forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
LinearAlgebra.cpp
3137 lines (2741 loc) · 114 KB
/
LinearAlgebra.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
#include <ATen/ATen.h>
#include <ATen/core/grad_mode.h>
#include <ATen/Dispatch.h>
#include <ATen/ExpandUtils.h>
#include <ATen/NamedTensorUtils.h>
#include <ATen/OpMathType.h>
#include <ATen/native/mkldnn/Matmul.h>
#include <ATen/native/CPUBlas.h>
#include <ATen/native/IndexingUtils.h>
#include <ATen/native/LinearAlgebra.h>
#include <ATen/native/LinearAlgebraUtils.h>
#include <ATen/native/ReduceOps.h>
#include <ATen/native/ReduceOpsUtils.h>
#include <ATen/native/Resize.h>
#include <ATen/native/TensorIterator.h>
#include <ATen/NativeFunctions.h>
#include <ATen/Parallel.h>
#include <ATen/TensorUtils.h>
#include <ATen/Utils.h>
#include <c10/util/accumulate.h>
#include <c10/util/irange.h>
#include <c10/util/variant.h>
#include <functional>
#include <limits>
#include <numeric>
#include <string>
#include <tuple>
namespace at {
namespace meta {
TORCH_META_FUNC(addmm)(const Tensor& self, const Tensor& mat1, const Tensor& mat2, const Scalar& beta, const Scalar& alpha) {
TORCH_CHECK(mat1.dim() == 2, "mat1 must be a matrix, got ", mat1.dim(), "-D tensor");
TORCH_CHECK(mat2.dim() == 2, "mat2 must be a matrix, got ", mat2.dim(), "-D tensor");
TORCH_CHECK(
mat1.sizes()[1] == mat2.sizes()[0], "mat1 and mat2 shapes cannot be multiplied (",
mat1.sizes()[0], "x", mat1.sizes()[1], " and ", mat2.sizes()[0], "x", mat2.sizes()[1], ")");
auto names = at::namedinference::propagate_names_for_addmm(mat1, mat2, self);
set_output(0, {mat1.sizes()[0], mat2.sizes()[1]}, {}, self.options(), names);
}
TORCH_META_FUNC(mm)(const Tensor & self, const Tensor & mat2) {
TORCH_CHECK(self.dim() == 2, "self must be a matrix");
TORCH_CHECK(mat2.dim() == 2, "mat2 must be a matrix");
TORCH_CHECK(
self.sizes()[1] == mat2.sizes()[0], "mat1 and mat2 shapes cannot be multiplied (",
self.sizes()[0], "x", self.sizes()[1], " and ", mat2.sizes()[0], "x", mat2.sizes()[1], ")");
auto names = at::namedinference::compute_matmul_outnames(self, mat2);
set_output(0, {self.sizes()[0], mat2.sizes()[1]}, {}, self.options(), names);
}
template <typename Meta>
void common_checks_baddbmm_bmm(Meta& meta, const Tensor& batch1, const Tensor& batch2, const Scalar& beta, const Scalar& alpha, bool is_bmm, const c10::optional<Tensor>& self_baddbmm = nullopt) {
TORCH_CHECK(batch1.dim() == 3, "batch1 must be a 3D tensor");
TORCH_CHECK(batch2.dim() == 3, "batch2 must be a 3D tensor");
const auto batch1_sizes = batch1.sizes();
const auto batch2_sizes = batch2.sizes();
int64_t bs = batch1_sizes[0];
int64_t contraction_size = batch1_sizes[2];
int64_t res_rows = batch1_sizes[1];
int64_t res_cols = batch2_sizes[2];
std::vector<int64_t> output_size {bs, res_rows, res_cols};
TORCH_CHECK(batch2_sizes[0] == bs && batch2_sizes[1] == contraction_size,
"Expected size for first two dimensions of batch2 tensor to be: [",
bs, ", ", contraction_size, "] but got: [", batch2_sizes[0], ", ", batch2_sizes[1], "].");
auto& result = meta.maybe_get_output(0);
// 'set_output' does not resize for in-place calls
meta.set_output(output_size, batch2.options());
const auto result_sizes = result.sizes();
// Error is raised if called from in-place overload with incorrect shape
TORCH_CHECK(result_sizes == output_size,
"Expected an output tensor with shape [", output_size, "] but got shape ", result_sizes);
std::vector<Dimname> outnames = {};
if (!is_bmm) {
if (self_baddbmm.has_value()) {
const auto& self = self_baddbmm.value();
if (beta.toComplexDouble() != 0.0) result.copy_(self);
TORCH_CHECK(self.dim() == 3, "self must be a 3D tensor");
const auto self_sizes = self.sizes();
TORCH_CHECK(self_sizes == output_size,
"Expected an input tensor shape with shape ", output_size, " but got shape: ", self_sizes);
outnames = namedinference::compute_baddbmm_outnames(result, batch1, batch2, self);
}
} else {
outnames = namedinference::compute_bmm_outnames(result, batch1, batch2);
}
namedinference::propagate_names_if_nonempty(
result,
outnames
);
}
TORCH_META_FUNC(bmm)(const Tensor& self, const Tensor& mat2) {
common_checks_baddbmm_bmm(*this, self, mat2, Scalar(0.0), Scalar(1.0), true);
}
TORCH_META_FUNC(baddbmm)(const Tensor& self, const Tensor& batch1, const Tensor& batch2, const Scalar& beta, const Scalar& alpha) {
auto self_ = expand_size(self, {batch1.size(0), batch1.size(1), batch2.size(2)}, "baddbmm");
common_checks_baddbmm_bmm(*this, batch1, batch2, beta, alpha, false, *self_);
}
} // namespace meta
namespace native {
DEFINE_DISPATCH(addr_stub);
DEFINE_DISPATCH(linalg_vector_norm_stub);
// As P is a permutation matrix
// det(P) = 1 if it's an even permutation and det(P) = -1 if it's an odd permutation
static inline Tensor _lu_det_P(const Tensor& lu, const Tensor& pivs) {
const auto n = lu.size(-1);
auto det_P = (at::arange(1, n + 1, pivs.options()) != pivs)
.sum(-1, /*keepdim=*/false, /*dtype=*/at::kLong)
.fmod_(2)
// take 0 to 1 and 1 to -1
.mul_(-2)
.add_(1);
return det_P;
}
// Given a pivoted LU factorization A = P L U,
// det(A) = det(P) * det(L) * det(U).
std::tuple<Tensor, Tensor, Tensor> _det_lu_based_helper(const Tensor& self) {
Tensor pivs, lu;
std::tie(lu, pivs, std::ignore) = at::linalg_lu_factor_ex(self);
const auto det_P = _lu_det_P(lu, pivs);
auto det = det_P * at::prod(lu.diagonal(0, -2 ,-1), /*dim=*/-1);
return std::make_tuple(std::move(det), std::move(lu), std::move(pivs));
}
// torch.det, alias for torch.linalg.det
Tensor det(const Tensor& self) {
return at::linalg_det(self);
}
Tensor linalg_det(const Tensor& self) {
squareCheckInputs(self, "linalg.det");
checkFloatingOrComplex(self, "linalg.det");
return std::get<0>(at::_det_lu_based_helper(self));
}
Tensor& linalg_det_out(const Tensor& self, Tensor& out) {
checkSameDevice("torch.linalg.det", out, self, "out");
checkLinalgCompatibleDtype("torch.linalg.det", out, self, "out");
IntArrayRef out_sizes(self.sizes().data(), self.dim() - 2);
at::native::resize_output(out, out_sizes);
auto det = at::native::linalg_det(self);
out.copy_(det);
return out;
}
Tensor logdet(const Tensor& self) {
squareCheckInputs(self, "logdet");
checkFloatingOrComplex(self, "logdet");
Tensor pivs, lu;
std::tie(lu, pivs, std::ignore) = at::linalg_lu_factor_ex(self);
const auto det_P = _lu_det_P(lu, pivs);
const auto diag_U = lu.diagonal(0, -2 ,-1);
const auto det_sign = diag_U.sign().prod(-1).mul_(det_P);
// If det_sign > 0, diag_U.abs_().log_().sum(-1) gives logdet (this means U is not singular).
// If det_sign <= 0, then we get proper nan (when det < 0, i.e., det_sign) or -inf (when det = 0, i.e., U is singular).
// U is singular when U(i, i) = 0 for some i in [1, self.size(-1)].
Tensor logdet_vals = diag_U.abs_().log_().sum(-1);
if (self.dim() > 2) {
auto indices = toListOfOptionalTensors((det_sign < 0).nonzero_numpy());
// NOLINTNEXTLINE(performance-move-const-arg)
logdet_vals.index_put_(std::move(indices), at::full({}, NAN, self.options()));
} else if (det_sign.item<double>() < 0) {
logdet_vals.fill_(NAN);
}
return logdet_vals;
}
std::tuple<Tensor, Tensor> linalg_slogdet(const Tensor& self) {
squareCheckInputs(self, "linalg.slogdet");
ScalarType t = self.scalar_type();
TORCH_CHECK(t == ScalarType::Double || t == ScalarType::Float || t == ScalarType::ComplexFloat || t == ScalarType::ComplexDouble,
"linalg.slogdet: expected a tensor of float, double, cfloat or cdouble types but got ", t);
Tensor pivs, lu;
std::tie(lu, pivs, std::ignore) = at::linalg_lu_factor_ex(self);
const auto det_P = _lu_det_P(lu, pivs);
const auto diag_U = lu.diagonal(0, -2 ,-1);
const auto det_sign = diag_U.sgn().prod(-1).mul_(det_P);
// abslogdet_val is -inf if U is singular, in which case diag_U.abs_().log_().sum(-1) will return -inf.
// U is singular when U(i, i) = 0 for some i in [1, self.size(-1)].
// Since abslogdet_val cannot take nan, no special case handling is required.
// in-place abs is not supported for complex tensors
auto abslogdet_val = isComplexType(t) ? diag_U.abs().log_().sum(-1) : diag_U.abs_().log_().sum(-1);
return std::make_tuple(det_sign, abslogdet_val);
}
// TODO: implement _out variant avoiding copy and using already allocated storage directly
std::tuple<Tensor&, Tensor&> linalg_slogdet_out(const Tensor& input, Tensor& sign, Tensor& logabsdet) {
checkSameDevice("linalg.slogdet", sign, input, "sign");
checkSameDevice("linalg.slogdet", logabsdet, input, "logabsdet");
checkLinalgCompatibleDtype("linalg.slogdet", sign, input, "sign");
ScalarType real_dtype = toRealValueType(input.scalar_type());
// logabsdet is always real-valued here
checkLinalgCompatibleDtype("linalg.slogdet", logabsdet.scalar_type(), real_dtype, "logabsdet");
Tensor sign_tmp, logabsdet_tmp;
std::tie(sign_tmp, logabsdet_tmp) = at::linalg_slogdet(input);
at::native::resize_output(sign, sign_tmp.sizes());
sign.copy_(sign_tmp);
at::native::resize_output(logabsdet, logabsdet_tmp.sizes());
logabsdet.copy_(logabsdet_tmp);
return std::tuple<Tensor&, Tensor&>(sign, logabsdet);
}
std::tuple<Tensor, Tensor> slogdet(const Tensor& self) {
return at::linalg_slogdet(self);
}
namespace {
// This function extracts the optional Tensors for atol and rtol
// Default value for atol is zero
// Default value for rtol is eps*max(rows, cols)
// If atol is specified and rtol is not specified then default value for rtol is zero
// It is used for matrix_rank and pinv
std::tuple<Tensor, Tensor> get_atol_rtol(
const Tensor& input,
const optional<Tensor>& atol_opt,
const optional<Tensor>& rtol_opt,
const c10::string_view function_name) {
auto options = input.options().dtype(ScalarType::Double);
auto atol = atol_opt.has_value() ? atol_opt.value() : at::zeros({}, options);
checkNotComplexTolerance(atol, function_name, "atol");
Tensor rtol;
if (rtol_opt.has_value()) {
rtol = rtol_opt.value();
checkNotComplexTolerance(rtol, function_name, "rtol");
} else {
ScalarType real_dtype = toRealValueType(input.scalar_type());
auto default_rtol = at::full({}, _get_epsilon(real_dtype) * std::max(input.size(-1), input.size(-2)), options);
rtol = atol_opt.has_value()
? at::where(atol_opt.value() > 0, at::zeros({}, options), default_rtol)
: default_rtol;
}
return std::make_tuple(atol, rtol);
}
std::tuple<Tensor, Tensor> get_atol_rtol(
const Tensor& input,
optional<double> atol_opt,
optional<double> rtol_opt) {
double atol = atol_opt.has_value() ? atol_opt.value() : 0.0;
double rtol;
if (rtol_opt.has_value()) {
rtol = rtol_opt.value();
} else {
ScalarType real_dtype = toRealValueType(input.scalar_type());
auto default_rtol = _get_epsilon(real_dtype) * std::max(input.size(-1), input.size(-2));
rtol = (atol_opt.has_value() && atol_opt.value() > 0.0)
? 0.0
: default_rtol;
}
auto options = input.options().dtype(ScalarType::Double);
auto atol_tensor = at::full({}, atol, options);
auto rtol_tensor = at::full({}, rtol, options);
return std::make_tuple(atol_tensor, rtol_tensor);
}
} // anonymous namespace
Tensor linalg_pinv(
const Tensor& input,
const optional<Tensor>& atol_opt,
const optional<Tensor>& rtol_opt,
bool hermitian) {
// FIXME: Whenever we have a nice lstsq, we should dispatch this function to simply be
// `torch.lstsq(A, torch.eye(A.shape[-1]), atol=atol, rtol=rtol)`
// with a driver that supports singular inputs
NoTF32Guard disable_tf32;
ScalarType t = input.scalar_type();
TORCH_CHECK((t == ScalarType::Double || t == ScalarType::Float || t == ScalarType::ComplexFloat || t == ScalarType::ComplexDouble)
&& input.dim() >= 2,
"linalg.pinv(", t, "{", input.sizes(), "}): expected a tensor with 2 or more dimensions "
"of float, double, cfloat or cdouble types");
Tensor atol, rtol;
std::tie(atol, rtol) = get_atol_rtol(input, atol_opt, rtol_opt, "torch.linalg.pinv");
if (input.numel() == 0) {
// The implementation below uses operations that do not work for zero numel tensors
// therefore we need this early return for 'input.numel() == 0' case
Tensor U, S, V;
// TODO: replace input.svd with linalg_svd when torch/xla can work with at::linalg_svd
std::tie(U, S, V) = input.svd();
return at::matmul(V * S.reciprocal().unsqueeze(-2), U.mH());
}
// If not Hermitian use singular value decomposition, else use eigenvalue decomposition
if (!hermitian) {
Tensor U, S, V;
// TODO: replace input.svd with linalg_svd
// using linalg_svd breaks pytorch/xla, see https://github.com/pytorch/xla/issues/2755
std::tie(U, S, V) = input.svd();
Tensor max_val = at::narrow(S, /*dim=*/-1, /*start=*/0, /*length=*/1); // singular values are sorted in descending order
Tensor tol = at::max(atol.unsqueeze(-1), rtol.unsqueeze(-1) * max_val);
Tensor S_pseudoinv = at::where(S > tol, S.reciprocal(), at::zeros({}, S.options())).to(input.dtype());
// computes V @ diag(S_pseudoinv) @ U.conj().T
return at::matmul(V * S_pseudoinv.unsqueeze(-2), U.mH());
} else {
Tensor S, U;
std::tie(S, U) = at::linalg_eigh(input);
// For Hermitian matrices, singular values equal to abs(eigenvalues)
Tensor S_abs = S.abs();
// eigenvalues are sorted in ascending order starting with negative values, we need a maximum value of abs(eigenvalues)
Tensor max_val = S_abs.amax(/*dim=*/-1, /*keepdim=*/true);
Tensor tol = at::max(atol.unsqueeze(-1), rtol.unsqueeze(-1) * max_val);
Tensor S_pseudoinv = at::where(S_abs > tol, S.reciprocal(), at::zeros({}, S.options())).to(input.dtype());
// computes U @ diag(S_pseudoinv) @ U.conj().T
return at::matmul(U * S_pseudoinv.unsqueeze(-2), U.mH());
}
}
Tensor linalg_pinv(const Tensor& input, optional<double> atol, optional<double> rtol, bool hermitian) {
Tensor atol_tensor, rtol_tensor;
std::tie(atol_tensor, rtol_tensor) = get_atol_rtol(input, atol, rtol);
return at::linalg_pinv(input, atol_tensor, rtol_tensor, hermitian);
}
Tensor linalg_pinv(const Tensor& input, const Tensor& rcond, bool hermitian) {
// For NumPy compatibility the rcond argument is used as relative tolerance
checkNotComplexTolerance(rcond, "torch.linalg.pinv", "rcond");
auto options = input.options().dtype(ScalarType::Double);
return at::linalg_pinv(input, at::zeros({}, options), rcond, hermitian);
}
Tensor linalg_pinv(const Tensor& input, double rcond, bool hermitian) {
// For NumPy compatibility the rcond argument is used as relative tolerance
return at::linalg_pinv(input, 0.0, rcond, hermitian);
}
// TODO: implement _out variant avoiding copy and using already allocated storage directly
Tensor& linalg_pinv_out(
const Tensor& input,
const optional<Tensor>& atol,
const optional<Tensor>& rtol,
bool hermitian,
Tensor& result) {
checkSameDevice("linalg.pinv", result, input);
checkLinalgCompatibleDtype("linalg.pinv", result, input);
Tensor result_tmp = at::linalg_pinv(input, atol, rtol, hermitian);
at::native::resize_output(result, result_tmp.sizes());
result.copy_(result_tmp);
return result;
}
Tensor& linalg_pinv_out(
const Tensor& input,
optional<double> atol,
optional<double> rtol,
bool hermitian,
Tensor& result) {
checkSameDevice("linalg.pinv", result, input);
checkLinalgCompatibleDtype("linalg.pinv", result, input);
Tensor result_tmp = at::linalg_pinv(input, atol, rtol, hermitian);
at::native::resize_output(result, result_tmp.sizes());
result.copy_(result_tmp);
return result;
}
Tensor& linalg_pinv_out(const Tensor& input, const Tensor& rcond, bool hermitian, Tensor& result) {
checkSameDevice("linalg.pinv", result, input);
checkLinalgCompatibleDtype("linalg.pinv", result, input);
Tensor result_tmp = at::linalg_pinv(input, rcond, hermitian);
at::native::resize_output(result, result_tmp.sizes());
result.copy_(result_tmp);
return result;
}
Tensor& linalg_pinv_out(const Tensor& input, double rcond, bool hermitian, Tensor& result) {
Tensor rcond_tensor = at::full({}, rcond, input.options().dtype(ScalarType::Double));
return at::linalg_pinv_out(result, input, rcond_tensor, hermitian);
}
Tensor pinverse(const Tensor& self, double rcond) {
return at::linalg_pinv(self, rcond, /*hermitian=*/false);
}
// matrix_power implementation
namespace {
/**
* @brief Raises the input matrix to the given power n
*
* If the exponent n is negative, the inverse of the input
* matrix will be raised to power abs(n).
*
* @param self (batched) square matrix to raise to power n
* @param n exponent to raise matrix (or matrices in batch) to
* @param _out optional tensor to write the output to
* @return Tensor input matrix raised to power n
*/
Tensor linalg_matrix_power_impl(
const Tensor& self,
int64_t n,
c10::optional<Tensor> _out) {
NoTF32Guard disable_tf32;
auto out = _out.value_or(Tensor());
squareCheckInputs(self, "linalg.matrix_power");
if (_out.has_value()) {
checkSameDevice("matrix_power", out, self);
checkLinalgCompatibleDtype("matrix_power", out, self);
at::native::resize_output(out, self.sizes());
}
// For n=0 we return the identity matrix of the same shape as input.
if (n == 0) {
if (!_out.has_value()) {
// Clone input to include result in the autograd graph
out = self.clone(at::MemoryFormat::Contiguous);
}
return out.copy_(at::eye(self.size(-2), self.options()));
}
if (n == 1) {
return _out.has_value() ? out.copy_(self)
: self.clone(at::MemoryFormat::Contiguous);
}
if (n == -1) {
return _out.has_value() ? at::linalg_inv_out(out, self)
: at::linalg_inv(self);
}
// For negative n we inverte the input matrix before raising to power abs(n)
auto a = n < 0 ? at::linalg_inv(self) : self;
n = std::abs(n);
// Fast paths for small powers
if (n == 2) {
return _out.has_value() ? at::matmul_out(out, a, a) : at::matmul(a, a);
}
if (n == 3) {
return _out.has_value() ? at::matmul_out(out, at::matmul(a, a), a)
: at::matmul(at::matmul(a, a), a);
}
// This is a binary decomposition of n.
// Moving from the least significant bit to the most significant bit
// This is done to reduce the number of matrix multiplications
// by raising the input matrix in powers of 2
// The total number of matrix multiplications are
// number of bits + number of bits that equal 1 ~ O(log n)
// instead of O(n)
Tensor z, result;
while (n > 0) {
const auto bit = n % 2;
n = n / 2;
z = z.defined() ? at::matmul(z, z) : a;
if (bit == 1) {
if (_out.has_value() && n <= 0) {
// Last multiplication can use the out version
return result.defined() ? at::matmul_out(out, result, z) : out.copy_(z);
}
result = result.defined() ? at::matmul(result, z) : z;
}
}
return result;
}
} // namespace
Tensor& linalg_matrix_power_out(const Tensor& self, int64_t n, Tensor& result) {
linalg_matrix_power_impl(self, n, result);
return result;
}
Tensor linalg_matrix_power(const Tensor& self, int64_t n) {
return linalg_matrix_power_impl(self, n, c10::nullopt);
}
Tensor& matrix_power_out(const Tensor& self, int64_t n, Tensor& result) {
return at::native::linalg_matrix_power_out(self, n, result);
}
Tensor matrix_power(const Tensor& self, int64_t n) {
return at::native::linalg_matrix_power(self, n);
}
// Computes the rank of 'input' and saves the result in-place in 'result'
// 'hermitian' controls whether SVD or eigendecomposition is used for computing the singular values
// 'atol' and 'rtol' are the absolute and relative tolerances, respectively.
Tensor& linalg_matrix_rank_out(
const Tensor& input,
const optional<Tensor>& atol_opt,
const optional<Tensor>& rtol_opt,
bool hermitian,
Tensor& result) {
Tensor atol, rtol;
std::tie(atol, rtol) = get_atol_rtol(input, atol_opt, rtol_opt, "torch.linalg.matrix_rank");
checkSameDevice("torch.linalg.matrix_rank", result, input);
checkSameDevice("torch.linalg.matrix_rank", atol, input, "atol");
checkSameDevice("torch.linalg.matrix_rank", rtol, input, "rtol");
ScalarType output_type = ScalarType::Long;
checkLinalgCompatibleDtype("torch.linalg.matrix_rank", result.scalar_type(), output_type);
// Matrices or batch of matrices are allowed
TORCH_CHECK(input.dim() >= 2, "torch.linalg.matrix_rank: Expected as input a matrix or a batch of matrices, but got a tensor of size: ", input.sizes());
checkNotComplexTolerance(atol, "torch.linalg.matrix_rank", "atol");
checkNotComplexTolerance(rtol, "torch.linalg.matrix_rank", "rtol");
// matrix_rank assigns a scalar value for each matrix in the batch so
// result's shape is equal to input.shape[0:input.ndim-2]
// for single matrix result_shape = {}
auto result_shape = IntArrayRef(input.sizes().cbegin(), input.sizes().cend() - 2);
at::native::resize_output(result, result_shape);
// NumPy doesn't take into account possible input with no elements and it errors on max not defined for this case
// Let's output 0 for this case, since that kind of matrices have zero number of non-zero rows, hence rank is 0.
if (input.numel() == 0) {
result.fill_(0);
return result;
}
// We compute matrix rank as the number of singular or absolute eigen values
// that are above max(atol, rtol * max(S)) threshold
Tensor S, max_S;
if (!hermitian) {
S = at::linalg_svdvals(input);
// singular values are sorted in descending order
max_S = at::narrow(S, /*dim=*/-1, /*start=*/0, /*length=*/1);
} else {
S = at::linalg_eigvalsh(input);
S = S.abs();
// eigenvalues are sorted in ascending order starting with negative values, we need a maximum value of abs(eigenvalues)
max_S = S.amax(/*dim=*/-1, /*keepdim=*/true);
}
Tensor tol = at::max(atol.unsqueeze(-1), rtol.unsqueeze(-1) * max_S);
result = at::sum_out(result, S > tol, /*dim=*/-1);
return result;
}
Tensor& linalg_matrix_rank_out(const Tensor& input, optional<double> atol, optional<double> rtol, bool hermitian, Tensor& result) {
Tensor atol_tensor, rtol_tensor;
std::tie(atol_tensor, rtol_tensor) = get_atol_rtol(input, atol, rtol);
result = linalg_matrix_rank_out(input, atol_tensor, rtol_tensor, hermitian, result);
return result;
}
Tensor linalg_matrix_rank(const Tensor& input, const optional<Tensor>& atol, const optional<Tensor>& rtol, bool hermitian) {
Tensor result = at::empty({0}, input.options().dtype(ScalarType::Long));
result = at::linalg_matrix_rank_outf(input, atol, rtol, hermitian, result);
return result;
}
Tensor linalg_matrix_rank(const Tensor& input, optional<double> atol, optional<double> rtol, bool hermitian) {
Tensor result = at::empty({0}, input.options().dtype(ScalarType::Long));
result = at::linalg_matrix_rank_outf(input, atol, rtol, hermitian, result);
return result;
}
Tensor& linalg_matrix_rank_out(const Tensor& input, const Tensor& tol, bool hermitian, Tensor& result) {
// For NumPy compatibility tol is not scaled with max(singular_value) if the value for tol is provided
// It is assumed that the provided value is the absolute tolerance
Tensor rtol = at::zeros({}, tol.options());
result = at::linalg_matrix_rank_outf(input, tol, rtol, hermitian, result);
return result;
}
Tensor& linalg_matrix_rank_out(const Tensor& input, double tol, bool hermitian, Tensor& result) {
// For NumPy compatibility tol is not scaled with max(singular_value) if the value for tol is provided
// It is assumed that the provided value is the absolute tolerance
result = at::linalg_matrix_rank_outf(input, tol, 0.0, hermitian, result);
return result;
}
Tensor linalg_matrix_rank(const Tensor& input, const Tensor& tol, bool hermitian) {
Tensor result = at::empty({0}, input.options().dtype(ScalarType::Long));
result = at::linalg_matrix_rank_outf(input, tol, hermitian, result);
return result;
}
Tensor linalg_matrix_rank(const Tensor& input, double tol, bool hermitian) {
Tensor result = at::empty({0}, input.options().dtype(ScalarType::Long));
result = at::linalg_matrix_rank_outf(input, tol, hermitian, result);
return result;
}
Tensor matrix_rank(const Tensor& self, double tol, bool symmetric) {
TORCH_WARN_ONCE(
"torch.matrix_rank is deprecated in favor of torch.linalg.matrix_rank",
"and will be removed in a future PyTorch release. The parameter 'symmetric' was ",
"renamed in torch.linalg.matrix_rank to 'hermitian'."
);
return at::linalg_matrix_rank(self, tol, symmetric);
}
Tensor matrix_rank(const Tensor& self, bool symmetric) {
TORCH_WARN_ONCE(
"torch.matrix_rank is deprecated in favor of torch.linalg.matrix_rank",
"and will be removed in a future PyTorch release. The parameter 'symmetric' was ",
"renamed in torch.linalg.matrix_rank to 'hermitian'."
);
return at::linalg_matrix_rank(self, 0.0, c10::nullopt, symmetric);
}
// multi_dot helper functions
namespace {
/**
* @brief Computes the optimal matrix chain multiplication order
*
* Follows the dynamic programming algorithm from Cormen et al,
* "Introduction to Algorithms, Third Edition", Chapter 15.2,
* p. 370-378. Note that the book uses 1-based indexing.
*
* The cost of multiplying two matrices with sizes p x q and q x r
* is defined here as p * q * r. The optimal multiplication order
* is the one that minimizes the total cost.
*
* @param tensors list of 2D tensors
* @return a 2D vector s used by #matrix_chain_multiplication to construct
* the optimal matrix multiplication order. The optimal multiplication
* order for multiplying tensors i...j is to multiply tensors i...s[i, j]
* and tensors (s[i, j] + 1)...j first and then the result of that.
*/
std::vector<std::vector<int64_t>> matrix_chain_order(TensorList tensors) {
const size_t n = tensors.size();
// Tensor i has dimensions p[i] x p[i + 1]
std::vector<int64_t> p(n + 1);
for (const auto i : c10::irange(n)) {
p[i] = tensors[i].size(0);
}
p[n] = tensors[n - 1].size(1);
// m[i, j] = k where k is the minimum cost for multiplying tensors i...j
std::vector<std::vector<int64_t>> m(n, std::vector<int64_t>(n, 0));
// s[i, j] = k where k is the index at which to split the list such that
// optimally multiplying matrices i...k and k...j first and then the resulting
// matrices is the optimal order for multiplying matrices i...j.
std::vector<std::vector<int64_t>> s(n, std::vector<int64_t>(n));
// Compute the optimal multiplication order
for (const auto l : c10::irange(1, n)) {
for (const auto i : c10::irange(n - l)) {
const auto j = i + l;
m[i][j] = std::numeric_limits<int64_t>::max();
for (const auto k : c10::irange(i, j)) {
const auto q = m[i][k] + m[k + 1][j] + p[i] * p[k + 1] * p[j + 1];
if (q < m[i][j]) {
m[i][j] = q;
s[i][j] = k;
}
}
}
}
return s;
}
/**
* @brief Recursively multiplies the tensors i...j using the given order
*
* @param tensors matrices to multiply togther
* @param order optimal chain multiplication order from #matrix_chain_order
* @param i index of first tensor to be multiplied
* @param j index of last tensor to be multiplied
* @return Tensor result of multiplying tensors[i...j] together.
*/
Tensor matrix_chain_multiplication(
TensorList tensors,
const std::vector<std::vector<int64_t>>& order,
int64_t i,
int64_t j) {
if (i == j) {
return tensors[i];
}
return at::mm(
matrix_chain_multiplication(tensors, order, i, order[i][j]),
matrix_chain_multiplication(tensors, order, order[i][j] + 1, j));
}
// Implements torch.linalg.multi_dot
Tensor multi_dot_impl(TensorList _tensors, c10::optional<Tensor> _out) {
const size_t n = _tensors.size();
TORCH_CHECK(n >= 2, "multi_dot(): expected at least 2 tensors but got ", n);
std::vector<int64_t> out_shape;
std::vector<Tensor> tensors(n);
// If the first tensor is 1D of size n view it as a row vector (1, n)
if (_tensors[0].dim() == 1) {
tensors[0] = _tensors[0].unsqueeze(0);
} else if (_tensors[0].dim() == 2) {
tensors[0] = _tensors[0];
out_shape.emplace_back(tensors[0].size(0));
} else {
TORCH_CHECK(
false,
"multi_dot(): the first tensor must be 1D or 2D but got ",
_tensors[0].dim(),
"D");
}
// If the last tensor is 1D of size n view it as a column vector (n, 1)
if (_tensors[n - 1].dim() == 1) {
tensors[n - 1] = _tensors[n - 1].unsqueeze(-1);
} else if (_tensors[n - 1].dim() == 2) {
tensors[n - 1] = _tensors[n - 1];
out_shape.emplace_back(tensors[n - 1].size(1));
} else {
TORCH_CHECK(
false,
"multi_dot(): the last tensor must be 1D or 2D but got ",
_tensors[n - 1].dim(),
"D");
}
// Ensure middle tensors are 2D
for (const auto i : c10::irange(1, n - 1)) {
TORCH_CHECK(
_tensors[i].dim() == 2,
"multi_dot(): tensor ",
i,
" must be 2D but got ",
_tensors[i].dim(),
"D");
tensors[i] = _tensors[i];
}
// Ensure all tensors have the same device and dtype and check
// that the shapes can be multiplied
const auto dtype = tensors[0].dtype();
const auto device = tensors[0].device();
for (const auto i : c10::irange(1, n)) {
TORCH_CHECK(
tensors[i].dtype() == dtype,
"multi_dot(): all tensors must have be the same dtype but tensor 0 is ",
dtype,
" and tensor ",
i,
" ",
tensors[i].dtype());
TORCH_CHECK(
tensors[i].device() == device,
"multi_dot(): all tensors must be on the same device but tensor 0 is on ",
device,
" and tensor ",
i,
" on ",
tensors[i].device());
TORCH_CHECK(
tensors[i - 1].size(-1) == tensors[i].size(0),
"multi_dot(): tensors ",
i - 1,
" and ",
i,
" with shapes ",
_tensors[i - 1].sizes(),
" and ",
_tensors[i].sizes(),
" cannot be multiplied")
}
Tensor result;
if (_out.has_value()) {
auto out = *_out;
TORCH_CHECK(
dtype == out.dtype(),
"multi_dot(): expected out tensor to have dtype ",
dtype,
" but got ",
out.dtype());
TORCH_CHECK(
device == out.device(),
"multi_dot(): expected out tensor to be on device ",
device,
" but got ",
out.device());
// If the last and last tensors have shapes (a, b) and (b, c) the
// output has shape (a, c). If either the first or last tensor is 1D
// a and/or c dimensions will be implicitely size 1 and will be ommited
// from the output. e.g. for inputs (a, b) x (b) the output has shape (a,).
at::native::resize_output(out, out_shape);
// View output as 2D for simplicity of computation.
result = out.view({tensors[0].size(0), tensors.back().size(-1)});
}
// The resize_ and view calls below are to ensure the
// output shape respects the original dimensionality of
// the first and last tensors which we are now viewed as 2D
if (tensors.size() == 2) {
return _out.has_value() ? at::mm_out(result, tensors[0], tensors[1])
: at::mm(tensors[0], tensors[1]).view(out_shape);
}
// Why the separate implementation for 3 matrices?
// The logic for three matrices is much faster when done directly
// Requires 1 comparison to 4 comparisons and fewer arithmetic operations
if (tensors.size() == 3) {
const auto a = tensors[0].size(0);
const auto b = tensors[1].size(0);
const auto c = tensors[2].size(0);
const auto d = tensors[2].size(1);
// The matrices are of size (a x b), (b x c), (c x d)
// cost_1 is the cost of parenthesizing (a x b) and (b x c) and then
// combining (c x d) cost_2 is the cost of parenthesizing (b x c) and (c x
// d) and then combining (a x b)
const auto cost_1 = (a * c) * (b + d);
const auto cost_2 = (b * d) * (a + c);
if (cost_1 > cost_2) {
return _out.has_value()
? at::mm_out(result, tensors[0], at::mm(tensors[1], tensors[2]))
: at::mm(tensors[0], at::mm(tensors[1], tensors[2])).view(out_shape);
} else {
return _out.has_value()
? at::mm_out(result, at::mm(tensors[0], tensors[1]), tensors[2])
: at::mm(at::mm(tensors[0], tensors[1]), tensors[2]).view(out_shape);
}
}
// Algorithm for multiplying 4 or more matrices
const auto order = matrix_chain_order(tensors);
const int64_t i = 0;
const int64_t j = n - 1;
if (_out.has_value()) {
// We manually implement the first recursive layer here so we can use mm_out
// for the final multiplication
return at::mm_out(
result,
matrix_chain_multiplication(tensors, order, i, order[i][j]),
matrix_chain_multiplication(tensors, order, order[i][j] + 1, j));
}
return matrix_chain_multiplication(tensors, order, i, j).view(out_shape);
}
} // namespace
Tensor linalg_multi_dot(TensorList tensors) {
return multi_dot_impl(tensors, c10::nullopt);
}
Tensor& linalg_multi_dot_out(TensorList tensors, Tensor& result) {
multi_dot_impl(tensors, result);
return result;
}
Tensor chain_matmul(TensorList matrices) {
TORCH_WARN_ONCE(
"torch.chain_matmul is deprecated and will be removed in a future PyTorch release. ",
"Use torch.linalg.multi_dot instead, which accepts a list of two or more tensors rather than ",
"multiple parameters."
);
checkAllSameDim(matrices, 2);
TORCH_CHECK(
matrices.size() > 0, "chain_matmul(): Expected one or more matrices");
if (matrices.size() == 1) {
return matrices[0].clone();
}
return at::native::linalg_multi_dot(matrices);
}
Tensor& chain_matmul_out(TensorList matrices, Tensor& result) {
TORCH_WARN_ONCE(
"torch.chain_matmul is deprecated and will be removed in a future PyTorch release. ",
"Use torch.linalg.multi_dot instead, which accepts a list of two or more tensors rather than ",
"multiple parameters."
);
checkAllSameDim(matrices, 2);
TORCH_CHECK(
matrices.size() > 0, "chain_matmul(): Expected one or more matrices");
if (matrices.size() == 1) {
at::native::resize_output(result, matrices[0].sizes());
return result.copy_(matrices[0]);
}
return at::native::linalg_multi_dot_out(matrices, result);
}
static void check_1d(const Tensor& t, const char* arg, const char* fn) {
TORCH_CHECK(t.dim() == 1, fn, ": Expected 1-D argument ", arg, ", but got ", t.dim(), "-D");
}
static void check_addr_scalar(const ScalarType dtype,
const Scalar& scalar,
const std::string& scalar_name) {
TORCH_CHECK(
!scalar.isBoolean() || dtype == ScalarType::Bool,
"Boolean ", scalar_name, " only supported for Boolean results.");
TORCH_CHECK(
isFloatingType(dtype) || isComplexType(dtype) || scalar.isIntegral(true),
"For integral input tensors, "
"argument ", scalar_name ," must not be a floating point number.");
}
static TensorIterator build_addr_iter(Tensor& result,
const Tensor& self,
const Tensor& vec1,
const Tensor& vec2) {
check_1d(vec1, "vec1", "addr");
check_1d(vec2, "vec2", "addr");
const auto vec1_size0 = vec1.sizes()[0];
const auto vec2_size0 = vec2.sizes()[0];
auto self_ = &result == &self
? c10::MaybeOwned<Tensor>::borrowed(self)
: expand_size(self, {vec1_size0, vec2_size0}, "addr");
TORCH_CHECK(
self_->dim() == 2,
"2D tensor expected, got ", self_->dim(), "D tensor for input"
);
TORCH_CHECK(
self_->sizes()[0] == vec1_size0 && self_->sizes()[1] == vec2_size0,
"size mismatch, input: ", self_->sizes(),
", v1: ", vec1.sizes(),
", v2: ", vec2.sizes()
);
auto iter = TensorIteratorConfig()
.set_check_mem_overlap(true)
.add_output(result)
.add_owned_input(*self_)
.add_owned_input(vec1.reshape({vec1_size0, 1}))
.add_input(vec2)
.allow_cpu_scalars(true)
.promote_inputs_to_common_dtype(true)
.cast_common_dtype_to_outputs(true)
.enforce_safe_casting_to_output(true)
.build();
return iter;
}
Tensor addr(const Tensor& self,
const Tensor& vec1, const Tensor& vec2,
const Scalar& beta, const Scalar& alpha) {
Tensor result;
auto iter = build_addr_iter(result, self, vec1, vec2);
check_addr_scalar(iter.dtype(), beta, "beta");
check_addr_scalar(iter.dtype(), alpha, "alpha");
addr_stub(iter.device_type(), iter, beta, alpha);
return iter.output();
}
Tensor& addr_(Tensor& self,
const Tensor& vec1, const Tensor& vec2,
const Scalar& beta, const Scalar& alpha) {
return at::addr_out(self, self, vec1, vec2, beta, alpha);
}
Tensor& addr_out(const Tensor& self,
const Tensor& vec1, const Tensor& vec2,
const Scalar& beta, const Scalar& alpha, Tensor &result) {
auto iter = build_addr_iter(result, self, vec1, vec2);
check_addr_scalar(iter.dtype(), beta, "beta");
check_addr_scalar(iter.dtype(), alpha, "alpha");
addr_stub(iter.device_type(), iter, beta, alpha);
return result;
}
// The math_addr and math_addr_out functions support backends
// other than CPU and CUDA, such as XLA.
// They are implemented using the composition of existing ops
Tensor math_addr(const Tensor& self,
const Tensor& vec1, const Tensor& vec2,
const Scalar& beta, const Scalar& alpha) {
// when beta==0, values in self should be ignored,