-
Notifications
You must be signed in to change notification settings - Fork 461
/
aten_xla_type.cpp
4236 lines (3844 loc) · 189 KB
/
aten_xla_type.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/ExpandUtils.h>
#include <ATen/FunctionalTensorWrapper.h>
#include <ATen/MetaFunctions.h>
#include <ATen/NativeFunctions.h>
#include <ATen/OpMathType.h>
#include <ATen/Operators.h>
#include <ATen/native/BinaryOps.h>
#include <ATen/native/CPUFallback.h>
#include <ATen/native/TypeProperties.h>
#include <ATen/ops/_embedding_bag_backward_native.h>
#include <ATen/ops/expand_copy.h>
#include <c10/core/Contiguity.h>
#include <torch/csrc/lazy/core/shape_inference.h>
#include <torch/csrc/lazy/core/tensor_util.h>
#include <torch/csrc/lazy/core/util.h>
#include <mutex>
#include <optional>
#include "torch/csrc/lazy/core/helpers.h"
#include "torch/csrc/lazy/core/shape_inference.h"
#include "torch/csrc/lazy/core/tensor_util.h"
#include "torch/csrc/lazy/core/util.h"
#include "torch_xla/csrc/LazyIr.h"
#include "torch_xla/csrc/XLANativeFunctions.h"
#include "torch_xla/csrc/aten_autograd_ops.h"
#include "torch_xla/csrc/aten_fallback.h"
#include "torch_xla/csrc/aten_xla_bridge.h"
#include "torch_xla/csrc/debug_util.h"
#include "torch_xla/csrc/device.h"
#include "torch_xla/csrc/dtype.h"
#include "torch_xla/csrc/helpers.h"
#include "torch_xla/csrc/ops/as_strided.h"
#include "torch_xla/csrc/ops/as_strided_view_update.h"
#include "torch_xla/csrc/ops/device_data.h"
#include "torch_xla/csrc/ops/diagonal_view_update.h"
#include "torch_xla/csrc/ops/einsum_utilities.h"
#include "torch_xla/csrc/ops/index_ops.h"
#include "torch_xla/csrc/ops/unselect.h"
#include "torch_xla/csrc/ops/update_slice.h"
#include "torch_xla/csrc/ops/view.h"
#include "torch_xla/csrc/pooling.h"
#include "torch_xla/csrc/runtime/debug_macros.h"
#include "torch_xla/csrc/runtime/metrics.h"
#include "torch_xla/csrc/runtime/sys_util.h"
#include "torch_xla/csrc/runtime/util.h"
#include "torch_xla/csrc/tensor_impl.h"
#include "torch_xla/csrc/tensor_methods.h"
#include "torch_xla/csrc/tensor_util.h"
#include "torch_xla/csrc/torch_util.h"
#include "torch_xla/csrc/xla_graph_executor.h"
#include "torch_xla/csrc/xla_sharding_util.h"
// [Implementation Guidelines]
// - If you want to call a at::func which doesn't have a kernel registered
// according to xla_native_functions.yaml,
// you can call a boxed CPU fallback kernel instead.
// E.g. don't call tensor.op() or at::op(tensor).
// use at::native::call_fallback_fn<&xla_fallback,
// ATEN_OP2(op_name, overload_name)>::call(args...)
// ATEN_OP accepts an operator name without an overload, and
// ATEN_OP2 accepts an operator name along with its overload name.
// The description of these macros can be found in
// https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/templates/Operators.h
// (You can find some examples below)
namespace torch_xla {
namespace {
using XLAInputVector = std::vector<XLATensorPtr>;
// Calls the inner function by spreading inputs in order, and adding the
// common data-type in the end.
template <class InnerFnType, size_t... Ints>
XLATensorPtr CallInner(const InnerFnType& inner, XLAInputVector inputs,
at::ScalarType common_dtype,
std::integer_sequence<size_t, Ints...> seq) {
return inner(inputs[Ints]..., common_dtype);
}
// Computes the number of XLATensorPtr arguments of a given function.
//
// This is used when calling tensor_methods functions, given a list of inputs.
// Specifically, in order to know how many inputs we should get from the list.
template <class T>
struct NumberOfXLATensorArgs {};
template <class... Args>
struct NumberOfXLATensorArgs<XLATensorPtr(Args...)> {
static constexpr size_t value =
(std::is_same_v<XLATensorPtr,
std::remove_cv_t<std::remove_reference_t<Args>>> +
...);
};
// Stateful configuration structure for pre/post-processing the inputs and the
// output.
//
// There are a few checks and preprocessing that PyTorch does, that we are
// mirroring with this class. This should help us get many data-type behavior
// right.
class OpConfig {
public:
using InputVector = std::vector<at::Tensor>;
using ImplFnType =
std::function<XLATensorPtr(const XLAInputVector&, at::ScalarType)>;
// Construct an instance from a function of exactly ImplFnType.
OpConfig(ImplFnType impl) : impl_(impl) {}
// Construct an instance from a function of the following type:
// XLATensorPtr(Tensor..., ScalarType)
//
// This is a convenience for wrapping tensor_methods functions.
template <class InnerFnType>
static OpConfig From(const InnerFnType& inner_impl) {
return OpConfig(
[&](const XLAInputVector& inputs, at::ScalarType common_dtype) {
constexpr size_t num_tensor_args =
NumberOfXLATensorArgs<std::remove_pointer_t<InnerFnType>>::value;
return CallInner(inner_impl, inputs, common_dtype,
std::make_index_sequence<num_tensor_args>{});
});
}
OpConfig& add_input(const at::Tensor& input) {
inputs_.push_back(input);
return *this;
}
OpConfig& cast_inputs_to_common_dtype() {
cast_inputs_to_common_dtype_ = true;
return *this;
}
OpConfig& use_opmathtype_for_compute() {
use_opmathtype_for_compute_ = true;
return *this;
}
// Pre-processes the inputs and post-processes the outputs depending on the
// configured state of this class.
//
// In summary, it will:
// - Compute the common data-type to be used
// - Cast the inputs to the common data-type
// - Cast the inputs to its OpMathType (for computation only)
// - Run the specified impl
// - Cast the output back to the common data-type
at::Tensor run() {
at::ScalarType common_dtype = at::native::result_type(inputs_);
at::ScalarType opmathtype = at::toOpMathType(common_dtype);
// Pre-process the inputs, given the specified configuration and
// common_dtype.
InputVector inputs = maybe_preprocess_inputs(common_dtype, opmathtype);
// Look for, at least, one tensor already in PyTorch/XLA.
InputVector::iterator it = std::find_if(
inputs.begin(), inputs.end(), [](const at::Tensor& tensor) {
return bridge::TryGetXlaTensor(tensor);
});
XLA_CHECK(it != inputs.end());
// Transform the inputs into a list of XLATensorPtr.
// For that, either get their corresponding XLATensorPtr, or use the found
// XLA tensor's BackendDevice for creating a new one.
torch::lazy::BackendDevice device = bridge::GetXlaTensor(*it)->GetDevice();
XLAInputVector xla_inputs(inputs.size());
std::transform(inputs.begin(), inputs.end(), xla_inputs.begin(),
[&](const at::Tensor& tensor) {
return bridge::GetOrCreateXlaTensor(tensor, device);
});
// Actually call the impl.
at::ScalarType inner_dtype =
(use_opmathtype_for_compute_) ? opmathtype : common_dtype;
XLATensorPtr xla_out = impl_(xla_inputs, inner_dtype);
at::Tensor out = bridge::AtenFromXlaTensor(xla_out);
// If we used OpMathType for the computation, cast the result back to its
// common_dtype.
if (use_opmathtype_for_compute_) {
out = out.to(common_dtype);
}
return out;
}
private:
// Pre-processes the inputs based on the state of this instance.
//
// In summary:
// - Cast the inputs to the common data-type (if
// cast_inputs_to_common_dtype_ is set)
//
// - Cast the inputs to the OpMathType data-type (if
// use_opmathtype_for_compute_ is set)
InputVector maybe_preprocess_inputs(at::ScalarType common_dtype,
at::ScalarType opmathtype) {
InputVector inputs = inputs_;
// Cast only once: either to the common dtype or to OpMathType.
if (use_opmathtype_for_compute_) {
std::transform(
inputs.begin(), inputs.end(), inputs.begin(),
[=](const at::Tensor& tensor) { return tensor.to(opmathtype); });
} else if (cast_inputs_to_common_dtype_) {
std::transform(
inputs.begin(), inputs.end(), inputs.begin(),
[=](const at::Tensor& tensor) { return tensor.to(common_dtype); });
}
return inputs;
}
// Actual implementation of the operation.
ImplFnType impl_;
// List of tensor inputs.
InputVector inputs_;
// Whether to cast every input to the common data-type.
// It's analogous to TensorIterator's flag. If the operation you are lowering
// uses TensorIterator in PyTorch, you can check whether to set this flag or
// not.
bool cast_inputs_to_common_dtype_ = false;
// Whether to use OpMathType for computation.
// This flag mimics the actual PyTorch kernel implementations. When lowering
// an operation, take a look at that for deciding whether to set this flag or
// not.
bool use_opmathtype_for_compute_ = false;
};
at::Tensor to_meta(const at::Tensor& tensor) {
// undefined tensors can't be converted to the meta device, since they don't
// have sizes/strides
if (!tensor.defined()) return tensor;
auto out = at::native::empty_strided_meta_symint(
tensor.sym_sizes(), tensor.sym_strides(),
/*dtype=*/std::make_optional(tensor.scalar_type()),
/*layout=*/std::make_optional(tensor.layout()),
/*device=*/std::make_optional(c10::Device(c10::kMeta)),
/*pin_memory=*/std::nullopt);
// needs to handle wrapped numbers, so dtype promotion works properly.
if (tensor.unsafeGetTensorImpl()->is_wrapped_number()) {
out.unsafeGetTensorImpl()->set_wrapped_number(true);
}
return out;
}
torch::lazy::BackendDevice GetXlaDeviceOrCurrent(
const std::optional<c10::Device>& device) {
auto xla_device_opt = bridge::GetXlaDevice(device);
return xla_device_opt ? *xla_device_opt : bridge::GetCurrentDevice();
}
bool IsOperationOnType(const std::optional<at::ScalarType>& opt_dtype,
at::ScalarType tensor_type, at::ScalarType type) {
if (opt_dtype && *opt_dtype == type) {
return true;
}
return tensor_type == type;
}
bool TensorsAreOfType(std::vector<XLATensorPtr> tensors, at::ScalarType type) {
for (const XLATensorPtr& tensor : tensors) {
if (IsOperationOnType(std::optional<at::ScalarType>(std::nullopt),
tensor->dtype(), type)) {
return true;
}
}
return false;
}
void CheckSubOperandTypes(at::ScalarType type1, at::ScalarType type2) {
XLA_CHECK(type1 != at::kBool || type2 != at::kBool)
<< "Subtraction, the `-` operator, with two bool tensors is not "
"supported. Use the `^` or `logical_xor()` operator instead.";
XLA_CHECK(type1 != at::kBool && type2 != at::kBool)
<< "Subtraction, the `-` operator, with a bool tensor is not "
"supported. If you are trying to invert a mask, use the `~` or "
"`logical_not()` operator instead.";
}
std::optional<at::ScalarType> PromoteIntegralType(
at::ScalarType src_dtype, const std::optional<at::ScalarType>& opt_dtype) {
return opt_dtype.has_value() ? opt_dtype.value()
: at::isIntegralType(src_dtype, /*includeBool=*/true) ? at::kLong
: opt_dtype;
}
bool IsTypeWithLargerRangeThanLong(torch::ScalarType dtype) {
return dtype == at::ScalarType::BFloat16 || dtype == at::ScalarType::Float ||
dtype == at::ScalarType::Double;
}
// Return the upper limit for a given type. For floating point typesreturn
// 2^mantissa to ensure that every value is representable.
int64_t GetIntegerUpperLimitForType(torch::ScalarType dtype) {
xla::PrimitiveType xla_type = XlaTypeFromTorchType(dtype);
switch (xla_type) {
case xla::PrimitiveType::F16:
return static_cast<int64_t>(1) << std::numeric_limits<xla::half>::digits;
case xla::PrimitiveType::BF16:
return static_cast<int64_t>(1)
<< std::numeric_limits<xla::bfloat16>::digits;
case xla::PrimitiveType::F32:
return static_cast<int64_t>(1) << std::numeric_limits<float>::digits;
case xla::PrimitiveType::F64:
return static_cast<int64_t>(1) << std::numeric_limits<double>::digits;
default:
return XlaHelpers::MinMaxValues(xla_type).max.toLong();
}
}
void CheckRangeValues(torch::ScalarType dtype, int64_t from, int64_t to) {
XlaHelpers::MinMax min_max;
// Bound the min_max by int64_t since types of "from" and "to" are int64.
if (IsTypeWithLargerRangeThanLong(dtype)) {
min_max = XlaHelpers::MinMaxValues(xla::PrimitiveType::S64);
} else {
min_max = XlaHelpers::MinMaxValues(XlaTypeFromTorchType(dtype));
}
XLA_CHECK_GE(from, min_max.min.toLong());
XLA_CHECK_LE(from, min_max.max.toLong());
XLA_CHECK_GE(to, min_max.min.toLong());
XLA_CHECK_LE(to, min_max.max.toLong());
}
std::pair<XLATensorPtr, XLATensorPtr> GetBinaryOperands(
const at::Tensor& self, const at::Tensor& other) {
XLATensorPtr self_tensor;
XLATensorPtr other_tensor;
auto self_xtensor = bridge::TryGetXlaTensor(self);
if (!self_xtensor) {
other_tensor = bridge::GetXlaTensor(other);
self_tensor = bridge::GetOrCreateXlaTensor(self, other_tensor->GetDevice());
} else {
self_tensor = self_xtensor;
other_tensor =
bridge::GetOrCreateXlaTensor(other, self_tensor->GetDevice());
}
return std::pair<XLATensorPtr, XLATensorPtr>(self_tensor, other_tensor);
}
// The input is in format of {N, C, H, W} and the output will be {H, W}.
std::vector<int64_t> GetOutputSizeWithScale(
absl::Span<const int64_t> input_size, const std::optional<double> scales_h,
const std::optional<double> scales_w,
const std::vector<int64_t>& output_size) {
XLA_CHECK(scales_h);
XLA_CHECK(scales_w);
// Calculate the output size from input_shape and scale_factors
XLA_CHECK_EQ(input_size.size(), 4);
int64_t output_h = input_size[2] * (*scales_h);
int64_t output_w = input_size[3] * (*scales_w);
return {output_h, output_w};
}
void CheckBinaryOpTypePromotion(const at::Tensor& out, const at::Tensor& self,
const at::Tensor& other) {
at::ScalarType resultType = at::result_type(self, other);
XLA_CHECK(at::canCast(/*from=*/resultType, /*to=*/out.scalar_type()));
}
void CheckBinaryOpTypePromotion(const at::Tensor& out, const at::Tensor& self,
const at::Scalar& other) {
at::ScalarType resultType = at::result_type(self, other);
XLA_CHECK(at::canCast(/*from=*/resultType, /*to=*/out.scalar_type()));
}
template <typename B>
at::Tensor DoBinaryOp(const at::Tensor& self, const at::Tensor& other,
const B& bin_op) {
at::ScalarType dtype = at::result_type(self, other);
std::pair<XLATensorPtr, XLATensorPtr> operands =
GetBinaryOperands(self, UnwrapNumber(other, dtype));
XLATensorPtr result = bin_op(operands.first, operands.second, dtype);
return bridge::AtenFromXlaTensor(result);
}
template <typename B>
at::Tensor DoBinaryOp(const at::Tensor& self, const at::Scalar& other,
const B& bin_op) {
at::ScalarType dtype = at::result_type(self, other);
XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
XLATensorPtr result = bin_op(self_tensor, other, dtype);
return bridge::AtenFromXlaTensor(result);
}
template <typename B>
at::Tensor DoBinaryOp(const at::Scalar& self, const at::Tensor& other,
const B& bin_op) {
at::ScalarType dtype = at::result_type(self, other);
XLATensorPtr other_tensor = bridge::GetXlaTensor(other);
XLATensorPtr result = bin_op(self, other_tensor, dtype);
return bridge::AtenFromXlaTensor(result);
}
template <typename B>
at::Tensor DoBinaryOpWithoutPromo(const at::Tensor& self,
const at::Tensor& other, const B& bin_op) {
at::ScalarType dtype = at::result_type(self, other);
std::pair<XLATensorPtr, XLATensorPtr> operands =
GetBinaryOperands(self, UnwrapNumber(other, dtype));
XLATensorPtr result = bin_op(operands.first, operands.second);
return bridge::AtenFromXlaTensor(result);
}
template <typename B>
at::Tensor DoBinaryOpWithoutPromo(const at::Tensor& self,
const at::Scalar& other, const B& bin_op) {
XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
XLATensorPtr result = bin_op(self_tensor, other);
return bridge::AtenFromXlaTensor(result);
}
template <typename B>
void DoBinaryOpOut(const at::Tensor& self, const at::Tensor& other,
at::Tensor& out, const B& bin_op_out) {
at::ScalarType dtype = at::result_type(self, other);
XLA_CHECK(at::canCast(/*from=*/dtype, /*to=*/out.scalar_type()));
std::pair<XLATensorPtr, XLATensorPtr> operands =
GetBinaryOperands(self, UnwrapNumber(other, dtype));
XLATensorPtr out_tensor = bridge::GetXlaTensor(out);
bin_op_out(operands.first, operands.second, out_tensor);
}
} // namespace
at::Tensor& XLANativeFunctions::__ilshift__(at::Tensor& self,
const at::Scalar& other) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
tensor_methods::__ilshift__(self_tensor, other);
return self;
}
at::Tensor& XLANativeFunctions::__ilshift__(at::Tensor& self,
const at::Tensor& other) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
CheckBinaryOpTypePromotion(self, self, other);
XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
tensor_methods::__ilshift__(self_tensor, bridge::GetXlaTensor(other));
return self;
}
at::Tensor& XLANativeFunctions::__irshift__(at::Tensor& self,
const at::Scalar& other) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
CheckBinaryOpTypePromotion(self, self, other);
XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
tensor_methods::__irshift__(self_tensor, other);
return self;
}
at::Tensor& XLANativeFunctions::__irshift__(at::Tensor& self,
const at::Tensor& other) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
CheckBinaryOpTypePromotion(self, self, other);
XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
tensor_methods::__irshift__(self_tensor, bridge::GetXlaTensor(other));
return self;
}
at::Tensor XLANativeFunctions::__lshift__(const at::Tensor& self,
const at::Scalar& other) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return DoBinaryOp(self, other,
[&](const XLATensorPtr& xself, const at::Scalar& other,
at::ScalarType dtype) {
return tensor_methods::__lshift__(xself, other, dtype);
});
}
at::Tensor XLANativeFunctions::__lshift__(const at::Tensor& self,
const at::Tensor& other) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return DoBinaryOp(self, other,
[&](const XLATensorPtr& xself, const XLATensorPtr& xother,
at::ScalarType dtype) {
return tensor_methods::__lshift__(xself, xother, dtype);
});
}
at::Tensor XLANativeFunctions::__rshift__(const at::Tensor& self,
const at::Scalar& other) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return DoBinaryOp(self, other,
[&](const XLATensorPtr& xself, const at::Scalar& other,
at::ScalarType dtype) {
return tensor_methods::__rshift__(xself, other, dtype);
});
}
at::Tensor XLANativeFunctions::__rshift__(const at::Tensor& self,
const at::Tensor& other) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return DoBinaryOp(self, other,
[&](const XLATensorPtr& xself, const XLATensorPtr& xother,
at::ScalarType dtype) {
return tensor_methods::__rshift__(xself, xother, dtype);
});
}
at::Tensor XLANativeFunctions::_adaptive_avg_pool3d(
const at::Tensor& self, at::IntArrayRef output_size) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
auto output_size_list = XlaHelpers::I64List(output_size);
if (!IsSupportedAdaptivePool(XlaHelpers::I64List(self.sizes()),
output_size_list, /*pool_dim=*/3)) {
return at::native::call_fallback_fn<
&xla_fallback, ATEN_OP(_adaptive_avg_pool3d)>::call(self, output_size);
}
auto common_device = torch_xla::bridge::GetXlaDevice(self);
XLA_CHECK(common_device);
torch::lazy::NodePtr node = torch_xla::MakeNode<AdaptiveAvgPool3d>(
bridge::GetXlaTensor(self)->GetIrValue(),
std::vector<int64_t>(output_size.begin(), output_size.end()));
return torch_xla::bridge::AtenFromXlaTensor(
torch_xla::XLATensor::Create(std::move(node), *common_device));
}
at::Tensor XLANativeFunctions::_adaptive_avg_pool3d_backward(
const at::Tensor& grad_output, const at::Tensor& self) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
int64_t rank = grad_output.dim();
std::vector<int64_t> output_size{grad_output.size(rank - 3),
grad_output.size(rank - 2),
grad_output.size(rank - 1)};
if (!IsSupportedAdaptivePool(XlaHelpers::I64List(self.sizes()), output_size,
/*pool_dim=*/3)) {
return at::native::call_fallback_fn<
&xla_fallback,
ATEN_OP(_adaptive_avg_pool3d_backward)>::call(grad_output, self);
}
auto common_device = torch_xla::bridge::GetXlaDevice(grad_output, self);
XLA_CHECK(common_device);
torch::lazy::NodePtr node = torch_xla::MakeNode<AdaptiveAvgPool3dBackward>(
bridge::GetXlaTensor(grad_output)->GetIrValue(),
bridge::GetXlaTensor(self)->GetIrValue());
return torch_xla::bridge::AtenFromXlaTensor(
torch_xla::XLATensor::Create(std::move(node), *common_device));
}
at::Tensor XLANativeFunctions::_adaptive_avg_pool2d(
const at::Tensor& self, at::IntArrayRef output_size) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
auto output_size_list = XlaHelpers::I64List(output_size);
if (!IsSupportedAdaptivePool(XlaHelpers::I64List(self.sizes()),
output_size_list, /*pool_dim=*/2)) {
return at::native::call_fallback_fn<
&xla_fallback, ATEN_OP(_adaptive_avg_pool2d)>::call(self, output_size);
}
return bridge::AtenFromXlaTensor(tensor_methods::_adaptive_avg_pool2d(
bridge::GetXlaTensor(self), output_size_list));
}
at::Tensor XLANativeFunctions::_adaptive_avg_pool2d_backward(
const at::Tensor& grad_output, const at::Tensor& self) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
int64_t rank = grad_output.dim();
std::vector<int64_t> output_size{grad_output.size(rank - 2),
grad_output.size(rank - 1)};
if (!IsSupportedAdaptivePool(XlaHelpers::I64List(self.sizes()), output_size,
/*pool_dim=*/2)) {
return at::native::call_fallback_fn<
&xla_fallback,
ATEN_OP(_adaptive_avg_pool2d_backward)>::call(grad_output, self);
}
return bridge::AtenFromXlaTensor(
tensor_methods::_adaptive_avg_pool2d_backward(
bridge::GetXlaTensor(grad_output), bridge::GetXlaTensor(self)));
}
std::tuple<at::Tensor, at::Tensor> XLANativeFunctions::adaptive_max_pool2d(
const at::Tensor& self, at::IntArrayRef output_size) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
auto output_size_list = XlaHelpers::I64List(output_size);
if (!IsSupportedAdaptivePool(XlaHelpers::I64List(self.sizes()),
output_size_list, /*pool_dim=*/2)) {
return at::native::call_fallback_fn<
&xla_fallback, ATEN_OP(adaptive_max_pool2d)>::call(self, output_size);
}
std::tuple<XLATensorPtr, XLATensorPtr> res =
tensor_methods::adaptive_max_pool2d(bridge::GetXlaTensor(self),
output_size_list);
return std::make_tuple(bridge::AtenFromXlaTensor(std::get<0>(res)),
bridge::AtenFromXlaTensor(std::get<1>(res)));
}
at::Tensor XLANativeFunctions::adaptive_max_pool2d_backward(
const at::Tensor& grad_output, const at::Tensor& self,
const at::Tensor& indices) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
int64_t rank = grad_output.dim();
std::vector<int64_t> output_size{grad_output.size(rank - 2),
grad_output.size(rank - 1)};
if (!IsSupportedAdaptivePool(XlaHelpers::I64List(self.sizes()), output_size,
/*pool_dim=*/2)) {
return at::native::call_fallback_fn<
&xla_fallback, ATEN_OP(adaptive_max_pool2d_backward)>::call(grad_output,
self,
indices);
}
return bridge::AtenFromXlaTensor(tensor_methods::adaptive_max_pool2d_backward(
bridge::GetXlaTensor(grad_output), bridge::GetXlaTensor(self)));
}
void XLANativeFunctions::_amp_foreach_non_finite_check_and_unscale_(
at::TensorList self, at::Tensor& found_inf, const at::Tensor& inv_scale) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
XLATensorPtr found_inf_tensor = bridge::GetXlaTensor(found_inf);
tensor_methods::_amp_foreach_non_finite_check_and_unscale_(
bridge::GetXlaTensors(self), found_inf_tensor,
bridge::GetXlaTensor(inv_scale));
}
at::Tensor& XLANativeFunctions::_amp_update_scale_(at::Tensor& current_scale,
at::Tensor& growth_tracker,
const at::Tensor& found_inf,
double scale_growth_factor,
double scale_backoff_factor,
int64_t growth_interval) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
XLATensorPtr growth_tracker_tensor = bridge::GetXlaTensor(growth_tracker);
XLATensorPtr current_scale_tensor = bridge::GetXlaTensor(current_scale);
tensor_methods::_amp_update_scale_(
growth_tracker_tensor, current_scale_tensor,
bridge::GetXlaTensor(found_inf), scale_growth_factor,
scale_backoff_factor, growth_interval);
return current_scale;
}
at::Tensor XLANativeFunctions::_copy_from(const at::Tensor& self,
const at::Tensor& dst,
bool /*non_blocking*/) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
auto dst_tensor = bridge::TryGetXlaTensor(dst);
auto self_tensor = bridge::TryGetXlaTensor(self);
if (!self_tensor) {
static bool sync_update =
runtime::sys_util::GetEnvBool("XLA_TENSOR_UPDATE_SYNC", true) &&
!UseVirtualDevice();
dst_tensor->UpdateFromTensor(self, /*sync=*/sync_update);
XLA_CHECK(dst_tensor);
} else if (!dst_tensor) {
at::Tensor tensor = self_tensor->ToTensor(/*detached=*/true);
at::Tensor typed_tensor =
torch::lazy::CopyTensor(tensor, dst.scalar_type(), /*copy=*/false);
dst.resize_as_(typed_tensor).copy_(typed_tensor);
} else {
tensor_methods::copy_(dst_tensor, self_tensor);
bridge::ReplaceXlaTensor(dst, dst_tensor);
}
return dst;
}
at::Tensor XLANativeFunctions::_copy_from_and_resize(const at::Tensor& self,
const at::Tensor& dst) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
auto dst_tensor = bridge::TryGetXlaTensor(dst);
auto self_tensor = bridge::TryGetXlaTensor(self);
if (!self_tensor) {
XLA_CHECK(dst_tensor);
dst_tensor->UpdateFromTensorOut(self);
} else if (!dst_tensor) {
at::Tensor tensor = self_tensor->ToTensor(/*detached=*/true);
at::Tensor typed_tensor =
torch::lazy::CopyTensor(tensor, dst.scalar_type(), /*copy=*/false);
dst.resize_as_(typed_tensor).copy_(typed_tensor);
} else {
// at this point we know dst is an XLA tensor
XLATensorImpl* dest_impl =
dynamic_cast<XLATensorImpl*>(dst.unsafeGetTensorImpl());
dest_impl->tensor()->UpdateFromTensorOut(self_tensor);
dest_impl->force_refresh_sizes();
}
return dst;
}
std::vector<at::Tensor> XLANativeFunctions::_to_cpu(at::TensorList tensors) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return bridge::XlaCreateTensorList(tensors);
}
// TODO(alanwaketan): Improve the error messages.
// Let's rewrite it without reusing other native functions.
at::Tensor XLANativeFunctions::_to_copy(
const at::Tensor& self, std::optional<at::ScalarType> dtype,
std::optional<at::Layout> layout, std::optional<at::Device> device,
std::optional<bool> pin_memory, bool non_blocking,
std::optional<at::MemoryFormat> memory_format) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
auto options = self.options();
// I put each of these setters in a conditional instead of doing
// `self.options().dtype(dtype).layout(layout)... because calling
// .dtype(nullopt) on an options() that already has dtype appears to wipe it
if (dtype) {
options = options.dtype(dtype);
}
if (layout) {
options = options.layout(layout);
}
if (device) {
options = options.device(device);
}
if (pin_memory) {
options = options.pinned_memory(pin_memory);
}
if (memory_format) {
options = options.memory_format(memory_format);
}
// Case 1: Materialize the tensor.
if (device && device->type() != c10::kXLA) {
XLA_CHECK(device->type() == c10::kCPU)
<< "only cpu device is supported in _to_copy.";
auto self_tensor = bridge::GetXlaTensor(self);
auto eager_tensor = self_tensor->ToTensor(/*detached=*/true);
// Use the eager .to on the eager tensor.
return eager_tensor.to(options, non_blocking, /*copy=*/true);
}
// Case 2: Create a new XLA tensor with the supplied data and options.
auto new_tensor =
empty_symint(self.sym_sizes(), at::typeMetaToScalarType(options.dtype()),
options.layout(), options.device(), options.pinned_memory(),
options.memory_format_opt());
return _copy_from(self, new_tensor, non_blocking);
}
at::Tensor& XLANativeFunctions::_index_put_impl_(
at::Tensor& self, const c10::List<std::optional<at::Tensor>>& indices,
const at::Tensor& values, bool accumulate, bool /* unsafe */) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return torch_xla::XLANativeFunctions::index_put_(self, indices, values,
accumulate);
}
std::tuple<at::Tensor, at::Tensor> XLANativeFunctions::_linalg_eigh(
const at::Tensor& self, c10::string_view uplo, bool compute_v) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
if (!compute_v) {
// Fallback to aten in case of `eigvalsh`.
return at::native::call_fallback_fn<&xla_fallback,
ATEN_OP(_linalg_eigh)>::call(self, uplo,
compute_v);
}
XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
auto outputs = tensor_methods::eigh(self_tensor, uplo);
return std::make_tuple(bridge::AtenFromXlaTensor(std::get<0>(outputs)),
bridge::AtenFromXlaTensor(std::get<1>(outputs)));
}
std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor>
XLANativeFunctions::_linalg_slogdet(const at::Tensor& self) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
auto outputs = tensor_methods::slogdet(self_tensor);
return std::make_tuple(bridge::AtenFromXlaTensor(std::get<0>(outputs)),
bridge::AtenFromXlaTensor(std::get<1>(outputs)),
bridge::AtenFromXlaTensor(XLATensorPtr()),
bridge::AtenFromXlaTensor(XLATensorPtr()));
}
at::Tensor XLANativeFunctions::_log_softmax(const at::Tensor& self, int64_t dim,
bool half_to_float) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
auto self_meta = to_meta(self);
auto out_meta = at::meta::_log_softmax(self_meta, dim, half_to_float);
std::vector<torch::lazy::Shape> shapes{
torch::lazy::Shape(out_meta.scalar_type(), out_meta.sizes().vec())};
return bridge::AtenFromXlaTensor(tensor_methods::log_softmax(
bridge::GetXlaTensor(self), dim, std::nullopt, std::move(shapes)));
}
at::Tensor XLANativeFunctions::_log_softmax_backward_data(
const at::Tensor& grad_output, const at::Tensor& output, int64_t dim,
at::ScalarType /* input_dtype */) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return bridge::AtenFromXlaTensor(tensor_methods::log_softmax_backward(
bridge::GetXlaTensor(grad_output), bridge::GetXlaTensor(output), dim));
}
std::tuple<at::Tensor, at::Tensor> XLANativeFunctions::_pack_padded_sequence(
const at::Tensor& input, const at::Tensor& lengths, bool batch_first) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
std::vector<at::Tensor> xla_tensors = {lengths};
auto cpu_tensors = bridge::XlaCreateTensorList(xla_tensors);
return at::native::_pack_padded_sequence(input, cpu_tensors[0], batch_first);
}
at::Tensor XLANativeFunctions::_softmax(const at::Tensor& self, int64_t dim,
bool /* half_to_float */) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return bridge::AtenFromXlaTensor(
tensor_methods::softmax(bridge::GetXlaTensor(self), dim, std::nullopt));
}
at::Tensor XLANativeFunctions::_softmax_backward_data(
const at::Tensor& grad_output, const at::Tensor& output, int64_t dim,
at::ScalarType input_dtype) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return bridge::AtenFromXlaTensor(tensor_methods::softmax_backward(
bridge::GetXlaTensor(grad_output), bridge::GetXlaTensor(output), dim));
}
at::Tensor XLANativeFunctions::_unsafe_view(const at::Tensor& self,
at::IntArrayRef size) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return view_copy_symint(self, c10::fromIntArrayRefSlow(size));
}
at::Tensor XLANativeFunctions::add(const at::Tensor& self,
const at::Tensor& other,
const at::Scalar& alpha) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
// Currently, we disallow the case when both operands contain dynamic
// dimensions. This is consistent with PyTorch's behavior.
XLA_CHECK(!(tensor_has_dym_dim(self) && tensor_has_dym_dim(other)))
<< "Both operands of torch.add cannot have dynamic dimensions at the "
"same time. This is not "
"supported in PyTorch/XLA.";
at::native::alpha_check(at::result_type(self, other), alpha);
return DoBinaryOp(self, other,
[&](const XLATensorPtr& xself, const XLATensorPtr& xother,
at::ScalarType dtype) {
return tensor_methods::add(xself, xother, alpha, dtype);
});
}
at::Tensor XLANativeFunctions::add(const at::Tensor& self,
const at::Scalar& other,
const at::Scalar& alpha) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return DoBinaryOp(self, other,
[&](const XLATensorPtr& xself, const at::Scalar& other,
at::ScalarType dtype) {
return tensor_methods::add(xself, other, alpha, dtype);
});
}
at::Tensor XLANativeFunctions::addmm(const at::Tensor& self,
const at::Tensor& mat1,
const at::Tensor& mat2,
const at::Scalar& beta,
const at::Scalar& alpha) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
if (beta.to<double>() != 1 || alpha.to<double>() != 1) {
return at::native::call_fallback_fn<&xla_fallback, ATEN_OP(addmm)>::call(
self, mat1, mat2, beta, alpha);
}
return bridge::AtenFromXlaTensor(
tensor_methods::addmm(bridge::GetXlaTensor(mat1),
/*weight=*/bridge::GetXlaTensor(mat2),
/*bias=*/bridge::GetXlaTensor(self)));
}
at::Tensor XLANativeFunctions::alias_copy(const at::Tensor& self) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
return bridge::AtenFromXlaTensor(
tensor_methods::alias(bridge::GetXlaTensor(self)));
}
at::Tensor& XLANativeFunctions::arange_out(const at::Scalar& start,
const at::Scalar& end,
const at::Scalar& step,
at::Tensor& out) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
XLATensorPtr out_tensor = bridge::GetXlaTensor(out);
tensor_methods::arange_out(out_tensor, start, end, step, out.scalar_type());
return out;
}
at::Tensor XLANativeFunctions::as_strided_copy(
const at::Tensor& self, at::IntArrayRef size, at::IntArrayRef stride,
std::optional<int64_t> storage_offset) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
// Retrieve the base tensor, if there's one.
// This function actually operates on the tensor's storage. Since XLA does not
// expose the actual storage, we use the originally allocated tensor.
const at::Tensor& base = bridge::GetXlaTensor(self)->Base();
at::Tensor tensor = base.defined() ? base : self;
// Fast path: PyTorch/XLA implementation for as_strided works only with
// non-overlapping and dense tensors.
if (c10::_compute_non_overlapping_and_dense(size, stride)) {
// Sets the base tensor as tensor.
// Even though this function copies (without aliasing) tensor, it's still
// treated as a view function in the functionalization layer.
return bridge::AtenFromXlaTensor(bridge::SetBaseTensor(
tensor_methods::as_strided(bridge::GetXlaTensor(tensor),
XlaHelpers::I64List(size),
XlaHelpers::I64List(stride),
XlaHelpers::I64Optional(storage_offset)),
tensor));
}
// Slow path: decompose as_strided into indexing (we use take, though)
// operations. We pre-compute the index on CPU, so as to avoid runtime
// overhead.
auto dim = size.size();
auto itemsize = tensor.dtype().itemsize();
int64_t storage_size =
at::detail::computeStorageNbytes(size, stride, itemsize);
XLA_CHECK(tensor.numel() * itemsize >= storage_size)
<< "as_strided: storage not big enough for size " << size << ": "
<< storage_size << " (needed) vs " << tensor.numel() << " (actual).";
if (dim == 0 && tensor.numel() > 0) {
// If there's no specified dimension, return the first element of the
// storage. This behavior is consistent with eager.
return select_copy(view_copy_symint(tensor, {tensor.numel()}), 0, 0);
}
if (storage_size == 0) {
// Return an empty tensor, if no storage is actually needed.
return empty_symint(c10::fromIntArrayRefSlow(size), tensor.scalar_type(),
/* layout= */ std::nullopt, tensor.device(),
/* pin_memory= */ std::nullopt,
/* memory_format= */ std::nullopt);
}
// At this point, the following is true:
XLA_CHECK(storage_size > 0);
XLA_CHECK(tensor.numel() > 0);
XLA_CHECK(dim > 0);
// Index tensor for gathering the needed elements into contiguous data.
//
// PyTorch/XLA, by default, assumes dense and contiguous data. However, when
// specifying strides, that might not be the case.
//
// Therefore, we gather the elements selected by following the size, stride,
// and storage offset, materializing it into contiguous elements.
//
// In order to accomplish that, we create an index tensor. Specifically, we
// create an n-dimensional tensor (n is the number of dimensions of the
// output) of indices. Each element represent the at which position of the
// flattened tensor the desired element is in.
// Example: arange(13).as_strided((2, 2, 2), (3, 4, 5))
//
// Start with a 1-element n-dimensional tensor, initialized with 0:
//
// [[[0]]]
//
std::vector<int64_t> view_shape(dim, 1);
auto index_tensor =
at::tensor({storage_offset.value_or(self.storage_offset())},
at::TensorOptions().dtype(at::kLong))
.view(view_shape);
// Then, add to the index_tensor the offset value introduced for each possible
// index of that corresponding dimension.
//
// - Iteration i=0:
// [[[0]]] + [[[0 * 3]], [[1 * 3]]]
// = [[[0 * 3]], [[1 * 3]]]
// = [[[0]], [[3]]]
//
// - Iteration i=1:
// [[[0]], [[3]]] + [[[0 * 4], [1 * 4]]]
// = [[[0 + 0 * 4], [0 + 1 * 4]], [[3 + 0 * 4], [3 + 1 * 4]]]
// = [[[0], [4]], [[3], [7]]]
//
// - Iteration i=2:
// [[[0], [4]], [[3], [7]]] + [[[0 * 5, 1 * 5]]]
// =[[[0 + 0 * 5, 0 + 1 * 5], [4 + 0 * 5, 4 + 1 * 5]],
// [[3 + 0 * 5, 3 + 1 * 5], [7 + 0 * 5, 7 + 1 * 5]]]
// =[[[0, 5], [4, 9]], [[3, 8], [7, 12]]]
for (int i = 0; i < dim; i++) {
auto vshape = view_shape;
vshape[i] = size[i];
index_tensor =
index_tensor.add((at::arange(size[i]) * stride[i]).view(vshape));
}
// Finally, index the tensor with the computed indices.
return take(tensor, index_tensor.to(tensor.device()));
}
at::Tensor XLANativeFunctions::as_strided_scatter(
const at::Tensor& base, const at::Tensor& mutated_view,
at::IntArrayRef size, at::IntArrayRef stride,
std::optional<int64_t> storage_offset) {
TORCH_LAZY_FN_COUNTER_TIMED_TRACING("xla::");
auto base_ = bridge::GetXlaTensor(base);
auto xsize = XlaHelpers::I64List(size);
auto xstride = XlaHelpers::I64List(stride);
if (!AsStrided::StrideIsSupported(base_->shape(), xsize, xstride,
storage_offset.value_or(0))) {