forked from ROCm/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
_meta_registrations.py
6721 lines (5771 loc) · 202 KB
/
_meta_registrations.py
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
# mypy: allow-untyped-decorators
# mypy: allow-untyped-defs
import math
from enum import Enum
from functools import wraps
from typing import List, Optional, Sequence, Tuple, Union
import torch
import torch._prims_common as utils
from torch import SymBool, SymFloat, Tensor
from torch._decomp import (
_add_op_to_registry,
_convert_out_params,
global_decomposition_table,
meta_table,
)
from torch._ops import OpOverload
from torch._prims import _prim_elementwise_meta, ELEMENTWISE_PRIM_TYPE_PROMOTION_KIND
from torch._prims_common import (
BoolLike,
corresponding_complex_dtype,
corresponding_real_dtype,
elementwise_dtypes,
ELEMENTWISE_TYPE_PROMOTION_KIND,
FloatLike,
IntLike,
make_contiguous_strides_for,
Number,
TensorLike,
)
from torch._prims_common.wrappers import (
_maybe_convert_to_dtype,
_maybe_resize_out,
_resize_output_check,
_safe_copy_out,
out_wrapper,
)
from torch._refs import _broadcast_shapes, _maybe_broadcast
from torch.utils import _pytree as pytree
aten = torch.ops.aten
_meta_lib_dont_use_me_use_register_meta = torch.library.Library("aten", "IMPL", "Meta")
def register_meta(op):
def wrapper(fn):
fn = _convert_out_params(fn)
def register(op):
_add_op_to_registry(meta_table, op, fn)
pytree.tree_map_(register, op)
return fn
return wrapper
def elementwise_meta(
*args,
type_promotion: ELEMENTWISE_TYPE_PROMOTION_KIND,
):
# Perform type promotion, as this is expected from prim_metafunction
_, result_dtype = utils.elementwise_dtypes(
*args,
type_promotion_kind=type_promotion,
)
args = [_maybe_convert_to_dtype(x, result_dtype) for x in args]
# Broadcast
args = _maybe_broadcast(*args)
# Perform prim checks
return _prim_elementwise_meta(
*args, type_promotion=ELEMENTWISE_PRIM_TYPE_PROMOTION_KIND.DEFAULT
)
def toRealValueType(dtype):
from_complex = {
torch.complex32: torch.half,
torch.cfloat: torch.float,
torch.cdouble: torch.double,
}
return from_complex.get(dtype, dtype)
def check_inplace_broadcast(self_shape, *args_shape):
broadcasted_shape = tuple(_broadcast_shapes(self_shape, *args_shape))
torch._check(
broadcasted_shape == self_shape,
lambda: f"output with shape {self_shape} doesn't match the broadcast shape {broadcasted_shape}",
)
@register_meta([aten.linspace, aten.logspace])
@out_wrapper()
def meta_linspace_logspace(
start,
end,
steps,
base=None,
dtype=None,
device=None,
layout=torch.strided,
pin_memory=False,
requires_grad=False,
):
if isinstance(start, torch.Tensor):
torch._check(
start.dim() == 0,
lambda: "linspace only supports 0-dimensional start and end tensors",
)
if isinstance(end, torch.Tensor):
torch._check(
end.dim() == 0,
lambda: "linspace only supports 0-dimensional start and end tensors",
)
if any(isinstance(arg, complex) for arg in (start, end, steps)):
default_complex_dtype = utils.corresponding_complex_dtype(
torch.get_default_dtype()
)
if dtype is None:
dtype = default_complex_dtype
else:
torch._check(
utils.is_complex_dtype(dtype),
lambda: f"linspace(): inferred dtype {default_complex_dtype} can't be safely cast to passed dtype {dtype}",
)
else:
dtype = dtype or torch.get_default_dtype()
assert isinstance(dtype, torch.dtype)
# steps does not participate in the computation of the dtype
torch._check_type(
isinstance(steps, IntLike),
lambda: f"received an invalid combination of arguments - got \
({type(start).__name__}, {type(end).__name__}, {type(steps).__name__})",
)
assert isinstance(steps, IntLike) # for mypy
torch._check(steps >= 0, lambda: "number of steps must be non-negative")
return torch.empty(
(steps,), # type: ignore[arg-type]
dtype=dtype,
layout=layout,
device="meta",
pin_memory=pin_memory,
requires_grad=requires_grad,
)
@register_meta([aten.take.default, aten.take.out])
@out_wrapper()
def meta_take(self, index):
# Type and device checks
torch._check(
index.dtype == torch.long,
lambda: f"take(): Expected a long tensor for index, but got {index.dtype}",
)
# Index checks
torch._check_index(
not (self.numel() == 0 and index.numel() != 0),
lambda: "take(): tried to take from an empty tensor",
)
return self.new_empty(index.shape)
@register_meta([aten.linalg_cross.default, aten.linalg_cross.out])
@out_wrapper()
def linalg_cross(self, other, *, dim=-1):
x_d = self.ndim
y_d = other.ndim
torch._check(
x_d == y_d,
lambda: "linalg.cross: inputs must have the same number of dimensions.",
)
torch._check(
self.size(dim) == 3 and other.size(dim) == 3,
lambda: (
f"linalg.cross: inputs dimension {dim} must have length 3. "
f"Got {self.size(dim)} and {other.size(dim)}"
),
)
out_shape = _broadcast_shapes(self.shape, other.shape)
return self.new_empty(out_shape)
@register_meta(aten.linalg_matrix_exp)
@out_wrapper()
def linalg_matrix_exp(self):
squareCheckInputs(self, "linalg.matrix_exp")
checkFloatingOrComplex(self, "linalg.matrix_exp")
return torch.empty_like(self, memory_format=torch.contiguous_format)
@register_meta(
[aten.cummax.default, aten.cummax.out, aten.cummin.default, aten.cummin.out]
)
@out_wrapper("values", "indices")
def cummaxmin(self, dim):
values = torch.empty(self.shape, device=self.device, dtype=self.dtype)
indices = torch.empty(self.shape, device=self.device, dtype=torch.int64)
if self.numel() != 0 and self.ndim != 0:
# Checks that dim is within bounds
maybe_wrap_dim(dim, self.ndim)
return values, indices
@register_meta([aten.logcumsumexp.default, aten.logcumsumexp.out])
@out_wrapper()
def logcumsumexp(self, dim):
# Checks that dim is within bounds
maybe_wrap_dim(dim, self.ndim)
return torch.empty_like(self).contiguous()
# Stride-related code from _exec_fft in aten/src/ATen/native/cuda/SpectralOps.cpp
def _exec_fft(out, self, out_sizes, dim, forward):
ndim = self.ndim
signal_ndim = len(dim)
batch_dims = ndim - signal_ndim
# Permute dimensions so batch dimensions come first, and in stride order
dim_permute = list(range(ndim))
is_transformed_dim = [False for _ in range(ndim)]
for d in dim:
is_transformed_dim[d] = True
# std::partition
left, right = [], []
for d in dim_permute:
if not is_transformed_dim[d]:
left.append(d)
else:
right.append(d)
dim_permute = left + right
batch_end = len(left)
self_strides = self.stride()
tmp = dim_permute[:batch_end]
tmp.sort(key=lambda x: self_strides[x], reverse=True)
dim_permute = tmp + dim_permute[batch_end:]
input = self.permute(dim_permute)
# Collapse batch dimensions into a single dimension
batched_sizes = [-1] + list(input.shape[batch_dims:])
input = input.reshape(batched_sizes)
batch_size = input.size(0)
batched_sizes[0] = batch_size
batched_out_sizes = batched_sizes
for i in range(len(dim)):
batched_out_sizes[i + 1] = out_sizes[dim[i]]
out = out.reshape(batched_out_sizes)
# Reshaping to original batch shape and inverting the dimension permutation
out_strides = [0 for _ in range(ndim)]
batch_numel = 1
i = batch_dims - 1
while i >= 0:
out_strides[dim_permute[i]] = batch_numel * out.stride(0)
batch_numel *= out_sizes[dim_permute[i]]
i -= 1
for i in range(batch_dims, ndim):
out_strides[dim_permute[i]] = out.stride(1 + (i - batch_dims))
return out.as_strided(out_sizes, out_strides, out.storage_offset())
# See _fft_c2c_cufft in aten/src/ATen/native/cuda/SpectralOps.cpp
# and _fft_c2c_mkl in aten/src/ATen/native/mkl/SpectralOps.cpp
@register_meta([aten._fft_c2c.default, aten._fft_c2c.out])
@out_wrapper()
def meta_fft_c2c(self, dim, normalization, forward):
assert self.dtype.is_complex
out_sizes = self.shape
output = self.new_empty(out_sizes)
if not dim:
return output
sorted_dims = dim[:]
self_strides = self.stride()
sorted_dims.sort(key=lambda x: self_strides[x], reverse=True)
output = _exec_fft(output, self, out_sizes, sorted_dims, forward)
return output
@register_meta([aten._fft_r2c.default, aten._fft_r2c.out])
@out_wrapper()
def meta_fft_r2c(self, dim, normalization, onesided):
assert self.dtype.is_floating_point
output_sizes = list(self.size())
if onesided:
last_dim = dim[-1]
last_dim_halfsize = (output_sizes[last_dim] // 2) + 1
output_sizes[last_dim] = last_dim_halfsize
return self.new_empty(
output_sizes, dtype=utils.corresponding_complex_dtype(self.dtype)
)
@register_meta(aten.randperm.generator_out)
def meta_randperm(n, *, generator=None, out):
return _maybe_resize_out(out, torch.Size([n]))
@register_meta(aten.randperm.default)
def meta_randperm_default(
n,
*,
dtype=torch.long,
layout=None,
device=None,
pin_memory=None,
):
return torch.empty(
n, dtype=dtype, layout=layout, device=device, pin_memory=pin_memory
)
@register_meta([aten.randint.default, aten.randint.out])
@out_wrapper()
def meta_randint(
high,
size,
*,
dtype=torch.long,
layout=None,
device=None,
pin_memory=None,
):
return torch.empty(
size, dtype=dtype, layout=layout, device=device, pin_memory=pin_memory
)
@register_meta([aten.randint.low, aten.randint.low_out])
@out_wrapper()
def meta_randint_low(
low,
high,
size,
*,
dtype=torch.long,
layout=None,
device=None,
pin_memory=None,
):
return torch.empty(
size, dtype=dtype, layout=layout, device=device, pin_memory=pin_memory
)
@register_meta([aten.rand.default, aten.rand.out])
@out_wrapper()
def meta_rand_default(size, *, dtype=None, layout=None, device=None, pin_memory=None):
return torch.empty(
size, dtype=dtype, layout=layout, device=device, pin_memory=pin_memory
)
@register_meta([aten._fft_c2r.default, aten._fft_c2r.out])
@out_wrapper()
def meta_fft_c2r(self, dim, normalization, lastdim):
assert self.dtype.is_complex
output_sizes = list(self.size())
output_sizes[dim[-1]] = lastdim
return self.new_empty(output_sizes, dtype=toRealValueType(self.dtype))
@register_meta(aten.copy_.default)
def meta_copy_(self, src, non_blocking=False):
# This code simulates the original decomp from inductor,
# which runs most of the meta checks that we care about.
# In theory, we should make this more robust by carefully
# auditing our C++ copy_() kernel and copying the checks here.
from torch.fx.experimental.symbolic_shapes import free_unbacked_symbols
# TODO: Ideally, we'd insert a deferred runtime assert here, but if we are
# calling an actual copy_, you'll get that automatically
# https://github.com/pytorch/pytorch/issues/122477
if (
not free_unbacked_symbols(self) and torch._debug_has_internal_overlap(self) == 1
): # 1 == MemOverlap::Yes
raise RuntimeError(
"more than one element of the written-to tensor refers to a single memory location"
)
if isinstance(src, Tensor):
intermediate = src.to(self, non_blocking)
if self.size() != intermediate.size():
aten.expand_copy.default(intermediate, self.size())
return self
def inferUnsqueezeGeometry(tensor, dim):
result_sizes = list(tensor.size())
result_strides = list(tensor.stride())
new_stride = 1 if dim >= tensor.dim() else result_sizes[dim] * result_strides[dim]
result_sizes.insert(dim, 1)
result_strides.insert(dim, new_stride)
return result_sizes, result_strides
@register_meta(aten.unsqueeze_.default)
def meta_unsqueeze_(self, dim):
dim = maybe_wrap_dim(dim, self.dim() + 1)
g_sizes, g_strides = inferUnsqueezeGeometry(self, dim)
self.as_strided_(g_sizes, g_strides)
return self
@register_meta(aten._sparse_semi_structured_linear)
def meta_sparse_structured_linear(
input: Tensor,
weight: Tensor,
_meta: Tensor,
bias: Optional[Tensor] = None,
_activation_opt: Optional[str] = None,
out_dtype: Optional[torch.dtype] = None,
):
output_sizes = list(input.shape)
if bias is not None:
assert weight.size(0) == bias.size(0), "output size mismatch"
assert weight.size(1) == input.size(-1) / 2
output_sizes[-1] = weight.size(0)
# see: https://github.com/pytorch/pytorch/pull/114477#issuecomment-1830121375
# We assume that we have already squashed the inputs into a 2-D tensor
# Then, as the output is transposed, we need to propagate the transposed
# stride information to the output tensor
assert len(input.shape) == 2, "we can only handle the squashed input case"
transposed_strides = (1, input.size(0))
if out_dtype is not None:
assert (
input.dtype == torch.int8 and out_dtype == torch.int32
), "out_dtype is only supported for i8i8->i32 linear operator"
output = input.new_empty(
output_sizes,
dtype=input.dtype if out_dtype is None else out_dtype,
).as_strided(output_sizes, transposed_strides)
return output
@register_meta(aten._sparse_semi_structured_mm)
def meta_sparse_structured_mm(
mat1: Tensor,
mat1_meta: Tensor,
mat2: Tensor,
out_dtype: Optional[torch.dtype] = None,
):
assert len(mat1.shape) == 2
assert len(mat1_meta.shape) == 2
assert len(mat2.shape) == 2
assert mat1.size(1) == mat2.size(0) / 2
output_sizes = [mat1.size(0), mat2.size(1)]
if out_dtype is not None:
assert (
mat2.dtype == torch.int8 and out_dtype == torch.int32
), "out_dtype is only supported for i8i8->i32 linear operator"
output = mat2.new_empty(
output_sizes,
dtype=mat2.dtype if out_dtype is None else out_dtype,
)
return output
@register_meta(aten._sparse_semi_structured_addmm)
def meta_sparse_structured_addmm(
input: Tensor,
mat1: Tensor,
mat1_meta: Tensor,
mat2: Tensor,
*,
alpha=1,
beta=1,
out_dtype: Optional[torch.dtype] = None,
):
assert (
len(input.shape) == 1
), "only input broadcasted to columns of mat1 * mat2 product is supported"
assert len(mat1.shape) == 2
assert len(mat1_meta.shape) == 2
assert len(mat2.shape) == 2
assert input.size(0) == mat1.size(
0
), "only input broadcasted to columns of mat1 * mat2 product is supported"
assert mat1.size(1) == mat2.size(0) / 2
output_sizes = [mat1.size(0), mat2.size(1)]
if out_dtype is not None:
assert (
mat2.dtype == torch.int8 and out_dtype == torch.int32
), "out_dtype is only supported for i8i8->i32 linear operator"
output = mat2.new_empty(
output_sizes,
dtype=mat2.dtype if out_dtype is None else out_dtype,
)
return output
@register_meta(aten._cslt_sparse_mm)
def meta__cslt_sparse_mm(
compressed_A: torch.Tensor,
dense_B: torch.Tensor,
bias: Optional[Tensor] = None,
alpha: Optional[Tensor] = None,
out_dtype: Optional[torch.dtype] = None,
transpose_result: bool = False,
alg_id: int = 0,
split_k: int = 1,
split_k_one_kernel: bool = False,
):
assert dense_B.dtype in {
torch.float32,
torch.float16,
torch.bfloat16,
torch.int8,
torch.float8_e4m3fn,
}, "_cslt_sparse_mm only supports fp16, bf16, int8, and fp8e4m3"
assert compressed_A.dtype == dense_B.dtype, "inputs must have the same dtype"
assert len(dense_B.shape) == 2, "_cslt_sparse_mm only supports 2d inputs"
is_8bit_input_type = compressed_A.dtype in [torch.int8, torch.float8_e4m3fn]
compression_factor = 10 if is_8bit_input_type else 9
if is_8bit_input_type:
assert (
not dense_B.is_contiguous()
), "dense input must be transposed for 8bit dtypes"
k = dense_B.size(0)
n = dense_B.size(1)
m = (compressed_A.numel() * 16) // (compression_factor * k)
if bias is not None:
assert m == bias.size(0)
if out_dtype is not None:
assert (
is_8bit_input_type
and out_dtype
in {
torch.float16,
torch.bfloat16,
torch.int32,
torch.float8_e4m3fn,
}
), "out_dtype is not supported for {compressed_A.dtype} x {dense_B.dtype} -> {out_dtype} matmul!"
output_shape = (n, m) if transpose_result else (m, n)
return dense_B.new_empty(output_shape, dtype=out_dtype)
@register_meta(aten.index_reduce.default)
def meta_index_reduce(
self: Tensor,
dim: int,
index: Tensor,
source: torch.Tensor,
reduce: str,
*,
include_self: bool = True,
) -> Tensor:
return torch.empty_like(self, memory_format=torch.contiguous_format)
@register_meta(aten.index_reduce_.default)
def meta_index_reduce_(
self: Tensor,
dim: int,
index: Tensor,
source: torch.Tensor,
reduce: str,
*,
include_self: bool = True,
) -> Tensor:
return self
# Implementations below are taken from https://github.com/albanD/subclass_zoo/blob/main/python_meta_tensor.py
@out_wrapper()
@register_meta(aten.index_select.default)
def meta_index_select(self, dim, index):
result_size = list(self.size())
if self.dim() > 0:
result_size[dim] = index.numel()
return self.new_empty(result_size)
@register_meta(aten.segment_reduce.default)
def meta_segment_reduce(
data: Tensor,
reduce: str,
*,
lengths: Optional[Tensor] = None,
indices: Optional[Tensor] = None,
offsets: Optional[Tensor] = None,
axis: int = 0,
unsafe: bool = False,
initial=None,
) -> Tensor:
if indices is not None:
raise NotImplementedError(
"segment_reduce(): indices based reduction is not supported yet."
)
def segment_reduce_lengths_tensor(lengths_shape):
return torch.empty(
lengths_shape + data.shape[axis + 1 :],
dtype=data.dtype,
device="meta",
memory_format=torch.contiguous_format,
)
if lengths is not None:
return segment_reduce_lengths_tensor(lengths.shape)
# FIXME should probably check that lengths and offset aren't both set, but
# the ATen implementation neglects this too
if offsets is not None:
# lengths == torch.diff(offsets)
lengths_shape = offsets.shape[:-1] + (offsets.shape[-1] - 1,)
return segment_reduce_lengths_tensor(lengths_shape)
raise RuntimeError("segment_reduce(): Either lengths or offsets must be defined.")
@register_meta([aten.max.default, aten.max.unary_out])
@out_wrapper()
def meta_max(self):
return self.new_empty(())
@register_meta(aten.max.dim)
def meta_max_dim(self, dim, keepdim=False):
dim = utils.reduction_dims(self.shape, (dim,))
output_shape = _compute_reduction_shape(self, dim, keepdim)
return (
self.new_empty(output_shape),
self.new_empty(output_shape, dtype=torch.long),
)
@register_meta([aten.min.default, aten.min.unary_out])
@out_wrapper()
def meta_min(self):
return self.new_empty(())
@register_meta(aten.min.dim)
def meta_min_dim(self, dim, keepdim=False):
dim = utils.reduction_dims(self.shape, (dim,))
output_shape = _compute_reduction_shape(self, dim, keepdim)
return (
self.new_empty(output_shape),
self.new_empty(output_shape, dtype=torch.long),
)
@register_meta(aten.angle.default)
def meta_angle(self):
if self.is_complex():
result_dtype = corresponding_real_dtype(self.dtype)
else:
_, result_dtype = elementwise_dtypes(
self,
type_promotion_kind=ELEMENTWISE_TYPE_PROMOTION_KIND.INT_TO_FLOAT,
)
return torch.empty_like(self, dtype=result_dtype)
@register_meta(aten.angle.out)
def meta_angle_out(self, out):
torch._resize_output_(out, self.size(), self.device)
return out.copy_(torch.angle(self))
@register_meta(aten._assert_async.default)
def assert_async(val):
return
@register_meta(aten._assert_async.msg)
def assert_async_meta(val, assert_msg):
return
@register_meta(aten._print.default)
def print_meta(s):
return
@register_meta(aten._make_dep_token.default)
def make_dep_token(
*,
dtype=None,
layout=None,
device=None,
pin_memory=None,
memory_format=None,
):
return torch.empty(0, device="meta")
@register_meta(aten.sym_constrain_range.default)
def sym_constrain_range(size, min=None, max=None):
# Avoid importing sympy at a module level
from torch.fx.experimental.symbolic_shapes import constrain_range
if isinstance(size, (SymFloat, SymBool)):
raise ValueError("Constraining SymFloat or Symbool is nyi")
constrain_range(size, min=min, max=max)
@register_meta(aten._functional_sym_constrain_range.default)
def functional_sym_constrain_range(size, min=None, max=None, dep_token=None):
aten.sym_constrain_range(size, min=min, max=max)
return dep_token
@register_meta(aten.sym_constrain_range_for_size.default)
def sym_constrain_range_for_size(size, min=None, max=None):
# Avoid importing sympy at a module level
from torch.fx.experimental.symbolic_shapes import _constrain_range_for_size
if min is None and max is None:
torch._check_is_size(size)
return
if isinstance(size, (SymFloat, SymBool)):
raise ValueError("Constraining SymFloat or Symbool is nyi")
if type(size) is int:
if min is not None:
torch._check(size >= min)
if max is not None:
torch._check(size <= max)
return
_constrain_range_for_size(size, min=min, max=max)
@register_meta(aten._functional_sym_constrain_range_for_size.default)
def functional_sym_constrain_range_for_size(size, min, max, dep_token):
aten.sym_constrain_range_for_size(size, min=min, max=max)
return dep_token
@register_meta(aten._functional_assert_async.msg)
def functional_assert_async_meta(val, assert_msg, dep_token):
return dep_token
# From aten/src/ATen/native/LinearAlgebraUtils.h
def squareCheckInputs(self: Tensor, f_name: str):
assert (
self.dim() >= 2
), f"{f_name}: The input tensor must have at least 2 dimensions."
assert (
self.size(-1) == self.size(-2)
), f"{f_name}: A must be batches of square matrices, but they are {self.size(-2)} by {self.size(-1)} matrices"
# Validates input shapes and devices
# for linear solve methods (solve, cholesky_solve, lu_solve, triangular_solve)
# From aten/src/ATen/native/LinearAlgebraUtils.h
def linearSolveCheckInputs(self: Tensor, A: Tensor, name: str):
torch._check(
self.device == A.device,
lambda: (
f"Expected b and A to be on the same device, but found b on "
f"{self.device} and A on {A.device} instead."
),
)
torch._check(
self.dtype == A.dtype,
lambda: (
f"Expected b and A to have the same dtype, but found b of type "
f"{self.dtype} and A of type {A.dtype} instead."
),
)
torch._check(
A.size(-1) == A.size(-2),
lambda: (
f"A must be batches of square matrices, "
f"but they are {A.size(-2)} by {A.size(-1)} matrices"
),
)
torch._check(
A.size(-1) == self.size(-2),
lambda: (
f"Incompatible matrix sizes for {name}: each A "
f"matrix is {A.size(-1)} by {A.size(-1)}"
f" but each b matrix is {self.size(-2)} by {self.size(-1)}"
),
)
# From aten/src/ATen/native/LinearAlgebraUtils.h
def checkFloatingOrComplex(
t: Tensor,
f_name: str,
allow_low_precision_dtypes: bool = True,
):
dtype = t.dtype
torch._check(
t.is_floating_point() or t.is_complex(),
lambda: f"{f_name}: Expected a floating point or complex tensor as input. Got {dtype}",
)
if not allow_low_precision_dtypes:
torch._check(
dtype in (torch.float, torch.double, torch.cfloat, torch.cdouble),
lambda: f"{f_name}: Low precision dtypes not supported. Got {dtype}",
)
# From aten/src/ATen/native/LinearAlgebraUtils.h
def checkIsMatrix(A: Tensor, f_name: str, arg_name: str = "A"):
torch._check(
A.dim() >= 2,
lambda: f"{f_name}: The input tensor {arg_name} must have at least 2 dimensions.",
)
def checkInputsSolver(A: Tensor, B: Tensor, left: bool, f_name: str):
squareCheckInputs(A, f_name)
checkIsMatrix(B, f_name)
torch._check(
A.size(-2) == B.size(-2) if left else A.size(-1) == B.size(-1),
lambda: (
f"{f_name}: Incompatible shapes of A and B for the equation "
f"{'AX = B' if left else 'XA = B'}"
f" ({A.size(-2)}x{A.size(-1)} and {B.size(-2)}x{B.size(-1)})"
),
)
def checkSameDevice(
fn_name: str,
result: Tensor,
input: Tensor,
result_name: str = "result",
):
torch._check(
result.device == input.device,
lambda: (
f"{fn_name}: Expected {result_name} and input tensors to be on the same device, but got "
f"{result_name} on {result.device} and input on {input.device}"
),
)
def checkUplo(UPLO: str):
UPLO_uppercase = UPLO.upper()
torch._check(
len(UPLO) == 1 and (UPLO_uppercase == "U" or UPLO_uppercase == "L"),
lambda: f"Expected UPLO argument to be 'L' or 'U', but got {UPLO}",
)
@register_meta([aten._linalg_eigh.default, aten._linalg_eigh.eigenvalues])
@out_wrapper("eigenvalues", "eigenvectors")
def meta__linalg_eigh(A: Tensor, UPLO: str = "L", compute_v: bool = True):
squareCheckInputs(A, "linalg.eigh")
checkUplo(UPLO)
shape = list(A.shape)
if compute_v:
vecs = A.new_empty(shape)
vecs.as_strided_(shape, make_contiguous_strides_for(shape, row_major=False))
else:
vecs = A.new_empty([0])
shape.pop()
vals = A.new_empty(shape, dtype=toRealValueType(A.dtype))
return vals, vecs
@register_meta([aten._linalg_eigvals.default, aten.linalg_eigvals.out])
@out_wrapper()
def meta__linalg_eigvals(input: Tensor) -> Tensor:
squareCheckInputs(input, "linalg.eigvals")
complex_dtype = (
input.dtype
if utils.is_complex_dtype(input.dtype)
else utils.corresponding_complex_dtype(input.dtype)
)
return input.new_empty(input.shape[:-1], dtype=complex_dtype)
@register_meta([aten.linalg_eig])
@out_wrapper("eigenvalues", "eigenvectors")
def meta_linalg_eig(input: Tensor):
squareCheckInputs(input, "linalg.eig")
complex_dtype = (
input.dtype
if utils.is_complex_dtype(input.dtype)
else utils.corresponding_complex_dtype(input.dtype)
)
values = input.new_empty(input.shape[:-1], dtype=complex_dtype)
vectors = input.new_empty(input.shape, dtype=complex_dtype)
return values, vectors
def cloneBatchedColumnMajor(src: Tensor) -> Tensor:
return src.mT.clone(memory_format=torch.contiguous_format).transpose(-2, -1)
@register_meta(aten._cholesky_solve_helper)
@out_wrapper()
def _cholesky_solve_helper(self: Tensor, A: Tensor, upper: bool) -> Tensor:
return cloneBatchedColumnMajor(self)
@register_meta(aten.cholesky_solve)
@out_wrapper()
def cholesky_solve(self: Tensor, A: Tensor, upper: bool = False) -> Tensor:
torch._check(
self.ndim >= 2,
lambda: f"b should have at least 2 dimensions, but has {self.ndim} dimensions instead",
)
torch._check(
A.ndim >= 2,
lambda: f"u should have at least 2 dimensions, but has {A.ndim} dimensions instead",
)
self_broadcasted, A_broadcasted = _linalg_broadcast_batch_dims_name(
self, A, "cholesky_solve"
)
return _cholesky_solve_helper(self_broadcasted, A_broadcasted, upper)
@register_meta(aten.cholesky)
@out_wrapper()
def cholesky(self: Tensor, upper: bool = False) -> Tensor:
if self.numel() == 0:
return torch.empty_like(self, memory_format=torch.legacy_contiguous_format)
squareCheckInputs(self, "cholesky")
return cloneBatchedColumnMajor(self)
@register_meta(aten.cholesky_inverse)
@out_wrapper()
def cholesky_inverse(self: Tensor, upper: bool = False) -> Tensor:
squareCheckInputs(self, "cholesky_inverse")
return cloneBatchedColumnMajor(self)
# From aten/src/ATen/native/BatchLinearAlgebra.cpp
@register_meta(aten.linalg_cholesky_ex.default)
def linalg_cholesky_ex(A: Tensor, upper: bool = False, check_errors: bool = False):
squareCheckInputs(A, "linalg.cholesky")
checkFloatingOrComplex(A, "linalg.cholesky")
A_shape = A.shape
ndim = len(A_shape)
# L
L_strides = make_contiguous_strides_for(A_shape, False)
L = A.new_empty(A_shape)
L.as_strided_(A_shape, L_strides)
# infos
infos = A.new_empty(A_shape[0 : ndim - 2], dtype=torch.int32)
return L, infos
@register_meta(
[aten.linalg_householder_product.default, aten.linalg_householder_product.out]
)
@out_wrapper()
def linalg_householder_product(input: Tensor, tau: Tensor) -> Tensor:
torch._check(
input.ndim >= 2,
lambda: "torch.linalg.householder_product: input must have at least 2 dimensions.",
)
torch._check(
input.size(-2) >= input.size(-1),
lambda: "torch.linalg.householder_product: input.shape[-2] must be greater than or equal to input.shape[-1]",
)
torch._check(
input.size(-1) >= tau.size(-1),
lambda: "torch.linalg.householder_product: input.shape[-1] must be greater than or equal to tau.shape[-1]",
)
torch._check(
input.ndim - tau.ndim == 1,
lambda: (
f"torch.linalg.householder_product: Expected tau to have one dimension less than input, "