forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
group_norm_kernel.cpp
1412 lines (1340 loc) · 50.1 KB
/
group_norm_kernel.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
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
#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
#include <ATen/native/group_norm.h>
#include <algorithm>
#include <array>
#include <numeric>
#include <ATen/core/Tensor.h>
#include <ATen/Dispatch.h>
#include <ATen/cpu/vec/vec.h>
#include <ATen/cpu/vec/functional.h>
#include <ATen/native/cpu/utils.h>
#include <ATen/native/cpu/moments_utils.h>
#include <ATen/native/cpu/mixed_data_type.h>
#include <ATen/OpMathType.h>
#include <c10/util/irange.h>
#ifndef AT_PER_OPERATOR_HEADERS
#include <ATen/Functions.h>
#else
#include <ATen/ops/empty.h>
#endif
namespace at::native {
namespace {
template <typename T, typename PT>
void GroupNormKernelImplInternal(
const Tensor& X,
const Tensor& gamma,
const Tensor& beta,
int64_t N,
int64_t C,
int64_t HxW,
int64_t group,
double eps,
Tensor& Y,
Tensor& mean,
Tensor& rstd) {
TORCH_CHECK(X.numel() == N * C * HxW);
TORCH_CHECK(!gamma.defined() || gamma.numel() == C);
TORCH_CHECK(!beta.defined() || beta.numel() == C);
const int64_t G = group;
const int64_t D = C / G;
const T* X_data = X.data_ptr<T>();
const PT* gamma_data = gamma.defined() ? gamma.data_ptr<PT>() : nullptr;
const PT* beta_data = beta.defined() ? beta.data_ptr<PT>() : nullptr;
T* Y_data = Y.data_ptr<T>();
PT* mean_data = mean.data_ptr<PT>();
PT* rstd_data = rstd.data_ptr<PT>();
const bool gamma_null = (gamma_data == nullptr);
const bool beta_null = beta_data == nullptr;
const int64_t inner_size = D * HxW;
using T_ACC = at::opmath_type<T>;
at::parallel_for(0, N * G, 1, [&](int64_t start, int64_t end) {
for (const auto i : c10::irange(start, end)) {
const T* X_ptr = X_data + i * inner_size;
T_ACC mean_val;
T_ACC rstd_val;
std::tie(mean_val, rstd_val) = RowwiseMoments(X_ptr, inner_size);
rstd_val = T_ACC(1) / std::sqrt(std::max(rstd_val, T_ACC(0)) + eps);
if (gamma_null && beta_null) {
T* Y_ptr = Y_data + i * inner_size;
for (const auto j : c10::irange(inner_size)) {
Y_ptr[j] = (X_ptr[j] - mean_val) * rstd_val;
}
} else {
const int64_t g = i % G;
for (const auto j : c10::irange(D)) {
const int64_t c = g * D + j;
const T_ACC scale = rstd_val * (gamma_null ? T_ACC(1) : T_ACC(gamma_data[c]));
const T_ACC bias = -scale * mean_val + (beta_null ? T_ACC(0) : T_ACC(beta_data[c]));
X_ptr = X_data + (i * D + j) * HxW;
T* Y_ptr = Y_data + (i * D + j) * HxW;
for (const auto k : c10::irange(HxW)) {
Y_ptr[k] = scale * X_ptr[k] + bias;
}
}
}
mean_data[i] = mean_val;
rstd_data[i] = rstd_val;
}
});
}
template <typename T>
std::tuple<T, T> ColumnwiseMoments(
const T* X_data,
int64_t HxW,
int64_t C,
int64_t D) {
using Vec = vec::Vectorized<T>;
constexpr int64_t K = Vec::size();
const int64_t inner_size = D / K * K;
Vec acc0_vec{0}, acc1_vec{0};
for (const auto m : c10::irange(HxW)) {
const T* X_ptr = X_data + m * C;
int64_t d = 0;
for (; d < inner_size; d += K) {
Vec x_vec = Vec::loadu(X_ptr + d);
acc0_vec += x_vec;
acc1_vec += x_vec * x_vec;
}
if (D - d > 0) {
Vec x_vec = Vec::loadu(X_ptr + d, D - d);
acc0_vec += x_vec;
acc1_vec += x_vec * x_vec;
}
}
T mean_val = vec::vec_reduce_all([](Vec& x, Vec& y) { return x + y; }, acc0_vec);
T rstd_val = vec::vec_reduce_all([](Vec& x, Vec& y) { return x + y; }, acc1_vec);
return std::tuple<T, T>(mean_val, rstd_val);
}
template <typename T = BFloat16>
std::tuple<float, float> ColumnwiseMoments(
const BFloat16* X_data,
int64_t HxW,
int64_t C,
int64_t D) {
using bVec = vec::Vectorized<BFloat16>;
using fVec = vec::Vectorized<float>;
constexpr int64_t K = bVec::size();
const int64_t inner_size = D / K * K;
fVec acc0_fvec{0}, acc1_fvec{0}, zero{0};
for (const auto m : c10::irange(HxW)) {
const BFloat16* X_ptr = X_data + m * C;
int64_t d = 0;
for (; d < inner_size; d += K) {
bVec x_bvec = bVec::loadu(X_ptr + d);
fVec x_fvec0, x_fvec1;
std::tie(x_fvec0, x_fvec1) = convert_bfloat16_float(x_bvec);
acc0_fvec += x_fvec0 + x_fvec1;
acc1_fvec += x_fvec0 * x_fvec0 + x_fvec1 * x_fvec1;
}
if (D - d > 0) {
bVec x_bvec = bVec::loadu(X_ptr + d, D - d);
fVec x_fvec0, x_fvec1;
std::tie(x_fvec0, x_fvec1) = convert_bfloat16_float(x_bvec);
if (D - d > fVec::size()) {
x_fvec1 = fVec::set(zero, x_fvec1, D - d - fVec::size());
acc0_fvec += x_fvec0 + x_fvec1;
acc1_fvec += x_fvec0 * x_fvec0 + x_fvec1 * x_fvec1;
} else {
x_fvec0 = fVec::set(zero, x_fvec0, D - d);
acc0_fvec += x_fvec0;
acc1_fvec += x_fvec0 * x_fvec0;
}
}
}
float mean_val = vec::vec_reduce_all([](fVec& x, fVec& y) { return x + y; }, acc0_fvec);
float rstd_val = vec::vec_reduce_all([](fVec& x, fVec& y) { return x + y; }, acc1_fvec);
return std::tuple<float, float>(mean_val, rstd_val);
}
template <typename scalar_t, typename param_t>
inline void CalcMeanVar(
const scalar_t* X_ptr,
param_t* mean_ptr,
param_t* rstd_ptr,
int64_t C) {
using Vec = vec::Vectorized<scalar_t>;
vec::map2<scalar_t>(
[](Vec x, Vec y) { return x + y; },
mean_ptr,
X_ptr,
mean_ptr,
C);
vec::map2<scalar_t>(
[](Vec x, Vec y) { return x * x + y; },
rstd_ptr,
X_ptr,
rstd_ptr,
C);
}
template <>
inline void CalcMeanVar(
const BFloat16* X_ptr,
float* mean_ptr,
float* rstd_ptr,
int64_t C) {
using fVec = vec::Vectorized<float>;
using bVec = vec::Vectorized<BFloat16>;
int64_t d = 0;
for (; d < C - (C % bVec::size()); d += bVec::size()) {
bVec data_bvec = bVec::loadu(X_ptr + d);
fVec mean_fvec0 = fVec::loadu(mean_ptr + d);
fVec mean_fvec1 = fVec::loadu(mean_ptr + d + fVec::size());
fVec rstd_fvec0 = fVec::loadu(rstd_ptr + d);
fVec rstd_fvec1 = fVec::loadu(rstd_ptr + d + fVec::size());
fVec data_fvec0, data_fvec1;
std::tie(data_fvec0, data_fvec1) = convert_bfloat16_float(data_bvec);
mean_fvec0 = data_fvec0 + mean_fvec0;
mean_fvec1 = data_fvec1 + mean_fvec1;
rstd_fvec0 = data_fvec0 * data_fvec0 + rstd_fvec0;
rstd_fvec1 = data_fvec1 * data_fvec1 + rstd_fvec1;
mean_fvec0.store(mean_ptr + d);
mean_fvec1.store(mean_ptr + d + fVec::size());
rstd_fvec0.store(rstd_ptr + d);
rstd_fvec1.store(rstd_ptr + d + fVec::size());
}
if (C - d > 0) {
bVec data_bvec = bVec::loadu(X_ptr + d, C - d);
fVec mean_fvec0 = fVec::loadu(mean_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
fVec mean_fvec1 = fVec::loadu(mean_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
fVec rstd_fvec0 = fVec::loadu(rstd_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
fVec rstd_fvec1 = fVec::loadu(rstd_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
fVec data_fvec0, data_fvec1;
std::tie(data_fvec0, data_fvec1) = convert_bfloat16_float(data_bvec);
mean_fvec0 = data_fvec0 + mean_fvec0;
mean_fvec1 = data_fvec1 + mean_fvec1;
rstd_fvec0 = data_fvec0 * data_fvec0 + rstd_fvec0;
rstd_fvec1 = data_fvec1 * data_fvec1 + rstd_fvec1;
mean_fvec0.store(mean_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
mean_fvec1.store(mean_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
rstd_fvec0.store(rstd_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
rstd_fvec1.store(rstd_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
}
}
template <typename scalar_t, typename param_t>
inline void ApplyScaleBias(
scalar_t* Y_ptr,
const scalar_t* X_ptr,
const param_t* scale_ptr,
const param_t* bias_ptr,
int64_t C) {
using Vec = vec::Vectorized<scalar_t>;
vec::map3<scalar_t>(
[](Vec x, Vec scale, Vec bias) { return x * scale + bias; },
Y_ptr,
X_ptr,
scale_ptr,
bias_ptr,
C);
}
template <>
inline void ApplyScaleBias(
BFloat16* Y_ptr,
const BFloat16* X_ptr,
const float* scale_ptr,
const float* bias_ptr,
int64_t C) {
using fVec = vec::Vectorized<float>;
using bVec = vec::Vectorized<BFloat16>;
int64_t d = 0;
for (; d < C - (C % bVec::size()); d += bVec::size()) {
bVec data_bvec = bVec::loadu(X_ptr + d);
fVec scale_fvec0 = fVec::loadu(scale_ptr + d);
fVec scale_fvec1 = fVec::loadu(scale_ptr + d + fVec::size());
fVec bias_fvec0 = fVec::loadu(bias_ptr + d);
fVec bias_fvec1 = fVec::loadu(bias_ptr + d + fVec::size());
fVec data_fvec0, data_fvec1, out0, out1;
std::tie(data_fvec0, data_fvec1) = convert_bfloat16_float(data_bvec);
out0 = data_fvec0 * scale_fvec0 + bias_fvec0;
out1 = data_fvec1 * scale_fvec1 + bias_fvec1;
convert_float_bfloat16(out0, out1).store(Y_ptr + d);
}
if (C - d > 0) {
bVec data_bvec = bVec::loadu(X_ptr + d, C - d);
fVec scale_fvec0 = fVec::loadu(scale_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
fVec scale_fvec1 = fVec::loadu(scale_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
fVec bias_fvec0 = fVec::loadu(bias_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
fVec bias_fvec1 = fVec::loadu(bias_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
fVec data_fvec0, data_fvec1, out0, out1;
std::tie(data_fvec0, data_fvec1) = convert_bfloat16_float(data_bvec);
out0 = data_fvec0 * scale_fvec0 + bias_fvec0;
out1 = data_fvec1 * scale_fvec1 + bias_fvec1;
convert_float_bfloat16(out0, out1).store(Y_ptr + d, C - d);
}
}
template <typename T, typename PT>
void GroupNormKernelImplChannelsLastInternal(
const Tensor& X,
const Tensor& gamma,
const Tensor& beta,
int64_t N,
int64_t C,
int64_t HxW,
int64_t group,
double eps,
Tensor& Y,
Tensor& mean,
Tensor& rstd) {
TORCH_CHECK(X.numel() == N * C * HxW);
TORCH_CHECK(!gamma.defined() || gamma.numel() == C);
TORCH_CHECK(!beta.defined() || beta.numel() == C);
const int64_t G = group;
const int64_t D = C / G;
const T* X_data = X.data_ptr<T>();
const PT* gamma_data = gamma.defined() ? gamma.data_ptr<PT>() : nullptr;
const PT* beta_data = beta.defined() ? beta.data_ptr<PT>() : nullptr;
T* Y_data = Y.data_ptr<T>();
PT* mean_data = mean.data_ptr<PT>();
PT* rstd_data = rstd.data_ptr<PT>();
using T_ACC = at::opmath_type<T>;
const T_ACC s = T_ACC(1) / static_cast<T_ACC>(D * HxW);
const bool gamma_null = (gamma_data == nullptr);
const bool beta_null = beta_data == nullptr;
// NB: About algorithm choosen:
//
// On channels last, GroupNorm has a input shape of {N, H, W, GD},
// Mean and rstd are collected per each n and g, which involves reduction
// on non-adjacent dimensions. We can parallel in the following 2 impls:
//
// impl-1: parallel on N * G. Only need one omp session but memory access
// per thread is non-contiguous.
//
// impl-2: parallel on N * HxW. Memory access per thread is contiguous,
// but requires help of extra temp buffer of size {T, N, 2C}.
//
// Generally impl-2 has better performance when HxW is large enough, so that
// data per thread {NHWC / T} is much larger then temp buffer per thread {2NC}
//
constexpr int64_t feature_map_threshold = 1024;
if (HxW < feature_map_threshold) {
// impl-1: parallel on N * G.
//
// for each plain of HxW, scale and bias is calculated only once
Tensor buffer = at::empty({N * G, 2 * D}, X.options().dtype(c10::CppTypeToScalarType<T_ACC>::value));
T_ACC* buffer_data = buffer.data_ptr<T_ACC>();
at::parallel_for(0, N * G, 1, [&](int64_t begin, int64_t end) {
int64_t n{0}, g{0};
data_index_init(begin, n, N, g, G);
for (const auto i : c10::irange(begin, end)) {
// step-1: for each n and g, collect sum of x and x2
//
// Note that using vec::map_reduce_all here is simpler to write
// but it is slower since horizontal reduce from vec to scalar is slow.
// So it is better to reduce with a vec across all HxW plain,
// and do a horizontal add just once for each {n, g}.
//
T_ACC mean_val, rstd_val;
std::tie(mean_val, rstd_val) = ColumnwiseMoments(
X_data + n * HxW * C + g * D,
HxW,
C,
D);
mean_val *= s;
rstd_val = std::max(rstd_val * s - mean_val * mean_val, T_ACC(0));
rstd_val = T_ACC(1) / std::sqrt(rstd_val + eps);
mean_data[i] = mean_val;
rstd_data[i] = rstd_val;
// step-2: calculate scale and bias
T_ACC* scale_ptr = buffer_data + i * 2 * D;
T_ACC* bias_ptr = scale_ptr + D;
for (const auto d : c10::irange(D)) {
const int64_t c = g * D + d;
scale_ptr[d] = rstd_val * (gamma_null ? T_ACC(1) : T_ACC(gamma_data[c]));
bias_ptr[d] = -scale_ptr[d] * mean_val + (beta_null ? T_ACC(0) : T_ACC(beta_data[c]));
}
// step-3: apply scale and bias
for (const auto m : c10::irange(HxW)) {
const T* X_ptr = X_data + n * HxW * C + m * C + g * D;
T* Y_ptr = Y_data + n * HxW * C + m * C + g * D;
ApplyScaleBias<T, T_ACC>(Y_ptr, X_ptr, scale_ptr, bias_ptr, D);
}
data_index_step(n, N, g, G);
}
});
} else {
// impl-2: parallel on N * HxW.
//
// temp buffer holding x and x2
int num_threads = at::get_num_threads();
Tensor buffer = at::empty({num_threads, N, 2 * C},
X.options().dtype(c10::CppTypeToScalarType<T_ACC>::value)).zero_();
T_ACC* buffer_data = buffer.data_ptr<T_ACC>();
Tensor tmp_buffer = at::empty({N, 2 * G},
X.options().dtype(c10::CppTypeToScalarType<T_ACC>::value));
T_ACC* tmp_buffer_data = tmp_buffer.data_ptr<T_ACC>();
// step-1: accumulate on dimension of C
//
// In order to improve multi-core performance when N=1,
// we parallel on the all the outer dimensions of N and HxW,
// leaving the most inner dimension C for vectorization.
//
// Note that parallel on {N, HxW, G} is not feasible for some common configs,
// e.g. say input shape is {1, 32, h, w} and G = 8,
// this will give D = 4 which is unable to take full SIMD length.
//
// To avoid thread conflict, we make use of a temp buffer of {T, N, 2C},
// firstly, reduce from {N, HxW, C} to {T, N, 2C}
//
at::parallel_for(0, N * HxW, 1, [&](int64_t begin, int64_t end) {
int tid = at::get_thread_num();
T_ACC* buffer_ptr = buffer_data + tid * N * 2 * C;
int64_t n{0}, m{0};
data_index_init(begin, n, N, m, HxW);
for (const auto i : c10::irange(begin, end)) {
T_ACC* mean_ptr = buffer_ptr + n * 2 * C;
T_ACC* rstd_ptr = mean_ptr + C;
const T* X_ptr = X_data + i * C;
CalcMeanVar<T, T_ACC>(X_ptr, mean_ptr, rstd_ptr, C);
data_index_step(n, N, m, HxW);
}
});
// step-2: compute mean and rstd
for (const auto n : c10::irange(N)) {
for (const auto g : c10::irange(G)) {
T_ACC mean_val{0}, rstd_val{0};
for (const auto d : c10::irange(D)) {
for (const auto t : c10::irange(num_threads)) {
T_ACC* buffer_ptr = buffer_data + t * N * 2 * C + n * 2 * C;
mean_val += buffer_ptr[g * D + d];
rstd_val += buffer_ptr[g * D + d + C];
}
}
mean_val *= s;
rstd_val = std::max(rstd_val * s - mean_val * mean_val, T_ACC(0));
rstd_val = T_ACC(1) / std::sqrt(rstd_val + eps);
tmp_buffer_data[n * 2 * G + 2 * g] = mean_val;
tmp_buffer_data[n * 2 * G + 2 * g + 1] = rstd_val;
}
}
// step-3: compute scale and bias
//
// mean/rstd have shape of {N, G}, gamma/beta have shape of {G, D}.
// And scale/bias have shape of {N, C} so that we can directly vectorize on
// dimension of C in the final step.
//
// We could fuse step 3 and 4 into a single session but this way is better:
// a. D might be too small for vectorization;
// b. Avoid duplicate caculation of scale/bias, each HxW plain share the same scale/bias
//
for (const auto n : c10::irange(N)) {
for (const auto g : c10::irange(G)) {
T_ACC* scale_ptr = buffer_data + n * 2 * C;
T_ACC* bias_ptr = scale_ptr + C;
T_ACC mean_val = tmp_buffer_data[n * 2 * G + 2 * g];
T_ACC rstd_val = tmp_buffer_data[n * 2 * G + 2 * g + 1];
mean_data[n * G + g] = mean_val;
rstd_data[n * G + g] = rstd_val;
for (const auto d : c10::irange(D)) {
const int64_t c = g * D + d;
scale_ptr[c] = rstd_val * (gamma_null ? T_ACC(1) : T_ACC(gamma_data[c]));
bias_ptr[c] = -scale_ptr[c] * mean_val + (beta_null ? T_ACC(0) : T_ACC(beta_data[c]));
}
}
}
// step-4: apply scale and bias
//
// Parallel on on the all the outer dimensions of N and HxW
// and vectorize on C.
//
at::parallel_for(0, N * HxW, 1, [&](int64_t begin, int64_t end) {
int64_t n{0}, m{0};
data_index_init(begin, n, N, m, HxW);
for (const auto i : c10::irange(begin, end)) {
const T* X_ptr = X_data + i * C;
T* Y_ptr = Y_data + i * C;
T_ACC* scale_ptr = buffer_data + n * 2 * C;
T_ACC* bias_ptr = scale_ptr + C;
ApplyScaleBias<T, T_ACC>(Y_ptr, X_ptr, scale_ptr, bias_ptr, C);
data_index_step(n, N, m, HxW);
}
});
}
}
void GroupNormKernelImpl(
const Tensor& X,
const Tensor& gamma,
const Tensor& beta,
int64_t N,
int64_t C,
int64_t HxW,
int64_t group,
double eps,
Tensor& Y,
Tensor& mean,
Tensor& rstd) {
const bool mixed_type = is_mixed_type(X, gamma, beta);
switch (X.suggest_memory_format()) {
case at::MemoryFormat::Contiguous: {
AT_DISPATCH_FLOATING_TYPES_AND(ScalarType::BFloat16, X.scalar_type(), "GroupNormKernelImpl", [&]() {
using param_t = at::opmath_type<scalar_t>;
if (mixed_type) {
GroupNormKernelImplInternal<scalar_t, param_t>(
X, gamma, beta, N, C, HxW, group, eps, Y, mean, rstd);
} else {
GroupNormKernelImplInternal<scalar_t, scalar_t>(
X, gamma, beta, N, C, HxW, group, eps, Y, mean, rstd);
}
});
break;
}
case at::MemoryFormat::ChannelsLast:
case at::MemoryFormat::ChannelsLast3d: {
AT_DISPATCH_FLOATING_TYPES_AND(ScalarType::BFloat16, X.scalar_type(), "GroupNormKernelImpl", [&]() {
using param_t = at::opmath_type<scalar_t>;
if (mixed_type) {
GroupNormKernelImplChannelsLastInternal<scalar_t, param_t>(
X, gamma, beta, N, C, HxW, group, eps, Y, mean, rstd);
} else {
GroupNormKernelImplChannelsLastInternal<scalar_t, scalar_t>(
X, gamma, beta, N, C, HxW, group, eps, Y, mean, rstd);
}
});
break;
}
default:
TORCH_CHECK(false, "Unsupported memory format. Supports only ChannelsLast, ChannelsLast3d, Contiguous");
}
}
template <typename T, typename PT>
void ComputeInternalGradients(
int64_t N,
int64_t C,
int64_t HxW,
const T* dY,
const T* X,
PT* ds,
PT* db) {
at::parallel_for(0, N * C, 1, [=](int64_t start, int64_t end) {
constexpr int64_t K = vec::Vectorized<T>::size();
const int64_t inner_size = HxW / K * K;
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
std::array<PT, K> ds_arr;
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
std::array<PT, K> db_arr;
for (const auto i : c10::irange(start, end)) {
const T* dY_ptr = dY + i * HxW;
const T* X_ptr = X + i * HxW;
vec::Vectorized<PT> ds_vec(0);
vec::Vectorized<PT> db_vec(0);
for (int64_t j = 0; j < inner_size; j += K) {
const vec::Vectorized<T> dy_vec = vec::Vectorized<T>::loadu(dY_ptr + j);
const vec::Vectorized<T> x_vec = vec::Vectorized<T>::loadu(X_ptr + j);
ds_vec = ds_vec + dy_vec * x_vec;
db_vec = db_vec + dy_vec;
}
ds_vec.store(ds_arr.data());
db_vec.store(db_arr.data());
PT ds_val = std::accumulate(ds_arr.cbegin(), ds_arr.cend(), PT(0));
PT db_val = std::accumulate(db_arr.cbegin(), db_arr.cend(), PT(0));
for (const auto j : c10::irange(inner_size, HxW)) {
ds_val += dY_ptr[j] * X_ptr[j];
db_val += dY_ptr[j];
}
ds[i] = ds_val;
db[i] = db_val;
}
});
}
template <>
void ComputeInternalGradients(
int64_t N,
int64_t C,
int64_t HxW,
const BFloat16* dY,
const BFloat16* X,
float* ds,
float* db) {
using bVec = vec::Vectorized<BFloat16>;
using fVec = vec::Vectorized<float>;
at::parallel_for(0, N * C, 1, [=](int64_t start, int64_t end) {
constexpr int64_t K = bVec::size();
const int64_t inner_size = HxW / K * K;
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
std::array<float, K/2> ds_arr;
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
std::array<float, K/2> db_arr;
for (const auto i : c10::irange(start, end)) {
const BFloat16* dY_ptr = dY + i * HxW;
const BFloat16* X_ptr = X + i * HxW;
fVec ds_vec(0);
fVec db_vec(0);
for (int64_t j = 0; j < inner_size; j += K) {
const bVec dy_bvec = bVec::loadu(dY_ptr + j);
const bVec x_bvec = bVec::loadu(X_ptr + j);
fVec x_fvec0, x_fvec1, dy_fvec0, dy_fvec1;
std::tie(x_fvec0, x_fvec1) = convert_bfloat16_float(x_bvec);
std::tie(dy_fvec0, dy_fvec1) = convert_bfloat16_float(dy_bvec);
ds_vec = ds_vec + dy_fvec0 * x_fvec0;
ds_vec = ds_vec + dy_fvec1 * x_fvec1;
db_vec = db_vec + dy_fvec0 + dy_fvec1;
}
ds_vec.store(ds_arr.data());
db_vec.store(db_arr.data());
float ds_val = std::accumulate(ds_arr.cbegin(), ds_arr.cend(), float(0));
float db_val = std::accumulate(db_arr.cbegin(), db_arr.cend(), float(0));
for (const auto j : c10::irange(inner_size, HxW)) {
ds_val += float(dY_ptr[j]) * float(X_ptr[j]);
db_val += float(dY_ptr[j]);
}
ds[i] = ds_val;
db[i] = db_val;
}
});
}
template <typename PT>
inline void CalcDsDb(
const PT* ds_ptr,
const PT* db_ptr,
bool gamma_null,
const PT* gamma_ptr,
const int64_t d,
const int64_t K,
void* ds_arr,
void* db_arr) {
vec::Vectorized<PT> ds_vec(0);
vec::Vectorized<PT> db_vec(0);
for (int64_t j = 0; j < d; j += K) {
const vec::Vectorized<PT> gamma_vec = gamma_null
? vec::Vectorized<PT>(1)
: vec::Vectorized<PT>::loadu(gamma_ptr + j);
ds_vec = ds_vec + vec::Vectorized<PT>::loadu(ds_ptr + j) * gamma_vec;
db_vec = db_vec + vec::Vectorized<PT>::loadu(db_ptr + j) * gamma_vec;
}
ds_vec.store(ds_arr);
db_vec.store(db_arr);
}
template <typename T, typename PT>
void GroupNormInputBackward(
int64_t N,
int64_t C,
int64_t HxW,
int64_t group,
const T* dY,
const T* X,
const PT* mean,
const PT* rstd,
const PT* gamma,
const PT* ds,
const PT* db,
T* dX) {
const int64_t G = group;
const int64_t D = C / G;
const PT s = PT(1) / static_cast<PT>(D * HxW);
const bool gamma_null = (gamma == nullptr);
at::parallel_for(0, N * G, 1, [=](int64_t start, int64_t end) {
constexpr int64_t K = vec::Vectorized<PT>::size();
const int64_t d = D / K * K;
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
std::array<PT, K> ds_arr;
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
std::array<PT, K> db_arr;
for (const auto i : c10::irange(start, end)) {
const int64_t g = i % G;
const PT* ds_ptr = ds + i * D;
const PT* db_ptr = db + i * D;
const PT* gamma_ptr = gamma + g * D;
CalcDsDb(ds_ptr, db_ptr, gamma_null, gamma_ptr, d, K, ds_arr.data(), db_arr.data());
PT ds_val = std::accumulate(ds_arr.cbegin(), ds_arr.cend(), PT(0));
PT db_val = std::accumulate(db_arr.cbegin(), db_arr.cend(), PT(0));
for (const auto j : c10::irange(d, D)) {
const PT gamma_v = gamma_null ? PT(1) : PT(gamma[g * D + j]);
ds_val += ds_ptr[j] * gamma_v;
db_val += db_ptr[j] * gamma_v;
}
const PT c2 =
(db_val * PT(mean[i]) - ds_val) * PT(rstd[i]) * PT(rstd[i]) * PT(rstd[i]) * s;
const PT c3 = -c2 * PT(mean[i]) - db_val * PT(rstd[i]) * s;
for (const auto j : c10::irange(D)) {
const int64_t c = g * D + j;
const T* dY_ptr = dY + (i * D + j) * HxW;
const T* X_ptr = X + (i * D + j) * HxW;
T* dX_ptr = dX + (i * D + j) * HxW;
const PT c1 = PT(rstd[i]) * (gamma_null ? PT(1) : PT(gamma[c]));
for (const auto k : c10::irange(HxW)) {
dX_ptr[k] = c1 * PT(dY_ptr[k]) + c2 * PT(X_ptr[k]) + c3;
}
}
}
});
}
template <typename T>
void GammaBackward(
int64_t N,
int64_t C,
int64_t group,
const T* mean,
const T* rstd,
const T* ds,
const T* db,
T* dgamma) {
const int64_t G = group;
const int64_t D = C / G;
constexpr int64_t K = vec::Vectorized<T>::size();
at::parallel_for(0, D, K, [=](int64_t start, int64_t end) {
for (const auto i : c10::irange(G)) {
std::memset(dgamma + i * D + start, 0, (end - start) * sizeof(T));
}
for (int64_t i = 0; i < N * G; ++i) {
const T* ds_ptr = ds + i * D;
const T* db_ptr = db + i * D;
const int64_t g = i % G;
for (const auto j : c10::irange(start, end)) {
const int64_t c = g * D + j;
dgamma[c] += (ds_ptr[j] - db_ptr[j] * mean[i]) * rstd[i];
}
}
});
}
template <typename T>
void BetaBackward(int64_t N, int64_t C, const T* db, T* dbeta) {
constexpr int64_t K = vec::Vectorized<T>::size();
at::parallel_for(0, C, K, [=](int64_t start, int64_t end) {
std::memset(dbeta + start, 0, (end - start) * sizeof(T));
for (const auto i : c10::irange(N)) {
const T* db_ptr = db + i * C;
for (const auto j : c10::irange(start, end)) {
dbeta[j] += db_ptr[j];
}
}
});
}
template <typename T, typename PT>
void GroupNormBackwardKernelImplInternal(
const Tensor& dY,
const Tensor& X,
const Tensor& mean,
const Tensor& rstd,
const Tensor& gamma,
int64_t N,
int64_t C,
int64_t HxW,
int64_t group,
Tensor& dX,
Tensor& dgamma,
Tensor& dbeta) {
TORCH_CHECK(dY.numel() == N * C * HxW);
TORCH_CHECK(X.numel() == N * C * HxW);
TORCH_CHECK(mean.numel() == N * group);
TORCH_CHECK(rstd.numel() == N * group);
TORCH_CHECK(!gamma.defined() || gamma.numel() == C);
const T* dY_data = dY.data_ptr<T>();
const T* X_data = X.data_ptr<T>();
const PT* mean_data = mean.data_ptr<PT>();
const PT* rstd_data = rstd.data_ptr<PT>();
const PT* gamma_data = gamma.defined() ? gamma.data_ptr<PT>() : nullptr;
T* dX_data = dX.defined() ? dX.data_ptr<T>() : nullptr;
PT* dgamma_data = dgamma.defined() ? dgamma.data_ptr<PT>() : nullptr;
PT* dbeta_data = dbeta.defined() ? dbeta.data_ptr<PT>() : nullptr;
Tensor ds = at::empty({N, C}, X.options().dtype(c10::CppTypeToScalarType<PT>::value));
Tensor db = at::empty({N, C}, X.options().dtype(c10::CppTypeToScalarType<PT>::value));
PT* ds_data = ds.data_ptr<PT>();
PT* db_data = db.data_ptr<PT>();
ComputeInternalGradients<T, PT>(N, C, HxW, dY_data, X_data, ds_data, db_data);
if (dX_data != nullptr) {
GroupNormInputBackward<T, PT>(
N,
C,
HxW,
group,
dY_data,
X_data,
mean_data,
rstd_data,
gamma_data,
ds_data,
db_data,
dX_data);
}
if (dgamma_data != nullptr) {
GammaBackward(
N, C, group, mean_data, rstd_data, ds_data, db_data, dgamma_data);
}
if (dbeta_data != nullptr) {
BetaBackward(N, C, db_data, dbeta_data);
}
}
template <typename T, typename PT>
inline void DsDbRowwiseMomentsChannelsLast(
const T* dY_ptr,
const T* X_ptr,
PT* ds_ptr,
PT* db_ptr,
int64_t C) {
using Vec = vec::Vectorized<T>;
constexpr int64_t K = vec::Vectorized<T>::size();
const int64_t inner_size = C / K * K;
int64_t d = 0;
for (; d < inner_size; d+= K) {
Vec ds_dev = Vec::loadu(ds_ptr + d);
Vec db_vec = Vec::loadu(db_ptr + d);
Vec x_vec = Vec::loadu(X_ptr + d);
Vec dy_vec = Vec::loadu(dY_ptr + d);
ds_dev += x_vec * dy_vec;
db_vec += dy_vec;
ds_dev.store(ds_ptr + d);
db_vec.store(db_ptr + d);
}
if (C - d > 0) {
Vec ds_dev = Vec::loadu(ds_ptr + d, C - d);
Vec db_vec = Vec::loadu(db_ptr + d, C - d);
Vec x_vec = Vec::loadu(X_ptr + d, C - d);
Vec dy_vec = Vec::loadu(dY_ptr + d, C - d);
ds_dev += x_vec * dy_vec;
db_vec += dy_vec;
ds_dev.store(ds_ptr + d, C - d);
db_vec.store(db_ptr + d, C - d);
}
}
template <>
inline void DsDbRowwiseMomentsChannelsLast(
const BFloat16* dY_ptr,
const BFloat16* X_ptr,
float* ds_ptr,
float* db_ptr,
int64_t C) {
using fVec = vec::Vectorized<float>;
using bVec = vec::Vectorized<BFloat16>;
int64_t d = 0;
for (; d < C - (C % bVec::size()); d += bVec::size()) {
fVec ds_dev0 = fVec::loadu(ds_ptr + d);
fVec ds_dev1 = fVec::loadu(ds_ptr + d + fVec::size());
fVec db_vec0 = fVec::loadu(db_ptr + d);
fVec db_vec1 = fVec::loadu(db_ptr + d + fVec::size());
bVec x_vec = bVec::loadu(X_ptr + d);
bVec dy_vec = bVec::loadu(dY_ptr + d);
fVec x_vec0, x_vec1, dy_vec0, dy_vec1;
std::tie(x_vec0, x_vec1) = convert_bfloat16_float(x_vec);
std::tie(dy_vec0, dy_vec1) = convert_bfloat16_float(dy_vec);
ds_dev0 += x_vec0 * dy_vec0;
ds_dev1 += x_vec1 * dy_vec1;
db_vec0 += dy_vec0;
db_vec1 += dy_vec1;
ds_dev0.store(ds_ptr + d);
ds_dev1.store(ds_ptr + d + fVec::size());
db_vec0.store(db_ptr + d);
db_vec1.store(db_ptr + d + fVec::size());
}
if (C - d > 0) {
fVec ds_dev0 = fVec::loadu(ds_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
fVec ds_dev1 = fVec::loadu(ds_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
fVec db_vec0 = fVec::loadu(db_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
fVec db_vec1 = fVec::loadu(db_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
bVec x_vec = bVec::loadu(X_ptr + d, C - d);
bVec dy_vec = bVec::loadu(dY_ptr + d, C - d);
fVec x_vec0, x_vec1, dy_vec0, dy_vec1;
std::tie(x_vec0, x_vec1) = convert_bfloat16_float(x_vec);
std::tie(dy_vec0, dy_vec1) = convert_bfloat16_float(dy_vec);
ds_dev0 += x_vec0 * dy_vec0;
ds_dev1 += x_vec1 * dy_vec1;
db_vec0 += dy_vec0;
db_vec1 += dy_vec1;
ds_dev0.store(ds_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
ds_dev1.store(ds_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
db_vec0.store(db_ptr + d, (C - d) > fVec::size() ? fVec::size() : (C - d));
db_vec1.store(db_ptr + d + fVec::size(), (C - d) > fVec::size() ? (C - d - fVec::size()) : 0);
}
}
template <typename T, typename PT>
inline void ApplyInputGradientsChannelsLastColMov(
const T* dY_data,
const T* X_data,
T* dX_data,
const PT* rstd,
const PT* gamma,
PT c2,
PT c3,
int64_t HxW,
int64_t C,
int64_t D) {
const bool gamma_null = (gamma == nullptr);
int64_t d = 0;
auto K = vec::Vectorized<T>::size();
for (; d < D / K * K; d += K) {
for (const auto m : c10::irange(HxW)) {
const T* X_ptr = X_data + m * C;
const T* dY_ptr = dY_data + m * C;
T* dX_ptr = dX_data + m * C;
auto c1 = vec::Vectorized<T>(*rstd) *
(gamma_null ? vec::Vectorized<T>(1) : vec::Vectorized<T>::loadu(gamma + d));
auto dy_vec = vec::Vectorized<T>::loadu(dY_ptr + d);
auto x_vec = vec::Vectorized<T>::loadu(X_ptr + d);
auto dx_vec = c1 * dy_vec +
vec::Vectorized<T>(c2) * x_vec + vec::Vectorized<T>(c3);
dx_vec.store(dX_ptr + d);
}
}
if (D - d > 0) {
for (const auto m : c10::irange(HxW)) {
const T* X_ptr = X_data + m * C;
const T* dY_ptr = dY_data + m * C;
T* dX_ptr = dX_data + m * C;
auto c1 = vec::Vectorized<T>(*rstd) *
(gamma_null ? vec::Vectorized<T>(1) : vec::Vectorized<T>::loadu(gamma + d, D - d));
auto dy_vec = vec::Vectorized<T>::loadu(dY_ptr + d, D - d);
auto x_vec = vec::Vectorized<T>::loadu(X_ptr + d, D - d);
auto dx_vec = c1 * dy_vec +
vec::Vectorized<T>(c2) * x_vec + vec::Vectorized<T>(c3);
dx_vec.store(dX_ptr + d, D - d);
}
}
}
template <>
inline void ApplyInputGradientsChannelsLastColMov(
const BFloat16* dY_data,
const BFloat16* X_data,
BFloat16* dX_data,
const float* rstd,
const float* gamma,
float c2,
float c3,
int64_t HxW,
int64_t C,
int64_t D) {
using bVec = vec::Vectorized<BFloat16>;
using fVec = vec::Vectorized<float>;
const bool gamma_null = (gamma == nullptr);
auto K = bVec::size();
int64_t d = 0;
for (; d < D / K * K; d += K) {
for (const auto m : c10::irange(HxW)) {
const BFloat16* X_ptr = X_data + m * C;
const BFloat16* dY_ptr = dY_data + m * C;
BFloat16* dX_ptr = dX_data + m * C;
fVec c1_0 = fVec(*rstd) *
(gamma_null ? fVec(1) : fVec::loadu(gamma + d));
fVec c1_1 = fVec(*rstd) *
(gamma_null ? fVec(1) : fVec::loadu(gamma + d + fVec::size()));
bVec dy_vec = bVec::loadu(dY_ptr + d);
bVec x_vec = bVec::loadu(X_ptr + d);
fVec dy_vec0, dy_vec1, x_vec0, x_vec1;
std::tie(x_vec0, x_vec1) = convert_bfloat16_float(x_vec);
std::tie(dy_vec0, dy_vec1) = convert_bfloat16_float(dy_vec);
fVec dx_vec0 = c1_0 * dy_vec0 + fVec(c2) * x_vec0 + fVec(c3);
fVec dx_vec1 = c1_1 * dy_vec1 + fVec(c2) * x_vec1 + fVec(c3);
convert_float_bfloat16(dx_vec0, dx_vec1).store(dX_ptr + d);
}
}
if (D - d > 0) {
for (const auto m : c10::irange(HxW)) {
const BFloat16* X_ptr = X_data + m * C;
const BFloat16* dY_ptr = dY_data + m * C;
BFloat16* dX_ptr = dX_data + m * C;
fVec c1_0 = fVec(*rstd) *
(gamma_null ? fVec(1) : fVec::loadu(gamma + d, (D - d) > fVec::size() ? fVec::size() : (D - d)));
fVec c1_1 = fVec(*rstd) *
(gamma_null ? fVec(1) : fVec::loadu(gamma + d + fVec::size(), (D - d) > fVec::size() ? (D - d - fVec::size()) : 0));
bVec dy_vec = bVec::loadu(dY_ptr + d, D - d);
bVec x_vec = bVec::loadu(X_ptr + d, D - d);
fVec dy_vec0, dy_vec1, x_vec0, x_vec1;
std::tie(x_vec0, x_vec1) = convert_bfloat16_float(x_vec);
std::tie(dy_vec0, dy_vec1) = convert_bfloat16_float(dy_vec);
fVec dx_vec0 = c1_0 * dy_vec0 + fVec(c2) * x_vec0 + fVec(c3);
fVec dx_vec1 = c1_1 * dy_vec1 + fVec(c2) * x_vec1 + fVec(c3);
convert_float_bfloat16(dx_vec0, dx_vec1).store(dX_ptr + d, D - d);
}
}
}
template <typename T, typename PT>
inline void ApplyInputGradientsChannelsLastRowMov(
const T* dY_data,
const T* X_data,
T* dX_data,
const PT* rstd,
const PT* gamma,
PT c2,
PT c3,
int64_t HxW,
int64_t C,
int64_t D) {
const bool gamma_null = (gamma == nullptr);
int64_t d = 0;
auto K = vec::Vectorized<T>::size();
for (; d < D / K * K; d += K) {
auto c1 = vec::Vectorized<T>(*rstd) *
(gamma_null ? vec::Vectorized<T>(1) : vec::Vectorized<T>::loadu(gamma + d));