outer);
+
+                if (inner <= maxInnerIndex)
+                    return this->m_data.lower(m_rowStartIndex[outer] + (inner - outer));
+                else
+                    return Scalar(0);
+            }
+        }
+    }
+
+    inline Scalar& coeffRef(Index row, Index col) {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+
+        if (outer == inner)
+            return this->m_data.diag(outer);
+
+        if (IsRowMajor) {
+            if (col > row) //upper matrix
+            {
+                const Index minOuterIndex = inner - m_data.upperProfile(inner);
+                eigen_assert(outer >= minOuterIndex && "You tried to access a coeff that does not exist in the storage");
+                return this->m_data.upper(m_colStartIndex[inner] + outer - (inner - m_data.upperProfile(inner)));
+            }
+            if (col < row) //lower matrix
+            {
+                const Index minInnerIndex = outer - m_data.lowerProfile(outer);
+                eigen_assert(inner >= minInnerIndex && "You tried to access a coeff that does not exist in the storage");
+                return this->m_data.lower(m_rowStartIndex[outer] + inner - (outer - m_data.lowerProfile(outer)));
+            }
+        } else {
+            if (outer > inner) //upper matrix
+            {
+                const Index maxOuterIndex = inner + m_data.upperProfile(inner);
+                eigen_assert(outer <= maxOuterIndex && "You tried to access a coeff that does not exist in the storage");
+                return this->m_data.upper(m_colStartIndex[inner] + (outer - inner));
+            }
+            if (outer < inner) //lower matrix
+            {
+                const Index maxInnerIndex = outer + m_data.lowerProfile(outer);
+                eigen_assert(inner <= maxInnerIndex && "You tried to access a coeff that does not exist in the storage");
+                return this->m_data.lower(m_rowStartIndex[outer] + (inner - outer));
+            }
+        }
+    }
+
+    inline Scalar coeffDiag(Index idx) const {
+        eigen_assert(idx < outerSize());
+        eigen_assert(idx < innerSize());
+        return this->m_data.diag(idx);
+    }
+
+    inline Scalar coeffLower(Index row, Index col) const {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+        eigen_assert(inner != outer);
+
+        if (IsRowMajor) {
+            const Index minInnerIndex = outer - m_data.lowerProfile(outer);
+            if (inner >= minInnerIndex)
+                return this->m_data.lower(m_rowStartIndex[outer] + inner - (outer - m_data.lowerProfile(outer)));
+            else
+                return Scalar(0);
+
+        } else {
+            const Index maxInnerIndex = outer + m_data.lowerProfile(outer);
+            if (inner <= maxInnerIndex)
+                return this->m_data.lower(m_rowStartIndex[outer] + (inner - outer));
+            else
+                return Scalar(0);
+        }
+    }
+
+    inline Scalar coeffUpper(Index row, Index col) const {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+        eigen_assert(inner != outer);
+
+        if (IsRowMajor) {
+            const Index minOuterIndex = inner - m_data.upperProfile(inner);
+            if (outer >= minOuterIndex)
+                return this->m_data.upper(m_colStartIndex[inner] + outer - (inner - m_data.upperProfile(inner)));
+            else
+                return Scalar(0);
+        } else {
+            const Index maxOuterIndex = inner + m_data.upperProfile(inner);
+            if (outer <= maxOuterIndex)
+                return this->m_data.upper(m_colStartIndex[inner] + (outer - inner));
+            else
+                return Scalar(0);
+        }
+    }
+
+    inline Scalar& coeffRefDiag(Index idx) {
+        eigen_assert(idx < outerSize());
+        eigen_assert(idx < innerSize());
+        return this->m_data.diag(idx);
+    }
+
+    inline Scalar& coeffRefLower(Index row, Index col) {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+        eigen_assert(inner != outer);
+
+        if (IsRowMajor) {
+            const Index minInnerIndex = outer - m_data.lowerProfile(outer);
+            eigen_assert(inner >= minInnerIndex && "You tried to access a coeff that does not exist in the storage");
+            return this->m_data.lower(m_rowStartIndex[outer] + inner - (outer - m_data.lowerProfile(outer)));
+        } else {
+            const Index maxInnerIndex = outer + m_data.lowerProfile(outer);
+            eigen_assert(inner <= maxInnerIndex && "You tried to access a coeff that does not exist in the storage");
+            return this->m_data.lower(m_rowStartIndex[outer] + (inner - outer));
+        }
+    }
+
+    inline bool coeffExistLower(Index row, Index col) {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+        eigen_assert(inner != outer);
+
+        if (IsRowMajor) {
+            const Index minInnerIndex = outer - m_data.lowerProfile(outer);
+            return inner >= minInnerIndex;
+        } else {
+            const Index maxInnerIndex = outer + m_data.lowerProfile(outer);
+            return inner <= maxInnerIndex;
+        }
+    }
+
+    inline Scalar& coeffRefUpper(Index row, Index col) {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+        eigen_assert(inner != outer);
+
+        if (IsRowMajor) {
+            const Index minOuterIndex = inner - m_data.upperProfile(inner);
+            eigen_assert(outer >= minOuterIndex && "You tried to access a coeff that does not exist in the storage");
+            return this->m_data.upper(m_colStartIndex[inner] + outer - (inner - m_data.upperProfile(inner)));
+        } else {
+            const Index maxOuterIndex = inner + m_data.upperProfile(inner);
+            eigen_assert(outer <= maxOuterIndex && "You tried to access a coeff that does not exist in the storage");
+            return this->m_data.upper(m_colStartIndex[inner] + (outer - inner));
+        }
+    }
+
+    inline bool coeffExistUpper(Index row, Index col) {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+        eigen_assert(inner != outer);
+
+        if (IsRowMajor) {
+            const Index minOuterIndex = inner - m_data.upperProfile(inner);
+            return outer >= minOuterIndex;
+        } else {
+            const Index maxOuterIndex = inner + m_data.upperProfile(inner);
+            return outer <= maxOuterIndex;
+        }
+    }
+
+
+protected:
+
+public:
+    class InnerUpperIterator;
+    class InnerLowerIterator;
+
+    class OuterUpperIterator;
+    class OuterLowerIterator;
+
+    /** Removes all non zeros */
+    inline void setZero() {
+        m_data.clear();
+        memset(m_colStartIndex, 0, (m_outerSize + 1) * sizeof (Index));
+        memset(m_rowStartIndex, 0, (m_outerSize + 1) * sizeof (Index));
+    }
+
+    /** \returns the number of non zero coefficients */
+    inline Index nonZeros() const {
+        return m_data.diagSize() + m_data.upperSize() + m_data.lowerSize();
+    }
+
+    /** Preallocates \a reserveSize non zeros */
+    inline void reserve(Index reserveSize, Index reserveUpperSize, Index reserveLowerSize) {
+        m_data.reserve(reserveSize, reserveUpperSize, reserveLowerSize);
+    }
+
+    /** \returns a reference to a novel non zero coefficient with coordinates \a row x \a col.
+
+     *
+     * \warning This function can be extremely slow if the non zero coefficients
+     * are not inserted in a coherent order.
+     *
+     * After an insertion session, you should call the finalize() function.
+     */
+    EIGEN_DONT_INLINE Scalar & insert(Index row, Index col) {
+        const Index outer = IsRowMajor ? row : col;
+        const Index inner = IsRowMajor ? col : row;
+
+        eigen_assert(outer < outerSize());
+        eigen_assert(inner < innerSize());
+
+        if (outer == inner)
+            return m_data.diag(col);
+
+        if (IsRowMajor) {
+            if (outer < inner) //upper matrix
+            {
+                Index minOuterIndex = 0;
+                minOuterIndex = inner - m_data.upperProfile(inner);
+
+                if (outer < minOuterIndex) //The value does not yet exist
+                {
+                    const Index previousProfile = m_data.upperProfile(inner);
+
+                    m_data.upperProfile(inner) = inner - outer;
+
+
+                    const Index bandIncrement = m_data.upperProfile(inner) - previousProfile;
+                    //shift data stored after this new one
+                    const Index stop = m_colStartIndex[cols()];
+                    const Index start = m_colStartIndex[inner];
+
+
+                    for (Index innerIdx = stop; innerIdx >= start; innerIdx--) {
+                        m_data.upper(innerIdx + bandIncrement) = m_data.upper(innerIdx);
+                    }
+
+                    for (Index innerIdx = cols(); innerIdx > inner; innerIdx--) {
+                        m_colStartIndex[innerIdx] += bandIncrement;
+                    }
+
+                    //zeros new data
+                    memset(this->_upperPtr() + start, 0, (bandIncrement - 1) * sizeof (Scalar));
+
+                    return m_data.upper(m_colStartIndex[inner]);
+                } else {
+                    return m_data.upper(m_colStartIndex[inner] + outer - (inner - m_data.upperProfile(inner)));
+                }
+            }
+
+            if (outer > inner) //lower matrix
+            {
+                const Index minInnerIndex = outer - m_data.lowerProfile(outer);
+                if (inner < minInnerIndex) //The value does not yet exist
+                {
+                    const Index previousProfile = m_data.lowerProfile(outer);
+                    m_data.lowerProfile(outer) = outer - inner;
+
+                    const Index bandIncrement = m_data.lowerProfile(outer) - previousProfile;
+                    //shift data stored after this new one
+                    const Index stop = m_rowStartIndex[rows()];
+                    const Index start = m_rowStartIndex[outer];
+
+
+                    for (Index innerIdx = stop; innerIdx >= start; innerIdx--) {
+                        m_data.lower(innerIdx + bandIncrement) = m_data.lower(innerIdx);
+                    }
+
+                    for (Index innerIdx = rows(); innerIdx > outer; innerIdx--) {
+                        m_rowStartIndex[innerIdx] += bandIncrement;
+                    }
+
+                    //zeros new data
+                    memset(this->_lowerPtr() + start, 0, (bandIncrement - 1) * sizeof (Scalar));
+                    return m_data.lower(m_rowStartIndex[outer]);
+                } else {
+                    return m_data.lower(m_rowStartIndex[outer] + inner - (outer - m_data.lowerProfile(outer)));
+                }
+            }
+        } else {
+            if (outer > inner) //upper matrix
+            {
+                const Index maxOuterIndex = inner + m_data.upperProfile(inner);
+                if (outer > maxOuterIndex) //The value does not yet exist
+                {
+                    const Index previousProfile = m_data.upperProfile(inner);
+                    m_data.upperProfile(inner) = outer - inner;
+
+                    const Index bandIncrement = m_data.upperProfile(inner) - previousProfile;
+                    //shift data stored after this new one
+                    const Index stop = m_rowStartIndex[rows()];
+                    const Index start = m_rowStartIndex[inner + 1];
+
+                    for (Index innerIdx = stop; innerIdx >= start; innerIdx--) {
+                        m_data.upper(innerIdx + bandIncrement) = m_data.upper(innerIdx);
+                    }
+
+                    for (Index innerIdx = inner + 1; innerIdx < outerSize() + 1; innerIdx++) {
+                        m_rowStartIndex[innerIdx] += bandIncrement;
+                    }
+                    memset(this->_upperPtr() + m_rowStartIndex[inner] + previousProfile + 1, 0, (bandIncrement - 1) * sizeof (Scalar));
+                    return m_data.upper(m_rowStartIndex[inner] + m_data.upperProfile(inner));
+                } else {
+                    return m_data.upper(m_rowStartIndex[inner] + (outer - inner));
+                }
+            }
+
+            if (outer < inner) //lower matrix
+            {
+                const Index maxInnerIndex = outer + m_data.lowerProfile(outer);
+                if (inner > maxInnerIndex) //The value does not yet exist
+                {
+                    const Index previousProfile = m_data.lowerProfile(outer);
+                    m_data.lowerProfile(outer) = inner - outer;
+
+                    const Index bandIncrement = m_data.lowerProfile(outer) - previousProfile;
+                    //shift data stored after this new one
+                    const Index stop = m_colStartIndex[cols()];
+                    const Index start = m_colStartIndex[outer + 1];
+
+                    for (Index innerIdx = stop; innerIdx >= start; innerIdx--) {
+                        m_data.lower(innerIdx + bandIncrement) = m_data.lower(innerIdx);
+                    }
+
+                    for (Index innerIdx = outer + 1; innerIdx < outerSize() + 1; innerIdx++) {
+                        m_colStartIndex[innerIdx] += bandIncrement;
+                    }
+                    memset(this->_lowerPtr() + m_colStartIndex[outer] + previousProfile + 1, 0, (bandIncrement - 1) * sizeof (Scalar));
+                    return m_data.lower(m_colStartIndex[outer] + m_data.lowerProfile(outer));
+                } else {
+                    return m_data.lower(m_colStartIndex[outer] + (inner - outer));
+                }
+            }
+        }
+    }
+
+    /** Must be called after inserting a set of non zero entries.
+     */
+    inline void finalize() {
+        if (IsRowMajor) {
+            if (rows() > cols())
+                m_data.resize(cols(), cols(), rows(), m_colStartIndex[cols()] + 1, m_rowStartIndex[rows()] + 1);
+            else
+                m_data.resize(rows(), cols(), rows(), m_colStartIndex[cols()] + 1, m_rowStartIndex[rows()] + 1);
+
+            //            eigen_assert(rows() == cols() && "memory reorganisatrion only works with suare matrix");
+            //
+            //            Scalar* newArray = new Scalar[m_colStartIndex[cols()] + 1 + m_rowStartIndex[rows()] + 1];
+            //            Index dataIdx = 0;
+            //            for (Index row = 0; row < rows(); row++) {
+            //
+            //                const Index nbLowerElts = m_rowStartIndex[row + 1] - m_rowStartIndex[row];
+            //                //                std::cout << "nbLowerElts" << nbLowerElts << std::endl;
+            //                memcpy(newArray + dataIdx, m_data.m_lower + m_rowStartIndex[row], nbLowerElts * sizeof (Scalar));
+            //                m_rowStartIndex[row] = dataIdx;
+            //                dataIdx += nbLowerElts;
+            //
+            //                const Index nbUpperElts = m_colStartIndex[row + 1] - m_colStartIndex[row];
+            //                memcpy(newArray + dataIdx, m_data.m_upper + m_colStartIndex[row], nbUpperElts * sizeof (Scalar));
+            //                m_colStartIndex[row] = dataIdx;
+            //                dataIdx += nbUpperElts;
+            //
+            //
+            //            }
+            //            //todo : don't access m_data profile directly : add an accessor from SkylineMatrix
+            //            m_rowStartIndex[rows()] = m_rowStartIndex[rows()-1] + m_data.lowerProfile(rows()-1);
+            //            m_colStartIndex[cols()] = m_colStartIndex[cols()-1] + m_data.upperProfile(cols()-1);
+            //
+            //            delete[] m_data.m_lower;
+            //            delete[] m_data.m_upper;
+            //
+            //            m_data.m_lower = newArray;
+            //            m_data.m_upper = newArray;
+        } else {
+            if (rows() > cols())
+                m_data.resize(cols(), rows(), cols(), m_rowStartIndex[cols()] + 1, m_colStartIndex[cols()] + 1);
+            else
+                m_data.resize(rows(), rows(), cols(), m_rowStartIndex[rows()] + 1, m_colStartIndex[rows()] + 1);
+        }
+    }
+
+    inline void squeeze() {
+        finalize();
+        m_data.squeeze();
+    }
+
+    void prune(Scalar reference, RealScalar epsilon = dummy_precision<RealScalar > ()) {
+        //TODO
+    }
+
+    /** Resizes the matrix to a \a rows x \a cols matrix and initializes it to zero
+     * \sa resizeNonZeros(Index), reserve(), setZero()
+     */
+    void resize(size_t rows, size_t cols) {
+        const Index diagSize = rows > cols ? cols : rows;
+        m_innerSize = IsRowMajor ? cols : rows;
+
+        eigen_assert(rows == cols && "Skyline matrix must be square matrix");
+
+        if (diagSize % 2) { // diagSize is odd
+            const Index k = (diagSize - 1) / 2;
+
+            m_data.resize(diagSize, IsRowMajor ? cols : rows, IsRowMajor ? rows : cols,
+                    2 * k * k + k + 1,
+                    2 * k * k + k + 1);
+
+        } else // diagSize is even
+        {
+            const Index k = diagSize / 2;
+            m_data.resize(diagSize, IsRowMajor ? cols : rows, IsRowMajor ? rows : cols,
+                    2 * k * k - k + 1,
+                    2 * k * k - k + 1);
+        }
+
+        if (m_colStartIndex && m_rowStartIndex) {
+            delete[] m_colStartIndex;
+            delete[] m_rowStartIndex;
+        }
+        m_colStartIndex = new Index [cols + 1];
+        m_rowStartIndex = new Index [rows + 1];
+        m_outerSize = diagSize;
+
+        m_data.reset();
+        m_data.clear();
+
+        m_outerSize = diagSize;
+        memset(m_colStartIndex, 0, (cols + 1) * sizeof (Index));
+        memset(m_rowStartIndex, 0, (rows + 1) * sizeof (Index));
+    }
+
+    void resizeNonZeros(Index size) {
+        m_data.resize(size);
+    }
+
+    inline SkylineMatrix()
+    : m_outerSize(-1), m_innerSize(0), m_colStartIndex(0), m_rowStartIndex(0) {
+        resize(0, 0);
+    }
+
+    inline SkylineMatrix(size_t rows, size_t cols)
+    : m_outerSize(0), m_innerSize(0), m_colStartIndex(0), m_rowStartIndex(0) {
+        resize(rows, cols);
+    }
+
+    template<typename OtherDerived>
+    inline SkylineMatrix(const SkylineMatrixBase<OtherDerived>& other)
+    : m_outerSize(0), m_innerSize(0), m_colStartIndex(0), m_rowStartIndex(0) {
+        *this = other.derived();
+    }
+
+    inline SkylineMatrix(const SkylineMatrix & other)
+    : Base(), m_outerSize(0), m_innerSize(0), m_colStartIndex(0), m_rowStartIndex(0) {
+        *this = other.derived();
+    }
+
+    inline void swap(SkylineMatrix & other) {
+        //EIGEN_DBG_SKYLINE(std::cout << "SkylineMatrix:: swap\n");
+        std::swap(m_colStartIndex, other.m_colStartIndex);
+        std::swap(m_rowStartIndex, other.m_rowStartIndex);
+        std::swap(m_innerSize, other.m_innerSize);
+        std::swap(m_outerSize, other.m_outerSize);
+        m_data.swap(other.m_data);
+    }
+
+    inline SkylineMatrix & operator=(const SkylineMatrix & other) {
+        std::cout << "SkylineMatrix& operator=(const SkylineMatrix& other)\n";
+        if (other.isRValue()) {
+            swap(other.const_cast_derived());
+        } else {
+            resize(other.rows(), other.cols());
+            memcpy(m_colStartIndex, other.m_colStartIndex, (m_outerSize + 1) * sizeof (Index));
+            memcpy(m_rowStartIndex, other.m_rowStartIndex, (m_outerSize + 1) * sizeof (Index));
+            m_data = other.m_data;
+        }
+        return *this;
+    }
+
+    template<typename OtherDerived>
+            inline SkylineMatrix & operator=(const SkylineMatrixBase<OtherDerived>& other) {
+        const bool needToTranspose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit);
+        if (needToTranspose) {
+            //         TODO
+            //            return *this;
+        } else {
+            // there is no special optimization
+            return SkylineMatrixBase<SkylineMatrix>::operator=(other.derived());
+        }
+    }
+
+    friend std::ostream & operator <<(std::ostream & s, const SkylineMatrix & m) {
+
+        EIGEN_DBG_SKYLINE(
+        std::cout << "upper elements : " << std::endl;
+        for (Index i = 0; i < m.m_data.upperSize(); i++)
+            std::cout << m.m_data.upper(i) << "\t";
+        std::cout << std::endl;
+        std::cout << "upper profile : " << std::endl;
+        for (Index i = 0; i < m.m_data.upperProfileSize(); i++)
+            std::cout << m.m_data.upperProfile(i) << "\t";
+        std::cout << std::endl;
+        std::cout << "lower startIdx : " << std::endl;
+        for (Index i = 0; i < m.m_data.upperProfileSize(); i++)
+            std::cout << (IsRowMajor ? m.m_colStartIndex[i] : m.m_rowStartIndex[i]) << "\t";
+        std::cout << std::endl;
+
+
+        std::cout << "lower elements : " << std::endl;
+        for (Index i = 0; i < m.m_data.lowerSize(); i++)
+            std::cout << m.m_data.lower(i) << "\t";
+        std::cout << std::endl;
+        std::cout << "lower profile : " << std::endl;
+        for (Index i = 0; i < m.m_data.lowerProfileSize(); i++)
+            std::cout << m.m_data.lowerProfile(i) << "\t";
+        std::cout << std::endl;
+        std::cout << "lower startIdx : " << std::endl;
+        for (Index i = 0; i < m.m_data.lowerProfileSize(); i++)
+            std::cout << (IsRowMajor ? m.m_rowStartIndex[i] : m.m_colStartIndex[i]) << "\t";
+        std::cout << std::endl;
+        );
+        for (Index rowIdx = 0; rowIdx < m.rows(); rowIdx++) {
+            for (Index colIdx = 0; colIdx < m.cols(); colIdx++) {
+                s << m.coeff(rowIdx, colIdx) << "\t";
+            }
+            s << std::endl;
+        }
+        return s;
+    }
+
+    /** Destructor */
+    inline ~SkylineMatrix() {
+        delete[] m_colStartIndex;
+        delete[] m_rowStartIndex;
+    }
+
+    /** Overloaded for performance */
+    Scalar sum() const;
+};
+
+template<typename Scalar, int _Options>
+class SkylineMatrix<Scalar, _Options>::InnerUpperIterator {
+public:
+
+    InnerUpperIterator(const SkylineMatrix& mat, Index outer)
+    : m_matrix(mat), m_outer(outer),
+    m_id(_Options == RowMajor ? mat.m_colStartIndex[outer] : mat.m_rowStartIndex[outer] + 1),
+    m_start(m_id),
+    m_end(_Options == RowMajor ? mat.m_colStartIndex[outer + 1] : mat.m_rowStartIndex[outer + 1] + 1) {
+    }
+
+    inline InnerUpperIterator & operator++() {
+        m_id++;
+        return *this;
+    }
+
+    inline InnerUpperIterator & operator+=(Index shift) {
+        m_id += shift;
+        return *this;
+    }
+
+    inline Scalar value() const {
+        return m_matrix.m_data.upper(m_id);
+    }
+
+    inline Scalar* valuePtr() {
+        return const_cast<Scalar*> (&(m_matrix.m_data.upper(m_id)));
+    }
+
+    inline Scalar& valueRef() {
+        return const_cast<Scalar&> (m_matrix.m_data.upper(m_id));
+    }
+
+    inline Index index() const {
+        return IsRowMajor ? m_outer - m_matrix.m_data.upperProfile(m_outer) + (m_id - m_start) :
+                m_outer + (m_id - m_start) + 1;
+    }
+
+    inline Index row() const {
+        return IsRowMajor ? index() : m_outer;
+    }
+
+    inline Index col() const {
+        return IsRowMajor ? m_outer : index();
+    }
+
+    inline size_t size() const {
+        return m_matrix.m_data.upperProfile(m_outer);
+    }
+
+    inline operator bool() const {
+        return (m_id < m_end) && (m_id >= m_start);
+    }
+
+protected:
+    const SkylineMatrix& m_matrix;
+    const Index m_outer;
+    Index m_id;
+    const Index m_start;
+    const Index m_end;
+};
+
+template<typename Scalar, int _Options>
+class SkylineMatrix<Scalar, _Options>::InnerLowerIterator {
+public:
+
+    InnerLowerIterator(const SkylineMatrix& mat, Index outer)
+    : m_matrix(mat),
+    m_outer(outer),
+    m_id(_Options == RowMajor ? mat.m_rowStartIndex[outer] : mat.m_colStartIndex[outer] + 1),
+    m_start(m_id),
+    m_end(_Options == RowMajor ? mat.m_rowStartIndex[outer + 1] : mat.m_colStartIndex[outer + 1] + 1) {
+    }
+
+    inline InnerLowerIterator & operator++() {
+        m_id++;
+        return *this;
+    }
+
+    inline InnerLowerIterator & operator+=(Index shift) {
+        m_id += shift;
+        return *this;
+    }
+
+    inline Scalar value() const {
+        return m_matrix.m_data.lower(m_id);
+    }
+
+    inline Scalar* valuePtr() {
+        return const_cast<Scalar*> (&(m_matrix.m_data.lower(m_id)));
+    }
+
+    inline Scalar& valueRef() {
+        return const_cast<Scalar&> (m_matrix.m_data.lower(m_id));
+    }
+
+    inline Index index() const {
+        return IsRowMajor ? m_outer - m_matrix.m_data.lowerProfile(m_outer) + (m_id - m_start) :
+                m_outer + (m_id - m_start) + 1;
+        ;
+    }
+
+    inline Index row() const {
+        return IsRowMajor ? m_outer : index();
+    }
+
+    inline Index col() const {
+        return IsRowMajor ? index() : m_outer;
+    }
+
+    inline size_t size() const {
+        return m_matrix.m_data.lowerProfile(m_outer);
+    }
+
+    inline operator bool() const {
+        return (m_id < m_end) && (m_id >= m_start);
+    }
+
+protected:
+    const SkylineMatrix& m_matrix;
+    const Index m_outer;
+    Index m_id;
+    const Index m_start;
+    const Index m_end;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_SKYLINEMATRIX_H
diff --git a/src/EigenUnsupported/src/Skyline/SkylineMatrixBase.h b/src/EigenUnsupported/src/Skyline/SkylineMatrixBase.h
new file mode 100644
index 0000000..b0d5e10
--- /dev/null
+++ b/src/EigenUnsupported/src/Skyline/SkylineMatrixBase.h
@@ -0,0 +1,212 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Guillaume Saupin <guillaume.saupin@cea.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SKYLINEMATRIXBASE_H
+#define EIGEN_SKYLINEMATRIXBASE_H
+
+#include "SkylineUtil.h"
+
+namespace Eigen {
+
+/** \ingroup Skyline_Module
+ *
+ * \class SkylineMatrixBase
+ *
+ * \brief Base class of any skyline matrices or skyline expressions
+ *
+ * \param Derived
+ *
+ */
+template<typename Derived> class SkylineMatrixBase : public EigenBase<Derived> {
+public:
+
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::index<StorageKind>::type Index;
+
+    enum {
+        RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+        /**< The number of rows at compile-time. This is just a copy of the value provided
+         * by the \a Derived type. If a value is not known at compile-time,
+         * it is set to the \a Dynamic constant.
+         * \sa MatrixBase::rows(), MatrixBase::cols(), ColsAtCompileTime, SizeAtCompileTime */
+
+        ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+        /**< The number of columns at compile-time. This is just a copy of the value provided
+         * by the \a Derived type. If a value is not known at compile-time,
+         * it is set to the \a Dynamic constant.
+         * \sa MatrixBase::rows(), MatrixBase::cols(), RowsAtCompileTime, SizeAtCompileTime */
+
+
+        SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime,
+        internal::traits<Derived>::ColsAtCompileTime>::ret),
+        /**< This is equal to the number of coefficients, i.e. the number of
+         * rows times the number of columns, or to \a Dynamic if this is not
+         * known at compile-time. \sa RowsAtCompileTime, ColsAtCompileTime */
+
+        MaxRowsAtCompileTime = RowsAtCompileTime,
+        MaxColsAtCompileTime = ColsAtCompileTime,
+
+        MaxSizeAtCompileTime = (internal::size_at_compile_time<MaxRowsAtCompileTime,
+        MaxColsAtCompileTime>::ret),
+
+        IsVectorAtCompileTime = RowsAtCompileTime == 1 || ColsAtCompileTime == 1,
+        /**< This is set to true if either the number of rows or the number of
+         * columns is known at compile-time to be equal to 1. Indeed, in that case,
+         * we are dealing with a column-vector (if there is only one column) or with
+         * a row-vector (if there is only one row). */
+
+        Flags = internal::traits<Derived>::Flags,
+        /**< This stores expression \ref flags flags which may or may not be inherited by new expressions
+         * constructed from this one. See the \ref flags "list of flags".
+         */
+
+        CoeffReadCost = internal::traits<Derived>::CoeffReadCost,
+        /**< This is a rough measure of how expensive it is to read one coefficient from
+         * this expression.
+         */
+
+        IsRowMajor = Flags & RowMajorBit ? 1 : 0
+    };
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is the "real scalar" type; if the \a Scalar type is already real numbers
+     * (e.g. int, float or double) then \a RealScalar is just the same as \a Scalar. If
+     * \a Scalar is \a std::complex<T> then RealScalar is \a T.
+     *
+     * \sa class NumTraits
+     */
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    /** type of the equivalent square matrix */
+    typedef Matrix<Scalar, EIGEN_SIZE_MAX(RowsAtCompileTime, ColsAtCompileTime),
+                           EIGEN_SIZE_MAX(RowsAtCompileTime, ColsAtCompileTime) > SquareMatrixType;
+
+    inline const Derived& derived() const {
+        return *static_cast<const Derived*> (this);
+    }
+
+    inline Derived& derived() {
+        return *static_cast<Derived*> (this);
+    }
+
+    inline Derived& const_cast_derived() const {
+        return *static_cast<Derived*> (const_cast<SkylineMatrixBase*> (this));
+    }
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+    /** \returns the number of rows. \sa cols(), RowsAtCompileTime */
+    inline EIGEN_CONSTEXPR Index rows() const EIGEN_NOEXCEPT {
+        return derived().rows();
+    }
+
+    /** \returns the number of columns. \sa rows(), ColsAtCompileTime*/
+    inline EIGEN_CONSTEXPR Index cols() const EIGEN_NOEXCEPT {
+        return derived().cols();
+    }
+
+    /** \returns the number of coefficients, which is \a rows()*cols().
+     * \sa rows(), cols(), SizeAtCompileTime. */
+    inline EIGEN_CONSTEXPR Index size() const EIGEN_NOEXCEPT {
+        return rows() * cols();
+    }
+
+    /** \returns the number of nonzero coefficients which is in practice the number
+     * of stored coefficients. */
+    inline Index nonZeros() const {
+        return derived().nonZeros();
+    }
+
+    /** \returns the size of the storage major dimension,
+     * i.e., the number of columns for a columns major matrix, and the number of rows otherwise */
+    Index outerSize() const {
+        return (int(Flags) & RowMajorBit) ? this->rows() : this->cols();
+    }
+
+    /** \returns the size of the inner dimension according to the storage order,
+     * i.e., the number of rows for a columns major matrix, and the number of cols otherwise */
+    Index innerSize() const {
+        return (int(Flags) & RowMajorBit) ? this->cols() : this->rows();
+    }
+
+    bool isRValue() const {
+        return m_isRValue;
+    }
+
+    Derived& markAsRValue() {
+        m_isRValue = true;
+        return derived();
+    }
+
+    SkylineMatrixBase() : m_isRValue(false) {
+        /* TODO check flags */
+    }
+
+    inline Derived & operator=(const Derived& other) {
+        this->operator=<Derived > (other);
+        return derived();
+    }
+
+    template<typename OtherDerived>
+    inline void assignGeneric(const OtherDerived& other) {
+        derived().resize(other.rows(), other.cols());
+        for (Index row = 0; row < rows(); row++)
+            for (Index col = 0; col < cols(); col++) {
+                if (other.coeff(row, col) != Scalar(0))
+                    derived().insert(row, col) = other.coeff(row, col);
+            }
+        derived().finalize();
+    }
+
+    template<typename OtherDerived>
+            inline Derived & operator=(const SkylineMatrixBase<OtherDerived>& other) {
+        //TODO
+    }
+
+    template<typename Lhs, typename Rhs>
+            inline Derived & operator=(const SkylineProduct<Lhs, Rhs, SkylineTimeSkylineProduct>& product);
+
+    friend std::ostream & operator <<(std::ostream & s, const SkylineMatrixBase& m) {
+        s << m.derived();
+        return s;
+    }
+
+    template<typename OtherDerived>
+    const typename SkylineProductReturnType<Derived, OtherDerived>::Type
+    operator*(const MatrixBase<OtherDerived> &other) const;
+
+    /** \internal use operator= */
+    template<typename DenseDerived>
+    void evalTo(MatrixBase<DenseDerived>& dst) const {
+        dst.setZero();
+        for (Index i = 0; i < rows(); i++)
+            for (Index j = 0; j < rows(); j++)
+                dst(i, j) = derived().coeff(i, j);
+    }
+
+    Matrix<Scalar, RowsAtCompileTime, ColsAtCompileTime> toDense() const {
+        return derived();
+    }
+
+    /** \returns the matrix or vector obtained by evaluating this expression.
+     *
+     * Notice that in the case of a plain matrix or vector (not an expression) this function just returns
+     * a const reference, in order to avoid a useless copy.
+     */
+    EIGEN_STRONG_INLINE const typename internal::eval<Derived, IsSkyline>::type eval() const {
+        return typename internal::eval<Derived>::type(derived());
+    }
+
+protected:
+    bool m_isRValue;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_SKYLINEMATRIXBASE_H
diff --git a/src/EigenUnsupported/src/Skyline/SkylineProduct.h b/src/EigenUnsupported/src/Skyline/SkylineProduct.h
new file mode 100644
index 0000000..d9eb814
--- /dev/null
+++ b/src/EigenUnsupported/src/Skyline/SkylineProduct.h
@@ -0,0 +1,295 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Guillaume Saupin <guillaume.saupin@cea.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SKYLINEPRODUCT_H
+#define EIGEN_SKYLINEPRODUCT_H
+
+namespace Eigen { 
+
+template<typename Lhs, typename Rhs, int ProductMode>
+struct SkylineProductReturnType {
+    typedef const typename internal::nested_eval<Lhs, Rhs::RowsAtCompileTime>::type LhsNested;
+    typedef const typename internal::nested_eval<Rhs, Lhs::RowsAtCompileTime>::type RhsNested;
+
+    typedef SkylineProduct<LhsNested, RhsNested, ProductMode> Type;
+};
+
+template<typename LhsNested, typename RhsNested, int ProductMode>
+struct internal::traits<SkylineProduct<LhsNested, RhsNested, ProductMode> > {
+    // clean the nested types:
+    typedef typename internal::remove_all<LhsNested>::type _LhsNested;
+    typedef typename internal::remove_all<RhsNested>::type _RhsNested;
+    typedef typename _LhsNested::Scalar Scalar;
+
+    enum {
+        LhsCoeffReadCost = _LhsNested::CoeffReadCost,
+        RhsCoeffReadCost = _RhsNested::CoeffReadCost,
+        LhsFlags = _LhsNested::Flags,
+        RhsFlags = _RhsNested::Flags,
+
+        RowsAtCompileTime = _LhsNested::RowsAtCompileTime,
+        ColsAtCompileTime = _RhsNested::ColsAtCompileTime,
+        InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(_LhsNested::ColsAtCompileTime, _RhsNested::RowsAtCompileTime),
+
+        MaxRowsAtCompileTime = _LhsNested::MaxRowsAtCompileTime,
+        MaxColsAtCompileTime = _RhsNested::MaxColsAtCompileTime,
+
+        EvalToRowMajor = (RhsFlags & LhsFlags & RowMajorBit),
+        ResultIsSkyline = ProductMode == SkylineTimeSkylineProduct,
+
+        RemovedBits = ~((EvalToRowMajor ? 0 : RowMajorBit) | (ResultIsSkyline ? 0 : SkylineBit)),
+
+        Flags = (int(LhsFlags | RhsFlags) & HereditaryBits & RemovedBits)
+        | EvalBeforeAssigningBit
+        | EvalBeforeNestingBit,
+
+        CoeffReadCost = HugeCost
+    };
+
+    typedef typename internal::conditional<ResultIsSkyline,
+            SkylineMatrixBase<SkylineProduct<LhsNested, RhsNested, ProductMode> >,
+            MatrixBase<SkylineProduct<LhsNested, RhsNested, ProductMode> > >::type Base;
+};
+
+namespace internal {
+template<typename LhsNested, typename RhsNested, int ProductMode>
+class SkylineProduct : no_assignment_operator,
+public traits<SkylineProduct<LhsNested, RhsNested, ProductMode> >::Base {
+public:
+
+    EIGEN_GENERIC_PUBLIC_INTERFACE(SkylineProduct)
+
+private:
+
+    typedef typename traits<SkylineProduct>::_LhsNested _LhsNested;
+    typedef typename traits<SkylineProduct>::_RhsNested _RhsNested;
+
+public:
+
+    template<typename Lhs, typename Rhs>
+    EIGEN_STRONG_INLINE SkylineProduct(const Lhs& lhs, const Rhs& rhs)
+    : m_lhs(lhs), m_rhs(rhs) {
+        eigen_assert(lhs.cols() == rhs.rows());
+
+        enum {
+            ProductIsValid = _LhsNested::ColsAtCompileTime == Dynamic
+            || _RhsNested::RowsAtCompileTime == Dynamic
+            || int(_LhsNested::ColsAtCompileTime) == int(_RhsNested::RowsAtCompileTime),
+            AreVectors = _LhsNested::IsVectorAtCompileTime && _RhsNested::IsVectorAtCompileTime,
+            SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(_LhsNested, _RhsNested)
+        };
+        // note to the lost user:
+        //    * for a dot product use: v1.dot(v2)
+        //    * for a coeff-wise product use: v1.cwise()*v2
+        EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
+                INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
+                EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
+                INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
+                EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
+    }
+
+    EIGEN_STRONG_INLINE Index rows() const {
+        return m_lhs.rows();
+    }
+
+    EIGEN_STRONG_INLINE Index cols() const {
+        return m_rhs.cols();
+    }
+
+    EIGEN_STRONG_INLINE const _LhsNested& lhs() const {
+        return m_lhs;
+    }
+
+    EIGEN_STRONG_INLINE const _RhsNested& rhs() const {
+        return m_rhs;
+    }
+
+protected:
+    LhsNested m_lhs;
+    RhsNested m_rhs;
+};
+
+// dense = skyline * dense
+// Note that here we force no inlining and separate the setZero() because GCC messes up otherwise
+
+template<typename Lhs, typename Rhs, typename Dest>
+EIGEN_DONT_INLINE void skyline_row_major_time_dense_product(const Lhs& lhs, const Rhs& rhs, Dest& dst) {
+    typedef typename remove_all<Lhs>::type _Lhs;
+    typedef typename remove_all<Rhs>::type _Rhs;
+    typedef typename traits<Lhs>::Scalar Scalar;
+
+    enum {
+        LhsIsRowMajor = (_Lhs::Flags & RowMajorBit) == RowMajorBit,
+        LhsIsSelfAdjoint = (_Lhs::Flags & SelfAdjointBit) == SelfAdjointBit,
+        ProcessFirstHalf = LhsIsSelfAdjoint
+        && (((_Lhs::Flags & (UpperTriangularBit | LowerTriangularBit)) == 0)
+        || ((_Lhs::Flags & UpperTriangularBit) && !LhsIsRowMajor)
+        || ((_Lhs::Flags & LowerTriangularBit) && LhsIsRowMajor)),
+        ProcessSecondHalf = LhsIsSelfAdjoint && (!ProcessFirstHalf)
+    };
+
+    //Use matrix diagonal part <- Improvement : use inner iterator on dense matrix.
+    for (Index col = 0; col < rhs.cols(); col++) {
+        for (Index row = 0; row < lhs.rows(); row++) {
+            dst(row, col) = lhs.coeffDiag(row) * rhs(row, col);
+        }
+    }
+    //Use matrix lower triangular part
+    for (Index row = 0; row < lhs.rows(); row++) {
+        typename _Lhs::InnerLowerIterator lIt(lhs, row);
+        const Index stop = lIt.col() + lIt.size();
+        for (Index col = 0; col < rhs.cols(); col++) {
+
+            Index k = lIt.col();
+            Scalar tmp = 0;
+            while (k < stop) {
+                tmp +=
+                        lIt.value() *
+                        rhs(k++, col);
+                ++lIt;
+            }
+            dst(row, col) += tmp;
+            lIt += -lIt.size();
+        }
+
+    }
+
+    //Use matrix upper triangular part
+    for (Index lhscol = 0; lhscol < lhs.cols(); lhscol++) {
+        typename _Lhs::InnerUpperIterator uIt(lhs, lhscol);
+        const Index stop = uIt.size() + uIt.row();
+        for (Index rhscol = 0; rhscol < rhs.cols(); rhscol++) {
+
+
+            const Scalar rhsCoeff = rhs.coeff(lhscol, rhscol);
+            Index k = uIt.row();
+            while (k < stop) {
+                dst(k++, rhscol) +=
+                        uIt.value() *
+                        rhsCoeff;
+                ++uIt;
+            }
+            uIt += -uIt.size();
+        }
+    }
+
+}
+
+template<typename Lhs, typename Rhs, typename Dest>
+EIGEN_DONT_INLINE void skyline_col_major_time_dense_product(const Lhs& lhs, const Rhs& rhs, Dest& dst) {
+    typedef typename remove_all<Lhs>::type _Lhs;
+    typedef typename remove_all<Rhs>::type _Rhs;
+    typedef typename traits<Lhs>::Scalar Scalar;
+
+    enum {
+        LhsIsRowMajor = (_Lhs::Flags & RowMajorBit) == RowMajorBit,
+        LhsIsSelfAdjoint = (_Lhs::Flags & SelfAdjointBit) == SelfAdjointBit,
+        ProcessFirstHalf = LhsIsSelfAdjoint
+        && (((_Lhs::Flags & (UpperTriangularBit | LowerTriangularBit)) == 0)
+        || ((_Lhs::Flags & UpperTriangularBit) && !LhsIsRowMajor)
+        || ((_Lhs::Flags & LowerTriangularBit) && LhsIsRowMajor)),
+        ProcessSecondHalf = LhsIsSelfAdjoint && (!ProcessFirstHalf)
+    };
+
+    //Use matrix diagonal part <- Improvement : use inner iterator on dense matrix.
+    for (Index col = 0; col < rhs.cols(); col++) {
+        for (Index row = 0; row < lhs.rows(); row++) {
+            dst(row, col) = lhs.coeffDiag(row) * rhs(row, col);
+        }
+    }
+
+    //Use matrix upper triangular part
+    for (Index row = 0; row < lhs.rows(); row++) {
+        typename _Lhs::InnerUpperIterator uIt(lhs, row);
+        const Index stop = uIt.col() + uIt.size();
+        for (Index col = 0; col < rhs.cols(); col++) {
+
+            Index k = uIt.col();
+            Scalar tmp = 0;
+            while (k < stop) {
+                tmp +=
+                        uIt.value() *
+                        rhs(k++, col);
+                ++uIt;
+            }
+
+
+            dst(row, col) += tmp;
+            uIt += -uIt.size();
+        }
+    }
+
+    //Use matrix lower triangular part
+    for (Index lhscol = 0; lhscol < lhs.cols(); lhscol++) {
+        typename _Lhs::InnerLowerIterator lIt(lhs, lhscol);
+        const Index stop = lIt.size() + lIt.row();
+        for (Index rhscol = 0; rhscol < rhs.cols(); rhscol++) {
+
+            const Scalar rhsCoeff = rhs.coeff(lhscol, rhscol);
+            Index k = lIt.row();
+            while (k < stop) {
+                dst(k++, rhscol) +=
+                        lIt.value() *
+                        rhsCoeff;
+                ++lIt;
+            }
+            lIt += -lIt.size();
+        }
+    }
+
+}
+
+template<typename Lhs, typename Rhs, typename ResultType,
+        int LhsStorageOrder = traits<Lhs>::Flags&RowMajorBit>
+        struct skyline_product_selector;
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct skyline_product_selector<Lhs, Rhs, ResultType, RowMajor> {
+    typedef typename traits<typename remove_all<Lhs>::type>::Scalar Scalar;
+
+    static void run(const Lhs& lhs, const Rhs& rhs, ResultType & res) {
+        skyline_row_major_time_dense_product<Lhs, Rhs, ResultType > (lhs, rhs, res);
+    }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct skyline_product_selector<Lhs, Rhs, ResultType, ColMajor> {
+    typedef typename traits<typename remove_all<Lhs>::type>::Scalar Scalar;
+
+    static void run(const Lhs& lhs, const Rhs& rhs, ResultType & res) {
+        skyline_col_major_time_dense_product<Lhs, Rhs, ResultType > (lhs, rhs, res);
+    }
+};
+
+} // end namespace internal
+
+// template<typename Derived>
+// template<typename Lhs, typename Rhs >
+// Derived & MatrixBase<Derived>::lazyAssign(const SkylineProduct<Lhs, Rhs, SkylineTimeDenseProduct>& product) {
+//     typedef typename internal::remove_all<Lhs>::type _Lhs;
+//     internal::skyline_product_selector<typename internal::remove_all<Lhs>::type,
+//             typename internal::remove_all<Rhs>::type,
+//             Derived>::run(product.lhs(), product.rhs(), derived());
+// 
+//     return derived();
+// }
+
+// skyline * dense
+
+template<typename Derived>
+template<typename OtherDerived >
+EIGEN_STRONG_INLINE const typename SkylineProductReturnType<Derived, OtherDerived>::Type
+SkylineMatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const {
+
+    return typename SkylineProductReturnType<Derived, OtherDerived>::Type(derived(), other.derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SKYLINEPRODUCT_H
diff --git a/src/EigenUnsupported/src/Skyline/SkylineStorage.h b/src/EigenUnsupported/src/Skyline/SkylineStorage.h
new file mode 100644
index 0000000..cc7514f
--- /dev/null
+++ b/src/EigenUnsupported/src/Skyline/SkylineStorage.h
@@ -0,0 +1,259 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Guillaume Saupin <guillaume.saupin@cea.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SKYLINE_STORAGE_H
+#define EIGEN_SKYLINE_STORAGE_H
+
+namespace Eigen { 
+
+/** Stores a skyline set of values in three structures :
+ * The diagonal elements
+ * The upper elements
+ * The lower elements
+ *
+ */
+template<typename Scalar>
+class SkylineStorage {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef SparseIndex Index;
+public:
+
+    SkylineStorage()
+    : m_diag(0),
+    m_lower(0),
+    m_upper(0),
+    m_lowerProfile(0),
+    m_upperProfile(0),
+    m_diagSize(0),
+    m_upperSize(0),
+    m_lowerSize(0),
+    m_upperProfileSize(0),
+    m_lowerProfileSize(0),
+    m_allocatedSize(0) {
+    }
+
+    SkylineStorage(const SkylineStorage& other)
+    : m_diag(0),
+    m_lower(0),
+    m_upper(0),
+    m_lowerProfile(0),
+    m_upperProfile(0),
+    m_diagSize(0),
+    m_upperSize(0),
+    m_lowerSize(0),
+    m_upperProfileSize(0),
+    m_lowerProfileSize(0),
+    m_allocatedSize(0) {
+        *this = other;
+    }
+
+    SkylineStorage & operator=(const SkylineStorage& other) {
+        resize(other.diagSize(), other.m_upperProfileSize, other.m_lowerProfileSize, other.upperSize(), other.lowerSize());
+        memcpy(m_diag, other.m_diag, m_diagSize * sizeof (Scalar));
+        memcpy(m_upper, other.m_upper, other.upperSize() * sizeof (Scalar));
+        memcpy(m_lower, other.m_lower, other.lowerSize() * sizeof (Scalar));
+        memcpy(m_upperProfile, other.m_upperProfile, m_upperProfileSize * sizeof (Index));
+        memcpy(m_lowerProfile, other.m_lowerProfile, m_lowerProfileSize * sizeof (Index));
+        return *this;
+    }
+
+    void swap(SkylineStorage& other) {
+        std::swap(m_diag, other.m_diag);
+        std::swap(m_upper, other.m_upper);
+        std::swap(m_lower, other.m_lower);
+        std::swap(m_upperProfile, other.m_upperProfile);
+        std::swap(m_lowerProfile, other.m_lowerProfile);
+        std::swap(m_diagSize, other.m_diagSize);
+        std::swap(m_upperSize, other.m_upperSize);
+        std::swap(m_lowerSize, other.m_lowerSize);
+        std::swap(m_allocatedSize, other.m_allocatedSize);
+    }
+
+    ~SkylineStorage() {
+        delete[] m_diag;
+        delete[] m_upper;
+        if (m_upper != m_lower)
+            delete[] m_lower;
+        delete[] m_upperProfile;
+        delete[] m_lowerProfile;
+    }
+
+    void reserve(Index size, Index upperProfileSize, Index lowerProfileSize, Index upperSize, Index lowerSize) {
+        Index newAllocatedSize = size + upperSize + lowerSize;
+        if (newAllocatedSize > m_allocatedSize)
+            reallocate(size, upperProfileSize, lowerProfileSize, upperSize, lowerSize);
+    }
+
+    void squeeze() {
+        if (m_allocatedSize > m_diagSize + m_upperSize + m_lowerSize)
+            reallocate(m_diagSize, m_upperProfileSize, m_lowerProfileSize, m_upperSize, m_lowerSize);
+    }
+
+    void resize(Index diagSize, Index upperProfileSize, Index lowerProfileSize, Index upperSize, Index lowerSize, float reserveSizeFactor = 0) {
+        if (m_allocatedSize < diagSize + upperSize + lowerSize)
+            reallocate(diagSize, upperProfileSize, lowerProfileSize, upperSize + Index(reserveSizeFactor * upperSize), lowerSize + Index(reserveSizeFactor * lowerSize));
+        m_diagSize = diagSize;
+        m_upperSize = upperSize;
+        m_lowerSize = lowerSize;
+        m_upperProfileSize = upperProfileSize;
+        m_lowerProfileSize = lowerProfileSize;
+    }
+
+    inline Index diagSize() const {
+        return m_diagSize;
+    }
+
+    inline Index upperSize() const {
+        return m_upperSize;
+    }
+
+    inline Index lowerSize() const {
+        return m_lowerSize;
+    }
+
+    inline Index upperProfileSize() const {
+        return m_upperProfileSize;
+    }
+
+    inline Index lowerProfileSize() const {
+        return m_lowerProfileSize;
+    }
+
+    inline Index allocatedSize() const {
+        return m_allocatedSize;
+    }
+
+    inline void clear() {
+        m_diagSize = 0;
+    }
+
+    inline Scalar& diag(Index i) {
+        return m_diag[i];
+    }
+
+    inline const Scalar& diag(Index i) const {
+        return m_diag[i];
+    }
+
+    inline Scalar& upper(Index i) {
+        return m_upper[i];
+    }
+
+    inline const Scalar& upper(Index i) const {
+        return m_upper[i];
+    }
+
+    inline Scalar& lower(Index i) {
+        return m_lower[i];
+    }
+
+    inline const Scalar& lower(Index i) const {
+        return m_lower[i];
+    }
+
+    inline Index& upperProfile(Index i) {
+        return m_upperProfile[i];
+    }
+
+    inline const Index& upperProfile(Index i) const {
+        return m_upperProfile[i];
+    }
+
+    inline Index& lowerProfile(Index i) {
+        return m_lowerProfile[i];
+    }
+
+    inline const Index& lowerProfile(Index i) const {
+        return m_lowerProfile[i];
+    }
+
+    static SkylineStorage Map(Index* upperProfile, Index* lowerProfile, Scalar* diag, Scalar* upper, Scalar* lower, Index size, Index upperSize, Index lowerSize) {
+        SkylineStorage res;
+        res.m_upperProfile = upperProfile;
+        res.m_lowerProfile = lowerProfile;
+        res.m_diag = diag;
+        res.m_upper = upper;
+        res.m_lower = lower;
+        res.m_allocatedSize = res.m_diagSize = size;
+        res.m_upperSize = upperSize;
+        res.m_lowerSize = lowerSize;
+        return res;
+    }
+
+    inline void reset() {
+        memset(m_diag, 0, m_diagSize * sizeof (Scalar));
+        memset(m_upper, 0, m_upperSize * sizeof (Scalar));
+        memset(m_lower, 0, m_lowerSize * sizeof (Scalar));
+        memset(m_upperProfile, 0, m_diagSize * sizeof (Index));
+        memset(m_lowerProfile, 0, m_diagSize * sizeof (Index));
+    }
+
+    void prune(Scalar reference, RealScalar epsilon = dummy_precision<RealScalar>()) {
+        //TODO
+    }
+
+protected:
+
+    inline void reallocate(Index diagSize, Index upperProfileSize, Index lowerProfileSize, Index upperSize, Index lowerSize) {
+
+        Scalar* diag = new Scalar[diagSize];
+        Scalar* upper = new Scalar[upperSize];
+        Scalar* lower = new Scalar[lowerSize];
+        Index* upperProfile = new Index[upperProfileSize];
+        Index* lowerProfile = new Index[lowerProfileSize];
+
+        Index copyDiagSize = (std::min)(diagSize, m_diagSize);
+        Index copyUpperSize = (std::min)(upperSize, m_upperSize);
+        Index copyLowerSize = (std::min)(lowerSize, m_lowerSize);
+        Index copyUpperProfileSize = (std::min)(upperProfileSize, m_upperProfileSize);
+        Index copyLowerProfileSize = (std::min)(lowerProfileSize, m_lowerProfileSize);
+
+        // copy
+        memcpy(diag, m_diag, copyDiagSize * sizeof (Scalar));
+        memcpy(upper, m_upper, copyUpperSize * sizeof (Scalar));
+        memcpy(lower, m_lower, copyLowerSize * sizeof (Scalar));
+        memcpy(upperProfile, m_upperProfile, copyUpperProfileSize * sizeof (Index));
+        memcpy(lowerProfile, m_lowerProfile, copyLowerProfileSize * sizeof (Index));
+
+
+
+        // delete old stuff
+        delete[] m_diag;
+        delete[] m_upper;
+        delete[] m_lower;
+        delete[] m_upperProfile;
+        delete[] m_lowerProfile;
+        m_diag = diag;
+        m_upper = upper;
+        m_lower = lower;
+        m_upperProfile = upperProfile;
+        m_lowerProfile = lowerProfile;
+        m_allocatedSize = diagSize + upperSize + lowerSize;
+        m_upperSize = upperSize;
+        m_lowerSize = lowerSize;
+    }
+
+public:
+    Scalar* m_diag;
+    Scalar* m_upper;
+    Scalar* m_lower;
+    Index* m_upperProfile;
+    Index* m_lowerProfile;
+    Index m_diagSize;
+    Index m_upperSize;
+    Index m_lowerSize;
+    Index m_upperProfileSize;
+    Index m_lowerProfileSize;
+    Index m_allocatedSize;
+
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_SKYLINE_STORAGE_H
diff --git a/src/EigenUnsupported/src/Skyline/SkylineUtil.h b/src/EigenUnsupported/src/Skyline/SkylineUtil.h
new file mode 100644
index 0000000..75eb612
--- /dev/null
+++ b/src/EigenUnsupported/src/Skyline/SkylineUtil.h
@@ -0,0 +1,89 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Guillaume Saupin <guillaume.saupin@cea.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SKYLINEUTIL_H
+#define EIGEN_SKYLINEUTIL_H
+
+namespace Eigen { 
+
+#ifdef NDEBUG
+#define EIGEN_DBG_SKYLINE(X)
+#else
+#define EIGEN_DBG_SKYLINE(X) X
+#endif
+
+const unsigned int SkylineBit = 0x1200;
+template<typename Lhs, typename Rhs, int ProductMode> class SkylineProduct;
+enum AdditionalProductEvaluationMode {SkylineTimeDenseProduct, SkylineTimeSkylineProduct, DenseTimeSkylineProduct};
+enum {IsSkyline = SkylineBit};
+
+
+#define EIGEN_SKYLINE_INHERIT_ASSIGNMENT_OPERATOR(Derived, Op) \
+template<typename OtherDerived> \
+EIGEN_STRONG_INLINE Derived& operator Op(const Eigen::SkylineMatrixBase<OtherDerived>& other) \
+{ \
+  return Base::operator Op(other.derived()); \
+} \
+EIGEN_STRONG_INLINE Derived& operator Op(const Derived& other) \
+{ \
+  return Base::operator Op(other); \
+}
+
+#define EIGEN_SKYLINE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, Op) \
+template<typename Other> \
+EIGEN_STRONG_INLINE Derived& operator Op(const Other& scalar) \
+{ \
+  return Base::operator Op(scalar); \
+}
+
+#define EIGEN_SKYLINE_INHERIT_ASSIGNMENT_OPERATORS(Derived) \
+  EIGEN_SKYLINE_INHERIT_ASSIGNMENT_OPERATOR(Derived, =) \
+  EIGEN_SKYLINE_INHERIT_ASSIGNMENT_OPERATOR(Derived, +=) \
+  EIGEN_SKYLINE_INHERIT_ASSIGNMENT_OPERATOR(Derived, -=) \
+  EIGEN_SKYLINE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, *=) \
+  EIGEN_SKYLINE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=)
+
+#define _EIGEN_SKYLINE_GENERIC_PUBLIC_INTERFACE(Derived, BaseClass) \
+  typedef BaseClass Base; \
+  typedef typename Eigen::internal::traits<Derived>::Scalar Scalar; \
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; \
+  typedef typename Eigen::internal::traits<Derived>::StorageKind StorageKind; \
+  typedef typename Eigen::internal::index<StorageKind>::type Index; \
+  enum {  Flags = Eigen::internal::traits<Derived>::Flags, };
+
+#define EIGEN_SKYLINE_GENERIC_PUBLIC_INTERFACE(Derived) \
+  _EIGEN_SKYLINE_GENERIC_PUBLIC_INTERFACE(Derived, Eigen::SkylineMatrixBase<Derived>)
+
+template<typename Derived> class SkylineMatrixBase;
+template<typename _Scalar, int _Flags = 0> class SkylineMatrix;
+template<typename _Scalar, int _Flags = 0> class DynamicSkylineMatrix;
+template<typename _Scalar, int _Flags = 0> class SkylineVector;
+template<typename _Scalar, int _Flags = 0> class MappedSkylineMatrix;
+
+namespace internal {
+
+template<typename Lhs, typename Rhs> struct skyline_product_mode;
+template<typename Lhs, typename Rhs, int ProductMode = skyline_product_mode<Lhs,Rhs>::value> struct SkylineProductReturnType;
+
+template<typename T> class eval<T,IsSkyline>
+{
+    typedef typename traits<T>::Scalar _Scalar;
+    enum {
+          _Flags = traits<T>::Flags
+    };
+
+  public:
+    typedef SkylineMatrix<_Scalar, _Flags> type;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SKYLINEUTIL_H
diff --git a/src/EigenUnsupported/src/SparseExtra/BlockOfDynamicSparseMatrix.h b/src/EigenUnsupported/src/SparseExtra/BlockOfDynamicSparseMatrix.h
new file mode 100644
index 0000000..e9ec746
--- /dev/null
+++ b/src/EigenUnsupported/src/SparseExtra/BlockOfDynamicSparseMatrix.h
@@ -0,0 +1,122 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_BLOCKFORDYNAMICMATRIX_H
+#define EIGEN_SPARSE_BLOCKFORDYNAMICMATRIX_H
+
+namespace Eigen { 
+
+#if 0
+
+// NOTE Have to be reimplemented as a specialization of BlockImpl< DynamicSparseMatrix<_Scalar, _Options, _Index>, ... >
+// See SparseBlock.h for an example
+
+
+/***************************************************************************
+* specialisation for DynamicSparseMatrix
+***************************************************************************/
+
+template<typename _Scalar, int _Options, typename _Index, int Size>
+class SparseInnerVectorSet<DynamicSparseMatrix<_Scalar, _Options, _Index>, Size>
+  : public SparseMatrixBase<SparseInnerVectorSet<DynamicSparseMatrix<_Scalar, _Options, _Index>, Size> >
+{
+    typedef DynamicSparseMatrix<_Scalar, _Options, _Index> MatrixType;
+  public:
+
+    enum { IsRowMajor = internal::traits<SparseInnerVectorSet>::IsRowMajor };
+
+    EIGEN_SPARSE_PUBLIC_INTERFACE(SparseInnerVectorSet)
+    class InnerIterator: public MatrixType::InnerIterator
+    {
+      public:
+        inline InnerIterator(const SparseInnerVectorSet& xpr, Index outer)
+          : MatrixType::InnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer)
+        {}
+        inline Index row() const { return IsRowMajor ? m_outer : this->index(); }
+        inline Index col() const { return IsRowMajor ? this->index() : m_outer; }
+      protected:
+        Index m_outer;
+    };
+
+    inline SparseInnerVectorSet(const MatrixType& matrix, Index outerStart, Index outerSize)
+      : m_matrix(matrix), m_outerStart(outerStart), m_outerSize(outerSize)
+    {
+      eigen_assert( (outerStart>=0) && ((outerStart+outerSize)<=matrix.outerSize()) );
+    }
+
+    inline SparseInnerVectorSet(const MatrixType& matrix, Index outer)
+      : m_matrix(matrix), m_outerStart(outer), m_outerSize(Size)
+    {
+      eigen_assert(Size!=Dynamic);
+      eigen_assert( (outer>=0) && (outer<matrix.outerSize()) );
+    }
+
+    template<typename OtherDerived>
+    inline SparseInnerVectorSet& operator=(const SparseMatrixBase<OtherDerived>& other)
+    {
+      if (IsRowMajor != ((OtherDerived::Flags&RowMajorBit)==RowMajorBit))
+      {
+        // need to transpose => perform a block evaluation followed by a big swap
+        DynamicSparseMatrix<Scalar,IsRowMajor?RowMajorBit:0> aux(other);
+        *this = aux.markAsRValue();
+      }
+      else
+      {
+        // evaluate/copy vector per vector
+        for (Index j=0; j<m_outerSize.value(); ++j)
+        {
+          SparseVector<Scalar,IsRowMajor ? RowMajorBit : 0> aux(other.innerVector(j));
+          m_matrix.const_cast_derived()._data()[m_outerStart+j].swap(aux._data());
+        }
+      }
+      return *this;
+    }
+
+    inline SparseInnerVectorSet& operator=(const SparseInnerVectorSet& other)
+    {
+      return operator=<SparseInnerVectorSet>(other);
+    }
+
+    Index nonZeros() const
+    {
+      Index count = 0;
+      for (Index j=0; j<m_outerSize.value(); ++j)
+        count += m_matrix._data()[m_outerStart+j].size();
+      return count;
+    }
+
+    const Scalar& lastCoeff() const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(SparseInnerVectorSet);
+      eigen_assert(m_matrix.data()[m_outerStart].size()>0);
+      return m_matrix.data()[m_outerStart].vale(m_matrix.data()[m_outerStart].size()-1);
+    }
+
+//     template<typename Sparse>
+//     inline SparseInnerVectorSet& operator=(const SparseMatrixBase<OtherDerived>& other)
+//     {
+//       return *this;
+//     }
+
+    EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); }
+    EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); }
+
+  protected:
+
+    const typename MatrixType::Nested m_matrix;
+    Index m_outerStart;
+    const internal::variable_if_dynamic<Index, Size> m_outerSize;
+
+};
+
+#endif
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_BLOCKFORDYNAMICMATRIX_H
diff --git a/src/EigenUnsupported/src/SparseExtra/BlockSparseMatrix.h b/src/EigenUnsupported/src/SparseExtra/BlockSparseMatrix.h
new file mode 100644
index 0000000..536a0c3
--- /dev/null
+++ b/src/EigenUnsupported/src/SparseExtra/BlockSparseMatrix.h
@@ -0,0 +1,1079 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2013 Desire Nuentsa <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2013 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEBLOCKMATRIX_H
+#define EIGEN_SPARSEBLOCKMATRIX_H
+
+namespace Eigen { 
+/** \ingroup SparseCore_Module
+  *
+  * \class BlockSparseMatrix
+  *
+  * \brief A versatile sparse matrix representation where each element is a block
+  *
+  * This class provides routines to manipulate block sparse matrices stored in a
+  * BSR-like representation. There are two main types :
+  *
+  * 1. All blocks have the same number of rows and columns, called block size
+  * in the following. In this case, if this block size is known at compile time,
+  * it can be given as a template parameter like
+  * \code
+  * BlockSparseMatrix<Scalar, 3, ColMajor> bmat(b_rows, b_cols);
+  * \endcode
+  * Here, bmat is a b_rows x b_cols block sparse matrix
+  * where each coefficient is a 3x3 dense matrix.
+  * If the block size is fixed but will be given at runtime,
+  * \code
+  * BlockSparseMatrix<Scalar, Dynamic, ColMajor> bmat(b_rows, b_cols);
+  * bmat.setBlockSize(block_size);
+  * \endcode
+  *
+  * 2. The second case is for variable-block sparse matrices.
+  * Here each block has its own dimensions. The only restriction is that all the blocks
+  * in a row (resp. a column) should have the same number of rows (resp. of columns).
+  * It is thus required in this case to describe the layout of the matrix by calling
+  * setBlockLayout(rowBlocks, colBlocks).
+  *
+  * In any of the previous case, the matrix can be filled by calling setFromTriplets().
+  * A regular sparse matrix can be converted to a block sparse matrix and vice versa.
+  * It is obviously required to describe the block layout beforehand by calling either
+  * setBlockSize() for fixed-size blocks or setBlockLayout for variable-size blocks.
+  *
+  * \tparam _Scalar The Scalar type
+  * \tparam _BlockAtCompileTime The block layout option. It takes the following values
+  * Dynamic : block size known at runtime
+  * a numeric number : fixed-size block known at compile time
+  */
+template<typename _Scalar, int _BlockAtCompileTime=Dynamic, int _Options=ColMajor, typename _StorageIndex=int> class BlockSparseMatrix;
+
+template<typename BlockSparseMatrixT> class BlockSparseMatrixView;
+
+namespace internal {
+template<typename _Scalar, int _BlockAtCompileTime, int _Options, typename _Index>
+struct traits<BlockSparseMatrix<_Scalar,_BlockAtCompileTime,_Options, _Index> >
+{
+  typedef _Scalar Scalar;
+  typedef _Index Index;
+  typedef Sparse StorageKind; // FIXME Where is it used ??
+  typedef MatrixXpr XprKind;
+  enum {
+    RowsAtCompileTime = Dynamic,
+    ColsAtCompileTime = Dynamic,
+    MaxRowsAtCompileTime = Dynamic,
+    MaxColsAtCompileTime = Dynamic,
+    BlockSize = _BlockAtCompileTime,
+    Flags = _Options | NestByRefBit | LvalueBit,
+    CoeffReadCost = NumTraits<Scalar>::ReadCost,
+    SupportedAccessPatterns = InnerRandomAccessPattern
+  };
+};
+template<typename BlockSparseMatrixT>
+struct traits<BlockSparseMatrixView<BlockSparseMatrixT> >
+{
+  typedef Ref<Matrix<typename BlockSparseMatrixT::Scalar, BlockSparseMatrixT::BlockSize, BlockSparseMatrixT::BlockSize> > Scalar;
+  typedef Ref<Matrix<typename BlockSparseMatrixT::RealScalar, BlockSparseMatrixT::BlockSize, BlockSparseMatrixT::BlockSize> > RealScalar;
+
+};
+
+// Function object to sort a triplet list
+template<typename Iterator, bool IsColMajor>
+struct TripletComp
+{
+  typedef typename Iterator::value_type Triplet;
+  bool operator()(const Triplet& a, const Triplet& b)
+  { if(IsColMajor)
+      return ((a.col() == b.col() && a.row() < b.row()) || (a.col() < b.col()));
+    else
+      return ((a.row() == b.row() && a.col() < b.col()) || (a.row() < b.row()));
+  }
+};
+} // end namespace internal
+
+
+/* Proxy to view the block sparse matrix as a regular sparse matrix */
+template<typename BlockSparseMatrixT>
+class BlockSparseMatrixView : public SparseMatrixBase<BlockSparseMatrixT>
+{
+  public:
+    typedef Ref<typename BlockSparseMatrixT::BlockScalar> Scalar;
+    typedef Ref<typename BlockSparseMatrixT::BlockRealScalar> RealScalar;
+    typedef typename BlockSparseMatrixT::Index Index;
+    typedef  BlockSparseMatrixT Nested;
+    enum {
+      Flags = BlockSparseMatrixT::Options,
+      Options = BlockSparseMatrixT::Options,
+      RowsAtCompileTime = BlockSparseMatrixT::RowsAtCompileTime,
+      ColsAtCompileTime = BlockSparseMatrixT::ColsAtCompileTime,
+      MaxColsAtCompileTime = BlockSparseMatrixT::MaxColsAtCompileTime,
+      MaxRowsAtCompileTime = BlockSparseMatrixT::MaxRowsAtCompileTime
+    };
+  public:
+    BlockSparseMatrixView(const BlockSparseMatrixT& spblockmat)
+     : m_spblockmat(spblockmat)
+    {}
+
+    Index outerSize() const
+    {
+      return (Flags&RowMajorBit) == 1 ? this->rows() : this->cols();
+    }
+    Index cols() const
+    {
+      return m_spblockmat.blockCols();
+    }
+    Index rows() const
+    {
+      return m_spblockmat.blockRows();
+    }
+    Scalar coeff(Index row, Index col)
+    {
+      return m_spblockmat.coeff(row, col);
+    }
+    Scalar coeffRef(Index row, Index col)
+    {
+      return m_spblockmat.coeffRef(row, col);
+    }
+    // Wrapper to iterate over all blocks
+    class InnerIterator : public BlockSparseMatrixT::BlockInnerIterator
+    {
+      public:
+      InnerIterator(const BlockSparseMatrixView& mat, Index outer)
+          : BlockSparseMatrixT::BlockInnerIterator(mat.m_spblockmat, outer)
+      {}
+
+    };
+
+  protected:
+    const BlockSparseMatrixT& m_spblockmat;
+};
+
+// Proxy to view a regular vector as a block vector
+template<typename BlockSparseMatrixT, typename VectorType>
+class BlockVectorView
+{
+  public:
+    enum {
+      BlockSize = BlockSparseMatrixT::BlockSize,
+      ColsAtCompileTime = VectorType::ColsAtCompileTime,
+      RowsAtCompileTime = VectorType::RowsAtCompileTime,
+      Flags = VectorType::Flags
+    };
+    typedef Ref<const Matrix<typename BlockSparseMatrixT::Scalar, (RowsAtCompileTime==1)? 1 : BlockSize, (ColsAtCompileTime==1)? 1 : BlockSize> >Scalar;
+    typedef typename BlockSparseMatrixT::Index Index;
+  public:
+    BlockVectorView(const BlockSparseMatrixT& spblockmat, const VectorType& vec)
+    : m_spblockmat(spblockmat),m_vec(vec)
+    { }
+    inline Index cols() const
+    {
+      return m_vec.cols();
+    }
+    inline Index size() const
+    {
+      return m_spblockmat.blockRows();
+    }
+    inline Scalar coeff(Index bi) const
+    {
+      Index startRow = m_spblockmat.blockRowsIndex(bi);
+      Index rowSize = m_spblockmat.blockRowsIndex(bi+1) - startRow;
+      return m_vec.middleRows(startRow, rowSize);
+    }
+    inline Scalar coeff(Index bi, Index j) const
+    {
+      Index startRow = m_spblockmat.blockRowsIndex(bi);
+      Index rowSize = m_spblockmat.blockRowsIndex(bi+1) - startRow;
+      return m_vec.block(startRow, j, rowSize, 1);
+    }
+  protected:
+    const BlockSparseMatrixT& m_spblockmat;
+    const VectorType& m_vec;
+};
+
+template<typename VectorType, typename Index> class BlockVectorReturn;
+
+
+// Proxy to view a regular vector as a block vector
+template<typename BlockSparseMatrixT, typename VectorType>
+class BlockVectorReturn
+{
+  public:
+    enum {
+      ColsAtCompileTime = VectorType::ColsAtCompileTime,
+      RowsAtCompileTime = VectorType::RowsAtCompileTime,
+      Flags = VectorType::Flags
+    };
+    typedef Ref<Matrix<typename VectorType::Scalar, RowsAtCompileTime, ColsAtCompileTime> > Scalar;
+    typedef typename BlockSparseMatrixT::Index Index;
+  public:
+    BlockVectorReturn(const BlockSparseMatrixT& spblockmat, VectorType& vec)
+    : m_spblockmat(spblockmat),m_vec(vec)
+    { }
+    inline Index size() const
+    {
+      return m_spblockmat.blockRows();
+    }
+    inline Scalar coeffRef(Index bi)
+    {
+      Index startRow = m_spblockmat.blockRowsIndex(bi);
+      Index rowSize = m_spblockmat.blockRowsIndex(bi+1) - startRow;
+      return m_vec.middleRows(startRow, rowSize);
+    }
+    inline Scalar coeffRef(Index bi, Index j)
+    {
+      Index startRow = m_spblockmat.blockRowsIndex(bi);
+      Index rowSize = m_spblockmat.blockRowsIndex(bi+1) - startRow;
+      return m_vec.block(startRow, j, rowSize, 1);
+    }
+
+  protected:
+    const BlockSparseMatrixT& m_spblockmat;
+    VectorType& m_vec;
+};
+
+// Block version of the sparse dense product
+template<typename Lhs, typename Rhs>
+class BlockSparseTimeDenseProduct;
+
+namespace internal {
+
+template<typename BlockSparseMatrixT, typename VecType>
+struct traits<BlockSparseTimeDenseProduct<BlockSparseMatrixT, VecType> >
+{
+  typedef Dense StorageKind;
+  typedef MatrixXpr XprKind;
+  typedef typename BlockSparseMatrixT::Scalar Scalar;
+  typedef typename BlockSparseMatrixT::Index Index;
+  enum {
+    RowsAtCompileTime = Dynamic,
+    ColsAtCompileTime = Dynamic,
+    MaxRowsAtCompileTime = Dynamic,
+    MaxColsAtCompileTime = Dynamic,
+    Flags = 0,
+    CoeffReadCost = internal::traits<BlockSparseMatrixT>::CoeffReadCost
+  };
+};
+} // end namespace internal
+
+template<typename Lhs, typename Rhs>
+class BlockSparseTimeDenseProduct
+  : public ProductBase<BlockSparseTimeDenseProduct<Lhs,Rhs>, Lhs, Rhs>
+{
+  public:
+    EIGEN_PRODUCT_PUBLIC_INTERFACE(BlockSparseTimeDenseProduct)
+
+    BlockSparseTimeDenseProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
+    {}
+
+    template<typename Dest> void scaleAndAddTo(Dest& dest, const typename Rhs::Scalar& alpha) const
+    {
+      BlockVectorReturn<Lhs,Dest> tmpDest(m_lhs, dest);
+      internal::sparse_time_dense_product( BlockSparseMatrixView<Lhs>(m_lhs),  BlockVectorView<Lhs, Rhs>(m_lhs, m_rhs), tmpDest, alpha);
+    }
+
+  private:
+    BlockSparseTimeDenseProduct& operator=(const BlockSparseTimeDenseProduct&);
+};
+
+template<typename _Scalar, int _BlockAtCompileTime, int _Options, typename _StorageIndex>
+class BlockSparseMatrix : public SparseMatrixBase<BlockSparseMatrix<_Scalar,_BlockAtCompileTime, _Options,_StorageIndex> >
+{
+  public:
+    typedef _Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef _StorageIndex StorageIndex;
+    typedef typename internal::ref_selector<BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, _StorageIndex> >::type Nested;
+
+    enum {
+      Options = _Options,
+      Flags = Options,
+      BlockSize=_BlockAtCompileTime,
+      RowsAtCompileTime = Dynamic,
+      ColsAtCompileTime = Dynamic,
+      MaxRowsAtCompileTime = Dynamic,
+      MaxColsAtCompileTime = Dynamic,
+      IsVectorAtCompileTime = 0,
+      IsColMajor = Flags&RowMajorBit ? 0 : 1
+    };
+    typedef Matrix<Scalar, _BlockAtCompileTime, _BlockAtCompileTime,IsColMajor ? ColMajor : RowMajor> BlockScalar;
+    typedef Matrix<RealScalar, _BlockAtCompileTime, _BlockAtCompileTime,IsColMajor ? ColMajor : RowMajor> BlockRealScalar;
+    typedef typename internal::conditional<_BlockAtCompileTime==Dynamic, Scalar, BlockScalar>::type BlockScalarReturnType;
+    typedef BlockSparseMatrix<Scalar, BlockSize, IsColMajor ? ColMajor : RowMajor, StorageIndex> PlainObject;
+  public:
+    // Default constructor
+    BlockSparseMatrix()
+    : m_innerBSize(0),m_outerBSize(0),m_innerOffset(0),m_outerOffset(0),
+      m_nonzerosblocks(0),m_values(0),m_blockPtr(0),m_indices(0),
+      m_outerIndex(0),m_blockSize(BlockSize)
+    { }
+
+
+    /**
+     * \brief Construct and resize
+     *
+     */
+    BlockSparseMatrix(Index brow, Index bcol)
+      : m_innerBSize(IsColMajor ? brow : bcol),
+        m_outerBSize(IsColMajor ? bcol : brow),
+        m_innerOffset(0),m_outerOffset(0),m_nonzerosblocks(0),
+        m_values(0),m_blockPtr(0),m_indices(0),
+        m_outerIndex(0),m_blockSize(BlockSize)
+    { }
+
+    /**
+     * \brief Copy-constructor
+     */
+    BlockSparseMatrix(const BlockSparseMatrix& other)
+      : m_innerBSize(other.m_innerBSize),m_outerBSize(other.m_outerBSize),
+        m_nonzerosblocks(other.m_nonzerosblocks),m_nonzeros(other.m_nonzeros),
+        m_blockPtr(0),m_blockSize(other.m_blockSize)
+    {
+      // should we allow copying between variable-size blocks and fixed-size blocks ??
+      eigen_assert(m_blockSize == BlockSize && " CAN NOT COPY BETWEEN FIXED-SIZE AND VARIABLE-SIZE BLOCKS");
+
+      std::copy(other.m_innerOffset, other.m_innerOffset+m_innerBSize+1, m_innerOffset);
+      std::copy(other.m_outerOffset, other.m_outerOffset+m_outerBSize+1, m_outerOffset);
+      std::copy(other.m_values, other.m_values+m_nonzeros, m_values);
+
+      if(m_blockSize != Dynamic)
+        std::copy(other.m_blockPtr, other.m_blockPtr+m_nonzerosblocks, m_blockPtr);
+
+      std::copy(other.m_indices, other.m_indices+m_nonzerosblocks, m_indices);
+      std::copy(other.m_outerIndex, other.m_outerIndex+m_outerBSize, m_outerIndex);
+    }
+
+    friend void swap(BlockSparseMatrix& first, BlockSparseMatrix& second)
+    {
+      std::swap(first.m_innerBSize, second.m_innerBSize);
+      std::swap(first.m_outerBSize, second.m_outerBSize);
+      std::swap(first.m_innerOffset, second.m_innerOffset);
+      std::swap(first.m_outerOffset, second.m_outerOffset);
+      std::swap(first.m_nonzerosblocks, second.m_nonzerosblocks);
+      std::swap(first.m_nonzeros, second.m_nonzeros);
+      std::swap(first.m_values, second.m_values);
+      std::swap(first.m_blockPtr, second.m_blockPtr);
+      std::swap(first.m_indices, second.m_indices);
+      std::swap(first.m_outerIndex, second.m_outerIndex);
+      std::swap(first.m_BlockSize, second.m_blockSize);
+    }
+
+    BlockSparseMatrix& operator=(BlockSparseMatrix other)
+    {
+      //Copy-and-swap paradigm ... avoid leaked data if thrown
+      swap(*this, other);
+      return *this;
+    }
+
+    // Destructor
+    ~BlockSparseMatrix()
+    {
+      delete[] m_outerIndex;
+      delete[] m_innerOffset;
+      delete[] m_outerOffset;
+      delete[] m_indices;
+      delete[] m_blockPtr;
+      delete[] m_values;
+    }
+
+
+    /**
+      * \brief Constructor from a sparse matrix
+      *
+      */
+    template<typename MatrixType>
+    inline BlockSparseMatrix(const MatrixType& spmat) : m_blockSize(BlockSize)
+    {
+      EIGEN_STATIC_ASSERT((m_blockSize != Dynamic), THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE);
+
+      *this = spmat;
+    }
+
+    /**
+      * \brief Assignment from a sparse matrix with the same storage order
+      *
+      * Convert from a sparse matrix to block sparse matrix.
+      * \warning Before calling this function, tt is necessary to call
+      * either setBlockLayout() (matrices with variable-size blocks)
+      * or setBlockSize() (for fixed-size blocks).
+      */
+    template<typename MatrixType>
+    inline BlockSparseMatrix& operator=(const MatrixType& spmat)
+    {
+      eigen_assert((m_innerBSize != 0 && m_outerBSize != 0)
+                   && "Trying to assign to a zero-size matrix, call resize() first");
+      eigen_assert(((MatrixType::Options&RowMajorBit) != IsColMajor) && "Wrong storage order");
+      typedef SparseMatrix<bool,MatrixType::Options,typename MatrixType::Index> MatrixPatternType;
+      MatrixPatternType  blockPattern(blockRows(), blockCols());
+      m_nonzeros = 0;
+
+      // First, compute the number of nonzero blocks and their locations
+      for(StorageIndex bj = 0; bj < m_outerBSize; ++bj)
+      {
+        // Browse each outer block and compute the structure
+        std::vector<bool> nzblocksFlag(m_innerBSize,false);  // Record the existing blocks
+        blockPattern.startVec(bj);
+        for(StorageIndex j = blockOuterIndex(bj); j < blockOuterIndex(bj+1); ++j)
+        {
+          typename MatrixType::InnerIterator it_spmat(spmat, j);
+          for(; it_spmat; ++it_spmat)
+          {
+            StorageIndex bi = innerToBlock(it_spmat.index()); // Index of the current nonzero block
+            if(!nzblocksFlag[bi])
+            {
+              // Save the index of this nonzero block
+              nzblocksFlag[bi] = true;
+              blockPattern.insertBackByOuterInnerUnordered(bj, bi) = true;
+              // Compute the total number of nonzeros (including explicit zeros in blocks)
+              m_nonzeros += blockOuterSize(bj) * blockInnerSize(bi);
+            }
+          }
+        } // end current outer block
+      }
+      blockPattern.finalize();
+
+      // Allocate the internal arrays
+      setBlockStructure(blockPattern);
+
+      for(StorageIndex nz = 0; nz < m_nonzeros; ++nz) m_values[nz] = Scalar(0);
+      for(StorageIndex bj = 0; bj < m_outerBSize; ++bj)
+      {
+        // Now copy the values
+        for(StorageIndex j = blockOuterIndex(bj); j < blockOuterIndex(bj+1); ++j)
+        {
+          // Browse the outer block column by column (for column-major matrices)
+          typename MatrixType::InnerIterator it_spmat(spmat, j);
+          for(; it_spmat; ++it_spmat)
+          {
+            StorageIndex idx = 0; // Position of this block in the column block
+            StorageIndex bi = innerToBlock(it_spmat.index()); // Index of the current nonzero block
+            // Go to the inner block where this element belongs to
+            while(bi > m_indices[m_outerIndex[bj]+idx]) ++idx; // Not expensive for ordered blocks
+            StorageIndex idxVal;// Get the right position in the array of values for this element
+            if(m_blockSize == Dynamic)
+            {
+              // Offset from all blocks before ...
+              idxVal =  m_blockPtr[m_outerIndex[bj]+idx];
+              // ... and offset inside the block
+              idxVal += (j - blockOuterIndex(bj)) * blockOuterSize(bj) + it_spmat.index() - m_innerOffset[bi];
+            }
+            else
+            {
+              // All blocks before
+              idxVal = (m_outerIndex[bj] + idx) * m_blockSize * m_blockSize;
+              // inside the block
+              idxVal += (j - blockOuterIndex(bj)) * m_blockSize + (it_spmat.index()%m_blockSize);
+            }
+            // Insert the value
+            m_values[idxVal] = it_spmat.value();
+          } // end of this column
+        } // end of this block
+      } // end of this outer block
+
+      return *this;
+    }
+
+    /**
+      * \brief Set the nonzero block pattern of the matrix
+      *
+      * Given a sparse matrix describing the nonzero block pattern,
+      * this function prepares the internal pointers for values.
+      * After calling this function, any *nonzero* block (bi, bj) can be set
+      * with a simple call to coeffRef(bi,bj).
+      *
+      *
+      * \warning Before calling this function, tt is necessary to call
+      * either setBlockLayout() (matrices with variable-size blocks)
+      * or setBlockSize() (for fixed-size blocks).
+      *
+      * \param blockPattern Sparse matrix of boolean elements describing the block structure
+      *
+      * \sa setBlockLayout() \sa setBlockSize()
+      */
+    template<typename MatrixType>
+    void setBlockStructure(const MatrixType& blockPattern)
+    {
+      resize(blockPattern.rows(), blockPattern.cols());
+      reserve(blockPattern.nonZeros());
+
+      // Browse the block pattern and set up the various pointers
+      m_outerIndex[0] = 0;
+      if(m_blockSize == Dynamic) m_blockPtr[0] = 0;
+      for(StorageIndex nz = 0; nz < m_nonzeros; ++nz) m_values[nz] = Scalar(0);
+      for(StorageIndex bj = 0; bj < m_outerBSize; ++bj)
+      {
+        //Browse each outer block
+
+        //First, copy and save the indices of nonzero blocks
+        //FIXME : find a way to avoid this ...
+        std::vector<int> nzBlockIdx;
+        typename MatrixType::InnerIterator it(blockPattern, bj);
+        for(; it; ++it)
+        {
+          nzBlockIdx.push_back(it.index());
+        }
+        std::sort(nzBlockIdx.begin(), nzBlockIdx.end());
+
+        // Now, fill block indices and (eventually) pointers to blocks
+        for(StorageIndex idx = 0; idx < nzBlockIdx.size(); ++idx)
+        {
+          StorageIndex offset = m_outerIndex[bj]+idx; // offset in m_indices
+          m_indices[offset] = nzBlockIdx[idx];
+          if(m_blockSize == Dynamic)
+            m_blockPtr[offset] = m_blockPtr[offset-1] + blockInnerSize(nzBlockIdx[idx]) * blockOuterSize(bj);
+          // There is no blockPtr for fixed-size blocks... not needed !???
+        }
+        // Save the pointer to the next outer block
+        m_outerIndex[bj+1] = m_outerIndex[bj] + nzBlockIdx.size();
+      }
+    }
+
+    /**
+      * \brief Set the number of rows and columns blocks
+      */
+    inline void resize(Index brow, Index bcol)
+    {
+      m_innerBSize = IsColMajor ? brow : bcol;
+      m_outerBSize = IsColMajor ? bcol : brow;
+    }
+
+    /**
+      * \brief set the block size at runtime for fixed-size block layout
+      *
+      * Call this only for fixed-size blocks
+      */
+    inline void setBlockSize(Index blockSize)
+    {
+      m_blockSize = blockSize;
+    }
+
+    /**
+      * \brief Set the row and column block layouts,
+      *
+      * This function set the size of each row and column block.
+      * So this function should be used only for blocks with variable size.
+      * \param rowBlocks : Number of rows per row block
+      * \param colBlocks : Number of columns per column block
+      * \sa resize(), setBlockSize()
+      */
+    inline void setBlockLayout(const VectorXi& rowBlocks, const VectorXi& colBlocks)
+    {
+      const VectorXi& innerBlocks = IsColMajor ? rowBlocks : colBlocks;
+      const VectorXi& outerBlocks = IsColMajor ? colBlocks : rowBlocks;
+      eigen_assert(m_innerBSize == innerBlocks.size() && "CHECK THE NUMBER OF ROW OR COLUMN BLOCKS");
+      eigen_assert(m_outerBSize == outerBlocks.size() && "CHECK THE NUMBER OF ROW OR COLUMN BLOCKS");
+      m_outerBSize = outerBlocks.size();
+      //  starting index of blocks... cumulative sums
+      m_innerOffset = new StorageIndex[m_innerBSize+1];
+      m_outerOffset = new StorageIndex[m_outerBSize+1];
+      m_innerOffset[0] = 0;
+      m_outerOffset[0] = 0;
+      std::partial_sum(&innerBlocks[0], &innerBlocks[m_innerBSize-1]+1, &m_innerOffset[1]);
+      std::partial_sum(&outerBlocks[0], &outerBlocks[m_outerBSize-1]+1, &m_outerOffset[1]);
+
+      // Compute the total number of nonzeros
+      m_nonzeros = 0;
+      for(StorageIndex bj = 0; bj < m_outerBSize; ++bj)
+        for(StorageIndex bi = 0; bi < m_innerBSize; ++bi)
+          m_nonzeros += outerBlocks[bj] * innerBlocks[bi];
+
+    }
+
+    /**
+      * \brief Allocate the internal array of pointers to blocks and their inner indices
+      *
+      * \note For fixed-size blocks, call setBlockSize() to set the block.
+      * And For variable-size blocks, call setBlockLayout() before using this function
+      *
+      * \param nonzerosblocks Number of nonzero blocks. The total number of nonzeros is
+      * is computed in setBlockLayout() for variable-size blocks
+      * \sa setBlockSize()
+      */
+    inline void reserve(const Index nonzerosblocks)
+    {
+      eigen_assert((m_innerBSize != 0 && m_outerBSize != 0) &&
+          "TRYING TO RESERVE ZERO-SIZE MATRICES, CALL resize() first");
+
+      //FIXME Should free if already allocated
+      m_outerIndex = new StorageIndex[m_outerBSize+1];
+
+      m_nonzerosblocks = nonzerosblocks;
+      if(m_blockSize != Dynamic)
+      {
+        m_nonzeros = nonzerosblocks * (m_blockSize * m_blockSize);
+        m_blockPtr = 0;
+      }
+      else
+      {
+        // m_nonzeros  is already computed in setBlockLayout()
+        m_blockPtr = new StorageIndex[m_nonzerosblocks+1];
+      }
+      m_indices = new StorageIndex[m_nonzerosblocks+1];
+      m_values = new Scalar[m_nonzeros];
+    }
+
+
+    /**
+      * \brief Fill values in a matrix  from a triplet list.
+      *
+      * Each triplet item has a block stored in an Eigen dense matrix.
+      * The InputIterator class should provide the functions row(), col() and value()
+      *
+      * \note For fixed-size blocks, call setBlockSize() before this function.
+      *
+      * FIXME Do not accept duplicates
+      */
+    template<typename InputIterator>
+    void setFromTriplets(const InputIterator& begin, const InputIterator& end)
+    {
+      eigen_assert((m_innerBSize!=0 && m_outerBSize !=0) && "ZERO BLOCKS, PLEASE CALL resize() before");
+
+      /* First, sort the triplet list
+        * FIXME This can be unnecessarily expensive since only the inner indices have to be sorted
+        * The best approach is like in SparseMatrix::setFromTriplets()
+        */
+      internal::TripletComp<InputIterator, IsColMajor> tripletcomp;
+      std::sort(begin, end, tripletcomp);
+
+      /* Count the number of rows and column blocks,
+       * and the number of nonzero blocks per outer dimension
+       */
+      VectorXi rowBlocks(m_innerBSize); // Size of each block row
+      VectorXi colBlocks(m_outerBSize); // Size of each block column
+      rowBlocks.setZero(); colBlocks.setZero();
+      VectorXi nzblock_outer(m_outerBSize); // Number of nz blocks per outer vector
+      VectorXi nz_outer(m_outerBSize); // Number of nz per outer vector...for variable-size blocks
+      nzblock_outer.setZero();
+      nz_outer.setZero();
+      for(InputIterator it(begin); it !=end; ++it)
+      {
+        eigen_assert(it->row() >= 0 && it->row() < this->blockRows() && it->col() >= 0 && it->col() < this->blockCols());
+        eigen_assert((it->value().rows() == it->value().cols() && (it->value().rows() == m_blockSize))
+                     || (m_blockSize == Dynamic));
+
+        if(m_blockSize == Dynamic)
+        {
+          eigen_assert((rowBlocks[it->row()] == 0 || rowBlocks[it->row()] == it->value().rows()) &&
+              "NON CORRESPONDING SIZES FOR ROW BLOCKS");
+          eigen_assert((colBlocks[it->col()] == 0 || colBlocks[it->col()] == it->value().cols()) &&
+              "NON CORRESPONDING SIZES FOR COLUMN BLOCKS");
+          rowBlocks[it->row()] =it->value().rows();
+          colBlocks[it->col()] = it->value().cols();
+        }
+        nz_outer(IsColMajor ? it->col() : it->row()) += it->value().rows() * it->value().cols();
+        nzblock_outer(IsColMajor ? it->col() : it->row())++;
+      }
+      // Allocate member arrays
+      if(m_blockSize == Dynamic) setBlockLayout(rowBlocks, colBlocks);
+      StorageIndex nzblocks = nzblock_outer.sum();
+      reserve(nzblocks);
+
+       // Temporary markers
+      VectorXi block_id(m_outerBSize); // To be used as a block marker during insertion
+
+      // Setup outer index pointers and markers
+      m_outerIndex[0] = 0;
+      if (m_blockSize == Dynamic)  m_blockPtr[0] =  0;
+      for(StorageIndex bj = 0; bj < m_outerBSize; ++bj)
+      {
+        m_outerIndex[bj+1] = m_outerIndex[bj] + nzblock_outer(bj);
+        block_id(bj) = m_outerIndex[bj];
+        if(m_blockSize==Dynamic)
+        {
+          m_blockPtr[m_outerIndex[bj+1]] = m_blockPtr[m_outerIndex[bj]] + nz_outer(bj);
+        }
+      }
+
+      // Fill the matrix
+      for(InputIterator it(begin); it!=end; ++it)
+      {
+        StorageIndex outer = IsColMajor ? it->col() : it->row();
+        StorageIndex inner = IsColMajor ? it->row() : it->col();
+        m_indices[block_id(outer)] = inner;
+        StorageIndex block_size = it->value().rows()*it->value().cols();
+        StorageIndex nz_marker = blockPtr(block_id[outer]);
+        memcpy(&(m_values[nz_marker]), it->value().data(), block_size * sizeof(Scalar));
+        if(m_blockSize == Dynamic)
+        {
+          m_blockPtr[block_id(outer)+1] = m_blockPtr[block_id(outer)] + block_size;
+        }
+        block_id(outer)++;
+      }
+
+      // An alternative when the outer indices are sorted...no need to use an array of markers
+//      for(Index bcol = 0; bcol < m_outerBSize; ++bcol)
+//      {
+//      Index id = 0, id_nz = 0, id_nzblock = 0;
+//      for(InputIterator it(begin); it!=end; ++it)
+//      {
+//        while (id<bcol) // one pass should do the job unless there are empty columns
+//        {
+//          id++;
+//          m_outerIndex[id+1]=m_outerIndex[id];
+//        }
+//        m_outerIndex[id+1] += 1;
+//        m_indices[id_nzblock]=brow;
+//        Index block_size = it->value().rows()*it->value().cols();
+//        m_blockPtr[id_nzblock+1] = m_blockPtr[id_nzblock] + block_size;
+//        id_nzblock++;
+//        memcpy(&(m_values[id_nz]),it->value().data(), block_size*sizeof(Scalar));
+//        id_nz += block_size;
+//      }
+//      while(id < m_outerBSize-1) // Empty columns at the end
+//      {
+//        id++;
+//        m_outerIndex[id+1]=m_outerIndex[id];
+//      }
+//      }
+    }
+
+
+    /**
+      * \returns the number of rows
+      */
+    inline Index rows() const
+    {
+//      return blockRows();
+      return (IsColMajor ? innerSize() : outerSize());
+    }
+
+    /**
+      * \returns the number of cols
+      */
+    inline Index cols() const
+    {
+//      return blockCols();
+      return (IsColMajor ? outerSize() : innerSize());
+    }
+
+    inline Index innerSize() const
+    {
+      if(m_blockSize == Dynamic) return m_innerOffset[m_innerBSize];
+      else return  (m_innerBSize * m_blockSize) ;
+    }
+
+    inline Index outerSize() const
+    {
+      if(m_blockSize == Dynamic) return m_outerOffset[m_outerBSize];
+      else return  (m_outerBSize * m_blockSize) ;
+    }
+    /** \returns the number of rows grouped by blocks */
+    inline Index blockRows() const
+    {
+      return (IsColMajor ? m_innerBSize : m_outerBSize);
+    }
+    /** \returns the number of columns grouped by blocks */
+    inline Index blockCols() const
+    {
+      return (IsColMajor ? m_outerBSize : m_innerBSize);
+    }
+
+    inline Index outerBlocks() const { return m_outerBSize; }
+    inline Index innerBlocks() const { return m_innerBSize; }
+
+    /** \returns the block index where outer belongs to */
+    inline Index outerToBlock(Index outer) const
+    {
+      eigen_assert(outer < outerSize() && "OUTER INDEX OUT OF BOUNDS");
+
+      if(m_blockSize != Dynamic)
+        return (outer / m_blockSize); // Integer division
+
+      StorageIndex b_outer = 0;
+      while(m_outerOffset[b_outer] <= outer) ++b_outer;
+      return b_outer - 1;
+    }
+    /** \returns  the block index where inner belongs to */
+    inline Index innerToBlock(Index inner) const
+    {
+      eigen_assert(inner < innerSize() && "OUTER INDEX OUT OF BOUNDS");
+
+      if(m_blockSize != Dynamic)
+        return (inner / m_blockSize); // Integer division
+
+      StorageIndex b_inner = 0;
+      while(m_innerOffset[b_inner] <= inner) ++b_inner;
+      return b_inner - 1;
+    }
+
+    /**
+      *\returns a reference to the (i,j) block as an Eigen Dense Matrix
+      */
+    Ref<BlockScalar> coeffRef(Index brow, Index bcol)
+    {
+      eigen_assert(brow < blockRows() && "BLOCK ROW INDEX OUT OF BOUNDS");
+      eigen_assert(bcol < blockCols() && "BLOCK nzblocksFlagCOLUMN OUT OF BOUNDS");
+
+      StorageIndex rsize = IsColMajor ? blockInnerSize(brow): blockOuterSize(bcol);
+      StorageIndex csize = IsColMajor ? blockOuterSize(bcol) : blockInnerSize(brow);
+      StorageIndex inner = IsColMajor ? brow : bcol;
+      StorageIndex outer = IsColMajor ? bcol : brow;
+      StorageIndex offset = m_outerIndex[outer];
+      while(offset < m_outerIndex[outer+1] && m_indices[offset] != inner)
+        offset++;
+      if(m_indices[offset] == inner)
+      {
+        return Map<BlockScalar>(&(m_values[blockPtr(offset)]), rsize, csize);
+      }
+      else
+      {
+        //FIXME the block does not exist, Insert it !!!!!!!!!
+        eigen_assert("DYNAMIC INSERTION IS NOT YET SUPPORTED");
+      }
+    }
+
+    /**
+      * \returns the value of the (i,j) block as an Eigen Dense Matrix
+      */
+    Map<const BlockScalar> coeff(Index brow, Index bcol) const
+    {
+      eigen_assert(brow < blockRows() && "BLOCK ROW INDEX OUT OF BOUNDS");
+      eigen_assert(bcol < blockCols() && "BLOCK COLUMN OUT OF BOUNDS");
+
+      StorageIndex rsize = IsColMajor ? blockInnerSize(brow): blockOuterSize(bcol);
+      StorageIndex csize = IsColMajor ? blockOuterSize(bcol) : blockInnerSize(brow);
+      StorageIndex inner = IsColMajor ? brow : bcol;
+      StorageIndex outer = IsColMajor ? bcol : brow;
+      StorageIndex offset = m_outerIndex[outer];
+      while(offset < m_outerIndex[outer+1] && m_indices[offset] != inner) offset++;
+      if(m_indices[offset] == inner)
+      {
+        return Map<const BlockScalar> (&(m_values[blockPtr(offset)]), rsize, csize);
+      }
+      else
+//        return BlockScalar::Zero(rsize, csize);
+        eigen_assert("NOT YET SUPPORTED");
+    }
+
+    // Block Matrix times vector product
+    template<typename VecType>
+    BlockSparseTimeDenseProduct<BlockSparseMatrix, VecType> operator*(const VecType& lhs) const
+    {
+      return BlockSparseTimeDenseProduct<BlockSparseMatrix, VecType>(*this, lhs);
+    }
+
+    /** \returns the number of nonzero blocks */
+    inline Index nonZerosBlocks() const { return m_nonzerosblocks; }
+    /** \returns the total number of nonzero elements, including eventual explicit zeros in blocks */
+    inline Index nonZeros() const { return m_nonzeros; }
+
+    inline BlockScalarReturnType *valuePtr() {return static_cast<BlockScalarReturnType *>(m_values);}
+//    inline Scalar *valuePtr(){ return m_values; }
+    inline StorageIndex *innerIndexPtr() {return m_indices; }
+    inline const StorageIndex *innerIndexPtr() const {return m_indices; }
+    inline StorageIndex *outerIndexPtr() {return m_outerIndex; }
+    inline const StorageIndex* outerIndexPtr() const {return m_outerIndex; }
+
+    /** \brief for compatibility purposes with the SparseMatrix class */
+    inline bool isCompressed() const {return true;}
+    /**
+      * \returns the starting index of the bi row block
+      */
+    inline Index blockRowsIndex(Index bi) const
+    {
+      return IsColMajor ? blockInnerIndex(bi) : blockOuterIndex(bi);
+    }
+
+    /**
+      * \returns the starting index of the bj col block
+      */
+    inline Index blockColsIndex(Index bj) const
+    {
+      return IsColMajor ? blockOuterIndex(bj) : blockInnerIndex(bj);
+    }
+
+    inline Index blockOuterIndex(Index bj) const
+    {
+      return (m_blockSize == Dynamic) ? m_outerOffset[bj] : (bj * m_blockSize);
+    }
+    inline Index blockInnerIndex(Index bi) const
+    {
+      return (m_blockSize == Dynamic) ? m_innerOffset[bi] : (bi * m_blockSize);
+    }
+
+    // Not needed ???
+    inline Index blockInnerSize(Index bi) const
+    {
+      return (m_blockSize == Dynamic) ? (m_innerOffset[bi+1] - m_innerOffset[bi]) : m_blockSize;
+    }
+    inline Index blockOuterSize(Index bj) const
+    {
+      return (m_blockSize == Dynamic) ? (m_outerOffset[bj+1]- m_outerOffset[bj]) : m_blockSize;
+    }
+
+    /**
+      * \brief Browse the matrix by outer index
+      */
+    class InnerIterator; // Browse column by column
+
+    /**
+      * \brief Browse the matrix by block outer index
+      */
+    class BlockInnerIterator; // Browse block by block
+
+    friend std::ostream & operator << (std::ostream & s, const BlockSparseMatrix& m)
+    {
+      for (StorageIndex j = 0; j < m.outerBlocks(); ++j)
+      {
+        BlockInnerIterator itb(m, j);
+        for(; itb; ++itb)
+        {
+          s << "("<<itb.row() << ", " << itb.col() << ")\n";
+          s << itb.value() <<"\n";
+        }
+      }
+      s << std::endl;
+      return s;
+    }
+
+    /**
+      * \returns the starting position of the block \p id in the array of values
+      */
+    Index blockPtr(Index id) const
+    {
+      if(m_blockSize == Dynamic) return m_blockPtr[id];
+      else return id * m_blockSize * m_blockSize;
+      //return blockDynIdx(id, typename internal::conditional<(BlockSize==Dynamic), internal::true_type, internal::false_type>::type());
+    }
+
+
+  protected:
+//    inline Index blockDynIdx(Index id, internal::true_type) const
+//    {
+//      return m_blockPtr[id];
+//    }
+//    inline Index blockDynIdx(Index id, internal::false_type) const
+//    {
+//      return id * BlockSize * BlockSize;
+//    }
+
+    // To be implemented
+    // Insert a block at a particular location... need to make a room for that
+    Map<BlockScalar> insert(Index brow, Index bcol);
+
+    Index m_innerBSize; // Number of block rows
+    Index m_outerBSize; // Number of block columns
+    StorageIndex *m_innerOffset; // Starting index of each inner block (size m_innerBSize+1)
+    StorageIndex *m_outerOffset; // Starting index of each outer block (size m_outerBSize+1)
+    Index m_nonzerosblocks; // Total nonzeros blocks (lower than  m_innerBSize x m_outerBSize)
+    Index m_nonzeros; // Total nonzeros elements
+    Scalar *m_values; //Values stored block column after block column (size m_nonzeros)
+    StorageIndex *m_blockPtr; // Pointer to the beginning of each block in m_values, size m_nonzeroblocks ... null for fixed-size blocks
+    StorageIndex *m_indices; //Inner block indices, size m_nonzerosblocks ... OK
+    StorageIndex *m_outerIndex; // Starting pointer of each block column in m_indices (size m_outerBSize)... OK
+    Index m_blockSize; // Size of a block for fixed-size blocks, otherwise -1
+};
+
+template<typename _Scalar, int _BlockAtCompileTime, int _Options, typename _StorageIndex>
+class BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, _StorageIndex>::BlockInnerIterator
+{
+  public:
+
+    enum{
+      Flags = _Options
+    };
+
+    BlockInnerIterator(const BlockSparseMatrix& mat, const Index outer)
+    : m_mat(mat),m_outer(outer),
+      m_id(mat.m_outerIndex[outer]),
+      m_end(mat.m_outerIndex[outer+1])
+    {
+    }
+
+    inline BlockInnerIterator& operator++() {m_id++; return *this; }
+
+    inline const Map<const BlockScalar> value() const
+    {
+      return Map<const BlockScalar>(&(m_mat.m_values[m_mat.blockPtr(m_id)]),
+          rows(),cols());
+    }
+    inline Map<BlockScalar> valueRef()
+    {
+      return Map<BlockScalar>(&(m_mat.m_values[m_mat.blockPtr(m_id)]),
+          rows(),cols());
+    }
+    // Block inner index
+    inline Index index() const {return m_mat.m_indices[m_id]; }
+    inline Index outer() const { return m_outer; }
+    // block row index
+    inline Index row() const  {return index(); }
+    // block column index
+    inline Index col() const {return outer(); }
+    // FIXME Number of rows in the current block
+    inline Index rows() const { return (m_mat.m_blockSize==Dynamic) ? (m_mat.m_innerOffset[index()+1] - m_mat.m_innerOffset[index()]) : m_mat.m_blockSize; }
+    // Number of columns in the current block ...
+    inline Index cols() const { return (m_mat.m_blockSize==Dynamic) ? (m_mat.m_outerOffset[m_outer+1]-m_mat.m_outerOffset[m_outer]) : m_mat.m_blockSize;}
+    inline operator bool() const { return (m_id < m_end); }
+
+  protected:
+    const BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, StorageIndex>& m_mat;
+    const Index m_outer;
+    Index m_id;
+    Index m_end;
+};
+
+template<typename _Scalar, int _BlockAtCompileTime, int _Options, typename _StorageIndex>
+class BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, _StorageIndex>::InnerIterator
+{
+  public:
+    InnerIterator(const BlockSparseMatrix& mat, Index outer)
+    : m_mat(mat),m_outerB(mat.outerToBlock(outer)),m_outer(outer),
+      itb(mat, mat.outerToBlock(outer)),
+      m_offset(outer - mat.blockOuterIndex(m_outerB))
+     {
+        if (itb)
+        {
+          m_id = m_mat.blockInnerIndex(itb.index());
+          m_start = m_id;
+          m_end = m_mat.blockInnerIndex(itb.index()+1);
+        }
+     }
+    inline InnerIterator& operator++()
+    {
+      m_id++;
+      if (m_id >= m_end)
+      {
+        ++itb;
+        if (itb)
+        {
+          m_id = m_mat.blockInnerIndex(itb.index());
+          m_start = m_id;
+          m_end = m_mat.blockInnerIndex(itb.index()+1);
+        }
+      }
+      return *this;
+    }
+    inline const Scalar& value() const
+    {
+      return itb.value().coeff(m_id - m_start, m_offset);
+    }
+    inline Scalar& valueRef()
+    {
+      return itb.valueRef().coeff(m_id - m_start, m_offset);
+    }
+    inline Index index() const { return m_id; }
+    inline Index outer() const {return m_outer; }
+    inline Index col() const {return outer(); }
+    inline Index row() const { return index();}
+    inline operator bool() const
+    {
+      return itb;
+    }
+  protected:
+    const BlockSparseMatrix& m_mat;
+    const Index m_outer;
+    const Index m_outerB;
+    BlockInnerIterator itb; // Iterator through the blocks
+    const Index m_offset; // Position of this column in the block
+    Index m_start; // starting inner index of this block
+    Index m_id; // current inner index in the block
+    Index m_end; // starting inner index of the next block
+
+};
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEBLOCKMATRIX_H
diff --git a/src/EigenUnsupported/src/SparseExtra/DynamicSparseMatrix.h b/src/EigenUnsupported/src/SparseExtra/DynamicSparseMatrix.h
new file mode 100644
index 0000000..42c99e4
--- /dev/null
+++ b/src/EigenUnsupported/src/SparseExtra/DynamicSparseMatrix.h
@@ -0,0 +1,404 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DYNAMIC_SPARSEMATRIX_H
+#define EIGEN_DYNAMIC_SPARSEMATRIX_H
+
+namespace Eigen { 
+
+/** \deprecated use a SparseMatrix in an uncompressed mode
+  *
+  * \class DynamicSparseMatrix
+  *
+  * \brief A sparse matrix class designed for matrix assembly purpose
+  *
+  * \param _Scalar the scalar type, i.e. the type of the coefficients
+  *
+  * Unlike SparseMatrix, this class provides a much higher degree of flexibility. In particular, it allows
+  * random read/write accesses in log(rho*outer_size) where \c rho is the probability that a coefficient is
+  * nonzero and outer_size is the number of columns if the matrix is column-major and the number of rows
+  * otherwise.
+  *
+  * Internally, the data are stored as a std::vector of compressed vector. The performances of random writes might
+  * decrease as the number of nonzeros per inner-vector increase. In practice, we observed very good performance
+  * till about 100 nonzeros/vector, and the performance remains relatively good till 500 nonzeros/vectors.
+  *
+  * \see SparseMatrix
+  */
+
+namespace internal {
+template<typename _Scalar, int _Options, typename _StorageIndex>
+struct traits<DynamicSparseMatrix<_Scalar, _Options, _StorageIndex> >
+{
+  typedef _Scalar Scalar;
+  typedef _StorageIndex StorageIndex;
+  typedef Sparse StorageKind;
+  typedef MatrixXpr XprKind;
+  enum {
+    RowsAtCompileTime = Dynamic,
+    ColsAtCompileTime = Dynamic,
+    MaxRowsAtCompileTime = Dynamic,
+    MaxColsAtCompileTime = Dynamic,
+    Flags = _Options | NestByRefBit | LvalueBit,
+    CoeffReadCost = NumTraits<Scalar>::ReadCost,
+    SupportedAccessPatterns = OuterRandomAccessPattern
+  };
+};
+}
+
+template<typename _Scalar, int _Options, typename _StorageIndex>
+ class  DynamicSparseMatrix
+  : public SparseMatrixBase<DynamicSparseMatrix<_Scalar, _Options, _StorageIndex> >
+{
+    typedef SparseMatrixBase<DynamicSparseMatrix> Base;
+    using Base::convert_index;
+  public:
+    EIGEN_SPARSE_PUBLIC_INTERFACE(DynamicSparseMatrix)
+    // FIXME: why are these operator already alvailable ???
+    // EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(DynamicSparseMatrix, +=)
+    // EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(DynamicSparseMatrix, -=)
+    typedef MappedSparseMatrix<Scalar,Flags> Map;
+    using Base::IsRowMajor;
+    using Base::operator=;
+    enum {
+      Options = _Options
+    };
+
+  protected:
+
+    typedef DynamicSparseMatrix<Scalar,(Flags&~RowMajorBit)|(IsRowMajor?RowMajorBit:0), StorageIndex> TransposedSparseMatrix;
+
+    Index m_innerSize;
+    std::vector<internal::CompressedStorage<Scalar,StorageIndex> > m_data;
+
+  public:
+
+    inline Index rows() const { return IsRowMajor ? outerSize() : m_innerSize; }
+    inline Index cols() const { return IsRowMajor ? m_innerSize : outerSize(); }
+    inline Index innerSize() const { return m_innerSize; }
+    inline Index outerSize() const { return convert_index(m_data.size()); }
+    inline Index innerNonZeros(Index j) const { return m_data[j].size(); }
+
+    std::vector<internal::CompressedStorage<Scalar,StorageIndex> >& _data() { return m_data; }
+    const std::vector<internal::CompressedStorage<Scalar,StorageIndex> >& _data() const { return m_data; }
+
+    /** \returns the coefficient value at given position \a row, \a col
+      * This operation involes a log(rho*outer_size) binary search.
+      */
+    inline Scalar coeff(Index row, Index col) const
+    {
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+      return m_data[outer].at(inner);
+    }
+
+    /** \returns a reference to the coefficient value at given position \a row, \a col
+      * This operation involes a log(rho*outer_size) binary search. If the coefficient does not
+      * exist yet, then a sorted insertion into a sequential buffer is performed.
+      */
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+      return m_data[outer].atWithInsertion(inner);
+    }
+
+    class InnerIterator;
+    class ReverseInnerIterator;
+
+    void setZero()
+    {
+      for (Index j=0; j<outerSize(); ++j)
+        m_data[j].clear();
+    }
+
+    /** \returns the number of non zero coefficients */
+    Index nonZeros() const
+    {
+      Index res = 0;
+      for (Index j=0; j<outerSize(); ++j)
+        res += m_data[j].size();
+      return res;
+    }
+
+
+
+    void reserve(Index reserveSize = 1000)
+    {
+      if (outerSize()>0)
+      {
+        Index reserveSizePerVector = (std::max)(reserveSize/outerSize(),Index(4));
+        for (Index j=0; j<outerSize(); ++j)
+        {
+          m_data[j].reserve(reserveSizePerVector);
+        }
+      }
+    }
+
+    /** Does nothing: provided for compatibility with SparseMatrix */
+    inline void startVec(Index /*outer*/) {}
+
+    /** \returns a reference to the non zero coefficient at position \a row, \a col assuming that:
+      * - the nonzero does not already exist
+      * - the new coefficient is the last one of the given inner vector.
+      *
+      * \sa insert, insertBackByOuterInner */
+    inline Scalar& insertBack(Index row, Index col)
+    {
+      return insertBackByOuterInner(IsRowMajor?row:col, IsRowMajor?col:row);
+    }
+
+    /** \sa insertBack */
+    inline Scalar& insertBackByOuterInner(Index outer, Index inner)
+    {
+      eigen_assert(outer<Index(m_data.size()) && inner<m_innerSize && "out of range");
+      eigen_assert(((m_data[outer].size()==0) || (m_data[outer].index(m_data[outer].size()-1)<inner))
+                && "wrong sorted insertion");
+      m_data[outer].append(0, inner);
+      return m_data[outer].value(m_data[outer].size()-1);
+    }
+
+    inline Scalar& insert(Index row, Index col)
+    {
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+
+      Index startId = 0;
+      Index id = static_cast<Index>(m_data[outer].size()) - 1;
+      m_data[outer].resize(id+2,1);
+
+      while ( (id >= startId) && (m_data[outer].index(id) > inner) )
+      {
+        m_data[outer].index(id+1) = m_data[outer].index(id);
+        m_data[outer].value(id+1) = m_data[outer].value(id);
+        --id;
+      }
+      m_data[outer].index(id+1) = inner;
+      m_data[outer].value(id+1) = 0;
+      return m_data[outer].value(id+1);
+    }
+
+    /** Does nothing: provided for compatibility with SparseMatrix */
+    inline void finalize() {}
+
+    /** Suppress all nonzeros which are smaller than \a reference under the tolerance \a epsilon */
+    void prune(Scalar reference, RealScalar epsilon = NumTraits<RealScalar>::dummy_precision())
+    {
+      for (Index j=0; j<outerSize(); ++j)
+        m_data[j].prune(reference,epsilon);
+    }
+
+    /** Resize the matrix without preserving the data (the matrix is set to zero)
+      */
+    void resize(Index rows, Index cols)
+    {
+      const Index outerSize = IsRowMajor ? rows : cols;
+      m_innerSize = convert_index(IsRowMajor ? cols : rows);
+      setZero();
+      if (Index(m_data.size()) != outerSize)
+      {
+        m_data.resize(outerSize);
+      }
+    }
+
+    void resizeAndKeepData(Index rows, Index cols)
+    {
+      const Index outerSize = IsRowMajor ? rows : cols;
+      const Index innerSize = IsRowMajor ? cols : rows;
+      if (m_innerSize>innerSize)
+      {
+        // remove all coefficients with innerCoord>=innerSize
+        // TODO
+        //std::cerr << "not implemented yet\n";
+        exit(2);
+      }
+      if (m_data.size() != outerSize)
+      {
+        m_data.resize(outerSize);
+      }
+    }
+
+    /** The class DynamicSparseMatrix is deprecated */
+    EIGEN_DEPRECATED inline DynamicSparseMatrix()
+      : m_innerSize(0), m_data(0)
+    {
+      #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+        EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+      #endif
+      eigen_assert(innerSize()==0 && outerSize()==0);
+    }
+
+    /** The class DynamicSparseMatrix is deprecated */
+    EIGEN_DEPRECATED inline DynamicSparseMatrix(Index rows, Index cols)
+      : m_innerSize(0)
+    {
+      #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+        EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+      #endif
+      resize(rows, cols);
+    }
+
+    /** The class DynamicSparseMatrix is deprecated */
+    template<typename OtherDerived>
+    EIGEN_DEPRECATED explicit inline DynamicSparseMatrix(const SparseMatrixBase<OtherDerived>& other)
+      : m_innerSize(0)
+    {
+      #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+        EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+      #endif
+      Base::operator=(other.derived());
+    }
+
+    inline DynamicSparseMatrix(const DynamicSparseMatrix& other)
+      : Base(), m_innerSize(0)
+    {
+      #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+        EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+      #endif
+      *this = other.derived();
+    }
+
+    inline void swap(DynamicSparseMatrix& other)
+    {
+      //EIGEN_DBG_SPARSE(std::cout << "SparseMatrix:: swap\n");
+      std::swap(m_innerSize, other.m_innerSize);
+      //std::swap(m_outerSize, other.m_outerSize);
+      m_data.swap(other.m_data);
+    }
+
+    inline DynamicSparseMatrix& operator=(const DynamicSparseMatrix& other)
+    {
+      if (other.isRValue())
+      {
+        swap(other.const_cast_derived());
+      }
+      else
+      {
+        resize(other.rows(), other.cols());
+        m_data = other.m_data;
+      }
+      return *this;
+    }
+
+    /** Destructor */
+    inline ~DynamicSparseMatrix() {}
+
+  public:
+
+    /** \deprecated
+      * Set the matrix to zero and reserve the memory for \a reserveSize nonzero coefficients. */
+    EIGEN_DEPRECATED void startFill(Index reserveSize = 1000)
+    {
+      setZero();
+      reserve(reserveSize);
+    }
+
+    /** \deprecated use insert()
+      * inserts a nonzero coefficient at given coordinates \a row, \a col and returns its reference assuming that:
+      *  1 - the coefficient does not exist yet
+      *  2 - this the coefficient with greater inner coordinate for the given outer coordinate.
+      * In other words, assuming \c *this is column-major, then there must not exists any nonzero coefficient of coordinates
+      * \c i \c x \a col such that \c i >= \a row. Otherwise the matrix is invalid.
+      *
+      * \see fillrand(), coeffRef()
+      */
+    EIGEN_DEPRECATED Scalar& fill(Index row, Index col)
+    {
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+      return insertBack(outer,inner);
+    }
+
+    /** \deprecated use insert()
+      * Like fill() but with random inner coordinates.
+      * Compared to the generic coeffRef(), the unique limitation is that we assume
+      * the coefficient does not exist yet.
+      */
+    EIGEN_DEPRECATED Scalar& fillrand(Index row, Index col)
+    {
+      return insert(row,col);
+    }
+
+    /** \deprecated use finalize()
+      * Does nothing. Provided for compatibility with SparseMatrix. */
+    EIGEN_DEPRECATED void endFill() {}
+    
+#   ifdef EIGEN_DYNAMICSPARSEMATRIX_PLUGIN
+#     include EIGEN_DYNAMICSPARSEMATRIX_PLUGIN
+#   endif
+ };
+
+template<typename Scalar, int _Options, typename _StorageIndex>
+class DynamicSparseMatrix<Scalar,_Options,_StorageIndex>::InnerIterator : public SparseVector<Scalar,_Options,_StorageIndex>::InnerIterator
+{
+    typedef typename SparseVector<Scalar,_Options,_StorageIndex>::InnerIterator Base;
+  public:
+    InnerIterator(const DynamicSparseMatrix& mat, Index outer)
+      : Base(mat.m_data[outer]), m_outer(outer)
+    {}
+
+    inline Index row() const { return IsRowMajor ? m_outer : Base::index(); }
+    inline Index col() const { return IsRowMajor ? Base::index() : m_outer; }
+    inline Index outer() const { return m_outer; }
+
+  protected:
+    const Index m_outer;
+};
+
+template<typename Scalar, int _Options, typename _StorageIndex>
+class DynamicSparseMatrix<Scalar,_Options,_StorageIndex>::ReverseInnerIterator : public SparseVector<Scalar,_Options,_StorageIndex>::ReverseInnerIterator
+{
+    typedef typename SparseVector<Scalar,_Options,_StorageIndex>::ReverseInnerIterator Base;
+  public:
+    ReverseInnerIterator(const DynamicSparseMatrix& mat, Index outer)
+      : Base(mat.m_data[outer]), m_outer(outer)
+    {}
+
+    inline Index row() const { return IsRowMajor ? m_outer : Base::index(); }
+    inline Index col() const { return IsRowMajor ? Base::index() : m_outer; }
+    inline Index outer() const { return m_outer; }
+
+  protected:
+    const Index m_outer;
+};
+
+namespace internal {
+
+template<typename _Scalar, int _Options, typename _StorageIndex>
+struct evaluator<DynamicSparseMatrix<_Scalar,_Options,_StorageIndex> >
+  : evaluator_base<DynamicSparseMatrix<_Scalar,_Options,_StorageIndex> >
+{
+  typedef _Scalar Scalar;
+  typedef DynamicSparseMatrix<_Scalar,_Options,_StorageIndex> SparseMatrixType;
+  typedef typename SparseMatrixType::InnerIterator InnerIterator;
+  typedef typename SparseMatrixType::ReverseInnerIterator ReverseInnerIterator;
+  
+  enum {
+    CoeffReadCost = NumTraits<_Scalar>::ReadCost,
+    Flags = SparseMatrixType::Flags
+  };
+  
+  evaluator() : m_matrix(0) {}
+  evaluator(const SparseMatrixType &mat) : m_matrix(&mat) {}
+  
+  operator SparseMatrixType&() { return m_matrix->const_cast_derived(); }
+  operator const SparseMatrixType&() const { return *m_matrix; }
+  
+  Scalar coeff(Index row, Index col) const { return m_matrix->coeff(row,col); }
+  
+  Index nonZerosEstimate() const { return m_matrix->nonZeros(); }
+
+  const SparseMatrixType *m_matrix;
+};
+
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_DYNAMIC_SPARSEMATRIX_H
diff --git a/src/EigenUnsupported/src/SparseExtra/MarketIO.h b/src/EigenUnsupported/src/SparseExtra/MarketIO.h
new file mode 100644
index 0000000..dd786d5
--- /dev/null
+++ b/src/EigenUnsupported/src/SparseExtra/MarketIO.h
@@ -0,0 +1,282 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2012 Desire NUENTSA WAKAM <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_MARKET_IO_H
+#define EIGEN_SPARSE_MARKET_IO_H
+
+#include <iostream>
+#include <vector>
+
+namespace Eigen { 
+
+namespace internal 
+{
+  template <typename Scalar, typename StorageIndex>
+  inline void GetMarketLine (const char* line, StorageIndex& i, StorageIndex& j, Scalar& value)
+  {
+    std::stringstream sline(line);
+    sline >> i >> j >> value;
+  }
+
+  template<> inline void GetMarketLine (const char* line, int& i, int& j, float& value)
+  { std::sscanf(line, "%d %d %g", &i, &j, &value); }
+
+  template<> inline void GetMarketLine (const char* line, int& i, int& j, double& value)
+  { std::sscanf(line, "%d %d %lg", &i, &j, &value); }
+
+  template<> inline void GetMarketLine (const char* line, int& i, int& j, std::complex<float>& value)
+  { std::sscanf(line, "%d %d %g %g", &i, &j, &numext::real_ref(value), &numext::imag_ref(value)); }
+
+  template<> inline void GetMarketLine (const char* line, int& i, int& j, std::complex<double>& value)
+  { std::sscanf(line, "%d %d %lg %lg", &i, &j, &numext::real_ref(value), &numext::imag_ref(value)); }
+
+  template <typename Scalar, typename StorageIndex>
+  inline void GetMarketLine (const char* line, StorageIndex& i, StorageIndex& j, std::complex<Scalar>& value)
+  {
+    std::stringstream sline(line);
+    Scalar valR, valI;
+    sline >> i >> j >> valR >> valI;
+    value = std::complex<Scalar>(valR,valI);
+  }
+
+  template <typename RealScalar>
+  inline void  GetVectorElt (const std::string& line, RealScalar& val)
+  {
+    std::istringstream newline(line);
+    newline >> val;  
+  }
+
+  template <typename RealScalar>
+  inline void GetVectorElt (const std::string& line, std::complex<RealScalar>& val)
+  {
+    RealScalar valR, valI; 
+    std::istringstream newline(line);
+    newline >> valR >> valI; 
+    val = std::complex<RealScalar>(valR, valI);
+  }
+  
+  template<typename Scalar>
+  inline void putMarketHeader(std::string& header,int sym)
+  {
+    header= "%%MatrixMarket matrix coordinate ";
+    if(internal::is_same<Scalar, std::complex<float> >::value || internal::is_same<Scalar, std::complex<double> >::value)
+    {
+      header += " complex"; 
+      if(sym == Symmetric) header += " symmetric";
+      else if (sym == SelfAdjoint) header += " Hermitian";
+      else header += " general";
+    }
+    else
+    {
+      header += " real"; 
+      if(sym == Symmetric) header += " symmetric";
+      else header += " general";
+    }
+  }
+
+  template<typename Scalar, typename StorageIndex>
+  inline void PutMatrixElt(Scalar value, StorageIndex row, StorageIndex col, std::ofstream& out)
+  {
+    out << row << " "<< col << " " << value << "\n";
+  }
+  template<typename Scalar, typename StorageIndex>
+  inline void PutMatrixElt(std::complex<Scalar> value, StorageIndex row, StorageIndex col, std::ofstream& out)
+  {
+    out << row << " " << col << " " << value.real() << " " << value.imag() << "\n";
+  }
+
+
+  template<typename Scalar>
+  inline void putVectorElt(Scalar value, std::ofstream& out)
+  {
+    out << value << "\n"; 
+  }
+  template<typename Scalar>
+  inline void putVectorElt(std::complex<Scalar> value, std::ofstream& out)
+  {
+    out << value.real() << " " << value.imag()<< "\n"; 
+  }
+
+} // end namespace internal
+
+inline bool getMarketHeader(const std::string& filename, int& sym, bool& iscomplex, bool& isvector)
+{
+  sym = 0; 
+  iscomplex = false;
+  isvector = false;
+  std::ifstream in(filename.c_str(),std::ios::in);
+  if(!in)
+    return false;
+  
+  std::string line; 
+  // The matrix header is always the first line in the file 
+  std::getline(in, line); eigen_assert(in.good());
+  
+  std::stringstream fmtline(line); 
+  std::string substr[5];
+  fmtline>> substr[0] >> substr[1] >> substr[2] >> substr[3] >> substr[4];
+  if(substr[2].compare("array") == 0) isvector = true;
+  if(substr[3].compare("complex") == 0) iscomplex = true;
+  if(substr[4].compare("symmetric") == 0) sym = Symmetric;
+  else if (substr[4].compare("Hermitian") == 0) sym = SelfAdjoint;
+  
+  return true;
+}
+  
+template<typename SparseMatrixType>
+bool loadMarket(SparseMatrixType& mat, const std::string& filename)
+{
+  typedef typename SparseMatrixType::Scalar Scalar;
+  typedef typename SparseMatrixType::StorageIndex StorageIndex;
+  std::ifstream input(filename.c_str(),std::ios::in);
+  if(!input)
+    return false;
+
+  char rdbuffer[4096];
+  input.rdbuf()->pubsetbuf(rdbuffer, 4096);
+  
+  const int maxBuffersize = 2048;
+  char buffer[maxBuffersize];
+  
+  bool readsizes = false;
+
+  typedef Triplet<Scalar,StorageIndex> T;
+  std::vector<T> elements;
+  
+  Index M(-1), N(-1), NNZ(-1);
+  Index count = 0;
+  while(input.getline(buffer, maxBuffersize))
+  {
+    // skip comments   
+    //NOTE An appropriate test should be done on the header to get the  symmetry
+    if(buffer[0]=='%')
+      continue;
+
+    if(!readsizes)
+    {
+      std::stringstream line(buffer);
+      line >> M >> N >> NNZ;
+      if(M > 0 && N > 0)
+      {
+        readsizes = true;
+        mat.resize(M,N);
+        mat.reserve(NNZ);
+      }
+    }
+    else
+    { 
+      StorageIndex i(-1), j(-1);
+      Scalar value; 
+      internal::GetMarketLine(buffer, i, j, value);
+
+      i--;
+      j--;
+      if(i>=0 && j>=0 && i<M && j<N)
+      {
+        ++count;
+        elements.push_back(T(i,j,value));
+      }
+      else
+        std::cerr << "Invalid read: " << i << "," << j << "\n";        
+    }
+  }
+
+  mat.setFromTriplets(elements.begin(), elements.end());
+  if(count!=NNZ)
+    std::cerr << count << "!=" << NNZ << "\n";
+  
+  input.close();
+  return true;
+}
+
+template<typename VectorType>
+bool loadMarketVector(VectorType& vec, const std::string& filename)
+{
+   typedef typename VectorType::Scalar Scalar;
+  std::ifstream in(filename.c_str(), std::ios::in);
+  if(!in)
+    return false;
+  
+  std::string line; 
+  int n(0), col(0); 
+  do 
+  { // Skip comments
+    std::getline(in, line); eigen_assert(in.good());
+  } while (line[0] == '%');
+  std::istringstream newline(line);
+  newline  >> n >> col; 
+  eigen_assert(n>0 && col>0);
+  vec.resize(n);
+  int i = 0; 
+  Scalar value; 
+  while ( std::getline(in, line) && (i < n) ){
+    internal::GetVectorElt(line, value); 
+    vec(i++) = value; 
+  }
+  in.close();
+  if (i!=n){
+    std::cerr<< "Unable to read all elements from file " << filename << "\n";
+    return false;
+  }
+  return true;
+}
+
+template<typename SparseMatrixType>
+bool saveMarket(const SparseMatrixType& mat, const std::string& filename, int sym = 0)
+{
+  typedef typename SparseMatrixType::Scalar Scalar;
+  typedef typename SparseMatrixType::RealScalar RealScalar;
+  std::ofstream out(filename.c_str(),std::ios::out);
+  if(!out)
+    return false;
+  
+  out.flags(std::ios_base::scientific);
+  out.precision(std::numeric_limits<RealScalar>::digits10 + 2);
+  std::string header; 
+  internal::putMarketHeader<Scalar>(header, sym); 
+  out << header << std::endl; 
+  out << mat.rows() << " " << mat.cols() << " " << mat.nonZeros() << "\n";
+  int count = 0;
+  for(int j=0; j<mat.outerSize(); ++j)
+    for(typename SparseMatrixType::InnerIterator it(mat,j); it; ++it)
+    {
+      ++ count;
+      internal::PutMatrixElt(it.value(), it.row()+1, it.col()+1, out);
+    }
+  out.close();
+  return true;
+}
+
+template<typename VectorType>
+bool saveMarketVector (const VectorType& vec, const std::string& filename)
+{
+ typedef typename VectorType::Scalar Scalar;
+ typedef typename VectorType::RealScalar RealScalar;
+ std::ofstream out(filename.c_str(),std::ios::out);
+  if(!out)
+    return false;
+  
+  out.flags(std::ios_base::scientific);
+  out.precision(std::numeric_limits<RealScalar>::digits10 + 2);
+  if(internal::is_same<Scalar, std::complex<float> >::value || internal::is_same<Scalar, std::complex<double> >::value)
+      out << "%%MatrixMarket matrix array complex general\n"; 
+  else
+    out << "%%MatrixMarket matrix array real general\n"; 
+  out << vec.size() << " "<< 1 << "\n";
+  for (int i=0; i < vec.size(); i++){
+    internal::putVectorElt(vec(i), out); 
+  }
+  out.close();
+  return true; 
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_MARKET_IO_H
diff --git a/src/EigenUnsupported/src/SparseExtra/MatrixMarketIterator.h b/src/EigenUnsupported/src/SparseExtra/MatrixMarketIterator.h
new file mode 100644
index 0000000..02916ea
--- /dev/null
+++ b/src/EigenUnsupported/src/SparseExtra/MatrixMarketIterator.h
@@ -0,0 +1,247 @@
+
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Desire NUENTSA WAKAM <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BROWSE_MATRICES_H
+#define EIGEN_BROWSE_MATRICES_H
+
+namespace Eigen {
+
+enum {
+  SPD = 0x100,
+  NonSymmetric = 0x0
+}; 
+
+/** 
+ * @brief Iterator to browse matrices from a specified folder
+ * 
+ * This is used to load all the matrices from a folder. 
+ * The matrices should be in Matrix Market format
+ * It is assumed that the matrices are named as matname.mtx
+ * and matname_SPD.mtx if the matrix is Symmetric and positive definite (or Hermitian)
+ * The right hand side vectors are loaded as well, if they exist.
+ * They should be named as matname_b.mtx. 
+ * Note that the right hand side for a SPD matrix is named as matname_SPD_b.mtx
+ * 
+ * Sometimes a reference solution is available. In this case, it should be named as matname_x.mtx
+ * 
+ * Sample code
+ * \code
+ * 
+ * \endcode
+ * 
+ * \tparam Scalar The scalar type 
+ */
+template <typename Scalar>
+class MatrixMarketIterator 
+{
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+  public:
+    typedef Matrix<Scalar,Dynamic,1> VectorType; 
+    typedef SparseMatrix<Scalar,ColMajor> MatrixType; 
+  
+  public:
+    MatrixMarketIterator(const std::string &folder)
+      : m_sym(0), m_isvalid(false), m_matIsLoaded(false), m_hasRhs(false), m_hasrefX(false), m_folder(folder)
+    {
+      m_folder_id = opendir(folder.c_str());
+      if(m_folder_id)
+        Getnextvalidmatrix();
+    }
+    
+    ~MatrixMarketIterator()
+    {
+      if (m_folder_id) closedir(m_folder_id); 
+    }
+    
+    inline MatrixMarketIterator& operator++()
+    {
+      m_matIsLoaded = false;
+      m_hasrefX = false;
+      m_hasRhs = false;
+      Getnextvalidmatrix();
+      return *this;
+    }
+    inline operator bool() const { return m_isvalid;}
+    
+    /** Return the sparse matrix corresponding to the current file */
+    inline MatrixType& matrix() 
+    { 
+      // Read the matrix
+      if (m_matIsLoaded) return m_mat;
+      
+      std::string matrix_file = m_folder + "/" + m_matname + ".mtx";
+      if ( !loadMarket(m_mat, matrix_file)) 
+      {
+        std::cerr << "Warning loadMarket failed when loading \"" << matrix_file << "\"" << std::endl;
+        m_matIsLoaded = false;
+        return m_mat;
+      }
+      m_matIsLoaded = true; 
+
+      if (m_sym != NonSymmetric) 
+      {
+        // Check whether we need to restore a full matrix:
+        RealScalar diag_norm  = m_mat.diagonal().norm();
+        RealScalar lower_norm = m_mat.template triangularView<Lower>().norm();
+        RealScalar upper_norm = m_mat.template triangularView<Upper>().norm();
+        if(lower_norm>diag_norm && upper_norm==diag_norm)
+        {
+          // only the lower part is stored
+          MatrixType tmp(m_mat);
+          m_mat = tmp.template selfadjointView<Lower>();
+        }
+        else if(upper_norm>diag_norm && lower_norm==diag_norm)
+        {
+          // only the upper part is stored
+          MatrixType tmp(m_mat);
+          m_mat = tmp.template selfadjointView<Upper>();
+        }
+      }
+      return m_mat; 
+    }
+    
+    /** Return the right hand side corresponding to the current matrix. 
+     * If the rhs file is not provided, a random rhs is generated
+     */
+    inline VectorType& rhs() 
+    { 
+       // Get the right hand side
+      if (m_hasRhs) return m_rhs;
+      
+      std::string rhs_file;
+      rhs_file = m_folder + "/" + m_matname + "_b.mtx"; // The pattern is matname_b.mtx
+      m_hasRhs = Fileexists(rhs_file);
+      if (m_hasRhs)
+      {
+        m_rhs.resize(m_mat.cols());
+        m_hasRhs = loadMarketVector(m_rhs, rhs_file);
+      }
+      if (!m_hasRhs)
+      {
+        // Generate a random right hand side
+        if (!m_matIsLoaded) this->matrix(); 
+        m_refX.resize(m_mat.cols());
+        m_refX.setRandom();
+        m_rhs = m_mat * m_refX;
+        m_hasrefX = true;
+        m_hasRhs = true;
+      }
+      return m_rhs; 
+    }
+    
+    /** Return a reference solution
+     * If it is not provided and if the right hand side is not available
+     * then refX is randomly generated such that A*refX = b 
+     * where A and b are the matrix and the rhs. 
+     * Note that when a rhs is provided, refX is not available 
+     */
+    inline VectorType& refX() 
+    { 
+      // Check if a reference solution is provided
+      if (m_hasrefX) return m_refX;
+      
+      std::string lhs_file;
+      lhs_file = m_folder + "/" + m_matname + "_x.mtx"; 
+      m_hasrefX = Fileexists(lhs_file);
+      if (m_hasrefX)
+      {
+        m_refX.resize(m_mat.cols());
+        m_hasrefX = loadMarketVector(m_refX, lhs_file);
+      }
+      else
+        m_refX.resize(0);
+      return m_refX; 
+    }
+    
+    inline std::string& matname() { return m_matname; }
+    
+    inline int sym() { return m_sym; }
+    
+    bool hasRhs() {return m_hasRhs; }
+    bool hasrefX() {return m_hasrefX; }
+    bool isFolderValid() { return bool(m_folder_id); }
+    
+  protected:
+    
+    inline bool Fileexists(std::string file)
+    {
+      std::ifstream file_id(file.c_str());
+      if (!file_id.good() ) 
+      {
+        return false;
+      }
+      else 
+      {
+        file_id.close();
+        return true;
+      }
+    }
+    
+    void Getnextvalidmatrix( )
+    {
+      m_isvalid = false;
+      // Here, we return with the next valid matrix in the folder
+      while ( (m_curs_id = readdir(m_folder_id)) != NULL) {
+        m_isvalid = false;
+        std::string curfile;
+        curfile = m_folder + "/" + m_curs_id->d_name;
+        // Discard if it is a folder
+        if (m_curs_id->d_type == DT_DIR) continue; //FIXME This may not be available on non BSD systems
+//         struct stat st_buf; 
+//         stat (curfile.c_str(), &st_buf);
+//         if (S_ISDIR(st_buf.st_mode)) continue;
+        
+        // Determine from the header if it is a matrix or a right hand side 
+        bool isvector,iscomplex=false;
+        if(!getMarketHeader(curfile,m_sym,iscomplex,isvector)) continue;
+        if(isvector) continue;
+        if (!iscomplex)
+        {
+          if(internal::is_same<Scalar, std::complex<float> >::value || internal::is_same<Scalar, std::complex<double> >::value)
+            continue; 
+        }
+        if (iscomplex)
+        {
+          if(internal::is_same<Scalar, float>::value || internal::is_same<Scalar, double>::value)
+            continue; 
+        }
+        
+        
+        // Get the matrix name
+        std::string filename = m_curs_id->d_name;
+        m_matname = filename.substr(0, filename.length()-4); 
+        
+        // Find if the matrix is SPD 
+        size_t found = m_matname.find("SPD");
+        if( (found!=std::string::npos) && (m_sym != NonSymmetric) )
+          m_sym = SPD;
+       
+        m_isvalid = true;
+        break; 
+      }
+    }
+    int m_sym; // Symmetry of the matrix
+    MatrixType m_mat; // Current matrix  
+    VectorType m_rhs;  // Current vector
+    VectorType m_refX; // The reference solution, if exists
+    std::string m_matname; // Matrix Name
+    bool m_isvalid; 
+    bool m_matIsLoaded; // Determine if the matrix has already been loaded from the file
+    bool m_hasRhs; // The right hand side exists
+    bool m_hasrefX; // A reference solution is provided
+    std::string m_folder;
+    DIR * m_folder_id;
+    struct dirent *m_curs_id; 
+    
+};
+
+} // end namespace Eigen
+
+#endif
diff --git a/src/EigenUnsupported/src/SparseExtra/RandomSetter.h b/src/EigenUnsupported/src/SparseExtra/RandomSetter.h
new file mode 100644
index 0000000..985702b
--- /dev/null
+++ b/src/EigenUnsupported/src/SparseExtra/RandomSetter.h
@@ -0,0 +1,349 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_RANDOMSETTER_H
+#define EIGEN_RANDOMSETTER_H
+
+#if defined(EIGEN_GOOGLEHASH_SUPPORT)
+// Ensure the ::google namespace exists, required for checking existence of 
+// ::google::dense_hash_map and ::google::sparse_hash_map.
+namespace google {}
+#endif
+
+namespace Eigen {
+
+/** Represents a std::map
+  *
+  * \see RandomSetter
+  */
+template<typename Scalar> struct StdMapTraits
+{
+  typedef int KeyType;
+  typedef std::map<KeyType,Scalar> Type;
+  enum {
+    IsSorted = 1
+  };
+
+  static void setInvalidKey(Type&, const KeyType&) {}
+};
+
+#ifdef EIGEN_UNORDERED_MAP_SUPPORT
+/** Represents a std::unordered_map
+  *
+  * To use it you need to both define EIGEN_UNORDERED_MAP_SUPPORT and include the unordered_map header file
+  * yourself making sure that unordered_map is defined in the std namespace.
+  *
+  * For instance, with current version of gcc you can either enable C++0x standard (-std=c++0x) or do:
+  * \code
+  * #include <tr1/unordered_map>
+  * #define EIGEN_UNORDERED_MAP_SUPPORT
+  * namespace std {
+  *   using std::tr1::unordered_map;
+  * }
+  * \endcode
+  *
+  * \see RandomSetter
+  */
+template<typename Scalar> struct StdUnorderedMapTraits
+{
+  typedef int KeyType;
+  typedef std::unordered_map<KeyType,Scalar> Type;
+  enum {
+    IsSorted = 0
+  };
+
+  static void setInvalidKey(Type&, const KeyType&) {}
+};
+#endif // EIGEN_UNORDERED_MAP_SUPPORT
+
+#if defined(EIGEN_GOOGLEHASH_SUPPORT)
+
+namespace google {
+  
+// Namespace work-around, since sometimes dense_hash_map and sparse_hash_map
+// are in the global namespace, and other times they are under ::google.
+using namespace ::google;
+
+template<typename KeyType, typename Scalar>
+struct DenseHashMap {
+  typedef dense_hash_map<KeyType, Scalar> type;
+};
+
+template<typename KeyType, typename Scalar>
+struct SparseHashMap {
+  typedef sparse_hash_map<KeyType, Scalar> type;
+};
+
+} // namespace google
+
+/** Represents a google::dense_hash_map
+  *
+  * \see RandomSetter
+  */
+template<typename Scalar> struct GoogleDenseHashMapTraits
+{
+  typedef int KeyType;
+  typedef typename google::DenseHashMap<KeyType,Scalar>::type Type;
+  enum {
+    IsSorted = 0
+  };
+
+  static void setInvalidKey(Type& map, const KeyType& k)
+  { map.set_empty_key(k); }
+};
+
+/** Represents a google::sparse_hash_map
+  *
+  * \see RandomSetter
+  */
+template<typename Scalar> struct GoogleSparseHashMapTraits
+{
+  typedef int KeyType;
+  typedef typename google::SparseHashMap<KeyType,Scalar>::type Type;
+  enum {
+    IsSorted = 0
+  };
+
+  static void setInvalidKey(Type&, const KeyType&) {}
+};
+#endif
+
+/** \class RandomSetter
+  *
+  * \brief The RandomSetter is a wrapper object allowing to set/update a sparse matrix with random access
+  *
+  * \tparam SparseMatrixType the type of the sparse matrix we are updating
+  * \tparam MapTraits a traits class representing the map implementation used for the temporary sparse storage.
+  *                  Its default value depends on the system.
+  * \tparam OuterPacketBits defines the number of rows (or columns) manage by a single map object
+  *                        as a power of two exponent.
+  *
+  * This class temporarily represents a sparse matrix object using a generic map implementation allowing for
+  * efficient random access. The conversion from the compressed representation to a hash_map object is performed
+  * in the RandomSetter constructor, while the sparse matrix is updated back at destruction time. This strategy
+  * suggest the use of nested blocks as in this example:
+  *
+  * \code
+  * SparseMatrix<double> m(rows,cols);
+  * {
+  *   RandomSetter<SparseMatrix<double> > w(m);
+  *   // don't use m but w instead with read/write random access to the coefficients:
+  *   for(;;)
+  *     w(rand(),rand()) = rand;
+  * }
+  * // when w is deleted, the data are copied back to m
+  * // and m is ready to use.
+  * \endcode
+  *
+  * Since hash_map objects are not fully sorted, representing a full matrix as a single hash_map would
+  * involve a big and costly sort to update the compressed matrix back. To overcome this issue, a RandomSetter
+  * use multiple hash_map, each representing 2^OuterPacketBits columns or rows according to the storage order.
+  * To reach optimal performance, this value should be adjusted according to the average number of nonzeros
+  * per rows/columns.
+  *
+  * The possible values for the template parameter MapTraits are:
+  *  - \b StdMapTraits: corresponds to std::map. (does not perform very well)
+  *  - \b GnuHashMapTraits: corresponds to __gnu_cxx::hash_map (available only with GCC)
+  *  - \b GoogleDenseHashMapTraits: corresponds to google::dense_hash_map (best efficiency, reasonable memory consumption)
+  *  - \b GoogleSparseHashMapTraits: corresponds to google::sparse_hash_map (best memory consumption, relatively good performance)
+  *
+  * The default map implementation depends on the availability, and the preferred order is:
+  * GoogleSparseHashMapTraits, GnuHashMapTraits, and finally StdMapTraits.
+  *
+  * For performance and memory consumption reasons it is highly recommended to use one of
+  * Google's hash_map implementations. To enable the support for them, you must define
+  * EIGEN_GOOGLEHASH_SUPPORT. This will include both <google/dense_hash_map> and
+  * <google/sparse_hash_map> for you.
+  *
+  * \see https://github.com/sparsehash/sparsehash
+  */
+template<typename SparseMatrixType,
+         template <typename T> class MapTraits =
+#if defined(EIGEN_GOOGLEHASH_SUPPORT)
+          GoogleDenseHashMapTraits
+#elif defined(_HASH_MAP)
+          GnuHashMapTraits
+#else
+          StdMapTraits
+#endif
+         ,int OuterPacketBits = 6>
+class RandomSetter
+{
+    typedef typename SparseMatrixType::Scalar Scalar;
+    typedef typename SparseMatrixType::StorageIndex StorageIndex;
+
+    struct ScalarWrapper
+    {
+      ScalarWrapper() : value(0) {}
+      Scalar value;
+    };
+    typedef typename MapTraits<ScalarWrapper>::KeyType KeyType;
+    typedef typename MapTraits<ScalarWrapper>::Type HashMapType;
+    static const int OuterPacketMask = (1 << OuterPacketBits) - 1;
+    enum {
+      SwapStorage = 1 - MapTraits<ScalarWrapper>::IsSorted,
+      TargetRowMajor = (SparseMatrixType::Flags & RowMajorBit) ? 1 : 0,
+      SetterRowMajor = SwapStorage ? 1-TargetRowMajor : TargetRowMajor
+    };
+
+  public:
+
+    /** Constructs a random setter object from the sparse matrix \a target
+      *
+      * Note that the initial value of \a target are imported. If you want to re-set
+      * a sparse matrix from scratch, then you must set it to zero first using the
+      * setZero() function.
+      */
+    inline RandomSetter(SparseMatrixType& target)
+      : mp_target(&target)
+    {
+      const Index outerSize = SwapStorage ? target.innerSize() : target.outerSize();
+      const Index innerSize = SwapStorage ? target.outerSize() : target.innerSize();
+      m_outerPackets = outerSize >> OuterPacketBits;
+      if (outerSize&OuterPacketMask)
+        m_outerPackets += 1;
+      m_hashmaps = new HashMapType[m_outerPackets];
+      // compute number of bits needed to store inner indices
+      Index aux = innerSize - 1;
+      m_keyBitsOffset = 0;
+      while (aux)
+      {
+        ++m_keyBitsOffset;
+        aux = aux >> 1;
+      }
+      KeyType ik = (1<<(OuterPacketBits+m_keyBitsOffset));
+      for (Index k=0; k<m_outerPackets; ++k)
+        MapTraits<ScalarWrapper>::setInvalidKey(m_hashmaps[k],ik);
+
+      // insert current coeffs
+      for (Index j=0; j<mp_target->outerSize(); ++j)
+        for (typename SparseMatrixType::InnerIterator it(*mp_target,j); it; ++it)
+          (*this)(TargetRowMajor?j:it.index(), TargetRowMajor?it.index():j) = it.value();
+    }
+
+    /** Destructor updating back the sparse matrix target */
+    ~RandomSetter()
+    {
+      KeyType keyBitsMask = (1<<m_keyBitsOffset)-1;
+      if (!SwapStorage) // also means the map is sorted
+      {
+        mp_target->setZero();
+        mp_target->makeCompressed();
+        mp_target->reserve(nonZeros());
+        Index prevOuter = -1;
+        for (Index k=0; k<m_outerPackets; ++k)
+        {
+          const Index outerOffset = (1<<OuterPacketBits) * k;
+          typename HashMapType::iterator end = m_hashmaps[k].end();
+          for (typename HashMapType::iterator it = m_hashmaps[k].begin(); it!=end; ++it)
+          {
+            const Index outer = (it->first >> m_keyBitsOffset) + outerOffset;
+            const Index inner = it->first & keyBitsMask;
+            if (prevOuter!=outer)
+            {
+              for (Index j=prevOuter+1;j<=outer;++j)
+                mp_target->startVec(j);
+              prevOuter = outer;
+            }
+            mp_target->insertBackByOuterInner(outer, inner) = it->second.value;
+          }
+        }
+        mp_target->finalize();
+      }
+      else
+      {
+        VectorXi positions(mp_target->outerSize());
+        positions.setZero();
+        // pass 1
+        for (Index k=0; k<m_outerPackets; ++k)
+        {
+          typename HashMapType::iterator end = m_hashmaps[k].end();
+          for (typename HashMapType::iterator it = m_hashmaps[k].begin(); it!=end; ++it)
+          {
+            const Index outer = it->first & keyBitsMask;
+            ++positions[outer];
+          }
+        }
+        // prefix sum
+        StorageIndex count = 0;
+        for (Index j=0; j<mp_target->outerSize(); ++j)
+        {
+          StorageIndex tmp = positions[j];
+          mp_target->outerIndexPtr()[j] = count;
+          positions[j] = count;
+          count += tmp;
+        }
+        mp_target->makeCompressed();
+        mp_target->outerIndexPtr()[mp_target->outerSize()] = count;
+        mp_target->resizeNonZeros(count);
+        // pass 2
+        for (Index k=0; k<m_outerPackets; ++k)
+        {
+          const Index outerOffset = (1<<OuterPacketBits) * k;
+          typename HashMapType::iterator end = m_hashmaps[k].end();
+          for (typename HashMapType::iterator it = m_hashmaps[k].begin(); it!=end; ++it)
+          {
+            const Index inner = (it->first >> m_keyBitsOffset) + outerOffset;
+            const Index outer = it->first & keyBitsMask;
+            // sorted insertion
+            // Note that we have to deal with at most 2^OuterPacketBits unsorted coefficients,
+            // moreover those 2^OuterPacketBits coeffs are likely to be sparse, an so only a
+            // small fraction of them have to be sorted, whence the following simple procedure:
+            Index posStart = mp_target->outerIndexPtr()[outer];
+            Index i = (positions[outer]++) - 1;
+            while ( (i >= posStart) && (mp_target->innerIndexPtr()[i] > inner) )
+            {
+              mp_target->valuePtr()[i+1] = mp_target->valuePtr()[i];
+              mp_target->innerIndexPtr()[i+1] = mp_target->innerIndexPtr()[i];
+              --i;
+            }
+            mp_target->innerIndexPtr()[i+1] = internal::convert_index<StorageIndex>(inner);
+            mp_target->valuePtr()[i+1] = it->second.value;
+          }
+        }
+      }
+      delete[] m_hashmaps;
+    }
+
+    /** \returns a reference to the coefficient at given coordinates \a row, \a col */
+    Scalar& operator() (Index row, Index col)
+    {
+      const Index outer = SetterRowMajor ? row : col;
+      const Index inner = SetterRowMajor ? col : row;
+      const Index outerMajor = outer >> OuterPacketBits; // index of the packet/map
+      const Index outerMinor = outer & OuterPacketMask;  // index of the inner vector in the packet
+      const KeyType key = internal::convert_index<KeyType>((outerMinor<<m_keyBitsOffset) | inner);
+      return m_hashmaps[outerMajor][key].value;
+    }
+
+    /** \returns the number of non zero coefficients
+      *
+      * \note According to the underlying map/hash_map implementation,
+      * this function might be quite expensive.
+      */
+    Index nonZeros() const
+    {
+      Index nz = 0;
+      for (Index k=0; k<m_outerPackets; ++k)
+        nz += static_cast<Index>(m_hashmaps[k].size());
+      return nz;
+    }
+
+
+  protected:
+
+    HashMapType* m_hashmaps;
+    SparseMatrixType* mp_target;
+    Index m_outerPackets;
+    unsigned char m_keyBitsOffset;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_RANDOMSETTER_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsArrayAPI.h b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsArrayAPI.h
new file mode 100644
index 0000000..41d2bf6
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsArrayAPI.h
@@ -0,0 +1,286 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_BESSELFUNCTIONS_ARRAYAPI_H
+#define EIGEN_BESSELFUNCTIONS_ARRAYAPI_H
+
+namespace Eigen {
+
+/** \returns an expression of the coefficient-wise i0(\a x) to the given
+ * arrays.
+  *
+  * It returns the modified Bessel function of the first kind of order zero.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of i0(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_i0()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_i0_op<typename Derived::Scalar>, const Derived>
+bessel_i0(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_i0_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise i0e(\a x) to the given
+ * arrays.
+  *
+  * It returns the exponentially scaled modified Bessel
+  * function of the first kind of order zero.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of i0e(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_i0e()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_i0e_op<typename Derived::Scalar>, const Derived>
+bessel_i0e(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_i0e_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise i1(\a x) to the given
+ * arrays.
+  *
+  * It returns the modified Bessel function of the first kind of order one.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of i1(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_i1()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_i1_op<typename Derived::Scalar>, const Derived>
+bessel_i1(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_i1_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise i1e(\a x) to the given
+ * arrays.
+  *
+  * It returns the exponentially scaled modified Bessel
+  * function of the first kind of order one.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of i1e(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_i1e()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_i1e_op<typename Derived::Scalar>, const Derived>
+bessel_i1e(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_i1e_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise k0(\a x) to the given
+ * arrays.
+  *
+  * It returns the modified Bessel function of the second kind of order zero.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of k0(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_k0()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_k0_op<typename Derived::Scalar>, const Derived>
+bessel_k0(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_k0_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise k0e(\a x) to the given
+ * arrays.
+  *
+  * It returns the exponentially scaled modified Bessel
+  * function of the second kind of order zero.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of k0e(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_k0e()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_k0e_op<typename Derived::Scalar>, const Derived>
+bessel_k0e(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_k0e_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise k1(\a x) to the given
+ * arrays.
+  *
+  * It returns the modified Bessel function of the second kind of order one.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of k1(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_k1()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_k1_op<typename Derived::Scalar>, const Derived>
+bessel_k1(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_k1_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise k1e(\a x) to the given
+ * arrays.
+  *
+  * It returns the exponentially scaled modified Bessel
+  * function of the second kind of order one.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of k1e(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_k1e()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_k1e_op<typename Derived::Scalar>, const Derived>
+bessel_k1e(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_k1e_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise j0(\a x) to the given
+ * arrays.
+  *
+  * It returns the Bessel function of the first kind of order zero.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of j0(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_j0()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_j0_op<typename Derived::Scalar>, const Derived>
+bessel_j0(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_j0_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise y0(\a x) to the given
+ * arrays.
+  *
+  * It returns the Bessel function of the second kind of order zero.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of y0(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_y0()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_y0_op<typename Derived::Scalar>, const Derived>
+bessel_y0(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_y0_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise j1(\a x) to the given
+ * arrays.
+  *
+  * It returns the modified Bessel function of the first kind of order one.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of j1(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_j1()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_j1_op<typename Derived::Scalar>, const Derived>
+bessel_j1(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_j1_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+/** \returns an expression of the coefficient-wise y1(\a x) to the given
+ * arrays.
+  *
+  * It returns the Bessel function of the second kind of order one.
+  *
+  * \param x is the argument
+  *
+  * \note This function supports only float and double scalar types. To support
+  * other scalar types, the user has to provide implementations of y1(T) for
+  * any scalar type T to be supported.
+  *
+  * \sa ArrayBase::bessel_y1()
+  */
+template <typename Derived>
+EIGEN_STRONG_INLINE const Eigen::CwiseUnaryOp<
+    Eigen::internal::scalar_bessel_y1_op<typename Derived::Scalar>, const Derived>
+bessel_y1(const Eigen::ArrayBase<Derived>& x) {
+  return Eigen::CwiseUnaryOp<
+      Eigen::internal::scalar_bessel_y1_op<typename Derived::Scalar>,
+      const Derived>(x.derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_BESSELFUNCTIONS_ARRAYAPI_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsBFloat16.h b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsBFloat16.h
new file mode 100644
index 0000000..6049cc2
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsBFloat16.h
@@ -0,0 +1,68 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BESSELFUNCTIONS_BFLOAT16_H
+#define EIGEN_BESSELFUNCTIONS_BFLOAT16_H
+
+namespace Eigen {
+namespace numext {
+
+#if EIGEN_HAS_C99_MATH
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_i0(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_i0(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_i0e(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_i0e(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_i1(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_i1(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_i1e(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_i1e(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_j0(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_j0(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_j1(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_j1(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_y0(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_y0(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_y1(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_y1(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_k0(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_k0(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_k0e(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_k0e(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_k1(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_k1(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bessel_k1e(const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::bessel_k1e(static_cast<float>(x)));
+}
+#endif
+
+}  // end namespace numext
+}  // end namespace Eigen
+
+#endif  // EIGEN_BESSELFUNCTIONS_BFLOAT16_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsFunctors.h b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsFunctors.h
new file mode 100644
index 0000000..8606a9f
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsFunctors.h
@@ -0,0 +1,357 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com>
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BESSELFUNCTIONS_FUNCTORS_H
+#define EIGEN_BESSELFUNCTIONS_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal
+ * \brief Template functor to compute the modified Bessel function of the first
+ * kind of order zero.
+ * \sa class CwiseUnaryOp, Cwise::bessel_i0()
+ */
+template <typename Scalar>
+struct scalar_bessel_i0_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_i0_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_i0;
+    return bessel_i0(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_i0(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_i0_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=20 is computed.
+    // The cost is N multiplications and 2N additions. We also add
+    // the cost of an additional exp over i0e.
+    Cost = 28 * NumTraits<Scalar>::MulCost + 48 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the exponentially scaled modified Bessel
+ * function of the first kind of order zero
+ * \sa class CwiseUnaryOp, Cwise::bessel_i0e()
+ */
+template <typename Scalar>
+struct scalar_bessel_i0e_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_i0e_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_i0e;
+    return bessel_i0e(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_i0e(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_i0e_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=20 is computed.
+    // The cost is N multiplications and 2N additions.
+    Cost = 20 * NumTraits<Scalar>::MulCost + 40 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the modified Bessel function of the first
+ * kind of order one
+ * \sa class CwiseUnaryOp, Cwise::bessel_i1()
+ */
+template <typename Scalar>
+struct scalar_bessel_i1_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_i1_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_i1;
+    return bessel_i1(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_i1(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_i1_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=20 is computed.
+    // The cost is N multiplications and 2N additions. We also add
+    // the cost of an additional exp over i1e.
+    Cost = 28 * NumTraits<Scalar>::MulCost + 48 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the exponentially scaled modified Bessel
+ * function of the first kind of order zero
+ * \sa class CwiseUnaryOp, Cwise::bessel_i1e()
+ */
+template <typename Scalar>
+struct scalar_bessel_i1e_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_i1e_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_i1e;
+    return bessel_i1e(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_i1e(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_i1e_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=20 is computed.
+    // The cost is N multiplications and 2N additions.
+    Cost = 20 * NumTraits<Scalar>::MulCost + 40 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Bessel function of the second kind of
+ * order zero
+ * \sa class CwiseUnaryOp, Cwise::bessel_j0()
+ */
+template <typename Scalar>
+struct scalar_bessel_j0_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_j0_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_j0;
+    return bessel_j0(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_j0(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_j0_op<Scalar> > {
+  enum {
+    // 6 polynomial of order ~N=8 is computed.
+    // The cost is N multiplications and N additions each, along with a
+    // sine, cosine and rsqrt cost.
+    Cost = 63 * NumTraits<Scalar>::MulCost + 48 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Bessel function of the second kind of
+ * order zero
+ * \sa class CwiseUnaryOp, Cwise::bessel_y0()
+ */
+template <typename Scalar>
+struct scalar_bessel_y0_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_y0_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_y0;
+    return bessel_y0(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_y0(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_y0_op<Scalar> > {
+  enum {
+    // 6 polynomial of order ~N=8 is computed.
+    // The cost is N multiplications and N additions each, along with a
+    // sine, cosine, rsqrt and j0 cost.
+    Cost = 126 * NumTraits<Scalar>::MulCost + 96 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Bessel function of the first kind of
+ * order one
+ * \sa class CwiseUnaryOp, Cwise::bessel_j1()
+ */
+template <typename Scalar>
+struct scalar_bessel_j1_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_j1_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_j1;
+    return bessel_j1(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_j1(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_j1_op<Scalar> > {
+  enum {
+    // 6 polynomial of order ~N=8 is computed.
+    // The cost is N multiplications and N additions each, along with a
+    // sine, cosine and rsqrt cost.
+    Cost = 63 * NumTraits<Scalar>::MulCost + 48 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Bessel function of the second kind of
+ * order one
+ * \sa class CwiseUnaryOp, Cwise::bessel_j1e()
+ */
+template <typename Scalar>
+struct scalar_bessel_y1_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_y1_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_y1;
+    return bessel_y1(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_y1(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_y1_op<Scalar> > {
+  enum {
+    // 6 polynomial of order ~N=8 is computed.
+    // The cost is N multiplications and N additions each, along with a
+    // sine, cosine, rsqrt and j1 cost.
+    Cost = 126 * NumTraits<Scalar>::MulCost + 96 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the modified Bessel function of the second
+ * kind of order zero
+ * \sa class CwiseUnaryOp, Cwise::bessel_k0()
+ */
+template <typename Scalar>
+struct scalar_bessel_k0_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_k0_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_k0;
+    return bessel_k0(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_k0(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_k0_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=10 is computed.
+    // The cost is N multiplications and 2N additions. In addition we compute
+    // i0, a log, exp and prsqrt and sin and cos.
+    Cost = 68 * NumTraits<Scalar>::MulCost + 88 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the exponentially scaled modified Bessel
+ * function of the second kind of order zero
+ * \sa class CwiseUnaryOp, Cwise::bessel_k0e()
+ */
+template <typename Scalar>
+struct scalar_bessel_k0e_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_k0e_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_k0e;
+    return bessel_k0e(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_k0e(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_k0e_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=10 is computed.
+    // The cost is N multiplications and 2N additions. In addition we compute
+    // i0, a log, exp and prsqrt and sin and cos.
+    Cost = 68 * NumTraits<Scalar>::MulCost + 88 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the modified Bessel function of the
+ * second kind of order one
+ * \sa class CwiseUnaryOp, Cwise::bessel_k1()
+ */
+template <typename Scalar>
+struct scalar_bessel_k1_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_k1_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_k1;
+    return bessel_k1(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_k1(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_k1_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=10 is computed.
+    // The cost is N multiplications and 2N additions. In addition we compute
+    // i1, a log, exp and prsqrt and sin and cos.
+    Cost = 68 * NumTraits<Scalar>::MulCost + 88 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the exponentially scaled modified Bessel
+ * function of the second kind of order one
+ * \sa class CwiseUnaryOp, Cwise::bessel_k1e()
+ */
+template <typename Scalar>
+struct scalar_bessel_k1e_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_bessel_k1e_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& x) const {
+    using numext::bessel_k1e;
+    return bessel_k1e(x);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return internal::pbessel_k1e(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_bessel_k1e_op<Scalar> > {
+  enum {
+    // On average, a Chebyshev polynomial of order N=10 is computed.
+    // The cost is N multiplications and 2N additions. In addition we compute
+    // i1, a log, exp and prsqrt and sin and cos.
+    Cost = 68 * NumTraits<Scalar>::MulCost + 88 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBessel
+  };
+};
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BESSELFUNCTIONS_FUNCTORS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsHalf.h b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsHalf.h
new file mode 100644
index 0000000..8930d1a
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsHalf.h
@@ -0,0 +1,66 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BESSELFUNCTIONS_HALF_H
+#define EIGEN_BESSELFUNCTIONS_HALF_H
+
+namespace Eigen {
+namespace numext {
+
+#if EIGEN_HAS_C99_MATH
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_i0(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_i0(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_i0e(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_i0e(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_i1(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_i1(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_i1e(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_i1e(static_cast<float>(x)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_j0(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_j0(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_j1(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_j1(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_y0(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_y0(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_y1(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_y1(static_cast<float>(x)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_k0(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_k0(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_k0e(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_k0e(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_k1(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_k1(static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bessel_k1e(const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::bessel_k1e(static_cast<float>(x)));
+}
+#endif
+
+}  // end namespace numext
+}  // end namespace Eigen
+
+#endif  // EIGEN_BESSELFUNCTIONS_HALF_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsImpl.h b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsImpl.h
new file mode 100644
index 0000000..24812be
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsImpl.h
@@ -0,0 +1,1959 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Eugene Brevdo <ebrevdo@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BESSEL_FUNCTIONS_H
+#define EIGEN_BESSEL_FUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+//  Parts of this code are based on the Cephes Math Library.
+//
+//  Cephes Math Library Release 2.8:  June, 2000
+//  Copyright 1984, 1987, 1992, 2000 by Stephen L. Moshier
+//
+//  Permission has been kindly provided by the original author
+//  to incorporate the Cephes software into the Eigen codebase:
+//
+//    From: Stephen Moshier
+//    To: Eugene Brevdo
+//    Subject: Re: Permission to wrap several cephes functions in Eigen
+//
+//    Hello Eugene,
+//
+//    Thank you for writing.
+//
+//    If your licensing is similar to BSD, the formal way that has been
+//    handled is simply to add a statement to the effect that you are incorporating
+//    the Cephes software by permission of the author.
+//
+//    Good luck with your project,
+//    Steve
+
+
+/****************************************************************************
+ * Implementation of Bessel function, based on Cephes                       *
+ ****************************************************************************/
+
+template <typename Scalar>
+struct bessel_i0e_retval {
+  typedef Scalar type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_i0e {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_i0e<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  i0ef.c
+     *
+     *  Modified Bessel function of order zero,
+     *  exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, i0ef();
+     *
+     * y = i0ef( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of order zero of the argument.
+     *
+     * The function is defined as i0e(x) = exp(-|x|) j0( ix ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0,30        100000      3.7e-7      7.0e-8
+     * See i0f().
+     *
+     */
+
+    const float A[] = {-1.30002500998624804212E-8f, 6.04699502254191894932E-8f,
+                       -2.67079385394061173391E-7f, 1.11738753912010371815E-6f,
+                       -4.41673835845875056359E-6f, 1.64484480707288970893E-5f,
+                       -5.75419501008210370398E-5f, 1.88502885095841655729E-4f,
+                       -5.76375574538582365885E-4f, 1.63947561694133579842E-3f,
+                       -4.32430999505057594430E-3f, 1.05464603945949983183E-2f,
+                       -2.37374148058994688156E-2f, 4.93052842396707084878E-2f,
+                       -9.49010970480476444210E-2f, 1.71620901522208775349E-1f,
+                       -3.04682672343198398683E-1f, 6.76795274409476084995E-1f};
+
+    const float B[] = {3.39623202570838634515E-9f, 2.26666899049817806459E-8f,
+                       2.04891858946906374183E-7f, 2.89137052083475648297E-6f,
+                       6.88975834691682398426E-5f, 3.36911647825569408990E-3f,
+                       8.04490411014108831608E-1f};
+    T y = pabs(x);
+    T y_le_eight = internal::pchebevl<T, 18>::run(
+        pmadd(pset1<T>(0.5f), y, pset1<T>(-2.0f)), A);
+    T y_gt_eight = pmul(
+        internal::pchebevl<T, 7>::run(
+            psub(pdiv(pset1<T>(32.0f), y), pset1<T>(2.0f)), B),
+        prsqrt(y));
+    // TODO: Perhaps instead check whether all packet elements are in
+    // [-8, 8] and evaluate a branch based off of that. It's possible
+    // in practice most elements are in this region.
+    return pselect(pcmp_le(y, pset1<T>(8.0f)), y_le_eight, y_gt_eight);
+  }
+};
+
+template <typename T>
+struct generic_i0e<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  i0e.c
+     *
+     *  Modified Bessel function of order zero,
+     *  exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, i0e();
+     *
+     * y = i0e( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of order zero of the argument.
+     *
+     * The function is defined as i0e(x) = exp(-|x|) j0( ix ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0,30        30000       5.4e-16     1.2e-16
+     * See i0().
+     *
+     */
+
+    const double A[] = {-4.41534164647933937950E-18, 3.33079451882223809783E-17,
+                        -2.43127984654795469359E-16, 1.71539128555513303061E-15,
+                        -1.16853328779934516808E-14, 7.67618549860493561688E-14,
+                        -4.85644678311192946090E-13, 2.95505266312963983461E-12,
+                        -1.72682629144155570723E-11, 9.67580903537323691224E-11,
+                        -5.18979560163526290666E-10, 2.65982372468238665035E-9,
+                        -1.30002500998624804212E-8,  6.04699502254191894932E-8,
+                        -2.67079385394061173391E-7,  1.11738753912010371815E-6,
+                        -4.41673835845875056359E-6,  1.64484480707288970893E-5,
+                        -5.75419501008210370398E-5,  1.88502885095841655729E-4,
+                        -5.76375574538582365885E-4,  1.63947561694133579842E-3,
+                        -4.32430999505057594430E-3,  1.05464603945949983183E-2,
+                        -2.37374148058994688156E-2,  4.93052842396707084878E-2,
+                        -9.49010970480476444210E-2,  1.71620901522208775349E-1,
+                        -3.04682672343198398683E-1,  6.76795274409476084995E-1};
+    const double B[] = {
+        -7.23318048787475395456E-18, -4.83050448594418207126E-18,
+        4.46562142029675999901E-17,  3.46122286769746109310E-17,
+        -2.82762398051658348494E-16, -3.42548561967721913462E-16,
+        1.77256013305652638360E-15,  3.81168066935262242075E-15,
+        -9.55484669882830764870E-15, -4.15056934728722208663E-14,
+        1.54008621752140982691E-14,  3.85277838274214270114E-13,
+        7.18012445138366623367E-13,  -1.79417853150680611778E-12,
+        -1.32158118404477131188E-11, -3.14991652796324136454E-11,
+        1.18891471078464383424E-11,  4.94060238822496958910E-10,
+        3.39623202570838634515E-9,   2.26666899049817806459E-8,
+        2.04891858946906374183E-7,   2.89137052083475648297E-6,
+        6.88975834691682398426E-5,   3.36911647825569408990E-3,
+        8.04490411014108831608E-1};
+    T y = pabs(x);
+    T y_le_eight = internal::pchebevl<T, 30>::run(
+        pmadd(pset1<T>(0.5), y, pset1<T>(-2.0)), A);
+    T y_gt_eight = pmul(
+        internal::pchebevl<T, 25>::run(
+            psub(pdiv(pset1<T>(32.0), y), pset1<T>(2.0)), B),
+        prsqrt(y));
+    // TODO: Perhaps instead check whether all packet elements are in
+    // [-8, 8] and evaluate a branch based off of that. It's possible
+    // in practice most elements are in this region.
+    return pselect(pcmp_le(y, pset1<T>(8.0)), y_le_eight, y_gt_eight);
+  }
+};
+
+template <typename T>
+struct bessel_i0e_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_i0e<T>::run(x);
+  }
+};
+
+template <typename Scalar>
+struct bessel_i0_retval {
+  typedef Scalar type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_i0 {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    return pmul(
+        pexp(pabs(x)),
+        generic_i0e<T, ScalarType>::run(x));
+  }
+};
+
+template <typename T>
+struct bessel_i0_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_i0<T>::run(x);
+  }
+};
+
+template <typename Scalar>
+struct bessel_i1e_retval {
+  typedef Scalar type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type >
+struct generic_i1e {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_i1e<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /* i1ef.c
+     *
+     *  Modified Bessel function of order one,
+     *  exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, i1ef();
+     *
+     * y = i1ef( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of order one of the argument.
+     *
+     * The function is defined as i1(x) = -i exp(-|x|) j1( ix ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       1.5e-6      1.5e-7
+     * See i1().
+     *
+     */
+    const float A[] = {9.38153738649577178388E-9f, -4.44505912879632808065E-8f,
+                       2.00329475355213526229E-7f, -8.56872026469545474066E-7f,
+                       3.47025130813767847674E-6f, -1.32731636560394358279E-5f,
+                       4.78156510755005422638E-5f, -1.61760815825896745588E-4f,
+                       5.12285956168575772895E-4f, -1.51357245063125314899E-3f,
+                       4.15642294431288815669E-3f, -1.05640848946261981558E-2f,
+                       2.47264490306265168283E-2f, -5.29459812080949914269E-2f,
+                       1.02643658689847095384E-1f, -1.76416518357834055153E-1f,
+                       2.52587186443633654823E-1f};
+
+    const float B[] = {-3.83538038596423702205E-9f, -2.63146884688951950684E-8f,
+                       -2.51223623787020892529E-7f, -3.88256480887769039346E-6f,
+                       -1.10588938762623716291E-4f, -9.76109749136146840777E-3f,
+                       7.78576235018280120474E-1f};
+
+
+    T y = pabs(x);
+    T y_le_eight = pmul(y, internal::pchebevl<T, 17>::run(
+        pmadd(pset1<T>(0.5f), y, pset1<T>(-2.0f)), A));
+    T y_gt_eight = pmul(
+        internal::pchebevl<T, 7>::run(
+            psub(pdiv(pset1<T>(32.0f), y),
+                 pset1<T>(2.0f)), B),
+        prsqrt(y));
+    // TODO: Perhaps instead check whether all packet elements are in
+    // [-8, 8] and evaluate a branch based off of that. It's possible
+    // in practice most elements are in this region.
+    y = pselect(pcmp_le(y, pset1<T>(8.0f)), y_le_eight, y_gt_eight);
+    return pselect(pcmp_lt(x, pset1<T>(0.0f)), pnegate(y), y);
+  }
+};
+
+template <typename T>
+struct generic_i1e<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  i1e.c
+     *
+     *  Modified Bessel function of order one,
+     *  exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, i1e();
+     *
+     * y = i1e( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of order one of the argument.
+     *
+     * The function is defined as i1(x) = -i exp(-|x|) j1( ix ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       2.0e-15     2.0e-16
+     * See i1().
+     *
+     */
+    const double A[] = {2.77791411276104639959E-18, -2.11142121435816608115E-17,
+                        1.55363195773620046921E-16, -1.10559694773538630805E-15,
+                        7.60068429473540693410E-15, -5.04218550472791168711E-14,
+                        3.22379336594557470981E-13, -1.98397439776494371520E-12,
+                        1.17361862988909016308E-11, -6.66348972350202774223E-11,
+                        3.62559028155211703701E-10, -1.88724975172282928790E-9,
+                        9.38153738649577178388E-9,  -4.44505912879632808065E-8,
+                        2.00329475355213526229E-7,  -8.56872026469545474066E-7,
+                        3.47025130813767847674E-6,  -1.32731636560394358279E-5,
+                        4.78156510755005422638E-5,  -1.61760815825896745588E-4,
+                        5.12285956168575772895E-4,  -1.51357245063125314899E-3,
+                        4.15642294431288815669E-3,  -1.05640848946261981558E-2,
+                        2.47264490306265168283E-2,  -5.29459812080949914269E-2,
+                        1.02643658689847095384E-1,  -1.76416518357834055153E-1,
+                        2.52587186443633654823E-1};
+    const double B[] = {
+        7.51729631084210481353E-18,  4.41434832307170791151E-18,
+        -4.65030536848935832153E-17, -3.20952592199342395980E-17,
+        2.96262899764595013876E-16,  3.30820231092092828324E-16,
+        -1.88035477551078244854E-15, -3.81440307243700780478E-15,
+        1.04202769841288027642E-14,  4.27244001671195135429E-14,
+        -2.10154184277266431302E-14, -4.08355111109219731823E-13,
+        -7.19855177624590851209E-13, 2.03562854414708950722E-12,
+        1.41258074366137813316E-11,  3.25260358301548823856E-11,
+        -1.89749581235054123450E-11, -5.58974346219658380687E-10,
+        -3.83538038596423702205E-9,  -2.63146884688951950684E-8,
+        -2.51223623787020892529E-7,  -3.88256480887769039346E-6,
+        -1.10588938762623716291E-4,  -9.76109749136146840777E-3,
+        7.78576235018280120474E-1};
+    T y = pabs(x);
+    T y_le_eight = pmul(y, internal::pchebevl<T, 29>::run(
+        pmadd(pset1<T>(0.5), y, pset1<T>(-2.0)), A));
+    T y_gt_eight = pmul(
+        internal::pchebevl<T, 25>::run(
+            psub(pdiv(pset1<T>(32.0), y),
+                 pset1<T>(2.0)), B),
+        prsqrt(y));
+    // TODO: Perhaps instead check whether all packet elements are in
+    // [-8, 8] and evaluate a branch based off of that. It's possible
+    // in practice most elements are in this region.
+    y = pselect(pcmp_le(y, pset1<T>(8.0)), y_le_eight, y_gt_eight);
+    return pselect(pcmp_lt(x, pset1<T>(0.0)), pnegate(y), y);
+  }
+};
+
+template <typename T>
+struct bessel_i1e_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_i1e<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_i1_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_i1 {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    return pmul(
+        pexp(pabs(x)),
+        generic_i1e<T, ScalarType>::run(x));
+  }
+};
+
+template <typename T>
+struct bessel_i1_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_i1<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_k0e_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_k0e {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_k0e<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  k0ef.c
+     *	Modified Bessel function, third kind, order zero,
+     *	exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, k0ef();
+     *
+     * y = k0ef( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of the third kind of order zero of the argument.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       8.1e-7      7.8e-8
+     * See k0().
+     *
+     */
+
+    const float A[] = {1.90451637722020886025E-9f, 2.53479107902614945675E-7f,
+                       2.28621210311945178607E-5f, 1.26461541144692592338E-3f,
+                       3.59799365153615016266E-2f, 3.44289899924628486886E-1f,
+                       -5.35327393233902768720E-1f};
+
+    const float B[] = {-1.69753450938905987466E-9f, 8.57403401741422608519E-9f,
+                       -4.66048989768794782956E-8f, 2.76681363944501510342E-7f,
+                       -1.83175552271911948767E-6f, 1.39498137188764993662E-5f,
+                       -1.28495495816278026384E-4f, 1.56988388573005337491E-3f,
+                       -3.14481013119645005427E-2f, 2.44030308206595545468E0f};
+    const T MAXNUM = pset1<T>(NumTraits<float>::infinity());
+    const T two = pset1<T>(2.0);
+    T x_le_two = internal::pchebevl<T, 7>::run(
+        pmadd(x, x, pset1<T>(-2.0)), A);
+    x_le_two = pmadd(
+        generic_i0<T, float>::run(x), pnegate(
+            plog(pmul(pset1<T>(0.5), x))), x_le_two);
+    x_le_two = pmul(pexp(x), x_le_two);
+    T x_gt_two = pmul(
+            internal::pchebevl<T, 10>::run(
+                psub(pdiv(pset1<T>(8.0), x), two), B),
+            prsqrt(x));
+    return pselect(
+        pcmp_le(x, pset1<T>(0.0)),
+        MAXNUM,
+        pselect(pcmp_le(x, two), x_le_two, x_gt_two));
+  }
+};
+
+template <typename T>
+struct generic_k0e<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  k0e.c
+     *	Modified Bessel function, third kind, order zero,
+     *	exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, k0e();
+     *
+     * y = k0e( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of the third kind of order zero of the argument.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       1.4e-15     1.4e-16
+     * See k0().
+     *
+     */
+
+    const double A[] = {
+      1.37446543561352307156E-16,
+      4.25981614279661018399E-14,
+      1.03496952576338420167E-11,
+      1.90451637722020886025E-9,
+      2.53479107902614945675E-7,
+      2.28621210311945178607E-5,
+      1.26461541144692592338E-3,
+      3.59799365153615016266E-2,
+      3.44289899924628486886E-1,
+      -5.35327393233902768720E-1};
+    const double B[] = {
+       5.30043377268626276149E-18, -1.64758043015242134646E-17,
+       5.21039150503902756861E-17, -1.67823109680541210385E-16,
+       5.51205597852431940784E-16, -1.84859337734377901440E-15,
+       6.34007647740507060557E-15, -2.22751332699166985548E-14,
+       8.03289077536357521100E-14, -2.98009692317273043925E-13,
+       1.14034058820847496303E-12, -4.51459788337394416547E-12,
+       1.85594911495471785253E-11, -7.95748924447710747776E-11,
+       3.57739728140030116597E-10, -1.69753450938905987466E-9,
+       8.57403401741422608519E-9, -4.66048989768794782956E-8,
+       2.76681363944501510342E-7, -1.83175552271911948767E-6,
+       1.39498137188764993662E-5, -1.28495495816278026384E-4,
+       1.56988388573005337491E-3, -3.14481013119645005427E-2,
+       2.44030308206595545468E0
+    };
+    const T MAXNUM = pset1<T>(NumTraits<double>::infinity());
+    const T two = pset1<T>(2.0);
+    T x_le_two = internal::pchebevl<T, 10>::run(
+        pmadd(x, x, pset1<T>(-2.0)), A);
+    x_le_two = pmadd(
+        generic_i0<T, double>::run(x), pmul(
+            pset1<T>(-1.0), plog(pmul(pset1<T>(0.5), x))), x_le_two);
+    x_le_two = pmul(pexp(x), x_le_two);
+    x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
+    T x_gt_two = pmul(
+            internal::pchebevl<T, 25>::run(
+                psub(pdiv(pset1<T>(8.0), x), two), B),
+            prsqrt(x));
+    return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct bessel_k0e_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_k0e<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_k0_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_k0 {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_k0<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  k0f.c
+     *	Modified Bessel function, third kind, order zero
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, k0f();
+     *
+     * y = k0f( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns modified Bessel function of the third kind
+     * of order zero of the argument.
+     *
+     * The range is partitioned into the two intervals [0,8] and
+     * (8, infinity).  Chebyshev polynomial expansions are employed
+     * in each interval.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     * Tested at 2000 random points between 0 and 8.  Peak absolute
+     * error (relative when K0 > 1) was 1.46e-14; rms, 4.26e-15.
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       7.8e-7      8.5e-8
+     *
+     * ERROR MESSAGES:
+     *
+     *   message         condition      value returned
+     *  K0 domain          x <= 0          MAXNUM
+     *
+     */
+
+    const float A[] = {1.90451637722020886025E-9f, 2.53479107902614945675E-7f,
+                       2.28621210311945178607E-5f, 1.26461541144692592338E-3f,
+                       3.59799365153615016266E-2f, 3.44289899924628486886E-1f,
+                       -5.35327393233902768720E-1f};
+
+    const float B[] = {-1.69753450938905987466E-9f, 8.57403401741422608519E-9f,
+                       -4.66048989768794782956E-8f, 2.76681363944501510342E-7f,
+                       -1.83175552271911948767E-6f, 1.39498137188764993662E-5f,
+                       -1.28495495816278026384E-4f, 1.56988388573005337491E-3f,
+                       -3.14481013119645005427E-2f, 2.44030308206595545468E0f};
+    const T MAXNUM = pset1<T>(NumTraits<float>::infinity());
+    const T two = pset1<T>(2.0);
+    T x_le_two = internal::pchebevl<T, 7>::run(
+        pmadd(x, x, pset1<T>(-2.0)), A);
+    x_le_two = pmadd(
+        generic_i0<T, float>::run(x), pnegate(
+            plog(pmul(pset1<T>(0.5), x))), x_le_two);
+    x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
+    T x_gt_two = pmul(
+        pmul(
+            pexp(pnegate(x)),
+            internal::pchebevl<T, 10>::run(
+                psub(pdiv(pset1<T>(8.0), x), two), B)),
+        prsqrt(x));
+    return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct generic_k0<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*
+     *
+     *	Modified Bessel function, third kind, order zero,
+     *	exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, k0();
+     *
+     * y = k0( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of the third kind of order zero of the argument.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       1.4e-15     1.4e-16
+     * See k0().
+     *
+     */
+    const double A[] = {
+      1.37446543561352307156E-16,
+      4.25981614279661018399E-14,
+      1.03496952576338420167E-11,
+      1.90451637722020886025E-9,
+      2.53479107902614945675E-7,
+      2.28621210311945178607E-5,
+      1.26461541144692592338E-3,
+      3.59799365153615016266E-2,
+      3.44289899924628486886E-1,
+      -5.35327393233902768720E-1};
+    const double B[] = {
+       5.30043377268626276149E-18, -1.64758043015242134646E-17,
+       5.21039150503902756861E-17, -1.67823109680541210385E-16,
+       5.51205597852431940784E-16, -1.84859337734377901440E-15,
+       6.34007647740507060557E-15, -2.22751332699166985548E-14,
+       8.03289077536357521100E-14, -2.98009692317273043925E-13,
+       1.14034058820847496303E-12, -4.51459788337394416547E-12,
+       1.85594911495471785253E-11, -7.95748924447710747776E-11,
+       3.57739728140030116597E-10, -1.69753450938905987466E-9,
+       8.57403401741422608519E-9, -4.66048989768794782956E-8,
+       2.76681363944501510342E-7, -1.83175552271911948767E-6,
+       1.39498137188764993662E-5, -1.28495495816278026384E-4,
+       1.56988388573005337491E-3, -3.14481013119645005427E-2,
+       2.44030308206595545468E0
+    };
+    const T MAXNUM = pset1<T>(NumTraits<double>::infinity());
+    const T two = pset1<T>(2.0);
+    T x_le_two = internal::pchebevl<T, 10>::run(
+        pmadd(x, x, pset1<T>(-2.0)), A);
+    x_le_two = pmadd(
+        generic_i0<T, double>::run(x), pnegate(
+            plog(pmul(pset1<T>(0.5), x))), x_le_two);
+    x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
+    T x_gt_two = pmul(
+        pmul(
+            pexp(-x),
+            internal::pchebevl<T, 25>::run(
+                psub(pdiv(pset1<T>(8.0), x), two), B)),
+        prsqrt(x));
+    return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct bessel_k0_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_k0<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_k1e_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_k1e {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_k1e<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /* k1ef.c
+     *
+     *	Modified Bessel function, third kind, order one,
+     *	exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, k1ef();
+     *
+     * y = k1ef( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of the third kind of order one of the argument:
+     *
+     *      k1e(x) = exp(x) * k1(x).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       4.9e-7      6.7e-8
+     * See k1().
+     *
+     */
+
+    const float A[] = {-2.21338763073472585583E-8f, -2.43340614156596823496E-6f,
+                        -1.73028895751305206302E-4f, -6.97572385963986435018E-3f,
+                        -1.22611180822657148235E-1f, -3.53155960776544875667E-1f,
+                        1.52530022733894777053E0f};
+    const float B[] = {2.01504975519703286596E-9f, -1.03457624656780970260E-8f,
+                       5.74108412545004946722E-8f, -3.50196060308781257119E-7f,
+                       2.40648494783721712015E-6f, -1.93619797416608296024E-5f,
+                       1.95215518471351631108E-4f, -2.85781685962277938680E-3f,
+                       1.03923736576817238437E-1f, 2.72062619048444266945E0f};
+    const T MAXNUM = pset1<T>(NumTraits<float>::infinity());
+    const T two = pset1<T>(2.0);
+    T x_le_two = pdiv(internal::pchebevl<T, 7>::run(
+        pmadd(x, x, pset1<T>(-2.0)), A), x);
+    x_le_two = pmadd(
+        generic_i1<T, float>::run(x), plog(pmul(pset1<T>(0.5), x)), x_le_two);
+    x_le_two = pmul(x_le_two, pexp(x));
+    x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
+    T x_gt_two = pmul(
+        internal::pchebevl<T, 10>::run(
+            psub(pdiv(pset1<T>(8.0), x), two), B),
+        prsqrt(x));
+    return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct generic_k1e<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  k1e.c
+     *
+     *	Modified Bessel function, third kind, order one,
+     *	exponentially scaled
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, k1e();
+     *
+     * y = k1e( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns exponentially scaled modified Bessel function
+     * of the third kind of order one of the argument:
+     *
+     *      k1e(x) = exp(x) * k1(x).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       7.8e-16     1.2e-16
+     * See k1().
+     *
+     */
+    const double A[] = {-7.02386347938628759343E-18, -2.42744985051936593393E-15,
+                        -6.66690169419932900609E-13, -1.41148839263352776110E-10,
+                        -2.21338763073472585583E-8, -2.43340614156596823496E-6,
+                        -1.73028895751305206302E-4, -6.97572385963986435018E-3,
+                        -1.22611180822657148235E-1, -3.53155960776544875667E-1,
+                        1.52530022733894777053E0};
+    const double B[] = {-5.75674448366501715755E-18, 1.79405087314755922667E-17,
+                        -5.68946255844285935196E-17, 1.83809354436663880070E-16,
+                        -6.05704724837331885336E-16, 2.03870316562433424052E-15,
+                        -7.01983709041831346144E-15, 2.47715442448130437068E-14,
+                        -8.97670518232499435011E-14, 3.34841966607842919884E-13,
+                        -1.28917396095102890680E-12, 5.13963967348173025100E-12,
+                        -2.12996783842756842877E-11, 9.21831518760500529508E-11,
+                        -4.19035475934189648750E-10, 2.01504975519703286596E-9,
+                        -1.03457624656780970260E-8, 5.74108412545004946722E-8,
+                        -3.50196060308781257119E-7, 2.40648494783721712015E-6,
+                        -1.93619797416608296024E-5, 1.95215518471351631108E-4,
+                        -2.85781685962277938680E-3, 1.03923736576817238437E-1,
+                        2.72062619048444266945E0};
+    const T MAXNUM = pset1<T>(NumTraits<double>::infinity());
+    const T two = pset1<T>(2.0);
+    T x_le_two = pdiv(internal::pchebevl<T, 11>::run(
+        pmadd(x, x, pset1<T>(-2.0)), A), x);
+    x_le_two = pmadd(
+        generic_i1<T, double>::run(x), plog(pmul(pset1<T>(0.5), x)), x_le_two);
+    x_le_two = pmul(x_le_two, pexp(x));
+    x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
+    T x_gt_two = pmul(
+        internal::pchebevl<T, 25>::run(
+            psub(pdiv(pset1<T>(8.0), x), two), B),
+        prsqrt(x));
+    return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct bessel_k1e_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_k1e<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_k1_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_k1 {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_k1<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /* k1f.c
+     *	Modified Bessel function, third kind, order one
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, k1f();
+     *
+     * y = k1f( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Computes the modified Bessel function of the third kind
+     * of order one of the argument.
+     *
+     * The range is partitioned into the two intervals [0,2] and
+     * (2, infinity).  Chebyshev polynomial expansions are employed
+     * in each interval.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       4.6e-7      7.6e-8
+     *
+     * ERROR MESSAGES:
+     *
+     *   message         condition      value returned
+     * k1 domain          x <= 0          MAXNUM
+     *
+     */
+
+    const float A[] = {-2.21338763073472585583E-8f, -2.43340614156596823496E-6f,
+                        -1.73028895751305206302E-4f, -6.97572385963986435018E-3f,
+                        -1.22611180822657148235E-1f, -3.53155960776544875667E-1f,
+                        1.52530022733894777053E0f};
+    const float B[] = {2.01504975519703286596E-9f, -1.03457624656780970260E-8f,
+                       5.74108412545004946722E-8f, -3.50196060308781257119E-7f,
+                       2.40648494783721712015E-6f, -1.93619797416608296024E-5f,
+                       1.95215518471351631108E-4f, -2.85781685962277938680E-3f,
+                       1.03923736576817238437E-1f, 2.72062619048444266945E0f};
+    const T MAXNUM = pset1<T>(NumTraits<float>::infinity());
+    const T two = pset1<T>(2.0);
+    T x_le_two = pdiv(internal::pchebevl<T, 7>::run(
+        pmadd(x, x, pset1<T>(-2.0)), A), x);
+    x_le_two = pmadd(
+        generic_i1<T, float>::run(x), plog(pmul(pset1<T>(0.5), x)), x_le_two);
+    x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
+    T x_gt_two = pmul(
+        pexp(pnegate(x)),
+        pmul(
+            internal::pchebevl<T, 10>::run(
+                psub(pdiv(pset1<T>(8.0), x), two), B),
+            prsqrt(x)));
+    return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct generic_k1<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  k1.c
+     *	Modified Bessel function, third kind, order one
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, k1f();
+     *
+     * y = k1f( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Computes the modified Bessel function of the third kind
+     * of order one of the argument.
+     *
+     * The range is partitioned into the two intervals [0,2] and
+     * (2, infinity).  Chebyshev polynomial expansions are employed
+     * in each interval.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 30       30000       4.6e-7      7.6e-8
+     *
+     * ERROR MESSAGES:
+     *
+     *   message         condition      value returned
+     * k1 domain          x <= 0          MAXNUM
+     *
+     */
+    const double A[] = {-7.02386347938628759343E-18, -2.42744985051936593393E-15,
+                        -6.66690169419932900609E-13, -1.41148839263352776110E-10,
+                        -2.21338763073472585583E-8, -2.43340614156596823496E-6,
+                        -1.73028895751305206302E-4, -6.97572385963986435018E-3,
+                        -1.22611180822657148235E-1, -3.53155960776544875667E-1,
+                        1.52530022733894777053E0};
+    const double B[] = {-5.75674448366501715755E-18, 1.79405087314755922667E-17,
+                        -5.68946255844285935196E-17, 1.83809354436663880070E-16,
+                        -6.05704724837331885336E-16, 2.03870316562433424052E-15,
+                        -7.01983709041831346144E-15, 2.47715442448130437068E-14,
+                        -8.97670518232499435011E-14, 3.34841966607842919884E-13,
+                        -1.28917396095102890680E-12, 5.13963967348173025100E-12,
+                        -2.12996783842756842877E-11, 9.21831518760500529508E-11,
+                        -4.19035475934189648750E-10, 2.01504975519703286596E-9,
+                        -1.03457624656780970260E-8, 5.74108412545004946722E-8,
+                        -3.50196060308781257119E-7, 2.40648494783721712015E-6,
+                        -1.93619797416608296024E-5, 1.95215518471351631108E-4,
+                        -2.85781685962277938680E-3, 1.03923736576817238437E-1,
+                        2.72062619048444266945E0};
+    const T MAXNUM = pset1<T>(NumTraits<double>::infinity());
+    const T two = pset1<T>(2.0);
+    T x_le_two = pdiv(internal::pchebevl<T, 11>::run(
+        pmadd(x, x, pset1<T>(-2.0)), A), x);
+    x_le_two = pmadd(
+        generic_i1<T, double>::run(x), plog(pmul(pset1<T>(0.5), x)), x_le_two);
+    x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
+    T x_gt_two = pmul(
+        pexp(-x),
+        pmul(
+            internal::pchebevl<T, 25>::run(
+                psub(pdiv(pset1<T>(8.0), x), two), B),
+            prsqrt(x)));
+    return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct bessel_k1_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_k1<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_j0_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_j0 {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_j0<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /* j0f.c
+     *	Bessel function of order zero
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, j0f();
+     *
+     * y = j0f( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns Bessel function of order zero of the argument.
+     *
+     * The domain is divided into the intervals [0, 2] and
+     * (2, infinity). In the first interval the following polynomial
+     * approximation is used:
+     *
+     *
+     *        2         2         2
+     * (w - r  ) (w - r  ) (w - r  ) P(w)
+     *       1         2         3
+     *
+     *            2
+     * where w = x  and the three r's are zeros of the function.
+     *
+     * In the second interval, the modulus and phase are approximated
+     * by polynomials of the form Modulus(x) = sqrt(1/x) Q(1/x)
+     * and Phase(x) = x + 1/x R(1/x^2) - pi/4.  The function is
+     *
+     *   j0(x) = Modulus(x) cos( Phase(x) ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Absolute error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0, 2        100000      1.3e-7      3.6e-8
+     *    IEEE      2, 32       100000      1.9e-7      5.4e-8
+     *
+     */
+
+    const float JP[] = {-6.068350350393235E-008f, 6.388945720783375E-006f,
+                        -3.969646342510940E-004f, 1.332913422519003E-002f,
+                        -1.729150680240724E-001f};
+    const float MO[] = {-6.838999669318810E-002f, 1.864949361379502E-001f,
+                        -2.145007480346739E-001f, 1.197549369473540E-001f,
+                        -3.560281861530129E-003f, -4.969382655296620E-002f,
+                        -3.355424622293709E-006f, 7.978845717621440E-001f};
+    const float PH[] = {3.242077816988247E+001f, -3.630592630518434E+001f,
+                        1.756221482109099E+001f, -4.974978466280903E+000f,
+                        1.001973420681837E+000f, -1.939906941791308E-001f,
+                        6.490598792654666E-002f, -1.249992184872738E-001f};
+    const T DR1 =  pset1<T>(5.78318596294678452118f);
+    const T NEG_PIO4F = pset1<T>(-0.7853981633974483096f); /* -pi / 4 */
+    T y = pabs(x);
+    T z = pmul(y, y);
+    T y_le_two = pselect(
+        pcmp_lt(y, pset1<T>(1.0e-3f)),
+        pmadd(z, pset1<T>(-0.25f), pset1<T>(1.0f)),
+        pmul(psub(z, DR1), internal::ppolevl<T, 4>::run(z, JP)));
+    T q = pdiv(pset1<T>(1.0f), y);
+    T w = prsqrt(y);
+    T p = pmul(w, internal::ppolevl<T, 7>::run(q, MO));
+    w = pmul(q, q);
+    T yn = pmadd(q, internal::ppolevl<T, 7>::run(w, PH), NEG_PIO4F);
+    T y_gt_two = pmul(p, pcos(padd(yn, y)));
+    return pselect(pcmp_le(y, pset1<T>(2.0)), y_le_two, y_gt_two);
+  }
+};
+
+template <typename T>
+struct generic_j0<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  j0.c
+     *	Bessel function of order zero
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, j0();
+     *
+     * y = j0( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns Bessel function of order zero of the argument.
+     *
+     * The domain is divided into the intervals [0, 5] and
+     * (5, infinity). In the first interval the following rational
+     * approximation is used:
+     *
+     *
+     *        2         2
+     * (w - r  ) (w - r  ) P (w) / Q (w)
+     *       1         2    3       8
+     *
+     *            2
+     * where w = x  and the two r's are zeros of the function.
+     *
+     * In the second interval, the Hankel asymptotic expansion
+     * is employed with two rational functions of degree 6/6
+     * and 7/7.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Absolute error:
+     * arithmetic   domain     # trials      peak         rms
+     *    DEC       0, 30       10000       4.4e-17     6.3e-18
+     *    IEEE      0, 30       60000       4.2e-16     1.1e-16
+     *
+     */
+    const double PP[] = {7.96936729297347051624E-4, 8.28352392107440799803E-2,
+                        1.23953371646414299388E0, 5.44725003058768775090E0,
+                        8.74716500199817011941E0, 5.30324038235394892183E0,
+                        9.99999999999999997821E-1};
+    const double PQ[] = {9.24408810558863637013E-4, 8.56288474354474431428E-2,
+                         1.25352743901058953537E0, 5.47097740330417105182E0,
+                         8.76190883237069594232E0, 5.30605288235394617618E0,
+                         1.00000000000000000218E0};
+    const double QP[] = {-1.13663838898469149931E-2, -1.28252718670509318512E0,
+                         -1.95539544257735972385E1, -9.32060152123768231369E1,
+                         -1.77681167980488050595E2, -1.47077505154951170175E2,
+                         -5.14105326766599330220E1, -6.05014350600728481186E0};
+    const double QQ[] = {1.00000000000000000000E0, 6.43178256118178023184E1,
+                         8.56430025976980587198E2, 3.88240183605401609683E3,
+                         7.24046774195652478189E3, 5.93072701187316984827E3,
+                         2.06209331660327847417E3, 2.42005740240291393179E2};
+    const double RP[] = {-4.79443220978201773821E9, 1.95617491946556577543E12,
+                         -2.49248344360967716204E14, 9.70862251047306323952E15};
+    const double RQ[] = {1.00000000000000000000E0, 4.99563147152651017219E2,
+                         1.73785401676374683123E5, 4.84409658339962045305E7,
+                         1.11855537045356834862E10, 2.11277520115489217587E12,
+                         3.10518229857422583814E14, 3.18121955943204943306E16,
+                         1.71086294081043136091E18};
+    const T DR1 = pset1<T>(5.78318596294678452118E0);
+    const T DR2 = pset1<T>(3.04712623436620863991E1);
+    const T SQ2OPI = pset1<T>(7.9788456080286535587989E-1); /* sqrt(2 / pi) */
+    const T NEG_PIO4 = pset1<T>(-0.7853981633974483096); /* pi / 4 */
+
+    T y = pabs(x);
+    T z = pmul(y, y);
+    T y_le_five = pselect(
+        pcmp_lt(y, pset1<T>(1.0e-5)),
+        pmadd(z, pset1<T>(-0.25), pset1<T>(1.0)),
+        pmul(pmul(psub(z, DR1), psub(z, DR2)),
+             pdiv(internal::ppolevl<T, 3>::run(z, RP),
+                  internal::ppolevl<T, 8>::run(z, RQ))));
+    T s = pdiv(pset1<T>(25.0), z);
+    T p = pdiv(
+        internal::ppolevl<T, 6>::run(s, PP),
+        internal::ppolevl<T, 6>::run(s, PQ));
+    T q = pdiv(
+        internal::ppolevl<T, 7>::run(s, QP),
+        internal::ppolevl<T, 7>::run(s, QQ));
+    T yn = padd(y, NEG_PIO4);
+    T w = pdiv(pset1<T>(-5.0), y);
+    p = pmadd(p, pcos(yn), pmul(w, pmul(q, psin(yn))));
+    T y_gt_five = pmul(p, pmul(SQ2OPI, prsqrt(y)));
+    return pselect(pcmp_le(y, pset1<T>(5.0)), y_le_five, y_gt_five);
+  }
+};
+
+template <typename T>
+struct bessel_j0_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_j0<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_y0_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_y0 {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_y0<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /* j0f.c
+     * 	Bessel function of the second kind, order zero
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, y0f();
+     *
+     * y = y0f( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns Bessel function of the second kind, of order
+     * zero, of the argument.
+     *
+     * The domain is divided into the intervals [0, 2] and
+     * (2, infinity). In the first interval a rational approximation
+     * R(x) is employed to compute
+     *
+     *                  2         2         2
+     * y0(x)  =  (w - r  ) (w - r  ) (w - r  ) R(x)  +  2/pi ln(x) j0(x).
+     *                 1         2         3
+     *
+     * Thus a call to j0() is required.  The three zeros are removed
+     * from R(x) to improve its numerical stability.
+     *
+     * In the second interval, the modulus and phase are approximated
+     * by polynomials of the form Modulus(x) = sqrt(1/x) Q(1/x)
+     * and Phase(x) = x + 1/x S(1/x^2) - pi/4.  Then the function is
+     *
+     *   y0(x) = Modulus(x) sin( Phase(x) ).
+     *
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *  Absolute error, when y0(x) < 1; else relative error:
+     *
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0,  2       100000      2.4e-7      3.4e-8
+     *    IEEE      2, 32       100000      1.8e-7      5.3e-8
+     *
+     */
+
+    const float YP[] = {9.454583683980369E-008f, -9.413212653797057E-006f,
+                        5.344486707214273E-004f, -1.584289289821316E-002f,
+                        1.707584643733568E-001f};
+    const float MO[] = {-6.838999669318810E-002f, 1.864949361379502E-001f,
+                        -2.145007480346739E-001f, 1.197549369473540E-001f,
+                        -3.560281861530129E-003f, -4.969382655296620E-002f,
+                        -3.355424622293709E-006f, 7.978845717621440E-001f};
+    const float PH[] = {3.242077816988247E+001f, -3.630592630518434E+001f,
+                        1.756221482109099E+001f, -4.974978466280903E+000f,
+                        1.001973420681837E+000f, -1.939906941791308E-001f,
+                        6.490598792654666E-002f, -1.249992184872738E-001f};
+    const T YZ1 = pset1<T>(0.43221455686510834878f);
+    const T TWOOPI =  pset1<T>(0.636619772367581343075535f); /* 2 / pi */
+    const T NEG_PIO4F = pset1<T>(-0.7853981633974483096f); /* -pi / 4 */
+    const T NEG_MAXNUM = pset1<T>(-NumTraits<float>::infinity());
+    T z = pmul(x, x);
+    T x_le_two = pmul(TWOOPI, pmul(plog(x), generic_j0<T, float>::run(x)));
+    x_le_two = pmadd(
+        psub(z, YZ1), internal::ppolevl<T, 4>::run(z, YP), x_le_two);
+    x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), NEG_MAXNUM, x_le_two);
+    T q = pdiv(pset1<T>(1.0), x);
+    T w = prsqrt(x);
+    T p = pmul(w, internal::ppolevl<T, 7>::run(q, MO));
+    T u = pmul(q, q);
+    T xn = pmadd(q, internal::ppolevl<T, 7>::run(u, PH), NEG_PIO4F);
+    T x_gt_two = pmul(p, psin(padd(xn, x)));
+    return pselect(pcmp_le(x, pset1<T>(2.0)), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct generic_y0<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  j0.c
+     *	Bessel function of the second kind, order zero
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, y0();
+     *
+     * y = y0( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns Bessel function of the second kind, of order
+     * zero, of the argument.
+     *
+     * The domain is divided into the intervals [0, 5] and
+     * (5, infinity). In the first interval a rational approximation
+     * R(x) is employed to compute
+     *   y0(x)  = R(x)  +   2 * log(x) * j0(x) / PI.
+     * Thus a call to j0() is required.
+     *
+     * In the second interval, the Hankel asymptotic expansion
+     * is employed with two rational functions of degree 6/6
+     * and 7/7.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *  Absolute error, when y0(x) < 1; else relative error:
+     *
+     * arithmetic   domain     # trials      peak         rms
+     *    DEC       0, 30        9400       7.0e-17     7.9e-18
+     *    IEEE      0, 30       30000       1.3e-15     1.6e-16
+     *
+     */
+    const double PP[] = {7.96936729297347051624E-4, 8.28352392107440799803E-2,
+                        1.23953371646414299388E0, 5.44725003058768775090E0,
+                        8.74716500199817011941E0, 5.30324038235394892183E0,
+                        9.99999999999999997821E-1};
+    const double PQ[] = {9.24408810558863637013E-4, 8.56288474354474431428E-2,
+                         1.25352743901058953537E0, 5.47097740330417105182E0,
+                         8.76190883237069594232E0, 5.30605288235394617618E0,
+                         1.00000000000000000218E0};
+    const double QP[] = {-1.13663838898469149931E-2, -1.28252718670509318512E0,
+                         -1.95539544257735972385E1, -9.32060152123768231369E1,
+                         -1.77681167980488050595E2, -1.47077505154951170175E2,
+                         -5.14105326766599330220E1, -6.05014350600728481186E0};
+    const double QQ[] = {1.00000000000000000000E0, 6.43178256118178023184E1,
+                         8.56430025976980587198E2, 3.88240183605401609683E3,
+                         7.24046774195652478189E3, 5.93072701187316984827E3,
+                         2.06209331660327847417E3, 2.42005740240291393179E2};
+    const double YP[] = {1.55924367855235737965E4, -1.46639295903971606143E7,
+                         5.43526477051876500413E9, -9.82136065717911466409E11,
+                         8.75906394395366999549E13, -3.46628303384729719441E15,
+                         4.42733268572569800351E16, -1.84950800436986690637E16};
+    const double YQ[] = {1.00000000000000000000E0,  1.04128353664259848412E3,
+                         6.26107330137134956842E5, 2.68919633393814121987E8,
+                         8.64002487103935000337E10, 2.02979612750105546709E13,
+                         3.17157752842975028269E15, 2.50596256172653059228E17};
+    const T SQ2OPI = pset1<T>(7.9788456080286535587989E-1); /* sqrt(2 / pi) */
+    const T TWOOPI =  pset1<T>(0.636619772367581343075535); /* 2 / pi */
+    const T NEG_PIO4 = pset1<T>(-0.7853981633974483096); /* -pi / 4 */
+    const T NEG_MAXNUM = pset1<T>(-NumTraits<double>::infinity());
+
+    T z = pmul(x, x);
+    T x_le_five = pdiv(internal::ppolevl<T, 7>::run(z, YP),
+                       internal::ppolevl<T, 7>::run(z, YQ));
+    x_le_five = pmadd(
+        pmul(TWOOPI, plog(x)), generic_j0<T, double>::run(x), x_le_five);
+    x_le_five = pselect(pcmp_le(x, pset1<T>(0.0)), NEG_MAXNUM, x_le_five);
+    T s = pdiv(pset1<T>(25.0), z);
+    T p = pdiv(
+        internal::ppolevl<T, 6>::run(s, PP),
+        internal::ppolevl<T, 6>::run(s, PQ));
+    T q = pdiv(
+        internal::ppolevl<T, 7>::run(s, QP),
+        internal::ppolevl<T, 7>::run(s, QQ));
+    T xn = padd(x, NEG_PIO4);
+    T w = pdiv(pset1<T>(5.0), x);
+    p = pmadd(p, psin(xn), pmul(w, pmul(q, pcos(xn))));
+    T x_gt_five = pmul(p, pmul(SQ2OPI, prsqrt(x)));
+    return pselect(pcmp_le(x, pset1<T>(5.0)), x_le_five, x_gt_five);
+  }
+};
+
+template <typename T>
+struct bessel_y0_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_y0<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_j1_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_j1 {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_j1<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /* j1f.c
+     *	Bessel function of order one
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float x, y, j1f();
+     *
+     * y = j1f( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns Bessel function of order one of the argument.
+     *
+     * The domain is divided into the intervals [0, 2] and
+     * (2, infinity). In the first interval a polynomial approximation
+     *        2
+     * (w - r  ) x P(w)
+     *       1
+     *                     2
+     * is used, where w = x  and r is the first zero of the function.
+     *
+     * In the second interval, the modulus and phase are approximated
+     * by polynomials of the form Modulus(x) = sqrt(1/x) Q(1/x)
+     * and Phase(x) = x + 1/x R(1/x^2) - 3pi/4.  The function is
+     *
+     *   j0(x) = Modulus(x) cos( Phase(x) ).
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Absolute error:
+     * arithmetic   domain      # trials      peak       rms
+     *    IEEE      0,  2       100000       1.2e-7     2.5e-8
+     *    IEEE      2, 32       100000       2.0e-7     5.3e-8
+     *
+     *
+     */
+
+    const float JP[] = {-4.878788132172128E-009f, 6.009061827883699E-007f,
+                        -4.541343896997497E-005f, 1.937383947804541E-003f,
+                        -3.405537384615824E-002f};
+    const float MO1[] = {6.913942741265801E-002f, -2.284801500053359E-001f,
+                        3.138238455499697E-001f, -2.102302420403875E-001f,
+                        5.435364690523026E-003f, 1.493389585089498E-001f,
+                        4.976029650847191E-006f, 7.978845453073848E-001f};
+    const float PH1[] = {-4.497014141919556E+001f, 5.073465654089319E+001f,
+                        -2.485774108720340E+001f, 7.222973196770240E+000f,
+                        -1.544842782180211E+000f, 3.503787691653334E-001f,
+                        -1.637986776941202E-001f, 3.749989509080821E-001f};
+    const T Z1 = pset1<T>(1.46819706421238932572E1f);
+    const T NEG_THPIO4F = pset1<T>(-2.35619449019234492885f);    /* -3*pi/4 */
+
+    T y = pabs(x);
+    T z = pmul(y, y);
+    T y_le_two = pmul(
+        psub(z, Z1),
+        pmul(x, internal::ppolevl<T, 4>::run(z, JP)));
+    T q = pdiv(pset1<T>(1.0f), y);
+    T w = prsqrt(y);
+    T p = pmul(w, internal::ppolevl<T, 7>::run(q, MO1));
+    w = pmul(q, q);
+    T yn = pmadd(q, internal::ppolevl<T, 7>::run(w, PH1), NEG_THPIO4F);
+    T y_gt_two = pmul(p, pcos(padd(yn, y)));
+    // j1 is an odd function. This implementation differs from cephes to
+    // take this fact in to account. Cephes returns -j1(x) for y > 2 range.
+    y_gt_two = pselect(
+        pcmp_lt(x, pset1<T>(0.0f)), pnegate(y_gt_two), y_gt_two);
+    return pselect(pcmp_le(y, pset1<T>(2.0f)), y_le_two, y_gt_two);
+  }
+};
+
+template <typename T>
+struct generic_j1<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  j1.c
+     *	Bessel function of order one
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, j1();
+     *
+     * y = j1( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns Bessel function of order one of the argument.
+     *
+     * The domain is divided into the intervals [0, 8] and
+     * (8, infinity). In the first interval a 24 term Chebyshev
+     * expansion is used. In the second, the asymptotic
+     * trigonometric representation is employed using two
+     * rational functions of degree 5/5.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Absolute error:
+     * arithmetic   domain      # trials      peak         rms
+     *    DEC       0, 30       10000       4.0e-17     1.1e-17
+     *    IEEE      0, 30       30000       2.6e-16     1.1e-16
+     *
+     */
+    const double PP[] = {7.62125616208173112003E-4, 7.31397056940917570436E-2,
+                         1.12719608129684925192E0, 5.11207951146807644818E0,
+                         8.42404590141772420927E0, 5.21451598682361504063E0,
+                         1.00000000000000000254E0};
+    const double PQ[] = {5.71323128072548699714E-4, 6.88455908754495404082E-2,
+                         1.10514232634061696926E0, 5.07386386128601488557E0,
+                         8.39985554327604159757E0, 5.20982848682361821619E0,
+                         9.99999999999999997461E-1};
+    const double QP[] = {5.10862594750176621635E-2, 4.98213872951233449420E0,
+                         7.58238284132545283818E1, 3.66779609360150777800E2,
+                         7.10856304998926107277E2, 5.97489612400613639965E2,
+                         2.11688757100572135698E2, 2.52070205858023719784E1};
+    const double QQ[] = {1.00000000000000000000E0, 7.42373277035675149943E1,
+                         1.05644886038262816351E3, 4.98641058337653607651E3,
+                         9.56231892404756170795E3, 7.99704160447350683650E3,
+                         2.82619278517639096600E3, 3.36093607810698293419E2};
+    const double RP[] = {-8.99971225705559398224E8, 4.52228297998194034323E11,
+                         -7.27494245221818276015E13, 3.68295732863852883286E15};
+    const double RQ[] = {1.00000000000000000000E0, 6.20836478118054335476E2,
+                         2.56987256757748830383E5, 8.35146791431949253037E7,
+                         2.21511595479792499675E10, 4.74914122079991414898E12,
+                         7.84369607876235854894E14, 8.95222336184627338078E16,
+                         5.32278620332680085395E18};
+    const T Z1 = pset1<T>(1.46819706421238932572E1);
+    const T Z2 = pset1<T>(4.92184563216946036703E1);
+    const T NEG_THPIO4 = pset1<T>(-2.35619449019234492885);    /* -3*pi/4 */
+    const T SQ2OPI = pset1<T>(7.9788456080286535587989E-1); /* sqrt(2 / pi) */
+    T y = pabs(x);
+    T z = pmul(y, y);
+    T y_le_five = pdiv(internal::ppolevl<T, 3>::run(z, RP),
+                       internal::ppolevl<T, 8>::run(z, RQ));
+    y_le_five = pmul(pmul(pmul(y_le_five, x), psub(z, Z1)), psub(z, Z2));
+    T s = pdiv(pset1<T>(25.0), z);
+    T p = pdiv(
+        internal::ppolevl<T, 6>::run(s, PP),
+        internal::ppolevl<T, 6>::run(s, PQ));
+    T q = pdiv(
+        internal::ppolevl<T, 7>::run(s, QP),
+        internal::ppolevl<T, 7>::run(s, QQ));
+    T yn = padd(y, NEG_THPIO4);
+    T w = pdiv(pset1<T>(-5.0), y);
+    p = pmadd(p, pcos(yn), pmul(w, pmul(q, psin(yn))));
+    T y_gt_five = pmul(p, pmul(SQ2OPI, prsqrt(y)));
+    // j1 is an odd function. This implementation differs from cephes to
+    // take this fact in to account. Cephes returns -j1(x) for y > 5 range.
+    y_gt_five = pselect(
+        pcmp_lt(x, pset1<T>(0.0)), pnegate(y_gt_five), y_gt_five);
+    return pselect(pcmp_le(y, pset1<T>(5.0)), y_le_five, y_gt_five);
+  }
+};
+
+template <typename T>
+struct bessel_j1_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_j1<T>::run(x);
+  }
+};
+
+template <typename T>
+struct bessel_y1_retval {
+  typedef T type;
+};
+
+template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
+struct generic_y1 {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T&) {
+    EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return ScalarType(0);
+  }
+};
+
+template <typename T>
+struct generic_y1<T, float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /* j1f.c
+     *	Bessel function of second kind of order one
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, y1();
+     *
+     * y = y1( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns Bessel function of the second kind of order one
+     * of the argument.
+     *
+     * The domain is divided into the intervals [0, 2] and
+     * (2, infinity). In the first interval a rational approximation
+     * R(x) is employed to compute
+     *
+     *                  2
+     * y0(x)  =  (w - r  ) x R(x^2)  +  2/pi (ln(x) j1(x) - 1/x) .
+     *                 1
+     *
+     * Thus a call to j1() is required.
+     *
+     * In the second interval, the modulus and phase are approximated
+     * by polynomials of the form Modulus(x) = sqrt(1/x) Q(1/x)
+     * and Phase(x) = x + 1/x S(1/x^2) - 3pi/4.  Then the function is
+     *
+     *   y0(x) = Modulus(x) sin( Phase(x) ).
+     *
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Absolute error:
+     * arithmetic   domain      # trials      peak         rms
+     *    IEEE      0,  2       100000       2.2e-7     4.6e-8
+     *    IEEE      2, 32       100000       1.9e-7     5.3e-8
+     *
+     * (error criterion relative when |y1| > 1).
+     *
+     */
+
+    const float YP[] = {8.061978323326852E-009f, -9.496460629917016E-007f,
+                        6.719543806674249E-005f, -2.641785726447862E-003f,
+                        4.202369946500099E-002f};
+    const float MO1[] = {6.913942741265801E-002f, -2.284801500053359E-001f,
+                        3.138238455499697E-001f, -2.102302420403875E-001f,
+                        5.435364690523026E-003f, 1.493389585089498E-001f,
+                        4.976029650847191E-006f, 7.978845453073848E-001f};
+    const float PH1[] = {-4.497014141919556E+001f, 5.073465654089319E+001f,
+                        -2.485774108720340E+001f, 7.222973196770240E+000f,
+                        -1.544842782180211E+000f, 3.503787691653334E-001f,
+                        -1.637986776941202E-001f, 3.749989509080821E-001f};
+    const T YO1 = pset1<T>(4.66539330185668857532f);
+    const T NEG_THPIO4F = pset1<T>(-2.35619449019234492885f);    /* -3*pi/4 */
+    const T TWOOPI = pset1<T>(0.636619772367581343075535f); /* 2/pi */
+    const T NEG_MAXNUM = pset1<T>(-NumTraits<float>::infinity());
+
+    T z = pmul(x, x);
+    T x_le_two = pmul(psub(z, YO1), internal::ppolevl<T, 4>::run(z, YP));
+    x_le_two = pmadd(
+       x_le_two, x,
+       pmul(TWOOPI, pmadd(
+           generic_j1<T, float>::run(x), plog(x),
+           pdiv(pset1<T>(-1.0f), x))));
+    x_le_two = pselect(pcmp_lt(x, pset1<T>(0.0f)), NEG_MAXNUM, x_le_two);
+
+    T q = pdiv(pset1<T>(1.0), x);
+    T w = prsqrt(x);
+    T p = pmul(w, internal::ppolevl<T, 7>::run(q, MO1));
+    w = pmul(q, q);
+    T xn = pmadd(q, internal::ppolevl<T, 7>::run(w, PH1), NEG_THPIO4F);
+    T x_gt_two = pmul(p, psin(padd(xn, x)));
+    return pselect(pcmp_le(x, pset1<T>(2.0)), x_le_two, x_gt_two);
+  }
+};
+
+template <typename T>
+struct generic_y1<T, double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    /*  j1.c
+     *	Bessel function of second kind of order one
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, y1();
+     *
+     * y = y1( x );
+     *
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns Bessel function of the second kind of order one
+     * of the argument.
+     *
+     * The domain is divided into the intervals [0, 8] and
+     * (8, infinity). In the first interval a 25 term Chebyshev
+     * expansion is used, and a call to j1() is required.
+     * In the second, the asymptotic trigonometric representation
+     * is employed using two rational functions of degree 5/5.
+     *
+     *
+     *
+     * ACCURACY:
+     *
+     *                      Absolute error:
+     * arithmetic   domain      # trials      peak         rms
+     *    DEC       0, 30       10000       8.6e-17     1.3e-17
+     *    IEEE      0, 30       30000       1.0e-15     1.3e-16
+     *
+     * (error criterion relative when |y1| > 1).
+     *
+     */
+    const double PP[] = {7.62125616208173112003E-4, 7.31397056940917570436E-2,
+                         1.12719608129684925192E0, 5.11207951146807644818E0,
+                         8.42404590141772420927E0, 5.21451598682361504063E0,
+                         1.00000000000000000254E0};
+    const double PQ[] = {5.71323128072548699714E-4, 6.88455908754495404082E-2,
+                         1.10514232634061696926E0, 5.07386386128601488557E0,
+                         8.39985554327604159757E0, 5.20982848682361821619E0,
+                         9.99999999999999997461E-1};
+    const double QP[] = {5.10862594750176621635E-2, 4.98213872951233449420E0,
+                         7.58238284132545283818E1, 3.66779609360150777800E2,
+                         7.10856304998926107277E2, 5.97489612400613639965E2,
+                         2.11688757100572135698E2, 2.52070205858023719784E1};
+    const double QQ[] = {1.00000000000000000000E0, 7.42373277035675149943E1,
+                         1.05644886038262816351E3, 4.98641058337653607651E3,
+                         9.56231892404756170795E3, 7.99704160447350683650E3,
+                         2.82619278517639096600E3, 3.36093607810698293419E2};
+    const double YP[] = {1.26320474790178026440E9, -6.47355876379160291031E11,
+                         1.14509511541823727583E14, -8.12770255501325109621E15,
+                         2.02439475713594898196E17, -7.78877196265950026825E17};
+    const double YQ[] = {1.00000000000000000000E0, 5.94301592346128195359E2,
+                         2.35564092943068577943E5, 7.34811944459721705660E7,
+                         1.87601316108706159478E10, 3.88231277496238566008E12,
+                         6.20557727146953693363E14, 6.87141087355300489866E16,
+                         3.97270608116560655612E18};
+    const T SQ2OPI = pset1<T>(.79788456080286535588);
+    const T NEG_THPIO4 = pset1<T>(-2.35619449019234492885);    /* -3*pi/4 */
+    const T TWOOPI = pset1<T>(0.636619772367581343075535); /* 2/pi */
+    const T NEG_MAXNUM = pset1<T>(-NumTraits<double>::infinity());
+
+    T z = pmul(x, x);
+    T x_le_five = pdiv(internal::ppolevl<T, 5>::run(z, YP),
+                   internal::ppolevl<T, 8>::run(z, YQ));
+    x_le_five = pmadd(
+        x_le_five, x, pmul(
+            TWOOPI, pmadd(generic_j1<T, double>::run(x), plog(x),
+                          pdiv(pset1<T>(-1.0), x))));
+
+    x_le_five = pselect(pcmp_le(x, pset1<T>(0.0)), NEG_MAXNUM, x_le_five);
+    T s = pdiv(pset1<T>(25.0), z);
+    T p = pdiv(
+        internal::ppolevl<T, 6>::run(s, PP),
+        internal::ppolevl<T, 6>::run(s, PQ));
+    T q = pdiv(
+        internal::ppolevl<T, 7>::run(s, QP),
+        internal::ppolevl<T, 7>::run(s, QQ));
+    T xn = padd(x, NEG_THPIO4);
+    T w = pdiv(pset1<T>(5.0), x);
+    p = pmadd(p, psin(xn), pmul(w, pmul(q, pcos(xn))));
+    T x_gt_five = pmul(p, pmul(SQ2OPI, prsqrt(x)));
+    return pselect(pcmp_le(x, pset1<T>(5.0)), x_le_five, x_gt_five);
+  }
+};
+
+template <typename T>
+struct bessel_y1_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T x) {
+    return generic_y1<T>::run(x);
+  }
+};
+
+}  // end namespace internal
+
+namespace numext {
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_i0, Scalar)
+    bessel_i0(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_i0, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_i0e, Scalar)
+    bessel_i0e(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_i0e, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_i1, Scalar)
+    bessel_i1(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_i1, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_i1e, Scalar)
+    bessel_i1e(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_i1e, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_k0, Scalar)
+    bessel_k0(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_k0, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_k0e, Scalar)
+    bessel_k0e(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_k0e, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_k1, Scalar)
+    bessel_k1(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_k1, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_k1e, Scalar)
+    bessel_k1e(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_k1e, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_j0, Scalar)
+    bessel_j0(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_j0, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_y0, Scalar)
+    bessel_y0(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_y0, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_j1, Scalar)
+    bessel_j1(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_j1, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_y1, Scalar)
+    bessel_y1(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(bessel_y1, Scalar)::run(x);
+}
+
+}  // end namespace numext
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_BESSEL_FUNCTIONS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsPacketMath.h b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsPacketMath.h
new file mode 100644
index 0000000..943d10f
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/BesselFunctionsPacketMath.h
@@ -0,0 +1,118 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BESSELFUNCTIONS_PACKETMATH_H
+#define EIGEN_BESSELFUNCTIONS_PACKETMATH_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order zero i0(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_i0(const Packet& x) {
+  return numext::bessel_i0(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order zero i0e(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_i0e(const Packet& x) {
+  return numext::bessel_i0e(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order one i1(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_i1(const Packet& x) {
+  return numext::bessel_i1(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order one i1e(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_i1e(const Packet& x) {
+  return numext::bessel_i1e(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order zero j0(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_j0(const Packet& x) {
+  return numext::bessel_j0(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order zero j1(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_j1(const Packet& x) {
+  return numext::bessel_j1(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order one y0(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_y0(const Packet& x) {
+  return numext::bessel_y0(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order one y1(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_y1(const Packet& x) {
+  return numext::bessel_y1(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order zero k0(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_k0(const Packet& x) {
+  return numext::bessel_k0(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order zero k0e(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_k0e(const Packet& x) {
+  return numext::bessel_k0e(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order one k1e(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_k1(const Packet& x) {
+  return numext::bessel_k1(x);
+}
+
+/** \internal \returns the exponentially scaled modified Bessel function of
+ * order one k1e(\a a) (coeff-wise) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pbessel_k1e(const Packet& x) {
+  return numext::bessel_k1e(x);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BESSELFUNCTIONS_PACKETMATH_H
+
diff --git a/src/EigenUnsupported/src/SpecialFunctions/HipVectorCompatibility.h b/src/EigenUnsupported/src/SpecialFunctions/HipVectorCompatibility.h
new file mode 100644
index 0000000..d7b231a
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/HipVectorCompatibility.h
@@ -0,0 +1,67 @@
+#ifndef HIP_VECTOR_COMPATIBILITY_H
+#define HIP_VECTOR_COMPATIBILITY_H
+
+namespace hip_impl {
+  template <typename, typename, unsigned int> struct Scalar_accessor;
+}   // end namespace hip_impl
+
+namespace Eigen {
+namespace internal {
+
+#define HIP_SCALAR_ACCESSOR_BUILDER(NAME) \
+template <typename T, typename U, unsigned int n> \
+struct NAME <hip_impl::Scalar_accessor<T, U, n>> : NAME <T> {};
+
+#define HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(NAME) \
+template <typename T, typename U, unsigned int n> \
+struct NAME##_impl <hip_impl::Scalar_accessor<T, U, n>> : NAME##_impl <T> {}; \
+template <typename T, typename U, unsigned int n> \
+struct NAME##_retval <hip_impl::Scalar_accessor<T, U, n>> : NAME##_retval <T> {};
+
+#define HIP_SCALAR_ACCESSOR_BUILDER_IGAMMA(NAME) \
+template <typename T, typename U, unsigned int n, IgammaComputationMode mode> \
+struct NAME <hip_impl::Scalar_accessor<T, U, n>, mode> : NAME <T, mode> {};
+
+#if EIGEN_HAS_C99_MATH
+HIP_SCALAR_ACCESSOR_BUILDER(betainc_helper)
+HIP_SCALAR_ACCESSOR_BUILDER(incbeta_cfe)
+
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(erf)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(erfc)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(igammac)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(lgamma)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(ndtri)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(polygamma)
+
+HIP_SCALAR_ACCESSOR_BUILDER_IGAMMA(igamma_generic_impl)
+#endif
+
+HIP_SCALAR_ACCESSOR_BUILDER(digamma_impl_maybe_poly)
+HIP_SCALAR_ACCESSOR_BUILDER(zeta_impl_series)
+
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_i0)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_i0e)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_i1)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_i1e)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_j0)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_j1)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_k0)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_k0e)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_k1)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_k1e)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_y0)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(bessel_y1)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(betainc)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(digamma)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(gamma_sample_der_alpha)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(igamma_der_a)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(igamma)
+HIP_SCALAR_ACCESSOR_BUILDER_RETVAL(zeta)
+
+HIP_SCALAR_ACCESSOR_BUILDER_IGAMMA(igamma_series_impl)
+HIP_SCALAR_ACCESSOR_BUILDER_IGAMMA(igammac_cf_impl)
+
+}  // end namespace internal
+}  // end namespace Eigen
+
+#endif  // HIP_VECTOR_COMPATIBILITY_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsArrayAPI.h b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
new file mode 100644
index 0000000..691ff4d
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsArrayAPI.h
@@ -0,0 +1,167 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_SPECIALFUNCTIONS_ARRAYAPI_H
+#define EIGEN_SPECIALFUNCTIONS_ARRAYAPI_H
+
+namespace Eigen {
+
+/** \cpp11 \returns an expression of the coefficient-wise igamma(\a a, \a x) to the given arrays.
+  *
+  * This function computes the coefficient-wise incomplete gamma function.
+  *
+  * \note This function supports only float and double scalar types in c++11 mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations of igammac(T,T) for any scalar
+  * type T to be supported.
+  *
+  * \sa Eigen::igammac(), Eigen::lgamma()
+  */
+template<typename Derived,typename ExponentDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
+igamma(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerived>& x)
+{
+  return Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_op<typename Derived::Scalar>, const Derived, const ExponentDerived>(
+    a.derived(),
+    x.derived()
+  );
+}
+
+/** \cpp11 \returns an expression of the coefficient-wise igamma_der_a(\a a, \a x) to the given arrays.
+  *
+  * This function computes the coefficient-wise derivative of the incomplete
+  * gamma function with respect to the parameter a.
+  *
+  * \note This function supports only float and double scalar types in c++11
+  * mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations
+  * of igamma_der_a(T,T) for any scalar
+  * type T to be supported.
+  *
+  * \sa Eigen::igamma(), Eigen::lgamma()
+  */
+template <typename Derived, typename ExponentDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_der_a_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
+igamma_der_a(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerived>& x) {
+  return Eigen::CwiseBinaryOp<Eigen::internal::scalar_igamma_der_a_op<typename Derived::Scalar>, const Derived, const ExponentDerived>(
+    a.derived(),
+    x.derived());
+}
+
+/** \cpp11 \returns an expression of the coefficient-wise gamma_sample_der_alpha(\a alpha, \a sample) to the given arrays.
+  *
+  * This function computes the coefficient-wise derivative of the sample
+  * of a Gamma(alpha, 1) random variable with respect to the parameter alpha.
+  *
+  * \note This function supports only float and double scalar types in c++11
+  * mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations
+  * of gamma_sample_der_alpha(T,T) for any scalar
+  * type T to be supported.
+  *
+  * \sa Eigen::igamma(), Eigen::lgamma()
+  */
+template <typename AlphaDerived, typename SampleDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_gamma_sample_der_alpha_op<typename AlphaDerived::Scalar>, const AlphaDerived, const SampleDerived>
+gamma_sample_der_alpha(const Eigen::ArrayBase<AlphaDerived>& alpha, const Eigen::ArrayBase<SampleDerived>& sample) {
+  return Eigen::CwiseBinaryOp<Eigen::internal::scalar_gamma_sample_der_alpha_op<typename AlphaDerived::Scalar>, const AlphaDerived, const SampleDerived>(
+      alpha.derived(),
+      sample.derived());
+}
+
+/** \cpp11 \returns an expression of the coefficient-wise igammac(\a a, \a x) to the given arrays.
+  *
+  * This function computes the coefficient-wise complementary incomplete gamma function.
+  *
+  * \note This function supports only float and double scalar types in c++11 mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations of igammac(T,T) for any scalar
+  * type T to be supported.
+  *
+  * \sa Eigen::igamma(), Eigen::lgamma()
+  */
+template<typename Derived,typename ExponentDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_igammac_op<typename Derived::Scalar>, const Derived, const ExponentDerived>
+igammac(const Eigen::ArrayBase<Derived>& a, const Eigen::ArrayBase<ExponentDerived>& x)
+{
+  return Eigen::CwiseBinaryOp<Eigen::internal::scalar_igammac_op<typename Derived::Scalar>, const Derived, const ExponentDerived>(
+    a.derived(),
+    x.derived()
+  );
+}
+
+/** \cpp11 \returns an expression of the coefficient-wise polygamma(\a n, \a x) to the given arrays.
+  *
+  * It returns the \a n -th derivative of the digamma(psi) evaluated at \c x.
+  *
+  * \note This function supports only float and double scalar types in c++11 mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations of polygamma(T,T) for any scalar
+  * type T to be supported.
+  *
+  * \sa Eigen::digamma()
+  */
+// * \warning Be careful with the order of the parameters: x.polygamma(n) is equivalent to polygamma(n,x)
+// * \sa ArrayBase::polygamma()
+template<typename DerivedN,typename DerivedX>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_polygamma_op<typename DerivedX::Scalar>, const DerivedN, const DerivedX>
+polygamma(const Eigen::ArrayBase<DerivedN>& n, const Eigen::ArrayBase<DerivedX>& x)
+{
+  return Eigen::CwiseBinaryOp<Eigen::internal::scalar_polygamma_op<typename DerivedX::Scalar>, const DerivedN, const DerivedX>(
+    n.derived(),
+    x.derived()
+  );
+}
+
+/** \cpp11 \returns an expression of the coefficient-wise betainc(\a x, \a a, \a b) to the given arrays.
+  *
+  * This function computes the regularized incomplete beta function (integral).
+  *
+  * \note This function supports only float and double scalar types in c++11 mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations of betainc(T,T,T) for any scalar
+  * type T to be supported.
+  *
+  * \sa Eigen::betainc(), Eigen::lgamma()
+  */
+template<typename ArgADerived, typename ArgBDerived, typename ArgXDerived>
+EIGEN_STRONG_INLINE const Eigen::CwiseTernaryOp<Eigen::internal::scalar_betainc_op<typename ArgXDerived::Scalar>, const ArgADerived, const ArgBDerived, const ArgXDerived>
+betainc(const Eigen::ArrayBase<ArgADerived>& a, const Eigen::ArrayBase<ArgBDerived>& b, const Eigen::ArrayBase<ArgXDerived>& x)
+{
+  return Eigen::CwiseTernaryOp<Eigen::internal::scalar_betainc_op<typename ArgXDerived::Scalar>, const ArgADerived, const ArgBDerived, const ArgXDerived>(
+    a.derived(),
+    b.derived(),
+    x.derived()
+  );
+}
+
+
+/** \returns an expression of the coefficient-wise zeta(\a x, \a q) to the given arrays.
+  *
+  * It returns the Riemann zeta function of two arguments \a x and \a q:
+  *
+  * \param x is the exponent, it must be > 1
+  * \param q is the shift, it must be > 0
+  *
+  * \note This function supports only float and double scalar types. To support other scalar types, the user has
+  * to provide implementations of zeta(T,T) for any scalar type T to be supported.
+  *
+  * \sa ArrayBase::zeta()
+  */
+template<typename DerivedX,typename DerivedQ>
+EIGEN_STRONG_INLINE const Eigen::CwiseBinaryOp<Eigen::internal::scalar_zeta_op<typename DerivedX::Scalar>, const DerivedX, const DerivedQ>
+zeta(const Eigen::ArrayBase<DerivedX>& x, const Eigen::ArrayBase<DerivedQ>& q)
+{
+  return Eigen::CwiseBinaryOp<Eigen::internal::scalar_zeta_op<typename DerivedX::Scalar>, const DerivedX, const DerivedQ>(
+    x.derived(),
+    q.derived()
+  );
+}
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPECIALFUNCTIONS_ARRAYAPI_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsBFloat16.h b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsBFloat16.h
new file mode 100644
index 0000000..2d94231
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsBFloat16.h
@@ -0,0 +1,58 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPECIALFUNCTIONS_BFLOAT16_H
+#define EIGEN_SPECIALFUNCTIONS_BFLOAT16_H
+
+namespace Eigen {
+namespace numext {
+
+#if EIGEN_HAS_C99_MATH
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 lgamma(const Eigen::bfloat16& a) {
+  return Eigen::bfloat16(Eigen::numext::lgamma(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 digamma(const Eigen::bfloat16& a) {
+  return Eigen::bfloat16(Eigen::numext::digamma(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 zeta(const Eigen::bfloat16& x, const Eigen::bfloat16& q) {
+  return Eigen::bfloat16(Eigen::numext::zeta(static_cast<float>(x), static_cast<float>(q)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 polygamma(const Eigen::bfloat16& n, const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::polygamma(static_cast<float>(n), static_cast<float>(x)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 erf(const Eigen::bfloat16& a) {
+  return Eigen::bfloat16(Eigen::numext::erf(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 erfc(const Eigen::bfloat16& a) {
+  return Eigen::bfloat16(Eigen::numext::erfc(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 ndtri(const Eigen::bfloat16& a) {
+  return Eigen::bfloat16(Eigen::numext::ndtri(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 igamma(const Eigen::bfloat16& a, const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::igamma(static_cast<float>(a), static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 igamma_der_a(const Eigen::bfloat16& a, const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::igamma_der_a(static_cast<float>(a), static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 gamma_sample_der_alpha(const Eigen::bfloat16& alpha, const Eigen::bfloat16& sample) {
+  return Eigen::bfloat16(Eigen::numext::gamma_sample_der_alpha(static_cast<float>(alpha), static_cast<float>(sample)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 igammac(const Eigen::bfloat16& a, const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::igammac(static_cast<float>(a), static_cast<float>(x)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 betainc(const Eigen::bfloat16& a, const Eigen::bfloat16& b, const Eigen::bfloat16& x) {
+  return Eigen::bfloat16(Eigen::numext::betainc(static_cast<float>(a), static_cast<float>(b), static_cast<float>(x)));
+}
+#endif
+
+}  // end namespace numext
+}  // end namespace Eigen
+
+#endif  // EIGEN_SPECIALFUNCTIONS_BFLOAT16_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsFunctors.h b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsFunctors.h
new file mode 100644
index 0000000..abefe99
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsFunctors.h
@@ -0,0 +1,330 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com>
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPECIALFUNCTIONS_FUNCTORS_H
+#define EIGEN_SPECIALFUNCTIONS_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+
+/** \internal
+  * \brief Template functor to compute the incomplete gamma function igamma(a, x)
+  *
+  * \sa class CwiseBinaryOp, Cwise::igamma
+  */
+template<typename Scalar> struct scalar_igamma_op : binary_op_base<Scalar,Scalar>
+{
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_igamma_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& x) const {
+    using numext::igamma; return igamma(a, x);
+  }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& x) const {
+    return internal::pigamma(a, x);
+  }
+};
+template<typename Scalar>
+struct functor_traits<scalar_igamma_op<Scalar> > {
+  enum {
+    // Guesstimate
+    Cost = 20 * NumTraits<Scalar>::MulCost + 10 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasIGamma
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the derivative of the incomplete gamma
+  * function igamma_der_a(a, x)
+  *
+  * \sa class CwiseBinaryOp, Cwise::igamma_der_a
+  */
+template <typename Scalar>
+struct scalar_igamma_der_a_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_igamma_der_a_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& a, const Scalar& x) const {
+    using numext::igamma_der_a;
+    return igamma_der_a(a, x);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& x) const {
+    return internal::pigamma_der_a(a, x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_igamma_der_a_op<Scalar> > {
+  enum {
+    // 2x the cost of igamma
+    Cost = 40 * NumTraits<Scalar>::MulCost + 20 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasIGammaDerA
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the derivative of the sample
+  * of a Gamma(alpha, 1) random variable with respect to the parameter alpha
+  * gamma_sample_der_alpha(alpha, sample)
+  *
+  * \sa class CwiseBinaryOp, Cwise::gamma_sample_der_alpha
+  */
+template <typename Scalar>
+struct scalar_gamma_sample_der_alpha_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_gamma_sample_der_alpha_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator()(const Scalar& alpha, const Scalar& sample) const {
+    using numext::gamma_sample_der_alpha;
+    return gamma_sample_der_alpha(alpha, sample);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& alpha, const Packet& sample) const {
+    return internal::pgamma_sample_der_alpha(alpha, sample);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_gamma_sample_der_alpha_op<Scalar> > {
+  enum {
+    // 2x the cost of igamma, minus the lgamma cost (the lgamma cancels out)
+    Cost = 30 * NumTraits<Scalar>::MulCost + 15 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasGammaSampleDerAlpha
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the complementary incomplete gamma function igammac(a, x)
+  *
+  * \sa class CwiseBinaryOp, Cwise::igammac
+  */
+template<typename Scalar> struct scalar_igammac_op : binary_op_base<Scalar,Scalar>
+{
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_igammac_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& x) const {
+    using numext::igammac; return igammac(a, x);
+  }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& x) const
+  {
+    return internal::pigammac(a, x);
+  }
+};
+template<typename Scalar>
+struct functor_traits<scalar_igammac_op<Scalar> > {
+  enum {
+    // Guesstimate
+    Cost = 20 * NumTraits<Scalar>::MulCost + 10 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasIGammac
+  };
+};
+
+
+/** \internal
+  * \brief Template functor to compute the incomplete beta integral betainc(a, b, x)
+  *
+  */
+template<typename Scalar> struct scalar_betainc_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_betainc_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& x, const Scalar& a, const Scalar& b) const {
+    using numext::betainc; return betainc(x, a, b);
+  }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& x, const Packet& a, const Packet& b) const
+  {
+    return internal::pbetainc(x, a, b);
+  }
+};
+template<typename Scalar>
+struct functor_traits<scalar_betainc_op<Scalar> > {
+  enum {
+    // Guesstimate
+    Cost = 400 * NumTraits<Scalar>::MulCost + 400 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasBetaInc
+  };
+};
+
+
+/** \internal
+ * \brief Template functor to compute the natural log of the absolute
+ * value of Gamma of a scalar
+ * \sa class CwiseUnaryOp, Cwise::lgamma()
+ */
+template<typename Scalar> struct scalar_lgamma_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_lgamma_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const {
+    using numext::lgamma; return lgamma(a);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a) const { return internal::plgamma(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_lgamma_op<Scalar> >
+{
+  enum {
+    // Guesstimate
+    Cost = 10 * NumTraits<Scalar>::MulCost + 5 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasLGamma
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute psi, the derivative of lgamma of a scalar.
+ * \sa class CwiseUnaryOp, Cwise::digamma()
+ */
+template<typename Scalar> struct scalar_digamma_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_digamma_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const {
+    using numext::digamma; return digamma(a);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a) const { return internal::pdigamma(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_digamma_op<Scalar> >
+{
+  enum {
+    // Guesstimate
+    Cost = 10 * NumTraits<Scalar>::MulCost + 5 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasDiGamma
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Riemann Zeta function of two arguments.
+ * \sa class CwiseUnaryOp, Cwise::zeta()
+ */
+template<typename Scalar> struct scalar_zeta_op {
+    EIGEN_EMPTY_STRUCT_CTOR(scalar_zeta_op)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& x, const Scalar& q) const {
+        using numext::zeta; return zeta(x, q);
+    }
+    typedef typename packet_traits<Scalar>::type Packet;
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x, const Packet& q) const { return internal::pzeta(x, q); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_zeta_op<Scalar> >
+{
+    enum {
+        // Guesstimate
+        Cost = 10 * NumTraits<Scalar>::MulCost + 5 * NumTraits<Scalar>::AddCost,
+        PacketAccess = packet_traits<Scalar>::HasZeta
+    };
+};
+
+/** \internal
+ * \brief Template functor to compute the polygamma function.
+ * \sa class CwiseUnaryOp, Cwise::polygamma()
+ */
+template<typename Scalar> struct scalar_polygamma_op {
+    EIGEN_EMPTY_STRUCT_CTOR(scalar_polygamma_op)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& n, const Scalar& x) const {
+        using numext::polygamma; return polygamma(n, x);
+    }
+    typedef typename packet_traits<Scalar>::type Packet;
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& n, const Packet& x) const { return internal::ppolygamma(n, x); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_polygamma_op<Scalar> >
+{
+    enum {
+        // Guesstimate
+        Cost = 10 * NumTraits<Scalar>::MulCost + 5 * NumTraits<Scalar>::AddCost,
+        PacketAccess = packet_traits<Scalar>::HasPolygamma
+    };
+};
+
+/** \internal
+ * \brief Template functor to compute the error function of a scalar
+ * \sa class CwiseUnaryOp, ArrayBase::erf()
+ */
+template<typename Scalar> struct scalar_erf_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_erf_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar
+  operator()(const Scalar& a) const {
+    return numext::erf(a);
+  }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& x) const {
+    return perf(x);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_erf_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasErf,
+    Cost =
+        (PacketAccess
+#ifdef EIGEN_VECTORIZE_FMA
+             // TODO(rmlarsen): Move the FMA cost model to a central location.
+             // Haswell can issue 2 add/mul/madd per cycle.
+             // 10 pmadd, 2 pmul, 1 div, 2 other
+             ? (2 * NumTraits<Scalar>::AddCost +
+                7 * NumTraits<Scalar>::MulCost +
+                scalar_div_cost<Scalar, packet_traits<Scalar>::HasDiv>::value)
+#else
+             ? (12 * NumTraits<Scalar>::AddCost +
+                12 * NumTraits<Scalar>::MulCost +
+                scalar_div_cost<Scalar, packet_traits<Scalar>::HasDiv>::value)
+#endif
+             // Assume for simplicity that this is as expensive as an exp().
+             : (functor_traits<scalar_exp_op<Scalar> >::Cost))
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Complementary Error Function
+ * of a scalar
+ * \sa class CwiseUnaryOp, Cwise::erfc()
+ */
+template<typename Scalar> struct scalar_erfc_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_erfc_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const {
+    using numext::erfc; return erfc(a);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a) const { return internal::perfc(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_erfc_op<Scalar> >
+{
+  enum {
+    // Guesstimate
+    Cost = 10 * NumTraits<Scalar>::MulCost + 5 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasErfc
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Inverse of the normal distribution
+ * function of a scalar
+ * \sa class CwiseUnaryOp, Cwise::ndtri()
+ */
+template<typename Scalar> struct scalar_ndtri_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_ndtri_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const {
+    using numext::ndtri; return ndtri(a);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a) const { return internal::pndtri(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_ndtri_op<Scalar> >
+{
+  enum {
+    // On average, We are evaluating rational functions with degree N=9 in the
+    // numerator and denominator. This results in 2*N additions and 2*N
+    // multiplications.
+    Cost = 18 * NumTraits<Scalar>::MulCost + 18 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasNdtri
+  };
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPECIALFUNCTIONS_FUNCTORS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsHalf.h b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsHalf.h
new file mode 100644
index 0000000..2a3a531
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsHalf.h
@@ -0,0 +1,58 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPECIALFUNCTIONS_HALF_H
+#define EIGEN_SPECIALFUNCTIONS_HALF_H
+
+namespace Eigen {
+namespace numext {
+
+#if EIGEN_HAS_C99_MATH
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half lgamma(const Eigen::half& a) {
+  return Eigen::half(Eigen::numext::lgamma(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half digamma(const Eigen::half& a) {
+  return Eigen::half(Eigen::numext::digamma(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half zeta(const Eigen::half& x, const Eigen::half& q) {
+  return Eigen::half(Eigen::numext::zeta(static_cast<float>(x), static_cast<float>(q)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half polygamma(const Eigen::half& n, const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::polygamma(static_cast<float>(n), static_cast<float>(x)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half erf(const Eigen::half& a) {
+  return Eigen::half(Eigen::numext::erf(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half erfc(const Eigen::half& a) {
+  return Eigen::half(Eigen::numext::erfc(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half ndtri(const Eigen::half& a) {
+  return Eigen::half(Eigen::numext::ndtri(static_cast<float>(a)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half igamma(const Eigen::half& a, const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::igamma(static_cast<float>(a), static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half igamma_der_a(const Eigen::half& a, const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::igamma_der_a(static_cast<float>(a), static_cast<float>(x)));
+}
+template <>
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half gamma_sample_der_alpha(const Eigen::half& alpha, const Eigen::half& sample) {
+  return Eigen::half(Eigen::numext::gamma_sample_der_alpha(static_cast<float>(alpha), static_cast<float>(sample)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half igammac(const Eigen::half& a, const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::igammac(static_cast<float>(a), static_cast<float>(x)));
+}
+template<> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half betainc(const Eigen::half& a, const Eigen::half& b, const Eigen::half& x) {
+  return Eigen::half(Eigen::numext::betainc(static_cast<float>(a), static_cast<float>(b), static_cast<float>(x)));
+}
+#endif
+
+}  // end namespace numext
+}  // end namespace Eigen
+
+#endif  // EIGEN_SPECIALFUNCTIONS_HALF_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsImpl.h b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsImpl.h
new file mode 100644
index 0000000..f1c260e
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -0,0 +1,2045 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Eugene Brevdo <ebrevdo@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPECIAL_FUNCTIONS_H
+#define EIGEN_SPECIAL_FUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+//  Parts of this code are based on the Cephes Math Library.
+//
+//  Cephes Math Library Release 2.8:  June, 2000
+//  Copyright 1984, 1987, 1992, 2000 by Stephen L. Moshier
+//
+//  Permission has been kindly provided by the original author
+//  to incorporate the Cephes software into the Eigen codebase:
+//
+//    From: Stephen Moshier
+//    To: Eugene Brevdo
+//    Subject: Re: Permission to wrap several cephes functions in Eigen
+//
+//    Hello Eugene,
+//
+//    Thank you for writing.
+//
+//    If your licensing is similar to BSD, the formal way that has been
+//    handled is simply to add a statement to the effect that you are incorporating
+//    the Cephes software by permission of the author.
+//
+//    Good luck with your project,
+//    Steve
+
+
+/****************************************************************************
+ * Implementation of lgamma, requires C++11/C99                             *
+ ****************************************************************************/
+
+template <typename Scalar>
+struct lgamma_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+template <typename Scalar>
+struct lgamma_retval {
+  typedef Scalar type;
+};
+
+#if EIGEN_HAS_C99_MATH
+// Since glibc 2.19
+#if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 19) || __GLIBC__>2) \
+ && (defined(_DEFAULT_SOURCE) || defined(_BSD_SOURCE) || defined(_SVID_SOURCE))
+#define EIGEN_HAS_LGAMMA_R
+#endif
+
+// Glibc versions before 2.19
+#if defined(__GLIBC__) && ((__GLIBC__==2 && __GLIBC_MINOR__ < 19) || __GLIBC__<2) \
+ && (defined(_BSD_SOURCE) || defined(_SVID_SOURCE))
+#define EIGEN_HAS_LGAMMA_R
+#endif
+
+template <>
+struct lgamma_impl<float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float run(float x) {
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && defined (EIGEN_HAS_LGAMMA_R) && !defined(__APPLE__)
+    int dummy;
+    return ::lgammaf_r(x, &dummy);
+#elif defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::lgamma(x);
+#else
+    return ::lgammaf(x);
+#endif
+  }
+};
+
+template <>
+struct lgamma_impl<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(double x) {
+#if !defined(EIGEN_GPU_COMPILE_PHASE) && defined(EIGEN_HAS_LGAMMA_R) && !defined(__APPLE__)
+    int dummy;
+    return ::lgamma_r(x, &dummy);
+#elif defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::lgamma(x);
+#else
+    return ::lgamma(x);
+#endif
+  }
+};
+
+#undef EIGEN_HAS_LGAMMA_R
+#endif
+
+/****************************************************************************
+ * Implementation of digamma (psi), based on Cephes                         *
+ ****************************************************************************/
+
+template <typename Scalar>
+struct digamma_retval {
+  typedef Scalar type;
+};
+
+/*
+ *
+ * Polynomial evaluation helper for the Psi (digamma) function.
+ *
+ * digamma_impl_maybe_poly::run(s) evaluates the asymptotic Psi expansion for
+ * input Scalar s, assuming s is above 10.0.
+ *
+ * If s is above a certain threshold for the given Scalar type, zero
+ * is returned.  Otherwise the polynomial is evaluated with enough
+ * coefficients for results matching Scalar machine precision.
+ *
+ *
+ */
+template <typename Scalar>
+struct digamma_impl_maybe_poly {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+
+template <>
+struct digamma_impl_maybe_poly<float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float run(const float s) {
+    const float A[] = {
+      -4.16666666666666666667E-3f,
+      3.96825396825396825397E-3f,
+      -8.33333333333333333333E-3f,
+      8.33333333333333333333E-2f
+    };
+
+    float z;
+    if (s < 1.0e8f) {
+      z = 1.0f / (s * s);
+      return z * internal::ppolevl<float, 3>::run(z, A);
+    } else return 0.0f;
+  }
+};
+
+template <>
+struct digamma_impl_maybe_poly<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(const double s) {
+    const double A[] = {
+      8.33333333333333333333E-2,
+      -2.10927960927960927961E-2,
+      7.57575757575757575758E-3,
+      -4.16666666666666666667E-3,
+      3.96825396825396825397E-3,
+      -8.33333333333333333333E-3,
+      8.33333333333333333333E-2
+    };
+
+    double z;
+    if (s < 1.0e17) {
+      z = 1.0 / (s * s);
+      return z * internal::ppolevl<double, 6>::run(z, A);
+    }
+    else return 0.0;
+  }
+};
+
+template <typename Scalar>
+struct digamma_impl {
+  EIGEN_DEVICE_FUNC
+  static Scalar run(Scalar x) {
+    /*
+     *
+     *     Psi (digamma) function (modified for Eigen)
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double x, y, psi();
+     *
+     * y = psi( x );
+     *
+     *
+     * DESCRIPTION:
+     *
+     *              d      -
+     *   psi(x)  =  -- ln | (x)
+     *              dx
+     *
+     * is the logarithmic derivative of the gamma function.
+     * For integer x,
+     *                   n-1
+     *                    -
+     * psi(n) = -EUL  +   >  1/k.
+     *                    -
+     *                   k=1
+     *
+     * If x is negative, it is transformed to a positive argument by the
+     * reflection formula  psi(1-x) = psi(x) + pi cot(pi x).
+     * For general positive x, the argument is made greater than 10
+     * using the recurrence  psi(x+1) = psi(x) + 1/x.
+     * Then the following asymptotic expansion is applied:
+     *
+     *                           inf.   B
+     *                            -      2k
+     * psi(x) = log(x) - 1/2x -   >   -------
+     *                            -        2k
+     *                           k=1   2k x
+     *
+     * where the B2k are Bernoulli numbers.
+     *
+     * ACCURACY (float):
+     *    Relative error (except absolute when |psi| < 1):
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0,30        30000       1.3e-15     1.4e-16
+     *    IEEE      -30,0       40000       1.5e-15     2.2e-16
+     *
+     * ACCURACY (double):
+     *    Absolute error,  relative when |psi| > 1 :
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      -33,0        30000      8.2e-7      1.2e-7
+     *    IEEE      0,33        100000      7.3e-7      7.7e-8
+     *
+     * ERROR MESSAGES:
+     *     message         condition      value returned
+     * psi singularity    x integer <=0      INFINITY
+     */
+
+    Scalar p, q, nz, s, w, y;
+    bool negative = false;
+
+    const Scalar nan = NumTraits<Scalar>::quiet_NaN();
+    const Scalar m_pi = Scalar(EIGEN_PI);
+
+    const Scalar zero = Scalar(0);
+    const Scalar one = Scalar(1);
+    const Scalar half = Scalar(0.5);
+    nz = zero;
+
+    if (x <= zero) {
+      negative = true;
+      q = x;
+      p = numext::floor(q);
+      if (p == q) {
+        return nan;
+      }
+      /* Remove the zeros of tan(m_pi x)
+       * by subtracting the nearest integer from x
+       */
+      nz = q - p;
+      if (nz != half) {
+        if (nz > half) {
+          p += one;
+          nz = q - p;
+        }
+        nz = m_pi / numext::tan(m_pi * nz);
+      }
+      else {
+        nz = zero;
+      }
+      x = one - x;
+    }
+
+    /* use the recurrence psi(x+1) = psi(x) + 1/x. */
+    s = x;
+    w = zero;
+    while (s < Scalar(10)) {
+      w += one / s;
+      s += one;
+    }
+
+    y = digamma_impl_maybe_poly<Scalar>::run(s);
+
+    y = numext::log(s) - (half / s) - y - w;
+
+    return (negative) ? y - nz : y;
+  }
+};
+
+/****************************************************************************
+ * Implementation of erf, requires C++11/C99                                *
+ ****************************************************************************/
+
+/** \internal \returns the error function of \a a (coeff-wise)
+    Doesn't do anything fancy, just a 13/8-degree rational interpolant which
+    is accurate up to a couple of ulp in the range [-4, 4], outside of which
+    fl(erf(x)) = +/-1.
+
+    This implementation works on both scalars and Ts.
+*/
+template <typename T>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T generic_fast_erf_float(const T& a_x) {
+  // Clamp the inputs to the range [-4, 4] since anything outside
+  // this range is +/-1.0f in single-precision.
+  const T plus_4 = pset1<T>(4.f);
+  const T minus_4 = pset1<T>(-4.f);
+  const T x = pmax(pmin(a_x, plus_4), minus_4);
+  // The monomial coefficients of the numerator polynomial (odd).
+  const T alpha_1 = pset1<T>(-1.60960333262415e-02f);
+  const T alpha_3 = pset1<T>(-2.95459980854025e-03f);
+  const T alpha_5 = pset1<T>(-7.34990630326855e-04f);
+  const T alpha_7 = pset1<T>(-5.69250639462346e-05f);
+  const T alpha_9 = pset1<T>(-2.10102402082508e-06f);
+  const T alpha_11 = pset1<T>(2.77068142495902e-08f);
+  const T alpha_13 = pset1<T>(-2.72614225801306e-10f);
+
+  // The monomial coefficients of the denominator polynomial (even).
+  const T beta_0 = pset1<T>(-1.42647390514189e-02f);
+  const T beta_2 = pset1<T>(-7.37332916720468e-03f);
+  const T beta_4 = pset1<T>(-1.68282697438203e-03f);
+  const T beta_6 = pset1<T>(-2.13374055278905e-04f);
+  const T beta_8 = pset1<T>(-1.45660718464996e-05f);
+
+  // Since the polynomials are odd/even, we need x^2.
+  const T x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial p.
+  T p = pmadd(x2, alpha_13, alpha_11);
+  p = pmadd(x2, p, alpha_9);
+  p = pmadd(x2, p, alpha_7);
+  p = pmadd(x2, p, alpha_5);
+  p = pmadd(x2, p, alpha_3);
+  p = pmadd(x2, p, alpha_1);
+  p = pmul(x, p);
+
+  // Evaluate the denominator polynomial p.
+  T q = pmadd(x2, beta_8, beta_6);
+  q = pmadd(x2, q, beta_4);
+  q = pmadd(x2, q, beta_2);
+  q = pmadd(x2, q, beta_0);
+
+  // Divide the numerator by the denominator.
+  return pdiv(p, q);
+}
+
+template <typename T>
+struct erf_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE T run(const T& x) {
+    return generic_fast_erf_float(x);
+  }
+};
+
+template <typename Scalar>
+struct erf_retval {
+  typedef Scalar type;
+};
+
+#if EIGEN_HAS_C99_MATH
+template <>
+struct erf_impl<float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float run(float x) {
+#if defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::erf(x);
+#else
+    return generic_fast_erf_float(x);
+#endif
+  }
+};
+
+template <>
+struct erf_impl<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(double x) {
+#if defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::erf(x);
+#else
+    return ::erf(x);
+#endif
+  }
+};
+#endif  // EIGEN_HAS_C99_MATH
+
+/***************************************************************************
+* Implementation of erfc, requires C++11/C99                               *
+****************************************************************************/
+
+template <typename Scalar>
+struct erfc_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+template <typename Scalar>
+struct erfc_retval {
+  typedef Scalar type;
+};
+
+#if EIGEN_HAS_C99_MATH
+template <>
+struct erfc_impl<float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float run(const float x) {
+#if defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::erfc(x);
+#else
+    return ::erfcf(x);
+#endif
+  }
+};
+
+template <>
+struct erfc_impl<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(const double x) {
+#if defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::erfc(x);
+#else
+    return ::erfc(x);
+#endif
+  }
+};
+#endif  // EIGEN_HAS_C99_MATH
+
+
+/***************************************************************************
+* Implementation of ndtri.                                                 *
+****************************************************************************/
+
+/* Inverse of Normal distribution function (modified for Eigen).
+ *
+ *
+ * SYNOPSIS:
+ *
+ * double x, y, ndtri();
+ *
+ * x = ndtri( y );
+ *
+ *
+ *
+ * DESCRIPTION:
+ *
+ * Returns the argument, x, for which the area under the
+ * Gaussian probability density function (integrated from
+ * minus infinity to x) is equal to y.
+ *
+ *
+ * For small arguments 0 < y < exp(-2), the program computes
+ * z = sqrt( -2.0 * log(y) );  then the approximation is
+ * x = z - log(z)/z  - (1/z) P(1/z) / Q(1/z).
+ * There are two rational functions P/Q, one for 0 < y < exp(-32)
+ * and the other for y up to exp(-2).  For larger arguments,
+ * w = y - 0.5, and  x/sqrt(2pi) = w + w**3 R(w**2)/S(w**2)).
+ *
+ *
+ * ACCURACY:
+ *
+ *                      Relative error:
+ * arithmetic   domain        # trials      peak         rms
+ *    DEC      0.125, 1         5500       9.5e-17     2.1e-17
+ *    DEC      6e-39, 0.135     3500       5.7e-17     1.3e-17
+ *    IEEE     0.125, 1        20000       7.2e-16     1.3e-16
+ *    IEEE     3e-308, 0.135   50000       4.6e-16     9.8e-17
+ *
+ *
+ * ERROR MESSAGES:
+ *
+ *   message         condition    value returned
+ * ndtri domain       x <= 0        -MAXNUM
+ * ndtri domain       x >= 1         MAXNUM
+ *
+ */
+ /*
+   Cephes Math Library Release 2.2: June, 1992
+   Copyright 1985, 1987, 1992 by Stephen L. Moshier
+   Direct inquiries to 30 Frost Street, Cambridge, MA 02140
+ */
+
+
+// TODO: Add a cheaper approximation for float.
+
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T flipsign(
+    const T& should_flipsign, const T& x) {
+  typedef typename unpacket_traits<T>::type Scalar;
+  const T sign_mask = pset1<T>(Scalar(-0.0));
+  T sign_bit = pand<T>(should_flipsign, sign_mask);
+  return pxor<T>(sign_bit, x);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double flipsign<double>(
+    const double& should_flipsign, const double& x) {
+  return should_flipsign == 0 ? x : -x;
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float flipsign<float>(
+    const float& should_flipsign, const float& x) {
+  return should_flipsign == 0 ? x : -x;
+}
+
+// We split this computation in to two so that in the scalar path
+// only one branch is evaluated (due to our template specialization of pselect
+// being an if statement.)
+
+template <typename T, typename ScalarType>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T generic_ndtri_gt_exp_neg_two(const T& b) {
+  const ScalarType p0[] = {
+    ScalarType(-5.99633501014107895267e1),
+    ScalarType(9.80010754185999661536e1),
+    ScalarType(-5.66762857469070293439e1),
+    ScalarType(1.39312609387279679503e1),
+    ScalarType(-1.23916583867381258016e0)
+  };
+  const ScalarType q0[] = {
+    ScalarType(1.0),
+    ScalarType(1.95448858338141759834e0),
+    ScalarType(4.67627912898881538453e0),
+    ScalarType(8.63602421390890590575e1),
+    ScalarType(-2.25462687854119370527e2),
+    ScalarType(2.00260212380060660359e2),
+    ScalarType(-8.20372256168333339912e1),
+    ScalarType(1.59056225126211695515e1),
+    ScalarType(-1.18331621121330003142e0)
+  };
+  const T sqrt2pi = pset1<T>(ScalarType(2.50662827463100050242e0));
+  const T half = pset1<T>(ScalarType(0.5));
+  T c, c2, ndtri_gt_exp_neg_two;
+
+  c = psub(b, half);
+  c2 = pmul(c, c);
+  ndtri_gt_exp_neg_two = pmadd(c, pmul(
+      c2, pdiv(
+          internal::ppolevl<T, 4>::run(c2, p0),
+          internal::ppolevl<T, 8>::run(c2, q0))), c);
+  return pmul(ndtri_gt_exp_neg_two, sqrt2pi);
+}
+
+template <typename T, typename ScalarType>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T generic_ndtri_lt_exp_neg_two(
+    const T& b, const T& should_flipsign) {
+  /* Approximation for interval z = sqrt(-2 log a ) between 2 and 8
+   * i.e., a between exp(-2) = .135 and exp(-32) = 1.27e-14.
+   */
+  const ScalarType p1[] = {
+    ScalarType(4.05544892305962419923e0),
+    ScalarType(3.15251094599893866154e1),
+    ScalarType(5.71628192246421288162e1),
+    ScalarType(4.40805073893200834700e1),
+    ScalarType(1.46849561928858024014e1),
+    ScalarType(2.18663306850790267539e0),
+    ScalarType(-1.40256079171354495875e-1),
+    ScalarType(-3.50424626827848203418e-2),
+    ScalarType(-8.57456785154685413611e-4)
+  };
+  const ScalarType q1[] = {
+    ScalarType(1.0),
+    ScalarType(1.57799883256466749731e1),
+    ScalarType(4.53907635128879210584e1),
+    ScalarType(4.13172038254672030440e1),
+    ScalarType(1.50425385692907503408e1),
+    ScalarType(2.50464946208309415979e0),
+    ScalarType(-1.42182922854787788574e-1),
+    ScalarType(-3.80806407691578277194e-2),
+    ScalarType(-9.33259480895457427372e-4)
+  };
+  /* Approximation for interval z = sqrt(-2 log a ) between 8 and 64
+   * i.e., a between exp(-32) = 1.27e-14 and exp(-2048) = 3.67e-890.
+   */
+  const ScalarType p2[] = {
+    ScalarType(3.23774891776946035970e0),
+    ScalarType(6.91522889068984211695e0),
+    ScalarType(3.93881025292474443415e0),
+    ScalarType(1.33303460815807542389e0),
+    ScalarType(2.01485389549179081538e-1),
+    ScalarType(1.23716634817820021358e-2),
+    ScalarType(3.01581553508235416007e-4),
+    ScalarType(2.65806974686737550832e-6),
+    ScalarType(6.23974539184983293730e-9)
+  };
+  const ScalarType q2[] = {
+    ScalarType(1.0),
+    ScalarType(6.02427039364742014255e0),
+    ScalarType(3.67983563856160859403e0),
+    ScalarType(1.37702099489081330271e0),
+    ScalarType(2.16236993594496635890e-1),
+    ScalarType(1.34204006088543189037e-2),
+    ScalarType(3.28014464682127739104e-4),
+    ScalarType(2.89247864745380683936e-6),
+    ScalarType(6.79019408009981274425e-9)
+  };
+  const T eight = pset1<T>(ScalarType(8.0));
+  const T one = pset1<T>(ScalarType(1));
+  const T neg_two = pset1<T>(ScalarType(-2));
+  T x, x0, x1, z;
+
+  x = psqrt(pmul(neg_two, plog(b)));
+  x0 = psub(x, pdiv(plog(x), x));
+  z = pdiv(one, x);
+  x1 = pmul(
+      z, pselect(
+          pcmp_lt(x, eight),
+          pdiv(internal::ppolevl<T, 8>::run(z, p1),
+               internal::ppolevl<T, 8>::run(z, q1)),
+          pdiv(internal::ppolevl<T, 8>::run(z, p2),
+               internal::ppolevl<T, 8>::run(z, q2))));
+  return flipsign(should_flipsign, psub(x0, x1));
+}
+
+template <typename T, typename ScalarType>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T generic_ndtri(const T& a) {
+  const T maxnum = pset1<T>(NumTraits<ScalarType>::infinity());
+  const T neg_maxnum = pset1<T>(-NumTraits<ScalarType>::infinity());
+
+  const T zero = pset1<T>(ScalarType(0));
+  const T one = pset1<T>(ScalarType(1));
+  // exp(-2)
+  const T exp_neg_two = pset1<T>(ScalarType(0.13533528323661269189));
+  T b, ndtri, should_flipsign;
+
+  should_flipsign = pcmp_le(a, psub(one, exp_neg_two));
+  b = pselect(should_flipsign, a, psub(one, a));
+
+  ndtri = pselect(
+      pcmp_lt(exp_neg_two, b),
+      generic_ndtri_gt_exp_neg_two<T, ScalarType>(b),
+      generic_ndtri_lt_exp_neg_two<T, ScalarType>(b, should_flipsign));
+
+  return pselect(
+      pcmp_le(a, zero), neg_maxnum,
+      pselect(pcmp_le(one, a), maxnum, ndtri));
+}
+
+template <typename Scalar>
+struct ndtri_retval {
+  typedef Scalar type;
+};
+
+#if !EIGEN_HAS_C99_MATH
+
+template <typename Scalar>
+struct ndtri_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+# else
+
+template <typename Scalar>
+struct ndtri_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar x) {
+    return generic_ndtri<Scalar, Scalar>(x);
+  }
+};
+
+#endif  // EIGEN_HAS_C99_MATH
+
+
+/**************************************************************************************************************
+ * Implementation of igammac (complemented incomplete gamma integral), based on Cephes but requires C++11/C99 *
+ **************************************************************************************************************/
+
+template <typename Scalar>
+struct igammac_retval {
+  typedef Scalar type;
+};
+
+// NOTE: cephes_helper is also used to implement zeta
+template <typename Scalar>
+struct cephes_helper {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar machep() { assert(false && "machep not supported for this type"); return 0.0; }
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar big() { assert(false && "big not supported for this type"); return 0.0; }
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar biginv() { assert(false && "biginv not supported for this type"); return 0.0; }
+};
+
+template <>
+struct cephes_helper<float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float machep() {
+    return NumTraits<float>::epsilon() / 2;  // 1.0 - machep == 1.0
+  }
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float big() {
+    // use epsneg (1.0 - epsneg == 1.0)
+    return 1.0f / (NumTraits<float>::epsilon() / 2);
+  }
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float biginv() {
+    // epsneg
+    return machep();
+  }
+};
+
+template <>
+struct cephes_helper<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double machep() {
+    return NumTraits<double>::epsilon() / 2;  // 1.0 - machep == 1.0
+  }
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double big() {
+    return 1.0 / NumTraits<double>::epsilon();
+  }
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double biginv() {
+    // inverse of eps
+    return NumTraits<double>::epsilon();
+  }
+};
+
+enum IgammaComputationMode { VALUE, DERIVATIVE, SAMPLE_DERIVATIVE };
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC
+static EIGEN_STRONG_INLINE Scalar main_igamma_term(Scalar a, Scalar x) {
+    /* Compute  x**a * exp(-x) / gamma(a)  */
+    Scalar logax = a * numext::log(x) - x - lgamma_impl<Scalar>::run(a);
+    if (logax < -numext::log(NumTraits<Scalar>::highest()) ||
+        // Assuming x and a aren't Nan.
+        (numext::isnan)(logax)) {
+      return Scalar(0);
+    }
+    return numext::exp(logax);
+}
+
+template <typename Scalar, IgammaComputationMode mode>
+EIGEN_DEVICE_FUNC
+int igamma_num_iterations() {
+  /* Returns the maximum number of internal iterations for igamma computation.
+   */
+  if (mode == VALUE) {
+    return 2000;
+  }
+
+  if (internal::is_same<Scalar, float>::value) {
+    return 200;
+  } else if (internal::is_same<Scalar, double>::value) {
+    return 500;
+  } else {
+    return 2000;
+  }
+}
+
+template <typename Scalar, IgammaComputationMode mode>
+struct igammac_cf_impl {
+  /* Computes igamc(a, x) or derivative (depending on the mode)
+   * using the continued fraction expansion of the complementary
+   * incomplete Gamma function.
+   *
+   * Preconditions:
+   *   a > 0
+   *   x >= 1
+   *   x >= a
+   */
+  EIGEN_DEVICE_FUNC
+  static Scalar run(Scalar a, Scalar x) {
+    const Scalar zero = 0;
+    const Scalar one = 1;
+    const Scalar two = 2;
+    const Scalar machep = cephes_helper<Scalar>::machep();
+    const Scalar big = cephes_helper<Scalar>::big();
+    const Scalar biginv = cephes_helper<Scalar>::biginv();
+
+    if ((numext::isinf)(x)) {
+      return zero;
+    }
+
+    Scalar ax = main_igamma_term<Scalar>(a, x);
+    // This is independent of mode. If this value is zero,
+    // then the function value is zero. If the function value is zero,
+    // then we are in a neighborhood where the function value evalutes to zero,
+    // so the derivative is zero.
+    if (ax == zero) {
+      return zero;
+    }
+
+    // continued fraction
+    Scalar y = one - a;
+    Scalar z = x + y + one;
+    Scalar c = zero;
+    Scalar pkm2 = one;
+    Scalar qkm2 = x;
+    Scalar pkm1 = x + one;
+    Scalar qkm1 = z * x;
+    Scalar ans = pkm1 / qkm1;
+
+    Scalar dpkm2_da = zero;
+    Scalar dqkm2_da = zero;
+    Scalar dpkm1_da = zero;
+    Scalar dqkm1_da = -x;
+    Scalar dans_da = (dpkm1_da - ans * dqkm1_da) / qkm1;
+
+    for (int i = 0; i < igamma_num_iterations<Scalar, mode>(); i++) {
+      c += one;
+      y += one;
+      z += two;
+
+      Scalar yc = y * c;
+      Scalar pk = pkm1 * z - pkm2 * yc;
+      Scalar qk = qkm1 * z - qkm2 * yc;
+
+      Scalar dpk_da = dpkm1_da * z - pkm1 - dpkm2_da * yc + pkm2 * c;
+      Scalar dqk_da = dqkm1_da * z - qkm1 - dqkm2_da * yc + qkm2 * c;
+
+      if (qk != zero) {
+        Scalar ans_prev = ans;
+        ans = pk / qk;
+
+        Scalar dans_da_prev = dans_da;
+        dans_da = (dpk_da - ans * dqk_da) / qk;
+
+        if (mode == VALUE) {
+          if (numext::abs(ans_prev - ans) <= machep * numext::abs(ans)) {
+            break;
+          }
+        } else {
+          if (numext::abs(dans_da - dans_da_prev) <= machep) {
+            break;
+          }
+        }
+      }
+
+      pkm2 = pkm1;
+      pkm1 = pk;
+      qkm2 = qkm1;
+      qkm1 = qk;
+
+      dpkm2_da = dpkm1_da;
+      dpkm1_da = dpk_da;
+      dqkm2_da = dqkm1_da;
+      dqkm1_da = dqk_da;
+
+      if (numext::abs(pk) > big) {
+        pkm2 *= biginv;
+        pkm1 *= biginv;
+        qkm2 *= biginv;
+        qkm1 *= biginv;
+
+        dpkm2_da *= biginv;
+        dpkm1_da *= biginv;
+        dqkm2_da *= biginv;
+        dqkm1_da *= biginv;
+      }
+    }
+
+    /* Compute  x**a * exp(-x) / gamma(a)  */
+    Scalar dlogax_da = numext::log(x) - digamma_impl<Scalar>::run(a);
+    Scalar dax_da = ax * dlogax_da;
+
+    switch (mode) {
+      case VALUE:
+        return ans * ax;
+      case DERIVATIVE:
+        return ans * dax_da + dans_da * ax;
+      case SAMPLE_DERIVATIVE:
+      default: // this is needed to suppress clang warning
+        return -(dans_da + ans * dlogax_da) * x;
+    }
+  }
+};
+
+template <typename Scalar, IgammaComputationMode mode>
+struct igamma_series_impl {
+  /* Computes igam(a, x) or its derivative (depending on the mode)
+   * using the series expansion of the incomplete Gamma function.
+   *
+   * Preconditions:
+   *   x > 0
+   *   a > 0
+   *   !(x > 1 && x > a)
+   */
+  EIGEN_DEVICE_FUNC
+  static Scalar run(Scalar a, Scalar x) {
+    const Scalar zero = 0;
+    const Scalar one = 1;
+    const Scalar machep = cephes_helper<Scalar>::machep();
+
+    Scalar ax = main_igamma_term<Scalar>(a, x);
+
+    // This is independent of mode. If this value is zero,
+    // then the function value is zero. If the function value is zero,
+    // then we are in a neighborhood where the function value evalutes to zero,
+    // so the derivative is zero.
+    if (ax == zero) {
+      return zero;
+    }
+
+    ax /= a;
+
+    /* power series */
+    Scalar r = a;
+    Scalar c = one;
+    Scalar ans = one;
+
+    Scalar dc_da = zero;
+    Scalar dans_da = zero;
+
+    for (int i = 0; i < igamma_num_iterations<Scalar, mode>(); i++) {
+      r += one;
+      Scalar term = x / r;
+      Scalar dterm_da = -x / (r * r);
+      dc_da = term * dc_da + dterm_da * c;
+      dans_da += dc_da;
+      c *= term;
+      ans += c;
+
+      if (mode == VALUE) {
+        if (c <= machep * ans) {
+          break;
+        }
+      } else {
+        if (numext::abs(dc_da) <= machep * numext::abs(dans_da)) {
+          break;
+        }
+      }
+    }
+
+    Scalar dlogax_da = numext::log(x) - digamma_impl<Scalar>::run(a + one);
+    Scalar dax_da = ax * dlogax_da;
+
+    switch (mode) {
+      case VALUE:
+        return ans * ax;
+      case DERIVATIVE:
+        return ans * dax_da + dans_da * ax;
+      case SAMPLE_DERIVATIVE:
+      default: // this is needed to suppress clang warning
+        return -(dans_da + ans * dlogax_da) * x / a;
+    }
+  }
+};
+
+#if !EIGEN_HAS_C99_MATH
+
+template <typename Scalar>
+struct igammac_impl {
+  EIGEN_DEVICE_FUNC
+  static Scalar run(Scalar a, Scalar x) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+#else
+
+template <typename Scalar>
+struct igammac_impl {
+  EIGEN_DEVICE_FUNC
+  static Scalar run(Scalar a, Scalar x) {
+    /*  igamc()
+     *
+     *	Incomplete gamma integral (modified for Eigen)
+     *
+     *
+     *
+     * SYNOPSIS:
+     *
+     * double a, x, y, igamc();
+     *
+     * y = igamc( a, x );
+     *
+     * DESCRIPTION:
+     *
+     * The function is defined by
+     *
+     *
+     *  igamc(a,x)   =   1 - igam(a,x)
+     *
+     *                            inf.
+     *                              -
+     *                     1       | |  -t  a-1
+     *               =   -----     |   e   t   dt.
+     *                    -      | |
+     *                   | (a)    -
+     *                             x
+     *
+     *
+     * In this implementation both arguments must be positive.
+     * The integral is evaluated by either a power series or
+     * continued fraction expansion, depending on the relative
+     * values of a and x.
+     *
+     * ACCURACY (float):
+     *
+     *                      Relative error:
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0,30        30000       7.8e-6      5.9e-7
+     *
+     *
+     * ACCURACY (double):
+     *
+     * Tested at random a, x.
+     *                a         x                      Relative error:
+     * arithmetic   domain   domain     # trials      peak         rms
+     *    IEEE     0.5,100   0,100      200000       1.9e-14     1.7e-15
+     *    IEEE     0.01,0.5  0,100      200000       1.4e-13     1.6e-15
+     *
+     */
+    /*
+      Cephes Math Library Release 2.2: June, 1992
+      Copyright 1985, 1987, 1992 by Stephen L. Moshier
+      Direct inquiries to 30 Frost Street, Cambridge, MA 02140
+    */
+    const Scalar zero = 0;
+    const Scalar one = 1;
+    const Scalar nan = NumTraits<Scalar>::quiet_NaN();
+
+    if ((x < zero) || (a <= zero)) {
+      // domain error
+      return nan;
+    }
+
+    if ((numext::isnan)(a) || (numext::isnan)(x)) {  // propagate nans
+      return nan;
+    }
+
+    if ((x < one) || (x < a)) {
+      return (one - igamma_series_impl<Scalar, VALUE>::run(a, x));
+    }
+
+    return igammac_cf_impl<Scalar, VALUE>::run(a, x);
+  }
+};
+
+#endif  // EIGEN_HAS_C99_MATH
+
+/************************************************************************************************
+ * Implementation of igamma (incomplete gamma integral), based on Cephes but requires C++11/C99 *
+ ************************************************************************************************/
+
+#if !EIGEN_HAS_C99_MATH
+
+template <typename Scalar, IgammaComputationMode mode>
+struct igamma_generic_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(Scalar a, Scalar x) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+#else
+
+template <typename Scalar, IgammaComputationMode mode>
+struct igamma_generic_impl {
+  EIGEN_DEVICE_FUNC
+  static Scalar run(Scalar a, Scalar x) {
+    /* Depending on the mode, returns
+     * - VALUE: incomplete Gamma function igamma(a, x)
+     * - DERIVATIVE: derivative of incomplete Gamma function d/da igamma(a, x)
+     * - SAMPLE_DERIVATIVE: implicit derivative of a Gamma random variable
+     * x ~ Gamma(x | a, 1), dx/da = -1 / Gamma(x | a, 1) * d igamma(a, x) / dx
+     *
+     * Derivatives are implemented by forward-mode differentiation.
+     */
+    const Scalar zero = 0;
+    const Scalar one = 1;
+    const Scalar nan = NumTraits<Scalar>::quiet_NaN();
+
+    if (x == zero) return zero;
+
+    if ((x < zero) || (a <= zero)) {  // domain error
+      return nan;
+    }
+
+    if ((numext::isnan)(a) || (numext::isnan)(x)) {  // propagate nans
+      return nan;
+    }
+
+    if ((x > one) && (x > a)) {
+      Scalar ret = igammac_cf_impl<Scalar, mode>::run(a, x);
+      if (mode == VALUE) {
+        return one - ret;
+      } else {
+        return -ret;
+      }
+    }
+
+    return igamma_series_impl<Scalar, mode>::run(a, x);
+  }
+};
+
+#endif  // EIGEN_HAS_C99_MATH
+
+template <typename Scalar>
+struct igamma_retval {
+  typedef Scalar type;
+};
+
+template <typename Scalar>
+struct igamma_impl : igamma_generic_impl<Scalar, VALUE> {
+  /* igam()
+   * Incomplete gamma integral.
+   *
+   * The CDF of Gamma(a, 1) random variable at the point x.
+   *
+   * Accuracy estimation. For each a in [10^-2, 10^-1...10^3] we sample
+   * 50 Gamma random variables x ~ Gamma(x | a, 1), a total of 300 points.
+   * The ground truth is computed by mpmath. Mean absolute error:
+   * float: 1.26713e-05
+   * double: 2.33606e-12
+   *
+   * Cephes documentation below.
+   *
+   * SYNOPSIS:
+   *
+   * double a, x, y, igam();
+   *
+   * y = igam( a, x );
+   *
+   * DESCRIPTION:
+   *
+   * The function is defined by
+   *
+   *                           x
+   *                            -
+   *                   1       | |  -t  a-1
+   *  igam(a,x)  =   -----     |   e   t   dt.
+   *                  -      | |
+   *                 | (a)    -
+   *                           0
+   *
+   *
+   * In this implementation both arguments must be positive.
+   * The integral is evaluated by either a power series or
+   * continued fraction expansion, depending on the relative
+   * values of a and x.
+   *
+   * ACCURACY (double):
+   *
+   *                      Relative error:
+   * arithmetic   domain     # trials      peak         rms
+   *    IEEE      0,30       200000       3.6e-14     2.9e-15
+   *    IEEE      0,100      300000       9.9e-14     1.5e-14
+   *
+   *
+   * ACCURACY (float):
+   *
+   *                      Relative error:
+   * arithmetic   domain     # trials      peak         rms
+   *    IEEE      0,30        20000       7.8e-6      5.9e-7
+   *
+   */
+  /*
+    Cephes Math Library Release 2.2: June, 1992
+    Copyright 1985, 1987, 1992 by Stephen L. Moshier
+    Direct inquiries to 30 Frost Street, Cambridge, MA 02140
+  */
+
+  /* left tail of incomplete gamma function:
+   *
+   *          inf.      k
+   *   a  -x   -       x
+   *  x  e     >   ----------
+   *           -     -
+   *          k=0   | (a+k+1)
+   *
+   */
+};
+
+template <typename Scalar>
+struct igamma_der_a_retval : igamma_retval<Scalar> {};
+
+template <typename Scalar>
+struct igamma_der_a_impl : igamma_generic_impl<Scalar, DERIVATIVE> {
+  /* Derivative of the incomplete Gamma function with respect to a.
+   *
+   * Computes d/da igamma(a, x) by forward differentiation of the igamma code.
+   *
+   * Accuracy estimation. For each a in [10^-2, 10^-1...10^3] we sample
+   * 50 Gamma random variables x ~ Gamma(x | a, 1), a total of 300 points.
+   * The ground truth is computed by mpmath. Mean absolute error:
+   * float: 6.17992e-07
+   * double: 4.60453e-12
+   *
+   * Reference:
+   * R. Moore. "Algorithm AS 187: Derivatives of the incomplete gamma
+   * integral". Journal of the Royal Statistical Society. 1982
+   */
+};
+
+template <typename Scalar>
+struct gamma_sample_der_alpha_retval : igamma_retval<Scalar> {};
+
+template <typename Scalar>
+struct gamma_sample_der_alpha_impl
+    : igamma_generic_impl<Scalar, SAMPLE_DERIVATIVE> {
+  /* Derivative of a Gamma random variable sample with respect to alpha.
+   *
+   * Consider a sample of a Gamma random variable with the concentration
+   * parameter alpha: sample ~ Gamma(alpha, 1). The reparameterization
+   * derivative that we want to compute is dsample / dalpha =
+   * d igammainv(alpha, u) / dalpha, where u = igamma(alpha, sample).
+   * However, this formula is numerically unstable and expensive, so instead
+   * we use implicit differentiation:
+   *
+   * igamma(alpha, sample) = u, where u ~ Uniform(0, 1).
+   * Apply d / dalpha to both sides:
+   * d igamma(alpha, sample) / dalpha
+   *     + d igamma(alpha, sample) / dsample * dsample/dalpha  = 0
+   * d igamma(alpha, sample) / dalpha
+   *     + Gamma(sample | alpha, 1) dsample / dalpha = 0
+   * dsample/dalpha = - (d igamma(alpha, sample) / dalpha)
+   *                   / Gamma(sample | alpha, 1)
+   *
+   * Here Gamma(sample | alpha, 1) is the PDF of the Gamma distribution
+   * (note that the derivative of the CDF w.r.t. sample is the PDF).
+   * See the reference below for more details.
+   *
+   * The derivative of igamma(alpha, sample) is computed by forward
+   * differentiation of the igamma code. Division by the Gamma PDF is performed
+   * in the same code, increasing the accuracy and speed due to cancellation
+   * of some terms.
+   *
+   * Accuracy estimation. For each alpha in [10^-2, 10^-1...10^3] we sample
+   * 50 Gamma random variables sample ~ Gamma(sample | alpha, 1), a total of 300
+   * points. The ground truth is computed by mpmath. Mean absolute error:
+   * float: 2.1686e-06
+   * double: 1.4774e-12
+   *
+   * Reference:
+   * M. Figurnov, S. Mohamed, A. Mnih "Implicit Reparameterization Gradients".
+   * 2018
+   */
+};
+
+/*****************************************************************************
+ * Implementation of Riemann zeta function of two arguments, based on Cephes *
+ *****************************************************************************/
+
+template <typename Scalar>
+struct zeta_retval {
+    typedef Scalar type;
+};
+
+template <typename Scalar>
+struct zeta_impl_series {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+template <>
+struct zeta_impl_series<float> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE bool run(float& a, float& b, float& s, const float x, const float machep) {
+    int i = 0;
+    while(i < 9)
+    {
+        i += 1;
+        a += 1.0f;
+        b = numext::pow( a, -x );
+        s += b;
+        if( numext::abs(b/s) < machep )
+            return true;
+    }
+
+    //Return whether we are done
+    return false;
+  }
+};
+
+template <>
+struct zeta_impl_series<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE bool run(double& a, double& b, double& s, const double x, const double machep) {
+    int i = 0;
+    while( (i < 9) || (a <= 9.0) )
+    {
+        i += 1;
+        a += 1.0;
+        b = numext::pow( a, -x );
+        s += b;
+        if( numext::abs(b/s) < machep )
+            return true;
+    }
+
+    //Return whether we are done
+    return false;
+  }
+};
+
+template <typename Scalar>
+struct zeta_impl {
+    EIGEN_DEVICE_FUNC
+    static Scalar run(Scalar x, Scalar q) {
+        /*							zeta.c
+         *
+         *	Riemann zeta function of two arguments
+         *
+         *
+         *
+         * SYNOPSIS:
+         *
+         * double x, q, y, zeta();
+         *
+         * y = zeta( x, q );
+         *
+         *
+         *
+         * DESCRIPTION:
+         *
+         *
+         *
+         *                 inf.
+         *                  -        -x
+         *   zeta(x,q)  =   >   (k+q)
+         *                  -
+         *                 k=0
+         *
+         * where x > 1 and q is not a negative integer or zero.
+         * The Euler-Maclaurin summation formula is used to obtain
+         * the expansion
+         *
+         *                n
+         *                -       -x
+         * zeta(x,q)  =   >  (k+q)
+         *                -
+         *               k=1
+         *
+         *           1-x                 inf.  B   x(x+1)...(x+2j)
+         *      (n+q)           1         -     2j
+         *  +  ---------  -  -------  +   >    --------------------
+         *        x-1              x      -                   x+2j+1
+         *                   2(n+q)      j=1       (2j)! (n+q)
+         *
+         * where the B2j are Bernoulli numbers.  Note that (see zetac.c)
+         * zeta(x,1) = zetac(x) + 1.
+         *
+         *
+         *
+         * ACCURACY:
+         *
+         * Relative error for single precision:
+         * arithmetic   domain     # trials      peak         rms
+         *    IEEE      0,25        10000       6.9e-7      1.0e-7
+         *
+         * Large arguments may produce underflow in powf(), in which
+         * case the results are inaccurate.
+         *
+         * REFERENCE:
+         *
+         * Gradshteyn, I. S., and I. M. Ryzhik, Tables of Integrals,
+         * Series, and Products, p. 1073; Academic Press, 1980.
+         *
+         */
+
+        int i;
+        Scalar p, r, a, b, k, s, t, w;
+
+        const Scalar A[] = {
+            Scalar(12.0),
+            Scalar(-720.0),
+            Scalar(30240.0),
+            Scalar(-1209600.0),
+            Scalar(47900160.0),
+            Scalar(-1.8924375803183791606e9), /*1.307674368e12/691*/
+            Scalar(7.47242496e10),
+            Scalar(-2.950130727918164224e12), /*1.067062284288e16/3617*/
+            Scalar(1.1646782814350067249e14), /*5.109094217170944e18/43867*/
+            Scalar(-4.5979787224074726105e15), /*8.028576626982912e20/174611*/
+            Scalar(1.8152105401943546773e17), /*1.5511210043330985984e23/854513*/
+            Scalar(-7.1661652561756670113e18) /*1.6938241367317436694528e27/236364091*/
+            };
+
+        const Scalar maxnum = NumTraits<Scalar>::infinity();
+        const Scalar zero = 0.0, half = 0.5, one = 1.0;
+        const Scalar machep = cephes_helper<Scalar>::machep();
+        const Scalar nan = NumTraits<Scalar>::quiet_NaN();
+
+        if( x == one )
+            return maxnum;
+
+        if( x < one )
+        {
+            return nan;
+        }
+
+        if( q <= zero )
+        {
+            if(q == numext::floor(q))
+            {
+                if (x == numext::floor(x) && long(x) % 2 == 0) {
+                    return maxnum;
+                }
+                else {
+                    return nan;
+                }
+            }
+            p = x;
+            r = numext::floor(p);
+            if (p != r)
+                return nan;
+        }
+
+        /* Permit negative q but continue sum until n+q > +9 .
+         * This case should be handled by a reflection formula.
+         * If q<0 and x is an integer, there is a relation to
+         * the polygamma function.
+         */
+        s = numext::pow( q, -x );
+        a = q;
+        b = zero;
+        // Run the summation in a helper function that is specific to the floating precision
+        if (zeta_impl_series<Scalar>::run(a, b, s, x, machep)) {
+            return s;
+        }
+
+        w = a;
+        s += b*w/(x-one);
+        s -= half * b;
+        a = one;
+        k = zero;
+        for( i=0; i<12; i++ )
+        {
+            a *= x + k;
+            b /= w;
+            t = a*b/A[i];
+            s = s + t;
+            t = numext::abs(t/s);
+            if( t < machep ) {
+              break;
+            }
+            k += one;
+            a *= x + k;
+            b /= w;
+            k += one;
+        }
+        return s;
+  }
+};
+
+/****************************************************************************
+ * Implementation of polygamma function, requires C++11/C99                 *
+ ****************************************************************************/
+
+template <typename Scalar>
+struct polygamma_retval {
+    typedef Scalar type;
+};
+
+#if !EIGEN_HAS_C99_MATH
+
+template <typename Scalar>
+struct polygamma_impl {
+    EIGEN_DEVICE_FUNC
+    static EIGEN_STRONG_INLINE Scalar run(Scalar n, Scalar x) {
+        EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                            THIS_TYPE_IS_NOT_SUPPORTED);
+        return Scalar(0);
+    }
+};
+
+#else
+
+template <typename Scalar>
+struct polygamma_impl {
+    EIGEN_DEVICE_FUNC
+    static Scalar run(Scalar n, Scalar x) {
+        Scalar zero = 0.0, one = 1.0;
+        Scalar nplus = n + one;
+        const Scalar nan = NumTraits<Scalar>::quiet_NaN();
+
+        // Check that n is a non-negative integer
+        if (numext::floor(n) != n || n < zero) {
+            return nan;
+        }
+        // Just return the digamma function for n = 0
+        else if (n == zero) {
+            return digamma_impl<Scalar>::run(x);
+        }
+        // Use the same implementation as scipy
+        else {
+            Scalar factorial = numext::exp(lgamma_impl<Scalar>::run(nplus));
+            return numext::pow(-one, nplus) * factorial * zeta_impl<Scalar>::run(nplus, x);
+        }
+  }
+};
+
+#endif  // EIGEN_HAS_C99_MATH
+
+/************************************************************************************************
+ * Implementation of betainc (incomplete beta integral), based on Cephes but requires C++11/C99 *
+ ************************************************************************************************/
+
+template <typename Scalar>
+struct betainc_retval {
+  typedef Scalar type;
+};
+
+#if !EIGEN_HAS_C99_MATH
+
+template <typename Scalar>
+struct betainc_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(Scalar a, Scalar b, Scalar x) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+#else
+
+template <typename Scalar>
+struct betainc_impl {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(Scalar, Scalar, Scalar) {
+    /*	betaincf.c
+     *
+     *	Incomplete beta integral
+     *
+     *
+     * SYNOPSIS:
+     *
+     * float a, b, x, y, betaincf();
+     *
+     * y = betaincf( a, b, x );
+     *
+     *
+     * DESCRIPTION:
+     *
+     * Returns incomplete beta integral of the arguments, evaluated
+     * from zero to x.  The function is defined as
+     *
+     *                  x
+     *     -            -
+     *    | (a+b)      | |  a-1     b-1
+     *  -----------    |   t   (1-t)   dt.
+     *   -     -     | |
+     *  | (a) | (b)   -
+     *                 0
+     *
+     * The domain of definition is 0 <= x <= 1.  In this
+     * implementation a and b are restricted to positive values.
+     * The integral from x to 1 may be obtained by the symmetry
+     * relation
+     *
+     *    1 - betainc( a, b, x )  =  betainc( b, a, 1-x ).
+     *
+     * The integral is evaluated by a continued fraction expansion.
+     * If a < 1, the function calls itself recursively after a
+     * transformation to increase a to a+1.
+     *
+     * ACCURACY (float):
+     *
+     * Tested at random points (a,b,x) with a and b in the indicated
+     * interval and x between 0 and 1.
+     *
+     * arithmetic   domain     # trials      peak         rms
+     * Relative error:
+     *    IEEE       0,30       10000       3.7e-5      5.1e-6
+     *    IEEE       0,100      10000       1.7e-4      2.5e-5
+     * The useful domain for relative error is limited by underflow
+     * of the single precision exponential function.
+     * Absolute error:
+     *    IEEE       0,30      100000       2.2e-5      9.6e-7
+     *    IEEE       0,100      10000       6.5e-5      3.7e-6
+     *
+     * Larger errors may occur for extreme ratios of a and b.
+     *
+     * ACCURACY (double):
+     * arithmetic   domain     # trials      peak         rms
+     *    IEEE      0,5         10000       6.9e-15     4.5e-16
+     *    IEEE      0,85       250000       2.2e-13     1.7e-14
+     *    IEEE      0,1000      30000       5.3e-12     6.3e-13
+     *    IEEE      0,10000    250000       9.3e-11     7.1e-12
+     *    IEEE      0,100000    10000       8.7e-10     4.8e-11
+     * Outputs smaller than the IEEE gradual underflow threshold
+     * were excluded from these statistics.
+     *
+     * ERROR MESSAGES:
+     *   message         condition      value returned
+     * incbet domain      x<0, x>1          nan
+     * incbet underflow                     nan
+     */
+
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+/* Continued fraction expansion #1 for incomplete beta integral (small_branch = True)
+ * Continued fraction expansion #2 for incomplete beta integral (small_branch = False)
+ */
+template <typename Scalar>
+struct incbeta_cfe {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(Scalar a, Scalar b, Scalar x, bool small_branch) {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, float>::value ||
+                         internal::is_same<Scalar, double>::value),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    const Scalar big = cephes_helper<Scalar>::big();
+    const Scalar machep = cephes_helper<Scalar>::machep();
+    const Scalar biginv = cephes_helper<Scalar>::biginv();
+
+    const Scalar zero = 0;
+    const Scalar one = 1;
+    const Scalar two = 2;
+
+    Scalar xk, pk, pkm1, pkm2, qk, qkm1, qkm2;
+    Scalar k1, k2, k3, k4, k5, k6, k7, k8, k26update;
+    Scalar ans;
+    int n;
+
+    const int num_iters = (internal::is_same<Scalar, float>::value) ? 100 : 300;
+    const Scalar thresh =
+        (internal::is_same<Scalar, float>::value) ? machep : Scalar(3) * machep;
+    Scalar r = (internal::is_same<Scalar, float>::value) ? zero : one;
+
+    if (small_branch) {
+      k1 = a;
+      k2 = a + b;
+      k3 = a;
+      k4 = a + one;
+      k5 = one;
+      k6 = b - one;
+      k7 = k4;
+      k8 = a + two;
+      k26update = one;
+    } else {
+      k1 = a;
+      k2 = b - one;
+      k3 = a;
+      k4 = a + one;
+      k5 = one;
+      k6 = a + b;
+      k7 = a + one;
+      k8 = a + two;
+      k26update = -one;
+      x = x / (one - x);
+    }
+
+    pkm2 = zero;
+    qkm2 = one;
+    pkm1 = one;
+    qkm1 = one;
+    ans = one;
+    n = 0;
+
+    do {
+      xk = -(x * k1 * k2) / (k3 * k4);
+      pk = pkm1 + pkm2 * xk;
+      qk = qkm1 + qkm2 * xk;
+      pkm2 = pkm1;
+      pkm1 = pk;
+      qkm2 = qkm1;
+      qkm1 = qk;
+
+      xk = (x * k5 * k6) / (k7 * k8);
+      pk = pkm1 + pkm2 * xk;
+      qk = qkm1 + qkm2 * xk;
+      pkm2 = pkm1;
+      pkm1 = pk;
+      qkm2 = qkm1;
+      qkm1 = qk;
+
+      if (qk != zero) {
+        r = pk / qk;
+        if (numext::abs(ans - r) < numext::abs(r) * thresh) {
+          return r;
+        }
+        ans = r;
+      }
+
+      k1 += one;
+      k2 += k26update;
+      k3 += two;
+      k4 += two;
+      k5 += one;
+      k6 -= k26update;
+      k7 += two;
+      k8 += two;
+
+      if ((numext::abs(qk) + numext::abs(pk)) > big) {
+        pkm2 *= biginv;
+        pkm1 *= biginv;
+        qkm2 *= biginv;
+        qkm1 *= biginv;
+      }
+      if ((numext::abs(qk) < biginv) || (numext::abs(pk) < biginv)) {
+        pkm2 *= big;
+        pkm1 *= big;
+        qkm2 *= big;
+        qkm1 *= big;
+      }
+    } while (++n < num_iters);
+
+    return ans;
+  }
+};
+
+/* Helper functions depending on the Scalar type */
+template <typename Scalar>
+struct betainc_helper {};
+
+template <>
+struct betainc_helper<float> {
+  /* Core implementation, assumes a large (> 1.0) */
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE float incbsa(float aa, float bb,
+                                                            float xx) {
+    float ans, a, b, t, x, onemx;
+    bool reversed_a_b = false;
+
+    onemx = 1.0f - xx;
+
+    /* see if x is greater than the mean */
+    if (xx > (aa / (aa + bb))) {
+      reversed_a_b = true;
+      a = bb;
+      b = aa;
+      t = xx;
+      x = onemx;
+    } else {
+      a = aa;
+      b = bb;
+      t = onemx;
+      x = xx;
+    }
+
+    /* Choose expansion for optimal convergence */
+    if (b > 10.0f) {
+      if (numext::abs(b * x / a) < 0.3f) {
+        t = betainc_helper<float>::incbps(a, b, x);
+        if (reversed_a_b) t = 1.0f - t;
+        return t;
+      }
+    }
+
+    ans = x * (a + b - 2.0f) / (a - 1.0f);
+    if (ans < 1.0f) {
+      ans = incbeta_cfe<float>::run(a, b, x, true /* small_branch */);
+      t = b * numext::log(t);
+    } else {
+      ans = incbeta_cfe<float>::run(a, b, x, false /* small_branch */);
+      t = (b - 1.0f) * numext::log(t);
+    }
+
+    t += a * numext::log(x) + lgamma_impl<float>::run(a + b) -
+         lgamma_impl<float>::run(a) - lgamma_impl<float>::run(b);
+    t += numext::log(ans / a);
+    t = numext::exp(t);
+
+    if (reversed_a_b) t = 1.0f - t;
+    return t;
+  }
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float incbps(float a, float b, float x) {
+    float t, u, y, s;
+    const float machep = cephes_helper<float>::machep();
+
+    y = a * numext::log(x) + (b - 1.0f) * numext::log1p(-x) - numext::log(a);
+    y -= lgamma_impl<float>::run(a) + lgamma_impl<float>::run(b);
+    y += lgamma_impl<float>::run(a + b);
+
+    t = x / (1.0f - x);
+    s = 0.0f;
+    u = 1.0f;
+    do {
+      b -= 1.0f;
+      if (b == 0.0f) {
+        break;
+      }
+      a += 1.0f;
+      u *= t * b / a;
+      s += u;
+    } while (numext::abs(u) > machep);
+
+    return numext::exp(y) * (1.0f + s);
+  }
+};
+
+template <>
+struct betainc_impl<float> {
+  EIGEN_DEVICE_FUNC
+  static float run(float a, float b, float x) {
+    const float nan = NumTraits<float>::quiet_NaN();
+    float ans, t;
+
+    if (a <= 0.0f) return nan;
+    if (b <= 0.0f) return nan;
+    if ((x <= 0.0f) || (x >= 1.0f)) {
+      if (x == 0.0f) return 0.0f;
+      if (x == 1.0f) return 1.0f;
+      // mtherr("betaincf", DOMAIN);
+      return nan;
+    }
+
+    /* transformation for small aa */
+    if (a <= 1.0f) {
+      ans = betainc_helper<float>::incbsa(a + 1.0f, b, x);
+      t = a * numext::log(x) + b * numext::log1p(-x) +
+          lgamma_impl<float>::run(a + b) - lgamma_impl<float>::run(a + 1.0f) -
+          lgamma_impl<float>::run(b);
+      return (ans + numext::exp(t));
+    } else {
+      return betainc_helper<float>::incbsa(a, b, x);
+    }
+  }
+};
+
+template <>
+struct betainc_helper<double> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double incbps(double a, double b, double x) {
+    const double machep = cephes_helper<double>::machep();
+
+    double s, t, u, v, n, t1, z, ai;
+
+    ai = 1.0 / a;
+    u = (1.0 - b) * x;
+    v = u / (a + 1.0);
+    t1 = v;
+    t = u;
+    n = 2.0;
+    s = 0.0;
+    z = machep * ai;
+    while (numext::abs(v) > z) {
+      u = (n - b) * x / n;
+      t *= u;
+      v = t / (a + n);
+      s += v;
+      n += 1.0;
+    }
+    s += t1;
+    s += ai;
+
+    u = a * numext::log(x);
+    // TODO: gamma() is not directly implemented in Eigen.
+    /*
+    if ((a + b) < maxgam && numext::abs(u) < maxlog) {
+      t = gamma(a + b) / (gamma(a) * gamma(b));
+      s = s * t * pow(x, a);
+    }
+    */
+    t = lgamma_impl<double>::run(a + b) - lgamma_impl<double>::run(a) -
+        lgamma_impl<double>::run(b) + u + numext::log(s);
+    return s = numext::exp(t);
+  }
+};
+
+template <>
+struct betainc_impl<double> {
+  EIGEN_DEVICE_FUNC
+  static double run(double aa, double bb, double xx) {
+    const double nan = NumTraits<double>::quiet_NaN();
+    const double machep = cephes_helper<double>::machep();
+    // const double maxgam = 171.624376956302725;
+
+    double a, b, t, x, xc, w, y;
+    bool reversed_a_b = false;
+
+    if (aa <= 0.0 || bb <= 0.0) {
+      return nan;  // goto domerr;
+    }
+
+    if ((xx <= 0.0) || (xx >= 1.0)) {
+      if (xx == 0.0) return (0.0);
+      if (xx == 1.0) return (1.0);
+      // mtherr("incbet", DOMAIN);
+      return nan;
+    }
+
+    if ((bb * xx) <= 1.0 && xx <= 0.95) {
+      return betainc_helper<double>::incbps(aa, bb, xx);
+    }
+
+    w = 1.0 - xx;
+
+    /* Reverse a and b if x is greater than the mean. */
+    if (xx > (aa / (aa + bb))) {
+      reversed_a_b = true;
+      a = bb;
+      b = aa;
+      xc = xx;
+      x = w;
+    } else {
+      a = aa;
+      b = bb;
+      xc = w;
+      x = xx;
+    }
+
+    if (reversed_a_b && (b * x) <= 1.0 && x <= 0.95) {
+      t = betainc_helper<double>::incbps(a, b, x);
+      if (t <= machep) {
+        t = 1.0 - machep;
+      } else {
+        t = 1.0 - t;
+      }
+      return t;
+    }
+
+    /* Choose expansion for better convergence. */
+    y = x * (a + b - 2.0) - (a - 1.0);
+    if (y < 0.0) {
+      w = incbeta_cfe<double>::run(a, b, x, true /* small_branch */);
+    } else {
+      w = incbeta_cfe<double>::run(a, b, x, false /* small_branch */) / xc;
+    }
+
+    /* Multiply w by the factor
+         a      b   _             _     _
+        x  (1-x)   | (a+b) / ( a | (a) | (b) ) .   */
+
+    y = a * numext::log(x);
+    t = b * numext::log(xc);
+    // TODO: gamma is not directly implemented in Eigen.
+    /*
+    if ((a + b) < maxgam && numext::abs(y) < maxlog && numext::abs(t) < maxlog)
+    {
+      t = pow(xc, b);
+      t *= pow(x, a);
+      t /= a;
+      t *= w;
+      t *= gamma(a + b) / (gamma(a) * gamma(b));
+    } else {
+    */
+    /* Resort to logarithms.  */
+    y += t + lgamma_impl<double>::run(a + b) - lgamma_impl<double>::run(a) -
+         lgamma_impl<double>::run(b);
+    y += numext::log(w / a);
+    t = numext::exp(y);
+
+    /* } */
+    // done:
+
+    if (reversed_a_b) {
+      if (t <= machep) {
+        t = 1.0 - machep;
+      } else {
+        t = 1.0 - t;
+      }
+    }
+    return t;
+  }
+};
+
+#endif  // EIGEN_HAS_C99_MATH
+
+}  // end namespace internal
+
+namespace numext {
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(lgamma, Scalar)
+    lgamma(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(lgamma, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(digamma, Scalar)
+    digamma(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(digamma, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(zeta, Scalar)
+zeta(const Scalar& x, const Scalar& q) {
+    return EIGEN_MATHFUNC_IMPL(zeta, Scalar)::run(x, q);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(polygamma, Scalar)
+polygamma(const Scalar& n, const Scalar& x) {
+    return EIGEN_MATHFUNC_IMPL(polygamma, Scalar)::run(n, x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(erf, Scalar)
+    erf(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(erf, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(erfc, Scalar)
+    erfc(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(erfc, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(ndtri, Scalar)
+    ndtri(const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(ndtri, Scalar)::run(x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(igamma, Scalar)
+    igamma(const Scalar& a, const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(igamma, Scalar)::run(a, x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(igamma_der_a, Scalar)
+    igamma_der_a(const Scalar& a, const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(igamma_der_a, Scalar)::run(a, x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(gamma_sample_der_alpha, Scalar)
+    gamma_sample_der_alpha(const Scalar& a, const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(gamma_sample_der_alpha, Scalar)::run(a, x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(igammac, Scalar)
+    igammac(const Scalar& a, const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(igammac, Scalar)::run(a, x);
+}
+
+template <typename Scalar>
+EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(betainc, Scalar)
+    betainc(const Scalar& a, const Scalar& b, const Scalar& x) {
+  return EIGEN_MATHFUNC_IMPL(betainc, Scalar)::run(a, b, x);
+}
+
+}  // end namespace numext
+}  // end namespace Eigen
+
+#endif  // EIGEN_SPECIAL_FUNCTIONS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsPacketMath.h b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsPacketMath.h
new file mode 100644
index 0000000..2bb0179
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/SpecialFunctionsPacketMath.h
@@ -0,0 +1,79 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPECIALFUNCTIONS_PACKETMATH_H
+#define EIGEN_SPECIALFUNCTIONS_PACKETMATH_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal \returns the ln(|gamma(\a a)|) (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plgamma(const Packet& a) { using numext::lgamma; return lgamma(a); }
+
+/** \internal \returns the derivative of lgamma, psi(\a a) (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pdigamma(const Packet& a) { using numext::digamma; return digamma(a); }
+
+/** \internal \returns the zeta function of two arguments (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pzeta(const Packet& x, const Packet& q) { using numext::zeta; return zeta(x, q); }
+
+/** \internal \returns the polygamma function (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet ppolygamma(const Packet& n, const Packet& x) { using numext::polygamma; return polygamma(n, x); }
+
+/** \internal \returns the erf(\a a) (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet perf(const Packet& a) { using numext::erf; return erf(a); }
+
+/** \internal \returns the erfc(\a a) (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet perfc(const Packet& a) { using numext::erfc; return erfc(a); }
+
+/** \internal \returns the ndtri(\a a) (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pndtri(const Packet& a) {
+  typedef typename unpacket_traits<Packet>::type ScalarType;
+  using internal::generic_ndtri; return generic_ndtri<Packet, ScalarType>(a);
+}
+
+/** \internal \returns the incomplete gamma function igamma(\a a, \a x) */
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+Packet pigamma(const Packet& a, const Packet& x) { using numext::igamma; return igamma(a, x); }
+
+/** \internal \returns the derivative of the incomplete gamma function
+ * igamma_der_a(\a a, \a x) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet pigamma_der_a(const Packet& a, const Packet& x) {
+  using numext::igamma_der_a; return igamma_der_a(a, x);
+}
+
+/** \internal \returns compute the derivative of the sample
+  * of Gamma(alpha, 1) random variable with respect to the parameter a
+  * gamma_sample_der_alpha(\a alpha, \a sample) */
+template <typename Packet>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet pgamma_sample_der_alpha(const Packet& alpha, const Packet& sample) {
+  using numext::gamma_sample_der_alpha; return gamma_sample_der_alpha(alpha, sample);
+}
+
+/** \internal \returns the complementary incomplete gamma function igammac(\a a, \a x) */
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+Packet pigammac(const Packet& a, const Packet& x) { using numext::igammac; return igammac(a, x); }
+
+/** \internal \returns the complementary incomplete gamma function betainc(\a a, \a b, \a x) */
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+Packet pbetainc(const Packet& a, const Packet& b,const Packet& x) { using numext::betainc; return betainc(a, b, x); }
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPECIALFUNCTIONS_PACKETMATH_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/arch/AVX/BesselFunctions.h b/src/EigenUnsupported/src/SpecialFunctions/arch/AVX/BesselFunctions.h
new file mode 100644
index 0000000..2d76692
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/arch/AVX/BesselFunctions.h
@@ -0,0 +1,46 @@
+#ifndef EIGEN_AVX_BESSELFUNCTIONS_H
+#define EIGEN_AVX_BESSELFUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_i0)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_i0)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_i0e)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_i0e)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_i1)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_i1)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_i1e)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_i1e)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_j0)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_j0)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_j1)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_j1)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_k0)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_k0)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_k0e)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_k0e)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_k1)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_k1)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_k1e)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_k1e)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_y0)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_y0)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pbessel_y1)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pbessel_y1)
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_AVX_BESSELFUNCTIONS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/arch/AVX/SpecialFunctions.h b/src/EigenUnsupported/src/SpecialFunctions/arch/AVX/SpecialFunctions.h
new file mode 100644
index 0000000..35e62a8
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/arch/AVX/SpecialFunctions.h
@@ -0,0 +1,16 @@
+#ifndef EIGEN_AVX_SPECIALFUNCTIONS_H
+#define EIGEN_AVX_SPECIALFUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, perf)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, perf)
+
+F16_PACKET_FUNCTION(Packet8f, Packet8h, pndtri)
+BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pndtri)
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_AVX_SPECIAL_FUNCTIONS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/arch/AVX512/BesselFunctions.h b/src/EigenUnsupported/src/SpecialFunctions/arch/AVX512/BesselFunctions.h
new file mode 100644
index 0000000..7dd3c3e
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/arch/AVX512/BesselFunctions.h
@@ -0,0 +1,46 @@
+#ifndef EIGEN_AVX512_BESSELFUNCTIONS_H
+#define EIGEN_AVX512_BESSELFUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_i0)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_i0)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_i0e)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_i0e)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_i1)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_i1)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_i1e)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_i1e)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_j0)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_j0)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_j1)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_j1)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_k0)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_k0)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_k0e)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_k0e)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_k1)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_k1)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_k1e)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_k1e)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_y0)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_y0)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_y1)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_y1)
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_AVX512_BESSELFUNCTIONS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/arch/AVX512/SpecialFunctions.h b/src/EigenUnsupported/src/SpecialFunctions/arch/AVX512/SpecialFunctions.h
new file mode 100644
index 0000000..79878f2
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/arch/AVX512/SpecialFunctions.h
@@ -0,0 +1,16 @@
+#ifndef EIGEN_AVX512_SPECIALFUNCTIONS_H
+#define EIGEN_AVX512_SPECIALFUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, perf)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, perf)
+
+F16_PACKET_FUNCTION(Packet16f, Packet16h, pndtri)
+BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pndtri)
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_AVX512_SPECIAL_FUNCTIONS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/arch/GPU/SpecialFunctions.h b/src/EigenUnsupported/src/SpecialFunctions/arch/GPU/SpecialFunctions.h
new file mode 100644
index 0000000..dd3bf4d
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/arch/GPU/SpecialFunctions.h
@@ -0,0 +1,369 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GPU_SPECIALFUNCTIONS_H
+#define EIGEN_GPU_SPECIALFUNCTIONS_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Make sure this is only available when targeting a GPU: we don't want to
+// introduce conflicts between these packet_traits definitions and the ones
+// we'll use on the host side (SSE, AVX, ...)
+#if defined(EIGEN_GPUCC) && defined(EIGEN_USE_GPU)
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 plgamma<float4>(const float4& a)
+{
+  return make_float4(lgammaf(a.x), lgammaf(a.y), lgammaf(a.z), lgammaf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 plgamma<double2>(const double2& a)
+{
+  using numext::lgamma;
+  return make_double2(lgamma(a.x), lgamma(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pdigamma<float4>(const float4& a)
+{
+  using numext::digamma;
+  return make_float4(digamma(a.x), digamma(a.y), digamma(a.z), digamma(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pdigamma<double2>(const double2& a)
+{
+  using numext::digamma;
+  return make_double2(digamma(a.x), digamma(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pzeta<float4>(const float4& x, const float4& q)
+{
+    using numext::zeta;
+    return make_float4(zeta(x.x, q.x), zeta(x.y, q.y), zeta(x.z, q.z), zeta(x.w, q.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pzeta<double2>(const double2& x, const double2& q)
+{
+    using numext::zeta;
+    return make_double2(zeta(x.x, q.x), zeta(x.y, q.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 ppolygamma<float4>(const float4& n, const float4& x)
+{
+    using numext::polygamma;
+    return make_float4(polygamma(n.x, x.x), polygamma(n.y, x.y), polygamma(n.z, x.z), polygamma(n.w, x.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 ppolygamma<double2>(const double2& n, const double2& x)
+{
+    using numext::polygamma;
+    return make_double2(polygamma(n.x, x.x), polygamma(n.y, x.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 perf<float4>(const float4& a)
+{
+  return make_float4(erff(a.x), erff(a.y), erff(a.z), erff(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 perf<double2>(const double2& a)
+{
+  using numext::erf;
+  return make_double2(erf(a.x), erf(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 perfc<float4>(const float4& a)
+{
+  using numext::erfc;
+  return make_float4(erfc(a.x), erfc(a.y), erfc(a.z), erfc(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 perfc<double2>(const double2& a)
+{
+  using numext::erfc;
+  return make_double2(erfc(a.x), erfc(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pndtri<float4>(const float4& a)
+{
+  using numext::ndtri;
+  return make_float4(ndtri(a.x), ndtri(a.y), ndtri(a.z), ndtri(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pndtri<double2>(const double2& a)
+{
+  using numext::ndtri;
+  return make_double2(ndtri(a.x), ndtri(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pigamma<float4>(const float4& a, const float4& x)
+{
+  using numext::igamma;
+  return make_float4(
+      igamma(a.x, x.x),
+      igamma(a.y, x.y),
+      igamma(a.z, x.z),
+      igamma(a.w, x.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pigamma<double2>(const double2& a, const double2& x)
+{
+  using numext::igamma;
+  return make_double2(igamma(a.x, x.x), igamma(a.y, x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pigamma_der_a<float4>(
+    const float4& a, const float4& x) {
+  using numext::igamma_der_a;
+  return make_float4(igamma_der_a(a.x, x.x), igamma_der_a(a.y, x.y),
+                     igamma_der_a(a.z, x.z), igamma_der_a(a.w, x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pigamma_der_a<double2>(const double2& a, const double2& x) {
+  using numext::igamma_der_a;
+  return make_double2(igamma_der_a(a.x, x.x), igamma_der_a(a.y, x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pgamma_sample_der_alpha<float4>(
+    const float4& alpha, const float4& sample) {
+  using numext::gamma_sample_der_alpha;
+  return make_float4(
+      gamma_sample_der_alpha(alpha.x, sample.x),
+      gamma_sample_der_alpha(alpha.y, sample.y),
+      gamma_sample_der_alpha(alpha.z, sample.z),
+      gamma_sample_der_alpha(alpha.w, sample.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pgamma_sample_der_alpha<double2>(const double2& alpha, const double2& sample) {
+  using numext::gamma_sample_der_alpha;
+  return make_double2(
+      gamma_sample_der_alpha(alpha.x, sample.x),
+      gamma_sample_der_alpha(alpha.y, sample.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pigammac<float4>(const float4& a, const float4& x)
+{
+  using numext::igammac;
+  return make_float4(
+      igammac(a.x, x.x),
+      igammac(a.y, x.y),
+      igammac(a.z, x.z),
+      igammac(a.w, x.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pigammac<double2>(const double2& a, const double2& x)
+{
+  using numext::igammac;
+  return make_double2(igammac(a.x, x.x), igammac(a.y, x.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pbetainc<float4>(const float4& a, const float4& b, const float4& x)
+{
+  using numext::betainc;
+  return make_float4(
+      betainc(a.x, b.x, x.x),
+      betainc(a.y, b.y, x.y),
+      betainc(a.z, b.z, x.z),
+      betainc(a.w, b.w, x.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pbetainc<double2>(const double2& a, const double2& b, const double2& x)
+{
+  using numext::betainc;
+  return make_double2(betainc(a.x, b.x, x.x), betainc(a.y, b.y, x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_i0e<float4>(const float4& x) {
+  using numext::bessel_i0e;
+  return make_float4(bessel_i0e(x.x), bessel_i0e(x.y), bessel_i0e(x.z), bessel_i0e(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_i0e<double2>(const double2& x) {
+  using numext::bessel_i0e;
+  return make_double2(bessel_i0e(x.x), bessel_i0e(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_i0<float4>(const float4& x) {
+  using numext::bessel_i0;
+  return make_float4(bessel_i0(x.x), bessel_i0(x.y), bessel_i0(x.z), bessel_i0(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_i0<double2>(const double2& x) {
+  using numext::bessel_i0;
+  return make_double2(bessel_i0(x.x), bessel_i0(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_i1e<float4>(const float4& x) {
+  using numext::bessel_i1e;
+  return make_float4(bessel_i1e(x.x), bessel_i1e(x.y), bessel_i1e(x.z), bessel_i1e(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_i1e<double2>(const double2& x) {
+  using numext::bessel_i1e;
+  return make_double2(bessel_i1e(x.x), bessel_i1e(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_i1<float4>(const float4& x) {
+  using numext::bessel_i1;
+  return make_float4(bessel_i1(x.x), bessel_i1(x.y), bessel_i1(x.z), bessel_i1(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_i1<double2>(const double2& x) {
+  using numext::bessel_i1;
+  return make_double2(bessel_i1(x.x), bessel_i1(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_k0e<float4>(const float4& x) {
+  using numext::bessel_k0e;
+  return make_float4(bessel_k0e(x.x), bessel_k0e(x.y), bessel_k0e(x.z), bessel_k0e(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_k0e<double2>(const double2& x) {
+  using numext::bessel_k0e;
+  return make_double2(bessel_k0e(x.x), bessel_k0e(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_k0<float4>(const float4& x) {
+  using numext::bessel_k0;
+  return make_float4(bessel_k0(x.x), bessel_k0(x.y), bessel_k0(x.z), bessel_k0(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_k0<double2>(const double2& x) {
+  using numext::bessel_k0;
+  return make_double2(bessel_k0(x.x), bessel_k0(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_k1e<float4>(const float4& x) {
+  using numext::bessel_k1e;
+  return make_float4(bessel_k1e(x.x), bessel_k1e(x.y), bessel_k1e(x.z), bessel_k1e(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_k1e<double2>(const double2& x) {
+  using numext::bessel_k1e;
+  return make_double2(bessel_k1e(x.x), bessel_k1e(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_k1<float4>(const float4& x) {
+  using numext::bessel_k1;
+  return make_float4(bessel_k1(x.x), bessel_k1(x.y), bessel_k1(x.z), bessel_k1(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_k1<double2>(const double2& x) {
+  using numext::bessel_k1;
+  return make_double2(bessel_k1(x.x), bessel_k1(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_j0<float4>(const float4& x) {
+  using numext::bessel_j0;
+  return make_float4(bessel_j0(x.x), bessel_j0(x.y), bessel_j0(x.z), bessel_j0(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_j0<double2>(const double2& x) {
+  using numext::bessel_j0;
+  return make_double2(bessel_j0(x.x), bessel_j0(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_j1<float4>(const float4& x) {
+  using numext::bessel_j1;
+  return make_float4(bessel_j1(x.x), bessel_j1(x.y), bessel_j1(x.z), bessel_j1(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_j1<double2>(const double2& x) {
+  using numext::bessel_j1;
+  return make_double2(bessel_j1(x.x), bessel_j1(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_y0<float4>(const float4& x) {
+  using numext::bessel_y0;
+  return make_float4(bessel_y0(x.x), bessel_y0(x.y), bessel_y0(x.z), bessel_y0(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_y0<double2>(const double2& x) {
+  using numext::bessel_y0;
+  return make_double2(bessel_y0(x.x), bessel_y0(x.y));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pbessel_y1<float4>(const float4& x) {
+  using numext::bessel_y1;
+  return make_float4(bessel_y1(x.x), bessel_y1(x.y), bessel_y1(x.z), bessel_y1(x.w));
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2
+pbessel_y1<double2>(const double2& x) {
+  using numext::bessel_y1;
+  return make_double2(bessel_y1(x.x), bessel_y1(x.y));
+}
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GPU_SPECIALFUNCTIONS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/arch/NEON/BesselFunctions.h b/src/EigenUnsupported/src/SpecialFunctions/arch/NEON/BesselFunctions.h
new file mode 100644
index 0000000..67433b0
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/arch/NEON/BesselFunctions.h
@@ -0,0 +1,54 @@
+#ifndef EIGEN_NEON_BESSELFUNCTIONS_H
+#define EIGEN_NEON_BESSELFUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+#if EIGEN_HAS_ARM64_FP16_VECTOR_ARITHMETIC
+
+#define NEON_HALF_TO_FLOAT_FUNCTIONS(METHOD)                            \
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE                       \
+Packet8hf METHOD<Packet8hf>(const Packet8hf& x) {                       \
+  const Packet4f lo = METHOD<Packet4f>(vcvt_f32_f16(vget_low_f16(x)));  \
+  const Packet4f hi = METHOD<Packet4f>(vcvt_f32_f16(vget_high_f16(x))); \
+  return vcombine_f16(vcvt_f16_f32(lo), vcvt_f16_f32(hi));              \
+}                                                                       \
+                                                                        \
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE                       \
+Packet4hf METHOD<Packet4hf>(const Packet4hf& x) {                       \
+  return vcvt_f16_f32(METHOD<Packet4f>(vcvt_f32_f16(x)));               \
+}
+
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_i0)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_i0e)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_i1)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_i1e)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_j0)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_j1)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_k0)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_k0e)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_k1)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_k1e)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_y0)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pbessel_y1)
+
+#undef NEON_HALF_TO_FLOAT_FUNCTIONS
+#endif
+
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_i0)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_i0e)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_i1)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_i1e)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_j0)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_j1)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_k0)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_k0e)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_k1)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_k1e)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_y0)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pbessel_y1)
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_NEON_BESSELFUNCTIONS_H
diff --git a/src/EigenUnsupported/src/SpecialFunctions/arch/NEON/SpecialFunctions.h b/src/EigenUnsupported/src/SpecialFunctions/arch/NEON/SpecialFunctions.h
new file mode 100644
index 0000000..ec92951
--- /dev/null
+++ b/src/EigenUnsupported/src/SpecialFunctions/arch/NEON/SpecialFunctions.h
@@ -0,0 +1,34 @@
+#ifndef EIGEN_NEON_SPECIALFUNCTIONS_H
+#define EIGEN_NEON_SPECIALFUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+#if EIGEN_HAS_ARM64_FP16_VECTOR_ARITHMETIC
+
+#define NEON_HALF_TO_FLOAT_FUNCTIONS(METHOD)                            \
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE                       \
+Packet8hf METHOD<Packet8hf>(const Packet8hf& x) {                       \
+  const Packet4f lo = METHOD<Packet4f>(vcvt_f32_f16(vget_low_f16(x)));  \
+  const Packet4f hi = METHOD<Packet4f>(vcvt_f32_f16(vget_high_f16(x))); \
+  return vcombine_f16(vcvt_f16_f32(lo), vcvt_f16_f32(hi));              \
+}                                                                       \
+                                                                        \
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE                       \
+Packet4hf METHOD<Packet4hf>(const Packet4hf& x) {                       \
+  return vcvt_f16_f32(METHOD<Packet4f>(vcvt_f32_f16(x)));               \
+}
+
+NEON_HALF_TO_FLOAT_FUNCTIONS(perf)
+NEON_HALF_TO_FLOAT_FUNCTIONS(pndtri)
+
+#undef NEON_HALF_TO_FLOAT_FUNCTIONS
+#endif
+
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, perf)
+BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pndtri)
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_NEON_SPECIALFUNCTIONS_H
diff --git a/src/EigenUnsupported/src/Splines/Spline.h b/src/EigenUnsupported/src/Splines/Spline.h
new file mode 100644
index 0000000..79edd52
--- /dev/null
+++ b/src/EigenUnsupported/src/Splines/Spline.h
@@ -0,0 +1,507 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 20010-2011 Hauke Heibel <hauke.heibel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPLINE_H
+#define EIGEN_SPLINE_H
+
+#include "SplineFwd.h"
+
+namespace Eigen
+{
+    /**
+     * \ingroup Splines_Module
+     * \class Spline
+     * \brief A class representing multi-dimensional spline curves.
+     *
+     * The class represents B-splines with non-uniform knot vectors. Each control
+     * point of the B-spline is associated with a basis function
+     * \f{align*}
+     *   C(u) & = \sum_{i=0}^{n}N_{i,p}(u)P_i
+     * \f}
+     *
+     * \tparam _Scalar The underlying data type (typically float or double)
+     * \tparam _Dim The curve dimension (e.g. 2 or 3)
+     * \tparam _Degree Per default set to Dynamic; could be set to the actual desired
+     *                degree for optimization purposes (would result in stack allocation
+     *                of several temporary variables).
+     **/
+  template <typename _Scalar, int _Dim, int _Degree>
+  class Spline
+  {
+  public:
+    typedef _Scalar Scalar; /*!< The spline curve's scalar type. */
+    enum { Dimension = _Dim /*!< The spline curve's dimension. */ };
+    enum { Degree = _Degree /*!< The spline curve's degree. */ };
+
+    /** \brief The point type the spline is representing. */
+    typedef typename SplineTraits<Spline>::PointType PointType;
+    
+    /** \brief The data type used to store knot vectors. */
+    typedef typename SplineTraits<Spline>::KnotVectorType KnotVectorType;
+
+    /** \brief The data type used to store parameter vectors. */
+    typedef typename SplineTraits<Spline>::ParameterVectorType ParameterVectorType;
+    
+    /** \brief The data type used to store non-zero basis functions. */
+    typedef typename SplineTraits<Spline>::BasisVectorType BasisVectorType;
+
+    /** \brief The data type used to store the values of the basis function derivatives. */
+    typedef typename SplineTraits<Spline>::BasisDerivativeType BasisDerivativeType;
+    
+    /** \brief The data type representing the spline's control points. */
+    typedef typename SplineTraits<Spline>::ControlPointVectorType ControlPointVectorType;
+    
+    /**
+    * \brief Creates a (constant) zero spline.
+    * For Splines with dynamic degree, the resulting degree will be 0.
+    **/
+    Spline() 
+    : m_knots(1, (Degree==Dynamic ? 2 : 2*Degree+2))
+    , m_ctrls(ControlPointVectorType::Zero(Dimension,(Degree==Dynamic ? 1 : Degree+1))) 
+    {
+      // in theory this code can go to the initializer list but it will get pretty
+      // much unreadable ...
+      enum { MinDegree = (Degree==Dynamic ? 0 : Degree) };
+      m_knots.template segment<MinDegree+1>(0) = Array<Scalar,1,MinDegree+1>::Zero();
+      m_knots.template segment<MinDegree+1>(MinDegree+1) = Array<Scalar,1,MinDegree+1>::Ones();
+    }
+
+    /**
+    * \brief Creates a spline from a knot vector and control points.
+    * \param knots The spline's knot vector.
+    * \param ctrls The spline's control point vector.
+    **/
+    template <typename OtherVectorType, typename OtherArrayType>
+    Spline(const OtherVectorType& knots, const OtherArrayType& ctrls) : m_knots(knots), m_ctrls(ctrls) {}
+
+    /**
+    * \brief Copy constructor for splines.
+    * \param spline The input spline.
+    **/
+    template <int OtherDegree>
+    Spline(const Spline<Scalar, Dimension, OtherDegree>& spline) : 
+    m_knots(spline.knots()), m_ctrls(spline.ctrls()) {}
+
+    /**
+     * \brief Returns the knots of the underlying spline.
+     **/
+    const KnotVectorType& knots() const { return m_knots; }
+    
+    /**
+     * \brief Returns the ctrls of the underlying spline.
+     **/    
+    const ControlPointVectorType& ctrls() const { return m_ctrls; }
+
+    /**
+     * \brief Returns the spline value at a given site \f$u\f$.
+     *
+     * The function returns
+     * \f{align*}
+     *   C(u) & = \sum_{i=0}^{n}N_{i,p}P_i
+     * \f}
+     *
+     * \param u Parameter \f$u \in [0;1]\f$ at which the spline is evaluated.
+     * \return The spline value at the given location \f$u\f$.
+     **/
+    PointType operator()(Scalar u) const;
+
+    /**
+     * \brief Evaluation of spline derivatives of up-to given order.
+     *
+     * The function returns
+     * \f{align*}
+     *   \frac{d^i}{du^i}C(u) & = \sum_{i=0}^{n} \frac{d^i}{du^i} N_{i,p}(u)P_i
+     * \f}
+     * for i ranging between 0 and order.
+     *
+     * \param u Parameter \f$u \in [0;1]\f$ at which the spline derivative is evaluated.
+     * \param order The order up to which the derivatives are computed.
+     **/
+    typename SplineTraits<Spline>::DerivativeType
+      derivatives(Scalar u, DenseIndex order) const;
+
+    /**
+     * \copydoc Spline::derivatives
+     * Using the template version of this function is more efficieent since
+     * temporary objects are allocated on the stack whenever this is possible.
+     **/    
+    template <int DerivativeOrder>
+    typename SplineTraits<Spline,DerivativeOrder>::DerivativeType
+      derivatives(Scalar u, DenseIndex order = DerivativeOrder) const;
+
+    /**
+     * \brief Computes the non-zero basis functions at the given site.
+     *
+     * Splines have local support and a point from their image is defined
+     * by exactly \f$p+1\f$ control points \f$P_i\f$ where \f$p\f$ is the
+     * spline degree.
+     *
+     * This function computes the \f$p+1\f$ non-zero basis function values
+     * for a given parameter value \f$u\f$. It returns
+     * \f{align*}{
+     *   N_{i,p}(u), \hdots, N_{i+p+1,p}(u)
+     * \f}
+     *
+     * \param u Parameter \f$u \in [0;1]\f$ at which the non-zero basis functions 
+     *          are computed.
+     **/
+    typename SplineTraits<Spline>::BasisVectorType
+      basisFunctions(Scalar u) const;
+
+    /**
+     * \brief Computes the non-zero spline basis function derivatives up to given order.
+     *
+     * The function computes
+     * \f{align*}{
+     *   \frac{d^i}{du^i} N_{i,p}(u), \hdots, \frac{d^i}{du^i} N_{i+p+1,p}(u)
+     * \f}
+     * with i ranging from 0 up to the specified order.
+     *
+     * \param u Parameter \f$u \in [0;1]\f$ at which the non-zero basis function
+     *          derivatives are computed.
+     * \param order The order up to which the basis function derivatives are computes.
+     **/
+    typename SplineTraits<Spline>::BasisDerivativeType
+      basisFunctionDerivatives(Scalar u, DenseIndex order) const;
+
+    /**
+     * \copydoc Spline::basisFunctionDerivatives
+     * Using the template version of this function is more efficieent since
+     * temporary objects are allocated on the stack whenever this is possible.
+     **/    
+    template <int DerivativeOrder>
+    typename SplineTraits<Spline,DerivativeOrder>::BasisDerivativeType
+      basisFunctionDerivatives(Scalar u, DenseIndex order = DerivativeOrder) const;
+
+    /**
+     * \brief Returns the spline degree.
+     **/ 
+    DenseIndex degree() const;
+
+    /** 
+     * \brief Returns the span within the knot vector in which u is falling.
+     * \param u The site for which the span is determined.
+     **/
+    DenseIndex span(Scalar u) const;
+
+    /**
+     * \brief Computes the span within the provided knot vector in which u is falling.
+     **/
+    static DenseIndex Span(typename SplineTraits<Spline>::Scalar u, DenseIndex degree, const typename SplineTraits<Spline>::KnotVectorType& knots);
+    
+    /**
+     * \brief Returns the spline's non-zero basis functions.
+     *
+     * The function computes and returns
+     * \f{align*}{
+     *   N_{i,p}(u), \hdots, N_{i+p+1,p}(u)
+     * \f}
+     *
+     * \param u The site at which the basis functions are computed.
+     * \param degree The degree of the underlying spline.
+     * \param knots The underlying spline's knot vector.
+     **/
+    static BasisVectorType BasisFunctions(Scalar u, DenseIndex degree, const KnotVectorType& knots);
+
+    /**
+     * \copydoc Spline::basisFunctionDerivatives
+     * \param degree The degree of the underlying spline
+     * \param knots The underlying spline's knot vector.
+     **/    
+    static BasisDerivativeType BasisFunctionDerivatives(
+      const Scalar u, const DenseIndex order, const DenseIndex degree, const KnotVectorType& knots);
+
+  private:
+    KnotVectorType m_knots; /*!< Knot vector. */
+    ControlPointVectorType  m_ctrls; /*!< Control points. */
+
+    template <typename DerivativeType>
+    static void BasisFunctionDerivativesImpl(
+      const typename Spline<_Scalar, _Dim, _Degree>::Scalar u,
+      const DenseIndex order,
+      const DenseIndex p, 
+      const typename Spline<_Scalar, _Dim, _Degree>::KnotVectorType& U,
+      DerivativeType& N_);
+  };
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  DenseIndex Spline<_Scalar, _Dim, _Degree>::Span(
+    typename SplineTraits< Spline<_Scalar, _Dim, _Degree> >::Scalar u,
+    DenseIndex degree,
+    const typename SplineTraits< Spline<_Scalar, _Dim, _Degree> >::KnotVectorType& knots)
+  {
+    // Piegl & Tiller, "The NURBS Book", A2.1 (p. 68)
+    if (u <= knots(0)) return degree;
+    const Scalar* pos = std::upper_bound(knots.data()+degree-1, knots.data()+knots.size()-degree-1, u);
+    return static_cast<DenseIndex>( std::distance(knots.data(), pos) - 1 );
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  typename Spline<_Scalar, _Dim, _Degree>::BasisVectorType
+    Spline<_Scalar, _Dim, _Degree>::BasisFunctions(
+    typename Spline<_Scalar, _Dim, _Degree>::Scalar u,
+    DenseIndex degree,
+    const typename Spline<_Scalar, _Dim, _Degree>::KnotVectorType& knots)
+  {
+    const DenseIndex p = degree;
+    const DenseIndex i = Spline::Span(u, degree, knots);
+
+    const KnotVectorType& U = knots;
+
+    BasisVectorType left(p+1); left(0) = Scalar(0);
+    BasisVectorType right(p+1); right(0) = Scalar(0);
+
+    VectorBlock<BasisVectorType,Degree>(left,1,p) = u - VectorBlock<const KnotVectorType,Degree>(U,i+1-p,p).reverse();
+    VectorBlock<BasisVectorType,Degree>(right,1,p) = VectorBlock<const KnotVectorType,Degree>(U,i+1,p) - u;
+
+    BasisVectorType N(1,p+1);
+    N(0) = Scalar(1);
+    for (DenseIndex j=1; j<=p; ++j)
+    {
+      Scalar saved = Scalar(0);
+      for (DenseIndex r=0; r<j; r++)
+      {
+        const Scalar tmp = N(r)/(right(r+1)+left(j-r));
+        N[r] = saved + right(r+1)*tmp;
+        saved = left(j-r)*tmp;
+      }
+      N(j) = saved;
+    }
+    return N;
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  DenseIndex Spline<_Scalar, _Dim, _Degree>::degree() const
+  {
+    if (_Degree == Dynamic)
+      return m_knots.size() - m_ctrls.cols() - 1;
+    else
+      return _Degree;
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  DenseIndex Spline<_Scalar, _Dim, _Degree>::span(Scalar u) const
+  {
+    return Spline::Span(u, degree(), knots());
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  typename Spline<_Scalar, _Dim, _Degree>::PointType Spline<_Scalar, _Dim, _Degree>::operator()(Scalar u) const
+  {
+    enum { Order = SplineTraits<Spline>::OrderAtCompileTime };
+
+    const DenseIndex span = this->span(u);
+    const DenseIndex p = degree();
+    const BasisVectorType basis_funcs = basisFunctions(u);
+
+    const Replicate<BasisVectorType,Dimension,1> ctrl_weights(basis_funcs);
+    const Block<const ControlPointVectorType,Dimension,Order> ctrl_pts(ctrls(),0,span-p,Dimension,p+1);
+    return (ctrl_weights * ctrl_pts).rowwise().sum();
+  }
+
+  /* --------------------------------------------------------------------------------------------- */
+
+  template <typename SplineType, typename DerivativeType>
+  void derivativesImpl(const SplineType& spline, typename SplineType::Scalar u, DenseIndex order, DerivativeType& der)
+  {    
+    enum { Dimension = SplineTraits<SplineType>::Dimension };
+    enum { Order = SplineTraits<SplineType>::OrderAtCompileTime };
+    enum { DerivativeOrder = DerivativeType::ColsAtCompileTime };
+
+    typedef typename SplineTraits<SplineType>::ControlPointVectorType ControlPointVectorType;
+    typedef typename SplineTraits<SplineType,DerivativeOrder>::BasisDerivativeType BasisDerivativeType;
+    typedef typename BasisDerivativeType::ConstRowXpr BasisDerivativeRowXpr;    
+
+    const DenseIndex p = spline.degree();
+    const DenseIndex span = spline.span(u);
+
+    const DenseIndex n = (std::min)(p, order);
+
+    der.resize(Dimension,n+1);
+
+    // Retrieve the basis function derivatives up to the desired order...    
+    const BasisDerivativeType basis_func_ders = spline.template basisFunctionDerivatives<DerivativeOrder>(u, n+1);
+
+    // ... and perform the linear combinations of the control points.
+    for (DenseIndex der_order=0; der_order<n+1; ++der_order)
+    {
+      const Replicate<BasisDerivativeRowXpr,Dimension,1> ctrl_weights( basis_func_ders.row(der_order) );
+      const Block<const ControlPointVectorType,Dimension,Order> ctrl_pts(spline.ctrls(),0,span-p,Dimension,p+1);
+      der.col(der_order) = (ctrl_weights * ctrl_pts).rowwise().sum();
+    }
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  typename SplineTraits< Spline<_Scalar, _Dim, _Degree> >::DerivativeType
+    Spline<_Scalar, _Dim, _Degree>::derivatives(Scalar u, DenseIndex order) const
+  {
+    typename SplineTraits< Spline >::DerivativeType res;
+    derivativesImpl(*this, u, order, res);
+    return res;
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  template <int DerivativeOrder>
+  typename SplineTraits< Spline<_Scalar, _Dim, _Degree>, DerivativeOrder >::DerivativeType
+    Spline<_Scalar, _Dim, _Degree>::derivatives(Scalar u, DenseIndex order) const
+  {
+    typename SplineTraits< Spline, DerivativeOrder >::DerivativeType res;
+    derivativesImpl(*this, u, order, res);
+    return res;
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  typename SplineTraits< Spline<_Scalar, _Dim, _Degree> >::BasisVectorType
+    Spline<_Scalar, _Dim, _Degree>::basisFunctions(Scalar u) const
+  {
+    return Spline::BasisFunctions(u, degree(), knots());
+  }
+
+  /* --------------------------------------------------------------------------------------------- */
+  
+  
+  template <typename _Scalar, int _Dim, int _Degree>
+  template <typename DerivativeType>
+  void Spline<_Scalar, _Dim, _Degree>::BasisFunctionDerivativesImpl(
+    const typename Spline<_Scalar, _Dim, _Degree>::Scalar u,
+    const DenseIndex order,
+    const DenseIndex p, 
+    const typename Spline<_Scalar, _Dim, _Degree>::KnotVectorType& U,
+    DerivativeType& N_)
+  {
+    typedef Spline<_Scalar, _Dim, _Degree> SplineType;
+    enum { Order = SplineTraits<SplineType>::OrderAtCompileTime };
+
+    const DenseIndex span = SplineType::Span(u, p, U);
+
+    const DenseIndex n = (std::min)(p, order);
+
+    N_.resize(n+1, p+1);
+
+    BasisVectorType left = BasisVectorType::Zero(p+1);
+    BasisVectorType right = BasisVectorType::Zero(p+1);
+
+    Matrix<Scalar,Order,Order> ndu(p+1,p+1);
+
+    Scalar saved, temp; // FIXME These were double instead of Scalar. Was there a reason for that?
+
+    ndu(0,0) = 1.0;
+
+    DenseIndex j;
+    for (j=1; j<=p; ++j)
+    {
+      left[j] = u-U[span+1-j];
+      right[j] = U[span+j]-u;
+      saved = 0.0;
+
+      for (DenseIndex r=0; r<j; ++r)
+      {
+        /* Lower triangle */
+        ndu(j,r) = right[r+1]+left[j-r];
+        temp = ndu(r,j-1)/ndu(j,r);
+        /* Upper triangle */
+        ndu(r,j) = static_cast<Scalar>(saved+right[r+1] * temp);
+        saved = left[j-r] * temp;
+      }
+
+      ndu(j,j) = static_cast<Scalar>(saved);
+    }
+
+    for (j = p; j>=0; --j) 
+      N_(0,j) = ndu(j,p);
+
+    // Compute the derivatives
+    DerivativeType a(n+1,p+1);
+    DenseIndex r=0;
+    for (; r<=p; ++r)
+    {
+      DenseIndex s1,s2;
+      s1 = 0; s2 = 1; // alternate rows in array a
+      a(0,0) = 1.0;
+
+      // Compute the k-th derivative
+      for (DenseIndex k=1; k<=static_cast<DenseIndex>(n); ++k)
+      {
+        Scalar d = 0.0;
+        DenseIndex rk,pk,j1,j2;
+        rk = r-k; pk = p-k;
+
+        if (r>=k)
+        {
+          a(s2,0) = a(s1,0)/ndu(pk+1,rk);
+          d = a(s2,0)*ndu(rk,pk);
+        }
+
+        if (rk>=-1) j1 = 1;
+        else        j1 = -rk;
+
+        if (r-1 <= pk) j2 = k-1;
+        else           j2 = p-r;
+
+        for (j=j1; j<=j2; ++j)
+        {
+          a(s2,j) = (a(s1,j)-a(s1,j-1))/ndu(pk+1,rk+j);
+          d += a(s2,j)*ndu(rk+j,pk);
+        }
+
+        if (r<=pk)
+        {
+          a(s2,k) = -a(s1,k-1)/ndu(pk+1,r);
+          d += a(s2,k)*ndu(r,pk);
+        }
+
+        N_(k,r) = static_cast<Scalar>(d);
+        j = s1; s1 = s2; s2 = j; // Switch rows
+      }
+    }
+
+    /* Multiply through by the correct factors */
+    /* (Eq. [2.9])                             */
+    r = p;
+    for (DenseIndex k=1; k<=static_cast<DenseIndex>(n); ++k)
+    {
+      for (j=p; j>=0; --j) N_(k,j) *= r;
+      r *= p-k;
+    }
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  typename SplineTraits< Spline<_Scalar, _Dim, _Degree> >::BasisDerivativeType
+    Spline<_Scalar, _Dim, _Degree>::basisFunctionDerivatives(Scalar u, DenseIndex order) const
+  {
+    typename SplineTraits<Spline<_Scalar, _Dim, _Degree> >::BasisDerivativeType der;
+    BasisFunctionDerivativesImpl(u, order, degree(), knots(), der);
+    return der;
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  template <int DerivativeOrder>
+  typename SplineTraits< Spline<_Scalar, _Dim, _Degree>, DerivativeOrder >::BasisDerivativeType
+    Spline<_Scalar, _Dim, _Degree>::basisFunctionDerivatives(Scalar u, DenseIndex order) const
+  {
+    typename SplineTraits< Spline<_Scalar, _Dim, _Degree>, DerivativeOrder >::BasisDerivativeType der;
+    BasisFunctionDerivativesImpl(u, order, degree(), knots(), der);
+    return der;
+  }
+
+  template <typename _Scalar, int _Dim, int _Degree>
+  typename SplineTraits<Spline<_Scalar, _Dim, _Degree> >::BasisDerivativeType
+  Spline<_Scalar, _Dim, _Degree>::BasisFunctionDerivatives(
+    const typename Spline<_Scalar, _Dim, _Degree>::Scalar u,
+    const DenseIndex order,
+    const DenseIndex degree,
+    const typename Spline<_Scalar, _Dim, _Degree>::KnotVectorType& knots)
+  {
+    typename SplineTraits<Spline>::BasisDerivativeType der;
+    BasisFunctionDerivativesImpl(u, order, degree, knots, der);
+    return der;
+  }
+}
+
+#endif // EIGEN_SPLINE_H
diff --git a/src/EigenUnsupported/src/Splines/SplineFitting.h b/src/EigenUnsupported/src/Splines/SplineFitting.h
new file mode 100644
index 0000000..9f6e8af
--- /dev/null
+++ b/src/EigenUnsupported/src/Splines/SplineFitting.h
@@ -0,0 +1,431 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 20010-2011 Hauke Heibel <hauke.heibel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPLINE_FITTING_H
+#define EIGEN_SPLINE_FITTING_H
+
+#include <algorithm>
+#include <functional>
+#include <numeric>
+#include <vector>
+
+#include "SplineFwd.h"
+
+#include "../../../../Eigen/LU"
+#include "../../../../Eigen/QR"
+
+namespace Eigen
+{
+  /**
+   * \brief Computes knot averages.
+   * \ingroup Splines_Module
+   *
+   * The knots are computed as
+   * \f{align*}
+   *  u_0 & = \hdots = u_p = 0 \\
+   *  u_{m-p} & = \hdots = u_{m} = 1 \\
+   *  u_{j+p} & = \frac{1}{p}\sum_{i=j}^{j+p-1}\bar{u}_i \quad\quad j=1,\hdots,n-p
+   * \f}
+   * where \f$p\f$ is the degree and \f$m+1\f$ the number knots
+   * of the desired interpolating spline.
+   *
+   * \param[in] parameters The input parameters. During interpolation one for each data point.
+   * \param[in] degree The spline degree which is used during the interpolation.
+   * \param[out] knots The output knot vector.
+   *
+   * \sa Les Piegl and Wayne Tiller, The NURBS book (2nd ed.), 1997, 9.2.1 Global Curve Interpolation to Point Data
+   **/
+  template <typename KnotVectorType>
+  void KnotAveraging(const KnotVectorType& parameters, DenseIndex degree, KnotVectorType& knots)
+  {
+    knots.resize(parameters.size()+degree+1);      
+
+    for (DenseIndex j=1; j<parameters.size()-degree; ++j)
+      knots(j+degree) = parameters.segment(j,degree).mean();
+
+    knots.segment(0,degree+1) = KnotVectorType::Zero(degree+1);
+    knots.segment(knots.size()-degree-1,degree+1) = KnotVectorType::Ones(degree+1);
+  }
+
+  /**
+   * \brief Computes knot averages when derivative constraints are present.
+   * Note that this is a technical interpretation of the referenced article
+   * since the algorithm contained therein is incorrect as written.
+   * \ingroup Splines_Module
+   *
+   * \param[in] parameters The parameters at which the interpolation B-Spline
+   *            will intersect the given interpolation points. The parameters
+   *            are assumed to be a non-decreasing sequence.
+   * \param[in] degree The degree of the interpolating B-Spline. This must be
+   *            greater than zero.
+   * \param[in] derivativeIndices The indices corresponding to parameters at
+   *            which there are derivative constraints. The indices are assumed
+   *            to be a non-decreasing sequence.
+   * \param[out] knots The calculated knot vector. These will be returned as a
+   *             non-decreasing sequence
+   *
+   * \sa Les A. Piegl, Khairan Rajab, Volha Smarodzinana. 2008.
+   * Curve interpolation with directional constraints for engineering design. 
+   * Engineering with Computers
+   **/
+  template <typename KnotVectorType, typename ParameterVectorType, typename IndexArray>
+  void KnotAveragingWithDerivatives(const ParameterVectorType& parameters,
+                                    const unsigned int degree,
+                                    const IndexArray& derivativeIndices,
+                                    KnotVectorType& knots)
+  {
+    typedef typename ParameterVectorType::Scalar Scalar;
+
+    DenseIndex numParameters = parameters.size();
+    DenseIndex numDerivatives = derivativeIndices.size();
+
+    if (numDerivatives < 1)
+    {
+      KnotAveraging(parameters, degree, knots);
+      return;
+    }
+
+    DenseIndex startIndex;
+    DenseIndex endIndex;
+  
+    DenseIndex numInternalDerivatives = numDerivatives;
+    
+    if (derivativeIndices[0] == 0)
+    {
+      startIndex = 0;
+      --numInternalDerivatives;
+    }
+    else
+    {
+      startIndex = 1;
+    }
+    if (derivativeIndices[numDerivatives - 1] == numParameters - 1)
+    {
+      endIndex = numParameters - degree;
+      --numInternalDerivatives;
+    }
+    else
+    {
+      endIndex = numParameters - degree - 1;
+    }
+
+    // There are (endIndex - startIndex + 1) knots obtained from the averaging
+    // and 2 for the first and last parameters.
+    DenseIndex numAverageKnots = endIndex - startIndex + 3;
+    KnotVectorType averageKnots(numAverageKnots);
+    averageKnots[0] = parameters[0];
+
+    int newKnotIndex = 0;
+    for (DenseIndex i = startIndex; i <= endIndex; ++i)
+      averageKnots[++newKnotIndex] = parameters.segment(i, degree).mean();
+    averageKnots[++newKnotIndex] = parameters[numParameters - 1];
+
+    newKnotIndex = -1;
+  
+    ParameterVectorType temporaryParameters(numParameters + 1);
+    KnotVectorType derivativeKnots(numInternalDerivatives);
+    for (DenseIndex i = 0; i < numAverageKnots - 1; ++i)
+    {
+      temporaryParameters[0] = averageKnots[i];
+      ParameterVectorType parameterIndices(numParameters);
+      int temporaryParameterIndex = 1;
+      for (DenseIndex j = 0; j < numParameters; ++j)
+      {
+        Scalar parameter = parameters[j];
+        if (parameter >= averageKnots[i] && parameter < averageKnots[i + 1])
+        {
+          parameterIndices[temporaryParameterIndex] = j;
+          temporaryParameters[temporaryParameterIndex++] = parameter;
+        }
+      }
+      temporaryParameters[temporaryParameterIndex] = averageKnots[i + 1];
+
+      for (int j = 0; j <= temporaryParameterIndex - 2; ++j)
+      {
+        for (DenseIndex k = 0; k < derivativeIndices.size(); ++k)
+        {
+          if (parameterIndices[j + 1] == derivativeIndices[k]
+              && parameterIndices[j + 1] != 0
+              && parameterIndices[j + 1] != numParameters - 1)
+          {
+            derivativeKnots[++newKnotIndex] = temporaryParameters.segment(j, 3).mean();
+            break;
+          }
+        }
+      }
+    }
+    
+    KnotVectorType temporaryKnots(averageKnots.size() + derivativeKnots.size());
+
+    std::merge(averageKnots.data(), averageKnots.data() + averageKnots.size(),
+               derivativeKnots.data(), derivativeKnots.data() + derivativeKnots.size(),
+               temporaryKnots.data());
+
+    // Number of knots (one for each point and derivative) plus spline order.
+    DenseIndex numKnots = numParameters + numDerivatives + degree + 1;
+    knots.resize(numKnots);
+
+    knots.head(degree).fill(temporaryKnots[0]);
+    knots.tail(degree).fill(temporaryKnots.template tail<1>()[0]);
+    knots.segment(degree, temporaryKnots.size()) = temporaryKnots;
+  }
+
+  /**
+   * \brief Computes chord length parameters which are required for spline interpolation.
+   * \ingroup Splines_Module
+   *
+   * \param[in] pts The data points to which a spline should be fit.
+   * \param[out] chord_lengths The resulting chord length vector.
+   *
+   * \sa Les Piegl and Wayne Tiller, The NURBS book (2nd ed.), 1997, 9.2.1 Global Curve Interpolation to Point Data
+   **/   
+  template <typename PointArrayType, typename KnotVectorType>
+  void ChordLengths(const PointArrayType& pts, KnotVectorType& chord_lengths)
+  {
+    typedef typename KnotVectorType::Scalar Scalar;
+
+    const DenseIndex n = pts.cols();
+
+    // 1. compute the column-wise norms
+    chord_lengths.resize(pts.cols());
+    chord_lengths[0] = 0;
+    chord_lengths.rightCols(n-1) = (pts.array().leftCols(n-1) - pts.array().rightCols(n-1)).matrix().colwise().norm();
+
+    // 2. compute the partial sums
+    std::partial_sum(chord_lengths.data(), chord_lengths.data()+n, chord_lengths.data());
+
+    // 3. normalize the data
+    chord_lengths /= chord_lengths(n-1);
+    chord_lengths(n-1) = Scalar(1);
+  }
+
+  /**
+   * \brief Spline fitting methods.
+   * \ingroup Splines_Module
+   **/     
+  template <typename SplineType>
+  struct SplineFitting
+  {
+    typedef typename SplineType::KnotVectorType KnotVectorType;
+    typedef typename SplineType::ParameterVectorType ParameterVectorType;
+
+    /**
+     * \brief Fits an interpolating Spline to the given data points.
+     *
+     * \param pts The points for which an interpolating spline will be computed.
+     * \param degree The degree of the interpolating spline.
+     *
+     * \returns A spline interpolating the initially provided points.
+     **/
+    template <typename PointArrayType>
+    static SplineType Interpolate(const PointArrayType& pts, DenseIndex degree);
+
+    /**
+     * \brief Fits an interpolating Spline to the given data points.
+     *
+     * \param pts The points for which an interpolating spline will be computed.
+     * \param degree The degree of the interpolating spline.
+     * \param knot_parameters The knot parameters for the interpolation.
+     *
+     * \returns A spline interpolating the initially provided points.
+     **/
+    template <typename PointArrayType>
+    static SplineType Interpolate(const PointArrayType& pts, DenseIndex degree, const KnotVectorType& knot_parameters);
+
+    /**
+     * \brief Fits an interpolating spline to the given data points and
+     * derivatives.
+     * 
+     * \param points The points for which an interpolating spline will be computed.
+     * \param derivatives The desired derivatives of the interpolating spline at interpolation
+     *                    points.
+     * \param derivativeIndices An array indicating which point each derivative belongs to. This
+     *                          must be the same size as @a derivatives.
+     * \param degree The degree of the interpolating spline.
+     *
+     * \returns A spline interpolating @a points with @a derivatives at those points.
+     *
+     * \sa Les A. Piegl, Khairan Rajab, Volha Smarodzinana. 2008.
+     * Curve interpolation with directional constraints for engineering design. 
+     * Engineering with Computers
+     **/
+    template <typename PointArrayType, typename IndexArray>
+    static SplineType InterpolateWithDerivatives(const PointArrayType& points,
+                                                 const PointArrayType& derivatives,
+                                                 const IndexArray& derivativeIndices,
+                                                 const unsigned int degree);
+
+    /**
+     * \brief Fits an interpolating spline to the given data points and derivatives.
+     * 
+     * \param points The points for which an interpolating spline will be computed.
+     * \param derivatives The desired derivatives of the interpolating spline at interpolation points.
+     * \param derivativeIndices An array indicating which point each derivative belongs to. This
+     *                          must be the same size as @a derivatives.
+     * \param degree The degree of the interpolating spline.
+     * \param parameters The parameters corresponding to the interpolation points.
+     *
+     * \returns A spline interpolating @a points with @a derivatives at those points.
+     *
+     * \sa Les A. Piegl, Khairan Rajab, Volha Smarodzinana. 2008.
+     * Curve interpolation with directional constraints for engineering design. 
+     * Engineering with Computers
+     */
+    template <typename PointArrayType, typename IndexArray>
+    static SplineType InterpolateWithDerivatives(const PointArrayType& points,
+                                                 const PointArrayType& derivatives,
+                                                 const IndexArray& derivativeIndices,
+                                                 const unsigned int degree,
+                                                 const ParameterVectorType& parameters);
+  };
+
+  template <typename SplineType>
+  template <typename PointArrayType>
+  SplineType SplineFitting<SplineType>::Interpolate(const PointArrayType& pts, DenseIndex degree, const KnotVectorType& knot_parameters)
+  {
+    typedef typename SplineType::KnotVectorType::Scalar Scalar;      
+    typedef typename SplineType::ControlPointVectorType ControlPointVectorType;      
+
+    typedef Matrix<Scalar,Dynamic,Dynamic> MatrixType;
+
+    KnotVectorType knots;
+    KnotAveraging(knot_parameters, degree, knots);
+
+    DenseIndex n = pts.cols();
+    MatrixType A = MatrixType::Zero(n,n);
+    for (DenseIndex i=1; i<n-1; ++i)
+    {
+      const DenseIndex span = SplineType::Span(knot_parameters[i], degree, knots);
+
+      // The segment call should somehow be told the spline order at compile time.
+      A.row(i).segment(span-degree, degree+1) = SplineType::BasisFunctions(knot_parameters[i], degree, knots);
+    }
+    A(0,0) = 1.0;
+    A(n-1,n-1) = 1.0;
+
+    HouseholderQR<MatrixType> qr(A);
+
+    // Here, we are creating a temporary due to an Eigen issue.
+    ControlPointVectorType ctrls = qr.solve(MatrixType(pts.transpose())).transpose();
+
+    return SplineType(knots, ctrls);
+  }
+
+  template <typename SplineType>
+  template <typename PointArrayType>
+  SplineType SplineFitting<SplineType>::Interpolate(const PointArrayType& pts, DenseIndex degree)
+  {
+    KnotVectorType chord_lengths; // knot parameters
+    ChordLengths(pts, chord_lengths);
+    return Interpolate(pts, degree, chord_lengths);
+  }
+  
+  template <typename SplineType>
+  template <typename PointArrayType, typename IndexArray>
+  SplineType 
+  SplineFitting<SplineType>::InterpolateWithDerivatives(const PointArrayType& points,
+                                                        const PointArrayType& derivatives,
+                                                        const IndexArray& derivativeIndices,
+                                                        const unsigned int degree,
+                                                        const ParameterVectorType& parameters)
+  {
+    typedef typename SplineType::KnotVectorType::Scalar Scalar;      
+    typedef typename SplineType::ControlPointVectorType ControlPointVectorType;
+
+    typedef Matrix<Scalar, Dynamic, Dynamic> MatrixType;
+
+    const DenseIndex n = points.cols() + derivatives.cols();
+    
+    KnotVectorType knots;
+
+    KnotAveragingWithDerivatives(parameters, degree, derivativeIndices, knots);
+    
+    // fill matrix
+    MatrixType A = MatrixType::Zero(n, n);
+
+    // Use these dimensions for quicker populating, then transpose for solving.
+    MatrixType b(points.rows(), n);
+
+    DenseIndex startRow;
+    DenseIndex derivativeStart;
+
+    // End derivatives.
+    if (derivativeIndices[0] == 0)
+    {
+      A.template block<1, 2>(1, 0) << -1, 1;
+      
+      Scalar y = (knots(degree + 1) - knots(0)) / degree;
+      b.col(1) = y*derivatives.col(0);
+      
+      startRow = 2;
+      derivativeStart = 1;
+    }
+    else
+    {
+      startRow = 1;
+      derivativeStart = 0;
+    }
+    if (derivativeIndices[derivatives.cols() - 1] == points.cols() - 1)
+    {
+      A.template block<1, 2>(n - 2, n - 2) << -1, 1;
+
+      Scalar y = (knots(knots.size() - 1) - knots(knots.size() - (degree + 2))) / degree;
+      b.col(b.cols() - 2) = y*derivatives.col(derivatives.cols() - 1);
+    }
+    
+    DenseIndex row = startRow;
+    DenseIndex derivativeIndex = derivativeStart;
+    for (DenseIndex i = 1; i < parameters.size() - 1; ++i)
+    {
+      const DenseIndex span = SplineType::Span(parameters[i], degree, knots);
+
+      if (derivativeIndex < derivativeIndices.size() && derivativeIndices[derivativeIndex] == i)
+      {
+        A.block(row, span - degree, 2, degree + 1)
+          = SplineType::BasisFunctionDerivatives(parameters[i], 1, degree, knots);
+
+        b.col(row++) = points.col(i);
+        b.col(row++) = derivatives.col(derivativeIndex++);
+      }
+      else
+      {
+        A.row(row).segment(span - degree, degree + 1)
+          = SplineType::BasisFunctions(parameters[i], degree, knots);
+        b.col(row++) = points.col(i);
+      }
+    }
+    b.col(0) = points.col(0);
+    b.col(b.cols() - 1) = points.col(points.cols() - 1);
+    A(0,0) = 1;
+    A(n - 1, n - 1) = 1;
+    
+    // Solve
+    FullPivLU<MatrixType> lu(A);
+    ControlPointVectorType controlPoints = lu.solve(MatrixType(b.transpose())).transpose();
+
+    SplineType spline(knots, controlPoints);
+    
+    return spline;
+  }
+  
+  template <typename SplineType>
+  template <typename PointArrayType, typename IndexArray>
+  SplineType
+  SplineFitting<SplineType>::InterpolateWithDerivatives(const PointArrayType& points,
+                                                        const PointArrayType& derivatives,
+                                                        const IndexArray& derivativeIndices,
+                                                        const unsigned int degree)
+  {
+    ParameterVectorType parameters;
+    ChordLengths(points, parameters);
+    return InterpolateWithDerivatives(points, derivatives, derivativeIndices, degree, parameters);
+  }
+}
+
+#endif // EIGEN_SPLINE_FITTING_H
diff --git a/src/EigenUnsupported/src/Splines/SplineFwd.h b/src/EigenUnsupported/src/Splines/SplineFwd.h
new file mode 100644
index 0000000..00d6b49
--- /dev/null
+++ b/src/EigenUnsupported/src/Splines/SplineFwd.h
@@ -0,0 +1,93 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 20010-2011 Hauke Heibel <hauke.heibel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPLINES_FWD_H
+#define EIGEN_SPLINES_FWD_H
+
+#include "../../../../Eigen/Core"
+
+namespace Eigen
+{
+    template <typename Scalar, int Dim, int Degree = Dynamic> class Spline;
+
+    template < typename SplineType, int DerivativeOrder = Dynamic > struct SplineTraits {};
+
+    /**
+     * \ingroup Splines_Module
+     * \brief Compile-time attributes of the Spline class for Dynamic degree.
+     **/
+    template <typename _Scalar, int _Dim, int _Degree>
+    struct SplineTraits< Spline<_Scalar, _Dim, _Degree>, Dynamic >
+    {
+      typedef _Scalar Scalar; /*!< The spline curve's scalar type. */
+      enum { Dimension = _Dim /*!< The spline curve's dimension. */ };
+      enum { Degree = _Degree /*!< The spline curve's degree. */ };
+
+      enum { OrderAtCompileTime = _Degree==Dynamic ? Dynamic : _Degree+1 /*!< The spline curve's order at compile-time. */ };
+      enum { NumOfDerivativesAtCompileTime = OrderAtCompileTime /*!< The number of derivatives defined for the current spline. */ };
+      
+      enum { DerivativeMemoryLayout = Dimension==1 ? RowMajor : ColMajor /*!< The derivative type's memory layout. */ };
+
+      /** \brief The data type used to store non-zero basis functions. */
+      typedef Array<Scalar,1,OrderAtCompileTime> BasisVectorType;
+
+      /** \brief The data type used to store the values of the basis function derivatives. */
+      typedef Array<Scalar,Dynamic,Dynamic,RowMajor,NumOfDerivativesAtCompileTime,OrderAtCompileTime> BasisDerivativeType;
+      
+      /** \brief The data type used to store the spline's derivative values. */
+      typedef Array<Scalar,Dimension,Dynamic,DerivativeMemoryLayout,Dimension,NumOfDerivativesAtCompileTime> DerivativeType;
+
+      /** \brief The point type the spline is representing. */
+      typedef Array<Scalar,Dimension,1> PointType;
+      
+      /** \brief The data type used to store knot vectors. */
+      typedef Array<Scalar,1,Dynamic> KnotVectorType;
+
+      /** \brief The data type used to store parameter vectors. */
+      typedef Array<Scalar,1,Dynamic> ParameterVectorType;
+      
+      /** \brief The data type representing the spline's control points. */
+      typedef Array<Scalar,Dimension,Dynamic> ControlPointVectorType;
+    };
+
+    /**
+     * \ingroup Splines_Module
+     * \brief Compile-time attributes of the Spline class for fixed degree.
+     *
+     * The traits class inherits all attributes from the SplineTraits of Dynamic degree.
+     **/
+    template < typename _Scalar, int _Dim, int _Degree, int _DerivativeOrder >
+    struct SplineTraits< Spline<_Scalar, _Dim, _Degree>, _DerivativeOrder > : public SplineTraits< Spline<_Scalar, _Dim, _Degree> >
+    {
+      enum { OrderAtCompileTime = _Degree==Dynamic ? Dynamic : _Degree+1 /*!< The spline curve's order at compile-time. */ };
+      enum { NumOfDerivativesAtCompileTime = _DerivativeOrder==Dynamic ? Dynamic : _DerivativeOrder+1 /*!< The number of derivatives defined for the current spline. */ };
+      
+      enum { DerivativeMemoryLayout = _Dim==1 ? RowMajor : ColMajor /*!< The derivative type's memory layout. */ };
+
+      /** \brief The data type used to store the values of the basis function derivatives. */
+      typedef Array<_Scalar,Dynamic,Dynamic,RowMajor,NumOfDerivativesAtCompileTime,OrderAtCompileTime> BasisDerivativeType;
+      
+      /** \brief The data type used to store the spline's derivative values. */      
+      typedef Array<_Scalar,_Dim,Dynamic,DerivativeMemoryLayout,_Dim,NumOfDerivativesAtCompileTime> DerivativeType;
+    };
+
+    /** \brief 2D float B-spline with dynamic degree. */
+    typedef Spline<float,2> Spline2f;
+    
+    /** \brief 3D float B-spline with dynamic degree. */
+    typedef Spline<float,3> Spline3f;
+
+    /** \brief 2D double B-spline with dynamic degree. */
+    typedef Spline<double,2> Spline2d;
+    
+    /** \brief 3D double B-spline with dynamic degree. */
+    typedef Spline<double,3> Spline3d;
+}
+
+#endif // EIGEN_SPLINES_FWD_H