From eda5bc26f44ee9a6f83dcf8c91f17296d7fc509d Mon Sep 17 00:00:00 2001 From: Nao Pross Date: Mon, 12 Feb 2024 14:52:43 +0100 Subject: Move into version control --- .../include/armadillo_bits/mul_gemm_mixed.hpp | 291 +++++++++++++++++++++ 1 file changed, 291 insertions(+) create mode 100644 src/armadillo/include/armadillo_bits/mul_gemm_mixed.hpp (limited to 'src/armadillo/include/armadillo_bits/mul_gemm_mixed.hpp') diff --git a/src/armadillo/include/armadillo_bits/mul_gemm_mixed.hpp b/src/armadillo/include/armadillo_bits/mul_gemm_mixed.hpp new file mode 100644 index 0000000..749cdb1 --- /dev/null +++ b/src/armadillo/include/armadillo_bits/mul_gemm_mixed.hpp @@ -0,0 +1,291 @@ +// SPDX-License-Identifier: Apache-2.0 +// +// Copyright 2008-2016 Conrad Sanderson (http://conradsanderson.id.au) +// Copyright 2008-2016 National ICT Australia (NICTA) +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +// ------------------------------------------------------------------------ + + +//! \addtogroup gemm_mixed +//! @{ + + + +//! \brief +//! Matrix multplication where the matrices have differing element types. +//! Uses caching for speedup. +//! Matrix 'C' is assumed to have been set to the correct size (ie. taking into account transposes) + +template +class gemm_mixed_large + { + public: + + template + arma_hot + inline + static + void + apply + ( + Mat& C, + const Mat& A, + const Mat& B, + const out_eT alpha = out_eT(1), + const out_eT beta = out_eT(0) + ) + { + arma_extra_debug_sigprint(); + + const uword A_n_rows = A.n_rows; + const uword A_n_cols = A.n_cols; + + const uword B_n_rows = B.n_rows; + const uword B_n_cols = B.n_cols; + + if( (do_trans_A == false) && (do_trans_B == false) ) + { + podarray tmp(A_n_cols); + in_eT1* A_rowdata = tmp.memptr(); + + #if defined(ARMA_USE_OPENMP) + const bool use_mp = (B_n_cols >= 2) && (B.n_elem >= 8192) && (mp_thread_limit::in_parallel() == false); + #else + const bool use_mp = false; + #endif + + if(use_mp) + { + #if defined(ARMA_USE_OPENMP) + { + const int n_threads = int( (std::min)( uword(mp_thread_limit::get()), uword(B_n_cols) ) ); + + for(uword row_A=0; row_A < A_n_rows; ++row_A) + { + tmp.copy_row(A, row_A); + + #pragma omp parallel for schedule(static) num_threads(n_threads) + for(uword col_B=0; col_B < B_n_cols; ++col_B) + { + const in_eT2* B_coldata = B.colptr(col_B); + + out_eT acc = out_eT(0); + for(uword i=0; i < B_n_rows; ++i) + { + acc += upgrade_val::apply(A_rowdata[i]) * upgrade_val::apply(B_coldata[i]); + } + + if( (use_alpha == false) && (use_beta == false) ) { C.at(row_A,col_B) = acc; } + else if( (use_alpha == true ) && (use_beta == false) ) { C.at(row_A,col_B) = alpha*acc; } + else if( (use_alpha == false) && (use_beta == true ) ) { C.at(row_A,col_B) = acc + beta*C.at(row_A,col_B); } + else if( (use_alpha == true ) && (use_beta == true ) ) { C.at(row_A,col_B) = alpha*acc + beta*C.at(row_A,col_B); } + } + } + } + #endif + } + else + { + for(uword row_A=0; row_A < A_n_rows; ++row_A) + { + tmp.copy_row(A, row_A); + + for(uword col_B=0; col_B < B_n_cols; ++col_B) + { + const in_eT2* B_coldata = B.colptr(col_B); + + out_eT acc = out_eT(0); + for(uword i=0; i < B_n_rows; ++i) + { + acc += upgrade_val::apply(A_rowdata[i]) * upgrade_val::apply(B_coldata[i]); + } + + if( (use_alpha == false) && (use_beta == false) ) { C.at(row_A,col_B) = acc; } + else if( (use_alpha == true ) && (use_beta == false) ) { C.at(row_A,col_B) = alpha*acc; } + else if( (use_alpha == false) && (use_beta == true ) ) { C.at(row_A,col_B) = acc + beta*C.at(row_A,col_B); } + else if( (use_alpha == true ) && (use_beta == true ) ) { C.at(row_A,col_B) = alpha*acc + beta*C.at(row_A,col_B); } + } + } + } + } + else + if( (do_trans_A == true) && (do_trans_B == false) ) + { + #if defined(ARMA_USE_OPENMP) + const bool use_mp = (B_n_cols >= 2) && (B.n_elem >= 8192) && (mp_thread_limit::in_parallel() == false); + #else + const bool use_mp = false; + #endif + + if(use_mp) + { + #if defined(ARMA_USE_OPENMP) + { + const int n_threads = int( (std::min)( uword(mp_thread_limit::get()), uword(B_n_cols) ) ); + + for(uword col_A=0; col_A < A_n_cols; ++col_A) + { + // col_A is interpreted as row_A when storing the results in matrix C + + const in_eT1* A_coldata = A.colptr(col_A); + + #pragma omp parallel for schedule(static) num_threads(n_threads) + for(uword col_B=0; col_B < B_n_cols; ++col_B) + { + const in_eT2* B_coldata = B.colptr(col_B); + + out_eT acc = out_eT(0); + for(uword i=0; i < B_n_rows; ++i) + { + acc += upgrade_val::apply(A_coldata[i]) * upgrade_val::apply(B_coldata[i]); + } + + if( (use_alpha == false) && (use_beta == false) ) { C.at(col_A,col_B) = acc; } + else if( (use_alpha == true ) && (use_beta == false) ) { C.at(col_A,col_B) = alpha*acc; } + else if( (use_alpha == false) && (use_beta == true ) ) { C.at(col_A,col_B) = acc + beta*C.at(col_A,col_B); } + else if( (use_alpha == true ) && (use_beta == true ) ) { C.at(col_A,col_B) = alpha*acc + beta*C.at(col_A,col_B); } + } + } + } + #endif + } + else + { + for(uword col_A=0; col_A < A_n_cols; ++col_A) + { + // col_A is interpreted as row_A when storing the results in matrix C + + const in_eT1* A_coldata = A.colptr(col_A); + + for(uword col_B=0; col_B < B_n_cols; ++col_B) + { + const in_eT2* B_coldata = B.colptr(col_B); + + out_eT acc = out_eT(0); + for(uword i=0; i < B_n_rows; ++i) + { + acc += upgrade_val::apply(A_coldata[i]) * upgrade_val::apply(B_coldata[i]); + } + + if( (use_alpha == false) && (use_beta == false) ) { C.at(col_A,col_B) = acc; } + else if( (use_alpha == true ) && (use_beta == false) ) { C.at(col_A,col_B) = alpha*acc; } + else if( (use_alpha == false) && (use_beta == true ) ) { C.at(col_A,col_B) = acc + beta*C.at(col_A,col_B); } + else if( (use_alpha == true ) && (use_beta == true ) ) { C.at(col_A,col_B) = alpha*acc + beta*C.at(col_A,col_B); } + } + } + } + } + else + if( (do_trans_A == false) && (do_trans_B == true) ) + { + Mat B_tmp; + + op_strans::apply_mat_noalias(B_tmp, B); + + gemm_mixed_large::apply(C, A, B_tmp, alpha, beta); + } + else + if( (do_trans_A == true) && (do_trans_B == true) ) + { + // mat B_tmp = trans(B); + // dgemm_arma::apply(C, A, B_tmp, alpha, beta); + + + // By using the trans(A)*trans(B) = trans(B*A) equivalency, + // transpose operations are not needed + + podarray tmp(B_n_cols); + in_eT2* B_rowdata = tmp.memptr(); + + for(uword row_B=0; row_B < B_n_rows; ++row_B) + { + tmp.copy_row(B, row_B); + + for(uword col_A=0; col_A < A_n_cols; ++col_A) + { + const in_eT1* A_coldata = A.colptr(col_A); + + out_eT acc = out_eT(0); + for(uword i=0; i < A_n_rows; ++i) + { + acc += upgrade_val::apply(B_rowdata[i]) * upgrade_val::apply(A_coldata[i]); + } + + if( (use_alpha == false) && (use_beta == false) ) { C.at(col_A,row_B) = acc; } + else if( (use_alpha == true ) && (use_beta == false) ) { C.at(col_A,row_B) = alpha*acc; } + else if( (use_alpha == false) && (use_beta == true ) ) { C.at(col_A,row_B) = acc + beta*C.at(col_A,row_B); } + else if( (use_alpha == true ) && (use_beta == true ) ) { C.at(col_A,row_B) = alpha*acc + beta*C.at(col_A,row_B); } + } + } + + } + } + + }; + + + +//! \brief +//! Matrix multplication where the matrices have differing element types. + +template +class gemm_mixed + { + public: + + //! immediate multiplication of matrices A and B, storing the result in C + template + inline + static + void + apply + ( + Mat& C, + const Mat& A, + const Mat& B, + const out_eT alpha = out_eT(1), + const out_eT beta = out_eT(0) + ) + { + arma_extra_debug_sigprint(); + + if((is_cx::yes && do_trans_A) || (is_cx::yes && do_trans_B)) + { + // better-than-nothing handling of hermitian transpose + + Mat tmp_A; + Mat tmp_B; + + const bool predo_trans_A = ( (do_trans_A == true) && (is_cx::yes) ); + const bool predo_trans_B = ( (do_trans_B == true) && (is_cx::yes) ); + + if(predo_trans_A) { op_htrans::apply_mat_noalias(tmp_A, A); } + if(predo_trans_B) { op_htrans::apply_mat_noalias(tmp_B, B); } + + const Mat& AA = (predo_trans_A == false) ? A : tmp_A; + const Mat& BB = (predo_trans_B == false) ? B : tmp_B; + + gemm_mixed_large<((predo_trans_A) ? false : do_trans_A), ((predo_trans_B) ? false : do_trans_B), use_alpha, use_beta>::apply(C, AA, BB, alpha, beta); + } + else + { + gemm_mixed_large::apply(C, A, B, alpha, beta); + } + } + + + }; + + + +//! @} -- cgit v1.2.1