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
|
// RUN: mlir-opt -fold-memref-alias-ops -split-input-file %s | FileCheck %s
func.func @fold_static_stride_subview_with_load(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index) -> f32 {
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][2, 3] : memref<12x32xf32> to memref<4x4xf32, strided<[64, 3], offset: ?>>
%1 = memref.load %0[%arg3, %arg4] : memref<4x4xf32, strided<[64, 3], offset: ?>>
return %1 : f32
}
// CHECK-DAG: #[[MAP0:.+]] = affine_map<()[s0, s1] -> (s0 + s1 * 2)>
// CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0, s1] -> (s0 + s1 * 3)>
// CHECK: func @fold_static_stride_subview_with_load
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: memref<12x32xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9_]+]]: index
// CHECK-DAG: %[[I1:.+]] = affine.apply #[[MAP0]]()[%[[ARG1]], %[[ARG3]]]
// CHECK-DAG: %[[I2:.+]] = affine.apply #[[MAP1]]()[%[[ARG2]], %[[ARG4]]]
// CHECK: memref.load %[[ARG0]][%[[I1]], %[[I2]]]
// -----
func.func @fold_dynamic_stride_subview_with_load(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index, %arg5 : index, %arg6 : index) -> f32 {
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][%arg5, %arg6] :
memref<12x32xf32> to memref<4x4xf32, strided<[?, ?], offset: ?>>
%1 = memref.load %0[%arg3, %arg4] : memref<4x4xf32, strided<[?, ?], offset: ?>>
return %1 : f32
}
// CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0, s1, s2] -> (s0 + s1 * s2)>
// CHECK: func @fold_dynamic_stride_subview_with_load
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: memref<12x32xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG5:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG6:[a-zA-Z0-9_]+]]: index
// CHECK-DAG: %[[I1:.+]] = affine.apply #[[MAP]]()[%[[ARG1]], %[[ARG3]], %[[ARG5]]]
// CHECK-DAG: %[[I2:.+]] = affine.apply #[[MAP]]()[%[[ARG2]], %[[ARG4]], %[[ARG6]]]
// CHECK: memref.load %[[ARG0]][%[[I1]], %[[I2]]]
// -----
func.func @fold_static_stride_subview_with_store(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index, %arg5 : f32) {
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][2, 3] :
memref<12x32xf32> to memref<4x4xf32, strided<[64, 3], offset: ?>>
memref.store %arg5, %0[%arg3, %arg4] : memref<4x4xf32, strided<[64, 3], offset: ?>>
return
}
// CHECK-DAG: #[[MAP0:.+]] = affine_map<()[s0, s1] -> (s0 + s1 * 2)>
// CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0, s1] -> (s0 + s1 * 3)>
// CHECK: func @fold_static_stride_subview_with_store
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: memref<12x32xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9_]+]]: index
// CHECK-DAG: %[[I1:.+]] = affine.apply #[[MAP0]]()[%[[ARG1]], %[[ARG3]]]
// CHECK-DAG: %[[I2:.+]] = affine.apply #[[MAP1]]()[%[[ARG2]], %[[ARG4]]]
// CHECK: memref.store %{{.+}}, %[[ARG0]][%[[I1]], %[[I2]]]
// -----
func.func @fold_dynamic_stride_subview_with_store(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index, %arg5 : index, %arg6 : index, %arg7 : f32) {
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][%arg5, %arg6] :
memref<12x32xf32> to memref<4x4xf32, strided<[?, ?], offset: ?>>
memref.store %arg7, %0[%arg3, %arg4] : memref<4x4xf32, strided<[?, ?], offset: ?>>
return
}
// CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0, s1, s2] -> (s0 + s1 * s2)>
// CHECK: func @fold_dynamic_stride_subview_with_store
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: memref<12x32xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG5:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG6:[a-zA-Z0-9_]+]]: index
// CHECK-DAG: %[[I1:.+]] = affine.apply #[[MAP]]()[%[[ARG1]], %[[ARG3]], %[[ARG5]]]
// CHECK-DAG: %[[I2:.+]] = affine.apply #[[MAP]]()[%[[ARG2]], %[[ARG4]], %[[ARG6]]]
// CHECK: memref.store %{{.+}}, %[[ARG0]][%[[I1]], %[[I2]]]
// -----
func.func @fold_subview_with_transfer_read_0d(
%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index)
-> vector<f32> {
%f1 = arith.constant 1.0 : f32
%0 = memref.subview %arg0[%arg1, %arg2][1, 1][1, 1] : memref<12x32xf32> to memref<f32, strided<[], offset: ?>>
%1 = vector.transfer_read %0[], %f1 : memref<f32, strided<[], offset: ?>>, vector<f32>
return %1 : vector<f32>
}
// CHECK: func @fold_subview_with_transfer_read_0d
// CHECK-SAME: %[[MEM:[a-zA-Z0-9_]+]]: memref<12x32xf32>
// CHECK-SAME: %[[SZ0:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[SZ1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ST1:[a-zA-Z0-9_]+]]: index
// CHECK: vector.transfer_read %[[MEM]][%[[SZ0]], %[[SZ1]]]
// -----
func.func @fold_subview_with_transfer_read(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index, %arg5 : index, %arg6 : index) -> vector<4xf32> {
%f1 = arith.constant 1.0 : f32
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][%arg5, %arg6] : memref<12x32xf32> to memref<4x4xf32, strided<[?, ?], offset: ?>>
%1 = vector.transfer_read %0[%arg3, %arg4], %f1 {in_bounds = [true]} : memref<4x4xf32, strided<[?, ?], offset: ?>>, vector<4xf32>
return %1 : vector<4xf32>
}
// CHECK: func @fold_subview_with_transfer_read
// Can't fold this atm since we don't emit the proper vector.extract_strided_slice.
// CHECK: memref.subview
// -----
func.func @fold_static_stride_subview_with_transfer_write_0d(
%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index,
%v : vector<f32>) {
%f1 = arith.constant 1.0 : f32
%0 = memref.subview %arg0[%arg1, %arg2][1, 1][1, 1] : memref<12x32xf32> to memref<f32, strided<[], offset: ?>>
vector.transfer_write %v, %0[] {in_bounds = []} : vector<f32>, memref<f32, strided<[], offset: ?>>
return
}
// CHECK: func @fold_static_stride_subview_with_transfer_write_0d
// CHECK-SAME: %[[MEM:[a-zA-Z0-9_]+]]: memref<12x32xf32>
// CHECK-SAME: %[[SZ0:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[SZ1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ST1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[V:[a-zA-Z0-9_]+]]: vector<f32>
// CHECK: vector.transfer_write %[[V]], %[[MEM]][%[[SZ0]], %[[SZ1]]]
// -----
func.func @fold_static_stride_subview_with_transfer_write(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index, %arg5: index, %arg6 : index, %arg7 : vector<4xf32>) {
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][%arg5, %arg6] :
memref<12x32xf32> to memref<4x4xf32, strided<[?, ?], offset: ?>>
vector.transfer_write %arg7, %0[%arg3, %arg4] {in_bounds = [true]} : vector<4xf32>, memref<4x4xf32, strided<[?, ?], offset: ?>>
return
}
// CHECK: func @fold_static_stride_subview_with_transfer_write
// Can't fold this atm since we don't emit the proper vector.extract_strided_slice.
// CHECK: memref.subview
// -----
func.func @fold_rank_reducing_subview_with_load
(%arg0 : memref<?x?x?x?x?x?xf32>, %arg1 : index, %arg2 : index,
%arg3 : index, %arg4 : index, %arg5 : index, %arg6 : index,
%arg7 : index, %arg8 : index, %arg9 : index, %arg10: index,
%arg11 : index, %arg12 : index, %arg13 : index, %arg14: index,
%arg15 : index, %arg16 : index) -> f32 {
%0 = memref.subview %arg0[%arg1, %arg2, %arg3, %arg4, %arg5, %arg6][4, 1, 1, 4, 1, 1][%arg7, %arg8, %arg9, %arg10, %arg11, %arg12] : memref<?x?x?x?x?x?xf32> to memref<4x1x4x1xf32, strided<[?, ?, ?, ?], offset: ?>>
%1 = memref.load %0[%arg13, %arg14, %arg15, %arg16] : memref<4x1x4x1xf32, strided<[?, ?, ?, ?], offset: ?>>
return %1 : f32
}
// CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0, s1, s2] -> (s0 + s1 * s2)>
// CHECK: func @fold_rank_reducing_subview_with_load
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: memref<?x?x?x?x?x?xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG5:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG6:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG7:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG8:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG9:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG10:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG11:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG12:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG13:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG14:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG15:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG16:[a-zA-Z0-9_]+]]: index
// CHECK-DAG: %[[I0:.+]] = affine.apply #[[MAP]]()[%[[ARG1]], %[[ARG13]], %[[ARG7]]]
// CHECK-DAG: %[[I2:.+]] = affine.apply #[[MAP]]()[%[[ARG3]], %[[ARG14]], %[[ARG9]]]
// CHECK-DAG: %[[I3:.+]] = affine.apply #[[MAP]]()[%[[ARG4]], %[[ARG15]], %[[ARG10]]]
// CHECK-DAG: %[[I4:.+]] = affine.apply #[[MAP]]()[%[[ARG5]], %[[ARG16]], %[[ARG11]]]
// CHECK: memref.load %[[ARG0]][%[[I0]], %[[ARG2]], %[[I2]], %[[I3]], %[[I4]], %[[ARG6]]]
// -----
func.func @fold_vector_transfer_read_with_rank_reduced_subview(
%arg0 : memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>>,
%arg1: index, %arg2 : index, %arg3 : index, %arg4: index, %arg5 : index,
%arg6 : index) -> vector<4xf32> {
%cst = arith.constant 0.0 : f32
%0 = memref.subview %arg0[0, %arg1, %arg2] [1, %arg3, %arg4] [1, 1, 1]
: memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>> to
memref<?x?xf32, strided<[?, ?], offset: ?>>
%1 = vector.transfer_read %0[%arg5, %arg6], %cst {in_bounds = [true]}
: memref<?x?xf32, strided<[?, ?], offset: ?>>, vector<4xf32>
return %1 : vector<4xf32>
}
// CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK: func @fold_vector_transfer_read_with_rank_reduced_subview
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG5:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG6:[a-zA-Z0-9]+]]: index
// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
// CHECK-DAG: %[[IDX0:.+]] = affine.apply #[[MAP1]]()[%[[ARG1]], %[[ARG5]]]
// CHECK-DAG: %[[IDX1:.+]] = affine.apply #[[MAP1]]()[%[[ARG2]], %[[ARG6]]]
// CHECK: vector.transfer_read %[[ARG0]][%[[C0]], %[[IDX0]], %[[IDX1]]], %{{.*}} : memref<?x?x?xf32
// -----
func.func @fold_vector_transfer_write_with_rank_reduced_subview(
%arg0 : memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>>,
%arg1 : vector<4xf32>, %arg2: index, %arg3 : index, %arg4 : index,
%arg5: index, %arg6 : index, %arg7 : index) {
%cst = arith.constant 0.0 : f32
%0 = memref.subview %arg0[0, %arg2, %arg3] [1, %arg4, %arg5] [1, 1, 1]
: memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>> to
memref<?x?xf32, strided<[?, ?], offset: ?>>
vector.transfer_write %arg1, %0[%arg6, %arg7] {in_bounds = [true]}
: vector<4xf32>, memref<?x?xf32, strided<[?, ?], offset: ?>>
return
}
// CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK: func @fold_vector_transfer_write_with_rank_reduced_subview
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: vector<4xf32>
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG5:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG6:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG7:[a-zA-Z0-9]+]]: index
// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
// CHECK-DAG: %[[IDX0:.+]] = affine.apply #[[MAP1]]()[%[[ARG2]], %[[ARG6]]]
// CHECK-DAG: %[[IDX1:.+]] = affine.apply #[[MAP1]]()[%[[ARG3]], %[[ARG7]]]
// CHECK-DAG: vector.transfer_write %[[ARG1]], %[[ARG0]][%[[C0]], %[[IDX0]], %[[IDX1]]] {in_bounds = [true]} : vector<4xf32>, memref<?x?x?xf32
// -----
func.func @fold_vector_transfer_write_with_inner_rank_reduced_subview(
%arg0 : memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>>,
%arg1 : vector<4xf32>, %arg2: index, %arg3 : index, %arg4 : index,
%arg5: index, %arg6 : index, %arg7 : index) {
%cst = arith.constant 0.0 : f32
%0 = memref.subview %arg0[%arg2, %arg3, 0] [%arg4, %arg5, 1] [1, 1, 1]
: memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>> to
memref<?x?xf32, strided<[?, ?], offset: ?>>
vector.transfer_write %arg1, %0[%arg6, %arg7] {in_bounds = [true]}
: vector<4xf32>, memref<?x?xf32, strided<[?, ?], offset: ?>>
return
}
// CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2) -> (d1)>
// CHECK: func @fold_vector_transfer_write_with_inner_rank_reduced_subview
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: memref<?x?x?xf32, strided<[?, ?, ?], offset: ?>>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: vector<4xf32>
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG5:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG6:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG7:[a-zA-Z0-9]+]]: index
// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
// CHECK-DAG: %[[IDX0:.+]] = affine.apply #[[MAP1]]()[%[[ARG2]], %[[ARG6]]]
// CHECK-DAG: %[[IDX1:.+]] = affine.apply #[[MAP1]]()[%[[ARG3]], %[[ARG7]]]
// CHECK-DAG: vector.transfer_write %[[ARG1]], %[[ARG0]][%[[IDX0]], %[[IDX1]], %[[C0]]]
// CHECK-SAME: {in_bounds = [true], permutation_map = #[[MAP2]]} : vector<4xf32>, memref<?x?x?xf32
// -----
// Test with affine.load/store ops. We only do a basic test here since the
// logic is identical to that with memref.load/store ops. The same affine.apply
// ops would be generated.
// CHECK-LABEL: func @fold_static_stride_subview_with_affine_load_store
func.func @fold_static_stride_subview_with_affine_load_store(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index) -> f32 {
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][2, 3] : memref<12x32xf32> to memref<4x4xf32, strided<[64, 3], offset: ?>>
%1 = affine.load %0[%arg3, %arg4] : memref<4x4xf32, strided<[64, 3], offset: ?>>
// CHECK-NEXT: affine.apply
// CHECK-NEXT: affine.apply
// CHECK-NEXT: affine.load
affine.store %1, %0[%arg3, %arg4] : memref<4x4xf32, strided<[64, 3], offset: ?>>
// CHECK-NEXT: affine.apply
// CHECK-NEXT: affine.apply
// CHECK-NEXT: affine.store
// CHECK-NEXT: return
return %1 : f32
}
// -----
// CHECK-DAG: #[[$MAP:.*]] = affine_map<()[s0, s1] -> (s0 * 6 + s1)>
// CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_expand_shape
// CHECK-SAME: (%[[ARG0:.*]]: memref<12x32xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index) -> f32 {
func.func @fold_static_stride_subview_with_affine_load_store_expand_shape(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index) -> f32 {
%0 = memref.expand_shape %arg0 [[0, 1], [2]] : memref<12x32xf32> into memref<2x6x32xf32>
%1 = affine.load %0[%arg1, %arg2, %arg3] : memref<2x6x32xf32>
return %1 : f32
}
// CHECK: %[[INDEX:.*]] = affine.apply #[[$MAP]]()[%[[ARG1]], %[[ARG2]]]
// CHECK-NEXT: %[[RESULT:.*]] = affine.load %[[ARG0]][%[[INDEX]], %[[ARG3]]] : memref<12x32xf32>
// CHECK-NEXT: return %[[RESULT]] : f32
// -----
// CHECK-DAG: #[[$MAP0:.*]] = affine_map<()[s0] -> (s0 floordiv 6)>
// CHECK-DAG: #[[$MAP1:.*]] = affine_map<()[s0] -> (s0 mod 6)>
// CHECK-LABEL: @fold_static_stride_subview_with_affine_load_store_collapse_shape
// CHECK-SAME: (%[[ARG0:.*]]: memref<2x6x32xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index)
func.func @fold_static_stride_subview_with_affine_load_store_collapse_shape(%arg0 : memref<2x6x32xf32>, %arg1 : index, %arg2 : index) -> f32 {
%0 = memref.collapse_shape %arg0 [[0, 1], [2]] : memref<2x6x32xf32> into memref<12x32xf32>
%1 = affine.load %0[%arg1, %arg2] : memref<12x32xf32>
return %1 : f32
}
// CHECK-NEXT: %[[MODIFIED_INDEX0:.*]] = affine.apply #[[$MAP0]]()[%[[ARG1]]]
// CHECK-NEXT: %[[MODIFIED_INDEX1:.*]] = affine.apply #[[$MAP1]]()[%[[ARG1]]]
// CHECK-NEXT: %[[RESULT:.*]] = affine.load %[[ARG0]][%[[MODIFIED_INDEX0]], %[[MODIFIED_INDEX1]], %[[ARG2]]] : memref<2x6x32xf32>
// CHECK-NEXT: return %[[RESULT]] : f32
// -----
// CHECK-DAG: #[[$MAP:.*]] = affine_map<()[s0, s1, s2] -> (s0 * 6 + s1 * 3 + s2)>
// CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_expand_shape_3d
// CHECK-SAME: (%[[ARG0:.*]]: memref<12x32xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index, %[[ARG4:.*]]: index) -> f32 {
func.func @fold_static_stride_subview_with_affine_load_store_expand_shape_3d(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4: index) -> f32 {
%0 = memref.expand_shape %arg0 [[0, 1, 2], [3]] : memref<12x32xf32> into memref<2x2x3x32xf32>
%1 = affine.load %0[%arg1, %arg2, %arg3, %arg4] : memref<2x2x3x32xf32>
return %1 : f32
}
// CHECK: %[[INDEX:.*]] = affine.apply #[[$MAP]]()[%[[ARG1]], %[[ARG2]], %[[ARG3]]]
// CHECK-NEXT: %[[RESULT:.*]] = affine.load %[[ARG0]][%[[INDEX]], %[[ARG4]]] : memref<12x32xf32>
// CHECK-NEXT: return %[[RESULT]] : f32
// -----
// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0 * 1024 + d1)>
// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0 + d1)>
// CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_expand_shape
// CHECK-SAME: (%[[ARG0:.*]]: memref<1024x1024xf32>, %[[ARG1:.*]]: memref<1xf32>, %[[ARG2:.*]]: index)
func.func @fold_static_stride_subview_with_affine_load_store_expand_shape(%arg0: memref<1024x1024xf32>, %arg1: memref<1xf32>, %arg2: index) -> f32 {
%0 = memref.expand_shape %arg0 [[0, 1], [2, 3]] : memref<1024x1024xf32> into memref<1x1024x1024x1xf32>
affine.for %arg3 = 0 to 1 {
affine.for %arg4 = 0 to 1024 {
affine.for %arg5 = 0 to 1020 {
affine.for %arg6 = 0 to 1 {
%1 = affine.load %0[%arg3, %arg4, %arg5, %arg6] : memref<1x1024x1024x1xf32>
affine.store %1, %arg1[%arg2] : memref<1xf32>
}
}
}
}
%2 = affine.load %arg1[%arg2] : memref<1xf32>
return %2 : f32
}
// CHECK-NEXT: affine.for %[[ARG3:.*]] = 0 to 1 {
// CHECK-NEXT: affine.for %[[ARG4:.*]] = 0 to 1024 {
// CHECK-NEXT: affine.for %[[ARG5:.*]] = 0 to 1020 {
// CHECK-NEXT: affine.for %[[ARG6:.*]] = 0 to 1 {
// CHECK-NEXT: %[[IDX1:.*]] = affine.apply #[[$MAP0]](%[[ARG3]], %[[ARG4]])
// CHECK-NEXT: %[[IDX2:.*]] = affine.apply #[[$MAP1]](%[[ARG5]], %[[ARG6]])
// CHECK-NEXT: affine.load %[[ARG0]][%[[IDX1]], %[[IDX2]]] : memref<1024x1024xf32>
// -----
// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0 * 1025 + d1)>
// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0 + d1)>
// CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_expand_shape_when_access_index_is_an_expression
// CHECK-SAME: (%[[ARG0:.*]]: memref<1024x1024xf32>, %[[ARG1:.*]]: memref<1xf32>, %[[ARG2:.*]]: index)
func.func @fold_static_stride_subview_with_affine_load_store_expand_shape_when_access_index_is_an_expression(%arg0: memref<1024x1024xf32>, %arg1: memref<1xf32>, %arg2: index) -> f32 {
%0 = memref.expand_shape %arg0 [[0, 1], [2, 3]] : memref<1024x1024xf32> into memref<1x1024x1024x1xf32>
affine.for %arg3 = 0 to 1 {
affine.for %arg4 = 0 to 1024 {
affine.for %arg5 = 0 to 1020 {
affine.for %arg6 = 0 to 1 {
%1 = affine.load %0[%arg3, %arg4 + %arg3, %arg5, %arg6] : memref<1x1024x1024x1xf32>
affine.store %1, %arg1[%arg2] : memref<1xf32>
}
}
}
}
%2 = affine.load %arg1[%arg2] : memref<1xf32>
return %2 : f32
}
// CHECK-NEXT: affine.for %[[ARG3:.*]] = 0 to 1 {
// CHECK-NEXT: affine.for %[[ARG4:.*]] = 0 to 1024 {
// CHECK-NEXT: affine.for %[[ARG5:.*]] = 0 to 1020 {
// CHECK-NEXT: affine.for %[[ARG6:.*]] = 0 to 1 {
// CHECK-NEXT: %[[TMP1:.*]] = affine.apply #[[$MAP0]](%[[ARG3]], %[[ARG4]])
// CHECK-NEXT: %[[TMP3:.*]] = affine.apply #[[$MAP1]](%[[ARG5]], %[[ARG6]])
// CHECK-NEXT: affine.load %[[ARG0]][%[[TMP1]], %[[TMP3]]] : memref<1024x1024xf32>
// -----
// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0) -> (d0 * 1024)>
// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0 + d1)>
// CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_expand_shape_with_constant_access_index
// CHECK-SAME: (%[[ARG0:.*]]: memref<1024x1024xf32>, %[[ARG1:.*]]: memref<1xf32>, %[[ARG2:.*]]: index)
func.func @fold_static_stride_subview_with_affine_load_store_expand_shape_with_constant_access_index(%arg0: memref<1024x1024xf32>, %arg1: memref<1xf32>, %arg2: index) -> f32 {
%0 = memref.expand_shape %arg0 [[0, 1], [2, 3]] : memref<1024x1024xf32> into memref<1x1024x1024x1xf32>
%cst = arith.constant 0 : index
affine.for %arg3 = 0 to 1 {
affine.for %arg4 = 0 to 1024 {
affine.for %arg5 = 0 to 1020 {
affine.for %arg6 = 0 to 1 {
%1 = memref.load %0[%arg3, %cst, %arg5, %arg6] : memref<1x1024x1024x1xf32>
memref.store %1, %arg1[%arg2] : memref<1xf32>
}
}
}
}
%2 = memref.load %arg1[%arg2] : memref<1xf32>
return %2 : f32
}
// CHECK-NEXT: affine.for %[[ARG3:.*]] = 0 to 1 {
// CHECK-NEXT: affine.for %[[ARG4:.*]] = 0 to 1024 {
// CHECK-NEXT: affine.for %[[ARG5:.*]] = 0 to 1020 {
// CHECK-NEXT: affine.for %[[ARG6:.*]] = 0 to 1 {
// CHECK-NEXT: %[[TMP1:.*]] = affine.apply #[[$MAP0]](%[[ARG3]])
// CHECK-NEXT: %[[TMP2:.*]] = affine.apply #[[$MAP1]](%[[ARG5]], %[[ARG6]])
// CHECK-NEXT: memref.load %[[ARG0]][%[[TMP1]], %[[TMP2]]] : memref<1024x1024xf32>
// -----
// CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_collapse_shape_with_0d_result
// CHECK-SAME: (%[[ARG0:.*]]: memref<1xf32>, %[[ARG1:.*]]: memref<1xf32>)
func.func @fold_static_stride_subview_with_affine_load_store_collapse_shape_with_0d_result(%arg0: memref<1xf32>, %arg1: memref<1xf32>) -> memref<1xf32> {
%0 = memref.collapse_shape %arg0 [] : memref<1xf32> into memref<f32>
affine.for %arg2 = 0 to 3 {
%1 = affine.load %0[] : memref<f32>
affine.store %1, %arg1[0] : memref<1xf32>
}
return %arg1 : memref<1xf32>
}
// CHECK-NEXT: %[[ZERO:.*]] = arith.constant 0 : index
// CHECK-NEXT: affine.for %{{.*}} = 0 to 3 {
// CHECK-NEXT: affine.load %[[ARG0]][%[[ZERO]]] : memref<1xf32>
// -----
// CHECK: #[[$map:.*]] = affine_map<()[s0] -> (s0 + 2)>
// CHECK-LABEL: func @subview_of_subview(
// CHECK-SAME: %[[m:.*]]: memref<1x1024xf32, 3>, %[[pos:.*]]: index
// CHECK: %[[add:.*]] = affine.apply #[[$map]]()[%arg1]
// CHECK: memref.subview %arg0[4, %[[add]]] [1, 1] [1, 1] : memref<1x1024xf32, 3> to memref<f32, strided<[], offset: ?>, 3>
func.func @subview_of_subview(%m: memref<1x1024xf32, 3>, %pos: index)
-> memref<f32, strided<[], offset: ?>, 3>
{
%0 = memref.subview %m[3, %pos] [1, 2] [1, 1]
: memref<1x1024xf32, 3>
to memref<1x2xf32, strided<[1024, 2], offset: ?>, 3>
%1 = memref.subview %0[1, 2] [1, 1] [1, 1]
: memref<1x2xf32, strided<[1024, 2], offset: ?>, 3>
to memref<f32, strided<[], offset: ?>, 3>
return %1 : memref<f32, strided<[], offset: ?>, 3>
}
// -----
// CHECK-LABEL: func @subview_of_subview_rank_reducing(
// CHECK-SAME: %[[m:.*]]: memref<?x?x?xf32>
// CHECK: memref.subview %arg0[3, 7, 8] [1, 1, 1] [1, 1, 1] : memref<?x?x?xf32> to memref<f32, strided<[], offset: ?>>
func.func @subview_of_subview_rank_reducing(%m: memref<?x?x?xf32>,
%sz: index, %pos: index)
-> memref<f32, strided<[], offset: ?>>
{
%0 = memref.subview %m[3, 1, 8] [1, %sz, 1] [1, 1, 1]
: memref<?x?x?xf32>
to memref<?xf32, strided<[1], offset: ?>>
%1 = memref.subview %0[6] [1] [1]
: memref<?xf32, strided<[1], offset: ?>>
to memref<f32, strided<[], offset: ?>>
return %1 : memref<f32, strided<[], offset: ?>>
}
// -----
// CHECK-LABEL: func @fold_load_keep_nontemporal(
// CHECK: memref.load %{{.+}}[%{{.+}}, %{{.+}}] {nontemporal = true}
func.func @fold_load_keep_nontemporal(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index) -> f32 {
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][2, 3] : memref<12x32xf32> to memref<4x4xf32, strided<[64, 3], offset: ?>>
%1 = memref.load %0[%arg3, %arg4] {nontemporal = true }: memref<4x4xf32, strided<[64, 3], offset: ?>>
return %1 : f32
}
// -----
// CHECK-LABEL: func @fold_store_keep_nontemporal(
// CHECK: memref.store %{{.+}}, %{{.+}}[%{{.+}}, %{{.+}}] {nontemporal = true} : memref<12x32xf32>
func.func @fold_store_keep_nontemporal(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index, %arg5 : f32) {
%0 = memref.subview %arg0[%arg1, %arg2][4, 4][2, 3] :
memref<12x32xf32> to memref<4x4xf32, strided<[64, 3], offset: ?>>
memref.store %arg5, %0[%arg3, %arg4] {nontemporal=true}: memref<4x4xf32, strided<[64, 3], offset: ?>>
return
}
// -----
func.func @fold_gpu_subgroup_mma_load_matrix_1d(%src: memref<?xvector<4xf32>>, %offset: index, %i: index) -> !gpu.mma_matrix<16x16xf16, "COp"> {
%subview = memref.subview %src[%offset] [81920] [1] : memref<?xvector<4xf32>> to memref<81920xvector<4xf32>, strided<[1], offset: ?>>
%matrix = gpu.subgroup_mma_load_matrix %subview[%i] {leadDimension = 160 : index} : memref<81920xvector<4xf32>, strided<[1], offset: ?>> -> !gpu.mma_matrix<16x16xf16, "COp">
return %matrix: !gpu.mma_matrix<16x16xf16, "COp">
}
// CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK: func.func @fold_gpu_subgroup_mma_load_matrix_1d
// CHECK-SAME: (%[[SRC:.+]]: memref<?xvector<4xf32>>, %[[OFFSET:.+]]: index, %[[I:.+]]: index)
// CHECK: %[[APPLY:.+]] = affine.apply #[[MAP]]()[%[[OFFSET]], %[[I]]]
// CHECK: %[[LOAD:.+]] = gpu.subgroup_mma_load_matrix %[[SRC]][%[[APPLY]]] {leadDimension = 160 : index} : memref<?xvector<4xf32>> -> !gpu.mma_matrix<16x16xf16, "COp">
// CHECK: return %[[LOAD]]
// -----
func.func @fold_gpu_subgroup_mma_store_matrix_1d(%dst: memref<?xvector<4xf32>>, %offset: index, %i: index, %matrix: !gpu.mma_matrix<16x16xf16, "COp">) {
%subview = memref.subview %dst[%offset] [81920] [1] : memref<?xvector<4xf32>> to memref<81920xvector<4xf32>, strided<[1], offset: ?>>
gpu.subgroup_mma_store_matrix %matrix, %subview[%i] {leadDimension = 160 : index} : !gpu.mma_matrix<16x16xf16, "COp">, memref<81920xvector<4xf32>, strided<[1], offset: ?>>
return
}
// CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK: func.func @fold_gpu_subgroup_mma_store_matrix_1d
// CHECK-SAME: (%[[DST:.+]]: memref<?xvector<4xf32>>, %[[OFFSET:.+]]: index, %[[I0:.+]]: index, %[[VAL:.+]]: !gpu.mma_matrix<16x16xf16, "COp">)
// CHECK: %[[APPLY:.+]] = affine.apply #[[MAP]]()[%[[OFFSET]], %[[I0]]]
// CHECK: gpu.subgroup_mma_store_matrix %[[VAL]], %[[DST]][%[[APPLY]]] {leadDimension = 160 : index} : !gpu.mma_matrix<16x16xf16, "COp">, memref<?xvector<4xf32>>
// -----
// CHECK-LABEL: func.func @fold_gpu_subgroup_mma_load_matrix_2d
// CHECK-SAME: %[[SRC:.+]]: memref<128x128xf32>
func.func @fold_gpu_subgroup_mma_load_matrix_2d(%arg0 : memref<128x128xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index) -> !gpu.mma_matrix<16x16xf16, "COp"> {
%subview = memref.subview %arg0[%arg1, %arg2][64, 32][2, 1] : memref<128x128xf32> to memref<64x32xf32, strided<[64, 1], offset: ?>>
// CHECK: gpu.subgroup_mma_load_matrix %[[SRC]][{{.+}}] {leadDimension = 32 : index} : memref<128x128xf32> -> !gpu.mma_matrix<16x16xf16, "COp">
%matrix = gpu.subgroup_mma_load_matrix %subview[%arg3, %arg4] {leadDimension = 32 : index} : memref<64x32xf32, strided<[64, 1], offset: ?>> -> !gpu.mma_matrix<16x16xf16, "COp">
return %matrix : !gpu.mma_matrix<16x16xf16, "COp">
}
// -----
// CHECK-LABEL: func.func @fold_gpu_subgroup_mma_load_matrix_2d
// CHECK-SAME: %[[DST:.+]]: memref<128x128xf32>
func.func @fold_gpu_subgroup_mma_load_matrix_2d(%arg0 : memref<128x128xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index, %matrix: !gpu.mma_matrix<16x16xf16, "COp">) {
%subview = memref.subview %arg0[%arg1, %arg2][64, 32][2, 1] : memref<128x128xf32> to memref<64x32xf32, strided<[64, 1], offset: ?>>
// CHECK: gpu.subgroup_mma_store_matrix %{{.+}}, %[[DST]][{{.+}}] {leadDimension = 32 : index} : !gpu.mma_matrix<16x16xf16, "COp">, memref<128x128xf32>
gpu.subgroup_mma_store_matrix %matrix, %subview[%arg3, %arg4] {leadDimension = 32 : index} : !gpu.mma_matrix<16x16xf16, "COp">, memref<64x32xf32, strided<[64, 1], offset: ?>>
return
}
// -----
func.func @fold_nvgpu_device_async_copy_zero_sub_idx(%gmem_memref_3d : memref<2x128x768xf16>, %idx_1 : index, %idx_2 : index, %idx_3 : index) {
%c0 = arith.constant 0 : index
%smem_memref_4d = memref.alloc() : memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
%gmem_memref_subview_2d = memref.subview %gmem_memref_3d[%idx_1, %idx_2, %idx_3] [1, 1, 8] [1, 1, 1] : memref<2x128x768xf16> to memref<1x8xf16, strided<[98304, 1], offset: ?>>
%async_token = nvgpu.device_async_copy %gmem_memref_subview_2d[%c0, %c0], %smem_memref_4d[%c0, %c0, %c0, %c0], 8 {bypassL1} : memref<1x8xf16, strided<[98304, 1], offset: ?>> to memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
return
}
// CHECK-LABEL: func.func @fold_nvgpu_device_async_copy_zero_sub_idx
// CHECK-SAME: (%[[GMEM_MEMREF_3d:.+]]: memref<2x128x768xf16>, %[[IDX_1:.+]]: index, %[[IDX_2:.+]]: index, %[[IDX_3:.+]]: index)
// CHECK-DAG: %[[c0:.+]] = arith.constant 0 : index
// CHECK-DAG: %[[SMEM_MEMREF_4d:.+]] = memref.alloc() : memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
// CHECK: nvgpu.device_async_copy %[[GMEM_MEMREF_3d]][%[[IDX_1]], %[[IDX_2]], %[[IDX_3]]], %[[SMEM_MEMREF_4d]][%[[c0]], %[[c0]], %[[c0]], %[[c0]]], 8 {bypassL1} : memref<2x128x768xf16> to memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
// -----
func.func @fold_src_nvgpu_device_async_copy(%gmem_memref_3d : memref<2x128x768xf16>, %src_idx_0 : index, %src_idx_1 : index, %src_idx_2 : index, %src_sub_idx_0 : index, %src_sub_idx_1 : index) {
%c0 = arith.constant 0 : index
%smem_memref_4d = memref.alloc() : memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
%gmem_memref_subview_2d = memref.subview %gmem_memref_3d[%src_idx_0, %src_idx_1, %src_idx_2] [1, 1, 8] [1, 1, 1] : memref<2x128x768xf16> to memref<1x8xf16, strided<[98304, 1], offset: ?>>
%async_token = nvgpu.device_async_copy %gmem_memref_subview_2d[%src_sub_idx_0, %src_sub_idx_1], %smem_memref_4d[%c0, %c0, %c0, %c0], 8 {bypassL1} : memref<1x8xf16, strided<[98304, 1], offset: ?>> to memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
return
}
// CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK: func.func @fold_src_nvgpu_device_async_copy
// CHECK-SAME: (%[[GMEM_MEMREF_3d:.+]]: memref<2x128x768xf16>, %[[SRC_IDX_0:.+]]: index, %[[SRC_IDX_1:.+]]: index, %[[SRC_IDX_2:.+]]: index, %[[SRC_SUB_IDX_0:.+]]: index, %[[SRC_SUB_IDX_1:.+]]: index)
// CHECK-DAG: %[[c0:.+]] = arith.constant 0 : index
// CHECK-DAG: %[[RESOLVED_SRC_IDX_0:.+]] = affine.apply #[[MAP]]()[%[[SRC_IDX_0]], %[[SRC_SUB_IDX_0]]]
// CHECK-DAG: %[[RESOLVED_SRC_IDX_1:.+]] = affine.apply #[[MAP]]()[%[[SRC_IDX_2]], %[[SRC_SUB_IDX_1]]]
// CHECK-DAG: nvgpu.device_async_copy %[[GMEM_MEMREF_3d]][%[[RESOLVED_SRC_IDX_0]], %[[SRC_IDX_1]], %[[RESOLVED_SRC_IDX_1]]], %[[SMEM_MEMREF_4d]][%[[c0]], %[[c0]], %[[c0]], %[[c0]]], 8 {bypassL1} : memref<2x128x768xf16> to memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
// -----
func.func @fold_src_fold_dest_nvgpu_device_async_copy(%gmem_memref_3d : memref<2x128x768xf16>, %src_idx_0 : index, %src_idx_1 : index, %src_idx_2 : index, %src_sub_idx_0 : index, %src_sub_idx_1 : index, %dest_idx_0 : index, %dest_idx_1 : index, %dest_idx_2 : index, %dest_idx_3 : index, %dest_sub_idx_0 : index, %dest_sub_idx_1 : index) {
%c0 = arith.constant 0 : index
%smem_memref_4d = memref.alloc() : memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
%gmem_memref_subview_2d = memref.subview %gmem_memref_3d[%src_idx_0, %src_idx_1, %src_idx_2] [1, 1, 8] [1, 1, 1] : memref<2x128x768xf16> to memref<1x8xf16, strided<[98304, 1], offset: ?>>
%smem_memref_2d = memref.subview %smem_memref_4d[%dest_idx_0, %dest_idx_1, %dest_idx_2, %dest_idx_3] [1, 1, 1, 8] [1, 1, 1, 1] : memref<5x1x64x64xf16, #gpu.address_space<workgroup>> to memref<1x8xf16, strided<[4096, 1], offset: ?>, #gpu.address_space<workgroup>>
%async_token = nvgpu.device_async_copy %gmem_memref_subview_2d[%src_sub_idx_0, %src_sub_idx_1], %smem_memref_2d[%dest_sub_idx_0, %dest_sub_idx_1], 8 {bypassL1} : memref<1x8xf16, strided<[98304, 1], offset: ?>> to memref<1x8xf16, strided<[4096, 1], offset: ?>, #gpu.address_space<workgroup>>
return
}
// CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0, s1] -> (s0 + s1)>
// CHECK: func.func @fold_src_fold_dest_nvgpu_device_async_copy
// CHECK-SAME: (%[[GMEM_MEMREF_3d:.+]]: memref<2x128x768xf16>, %[[SRC_IDX_0:.+]]: index, %[[SRC_IDX_1:.+]]: index, %[[SRC_IDX_2:.+]]: index, %[[SRC_SUB_IDX_0:.+]]: index, %[[SRC_SUB_IDX_1:.+]]: index, %[[DEST_IDX_0:.+]]: index, %[[DEST_IDX_1:.+]]: index, %[[DEST_IDX_2:.+]]: index, %[[DEST_IDX_3:.+]]: index, %[[DEST_SUB_IDX_0:.+]]: index, %[[DEST_SUB_IDX_1:.+]]: index)
// CHECK-DAG: %[[RESOLVED_SRC_IDX_0:.+]] = affine.apply #[[MAP]]()[%[[SRC_IDX_0]], %[[SRC_SUB_IDX_0]]]
// CHECK-DAG: %[[RESOLVED_SRC_IDX_1:.+]] = affine.apply #[[MAP]]()[%[[SRC_IDX_2]], %[[SRC_SUB_IDX_1]]]
// CHECK-DAG: %[[RESOLVED_DST_IDX_1:.+]] = affine.apply #[[MAP]]()[%[[DEST_IDX_1]], %[[DEST_SUB_IDX_0]]]
// CHECK-DAG: %[[RESOLVED_DST_IDX_3:.+]] = affine.apply #[[MAP]]()[%[[DEST_IDX_3]], %[[DEST_SUB_IDX_1]]]
// CHECK-DAG: nvgpu.device_async_copy %[[GMEM_MEMREF_3d]][%[[RESOLVED_SRC_IDX_0]], %[[SRC_IDX_1]], %[[RESOLVED_SRC_IDX_1]]], %[[SMEM_MEMREF_4d]][%[[DEST_IDX_0]], %[[RESOLVED_DST_IDX_1]], %[[DEST_IDX_2]], %[[RESOLVED_DST_IDX_3]]], 8 {bypassL1} : memref<2x128x768xf16> to memref<5x1x64x64xf16, #gpu.address_space<workgroup>>
// -----
#map = affine_map<()[s0] -> (-s0 + 4)>
#map1 = affine_map<()[s0] -> (-s0 + 32)>
func.func @test_ldmatrix(%arg0: memref<4x32x32xf16, 3>, %arg1: index, %arg2: index, %arg3: index) -> vector<4x2xf16> {
%c0 = arith.constant 0 : index
%0 = affine.apply #map()[%arg1]
%1 = affine.apply #map1()[%arg2]
%2 = affine.apply #map1()[%arg3]
%subview = memref.subview %arg0[%arg1, %arg2, %arg3] [%0, %1, %2] [1, 1, 1] : memref<4x32x32xf16, 3> to memref<?x?x?xf16, strided<[1024, 32, 1], offset: ?>, 3>
%3 = nvgpu.ldmatrix %subview[%c0, %c0, %c0] {numTiles = 4 : i32, transpose = false} : memref<?x?x?xf16, strided<[1024, 32, 1], offset: ?>, 3> -> vector<4x2xf16>
return %3 : vector<4x2xf16>
}
// CHECK: func @test_ldmatrix
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: memref<4x32x32xf16, 3>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9_]+]]: index
// CHECK: nvgpu.ldmatrix %[[ARG0]][%[[ARG1]], %[[ARG2]], %[[ARG3]]] {numTiles = 4 : i32, transpose = false} : memref<4x32x32xf16, 3> -> vector<4x2xf16>
// -----
func.func @fold_vector_load(
%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index) -> vector<12x32xf32> {
%0 = memref.subview %arg0[%arg1, %arg2][1, 1][1, 1] : memref<12x32xf32> to memref<f32, strided<[], offset: ?>>
%1 = vector.load %0[] : memref<f32, strided<[], offset: ?>>, vector<12x32xf32>
return %1 : vector<12x32xf32>
}
// CHECK: func @fold_vector_load
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: memref<12x32xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: index
// CHECK: vector.load %[[ARG0]][%[[ARG1]], %[[ARG2]]] : memref<12x32xf32>, vector<12x32xf32>
|
