File: test_functionalization.py

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (1137 lines) | stat: -rw-r--r-- 47,610 bytes parent folder | download
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
# Owner(s): ["module: codegen"]

import torch
from torch.testing._internal.common_utils import TestCase, run_tests, skipIfTorchDynamo, TEST_WITH_TORCHDYNAMO
from torch.testing._internal.logging_tensor import LoggingTensor, capture_logs
from torch.utils._pytree import tree_map
from torch.fx.experimental.proxy_tensor import make_fx
from torch.fx.passes.reinplace import reinplace

import unittest

def are_aliased(x, y):
    if x._base is None and y._base is None:
        return False
    if x._base is not None and y._base is None:
        return x._base is y
    if x._base is None and y._base is not None:
        return y._base is x
    return x._base is y._base

# We can unify testing and use functionalize() here instead
# if/when functorch moves into core.
# This is basically a crappy version of `functionalize()` for single-tensor-arg inputs.
def _functionalize(f, *, reapply_views: bool):
    def wrapped(a):
        input_functional = torch._to_functional_tensor(a)
        torch._enable_functionalization(reapply_views=reapply_views)
        try:
            out = f(input_functional)
        finally:
            torch._disable_functionalization()
        torch._sync(input_functional)
        inpt_new = torch._from_functional_tensor(input_functional)
        if inpt_new is not a:
            # Existing deficiency in functionalize():
            # we don't correctly mutate input metadata (yet?)
            if inpt_new.shape == a.shape:
                a.copy_(inpt_new)
        tree_map(torch._sync, out)
        out_unwrapped = tree_map(torch._from_functional_tensor, out)
        return out_unwrapped

    return wrapped

@unittest.skipIf(TEST_WITH_TORCHDYNAMO, "https://github.com/pytorch/pytorch/issues/81457")
class TestFunctionalization(TestCase):

    def get_logs(self, func, inpt, *, reapply_views=False, run_reinplace=False):
        inpt_clone = inpt.clone()
        traced_f = make_fx(_functionalize(func, reapply_views=reapply_views))(inpt)
        if run_reinplace:
            traced_f = reinplace(traced_f, inpt_clone)
        return traced_f.code

    def assert_functionalization(self, func, inpt, *, reapply_views=False, mutated_input_metadata=False):
        input_clone = inpt.clone()
        input_clone2 = inpt.clone()
        input_clone3 = inpt.clone()

        # Compare outputs (and mutated inputs), with and without functionalization.
        out_ref = func(inpt)
        out_functional = _functionalize(func, reapply_views=reapply_views)(input_clone)
        # The reinplacing pass is only valid to run with reapply_views=True.
        functional_func = make_fx(_functionalize(func, reapply_views=True))(input_clone2)
        reinplace_func = reinplace(make_fx(_functionalize(func, reapply_views=True))(input_clone2), input_clone2)

        # NOTE: for now, need to pass in fresh inputs here, because make_fx
        # will directly mutate the inputs that you trace with.
        # Once this is fixed we can clean this up.
        out_reinplace = reinplace_func(input_clone3)

        # functionalize() deficiency: input metadata mutations aren't propagated properly,
        # so we just need to skip checks here for the tests that exercise that.
        if not mutated_input_metadata:
            self.assertEqual(inpt, input_clone)  # input mutations should still occur
            self.assertEqual(inpt, input_clone3)

        # Handle tests with multi-tensor outputs
        if isinstance(out_ref, tuple):
            out_refs, out_functionals, out_reinplaces = list(out_ref), list(out_functional), list(out_reinplace)
        else:
            out_refs, out_functionals, out_reinplaces = [out_ref], [out_functional], [out_reinplace]

        for out_ref_, out_functional_, out_reinplace_ in zip(out_refs, out_functionals, out_reinplaces):
            self.assertEqual(out_ref_, out_functional_)
            self.assertEqual(out_ref_, out_reinplace_)

    def test_save_for_backwards_segfault(self):
        inp = torch._to_functional_tensor(LoggingTensor(torch.randn(2, 2))).requires_grad_(True)
        inp.exp()

    def test_multiple_views_of_same_base(self):
        def f(x):
            y = x.view(-1)
            z = x.view(-1)
            x.add_(1)
            # y should have been updated.
            y2 = y + 1
            # z should have been updated too.
            z2 = z + 1
            return z2
        self.assert_functionalization(f, torch.ones(4))

    def test_simple(self):
        def f(x):
            # simple test: 1 view op, 1 inplace op
            tmp = torch.ones(4, 2)
            y = x.view(4, 2)
            y.add_(tmp)
            z = x * x
            return y
        self.assert_functionalization(f, torch.ones(4, 2))
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4, 2], device = device(type='cpu'), pin_memory = False)
    view_copy = torch.ops.aten.view_copy.default(a_1, [4, 2])
    add = torch.ops.aten.add.Tensor(view_copy, ones);  view_copy = ones = None
    view_copy_1 = torch.ops.aten.view_copy.default(add, [4, 2])
    mul = torch.ops.aten.mul.Tensor(view_copy_1, view_copy_1)
    copy_ = torch.ops.aten.copy_.default(a_1, view_copy_1);  a_1 = view_copy_1 = None
    return add
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(4, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4, 2], device = device(type='cpu'), pin_memory = False)
    view = torch.ops.aten.view.default(a_1, [4, 2])
    add = torch.ops.aten.add.Tensor(view, ones);  view = ones = None
    view_1 = torch.ops.aten.view.default(add, [4, 2])
    mul = torch.ops.aten.mul.Tensor(view_1, view_1)
    copy_ = torch.ops.aten.copy_.default(a_1, view_1);  a_1 = view_1 = None
    return add
    """)

    def test_simple_out(self):
        def f(x):
            tmp = torch.ones(4, 2)
            y = x.view(4, 2)
            # the out= tensor will get resized, since it has size=0 to start.
            z = torch.empty(())
            torch.add(y, tmp, out=z)
            w = z * z
            return w
        self.assert_functionalization(f, torch.ones(4, 2))
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4, 2], device = device(type='cpu'), pin_memory = False)
    view_copy = torch.ops.aten.view_copy.default(a_1, [4, 2]);  a_1 = None
    empty = torch.ops.aten.empty.memory_format([], device = device(type='cpu'), pin_memory = False)
    add = torch.ops.aten.add.Tensor(view_copy, ones);  view_copy = ones = None
    mul = torch.ops.aten.mul.Tensor(add, add);  add = None
    return mul
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(4, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4, 2], device = device(type='cpu'), pin_memory = False)
    view = torch.ops.aten.view.default(a_1, [4, 2]);  a_1 = None
    empty = torch.ops.aten.empty.memory_format([], device = device(type='cpu'), pin_memory = False)
    add = torch.ops.aten.add.Tensor(view, ones);  view = ones = None
    mul = torch.ops.aten.mul.Tensor(add, add);  add = None
    return mul
    """)

    def test_multi_out(self):
        def f(x):
            # aminmax.out returns a tuple of tensors.
            # functionalization should properly handle the tuple.
            out_min = torch.empty(4)
            out_max = torch.empty(4)
            torch.aminmax(x, dim=0, out=(out_max, out_min))
            return out_max
        self.assert_functionalization(f, torch.arange(8, dtype=torch.float32))
        logs = self.get_logs(f, torch.arange(8, dtype=torch.float32))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    empty = torch.ops.aten.empty.memory_format([4], device = device(type='cpu'), pin_memory = False)
    empty_1 = torch.ops.aten.empty.memory_format([4], device = device(type='cpu'), pin_memory = False)
    aminmax = torch.ops.aten.aminmax.default(a_1, dim = 0);  a_1 = None
    getitem = aminmax[0]
    getitem_1 = aminmax[1];  aminmax = None
    return getitem
    """)

        reinplaced_logs = self.get_logs(f, torch.arange(8, dtype=torch.float32), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    empty = torch.ops.aten.empty.memory_format([4], device = device(type='cpu'), pin_memory = False)
    empty_1 = torch.ops.aten.empty.memory_format([4], device = device(type='cpu'), pin_memory = False)
    aminmax = torch.ops.aten.aminmax.default(a_1, dim = 0);  a_1 = None
    getitem = aminmax[0]
    getitem_1 = aminmax[1];  aminmax = None
    return getitem
    """)

    def test_tensor_ctr(self):
        def f(x):
            y = torch.tensor((1, 2, 3))
            z = y.view(-1)
            z.add_(1)
            return y

        inpt = torch.arange(3, dtype=torch.float32)
        self.assert_functionalization(f, inpt)

        logs = self.get_logs(f, inpt)
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    _tensor_constant0 = self._tensor_constant0
    lift_fresh_copy = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    view_copy = torch.ops.aten.view_copy.default(lift_fresh_copy, [-1]);  lift_fresh_copy = None
    add = torch.ops.aten.add.Tensor(view_copy, 1);  view_copy = None
    view_copy_1 = torch.ops.aten.view_copy.default(add, [3]);  add = None
    return view_copy_1
    """)

        reinplaced_logs = self.get_logs(f, inpt, reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    _tensor_constant0 = self._tensor_constant0
    lift_fresh_copy = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    view = torch.ops.aten.view.default(lift_fresh_copy, [-1]);  lift_fresh_copy = None
    add = torch.ops.aten.add_.Tensor(view, 1)
    view_1 = torch.ops.aten.view.default(view, [3]);  view = None
    return view_1
    """)


    def test_tensor_list_mixed_functional_nonfunctional(self):
        nonfunctional_tensor = torch.ones(2, dtype=torch.long)

        def f(x):
            # simple test: 1 view op, 1 inplace op
            functional_tensor = torch.ones(2, dtype=torch.long)
            out = x[functional_tensor, nonfunctional_tensor]
            return out
        out = f(torch.ones(2, 2))
        out_functional = _functionalize(f, reapply_views=True)(torch.ones(2, 2))
        self.assertEqual(out, out_functional)

    def test_inplace_on_non_view(self):
        def f(x):
            # test for the case where we functionalize an inplace op on the other tensor - not a view.
            # This is worth checking because the tensor will have an empty ViewMeta stack, which needs to be special cased.
            tmp = torch.ones(4, 2)
            y = x.view(4, 2)
            x.add_(tmp)
            return y
        self.assert_functionalization(f, torch.ones(4, 2))
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4, 2], device = device(type='cpu'), pin_memory = False)
    view_copy = torch.ops.aten.view_copy.default(a_1, [4, 2])
    add = torch.ops.aten.add.Tensor(a_1, ones);  ones = None
    copy_ = torch.ops.aten.copy_.default(a_1, add);  a_1 = None
    view_copy_1 = torch.ops.aten.view_copy.default(add, [4, 2]);  add = None
    return view_copy_1
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(4, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4, 2], device = device(type='cpu'), pin_memory = False)
    view = torch.ops.aten.view.default(a_1, [4, 2])
    add = torch.ops.aten.add.Tensor(a_1, ones);  ones = None
    copy_ = torch.ops.aten.copy_.default(a_1, add);  a_1 = None
    view_1 = torch.ops.aten.view.default(add, [4, 2]);  add = None
    return view_1
    """)

    # Some ops that are mutable are neither inplace nor out= ops.
    # They also need special handling.
    def test_mutable_op_not_inplace_or_other(self):
        def f(x):
            return torch._fused_moving_avg_obs_fq_helper(x, x, x, x, x, x, x, 1.0, 0, 1, 0)

        logs = self.get_logs(f, torch.ones(1))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    _fused_moving_avg_obs_fq_helper_functional = torch.ops.aten._fused_moving_avg_obs_fq_helper_functional.default(a_1, a_1, a_1, a_1, a_1, a_1, a_1, 1.0, 0, 1, 0)
    getitem = _fused_moving_avg_obs_fq_helper_functional[0]
    getitem_1 = _fused_moving_avg_obs_fq_helper_functional[1]
    getitem_2 = _fused_moving_avg_obs_fq_helper_functional[2]
    getitem_3 = _fused_moving_avg_obs_fq_helper_functional[3]
    getitem_4 = _fused_moving_avg_obs_fq_helper_functional[4]
    getitem_5 = _fused_moving_avg_obs_fq_helper_functional[5];  _fused_moving_avg_obs_fq_helper_functional = None
    copy_ = torch.ops.aten.copy_.default(a_1, getitem_5);  a_1 = getitem_5 = None
    return (getitem, getitem_1)
    """)  # noqa: B950

    def test_as_strided(self):
        def f(x):
            y = x.as_strided((2,), (2,), 1)
            y.add_(1)
            return x
        self.assert_functionalization(f, torch.ones(9))
        logs = self.get_logs(f, torch.ones(9))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    as_strided_copy = torch.ops.aten.as_strided_copy.default(a_1, [2], [2], 1)
    add = torch.ops.aten.add.Tensor(as_strided_copy, 1);  as_strided_copy = None
    as_strided_scatter = torch.ops.aten.as_strided_scatter.default(a_1, add, [2], [2], 1);  add = None
    copy_ = torch.ops.aten.copy_.default(a_1, as_strided_scatter);  a_1 = None
    return as_strided_scatter
    """)

    def test_tensor_list_composite(self):
        def f(x):
            # Test an op with TensorList input
            y = torch.block_diag(x, x)
            return y
        self.assert_functionalization(f, torch.ones(2, 2))
        logs = self.get_logs(f, torch.ones(2, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    block_diag = torch.ops.aten.block_diag.default([a_1, a_1]);  a_1 = None
    return block_diag
    """)

    def test_cat(self):
        def f(x):
            out = torch.empty(0)
            torch.cat((x,), out=out)
            return out
        self.assert_functionalization(f, torch.ones(2, 2))
        logs = self.get_logs(f, torch.ones(2, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    empty = torch.ops.aten.empty.memory_format([0], device = device(type='cpu'), pin_memory = False)
    cat = torch.ops.aten.cat.default([a_1]);  a_1 = None
    return cat
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(2, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    empty = torch.ops.aten.empty.memory_format([0], device = device(type='cpu'), pin_memory = False)
    cat = torch.ops.aten.cat.default([a_1]);  a_1 = None
    return cat
    """)


    def test_diagonal(self):
        def f(x):
            # test: view ops that take a subset of the original tensor (select/diagonal)
            tmp = torch.ones(2)
            y = x.clone().diagonal()
            y.add_(tmp)
            z = x * x
            return z
        self.assert_functionalization(f, torch.ones(2, 2))
        logs = self.get_logs(f, torch.ones(2, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([2], device = device(type='cpu'), pin_memory = False)
    clone = torch.ops.aten.clone.default(a_1)
    diagonal_copy = torch.ops.aten.diagonal_copy.default(clone);  clone = None
    add = torch.ops.aten.add.Tensor(diagonal_copy, ones);  diagonal_copy = ones = None
    mul = torch.ops.aten.mul.Tensor(a_1, a_1);  a_1 = None
    return mul
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(2, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([2], device = device(type='cpu'), pin_memory = False)
    clone = torch.ops.aten.clone.default(a_1)
    diagonal = torch.ops.aten.diagonal.default(clone);  clone = None
    add = torch.ops.aten.add_.Tensor(diagonal, ones);  diagonal = ones = None
    mul = torch.ops.aten.mul.Tensor(a_1, a_1);  a_1 = None
    return mul
    """)

    def test_diagonal_mutated_input(self):
        def f(x):
            # simple test: there are pending updates afterwards, which the test syncs manually
            tmp = torch.ones(2)
            y = x.diagonal()
            y.add_(tmp)
            return x
        x = torch.ones(2, 2)
        self.assert_functionalization(f, x)
        logs = self.get_logs(f, torch.ones(2, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([2], device = device(type='cpu'), pin_memory = False)
    diagonal_copy = torch.ops.aten.diagonal_copy.default(a_1)
    add = torch.ops.aten.add.Tensor(diagonal_copy, ones);  diagonal_copy = ones = None
    diagonal_scatter = torch.ops.aten.diagonal_scatter.default(a_1, add);  add = None
    copy_ = torch.ops.aten.copy_.default(a_1, diagonal_scatter);  a_1 = None
    return diagonal_scatter
    """)

    def test_split(self):
        def f(x):
            # test: view ops that return multiple tensors (split)
            tmp = torch.ones(2)
            y1, y2 = x.split(2)
            y3 = y2.diagonal()
            y3.add_(tmp)
            z = x * x
            return y3
        self.assert_functionalization(f, torch.ones(4, 2))
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([2], device = device(type='cpu'), pin_memory = False)
    split_copy = torch.ops.aten.split_copy.Tensor(a_1, 2)
    getitem = split_copy[0]
    getitem_1 = split_copy[1];  split_copy = None
    diagonal_copy = torch.ops.aten.diagonal_copy.default(getitem_1);  getitem_1 = None
    add = torch.ops.aten.add.Tensor(diagonal_copy, ones);  diagonal_copy = ones = None
    split_copy_1 = torch.ops.aten.split_copy.Tensor(a_1, 2)
    getitem_2 = split_copy_1[0]
    getitem_3 = split_copy_1[1];  split_copy_1 = None
    diagonal_scatter = torch.ops.aten.diagonal_scatter.default(getitem_3, add);  getitem_3 = None
    slice_scatter = torch.ops.aten.slice_scatter.default(a_1, diagonal_scatter, 0, 2, 4);  diagonal_scatter = None
    mul = torch.ops.aten.mul.Tensor(slice_scatter, slice_scatter)
    copy_ = torch.ops.aten.copy_.default(a_1, slice_scatter);  a_1 = slice_scatter = None
    return add
    """)  # noqa: B950

    def test_view_inplace(self):
        def f(x):
            # test: view + inplace op (transpose_)
            tmp = torch.ones(4)
            x.transpose_(1, 0)
            y = x[0]
            y.add_(tmp)
            return x
        self.assert_functionalization(f, torch.ones(4, 2), mutated_input_metadata=True)
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4], device = device(type='cpu'), pin_memory = False)
    transpose_copy = torch.ops.aten.transpose_copy.int(a_1, 1, 0)
    select_copy = torch.ops.aten.select_copy.int(transpose_copy, 0, 0);  transpose_copy = None
    add = torch.ops.aten.add.Tensor(select_copy, ones);  select_copy = ones = None
    transpose_copy_1 = torch.ops.aten.transpose_copy.int(a_1, 1, 0);  a_1 = None
    select_scatter = torch.ops.aten.select_scatter.default(transpose_copy_1, add, 0, 0);  transpose_copy_1 = add = None
    transpose_copy_2 = torch.ops.aten.transpose_copy.int(select_scatter, 1, 0);  select_scatter = None
    transpose_copy_3 = torch.ops.aten.transpose_copy.int(transpose_copy_2, 1, 0);  transpose_copy_2 = None
    return transpose_copy_3
    """)  # noqa: B950

    def test_optional_tensor_list(self):
        def f(x):
            # test: an operator that takes in a List[Optional[Tensor]] argument
            # (index_put)
            y = x.view(8)
            indices = torch.arange(4)
            values = torch.arange(4, dtype=y.dtype)
            y.index_put_((indices,), values, accumulate=False)
            return y
        self.assert_functionalization(f, torch.ones(4, 2))
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    view_copy = torch.ops.aten.view_copy.default(a_1, [8])
    arange = torch.ops.aten.arange.default(4, device = device(type='cpu'), pin_memory = False)
    arange_1 = torch.ops.aten.arange.default(4, dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
    index_put = torch.ops.aten.index_put.default(view_copy, [arange], arange_1);  view_copy = arange = arange_1 = None
    view_copy_1 = torch.ops.aten.view_copy.default(index_put, [4, 2])
    copy_ = torch.ops.aten.copy_.default(a_1, view_copy_1);  a_1 = view_copy_1 = None
    return index_put
    """)  # noqa: B950

    def test_scalars(self):
        def f(x):
            # test: the pass can handle scalar inputs properly
            tmp = torch.ones(4, 2)
            y = x.view(4, 2)
            y.add_(1)
            z = 2 * y
            z.div_(1)
            return z
        self.assert_functionalization(f, torch.ones(4, 2))
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4, 2], device = device(type='cpu'), pin_memory = False)
    view_copy = torch.ops.aten.view_copy.default(a_1, [4, 2])
    add = torch.ops.aten.add.Tensor(view_copy, 1);  view_copy = None
    mul = torch.ops.aten.mul.Tensor(add, 2)
    div = torch.ops.aten.div.Tensor(mul, 1);  mul = None
    view_copy_1 = torch.ops.aten.view_copy.default(add, [4, 2]);  add = None
    copy_ = torch.ops.aten.copy_.default(a_1, view_copy_1);  a_1 = view_copy_1 = None
    return div
    """)

    @skipIfTorchDynamo("Test does not work with TorchDynamo")
    def test_metadata_change(self):
        def f(x):
            # ops like ge_() are allowed to change the dtype of the input.
            # functionalization should pick up on that.
            y = x.clone()
            out = y.ge_(0)
            return out
        self.assert_functionalization(f, torch.ones(4, 2))
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    clone = torch.ops.aten.clone.default(a_1);  a_1 = None
    ge = torch.ops.aten.ge.Scalar(clone, 0);  clone = None
    _to_copy = torch.ops.aten._to_copy.default(ge, dtype = torch.float32, layout = torch.strided);  ge = None
    return _to_copy
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(2, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    clone = torch.ops.aten.clone.default(a_1);  a_1 = None
    ge = torch.ops.aten.ge.Scalar(clone, 0);  clone = None
    _to_copy = torch.ops.aten._to_copy.default(ge, dtype = torch.float32, layout = torch.strided);  ge = None
    return _to_copy
    """)  # noqa: B950

    @skipIfTorchDynamo("Test does not work with TorchDynamo")
    def test_metadata_change_out_op(self):
        def f(t, y):
            out_1 = torch.ones(1)
            return torch.add(t, y, out=out_1)

        inpt1, inpt2 = torch.tensor([1]), torch.tensor([1])
        inpt1_func, inpt2_func = torch._to_functional_tensor(inpt1), torch._to_functional_tensor(inpt2)

        out_ref = f(inpt1, inpt2)
        torch._enable_functionalization(reapply_views=True)
        try:
            out_functional = f(inpt1_func, inpt2_func)
        finally:
            torch._disable_functionalization()
        self.assertEqual(out_ref, torch._from_functional_tensor(out_functional))


    def test_only_one_view(self):
        def f(x):
            # This tests that we don't have any unnecessary views in the trace.
            # If the input wasn't mutated, we don't need to regenerate it,
            # so there should be a total of 1 op in the output trace.
            return x.view(4, 2)
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    view_copy = torch.ops.aten.view_copy.default(a_1, [4, 2]);  a_1 = None
    return view_copy
    """)

    def test_everything(self):
        def f(x):
            # test: everything
            tmp = torch.ones(2, 2)
            x2 = x + x
            y = x2.view(8)
            z0 = y.reshape(2, 4)
            z1 = z0.transpose(1, 0)
            z1.unsqueeze_(0)
            z1.squeeze_()
            z2, z3 = z1.split(2)
            z2.add_(tmp)
            z4 = z0[0] + z2.reshape(4)
            return z2
        self.assert_functionalization(f, torch.ones(4, 2))
        logs = self.get_logs(f, torch.ones(4, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([2, 2], device = device(type='cpu'), pin_memory = False)
    add = torch.ops.aten.add.Tensor(a_1, a_1);  a_1 = None
    view_copy = torch.ops.aten.view_copy.default(add, [8])
    _reshape_alias_copy = torch.ops.aten._reshape_alias_copy.default(view_copy, [2, 4], [4, 1]);  view_copy = None
    transpose_copy = torch.ops.aten.transpose_copy.int(_reshape_alias_copy, 1, 0)
    unsqueeze_copy = torch.ops.aten.unsqueeze_copy.default(transpose_copy, 0);  transpose_copy = None
    squeeze_copy = torch.ops.aten.squeeze_copy.default(unsqueeze_copy);  unsqueeze_copy = None
    split_copy = torch.ops.aten.split_copy.Tensor(squeeze_copy, 2);  squeeze_copy = None
    getitem = split_copy[0]
    getitem_1 = split_copy[1];  split_copy = None
    add_1 = torch.ops.aten.add.Tensor(getitem, ones);  getitem = ones = None
    select_copy = torch.ops.aten.select_copy.int(_reshape_alias_copy, 0, 0);  _reshape_alias_copy = None
    clone = torch.ops.aten.clone.default(add_1, memory_format = torch.contiguous_format)
    _unsafe_view = torch.ops.aten._unsafe_view.default(clone, [4]);  clone = None
    view_copy_1 = torch.ops.aten.view_copy.default(add, [8]);  add = None
    _reshape_alias_copy_1 = torch.ops.aten._reshape_alias_copy.default(view_copy_1, [2, 4], [4, 1]);  view_copy_1 = None
    transpose_copy_1 = torch.ops.aten.transpose_copy.int(_reshape_alias_copy_1, 1, 0);  _reshape_alias_copy_1 = None
    unsqueeze_copy_1 = torch.ops.aten.unsqueeze_copy.default(transpose_copy_1, 0);  transpose_copy_1 = None
    squeeze_copy_1 = torch.ops.aten.squeeze_copy.default(unsqueeze_copy_1);  unsqueeze_copy_1 = None
    slice_scatter = torch.ops.aten.slice_scatter.default(squeeze_copy_1, add_1, 0, 0, 2);  squeeze_copy_1 = None
    unsqueeze_copy_2 = torch.ops.aten.unsqueeze_copy.default(slice_scatter, 0);  slice_scatter = None
    squeeze_copy_2 = torch.ops.aten.squeeze_copy.dim(unsqueeze_copy_2, 0);  unsqueeze_copy_2 = None
    transpose_copy_2 = torch.ops.aten.transpose_copy.int(squeeze_copy_2, 1, 0);  squeeze_copy_2 = None
    _reshape_alias_copy_2 = torch.ops.aten._reshape_alias_copy.default(transpose_copy_2, [8], [1]);  transpose_copy_2 = None
    view_copy_2 = torch.ops.aten.view_copy.default(_reshape_alias_copy_2, [4, 2]);  _reshape_alias_copy_2 = None
    view_copy_3 = torch.ops.aten.view_copy.default(view_copy_2, [8]);  view_copy_2 = None
    _reshape_alias_copy_3 = torch.ops.aten._reshape_alias_copy.default(view_copy_3, [2, 4], [4, 1]);  view_copy_3 = None
    select_copy_1 = torch.ops.aten.select_copy.int(_reshape_alias_copy_3, 0, 0);  _reshape_alias_copy_3 = None
    add_2 = torch.ops.aten.add.Tensor(select_copy_1, _unsafe_view);  select_copy_1 = _unsafe_view = None
    return add_1
    """)  # noqa: B950

        reinplaced_logs = self.get_logs(f, torch.ones(4, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([2, 2], device = device(type='cpu'), pin_memory = False)
    add = torch.ops.aten.add.Tensor(a_1, a_1);  a_1 = None
    view = torch.ops.aten.view.default(add, [8])
    _reshape_alias = torch.ops.aten._reshape_alias.default(view, [2, 4], [4, 1]);  view = None
    transpose = torch.ops.aten.transpose.int(_reshape_alias, 1, 0)
    unsqueeze = torch.ops.aten.unsqueeze.default(transpose, 0);  transpose = None
    squeeze = torch.ops.aten.squeeze.default(unsqueeze);  unsqueeze = None
    split = torch.ops.aten.split.Tensor(squeeze, 2);  squeeze = None
    getitem = split[0]
    getitem_1 = split[1];  split = None
    add_1 = torch.ops.aten.add_.Tensor(getitem, ones);  ones = None
    select = torch.ops.aten.select.int(_reshape_alias, 0, 0);  _reshape_alias = None
    clone = torch.ops.aten.clone.default(getitem, memory_format = torch.contiguous_format)
    _unsafe_view = torch.ops.aten._unsafe_view.default(clone, [4]);  clone = None
    view_1 = torch.ops.aten.view.default(add, [8]);  add = None
    _reshape_alias_1 = torch.ops.aten._reshape_alias.default(view_1, [2, 4], [4, 1]);  view_1 = None
    transpose_1 = torch.ops.aten.transpose.int(_reshape_alias_1, 1, 0);  _reshape_alias_1 = None
    unsqueeze_1 = torch.ops.aten.unsqueeze.default(transpose_1, 0);  transpose_1 = None
    squeeze_1 = torch.ops.aten.squeeze.default(unsqueeze_1);  unsqueeze_1 = None
    unsqueeze_2 = torch.ops.aten.unsqueeze.default(squeeze_1, 0);  squeeze_1 = None
    squeeze_2 = torch.ops.aten.squeeze.dim(unsqueeze_2, 0);  unsqueeze_2 = None
    transpose_2 = torch.ops.aten.transpose.int(squeeze_2, 1, 0);  squeeze_2 = None
    _reshape_alias_2 = torch.ops.aten._reshape_alias.default(transpose_2, [8], [1]);  transpose_2 = None
    view_2 = torch.ops.aten.view.default(_reshape_alias_2, [4, 2]);  _reshape_alias_2 = None
    view_3 = torch.ops.aten.view.default(view_2, [8]);  view_2 = None
    _reshape_alias_3 = torch.ops.aten._reshape_alias.default(view_3, [2, 4], [4, 1]);  view_3 = None
    select_1 = torch.ops.aten.select.int(_reshape_alias_3, 0, 0);  _reshape_alias_3 = None
    add_2 = torch.ops.aten.add.Tensor(select_1, _unsafe_view);  select_1 = _unsafe_view = None
    return getitem
    """)

    def test_reapply_views_simple(self):
        def f(x):
            tmp = torch.ones(4, 2)
            y = x.view(4, 2)
            y.add_(tmp)
            z = x * x
            return y
        self.assert_functionalization(f, torch.ones(4, 2), reapply_views=True)
        logs = self.get_logs(f, torch.ones(4, 2), reapply_views=True)
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    ones = torch.ops.aten.ones.default([4, 2], device = device(type='cpu'), pin_memory = False)
    view = torch.ops.aten.view.default(a_1, [4, 2])
    add = torch.ops.aten.add.Tensor(view, ones);  view = ones = None
    view_1 = torch.ops.aten.view.default(add, [4, 2])
    mul = torch.ops.aten.mul.Tensor(view_1, view_1)
    copy_ = torch.ops.aten.copy_.default(a_1, view_1);  a_1 = view_1 = None
    return add
    """)

    def test_aliases_maintained_after_pass_when_reapplying_views(self):
        def f(x):
            tmp = torch.ones(4, 2)
            y = x.view(4, 2)
            z = x.view(4, 2)
            y.add_(tmp)
            return y, z

        input_functional = torch._to_functional_tensor(torch.ones(4, 2))
        torch._enable_functionalization(reapply_views=True)
        try:
            y, z = f(input_functional)
            torch._sync(y)
            torch._sync(z)
        finally:
            torch._disable_functionalization()

        # y and z are aliases inside of the function, and that aliasing relationship should be maintained.
        _y = torch._from_functional_tensor(y)
        _z = torch._from_functional_tensor(z)
        self.assertTrue(are_aliased(_y, _z))

    # copy_() gets its own test, because it is special cased in functionalization.
    # self.copy_(src) decomposes into src.to(self).expand_as(self).
    def test_copy_(self):
        def f(x):
            tmp = torch.zeros(2, 2)
            tmp_slice = tmp.diagonal()
            y = tmp_slice.copy_(x)
            z = y.add_(x)
            return z

        # Test 1: copy_() with same dtype and shape
        # to() is a composite op that noops when the dtype/shape match, so nothing gets logged.
        # self.assert_functionalization(f, torch.ones(2))
        logs = self.get_logs(f, torch.ones(2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
    diagonal_copy = torch.ops.aten.diagonal_copy.default(zeros);  zeros = None
    add = torch.ops.aten.add.Tensor(a_1, a_1);  a_1 = None
    return add
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
    diagonal = torch.ops.aten.diagonal.default(zeros);  zeros = None
    add = torch.ops.aten.add.Tensor(a_1, a_1);  a_1 = None
    return add
    """)

        # Test 2: copy_() with same dtype, different shape
        self.assert_functionalization(f, torch.ones(1))
        logs = self.get_logs(f, torch.ones(1))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
    diagonal_copy = torch.ops.aten.diagonal_copy.default(zeros);  zeros = None
    expand_copy = torch.ops.aten.expand_copy.default(a_1, [2])
    add = torch.ops.aten.add.Tensor(expand_copy, a_1);  expand_copy = a_1 = None
    return add
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(1), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
    diagonal = torch.ops.aten.diagonal.default(zeros);  zeros = None
    expand_copy = torch.ops.aten.expand_copy.default(a_1, [2])
    add = torch.ops.aten.add_.Tensor(expand_copy, a_1);  a_1 = None
    return expand_copy
    """)

        # Test 3: copy_() with different dtype, same shape
        self.assert_functionalization(f, torch.ones(2, dtype=torch.long))
        logs = self.get_logs(f, torch.ones(2, dtype=torch.long))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
    diagonal_copy = torch.ops.aten.diagonal_copy.default(zeros);  zeros = None
    _to_copy = torch.ops.aten._to_copy.default(a_1, dtype = torch.float32, layout = torch.strided, device = device(type='cpu'), pin_memory = False)
    add = torch.ops.aten.add.Tensor(_to_copy, a_1);  _to_copy = a_1 = None
    return add
    """)  # noqa: B950

        reinplaced_logs = self.get_logs(f, torch.ones(2, dtype=torch.long), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
    diagonal = torch.ops.aten.diagonal.default(zeros);  zeros = None
    _to_copy = torch.ops.aten._to_copy.default(a_1, dtype = torch.float32, layout = torch.strided, device = device(type='cpu'), pin_memory = False)
    add = torch.ops.aten.add_.Tensor(_to_copy, a_1);  a_1 = None
    return _to_copy
    """)  # noqa: B950

        # Test 4: copy_() with different dtype, different shape
        self.assert_functionalization(f, torch.ones(1, dtype=torch.long))
        logs = self.get_logs(f, torch.ones(1, dtype=torch.long))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
    diagonal_copy = torch.ops.aten.diagonal_copy.default(zeros);  zeros = None
    _to_copy = torch.ops.aten._to_copy.default(a_1, dtype = torch.float32, layout = torch.strided, device = device(type='cpu'), pin_memory = False)
    expand_copy = torch.ops.aten.expand_copy.default(_to_copy, [2]);  _to_copy = None
    add = torch.ops.aten.add.Tensor(expand_copy, a_1);  expand_copy = a_1 = None
    return add
    """)  # noqa: B950

        reinplaced_logs = self.get_logs(f, torch.ones(1, dtype=torch.long), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
    diagonal = torch.ops.aten.diagonal.default(zeros);  zeros = None
    _to_copy = torch.ops.aten._to_copy.default(a_1, dtype = torch.float32, layout = torch.strided, device = device(type='cpu'), pin_memory = False)
    expand_copy = torch.ops.aten.expand_copy.default(_to_copy, [2]);  _to_copy = None
    add = torch.ops.aten.add_.Tensor(expand_copy, a_1);  a_1 = None
    return expand_copy
    """)  # noqa: B950

    def test_expand_symint(self):
        # Once some existing SymInt bugs are ironed out, we should update
        # this test to plumb FakeSymbolicTensors through it
        def f(x):
            return x.expand(x.size(0), x.size(1))

        self.assert_functionalization(f, torch.ones(2, 2))
        logs = self.get_logs(f, torch.ones(2, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    expand_copy = torch.ops.aten.expand_copy.default(a_1, [2, 2]);  a_1 = None
    return expand_copy
    """)

    def test_fill_(self):
        def f(x):
            y = x + x
            z = y.diagonal()
            z.fill_(0)
            return y

        self.assert_functionalization(f, torch.ones(2, 2))
        logs = self.get_logs(f, torch.ones(2, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    add = torch.ops.aten.add.Tensor(a_1, a_1);  a_1 = None
    diagonal_copy = torch.ops.aten.diagonal_copy.default(add)
    fill = torch.ops.aten.fill.Scalar(diagonal_copy, 0);  diagonal_copy = None
    diagonal_scatter = torch.ops.aten.diagonal_scatter.default(add, fill);  add = fill = None
    return diagonal_scatter
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(2, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    add = torch.ops.aten.add.Tensor(a_1, a_1);  a_1 = None
    diagonal = torch.ops.aten.diagonal.default(add)
    fill = torch.ops.aten.fill_.Scalar(diagonal, 0);  diagonal = None
    return add
    """)

    def test_resize_smaller(self):
        def f(w):
            # Resizing to a smaller size doesn't affect storage
            x = w + 1
            y = x.view(4, 4)
            y.resize_(3, 3)
            y2 = y.view(-1)
            y2.add_(1)
            z = y + 1
            return z

        self.assert_functionalization(f, torch.ones(8, 2))
        logs = self.get_logs(f, torch.ones(8, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    add = torch.ops.aten.add.Tensor(a_1, 1);  a_1 = None
    view_copy = torch.ops.aten.view_copy.default(add, [4, 4])
    resize = torch.ops.aten.resize.default(view_copy, [3, 3])
    as_strided_copy = torch.ops.aten.as_strided_copy.default(view_copy, [3, 3], [3, 1]);  view_copy = None
    view_copy_1 = torch.ops.aten.view_copy.default(as_strided_copy, [-1]);  as_strided_copy = None
    add_1 = torch.ops.aten.add.Tensor(view_copy_1, 1);  view_copy_1 = None
    view_copy_2 = torch.ops.aten.view_copy.default(add, [4, 4]);  add = None
    as_strided_copy_1 = torch.ops.aten.as_strided_copy.default(view_copy_2, [3, 3], [3, 1])
    view_copy_3 = torch.ops.aten.view_copy.default(add_1, [3, 3]);  add_1 = None
    as_strided_scatter = torch.ops.aten.as_strided_scatter.default(view_copy_2, view_copy_3, [3, 3], [3, 1]);  view_copy_2 = view_copy_3 = None
    view_copy_4 = torch.ops.aten.view_copy.default(as_strided_scatter, [8, 2]);  as_strided_scatter = None
    view_copy_5 = torch.ops.aten.view_copy.default(view_copy_4, [4, 4]);  view_copy_4 = None
    as_strided_copy_2 = torch.ops.aten.as_strided_copy.default(view_copy_5, [3, 3], [3, 1]);  view_copy_5 = None
    add_2 = torch.ops.aten.add.Tensor(as_strided_copy_2, 1);  as_strided_copy_2 = None
    return add_2
    """)  # noqa: B950

        reinplaced_logs = self.get_logs(f, torch.ones(8, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    add = torch.ops.aten.add.Tensor(a_1, 1);  a_1 = None
    view = torch.ops.aten.view.default(add, [4, 4])
    resize = torch.ops.aten.resize.default(view, [3, 3])
    as_strided = torch.ops.aten.as_strided.default(view, [3, 3], [3, 1]);  view = None
    view_1 = torch.ops.aten.view.default(as_strided, [-1]);  as_strided = None
    add_1 = torch.ops.aten.add_.Tensor(view_1, 1)
    view_2 = torch.ops.aten.view.default(add, [4, 4]);  add = None
    as_strided_1 = torch.ops.aten.as_strided.default(view_2, [3, 3], [3, 1])
    view_3 = torch.ops.aten.view.default(view_1, [3, 3]);  view_1 = None
    view_4 = torch.ops.aten.view.default(view_2, [8, 2]);  view_2 = None
    view_5 = torch.ops.aten.view.default(view_4, [4, 4]);  view_4 = None
    as_strided_2 = torch.ops.aten.as_strided.default(view_5, [3, 3], [3, 1]);  view_5 = None
    add_2 = torch.ops.aten.add_.Tensor(as_strided_2, 1)
    return as_strided_2
    """)

    def test_resize_larger_valid(self):
        def f(x):
            y = x + 1
            # resizing a tensor to a larger size is only currently allowed
            # if the tensor-to-resize is not a view / has no outstanding views.
            # See Note [resize_() in functionalization pass]
            y.resize_(5, 5)
            y2 = y.view(25)
            # Do a mutation to ensure that aliases of the output of resize_()
            # propagate mutations correctly.
            # I'm using fill_ specifically because I want to guarantee that
            # none of the output has uninitialized memory at the end
            # (since these tests compare the data output against a reference impl)
            y2.fill_(1)
            out = y + 1
            return y, out

        self.assert_functionalization(f, torch.ones(8, 2))
        logs = self.get_logs(f, torch.ones(8, 2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    add = torch.ops.aten.add.Tensor(a_1, 1);  a_1 = None
    resize = torch.ops.aten.resize.default(add, [5, 5]);  add = None
    view_copy = torch.ops.aten.view_copy.default(resize, [25]);  resize = None
    fill = torch.ops.aten.fill.Scalar(view_copy, 1);  view_copy = None
    view_copy_1 = torch.ops.aten.view_copy.default(fill, [5, 5]);  fill = None
    add_1 = torch.ops.aten.add.Tensor(view_copy_1, 1)
    return (view_copy_1, add_1)
    """)

        reinplaced_logs = self.get_logs(f, torch.ones(8, 2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    add = torch.ops.aten.add.Tensor(a_1, 1);  a_1 = None
    resize = torch.ops.aten.resize_.default(add, [5, 5])
    view = torch.ops.aten.view.default(add, [25]);  add = None
    fill = torch.ops.aten.fill_.Scalar(view, 1)
    view_1 = torch.ops.aten.view.default(view, [5, 5]);  view = None
    add_1 = torch.ops.aten.add.Tensor(view_1, 1)
    return (view_1, add_1)
    """)

    def test_resize_larger_invalid(self):
        def f(x):
            y = x + 1
            z = y.view(4, 4)
            # resizing a tensor to a larger size is only currently allowed
            # if the tensor-to-resize is not a view / has no outstanding views.
            # See Note [resize_() in functionalization pass]
            # This should fail
            z.resize_(5, 5)
            z2 = z.view(25)
            z2.fill_(1)
            out = z + 1
            return y, out

        with self.assertRaisesRegex(
                RuntimeError,
                r'Attempted to resize a view tensor to a larger size. This is not allowed in the functionalization pass'):
            self.assert_functionalization(f, torch.ones(8, 2))

    def test_nested_functions_propagate_updates(self):
        def g(x):
            # Create a view of x
            y = x[0]
            y.add_(1)
            # The view, y, gets deallocated at the end of this function

        def f(x):
            # Calling g(x) should mutate x
            g(x)
            # We expect x to be synced here, even though the alias created in g() has been deallocated!
            y = x + x
            return y

        self.assert_functionalization(f, torch.ones(2, 2))

    def test_mixed_wrappers_valid(self):
        def f(x, y):
            z = x + y
            z.add_(1)
            return z

        x1_not_functional = LoggingTensor(torch.ones(4))
        x2_functional = torch._to_functional_tensor(LoggingTensor(torch.ones(4)))

        with capture_logs() as logs:
            y = f(x1_not_functional, x2_functional)

        # Make sure that functionalization ran the "+" kernel
        # with a functional + non-functional tensor, and wrapped the output appropriately.
        self.assertExpectedInline('\n'.join(logs), """\
$2 = torch._ops.aten.add.Tensor($0, $1)
$3 = torch._ops.aten.add.Tensor($2, 1)""")

    def test_mixed_wrappers_invalid(self):
        x1_not_functional = torch.ones(4)
        x2_functional = torch._to_functional_tensor(torch.ones(4))

        # When dealing with mixed functional + non functional tensors,
        # normal_tensor.add_(functional_tensor) is not valid
        # because normal_tensor would need to be "promoted" to a functional tensor.
        with self.assertRaises(RuntimeError):
            x1_not_functional.add_(x2_functional)

    def test_index_mutation_on_non_input(self):
        def f(x):
            tmp = torch.zeros(10)
            tmp[5].fill_(1)
            return tmp
        self.assert_functionalization(f, torch.ones(2))
        logs = self.get_logs(f, torch.ones(2))
        self.assertExpectedInline(logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([10], device = device(type='cpu'), pin_memory = False)
    select_copy = torch.ops.aten.select_copy.int(zeros, 0, 5)
    fill = torch.ops.aten.fill.Scalar(select_copy, 1);  select_copy = None
    select_scatter = torch.ops.aten.select_scatter.default(zeros, fill, 0, 5);  zeros = fill = None
    return select_scatter
    """)  # noqa: B950

        reinplaced_logs = self.get_logs(f, torch.ones(2), reapply_views=True, run_reinplace=True)
        self.assertExpectedInline(reinplaced_logs, """\



def forward(self, a_1):
    zeros = torch.ops.aten.zeros.default([10], device = device(type='cpu'), pin_memory = False)
    select = torch.ops.aten.select.int(zeros, 0, 5)
    fill = torch.ops.aten.fill_.Scalar(select, 1);  select = None
    return zeros
    """)

if __name__ == '__main__':
    run_tests()