forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
GridSampler.cpp
1057 lines (951 loc) · 49.1 KB
/
GridSampler.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/GridSampler.h>
#include <ATen/native/GridSamplerUtils.h>
#include <ATen/core/Tensor.h>
#include <ATen/Dispatch.h>
#include <ATen/Parallel.h>
#include <ATen/cpu/vec/vec.h>
#include <ATen/native/UpSample.h>
#include <ATen/native/cpu/GridSamplerKernel.h>
#include <c10/util/Exception.h>
#include <c10/util/irange.h>
#ifndef AT_PER_OPERATOR_HEADERS
#include <ATen/Functions.h>
#include <ATen/NativeFunctions.h>
#else
#include <ATen/ops/_empty_affine_quantized.h>
#include <ATen/ops/_grid_sampler_2d_cpu_fallback_backward_native.h>
#include <ATen/ops/_grid_sampler_2d_cpu_fallback_native.h>
#include <ATen/ops/cudnn_grid_sampler.h>
#include <ATen/ops/empty.h>
#include <ATen/ops/empty_like.h>
#include <ATen/ops/grid_sampler_2d.h>
#include <ATen/ops/grid_sampler_2d_backward_native.h>
#include <ATen/ops/grid_sampler_2d_native.h>
#include <ATen/ops/grid_sampler_3d.h>
#include <ATen/ops/grid_sampler_3d_backward_native.h>
#include <ATen/ops/grid_sampler_3d_native.h>
#include <ATen/ops/grid_sampler_native.h>
#include <ATen/ops/zeros_like.h>
#endif
namespace at { namespace native {
using at::native::detail::GridSamplerInterpolation;
using at::native::detail::GridSamplerPadding;
namespace {
template<typename scalar_t>
Tensor grid_sampler_3d_cpu_impl(const Tensor& input, const Tensor& grid,
GridSamplerInterpolation interpolation_mode,
GridSamplerPadding padding_mode,
bool align_corners) {
// See NOTE [ grid_sampler Native Functions ].
// Add checks here in case this is called instead of grid_sampler.
check_grid_sampler_common(input, grid);
check_grid_sampler_3d(
input, grid, static_cast<int64_t>(interpolation_mode));
int64_t N = input.size(0);
int64_t C = input.size(1);
int64_t inp_D = input.size(2);
int64_t inp_H = input.size(3);
int64_t inp_W = input.size(4);
int64_t out_D = grid.size(1);
int64_t out_H = grid.size(2);
int64_t out_W = grid.size(3);
auto output = at::empty({N, C, out_D, out_H, out_W}, input.options());
int64_t inp_sN = input.stride(0);
int64_t inp_sC = input.stride(1);
int64_t inp_sD = input.stride(2);
int64_t inp_sH = input.stride(3);
int64_t inp_sW = input.stride(4);
int64_t grid_sN = grid.stride(0);
int64_t grid_sD = grid.stride(1);
int64_t grid_sH = grid.stride(2);
int64_t grid_sW = grid.stride(3);
int64_t grid_sCoor = grid.stride(4);
int64_t out_sN = output.stride(0);
int64_t out_sC = output.stride(1);
int64_t out_sD = output.stride(2);
int64_t out_sH = output.stride(3);
int64_t out_sW = output.stride(4);
scalar_t *inp_ptr = input.data_ptr<scalar_t>();
scalar_t *out_ptr = output.data_ptr<scalar_t>();
scalar_t *grid_ptr = grid.data_ptr<scalar_t>();
// loop over each output pixel
at::parallel_for(0, N, 0, [&](int64_t start, int64_t end) {
for (const auto n : c10::irange(start, end)) {
scalar_t *grid_ptr_N = grid_ptr + n * grid_sN;
scalar_t *inp_ptr_N = inp_ptr + n * inp_sN;
for (const auto d : c10::irange(out_D)) {
for (const auto h : c10::irange(out_H)) {
for (const auto w : c10::irange(out_W)) {
// get the corresponding input x, y, z co-ordinates from grid
scalar_t *grid_ptr_NDHW = grid_ptr_N + d * grid_sD + h * grid_sH + w * grid_sW;
scalar_t ix = *grid_ptr_NDHW;
scalar_t iy = grid_ptr_NDHW[grid_sCoor];
scalar_t iz = grid_ptr_NDHW[2 * grid_sCoor];
ix = grid_sampler_compute_source_index(ix, inp_W, padding_mode, align_corners);
iy = grid_sampler_compute_source_index(iy, inp_H, padding_mode, align_corners);
iz = grid_sampler_compute_source_index(iz, inp_D, padding_mode, align_corners);
if (interpolation_mode == GridSamplerInterpolation::Bilinear) {
// get corner pixel values from (x, y, z)
// for 4d, we used north-east-south-west
// for 5d, we add top-bottom
int64_t ix_tnw = static_cast<int64_t>(std::floor(ix));
int64_t iy_tnw = static_cast<int64_t>(std::floor(iy));
int64_t iz_tnw = static_cast<int64_t>(std::floor(iz));
int64_t ix_tne = ix_tnw + 1;
int64_t iy_tne = iy_tnw;
int64_t iz_tne = iz_tnw;
int64_t ix_tsw = ix_tnw;
int64_t iy_tsw = iy_tnw + 1;
int64_t iz_tsw = iz_tnw;
int64_t ix_tse = ix_tnw + 1;
int64_t iy_tse = iy_tnw + 1;
int64_t iz_tse = iz_tnw;
int64_t ix_bnw = ix_tnw;
int64_t iy_bnw = iy_tnw;
int64_t iz_bnw = iz_tnw + 1;
int64_t ix_bne = ix_tnw + 1;
int64_t iy_bne = iy_tnw;
int64_t iz_bne = iz_tnw + 1;
int64_t ix_bsw = ix_tnw;
int64_t iy_bsw = iy_tnw + 1;
int64_t iz_bsw = iz_tnw + 1;
int64_t ix_bse = ix_tnw + 1;
int64_t iy_bse = iy_tnw + 1;
int64_t iz_bse = iz_tnw + 1;
// get surfaces to each neighbor:
scalar_t tnw = (ix_bse - ix) * (iy_bse - iy) * (iz_bse - iz);
scalar_t tne = (ix - ix_bsw) * (iy_bsw - iy) * (iz_bsw - iz);
scalar_t tsw = (ix_bne - ix) * (iy - iy_bne) * (iz_bne - iz);
scalar_t tse = (ix - ix_bnw) * (iy - iy_bnw) * (iz_bnw - iz);
scalar_t bnw = (ix_tse - ix) * (iy_tse - iy) * (iz - iz_tse);
scalar_t bne = (ix - ix_tsw) * (iy_tsw - iy) * (iz - iz_tsw);
scalar_t bsw = (ix_tne - ix) * (iy - iy_tne) * (iz - iz_tne);
scalar_t bse = (ix - ix_tnw) * (iy - iy_tnw) * (iz - iz_tnw);
// calculate bilinear weighted pixel value and set output pixel
scalar_t *out_ptr_NCDHW = out_ptr + n * out_sN + d * out_sD + h * out_sH + w * out_sW;
scalar_t *inp_ptr_NC = inp_ptr_N;
for (int64_t c = 0; c < C; ++c, out_ptr_NCDHW += out_sC, inp_ptr_NC += inp_sC) {
// (c, iz_tnw, iy_tnw, ix_tnw) * tnw + (c, iz_tne, iy_tne, ix_tne) * tne
// + (c, iz_tsw, iy_tsw, ix_tsw) * tsw + (c, iz_tse, iy_tse, ix_tse) * tse
// + (c, iz_bnw, iy_bnw, ix_bnw) * bnw + (c, iz_bne, iy_bne, ix_bne) * bne
// + (c, iz_bsw, iy_bsw, ix_bsw) * bsw + (c, iz_bse, iy_bse, ix_bse) * bse
*out_ptr_NCDHW = static_cast<scalar_t>(0);
if (within_bounds_3d(iz_tnw, iy_tnw, ix_tnw, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW += inp_ptr_NC[iz_tnw * inp_sD + iy_tnw * inp_sH + ix_tnw * inp_sW] * tnw;
}
if (within_bounds_3d(iz_tne, iy_tne, ix_tne, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW += inp_ptr_NC[iz_tne * inp_sD + iy_tne * inp_sH + ix_tne * inp_sW] * tne;
}
if (within_bounds_3d(iz_tsw, iy_tsw, ix_tsw, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW += inp_ptr_NC[iz_tsw * inp_sD + iy_tsw * inp_sH + ix_tsw * inp_sW] * tsw;
}
if (within_bounds_3d(iz_tse, iy_tse, ix_tse, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW += inp_ptr_NC[iz_tse * inp_sD + iy_tse * inp_sH + ix_tse * inp_sW] * tse;
}
if (within_bounds_3d(iz_bnw, iy_bnw, ix_bnw, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW += inp_ptr_NC[iz_bnw * inp_sD + iy_bnw * inp_sH + ix_bnw * inp_sW] * bnw;
}
if (within_bounds_3d(iz_bne, iy_bne, ix_bne, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW += inp_ptr_NC[iz_bne * inp_sD + iy_bne * inp_sH + ix_bne * inp_sW] * bne;
}
if (within_bounds_3d(iz_bsw, iy_bsw, ix_bsw, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW += inp_ptr_NC[iz_bsw * inp_sD + iy_bsw * inp_sH + ix_bsw * inp_sW] * bsw;
}
if (within_bounds_3d(iz_bse, iy_bse, ix_bse, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW += inp_ptr_NC[iz_bse * inp_sD + iy_bse * inp_sH + ix_bse * inp_sW] * bse;
}
}
} else if (interpolation_mode == GridSamplerInterpolation::Nearest) {
int64_t ix_nearest = static_cast<int64_t>(std::round(ix));
int64_t iy_nearest = static_cast<int64_t>(std::round(iy));
int64_t iz_nearest = static_cast<int64_t>(std::round(iz));
// assign nearest neighor pixel value to output pixel
scalar_t *out_ptr_NCDHW = out_ptr + n * out_sN + d * out_sD + h * out_sH + w * out_sW;
scalar_t *inp_ptr_NC = inp_ptr_N;
for (int64_t c = 0; c < C; ++c, out_ptr_NCDHW += out_sC, inp_ptr_NC += inp_sC) {
if (within_bounds_3d(iz_nearest, iy_nearest, ix_nearest, inp_D, inp_H, inp_W)) {
*out_ptr_NCDHW = inp_ptr_NC[iz_nearest * inp_sD + iy_nearest * inp_sH + ix_nearest * inp_sW];
} else {
*out_ptr_NCDHW = static_cast<scalar_t>(0);
}
}
}
}
}
}
}
});
return output;
}
template<typename scalar_t>
std::tuple<Tensor, Tensor>
grid_sampler_3d_backward_cpu_impl(const Tensor& grad_output,
const Tensor& input, const Tensor& grid,
GridSamplerInterpolation interpolation_mode,
GridSamplerPadding padding_mode,
bool align_corners, std::array<bool,2> output_mask) {
// See NOTE [ grid_sampler Native Functions ].
// Add checks here in case this is called instead of grid_sampler.
check_grid_sampler_common(input, grid);
check_grid_sampler_3d(
input, grid, static_cast<int64_t>(interpolation_mode));
auto input_requires_grad = output_mask[0];
Tensor grad_input = ([&]() {
if (input_requires_grad) {
return at::zeros_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
} else {
return Tensor();
}
})();
auto grad_grid = at::empty_like(grid, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
// If interpolation mode is Nearest, then grad_grid is not filled in the
// loop below.
if (interpolation_mode == GridSamplerInterpolation::Nearest) {
grad_grid.zero_();
}
int64_t N = input.size(0);
int64_t C = input.size(1);
int64_t inp_D = input.size(2);
int64_t inp_H = input.size(3);
int64_t inp_W = input.size(4);
int64_t out_D = grid.size(1);
int64_t out_H = grid.size(2);
int64_t out_W = grid.size(3);
int64_t inp_sN = input.stride(0);
int64_t inp_sC = input.stride(1);
int64_t inp_sD = input.stride(2);
int64_t inp_sH = input.stride(3);
int64_t inp_sW = input.stride(4);
int64_t grid_sN = grid.stride(0);
int64_t grid_sD = grid.stride(1);
int64_t grid_sH = grid.stride(2);
int64_t grid_sW = grid.stride(3);
int64_t grid_sCoor = grid.stride(4);
int64_t gOut_sN = grad_output.stride(0);
int64_t gOut_sC = grad_output.stride(1);
int64_t gOut_sD = grad_output.stride(2);
int64_t gOut_sH = grad_output.stride(3);
int64_t gOut_sW = grad_output.stride(4);
int64_t gInp_sN = 0;
int64_t gInp_sC = 0;
int64_t gInp_sD = 0;
int64_t gInp_sH = 0;
int64_t gInp_sW = 0;
if (input_requires_grad) {
gInp_sN = grad_input.stride(0);
gInp_sC = grad_input.stride(1);
gInp_sD = grad_input.stride(2);
gInp_sH = grad_input.stride(3);
gInp_sW = grad_input.stride(4);
}
int64_t gGrid_sN = grad_grid.stride(0);
int64_t gGrid_sW = grad_grid.stride(3);
scalar_t *inp_ptr = input.data_ptr<scalar_t>();
scalar_t *grid_ptr = grid.data_ptr<scalar_t>();
scalar_t *gOut_ptr = grad_output.data_ptr<scalar_t>();
scalar_t *gInp_ptr = nullptr;
if (input_requires_grad) {
gInp_ptr = grad_input.data_ptr<scalar_t>();
}
scalar_t *gGrid_ptr = grad_grid.data_ptr<scalar_t>();
// loop over each output pixel
at::parallel_for(0, N, 0, [&](int64_t start, int64_t end) {
for (const auto n : c10::irange(start, end)) {
scalar_t *grid_ptr_N = grid_ptr + n * grid_sN;
scalar_t *inp_ptr_N = inp_ptr + n * inp_sN;
scalar_t *gGrid_ptr_NDHW = gGrid_ptr + n * gGrid_sN;
for (const auto d : c10::irange(out_D)) {
for (const auto h : c10::irange(out_H)) {
for (int64_t w = 0; w < out_W; ++w, gGrid_ptr_NDHW += gGrid_sW /* grad_grid is contiguous */ ) {
// get the corresponding input x, y, z co-ordinates from grid
scalar_t *grid_ptr_NDHW = grid_ptr_N + d * grid_sD + h * grid_sH + w * grid_sW;
scalar_t ix = *grid_ptr_NDHW;
scalar_t iy = grid_ptr_NDHW[grid_sCoor];
scalar_t iz = grid_ptr_NDHW[2 * grid_sCoor];
// multipliers for gradients on ix, iy, and iz
scalar_t gix_mult, giy_mult, giz_mult;
ix = grid_sampler_compute_source_index_set_grad(ix, inp_W, padding_mode, align_corners, &gix_mult);
iy = grid_sampler_compute_source_index_set_grad(iy, inp_H, padding_mode, align_corners, &giy_mult);
iz = grid_sampler_compute_source_index_set_grad(iz, inp_D, padding_mode, align_corners, &giz_mult);
if (interpolation_mode == GridSamplerInterpolation::Bilinear) {
// get corner pixel values from (x, y, z)
// for 4d, we used north-east-south-west
// for 5d, we add top-bottom
int64_t ix_tnw = static_cast<int64_t>(std::floor(ix));
int64_t iy_tnw = static_cast<int64_t>(std::floor(iy));
int64_t iz_tnw = static_cast<int64_t>(std::floor(iz));
int64_t ix_tne = ix_tnw + 1;
int64_t iy_tne = iy_tnw;
int64_t iz_tne = iz_tnw;
int64_t ix_tsw = ix_tnw;
int64_t iy_tsw = iy_tnw + 1;
int64_t iz_tsw = iz_tnw;
int64_t ix_tse = ix_tnw + 1;
int64_t iy_tse = iy_tnw + 1;
int64_t iz_tse = iz_tnw;
int64_t ix_bnw = ix_tnw;
int64_t iy_bnw = iy_tnw;
int64_t iz_bnw = iz_tnw + 1;
int64_t ix_bne = ix_tnw + 1;
int64_t iy_bne = iy_tnw;
int64_t iz_bne = iz_tnw + 1;
int64_t ix_bsw = ix_tnw;
int64_t iy_bsw = iy_tnw + 1;
int64_t iz_bsw = iz_tnw + 1;
int64_t ix_bse = ix_tnw + 1;
int64_t iy_bse = iy_tnw + 1;
int64_t iz_bse = iz_tnw + 1;
// get surfaces to each neighbor:
scalar_t tnw = (ix_bse - ix) * (iy_bse - iy) * (iz_bse - iz);
scalar_t tne = (ix - ix_bsw) * (iy_bsw - iy) * (iz_bsw - iz);
scalar_t tsw = (ix_bne - ix) * (iy - iy_bne) * (iz_bne - iz);
scalar_t tse = (ix - ix_bnw) * (iy - iy_bnw) * (iz_bnw - iz);
scalar_t bnw = (ix_tse - ix) * (iy_tse - iy) * (iz - iz_tse);
scalar_t bne = (ix - ix_tsw) * (iy_tsw - iy) * (iz - iz_tsw);
scalar_t bsw = (ix_tne - ix) * (iy - iy_tne) * (iz - iz_tne);
scalar_t bse = (ix - ix_tnw) * (iy - iy_tnw) * (iz - iz_tnw);
scalar_t gix = static_cast<scalar_t>(0), giy = static_cast<scalar_t>(0), giz = static_cast<scalar_t>(0);
scalar_t *gOut_ptr_NCDHW = gOut_ptr + n * gOut_sN + d * gOut_sD + h * gOut_sH + w * gOut_sW;
scalar_t *inp_ptr_NC = inp_ptr_N;
scalar_t *gInp_ptr_NC = gInp_ptr + n * gInp_sN;
// calculate bilinear weighted pixel value and set output pixel
for (int64_t c = 0; c < C; ++c, gOut_ptr_NCDHW += gOut_sC, gInp_ptr_NC += gInp_sC, inp_ptr_NC += inp_sC) {
scalar_t gOut = *gOut_ptr_NCDHW;
// calculate and set grad_input
if (input_requires_grad) {
safe_add_3d(gInp_ptr_NC, iz_tnw, iy_tnw, ix_tnw, gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, tnw * gOut);
safe_add_3d(gInp_ptr_NC, iz_tne, iy_tne, ix_tne, gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, tne * gOut);
safe_add_3d(gInp_ptr_NC, iz_tsw, iy_tsw, ix_tsw, gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, tsw * gOut);
safe_add_3d(gInp_ptr_NC, iz_tse, iy_tse, ix_tse, gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, tse * gOut);
safe_add_3d(gInp_ptr_NC, iz_bnw, iy_bnw, ix_bnw, gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, bnw * gOut);
safe_add_3d(gInp_ptr_NC, iz_bne, iy_bne, ix_bne, gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, bne * gOut);
safe_add_3d(gInp_ptr_NC, iz_bsw, iy_bsw, ix_bsw, gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, bsw * gOut);
safe_add_3d(gInp_ptr_NC, iz_bse, iy_bse, ix_bse, gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, bse * gOut);
}
// calculate grad_grid
if (within_bounds_3d(iz_tnw, iy_tnw, ix_tnw, inp_D, inp_H, inp_W)) {
scalar_t tnw_val = inp_ptr_NC[iz_tnw * inp_sD + iy_tnw * inp_sH + ix_tnw * inp_sW];
gix -= tnw_val * (iy_bse - iy) * (iz_bse - iz) * gOut;
giy -= tnw_val * (ix_bse - ix) * (iz_bse - iz) * gOut;
giz -= tnw_val * (ix_bse - ix) * (iy_bse - iy) * gOut;
}
if (within_bounds_3d(iz_tne, iy_tne, ix_tne, inp_D, inp_H, inp_W)) {
scalar_t tne_val = inp_ptr_NC[iz_tne * inp_sD + iy_tne * inp_sH + ix_tne * inp_sW];
gix += tne_val * (iy_bsw - iy) * (iz_bsw - iz) * gOut;
giy -= tne_val * (ix - ix_bsw) * (iz_bsw - iz) * gOut;
giz -= tne_val * (ix - ix_bsw) * (iy_bsw - iy) * gOut;
}
if (within_bounds_3d(iz_tsw, iy_tsw, ix_tsw, inp_D, inp_H, inp_W)) {
scalar_t tsw_val = inp_ptr_NC[iz_tsw * inp_sD + iy_tsw * inp_sH + ix_tsw * inp_sW];
gix -= tsw_val * (iy - iy_bne) * (iz_bne - iz) * gOut;
giy += tsw_val * (ix_bne - ix) * (iz_bne - iz) * gOut;
giz -= tsw_val * (ix_bne - ix) * (iy - iy_bne) * gOut;
}
if (within_bounds_3d(iz_tse, iy_tse, ix_tse, inp_D, inp_H, inp_W)) {
scalar_t tse_val = inp_ptr_NC[iz_tse * inp_sD + iy_tse * inp_sH + ix_tse * inp_sW];
gix += tse_val * (iy - iy_bnw) * (iz_bnw - iz) * gOut;
giy += tse_val * (ix - ix_bnw) * (iz_bnw - iz) * gOut;
giz -= tse_val * (ix - ix_bnw) * (iy - iy_bnw) * gOut;
}
if (within_bounds_3d(iz_bnw, iy_bnw, ix_bnw, inp_D, inp_H, inp_W)) {
scalar_t bnw_val = inp_ptr_NC[iz_bnw * inp_sD + iy_bnw * inp_sH + ix_bnw * inp_sW];
gix -= bnw_val * (iy_tse - iy) * (iz - iz_tse) * gOut;
giy -= bnw_val * (ix_tse - ix) * (iz - iz_tse) * gOut;
giz += bnw_val * (ix_tse - ix) * (iy_tse - iy) * gOut;
}
if (within_bounds_3d(iz_bne, iy_bne, ix_bne, inp_D, inp_H, inp_W)) {
scalar_t bne_val = inp_ptr_NC[iz_bne * inp_sD + iy_bne * inp_sH + ix_bne * inp_sW];
gix += bne_val * (iy_tsw - iy) * (iz - iz_tsw) * gOut;
giy -= bne_val * (ix - ix_tsw) * (iz - iz_tsw) * gOut;
giz += bne_val * (ix - ix_tsw) * (iy_tsw - iy) * gOut;
}
if (within_bounds_3d(iz_bsw, iy_bsw, ix_bsw, inp_D, inp_H, inp_W)) {
scalar_t bsw_val = inp_ptr_NC[iz_bsw * inp_sD + iy_bsw * inp_sH + ix_bsw * inp_sW];
gix -= bsw_val * (iy - iy_tne) * (iz - iz_tne) * gOut;
giy += bsw_val * (ix_tne - ix) * (iz - iz_tne) * gOut;
giz += bsw_val * (ix_tne - ix) * (iy - iy_tne) * gOut;
}
if (within_bounds_3d(iz_bse, iy_bse, ix_bse, inp_D, inp_H, inp_W)) {
scalar_t bse_val = inp_ptr_NC[iz_bse * inp_sD + iy_bse * inp_sH + ix_bse * inp_sW];
gix += bse_val * (iy - iy_tnw) * (iz - iz_tnw) * gOut;
giy += bse_val * (ix - ix_tnw) * (iz - iz_tnw) * gOut;
giz += bse_val * (ix - ix_tnw) * (iy - iy_tnw) * gOut;
}
}
// assuming grad_grid is contiguous
gGrid_ptr_NDHW[0] = gix_mult * gix;
gGrid_ptr_NDHW[1] = giy_mult * giy;
gGrid_ptr_NDHW[2] = giz_mult * giz;
} else if (interpolation_mode == GridSamplerInterpolation::Nearest) {
int64_t ix_nearest = static_cast<int64_t>(std::round(ix));
int64_t iy_nearest = static_cast<int64_t>(std::round(iy));
int64_t iz_nearest = static_cast<int64_t>(std::round(iz));
// assign nearest neighor pixel value to output pixel
scalar_t *gOut_ptr_NCDHW = gOut_ptr + n * gOut_sN + d * gOut_sD + h * gOut_sH + w * gOut_sW;
if (input_requires_grad) {
scalar_t *gInp_ptr_NC = gInp_ptr + n * gInp_sN;
for (int64_t c = 0; c < C; ++c, gOut_ptr_NCDHW += gOut_sC, gInp_ptr_NC += gInp_sC) {
// calculate and set grad_input
safe_add_3d(gInp_ptr_NC, iz_nearest, iy_nearest, ix_nearest,
gInp_sD, gInp_sH, gInp_sW, inp_D, inp_H, inp_W, *gOut_ptr_NCDHW);
}
}
}
}
}
}
}
});
return std::make_tuple(grad_input, grad_grid);
}
} // namespace
Tensor _grid_sampler_2d_cpu_quantized(
const Tensor& input,
const Tensor& grid,
int64_t interpolation_mode_,
int64_t padding_mode_,
bool align_corners) {
// See NOTE [ grid_sampler Native Functions ].
// Add checks here in case this is called instead of grid_sampler.
check_grid_sampler_common(input, grid);
check_grid_sampler_2d(input, grid);
auto interpolation_mode =
static_cast<GridSamplerInterpolation>(interpolation_mode_);
/* Bilinear interpolation is supported using the fact that we can perform
* linear interpolations on quantized values without rescaling. */
TORCH_CHECK(
interpolation_mode == GridSamplerInterpolation::Bilinear,
"_grid_sampler_2d_cpu_quantized(): only bilinear interpolation supported")
auto padding_mode = static_cast<GridSamplerPadding>(padding_mode_);
int64_t N = input.size(0);
int64_t C = input.size(1);
int64_t inp_H = input.size(2);
int64_t inp_W = input.size(3);
int64_t out_H = grid.size(1);
int64_t out_W = grid.size(2);
uint8_t zero_point = input.q_zero_point();
auto output = at::_empty_affine_quantized(
{N, C, out_H, out_W},
at::device(c10::kCPU).dtype(c10::kQUInt8),
input.q_scale(),
zero_point);
int64_t inp_sN = input.stride(0);
int64_t inp_sC = input.stride(1);
int64_t inp_sH = input.stride(2);
int64_t inp_sW = input.stride(3);
int64_t grid_sN = grid.stride(0);
int64_t grid_sH = grid.stride(1);
int64_t grid_sW = grid.stride(2);
int64_t grid_sCoor = grid.stride(3);
int64_t out_sN = output.stride(0);
int64_t out_sC = output.stride(1);
int64_t out_sH = output.stride(2);
int64_t out_sW = output.stride(3);
uint8_t* inp_ptr = (uint8_t*)input.data_ptr<quint8>();
uint8_t* out_ptr = (uint8_t*)output.data_ptr<quint8>();
float* grid_ptr = grid.data_ptr<float>();
at::parallel_for(0, N, 0, [&](int64_t start, int64_t end) {
for (const auto n : c10::irange(start, end)) {
float* grid_ptr_N = grid_ptr + n * grid_sN;
uint8_t* inp_ptr_N = inp_ptr + n * inp_sN;
for (const auto h : c10::irange(out_H)) {
for (const auto w : c10::irange(out_W)) {
// get the corresponding input x, y, z co-ordinates from grid
float* grid_ptr_NHW = grid_ptr_N + h * grid_sH + w * grid_sW;
float x = *grid_ptr_NHW;
float y = grid_ptr_NHW[grid_sCoor];
float ix = grid_sampler_compute_source_index(
x, inp_W, padding_mode, align_corners);
float iy = grid_sampler_compute_source_index(
y, inp_H, padding_mode, align_corners);
// get corner pixel values from (x, y)
// for 4d, we use north-east-south-west
int64_t ix_nw = static_cast<int64_t>(std::floor(ix));
int64_t iy_nw = static_cast<int64_t>(std::floor(iy));
int64_t ix_ne = ix_nw + 1;
int64_t iy_ne = iy_nw;
int64_t ix_sw = ix_nw;
int64_t iy_sw = iy_nw + 1;
int64_t ix_se = ix_nw + 1;
int64_t iy_se = iy_nw + 1;
// get surfaces to each neighbor:
float nw = (ix_se - ix) * (iy_se - iy);
float ne = (ix - ix_sw) * (iy_sw - iy);
float sw = (ix_ne - ix) * (iy - iy_ne);
float se = (ix - ix_nw) * (iy - iy_nw);
// calculate bilinear weighted pixel value and set output pixel
uint8_t* inp_ptr_NC = inp_ptr_N;
uint8_t* out_ptr_NCHW =
out_ptr + n * out_sN + h * out_sH + w * out_sW;
for (int64_t c = 0; c < C;
++c, out_ptr_NCHW += out_sC, inp_ptr_NC += inp_sC) {
float res = 0;
res += within_bounds_2d(iy_nw, ix_nw, inp_H, inp_W)
? inp_ptr_NC[iy_nw * inp_sH + ix_nw * inp_sW] * nw
: zero_point * nw;
res += within_bounds_2d(iy_ne, ix_ne, inp_H, inp_W)
? inp_ptr_NC[iy_ne * inp_sH + ix_ne * inp_sW] * ne
: zero_point * ne;
res += within_bounds_2d(iy_sw, ix_sw, inp_H, inp_W)
? inp_ptr_NC[iy_sw * inp_sH + ix_sw * inp_sW] * sw
: zero_point * sw;
res += within_bounds_2d(iy_se, ix_se, inp_H, inp_W)
? inp_ptr_NC[iy_se * inp_sH + ix_se * inp_sW] * se
: zero_point * se;
*out_ptr_NCHW = std::round(res);
}
}
}
}
});
return output;
}
Tensor _grid_sampler_2d_cpu_fallback(const Tensor& input, const Tensor& grid,
int64_t interpolation_mode_,
int64_t padding_mode_,
bool align_corners) {
// See NOTE [ grid_sampler Native Functions ].
// Add checks here in case this is called instead of grid_sampler.
check_grid_sampler_common(input, grid);
check_grid_sampler_2d(input, grid);
auto interpolation_mode = static_cast<GridSamplerInterpolation>(interpolation_mode_);
auto padding_mode = static_cast<GridSamplerPadding>(padding_mode_);
using scalar_t = float;
int64_t N = input.size(0);
int64_t C = input.size(1);
int64_t inp_H = input.size(2);
int64_t inp_W = input.size(3);
int64_t out_H = grid.size(1);
int64_t out_W = grid.size(2);
auto output = at::empty({N, C, out_H, out_W}, input.options());
int64_t inp_sN = input.stride(0);
int64_t inp_sC = input.stride(1);
int64_t inp_sH = input.stride(2);
int64_t inp_sW = input.stride(3);
int64_t grid_sN = grid.stride(0);
int64_t grid_sH = grid.stride(1);
int64_t grid_sW = grid.stride(2);
int64_t grid_sCoor = grid.stride(3);
int64_t out_sN = output.stride(0);
int64_t out_sC = output.stride(1);
int64_t out_sH = output.stride(2);
int64_t out_sW = output.stride(3);
scalar_t *inp_ptr = input.data_ptr<scalar_t>();
scalar_t *out_ptr = output.data_ptr<scalar_t>();
scalar_t *grid_ptr = grid.data_ptr<scalar_t>();
// loop over each output pixel
at::parallel_for(0, N, 0, [&](int64_t start, int64_t end) {
for (const auto n : c10::irange(start, end)) {
scalar_t *grid_ptr_N = grid_ptr + n * grid_sN;
scalar_t *inp_ptr_N = inp_ptr + n * inp_sN;
for (const auto h : c10::irange(out_H)) {
for (const auto w : c10::irange(out_W)) {
// get the corresponding input x, y, z co-ordinates from grid
scalar_t *grid_ptr_NHW = grid_ptr_N + h * grid_sH + w * grid_sW;
scalar_t x = *grid_ptr_NHW;
scalar_t y = grid_ptr_NHW[grid_sCoor];
scalar_t ix = grid_sampler_compute_source_index(x, inp_W, padding_mode, align_corners);
scalar_t iy = grid_sampler_compute_source_index(y, inp_H, padding_mode, align_corners);
if (interpolation_mode == GridSamplerInterpolation::Bilinear) {
// get corner pixel values from (x, y)
// for 4d, we use north-east-south-west
int64_t ix_nw = static_cast<int64_t>(std::floor(ix));
int64_t iy_nw = static_cast<int64_t>(std::floor(iy));
int64_t ix_ne = ix_nw + 1;
int64_t iy_ne = iy_nw;
int64_t ix_sw = ix_nw;
int64_t iy_sw = iy_nw + 1;
int64_t ix_se = ix_nw + 1;
int64_t iy_se = iy_nw + 1;
// get surfaces to each neighbor:
scalar_t nw = (ix_se - ix) * (iy_se - iy);
scalar_t ne = (ix - ix_sw) * (iy_sw - iy);
scalar_t sw = (ix_ne - ix) * (iy - iy_ne);
scalar_t se = (ix - ix_nw) * (iy - iy_nw);
// calculate bilinear weighted pixel value and set output pixel
scalar_t *inp_ptr_NC = inp_ptr_N;
scalar_t *out_ptr_NCHW = out_ptr + n * out_sN + h * out_sH + w * out_sW;
for (int64_t c = 0; c < C; ++c, out_ptr_NCHW += out_sC, inp_ptr_NC += inp_sC) {
auto res = static_cast<scalar_t>(0);
if (within_bounds_2d(iy_nw, ix_nw, inp_H, inp_W)) {
res += inp_ptr_NC[iy_nw * inp_sH + ix_nw * inp_sW] * nw;
}
if (within_bounds_2d(iy_ne, ix_ne, inp_H, inp_W)) {
res += inp_ptr_NC[iy_ne * inp_sH + ix_ne * inp_sW] * ne;
}
if (within_bounds_2d(iy_sw, ix_sw, inp_H, inp_W)) {
res += inp_ptr_NC[iy_sw * inp_sH + ix_sw * inp_sW] * sw;
}
if (within_bounds_2d(iy_se, ix_se, inp_H, inp_W)) {
res += inp_ptr_NC[iy_se * inp_sH + ix_se * inp_sW] * se;
}
*out_ptr_NCHW = res;
}
} else if (interpolation_mode == GridSamplerInterpolation::Nearest) {
int64_t ix_nearest = static_cast<int64_t>(std::nearbyint(ix));
int64_t iy_nearest = static_cast<int64_t>(std::nearbyint(iy));
// assign nearest neighor pixel value to output pixel
scalar_t *out_ptr_NCHW = out_ptr + n * out_sN + h * out_sH + w * out_sW;
scalar_t *inp_ptr_NC = inp_ptr_N;
for (int64_t c = 0; c < C; ++c, out_ptr_NCHW += out_sC, inp_ptr_NC += inp_sC) {
if (within_bounds_2d(iy_nearest, ix_nearest, inp_H, inp_W)) {
*out_ptr_NCHW = inp_ptr_NC[iy_nearest * inp_sH + ix_nearest * inp_sW];
} else {
*out_ptr_NCHW = static_cast<scalar_t>(0);
}
}
} else if (interpolation_mode == GridSamplerInterpolation::Bicubic) {
// grid_sampler_compute_source_index will "clip the value" of idx depends on the padding,
// which would cause calculation to be wrong,
// for example x = -0.1 -> ix = 0 for zero padding, but in bicubic ix = floor(x) = -1
// There would be more problem in reflection padding, since the -1 and +1 direction is not fixed in boundary condition
ix = grid_sampler_unnormalize(x, inp_W, align_corners);
iy = grid_sampler_unnormalize(y, inp_H, align_corners);
scalar_t ix_nw = std::floor(ix);
scalar_t iy_nw = std::floor(iy);
const scalar_t tx = ix - ix_nw;
const scalar_t ty = iy - iy_nw;
scalar_t *inp_ptr_NC = inp_ptr_N;
scalar_t *out_ptr_NCHW = out_ptr + n * out_sN + h * out_sH + w * out_sW;
for (int64_t c = 0; c < C; ++c, out_ptr_NCHW += out_sC, inp_ptr_NC += inp_sC) {
// NOLINTNEXTLINE(modernize-avoid-c-arrays,cppcoreguidelines-avoid-c-arrays)
scalar_t coefficients[4];
// Interpolate 4 values in the x directon
for (const auto i : c10::irange(4)) {
coefficients[i] = cubic_interp1d<scalar_t>(
get_value_bounded<scalar_t>(inp_ptr_NC, ix_nw - 1, iy_nw - 1 + i, inp_W, inp_H, inp_sW, inp_sH, padding_mode, align_corners),
get_value_bounded<scalar_t>(inp_ptr_NC, ix_nw + 0, iy_nw - 1 + i, inp_W, inp_H, inp_sW, inp_sH, padding_mode, align_corners),
get_value_bounded<scalar_t>(inp_ptr_NC, ix_nw + 1, iy_nw - 1 + i, inp_W, inp_H, inp_sW, inp_sH, padding_mode, align_corners),
get_value_bounded<scalar_t>(inp_ptr_NC, ix_nw + 2, iy_nw - 1 + i, inp_W, inp_H, inp_sW, inp_sH, padding_mode, align_corners),
tx);
}
// Interpolate in the y direction
*out_ptr_NCHW = cubic_interp1d<scalar_t>(
coefficients[0],
coefficients[1],
coefficients[2],
coefficients[3],
ty);
}
}
}
}
}
});
return output;
}
std::tuple<Tensor, Tensor>
_grid_sampler_2d_cpu_fallback_backward(const Tensor& grad_output,
const Tensor& input, const Tensor& grid,
int64_t interpolation_mode_,
int64_t padding_mode_,
bool align_corners) {
// See NOTE [ grid_sampler Native Functions ].
// Add checks here in case this is called instead of grid_sampler.
check_grid_sampler_common(input, grid);
check_grid_sampler_2d(input, grid);
const auto interpolation_mode = static_cast<GridSamplerInterpolation>(interpolation_mode_);
const auto padding_mode = static_cast<GridSamplerPadding>(padding_mode_);
using scalar_t = float;
auto grad_input = at::zeros_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
auto grad_grid = at::empty_like(grid, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
// If interpolation mode is Nearest, then grad_grid is not filled in the
// loop below.
if (interpolation_mode == GridSamplerInterpolation::Nearest) {
grad_grid.zero_();
}
int64_t N = input.size(0);
int64_t C = input.size(1);
int64_t inp_H = input.size(2);
int64_t inp_W = input.size(3);
int64_t out_H = grid.size(1);
int64_t out_W = grid.size(2);
int64_t inp_sN = input.stride(0);
int64_t inp_sC = input.stride(1);
int64_t inp_sH = input.stride(2);
int64_t inp_sW = input.stride(3);
int64_t grid_sN = grid.stride(0);
int64_t grid_sH = grid.stride(1);
int64_t grid_sW = grid.stride(2);
int64_t grid_sCoor = grid.stride(3);
int64_t gOut_sN = grad_output.stride(0);
int64_t gOut_sC = grad_output.stride(1);
int64_t gOut_sH = grad_output.stride(2);
int64_t gOut_sW = grad_output.stride(3);
int64_t gInp_sN = grad_input.stride(0);
int64_t gInp_sC = grad_input.stride(1);
int64_t gInp_sH = grad_input.stride(2);
int64_t gInp_sW = grad_input.stride(3);
int64_t gGrid_sN = grad_grid.stride(0);
int64_t gGrid_sW = grad_grid.stride(2);
scalar_t *inp_ptr = input.data_ptr<scalar_t>();
scalar_t *grid_ptr = grid.data_ptr<scalar_t>();
scalar_t *gOut_ptr = grad_output.data_ptr<scalar_t>();
scalar_t *gInp_ptr = grad_input.data_ptr<scalar_t>();
scalar_t *gGrid_ptr = grad_grid.data_ptr<scalar_t>();
// loop over each output pixel
at::parallel_for(0, N, 0, [&](int64_t start, int64_t end) {
for (const auto n : c10::irange(start, end)) {
scalar_t *grid_ptr_N = grid_ptr + n * grid_sN;
scalar_t *inp_ptr_N = inp_ptr + n * inp_sN;
scalar_t *gGrid_ptr_NHW = gGrid_ptr + n * gGrid_sN;
for (const auto h : c10::irange(out_H)) {
for (int64_t w = 0; w < out_W; ++w, gGrid_ptr_NHW += gGrid_sW /* grad_grid is contiguous */ ) {
// get the corresponding input x, y co-ordinates from grid
scalar_t *grid_ptr_NHW = grid_ptr_N + h * grid_sH + w * grid_sW;
scalar_t x = *grid_ptr_NHW;
scalar_t y = grid_ptr_NHW[grid_sCoor];
// multipliers for gradients on ix, iy
// NOLINTNEXTLINE(cppcoreguidelines-init-variables)
scalar_t gix_mult, giy_mult;
scalar_t ix = grid_sampler_compute_source_index_set_grad(x, inp_W, padding_mode, align_corners, &gix_mult);
scalar_t iy = grid_sampler_compute_source_index_set_grad(y, inp_H, padding_mode, align_corners, &giy_mult);
if (interpolation_mode == GridSamplerInterpolation::Bilinear) {
// get corner pixel values from (x, y)
// for 4d, we use north-east-south-west
int64_t ix_nw = static_cast<int64_t>(std::floor(ix));
int64_t iy_nw = static_cast<int64_t>(std::floor(iy));
int64_t ix_ne = ix_nw + 1;
int64_t iy_ne = iy_nw;
int64_t ix_sw = ix_nw;
int64_t iy_sw = iy_nw + 1;
int64_t ix_se = ix_nw + 1;
int64_t iy_se = iy_nw + 1;
// get surfaces to each neighbor:
scalar_t nw = (ix_se - ix) * (iy_se - iy);
scalar_t ne = (ix - ix_sw) * (iy_sw - iy);
scalar_t sw = (ix_ne - ix) * (iy - iy_ne);
scalar_t se = (ix - ix_nw) * (iy - iy_nw);
scalar_t gix = static_cast<scalar_t>(0), giy = static_cast<scalar_t>(0);
scalar_t *gOut_ptr_NCHW = gOut_ptr + n * gOut_sN + h * gOut_sH + w * gOut_sW;
scalar_t *gInp_ptr_NC = gInp_ptr + n * gInp_sN;
scalar_t *inp_ptr_NC = inp_ptr_N;
// calculate bilinear weighted pixel value and set output pixel
for (int64_t c = 0; c < C; ++c, gOut_ptr_NCHW += gOut_sC, gInp_ptr_NC += gInp_sC, inp_ptr_NC += inp_sC) {
scalar_t gOut = *gOut_ptr_NCHW;
// calculate and set grad_input
safe_add_2d(gInp_ptr_NC, iy_nw, ix_nw, gInp_sH, gInp_sW, inp_H, inp_W, nw * gOut);
safe_add_2d(gInp_ptr_NC, iy_ne, ix_ne, gInp_sH, gInp_sW, inp_H, inp_W, ne * gOut);
safe_add_2d(gInp_ptr_NC, iy_sw, ix_sw, gInp_sH, gInp_sW, inp_H, inp_W, sw * gOut);
safe_add_2d(gInp_ptr_NC, iy_se, ix_se, gInp_sH, gInp_sW, inp_H, inp_W, se * gOut);
// calculate grad_grid
if (within_bounds_2d(iy_nw, ix_nw, inp_H, inp_W)) {
scalar_t nw_val = inp_ptr_NC[iy_nw * inp_sH + ix_nw * inp_sW];
gix -= nw_val * (iy_se - iy) * gOut;
giy -= nw_val * (ix_se - ix) * gOut;
}
if (within_bounds_2d(iy_ne, ix_ne, inp_H, inp_W)) {
scalar_t ne_val = inp_ptr_NC[iy_ne * inp_sH + ix_ne * inp_sW];
gix += ne_val * (iy_sw - iy) * gOut;
giy -= ne_val * (ix - ix_sw) * gOut;
}
if (within_bounds_2d(iy_sw, ix_sw, inp_H, inp_W)) {
scalar_t sw_val = inp_ptr_NC[iy_sw * inp_sH + ix_sw * inp_sW];
gix -= sw_val * (iy - iy_ne) * gOut;
giy += sw_val * (ix_ne - ix) * gOut;
}
if (within_bounds_2d(iy_se, ix_se, inp_H, inp_W)) {
scalar_t se_val = inp_ptr_NC[iy_se * inp_sH + ix_se * inp_sW];
gix += se_val * (iy - iy_nw) * gOut;
giy += se_val * (ix - ix_nw) * gOut;
}
}
// assuming grad_grid is contiguous
gGrid_ptr_NHW[0] = gix_mult * gix;
gGrid_ptr_NHW[1] = giy_mult * giy;
} else if (interpolation_mode == GridSamplerInterpolation::Nearest) {
int64_t ix_nearest = static_cast<int64_t>(std::nearbyint(ix));
int64_t iy_nearest = static_cast<int64_t>(std::nearbyint(iy));
// assign nearest neighor pixel value to output pixel
scalar_t *gOut_ptr_NCHW = gOut_ptr + n * gOut_sN + h * gOut_sH + w * gOut_sW;
scalar_t *gInp_ptr_NC = gInp_ptr + n * gInp_sN;
for (int64_t c = 0; c < C; ++c, gOut_ptr_NCHW += gOut_sC, gInp_ptr_NC += gInp_sC) {
// calculate and set grad_input
safe_add_2d(gInp_ptr_NC, iy_nearest, ix_nearest, gInp_sH, gInp_sW,
inp_H, inp_W, *gOut_ptr_NCHW);
}
} else if (interpolation_mode == GridSamplerInterpolation::Bicubic) {
ix = grid_sampler_unnormalize_set_grad(x, inp_W, align_corners, &gix_mult);
iy = grid_sampler_unnormalize_set_grad(y, inp_H, align_corners, &giy_mult);
scalar_t ix_nw = std::floor(ix);
scalar_t iy_nw = std::floor(iy);
const scalar_t tx = ix - ix_nw;
const scalar_t ty = iy - iy_nw;
// NOLINTNEXTLINE(modernize-avoid-c-arrays,cppcoreguidelines-avoid-c-arrays)
scalar_t x_coeffs[4];
// NOLINTNEXTLINE(modernize-avoid-c-arrays,cppcoreguidelines-avoid-c-arrays)
scalar_t y_coeffs[4];
// NOLINTNEXTLINE(modernize-avoid-c-arrays,cppcoreguidelines-avoid-c-arrays)
scalar_t x_coeffs_grad[4];
// NOLINTNEXTLINE(modernize-avoid-c-arrays,cppcoreguidelines-avoid-c-arrays)
scalar_t y_coeffs_grad[4];
get_cubic_upsample_coefficients<scalar_t>(x_coeffs, tx);
get_cubic_upsample_coefficients<scalar_t>(y_coeffs, ty);
get_cubic_coefficients_grad<scalar_t>(x_coeffs_grad, tx);
get_cubic_coefficients_grad<scalar_t>(y_coeffs_grad, ty);
scalar_t gix = static_cast<scalar_t>(0);
scalar_t giy = static_cast<scalar_t>(0);
scalar_t *gOut_ptr_NCHW = gOut_ptr + n * gOut_sN + h * gOut_sH + w * gOut_sW;
scalar_t *gInp_ptr_NC = gInp_ptr + n * gInp_sN;
scalar_t *inp_ptr_NC = inp_ptr_N;
for (int64_t c = 0; c < C; ++c, gOut_ptr_NCHW += gOut_sC, gInp_ptr_NC += gInp_sC, inp_ptr_NC+= inp_sC) {
scalar_t gOut = *gOut_ptr_NCHW;
for (const auto i : c10::irange(4)) {
for (const auto j : c10::irange(4)) {
// set input gradient
add_value_bounded<scalar_t>(gInp_ptr_NC, ix_nw - 1 + i, iy_nw - 1 + j,
inp_W, inp_H, gInp_sW, gInp_sH, gOut * x_coeffs[i] * y_coeffs[j], padding_mode, align_corners);
// set grid gradient
scalar_t val = get_value_bounded<scalar_t>(inp_ptr_NC, ix_nw - 1 + i, iy_nw - 1 + j,
inp_W, inp_H, inp_sW, inp_sH, padding_mode, align_corners);
gix -= val * x_coeffs_grad[i] * y_coeffs[j] * gOut;
giy -= val * y_coeffs_grad[j] * x_coeffs[i] * gOut;
}
}
}
gGrid_ptr_NHW[0] = gix_mult * gix;
gGrid_ptr_NHW[1] = giy_mult * giy;
}
}
}
}
});
return std::make_tuple(grad_input, grad_grid);
}
Tensor grid_sampler_2d_cpu(const Tensor& input, const Tensor& grid,
int64_t interpolation_mode, int64_t padding_mode,
bool align_corners) {
// See NOTE [ grid_sampler Native Functions ].
// Add checks here in case this is called instead of grid_sampler.
check_grid_sampler_common(input, grid);
check_grid_sampler_2d(input, grid);
if (input.scalar_type() == kQUInt8) {
return native::_grid_sampler_2d_cpu_quantized(
input, grid, interpolation_mode, padding_mode, align_corners);
}
// AVX gather instructions use signed 32-bit offsets to gather float values.
// Check for possible overflow and fallback to scalar implementation
if (input.scalar_type() != kDouble) {
TORCH_CHECK(input.scalar_type() == kFloat,
"grid_sampler_2d_cpu not implemented for ", input.scalar_type());
auto sizes = input.sizes();
auto strides = input.strides();
const auto grid_sW = grid.strides()[2];
// NOTE: Gather offsets are only used for the input H, W dimensions
// or only for strided access to the grid tensor
auto max_gather_offset = std::max(
(sizes[2] - 1) * strides[2] + (sizes[3] - 1) * strides[3],
grid_sW * (vec::Vectorized<float>::size() - 1));
if (max_gather_offset > std::numeric_limits<int32_t>::max()) {
return native::_grid_sampler_2d_cpu_fallback(
input, grid, interpolation_mode, padding_mode, align_corners);
}
}
auto in_size = input.sizes();
auto grid_size = grid.sizes();
auto output = at::empty(
{in_size[0], in_size[1], grid_size[1], grid_size[2]}, input.options());
grid_sampler_2d_cpu_kernel(
kCPU, output, input, grid, interpolation_mode, padding_mode, align_corners);
return output;
}
DEFINE_DISPATCH(grid_sampler_2d_cpu_kernel);
Tensor grid_sampler_3d_cpu(const Tensor& input, const Tensor& grid,
int64_t interpolation_mode, int64_t padding_mode,
bool align_corners) {
// See NOTE [ grid_sampler Native Functions ].
// Add checks here in case this is called instead of grid_sampler.
check_grid_sampler_common(input, grid);
check_grid_sampler_3d(input, grid, interpolation_mode);
return AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "grid_sampler3d_cpu", [&] {
return grid_sampler_3d_cpu_impl<scalar_t>(
input, grid, static_cast<GridSamplerInterpolation>(interpolation_mode),
static_cast<GridSamplerPadding>(padding_mode), align_corners);
});
}
std::tuple<Tensor, Tensor>
grid_sampler_2d_backward_cpu(const Tensor& grad_output, const Tensor& input, const Tensor& grid,
int64_t interpolation_mode, int64_t padding_mode, bool align_corners,
std::array<bool,2> output_mask) {
// See NOTE [ grid_sampler Native Functions ].
// Add checks here in case this is called instead of grid_sampler.
check_grid_sampler_common(input, grid);
check_grid_sampler_2d(input, grid);
// AVX gather instructions use signed 32-bit offsets to gather float values.
// Check for possible overflow and fallback to scalar implementation
if (input.scalar_type() != kDouble) {
TORCH_CHECK(input.scalar_type() == kFloat,
"grid_sampler_2d_backward_cpu not implemented for ", input.scalar_type());
auto isizes = input.sizes();
auto istrides = input.strides();
auto gsizes = grad_output.sizes();
auto gstrides = grad_output.strides();
const auto grid_sW = grid.strides()[2];
// NOTE: Gather offsets are only used for the height and width dimensions
auto max_gather_offset = std::max(
std::max(
(isizes[2] - 1) * istrides[2] + (isizes[3] - 1) * istrides[3],
(gsizes[2] - 1) * gstrides[2] + (gsizes[3] - 1) * gstrides[3]),
grid_sW * (vec::Vectorized<float>::size() - 1));
if (max_gather_offset > std::numeric_limits<int32_t>::max()) {
return native::_grid_sampler_2d_cpu_fallback_backward(
grad_output, input, grid, interpolation_mode, padding_mode, align_corners);
}
}
auto input_requires_grad = output_mask[0];
Tensor grad_input = ([&]() {
if (input_requires_grad) {
return at::zeros_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
} else {