This repository has been archived by the owner on Apr 23, 2021. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 257
/
AffineOps.cpp
1764 lines (1533 loc) · 68.4 KB
/
AffineOps.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
//===- AffineOps.cpp - MLIR Affine Operations -----------------------------===//
//
// Copyright 2019 The MLIR Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// =============================================================================
#include "mlir/AffineOps/AffineOps.h"
#include "mlir/IR/Block.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/Function.h"
#include "mlir/IR/IntegerSet.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/OpImplementation.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/StandardOps/Ops.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallBitVector.h"
#include "llvm/Support/Debug.h"
using namespace mlir;
using llvm::dbgs;
#define DEBUG_TYPE "affine-analysis"
//===----------------------------------------------------------------------===//
// AffineOpsDialect
//===----------------------------------------------------------------------===//
AffineOpsDialect::AffineOpsDialect(MLIRContext *context)
: Dialect(getDialectNamespace(), context) {
addOperations<AffineApplyOp, AffineDmaStartOp, AffineDmaWaitOp, AffineLoadOp,
AffineStoreOp,
#define GET_OP_LIST
#include "mlir/AffineOps/AffineOps.cpp.inc"
>();
}
/// A utility function to check if a given region is attached to a function.
static bool isFunctionRegion(Region *region) {
return llvm::isa<FuncOp>(region->getContainingOp());
}
/// A utility function to check if a value is defined at the top level of a
/// function. A value defined at the top level is always a valid symbol.
bool mlir::isTopLevelSymbol(Value *value) {
if (auto *arg = dyn_cast<BlockArgument>(value))
return isFunctionRegion(arg->getOwner()->getParent());
return isFunctionRegion(value->getDefiningOp()->getContainingRegion());
}
// Value can be used as a dimension id if it is valid as a symbol, or
// it is an induction variable, or it is a result of affine apply operation
// with dimension id arguments.
bool mlir::isValidDim(Value *value) {
// The value must be an index type.
if (!value->getType().isIndex())
return false;
if (auto *op = value->getDefiningOp()) {
// Top level operation or constant operation is ok.
if (isFunctionRegion(op->getContainingRegion()) || isa<ConstantOp>(op))
return true;
// Affine apply operation is ok if all of its operands are ok.
if (auto applyOp = dyn_cast<AffineApplyOp>(op))
return applyOp.isValidDim();
// The dim op is okay if its operand memref/tensor is defined at the top
// level.
if (auto dimOp = dyn_cast<DimOp>(op))
return isTopLevelSymbol(dimOp.getOperand());
return false;
}
// This value is a block argument (which also includes 'affine.for' loop IVs).
return true;
}
// Value can be used as a symbol if it is a constant, or it is defined at
// the top level, or it is a result of affine apply operation with symbol
// arguments.
bool mlir::isValidSymbol(Value *value) {
// The value must be an index type.
if (!value->getType().isIndex())
return false;
if (auto *op = value->getDefiningOp()) {
// Top level operation or constant operation is ok.
if (isFunctionRegion(op->getContainingRegion()) || isa<ConstantOp>(op))
return true;
// Affine apply operation is ok if all of its operands are ok.
if (auto applyOp = dyn_cast<AffineApplyOp>(op))
return applyOp.isValidSymbol();
// The dim op is okay if its operand memref/tensor is defined at the top
// level.
if (auto dimOp = dyn_cast<DimOp>(op))
return isTopLevelSymbol(dimOp.getOperand());
return false;
}
// Otherwise, check that the value is a top level symbol.
return isTopLevelSymbol(value);
}
/// Utility function to verify that a set of operands are valid dimension and
/// symbol identifiers. The operands should be layed out such that the dimension
/// operands are before the symbol operands. This function returns failure if
/// there was an invalid operand. An operation is provided to emit any necessary
/// errors.
template <typename OpTy>
static LogicalResult
verifyDimAndSymbolIdentifiers(OpTy &op, Operation::operand_range operands,
unsigned numDims) {
unsigned opIt = 0;
for (auto *operand : operands) {
if (opIt++ < numDims) {
if (!isValidDim(operand))
return op.emitOpError("operand cannot be used as a dimension id");
} else if (!isValidSymbol(operand)) {
return op.emitOpError("operand cannot be used as a symbol");
}
}
return success();
}
//===----------------------------------------------------------------------===//
// AffineApplyOp
//===----------------------------------------------------------------------===//
void AffineApplyOp::build(Builder *builder, OperationState *result,
AffineMap map, ArrayRef<Value *> operands) {
result->addOperands(operands);
result->types.append(map.getNumResults(), builder->getIndexType());
result->addAttribute("map", builder->getAffineMapAttr(map));
}
ParseResult AffineApplyOp::parse(OpAsmParser *parser, OperationState *result) {
auto &builder = parser->getBuilder();
auto affineIntTy = builder.getIndexType();
AffineMapAttr mapAttr;
unsigned numDims;
if (parser->parseAttribute(mapAttr, "map", result->attributes) ||
parseDimAndSymbolList(parser, result->operands, numDims) ||
parser->parseOptionalAttributeDict(result->attributes))
return failure();
auto map = mapAttr.getValue();
if (map.getNumDims() != numDims ||
numDims + map.getNumSymbols() != result->operands.size()) {
return parser->emitError(parser->getNameLoc(),
"dimension or symbol index mismatch");
}
result->types.append(map.getNumResults(), affineIntTy);
return success();
}
void AffineApplyOp::print(OpAsmPrinter *p) {
*p << "affine.apply " << getAttr("map");
printDimAndSymbolList(operand_begin(), operand_end(),
getAffineMap().getNumDims(), p);
p->printOptionalAttrDict(getAttrs(), /*elidedAttrs=*/{"map"});
}
LogicalResult AffineApplyOp::verify() {
// Check that affine map attribute was specified.
auto affineMapAttr = getAttrOfType<AffineMapAttr>("map");
if (!affineMapAttr)
return emitOpError("requires an affine map");
// Check input and output dimensions match.
auto map = affineMapAttr.getValue();
// Verify that operand count matches affine map dimension and symbol count.
if (getNumOperands() != map.getNumDims() + map.getNumSymbols())
return emitOpError(
"operand count and affine map dimension and symbol count must match");
// Verify that all operands are of `index` type.
for (Type t : getOperandTypes()) {
if (!t.isIndex())
return emitOpError("operands must be of type 'index'");
}
if (!getResult()->getType().isIndex())
return emitOpError("result must be of type 'index'");
// Verify that the operands are valid dimension and symbol identifiers.
if (failed(verifyDimAndSymbolIdentifiers(*this, getOperands(),
map.getNumDims())))
return failure();
// Verify that the map only produces one result.
if (map.getNumResults() != 1)
return emitOpError("mapping must produce one value");
return success();
}
// The result of the affine apply operation can be used as a dimension id if it
// is a CFG value or if it is an Value, and all the operands are valid
// dimension ids.
bool AffineApplyOp::isValidDim() {
return llvm::all_of(getOperands(),
[](Value *op) { return mlir::isValidDim(op); });
}
// The result of the affine apply operation can be used as a symbol if it is
// a CFG value or if it is an Value, and all the operands are symbols.
bool AffineApplyOp::isValidSymbol() {
return llvm::all_of(getOperands(),
[](Value *op) { return mlir::isValidSymbol(op); });
}
OpFoldResult AffineApplyOp::fold(ArrayRef<Attribute> operands) {
auto map = getAffineMap();
// Fold dims and symbols to existing values.
auto expr = map.getResult(0);
if (auto dim = expr.dyn_cast<AffineDimExpr>())
return getOperand(dim.getPosition());
if (auto sym = expr.dyn_cast<AffineSymbolExpr>())
return getOperand(map.getNumDims() + sym.getPosition());
// Otherwise, default to folding the map.
SmallVector<Attribute, 1> result;
if (failed(map.constantFold(operands, result)))
return {};
return result[0];
}
namespace {
/// An `AffineApplyNormalizer` is a helper class that is not visible to the user
/// and supports renumbering operands of AffineApplyOp. This acts as a
/// reindexing map of Value* to positional dims or symbols and allows
/// simplifications such as:
///
/// ```mlir
/// %1 = affine.apply (d0, d1) -> (d0 - d1) (%0, %0)
/// ```
///
/// into:
///
/// ```mlir
/// %1 = affine.apply () -> (0)
/// ```
struct AffineApplyNormalizer {
AffineApplyNormalizer(AffineMap map, ArrayRef<Value *> operands);
/// Returns the AffineMap resulting from normalization.
AffineMap getAffineMap() { return affineMap; }
SmallVector<Value *, 8> getOperands() {
SmallVector<Value *, 8> res(reorderedDims);
res.append(concatenatedSymbols.begin(), concatenatedSymbols.end());
return res;
}
private:
/// Helper function to insert `v` into the coordinate system of the current
/// AffineApplyNormalizer. Returns the AffineDimExpr with the corresponding
/// renumbered position.
AffineDimExpr renumberOneDim(Value *v);
/// Given an `other` normalizer, this rewrites `other.affineMap` in the
/// coordinate system of the current AffineApplyNormalizer.
/// Returns the rewritten AffineMap and updates the dims and symbols of
/// `this`.
AffineMap renumber(const AffineApplyNormalizer &other);
/// Maps of Value* to position in `affineMap`.
DenseMap<Value *, unsigned> dimValueToPosition;
/// Ordered dims and symbols matching positional dims and symbols in
/// `affineMap`.
SmallVector<Value *, 8> reorderedDims;
SmallVector<Value *, 8> concatenatedSymbols;
AffineMap affineMap;
/// Used with RAII to control the depth at which AffineApply are composed
/// recursively. Only accepts depth 1 for now to allow a behavior where a
/// newly composed AffineApplyOp does not increase the length of the chain of
/// AffineApplyOps. Full composition is implemented iteratively on top of
/// this behavior.
static unsigned &affineApplyDepth() {
static thread_local unsigned depth = 0;
return depth;
}
static constexpr unsigned kMaxAffineApplyDepth = 1;
AffineApplyNormalizer() { affineApplyDepth()++; }
public:
~AffineApplyNormalizer() { affineApplyDepth()--; }
};
} // end anonymous namespace.
AffineDimExpr AffineApplyNormalizer::renumberOneDim(Value *v) {
DenseMap<Value *, unsigned>::iterator iterPos;
bool inserted = false;
std::tie(iterPos, inserted) =
dimValueToPosition.insert(std::make_pair(v, dimValueToPosition.size()));
if (inserted) {
reorderedDims.push_back(v);
}
return getAffineDimExpr(iterPos->second, v->getContext())
.cast<AffineDimExpr>();
}
AffineMap AffineApplyNormalizer::renumber(const AffineApplyNormalizer &other) {
SmallVector<AffineExpr, 8> dimRemapping;
for (auto *v : other.reorderedDims) {
auto kvp = other.dimValueToPosition.find(v);
if (dimRemapping.size() <= kvp->second)
dimRemapping.resize(kvp->second + 1);
dimRemapping[kvp->second] = renumberOneDim(kvp->first);
}
unsigned numSymbols = concatenatedSymbols.size();
unsigned numOtherSymbols = other.concatenatedSymbols.size();
SmallVector<AffineExpr, 8> symRemapping(numOtherSymbols);
for (unsigned idx = 0; idx < numOtherSymbols; ++idx) {
symRemapping[idx] =
getAffineSymbolExpr(idx + numSymbols, other.affineMap.getContext());
}
concatenatedSymbols.insert(concatenatedSymbols.end(),
other.concatenatedSymbols.begin(),
other.concatenatedSymbols.end());
auto map = other.affineMap;
return map.replaceDimsAndSymbols(dimRemapping, symRemapping,
dimRemapping.size(), symRemapping.size());
}
// Gather the positions of the operands that are produced by an AffineApplyOp.
static llvm::SetVector<unsigned>
indicesFromAffineApplyOp(ArrayRef<Value *> operands) {
llvm::SetVector<unsigned> res;
for (auto en : llvm::enumerate(operands))
if (isa_and_nonnull<AffineApplyOp>(en.value()->getDefiningOp()))
res.insert(en.index());
return res;
}
// Support the special case of a symbol coming from an AffineApplyOp that needs
// to be composed into the current AffineApplyOp.
// This case is handled by rewriting all such symbols into dims for the purpose
// of allowing mathematical AffineMap composition.
// Returns an AffineMap where symbols that come from an AffineApplyOp have been
// rewritten as dims and are ordered after the original dims.
// TODO(andydavis,ntv): This promotion makes AffineMap lose track of which
// symbols are represented as dims. This loss is static but can still be
// recovered dynamically (with `isValidSymbol`). Still this is annoying for the
// semi-affine map case. A dynamic canonicalization of all dims that are valid
// symbols (a.k.a `canonicalizePromotedSymbols`) into symbols helps and even
// results in better simplifications and foldings. But we should evaluate
// whether this behavior is what we really want after using more.
static AffineMap promoteComposedSymbolsAsDims(AffineMap map,
ArrayRef<Value *> symbols) {
if (symbols.empty()) {
return map;
}
// Sanity check on symbols.
for (auto *sym : symbols) {
assert(isValidSymbol(sym) && "Expected only valid symbols");
(void)sym;
}
// Extract the symbol positions that come from an AffineApplyOp and
// needs to be rewritten as dims.
auto symPositions = indicesFromAffineApplyOp(symbols);
if (symPositions.empty()) {
return map;
}
// Create the new map by replacing each symbol at pos by the next new dim.
unsigned numDims = map.getNumDims();
unsigned numSymbols = map.getNumSymbols();
unsigned numNewDims = 0;
unsigned numNewSymbols = 0;
SmallVector<AffineExpr, 8> symReplacements(numSymbols);
for (unsigned i = 0; i < numSymbols; ++i) {
symReplacements[i] =
symPositions.count(i) > 0
? getAffineDimExpr(numDims + numNewDims++, map.getContext())
: getAffineSymbolExpr(numNewSymbols++, map.getContext());
}
assert(numSymbols >= numNewDims);
AffineMap newMap = map.replaceDimsAndSymbols(
{}, symReplacements, numDims + numNewDims, numNewSymbols);
return newMap;
}
/// The AffineNormalizer composes AffineApplyOp recursively. Its purpose is to
/// keep a correspondence between the mathematical `map` and the `operands` of
/// a given AffineApplyOp. This correspondence is maintained by iterating over
/// the operands and forming an `auxiliaryMap` that can be composed
/// mathematically with `map`. To keep this correspondence in cases where
/// symbols are produced by affine.apply operations, we perform a local rewrite
/// of symbols as dims.
///
/// Rationale for locally rewriting symbols as dims:
/// ================================================
/// The mathematical composition of AffineMap must always concatenate symbols
/// because it does not have enough information to do otherwise. For example,
/// composing `(d0)[s0] -> (d0 + s0)` with itself must produce
/// `(d0)[s0, s1] -> (d0 + s0 + s1)`.
///
/// The result is only equivalent to `(d0)[s0] -> (d0 + 2 * s0)` when
/// applied to the same mlir::Value* for both s0 and s1.
/// As a consequence mathematical composition of AffineMap always concatenates
/// symbols.
///
/// When AffineMaps are used in AffineApplyOp however, they may specify
/// composition via symbols, which is ambiguous mathematically. This corner case
/// is handled by locally rewriting such symbols that come from AffineApplyOp
/// into dims and composing through dims.
/// TODO(andydavis, ntv): Composition via symbols comes at a significant code
/// complexity. Alternatively we should investigate whether we want to
/// explicitly disallow symbols coming from affine.apply and instead force the
/// user to compose symbols beforehand. The annoyances may be small (i.e. 1 or 2
/// extra API calls for such uses, which haven't popped up until now) and the
/// benefit potentially big: simpler and more maintainable code for a
/// non-trivial, recursive, procedure.
AffineApplyNormalizer::AffineApplyNormalizer(AffineMap map,
ArrayRef<Value *> operands)
: AffineApplyNormalizer() {
static_assert(kMaxAffineApplyDepth > 0, "kMaxAffineApplyDepth must be > 0");
assert(map.getNumInputs() == operands.size() &&
"number of operands does not match the number of map inputs");
LLVM_DEBUG(map.print(dbgs() << "\nInput map: "));
// Promote symbols that come from an AffineApplyOp to dims by rewriting the
// map to always refer to:
// (dims, symbols coming from AffineApplyOp, other symbols).
// The order of operands can remain unchanged.
// This is a simplification that relies on 2 ordering properties:
// 1. rewritten symbols always appear after the original dims in the map;
// 2. operands are traversed in order and either dispatched to:
// a. auxiliaryExprs (dims and symbols rewritten as dims);
// b. concatenatedSymbols (all other symbols)
// This allows operand order to remain unchanged.
unsigned numDimsBeforeRewrite = map.getNumDims();
map = promoteComposedSymbolsAsDims(map,
operands.take_back(map.getNumSymbols()));
LLVM_DEBUG(map.print(dbgs() << "\nRewritten map: "));
SmallVector<AffineExpr, 8> auxiliaryExprs;
bool furtherCompose = (affineApplyDepth() <= kMaxAffineApplyDepth);
// We fully spell out the 2 cases below. In this particular instance a little
// code duplication greatly improves readability.
// Note that the first branch would disappear if we only supported full
// composition (i.e. infinite kMaxAffineApplyDepth).
if (!furtherCompose) {
// 1. Only dispatch dims or symbols.
for (auto en : llvm::enumerate(operands)) {
auto *t = en.value();
assert(t->getType().isIndex());
bool isDim = (en.index() < map.getNumDims());
if (isDim) {
// a. The mathematical composition of AffineMap composes dims.
auxiliaryExprs.push_back(renumberOneDim(t));
} else {
// b. The mathematical composition of AffineMap concatenates symbols.
// We do the same for symbol operands.
concatenatedSymbols.push_back(t);
}
}
} else {
assert(numDimsBeforeRewrite <= operands.size());
// 2. Compose AffineApplyOps and dispatch dims or symbols.
for (unsigned i = 0, e = operands.size(); i < e; ++i) {
auto *t = operands[i];
auto affineApply = dyn_cast_or_null<AffineApplyOp>(t->getDefiningOp());
if (affineApply) {
// a. Compose affine.apply operations.
LLVM_DEBUG(affineApply.getOperation()->print(
dbgs() << "\nCompose AffineApplyOp recursively: "));
AffineMap affineApplyMap = affineApply.getAffineMap();
SmallVector<Value *, 8> affineApplyOperands(
affineApply.getOperands().begin(), affineApply.getOperands().end());
AffineApplyNormalizer normalizer(affineApplyMap, affineApplyOperands);
LLVM_DEBUG(normalizer.affineMap.print(
dbgs() << "\nRenumber into current normalizer: "));
auto renumberedMap = renumber(normalizer);
LLVM_DEBUG(
renumberedMap.print(dbgs() << "\nRecursive composition yields: "));
auxiliaryExprs.push_back(renumberedMap.getResult(0));
} else {
if (i < numDimsBeforeRewrite) {
// b. The mathematical composition of AffineMap composes dims.
auxiliaryExprs.push_back(renumberOneDim(t));
} else {
// c. The mathematical composition of AffineMap concatenates symbols.
// We do the same for symbol operands.
concatenatedSymbols.push_back(t);
}
}
}
}
// Early exit if `map` is already composed.
if (auxiliaryExprs.empty()) {
affineMap = map;
return;
}
assert(concatenatedSymbols.size() >= map.getNumSymbols() &&
"Unexpected number of concatenated symbols");
auto numDims = dimValueToPosition.size();
auto numSymbols = concatenatedSymbols.size() - map.getNumSymbols();
auto auxiliaryMap = AffineMap::get(numDims, numSymbols, auxiliaryExprs);
LLVM_DEBUG(map.print(dbgs() << "\nCompose map: "));
LLVM_DEBUG(auxiliaryMap.print(dbgs() << "\nWith map: "));
LLVM_DEBUG(map.compose(auxiliaryMap).print(dbgs() << "\nResult: "));
// TODO(andydavis,ntv): Disabling simplification results in major speed gains.
// Another option is to cache the results as it is expected a lot of redundant
// work is performed in practice.
affineMap = simplifyAffineMap(map.compose(auxiliaryMap));
LLVM_DEBUG(affineMap.print(dbgs() << "\nSimplified result: "));
LLVM_DEBUG(dbgs() << "\n");
}
/// Implements `map` and `operands` composition and simplification to support
/// `makeComposedAffineApply`. This can be called to achieve the same effects
/// on `map` and `operands` without creating an AffineApplyOp that needs to be
/// immediately deleted.
static void composeAffineMapAndOperands(AffineMap *map,
SmallVectorImpl<Value *> *operands) {
AffineApplyNormalizer normalizer(*map, *operands);
auto normalizedMap = normalizer.getAffineMap();
auto normalizedOperands = normalizer.getOperands();
canonicalizeMapAndOperands(&normalizedMap, &normalizedOperands);
*map = normalizedMap;
*operands = normalizedOperands;
assert(*map);
}
void mlir::fullyComposeAffineMapAndOperands(
AffineMap *map, SmallVectorImpl<Value *> *operands) {
while (llvm::any_of(*operands, [](Value *v) {
return isa_and_nonnull<AffineApplyOp>(v->getDefiningOp());
})) {
composeAffineMapAndOperands(map, operands);
}
}
AffineApplyOp mlir::makeComposedAffineApply(OpBuilder &b, Location loc,
AffineMap map,
ArrayRef<Value *> operands) {
AffineMap normalizedMap = map;
SmallVector<Value *, 8> normalizedOperands(operands.begin(), operands.end());
composeAffineMapAndOperands(&normalizedMap, &normalizedOperands);
assert(normalizedMap);
return b.create<AffineApplyOp>(loc, normalizedMap, normalizedOperands);
}
// A symbol may appear as a dim in affine.apply operations. This function
// canonicalizes dims that are valid symbols into actual symbols.
static void
canonicalizePromotedSymbols(AffineMap *map,
llvm::SmallVectorImpl<Value *> *operands) {
if (!map || operands->empty())
return;
assert(map->getNumInputs() == operands->size() &&
"map inputs must match number of operands");
auto *context = map->getContext();
SmallVector<Value *, 8> resultOperands;
resultOperands.reserve(operands->size());
SmallVector<Value *, 8> remappedSymbols;
remappedSymbols.reserve(operands->size());
unsigned nextDim = 0;
unsigned nextSym = 0;
unsigned oldNumSyms = map->getNumSymbols();
SmallVector<AffineExpr, 8> dimRemapping(map->getNumDims());
for (unsigned i = 0, e = map->getNumInputs(); i != e; ++i) {
if (i < map->getNumDims()) {
if (isValidSymbol((*operands)[i])) {
// This is a valid symbols that appears as a dim, canonicalize it.
dimRemapping[i] = getAffineSymbolExpr(oldNumSyms + nextSym++, context);
remappedSymbols.push_back((*operands)[i]);
} else {
dimRemapping[i] = getAffineDimExpr(nextDim++, context);
resultOperands.push_back((*operands)[i]);
}
} else {
resultOperands.push_back((*operands)[i]);
}
}
resultOperands.append(remappedSymbols.begin(), remappedSymbols.end());
*operands = resultOperands;
*map = map->replaceDimsAndSymbols(dimRemapping, {}, nextDim,
oldNumSyms + nextSym);
assert(map->getNumInputs() == operands->size() &&
"map inputs must match number of operands");
}
void mlir::canonicalizeMapAndOperands(
AffineMap *map, llvm::SmallVectorImpl<Value *> *operands) {
if (!map || operands->empty())
return;
assert(map->getNumInputs() == operands->size() &&
"map inputs must match number of operands");
canonicalizePromotedSymbols(map, operands);
// Check to see what dims are used.
llvm::SmallBitVector usedDims(map->getNumDims());
llvm::SmallBitVector usedSyms(map->getNumSymbols());
map->walkExprs([&](AffineExpr expr) {
if (auto dimExpr = expr.dyn_cast<AffineDimExpr>())
usedDims[dimExpr.getPosition()] = true;
else if (auto symExpr = expr.dyn_cast<AffineSymbolExpr>())
usedSyms[symExpr.getPosition()] = true;
});
auto *context = map->getContext();
SmallVector<Value *, 8> resultOperands;
resultOperands.reserve(operands->size());
llvm::SmallDenseMap<Value *, AffineExpr, 8> seenDims;
SmallVector<AffineExpr, 8> dimRemapping(map->getNumDims());
unsigned nextDim = 0;
for (unsigned i = 0, e = map->getNumDims(); i != e; ++i) {
if (usedDims[i]) {
auto it = seenDims.find((*operands)[i]);
if (it == seenDims.end()) {
dimRemapping[i] = getAffineDimExpr(nextDim++, context);
resultOperands.push_back((*operands)[i]);
seenDims.insert(std::make_pair((*operands)[i], dimRemapping[i]));
} else {
dimRemapping[i] = it->second;
}
}
}
llvm::SmallDenseMap<Value *, AffineExpr, 8> seenSymbols;
SmallVector<AffineExpr, 8> symRemapping(map->getNumSymbols());
unsigned nextSym = 0;
for (unsigned i = 0, e = map->getNumSymbols(); i != e; ++i) {
if (usedSyms[i]) {
auto it = seenSymbols.find((*operands)[i + map->getNumDims()]);
if (it == seenSymbols.end()) {
symRemapping[i] = getAffineSymbolExpr(nextSym++, context);
resultOperands.push_back((*operands)[i + map->getNumDims()]);
seenSymbols.insert(std::make_pair((*operands)[i + map->getNumDims()],
symRemapping[i]));
} else {
symRemapping[i] = it->second;
}
}
}
*map =
map->replaceDimsAndSymbols(dimRemapping, symRemapping, nextDim, nextSym);
*operands = resultOperands;
}
namespace {
/// Simplify AffineApply operations.
///
struct SimplifyAffineApply : public OpRewritePattern<AffineApplyOp> {
using OpRewritePattern<AffineApplyOp>::OpRewritePattern;
PatternMatchResult matchAndRewrite(AffineApplyOp apply,
PatternRewriter &rewriter) const override {
auto map = apply.getAffineMap();
AffineMap oldMap = map;
SmallVector<Value *, 8> resultOperands(apply.getOperands());
composeAffineMapAndOperands(&map, &resultOperands);
if (map == oldMap)
return matchFailure();
rewriter.replaceOpWithNewOp<AffineApplyOp>(apply, map, resultOperands);
return matchSuccess();
}
};
} // end anonymous namespace.
void AffineApplyOp::getCanonicalizationPatterns(
OwningRewritePatternList &results, MLIRContext *context) {
results.push_back(llvm::make_unique<SimplifyAffineApply>(context));
}
//===----------------------------------------------------------------------===//
// Common canonicalization pattern support logic
//===----------------------------------------------------------------------===//
namespace {
/// This is a common class used for patterns of the form
/// "someop(memrefcast) -> someop". It folds the source of any memref_cast
/// into the root operation directly.
struct MemRefCastFolder : public RewritePattern {
/// The rootOpName is the name of the root operation to match against.
MemRefCastFolder(StringRef rootOpName, MLIRContext *context)
: RewritePattern(rootOpName, 1, context) {}
PatternMatchResult match(Operation *op) const override {
for (auto *operand : op->getOperands())
if (matchPattern(operand, m_Op<MemRefCastOp>()))
return matchSuccess();
return matchFailure();
}
void rewrite(Operation *op, PatternRewriter &rewriter) const override {
for (unsigned i = 0, e = op->getNumOperands(); i != e; ++i)
if (auto *memref = op->getOperand(i)->getDefiningOp())
if (auto cast = dyn_cast<MemRefCastOp>(memref))
op->setOperand(i, cast.getOperand());
rewriter.updatedRootInPlace(op);
}
};
} // end anonymous namespace.
//===----------------------------------------------------------------------===//
// AffineDmaStartOp
//===----------------------------------------------------------------------===//
// TODO(b/133776335) Check that map operands are loop IVs or symbols.
void AffineDmaStartOp::build(Builder *builder, OperationState *result,
Value *srcMemRef, AffineMap srcMap,
ArrayRef<Value *> srcIndices, Value *destMemRef,
AffineMap dstMap, ArrayRef<Value *> destIndices,
Value *tagMemRef, AffineMap tagMap,
ArrayRef<Value *> tagIndices, Value *numElements,
Value *stride, Value *elementsPerStride) {
result->addOperands(srcMemRef);
result->addAttribute(getSrcMapAttrName(), builder->getAffineMapAttr(srcMap));
result->addOperands(srcIndices);
result->addOperands(destMemRef);
result->addAttribute(getDstMapAttrName(), builder->getAffineMapAttr(dstMap));
result->addOperands(destIndices);
result->addOperands(tagMemRef);
result->addAttribute(getTagMapAttrName(), builder->getAffineMapAttr(tagMap));
result->addOperands(tagIndices);
result->addOperands(numElements);
if (stride) {
result->addOperands({stride, elementsPerStride});
}
}
void AffineDmaStartOp::print(OpAsmPrinter *p) {
*p << "affine.dma_start " << *getSrcMemRef() << '[';
SmallVector<Value *, 8> operands(getSrcIndices());
p->printAffineMapOfSSAIds(getSrcMapAttr(), operands);
*p << "], " << *getDstMemRef() << '[';
operands.assign(getDstIndices().begin(), getDstIndices().end());
p->printAffineMapOfSSAIds(getDstMapAttr(), operands);
*p << "], " << *getTagMemRef() << '[';
operands.assign(getTagIndices().begin(), getTagIndices().end());
p->printAffineMapOfSSAIds(getTagMapAttr(), operands);
*p << "], " << *getNumElements();
if (isStrided()) {
*p << ", " << *getStride();
*p << ", " << *getNumElementsPerStride();
}
*p << " : " << getSrcMemRefType() << ", " << getDstMemRefType() << ", "
<< getTagMemRefType();
}
// Parse AffineDmaStartOp.
// Ex:
// affine.dma_start %src[%i, %j], %dst[%k, %l], %tag[%index], %size,
// %stride, %num_elt_per_stride
// : memref<3076 x f32, 0>, memref<1024 x f32, 2>, memref<1 x i32>
//
ParseResult AffineDmaStartOp::parse(OpAsmParser *parser,
OperationState *result) {
OpAsmParser::OperandType srcMemRefInfo;
AffineMapAttr srcMapAttr;
SmallVector<OpAsmParser::OperandType, 4> srcMapOperands;
OpAsmParser::OperandType dstMemRefInfo;
AffineMapAttr dstMapAttr;
SmallVector<OpAsmParser::OperandType, 4> dstMapOperands;
OpAsmParser::OperandType tagMemRefInfo;
AffineMapAttr tagMapAttr;
SmallVector<OpAsmParser::OperandType, 4> tagMapOperands;
OpAsmParser::OperandType numElementsInfo;
SmallVector<OpAsmParser::OperandType, 2> strideInfo;
SmallVector<Type, 3> types;
auto indexType = parser->getBuilder().getIndexType();
// Parse and resolve the following list of operands:
// *) dst memref followed by its affine maps operands (in square brackets).
// *) src memref followed by its affine map operands (in square brackets).
// *) tag memref followed by its affine map operands (in square brackets).
// *) number of elements transferred by DMA operation.
if (parser->parseOperand(srcMemRefInfo) ||
parser->parseAffineMapOfSSAIds(srcMapOperands, srcMapAttr,
getSrcMapAttrName(), result->attributes) ||
parser->parseComma() || parser->parseOperand(dstMemRefInfo) ||
parser->parseAffineMapOfSSAIds(dstMapOperands, dstMapAttr,
getDstMapAttrName(), result->attributes) ||
parser->parseComma() || parser->parseOperand(tagMemRefInfo) ||
parser->parseAffineMapOfSSAIds(tagMapOperands, tagMapAttr,
getTagMapAttrName(), result->attributes) ||
parser->parseComma() || parser->parseOperand(numElementsInfo))
return failure();
// Parse optional stride and elements per stride.
if (parser->parseTrailingOperandList(strideInfo)) {
return failure();
}
if (!strideInfo.empty() && strideInfo.size() != 2) {
return parser->emitError(parser->getNameLoc(),
"expected two stride related operands");
}
bool isStrided = strideInfo.size() == 2;
if (parser->parseColonTypeList(types))
return failure();
if (types.size() != 3)
return parser->emitError(parser->getNameLoc(), "expected three types");
if (parser->resolveOperand(srcMemRefInfo, types[0], result->operands) ||
parser->resolveOperands(srcMapOperands, indexType, result->operands) ||
parser->resolveOperand(dstMemRefInfo, types[1], result->operands) ||
parser->resolveOperands(dstMapOperands, indexType, result->operands) ||
parser->resolveOperand(tagMemRefInfo, types[2], result->operands) ||
parser->resolveOperands(tagMapOperands, indexType, result->operands) ||
parser->resolveOperand(numElementsInfo, indexType, result->operands))
return failure();
if (isStrided) {
if (parser->resolveOperands(strideInfo, indexType, result->operands))
return failure();
}
// Check that src/dst/tag operand counts match their map.numInputs.
if (srcMapOperands.size() != srcMapAttr.getValue().getNumInputs() ||
dstMapOperands.size() != dstMapAttr.getValue().getNumInputs() ||
tagMapOperands.size() != tagMapAttr.getValue().getNumInputs())
return parser->emitError(parser->getNameLoc(),
"memref operand count not equal to map.numInputs");
return success();
}
LogicalResult AffineDmaStartOp::verify() {
if (!getOperand(getSrcMemRefOperandIndex())->getType().isa<MemRefType>())
return emitOpError("expected DMA source to be of memref type");
if (!getOperand(getDstMemRefOperandIndex())->getType().isa<MemRefType>())
return emitOpError("expected DMA destination to be of memref type");
if (!getOperand(getTagMemRefOperandIndex())->getType().isa<MemRefType>())
return emitOpError("expected DMA tag to be of memref type");
// DMAs from different memory spaces supported.
if (getSrcMemorySpace() == getDstMemorySpace()) {
return emitOpError("DMA should be between different memory spaces");
}
unsigned numInputsAllMaps = getSrcMap().getNumInputs() +
getDstMap().getNumInputs() +
getTagMap().getNumInputs();
if (getNumOperands() != numInputsAllMaps + 3 + 1 &&
getNumOperands() != numInputsAllMaps + 3 + 1 + 2) {
return emitOpError("incorrect number of operands");
}
return success();
}
void AffineDmaStartOp::getCanonicalizationPatterns(
OwningRewritePatternList &results, MLIRContext *context) {
/// dma_start(memrefcast) -> dma_start
results.push_back(
llvm::make_unique<MemRefCastFolder>(getOperationName(), context));
}
//===----------------------------------------------------------------------===//
// AffineDmaWaitOp
//===----------------------------------------------------------------------===//
// TODO(b/133776335) Check that map operands are loop IVs or symbols.
void AffineDmaWaitOp::build(Builder *builder, OperationState *result,
Value *tagMemRef, AffineMap tagMap,
ArrayRef<Value *> tagIndices, Value *numElements) {
result->addOperands(tagMemRef);
result->addAttribute(getTagMapAttrName(), builder->getAffineMapAttr(tagMap));
result->addOperands(tagIndices);
result->addOperands(numElements);
}
void AffineDmaWaitOp::print(OpAsmPrinter *p) {
*p << "affine.dma_wait " << *getTagMemRef() << '[';
SmallVector<Value *, 2> operands(getTagIndices());
p->printAffineMapOfSSAIds(getTagMapAttr(), operands);
*p << "], ";
p->printOperand(getNumElements());
*p << " : " << getTagMemRef()->getType();
}
// Parse AffineDmaWaitOp.
// Eg:
// affine.dma_wait %tag[%index], %num_elements
// : memref<1 x i32, (d0) -> (d0), 4>
//
ParseResult AffineDmaWaitOp::parse(OpAsmParser *parser,
OperationState *result) {
OpAsmParser::OperandType tagMemRefInfo;
AffineMapAttr tagMapAttr;
SmallVector<OpAsmParser::OperandType, 2> tagMapOperands;
Type type;
auto indexType = parser->getBuilder().getIndexType();
OpAsmParser::OperandType numElementsInfo;
// Parse tag memref, its map operands, and dma size.
if (parser->parseOperand(tagMemRefInfo) ||
parser->parseAffineMapOfSSAIds(tagMapOperands, tagMapAttr,
getTagMapAttrName(), result->attributes) ||
parser->parseComma() || parser->parseOperand(numElementsInfo) ||
parser->parseColonType(type) ||
parser->resolveOperand(tagMemRefInfo, type, result->operands) ||
parser->resolveOperands(tagMapOperands, indexType, result->operands) ||
parser->resolveOperand(numElementsInfo, indexType, result->operands))
return failure();
if (!type.isa<MemRefType>())
return parser->emitError(parser->getNameLoc(),
"expected tag to be of memref type");
if (tagMapOperands.size() != tagMapAttr.getValue().getNumInputs())
return parser->emitError(parser->getNameLoc(),
"tag memref operand count != to map.numInputs");
return success();
}
LogicalResult AffineDmaWaitOp::verify() {
if (!getOperand(0)->getType().isa<MemRefType>())
return emitOpError("expected DMA tag to be of memref type");
return success();
}
void AffineDmaWaitOp::getCanonicalizationPatterns(
OwningRewritePatternList &results, MLIRContext *context) {
/// dma_wait(memrefcast) -> dma_wait
results.push_back(
llvm::make_unique<MemRefCastFolder>(getOperationName(), context));
}
//===----------------------------------------------------------------------===//
// AffineForOp
//===----------------------------------------------------------------------===//
// Check that if a "block" has a terminator, it is an `AffineTerminatorOp`.
static LogicalResult checkHasAffineTerminator(OpState &op, Block &block) {
if (block.empty() || isa<AffineTerminatorOp>(block.back()))
return success();
return op.emitOpError("expects regions to end with '" +
AffineTerminatorOp::getOperationName() + "'")
.attachNote()
<< "in custom textual format, the absence of terminator implies '"
<< AffineTerminatorOp::getOperationName() << "'";
}
// Insert `affine.terminator` at the end of the region's only block if it does
// not have a terminator already. If the region is empty, insert a new block
// first.
static void ensureAffineTerminator(Region ®ion, Builder &builder,
Location loc) {
impl::ensureRegionTerminator<AffineTerminatorOp>(region, builder, loc);
}
void AffineForOp::build(Builder *builder, OperationState *result,
ArrayRef<Value *> lbOperands, AffineMap lbMap,
ArrayRef<Value *> ubOperands, AffineMap ubMap,
int64_t step) {
assert(((!lbMap && lbOperands.empty()) ||
lbOperands.size() == lbMap.getNumInputs()) &&
"lower bound operand count does not match the affine map");
assert(((!ubMap && ubOperands.empty()) ||
ubOperands.size() == ubMap.getNumInputs()) &&
"upper bound operand count does not match the affine map");
assert(step > 0 && "step has to be a positive integer constant");
// Add an attribute for the step.