forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest_memdependency.cpp
3150 lines (2565 loc) · 96.7 KB
/
test_memdependency.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 <gtest/gtest.h>
#include <test/cpp/tensorexpr/test_base.h>
#include <torch/csrc/jit/tensorexpr/bounds_overlap.h>
#include <torch/csrc/jit/tensorexpr/ir.h>
#include <torch/csrc/jit/tensorexpr/ir_printer.h>
#include <torch/csrc/jit/tensorexpr/ir_simplifier.h>
#include <torch/csrc/jit/tensorexpr/loopnest.h>
#include <torch/csrc/jit/tensorexpr/mem_dependency_checker.h>
#include <torch/csrc/jit/tensorexpr/tensor.h>
namespace torch {
namespace jit {
using namespace torch::jit::tensorexpr;
// Test helper function used to determine if two regions of a buffer have an
// overlap. No Overlap & partial overlap is obvious. Contains means A is
// larger and fully encloses B, while ContainedOrEqual is the reverse. Equal
// ranges are ContainedOrEqual.
TEST(MemDependency, BoundOverlap) {
KernelScope kernel_scope;
using namespace analysis;
auto CB = [](int s, int e) { return Bound(new IntImm(s), new IntImm(e)); };
// Sanity check 3 overlap cases.
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(0, 0), CB(0, 0)));
ASSERT_EQ(PartialOverlap, boundOverlap(CB(0, 3), CB(2, 5)));
ASSERT_EQ(NoOverlap, boundOverlap(CB(0, 0), CB(1, 1)));
// Partial overlap works in either order.
ASSERT_EQ(PartialOverlap, boundOverlap(CB(0, 10), CB(7, 14)));
ASSERT_EQ(PartialOverlap, boundOverlap(CB(7, 14), CB(0, 10)));
// Total Overlap works when one bound encloses the other, and returns which.
ASSERT_EQ(Contains, boundOverlap(CB(2, 15), CB(7, 9)));
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(2, 15), CB(0, 16)));
// Total overlap works when the bounds are an identical range, returns
// ContainedOrEqual.
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(2, 15), CB(2, 15)));
// Total overlap when only one end of the bound matches.
ASSERT_EQ(Contains, boundOverlap(CB(2, 15), CB(2, 10)));
ASSERT_EQ(Contains, boundOverlap(CB(2, 15), CB(3, 15)));
ASSERT_EQ(Contains, boundOverlap(CB(0, 10), CB(0, 9)));
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(2, 10), CB(2, 15)));
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(3, 15), CB(2, 15)));
// No overlap when a < b.
ASSERT_EQ(NoOverlap, boundOverlap(CB(0, 2), CB(5, 10)));
ASSERT_EQ(NoOverlap, boundOverlap(CB(2, 2), CB(3, 3)));
ASSERT_EQ(NoOverlap, boundOverlap(CB(100, 120), CB(130, 130)));
// No overlap when a > b.
ASSERT_EQ(NoOverlap, boundOverlap(CB(5, 10), CB(0, 2)));
ASSERT_EQ(NoOverlap, boundOverlap(CB(3, 3), CB(2, 2)));
ASSERT_EQ(NoOverlap, boundOverlap(CB(130, 130), CB(100, 120)));
// No overlap when adjacent.
ASSERT_EQ(NoOverlap, boundOverlap(CB(0, 100), CB(101, 120)));
ASSERT_EQ(NoOverlap, boundOverlap(CB(2, 3), CB(0, 1)));
// Partial overlap when middle bounds match.
ASSERT_EQ(PartialOverlap, boundOverlap(CB(0, 100), CB(100, 120)));
ASSERT_EQ(PartialOverlap, boundOverlap(CB(0, 2), CB(2, 4)));
ASSERT_EQ(PartialOverlap, boundOverlap(CB(100, 120), CB(0, 100)));
ASSERT_EQ(PartialOverlap, boundOverlap(CB(2, 3), CB(1, 2)));
// Total overlap when one bound is single length over one end of the other.
ASSERT_EQ(Contains, boundOverlap(CB(2, 15), CB(15, 15)));
ASSERT_EQ(Contains, boundOverlap(CB(2, 15), CB(2, 2)));
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(2, 2), CB(2, 15)));
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(15, 15), CB(2, 15)));
}
TEST(MemDependency, BoundOverlapSymbolic) {
KernelScope kernel_scope;
VarHandle x("x", kInt);
VarHandle y("y", kInt);
VarHandle z("z", kInt);
VarHandle w("w", kInt);
using namespace analysis;
auto CB = [](ExprHandle s, ExprHandle e) {
return Bound(s.node(), e.node());
};
// Sanity check cases where the start and end is symbolic but the diff is
// constant.
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(x, x), CB(x, x)));
ASSERT_EQ(PartialOverlap, boundOverlap(CB(x, x + 3), CB(x + 2, x + 5)));
ASSERT_EQ(NoOverlap, boundOverlap(CB(x, x), CB(x + 1, x + 1)));
// We can't infer the sign of y, so cannot tell whether adding y is larger or
// smaller than y/2.
ASSERT_EQ(PartialOverlap, boundOverlap(CB(x, x + y), CB(x, x + y / 2)));
// No information about this bound, have to take the most conservative option:
// there may be an overlap.
ASSERT_EQ(PartialOverlap, boundOverlap(CB(x, y), CB(z, w)));
// Math on opaque terms works.
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(x + w, y - z), CB(x + w, y - z)));
// Even requiring simplification.
ASSERT_EQ(ContainedOrEqual, boundOverlap(CB(x - w - w, y), CB(x - w * 2, y)));
}
// Tests the helper function for overlap of multi dimensional indices bounds.
// This uses boundOverlap on each dimension and return the "lowest" kind of
// overlap.
TEST(MemDependency, BoundOverlapMultiDim) {
KernelScope kernel_scope;
using namespace analysis;
auto CB = [](int s, int e) { return Bound(new IntImm(s), new IntImm(e)); };
// Sanity check one dimensional cases.
ASSERT_EQ(ContainedOrEqual, overlaps({CB(0, 0)}, {CB(0, 0)}));
ASSERT_EQ(NoOverlap, overlaps({CB(0, 2)}, {CB(5, 10)}));
ASSERT_EQ(PartialOverlap, overlaps({CB(0, 100)}, {CB(100, 120)}));
// Total overlap in 3 dims.
ASSERT_EQ(
ContainedOrEqual,
overlaps({CB(0, 2), CB(0, 5), CB(0, 4)}, {CB(0, 2), CB(0, 5), CB(0, 4)}));
ASSERT_EQ(
ContainedOrEqual,
overlaps(
{CB(0, 2), CB(0, 5), CB(0, 4)}, {CB(0, 2), CB(0, 5), CB(0, 10)}));
// Total overlap in 2 dims, no overlap in another.
ASSERT_EQ(
NoOverlap,
overlaps(
{CB(0, 2), CB(0, 5), CB(0, 4)}, {CB(0, 2), CB(0, 5), CB(5, 10)}));
// Total overlap in 2 dims, partial overlap in another.
ASSERT_EQ(
PartialOverlap,
overlaps(
{CB(0, 2), CB(0, 5), CB(0, 5)}, {CB(0, 2), CB(0, 5), CB(5, 10)}));
// This case is most important, so verify the overlap in any dim. (dim 2)
ASSERT_EQ(
PartialOverlap,
overlaps({CB(0, 2), CB(0, 5), CB(0, 5)}, {CB(0, 2), CB(2, 6), CB(0, 5)}));
// Dim 1.
ASSERT_EQ(
PartialOverlap,
overlaps({CB(0, 2), CB(0, 5), CB(0, 5)}, {CB(1, 3), CB(0, 5), CB(0, 5)}));
// Total overlap in 1 dim, partial in 2.
ASSERT_EQ(
PartialOverlap,
overlaps(
{CB(0, 2), CB(0, 5), CB(0, 5)}, {CB(2, 6), CB(0, 5), CB(5, 10)}));
// Total overlap, partial overlap, no overlap.
ASSERT_EQ(
NoOverlap,
overlaps(
{CB(0, 2), CB(0, 5), CB(0, 5)}, {CB(2, 6), CB(11, 15), CB(0, 5)}));
// Total overlap (B) in 2 dims, total overlap (A) in another.
ASSERT_EQ(
Contains,
overlaps({CB(0, 2), CB(0, 5), CB(0, 4)}, {CB(0, 2), CB(0, 3), CB(0, 4)}));
// Total overlap (A) in 2 dims, total overlap (B) in another.
ASSERT_EQ(
Contains,
overlaps(
{CB(0, 12), CB(0, 15), CB(0, 4)}, {CB(0, 2), CB(0, 3), CB(0, 14)}));
// Total (B), No Overlap, Total (A).
ASSERT_EQ(
NoOverlap,
overlaps(
{CB(0, 2), CB(0, 5), CB(0, 5)}, {CB(0, 6), CB(11, 15), CB(1, 2)}));
}
// Test the helper we use to subtract bounds: returns the regions(s) of A which
// remain after removing the region of B.
TEST(MemDependency, BoundSubtract) {
KernelScope kernel_scope;
using namespace analysis;
auto CB = [](int s, int e) { return Bound(new IntImm(s), new IntImm(e)); };
auto EQ = [](const IndexBounds& x, const IndexBounds& y) {
return indexBoundsEquals(x, y);
};
// One element subtract.
ASSERT_EQ(subtractBound(CB(0, 0), CB(0, 0)).size(), 0);
ASSERT_EQ(subtractBound(CB(5, 5), CB(5, 5)).size(), 0);
// No Overlap.
ASSERT_TRUE(EQ(subtractBound(CB(5, 5), CB(2, 2)), {CB(5, 5)}));
ASSERT_TRUE(EQ(subtractBound(CB(5, 5), CB(0, 4)), {CB(5, 5)}));
// one side overlap.
ASSERT_TRUE(EQ(subtractBound(CB(1, 5), CB(4, 7)), {CB(1, 3)}));
ASSERT_TRUE(EQ(subtractBound(CB(0, 5), CB(5, 7)), {CB(0, 4)}));
ASSERT_TRUE(EQ(subtractBound(CB(4, 5), CB(1, 4)), {CB(5, 5)}));
ASSERT_TRUE(EQ(subtractBound(CB(1, 5), CB(0, 4)), {CB(5, 5)}));
// both sides overlap.
ASSERT_TRUE(EQ(subtractBound(CB(1, 5), CB(0, 7)), {}));
ASSERT_TRUE(EQ(subtractBound(CB(5, 5), CB(5, 7)), {}));
// internal overlap.
ASSERT_TRUE(EQ(subtractBound(CB(1, 5), CB(2, 3)), {CB(1, 1), CB(4, 5)}));
ASSERT_TRUE(EQ(subtractBound(CB(0, 5), CB(2, 4)), {CB(0, 1), CB(5, 5)}));
}
TEST(MemDependency, BoundSubtractSymbolic) {
KernelScope kernel_scope;
VarHandle x("x", kInt);
VarHandle y("y", kInt);
VarHandle z("z", kInt);
VarHandle w("w", kInt);
using namespace analysis;
auto CB = [](ExprHandle s, ExprHandle e) {
return Bound(s.node(), e.node());
};
auto EQ = [](const IndexBounds& x, const IndexBounds& y) {
return indexBoundsEquals(x, y);
};
// One element subtract.
ASSERT_TRUE(EQ(subtractBound(CB(x, x), CB(x, x)), {}));
ASSERT_TRUE(EQ(subtractBound(CB(x + 1, x + 1), CB(x + 1, x + 1)), {}));
ASSERT_TRUE(EQ(subtractBound(CB(x * 2, x * 2), CB(x * 2, x * 2)), {}));
// Subtract constant range low.
ASSERT_TRUE(
EQ(subtractBound(CB(x, x + 10), CB(x, x + 4)), {CB(x + 5, x + 10)}));
// Subtract constant range high.
ASSERT_TRUE(
EQ(subtractBound(CB(x, x + 10), CB(x + 6, x + 12)), {CB(x, x + 5)}));
// Subtract constant range total overlap.
ASSERT_TRUE(EQ(subtractBound(CB(x, x + 10), CB(x, x + 10)), {}));
ASSERT_TRUE(EQ(subtractBound(CB(x + 2, x + 10), CB(x, x + 12)), {}));
// Subtract constant range internal.
ASSERT_TRUE(
EQ(subtractBound(CB(x, x + 10), CB(x + 3, x + 7)),
{CB(x, x + 2), CB(x + 8, x + 10)}));
// Size is inferable but not constant, only works with a single var.
ASSERT_TRUE(EQ(subtractBound(CB(0, x), CB(0, x * 2)), {}));
ASSERT_TRUE(EQ(subtractBound(CB(0, x * 2), CB(0, x - 1)), {CB(x, x * 2)}));
// Size is not inferable.
ASSERT_TRUE(EQ(subtractBound(CB(x, y), CB(z, w)), {CB(x, y)}));
ASSERT_TRUE(EQ(subtractBound(CB(x, y), CB(x, z)), {CB(x, y)}));
ASSERT_TRUE(EQ(subtractBound(CB(x, y), CB(0, x)), {CB(x, y)}));
ASSERT_TRUE(EQ(subtractBound(CB(x, x), CB(0, 0)), {CB(x, x)}));
}
// Tests the helper function that does subtraction, but for multi dimensional
// indices bounds.
TEST(MemDependency, BoundSubtractMultiDim) {
KernelScope kernel_scope;
using namespace analysis;
auto CB = [](int s, int e) { return Bound(new IntImm(s), new IntImm(e)); };
auto EQ = [](std::vector<IndexBounds> x, std::vector<IndexBounds> y) {
if (x.size() != y.size()) {
return false;
}
for (auto i = 0; i < x.size(); ++i) {
if (!indexBoundsEquals(x[i], y[i])) {
return false;
}
}
return true;
};
// sanity check one dimension.
ASSERT_TRUE(EQ(subtractIndicesBounds({CB(0, 9)}, {CB(0, 9)}), {}));
ASSERT_TRUE(EQ(subtractIndicesBounds({CB(3, 9)}, {CB(0, 12)}), {}));
ASSERT_TRUE(
EQ(subtractIndicesBounds({CB(0, 12)}, {CB(0, 9)}), {{CB(10, 12)}}));
ASSERT_TRUE(
EQ(subtractIndicesBounds({CB(0, 12)}, {CB(3, 12)}), {{CB(0, 2)}}));
ASSERT_TRUE(EQ(
subtractIndicesBounds({CB(0, 9)}, {CB(1, 8)}), {{CB(0, 0)}, {CB(9, 9)}}));
// Multi dim total overlap.
ASSERT_TRUE(EQ(
subtractIndicesBounds({CB(0, 9), CB(0, 2)}, {CB(0, 9), CB(0, 2)}), {}));
ASSERT_TRUE(EQ(
subtractIndicesBounds({CB(0, 9), CB(0, 2)}, {CB(0, 10), CB(0, 20)}), {}));
// Mutli dim one way partial in dim 1.
ASSERT_TRUE(
EQ(subtractIndicesBounds({CB(0, 9), CB(0, 2)}, {CB(0, 3), CB(0, 2)}),
{{CB(4, 9), CB(0, 2)}}));
// Mutli dim one way partial in dim 2.
ASSERT_TRUE(
EQ(subtractIndicesBounds({CB(0, 9), CB(0, 20)}, {CB(0, 9), CB(0, 10)}),
{{CB(0, 9), CB(11, 20)}}));
// Partial overlap in 2 dims.
ASSERT_TRUE(
EQ(subtractIndicesBounds({CB(0, 5), CB(0, 5)}, {CB(2, 8), CB(2, 8)}),
{{CB(0, 1), CB(0, 5)}, {CB(2, 5), CB(0, 1)}}));
// Partial overlap in 3 dims.
ASSERT_TRUE(
EQ(subtractIndicesBounds(
{CB(0, 5), CB(0, 5), CB(0, 5)}, {CB(2, 8), CB(2, 8), CB(2, 8)}),
{{CB(0, 1), CB(0, 5), CB(0, 5)},
{CB(2, 5), CB(0, 1), CB(0, 5)},
{CB(2, 5), CB(2, 5), CB(0, 1)}}));
}
// Tests the multi dimensional subtraction code for bounds that cannot be fully
// materialized.
TEST(MemDependency, BoundSubtractMultiDimSymbolic) {
KernelScope kernel_scope;
VarHandle x("x", kInt);
VarHandle y("y", kInt);
using namespace analysis;
auto CB = [](ExprHandle s, ExprHandle e) {
return Bound(s.node(), e.node());
};
auto EQ = [](std::vector<IndexBounds> x, std::vector<IndexBounds> y) {
if (x.size() != y.size()) {
return false;
}
for (auto i = 0; i < x.size(); ++i) {
if (!indexBoundsEquals(x[i], y[i])) {
return false;
}
}
return true;
};
// Cannot determine overlaps.
ASSERT_TRUE(EQ(subtractIndicesBounds({CB(x, x)}, {CB(0, 0)}), {{CB(x, x)}}));
// Various total Overlaps.
ASSERT_TRUE(EQ(
subtractIndicesBounds({CB(x, x), CB(x, x)}, {CB(x, x), CB(x, x)}), {}));
ASSERT_TRUE(EQ(
subtractIndicesBounds({CB(x, y), CB(x, y)}, {CB(x, y), CB(x, y)}), {}));
ASSERT_TRUE(EQ(
subtractIndicesBounds({CB(x, x), CB(y, y)}, {CB(x, x), CB(y, y)}), {}));
ASSERT_TRUE(EQ(
subtractIndicesBounds({CB(0, x), CB(0, y)}, {CB(0, x), CB(0, y)}), {}));
// one-way overlap in first dim.
ASSERT_TRUE(
EQ(subtractIndicesBounds({CB(0, x), CB(0, y)}, {CB(0, x - 5), CB(0, y)}),
{{CB(x - 4, x), CB(0, y)}}));
// second dim.
ASSERT_TRUE(
EQ(subtractIndicesBounds({CB(0, x), CB(0, y)}, {CB(0, x), CB(5, y)}),
{{CB(0, x), CB(0, 4)}}));
// Internal overlap in first dim.
ASSERT_TRUE(
EQ(subtractIndicesBounds({CB(0, x), CB(0, y)}, {CB(2, x - 5), CB(0, y)}),
{{CB(0, 1), CB(0, y)}, {CB(x - 4, x), CB(0, y)}}));
// second dim.
ASSERT_TRUE(EQ(
subtractIndicesBounds({CB(0, x), CB(0, y)}, {CB(0, x), CB(10, y - 10)}),
{{CB(0, x), CB(0, 9)}, {CB(0, x), CB(y - 9, y)}}));
// Overlap in both dimensions.
ASSERT_TRUE(
EQ(subtractIndicesBounds(
{CB(0, x), CB(0, y)}, {CB(5, x - 5), CB(10, y - 10)}),
{
{CB(0, 4), CB(0, y)},
{CB(x - 4, x), CB(0, y)},
{CB(0, x), CB(0, 9)},
{CB(0, x), CB(y - 9, y)},
}));
}
// Simple check that the analyzer does anything at all...
TEST(MemDependency, MemDependencyCheckerSimple) {
KernelScope kernel_scope;
BufHandle a("A", {1}, kInt);
BufHandle b("B", {1}, kInt);
analysis::MemDependencyChecker analyzer;
/*
* A[0] = 3;
* B[0] = A[0] + 1;
*/
Store* aStore = Store::make(a, {0}, 3, 1);
Store* bStore = Store::make(b, {0}, Add::make(Load::make(a, {0}, 1), 1), 1);
Stmt* stmt = Block::make({aStore, bStore});
stmt->accept(&analyzer);
ASSERT_TRUE(analyzer.dependsDirectly(bStore, aStore));
ASSERT_FALSE(analyzer.dependsIndirectly(aStore, bStore));
// sanity check, but anything that depends directly must depend indirectly.
ASSERT_TRUE(analyzer.dependsIndirectly(bStore, aStore));
}
// Check that there is a difference between direct and indirect dependence.
TEST(MemDependency, MemDependencyCheckerMultiStmt) {
KernelScope kernel_scope;
BufHandle a("A", {1}, kInt);
BufHandle b("B", {1}, kInt);
BufHandle c("C", {1}, kInt);
analysis::MemDependencyChecker analyzer;
/*
* A[0] = 3;
* B[0] = A[0];
* C[0] = B[0] + 1;
*/
Store* aStore = Store::make(a, {0}, 3, 1);
Store* bStore = Store::make(b, {0}, Load::make(a, {0}, 1), 1);
Store* cStore = Store::make(c, {0}, Add::make(Load::make(b, {0}, 1), 1), 1);
Stmt* stmt = Block::make({aStore, bStore, cStore});
stmt->accept(&analyzer);
// C depends on A indirectly.
ASSERT_FALSE(analyzer.dependsDirectly(cStore, aStore));
ASSERT_TRUE(analyzer.dependsIndirectly(cStore, aStore));
// C depends on B directly, which depends on A directly.
ASSERT_TRUE(analyzer.dependsDirectly(cStore, bStore));
ASSERT_TRUE(analyzer.dependsDirectly(bStore, aStore));
// Dependency goes top to bottom only.
ASSERT_FALSE(analyzer.dependsIndirectly(bStore, cStore));
ASSERT_FALSE(analyzer.dependsIndirectly(aStore, bStore));
ASSERT_FALSE(analyzer.dependsIndirectly(aStore, cStore));
}
// Verify that we do filter writes that are totally overlapped by later writes.
TEST(MemDependency, MemDependencyCheckerOverlap) {
KernelScope kernel_scope;
BufHandle a("A", {1}, kInt);
BufHandle b("B", {1}, kInt);
analysis::MemDependencyChecker analyzer;
/*
* A[0] = 3;
* A[0] = 6;
* B[0] = A[0] + 1;
*/
Store* aStore = Store::make(a, {0}, 3, 1);
Store* a2Store = Store::make(a, {0}, 6, 1);
Store* bStore = Store::make(b, {0}, Add::make(Load::make(a, {0}, 1), 1), 1);
Stmt* stmt = Block::make({aStore, a2Store, bStore});
stmt->accept(&analyzer);
// B store depends on second A store but not first since it is completely
// overlapped.
ASSERT_TRUE(analyzer.dependsIndirectly(bStore, a2Store));
ASSERT_FALSE(analyzer.dependsIndirectly(bStore, aStore));
// No dependency between either A store.
ASSERT_FALSE(analyzer.dependsIndirectly(aStore, a2Store));
ASSERT_FALSE(analyzer.dependsIndirectly(a2Store, aStore));
}
// Verify that bounds match loop iterations, and that dependencies progress
// across loop scopes.
TEST(MemDependency, MemDependencyCheckerLoop) {
KernelScope kernel_scope;
BufHandle a("A", {1}, kInt);
BufHandle b("B", {1}, kInt);
VarHandle x("x", kInt);
using namespace analysis;
MemDependencyChecker analyzer;
/*
* for (int x = 0; x < 10; ++x) {
* A[x] = x;
* }
* B[0] = A[0] + 1;
*/
Store* aStore = Store::make(a, {x}, x, 1);
Stmt* loop = For::make(x, 0, 10, aStore);
Store* bStore = Store::make(b, {0}, Add::make(Load::make(a, {4}, 1), 1), 1);
Stmt* stmt = Block::make({loop, bStore});
stmt->accept(&analyzer);
// Same A->B dependency.
ASSERT_TRUE(analyzer.dependsDirectly(bStore, aStore));
// B depends on the loop.
ASSERT_TRUE(analyzer.dependsDirectly(bStore, loop));
// A is in the loop but does not depend on any loop iteration.
ASSERT_FALSE(analyzer.dependsIndirectly(aStore, loop));
auto aStoreAccess = analyzer.accessFor(aStore);
ASSERT_NE(aStoreAccess, nullptr);
// It should have bounds covering the range of x: 0 <= x < 10.
ASSERT_TRUE(indexBoundsEquals(
aStoreAccess->bounds(), {Bound(new IntImm(0), new IntImm(9))}));
}
// Reductions should promote dependencies as well.
TEST(MemDependency, MemDependencyCheckerLoopReduce) {
KernelScope kernel_scope;
BufHandle a("A", {10}, kInt);
BufHandle b("B", {10}, kInt);
VarHandle x("x", kInt);
using namespace analysis;
MemDependencyChecker analyzer;
/*
* A[0] = 0;
* for (int x = 0; x < 10; ++x) {
* A[0] = A[x] + 1;
* }
* B[0] = A[0];
*/
Store* aInit = Store::make(a, {0}, 0, 1);
ExprHandle reduce =
ExprHandle(Sum()(a.node(), ExprHandle(1), {x.node()}, {x.node()}));
Store* aReduce = Store::make(a, {0}, reduce, 1);
Stmt* loop = For::make(x, 0, 10, aReduce);
Store* bStore = Store::make(b, {0}, Load::make(a, {0}, 1), 1);
Stmt* stmt = Block::make({aInit, loop, bStore});
stmt->accept(&analyzer);
// B -> A.
ASSERT_TRUE(analyzer.dependsDirectly(bStore, aReduce));
// B depends indirectly on the intializer of A, since the reduction depends
// on it.
ASSERT_FALSE(analyzer.dependsDirectly(bStore, aInit));
ASSERT_TRUE(analyzer.dependsIndirectly(bStore, aInit));
ASSERT_TRUE(analyzer.dependsDirectly(aReduce, aInit));
// B depends on the loop.
ASSERT_TRUE(analyzer.dependsDirectly(bStore, loop));
// A is in the loop and depends on other iterations.
ASSERT_TRUE(analyzer.dependsDirectly(aReduce, loop));
// The loop contents depend on the initializer too.
ASSERT_TRUE(analyzer.dependsDirectly(loop, aInit));
// Find loads within the reduction:
auto reduceLoads = NodeFinder<Load>::find(reduce.node());
// Pull out the access for the load inside the loop.
for (auto* load : reduceLoads) {
auto loopLoad = analyzer.accessFor(load);
// It should have 10 element long bounds.
ASSERT_TRUE(indexBoundsEquals(
loopLoad->bounds(), {Bound(new IntImm(0), new IntImm(9))}));
}
}
// Lowering a reduction doesn't affect dependency analysis.
TEST(MemDependency, MemDependencyCheckerLoopReduceExpanded) {
KernelScope kernel_scope;
BufHandle a("A", {10}, kInt);
BufHandle b("B", {10}, kInt);
VarHandle x("x", kInt);
using namespace analysis;
MemDependencyChecker analyzer;
/*
* A[0] = 0;
* for (int x = 0; x < 10; ++x) {
* A[0] = A[x] + 1;
* }
* B[0] = A[0];
*/
Store* aInit = Store::make(a, {0}, 0, 1);
ExprHandle aLoad = Load::make(a, {x}, 1);
Store* aReduce = Store::make(a, {0}, Add::make(aLoad, 1));
Stmt* loop = For::make(x, 0, 10, aReduce);
Store* bStore = Store::make(b, {0}, Load::make(a, {0}, 1), 1);
Stmt* stmt = Block::make({aInit, loop, bStore});
stmt->accept(&analyzer);
// B -> A.
ASSERT_TRUE(analyzer.dependsDirectly(bStore, aReduce));
// B depends indirectly on the intializer of A, since the reduction depends
// on it.
ASSERT_FALSE(analyzer.dependsDirectly(bStore, aInit));
ASSERT_TRUE(analyzer.dependsIndirectly(bStore, aInit));
ASSERT_TRUE(analyzer.dependsDirectly(aReduce, aInit));
// B depends on the loop.
ASSERT_TRUE(analyzer.dependsDirectly(bStore, loop));
// A is in the loop and depends on other iterations.
ASSERT_TRUE(analyzer.dependsDirectly(aReduce, loop));
// The loop contents depend on the initializer too.
ASSERT_TRUE(analyzer.dependsDirectly(loop, aInit));
// Pull out the access for the store inside the loop.
auto loopLoad = analyzer.accessFor(aLoad.node());
// It should have 10 element long bounds.
ASSERT_TRUE(indexBoundsEquals(
loopLoad->bounds(), {Bound(new IntImm(0), new IntImm(9))}));
}
// Can determine dependencies of outputs, through to inputs.
TEST(MemDependency, MemDependencyCheckerInputsOutputs) {
KernelScope kernel_scope;
BufHandle a("A", {10}, kInt);
BufHandle b("B", {10}, kInt);
VarHandle x("x", kInt);
// initialize analyzer with inputs and outputs.
analysis::MemDependencyChecker analyzer({a}, {b});
// Here's a Relu.
/*
* for (int x = 0; x < 10; ++x) {
* B[x] = Max(A[x], 0);
* }
*/
ExprHandle aLoad = Load::make(a, {x}, 1);
Store* bStore = Store::make(b, {x}, Max::make(aLoad, 0, true), 1);
Stmt* loop = For::make(x, 0, 10, bStore);
Stmt* stmt = Block::make({loop});
stmt->accept(&analyzer);
// Output depends indirectly on input.
ASSERT_TRUE(analyzer.dependsIndirectly(b.node(), a.node()));
// aLoad depends directly on the input A.
ASSERT_TRUE(analyzer.dependsDirectly(aLoad.node(), a.node()));
// bStore therefore depends directly on the input A.
ASSERT_TRUE(analyzer.dependsDirectly(bStore, a.node()));
// The output depends directly on the store.
ASSERT_TRUE(analyzer.dependsDirectly(b.node(), bStore));
// Check AccessInfo based overloads.
auto input = analyzer.input(a.node());
auto output = analyzer.output(b.node());
// Output depends indirectly on input.
ASSERT_TRUE(analyzer.dependsIndirectly(output, input));
// Not directly.
ASSERT_FALSE(analyzer.dependsDirectly(output, input));
// Not in reverse order.
ASSERT_FALSE(analyzer.dependsIndirectly(input, output));
// output -> bStore -> bLoad -> input.
auto storeAccess = analyzer.accessFor(bStore);
auto loadAccess = analyzer.accessFor(aLoad.node());
ASSERT_TRUE(analyzer.dependsDirectly(output, storeAccess));
ASSERT_TRUE(analyzer.dependsDirectly(loadAccess, input));
}
// Can tell if an output does not depend on an input.
TEST(MemDependency, MemDependencyCheckerOutputDoesntDepend) {
KernelScope kernel_scope;
BufHandle a("A", {10}, kInt);
BufHandle b("B", {10}, kInt);
VarHandle x("x", kInt);
// initialize analyzer with inputs and outputs.
analysis::MemDependencyChecker analyzer({a}, {b});
// Here's a dumb Relu.
/*
* for (int x = 0; x < 10; ++x) {
* B[x] = Max(x, 0);
* }
*/
Store* bStore = Store::make(b, {x}, Max::make(x, 0, true), 1);
Stmt* loop = For::make(x, 0, 10, bStore);
Stmt* stmt = Block::make({loop});
stmt->accept(&analyzer);
// Output does not depend indirectly on input.
ASSERT_FALSE(analyzer.dependsIndirectly(b.node(), a.node()));
// The output still depends directly on the store.
ASSERT_TRUE(analyzer.dependsDirectly(b.node(), bStore));
// Check AccessInfo based overloads.
auto input = analyzer.input(a.node());
auto output = analyzer.output(b.node());
// Output does not depend indirectly on input.
ASSERT_FALSE(analyzer.dependsIndirectly(output, input));
}
// Verify different loop extents produce accesses with different bounds, and
// that later accesses find dependencies that overlap their entire bound range.
TEST(MemDependency, MemDependencyCheckerLoopBounds) {
KernelScope kernel_scope;
BufHandle a("A", {10}, kInt);
BufHandle b("B", {10}, kInt);
BufHandle c("C", {10}, kInt);
VarHandle x("x", kInt);
using namespace analysis;
MemDependencyChecker analyzer({a}, {c});
// This enables using the execution order of the loops to determine if some
// loops are self dependent or not.
analyzer.allowLoopExecutionOrderAnalysis();
/*
* for (int x = 1; x < 10; ++x) {
* B[x] = A[x];
* }
* for (int x = 1; x < 9; ++x) {
* B[x] = B[x] * 2;
* }
* for (int x = 3; x < 4; ++x) {
* C[x] = A[x];
* }
* for (int x = 0; x < 10; ++x) {
* C[x] = B[x];
* }
*/
std::vector<Stmt*> stmts(
{For::make(x, 1, 10, Store::make(b, {x}, Load::make(a, {x}, 1), 1)),
For::make(
x,
1,
9,
Store::make(b, {x}, Mul::make(Load::make(b, {x}, 1), 2), 1)),
For::make(x, 3, 4, Store::make(c, {x}, Load::make(a, {x}, 1), 1)),
For::make(x, 0, 10, Store::make(c, {x}, Load::make(b, {x}, 1), 1))});
Stmt* stmt = Block::make(stmts);
stmt->accept(&analyzer);
auto input = analyzer.input(a.node());
auto output = analyzer.output(c.node());
// sanity check Output -> Input.
ASSERT_TRUE(analyzer.dependsIndirectly(output, input));
// Check the For loop dependencies:
// Last write to C depends on both writes to B since they contain the last
// write to at least one element.
ASSERT_TRUE(analyzer.dependsIndirectly(stmts[3], stmts[1]));
ASSERT_TRUE(analyzer.dependsIndirectly(stmts[3], stmts[0]));
// The last write to C does not depend on the other write to C.
ASSERT_FALSE(analyzer.dependsIndirectly(stmts[3], stmts[2]));
auto CB = [](int s, int e) { return Bound(new IntImm(s), new IntImm(e)); };
auto EQ = [](const IndexBounds& x, const IndexBounds& y) {
return indexBoundsEquals(x, y);
};
/* 0. Input: A[(0, 9)] - dependents: 1 5
* 1. Load: A[(1, 9)] - depends on: 0 - dependents: 2
* 2. Store: B[(1, 9)] - depends on: 1 - dependents: 3 7
* 3. Load: B[(1, 8)] - depends on: 2 - dependents: 4
* 4. Store: B[(1, 8)] - depends on: 3 - dependents: 7
* 5. Load: A[(3, 3)] - depends on: 0 - dependents: 6
* 6. Store: C[(3, 3)] - depends on: 5
* 7. Load: B[(0, 9)] - depends on: 2 4 - dependents: 8
* 8. Store: C[(0, 9)] - depends on: 7 - dependents: 9
* 9. Output: C[(0, 9)] - depends on: 8
*/
// Now let's look at the bounds of each access.
// There are 9 accesses in this Stmt, so this is exhaustive, we wont do this
// much.
auto history = analyzer.getHistory();
ASSERT_EQ(history.size(), 10);
const Var* aVar = a.node()->base_handle();
const Var* bVar = b.node()->base_handle();
const Var* cVar = c.node()->base_handle();
// The first access is the input A.
ASSERT_EQ(history[0]->type(), AccessType::Input);
ASSERT_EQ(history[0]->var(), aVar);
// It has the bounds of the producing Input.
ASSERT_TRUE(EQ(history[0]->bounds(), {CB(0, 9)}));
// sanity check the input we retrieved earlier matches.
ASSERT_EQ(history[0], input);
// The second access is the load of A in the first loop.
ASSERT_EQ(history[1]->type(), AccessType::Load);
ASSERT_EQ(history[1]->var(), aVar);
// It has the bounds of the loop, i.e. start == 1.
ASSERT_TRUE(EQ(history[1]->bounds(), {CB(1, 9)}));
// It reads from A, so it should have a dependency on the last write to this
// range - with is the input.
ASSERT_EQ(history[1]->dependencies().size(), 1);
ASSERT_TRUE(history[1]->hasDependency(history[0]));
// The third access is the store into B in the first loop.
ASSERT_EQ(history[2]->type(), AccessType::Store);
ASSERT_EQ(history[2]->var(), bVar);
// It also has the bounds of the loop, i.e. start == 1.
ASSERT_TRUE(EQ(history[2]->bounds(), {CB(1, 9)}));
// The previous load is in its RHS, so it depends on it.
ASSERT_EQ(history[2]->dependencies().size(), 1);
ASSERT_TRUE(history[2]->hasDependency(history[1]));
// The third access is the load from B in the second loop.
ASSERT_EQ(history[3]->type(), AccessType::Load);
ASSERT_EQ(history[3]->var(), bVar);
// It has the bounds of the second loop, i.e. >= 1 < 9.
ASSERT_TRUE(EQ(history[3]->bounds(), {CB(1, 8)}));
// It reads from B in a smaller range, so should depend on the previous
// store.
ASSERT_EQ(history[3]->dependencies().size(), 1);
ASSERT_TRUE(history[3]->hasDependency(history[2]));
// The fourth: the store to B in the second loop.
ASSERT_EQ(history[4]->type(), AccessType::Store);
ASSERT_EQ(history[4]->var(), bVar);
// It also has the bounds of the second loop.
ASSERT_TRUE(EQ(history[4]->bounds(), {CB(1, 8)}));
// The previous load is in its RHS, so it depends on it as before.
ASSERT_EQ(history[4]->dependencies().size(), 1);
ASSERT_TRUE(history[4]->hasDependency(history[3]));
// The fifth access is the load is from the 3rd loop, and skips previous B
// accesses.
ASSERT_EQ(history[5]->type(), AccessType::Load);
ASSERT_EQ(history[5]->var(), aVar);
// It has the bounds of the third loop: >= 3 < 4.
ASSERT_TRUE(EQ(history[5]->bounds(), {CB(3, 3)}));
// It depends on the last thing to write to A, which is the A input.
ASSERT_EQ(history[5]->dependencies().size(), 1);
ASSERT_TRUE(history[5]->hasDependency(history[0]));
// Sixth: the store into the output C.
ASSERT_EQ(history[6]->type(), AccessType::Store);
ASSERT_EQ(history[6]->var(), cVar);
// It also has the bounds of the third loop.
ASSERT_TRUE(EQ(history[6]->bounds(), {CB(3, 3)}));
// The previous load is in its RHS, so it depends on it as always.
ASSERT_EQ(history[6]->dependencies().size(), 1);
ASSERT_TRUE(history[6]->hasDependency(history[5]));
// The seventh access is the load of B in the fourth loop.
ASSERT_EQ(history[7]->type(), AccessType::Load);
ASSERT_EQ(history[7]->var(), bVar);
// It has the bounds of the final loop, >= 0 < 10
ASSERT_TRUE(EQ(history[7]->bounds(), {CB(0, 9)}));
// The bounds of this read are larger than the bounds of the previous write,
// so it depends on both previous Stores to B.
ASSERT_EQ(history[7]->dependencies().size(), 2);
ASSERT_TRUE(history[7]->hasDependency(history[2]));
ASSERT_TRUE(history[7]->hasDependency(history[4]));
// Eight: the final store into the output C.
ASSERT_EQ(history[8]->type(), AccessType::Store);
ASSERT_EQ(history[8]->var(), cVar);
// It also has the bounds of the final loop.
ASSERT_TRUE(EQ(history[8]->bounds(), {CB(0, 9)}));
// The previous load is in its RHS, so it depends on it as always.
ASSERT_EQ(history[8]->dependencies().size(), 1);
ASSERT_TRUE(history[8]->hasDependency(history[7]));
// The last access represents the output Buf.
ASSERT_EQ(history[9]->type(), AccessType::Output);
ASSERT_EQ(history[9]->var(), cVar);
// It has the bounds of the output Buf.
ASSERT_TRUE(EQ(history[9]->bounds(), {CB(0, 9)}));
// sanity check the input we retrieved earlier matches.
ASSERT_EQ(history[9], output);
// It depends on the last write to C only.
ASSERT_EQ(history[9]->dependencies().size(), 1);
ASSERT_TRUE(history[9]->hasDependency(history[8]));
}
// Verify that we can still infer bounds when the loop var is offset.
TEST(MemDependency, MemDependencyCheckerLoopBoundsIndexShift) {
KernelScope kernel_scope;
BufHandle a("A", {10}, kInt);
BufHandle b("B", {10}, kInt);
VarHandle x("x", kInt);
using namespace analysis;
MemDependencyChecker analyzer({a}, {b});
// This enables using the execution order of the loops to determine if some
// loops are self dependent or not.
analyzer.allowLoopExecutionOrderAnalysis();
/*
* for (int x = 1; x < 10; x++) {
* A[x] = A[x - 1];
* }
* for (int x = 0; x < 9; x++) {
* A[x] = A[x + 1];
* }
* for (int x = 0; x < 9; x++) {
* A[9 - x] = A[8 - x];
* }
* for (int x = 0; x < 10; x++) {
* A[x] = A[9 - x];
* }
* for (int x = 0; x < 10; x++) {
* B[x] = A[x];
* }
*/
Stmt* stmt = Block::make(
{For::make(x, 1, 10, Store::make(a, {x}, Load::make(a, {x - 1}, 1), 1)),
For::make(x, 0, 9, Store::make(a, {x}, Load::make(a, {x + 1}, 1), 1)),
For::make(
x,
0,
9,
Store::make(
a,
{ExprHandle(9) - x},
Load::make(a, {ExprHandle(8) - x}, 1),
1)),
For::make(
x,
0,
10,
Store::make(a, {x}, Load::make(a, {ExprHandle(9) - x}, 1), 1)),
For::make(x, 0, 10, Store::make(b, {x}, Load::make(a, {x}, 1), 1))});
stmt->accept(&analyzer);
// Sanity check output depends on Input.
ASSERT_TRUE(analyzer.dependsIndirectly(b.node(), a.node()));
auto CB = [](int s, int e) { return Bound(new IntImm(s), new IntImm(e)); };
auto EQ = [](const IndexBounds& x, const IndexBounds& y) {
return indexBoundsEquals(x, y);
};
/* 0. Input: A[(0, 9)] - dependents: 1
* 1. Load: A[(0, 8)] - depends on: 0 2 - dependents: 2
* 2. Store: A[(1, 9)] - depends on: 1 - dependents: 1 3
* 3. Load: A[(1, 9)] - depends on: 2 - dependents: 4
* 4. Store: A[(0, 8)] - depends on: 3 - dependents: 5 7
* 5. Load: A[(0, 8)] - depends on: 4 - dependents: 6
* 6. Store: A[(1, 9)] - depends on: 5 - dependents: 7
* 7. Load: A[(0, 9)] - depends on: 4 6 8 - dependents: 8
* 8. Store: A[(0, 9)] - depends on: 7 - dependents: 7 9
* 9. Load: A[(0, 9)] - depends on: 8 - dependents: 10
* 10. Store: B[(0, 9)] - depends on: 9 - dependents: 11
* 11. Output: B[(0, 9)] - depends on: 10
*/
// Now let's look at the bounds of each access.
auto history = analyzer.getHistory();
ASSERT_EQ(history.size(), 12);
const Var* aVar = a.node()->base_handle();
const Var* bVar = b.node()->base_handle();