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
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
|
// 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
//! @{
//! for tiny square matrices, size <= 4x4
template<const bool do_trans_A=false, const bool use_alpha=false, const bool use_beta=false>
class gemm_emul_tinysq
{
public:
template<typename eT, typename TA, typename TB>
arma_cold
inline
static
void
apply
(
Mat<eT>& C,
const TA& A,
const TB& B,
const eT alpha = eT(1),
const eT beta = eT(0)
)
{
arma_extra_debug_sigprint();
switch(A.n_rows)
{
case 4: gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(3), A, B.colptr(3), alpha, beta );
// fallthrough
case 3: gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(2), A, B.colptr(2), alpha, beta );
// fallthrough
case 2: gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(1), A, B.colptr(1), alpha, beta );
// fallthrough
case 1: gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(0), A, B.colptr(0), alpha, beta );
// fallthrough
default: ;
}
}
};
//! emulation of gemm(), for non-complex matrices only, as it assumes only simple transposes (ie. doesn't do hermitian transposes)
template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
class gemm_emul_large
{
public:
template<typename eT, typename TA, typename TB>
arma_hot
inline
static
void
apply
(
Mat<eT>& C,
const TA& A,
const TB& B,
const eT alpha = eT(1),
const eT beta = 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) )
{
arma_aligned podarray<eT> tmp(A_n_cols);
eT* A_rowdata = tmp.memptr();
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 eT acc = op_dot::direct_dot_arma(B_n_rows, A_rowdata, B.colptr(col_B));
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) )
{
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 eT* A_coldata = A.colptr(col_A);
for(uword col_B=0; col_B < B_n_cols; ++col_B)
{
const eT acc = op_dot::direct_dot_arma(B_n_rows, A_coldata, B.colptr(col_B));
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<eT> BB;
op_strans::apply_mat_noalias(BB, B);
gemm_emul_large<false, false, use_alpha, use_beta>::apply(C, A, BB, alpha, beta);
}
else
if( (do_trans_A == true) && (do_trans_B == true) )
{
// mat B_tmp = trans(B);
// dgemm_arma<true, false, use_alpha, use_beta>::apply(C, A, B_tmp, alpha, beta);
// By using the trans(A)*trans(B) = trans(B*A) equivalency,
// transpose operations are not needed
arma_aligned podarray<eT> tmp(B.n_cols);
eT* 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 eT acc = op_dot::direct_dot_arma(A_n_rows, B_rowdata, A.colptr(col_A));
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); }
}
}
}
}
};
template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
class gemm_emul
{
public:
template<typename eT, typename TA, typename TB>
arma_hot
inline
static
void
apply
(
Mat<eT>& C,
const TA& A,
const TB& B,
const eT alpha = eT(1),
const eT beta = eT(0),
const typename arma_not_cx<eT>::result* junk = nullptr
)
{
arma_extra_debug_sigprint();
arma_ignore(junk);
gemm_emul_large<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C, A, B, alpha, beta);
}
template<typename eT>
arma_hot
inline
static
void
apply
(
Mat<eT>& C,
const Mat<eT>& A,
const Mat<eT>& B,
const eT alpha = eT(1),
const eT beta = eT(0),
const typename arma_cx_only<eT>::result* junk = nullptr
)
{
arma_extra_debug_sigprint();
arma_ignore(junk);
// "better than nothing" handling of hermitian transposes for complex number matrices
Mat<eT> tmp_A;
Mat<eT> tmp_B;
if(do_trans_A) { op_htrans::apply_mat_noalias(tmp_A, A); }
if(do_trans_B) { op_htrans::apply_mat_noalias(tmp_B, B); }
const Mat<eT>& AA = (do_trans_A == false) ? A : tmp_A;
const Mat<eT>& BB = (do_trans_B == false) ? B : tmp_B;
gemm_emul_large<false, false, use_alpha, use_beta>::apply(C, AA, BB, alpha, beta);
}
};
//! \brief
//! Wrapper for BLAS dgemm function, using template arguments to control the arguments passed to dgemm.
//! Matrix 'C' is assumed to have been set to the correct size (ie. taking into account transposes)
template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
class gemm
{
public:
template<typename eT, typename TA, typename TB>
inline
static
void
apply_blas_type( Mat<eT>& C, const TA& A, const TB& B, const eT alpha = eT(1), const eT beta = eT(0) )
{
arma_extra_debug_sigprint();
if( (A.n_rows <= 4) && (A.n_rows == A.n_cols) && (A.n_rows == B.n_rows) && (B.n_rows == B.n_cols) && (is_cx<eT>::no) )
{
if(do_trans_B == false)
{
gemm_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply(C, A, B, alpha, beta);
}
else
{
Mat<eT> BB(B.n_rows, B.n_rows, arma_nozeros_indicator());
op_strans::apply_mat_noalias_tinysq(BB, B);
gemm_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply(C, A, BB, alpha, beta);
}
}
else
{
#if defined(ARMA_USE_ATLAS)
{
arma_extra_debug_print("atlas::cblas_gemm()");
arma_debug_assert_atlas_size(A,B);
atlas::cblas_gemm<eT>
(
atlas_CblasColMajor,
(do_trans_A) ? ( is_cx<eT>::yes ? atlas_CblasConjTrans : atlas_CblasTrans ) : atlas_CblasNoTrans,
(do_trans_B) ? ( is_cx<eT>::yes ? atlas_CblasConjTrans : atlas_CblasTrans ) : atlas_CblasNoTrans,
C.n_rows,
C.n_cols,
(do_trans_A) ? A.n_rows : A.n_cols,
(use_alpha) ? alpha : eT(1),
A.mem,
(do_trans_A) ? A.n_rows : C.n_rows,
B.mem,
(do_trans_B) ? C.n_cols : ( (do_trans_A) ? A.n_rows : A.n_cols ),
(use_beta) ? beta : eT(0),
C.memptr(),
C.n_rows
);
}
#elif defined(ARMA_USE_BLAS)
{
arma_extra_debug_print("blas::gemm()");
arma_debug_assert_blas_size(A,B);
const char trans_A = (do_trans_A) ? ( is_cx<eT>::yes ? 'C' : 'T' ) : 'N';
const char trans_B = (do_trans_B) ? ( is_cx<eT>::yes ? 'C' : 'T' ) : 'N';
const blas_int m = blas_int(C.n_rows);
const blas_int n = blas_int(C.n_cols);
const blas_int k = (do_trans_A) ? blas_int(A.n_rows) : blas_int(A.n_cols);
const eT local_alpha = (use_alpha) ? alpha : eT(1);
const blas_int lda = (do_trans_A) ? k : m;
const blas_int ldb = (do_trans_B) ? n : k;
const eT local_beta = (use_beta) ? beta : eT(0);
arma_extra_debug_print( arma_str::format("blas::gemm(): trans_A = %c") % trans_A );
arma_extra_debug_print( arma_str::format("blas::gemm(): trans_B = %c") % trans_B );
blas::gemm<eT>
(
&trans_A,
&trans_B,
&m,
&n,
&k,
&local_alpha,
A.mem,
&lda,
B.mem,
&ldb,
&local_beta,
C.memptr(),
&m
);
}
#else
{
gemm_emul<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C,A,B,alpha,beta);
}
#endif
}
}
//! immediate multiplication of matrices A and B, storing the result in C
template<typename eT, typename TA, typename TB>
inline
static
void
apply( Mat<eT>& C, const TA& A, const TB& B, const eT alpha = eT(1), const eT beta = eT(0) )
{
gemm_emul<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C,A,B,alpha,beta);
}
template<typename TA, typename TB>
arma_inline
static
void
apply
(
Mat<float>& C,
const TA& A,
const TB& B,
const float alpha = float(1),
const float beta = float(0)
)
{
gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
}
template<typename TA, typename TB>
arma_inline
static
void
apply
(
Mat<double>& C,
const TA& A,
const TB& B,
const double alpha = double(1),
const double beta = double(0)
)
{
gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
}
template<typename TA, typename TB>
arma_inline
static
void
apply
(
Mat< std::complex<float> >& C,
const TA& A,
const TB& B,
const std::complex<float> alpha = std::complex<float>(1),
const std::complex<float> beta = std::complex<float>(0)
)
{
gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
}
template<typename TA, typename TB>
arma_inline
static
void
apply
(
Mat< std::complex<double> >& C,
const TA& A,
const TB& B,
const std::complex<double> alpha = std::complex<double>(1),
const std::complex<double> beta = std::complex<double>(0)
)
{
gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
}
};
//! @}
|
