forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
CPUBlas.cpp
824 lines (778 loc) · 25.4 KB
/
CPUBlas.cpp
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
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
#include <ATen/native/CPUBlas.h>
#include <ATen/native/mkl/LinearAlgebra.h>
#include <ATen/native/mkldnn/Matmul.h>
#include <ATen/Config.h>
#include <c10/util/SmallBuffer.h>
#include <c10/util/irange.h>
#include <climits>
#if AT_BUILD_WITH_BLAS()
#if C10_IOS
#include <Accelerate/Accelerate.h>
#else
extern "C" void dgemm_(char *transa, char *transb, int *m, int *n, int *k, double *alpha, const double *a, int *lda, const double *b, int *ldb, double *beta, double *c, int *ldc);
extern "C" void sgemm_(char *transa, char *transb, int *m, int *n, int *k, float *alpha, const float *a, int *lda, const float *b, int *ldb, float *beta, float *c, int *ldc);
extern "C" void cgemm_(char *transa, char *transb, int *m, int *n, int *k, void *alpha, const void *a, int *lda, const void *b, int *ldb, void *beta, void *c, int *ldc);
extern "C" void zgemm_(char *transa, char *transb, int *m, int *n, int *k, void *alpha, const void *a, int *lda, const void *b, int *ldb, void *beta, void *c, int *ldc);
#ifdef BLAS_HAS_SBGEMM
extern "C" void sbgemm_(char *transa, char *transb, int *m, int *n, int *k,
float *alpha,
const at::BFloat16 *a, int *lda,
const at::BFloat16 *b, int *ldb,
float *beta,
float *c, int *ldc);
#endif // BLAS_HAS_SBGEMM
extern "C" void cswap_(int *n, const void *x, int *incx, void *y, int *incy);
extern "C" void dcopy_(int *n, const double *x, int *incx, double *y, int *incy);
extern "C" void scopy_(int *n, const float *x, int *incx, float *y, int *incy);
extern "C" void zcopy_(int *n, const void *x, int *incx, void *y, int *incy);
extern "C" void ccopy_(int *n, const void *x, int *incx, void *y, int *incy);
extern "C" void daxpy_(int *n, double *a, const double *x, int *incx, double *y, int *incy);
extern "C" void saxpy_(int *n, float *a, const float *x, int *incx, float *y, int *incy);
extern "C" void caxpy_(int *n, void *a, const void *x, int *incx, void *y, int *incy);
extern "C" void zaxpy_(int *n, void *a, const void *x, int *incx, void *y, int *incy);
#endif // C10_IOS
#endif // AT_BUILD_WITH_BLAS
#ifdef USE_FBGEMM
#include <fbgemm/FbgemmI64.h>
#endif // USE_FBGEMM
namespace at::native::cpublas {
namespace internal {
void normalize_last_dims(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
int64_t *lda, int64_t *ldb, int64_t *ldc) {
if (n == 1) {
*ldc = m;
}
if(transa != TransposeType::NoTranspose) {
if (m == 1) {
*lda = k;
}
} else if(k == 1) {
*lda = m;
}
if(transb != TransposeType::NoTranspose) {
if (k == 1) {
*ldb = n;
}
} else if (n == 1) {
*ldb = k;
}
}
} // namespace internal
namespace {
bool use_blas_gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
int64_t lda, int64_t ldb, int64_t ldc) {
const bool transa_ = transa != TransposeType::NoTranspose;
const bool transb_ = transb != TransposeType::NoTranspose;
return (
(m <= INT_MAX) && (n <= INT_MAX) && (k <= INT_MAX) &&
(lda <= INT_MAX) && (ldb <= INT_MAX) && (ldc <= INT_MAX) &&
(lda >= std::max(int64_t{1}, (transa_ ? k : m))) &&
(ldb >= std::max(int64_t{1}, (transb_ ? n : k))) &&
(ldc >= std::max(int64_t{1}, m)));
}
#ifdef USE_FBGEMM
fbgemm::matrix_op_t to_fbgemm(TransposeType trans) {
switch (trans) {
case TransposeType::Transpose: return fbgemm::matrix_op_t::Transpose;
case TransposeType::NoTranspose: return fbgemm::matrix_op_t::NoTranspose;
case TransposeType::ConjTranspose: TORCH_INTERNAL_ASSERT(false, "ConjTranspose type is not supported in fbgemm");
}
TORCH_INTERNAL_ASSERT(false, "Invalid transpose type");
}
#endif // USE_FBGEMM
#if (AT_BUILD_WITH_BLAS() && C10_IOS)
CBLAS_TRANSPOSE to_apple_accelerate_transpose(TransposeType trans) {
switch (trans) {
case TransposeType::Transpose: return CblasTrans;
case TransposeType::NoTranspose: return CblasNoTrans;
case TransposeType::ConjTranspose: return CblasConjTrans;
}
TORCH_INTERNAL_ASSERT(false, "Invalid transpose type");
}
#endif
} // namespace (anonymous)
DEFINE_DISPATCH(gemm_stub);
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const double alpha,
const double *a, int64_t lda,
const double *b, int64_t ldb,
const double beta,
double *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#if AT_BUILD_WITH_BLAS()
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
int m_ = m, n_ = n, k_ = k, lda_ = lda, ldb_ = ldb, ldc_ = ldc;
double alpha_ = alpha, beta_ = beta;
#if C10_IOS
CBLAS_TRANSPOSE transa_ = to_apple_accelerate_transpose(transa);
CBLAS_TRANSPOSE transb_ = to_apple_accelerate_transpose(transb);
cblas_dgemm(CblasColMajor,
transa_, transb_,
m_, n_, k_,
alpha_,
a, lda_,
b, ldb_,
beta_,
c, ldc_);
#else
char transa_ = to_blas(transa), transb_ = to_blas(transb);
dgemm_(
&transa_, &transb_,
&m_, &n_, &k_,
&alpha_,
a, &lda_,
b, &ldb_,
&beta_,
c, &ldc_);
#endif
return;
}
#endif
gemm_stub(
at::kCPU, at::kDouble,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const float alpha,
const float *a, int64_t lda,
const float *b, int64_t ldb,
const float beta,
float *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#if AT_MKLDNN_ENABLED()
if (mkldnn_bf32_gemm(transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc)) {
return;
}
#endif
#if AT_BUILD_WITH_BLAS()
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
int m_ = m, n_ = n, k_ = k, lda_ = lda, ldb_ = ldb, ldc_ = ldc;
float alpha_ = alpha, beta_ = beta;
#if C10_IOS
CBLAS_TRANSPOSE transa_ = to_apple_accelerate_transpose(transa);
CBLAS_TRANSPOSE transb_ = to_apple_accelerate_transpose(transb);
cblas_sgemm(CblasColMajor,
transa_, transb_,
m_, n_, k_,
alpha_,
a, lda_,
b, ldb_,
beta_,
c, ldc_);
#else
char transa_ = to_blas(transa), transb_ = to_blas(transb);
sgemm_(
&transa_, &transb_,
&m_, &n_, &k_,
&alpha_,
a, &lda_,
b, &ldb_,
&beta_,
c, &ldc_);
#endif
return;
}
#endif
gemm_stub(
at::kCPU, at::kFloat,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const c10::complex<double> alpha,
const c10::complex<double> *a, int64_t lda,
const c10::complex<double> *b, int64_t ldb,
const c10::complex<double> beta,
c10::complex<double> *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#if AT_BUILD_WITH_BLAS()
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
int m_ = m, n_ = n, k_ = k, lda_ = lda, ldb_ = ldb, ldc_ = ldc;
c10::complex<double> alpha_ = alpha, beta_ = beta;
#if C10_IOS
CBLAS_TRANSPOSE transa_ = to_apple_accelerate_transpose(transa);
CBLAS_TRANSPOSE transb_ = to_apple_accelerate_transpose(transb);
cblas_zgemm(CblasColMajor,
transa_, transb_,
m_, n_, k_,
&alpha_,
a, lda_,
b, ldb_,
&beta_,
c, ldc_);
#else
char transa_ = to_blas(transa), transb_ = to_blas(transb);
zgemm_(
&transa_, &transb_,
&m_, &n_, &k_,
&alpha_,
a, &lda_,
b, &ldb_,
&beta_,
c, &ldc_);
#endif
return;
}
#endif
gemm_stub(
at::kCPU, at::kComplexDouble,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const c10::complex<float> alpha,
const c10::complex<float> *a, int64_t lda,
const c10::complex<float> *b, int64_t ldb,
const c10::complex<float> beta,
c10::complex<float> *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#if AT_BUILD_WITH_BLAS()
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
int m_ = m, n_ = n, k_ = k, lda_ = lda, ldb_ = ldb, ldc_ = ldc;
c10::complex<float> alpha_ = alpha, beta_ = beta;
#if C10_IOS
CBLAS_TRANSPOSE transa_ = to_apple_accelerate_transpose(transa);
CBLAS_TRANSPOSE transb_ = to_apple_accelerate_transpose(transb);
cblas_cgemm(CblasColMajor,
transa_, transb_,
m_, n_, k_,
&alpha_,
a, lda_,
b, ldb_,
&beta_,
c, ldc_);
#else
char transa_ = to_blas(transa), transb_ = to_blas(transb);
cgemm_(
&transa_, &transb_,
&m_, &n_, &k_,
&alpha_,
a, &lda_,
b, &ldb_,
&beta_,
c, &ldc_);
#endif
return;
}
#endif
gemm_stub(
at::kCPU, at::kComplexFloat,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const float alpha,
const at::BFloat16 *a, int64_t lda,
const at::BFloat16 *b, int64_t ldb,
const float beta,
at::BFloat16 *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#if AT_BUILD_WITH_BLAS() && defined(BLAS_HAS_SBGEMM)
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
int m_ = m, n_ = n, k_ = k, lda_ = lda, ldb_ = ldb, ldc_ = ldc;
char transa_ = to_blas(transa), transb_ = to_blas(transb);
float alpha_ = alpha, beta_ = beta;
int c_size = n_ * ldc_;
// C matrix in OpenBLAS sbgemm are of type "float" so we have to convert, copy and copy back.
std::vector<float> float_v(c, c + c_size);
sbgemm_(&transa_, &transb_,
&m_, &n_, &k_,
&alpha_,
a, &lda_,
b, &ldb_,
&beta_,
float_v.data(), &ldc_);
for (auto cv: float_v) {
*(c++) = c10::convert<at::BFloat16>(cv);
}
return;
}
#endif
#if AT_MKLDNN_ENABLED()
if (mkldnn_bf16_gemm(transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc)) {
return;
}
#endif
gemm_stub(
at::kCPU, at::kBFloat16,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const float alpha,
const at::Half *a, int64_t lda,
const at::Half *b, int64_t ldb,
const float beta,
at::Half *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#if AT_MKLDNN_ENABLED()
if (mkldnn_fp16_gemm(transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc)) {
return;
}
#endif
gemm_stub(
at::kCPU, at::kHalf,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const float alpha,
const at::BFloat16 *a, int64_t lda,
const at::BFloat16 *b, int64_t ldb,
const float beta,
float *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#if AT_BUILD_WITH_BLAS() && defined(BLAS_HAS_SBGEMM)
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
int m_ = m, n_ = n, k_ = k, lda_ = lda, ldb_ = ldb, ldc_ = ldc;
char transa_ = to_blas(transa), transb_ = to_blas(transb);
float alpha_ = alpha, beta_ = beta;
sbgemm_(&transa_, &transb_,
&m_, &n_, &k_,
&alpha_,
a, &lda_,
b, &ldb_,
&beta_,
c, &ldc_);
return;
}
#endif
#ifdef MKL_HAS_SBGEMM
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
int m_ = m, n_ = n, k_ = k, lda_ = lda, ldb_ = ldb, ldc_ = ldc;
mkl_gemm_bf16bf16f32(transa, transb, m_, n_, k_, alpha, a, lda_, b, ldb_, beta, c, ldc_);
return;
}
#endif
// for the fallback path, first compute gemm with beta = 0,
// and then add c in full precision.
int64_t c_size = n * m;
std::vector<at::BFloat16> bfloat_c(c_size, 0.f);
gemm_stub(
at::kCPU, at::kBFloat16,
transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, bfloat_c.data(), m);
for (const auto j : c10::irange(n)) {
for (const auto i : c10::irange(m)) {
auto offset = j * ldc + i;
// beta == 0 won't propagate NaN from C
if (beta == 0.f) {
c[offset] = c10::convert<float>(bfloat_c[j * m + i]);
} else {
c[offset] = beta * c[offset] + c10::convert<float>(bfloat_c[j * m + i]);
}
}
}
}
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const float alpha,
const at::Half *a, int64_t lda,
const at::Half *b, int64_t ldb,
const float beta,
float *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#ifdef MKL_HAS_SHGEMM
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
int m_ = m, n_ = n, k_ = k, lda_ = lda, ldb_ = ldb, ldc_ = ldc;
mkl_gemm_f16f16f32(transa, transb, m_, n_, k_, alpha, a, lda_, b, ldb_, beta, c, ldc_);
return;
}
#endif
// for the fallback path, first compute gemm with beta = 0,
// and then add c in full precision.
int64_t c_size = n * m;
std::vector<at::Half> float16_c(c_size, 0.f);
gemm_stub(
at::kCPU, at::kHalf,
transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float16_c.data(), m);
for (const auto j : c10::irange(n)) {
for (const auto i : c10::irange(m)) {
auto offset = j * ldc + i;
// beta == 0 won't propagate NaN from C
if (beta == 0.f) {
c[offset] = c10::convert<float>(float16_c[j * m + i]);
} else {
c[offset] = beta * c[offset] + c10::convert<float>(float16_c[j * m + i]);
}
}
}
}
void gemm(
TransposeType transa, TransposeType transb,
int64_t m, int64_t n, int64_t k,
const int64_t alpha,
const int64_t *a, int64_t lda,
const int64_t *b, int64_t ldb,
const int64_t beta,
int64_t *c, int64_t ldc) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
#ifdef USE_FBGEMM
if (alpha == 1 && (beta == 0 || beta == 1)) {
// In FBGEMM, we assume row-major ordering; However, here we assume the
// column-major ordering following the FORTRAN tradition in BLAS interface
// in this function: we can configure the layout (row/column-major ordering)
// of A and B by changing transa_ and transb_, but we cannot change the
// layout of C with this FORTRAN-style BLAS interface.
//
// The workaround is that we compute
// C^T (n x m) = B^T (n x k) * A^T (k x m) instead.
//
// In this way we view C^T as the row-major ordering when passing to FBGEMM.
fbgemm::cblas_gemm_i64_i64acc(
to_fbgemm(transb),
to_fbgemm(transa),
n,
m,
k,
b,
ldb,
a,
lda,
beta == 1,
c,
ldc);
return;
}
#endif
gemm_stub(
kCPU, kLong,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
template <typename scalar_t>
static void gemm_batched_mkl_impl(
TransposeType transa, TransposeType transb,
int64_t batch_size, int64_t m, int64_t n, int64_t k,
scalar_t alpha,
const scalar_t **a, int64_t lda,
const scalar_t **b, int64_t ldb,
scalar_t beta,
scalar_t **c, int64_t ldc) {
for (int64_t i = 0; i < batch_size;) {
int sub_batch = std::min(batch_size - i, int64_t{INT_MAX});
mkl_gemm_batched(transa, transb, sub_batch, m, n, k, alpha,
&a[i], lda, &b[i], ldb, beta, &c[i], ldc);
i += sub_batch;
}
}
template <typename scalar_t>
using is_blas_library_type = std::integral_constant<bool,
std::is_same<scalar_t, double>::value ||
std::is_same<scalar_t, float>::value ||
std::is_same<scalar_t, c10::complex<double>>::value ||
std::is_same<scalar_t, c10::complex<float>>::value>;
template <typename scalar_t>
void gemm_batched_generic(
TransposeType transa, TransposeType transb,
int64_t batch_size, int64_t m, int64_t n, int64_t k,
scalar_t alpha,
const scalar_t **a, int64_t lda,
const scalar_t **b, int64_t ldb,
scalar_t beta,
scalar_t **c, int64_t ldc) {
for (const auto batch : c10::irange(batch_size)) {
gemm(transa, transb, m, n, k, alpha, a[batch], lda, b[batch], ldb, beta, c[batch], ldc);
}
}
template <typename scalar_t>
void gemm_batched(
TransposeType transa, TransposeType transb,
int64_t batch_size, int64_t m, int64_t n, int64_t k,
scalar_t alpha,
const scalar_t **a, int64_t lda,
const scalar_t **b, int64_t ldb,
scalar_t beta,
scalar_t **c, int64_t ldc) {
if (batch_size == 1) {
return gemm(transa, transb, m, n, k, alpha, a[0], lda, b[0], ldb, beta, c[0], ldc);
}
if constexpr (AT_MKL_ENABLED() && is_blas_library_type<scalar_t>::value) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
gemm_batched_mkl_impl(
transa, transb, batch_size, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
} else {
gemm_batched_generic(
transa, transb, batch_size, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
} else {
gemm_batched_generic(
transa, transb, batch_size, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
}
template <typename scalar_t>
void gemm_batched_with_stride_generic(
TransposeType transa, TransposeType transb,
int64_t batch_size, int64_t m, int64_t n, int64_t k,
scalar_t alpha,
const scalar_t *a, int64_t lda, int64_t batch_stride_a,
const scalar_t *b, int64_t ldb, int64_t batch_stride_b,
scalar_t beta,
scalar_t *c, int64_t ldc, int64_t batch_stride_c) {
for (const auto batch : c10::irange(batch_size)) {
const auto a_batch = a + batch_stride_a * batch;
const auto b_batch = b + batch_stride_b * batch;
const auto c_batch = c + batch_stride_c * batch;
gemm(transa, transb, m, n, k, alpha, a_batch, lda, b_batch, ldb, beta, c_batch, ldc);
}
}
template <typename scalar_t>
void gemm_batched_with_stride(
TransposeType transa, TransposeType transb,
int64_t batch_size, int64_t m, int64_t n, int64_t k,
scalar_t alpha,
const scalar_t *a, int64_t lda, int64_t batch_stride_a,
const scalar_t *b, int64_t ldb, int64_t batch_stride_b,
scalar_t beta,
scalar_t *c, int64_t ldc, int64_t batch_stride_c) {
if (batch_size == 1) {
return gemm(transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
}
if constexpr (AT_MKL_ENABLED() && is_blas_library_type<scalar_t>::value) {
internal::normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
if (use_blas_gemm(transa, transb, m, n, k, lda, ldb, ldc)) {
c10::SmallBuffer<const scalar_t*, 16> a_ptrs(batch_size);
c10::SmallBuffer<const scalar_t*, 16> b_ptrs(batch_size);
c10::SmallBuffer<scalar_t*, 16> c_ptrs(batch_size);
for (const auto batch : c10::irange(batch_size)) {
a_ptrs[batch] = a + batch_stride_a * batch;
b_ptrs[batch] = b + batch_stride_b * batch;
c_ptrs[batch] = c + batch_stride_c * batch;
}
gemm_batched_mkl_impl(
transa, transb, batch_size, m, n, k, alpha, a_ptrs.data(), lda,
b_ptrs.data(), ldb, beta, c_ptrs.data(), ldc);
} else {
gemm_batched_with_stride_generic(
transa, transb, batch_size, m, n, k, alpha, a, lda, batch_stride_a,
b, ldb, batch_stride_b, beta, c, ldc, batch_stride_c);
}
} else {
gemm_batched_with_stride_generic(transa, transb, batch_size, m, n, k, alpha,
a, lda, batch_stride_a, b, ldb, batch_stride_b,
beta, c, ldc, batch_stride_c);
}
}
#define INSTANTIATE_BATCHED_GEMM(scalar_t, DType) \
template void gemm_batched( \
TransposeType transa, TransposeType transb, \
int64_t batch_size, int64_t m, int64_t n, int64_t k, \
scalar_t alpha, \
const scalar_t **a, int64_t lda, \
const scalar_t **b, int64_t ldb, \
scalar_t beta, \
scalar_t **c, int64_t ldc); \
template void gemm_batched_with_stride( \
TransposeType transa, TransposeType transb, \
int64_t batch_size, int64_t m, int64_t n, int64_t k, \
scalar_t alpha, \
const scalar_t *a, int64_t lda, int64_t batch_stride_a, \
const scalar_t *b, int64_t ldb, int64_t batch_stride_b, \
scalar_t beta, \
scalar_t *c, int64_t ldc, int64_t batch_stride_c);
AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_EXCEPT_COMPLEX_HALF_F8NZ(INSTANTIATE_BATCHED_GEMM)
DEFINE_DISPATCH(axpy_stub);
void axpy(int64_t n, double a, const double *x, int64_t incx, double *y, int64_t incy) {
if(n == 1)
{
incx = 1;
incy = 1;
}
#if AT_BUILD_WITH_BLAS()
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) )
{
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if C10_IOS
cblas_daxpy(i_n, a, x, i_incx, y, i_incy);
#else
daxpy_(&i_n, &a, x, &i_incx, y, &i_incy);
#endif
return;
}
#endif
axpy_stub(
kCPU, at::kDouble,
n, a, x, incx, y, incy);
}
void axpy(int64_t n, float a, const float *x, int64_t incx, float *y, int64_t incy) {
if(n == 1)
{
incx = 1;
incy = 1;
}
#if AT_BUILD_WITH_BLAS()
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) )
{
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if C10_IOS
cblas_saxpy(i_n, a, x, i_incx, y, i_incy);
#else
saxpy_(&i_n, &a, x, &i_incx, y, &i_incy);
#endif
return;
}
#endif
axpy_stub(
kCPU, at::kFloat,
n, a, x, incx, y, incy);
}
void axpy(int64_t n, c10::complex<double> a, const c10::complex<double> *x, int64_t incx, c10::complex<double> *y, int64_t incy) {
if(n == 1)
{
incx = 1;
incy = 1;
}
#if AT_BUILD_WITH_BLAS()
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) )
{
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if C10_IOS
cblas_zaxpy(i_n, &a, x, i_incx, y, i_incy);
#else
zaxpy_(&i_n, &a, x, &i_incx, y, &i_incy);
#endif
return;
}
#endif
axpy_stub(
kCPU, at::kComplexDouble,
n, a, x, incx, y, incy);
}
void axpy(int64_t n, c10::complex<float> a, const c10::complex<float> *x, int64_t incx, c10::complex<float> *y, int64_t incy) {
if(n == 1)
{
incx = 1;
incy = 1;
}
#if AT_BUILD_WITH_BLAS()
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) )
{
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if C10_IOS
cblas_caxpy(i_n, &a, x, i_incx, y, i_incy);
#else
caxpy_(&i_n, &a, x, &i_incx, y, &i_incy);
#endif
return;
}
#endif
axpy_stub(
kCPU, at::kComplexFloat,
n, a, x, incx, y, incy);
}
DEFINE_DISPATCH(copy_stub);
void copy(int64_t n, const double *x, int64_t incx, double *y, int64_t incy) {
if(n == 1)
{
incx = 1;
incy = 1;
}
#if AT_BUILD_WITH_BLAS()
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) ) {
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if C10_IOS
cblas_dcopy(i_n, x, i_incx, y, i_incy);
#else
dcopy_(&i_n, x, &i_incx, y, &i_incy);
#endif
return;
}
#endif
copy_stub(
kCPU, at::kDouble,
n, x, incx, y, incy);
}
void copy(int64_t n, const float *x, int64_t incx, float *y, int64_t incy) {
if(n == 1)
{
incx = 1;
incy = 1;
}
#if AT_BUILD_WITH_BLAS()
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) ) {
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if C10_IOS
cblas_scopy(i_n, x, i_incx, y, i_incy);
#else
scopy_(&i_n, x, &i_incx, y, &i_incy);
#endif
return;
}
#endif
copy_stub(
kCPU, at::kFloat,
n, x, incx, y, incy);
}
void copy(int64_t n, const c10::complex<double> *x, int64_t incx, c10::complex<double> *y, int64_t incy) {
if(n == 1)
{
incx = 1;
incy = 1;
}
#if AT_BUILD_WITH_BLAS()
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) ) {
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if C10_IOS
cblas_zcopy(i_n, x, i_incx, y, i_incy);
#else
zcopy_(&i_n, x, &i_incx, y, &i_incy);
#endif
return;
}
#endif
copy_stub(
kCPU, at::kComplexDouble,
n, x, incx, y, incy);
}
void copy(int64_t n, const c10::complex<float> *x, int64_t incx, c10::complex<float> *y, int64_t incy){
if(n == 1)
{
incx = 1;
incy = 1;
}
#if AT_BUILD_WITH_BLAS()
if( (n <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) ) {
int i_n = (int)n;
int i_incx = (int)incx;
int i_incy = (int)incy;
#if C10_IOS
cblas_ccopy(i_n, &x, i_incx, y, i_incy);
#else
ccopy_(&i_n, x, &i_incx, y, &i_incy);
#endif
return;
}
#endif
copy_stub(
kCPU, at::kComplexFloat,
n, x, incx, y, incy);
}
} // namespace at::native::cpublas