-
Notifications
You must be signed in to change notification settings - Fork 74k
/
sparse_matmul_op.cc
1402 lines (1280 loc) · 51.2 KB
/
sparse_matmul_op.cc
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
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
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.
==============================================================================*/
// See docs in ../ops/math_ops.cc.
#define EIGEN_USE_THREADS
#include "tensorflow/core/kernels/sparse_matmul_op.h"
#include <map>
#include <memory>
#include <vector>
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/framework/bfloat16.h"
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/kernels/fill_functor.h"
#include "tensorflow/core/lib/core/threadpool.h"
#include "tensorflow/core/platform/blocking_counter.h"
#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/macros.h"
#include "tensorflow/core/platform/mutex.h"
#include "tensorflow/core/platform/thread_annotations.h"
#include "tensorflow/core/platform/types.h"
#if defined(TENSORFLOW_USE_CUSTOM_CONTRACTION_KERNEL)
#include "tensorflow/tsl/framework/contraction/eigen_contraction_kernel.h"
#endif
#define ALWAYS_INLINE EIGEN_ALWAYS_INLINE
namespace tensorflow {
namespace {
template <typename T>
using BasicMatrix = Eigen::Tensor<T, 2, Eigen::RowMajor>;
template <typename T>
using BasicMatrixMap =
Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor>, Eigen::Aligned>;
using Matrix = BasicMatrix<float>;
using MatrixMap = BasicMatrixMap<float>;
using CPUDevice = Eigen::ThreadPoolDevice;
using DSizes = Eigen::DSizes<Eigen::DenseIndex, 2>;
// Two commonly used static dsizes. We use Eigen::type2index to allow as much
// compile time optimization as possible.
inline Eigen::IndexList<Eigen::type2index<0>, Eigen::type2index<0>>
dsizes_00() {
return Eigen::IndexList<Eigen::type2index<0>, Eigen::type2index<0>>();
}
inline Eigen::IndexList<Eigen::type2index<1>, Eigen::type2index<0>>
dsizes_10() {
return Eigen::IndexList<Eigen::type2index<1>, Eigen::type2index<0>>();
}
// Blocksizes
// TODO(agarwal): compute these sizes based on cache sizes.
const int K = 64;
const int M = 64;
const int N = 128;
// This stores a sparse representation of a slice of a matrix with size
// (num_rows, num_cols). The slice is represented as a series of blocks of size
// (num_rows, b), where b = block_size for all but the last block, which may
// have fewer columns.
//
// num_rows and block_size are assumed to be <= 256. This allows storing
// different indices as uint8.
//
// For each block, we store all the non zero entries in data/data3 vector and
// the corresponding coordinates of the element in index/index3 vectors. index3
// vector stores index of 3 elements in the same row so that these elements can
// share the same row coordinate. Each entry in Index3 corresponds to 3 entries
// in data3.
//
// Note that all the data/indices of all the blocks are stored in the same
// vectors respectively. To identify block boundaries, we store the block
// offsets using index3_offset/index_offset. If there are n blocks in the slice,
// index3_offset and index_offset have n entries. The indices for the ith block
// are the values in the following range:
// [index3[index3_offset[i-1]], index3[index3_offset[i]]). Similarly for
// index_offset.
template <typename T>
struct SparseSlice {
using ConstMatrixMap = BasicMatrixMap<const T>;
public:
// Indices of three elements on the same row.
struct Index3 {
uint8 m; // row
// columns
uint8 k1;
uint8 k2;
uint8 k3;
};
// Index of one element.
struct Index {
uint8 m;
uint8 k;
};
SparseSlice(int nrows, int ncols, int bsize)
: num_rows(nrows), num_cols(ncols), block_size(bsize) {
DCHECK_LE(nrows, 256);
DCHECK_LE(block_size, 256);
}
// Initializes the slice with data starting at mat(0, col_offset) and with
// size (num_rows, num_cols).
// If Transpose is true, implicitly transposes mat.
template <bool Transpose = false>
void Initialize(const ConstMatrixMap& mat, int col_offset);
void Clear();
// See comments above.
std::vector<int> index3_offset;
std::vector<Index3> index3;
std::vector<T> data3;
// See comments above. Similar to "index3" except that each element in "index"
// corresponds to one element in data.
std::vector<int> index_offset;
std::vector<Index> index;
std::vector<T> data;
// Number of rows and columns for the slice.
const int num_rows;
const int num_cols;
// Block size used to initialize from a matrix.
const int block_size;
};
template <typename T>
bool IsZero(T v);
template <>
ALWAYS_INLINE bool IsZero(bfloat16 v) {
return !static_cast<bool>(v);
}
template <>
ALWAYS_INLINE bool IsZero(float v) {
return v == 0.0f;
}
template <typename T>
template <bool Transpose>
void SparseSlice<T>::Initialize(
const typename SparseSlice<T>::ConstMatrixMap& mat, int col_offset) {
const int mat_rows = Transpose ? mat.dimension(1) : mat.dimension(0);
const int mat_cols = Transpose ? mat.dimension(0) : mat.dimension(1);
DCHECK_LE(num_rows, mat_rows);
DCHECK_LE(num_cols + col_offset, mat_cols);
int num_blocks = (num_cols + block_size - 1) / block_size;
int mat_size = num_rows * num_cols;
index3_offset.reserve(num_blocks);
data3.reserve(mat_size);
index3.reserve(mat_size / 3);
index_offset.reserve(num_blocks);
data.reserve(num_blocks * num_rows * 2);
index.reserve(num_blocks * num_rows * 2);
Index3 idx3;
const int stride = Transpose ? mat.dimension(1) : 1;
for (int i = 0; i < num_blocks; ++i) {
int num_block_cols = std::min(block_size, num_cols - block_size * i);
for (int row = 0; row < num_rows; ++row) {
idx3.m = static_cast<uint8>(row);
// Safety note: The following code has a race, since it checks whether
// *curr is nonzero and then reads it again on use. However, the result
// of the race is only that some of the "nonzeros" in the resulting sparse
// representation may actually be zero, which is harmless.
const auto* start =
Transpose ? &mat(col_offset, row) : &mat(row, col_offset);
const auto* curr = start;
const auto* end = start + stride * num_block_cols;
uint8 k = 0;
#define NEXT_ELEM \
curr += stride; \
++k;
#define EAT_ZEROS \
while (curr < end && IsZero<T>(*curr)) { \
NEXT_ELEM; \
}
while (true) {
EAT_ZEROS
if (curr >= end) break;
idx3.k1 = k;
const T value1 = *curr;
NEXT_ELEM;
EAT_ZEROS
if (curr >= end) {
data.push_back(value1);
index.push_back({idx3.m, idx3.k1});
break;
}
idx3.k2 = k;
const T value2 = *curr;
NEXT_ELEM;
EAT_ZEROS
if (curr >= end) {
data.push_back(value2);
index.push_back({idx3.m, idx3.k2});
data.push_back(value1);
index.push_back({idx3.m, idx3.k1});
break;
}
idx3.k3 = k;
data3.push_back(value1);
data3.push_back(value2);
data3.push_back(*curr);
NEXT_ELEM;
index3.push_back(idx3);
#undef NEXT_ELEM
#undef EAT_ZEROS
}
}
col_offset += block_size;
index3_offset.push_back(index3.size());
index_offset.push_back(index.size());
}
DCHECK_EQ(index3_offset.size(), num_blocks);
DCHECK_EQ(index_offset.size(), num_blocks);
DCHECK_EQ(3 * index3.size(), data3.size());
DCHECK_EQ(index.size(), data.size());
}
template <typename T>
void SparseSlice<T>::Clear() {
index3_offset.clear();
index3.clear();
data3.clear();
index_offset.clear();
index.clear();
data.clear();
}
using Packet = Eigen::internal::packet_traits<float>::type;
const int kNumOperands = (sizeof(Packet) / sizeof(float));
#define LOAD(x) Eigen::internal::pload<Packet>(x);
#define EXPAND_BFLOAT_L(x, y) \
const auto y = Eigen::internal::pexpand_bf16_l<Packet>(x);
#define EXPAND_BFLOAT_U(x, y) \
const auto y = Eigen::internal::pexpand_bf16_u<Packet>(x);
#define STORE(x, y) Eigen::internal::pstore<float>(x, y);
#define FMA(a, b, c, d) d = Eigen::internal::pmadd<Packet>(a, b, c);
ALWAYS_INLINE float ConvertBfloat16ToFloat(const bfloat16* src) {
float out = 0;
auto tmp = reinterpret_cast<bfloat16*>(&out);
#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
tmp[0] = *src;
#else
tmp[1] = *src;
#endif
return out;
}
ALWAYS_INLINE Packet ConvertFourBfloat16ToFloat(const bfloat16* src) {
return Eigen::internal::pload4bf16<Packet>(
reinterpret_cast<const float*>(src));
}
ALWAYS_INLINE Packet ConvertTwoBfloat16ToFloat(const bfloat16* src) {
return Eigen::internal::pload2bf16<Packet>(
reinterpret_cast<const float*>(src));
}
ALWAYS_INLINE void ScalarMulAdd(const float a, const float** inp, float** out) {
**out += a * **inp;
++*inp;
++*out;
}
ALWAYS_INLINE void ScalarMulAdd(const float a, const bfloat16** inp,
float** out) {
float inp_f = ConvertBfloat16ToFloat(*inp);
**out += a * inp_f;
++*inp;
++*out;
}
ALWAYS_INLINE void ScalarMulAdd3Way(const float a1, const float a2,
const float a3, const bfloat16** inp1,
const bfloat16** inp2,
const bfloat16** inp3, float** out) {
float inp1_f = ConvertBfloat16ToFloat(*inp1);
float inp2_f = ConvertBfloat16ToFloat(*inp2);
float inp3_f = ConvertBfloat16ToFloat(*inp3);
**out += a1 * inp1_f + a2 * inp2_f + a3 * inp3_f;
++*out;
++*inp1;
++*inp2;
++*inp3;
}
ALWAYS_INLINE void ScalarMulAdd3Way(const float a1, const float a2,
const float a3, const float** inp1,
const float** inp2, const float** inp3,
float** out) {
**out += a1 * **inp1 + a2 * **inp2 + a3 * **inp3;
++*out;
++*inp1;
++*inp2;
++*inp3;
}
ALWAYS_INLINE void LoadSingleScalar(const bfloat16** data, Packet* l) {
auto tmp = ConvertBfloat16ToFloat(*data);
*l = Eigen::internal::pset1<Packet>(tmp);
++*data;
}
ALWAYS_INLINE void LoadTwoScalars(const bfloat16** data, Packet* l1,
Packet* l2) {
if (kNumOperands >= 2) {
auto tmp = ConvertTwoBfloat16ToFloat(*data);
*l1 = Eigen::internal::pbroadcast_first<Packet>(tmp);
*l2 = Eigen::internal::pbroadcast_second<Packet>(tmp);
*data += 2;
} else {
LoadSingleScalar(data, l1);
LoadSingleScalar(data, l2);
}
}
ALWAYS_INLINE void LoadFourScalars(const bfloat16** data, Packet* l1,
Packet* l2, Packet* l3, Packet* l4) {
if (kNumOperands >= 4) {
auto tmp = ConvertFourBfloat16ToFloat(*data);
*l1 = Eigen::internal::pbroadcast_first<Packet>(tmp);
*l2 = Eigen::internal::pbroadcast_second<Packet>(tmp);
*l3 = Eigen::internal::pbroadcast_third<Packet>(tmp);
*l4 = Eigen::internal::pbroadcast_fourth<Packet>(tmp);
*data += 4;
} else {
LoadTwoScalars(data, l1, l2);
LoadTwoScalars(data, l3, l4);
}
}
ALWAYS_INLINE void LoadSingleScalar(const float** data, Packet* l) {
*l = Eigen::internal::pload1<Packet>(*data);
++(*data);
}
ALWAYS_INLINE void LoadTwoScalars(const float** data, Packet* l1, Packet* l2) {
LoadSingleScalar(data, l1);
LoadSingleScalar(data, l2);
}
ALWAYS_INLINE void LoadFourScalars(const float** data, Packet* l1, Packet* l2,
Packet* l3, Packet* l4) {
LoadTwoScalars(data, l1, l2);
LoadTwoScalars(data, l3, l4);
}
template <typename T>
ALWAYS_INLINE void LoadThreeScalars(const T** data, Packet* l1, Packet* l2,
Packet* l3) {
LoadTwoScalars(data, l1, l2);
LoadSingleScalar(data, l3);
}
template <typename T>
ALWAYS_INLINE void LoadSixScalars(const T** data, Packet* l1, Packet* l2,
Packet* l3, Packet* l4, Packet* l5,
Packet* l6) {
LoadFourScalars(data, l1, l2, l3, l4);
LoadTwoScalars(data, l5, l6);
}
// Vectorized version of ScalarMulAdd.
ALWAYS_INLINE void MulAdd(const Packet a, const bfloat16** binp, float** out) {
auto inp = reinterpret_cast<const float*>(*binp);
const auto b = LOAD(inp);
EXPAND_BFLOAT_L(b, b_0);
EXPAND_BFLOAT_U(b, b_1);
*binp += 2 * kNumOperands;
auto c1 = LOAD(*out);
auto c2 = LOAD(*out + kNumOperands);
FMA(a, b_0, c1, c1);
FMA(a, b_1, c2, c2);
STORE(*out, c1);
STORE(*out + kNumOperands, c2);
*out += 2 * kNumOperands;
}
// Vectorized version of ScalarMulAdd3Way.
ALWAYS_INLINE void MulAdd3Way(const Packet a1, const Packet a2, const Packet a3,
const bfloat16** binp1, const bfloat16** binp2,
const bfloat16** binp3, float** out) {
auto inp1 = reinterpret_cast<const float*>(*binp1);
auto inp2 = reinterpret_cast<const float*>(*binp2);
auto inp3 = reinterpret_cast<const float*>(*binp3);
auto c1 = LOAD(*out);
auto c2 = LOAD(*out + kNumOperands);
const auto b1 = LOAD(inp1);
EXPAND_BFLOAT_L(b1, b1_0);
EXPAND_BFLOAT_U(b1, b1_1);
*binp1 += 2 * kNumOperands;
const auto b2 = LOAD(inp2);
EXPAND_BFLOAT_L(b2, b2_0);
EXPAND_BFLOAT_U(b2, b2_1);
*binp2 += 2 * kNumOperands;
const auto b3 = LOAD(inp3);
EXPAND_BFLOAT_L(b3, b3_0);
EXPAND_BFLOAT_U(b3, b3_1);
*binp3 += 2 * kNumOperands;
FMA(a1, b1_0, c1, c1);
FMA(a1, b1_1, c2, c2);
FMA(a2, b2_0, c1, c1);
FMA(a2, b2_1, c2, c2);
FMA(a3, b3_0, c1, c1);
FMA(a3, b3_1, c2, c2);
STORE(*out, c1);
STORE(*out + kNumOperands, c2);
*out += 2 * kNumOperands;
}
// Unroll MulAdd3Way for two iterations
ALWAYS_INLINE void TwoMulAdd3Way(const Packet a1, const Packet a2,
const Packet a3, const bfloat16** binp1,
const bfloat16** binp2, const bfloat16** binp3,
float** out) {
auto inp1 = reinterpret_cast<const float*>(*binp1);
auto inp2 = reinterpret_cast<const float*>(*binp2);
auto inp3 = reinterpret_cast<const float*>(*binp3);
auto c1 = LOAD(*out);
auto c2 = LOAD(*out + kNumOperands);
const auto b1 = LOAD(inp1);
const auto b2 = LOAD(inp2);
const auto b3 = LOAD(inp3);
EXPAND_BFLOAT_L(b1, b1_0);
EXPAND_BFLOAT_U(b1, b1_1);
EXPAND_BFLOAT_L(b2, b2_0);
EXPAND_BFLOAT_U(b2, b2_1);
EXPAND_BFLOAT_L(b3, b3_0);
EXPAND_BFLOAT_U(b3, b3_1);
auto c3 = LOAD(*out + 2 * kNumOperands);
auto c4 = LOAD(*out + 3 * kNumOperands);
const auto b4 = LOAD(inp1 + kNumOperands);
const auto b5 = LOAD(inp2 + kNumOperands);
const auto b6 = LOAD(inp3 + kNumOperands);
EXPAND_BFLOAT_L(b4, b4_0);
EXPAND_BFLOAT_U(b4, b4_1);
EXPAND_BFLOAT_L(b5, b5_0);
EXPAND_BFLOAT_U(b5, b5_1);
EXPAND_BFLOAT_L(b6, b6_0);
EXPAND_BFLOAT_U(b6, b6_1);
FMA(a1, b1_0, c1, c1);
FMA(a1, b1_1, c2, c2);
FMA(a1, b4_0, c3, c3);
FMA(a1, b4_1, c4, c4);
FMA(a2, b2_0, c1, c1);
FMA(a2, b2_1, c2, c2);
FMA(a2, b5_0, c3, c3);
FMA(a2, b5_1, c4, c4);
FMA(a3, b3_0, c1, c1);
FMA(a3, b3_1, c2, c2);
FMA(a3, b6_0, c3, c3);
FMA(a3, b6_1, c4, c4);
STORE(*out, c1);
STORE(*out + kNumOperands, c2);
STORE(*out + 2 * kNumOperands, c3);
STORE(*out + 3 * kNumOperands, c4);
*out += 4 * kNumOperands;
*binp1 += 4 * kNumOperands;
*binp2 += 4 * kNumOperands;
*binp3 += 4 * kNumOperands;
}
// Apply MulAdd3Way on 128 operands.
ALWAYS_INLINE void MulAdd3Way128(const Packet a1, const Packet a2,
const Packet a3, const bfloat16** inp1,
const bfloat16** inp2, const bfloat16** inp3,
float** out) {
for (int k = 0; k < 128 / (8 * kNumOperands); ++k) {
TwoMulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
TwoMulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
}
}
// Vectorized version of ScalarMulAdd
ALWAYS_INLINE void MulAdd(const Packet a, const float** inp, float** out) {
const auto b = LOAD(*inp);
*inp += kNumOperands;
auto c = LOAD(*out);
FMA(a, b, c, c);
STORE(*out, c);
*out += kNumOperands;
}
// Vectorized version of ScalarMulAdd3Way
ALWAYS_INLINE void MulAdd3Way(const Packet a1, const Packet a2, const Packet a3,
const float** inp1, const float** inp2,
const float** inp3, float** out) {
auto c = LOAD(*out);
const auto b1 = LOAD(*inp1);
*inp1 += kNumOperands;
const auto b2 = LOAD(*inp2);
*inp2 += kNumOperands;
const auto b3 = LOAD(*inp3);
*inp3 += kNumOperands;
FMA(a1, b1, c, c);
FMA(a2, b2, c, c);
FMA(a3, b3, c, c);
STORE(*out, c);
*out += kNumOperands;
}
// Unroll MulAdd3Way for two iterations
ALWAYS_INLINE void TwoMulAdd3Way(const Packet a1, const Packet a2,
const Packet a3, const float** inp1,
const float** inp2, const float** inp3,
float** out) {
auto c1 = LOAD(*out);
const auto b1 = LOAD(*inp1);
const auto b2 = LOAD(*inp2);
const auto b3 = LOAD(*inp3);
auto c2 = LOAD(*out + kNumOperands);
const auto b4 = LOAD(*inp1 + kNumOperands);
const auto b5 = LOAD(*inp2 + kNumOperands);
const auto b6 = LOAD(*inp3 + kNumOperands);
FMA(a1, b1, c1, c1);
FMA(a1, b4, c2, c2);
FMA(a2, b2, c1, c1);
FMA(a2, b5, c2, c2);
FMA(a3, b3, c1, c1);
FMA(a3, b6, c2, c2);
STORE(*out, c1);
STORE(*out + kNumOperands, c2);
*out += 2 * kNumOperands;
*inp1 += 2 * kNumOperands;
*inp2 += 2 * kNumOperands;
*inp3 += 2 * kNumOperands;
}
// Unroll MulAdd3Way for four iterations
ALWAYS_INLINE void FourMulAdd3Way(const Packet a1, const Packet a2,
const Packet a3, const float** inp1,
const float** inp2, const float** inp3,
float** out) {
TwoMulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
TwoMulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
}
// Apply MulAdd3Way on 128 operands.
ALWAYS_INLINE void MulAdd3Way128(const Packet a1, const Packet a2,
const Packet a3, const float** inp1,
const float** inp2, const float** inp3,
float** out) {
if (kNumOperands == 8) {
FourMulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
FourMulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
FourMulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
FourMulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
} else {
DCHECK_LE(4 * kNumOperands, 128);
for (int i = 0; i < 128 / (4 * kNumOperands); ++i) {
MulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
MulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
MulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
MulAdd3Way(a1, a2, a3, inp1, inp2, inp3, out);
}
}
}
// Computes product of "left_slices" with "num_cols" columns of "right", and
// stores the output in *"output".
// Note that left_slices is a list of SparseSlices, which are conceptually
// assumed to be concatenated along the column dimension. Also each SparseSlice
// is encoded as a list of blocks with upto N columns. See SparseSlice for more
// details.
template <typename TL, typename TR, int Cols>
inline void GEPP(
const std::vector<SparseSlice<TL>*>& left_slices,
const Eigen::TensorMap<Eigen::Tensor<const TR, 2, Eigen::RowMajor>,
Eigen::Aligned>& right,
const int num_cols, Matrix* output) {
const int cols = (Cols == -1) ? num_cols : Cols;
DCHECK_EQ(num_cols, cols);
const int right_num_cols = right.dimension(1);
const int output_num_cols = output->dimension(1);
static const int kNumOperandsR = kNumOperands * sizeof(float) / sizeof(TR);
const int cols_mod = cols % kNumOperandsR;
int k_offset = 0;
// Pre-compute pointers for output matrix.
float* out_ptrs[M];
float* const out_start = &(*output)(0, 0);
for (int j = 0; j < M; ++j) {
out_ptrs[j] = out_start + output_num_cols * j;
}
for (const auto* left_slice : left_slices) {
const auto& left = *left_slice;
const auto* data3 = (!left.data3.empty()) ? &left.data3[0] : nullptr;
const auto* data = (!left.data.empty()) ? &left.data[0] : nullptr;
const int num_blocks = left.index3_offset.size();
int begin3 = 0;
int begin = 0;
for (int i = 0; i < num_blocks; ++i) {
// Pre-compute pointers for right matrix
const TR* right_ptrs[K];
const auto* const right_start = &right(k_offset, 0);
DCHECK_LT(k_offset, right.dimension(0));
for (int j = 0; j < K; ++j) {
right_ptrs[j] = right_start + right_num_cols * j;
}
const int end3 = left.index3_offset[i];
int j = begin3;
// Loop unrolled for 2 iterations.
for (; j + 1 < end3; j += 2) {
Packet l1, l2, l3, nl1, nl2, nl3;
LoadSixScalars(&data3, &l1, &l2, &l3, &nl1, &nl2, &nl3);
const auto& index = left.index3[j];
const auto& nindex = left.index3[j + 1];
float* out = out_ptrs[index.m];
float* nout = out_ptrs[nindex.m];
const auto* r1 = right_ptrs[index.k1];
const auto* r2 = right_ptrs[index.k2];
const auto* r3 = right_ptrs[index.k3];
const auto* nr1 = right_ptrs[nindex.k1];
const auto* nr2 = right_ptrs[nindex.k2];
const auto* nr3 = right_ptrs[nindex.k3];
if (cols == 128) {
MulAdd3Way128(l1, l2, l3, &r1, &r2, &r3, &out);
MulAdd3Way128(nl1, nl2, nl3, &nr1, &nr2, &nr3, &nout);
} else {
for (int n = 0; n < cols / kNumOperandsR; ++n) {
MulAdd3Way(l1, l2, l3, &r1, &r2, &r3, &out);
MulAdd3Way(nl1, nl2, nl3, &nr1, &nr2, &nr3, &nout);
}
const float sl1 = Eigen::internal::pfirst<Packet>(l1);
const float sl2 = Eigen::internal::pfirst<Packet>(l2);
const float sl3 = Eigen::internal::pfirst<Packet>(l3);
const float nsl1 = Eigen::internal::pfirst<Packet>(nl1);
const float nsl2 = Eigen::internal::pfirst<Packet>(nl2);
const float nsl3 = Eigen::internal::pfirst<Packet>(nl3);
for (int k = 0; k < cols_mod; ++k) {
ScalarMulAdd3Way(sl1, sl2, sl3, &r1, &r2, &r3, &out);
ScalarMulAdd3Way(nsl1, nsl2, nsl3, &nr1, &nr2, &nr3, &nout);
}
}
}
if (j < end3) {
Packet l1, l2, l3;
LoadThreeScalars(&data3, &l1, &l2, &l3);
const auto& index = left.index3[j];
float* out = out_ptrs[index.m];
const auto* r1 = right_ptrs[index.k1];
const auto* r2 = right_ptrs[index.k2];
const auto* r3 = right_ptrs[index.k3];
if (cols == 128) {
MulAdd3Way128(l1, l2, l3, &r1, &r2, &r3, &out);
} else {
for (int n = 0; n < cols / kNumOperandsR; ++n) {
MulAdd3Way(l1, l2, l3, &r1, &r2, &r3, &out);
}
const float sl1 = Eigen::internal::pfirst<Packet>(l1);
const float sl2 = Eigen::internal::pfirst<Packet>(l2);
const float sl3 = Eigen::internal::pfirst<Packet>(l3);
for (int k = 0; k < cols_mod; ++k) {
ScalarMulAdd3Way(sl1, sl2, sl3, &r1, &r2, &r3, &out);
}
}
}
begin3 = end3;
int end = left.index_offset[i];
// Loop unrolled for 4 iterations.
j = begin;
for (; j + 3 < end; j += 4) {
Packet l, nl, n2l, n3l;
LoadFourScalars(&data, &l, &nl, &n2l, &n3l);
const auto& index = left.index[j];
const auto& nindex = left.index[j + 1];
const auto& n2index = left.index[j + 2];
const auto& n3index = left.index[j + 3];
const auto* r = right_ptrs[index.k];
const auto* nr = right_ptrs[nindex.k];
const auto* n2r = right_ptrs[n2index.k];
const auto* n3r = right_ptrs[n3index.k];
float* out = out_ptrs[index.m];
float* nout = out_ptrs[nindex.m];
float* n2out = out_ptrs[n2index.m];
float* n3out = out_ptrs[n3index.m];
for (int n = 0; n < cols / kNumOperandsR; ++n) {
MulAdd(l, &r, &out);
MulAdd(nl, &nr, &nout);
MulAdd(n2l, &n2r, &n2out);
MulAdd(n3l, &n3r, &n3out);
}
const float sl1 = Eigen::internal::pfirst<Packet>(l);
const float sl2 = Eigen::internal::pfirst<Packet>(nl);
const float sl3 = Eigen::internal::pfirst<Packet>(n2l);
const float sl4 = Eigen::internal::pfirst<Packet>(n3l);
for (int k = 0; k < cols_mod; ++k) {
ScalarMulAdd(sl1, &r, &out);
ScalarMulAdd(sl2, &nr, &nout);
ScalarMulAdd(sl3, &n2r, &n2out);
ScalarMulAdd(sl4, &n3r, &n3out);
}
}
while (j < end) {
Packet l;
LoadSingleScalar(&data, &l);
const auto& index = left.index[j];
const auto* r = right_ptrs[index.k];
float* out = out_ptrs[index.m];
for (int n = 0; n < cols / kNumOperandsR; ++n) {
MulAdd(l, &r, &out);
}
const float sl = Eigen::internal::pfirst<Packet>(l);
for (int k = 0; k < cols_mod; ++k) {
ScalarMulAdd(sl, &r, &out);
}
j++;
}
k_offset += left.block_size;
begin = end;
}
}
}
#undef LOAD
#undef EXPAND_BFLOAT_L
#undef EXPAND_BFLOAT_U
#undef STORE
#undef FMA
} // namespace
template <typename TL, typename TR>
class SparseMatMul {
using MatrixL = BasicMatrix<TL>;
using MatrixR = BasicMatrix<TR>;
using ConstMatrixMapL = BasicMatrixMap<const TL>;
using ConstMatrixMapR = BasicMatrixMap<const TR>;
using MatrixMapR = BasicMatrixMap<TR>;
public:
// Perform matrix multiplication of "left" and "right", and store the result
// in *"output".
public:
static inline void Compute(const ConstMatrixMapL& left,
const ConstMatrixMapR& right, bool transpose_left,
const DeviceBase::CpuWorkerThreads* thread_pool,
bool transpose_output, MatrixMap* output);
private:
// Computes multiplication of left and num_cols columns of right, and stores
// the output block in *"output" at offsets "output_row_offset" and
// "output_col_offset". If assign is true, assigns the value to that block,
// else adds the values to the existing values.
static inline void ComputeOutputBlock(
const std::vector<SparseSlice<TL>*>& left, const ConstMatrixMapR& right,
int num_cols, int output_row_offset, int output_col_offset, bool assign,
bool transpose_output, MatrixMap* output);
// Encodes "mat" using a sparse representation and stores that in
// "mat_slices". "mat" is broken into a grid with sizes "slice_num_rows" and
// "slice_num_cols", each grid element is converted into a SparseSlice and
// stored in mat_slices. "slice_block_size" is used to perform further column
// blocking of each slice.
static inline std::unique_ptr<BlockingCounter> CreateSparseSlices(
const ConstMatrixMapL& mat, bool transpose, int slice_num_rows,
int slice_block_size, int slice_num_cols,
std::vector<std::vector<SparseSlice<TL>*>>* mat_slices,
const DeviceBase::CpuWorkerThreads* thread_pool);
// This function chops "mat" along column dimension into pieces with at most N
// columns, and concatenates the pieces one after the other in "buffer". It
// returns the list of the pieces in "slices". It returns a BlockingCounter
// which should be used to wait for the shuffle operations to complete.
static inline std::unique_ptr<BlockingCounter> CreateDenseSlices(
const ConstMatrixMapR& mat, int row_start, int num_rows, int col_start,
int num_cols, const DeviceBase::CpuWorkerThreads* thread_pool,
MatrixR* buffer, std::vector<ConstMatrixMapR*>* slices);
// Helper function for CreateDenseSlices to move the data around. It returns a
// BlockingCounter which should be used to wait for the shuffle operations to
// complete.
static inline BlockingCounter* ShuffleMatrix(
const ConstMatrixMapR& mat, int slice_row_start, int slice_num_rows,
int slice_col_start, int slice_num_cols, const int N,
const DeviceBase::CpuWorkerThreads* thread_pool, MatrixR* buffer);
// Helper function for CreateDenseSlices to create slices.
static inline void SliceMatrix(const MatrixR& mat, const int num_rows,
const int num_slices,
std::vector<ConstMatrixMapR*>* slices);
// Heuristics to compute various block sizes.
// KR, NR: block sizes for "right". We run blocking iterations that operate on
// matrices with at most this size.
// KL: grid size along the column dimension used while encoding left.
// IB, JB: number of left and right slices to multiply together. This is used
// for ordering different ComputeBlockOutput operations inside each blocking
// iteration so as to potentially reduce the working set size.
static inline void ComputeBlockSizes(const ConstMatrixMapL& left,
const ConstMatrixMapR& right,
bool transpose_left, int num_threads,
int* KR, int* NR, int* KL, int* JB,
int* IB);
TF_DISALLOW_COPY_AND_ASSIGN(SparseMatMul);
};
template <typename TL, typename TR,
template <typename TL2, typename TR2> class DoMatMul>
class SparseMatMulOp : public OpKernel {
using MatrixR = BasicMatrix<TR>;
using ConstMatrixMapR = BasicMatrixMap<const TR>;
public:
explicit SparseMatMulOp(OpKernelConstruction* ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("transpose_a", &transpose_a_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("transpose_b", &transpose_b_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("a_is_sparse", &a_is_sparse_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("b_is_sparse", &b_is_sparse_));
}
void Compute(OpKernelContext* ctx) override {
const Tensor& a = ctx->input(0);
const Tensor& b = ctx->input(1);
OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(a.shape()),
errors::InvalidArgument("a is not a matrix"));
OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(b.shape()),
errors::InvalidArgument("b is not a matrix"));
const int m = transpose_a_ ? a.dim_size(1) : a.dim_size(0);
const int k = transpose_a_ ? a.dim_size(0) : a.dim_size(1);
const int n = transpose_b_ ? b.dim_size(0) : b.dim_size(1);
const int k2 = transpose_b_ ? b.dim_size(1) : b.dim_size(0);
OP_REQUIRES(ctx, k == k2,
errors::InvalidArgument(
"Matrix size incompatible: a: ", a.shape().DebugString(),
", b: ", b.shape().DebugString()));
OP_REQUIRES(ctx, m >= 0 && n >= 0 && k >= 0,
errors::InvalidArgument(
"Matrix dimensions cannot be negative: a: ",
a.shape().DebugString(), ", b: ", b.shape().DebugString()));
Tensor* output = nullptr;
OP_REQUIRES_OK(ctx, ctx->allocate_output(0, TensorShape({m, n}), &output));
// Return early if at least one of the output dimension size is 0.
if (m == 0 || n == 0) {
return;
}
if (k == 0) {
// If the inner dimension k in the matrix multiplication is zero, we fill
// the output with zeros.
functor::SetZeroFunctor<CPUDevice, float> f;
f(ctx->eigen_device<CPUDevice>(), output->flat<float>());
return;
}
auto out = output->matrix<float>();
std::unique_ptr<Tensor> a_float;
std::unique_ptr<Tensor> b_float;
if (!a_is_sparse_ && !b_is_sparse_) {
auto left = &a;
auto right = &b;
// TODO(agarwal): multi-thread the conversions from bfloat16 to float.
if (std::is_same<TL, bfloat16>::value) {
a_float.reset(new Tensor(DT_FLOAT, a.shape()));
BFloat16ToFloat(a.flat<bfloat16>().data(),
a_float->flat<float>().data(), a.NumElements());
left = a_float.get();
}
if (std::is_same<TR, bfloat16>::value) {
b_float.reset(new Tensor(DT_FLOAT, b.shape()));
BFloat16ToFloat(b.flat<bfloat16>().data(),
b_float->flat<float>().data(), b.NumElements());
right = b_float.get();
}
Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair;
dim_pair[0].first = transpose_a_ ? 0 : 1;
dim_pair[0].second = transpose_b_ ? 1 : 0;
out.device(ctx->template eigen_device<CPUDevice>()) =
left->matrix<float>().contract(right->matrix<float>(), dim_pair);
return;
}
auto left = &a;
auto right = &b;
bool transpose_output = false;
bool transpose_a = transpose_a_;
bool transpose_b = transpose_b_;
if (!a_is_sparse_) {
// Swap the order of multiplications using the identity:
// A * B = (B' * A')'.
std::swap(left, right);
std::swap(transpose_a, transpose_b);
transpose_a = !transpose_a;
transpose_b = !transpose_b;
transpose_output = !transpose_output;
}
std::unique_ptr<Tensor> right_tr;
if (transpose_b) {
// TODO(agarwal): avoid transposing the matrix here and directly handle
// transpose in CreateDenseSlices.
OP_REQUIRES(ctx, right->dim_size(0) != 0,
errors::InvalidArgument("b has an entry 0 in it's shape."));
OP_REQUIRES(ctx, right->dim_size(1) != 0,
errors::InvalidArgument("b has an entry 0 in it's shape."));
right_tr.reset(
new Tensor(right->dtype(),
TensorShape({right->dim_size(1), right->dim_size(0)})));
const auto perm = dsizes_10();
if (transpose_output) {
right_tr->matrix<TL>().device(ctx->template eigen_device<CPUDevice>()) =
right->matrix<TL>().shuffle(perm);
} else {
right_tr->matrix<TR>().device(ctx->template eigen_device<CPUDevice>()) =
right->matrix<TR>().shuffle(perm);
}
right = right_tr.get();
}
if (transpose_output) {
DoMatMul<TR, TL>::Compute(left->matrix<TR>(), right->matrix<TL>(),
transpose_a,
ctx->device()->tensorflow_cpu_worker_threads(),
transpose_output, &out);
} else {
DoMatMul<TL, TR>::Compute(left->matrix<TL>(), right->matrix<TR>(),
transpose_a,
ctx->device()->tensorflow_cpu_worker_threads(),
transpose_output, &out);
}
}
private:
bool transpose_a_;
bool transpose_b_;
bool a_is_sparse_;
bool b_is_sparse_;
TF_DISALLOW_COPY_AND_ASSIGN(SparseMatMulOp);
};
template <typename TL, typename TR>
inline void SparseMatMul<TL, TR>::ComputeOutputBlock(
const std::vector<SparseSlice<TL>*>& left,
const typename SparseMatMul<TL, TR>::ConstMatrixMapR& right, int num_cols,
int output_row_offset, int output_col_offset, bool assign,
bool transpose_output, MatrixMap* output) {
const auto perm = dsizes_10();
int num_rows = left[0]->num_rows;
const int rhs_num_cols = right.dimension(1);
DCHECK_LE(num_cols, rhs_num_cols);
Matrix out(num_rows, rhs_num_cols);
out.setZero();
if (num_cols == N) {
GEPP<TL, TR, N>(left, right, num_cols, &out);
} else {
GEPP<TL, TR, -1>(left, right, num_cols, &out);