1 //===- VectorOps.cpp - MLIR Vector Dialect Operations ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements convenience types for working with super-vectorization
10 // operations, in particular super-vector loads and stores.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "mlir/Dialect/Vector/IR/VectorOps.h"
15
16 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
17 #include "mlir/Dialect/Arithmetic/Utils/Utils.h"
18 #include "mlir/Dialect/MemRef/IR/MemRef.h"
19 #include "mlir/Dialect/Tensor/IR/Tensor.h"
20 #include "mlir/Dialect/Utils/IndexingUtils.h"
21 #include "mlir/Dialect/Utils/StructuredOpsUtils.h"
22 #include "mlir/IR/AffineExpr.h"
23 #include "mlir/IR/AffineMap.h"
24 #include "mlir/IR/BlockAndValueMapping.h"
25 #include "mlir/IR/Builders.h"
26 #include "mlir/IR/BuiltinOps.h"
27 #include "mlir/IR/BuiltinTypes.h"
28 #include "mlir/IR/DialectImplementation.h"
29 #include "mlir/IR/OpImplementation.h"
30 #include "mlir/IR/PatternMatch.h"
31 #include "mlir/IR/TypeUtilities.h"
32 #include "mlir/Support/LLVM.h"
33 #include "mlir/Support/MathExtras.h"
34 #include "llvm/ADT/StringSet.h"
35 #include "llvm/ADT/bit.h"
36 #include <numeric>
37
38 #include "mlir/Dialect/Vector/IR/VectorOpsDialect.cpp.inc"
39 // Pull in all enum type and utility function definitions.
40 #include "mlir/Dialect/Vector/IR/VectorOpsEnums.cpp.inc"
41
42 using namespace mlir;
43 using namespace mlir::vector;
44
45 /// Helper enum to classify mask value.
46 enum class MaskFormat {
47 AllTrue = 0,
48 AllFalse = 1,
49 Unknown = 2,
50 };
51
52 /// Helper method to classify a 1-D mask value. Currently, the method
53 /// looks "under the hood" of a constant value with dense attributes
54 /// and a constant mask operation (since the client may be called at
55 /// various stages during progressive lowering).
get1DMaskFormat(Value mask)56 static MaskFormat get1DMaskFormat(Value mask) {
57 if (auto c = mask.getDefiningOp<arith::ConstantOp>()) {
58 // Inspect constant dense values. We count up for bits that
59 // are set, count down for bits that are cleared, and bail
60 // when a mix is detected.
61 if (auto denseElts = c.getValue().dyn_cast<DenseIntElementsAttr>()) {
62 int64_t val = 0;
63 for (bool b : denseElts.getValues<bool>())
64 if (b && val >= 0)
65 val++;
66 else if (!b && val <= 0)
67 val--;
68 else
69 return MaskFormat::Unknown;
70 if (val > 0)
71 return MaskFormat::AllTrue;
72 if (val < 0)
73 return MaskFormat::AllFalse;
74 }
75 } else if (auto m = mask.getDefiningOp<ConstantMaskOp>()) {
76 // Inspect constant mask index. If the index exceeds the
77 // dimension size, all bits are set. If the index is zero
78 // or less, no bits are set.
79 ArrayAttr masks = m.getMaskDimSizes();
80 assert(masks.size() == 1);
81 int64_t i = masks[0].cast<IntegerAttr>().getInt();
82 int64_t u = m.getType().getDimSize(0);
83 if (i >= u)
84 return MaskFormat::AllTrue;
85 if (i <= 0)
86 return MaskFormat::AllFalse;
87 }
88 return MaskFormat::Unknown;
89 }
90
91 // Helper for verifying combining kinds in contractions and reductions.
isSupportedCombiningKind(CombiningKind combiningKind,Type elementType)92 static bool isSupportedCombiningKind(CombiningKind combiningKind,
93 Type elementType) {
94 switch (combiningKind) {
95 case CombiningKind::ADD:
96 case CombiningKind::MUL:
97 return elementType.isIntOrIndexOrFloat();
98 case CombiningKind::MINUI:
99 case CombiningKind::MINSI:
100 case CombiningKind::MAXUI:
101 case CombiningKind::MAXSI:
102 case CombiningKind::AND:
103 case CombiningKind::OR:
104 case CombiningKind::XOR:
105 return elementType.isIntOrIndex();
106 case CombiningKind::MINF:
107 case CombiningKind::MAXF:
108 return elementType.isa<FloatType>();
109 }
110 return false;
111 }
112
113 /// Return true if the last dimension of the MemRefType has unit stride. Also
114 /// return true for memrefs with no strides.
isLastMemrefDimUnitStride(MemRefType type)115 bool mlir::vector::isLastMemrefDimUnitStride(MemRefType type) {
116 int64_t offset;
117 SmallVector<int64_t> strides;
118 auto successStrides = getStridesAndOffset(type, strides, offset);
119 return succeeded(successStrides) && (strides.empty() || strides.back() == 1);
120 }
121
getTransferMinorIdentityMap(ShapedType shapedType,VectorType vectorType)122 AffineMap mlir::vector::getTransferMinorIdentityMap(ShapedType shapedType,
123 VectorType vectorType) {
124 int64_t elementVectorRank = 0;
125 VectorType elementVectorType =
126 shapedType.getElementType().dyn_cast<VectorType>();
127 if (elementVectorType)
128 elementVectorRank += elementVectorType.getRank();
129 // 0-d transfers are to/from tensor<t>/memref<t> and vector<1xt>.
130 // TODO: replace once we have 0-d vectors.
131 if (shapedType.getRank() == 0 &&
132 vectorType.getShape() == ArrayRef<int64_t>{1})
133 return AffineMap::get(
134 /*numDims=*/0, /*numSymbols=*/0,
135 getAffineConstantExpr(0, shapedType.getContext()));
136 return AffineMap::getMinorIdentityMap(
137 shapedType.getRank(), vectorType.getRank() - elementVectorRank,
138 shapedType.getContext());
139 }
140
checkSameValueRAW(vector::TransferWriteOp defWrite,vector::TransferReadOp read)141 bool mlir::vector::checkSameValueRAW(vector::TransferWriteOp defWrite,
142 vector::TransferReadOp read) {
143 return !defWrite.hasOutOfBoundsDim() && !defWrite.getMask() &&
144 !read.getMask() && defWrite.getIndices() == read.getIndices() &&
145 defWrite.getVectorType() == read.getVectorType() &&
146 defWrite.getPermutationMap() == read.getPermutationMap();
147 }
148
checkSameValueWAW(vector::TransferWriteOp write,vector::TransferWriteOp priorWrite)149 bool mlir::vector::checkSameValueWAW(vector::TransferWriteOp write,
150 vector::TransferWriteOp priorWrite) {
151 return priorWrite.getIndices() == write.getIndices() &&
152 priorWrite.getMask() == write.getMask() &&
153 priorWrite.getVectorType() == write.getVectorType() &&
154 priorWrite.getPermutationMap() == write.getPermutationMap();
155 }
156
isDisjointTransferIndices(VectorTransferOpInterface transferA,VectorTransferOpInterface transferB)157 bool mlir::vector::isDisjointTransferIndices(
158 VectorTransferOpInterface transferA, VectorTransferOpInterface transferB) {
159 // For simplicity only look at transfer of same type.
160 if (transferA.getVectorType() != transferB.getVectorType())
161 return false;
162 unsigned rankOffset = transferA.getLeadingShapedRank();
163 for (unsigned i = 0, e = transferA.indices().size(); i < e; i++) {
164 auto indexA = transferA.indices()[i].getDefiningOp<arith::ConstantOp>();
165 auto indexB = transferB.indices()[i].getDefiningOp<arith::ConstantOp>();
166 // If any of the indices are dynamic we cannot prove anything.
167 if (!indexA || !indexB)
168 continue;
169
170 if (i < rankOffset) {
171 // For leading dimensions, if we can prove that index are different we
172 // know we are accessing disjoint slices.
173 if (indexA.getValue().cast<IntegerAttr>().getInt() !=
174 indexB.getValue().cast<IntegerAttr>().getInt())
175 return true;
176 } else {
177 // For this dimension, we slice a part of the memref we need to make sure
178 // the intervals accessed don't overlap.
179 int64_t distance =
180 std::abs(indexA.getValue().cast<IntegerAttr>().getInt() -
181 indexB.getValue().cast<IntegerAttr>().getInt());
182 if (distance >= transferA.getVectorType().getDimSize(i - rankOffset))
183 return true;
184 }
185 }
186 return false;
187 }
188
isDisjointTransferSet(VectorTransferOpInterface transferA,VectorTransferOpInterface transferB)189 bool mlir::vector::isDisjointTransferSet(VectorTransferOpInterface transferA,
190 VectorTransferOpInterface transferB) {
191 if (transferA.source() != transferB.source())
192 return false;
193 return isDisjointTransferIndices(transferA, transferB);
194 }
195
196 //===----------------------------------------------------------------------===//
197 // CombiningKindAttr
198 //===----------------------------------------------------------------------===//
199
200 namespace mlir {
201 namespace vector {
202 namespace detail {
203 struct BitmaskEnumStorage : public AttributeStorage {
204 using KeyTy = uint64_t;
205
BitmaskEnumStoragemlir::vector::detail::BitmaskEnumStorage206 BitmaskEnumStorage(KeyTy val) : value(val) {}
207
operator ==mlir::vector::detail::BitmaskEnumStorage208 bool operator==(const KeyTy &key) const { return value == key; }
209
constructmlir::vector::detail::BitmaskEnumStorage210 static BitmaskEnumStorage *construct(AttributeStorageAllocator &allocator,
211 const KeyTy &key) {
212 return new (allocator.allocate<BitmaskEnumStorage>())
213 BitmaskEnumStorage(key);
214 }
215
216 KeyTy value = 0;
217 };
218 } // namespace detail
219 } // namespace vector
220 } // namespace mlir
221
get(CombiningKind kind,MLIRContext * context)222 CombiningKindAttr CombiningKindAttr::get(CombiningKind kind,
223 MLIRContext *context) {
224 return Base::get(context, static_cast<uint64_t>(kind));
225 }
226
getKind() const227 CombiningKind CombiningKindAttr::getKind() const {
228 return static_cast<CombiningKind>(getImpl()->value);
229 }
230
231 static constexpr const CombiningKind combiningKindsList[] = {
232 // clang-format off
233 CombiningKind::ADD,
234 CombiningKind::MUL,
235 CombiningKind::MINUI,
236 CombiningKind::MINSI,
237 CombiningKind::MINF,
238 CombiningKind::MAXUI,
239 CombiningKind::MAXSI,
240 CombiningKind::MAXF,
241 CombiningKind::AND,
242 CombiningKind::OR,
243 CombiningKind::XOR,
244 // clang-format on
245 };
246
print(AsmPrinter & printer) const247 void CombiningKindAttr::print(AsmPrinter &printer) const {
248 printer << "<";
249 auto kinds = llvm::make_filter_range(combiningKindsList, [&](auto kind) {
250 return bitEnumContains(this->getKind(), kind);
251 });
252 llvm::interleaveComma(kinds, printer,
253 [&](auto kind) { printer << stringifyEnum(kind); });
254 printer << ">";
255 }
256
parse(AsmParser & parser,Type type)257 Attribute CombiningKindAttr::parse(AsmParser &parser, Type type) {
258 if (failed(parser.parseLess()))
259 return {};
260
261 StringRef elemName;
262 if (failed(parser.parseKeyword(&elemName)))
263 return {};
264
265 auto kind = symbolizeCombiningKind(elemName);
266 if (!kind) {
267 parser.emitError(parser.getNameLoc(), "Unknown combining kind: ")
268 << elemName;
269 return {};
270 }
271
272 if (failed(parser.parseGreater()))
273 return {};
274
275 return CombiningKindAttr::get(*kind, parser.getContext());
276 }
277
parseAttribute(DialectAsmParser & parser,Type type) const278 Attribute VectorDialect::parseAttribute(DialectAsmParser &parser,
279 Type type) const {
280 StringRef attrKind;
281 if (parser.parseKeyword(&attrKind))
282 return {};
283
284 if (attrKind == "kind")
285 return CombiningKindAttr::parse(parser, {});
286
287 parser.emitError(parser.getNameLoc(), "Unknown attribute type: ") << attrKind;
288 return {};
289 }
290
printAttribute(Attribute attr,DialectAsmPrinter & os) const291 void VectorDialect::printAttribute(Attribute attr,
292 DialectAsmPrinter &os) const {
293 if (auto ck = attr.dyn_cast<CombiningKindAttr>()) {
294 os << "kind";
295 ck.print(os);
296 return;
297 }
298 llvm_unreachable("Unknown attribute type");
299 }
300
301 //===----------------------------------------------------------------------===//
302 // VectorDialect
303 //===----------------------------------------------------------------------===//
304
initialize()305 void VectorDialect::initialize() {
306 addAttributes<CombiningKindAttr>();
307
308 addOperations<
309 #define GET_OP_LIST
310 #include "mlir/Dialect/Vector/IR/VectorOps.cpp.inc"
311 >();
312 }
313
314 /// Materialize a single constant operation from a given attribute value with
315 /// the desired resultant type.
materializeConstant(OpBuilder & builder,Attribute value,Type type,Location loc)316 Operation *VectorDialect::materializeConstant(OpBuilder &builder,
317 Attribute value, Type type,
318 Location loc) {
319 return builder.create<arith::ConstantOp>(loc, type, value);
320 }
321
getVectorSubscriptType(Builder & builder)322 IntegerType vector::getVectorSubscriptType(Builder &builder) {
323 return builder.getIntegerType(64);
324 }
325
getVectorSubscriptAttr(Builder & builder,ArrayRef<int64_t> values)326 ArrayAttr vector::getVectorSubscriptAttr(Builder &builder,
327 ArrayRef<int64_t> values) {
328 return builder.getI64ArrayAttr(values);
329 }
330
331 //===----------------------------------------------------------------------===//
332 // MultiDimReductionOp
333 //===----------------------------------------------------------------------===//
334
build(OpBuilder & builder,OperationState & result,Value source,Value acc,ArrayRef<bool> reductionMask,CombiningKind kind)335 void vector::MultiDimReductionOp::build(OpBuilder &builder,
336 OperationState &result, Value source,
337 Value acc, ArrayRef<bool> reductionMask,
338 CombiningKind kind) {
339 SmallVector<int64_t> reductionDims;
340 for (const auto &en : llvm::enumerate(reductionMask))
341 if (en.value())
342 reductionDims.push_back(en.index());
343 build(builder, result, kind, source, acc,
344 builder.getI64ArrayAttr(reductionDims));
345 }
346
fold(ArrayRef<Attribute> operands)347 OpFoldResult MultiDimReductionOp::fold(ArrayRef<Attribute> operands) {
348 // Single parallel dim, this is a noop.
349 if (getSourceVectorType().getRank() == 1 && !isReducedDim(0))
350 return getSource();
351 return {};
352 }
353
getShapeForUnroll()354 Optional<SmallVector<int64_t, 4>> MultiDimReductionOp::getShapeForUnroll() {
355 return llvm::to_vector<4>(getSourceVectorType().getShape());
356 }
357
verify()358 LogicalResult MultiDimReductionOp::verify() {
359 SmallVector<int64_t> targetShape;
360 Type inferredReturnType;
361 for (auto it : llvm::enumerate(getSourceVectorType().getShape()))
362 if (!llvm::any_of(getReductionDims().getValue(), [&](Attribute attr) {
363 return attr.cast<IntegerAttr>().getValue() == it.index();
364 }))
365 targetShape.push_back(it.value());
366 // TODO: update to also allow 0-d vectors when available.
367 if (targetShape.empty())
368 inferredReturnType = getSourceVectorType().getElementType();
369 else
370 inferredReturnType =
371 VectorType::get(targetShape, getSourceVectorType().getElementType());
372 if (getType() != inferredReturnType)
373 return emitOpError() << "destination type " << getType()
374 << " is incompatible with source type "
375 << getSourceVectorType();
376
377 return success();
378 }
379
380 //===----------------------------------------------------------------------===//
381 // ReductionOp
382 //===----------------------------------------------------------------------===//
383
build(OpBuilder & builder,OperationState & result,CombiningKind kind,Value vector)384 void vector::ReductionOp::build(OpBuilder &builder, OperationState &result,
385 CombiningKind kind, Value vector) {
386 build(builder, result, kind, vector, /*acc=*/Value());
387 }
388
build(OpBuilder & builder,OperationState & result,CombiningKind kind,Value vector,Value acc)389 void vector::ReductionOp::build(OpBuilder &builder, OperationState &result,
390 CombiningKind kind, Value vector, Value acc) {
391 build(builder, result, vector.getType().cast<VectorType>().getElementType(),
392 kind, vector, acc);
393 }
394
verify()395 LogicalResult ReductionOp::verify() {
396 // Verify for 1-D vector.
397 int64_t rank = getVectorType().getRank();
398 if (rank != 1)
399 return emitOpError("unsupported reduction rank: ") << rank;
400
401 // Verify supported reduction kind.
402 Type eltType = getDest().getType();
403 if (!isSupportedCombiningKind(getKind(), eltType))
404 return emitOpError("unsupported reduction type '")
405 << eltType << "' for kind '" << stringifyCombiningKind(getKind())
406 << "'";
407
408 return success();
409 }
410
parse(OpAsmParser & parser,OperationState & result)411 ParseResult ReductionOp::parse(OpAsmParser &parser, OperationState &result) {
412 SmallVector<OpAsmParser::UnresolvedOperand, 2> operandsInfo;
413 Type redType;
414 Type resType;
415 CombiningKindAttr kindAttr;
416 if (parser.parseCustomAttributeWithFallback(kindAttr, Type{}, "kind",
417 result.attributes) ||
418 parser.parseComma() || parser.parseOperandList(operandsInfo) ||
419 parser.parseColonType(redType) ||
420 parser.parseKeywordType("into", resType) ||
421 (!operandsInfo.empty() &&
422 parser.resolveOperand(operandsInfo[0], redType, result.operands)) ||
423 (operandsInfo.size() > 1 &&
424 parser.resolveOperand(operandsInfo[1], resType, result.operands)) ||
425 parser.addTypeToList(resType, result.types))
426 return failure();
427 if (operandsInfo.empty() || operandsInfo.size() > 2)
428 return parser.emitError(parser.getNameLoc(),
429 "unsupported number of operands");
430 return success();
431 }
432
print(OpAsmPrinter & p)433 void ReductionOp::print(OpAsmPrinter &p) {
434 p << " ";
435 getKindAttr().print(p);
436 p << ", " << getVector();
437 if (getAcc())
438 p << ", " << getAcc();
439 p << " : " << getVector().getType() << " into " << getDest().getType();
440 }
441
getVectorReductionOp(arith::AtomicRMWKind op,OpBuilder & builder,Location loc,Value vector)442 Value mlir::vector::getVectorReductionOp(arith::AtomicRMWKind op,
443 OpBuilder &builder, Location loc,
444 Value vector) {
445 switch (op) {
446 case arith::AtomicRMWKind::addf:
447 case arith::AtomicRMWKind::addi:
448 return builder.create<vector::ReductionOp>(vector.getLoc(),
449 CombiningKind::ADD, vector);
450 case arith::AtomicRMWKind::mulf:
451 case arith::AtomicRMWKind::muli:
452 return builder.create<vector::ReductionOp>(vector.getLoc(),
453 CombiningKind::MUL, vector);
454 case arith::AtomicRMWKind::minf:
455 return builder.create<vector::ReductionOp>(vector.getLoc(),
456 CombiningKind::MINF, vector);
457 case arith::AtomicRMWKind::mins:
458 return builder.create<vector::ReductionOp>(vector.getLoc(),
459 CombiningKind::MINSI, vector);
460 case arith::AtomicRMWKind::minu:
461 return builder.create<vector::ReductionOp>(vector.getLoc(),
462 CombiningKind::MINUI, vector);
463 case arith::AtomicRMWKind::maxf:
464 return builder.create<vector::ReductionOp>(vector.getLoc(),
465 CombiningKind::MAXF, vector);
466 case arith::AtomicRMWKind::maxs:
467 return builder.create<vector::ReductionOp>(vector.getLoc(),
468 CombiningKind::MAXSI, vector);
469 case arith::AtomicRMWKind::maxu:
470 return builder.create<vector::ReductionOp>(vector.getLoc(),
471 CombiningKind::MAXUI, vector);
472 case arith::AtomicRMWKind::andi:
473 return builder.create<vector::ReductionOp>(vector.getLoc(),
474 CombiningKind::AND, vector);
475 case arith::AtomicRMWKind::ori:
476 return builder.create<vector::ReductionOp>(vector.getLoc(),
477 CombiningKind::OR, vector);
478 // TODO: Add remaining reduction operations.
479 default:
480 (void)emitOptionalError(loc, "Reduction operation type not supported");
481 break;
482 }
483 return nullptr;
484 }
485
getShapeForUnroll()486 Optional<SmallVector<int64_t, 4>> ReductionOp::getShapeForUnroll() {
487 return llvm::to_vector<4>(getVectorType().getShape());
488 }
489
490 namespace {
491 struct ElideSingleElementReduction : public OpRewritePattern<ReductionOp> {
492 using OpRewritePattern::OpRewritePattern;
493
matchAndRewrite__anon088a7a4f0411::ElideSingleElementReduction494 LogicalResult matchAndRewrite(ReductionOp reductionOp,
495 PatternRewriter &rewriter) const override {
496 if (reductionOp.getVectorType().getDimSize(0) != 1)
497 return failure();
498
499 Location loc = reductionOp.getLoc();
500 Value result = rewriter.create<ExtractOp>(loc, reductionOp.getType(),
501 reductionOp.getVector(),
502 rewriter.getI64ArrayAttr(0));
503
504 if (Value acc = reductionOp.getAcc())
505 result = vector::makeArithReduction(rewriter, loc, reductionOp.getKind(),
506 result, acc);
507
508 rewriter.replaceOp(reductionOp, result);
509 return success();
510 }
511 };
512 } // namespace
513
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)514 void ReductionOp::getCanonicalizationPatterns(RewritePatternSet &results,
515 MLIRContext *context) {
516 results.add<ElideSingleElementReduction>(context);
517 }
518
519 //===----------------------------------------------------------------------===//
520 // ContractionOp
521 //===----------------------------------------------------------------------===//
522
build(OpBuilder & builder,OperationState & result,Value lhs,Value rhs,Value acc,ArrayRef<ArrayRef<AffineExpr>> indexingExprs,ArrayRef<StringRef> iteratorTypes)523 void vector::ContractionOp::build(OpBuilder &builder, OperationState &result,
524 Value lhs, Value rhs, Value acc,
525 ArrayRef<ArrayRef<AffineExpr>> indexingExprs,
526 ArrayRef<StringRef> iteratorTypes) {
527 result.addOperands({lhs, rhs, acc});
528 result.addTypes(acc.getType());
529 result.addAttribute(::mlir::getIndexingMapsAttrName(),
530 builder.getAffineMapArrayAttr(
531 AffineMap::inferFromExprList(indexingExprs)));
532 result.addAttribute(::mlir::getIteratorTypesAttrName(),
533 builder.getStrArrayAttr(iteratorTypes));
534 }
535
build(OpBuilder & builder,OperationState & result,Value lhs,Value rhs,Value acc,ArrayAttr indexingMaps,ArrayAttr iteratorTypes)536 void vector::ContractionOp::build(OpBuilder &builder, OperationState &result,
537 Value lhs, Value rhs, Value acc,
538 ArrayAttr indexingMaps,
539 ArrayAttr iteratorTypes) {
540 build(builder, result, lhs, rhs, acc, indexingMaps, iteratorTypes,
541 ContractionOp::getDefaultKind());
542 }
543
build(OpBuilder & builder,OperationState & result,Value lhs,Value rhs,Value acc,ArrayAttr indexingMaps,ArrayAttr iteratorTypes,CombiningKind kind)544 void vector::ContractionOp::build(OpBuilder &builder, OperationState &result,
545 Value lhs, Value rhs, Value acc,
546 ArrayAttr indexingMaps,
547 ArrayAttr iteratorTypes, CombiningKind kind) {
548 result.addOperands({lhs, rhs, acc});
549 result.addTypes(acc.getType());
550 result.addAttribute(::mlir::getIndexingMapsAttrName(), indexingMaps);
551 result.addAttribute(::mlir::getIteratorTypesAttrName(), iteratorTypes);
552 result.addAttribute(ContractionOp::getKindAttrStrName(),
553 CombiningKindAttr::get(kind, builder.getContext()));
554 }
555
parse(OpAsmParser & parser,OperationState & result)556 ParseResult ContractionOp::parse(OpAsmParser &parser, OperationState &result) {
557 OpAsmParser::UnresolvedOperand lhsInfo;
558 OpAsmParser::UnresolvedOperand rhsInfo;
559 OpAsmParser::UnresolvedOperand accInfo;
560 SmallVector<OpAsmParser::UnresolvedOperand, 2> masksInfo;
561 SmallVector<Type, 2> types;
562 Type resultType;
563 auto loc = parser.getCurrentLocation();
564 DictionaryAttr dictAttr;
565 // TODO: Unify linalg op attribute parsing.
566 if (parser.parseAttribute(dictAttr, "_", result.attributes) ||
567 parser.parseOperand(lhsInfo) || parser.parseComma() ||
568 parser.parseOperand(rhsInfo) || parser.parseComma() ||
569 parser.parseOperand(accInfo) ||
570 parser.parseTrailingOperandList(masksInfo) ||
571 parser.parseOptionalAttrDict(result.attributes) ||
572 parser.parseColonTypeList(types) ||
573 parser.parseKeywordType("into", resultType) ||
574 parser.resolveOperand(lhsInfo, types[0], result.operands) ||
575 parser.resolveOperand(rhsInfo, types[1], result.operands) ||
576 parser.resolveOperand(accInfo, resultType, result.operands) ||
577 parser.addTypeToList(resultType, result.types))
578 return failure();
579 result.attributes.assign(dictAttr.getValue().begin(),
580 dictAttr.getValue().end());
581 if (!result.attributes.get(ContractionOp::getKindAttrStrName())) {
582 result.addAttribute(ContractionOp::getKindAttrStrName(),
583 CombiningKindAttr::get(ContractionOp::getDefaultKind(),
584 result.getContext()));
585 }
586 if (masksInfo.empty())
587 return success();
588 if (masksInfo.size() != 2)
589 return parser.emitError(parser.getNameLoc(),
590 "expected zero or exactly 2 vector mask operands");
591 auto lhsType = types[0].cast<VectorType>();
592 auto rhsType = types[1].cast<VectorType>();
593 auto maskElementType = parser.getBuilder().getI1Type();
594 std::array<Type, 2> maskTypes = {
595 VectorType::Builder(lhsType).setElementType(maskElementType),
596 VectorType::Builder(rhsType).setElementType(maskElementType)};
597 if (parser.resolveOperands(masksInfo, maskTypes, loc, result.operands))
598 return failure();
599 return success();
600 }
601
print(OpAsmPrinter & p)602 void ContractionOp::print(OpAsmPrinter &p) {
603 // TODO: Unify printing code with linalg ops.
604 auto attrNames = getTraitAttrNames();
605 llvm::StringSet<> traitAttrsSet;
606 traitAttrsSet.insert(attrNames.begin(), attrNames.end());
607 SmallVector<NamedAttribute, 8> attrs;
608 for (auto attr : (*this)->getAttrs())
609 if (traitAttrsSet.count(attr.getName().strref()) > 0)
610 attrs.push_back(attr);
611
612 auto dictAttr = DictionaryAttr::get(getContext(), attrs);
613 p << " " << dictAttr << " " << getLhs() << ", ";
614 p << getRhs() << ", " << getAcc();
615 if (getMasks().size() == 2)
616 p << ", " << getMasks();
617
618 p.printOptionalAttrDict((*this)->getAttrs(), attrNames);
619 p << " : " << getLhs().getType() << ", " << getRhs().getType() << " into "
620 << getResultType();
621 }
622
verifyDimMap(VectorType lhsType,VectorType rhsType,const std::vector<std::pair<int64_t,int64_t>> & map)623 static bool verifyDimMap(VectorType lhsType, VectorType rhsType,
624 const std::vector<std::pair<int64_t, int64_t>> &map) {
625 for (auto &dimPair : map) {
626 if (dimPair.first < 0 || dimPair.first >= lhsType.getRank() ||
627 dimPair.second < 0 || dimPair.second >= rhsType.getRank() ||
628 lhsType.getDimSize(dimPair.first) != rhsType.getDimSize(dimPair.second))
629 return false;
630 }
631 return true;
632 }
633
verifyOutputShape(ContractionOp op,VectorType lhsType,VectorType rhsType,Type accType,Type resType,const std::vector<std::pair<int64_t,int64_t>> & contractingDimMap,const std::vector<std::pair<int64_t,int64_t>> & batchDimMap)634 static LogicalResult verifyOutputShape(
635 ContractionOp op, VectorType lhsType, VectorType rhsType, Type accType,
636 Type resType,
637 const std::vector<std::pair<int64_t, int64_t>> &contractingDimMap,
638 const std::vector<std::pair<int64_t, int64_t>> &batchDimMap) {
639 DenseSet<int64_t> lhsContractingDimSet;
640 DenseSet<int64_t> rhsContractingDimSet;
641 for (auto &dimPair : contractingDimMap) {
642 lhsContractingDimSet.insert(dimPair.first);
643 rhsContractingDimSet.insert(dimPair.second);
644 }
645 DenseSet<int64_t> rhsBatchDimSet;
646 for (auto &dimPair : batchDimMap)
647 rhsBatchDimSet.insert(dimPair.second);
648
649 // Add free and batch dimensions from 'lhsType' to 'expectedResultDims'.
650 SmallVector<int64_t, 4> expectedResultDims;
651 for (int64_t i = 0, e = lhsType.getRank(); i < e; ++i) {
652 if (lhsContractingDimSet.count(i) > 0)
653 continue;
654 expectedResultDims.push_back(lhsType.getDimSize(i));
655 }
656
657 // Add free dimensions from 'rhsType' to 'expectedResultDims'.
658 for (int64_t i = 0, e = rhsType.getRank(); i < e; ++i) {
659 if (rhsContractingDimSet.count(i) > 0 || rhsBatchDimSet.count(i) > 0)
660 continue;
661 expectedResultDims.push_back(rhsType.getDimSize(i));
662 }
663
664 // Verify 'expectedResultDims'.
665 if (expectedResultDims.empty()) {
666 // No batch or free dimension implies a scalar result.
667 if (resType.isa<VectorType>() || accType.isa<VectorType>())
668 return op.emitOpError("invalid accumulator/result vector shape");
669 } else {
670 // At least one batch or free dimension implies a vector result.
671 auto resVectorType = resType.dyn_cast<VectorType>();
672 auto accVectorType = accType.dyn_cast<VectorType>();
673 if (!resVectorType || !accVectorType)
674 return op.emitOpError("invalid accumulator/result vector shape");
675
676 // Infer expected result vector type. Lhs + rhs map and lhs + rhs vector
677 // types fully define the result vector type. This assumes the affine maps
678 // are well-formed, which must have been verified already.
679 MLIRContext *ctx = op.getContext();
680 AffineMap lhsMap = op.getIndexingMapsArray()[0];
681 AffineMap rhsMap = op.getIndexingMapsArray()[1];
682 if (getUnusedDimsBitVector({lhsMap, rhsMap}).any())
683 return op.emitOpError(
684 "expected all dimensions to be either a LHS or a RHS dimension");
685 SmallVector<AffineExpr, 4> extents(lhsMap.getNumInputs());
686 for (auto pair :
687 {std::make_pair(lhsType, lhsMap), std::make_pair(rhsType, rhsMap)}) {
688 VectorType v = pair.first;
689 auto map = pair.second;
690 for (unsigned idx = 0, e = v.getRank(); idx < e; ++idx) {
691 unsigned pos = map.getDimPosition(idx);
692 if (!extents[pos])
693 extents[pos] = getAffineConstantExpr(v.getShape()[idx], ctx);
694 }
695 }
696 if (!llvm::all_of(extents, [](AffineExpr e) { return e; }))
697 return op.emitOpError("expected all dimensions to get an extent as "
698 "either a LHS or a RHS dimension");
699
700 AffineMap resMap = op.getIndexingMapsArray()[2];
701 auto extentsMap = AffineMap::get(/*dimCount=*/extents.size(),
702 /*symCount=*/0, extents, ctx);
703 // Compose the resMap with the extentsMap, which is a constant map.
704 AffineMap expectedMap = simplifyAffineMap(resMap.compose(extentsMap));
705 assert(llvm::all_of(
706 expectedMap.getResults(),
707 [](AffineExpr e) { return e.isa<AffineConstantExpr>(); }) &&
708 "expected constant extent along all dimensions.");
709 // Extract the expected shape and build the type.
710 auto expectedShape = llvm::to_vector<4>(
711 llvm::map_range(expectedMap.getResults(), [](AffineExpr e) {
712 return e.cast<AffineConstantExpr>().getValue();
713 }));
714 auto expected =
715 VectorType::get(expectedShape, resVectorType.getElementType());
716 if (resVectorType != expected || accVectorType != expected)
717 return op.emitOpError(
718 "invalid accumulator/result vector shape, expected: ")
719 << expected;
720 }
721 return success();
722 }
723
verify()724 LogicalResult ContractionOp::verify() {
725 auto lhsType = getLhsType();
726 auto rhsType = getRhsType();
727 auto accType = getAccType();
728 auto resType = getResultType();
729
730 // Verify that an indexing map was specified for each vector operand.
731 if (getIndexingMapsArray().size() != 3)
732 return emitOpError("expected an indexing map for each vector operand");
733
734 // Verify that each index map has 'numIterators' inputs, no symbols, and
735 // that the number of map outputs equals the rank of its associated
736 // vector operand.
737 unsigned numIterators = getIteratorTypes().getValue().size();
738 for (const auto &it : llvm::enumerate(getIndexingMapsArray())) {
739 auto index = it.index();
740 auto map = it.value();
741 if (map.getNumSymbols() != 0)
742 return emitOpError("expected indexing map ")
743 << index << " to have no symbols";
744 auto vectorType = getOperand(index).getType().dyn_cast<VectorType>();
745 unsigned rank = vectorType ? vectorType.getShape().size() : 0;
746 // Verify that the map has the right number of inputs, outputs, and indices.
747 // This also correctly accounts for (..) -> () for rank-0 results.
748 if (map.getNumDims() != numIterators)
749 return emitOpError("expected indexing map ")
750 << index << " to have " << numIterators << " number of inputs";
751 if (map.getNumResults() != rank)
752 return emitOpError("expected indexing map ")
753 << index << " to have " << rank << " number of outputs";
754 if (!map.isProjectedPermutation())
755 return emitOpError("expected indexing map ")
756 << index << " to be a projected permutation of its inputs";
757 }
758
759 auto contractingDimMap = getContractingDimMap();
760 auto batchDimMap = getBatchDimMap();
761
762 // Verify at least one contracting dimension pair was specified.
763 if (contractingDimMap.empty())
764 return emitOpError("expected at least one contracting dimension pair");
765
766 // Verify contracting dimension map was properly constructed.
767 if (!verifyDimMap(lhsType, rhsType, contractingDimMap))
768 return emitOpError("invalid contracting dimension map");
769
770 // Verify batch dimension map was properly constructed.
771 if (!verifyDimMap(lhsType, rhsType, batchDimMap))
772 return emitOpError("invalid batch dimension map");
773
774 // Verify 'accType' and 'resType' shape.
775 if (failed(verifyOutputShape(*this, lhsType, rhsType, accType, resType,
776 contractingDimMap, batchDimMap)))
777 return failure();
778
779 // Verify that either two vector masks are set or none are set.
780 auto lhsMaskType = getLHSVectorMaskType();
781 auto rhsMaskType = getRHSVectorMaskType();
782 if ((lhsMaskType && !rhsMaskType) || (!lhsMaskType && rhsMaskType))
783 return emitOpError("invalid number of vector masks specified");
784 if (lhsMaskType && rhsMaskType) {
785 // Verify mask rank == argument rank.
786 if (lhsMaskType.getShape().size() != lhsType.getShape().size() ||
787 rhsMaskType.getShape().size() != rhsType.getShape().size())
788 return emitOpError("invalid vector mask rank");
789 }
790
791 // Verify supported combining kind.
792 auto vectorType = resType.dyn_cast<VectorType>();
793 auto elementType = vectorType ? vectorType.getElementType() : resType;
794 if (!isSupportedCombiningKind(getKind(), elementType))
795 return emitOpError("unsupported contraction type");
796
797 return success();
798 }
799
getTraitAttrNames()800 ArrayRef<StringRef> ContractionOp::getTraitAttrNames() {
801 static constexpr StringRef names[3] = {::mlir::getIndexingMapsAttrName(),
802 ::mlir::getIteratorTypesAttrName(),
803 ContractionOp::getKindAttrStrName()};
804 return llvm::makeArrayRef(names);
805 }
806
getResultIndex(AffineMap map,AffineExpr targetExpr)807 static int64_t getResultIndex(AffineMap map, AffineExpr targetExpr) {
808 for (int64_t i = 0, e = map.getNumResults(); i < e; ++i)
809 if (targetExpr == map.getResult(i))
810 return i;
811 return -1;
812 }
813
814 static std::vector<std::pair<int64_t, int64_t>>
getDimMap(ArrayRef<AffineMap> indexingMaps,ArrayAttr iteratorTypes,StringRef targetIteratorTypeName,MLIRContext * context)815 getDimMap(ArrayRef<AffineMap> indexingMaps, ArrayAttr iteratorTypes,
816 StringRef targetIteratorTypeName, MLIRContext *context) {
817 std::vector<std::pair<int64_t, int64_t>> dimMap;
818 for (const auto &it : llvm::enumerate(iteratorTypes)) {
819 auto iteratorTypeName = it.value().cast<StringAttr>().getValue();
820 if (iteratorTypeName != targetIteratorTypeName)
821 continue;
822 // Search lhs/rhs map results for 'targetExpr'.
823 auto targetExpr = getAffineDimExpr(it.index(), context);
824 int64_t lhsDim = getResultIndex(indexingMaps[0], targetExpr);
825 int64_t rhsDim = getResultIndex(indexingMaps[1], targetExpr);
826 if (lhsDim >= 0 && rhsDim >= 0)
827 dimMap.emplace_back(lhsDim, rhsDim);
828 }
829 return dimMap;
830 }
831
getIterationBounds(SmallVectorImpl<int64_t> & iterationBounds)832 void ContractionOp::getIterationBounds(
833 SmallVectorImpl<int64_t> &iterationBounds) {
834 auto lhsShape = getLhsType().getShape();
835 auto resVectorType = getResultType().dyn_cast<VectorType>();
836 SmallVector<AffineMap, 4> indexingMaps(getIndexingMapsArray());
837 SmallVector<int64_t, 2> iterationShape;
838 for (const auto &it : llvm::enumerate(getIteratorTypes())) {
839 // Search lhs/rhs map results for 'targetExpr'.
840 auto targetExpr = getAffineDimExpr(it.index(), getContext());
841 auto iteratorTypeName = it.value().cast<StringAttr>().getValue();
842 if (iteratorTypeName == getReductionIteratorTypeName()) {
843 // Get reduction dim size from lhs shape (same size in rhsShape).
844 int64_t lhsDimIndex = getResultIndex(indexingMaps[0], targetExpr);
845 assert(lhsDimIndex >= 0);
846 iterationBounds.push_back(lhsShape[lhsDimIndex]);
847 continue;
848 }
849 // Get parallel dimension size from result shape.
850 int64_t resDimIndex = getResultIndex(indexingMaps[2], targetExpr);
851 assert(resDimIndex >= 0);
852 assert(resVectorType != nullptr);
853 iterationBounds.push_back(resVectorType.getShape()[resDimIndex]);
854 }
855 }
856
getIterationIndexMap(std::vector<DenseMap<int64_t,int64_t>> & iterationIndexMap)857 void ContractionOp::getIterationIndexMap(
858 std::vector<DenseMap<int64_t, int64_t>> &iterationIndexMap) {
859 unsigned numMaps = getIndexingMapsArray().size();
860 iterationIndexMap.resize(numMaps);
861 for (const auto &it : llvm::enumerate(getIndexingMapsArray())) {
862 auto index = it.index();
863 auto map = it.value();
864 for (unsigned i = 0, e = map.getNumResults(); i < e; ++i) {
865 auto dim = map.getResult(i).cast<AffineDimExpr>();
866 iterationIndexMap[index][dim.getPosition()] = i;
867 }
868 }
869 }
870
getContractingDimMap()871 std::vector<std::pair<int64_t, int64_t>> ContractionOp::getContractingDimMap() {
872 SmallVector<AffineMap, 4> indexingMaps(getIndexingMapsArray());
873 return getDimMap(indexingMaps, getIteratorTypes(),
874 getReductionIteratorTypeName(), getContext());
875 }
876
getBatchDimMap()877 std::vector<std::pair<int64_t, int64_t>> ContractionOp::getBatchDimMap() {
878 SmallVector<AffineMap, 4> indexingMaps(getIndexingMapsArray());
879 return getDimMap(indexingMaps, getIteratorTypes(),
880 getParallelIteratorTypeName(), getContext());
881 }
882
getShapeForUnroll()883 Optional<SmallVector<int64_t, 4>> ContractionOp::getShapeForUnroll() {
884 SmallVector<int64_t, 4> shape;
885 getIterationBounds(shape);
886 return shape;
887 }
888
889 /// Return a fused vector::ContractionOp which represents a patterns such as:
890 ///
891 /// ```mlir
892 /// %c0 = vector.constant 0: ...
893 /// %c = vector.contract %a, %b, %c0: ...
894 /// %e = add %c, %d: ...
895 /// ```
896 ///
897 /// by:
898 ///
899 /// ```mlir
900 /// %e = vector.contract %a, %b, %d: ...
901 /// ```
902 ///
903 /// Return null if the canonicalization does not apply.
904 // TODO: This should be a folding of Add into Contract in core but while they
905 // live in different dialects, it is not possible without unnatural
906 // dependencies.
907 template <typename AddOpType>
908 struct CanonicalizeContractAdd : public OpRewritePattern<AddOpType> {
909 using OpRewritePattern<AddOpType>::OpRewritePattern;
910
matchAndRewriteCanonicalizeContractAdd911 LogicalResult matchAndRewrite(AddOpType addOp,
912 PatternRewriter &rewriter) const override {
913 auto canonicalize = [&](Value maybeContraction,
914 Value otherOperand) -> vector::ContractionOp {
915 vector::ContractionOp contractionOp =
916 dyn_cast_or_null<vector::ContractionOp>(
917 maybeContraction.getDefiningOp());
918 if (!contractionOp)
919 return vector::ContractionOp();
920 if (auto maybeZero = dyn_cast_or_null<arith::ConstantOp>(
921 contractionOp.getAcc().getDefiningOp())) {
922 if (maybeZero.getValue() ==
923 rewriter.getZeroAttr(contractionOp.getAcc().getType())) {
924 BlockAndValueMapping bvm;
925 bvm.map(contractionOp.getAcc(), otherOperand);
926 auto newContraction =
927 cast<vector::ContractionOp>(rewriter.clone(*contractionOp, bvm));
928 rewriter.replaceOp(addOp, newContraction.getResult());
929 return newContraction;
930 }
931 }
932 return vector::ContractionOp();
933 };
934
935 Value a = addOp->getOperand(0), b = addOp->getOperand(1);
936 vector::ContractionOp contract = canonicalize(a, b);
937 contract = contract ? contract : canonicalize(b, a);
938 return contract ? success() : failure();
939 }
940 };
941
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)942 void ContractionOp::getCanonicalizationPatterns(RewritePatternSet &results,
943 MLIRContext *context) {
944 results.add<CanonicalizeContractAdd<arith::AddIOp>,
945 CanonicalizeContractAdd<arith::AddFOp>>(context);
946 }
947
948 //===----------------------------------------------------------------------===//
949 // ExtractElementOp
950 //===----------------------------------------------------------------------===//
951
build(OpBuilder & builder,OperationState & result,Value source)952 void vector::ExtractElementOp::build(OpBuilder &builder, OperationState &result,
953 Value source) {
954 result.addOperands({source});
955 result.addTypes(source.getType().cast<VectorType>().getElementType());
956 }
957
build(OpBuilder & builder,OperationState & result,Value source,Value position)958 void vector::ExtractElementOp::build(OpBuilder &builder, OperationState &result,
959 Value source, Value position) {
960 result.addOperands({source, position});
961 result.addTypes(source.getType().cast<VectorType>().getElementType());
962 }
963
verify()964 LogicalResult vector::ExtractElementOp::verify() {
965 VectorType vectorType = getVectorType();
966 if (vectorType.getRank() == 0) {
967 if (getPosition())
968 return emitOpError("expected position to be empty with 0-D vector");
969 return success();
970 }
971 if (vectorType.getRank() != 1)
972 return emitOpError("unexpected >1 vector rank");
973 if (!getPosition())
974 return emitOpError("expected position for 1-D vector");
975 return success();
976 }
977
fold(ArrayRef<Attribute> operands)978 OpFoldResult vector::ExtractElementOp::fold(ArrayRef<Attribute> operands) {
979 // Skip the 0-D vector here now.
980 if (operands.size() < 2)
981 return {};
982
983 Attribute src = operands[0];
984 Attribute pos = operands[1];
985
986 // Fold extractelement (splat X) -> X.
987 if (auto splat = getVector().getDefiningOp<vector::SplatOp>())
988 return splat.getInput();
989
990 if (!pos || !src)
991 return {};
992
993 auto srcElements = src.cast<DenseElementsAttr>().getValues<Attribute>();
994
995 auto attr = pos.dyn_cast<IntegerAttr>();
996 uint64_t posIdx = attr.getInt();
997
998 return srcElements[posIdx];
999 }
1000
1001 //===----------------------------------------------------------------------===//
1002 // ExtractOp
1003 //===----------------------------------------------------------------------===//
1004
build(OpBuilder & builder,OperationState & result,Value source,ArrayRef<int64_t> position)1005 void vector::ExtractOp::build(OpBuilder &builder, OperationState &result,
1006 Value source, ArrayRef<int64_t> position) {
1007 build(builder, result, source, getVectorSubscriptAttr(builder, position));
1008 }
1009
1010 // Convenience builder which assumes the values are constant indices.
build(OpBuilder & builder,OperationState & result,Value source,ValueRange position)1011 void vector::ExtractOp::build(OpBuilder &builder, OperationState &result,
1012 Value source, ValueRange position) {
1013 SmallVector<int64_t, 4> positionConstants =
1014 llvm::to_vector<4>(llvm::map_range(position, [](Value pos) {
1015 return pos.getDefiningOp<arith::ConstantIndexOp>().value();
1016 }));
1017 build(builder, result, source, positionConstants);
1018 }
1019
1020 LogicalResult
inferReturnTypes(MLIRContext *,Optional<Location>,ValueRange operands,DictionaryAttr attributes,RegionRange,SmallVectorImpl<Type> & inferredReturnTypes)1021 ExtractOp::inferReturnTypes(MLIRContext *, Optional<Location>,
1022 ValueRange operands, DictionaryAttr attributes,
1023 RegionRange,
1024 SmallVectorImpl<Type> &inferredReturnTypes) {
1025 ExtractOp::Adaptor op(operands, attributes);
1026 auto vectorType = op.getVector().getType().cast<VectorType>();
1027 if (static_cast<int64_t>(op.getPosition().size()) == vectorType.getRank()) {
1028 inferredReturnTypes.push_back(vectorType.getElementType());
1029 } else {
1030 auto n =
1031 std::min<size_t>(op.getPosition().size(), vectorType.getRank() - 1);
1032 inferredReturnTypes.push_back(VectorType::get(
1033 vectorType.getShape().drop_front(n), vectorType.getElementType()));
1034 }
1035 return success();
1036 }
1037
isCompatibleReturnTypes(TypeRange l,TypeRange r)1038 bool ExtractOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
1039 // Allow extracting 1-element vectors instead of scalars.
1040 auto isCompatible = [](TypeRange l, TypeRange r) {
1041 auto vectorType = l.front().dyn_cast<VectorType>();
1042 return vectorType && vectorType.getShape().equals({1}) &&
1043 vectorType.getElementType() == r.front();
1044 };
1045 if (l.size() == 1 && r.size() == 1 &&
1046 (isCompatible(l, r) || isCompatible(r, l)))
1047 return true;
1048 return l == r;
1049 }
1050
verify()1051 LogicalResult vector::ExtractOp::verify() {
1052 auto positionAttr = getPosition().getValue();
1053 if (positionAttr.size() > static_cast<unsigned>(getVectorType().getRank()))
1054 return emitOpError(
1055 "expected position attribute of rank smaller than vector rank");
1056 for (const auto &en : llvm::enumerate(positionAttr)) {
1057 auto attr = en.value().dyn_cast<IntegerAttr>();
1058 if (!attr || attr.getInt() < 0 ||
1059 attr.getInt() >= getVectorType().getDimSize(en.index()))
1060 return emitOpError("expected position attribute #")
1061 << (en.index() + 1)
1062 << " to be a non-negative integer smaller than the corresponding "
1063 "vector dimension";
1064 }
1065 return success();
1066 }
1067
1068 template <typename IntType>
extractVector(ArrayAttr arrayAttr)1069 static SmallVector<IntType> extractVector(ArrayAttr arrayAttr) {
1070 return llvm::to_vector<4>(llvm::map_range(
1071 arrayAttr.getAsRange<IntegerAttr>(),
1072 [](IntegerAttr attr) { return static_cast<IntType>(attr.getInt()); }));
1073 }
1074
1075 /// Fold the result of chains of ExtractOp in place by simply concatenating the
1076 /// positions.
foldExtractOpFromExtractChain(ExtractOp extractOp)1077 static LogicalResult foldExtractOpFromExtractChain(ExtractOp extractOp) {
1078 if (!extractOp.getVector().getDefiningOp<ExtractOp>())
1079 return failure();
1080
1081 SmallVector<int64_t, 4> globalPosition;
1082 ExtractOp currentOp = extractOp;
1083 auto extrPos = extractVector<int64_t>(currentOp.getPosition());
1084 globalPosition.append(extrPos.rbegin(), extrPos.rend());
1085 while (ExtractOp nextOp = currentOp.getVector().getDefiningOp<ExtractOp>()) {
1086 currentOp = nextOp;
1087 auto extrPos = extractVector<int64_t>(currentOp.getPosition());
1088 globalPosition.append(extrPos.rbegin(), extrPos.rend());
1089 }
1090 extractOp.setOperand(currentOp.getVector());
1091 // OpBuilder is only used as a helper to build an I64ArrayAttr.
1092 OpBuilder b(extractOp.getContext());
1093 std::reverse(globalPosition.begin(), globalPosition.end());
1094 extractOp->setAttr(ExtractOp::getPositionAttrStrName(),
1095 b.getI64ArrayAttr(globalPosition));
1096 return success();
1097 }
1098
1099 namespace {
1100 /// Fold an ExtractOp that is fed by a chain of InsertOps and TransposeOps.
1101 /// Walk back a chain of InsertOp/TransposeOp until we hit a match.
1102 /// Compose TransposeOp permutations as we walk back.
1103 /// This helper class keeps an updated extraction position `extractPosition`
1104 /// with extra trailing sentinels.
1105 /// The sentinels encode the internal transposition status of the result vector.
1106 /// As we iterate, extractPosition is permuted and updated.
1107 class ExtractFromInsertTransposeChainState {
1108 public:
1109 ExtractFromInsertTransposeChainState(ExtractOp e);
1110
1111 /// Iterate over producing insert and transpose ops until we find a fold.
1112 Value fold();
1113
1114 private:
1115 /// Return true if the vector at position `a` is contained within the vector
1116 /// at position `b`. Under insert/extract semantics, this is the same as `a`
1117 /// is a prefix of `b`.
1118 template <typename ContainerA, typename ContainerB>
isContainedWithin(const ContainerA & a,const ContainerB & b)1119 bool isContainedWithin(const ContainerA &a, const ContainerB &b) {
1120 return a.size() <= b.size() &&
1121 std::equal(a.begin(), a.begin() + a.size(), b.begin());
1122 }
1123
1124 /// Return true if the vector at position `a` intersects the vector at
1125 /// position `b`. Under insert/extract semantics, this is the same as equality
1126 /// of all entries of `a` that are >=0 with the corresponding entries of b.
1127 /// Comparison is on the common prefix (i.e. zip).
1128 template <typename ContainerA, typename ContainerB>
intersectsWhereNonNegative(const ContainerA & a,const ContainerB & b)1129 bool intersectsWhereNonNegative(const ContainerA &a, const ContainerB &b) {
1130 for (auto it : llvm::zip(a, b)) {
1131 if (std::get<0>(it) < 0 || std::get<0>(it) < 0)
1132 continue;
1133 if (std::get<0>(it) != std::get<1>(it))
1134 return false;
1135 }
1136 return true;
1137 }
1138
1139 /// Folding is only possible in the absence of an internal permutation in the
1140 /// result vector.
canFold()1141 bool canFold() {
1142 return (sentinels ==
1143 makeArrayRef(extractPosition).drop_front(extractedRank));
1144 }
1145
1146 // Helper to get the next defining op of interest.
updateStateForNextIteration(Value v)1147 void updateStateForNextIteration(Value v) {
1148 nextInsertOp = v.getDefiningOp<vector::InsertOp>();
1149 nextTransposeOp = v.getDefiningOp<vector::TransposeOp>();
1150 };
1151
1152 // Case 1. If we hit a transpose, just compose the map and iterate.
1153 // Invariant: insert + transpose do not change rank, we can always compose.
1154 LogicalResult handleTransposeOp();
1155
1156 // Case 2: the insert position matches extractPosition exactly, early return.
1157 LogicalResult handleInsertOpWithMatchingPos(Value &res);
1158
1159 /// Case 3: if the insert position is a prefix of extractPosition, extract a
1160 /// portion of the source of the insert.
1161 /// Example:
1162 /// ```
1163 /// %ins = vector.insert %source, %vest[1]: vector<3x4> into vector<2x3x4x5>
1164 /// // extractPosition == [1, 2, 3]
1165 /// %ext = vector.extract %ins[1, 0]: vector<3x4x5>
1166 /// // can fold to vector.extract %source[0, 3]
1167 /// %ext = vector.extract %source[3]: vector<5x6>
1168 /// ```
1169 /// To traverse through %source, we need to set the leading dims to 0 and
1170 /// drop the extra leading dims.
1171 /// This method updates the internal state.
1172 LogicalResult handleInsertOpWithPrefixPos(Value &res);
1173
1174 /// Try to fold in place to extract(source, extractPosition) and return the
1175 /// folded result. Return null if folding is not possible (e.g. due to an
1176 /// internal tranposition in the result).
1177 Value tryToFoldExtractOpInPlace(Value source);
1178
1179 ExtractOp extractOp;
1180 int64_t vectorRank;
1181 int64_t extractedRank;
1182
1183 InsertOp nextInsertOp;
1184 TransposeOp nextTransposeOp;
1185
1186 /// Sentinel values that encode the internal permutation status of the result.
1187 /// They are set to (-1, ... , -k) at the beginning and appended to
1188 /// `extractPosition`.
1189 /// In the end, the tail of `extractPosition` must be exactly `sentinels` to
1190 /// ensure that there is no internal transposition.
1191 /// Internal transposition cannot be accounted for with a folding pattern.
1192 // TODO: We could relax the internal transposition with an extra transposition
1193 // operation in a future canonicalizer.
1194 SmallVector<int64_t> sentinels;
1195 SmallVector<int64_t> extractPosition;
1196 };
1197 } // namespace
1198
ExtractFromInsertTransposeChainState(ExtractOp e)1199 ExtractFromInsertTransposeChainState::ExtractFromInsertTransposeChainState(
1200 ExtractOp e)
1201 : extractOp(e), vectorRank(extractOp.getVectorType().getRank()),
1202 extractedRank(extractOp.getPosition().size()) {
1203 assert(vectorRank >= extractedRank && "extracted pos overflow");
1204 sentinels.reserve(vectorRank - extractedRank);
1205 for (int64_t i = 0, e = vectorRank - extractedRank; i < e; ++i)
1206 sentinels.push_back(-(i + 1));
1207 extractPosition = extractVector<int64_t>(extractOp.getPosition());
1208 llvm::append_range(extractPosition, sentinels);
1209 }
1210
1211 // Case 1. If we hit a transpose, just compose the map and iterate.
1212 // Invariant: insert + transpose do not change rank, we can always compose.
handleTransposeOp()1213 LogicalResult ExtractFromInsertTransposeChainState::handleTransposeOp() {
1214 if (!nextTransposeOp)
1215 return failure();
1216 auto permutation = extractVector<unsigned>(nextTransposeOp.getTransp());
1217 AffineMap m = inversePermutation(
1218 AffineMap::getPermutationMap(permutation, extractOp.getContext()));
1219 extractPosition = applyPermutationMap(m, makeArrayRef(extractPosition));
1220 return success();
1221 }
1222
1223 // Case 2: the insert position matches extractPosition exactly, early return.
1224 LogicalResult
handleInsertOpWithMatchingPos(Value & res)1225 ExtractFromInsertTransposeChainState::handleInsertOpWithMatchingPos(
1226 Value &res) {
1227 auto insertedPos = extractVector<int64_t>(nextInsertOp.getPosition());
1228 if (makeArrayRef(insertedPos) !=
1229 llvm::makeArrayRef(extractPosition).take_front(extractedRank))
1230 return failure();
1231 // Case 2.a. early-exit fold.
1232 res = nextInsertOp.getSource();
1233 // Case 2.b. if internal transposition is present, canFold will be false.
1234 return success();
1235 }
1236
1237 /// Case 3: if inserted position is a prefix of extractPosition,
1238 /// extract a portion of the source of the insertion.
1239 /// This method updates the internal state.
1240 LogicalResult
handleInsertOpWithPrefixPos(Value & res)1241 ExtractFromInsertTransposeChainState::handleInsertOpWithPrefixPos(Value &res) {
1242 auto insertedPos = extractVector<int64_t>(nextInsertOp.getPosition());
1243 if (!isContainedWithin(insertedPos, extractPosition))
1244 return failure();
1245 // Set leading dims to zero.
1246 std::fill_n(extractPosition.begin(), insertedPos.size(), 0);
1247 // Drop extra leading dims.
1248 extractPosition.erase(extractPosition.begin(),
1249 extractPosition.begin() + insertedPos.size());
1250 extractedRank = extractPosition.size() - sentinels.size();
1251 // Case 3.a. early-exit fold (break and delegate to post-while path).
1252 res = nextInsertOp.getSource();
1253 // Case 3.b. if internal transposition is present, canFold will be false.
1254 return success();
1255 }
1256
1257 /// Try to fold in place to extract(source, extractPosition) and return the
1258 /// folded result. Return null if folding is not possible (e.g. due to an
1259 /// internal tranposition in the result).
tryToFoldExtractOpInPlace(Value source)1260 Value ExtractFromInsertTransposeChainState::tryToFoldExtractOpInPlace(
1261 Value source) {
1262 // If we can't fold (either internal transposition, or nothing to fold), bail.
1263 bool nothingToFold = (source == extractOp.getVector());
1264 if (nothingToFold || !canFold())
1265 return Value();
1266 // Otherwise, fold by updating the op inplace and return its result.
1267 OpBuilder b(extractOp.getContext());
1268 extractOp->setAttr(
1269 extractOp.getPositionAttrName(),
1270 b.getI64ArrayAttr(
1271 makeArrayRef(extractPosition).take_front(extractedRank)));
1272 extractOp.getVectorMutable().assign(source);
1273 return extractOp.getResult();
1274 }
1275
1276 /// Iterate over producing insert and transpose ops until we find a fold.
fold()1277 Value ExtractFromInsertTransposeChainState::fold() {
1278 Value valueToExtractFrom = extractOp.getVector();
1279 updateStateForNextIteration(valueToExtractFrom);
1280 while (nextInsertOp || nextTransposeOp) {
1281 // Case 1. If we hit a transpose, just compose the map and iterate.
1282 // Invariant: insert + transpose do not change rank, we can always compose.
1283 if (succeeded(handleTransposeOp())) {
1284 valueToExtractFrom = nextTransposeOp.getVector();
1285 updateStateForNextIteration(valueToExtractFrom);
1286 continue;
1287 }
1288
1289 Value result;
1290 // Case 2: the position match exactly.
1291 if (succeeded(handleInsertOpWithMatchingPos(result)))
1292 return result;
1293
1294 // Case 3: if the inserted position is a prefix of extractPosition, we can
1295 // just extract a portion of the source of the insert.
1296 if (succeeded(handleInsertOpWithPrefixPos(result)))
1297 return tryToFoldExtractOpInPlace(result);
1298
1299 // Case 4: extractPositionRef intersects insertedPosRef on non-sentinel
1300 // values. This is a more difficult case and we bail.
1301 auto insertedPos = extractVector<int64_t>(nextInsertOp.getPosition());
1302 if (isContainedWithin(extractPosition, insertedPos) ||
1303 intersectsWhereNonNegative(extractPosition, insertedPos))
1304 return Value();
1305
1306 // Case 5: No intersection, we forward the extract to insertOp.dest().
1307 valueToExtractFrom = nextInsertOp.getDest();
1308 updateStateForNextIteration(valueToExtractFrom);
1309 }
1310 // If after all this we can fold, go for it.
1311 return tryToFoldExtractOpInPlace(valueToExtractFrom);
1312 }
1313
1314 /// Fold extractOp with scalar result coming from BroadcastOp or SplatOp.
foldExtractFromBroadcast(ExtractOp extractOp)1315 static Value foldExtractFromBroadcast(ExtractOp extractOp) {
1316 Operation *defOp = extractOp.getVector().getDefiningOp();
1317 if (!defOp || !isa<vector::BroadcastOp, SplatOp>(defOp))
1318 return Value();
1319 Value source = defOp->getOperand(0);
1320 if (extractOp.getType() == source.getType())
1321 return source;
1322 auto getRank = [](Type type) {
1323 return type.isa<VectorType>() ? type.cast<VectorType>().getRank() : 0;
1324 };
1325 unsigned broadcastSrcRank = getRank(source.getType());
1326 unsigned extractResultRank = getRank(extractOp.getType());
1327 if (extractResultRank >= broadcastSrcRank)
1328 return Value();
1329 // Check that the dimension of the result haven't been broadcasted.
1330 auto extractVecType = extractOp.getType().dyn_cast<VectorType>();
1331 auto broadcastVecType = source.getType().dyn_cast<VectorType>();
1332 if (extractVecType && broadcastVecType &&
1333 extractVecType.getShape() !=
1334 broadcastVecType.getShape().take_back(extractResultRank))
1335 return Value();
1336 auto extractPos = extractVector<int64_t>(extractOp.getPosition());
1337 unsigned rankDiff = broadcastSrcRank - extractResultRank;
1338 extractPos.erase(extractPos.begin(),
1339 std::next(extractPos.begin(), extractPos.size() - rankDiff));
1340 extractOp.setOperand(source);
1341 // OpBuilder is only used as a helper to build an I64ArrayAttr.
1342 OpBuilder b(extractOp.getContext());
1343 extractOp->setAttr(ExtractOp::getPositionAttrStrName(),
1344 b.getI64ArrayAttr(extractPos));
1345 return extractOp.getResult();
1346 }
1347
1348 // Fold extractOp with source coming from ShapeCast op.
foldExtractFromShapeCast(ExtractOp extractOp)1349 static Value foldExtractFromShapeCast(ExtractOp extractOp) {
1350 auto shapeCastOp = extractOp.getVector().getDefiningOp<vector::ShapeCastOp>();
1351 if (!shapeCastOp)
1352 return Value();
1353 // Get the nth dimension size starting from lowest dimension.
1354 auto getDimReverse = [](VectorType type, int64_t n) {
1355 return type.getShape().take_back(n + 1).front();
1356 };
1357 int64_t destinationRank =
1358 extractOp.getType().isa<VectorType>()
1359 ? extractOp.getType().cast<VectorType>().getRank()
1360 : 0;
1361 if (destinationRank > shapeCastOp.getSourceVectorType().getRank())
1362 return Value();
1363 if (destinationRank > 0) {
1364 auto destinationType = extractOp.getResult().getType().cast<VectorType>();
1365 for (int64_t i = 0; i < destinationRank; i++) {
1366 // The lowest dimension of of the destination must match the lowest
1367 // dimension of the shapecast op source.
1368 // TODO: This case could be support in a canonicalization pattern.
1369 if (getDimReverse(shapeCastOp.getSourceVectorType(), i) !=
1370 getDimReverse(destinationType, i))
1371 return Value();
1372 }
1373 }
1374 // Extract the strides associated with the extract op vector source. Then use
1375 // this to calculate a linearized position for the extract.
1376 auto extractedPos = extractVector<int64_t>(extractOp.getPosition());
1377 std::reverse(extractedPos.begin(), extractedPos.end());
1378 SmallVector<int64_t, 4> strides;
1379 int64_t stride = 1;
1380 for (int64_t i = 0, e = extractedPos.size(); i < e; i++) {
1381 strides.push_back(stride);
1382 stride *= getDimReverse(extractOp.getVectorType(), i + destinationRank);
1383 }
1384
1385 int64_t position = linearize(extractedPos, strides);
1386 // Then extract the strides associated to the shapeCast op vector source and
1387 // delinearize the position using those strides.
1388 SmallVector<int64_t, 4> newStrides;
1389 int64_t numDimension =
1390 shapeCastOp.getSourceVectorType().getRank() - destinationRank;
1391 stride = 1;
1392 for (int64_t i = 0; i < numDimension; i++) {
1393 newStrides.push_back(stride);
1394 stride *=
1395 getDimReverse(shapeCastOp.getSourceVectorType(), i + destinationRank);
1396 }
1397 std::reverse(newStrides.begin(), newStrides.end());
1398 SmallVector<int64_t, 4> newPosition = delinearize(newStrides, position);
1399 // OpBuilder is only used as a helper to build an I64ArrayAttr.
1400 OpBuilder b(extractOp.getContext());
1401 extractOp->setAttr(ExtractOp::getPositionAttrStrName(),
1402 b.getI64ArrayAttr(newPosition));
1403 extractOp.setOperand(shapeCastOp.getSource());
1404 return extractOp.getResult();
1405 }
1406
1407 /// Fold an ExtractOp from ExtractStridedSliceOp.
foldExtractFromExtractStrided(ExtractOp extractOp)1408 static Value foldExtractFromExtractStrided(ExtractOp extractOp) {
1409 auto extractStridedSliceOp =
1410 extractOp.getVector().getDefiningOp<vector::ExtractStridedSliceOp>();
1411 if (!extractStridedSliceOp)
1412 return Value();
1413 // Return if 'extractStridedSliceOp' has non-unit strides.
1414 if (extractStridedSliceOp.hasNonUnitStrides())
1415 return Value();
1416
1417 // Trim offsets for dimensions fully extracted.
1418 auto sliceOffsets =
1419 extractVector<int64_t>(extractStridedSliceOp.getOffsets());
1420 while (!sliceOffsets.empty()) {
1421 size_t lastOffset = sliceOffsets.size() - 1;
1422 if (sliceOffsets.back() != 0 ||
1423 extractStridedSliceOp.getType().getDimSize(lastOffset) !=
1424 extractStridedSliceOp.getVectorType().getDimSize(lastOffset))
1425 break;
1426 sliceOffsets.pop_back();
1427 }
1428 unsigned destinationRank = 0;
1429 if (auto vecType = extractOp.getType().dyn_cast<VectorType>())
1430 destinationRank = vecType.getRank();
1431 // The dimensions of the result need to be untouched by the
1432 // extractStridedSlice op.
1433 if (destinationRank >
1434 extractStridedSliceOp.getVectorType().getRank() - sliceOffsets.size())
1435 return Value();
1436 auto extractedPos = extractVector<int64_t>(extractOp.getPosition());
1437 assert(extractedPos.size() >= sliceOffsets.size());
1438 for (size_t i = 0, e = sliceOffsets.size(); i < e; i++)
1439 extractedPos[i] = extractedPos[i] + sliceOffsets[i];
1440 extractOp.getVectorMutable().assign(extractStridedSliceOp.getVector());
1441 // OpBuilder is only used as a helper to build an I64ArrayAttr.
1442 OpBuilder b(extractOp.getContext());
1443 extractOp->setAttr(ExtractOp::getPositionAttrStrName(),
1444 b.getI64ArrayAttr(extractedPos));
1445 return extractOp.getResult();
1446 }
1447
1448 /// Fold extract_op fed from a chain of insertStridedSlice ops.
foldExtractStridedOpFromInsertChain(ExtractOp op)1449 static Value foldExtractStridedOpFromInsertChain(ExtractOp op) {
1450 int64_t destinationRank = op.getType().isa<VectorType>()
1451 ? op.getType().cast<VectorType>().getRank()
1452 : 0;
1453 auto insertOp = op.getVector().getDefiningOp<InsertStridedSliceOp>();
1454 while (insertOp) {
1455 int64_t insertRankDiff = insertOp.getDestVectorType().getRank() -
1456 insertOp.getSourceVectorType().getRank();
1457 if (destinationRank > insertOp.getSourceVectorType().getRank())
1458 return Value();
1459 auto insertOffsets = extractVector<int64_t>(insertOp.getOffsets());
1460 auto extractOffsets = extractVector<int64_t>(op.getPosition());
1461
1462 if (llvm::any_of(insertOp.getStrides(), [](Attribute attr) {
1463 return attr.cast<IntegerAttr>().getInt() != 1;
1464 }))
1465 return Value();
1466 bool disjoint = false;
1467 SmallVector<int64_t, 4> offsetDiffs;
1468 for (unsigned dim = 0, e = extractOffsets.size(); dim < e; ++dim) {
1469 int64_t start = insertOffsets[dim];
1470 int64_t size =
1471 (dim < insertRankDiff)
1472 ? 1
1473 : insertOp.getSourceVectorType().getDimSize(dim - insertRankDiff);
1474 int64_t end = start + size;
1475 int64_t offset = extractOffsets[dim];
1476 // Check if the start of the extract offset is in the interval inserted.
1477 if (start <= offset && offset < end) {
1478 if (dim >= insertRankDiff)
1479 offsetDiffs.push_back(offset - start);
1480 continue;
1481 }
1482 disjoint = true;
1483 break;
1484 }
1485 // The extract element chunk overlap with the vector inserted.
1486 if (!disjoint) {
1487 // If any of the inner dimensions are only partially inserted we have a
1488 // partial overlap.
1489 int64_t srcRankDiff =
1490 insertOp.getSourceVectorType().getRank() - destinationRank;
1491 for (int64_t i = 0; i < destinationRank; i++) {
1492 if (insertOp.getSourceVectorType().getDimSize(i + srcRankDiff) !=
1493 insertOp.getDestVectorType().getDimSize(i + srcRankDiff +
1494 insertRankDiff))
1495 return Value();
1496 }
1497 op.getVectorMutable().assign(insertOp.getSource());
1498 // OpBuilder is only used as a helper to build an I64ArrayAttr.
1499 OpBuilder b(op.getContext());
1500 op->setAttr(ExtractOp::getPositionAttrStrName(),
1501 b.getI64ArrayAttr(offsetDiffs));
1502 return op.getResult();
1503 }
1504 // If the chunk extracted is disjoint from the chunk inserted, keep
1505 // looking in the insert chain.
1506 insertOp = insertOp.getDest().getDefiningOp<InsertStridedSliceOp>();
1507 }
1508 return Value();
1509 }
1510
fold(ArrayRef<Attribute>)1511 OpFoldResult ExtractOp::fold(ArrayRef<Attribute>) {
1512 if (getPosition().empty())
1513 return getVector();
1514 if (succeeded(foldExtractOpFromExtractChain(*this)))
1515 return getResult();
1516 if (auto res = ExtractFromInsertTransposeChainState(*this).fold())
1517 return res;
1518 if (auto res = foldExtractFromBroadcast(*this))
1519 return res;
1520 if (auto res = foldExtractFromShapeCast(*this))
1521 return res;
1522 if (auto val = foldExtractFromExtractStrided(*this))
1523 return val;
1524 if (auto val = foldExtractStridedOpFromInsertChain(*this))
1525 return val;
1526 return OpFoldResult();
1527 }
1528
1529 namespace {
1530
1531 // Pattern to rewrite a ExtractOp(Broadcast) -> Broadcast.
1532 class ExtractOpFromBroadcast final : public OpRewritePattern<ExtractOp> {
1533 public:
1534 using OpRewritePattern<ExtractOp>::OpRewritePattern;
1535
matchAndRewrite(ExtractOp extractOp,PatternRewriter & rewriter) const1536 LogicalResult matchAndRewrite(ExtractOp extractOp,
1537 PatternRewriter &rewriter) const override {
1538 Operation *defOp = extractOp.getVector().getDefiningOp();
1539 if (!defOp || !isa<vector::BroadcastOp, SplatOp>(defOp))
1540 return failure();
1541
1542 Value source = defOp->getOperand(0);
1543 if (extractOp.getType() == source.getType())
1544 return failure();
1545 auto getRank = [](Type type) {
1546 return type.isa<VectorType>() ? type.cast<VectorType>().getRank() : 0;
1547 };
1548 unsigned broadcastSrcRank = getRank(source.getType());
1549 unsigned extractResultRank = getRank(extractOp.getType());
1550 // We only consider the case where the rank of the source is less than or
1551 // equal to the rank of the extract dst. The other cases are handled in the
1552 // folding patterns.
1553 if (extractResultRank < broadcastSrcRank)
1554 return failure();
1555 rewriter.replaceOpWithNewOp<vector::BroadcastOp>(
1556 extractOp, extractOp.getType(), source);
1557 return success();
1558 }
1559 };
1560
1561 // Pattern to rewrite a ExtractOp(splat ConstantOp) -> ConstantOp.
1562 class ExtractOpConstantFolder final : public OpRewritePattern<ExtractOp> {
1563 public:
1564 using OpRewritePattern<ExtractOp>::OpRewritePattern;
1565
matchAndRewrite(ExtractOp extractOp,PatternRewriter & rewriter) const1566 LogicalResult matchAndRewrite(ExtractOp extractOp,
1567 PatternRewriter &rewriter) const override {
1568 // Return if 'extractStridedSliceOp' operand is not defined by a
1569 // ConstantOp.
1570 auto constantOp = extractOp.getVector().getDefiningOp<arith::ConstantOp>();
1571 if (!constantOp)
1572 return failure();
1573 auto dense = constantOp.getValue().dyn_cast<SplatElementsAttr>();
1574 if (!dense)
1575 return failure();
1576 Attribute newAttr = dense.getSplatValue<Attribute>();
1577 if (auto vecDstType = extractOp.getType().dyn_cast<VectorType>())
1578 newAttr = DenseElementsAttr::get(vecDstType, newAttr);
1579 rewriter.replaceOpWithNewOp<arith::ConstantOp>(extractOp, newAttr);
1580 return success();
1581 }
1582 };
1583
1584 } // namespace
1585
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)1586 void ExtractOp::getCanonicalizationPatterns(RewritePatternSet &results,
1587 MLIRContext *context) {
1588 results.add<ExtractOpConstantFolder, ExtractOpFromBroadcast>(context);
1589 }
1590
populateFromInt64AttrArray(ArrayAttr arrayAttr,SmallVectorImpl<int64_t> & results)1591 static void populateFromInt64AttrArray(ArrayAttr arrayAttr,
1592 SmallVectorImpl<int64_t> &results) {
1593 for (auto attr : arrayAttr)
1594 results.push_back(attr.cast<IntegerAttr>().getInt());
1595 }
1596
1597 //===----------------------------------------------------------------------===//
1598 // ExtractMapOp
1599 //===----------------------------------------------------------------------===//
1600
build(OpBuilder & builder,OperationState & result,Value vector,ValueRange ids,ArrayRef<int64_t> multiplicity,AffineMap permutationMap)1601 void ExtractMapOp::build(OpBuilder &builder, OperationState &result,
1602 Value vector, ValueRange ids,
1603 ArrayRef<int64_t> multiplicity,
1604 AffineMap permutationMap) {
1605 assert(ids.size() == multiplicity.size() &&
1606 ids.size() == permutationMap.getNumResults());
1607 assert(permutationMap.isProjectedPermutation());
1608 VectorType type = vector.getType().cast<VectorType>();
1609 SmallVector<int64_t, 4> newShape(type.getShape().begin(),
1610 type.getShape().end());
1611 for (unsigned i = 0, e = permutationMap.getNumResults(); i < e; i++) {
1612 AffineExpr expr = permutationMap.getResult(i);
1613 auto dim = expr.cast<AffineDimExpr>();
1614 newShape[dim.getPosition()] = newShape[dim.getPosition()] / multiplicity[i];
1615 }
1616 VectorType resultType = VectorType::get(newShape, type.getElementType());
1617 ExtractMapOp::build(builder, result, resultType, vector, ids);
1618 }
1619
verify()1620 LogicalResult ExtractMapOp::verify() {
1621 if (getSourceVectorType().getRank() != getResultType().getRank())
1622 return emitOpError("expected source and destination vectors of same rank");
1623 unsigned numId = 0;
1624 for (unsigned i = 0, e = getSourceVectorType().getRank(); i < e; ++i) {
1625 if (getSourceVectorType().getDimSize(i) % getResultType().getDimSize(i) !=
1626 0)
1627 return emitOpError("source vector dimensions must be a multiple of "
1628 "destination vector dimensions");
1629 if (getSourceVectorType().getDimSize(i) != getResultType().getDimSize(i))
1630 numId++;
1631 }
1632 if (numId != getIds().size())
1633 return emitOpError("expected number of ids must match the number of "
1634 "dimensions distributed");
1635 return success();
1636 }
1637
fold(ArrayRef<Attribute> operands)1638 OpFoldResult ExtractMapOp::fold(ArrayRef<Attribute> operands) {
1639 auto insert = getVector().getDefiningOp<vector::InsertMapOp>();
1640 if (insert == nullptr || getType() != insert.getVector().getType() ||
1641 getIds() != insert.getIds())
1642 return {};
1643 return insert.getVector();
1644 }
1645
getMultiplicity(SmallVectorImpl<int64_t> & multiplicity)1646 void ExtractMapOp::getMultiplicity(SmallVectorImpl<int64_t> &multiplicity) {
1647 assert(multiplicity.empty());
1648 for (unsigned i = 0, e = getSourceVectorType().getRank(); i < e; i++) {
1649 if (getSourceVectorType().getDimSize(i) != getResultType().getDimSize(i))
1650 multiplicity.push_back(getSourceVectorType().getDimSize(i) /
1651 getResultType().getDimSize(i));
1652 }
1653 }
1654
1655 template <typename MapOp>
calculateImplicitMap(MapOp op)1656 AffineMap calculateImplicitMap(MapOp op) {
1657 SmallVector<AffineExpr, 4> perm;
1658 // Check which dimension have a multiplicity greater than 1 and associated
1659 // them to the IDs in order.
1660 for (unsigned i = 0, e = op.getSourceVectorType().getRank(); i < e; i++) {
1661 if (op.getSourceVectorType().getDimSize(i) !=
1662 op.getResultType().getDimSize(i))
1663 perm.push_back(getAffineDimExpr(i, op.getContext()));
1664 }
1665 auto map = AffineMap::get(op.getSourceVectorType().getRank(), 0, perm,
1666 op.getContext());
1667 return map;
1668 }
1669
map()1670 AffineMap ExtractMapOp::map() { return calculateImplicitMap(*this); }
1671
1672 //===----------------------------------------------------------------------===//
1673 // FmaOp
1674 //===----------------------------------------------------------------------===//
1675
getShapeForUnroll()1676 Optional<SmallVector<int64_t, 4>> FMAOp::getShapeForUnroll() {
1677 return llvm::to_vector<4>(getVectorType().getShape());
1678 }
1679
1680 //===----------------------------------------------------------------------===//
1681 // BroadcastOp
1682 //===----------------------------------------------------------------------===//
1683
1684 BroadcastableToResult
isBroadcastableTo(Type srcType,VectorType dstVectorType,std::pair<int,int> * mismatchingDims)1685 mlir::vector::isBroadcastableTo(Type srcType, VectorType dstVectorType,
1686 std::pair<int, int> *mismatchingDims) {
1687 // Broadcast scalar to vector of the same element type.
1688 if (srcType.isIntOrIndexOrFloat() && dstVectorType &&
1689 getElementTypeOrSelf(srcType) == getElementTypeOrSelf(dstVectorType))
1690 return BroadcastableToResult::Success;
1691 // From now on, only vectors broadcast.
1692 VectorType srcVectorType = srcType.dyn_cast<VectorType>();
1693 if (!srcVectorType)
1694 return BroadcastableToResult::SourceTypeNotAVector;
1695
1696 int64_t srcRank = srcVectorType.getRank();
1697 int64_t dstRank = dstVectorType.getRank();
1698 if (srcRank > dstRank)
1699 return BroadcastableToResult::SourceRankHigher;
1700 // Source has an exact match or singleton value for all trailing dimensions
1701 // (all leading dimensions are simply duplicated).
1702 int64_t lead = dstRank - srcRank;
1703 for (int64_t r = 0; r < srcRank; ++r) {
1704 int64_t srcDim = srcVectorType.getDimSize(r);
1705 int64_t dstDim = dstVectorType.getDimSize(lead + r);
1706 if (srcDim != 1 && srcDim != dstDim) {
1707 if (mismatchingDims) {
1708 mismatchingDims->first = srcDim;
1709 mismatchingDims->second = dstDim;
1710 }
1711 return BroadcastableToResult::DimensionMismatch;
1712 }
1713 }
1714
1715 return BroadcastableToResult::Success;
1716 }
1717
verify()1718 LogicalResult BroadcastOp::verify() {
1719 std::pair<int, int> mismatchingDims;
1720 BroadcastableToResult res =
1721 isBroadcastableTo(getSourceType(), getVectorType(), &mismatchingDims);
1722 if (res == BroadcastableToResult::Success)
1723 return success();
1724 if (res == BroadcastableToResult::SourceRankHigher)
1725 return emitOpError("source rank higher than destination rank");
1726 if (res == BroadcastableToResult::DimensionMismatch)
1727 return emitOpError("dimension mismatch (")
1728 << mismatchingDims.first << " vs. " << mismatchingDims.second << ")";
1729 if (res == BroadcastableToResult::SourceTypeNotAVector)
1730 return emitOpError("source type is not a vector");
1731 llvm_unreachable("unexpected vector.broadcast op error");
1732 }
1733
fold(ArrayRef<Attribute> operands)1734 OpFoldResult BroadcastOp::fold(ArrayRef<Attribute> operands) {
1735 if (getSourceType() == getVectorType())
1736 return getSource();
1737 if (!operands[0])
1738 return {};
1739 auto vectorType = getVectorType();
1740 if (operands[0].getType().isIntOrIndexOrFloat())
1741 return DenseElementsAttr::get(vectorType, operands[0]);
1742 if (auto attr = operands[0].dyn_cast<SplatElementsAttr>())
1743 return DenseElementsAttr::get(vectorType, attr.getSplatValue<Attribute>());
1744 return {};
1745 }
1746
1747 namespace {
1748
1749 // Fold broadcast1(broadcast2(x)) into broadcast1(x).
1750 struct BroadcastFolder : public OpRewritePattern<BroadcastOp> {
1751 using OpRewritePattern<BroadcastOp>::OpRewritePattern;
1752
matchAndRewrite__anon088a7a4f1211::BroadcastFolder1753 LogicalResult matchAndRewrite(BroadcastOp broadcastOp,
1754 PatternRewriter &rewriter) const override {
1755 auto srcBroadcast = broadcastOp.getSource().getDefiningOp<BroadcastOp>();
1756 if (!srcBroadcast)
1757 return failure();
1758 rewriter.replaceOpWithNewOp<BroadcastOp>(
1759 broadcastOp, broadcastOp.getVectorType(), srcBroadcast.getSource());
1760 return success();
1761 }
1762 };
1763 } // namespace
1764
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)1765 void BroadcastOp::getCanonicalizationPatterns(RewritePatternSet &results,
1766 MLIRContext *context) {
1767 // BroadcastToShapeCast is not a default canonicalization, it is opt-in by
1768 // calling `populateCastAwayVectorLeadingOneDimPatterns`
1769 results.add<BroadcastFolder>(context);
1770 }
1771
1772 //===----------------------------------------------------------------------===//
1773 // ShuffleOp
1774 //===----------------------------------------------------------------------===//
1775
build(OpBuilder & builder,OperationState & result,Value v1,Value v2,ArrayRef<int64_t> mask)1776 void ShuffleOp::build(OpBuilder &builder, OperationState &result, Value v1,
1777 Value v2, ArrayRef<int64_t> mask) {
1778 build(builder, result, v1, v2, getVectorSubscriptAttr(builder, mask));
1779 }
1780
verify()1781 LogicalResult ShuffleOp::verify() {
1782 VectorType resultType = getVectorType();
1783 VectorType v1Type = getV1VectorType();
1784 VectorType v2Type = getV2VectorType();
1785 // Verify ranks.
1786 int64_t resRank = resultType.getRank();
1787 int64_t v1Rank = v1Type.getRank();
1788 int64_t v2Rank = v2Type.getRank();
1789 if (resRank != v1Rank || v1Rank != v2Rank)
1790 return emitOpError("rank mismatch");
1791 // Verify all but leading dimension sizes.
1792 for (int64_t r = 1; r < v1Rank; ++r) {
1793 int64_t resDim = resultType.getDimSize(r);
1794 int64_t v1Dim = v1Type.getDimSize(r);
1795 int64_t v2Dim = v2Type.getDimSize(r);
1796 if (resDim != v1Dim || v1Dim != v2Dim)
1797 return emitOpError("dimension mismatch");
1798 }
1799 // Verify mask length.
1800 auto maskAttr = getMask().getValue();
1801 int64_t maskLength = maskAttr.size();
1802 if (maskLength <= 0)
1803 return emitOpError("invalid mask length");
1804 if (maskLength != resultType.getDimSize(0))
1805 return emitOpError("mask length mismatch");
1806 // Verify all indices.
1807 int64_t indexSize = v1Type.getDimSize(0) + v2Type.getDimSize(0);
1808 for (const auto &en : llvm::enumerate(maskAttr)) {
1809 auto attr = en.value().dyn_cast<IntegerAttr>();
1810 if (!attr || attr.getInt() < 0 || attr.getInt() >= indexSize)
1811 return emitOpError("mask index #") << (en.index() + 1) << " out of range";
1812 }
1813 return success();
1814 }
1815
1816 LogicalResult
inferReturnTypes(MLIRContext *,Optional<Location>,ValueRange operands,DictionaryAttr attributes,RegionRange,SmallVectorImpl<Type> & inferredReturnTypes)1817 ShuffleOp::inferReturnTypes(MLIRContext *, Optional<Location>,
1818 ValueRange operands, DictionaryAttr attributes,
1819 RegionRange,
1820 SmallVectorImpl<Type> &inferredReturnTypes) {
1821 ShuffleOp::Adaptor op(operands, attributes);
1822 auto v1Type = op.getV1().getType().cast<VectorType>();
1823 // Construct resulting type: leading dimension matches mask length,
1824 // all trailing dimensions match the operands.
1825 SmallVector<int64_t, 4> shape;
1826 shape.reserve(v1Type.getRank());
1827 shape.push_back(std::max<size_t>(1, op.getMask().size()));
1828 llvm::append_range(shape, v1Type.getShape().drop_front());
1829 inferredReturnTypes.push_back(
1830 VectorType::get(shape, v1Type.getElementType()));
1831 return success();
1832 }
1833
isStepIndexArray(ArrayAttr idxArr,uint64_t begin,size_t width)1834 static bool isStepIndexArray(ArrayAttr idxArr, uint64_t begin, size_t width) {
1835 uint64_t expected = begin;
1836 return idxArr.size() == width &&
1837 llvm::all_of(idxArr.getAsValueRange<IntegerAttr>(),
1838 [&expected](auto attr) {
1839 return attr.getZExtValue() == expected++;
1840 });
1841 }
1842
fold(ArrayRef<Attribute> operands)1843 OpFoldResult vector::ShuffleOp::fold(ArrayRef<Attribute> operands) {
1844 // fold shuffle V1, V2, [0, 1, 2, 3] : <4xi32>, <2xi32> -> V1
1845 if (!getV1VectorType().isScalable() &&
1846 isStepIndexArray(getMask(), 0, getV1VectorType().getDimSize(0)))
1847 return getV1();
1848 // fold shuffle V1, V2, [4, 5] : <4xi32>, <2xi32> -> V2
1849 if (!getV1VectorType().isScalable() && !getV2VectorType().isScalable() &&
1850 isStepIndexArray(getMask(), getV1VectorType().getDimSize(0),
1851 getV2VectorType().getDimSize(0)))
1852 return getV2();
1853
1854 Attribute lhs = operands.front(), rhs = operands.back();
1855 if (!lhs || !rhs)
1856 return {};
1857
1858 auto lhsType = lhs.getType().cast<VectorType>();
1859 // Only support 1-D for now to avoid complicated n-D DenseElementsAttr
1860 // manipulation.
1861 if (lhsType.getRank() != 1)
1862 return {};
1863 int64_t lhsSize = lhsType.getDimSize(0);
1864
1865 SmallVector<Attribute> results;
1866 auto lhsElements = lhs.cast<DenseElementsAttr>().getValues<Attribute>();
1867 auto rhsElements = rhs.cast<DenseElementsAttr>().getValues<Attribute>();
1868 for (const auto &index : this->getMask().getAsValueRange<IntegerAttr>()) {
1869 int64_t i = index.getZExtValue();
1870 if (i >= lhsSize) {
1871 results.push_back(rhsElements[i - lhsSize]);
1872 } else {
1873 results.push_back(lhsElements[i]);
1874 }
1875 }
1876
1877 return DenseElementsAttr::get(getVectorType(), results);
1878 }
1879
1880 namespace {
1881
1882 /// Pattern to rewrite a ShuffleOp(SplatOp, SplatOp) to SplatOp.
1883 class ShuffleSplat final : public OpRewritePattern<ShuffleOp> {
1884 public:
1885 using OpRewritePattern<ShuffleOp>::OpRewritePattern;
1886
matchAndRewrite(ShuffleOp op,PatternRewriter & rewriter) const1887 LogicalResult matchAndRewrite(ShuffleOp op,
1888 PatternRewriter &rewriter) const override {
1889 auto v1Splat = op.getV1().getDefiningOp<SplatOp>();
1890 auto v2Splat = op.getV2().getDefiningOp<SplatOp>();
1891
1892 if (!v1Splat || !v2Splat)
1893 return failure();
1894
1895 if (v1Splat.getInput() != v2Splat.getInput())
1896 return failure();
1897
1898 rewriter.replaceOpWithNewOp<SplatOp>(op, op.getType(), v1Splat.getInput());
1899 return success();
1900 }
1901 };
1902
1903 } // namespace
1904
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)1905 void ShuffleOp::getCanonicalizationPatterns(RewritePatternSet &results,
1906 MLIRContext *context) {
1907 results.add<ShuffleSplat>(context);
1908 }
1909
1910 //===----------------------------------------------------------------------===//
1911 // InsertElementOp
1912 //===----------------------------------------------------------------------===//
1913
build(OpBuilder & builder,OperationState & result,Value source,Value dest)1914 void InsertElementOp::build(OpBuilder &builder, OperationState &result,
1915 Value source, Value dest) {
1916 build(builder, result, source, dest, {});
1917 }
1918
verify()1919 LogicalResult InsertElementOp::verify() {
1920 auto dstVectorType = getDestVectorType();
1921 if (dstVectorType.getRank() == 0) {
1922 if (getPosition())
1923 return emitOpError("expected position to be empty with 0-D vector");
1924 return success();
1925 }
1926 if (dstVectorType.getRank() != 1)
1927 return emitOpError("unexpected >1 vector rank");
1928 if (!getPosition())
1929 return emitOpError("expected position for 1-D vector");
1930 return success();
1931 }
1932
fold(ArrayRef<Attribute> operands)1933 OpFoldResult vector::InsertElementOp::fold(ArrayRef<Attribute> operands) {
1934 // Skip the 0-D vector here.
1935 if (operands.size() < 3)
1936 return {};
1937
1938 Attribute src = operands[0];
1939 Attribute dst = operands[1];
1940 Attribute pos = operands[2];
1941 if (!src || !dst || !pos)
1942 return {};
1943
1944 auto dstElements = dst.cast<DenseElementsAttr>().getValues<Attribute>();
1945
1946 SmallVector<Attribute> results(dstElements);
1947
1948 auto attr = pos.dyn_cast<IntegerAttr>();
1949 uint64_t posIdx = attr.getInt();
1950
1951 results[posIdx] = src;
1952
1953 return DenseElementsAttr::get(getDestVectorType(), results);
1954 }
1955
1956 //===----------------------------------------------------------------------===//
1957 // InsertOp
1958 //===----------------------------------------------------------------------===//
1959
build(OpBuilder & builder,OperationState & result,Value source,Value dest,ArrayRef<int64_t> position)1960 void InsertOp::build(OpBuilder &builder, OperationState &result, Value source,
1961 Value dest, ArrayRef<int64_t> position) {
1962 result.addOperands({source, dest});
1963 auto positionAttr = getVectorSubscriptAttr(builder, position);
1964 result.addTypes(dest.getType());
1965 result.addAttribute(getPositionAttrStrName(), positionAttr);
1966 }
1967
1968 // Convenience builder which assumes the values are constant indices.
build(OpBuilder & builder,OperationState & result,Value source,Value dest,ValueRange position)1969 void InsertOp::build(OpBuilder &builder, OperationState &result, Value source,
1970 Value dest, ValueRange position) {
1971 SmallVector<int64_t, 4> positionConstants =
1972 llvm::to_vector<4>(llvm::map_range(position, [](Value pos) {
1973 return pos.getDefiningOp<arith::ConstantIndexOp>().value();
1974 }));
1975 build(builder, result, source, dest, positionConstants);
1976 }
1977
verify()1978 LogicalResult InsertOp::verify() {
1979 auto positionAttr = getPosition().getValue();
1980 auto destVectorType = getDestVectorType();
1981 if (positionAttr.size() > static_cast<unsigned>(destVectorType.getRank()))
1982 return emitOpError(
1983 "expected position attribute of rank smaller than dest vector rank");
1984 auto srcVectorType = getSourceType().dyn_cast<VectorType>();
1985 if (srcVectorType &&
1986 (static_cast<unsigned>(srcVectorType.getRank()) + positionAttr.size() !=
1987 static_cast<unsigned>(destVectorType.getRank())))
1988 return emitOpError("expected position attribute rank + source rank to "
1989 "match dest vector rank");
1990 if (!srcVectorType &&
1991 (positionAttr.size() != static_cast<unsigned>(destVectorType.getRank())))
1992 return emitOpError(
1993 "expected position attribute rank to match the dest vector rank");
1994 for (const auto &en : llvm::enumerate(positionAttr)) {
1995 auto attr = en.value().dyn_cast<IntegerAttr>();
1996 if (!attr || attr.getInt() < 0 ||
1997 attr.getInt() >= destVectorType.getDimSize(en.index()))
1998 return emitOpError("expected position attribute #")
1999 << (en.index() + 1)
2000 << " to be a non-negative integer smaller than the corresponding "
2001 "dest vector dimension";
2002 }
2003 return success();
2004 }
2005
2006 namespace {
2007
2008 // If insertOp is only inserting unit dimensions it can be transformed to a
2009 // broadcast.
2010 class InsertToBroadcast final : public OpRewritePattern<InsertOp> {
2011 public:
2012 using OpRewritePattern<InsertOp>::OpRewritePattern;
2013
matchAndRewrite(InsertOp insertOp,PatternRewriter & rewriter) const2014 LogicalResult matchAndRewrite(InsertOp insertOp,
2015 PatternRewriter &rewriter) const override {
2016 auto srcVecType = insertOp.getSourceType().dyn_cast<VectorType>();
2017 if (!srcVecType || insertOp.getDestVectorType().getNumElements() !=
2018 srcVecType.getNumElements())
2019 return failure();
2020 rewriter.replaceOpWithNewOp<BroadcastOp>(
2021 insertOp, insertOp.getDestVectorType(), insertOp.getSource());
2022 return success();
2023 }
2024 };
2025
2026 /// Pattern to rewrite a InsertOp(SplatOp, SplatOp) to SplatOp.
2027 class InsertSplatToSplat final : public OpRewritePattern<InsertOp> {
2028 public:
2029 using OpRewritePattern<InsertOp>::OpRewritePattern;
2030
matchAndRewrite(InsertOp op,PatternRewriter & rewriter) const2031 LogicalResult matchAndRewrite(InsertOp op,
2032 PatternRewriter &rewriter) const override {
2033 auto srcSplat = op.getSource().getDefiningOp<SplatOp>();
2034 auto dstSplat = op.getDest().getDefiningOp<SplatOp>();
2035
2036 if (!srcSplat || !dstSplat)
2037 return failure();
2038
2039 if (srcSplat.getInput() != dstSplat.getInput())
2040 return failure();
2041
2042 rewriter.replaceOpWithNewOp<SplatOp>(op, op.getType(), srcSplat.getInput());
2043 return success();
2044 }
2045 };
2046
2047 } // namespace
2048
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)2049 void InsertOp::getCanonicalizationPatterns(RewritePatternSet &results,
2050 MLIRContext *context) {
2051 results.add<InsertToBroadcast, BroadcastFolder, InsertSplatToSplat>(context);
2052 }
2053
2054 // Eliminates insert operations that produce values identical to their source
2055 // value. This happens when the source and destination vectors have identical
2056 // sizes.
fold(ArrayRef<Attribute> operands)2057 OpFoldResult vector::InsertOp::fold(ArrayRef<Attribute> operands) {
2058 if (getPosition().empty())
2059 return getSource();
2060 return {};
2061 }
2062
2063 //===----------------------------------------------------------------------===//
2064 // InsertMapOp
2065 //===----------------------------------------------------------------------===//
2066
verify()2067 LogicalResult InsertMapOp::verify() {
2068 if (getSourceVectorType().getRank() != getResultType().getRank())
2069 return emitOpError("expected source and destination vectors of same rank");
2070 unsigned numId = 0;
2071 for (unsigned i = 0, e = getResultType().getRank(); i < e; i++) {
2072 if (getResultType().getDimSize(i) % getSourceVectorType().getDimSize(i) !=
2073 0)
2074 return emitOpError(
2075 "destination vector size must be a multiple of source vector size");
2076 if (getResultType().getDimSize(i) != getSourceVectorType().getDimSize(i))
2077 numId++;
2078 }
2079 if (numId != getIds().size())
2080 return emitOpError("expected number of ids must match the number of "
2081 "dimensions distributed");
2082 return success();
2083 }
2084
map()2085 AffineMap InsertMapOp::map() { return calculateImplicitMap(*this); }
2086
2087 //===----------------------------------------------------------------------===//
2088 // InsertStridedSliceOp
2089 //===----------------------------------------------------------------------===//
2090
build(OpBuilder & builder,OperationState & result,Value source,Value dest,ArrayRef<int64_t> offsets,ArrayRef<int64_t> strides)2091 void InsertStridedSliceOp::build(OpBuilder &builder, OperationState &result,
2092 Value source, Value dest,
2093 ArrayRef<int64_t> offsets,
2094 ArrayRef<int64_t> strides) {
2095 result.addOperands({source, dest});
2096 auto offsetsAttr = getVectorSubscriptAttr(builder, offsets);
2097 auto stridesAttr = getVectorSubscriptAttr(builder, strides);
2098 result.addTypes(dest.getType());
2099 result.addAttribute(getOffsetsAttrStrName(), offsetsAttr);
2100 result.addAttribute(getStridesAttrStrName(), stridesAttr);
2101 }
2102
2103 // TODO: Should be moved to Tablegen Confined attributes.
2104 template <typename OpType>
isIntegerArrayAttrSmallerThanShape(OpType op,ArrayAttr arrayAttr,ArrayRef<int64_t> shape,StringRef attrName)2105 static LogicalResult isIntegerArrayAttrSmallerThanShape(OpType op,
2106 ArrayAttr arrayAttr,
2107 ArrayRef<int64_t> shape,
2108 StringRef attrName) {
2109 if (arrayAttr.size() > shape.size())
2110 return op.emitOpError("expected ")
2111 << attrName << " attribute of rank smaller than vector rank";
2112 return success();
2113 }
2114
2115 // Returns true if all integers in `arrayAttr` are in the half-open [min, max}
2116 // interval. If `halfOpen` is true then the admissible interval is [min, max).
2117 // Otherwise, the admissible interval is [min, max].
2118 template <typename OpType>
2119 static LogicalResult
isIntegerArrayAttrConfinedToRange(OpType op,ArrayAttr arrayAttr,int64_t min,int64_t max,StringRef attrName,bool halfOpen=true)2120 isIntegerArrayAttrConfinedToRange(OpType op, ArrayAttr arrayAttr, int64_t min,
2121 int64_t max, StringRef attrName,
2122 bool halfOpen = true) {
2123 for (auto attr : arrayAttr) {
2124 auto val = attr.cast<IntegerAttr>().getInt();
2125 auto upper = max;
2126 if (!halfOpen)
2127 upper += 1;
2128 if (val < min || val >= upper)
2129 return op.emitOpError("expected ") << attrName << " to be confined to ["
2130 << min << ", " << upper << ")";
2131 }
2132 return success();
2133 }
2134
2135 // Returns true if all integers in `arrayAttr` are in the half-open [min, max}
2136 // interval. If `halfOpen` is true then the admissible interval is [min, max).
2137 // Otherwise, the admissible interval is [min, max].
2138 template <typename OpType>
2139 static LogicalResult
isIntegerArrayAttrConfinedToShape(OpType op,ArrayAttr arrayAttr,ArrayRef<int64_t> shape,StringRef attrName,bool halfOpen=true,int64_t min=0)2140 isIntegerArrayAttrConfinedToShape(OpType op, ArrayAttr arrayAttr,
2141 ArrayRef<int64_t> shape, StringRef attrName,
2142 bool halfOpen = true, int64_t min = 0) {
2143 assert(arrayAttr.size() <= shape.size());
2144 unsigned index = 0;
2145 for (auto it : llvm::zip(arrayAttr, shape)) {
2146 auto val = std::get<0>(it).cast<IntegerAttr>().getInt();
2147 auto max = std::get<1>(it);
2148 if (!halfOpen)
2149 max += 1;
2150 if (val < min || val >= max)
2151 return op.emitOpError("expected ")
2152 << attrName << " dimension " << index << " to be confined to ["
2153 << min << ", " << max << ")";
2154 ++index;
2155 }
2156 return success();
2157 }
2158
2159 // Returns true if all integers in `arrayAttr` are in the interval [min, max}.
2160 // interval. If `halfOpen` is true then the admissible interval is [min, max).
2161 // Otherwise, the admissible interval is [min, max].
2162 template <typename OpType>
isSumOfIntegerArrayAttrConfinedToShape(OpType op,ArrayAttr arrayAttr1,ArrayAttr arrayAttr2,ArrayRef<int64_t> shape,StringRef attrName1,StringRef attrName2,bool halfOpen=true,int64_t min=1)2163 static LogicalResult isSumOfIntegerArrayAttrConfinedToShape(
2164 OpType op, ArrayAttr arrayAttr1, ArrayAttr arrayAttr2,
2165 ArrayRef<int64_t> shape, StringRef attrName1, StringRef attrName2,
2166 bool halfOpen = true, int64_t min = 1) {
2167 assert(arrayAttr1.size() <= shape.size());
2168 assert(arrayAttr2.size() <= shape.size());
2169 unsigned index = 0;
2170 for (auto it : llvm::zip(arrayAttr1, arrayAttr2, shape)) {
2171 auto val1 = std::get<0>(it).cast<IntegerAttr>().getInt();
2172 auto val2 = std::get<1>(it).cast<IntegerAttr>().getInt();
2173 auto max = std::get<2>(it);
2174 if (!halfOpen)
2175 max += 1;
2176 if (val1 + val2 < 0 || val1 + val2 >= max)
2177 return op.emitOpError("expected sum(")
2178 << attrName1 << ", " << attrName2 << ") dimension " << index
2179 << " to be confined to [" << min << ", " << max << ")";
2180 ++index;
2181 }
2182 return success();
2183 }
2184
makeI64ArrayAttr(ArrayRef<int64_t> values,MLIRContext * context)2185 static ArrayAttr makeI64ArrayAttr(ArrayRef<int64_t> values,
2186 MLIRContext *context) {
2187 auto attrs = llvm::map_range(values, [context](int64_t v) -> Attribute {
2188 return IntegerAttr::get(IntegerType::get(context, 64), APInt(64, v));
2189 });
2190 return ArrayAttr::get(context, llvm::to_vector<8>(attrs));
2191 }
2192
verify()2193 LogicalResult InsertStridedSliceOp::verify() {
2194 auto sourceVectorType = getSourceVectorType();
2195 auto destVectorType = getDestVectorType();
2196 auto offsets = getOffsetsAttr();
2197 auto strides = getStridesAttr();
2198 if (offsets.size() != static_cast<unsigned>(destVectorType.getRank()))
2199 return emitOpError(
2200 "expected offsets of same size as destination vector rank");
2201 if (strides.size() != static_cast<unsigned>(sourceVectorType.getRank()))
2202 return emitOpError("expected strides of same size as source vector rank");
2203 if (sourceVectorType.getRank() > destVectorType.getRank())
2204 return emitOpError(
2205 "expected source rank to be smaller than destination rank");
2206
2207 auto sourceShape = sourceVectorType.getShape();
2208 auto destShape = destVectorType.getShape();
2209 SmallVector<int64_t, 4> sourceShapeAsDestShape(
2210 destShape.size() - sourceShape.size(), 0);
2211 sourceShapeAsDestShape.append(sourceShape.begin(), sourceShape.end());
2212 auto offName = InsertStridedSliceOp::getOffsetsAttrName();
2213 auto stridesName = InsertStridedSliceOp::getStridesAttrName();
2214 if (failed(isIntegerArrayAttrConfinedToShape(*this, offsets, destShape,
2215 offName)) ||
2216 failed(isIntegerArrayAttrConfinedToRange(*this, strides, 1, 1,
2217 stridesName,
2218 /*halfOpen=*/false)) ||
2219 failed(isSumOfIntegerArrayAttrConfinedToShape(
2220 *this, offsets,
2221 makeI64ArrayAttr(sourceShapeAsDestShape, getContext()), destShape,
2222 offName, "source vector shape",
2223 /*halfOpen=*/false, /*min=*/1)))
2224 return failure();
2225
2226 return success();
2227 }
2228
2229 namespace {
2230 /// Pattern to rewrite an InsertStridedSliceOp(SplatOp(X):src_type,
2231 /// SplatOp(X):dst_type) to SplatOp(X):dst_type.
2232 class FoldInsertStridedSliceSplat final
2233 : public OpRewritePattern<InsertStridedSliceOp> {
2234 public:
2235 using OpRewritePattern<InsertStridedSliceOp>::OpRewritePattern;
2236
matchAndRewrite(InsertStridedSliceOp insertStridedSliceOp,PatternRewriter & rewriter) const2237 LogicalResult matchAndRewrite(InsertStridedSliceOp insertStridedSliceOp,
2238 PatternRewriter &rewriter) const override {
2239 auto srcSplatOp =
2240 insertStridedSliceOp.getSource().getDefiningOp<vector::SplatOp>();
2241 auto destSplatOp =
2242 insertStridedSliceOp.getDest().getDefiningOp<vector::SplatOp>();
2243
2244 if (!srcSplatOp || !destSplatOp)
2245 return failure();
2246
2247 if (srcSplatOp.getInput() != destSplatOp.getInput())
2248 return failure();
2249
2250 rewriter.replaceOp(insertStridedSliceOp, insertStridedSliceOp.getDest());
2251 return success();
2252 }
2253 };
2254
2255 /// Pattern to rewrite an InsertStridedSliceOp(ExtractStridedSliceOp(dst), dst)
2256 /// to dst.
2257 class FoldInsertStridedSliceOfExtract final
2258 : public OpRewritePattern<InsertStridedSliceOp> {
2259 public:
2260 using OpRewritePattern<InsertStridedSliceOp>::OpRewritePattern;
2261
matchAndRewrite(InsertStridedSliceOp insertStridedSliceOp,PatternRewriter & rewriter) const2262 LogicalResult matchAndRewrite(InsertStridedSliceOp insertStridedSliceOp,
2263 PatternRewriter &rewriter) const override {
2264 auto extractStridedSliceOp =
2265 insertStridedSliceOp.getSource()
2266 .getDefiningOp<vector::ExtractStridedSliceOp>();
2267
2268 if (!extractStridedSliceOp)
2269 return failure();
2270
2271 if (extractStridedSliceOp.getOperand() != insertStridedSliceOp.getDest())
2272 return failure();
2273
2274 // Check if have the same strides and offsets.
2275 if (extractStridedSliceOp.getStrides() !=
2276 insertStridedSliceOp.getStrides() ||
2277 extractStridedSliceOp.getOffsets() != insertStridedSliceOp.getOffsets())
2278 return failure();
2279
2280 rewriter.replaceOp(insertStridedSliceOp, insertStridedSliceOp.getDest());
2281 return success();
2282 }
2283 };
2284
2285 } // namespace
2286
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)2287 void vector::InsertStridedSliceOp::getCanonicalizationPatterns(
2288 RewritePatternSet &results, MLIRContext *context) {
2289 results.add<FoldInsertStridedSliceSplat, FoldInsertStridedSliceOfExtract>(
2290 context);
2291 }
2292
fold(ArrayRef<Attribute> operands)2293 OpFoldResult InsertStridedSliceOp::fold(ArrayRef<Attribute> operands) {
2294 if (getSourceVectorType() == getDestVectorType())
2295 return getSource();
2296 return {};
2297 }
2298
2299 //===----------------------------------------------------------------------===//
2300 // OuterProductOp
2301 //===----------------------------------------------------------------------===//
2302
2303 /// Build an op without mask, use the type of `acc` as the return type.
build(OpBuilder & builder,OperationState & result,Value lhs,Value rhs,Value acc)2304 void OuterProductOp::build(OpBuilder &builder, OperationState &result,
2305 Value lhs, Value rhs, Value acc) {
2306 result.addOperands({lhs, rhs, acc});
2307 result.addTypes(acc.getType());
2308 }
2309
print(OpAsmPrinter & p)2310 void OuterProductOp::print(OpAsmPrinter &p) {
2311 p << " " << getLhs() << ", " << getRhs();
2312 if (!getAcc().empty()) {
2313 p << ", " << getAcc();
2314 p.printOptionalAttrDict((*this)->getAttrs());
2315 }
2316 p << " : " << getLhs().getType() << ", " << getRhs().getType();
2317 }
2318
parse(OpAsmParser & parser,OperationState & result)2319 ParseResult OuterProductOp::parse(OpAsmParser &parser, OperationState &result) {
2320 SmallVector<OpAsmParser::UnresolvedOperand, 3> operandsInfo;
2321 Type tLHS, tRHS;
2322 if (parser.parseOperandList(operandsInfo) ||
2323 parser.parseOptionalAttrDict(result.attributes) ||
2324 parser.parseColonType(tLHS) || parser.parseComma() ||
2325 parser.parseType(tRHS))
2326 return failure();
2327 if (operandsInfo.size() < 2)
2328 return parser.emitError(parser.getNameLoc(),
2329 "expected at least 2 operands");
2330 VectorType vLHS = tLHS.dyn_cast<VectorType>();
2331 VectorType vRHS = tRHS.dyn_cast<VectorType>();
2332 if (!vLHS)
2333 return parser.emitError(parser.getNameLoc(),
2334 "expected vector type for operand #1");
2335 VectorType resType =
2336 vRHS ? VectorType::get({vLHS.getDimSize(0), vRHS.getDimSize(0)},
2337 vLHS.getElementType())
2338 : VectorType::get({vLHS.getDimSize(0)}, vLHS.getElementType());
2339
2340 if (!result.attributes.get(OuterProductOp::getKindAttrStrName())) {
2341 result.attributes.append(
2342 OuterProductOp::getKindAttrStrName(),
2343 CombiningKindAttr::get(OuterProductOp::getDefaultKind(),
2344 result.getContext()));
2345 }
2346
2347 return failure(
2348 parser.resolveOperand(operandsInfo[0], tLHS, result.operands) ||
2349 parser.resolveOperand(operandsInfo[1], tRHS, result.operands) ||
2350 (operandsInfo.size() > 2 &&
2351 parser.resolveOperand(operandsInfo[2], resType, result.operands)) ||
2352 parser.addTypeToList(resType, result.types));
2353 }
2354
verify()2355 LogicalResult OuterProductOp::verify() {
2356 Type tRHS = getOperandTypeRHS();
2357 VectorType vLHS = getOperandVectorTypeLHS(),
2358 vRHS = tRHS.dyn_cast<VectorType>(),
2359 vACC = getOperandVectorTypeACC(), vRES = getVectorType();
2360
2361 if (vLHS.getRank() != 1)
2362 return emitOpError("expected 1-d vector for operand #1");
2363
2364 if (vRHS) {
2365 // Proper OUTER operation.
2366 if (vRHS.getRank() != 1)
2367 return emitOpError("expected 1-d vector for operand #2");
2368 if (vRES.getRank() != 2)
2369 return emitOpError("expected 2-d vector result");
2370 if (vLHS.getDimSize(0) != vRES.getDimSize(0))
2371 return emitOpError("expected #1 operand dim to match result dim #1");
2372 if (vRHS.getDimSize(0) != vRES.getDimSize(1))
2373 return emitOpError("expected #2 operand dim to match result dim #2");
2374 } else {
2375 // An AXPY operation.
2376 if (vRES.getRank() != 1)
2377 return emitOpError("expected 1-d vector result");
2378 if (vLHS.getDimSize(0) != vRES.getDimSize(0))
2379 return emitOpError("expected #1 operand dim to match result dim #1");
2380 }
2381
2382 if (vACC && vACC != vRES)
2383 return emitOpError("expected operand #3 of same type as result type");
2384
2385 // Verify supported combining kind.
2386 if (!isSupportedCombiningKind(getKind(), vRES.getElementType()))
2387 return emitOpError("unsupported outerproduct type");
2388
2389 return success();
2390 }
2391
2392 //===----------------------------------------------------------------------===//
2393 // ReshapeOp
2394 //===----------------------------------------------------------------------===//
2395
verify()2396 LogicalResult ReshapeOp::verify() {
2397 // Verify that rank(numInputs/outputs) + numFixedVec dim matches vec rank.
2398 auto inputVectorType = getInputVectorType();
2399 auto outputVectorType = getOutputVectorType();
2400 int64_t inputShapeRank = getNumInputShapeSizes();
2401 int64_t outputShapeRank = getNumOutputShapeSizes();
2402 SmallVector<int64_t, 4> fixedVectorSizes;
2403 getFixedVectorSizes(fixedVectorSizes);
2404 int64_t numFixedVectorSizes = fixedVectorSizes.size();
2405
2406 if (inputVectorType.getRank() != inputShapeRank + numFixedVectorSizes)
2407 return emitError("invalid input shape for vector type ") << inputVectorType;
2408
2409 if (outputVectorType.getRank() != outputShapeRank + numFixedVectorSizes)
2410 return emitError("invalid output shape for vector type ")
2411 << outputVectorType;
2412
2413 // Verify that the 'fixedVectorSizes' match an input/output vector shape
2414 // suffix.
2415 unsigned inputVectorRank = inputVectorType.getRank();
2416 for (unsigned i = 0; i < numFixedVectorSizes; ++i) {
2417 unsigned index = inputVectorRank - numFixedVectorSizes - i;
2418 if (fixedVectorSizes[i] != inputVectorType.getShape()[index])
2419 return emitError("fixed vector size must match input vector for dim ")
2420 << i;
2421 }
2422
2423 unsigned outputVectorRank = outputVectorType.getRank();
2424 for (unsigned i = 0; i < numFixedVectorSizes; ++i) {
2425 unsigned index = outputVectorRank - numFixedVectorSizes - i;
2426 if (fixedVectorSizes[i] != outputVectorType.getShape()[index])
2427 return emitError("fixed vector size must match output vector for dim ")
2428 << i;
2429 }
2430
2431 // If all shape operands are produced by constant ops, verify that product
2432 // of dimensions for input/output shape match.
2433 auto isDefByConstant = [](Value operand) {
2434 return isa_and_nonnull<arith::ConstantIndexOp>(operand.getDefiningOp());
2435 };
2436 if (llvm::all_of(getInputShape(), isDefByConstant) &&
2437 llvm::all_of(getOutputShape(), isDefByConstant)) {
2438 int64_t numInputElements = 1;
2439 for (auto operand : getInputShape())
2440 numInputElements *=
2441 cast<arith::ConstantIndexOp>(operand.getDefiningOp()).value();
2442 int64_t numOutputElements = 1;
2443 for (auto operand : getOutputShape())
2444 numOutputElements *=
2445 cast<arith::ConstantIndexOp>(operand.getDefiningOp()).value();
2446 if (numInputElements != numOutputElements)
2447 return emitError("product of input and output shape sizes must match");
2448 }
2449 return success();
2450 }
2451
getFixedVectorSizes(SmallVectorImpl<int64_t> & results)2452 void ReshapeOp::getFixedVectorSizes(SmallVectorImpl<int64_t> &results) {
2453 populateFromInt64AttrArray(getFixedVectorSizes(), results);
2454 }
2455
2456 //===----------------------------------------------------------------------===//
2457 // ExtractStridedSliceOp
2458 //===----------------------------------------------------------------------===//
2459
2460 // Inference works as follows:
2461 // 1. Add 'sizes' from prefix of dims in 'offsets'.
2462 // 2. Add sizes from 'vectorType' for remaining dims.
inferStridedSliceOpResultType(VectorType vectorType,ArrayAttr offsets,ArrayAttr sizes,ArrayAttr strides)2463 static Type inferStridedSliceOpResultType(VectorType vectorType,
2464 ArrayAttr offsets, ArrayAttr sizes,
2465 ArrayAttr strides) {
2466 assert(offsets.size() == sizes.size() && offsets.size() == strides.size());
2467 SmallVector<int64_t, 4> shape;
2468 shape.reserve(vectorType.getRank());
2469 unsigned idx = 0;
2470 for (unsigned e = offsets.size(); idx < e; ++idx)
2471 shape.push_back(sizes[idx].cast<IntegerAttr>().getInt());
2472 for (unsigned e = vectorType.getShape().size(); idx < e; ++idx)
2473 shape.push_back(vectorType.getShape()[idx]);
2474
2475 return VectorType::get(shape, vectorType.getElementType());
2476 }
2477
build(OpBuilder & builder,OperationState & result,Value source,ArrayRef<int64_t> offsets,ArrayRef<int64_t> sizes,ArrayRef<int64_t> strides)2478 void ExtractStridedSliceOp::build(OpBuilder &builder, OperationState &result,
2479 Value source, ArrayRef<int64_t> offsets,
2480 ArrayRef<int64_t> sizes,
2481 ArrayRef<int64_t> strides) {
2482 result.addOperands(source);
2483 auto offsetsAttr = getVectorSubscriptAttr(builder, offsets);
2484 auto sizesAttr = getVectorSubscriptAttr(builder, sizes);
2485 auto stridesAttr = getVectorSubscriptAttr(builder, strides);
2486 result.addTypes(
2487 inferStridedSliceOpResultType(source.getType().cast<VectorType>(),
2488 offsetsAttr, sizesAttr, stridesAttr));
2489 result.addAttribute(getOffsetsAttrStrName(), offsetsAttr);
2490 result.addAttribute(getSizesAttrStrName(), sizesAttr);
2491 result.addAttribute(getStridesAttrStrName(), stridesAttr);
2492 }
2493
verify()2494 LogicalResult ExtractStridedSliceOp::verify() {
2495 auto type = getVectorType();
2496 auto offsets = getOffsetsAttr();
2497 auto sizes = getSizesAttr();
2498 auto strides = getStridesAttr();
2499 if (offsets.size() != sizes.size() || offsets.size() != strides.size())
2500 return emitOpError(
2501 "expected offsets, sizes and strides attributes of same size");
2502
2503 auto shape = type.getShape();
2504 auto offName = getOffsetsAttrName();
2505 auto sizesName = getSizesAttrName();
2506 auto stridesName = getStridesAttrName();
2507 if (failed(
2508 isIntegerArrayAttrSmallerThanShape(*this, offsets, shape, offName)) ||
2509 failed(
2510 isIntegerArrayAttrSmallerThanShape(*this, sizes, shape, sizesName)) ||
2511 failed(isIntegerArrayAttrSmallerThanShape(*this, strides, shape,
2512 stridesName)) ||
2513 failed(
2514 isIntegerArrayAttrConfinedToShape(*this, offsets, shape, offName)) ||
2515 failed(isIntegerArrayAttrConfinedToShape(*this, sizes, shape, sizesName,
2516 /*halfOpen=*/false,
2517 /*min=*/1)) ||
2518 failed(isIntegerArrayAttrConfinedToRange(*this, strides, 1, 1,
2519 stridesName,
2520 /*halfOpen=*/false)) ||
2521 failed(isSumOfIntegerArrayAttrConfinedToShape(*this, offsets, sizes,
2522 shape, offName, sizesName,
2523 /*halfOpen=*/false)))
2524 return failure();
2525
2526 auto resultType =
2527 inferStridedSliceOpResultType(getVectorType(), offsets, sizes, strides);
2528 if (getResult().getType() != resultType)
2529 return emitOpError("expected result type to be ") << resultType;
2530
2531 return success();
2532 }
2533
2534 // When the source of ExtractStrided comes from a chain of InsertStrided ops try
2535 // to use the source of the InsertStrided ops if we can detect that the
2536 // extracted vector is a subset of one of the vector inserted.
2537 static LogicalResult
foldExtractStridedOpFromInsertChain(ExtractStridedSliceOp op)2538 foldExtractStridedOpFromInsertChain(ExtractStridedSliceOp op) {
2539 // Helper to extract integer out of ArrayAttr.
2540 auto getElement = [](ArrayAttr array, int idx) {
2541 return array[idx].cast<IntegerAttr>().getInt();
2542 };
2543 ArrayAttr extractOffsets = op.getOffsets();
2544 ArrayAttr extractStrides = op.getStrides();
2545 ArrayAttr extractSizes = op.getSizes();
2546 auto insertOp = op.getVector().getDefiningOp<InsertStridedSliceOp>();
2547 while (insertOp) {
2548 if (op.getVectorType().getRank() !=
2549 insertOp.getSourceVectorType().getRank())
2550 return failure();
2551 ArrayAttr insertOffsets = insertOp.getOffsets();
2552 ArrayAttr insertStrides = insertOp.getStrides();
2553 // If the rank of extract is greater than the rank of insert, we are likely
2554 // extracting a partial chunk of the vector inserted.
2555 if (extractOffsets.size() > insertOffsets.size())
2556 return failure();
2557 bool patialoverlap = false;
2558 bool disjoint = false;
2559 SmallVector<int64_t, 4> offsetDiffs;
2560 for (unsigned dim = 0, e = extractOffsets.size(); dim < e; ++dim) {
2561 if (getElement(extractStrides, dim) != getElement(insertStrides, dim))
2562 return failure();
2563 int64_t start = getElement(insertOffsets, dim);
2564 int64_t end = start + insertOp.getSourceVectorType().getDimSize(dim);
2565 int64_t offset = getElement(extractOffsets, dim);
2566 int64_t size = getElement(extractSizes, dim);
2567 // Check if the start of the extract offset is in the interval inserted.
2568 if (start <= offset && offset < end) {
2569 // If the extract interval overlaps but is not fully included we may
2570 // have a partial overlap that will prevent any folding.
2571 if (offset + size > end)
2572 patialoverlap = true;
2573 offsetDiffs.push_back(offset - start);
2574 continue;
2575 }
2576 disjoint = true;
2577 break;
2578 }
2579 // The extract element chunk is a subset of the insert element.
2580 if (!disjoint && !patialoverlap) {
2581 op.setOperand(insertOp.getSource());
2582 // OpBuilder is only used as a helper to build an I64ArrayAttr.
2583 OpBuilder b(op.getContext());
2584 op->setAttr(ExtractStridedSliceOp::getOffsetsAttrStrName(),
2585 b.getI64ArrayAttr(offsetDiffs));
2586 return success();
2587 }
2588 // If the chunk extracted is disjoint from the chunk inserted, keep looking
2589 // in the insert chain.
2590 if (disjoint)
2591 insertOp = insertOp.getDest().getDefiningOp<InsertStridedSliceOp>();
2592 else {
2593 // The extracted vector partially overlap the inserted vector, we cannot
2594 // fold.
2595 return failure();
2596 }
2597 }
2598 return failure();
2599 }
2600
fold(ArrayRef<Attribute> operands)2601 OpFoldResult ExtractStridedSliceOp::fold(ArrayRef<Attribute> operands) {
2602 if (getVectorType() == getResult().getType())
2603 return getVector();
2604 if (succeeded(foldExtractStridedOpFromInsertChain(*this)))
2605 return getResult();
2606 return {};
2607 }
2608
getOffsets(SmallVectorImpl<int64_t> & results)2609 void ExtractStridedSliceOp::getOffsets(SmallVectorImpl<int64_t> &results) {
2610 populateFromInt64AttrArray(getOffsets(), results);
2611 }
2612
2613 namespace {
2614
2615 // Pattern to rewrite an ExtractStridedSliceOp(ConstantMaskOp) to
2616 // ConstantMaskOp.
2617 class StridedSliceConstantMaskFolder final
2618 : public OpRewritePattern<ExtractStridedSliceOp> {
2619 public:
2620 using OpRewritePattern<ExtractStridedSliceOp>::OpRewritePattern;
2621
matchAndRewrite(ExtractStridedSliceOp extractStridedSliceOp,PatternRewriter & rewriter) const2622 LogicalResult matchAndRewrite(ExtractStridedSliceOp extractStridedSliceOp,
2623 PatternRewriter &rewriter) const override {
2624 // Return if 'extractStridedSliceOp' operand is not defined by a
2625 // ConstantMaskOp.
2626 auto *defOp = extractStridedSliceOp.getVector().getDefiningOp();
2627 auto constantMaskOp = dyn_cast_or_null<ConstantMaskOp>(defOp);
2628 if (!constantMaskOp)
2629 return failure();
2630 // Return if 'extractStridedSliceOp' has non-unit strides.
2631 if (extractStridedSliceOp.hasNonUnitStrides())
2632 return failure();
2633 // Gather constant mask dimension sizes.
2634 SmallVector<int64_t, 4> maskDimSizes;
2635 populateFromInt64AttrArray(constantMaskOp.getMaskDimSizes(), maskDimSizes);
2636 // Gather strided slice offsets and sizes.
2637 SmallVector<int64_t, 4> sliceOffsets;
2638 populateFromInt64AttrArray(extractStridedSliceOp.getOffsets(),
2639 sliceOffsets);
2640 SmallVector<int64_t, 4> sliceSizes;
2641 populateFromInt64AttrArray(extractStridedSliceOp.getSizes(), sliceSizes);
2642
2643 // Compute slice of vector mask region.
2644 SmallVector<int64_t, 4> sliceMaskDimSizes;
2645 assert(sliceOffsets.size() == maskDimSizes.size());
2646 for (auto it : llvm::zip(maskDimSizes, sliceOffsets, sliceSizes)) {
2647 int64_t maskDimSize = std::get<0>(it);
2648 int64_t sliceOffset = std::get<1>(it);
2649 int64_t sliceSize = std::get<2>(it);
2650 int64_t sliceMaskDimSize = std::max(
2651 static_cast<int64_t>(0),
2652 std::min(sliceOffset + sliceSize, maskDimSize) - sliceOffset);
2653 sliceMaskDimSizes.push_back(sliceMaskDimSize);
2654 }
2655 // If any of 'sliceMaskDimSizes' are zero, then set all to zero (masked
2656 // region is a conjunction of mask dim intervals).
2657 if (llvm::is_contained(sliceMaskDimSizes, 0))
2658 sliceMaskDimSizes.assign(maskDimSizes.size(), 0);
2659
2660 // Replace 'extractStridedSliceOp' with ConstantMaskOp with sliced mask
2661 // region.
2662 rewriter.replaceOpWithNewOp<ConstantMaskOp>(
2663 extractStridedSliceOp, extractStridedSliceOp.getResult().getType(),
2664 vector::getVectorSubscriptAttr(rewriter, sliceMaskDimSizes));
2665 return success();
2666 }
2667 };
2668
2669 // Pattern to rewrite a ExtractStridedSliceOp(splat ConstantOp) -> ConstantOp.
2670 class StridedSliceConstantFolder final
2671 : public OpRewritePattern<ExtractStridedSliceOp> {
2672 public:
2673 using OpRewritePattern<ExtractStridedSliceOp>::OpRewritePattern;
2674
matchAndRewrite(ExtractStridedSliceOp extractStridedSliceOp,PatternRewriter & rewriter) const2675 LogicalResult matchAndRewrite(ExtractStridedSliceOp extractStridedSliceOp,
2676 PatternRewriter &rewriter) const override {
2677 // Return if 'extractStridedSliceOp' operand is not defined by a
2678 // ConstantOp.
2679 auto constantOp =
2680 extractStridedSliceOp.getVector().getDefiningOp<arith::ConstantOp>();
2681 if (!constantOp)
2682 return failure();
2683 auto dense = constantOp.getValue().dyn_cast<SplatElementsAttr>();
2684 if (!dense)
2685 return failure();
2686 auto newAttr = DenseElementsAttr::get(extractStridedSliceOp.getType(),
2687 dense.getSplatValue<Attribute>());
2688 rewriter.replaceOpWithNewOp<arith::ConstantOp>(extractStridedSliceOp,
2689 newAttr);
2690 return success();
2691 }
2692 };
2693
2694 // Pattern to rewrite an ExtractStridedSliceOp(BroadcastOp) to
2695 // BroadcastOp(ExtractStrideSliceOp).
2696 class StridedSliceBroadcast final
2697 : public OpRewritePattern<ExtractStridedSliceOp> {
2698 public:
2699 using OpRewritePattern<ExtractStridedSliceOp>::OpRewritePattern;
2700
matchAndRewrite(ExtractStridedSliceOp op,PatternRewriter & rewriter) const2701 LogicalResult matchAndRewrite(ExtractStridedSliceOp op,
2702 PatternRewriter &rewriter) const override {
2703 auto broadcast = op.getVector().getDefiningOp<BroadcastOp>();
2704 if (!broadcast)
2705 return failure();
2706 auto srcVecType = broadcast.getSource().getType().dyn_cast<VectorType>();
2707 unsigned srcRank = srcVecType ? srcVecType.getRank() : 0;
2708 auto dstVecType = op.getType().cast<VectorType>();
2709 unsigned dstRank = dstVecType.getRank();
2710 unsigned rankDiff = dstRank - srcRank;
2711 // Check if the most inner dimensions of the source of the broadcast are the
2712 // same as the destination of the extract. If this is the case we can just
2713 // use a broadcast as the original dimensions are untouched.
2714 bool lowerDimMatch = true;
2715 for (unsigned i = 0; i < srcRank; i++) {
2716 if (srcVecType.getDimSize(i) != dstVecType.getDimSize(i + rankDiff)) {
2717 lowerDimMatch = false;
2718 break;
2719 }
2720 }
2721 Value source = broadcast.getSource();
2722 // If the inner dimensions don't match, it means we need to extract from the
2723 // source of the orignal broadcast and then broadcast the extracted value.
2724 // We also need to handle degenerated cases where the source is effectively
2725 // just a single scalar.
2726 bool isScalarSrc = (srcRank == 0 || srcVecType.getNumElements() == 1);
2727 if (!lowerDimMatch && !isScalarSrc) {
2728 source = rewriter.create<ExtractStridedSliceOp>(
2729 op->getLoc(), source,
2730 getI64SubArray(op.getOffsets(), /* dropFront=*/rankDiff),
2731 getI64SubArray(op.getSizes(), /* dropFront=*/rankDiff),
2732 getI64SubArray(op.getStrides(), /* dropFront=*/rankDiff));
2733 }
2734 rewriter.replaceOpWithNewOp<BroadcastOp>(op, op.getType(), source);
2735 return success();
2736 }
2737 };
2738
2739 /// Pattern to rewrite an ExtractStridedSliceOp(SplatOp) to SplatOp.
2740 class StridedSliceSplat final : public OpRewritePattern<ExtractStridedSliceOp> {
2741 public:
2742 using OpRewritePattern<ExtractStridedSliceOp>::OpRewritePattern;
2743
matchAndRewrite(ExtractStridedSliceOp op,PatternRewriter & rewriter) const2744 LogicalResult matchAndRewrite(ExtractStridedSliceOp op,
2745 PatternRewriter &rewriter) const override {
2746 auto splat = op.getVector().getDefiningOp<SplatOp>();
2747 if (!splat)
2748 return failure();
2749 rewriter.replaceOpWithNewOp<SplatOp>(op, op.getType(), splat.getInput());
2750 return success();
2751 }
2752 };
2753
2754 } // namespace
2755
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)2756 void ExtractStridedSliceOp::getCanonicalizationPatterns(
2757 RewritePatternSet &results, MLIRContext *context) {
2758 // Pattern to rewrite a ExtractStridedSliceOp(ConstantMaskOp) ->
2759 // ConstantMaskOp and ExtractStridedSliceOp(ConstantOp) -> ConstantOp.
2760 results.add<StridedSliceConstantMaskFolder, StridedSliceConstantFolder,
2761 StridedSliceBroadcast, StridedSliceSplat>(context);
2762 }
2763
2764 //===----------------------------------------------------------------------===//
2765 // TransferReadOp
2766 //===----------------------------------------------------------------------===//
2767
2768 /// 1. Builder that sets padding to zero and an empty mask (variant with attrs).
build(OpBuilder & builder,OperationState & result,VectorType vectorType,Value source,ValueRange indices,AffineMapAttr permutationMapAttr,ArrayAttr inBoundsAttr)2769 void TransferReadOp::build(OpBuilder &builder, OperationState &result,
2770 VectorType vectorType, Value source,
2771 ValueRange indices, AffineMapAttr permutationMapAttr,
2772 /*optional*/ ArrayAttr inBoundsAttr) {
2773 Type elemType = source.getType().cast<ShapedType>().getElementType();
2774 Value padding = builder.create<arith::ConstantOp>(
2775 result.location, elemType, builder.getZeroAttr(elemType));
2776 build(builder, result, vectorType, source, indices, permutationMapAttr,
2777 padding, /*mask=*/Value(), inBoundsAttr);
2778 }
2779
2780 /// 2. Builder that sets padding to zero an empty mask (variant without attrs).
build(OpBuilder & builder,OperationState & result,VectorType vectorType,Value source,ValueRange indices,AffineMap permutationMap,Optional<ArrayRef<bool>> inBounds)2781 void TransferReadOp::build(OpBuilder &builder, OperationState &result,
2782 VectorType vectorType, Value source,
2783 ValueRange indices, AffineMap permutationMap,
2784 Optional<ArrayRef<bool>> inBounds) {
2785 auto permutationMapAttr = AffineMapAttr::get(permutationMap);
2786 auto inBoundsAttr = (inBounds && !inBounds.value().empty())
2787 ? builder.getBoolArrayAttr(inBounds.value())
2788 : ArrayAttr();
2789 build(builder, result, vectorType, source, indices, permutationMapAttr,
2790 inBoundsAttr);
2791 }
2792
2793 /// 3. Builder that sets permutation map to 'getMinorIdentityMap'.
build(OpBuilder & builder,OperationState & result,VectorType vectorType,Value source,ValueRange indices,Value padding,Optional<ArrayRef<bool>> inBounds)2794 void TransferReadOp::build(OpBuilder &builder, OperationState &result,
2795 VectorType vectorType, Value source,
2796 ValueRange indices, Value padding,
2797 Optional<ArrayRef<bool>> inBounds) {
2798 AffineMap permutationMap = getTransferMinorIdentityMap(
2799 source.getType().cast<ShapedType>(), vectorType);
2800 auto permutationMapAttr = AffineMapAttr::get(permutationMap);
2801 auto inBoundsAttr = (inBounds && !inBounds.value().empty())
2802 ? builder.getBoolArrayAttr(inBounds.value())
2803 : ArrayAttr();
2804 build(builder, result, vectorType, source, indices, permutationMapAttr,
2805 padding,
2806 /*mask=*/Value(), inBoundsAttr);
2807 }
2808
2809 /// 4. Builder that sets padding to zero and permutation map to
2810 /// 'getMinorIdentityMap'.
build(OpBuilder & builder,OperationState & result,VectorType vectorType,Value source,ValueRange indices,Optional<ArrayRef<bool>> inBounds)2811 void TransferReadOp::build(OpBuilder &builder, OperationState &result,
2812 VectorType vectorType, Value source,
2813 ValueRange indices,
2814 Optional<ArrayRef<bool>> inBounds) {
2815 Type elemType = source.getType().cast<ShapedType>().getElementType();
2816 Value padding = builder.create<arith::ConstantOp>(
2817 result.location, elemType, builder.getZeroAttr(elemType));
2818 build(builder, result, vectorType, source, indices, padding, inBounds);
2819 }
2820
2821 template <typename EmitFun>
verifyPermutationMap(AffineMap permutationMap,EmitFun emitOpError)2822 static LogicalResult verifyPermutationMap(AffineMap permutationMap,
2823 EmitFun emitOpError) {
2824 SmallVector<bool, 8> seen(permutationMap.getNumInputs(), false);
2825 for (auto expr : permutationMap.getResults()) {
2826 auto dim = expr.dyn_cast<AffineDimExpr>();
2827 auto zero = expr.dyn_cast<AffineConstantExpr>();
2828 if (zero) {
2829 if (zero.getValue() != 0) {
2830 return emitOpError(
2831 "requires a projected permutation_map (at most one dim or the zero "
2832 "constant can appear in each result)");
2833 }
2834 continue;
2835 }
2836 if (!dim) {
2837 return emitOpError("requires a projected permutation_map (at most one "
2838 "dim or the zero constant can appear in each result)");
2839 }
2840 if (seen[dim.getPosition()]) {
2841 return emitOpError(
2842 "requires a permutation_map that is a permutation (found one dim "
2843 "used more than once)");
2844 }
2845 seen[dim.getPosition()] = true;
2846 }
2847 return success();
2848 }
2849
2850 static LogicalResult
verifyTransferOp(VectorTransferOpInterface op,ShapedType shapedType,VectorType vectorType,VectorType maskType,AffineMap permutationMap,ArrayAttr inBounds)2851 verifyTransferOp(VectorTransferOpInterface op, ShapedType shapedType,
2852 VectorType vectorType, VectorType maskType,
2853 AffineMap permutationMap, ArrayAttr inBounds) {
2854 if (op->hasAttr("masked")) {
2855 return op->emitOpError("masked attribute has been removed. "
2856 "Use in_bounds instead.");
2857 }
2858
2859 if (!shapedType.isa<MemRefType, RankedTensorType>())
2860 return op->emitOpError(
2861 "requires source to be a memref or ranked tensor type");
2862
2863 auto elementType = shapedType.getElementType();
2864 DataLayout dataLayout = DataLayout::closest(op);
2865 if (auto vectorElementType = elementType.dyn_cast<VectorType>()) {
2866 // Memref or tensor has vector element type.
2867 unsigned sourceVecSize =
2868 dataLayout.getTypeSizeInBits(vectorElementType.getElementType()) *
2869 vectorElementType.getShape().back();
2870 unsigned resultVecSize =
2871 dataLayout.getTypeSizeInBits(vectorType.getElementType()) *
2872 vectorType.getShape().back();
2873 if (resultVecSize % sourceVecSize != 0)
2874 return op->emitOpError(
2875 "requires the bitwidth of the minor 1-D vector to be an integral "
2876 "multiple of the bitwidth of the minor 1-D vector of the source");
2877
2878 unsigned sourceVecEltRank = vectorElementType.getRank();
2879 unsigned resultVecRank = vectorType.getRank();
2880 if (sourceVecEltRank > resultVecRank)
2881 return op->emitOpError(
2882 "requires source vector element and vector result ranks to match.");
2883 unsigned rankOffset = resultVecRank - sourceVecEltRank;
2884 // Check that permutation map results match 'rankOffset' of vector type.
2885 if (permutationMap.getNumResults() != rankOffset)
2886 return op->emitOpError("requires a permutation_map with result dims of "
2887 "the same rank as the vector type");
2888
2889 if (maskType)
2890 return op->emitOpError("does not support masks with vector element type");
2891 } else {
2892 // Memref or tensor has scalar element type.
2893 unsigned minorSize =
2894 vectorType.getRank() == 0 ? 1 : vectorType.getShape().back();
2895 unsigned resultVecSize =
2896 dataLayout.getTypeSizeInBits(vectorType.getElementType()) * minorSize;
2897 if (resultVecSize % dataLayout.getTypeSizeInBits(elementType) != 0)
2898 return op->emitOpError(
2899 "requires the bitwidth of the minor 1-D vector to be an integral "
2900 "multiple of the bitwidth of the source element type");
2901
2902 // Check that permutation map results match rank of vector type.
2903 if (permutationMap.getNumResults() != vectorType.getRank())
2904 return op->emitOpError("requires a permutation_map with result dims of "
2905 "the same rank as the vector type");
2906
2907 VectorType expectedMaskType =
2908 vector::detail::transferMaskType(vectorType, permutationMap);
2909 if (maskType && expectedMaskType != maskType)
2910 return op->emitOpError("expects mask type consistent with permutation "
2911 "map: ")
2912 << maskType;
2913 }
2914
2915 if (permutationMap.getNumSymbols() != 0)
2916 return op->emitOpError("requires permutation_map without symbols");
2917
2918 if (permutationMap.getNumInputs() != shapedType.getRank())
2919 return op->emitOpError("requires a permutation_map with input dims of the "
2920 "same rank as the source type");
2921
2922 if (inBounds) {
2923 if (permutationMap.getNumResults() != static_cast<int64_t>(inBounds.size()))
2924 return op->emitOpError("expects the optional in_bounds attr of same rank "
2925 "as permutation_map results: ")
2926 << AffineMapAttr::get(permutationMap)
2927 << " vs inBounds of size: " << inBounds.size();
2928 for (unsigned int i = 0; i < permutationMap.getNumResults(); ++i)
2929 if (permutationMap.getResult(i).isa<AffineConstantExpr>() &&
2930 !inBounds.getValue()[i].cast<BoolAttr>().getValue())
2931 return op->emitOpError("requires broadcast dimensions to be in-bounds");
2932 }
2933
2934 return success();
2935 }
2936
printTransferAttrs(OpAsmPrinter & p,VectorTransferOpInterface op)2937 static void printTransferAttrs(OpAsmPrinter &p, VectorTransferOpInterface op) {
2938 SmallVector<StringRef, 3> elidedAttrs;
2939 elidedAttrs.push_back(TransferReadOp::getOperandSegmentSizeAttr());
2940 if (op.permutation_map().isMinorIdentity())
2941 elidedAttrs.push_back(op.getPermutationMapAttrStrName());
2942 bool elideInBounds = true;
2943 if (auto inBounds = op.in_bounds()) {
2944 for (auto attr : *inBounds) {
2945 if (attr.template cast<BoolAttr>().getValue()) {
2946 elideInBounds = false;
2947 break;
2948 }
2949 }
2950 }
2951 if (elideInBounds)
2952 elidedAttrs.push_back(op.getInBoundsAttrStrName());
2953 p.printOptionalAttrDict(op->getAttrs(), elidedAttrs);
2954 }
2955
print(OpAsmPrinter & p)2956 void TransferReadOp::print(OpAsmPrinter &p) {
2957 p << " " << getSource() << "[" << getIndices() << "], " << getPadding();
2958 if (getMask())
2959 p << ", " << getMask();
2960 printTransferAttrs(p, *this);
2961 p << " : " << getShapedType() << ", " << getVectorType();
2962 }
2963
parse(OpAsmParser & parser,OperationState & result)2964 ParseResult TransferReadOp::parse(OpAsmParser &parser, OperationState &result) {
2965 auto &builder = parser.getBuilder();
2966 SMLoc typesLoc;
2967 OpAsmParser::UnresolvedOperand sourceInfo;
2968 SmallVector<OpAsmParser::UnresolvedOperand, 8> indexInfo;
2969 OpAsmParser::UnresolvedOperand paddingInfo;
2970 SmallVector<Type, 2> types;
2971 OpAsmParser::UnresolvedOperand maskInfo;
2972 // Parsing with support for paddingValue.
2973 if (parser.parseOperand(sourceInfo) ||
2974 parser.parseOperandList(indexInfo, OpAsmParser::Delimiter::Square) ||
2975 parser.parseComma() || parser.parseOperand(paddingInfo))
2976 return failure();
2977 ParseResult hasMask = parser.parseOptionalComma();
2978 if (hasMask.succeeded()) {
2979 if (parser.parseOperand(maskInfo))
2980 return failure();
2981 }
2982 if (parser.parseOptionalAttrDict(result.attributes) ||
2983 parser.getCurrentLocation(&typesLoc) || parser.parseColonTypeList(types))
2984 return failure();
2985 if (types.size() != 2)
2986 return parser.emitError(typesLoc, "requires two types");
2987 auto indexType = builder.getIndexType();
2988 auto shapedType = types[0].dyn_cast<ShapedType>();
2989 if (!shapedType || !shapedType.isa<MemRefType, RankedTensorType>())
2990 return parser.emitError(typesLoc, "requires memref or ranked tensor type");
2991 VectorType vectorType = types[1].dyn_cast<VectorType>();
2992 if (!vectorType)
2993 return parser.emitError(typesLoc, "requires vector type");
2994 auto permutationAttrName = TransferReadOp::getPermutationMapAttrStrName();
2995 Attribute mapAttr = result.attributes.get(permutationAttrName);
2996 if (!mapAttr) {
2997 auto permMap = getTransferMinorIdentityMap(shapedType, vectorType);
2998 // Update `mapAttr` that is used later to determine mask type.
2999 mapAttr = AffineMapAttr::get(permMap);
3000 result.attributes.set(permutationAttrName, mapAttr);
3001 }
3002 if (parser.resolveOperand(sourceInfo, shapedType, result.operands) ||
3003 parser.resolveOperands(indexInfo, indexType, result.operands) ||
3004 parser.resolveOperand(paddingInfo, shapedType.getElementType(),
3005 result.operands))
3006 return failure();
3007 if (hasMask.succeeded()) {
3008 if (shapedType.getElementType().dyn_cast<VectorType>())
3009 return parser.emitError(
3010 maskInfo.location, "does not support masks with vector element type");
3011 auto map = mapAttr.dyn_cast<AffineMapAttr>().getValue();
3012 // Instead of adding the mask type as an op type, compute it based on the
3013 // vector type and the permutation map (to keep the type signature small).
3014 auto maskType = mlir::vector::detail::transferMaskType(vectorType, map);
3015 if (parser.resolveOperand(maskInfo, maskType, result.operands))
3016 return failure();
3017 }
3018 result.addAttribute(
3019 TransferReadOp::getOperandSegmentSizeAttr(),
3020 builder.getI32VectorAttr({1, static_cast<int32_t>(indexInfo.size()), 1,
3021 static_cast<int32_t>(hasMask.succeeded())}));
3022 return parser.addTypeToList(vectorType, result.types);
3023 }
3024
verify()3025 LogicalResult TransferReadOp::verify() {
3026 // Consistency of elemental types in source and vector.
3027 ShapedType shapedType = getShapedType();
3028 VectorType vectorType = getVectorType();
3029 VectorType maskType = getMaskType();
3030 auto paddingType = getPadding().getType();
3031 auto permutationMap = getPermutationMap();
3032 auto sourceElementType = shapedType.getElementType();
3033
3034 if (static_cast<int64_t>(getIndices().size()) != shapedType.getRank())
3035 return emitOpError("requires ") << shapedType.getRank() << " indices";
3036
3037 if (failed(verifyTransferOp(cast<VectorTransferOpInterface>(getOperation()),
3038 shapedType, vectorType, maskType, permutationMap,
3039 getInBounds() ? *getInBounds() : ArrayAttr())))
3040 return failure();
3041
3042 if (auto sourceVectorElementType = sourceElementType.dyn_cast<VectorType>()) {
3043 // Source has vector element type.
3044 // Check that 'sourceVectorElementType' and 'paddingType' types match.
3045 if (sourceVectorElementType != paddingType)
3046 return emitOpError(
3047 "requires source element type and padding type to match.");
3048
3049 } else {
3050 // Check that 'paddingType' is valid to store in a vector type.
3051 if (!VectorType::isValidElementType(paddingType))
3052 return emitOpError("requires valid padding vector elemental type");
3053
3054 // Check that padding type and vector element types match.
3055 if (paddingType != sourceElementType)
3056 return emitOpError(
3057 "requires formal padding and source of the same elemental type");
3058 }
3059
3060 return verifyPermutationMap(permutationMap,
3061 [&](Twine t) { return emitOpError(t); });
3062 }
3063
3064 /// This is a common class used for patterns of the form
3065 /// ```
3066 /// someop(memrefcast) -> someop
3067 /// ```
3068 /// It folds the source of the memref.cast into the root operation directly.
foldMemRefCast(Operation * op)3069 static LogicalResult foldMemRefCast(Operation *op) {
3070 bool folded = false;
3071 for (OpOperand &operand : op->getOpOperands()) {
3072 auto castOp = operand.get().getDefiningOp<memref::CastOp>();
3073 if (castOp && memref::CastOp::canFoldIntoConsumerOp(castOp)) {
3074 operand.set(castOp.getOperand());
3075 folded = true;
3076 }
3077 }
3078 return success(folded);
3079 }
3080
foldTensorCast(Operation * op)3081 static LogicalResult foldTensorCast(Operation *op) {
3082 bool folded = false;
3083 for (OpOperand &operand : op->getOpOperands()) {
3084 auto castOp = operand.get().getDefiningOp<tensor::CastOp>();
3085 if (castOp && tensor::canFoldIntoConsumerOp(castOp)) {
3086 operand.set(castOp.getOperand());
3087 folded = true;
3088 }
3089 }
3090 return success(folded);
3091 }
3092
3093 template <typename TransferOp>
isInBounds(TransferOp op,int64_t resultIdx,int64_t indicesIdx)3094 static bool isInBounds(TransferOp op, int64_t resultIdx, int64_t indicesIdx) {
3095 // TODO: support more aggressive createOrFold on:
3096 // `op.indices()[indicesIdx] + vectorType < dim(op.source(), indicesIdx)`
3097 if (op.getShapedType().isDynamicDim(indicesIdx))
3098 return false;
3099 Value index = op.getIndices()[indicesIdx];
3100 auto cstOp = index.getDefiningOp<arith::ConstantIndexOp>();
3101 if (!cstOp)
3102 return false;
3103
3104 int64_t sourceSize = op.getShapedType().getDimSize(indicesIdx);
3105 int64_t vectorSize = op.getVectorType().getDimSize(resultIdx);
3106
3107 return cstOp.value() + vectorSize <= sourceSize;
3108 }
3109
3110 template <typename TransferOp>
foldTransferInBoundsAttribute(TransferOp op)3111 static LogicalResult foldTransferInBoundsAttribute(TransferOp op) {
3112 // TODO: support 0-d corner case.
3113 // TODO: Be less conservative.
3114 if (op.getTransferRank() == 0)
3115 return failure();
3116 AffineMap permutationMap = op.getPermutationMap();
3117 bool changed = false;
3118 SmallVector<bool, 4> newInBounds;
3119 newInBounds.reserve(op.getTransferRank());
3120 for (unsigned i = 0; i < op.getTransferRank(); ++i) {
3121 // Already marked as in-bounds, nothing to see here.
3122 if (op.isDimInBounds(i)) {
3123 newInBounds.push_back(true);
3124 continue;
3125 }
3126 // Currently out-of-bounds, check whether we can statically determine it is
3127 // inBounds.
3128 auto dimExpr = permutationMap.getResult(i).dyn_cast<AffineDimExpr>();
3129 assert(dimExpr && "Broadcast dims must be in-bounds");
3130 auto inBounds =
3131 isInBounds(op, /*resultIdx=*/i, /*indicesIdx=*/dimExpr.getPosition());
3132 newInBounds.push_back(inBounds);
3133 // We commit the pattern if it is "more inbounds".
3134 changed |= inBounds;
3135 }
3136 if (!changed)
3137 return failure();
3138 // OpBuilder is only used as a helper to build an I64ArrayAttr.
3139 OpBuilder b(op.getContext());
3140 op->setAttr(TransferOp::getInBoundsAttrStrName(),
3141 b.getBoolArrayAttr(newInBounds));
3142 return success();
3143 }
3144
3145 /// ```
3146 /// %w0 = vector.transfer_write %v0, %arg0[%c1, %c0] {in_bounds = [true, true]}
3147 /// : vector<1x4xf32>, tensor<4x4xf32>
3148 /// %0 = vector.transfer_read %w0[%c1, %c0], %cf0 {in_bounds = [true, true]}
3149 /// : tensor<4x4xf32>, vector<1x4xf32>
3150 /// ```
3151 /// -> Folds into
3152 /// ```
3153 /// %v0
3154 /// ```
foldRAW(TransferReadOp readOp)3155 static Value foldRAW(TransferReadOp readOp) {
3156 if (!readOp.getShapedType().isa<RankedTensorType>())
3157 return {};
3158 auto defWrite = readOp.getSource().getDefiningOp<vector::TransferWriteOp>();
3159 while (defWrite) {
3160 if (checkSameValueRAW(defWrite, readOp))
3161 return defWrite.getVector();
3162 if (!isDisjointTransferIndices(
3163 cast<VectorTransferOpInterface>(defWrite.getOperation()),
3164 cast<VectorTransferOpInterface>(readOp.getOperation())))
3165 break;
3166 defWrite = defWrite.getSource().getDefiningOp<vector::TransferWriteOp>();
3167 }
3168 return {};
3169 }
3170
fold(ArrayRef<Attribute>)3171 OpFoldResult TransferReadOp::fold(ArrayRef<Attribute>) {
3172 if (Value vec = foldRAW(*this))
3173 return vec;
3174 /// transfer_read(memrefcast) -> transfer_read
3175 if (succeeded(foldTransferInBoundsAttribute(*this)))
3176 return getResult();
3177 if (succeeded(foldMemRefCast(*this)))
3178 return getResult();
3179 if (succeeded(foldTensorCast(*this)))
3180 return getResult();
3181 return OpFoldResult();
3182 }
3183
getShapeForUnroll()3184 Optional<SmallVector<int64_t, 4>> TransferReadOp::getShapeForUnroll() {
3185 return llvm::to_vector<4>(getVectorType().getShape());
3186 }
3187
getEffects(SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>> & effects)3188 void TransferReadOp::getEffects(
3189 SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>>
3190 &effects) {
3191 if (getShapedType().isa<MemRefType>())
3192 effects.emplace_back(MemoryEffects::Read::get(), getSource(),
3193 SideEffects::DefaultResource::get());
3194 }
3195
3196 namespace {
3197 /// Fold transfer_reads of a tensor.extract_slice op. E.g.:
3198 ///
3199 /// ```
3200 /// %0 = tensor.extract_slice %t[%a, %b] [%c, %d] [1, 1]
3201 /// : tensor<?x?xf32> to tensor<?x?xf32>
3202 /// %1 = vector.transfer_read %0[%e, %f], %cst {in_bounds = [true, true]}
3203 /// : tensor<?x?xf32>, vector<4x5xf32>
3204 /// ```
3205 /// is rewritten to:
3206 /// ```
3207 /// %p0 = arith.addi %a, %e : index
3208 /// %p1 = arith.addi %b, %f : index
3209 /// %1 = vector.transfer_read %t[%p0, %p1], %cst {in_bounds = [true, true]}
3210 /// : tensor<?x?xf32>, vector<4x5xf32>
3211 /// ```
3212 struct FoldExtractSliceIntoTransferRead
3213 : public OpRewritePattern<TransferReadOp> {
3214 public:
3215 using OpRewritePattern<TransferReadOp>::OpRewritePattern;
3216
matchAndRewrite__anon088a7a4f1d11::FoldExtractSliceIntoTransferRead3217 LogicalResult matchAndRewrite(TransferReadOp xferOp,
3218 PatternRewriter &rewriter) const override {
3219 // TODO: support 0-d corner case.
3220 if (xferOp.getTransferRank() == 0)
3221 return failure();
3222 if (xferOp.hasOutOfBoundsDim())
3223 return failure();
3224 if (!xferOp.getPermutationMap().isIdentity())
3225 return failure();
3226 if (xferOp.getMask())
3227 return failure();
3228 auto extractOp = xferOp.getSource().getDefiningOp<tensor::ExtractSliceOp>();
3229 if (!extractOp)
3230 return failure();
3231 if (!extractOp.hasUnitStride())
3232 return failure();
3233
3234 // Bail on illegal rank-reduction: we need to check that the rank-reduced
3235 // dims are exactly the leading dims. I.e. the following is illegal:
3236 // ```
3237 // %0 = tensor.extract_slice %t[0,0,0][2,1,4][1,1,1] :
3238 // tensor<2x1x4xf32> to tensor<2x4xf32>
3239 // %1 = vector.transfer_read %0[0,0], %cst :
3240 // tensor<2x4xf32>, vector<2x4xf32>
3241 // ```
3242 //
3243 // Cannot fold into:
3244 // ```
3245 // %0 = vector.transfer_read %t[0,0,0], %cst :
3246 // tensor<2x1x4xf32>, vector<2x4xf32>
3247 // ```
3248 // For this, check the trailing `vectorRank` dims of the extract_slice
3249 // result tensor match the trailing dims of the inferred result tensor.
3250 int64_t rankReduced =
3251 extractOp.getSourceType().getRank() - extractOp.getType().getRank();
3252 int64_t vectorRank = xferOp.getVectorType().getRank();
3253 RankedTensorType inferredDestTensorType =
3254 tensor::ExtractSliceOp::inferResultType(
3255 extractOp.getSourceType(), extractOp.getMixedOffsets(),
3256 extractOp.getMixedSizes(), extractOp.getMixedStrides());
3257 auto actualDestTensorShape = extractOp.getType().getShape();
3258 if (rankReduced > 0 &&
3259 actualDestTensorShape.take_back(vectorRank) !=
3260 inferredDestTensorType.getShape().take_back(vectorRank))
3261 return failure();
3262
3263 SmallVector<Value> newIndices;
3264 // In case this is a rank-reducing ExtractSliceOp, copy rank-reduced
3265 // indices first.
3266 for (int64_t i = 0; i < rankReduced; ++i) {
3267 OpFoldResult offset = extractOp.getMixedOffsets()[i];
3268 newIndices.push_back(getValueOrCreateConstantIndexOp(
3269 rewriter, extractOp.getLoc(), offset));
3270 }
3271 for (const auto &it : llvm::enumerate(xferOp.getIndices())) {
3272 OpFoldResult offset =
3273 extractOp.getMixedOffsets()[it.index() + rankReduced];
3274 newIndices.push_back(rewriter.create<arith::AddIOp>(
3275 xferOp->getLoc(), it.value(),
3276 getValueOrCreateConstantIndexOp(rewriter, extractOp.getLoc(),
3277 offset)));
3278 }
3279 SmallVector<bool> inBounds(xferOp.getTransferRank(), true);
3280 rewriter.replaceOpWithNewOp<TransferReadOp>(
3281 xferOp, xferOp.getVectorType(), extractOp.getSource(), newIndices,
3282 xferOp.getPadding(), ArrayRef<bool>{inBounds});
3283
3284 return success();
3285 }
3286 };
3287
3288 /// Store to load forwarding for transfer operations with permuation maps.
3289 /// Even if the permutation maps are different we can still propagate the store
3290 /// into the load if the size of the dimensions read and written match. Then we
3291 /// can replace the transfer_read + transfer_write by vector.broadcast and
3292 /// vector.transpose.
3293 /// Example:
3294 /// ```
3295 /// %w0 = vector.transfer_write %v0, %arg0[%c0, %c0, %c0]
3296 /// {in_bounds = [true, true],
3297 /// permutation_map = affine_map<(d0, d1, d2) -> (d2, d1)>} :
3298 /// vector<4x1xf32>, tensor<4x4x4xf32>
3299 /// %r = vector.transfer_read %w0[%c0, %c0, %c0], %cf0
3300 /// {in_bounds = [true, true, true, true],
3301 /// permutation_map = affine_map<(d0, d1, d2) -> (d1, 0, d2, 0)>} :
3302 /// tensor<4x4x4xf32>, vector<1x100x4x5xf32>
3303 /// ```
3304 /// To:
3305 /// ```
3306 /// %0 = vector.broadcast %arg1 : vector<4x1xf32> to vector<100x5x4x1xf32>
3307 /// %r = vector.transpose %0, [3, 0, 2, 1] :
3308 /// vector<100x5x4x1xf32> to vector<1x100x4x5xf32>
3309 /// ```
3310 struct TransferReadAfterWriteToBroadcast
3311 : public OpRewritePattern<TransferReadOp> {
3312 using OpRewritePattern<TransferReadOp>::OpRewritePattern;
3313
matchAndRewrite__anon088a7a4f1d11::TransferReadAfterWriteToBroadcast3314 LogicalResult matchAndRewrite(TransferReadOp readOp,
3315 PatternRewriter &rewriter) const override {
3316 if (readOp.hasOutOfBoundsDim() ||
3317 !readOp.getShapedType().isa<RankedTensorType>())
3318 return failure();
3319 auto defWrite = readOp.getSource().getDefiningOp<vector::TransferWriteOp>();
3320 if (!defWrite)
3321 return failure();
3322
3323 SmallVector<int64_t> readDims = readOp.getTransferChunkAccessed();
3324 Value vec;
3325 if (readOp.getIndices() == defWrite.getIndices() &&
3326 readOp.getMask() == defWrite.getMask()) {
3327 SmallVector<int64_t> writeDims = defWrite.getTransferChunkAccessed();
3328 // TODO: If the writeDim is a superset of the read dims we could do an
3329 // extract_strided_slice.
3330 if (writeDims == readDims)
3331 vec = defWrite.getVector();
3332 }
3333 // TODO: loop through the chain of transfer_write if we can prove that they
3334 // don't overlap with the transfer_read. This requires improving
3335 // `isDisjointTransferIndices` helper.
3336 if (!vec)
3337 return failure();
3338 SmallVector<unsigned> permutation;
3339 AffineMap readMap = compressUnusedDims(readOp.getPermutationMap());
3340 AffineMap writeMap = compressUnusedDims(defWrite.getPermutationMap());
3341 AffineMap map = readMap.compose(writeMap);
3342 if (map.getNumResults() == 0)
3343 return failure();
3344 // Calculate the permuation to apply to go from the vector stored to the
3345 // vector read.
3346 if (!map.isPermutationOfMinorIdentityWithBroadcasting(permutation))
3347 return failure();
3348
3349 Location loc = readOp.getLoc();
3350 // Calculate the broadcast shape by applying the reverse permuation to the
3351 // final shape we want.
3352 ArrayRef<int64_t> destShape = readOp.getVectorType().getShape();
3353 SmallVector<int64_t> broadcastShape(destShape.size());
3354 for (const auto &pos : llvm::enumerate(permutation))
3355 broadcastShape[pos.value()] = destShape[pos.index()];
3356 VectorType broadcastedType = VectorType::get(
3357 broadcastShape, defWrite.getVectorType().getElementType());
3358 vec = rewriter.create<vector::BroadcastOp>(loc, broadcastedType, vec);
3359 SmallVector<int64_t> transposePerm(permutation.begin(), permutation.end());
3360 rewriter.replaceOpWithNewOp<vector::TransposeOp>(readOp, vec,
3361 transposePerm);
3362 return success();
3363 }
3364 };
3365 } // namespace
3366
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)3367 void TransferReadOp::getCanonicalizationPatterns(RewritePatternSet &results,
3368 MLIRContext *context) {
3369 results
3370 .add<FoldExtractSliceIntoTransferRead, TransferReadAfterWriteToBroadcast>(
3371 context);
3372 }
3373
3374 //===----------------------------------------------------------------------===//
3375 // TransferWriteOp
3376 //===----------------------------------------------------------------------===//
3377
3378 /// 1. Builder with type inference.
build(OpBuilder & builder,OperationState & result,Value vector,Value dest,ValueRange indices,AffineMapAttr permutationMapAttr,Value mask,ArrayAttr inBoundsAttr)3379 void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
3380 Value vector, Value dest, ValueRange indices,
3381 AffineMapAttr permutationMapAttr,
3382 /*optional*/ Value mask,
3383 /*optional*/ ArrayAttr inBoundsAttr) {
3384 Type resultType = dest.getType().dyn_cast<RankedTensorType>();
3385 build(builder, result, resultType, vector, dest, indices, permutationMapAttr,
3386 mask, inBoundsAttr);
3387 }
3388
3389 /// 2. Builder with type inference that sets an empty mask (variant with attrs).
build(OpBuilder & builder,OperationState & result,Value vector,Value dest,ValueRange indices,AffineMapAttr permutationMapAttr,ArrayAttr inBoundsAttr)3390 void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
3391 Value vector, Value dest, ValueRange indices,
3392 AffineMapAttr permutationMapAttr,
3393 /*optional*/ ArrayAttr inBoundsAttr) {
3394 build(builder, result, vector, dest, indices, permutationMapAttr,
3395 /*mask=*/Value(), inBoundsAttr);
3396 }
3397
3398 /// 3. Builder with type inference that sets an empty mask (variant without
3399 /// attrs)
build(OpBuilder & builder,OperationState & result,Value vector,Value dest,ValueRange indices,AffineMap permutationMap,Optional<ArrayRef<bool>> inBounds)3400 void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
3401 Value vector, Value dest, ValueRange indices,
3402 AffineMap permutationMap,
3403 Optional<ArrayRef<bool>> inBounds) {
3404 auto permutationMapAttr = AffineMapAttr::get(permutationMap);
3405 auto inBoundsAttr = (inBounds && !inBounds.value().empty())
3406 ? builder.getBoolArrayAttr(inBounds.value())
3407 : ArrayAttr();
3408 build(builder, result, vector, dest, indices, permutationMapAttr,
3409 /*mask=*/Value(), inBoundsAttr);
3410 }
3411
3412 /// 4. Builder with type inference that sets an empty mask and sets permutation
3413 /// map to 'getMinorIdentityMap'.
build(OpBuilder & builder,OperationState & result,Value vector,Value dest,ValueRange indices,Optional<ArrayRef<bool>> inBounds)3414 void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
3415 Value vector, Value dest, ValueRange indices,
3416 Optional<ArrayRef<bool>> inBounds) {
3417 auto vectorType = vector.getType().cast<VectorType>();
3418 AffineMap permutationMap = getTransferMinorIdentityMap(
3419 dest.getType().cast<ShapedType>(), vectorType);
3420 build(builder, result, vector, dest, indices, permutationMap, inBounds);
3421 }
3422
parse(OpAsmParser & parser,OperationState & result)3423 ParseResult TransferWriteOp::parse(OpAsmParser &parser,
3424 OperationState &result) {
3425 auto &builder = parser.getBuilder();
3426 SMLoc typesLoc;
3427 OpAsmParser::UnresolvedOperand vectorInfo, sourceInfo;
3428 SmallVector<OpAsmParser::UnresolvedOperand, 8> indexInfo;
3429 SmallVector<Type, 2> types;
3430 OpAsmParser::UnresolvedOperand maskInfo;
3431 if (parser.parseOperand(vectorInfo) || parser.parseComma() ||
3432 parser.parseOperand(sourceInfo) ||
3433 parser.parseOperandList(indexInfo, OpAsmParser::Delimiter::Square))
3434 return failure();
3435 ParseResult hasMask = parser.parseOptionalComma();
3436 if (hasMask.succeeded() && parser.parseOperand(maskInfo))
3437 return failure();
3438 if (parser.parseOptionalAttrDict(result.attributes) ||
3439 parser.getCurrentLocation(&typesLoc) || parser.parseColonTypeList(types))
3440 return failure();
3441 if (types.size() != 2)
3442 return parser.emitError(typesLoc, "requires two types");
3443 auto indexType = builder.getIndexType();
3444 VectorType vectorType = types[0].dyn_cast<VectorType>();
3445 if (!vectorType)
3446 return parser.emitError(typesLoc, "requires vector type");
3447 ShapedType shapedType = types[1].dyn_cast<ShapedType>();
3448 if (!shapedType || !shapedType.isa<MemRefType, RankedTensorType>())
3449 return parser.emitError(typesLoc, "requires memref or ranked tensor type");
3450 auto permutationAttrName = TransferWriteOp::getPermutationMapAttrStrName();
3451 auto attr = result.attributes.get(permutationAttrName);
3452 if (!attr) {
3453 auto permMap = getTransferMinorIdentityMap(shapedType, vectorType);
3454 result.attributes.set(permutationAttrName, AffineMapAttr::get(permMap));
3455 }
3456 if (parser.resolveOperand(vectorInfo, vectorType, result.operands) ||
3457 parser.resolveOperand(sourceInfo, shapedType, result.operands) ||
3458 parser.resolveOperands(indexInfo, indexType, result.operands))
3459 return failure();
3460 if (hasMask.succeeded()) {
3461 if (shapedType.getElementType().dyn_cast<VectorType>())
3462 return parser.emitError(
3463 maskInfo.location, "does not support masks with vector element type");
3464 auto maskType = VectorType::get(vectorType.getShape(), builder.getI1Type());
3465 if (parser.resolveOperand(maskInfo, maskType, result.operands))
3466 return failure();
3467 }
3468 result.addAttribute(
3469 TransferWriteOp::getOperandSegmentSizeAttr(),
3470 builder.getI32VectorAttr({1, 1, static_cast<int32_t>(indexInfo.size()),
3471 static_cast<int32_t>(hasMask.succeeded())}));
3472 return failure(shapedType.isa<RankedTensorType>() &&
3473 parser.addTypeToList(shapedType, result.types));
3474 }
3475
print(OpAsmPrinter & p)3476 void TransferWriteOp::print(OpAsmPrinter &p) {
3477 p << " " << getVector() << ", " << getSource() << "[" << getIndices() << "]";
3478 if (getMask())
3479 p << ", " << getMask();
3480 printTransferAttrs(p, *this);
3481 p << " : " << getVectorType() << ", " << getShapedType();
3482 }
3483
verify()3484 LogicalResult TransferWriteOp::verify() {
3485 // Consistency of elemental types in shape and vector.
3486 ShapedType shapedType = getShapedType();
3487 VectorType vectorType = getVectorType();
3488 VectorType maskType = getMaskType();
3489 auto permutationMap = getPermutationMap();
3490
3491 if (llvm::size(getIndices()) != shapedType.getRank())
3492 return emitOpError("requires ") << shapedType.getRank() << " indices";
3493
3494 // We do not allow broadcast dimensions on TransferWriteOps for the moment,
3495 // as the semantics is unclear. This can be revisited later if necessary.
3496 if (hasBroadcastDim())
3497 return emitOpError("should not have broadcast dimensions");
3498
3499 if (failed(verifyTransferOp(cast<VectorTransferOpInterface>(getOperation()),
3500 shapedType, vectorType, maskType, permutationMap,
3501 getInBounds() ? *getInBounds() : ArrayAttr())))
3502 return failure();
3503
3504 return verifyPermutationMap(permutationMap,
3505 [&](Twine t) { return emitOpError(t); });
3506 }
3507
3508 /// Fold:
3509 /// ```
3510 /// %t1 = ...
3511 /// %v = vector.transfer_read %t0[%c0...], {in_bounds = [true...]} :
3512 /// tensor<static_sizesxf32>, vector<static_sizesxf32>
3513 /// %t2 = vector.transfer_write %v, %t1[%c0...] {in_bounds = [true...]} :
3514 /// vector<static_sizesxf32>, tensor<static_sizesxf32>
3515 /// ```
3516 ///
3517 /// into:
3518 ///
3519 /// ```
3520 /// %t0
3521 /// ```
3522 ///
3523 /// The producer of t1 may or may not be DCE'd depending on whether it is a
3524 /// block argument or has side effects.
foldReadInitWrite(TransferWriteOp write,ArrayRef<Attribute>,SmallVectorImpl<OpFoldResult> & results)3525 static LogicalResult foldReadInitWrite(TransferWriteOp write,
3526 ArrayRef<Attribute>,
3527 SmallVectorImpl<OpFoldResult> &results) {
3528 // TODO: support 0-d corner case.
3529 if (write.getTransferRank() == 0)
3530 return failure();
3531 auto rankedTensorType =
3532 write.getSource().getType().dyn_cast<RankedTensorType>();
3533 // If not operating on tensors, bail.
3534 if (!rankedTensorType)
3535 return failure();
3536 // If no read, bail.
3537 auto read = write.getVector().getDefiningOp<vector::TransferReadOp>();
3538 if (!read)
3539 return failure();
3540 // TODO: support 0-d corner case.
3541 if (read.getTransferRank() == 0)
3542 return failure();
3543 // For now, only accept minor identity. Future: composition is minor identity.
3544 if (!read.getPermutationMap().isMinorIdentity() ||
3545 !write.getPermutationMap().isMinorIdentity())
3546 return failure();
3547 // Bail on mismatching ranks.
3548 if (read.getTransferRank() != write.getTransferRank())
3549 return failure();
3550 // Bail on potential out-of-bounds accesses.
3551 if (read.hasOutOfBoundsDim() || write.hasOutOfBoundsDim())
3552 return failure();
3553 // Tensor types must be the same.
3554 if (read.getSource().getType() != rankedTensorType)
3555 return failure();
3556 // Vector types must be the same.
3557 if (read.getVectorType() != write.getVectorType())
3558 return failure();
3559 // Vector and Tensor shapes must match.
3560 if (read.getVectorType().getShape() != rankedTensorType.getShape())
3561 return failure();
3562 // If any index is nonzero.
3563 auto isNotConstantZero = [](Value v) {
3564 auto cstOp = v.getDefiningOp<arith::ConstantIndexOp>();
3565 return !cstOp || cstOp.value() != 0;
3566 };
3567 if (llvm::any_of(read.getIndices(), isNotConstantZero) ||
3568 llvm::any_of(write.getIndices(), isNotConstantZero))
3569 return failure();
3570 // Success.
3571 results.push_back(read.getSource());
3572 return success();
3573 }
3574
checkSameValueWAR(vector::TransferReadOp read,vector::TransferWriteOp write)3575 static bool checkSameValueWAR(vector::TransferReadOp read,
3576 vector::TransferWriteOp write) {
3577 return read.getSource() == write.getSource() &&
3578 read.getIndices() == write.getIndices() &&
3579 read.getPermutationMap() == write.getPermutationMap() &&
3580 read.getVectorType() == write.getVectorType() && !read.getMask() &&
3581 !write.getMask();
3582 }
3583 /// Fold transfer_write write after read:
3584 /// ```
3585 /// %t0 = ...
3586 /// %v = vector.transfer_read %t0[%c0...] :
3587 /// tensor<static_sizesxf32>, vector<static_sizesxf32>
3588 /// %t1 = vector.transfer_write %v, %t0[%c0...] :
3589 /// vector<static_sizesxf32>, tensor<static_sizesxf32>
3590 /// ```
3591 ///
3592 /// into:
3593 ///
3594 /// ```
3595 /// %t0
3596 /// ```
foldWAR(TransferWriteOp write,SmallVectorImpl<OpFoldResult> & results)3597 static LogicalResult foldWAR(TransferWriteOp write,
3598 SmallVectorImpl<OpFoldResult> &results) {
3599 if (!write.getSource().getType().isa<RankedTensorType>())
3600 return failure();
3601 auto read = write.getVector().getDefiningOp<vector::TransferReadOp>();
3602 if (!read)
3603 return failure();
3604
3605 if (!checkSameValueWAR(read, write))
3606 return failure();
3607 results.push_back(read.getSource());
3608 return success();
3609 }
3610
fold(ArrayRef<Attribute> operands,SmallVectorImpl<OpFoldResult> & results)3611 LogicalResult TransferWriteOp::fold(ArrayRef<Attribute> operands,
3612 SmallVectorImpl<OpFoldResult> &results) {
3613 if (succeeded(foldReadInitWrite(*this, operands, results)))
3614 return success();
3615 if (succeeded(foldWAR(*this, results)))
3616 return success();
3617 if (succeeded(foldTransferInBoundsAttribute(*this)))
3618 return success();
3619 return foldMemRefCast(*this);
3620 }
3621
getShapeForUnroll()3622 Optional<SmallVector<int64_t, 4>> TransferWriteOp::getShapeForUnroll() {
3623 return llvm::to_vector<4>(getVectorType().getShape());
3624 }
3625
getEffects(SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>> & effects)3626 void TransferWriteOp::getEffects(
3627 SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>>
3628 &effects) {
3629 if (getShapedType().isa<MemRefType>())
3630 effects.emplace_back(MemoryEffects::Write::get(), getSource(),
3631 SideEffects::DefaultResource::get());
3632 }
3633
3634 namespace {
3635 /// Remove dead transfer write from the SSA chain so that it an be eliminated by
3636 /// DCE
3637 /// ```
3638 /// %w0 = vector.transfer_write %v0, %arg0[%c1, %c0] {in_bounds = [true, true]}
3639 /// : vector<1x4xf32>, tensor<4x4xf32>
3640 /// %w1 = vector.transfer_write %v0, %w0[%c2, %c0] {in_bounds = [true, true]}
3641 /// : vector<1x4xf32>, tensor<4x4xf32>
3642 /// %w2 = vector.transfer_write %v1, %w1[%c1, %c0] {in_bounds = [true, true]}
3643 /// : vector<1x4xf32>, tensor<4x4xf32>
3644 /// ```
3645 ///
3646 /// into:
3647 ///
3648 /// ```
3649 /// %w0 = vector.transfer_write %v0, %arg0[%c1, %c0] {in_bounds = [true, true]}
3650 /// : vector<1x4xf32>, tensor<4x4xf32>
3651 /// %w1 = vector.transfer_write %v0, %arg0[%c2, %c0] {in_bounds = [true, true]}
3652 /// : vector<1x4xf32>, tensor<4x4xf32>
3653 /// %w2 = vector.transfer_write %v1, %w1[%c1, %c0] {in_bounds = [true, true]}
3654 /// : vector<1x4xf32>, tensor<4x4xf32>
3655 /// ```
3656 ///
3657 /// `%w0 = vector.transfer_write` op will be removed by DCE if it doesn't have
3658 /// any other uses.
3659 class FoldWaw final : public OpRewritePattern<TransferWriteOp> {
3660 public:
3661 using OpRewritePattern<TransferWriteOp>::OpRewritePattern;
matchAndRewrite(TransferWriteOp writeOp,PatternRewriter & rewriter) const3662 LogicalResult matchAndRewrite(TransferWriteOp writeOp,
3663 PatternRewriter &rewriter) const override {
3664 if (!writeOp.getShapedType().isa<RankedTensorType>())
3665 return failure();
3666 vector::TransferWriteOp writeToModify = writeOp;
3667
3668 auto defWrite =
3669 writeOp.getSource().getDefiningOp<vector::TransferWriteOp>();
3670 while (defWrite) {
3671 if (checkSameValueWAW(writeOp, defWrite)) {
3672 writeToModify.getSourceMutable().assign(defWrite.getSource());
3673 return success();
3674 }
3675 if (!isDisjointTransferIndices(
3676 cast<VectorTransferOpInterface>(defWrite.getOperation()),
3677 cast<VectorTransferOpInterface>(writeOp.getOperation())))
3678 break;
3679 // If the previous write op doesn't have any other use we an safely look
3680 // at the previous store to see if it can be removed.
3681 if (!defWrite->hasOneUse())
3682 break;
3683 writeToModify = defWrite;
3684 defWrite = defWrite.getSource().getDefiningOp<vector::TransferWriteOp>();
3685 }
3686 return failure();
3687 }
3688 };
3689
3690 /// Fold tensor.insert_slice into vector.transfer_write if the transfer_write
3691 /// could directly write to the insert_slice's destination. E.g.:
3692 ///
3693 /// ```
3694 /// %0 = vector.transfer_write %v, %t1[%c0, %c0] {in_bounds = [true, true]}
3695 /// : vector<4x5xf32>, tensor<4x5xf32>
3696 /// %1 = tensor.insert_slice %0 into %t2[%a, %b] [4, 5] [1, 1]
3697 /// : tensor<4x5xf32> into tensor<?x?xf32>
3698 /// ```
3699 /// is rewritten to:
3700 /// ```
3701 /// %1 = vector.transfer_write %v, %t2[%a, %b] {in_bounds = [true, true]}
3702 /// : vector<4x5xf32>, tensor<?x?xf32>
3703 /// ```
3704 struct FoldInsertSliceIntoTransferWrite
3705 : public OpRewritePattern<tensor::InsertSliceOp> {
3706 public:
3707 using OpRewritePattern<tensor::InsertSliceOp>::OpRewritePattern;
3708
matchAndRewrite__anon088a7a4f2011::FoldInsertSliceIntoTransferWrite3709 LogicalResult matchAndRewrite(tensor::InsertSliceOp insertOp,
3710 PatternRewriter &rewriter) const override {
3711 if (!insertOp.hasUnitStride())
3712 return failure();
3713
3714 auto xferOp = insertOp.getSource().getDefiningOp<TransferWriteOp>();
3715 if (!xferOp)
3716 return failure();
3717 // TODO: support 0-d corner case.
3718 if (xferOp.getTransferRank() == 0)
3719 return failure();
3720
3721 if (xferOp.hasOutOfBoundsDim())
3722 return failure();
3723 if (xferOp.getVectorType().getRank() != xferOp.getShapedType().getRank())
3724 return failure();
3725 if (xferOp.getMask())
3726 return failure();
3727 // Fold only if the TransferWriteOp completely overwrites the `source` with
3728 // a vector. I.e., the result of the TransferWriteOp is a new tensor whose
3729 // content is the data of the vector.
3730 if (!llvm::equal(xferOp.getVectorType().getShape(),
3731 xferOp.getShapedType().getShape()))
3732 return failure();
3733 if (!xferOp.getPermutationMap().isIdentity())
3734 return failure();
3735
3736 // Bail on illegal rank-reduction: we need to check that the rank-reduced
3737 // dims are exactly the leading dims. I.e. the following is illegal:
3738 // ```
3739 // %0 = vector.transfer_write %v, %t[0,0], %cst :
3740 // vector<2x4xf32>, tensor<2x4xf32>
3741 // %1 = tensor.insert_slice %0 into %tt[0,0,0][2,1,4][1,1,1] :
3742 // tensor<2x4xf32> into tensor<2x1x4xf32>
3743 // ```
3744 //
3745 // Cannot fold into:
3746 // ```
3747 // %0 = vector.transfer_write %v, %t[0,0,0], %cst :
3748 // vector<2x4xf32>, tensor<2x1x4xf32>
3749 // ```
3750 // For this, check the trailing `vectorRank` dims of the insert_slice result
3751 // tensor match the trailing dims of the inferred result tensor.
3752 int64_t rankReduced =
3753 insertOp.getType().getRank() - insertOp.getSourceType().getRank();
3754 int64_t vectorRank = xferOp.getVectorType().getRank();
3755 RankedTensorType inferredSourceTensorType =
3756 tensor::ExtractSliceOp::inferResultType(
3757 insertOp.getType(), insertOp.getMixedOffsets(),
3758 insertOp.getMixedSizes(), insertOp.getMixedStrides());
3759 auto actualSourceTensorShape = insertOp.getSourceType().getShape();
3760 if (rankReduced > 0 &&
3761 actualSourceTensorShape.take_back(vectorRank) !=
3762 inferredSourceTensorType.getShape().take_back(vectorRank))
3763 return failure();
3764
3765 SmallVector<Value> indices = getValueOrCreateConstantIndexOp(
3766 rewriter, insertOp.getLoc(), insertOp.getMixedOffsets());
3767 SmallVector<bool> inBounds(xferOp.getTransferRank(), true);
3768 rewriter.replaceOpWithNewOp<TransferWriteOp>(insertOp, xferOp.getVector(),
3769 insertOp.getDest(), indices,
3770 ArrayRef<bool>{inBounds});
3771 return success();
3772 }
3773 };
3774
3775 /// Rewrite tensor::ExtractSliceOp(vector::TransferWriteOp) to
3776 /// vector::TransferWriteOp(tensor::ExtractSliceOp) if the full slice is
3777 /// overwritten and inserted into another tensor. After this rewrite, the
3778 /// operations bufferize in-place since all of them work on the same slice.
3779 ///
3780 /// For example:
3781 /// ```mlir
3782 /// %0 = vector.transfer_write %vec, %init_tensor[%c0, %c0]
3783 /// : vector<8x16xf32>, tensor<8x16xf32>
3784 /// %1 = tensor.extract_slice %0[0, 0] [%sz0, %sz1] [1, 1]
3785 /// : tensor<8x16xf32> to tensor<?x?xf32>
3786 /// %r = tensor.insert_slice %1 into %iter_arg[%iv0, %iv1] [%sz0, %sz1] [1, 1]
3787 /// : tensor<?x?xf32> into tensor<27x37xf32>
3788 /// ```
3789 /// folds to
3790 /// ```mlir
3791 /// %0 = tensor.extract_slice %iter_arg[%iv0, %iv1] [%sz0, %sz1] [1, 1]
3792 /// : tensor<27x37xf32> to tensor<?x?xf32>
3793 /// %1 = vector.transfer_write %vec, %0[%c0, %c0]
3794 /// : vector<8x16xf32>, tensor<?x?xf32>
3795 /// %r = tensor.insert_slice %1 into %iter_arg[%iv0, %iv1] [%sz0, %sz1] [1, 1]
3796 /// : tensor<?x?xf32> into tensor<27x37xf32>
3797 /// ```
3798 struct SwapExtractSliceOfTransferWrite
3799 : public OpRewritePattern<tensor::InsertSliceOp> {
3800 public:
3801 using OpRewritePattern<tensor::InsertSliceOp>::OpRewritePattern;
3802
matchAndRewrite__anon088a7a4f2011::SwapExtractSliceOfTransferWrite3803 LogicalResult matchAndRewrite(tensor::InsertSliceOp insertOp,
3804 PatternRewriter &rewriter) const override {
3805 if (!insertOp.hasUnitStride())
3806 return failure();
3807 auto extractOp =
3808 insertOp.getSource().getDefiningOp<tensor::ExtractSliceOp>();
3809 if (!extractOp || !extractOp.hasUnitStride() || !extractOp->hasOneUse())
3810 return failure();
3811 auto transferOp = extractOp.getSource().getDefiningOp<TransferWriteOp>();
3812 if (!transferOp || !transferOp->hasOneUse())
3813 return failure();
3814
3815 // Fail if vector::TransferWriteOp or tensor::ExtractSliceOp is
3816 // rank-reducing.
3817 if (insertOp.getSourceType().getRank() != transferOp.getTransferRank()) {
3818 return rewriter.notifyMatchFailure(insertOp,
3819 "use-def chain is rank-reducing");
3820 }
3821
3822 // Fail if tensor::ExtractSliceOp has non-zero offset.
3823 if (!extractOp.hasZeroOffset()) {
3824 return rewriter.notifyMatchFailure(insertOp,
3825 "ExtractSliceOp has non-zero offset");
3826 }
3827
3828 // Fail if tensor::TransferWriteOp has non-zero offset.
3829 if (!llvm::all_of(transferOp.getIndices(), [](Value value) {
3830 return getConstantIntValue(value) == static_cast<int64_t>(0);
3831 })) {
3832 return rewriter.notifyMatchFailure(insertOp,
3833 "TranferWriteOp has non-zero offset");
3834 }
3835
3836 // Fail if tensor::ExtractSliceOp and tensor::InsertSliceOp sizes differ.
3837 for (const auto &it :
3838 llvm::zip(insertOp.getMixedSizes(), extractOp.getMixedSizes())) {
3839 if (!isEqualConstantIntOrValue(std::get<0>(it), std::get<1>(it))) {
3840 return rewriter.notifyMatchFailure(
3841 insertOp, "InsertSliceOp and ExtractSliceOp sizes differ");
3842 }
3843 }
3844
3845 // Fail if the vector::TransferWriteOp may not overwrite the full tensor.
3846 assert(transferOp.getVectorType().hasStaticShape() &&
3847 "expected vector to have a static shape");
3848 ArrayRef<int64_t> vectorShape = transferOp.getVectorType().getShape();
3849 SmallVector<int64_t> resultShape = applyPermutationMap(
3850 transferOp.getPermutationMap(), transferOp.getShapedType().getShape());
3851 if (transferOp.getMask() || !vectorShape.equals(resultShape)) {
3852 return rewriter.notifyMatchFailure(
3853 insertOp, "TransferWriteOp may not write the full tensor.");
3854 }
3855
3856 // Swap the tensor::ExtractSliceOp in front of the vector::TransferWriteOp.
3857 SmallVector<int64_t> newResultShape = applyPermutationMap(
3858 transferOp.getPermutationMap(), insertOp.getSourceType().getShape());
3859 SmallVector<bool> newInBounds;
3860 for (const auto &en : enumerate(newResultShape))
3861 newInBounds.push_back(en.value() == vectorShape[en.index()]);
3862 auto newExtractOp = rewriter.create<tensor::ExtractSliceOp>(
3863 extractOp.getLoc(), insertOp.getSourceType(), insertOp.getDest(),
3864 insertOp.getMixedOffsets(), insertOp.getMixedSizes(),
3865 insertOp.getMixedStrides());
3866 auto newTransferWriteOp = rewriter.create<TransferWriteOp>(
3867 transferOp.getLoc(), transferOp.getVector(), newExtractOp.getResult(),
3868 transferOp.getIndices(), transferOp.getPermutationMapAttr(),
3869 rewriter.getBoolArrayAttr(newInBounds));
3870 rewriter.updateRootInPlace(insertOp, [&]() {
3871 insertOp.getSourceMutable().assign(newTransferWriteOp.getResult());
3872 });
3873 return success();
3874 }
3875 };
3876
3877 } // namespace
3878
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)3879 void TransferWriteOp::getCanonicalizationPatterns(RewritePatternSet &results,
3880 MLIRContext *context) {
3881 results.add<FoldWaw, FoldInsertSliceIntoTransferWrite,
3882 SwapExtractSliceOfTransferWrite>(context);
3883 }
3884
3885 //===----------------------------------------------------------------------===//
3886 // LoadOp
3887 //===----------------------------------------------------------------------===//
3888
verifyLoadStoreMemRefLayout(Operation * op,MemRefType memRefTy)3889 static LogicalResult verifyLoadStoreMemRefLayout(Operation *op,
3890 MemRefType memRefTy) {
3891 if (!isLastMemrefDimUnitStride(memRefTy))
3892 return op->emitOpError("most minor memref dim must have unit stride");
3893 return success();
3894 }
3895
verify()3896 LogicalResult vector::LoadOp::verify() {
3897 VectorType resVecTy = getVectorType();
3898 MemRefType memRefTy = getMemRefType();
3899
3900 if (failed(verifyLoadStoreMemRefLayout(*this, memRefTy)))
3901 return failure();
3902
3903 // Checks for vector memrefs.
3904 Type memElemTy = memRefTy.getElementType();
3905 if (auto memVecTy = memElemTy.dyn_cast<VectorType>()) {
3906 if (memVecTy != resVecTy)
3907 return emitOpError("base memref and result vector types should match");
3908 memElemTy = memVecTy.getElementType();
3909 }
3910
3911 if (resVecTy.getElementType() != memElemTy)
3912 return emitOpError("base and result element types should match");
3913 if (llvm::size(getIndices()) != memRefTy.getRank())
3914 return emitOpError("requires ") << memRefTy.getRank() << " indices";
3915 return success();
3916 }
3917
fold(ArrayRef<Attribute>)3918 OpFoldResult LoadOp::fold(ArrayRef<Attribute>) {
3919 if (succeeded(foldMemRefCast(*this)))
3920 return getResult();
3921 return OpFoldResult();
3922 }
3923
3924 //===----------------------------------------------------------------------===//
3925 // StoreOp
3926 //===----------------------------------------------------------------------===//
3927
verify()3928 LogicalResult vector::StoreOp::verify() {
3929 VectorType valueVecTy = getVectorType();
3930 MemRefType memRefTy = getMemRefType();
3931
3932 if (failed(verifyLoadStoreMemRefLayout(*this, memRefTy)))
3933 return failure();
3934
3935 // Checks for vector memrefs.
3936 Type memElemTy = memRefTy.getElementType();
3937 if (auto memVecTy = memElemTy.dyn_cast<VectorType>()) {
3938 if (memVecTy != valueVecTy)
3939 return emitOpError(
3940 "base memref and valueToStore vector types should match");
3941 memElemTy = memVecTy.getElementType();
3942 }
3943
3944 if (valueVecTy.getElementType() != memElemTy)
3945 return emitOpError("base and valueToStore element type should match");
3946 if (llvm::size(getIndices()) != memRefTy.getRank())
3947 return emitOpError("requires ") << memRefTy.getRank() << " indices";
3948 return success();
3949 }
3950
fold(ArrayRef<Attribute> operands,SmallVectorImpl<OpFoldResult> & results)3951 LogicalResult StoreOp::fold(ArrayRef<Attribute> operands,
3952 SmallVectorImpl<OpFoldResult> &results) {
3953 return foldMemRefCast(*this);
3954 }
3955
3956 //===----------------------------------------------------------------------===//
3957 // MaskedLoadOp
3958 //===----------------------------------------------------------------------===//
3959
verify()3960 LogicalResult MaskedLoadOp::verify() {
3961 VectorType maskVType = getMaskVectorType();
3962 VectorType passVType = getPassThruVectorType();
3963 VectorType resVType = getVectorType();
3964 MemRefType memType = getMemRefType();
3965
3966 if (resVType.getElementType() != memType.getElementType())
3967 return emitOpError("base and result element type should match");
3968 if (llvm::size(getIndices()) != memType.getRank())
3969 return emitOpError("requires ") << memType.getRank() << " indices";
3970 if (resVType.getDimSize(0) != maskVType.getDimSize(0))
3971 return emitOpError("expected result dim to match mask dim");
3972 if (resVType != passVType)
3973 return emitOpError("expected pass_thru of same type as result type");
3974 return success();
3975 }
3976
3977 namespace {
3978 class MaskedLoadFolder final : public OpRewritePattern<MaskedLoadOp> {
3979 public:
3980 using OpRewritePattern<MaskedLoadOp>::OpRewritePattern;
matchAndRewrite(MaskedLoadOp load,PatternRewriter & rewriter) const3981 LogicalResult matchAndRewrite(MaskedLoadOp load,
3982 PatternRewriter &rewriter) const override {
3983 switch (get1DMaskFormat(load.getMask())) {
3984 case MaskFormat::AllTrue:
3985 rewriter.replaceOpWithNewOp<vector::LoadOp>(
3986 load, load.getType(), load.getBase(), load.getIndices());
3987 return success();
3988 case MaskFormat::AllFalse:
3989 rewriter.replaceOp(load, load.getPassThru());
3990 return success();
3991 case MaskFormat::Unknown:
3992 return failure();
3993 }
3994 llvm_unreachable("Unexpected 1DMaskFormat on MaskedLoad");
3995 }
3996 };
3997 } // namespace
3998
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)3999 void MaskedLoadOp::getCanonicalizationPatterns(RewritePatternSet &results,
4000 MLIRContext *context) {
4001 results.add<MaskedLoadFolder>(context);
4002 }
4003
fold(ArrayRef<Attribute>)4004 OpFoldResult MaskedLoadOp::fold(ArrayRef<Attribute>) {
4005 if (succeeded(foldMemRefCast(*this)))
4006 return getResult();
4007 return OpFoldResult();
4008 }
4009
4010 //===----------------------------------------------------------------------===//
4011 // MaskedStoreOp
4012 //===----------------------------------------------------------------------===//
4013
verify()4014 LogicalResult MaskedStoreOp::verify() {
4015 VectorType maskVType = getMaskVectorType();
4016 VectorType valueVType = getVectorType();
4017 MemRefType memType = getMemRefType();
4018
4019 if (valueVType.getElementType() != memType.getElementType())
4020 return emitOpError("base and valueToStore element type should match");
4021 if (llvm::size(getIndices()) != memType.getRank())
4022 return emitOpError("requires ") << memType.getRank() << " indices";
4023 if (valueVType.getDimSize(0) != maskVType.getDimSize(0))
4024 return emitOpError("expected valueToStore dim to match mask dim");
4025 return success();
4026 }
4027
4028 namespace {
4029 class MaskedStoreFolder final : public OpRewritePattern<MaskedStoreOp> {
4030 public:
4031 using OpRewritePattern<MaskedStoreOp>::OpRewritePattern;
matchAndRewrite(MaskedStoreOp store,PatternRewriter & rewriter) const4032 LogicalResult matchAndRewrite(MaskedStoreOp store,
4033 PatternRewriter &rewriter) const override {
4034 switch (get1DMaskFormat(store.getMask())) {
4035 case MaskFormat::AllTrue:
4036 rewriter.replaceOpWithNewOp<vector::StoreOp>(
4037 store, store.getValueToStore(), store.getBase(), store.getIndices());
4038 return success();
4039 case MaskFormat::AllFalse:
4040 rewriter.eraseOp(store);
4041 return success();
4042 case MaskFormat::Unknown:
4043 return failure();
4044 }
4045 llvm_unreachable("Unexpected 1DMaskFormat on MaskedStore");
4046 }
4047 };
4048 } // namespace
4049
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)4050 void MaskedStoreOp::getCanonicalizationPatterns(RewritePatternSet &results,
4051 MLIRContext *context) {
4052 results.add<MaskedStoreFolder>(context);
4053 }
4054
fold(ArrayRef<Attribute> operands,SmallVectorImpl<OpFoldResult> & results)4055 LogicalResult MaskedStoreOp::fold(ArrayRef<Attribute> operands,
4056 SmallVectorImpl<OpFoldResult> &results) {
4057 return foldMemRefCast(*this);
4058 }
4059
4060 //===----------------------------------------------------------------------===//
4061 // GatherOp
4062 //===----------------------------------------------------------------------===//
4063
verify()4064 LogicalResult GatherOp::verify() {
4065 VectorType indVType = getIndexVectorType();
4066 VectorType maskVType = getMaskVectorType();
4067 VectorType resVType = getVectorType();
4068 MemRefType memType = getMemRefType();
4069
4070 if (resVType.getElementType() != memType.getElementType())
4071 return emitOpError("base and result element type should match");
4072 if (llvm::size(getIndices()) != memType.getRank())
4073 return emitOpError("requires ") << memType.getRank() << " indices";
4074 if (resVType.getDimSize(0) != indVType.getDimSize(0))
4075 return emitOpError("expected result dim to match indices dim");
4076 if (resVType.getDimSize(0) != maskVType.getDimSize(0))
4077 return emitOpError("expected result dim to match mask dim");
4078 if (resVType != getPassThruVectorType())
4079 return emitOpError("expected pass_thru of same type as result type");
4080 return success();
4081 }
4082
4083 namespace {
4084 class GatherFolder final : public OpRewritePattern<GatherOp> {
4085 public:
4086 using OpRewritePattern<GatherOp>::OpRewritePattern;
matchAndRewrite(GatherOp gather,PatternRewriter & rewriter) const4087 LogicalResult matchAndRewrite(GatherOp gather,
4088 PatternRewriter &rewriter) const override {
4089 switch (get1DMaskFormat(gather.getMask())) {
4090 case MaskFormat::AllTrue:
4091 return failure(); // no unmasked equivalent
4092 case MaskFormat::AllFalse:
4093 rewriter.replaceOp(gather, gather.getPassThru());
4094 return success();
4095 case MaskFormat::Unknown:
4096 return failure();
4097 }
4098 llvm_unreachable("Unexpected 1DMaskFormat on GatherFolder");
4099 }
4100 };
4101 } // namespace
4102
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)4103 void GatherOp::getCanonicalizationPatterns(RewritePatternSet &results,
4104 MLIRContext *context) {
4105 results.add<GatherFolder>(context);
4106 }
4107
4108 //===----------------------------------------------------------------------===//
4109 // ScatterOp
4110 //===----------------------------------------------------------------------===//
4111
verify()4112 LogicalResult ScatterOp::verify() {
4113 VectorType indVType = getIndexVectorType();
4114 VectorType maskVType = getMaskVectorType();
4115 VectorType valueVType = getVectorType();
4116 MemRefType memType = getMemRefType();
4117
4118 if (valueVType.getElementType() != memType.getElementType())
4119 return emitOpError("base and valueToStore element type should match");
4120 if (llvm::size(getIndices()) != memType.getRank())
4121 return emitOpError("requires ") << memType.getRank() << " indices";
4122 if (valueVType.getDimSize(0) != indVType.getDimSize(0))
4123 return emitOpError("expected valueToStore dim to match indices dim");
4124 if (valueVType.getDimSize(0) != maskVType.getDimSize(0))
4125 return emitOpError("expected valueToStore dim to match mask dim");
4126 return success();
4127 }
4128
4129 namespace {
4130 class ScatterFolder final : public OpRewritePattern<ScatterOp> {
4131 public:
4132 using OpRewritePattern<ScatterOp>::OpRewritePattern;
matchAndRewrite(ScatterOp scatter,PatternRewriter & rewriter) const4133 LogicalResult matchAndRewrite(ScatterOp scatter,
4134 PatternRewriter &rewriter) const override {
4135 switch (get1DMaskFormat(scatter.getMask())) {
4136 case MaskFormat::AllTrue:
4137 return failure(); // no unmasked equivalent
4138 case MaskFormat::AllFalse:
4139 rewriter.eraseOp(scatter);
4140 return success();
4141 case MaskFormat::Unknown:
4142 return failure();
4143 }
4144 llvm_unreachable("Unexpected 1DMaskFormat on ScatterFolder");
4145 }
4146 };
4147 } // namespace
4148
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)4149 void ScatterOp::getCanonicalizationPatterns(RewritePatternSet &results,
4150 MLIRContext *context) {
4151 results.add<ScatterFolder>(context);
4152 }
4153
4154 //===----------------------------------------------------------------------===//
4155 // ExpandLoadOp
4156 //===----------------------------------------------------------------------===//
4157
verify()4158 LogicalResult ExpandLoadOp::verify() {
4159 VectorType maskVType = getMaskVectorType();
4160 VectorType passVType = getPassThruVectorType();
4161 VectorType resVType = getVectorType();
4162 MemRefType memType = getMemRefType();
4163
4164 if (resVType.getElementType() != memType.getElementType())
4165 return emitOpError("base and result element type should match");
4166 if (llvm::size(getIndices()) != memType.getRank())
4167 return emitOpError("requires ") << memType.getRank() << " indices";
4168 if (resVType.getDimSize(0) != maskVType.getDimSize(0))
4169 return emitOpError("expected result dim to match mask dim");
4170 if (resVType != passVType)
4171 return emitOpError("expected pass_thru of same type as result type");
4172 return success();
4173 }
4174
4175 namespace {
4176 class ExpandLoadFolder final : public OpRewritePattern<ExpandLoadOp> {
4177 public:
4178 using OpRewritePattern<ExpandLoadOp>::OpRewritePattern;
matchAndRewrite(ExpandLoadOp expand,PatternRewriter & rewriter) const4179 LogicalResult matchAndRewrite(ExpandLoadOp expand,
4180 PatternRewriter &rewriter) const override {
4181 switch (get1DMaskFormat(expand.getMask())) {
4182 case MaskFormat::AllTrue:
4183 rewriter.replaceOpWithNewOp<vector::LoadOp>(
4184 expand, expand.getType(), expand.getBase(), expand.getIndices());
4185 return success();
4186 case MaskFormat::AllFalse:
4187 rewriter.replaceOp(expand, expand.getPassThru());
4188 return success();
4189 case MaskFormat::Unknown:
4190 return failure();
4191 }
4192 llvm_unreachable("Unexpected 1DMaskFormat on ExpandLoadFolder");
4193 }
4194 };
4195 } // namespace
4196
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)4197 void ExpandLoadOp::getCanonicalizationPatterns(RewritePatternSet &results,
4198 MLIRContext *context) {
4199 results.add<ExpandLoadFolder>(context);
4200 }
4201
4202 //===----------------------------------------------------------------------===//
4203 // CompressStoreOp
4204 //===----------------------------------------------------------------------===//
4205
verify()4206 LogicalResult CompressStoreOp::verify() {
4207 VectorType maskVType = getMaskVectorType();
4208 VectorType valueVType = getVectorType();
4209 MemRefType memType = getMemRefType();
4210
4211 if (valueVType.getElementType() != memType.getElementType())
4212 return emitOpError("base and valueToStore element type should match");
4213 if (llvm::size(getIndices()) != memType.getRank())
4214 return emitOpError("requires ") << memType.getRank() << " indices";
4215 if (valueVType.getDimSize(0) != maskVType.getDimSize(0))
4216 return emitOpError("expected valueToStore dim to match mask dim");
4217 return success();
4218 }
4219
4220 namespace {
4221 class CompressStoreFolder final : public OpRewritePattern<CompressStoreOp> {
4222 public:
4223 using OpRewritePattern<CompressStoreOp>::OpRewritePattern;
matchAndRewrite(CompressStoreOp compress,PatternRewriter & rewriter) const4224 LogicalResult matchAndRewrite(CompressStoreOp compress,
4225 PatternRewriter &rewriter) const override {
4226 switch (get1DMaskFormat(compress.getMask())) {
4227 case MaskFormat::AllTrue:
4228 rewriter.replaceOpWithNewOp<vector::StoreOp>(
4229 compress, compress.getValueToStore(), compress.getBase(),
4230 compress.getIndices());
4231 return success();
4232 case MaskFormat::AllFalse:
4233 rewriter.eraseOp(compress);
4234 return success();
4235 case MaskFormat::Unknown:
4236 return failure();
4237 }
4238 llvm_unreachable("Unexpected 1DMaskFormat on CompressStoreFolder");
4239 }
4240 };
4241 } // namespace
4242
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)4243 void CompressStoreOp::getCanonicalizationPatterns(RewritePatternSet &results,
4244 MLIRContext *context) {
4245 results.add<CompressStoreFolder>(context);
4246 }
4247
4248 //===----------------------------------------------------------------------===//
4249 // ShapeCastOp
4250 //===----------------------------------------------------------------------===//
4251
4252 /// Returns true if each element of 'a' is equal to the product of a contiguous
4253 /// sequence of the elements of 'b'. Returns false otherwise.
isValidShapeCast(ArrayRef<int64_t> a,ArrayRef<int64_t> b)4254 static bool isValidShapeCast(ArrayRef<int64_t> a, ArrayRef<int64_t> b) {
4255 unsigned rankA = a.size();
4256 unsigned rankB = b.size();
4257 assert(rankA < rankB);
4258
4259 unsigned i = 0;
4260 unsigned j = 0;
4261 while (i < rankA && j < rankB) {
4262 int64_t dimA = a[i];
4263 int64_t dimB = 1;
4264 while (dimB < dimA && j < rankB)
4265 dimB *= b[j++];
4266 if (dimA != dimB)
4267 break;
4268 ++i;
4269
4270 // Handle the case when trailing dimensions are of size 1.
4271 // Include them into the contiguous sequence.
4272 auto isOne = [](int64_t v) { return v == 1; };
4273 if (i < rankA && llvm::all_of(a.slice(i), isOne))
4274 i = rankA;
4275 if (j < rankB && llvm::all_of(b.slice(j), isOne))
4276 j = rankB;
4277 }
4278
4279 return i == rankA && j == rankB;
4280 }
4281
verifyVectorShapeCast(Operation * op,VectorType sourceVectorType,VectorType resultVectorType)4282 static LogicalResult verifyVectorShapeCast(Operation *op,
4283 VectorType sourceVectorType,
4284 VectorType resultVectorType) {
4285 // Check that element type is the same.
4286 if (sourceVectorType.getElementType() != resultVectorType.getElementType())
4287 return op->emitOpError("source/result vectors must have same element type");
4288 auto sourceShape = sourceVectorType.getShape();
4289 auto resultShape = resultVectorType.getShape();
4290
4291 // Check that product of source dim sizes matches product of result dim sizes.
4292 int64_t sourceDimProduct = std::accumulate(
4293 sourceShape.begin(), sourceShape.end(), 1LL, std::multiplies<int64_t>{});
4294 int64_t resultDimProduct = std::accumulate(
4295 resultShape.begin(), resultShape.end(), 1LL, std::multiplies<int64_t>{});
4296 if (sourceDimProduct != resultDimProduct)
4297 return op->emitOpError("source/result number of elements must match");
4298
4299 // Check that expanding/contracting rank cases.
4300 unsigned sourceRank = sourceVectorType.getRank();
4301 unsigned resultRank = resultVectorType.getRank();
4302 if (sourceRank < resultRank) {
4303 if (!isValidShapeCast(sourceShape, resultShape))
4304 return op->emitOpError("invalid shape cast");
4305 } else if (sourceRank > resultRank) {
4306 if (!isValidShapeCast(resultShape, sourceShape))
4307 return op->emitOpError("invalid shape cast");
4308 }
4309 return success();
4310 }
4311
verify()4312 LogicalResult ShapeCastOp::verify() {
4313 auto sourceVectorType = getSource().getType().dyn_cast_or_null<VectorType>();
4314 auto resultVectorType = getResult().getType().dyn_cast_or_null<VectorType>();
4315
4316 // Check if source/result are of vector type.
4317 if (sourceVectorType && resultVectorType)
4318 return verifyVectorShapeCast(*this, sourceVectorType, resultVectorType);
4319
4320 return success();
4321 }
4322
fold(ArrayRef<Attribute> operands)4323 OpFoldResult ShapeCastOp::fold(ArrayRef<Attribute> operands) {
4324 // No-op shape cast.
4325 if (getSource().getType() == getResult().getType())
4326 return getSource();
4327
4328 // Canceling shape casts.
4329 if (auto otherOp = getSource().getDefiningOp<ShapeCastOp>()) {
4330 if (getResult().getType() == otherOp.getSource().getType())
4331 return otherOp.getSource();
4332
4333 // Only allows valid transitive folding.
4334 VectorType srcType = otherOp.getSource().getType().cast<VectorType>();
4335 VectorType resultType = getResult().getType().cast<VectorType>();
4336 if (srcType.getRank() < resultType.getRank()) {
4337 if (!isValidShapeCast(srcType.getShape(), resultType.getShape()))
4338 return {};
4339 } else if (srcType.getRank() > resultType.getRank()) {
4340 if (!isValidShapeCast(resultType.getShape(), srcType.getShape()))
4341 return {};
4342 } else {
4343 return {};
4344 }
4345
4346 setOperand(otherOp.getSource());
4347 return getResult();
4348 }
4349
4350 // Cancelling broadcast and shape cast ops.
4351 if (auto bcastOp = getSource().getDefiningOp<BroadcastOp>()) {
4352 if (bcastOp.getSourceType() == getType())
4353 return bcastOp.getSource();
4354 }
4355
4356 return {};
4357 }
4358
4359 namespace {
4360 // Pattern to rewrite a ShapeCast(splat ConstantOp) -> ConstantOp.
4361 class ShapeCastConstantFolder final : public OpRewritePattern<ShapeCastOp> {
4362 public:
4363 using OpRewritePattern<ShapeCastOp>::OpRewritePattern;
4364
matchAndRewrite(ShapeCastOp shapeCastOp,PatternRewriter & rewriter) const4365 LogicalResult matchAndRewrite(ShapeCastOp shapeCastOp,
4366 PatternRewriter &rewriter) const override {
4367 auto constantOp =
4368 shapeCastOp.getSource().getDefiningOp<arith::ConstantOp>();
4369 if (!constantOp)
4370 return failure();
4371 // Only handle splat for now.
4372 auto dense = constantOp.getValue().dyn_cast<SplatElementsAttr>();
4373 if (!dense)
4374 return failure();
4375 auto newAttr =
4376 DenseElementsAttr::get(shapeCastOp.getType().cast<VectorType>(),
4377 dense.getSplatValue<Attribute>());
4378 rewriter.replaceOpWithNewOp<arith::ConstantOp>(shapeCastOp, newAttr);
4379 return success();
4380 }
4381 };
4382
4383 /// Pattern to rewrite a ShapeCast(Broadcast) -> Broadcast.
4384 /// This only applies when the shape of the broadcast source is a suffix of the
4385 /// shape of the result (i.e. when broadcast without reshape is expressive
4386 /// enough to capture the result in a single op).
4387 class ShapeCastBroadcastFolder final : public OpRewritePattern<ShapeCastOp> {
4388 public:
4389 using OpRewritePattern<ShapeCastOp>::OpRewritePattern;
4390
matchAndRewrite(ShapeCastOp shapeCastOp,PatternRewriter & rewriter) const4391 LogicalResult matchAndRewrite(ShapeCastOp shapeCastOp,
4392 PatternRewriter &rewriter) const override {
4393 auto broadcastOp =
4394 shapeCastOp.getSource().getDefiningOp<vector::BroadcastOp>();
4395 if (!broadcastOp)
4396 return failure();
4397
4398 auto broadcastSourceVectorType =
4399 broadcastOp.getSourceType().dyn_cast<VectorType>();
4400 auto broadcastSourceShape = broadcastSourceVectorType
4401 ? broadcastSourceVectorType.getShape()
4402 : ArrayRef<int64_t>{};
4403 auto shapeCastTargetShape = shapeCastOp.getResultVectorType().getShape();
4404
4405 // Bail if `broadcastSourceShape` is not a suffix of the result.
4406 bool isSuffix = (broadcastSourceShape == shapeCastTargetShape.take_back(
4407 broadcastSourceShape.size()));
4408 if (!isSuffix)
4409 return failure();
4410
4411 rewriter.replaceOpWithNewOp<vector::BroadcastOp>(
4412 shapeCastOp, shapeCastOp.getResultVectorType(),
4413 broadcastOp.getSource());
4414 return success();
4415 }
4416 };
4417
4418 } // namespace
4419
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)4420 void ShapeCastOp::getCanonicalizationPatterns(RewritePatternSet &results,
4421 MLIRContext *context) {
4422 results.add<ShapeCastConstantFolder, ShapeCastBroadcastFolder>(context);
4423 }
4424
4425 //===----------------------------------------------------------------------===//
4426 // VectorBitCastOp
4427 //===----------------------------------------------------------------------===//
4428
verify()4429 LogicalResult BitCastOp::verify() {
4430 auto sourceVectorType = getSourceVectorType();
4431 auto resultVectorType = getResultVectorType();
4432
4433 for (int64_t i = 0, e = sourceVectorType.getRank() - 1; i < e; i++) {
4434 if (sourceVectorType.getDimSize(i) != resultVectorType.getDimSize(i))
4435 return emitOpError("dimension size mismatch at: ") << i;
4436 }
4437
4438 DataLayout dataLayout = DataLayout::closest(*this);
4439 auto sourceElementBits =
4440 dataLayout.getTypeSizeInBits(sourceVectorType.getElementType());
4441 auto resultElementBits =
4442 dataLayout.getTypeSizeInBits(resultVectorType.getElementType());
4443
4444 if (sourceVectorType.getRank() == 0) {
4445 if (sourceElementBits != resultElementBits)
4446 return emitOpError("source/result bitwidth of the 0-D vector element "
4447 "types must be equal");
4448 } else if (sourceElementBits * sourceVectorType.getShape().back() !=
4449 resultElementBits * resultVectorType.getShape().back()) {
4450 return emitOpError(
4451 "source/result bitwidth of the minor 1-D vectors must be equal");
4452 }
4453
4454 return success();
4455 }
4456
fold(ArrayRef<Attribute> operands)4457 OpFoldResult BitCastOp::fold(ArrayRef<Attribute> operands) {
4458 // Nop cast.
4459 if (getSource().getType() == getResult().getType())
4460 return getSource();
4461
4462 // Canceling bitcasts.
4463 if (auto otherOp = getSource().getDefiningOp<BitCastOp>()) {
4464 if (getResult().getType() == otherOp.getSource().getType())
4465 return otherOp.getSource();
4466
4467 setOperand(otherOp.getSource());
4468 return getResult();
4469 }
4470
4471 Attribute sourceConstant = operands.front();
4472 if (!sourceConstant)
4473 return {};
4474
4475 Type srcElemType = getSourceVectorType().getElementType();
4476 Type dstElemType = getResultVectorType().getElementType();
4477
4478 if (auto floatPack = sourceConstant.dyn_cast<DenseFPElementsAttr>()) {
4479 if (floatPack.isSplat()) {
4480 auto splat = floatPack.getSplatValue<FloatAttr>();
4481
4482 // Casting fp16 into fp32.
4483 if (srcElemType.isF16() && dstElemType.isF32()) {
4484 uint32_t bits = static_cast<uint32_t>(
4485 splat.getValue().bitcastToAPInt().getZExtValue());
4486 // Duplicate the 16-bit pattern.
4487 bits = (bits << 16) | (bits & 0xffff);
4488 APInt intBits(32, bits);
4489 APFloat floatBits(llvm::APFloat::IEEEsingle(), intBits);
4490 return DenseElementsAttr::get(getResultVectorType(), floatBits);
4491 }
4492 }
4493 }
4494
4495 return {};
4496 }
4497
4498 //===----------------------------------------------------------------------===//
4499 // TypeCastOp
4500 //===----------------------------------------------------------------------===//
4501
extractShape(MemRefType memRefType)4502 static SmallVector<int64_t, 8> extractShape(MemRefType memRefType) {
4503 auto vectorType = memRefType.getElementType().dyn_cast<VectorType>();
4504 SmallVector<int64_t, 8> res(memRefType.getShape().begin(),
4505 memRefType.getShape().end());
4506 if (vectorType)
4507 res.append(vectorType.getShape().begin(), vectorType.getShape().end());
4508 return res;
4509 }
4510
4511 /// Build the canonical memRefType with a single vector.
4512 /// E.g. memref<4 x 5 x vector<6 x f32>> -> memref<vector<4 x 5 x 6 x f32>>.
build(OpBuilder & builder,OperationState & result,Value source)4513 void TypeCastOp::build(OpBuilder &builder, OperationState &result,
4514 Value source) {
4515 result.addOperands(source);
4516 MemRefType memRefType = source.getType().cast<MemRefType>();
4517 VectorType vectorType =
4518 VectorType::get(extractShape(memRefType),
4519 getElementTypeOrSelf(getElementTypeOrSelf(memRefType)));
4520 result.addTypes(MemRefType::get({}, vectorType, MemRefLayoutAttrInterface(),
4521 memRefType.getMemorySpace()));
4522 }
4523
verify()4524 LogicalResult TypeCastOp::verify() {
4525 MemRefType canonicalType = canonicalizeStridedLayout(getMemRefType());
4526 if (!canonicalType.getLayout().isIdentity())
4527 return emitOpError("expects operand to be a memref with identity layout");
4528 if (!getResultMemRefType().getLayout().isIdentity())
4529 return emitOpError("expects result to be a memref with identity layout");
4530 if (getResultMemRefType().getMemorySpace() !=
4531 getMemRefType().getMemorySpace())
4532 return emitOpError("expects result in same memory space");
4533
4534 auto sourceType = getMemRefType();
4535 auto resultType = getResultMemRefType();
4536 if (getElementTypeOrSelf(getElementTypeOrSelf(sourceType)) !=
4537 getElementTypeOrSelf(getElementTypeOrSelf(resultType)))
4538 return emitOpError(
4539 "expects result and operand with same underlying scalar type: ")
4540 << resultType;
4541 if (extractShape(sourceType) != extractShape(resultType))
4542 return emitOpError(
4543 "expects concatenated result and operand shapes to be equal: ")
4544 << resultType;
4545 return success();
4546 }
4547
4548 //===----------------------------------------------------------------------===//
4549 // TransposeOp
4550 //===----------------------------------------------------------------------===//
4551
build(OpBuilder & builder,OperationState & result,Value vector,ArrayRef<int64_t> transp)4552 void vector::TransposeOp::build(OpBuilder &builder, OperationState &result,
4553 Value vector, ArrayRef<int64_t> transp) {
4554 VectorType vt = vector.getType().cast<VectorType>();
4555 SmallVector<int64_t, 4> transposedShape(vt.getRank());
4556 for (unsigned i = 0; i < transp.size(); ++i)
4557 transposedShape[i] = vt.getShape()[transp[i]];
4558
4559 result.addOperands(vector);
4560 result.addTypes(VectorType::get(transposedShape, vt.getElementType()));
4561 result.addAttribute(getTranspAttrStrName(), builder.getI64ArrayAttr(transp));
4562 }
4563
fold(ArrayRef<Attribute> operands)4564 OpFoldResult vector::TransposeOp::fold(ArrayRef<Attribute> operands) {
4565 // Eliminate splat constant transpose ops.
4566 if (auto attr = operands.front().dyn_cast_or_null<DenseElementsAttr>())
4567 if (attr.isSplat())
4568 return attr.reshape(getResultType());
4569
4570 // Eliminate identity transpose ops. This happens when the dimensions of the
4571 // input vector remain in their original order after the transpose operation.
4572 SmallVector<int64_t, 4> transp;
4573 getTransp(transp);
4574
4575 // Check if the permutation of the dimensions contains sequential values:
4576 // {0, 1, 2, ...}.
4577 for (int64_t i = 0, e = transp.size(); i < e; i++) {
4578 if (transp[i] != i)
4579 return {};
4580 }
4581
4582 return getVector();
4583 }
4584
verify()4585 LogicalResult vector::TransposeOp::verify() {
4586 VectorType vectorType = getVectorType();
4587 VectorType resultType = getResultType();
4588 int64_t rank = resultType.getRank();
4589 if (vectorType.getRank() != rank)
4590 return emitOpError("vector result rank mismatch: ") << rank;
4591 // Verify transposition array.
4592 auto transpAttr = getTransp().getValue();
4593 int64_t size = transpAttr.size();
4594 if (rank != size)
4595 return emitOpError("transposition length mismatch: ") << size;
4596 SmallVector<bool, 8> seen(rank, false);
4597 for (const auto &ta : llvm::enumerate(transpAttr)) {
4598 int64_t i = ta.value().cast<IntegerAttr>().getInt();
4599 if (i < 0 || i >= rank)
4600 return emitOpError("transposition index out of range: ") << i;
4601 if (seen[i])
4602 return emitOpError("duplicate position index: ") << i;
4603 seen[i] = true;
4604 if (resultType.getDimSize(ta.index()) != vectorType.getDimSize(i))
4605 return emitOpError("dimension size mismatch at: ") << i;
4606 }
4607 return success();
4608 }
4609
getShapeForUnroll()4610 Optional<SmallVector<int64_t, 4>> TransposeOp::getShapeForUnroll() {
4611 return llvm::to_vector<4>(getResultType().getShape());
4612 }
4613
4614 namespace {
4615
4616 // Rewrites two back-to-back TransposeOp operations into a single TransposeOp.
4617 class TransposeFolder final : public OpRewritePattern<vector::TransposeOp> {
4618 public:
4619 using OpRewritePattern<vector::TransposeOp>::OpRewritePattern;
4620
matchAndRewrite(vector::TransposeOp transposeOp,PatternRewriter & rewriter) const4621 LogicalResult matchAndRewrite(vector::TransposeOp transposeOp,
4622 PatternRewriter &rewriter) const override {
4623 // Wrapper around vector::TransposeOp::getTransp() for cleaner code.
4624 auto getPermutation = [](vector::TransposeOp transpose) {
4625 SmallVector<int64_t, 4> permutation;
4626 transpose.getTransp(permutation);
4627 return permutation;
4628 };
4629
4630 // Composes two permutations: result[i] = permutation1[permutation2[i]].
4631 auto composePermutations = [](ArrayRef<int64_t> permutation1,
4632 ArrayRef<int64_t> permutation2) {
4633 SmallVector<int64_t, 4> result;
4634 for (auto index : permutation2)
4635 result.push_back(permutation1[index]);
4636 return result;
4637 };
4638
4639 // Return if the input of 'transposeOp' is not defined by another transpose.
4640 vector::TransposeOp parentTransposeOp =
4641 transposeOp.getVector().getDefiningOp<vector::TransposeOp>();
4642 if (!parentTransposeOp)
4643 return failure();
4644
4645 SmallVector<int64_t, 4> permutation = composePermutations(
4646 getPermutation(parentTransposeOp), getPermutation(transposeOp));
4647 // Replace 'transposeOp' with a new transpose operation.
4648 rewriter.replaceOpWithNewOp<vector::TransposeOp>(
4649 transposeOp, transposeOp.getResult().getType(),
4650 parentTransposeOp.getVector(),
4651 vector::getVectorSubscriptAttr(rewriter, permutation));
4652 return success();
4653 }
4654 };
4655
4656 // Folds transpose(broadcast(<scalar>)) into brodcast(<scalar>).
4657 struct FoldTransposedScalarBroadcast final
4658 : public OpRewritePattern<vector::TransposeOp> {
4659 using OpRewritePattern::OpRewritePattern;
4660
matchAndRewrite__anon088a7a4f2b11::FoldTransposedScalarBroadcast4661 LogicalResult matchAndRewrite(vector::TransposeOp transposeOp,
4662 PatternRewriter &rewriter) const override {
4663 auto bcastOp = transposeOp.getVector().getDefiningOp<vector::BroadcastOp>();
4664 if (!bcastOp)
4665 return failure();
4666
4667 auto srcVectorType = bcastOp.getSourceType().dyn_cast<VectorType>();
4668 if (!srcVectorType || srcVectorType.getNumElements() == 1) {
4669 rewriter.replaceOpWithNewOp<vector::BroadcastOp>(
4670 transposeOp, transposeOp.getResultType(), bcastOp.getSource());
4671 return success();
4672 }
4673
4674 return failure();
4675 }
4676 };
4677
4678 // Folds transpose(splat x : src_type) : res_type into splat x : res_type.
4679 class FoldTransposeSplat final : public OpRewritePattern<TransposeOp> {
4680 public:
4681 using OpRewritePattern<TransposeOp>::OpRewritePattern;
4682
matchAndRewrite(TransposeOp transposeOp,PatternRewriter & rewriter) const4683 LogicalResult matchAndRewrite(TransposeOp transposeOp,
4684 PatternRewriter &rewriter) const override {
4685 auto splatOp = transposeOp.getVector().getDefiningOp<vector::SplatOp>();
4686 if (!splatOp)
4687 return failure();
4688
4689 rewriter.replaceOpWithNewOp<vector::SplatOp>(
4690 transposeOp, transposeOp.getResultType(), splatOp.getInput());
4691 return success();
4692 }
4693 };
4694
4695 } // namespace
4696
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)4697 void vector::TransposeOp::getCanonicalizationPatterns(
4698 RewritePatternSet &results, MLIRContext *context) {
4699 results
4700 .add<FoldTransposedScalarBroadcast, TransposeFolder, FoldTransposeSplat>(
4701 context);
4702 }
4703
getTransp(SmallVectorImpl<int64_t> & results)4704 void vector::TransposeOp::getTransp(SmallVectorImpl<int64_t> &results) {
4705 populateFromInt64AttrArray(getTransp(), results);
4706 }
4707
4708 //===----------------------------------------------------------------------===//
4709 // ConstantMaskOp
4710 //===----------------------------------------------------------------------===//
4711
verify()4712 LogicalResult ConstantMaskOp::verify() {
4713 auto resultType = getResult().getType().cast<VectorType>();
4714 // Check the corner case of 0-D vectors first.
4715 if (resultType.getRank() == 0) {
4716 if (getMaskDimSizes().size() != 1)
4717 return emitError("array attr must have length 1 for 0-D vectors");
4718 auto dim = getMaskDimSizes()[0].cast<IntegerAttr>().getInt();
4719 if (dim != 0 && dim != 1)
4720 return emitError("mask dim size must be either 0 or 1 for 0-D vectors");
4721 return success();
4722 }
4723
4724 // Verify that array attr size matches the rank of the vector result.
4725 if (static_cast<int64_t>(getMaskDimSizes().size()) != resultType.getRank())
4726 return emitOpError(
4727 "must specify array attr of size equal vector result rank");
4728 // Verify that each array attr element is in bounds of corresponding vector
4729 // result dimension size.
4730 auto resultShape = resultType.getShape();
4731 SmallVector<int64_t, 4> maskDimSizes;
4732 for (const auto &it : llvm::enumerate(getMaskDimSizes())) {
4733 int64_t attrValue = it.value().cast<IntegerAttr>().getInt();
4734 if (attrValue < 0 || attrValue > resultShape[it.index()])
4735 return emitOpError(
4736 "array attr of size out of bounds of vector result dimension size");
4737 maskDimSizes.push_back(attrValue);
4738 }
4739 // Verify that if one mask dim size is zero, they all should be zero (because
4740 // the mask region is a conjunction of each mask dimension interval).
4741 bool anyZeros = llvm::is_contained(maskDimSizes, 0);
4742 bool allZeros = llvm::all_of(maskDimSizes, [](int64_t s) { return s == 0; });
4743 if (anyZeros && !allZeros)
4744 return emitOpError("expected all mask dim sizes to be zeros, "
4745 "as a result of conjunction with zero mask dim");
4746 // Verify that if the mask type is scalable, dimensions should be zero because
4747 // constant scalable masks can only be defined for the "none set" or "all set"
4748 // cases, and there is no VLA way to define an "all set" case for
4749 // `vector.constant_mask`. In the future, a convention could be established
4750 // to decide if a specific dimension value could be considered as "all set".
4751 if (resultType.isScalable() &&
4752 getMaskDimSizes()[0].cast<IntegerAttr>().getInt() != 0)
4753 return emitOpError("expected mask dim sizes for scalable masks to be 0");
4754 return success();
4755 }
4756
4757 //===----------------------------------------------------------------------===//
4758 // CreateMaskOp
4759 //===----------------------------------------------------------------------===//
4760
verify()4761 LogicalResult CreateMaskOp::verify() {
4762 auto vectorType = getResult().getType().cast<VectorType>();
4763 // Verify that an operand was specified for each result vector each dimension.
4764 if (vectorType.getRank() == 0) {
4765 if (getNumOperands() != 1)
4766 return emitOpError(
4767 "must specify exactly one operand for 0-D create_mask");
4768 } else if (getNumOperands() !=
4769 getResult().getType().cast<VectorType>().getRank()) {
4770 return emitOpError(
4771 "must specify an operand for each result vector dimension");
4772 }
4773 return success();
4774 }
4775
4776 namespace {
4777
4778 // Pattern to rewrite a CreateMaskOp with a ConstantMaskOp.
4779 class CreateMaskFolder final : public OpRewritePattern<CreateMaskOp> {
4780 public:
4781 using OpRewritePattern<CreateMaskOp>::OpRewritePattern;
4782
matchAndRewrite(CreateMaskOp createMaskOp,PatternRewriter & rewriter) const4783 LogicalResult matchAndRewrite(CreateMaskOp createMaskOp,
4784 PatternRewriter &rewriter) const override {
4785 // Return if any of 'createMaskOp' operands are not defined by a constant.
4786 auto isNotDefByConstant = [](Value operand) {
4787 return !isa_and_nonnull<arith::ConstantIndexOp>(operand.getDefiningOp());
4788 };
4789 if (llvm::any_of(createMaskOp.operands(), isNotDefByConstant))
4790 return failure();
4791
4792 // CreateMaskOp for scalable vectors can be folded only if all dimensions
4793 // are negative or zero.
4794 if (auto vType = createMaskOp.getType().dyn_cast<VectorType>()) {
4795 if (vType.isScalable())
4796 for (auto opDim : createMaskOp.getOperands()) {
4797 APInt intVal;
4798 if (matchPattern(opDim, m_ConstantInt(&intVal)) &&
4799 intVal.isStrictlyPositive())
4800 return failure();
4801 }
4802 }
4803
4804 // Gather constant mask dimension sizes.
4805 SmallVector<int64_t, 4> maskDimSizes;
4806 for (auto it : llvm::zip(createMaskOp.operands(),
4807 createMaskOp.getType().getShape())) {
4808 auto *defOp = std::get<0>(it).getDefiningOp();
4809 int64_t maxDimSize = std::get<1>(it);
4810 int64_t dimSize = cast<arith::ConstantIndexOp>(defOp).value();
4811 dimSize = std::min(dimSize, maxDimSize);
4812 // If one of dim sizes is zero, set all dims to zero.
4813 if (dimSize <= 0) {
4814 maskDimSizes.assign(createMaskOp.getType().getRank(), 0);
4815 break;
4816 }
4817 maskDimSizes.push_back(dimSize);
4818 }
4819 // Replace 'createMaskOp' with ConstantMaskOp.
4820 rewriter.replaceOpWithNewOp<ConstantMaskOp>(
4821 createMaskOp, createMaskOp.getResult().getType(),
4822 vector::getVectorSubscriptAttr(rewriter, maskDimSizes));
4823 return success();
4824 }
4825 };
4826
4827 } // namespace
4828
getCanonicalizationPatterns(RewritePatternSet & results,MLIRContext * context)4829 void CreateMaskOp::getCanonicalizationPatterns(RewritePatternSet &results,
4830 MLIRContext *context) {
4831 results.add<CreateMaskFolder>(context);
4832 }
4833
4834 //===----------------------------------------------------------------------===//
4835 // ScanOp
4836 //===----------------------------------------------------------------------===//
4837
verify()4838 LogicalResult ScanOp::verify() {
4839 VectorType srcType = getSourceType();
4840 VectorType initialType = getInitialValueType();
4841 // Check reduction dimension < rank.
4842 int64_t srcRank = srcType.getRank();
4843 int64_t reductionDim = getReductionDim();
4844 if (reductionDim >= srcRank)
4845 return emitOpError("reduction dimension ")
4846 << reductionDim << " has to be less than " << srcRank;
4847
4848 // Check that rank(initial_value) = rank(src) - 1.
4849 int64_t initialValueRank = initialType.getRank();
4850 if (initialValueRank != srcRank - 1)
4851 return emitOpError("initial value rank ")
4852 << initialValueRank << " has to be equal to " << srcRank - 1;
4853
4854 // Check shapes of initial value and src.
4855 ArrayRef<int64_t> srcShape = srcType.getShape();
4856 ArrayRef<int64_t> initialValueShapes = initialType.getShape();
4857 SmallVector<int64_t> expectedShape;
4858 for (int i = 0; i < srcRank; i++) {
4859 if (i != reductionDim)
4860 expectedShape.push_back(srcShape[i]);
4861 }
4862 if (llvm::any_of(llvm::zip(initialValueShapes, expectedShape),
4863 [](std::tuple<int64_t, int64_t> s) {
4864 return std::get<0>(s) != std::get<1>(s);
4865 })) {
4866 return emitOpError("incompatible input/initial value shapes");
4867 }
4868
4869 // Verify supported reduction kind.
4870 Type eltType = getDestType().getElementType();
4871 if (!isSupportedCombiningKind(getKind(), eltType))
4872 return emitOpError("unsupported reduction type ")
4873 << eltType << " for kind '" << stringifyCombiningKind(getKind())
4874 << "'";
4875
4876 return success();
4877 }
4878
populateVectorToVectorCanonicalizationPatterns(RewritePatternSet & patterns)4879 void mlir::vector::populateVectorToVectorCanonicalizationPatterns(
4880 RewritePatternSet &patterns) {
4881 patterns
4882 .add<CreateMaskFolder, MaskedLoadFolder, MaskedStoreFolder, GatherFolder,
4883 ScatterFolder, ExpandLoadFolder, CompressStoreFolder,
4884 StridedSliceConstantMaskFolder, TransposeFolder>(
4885 patterns.getContext());
4886 }
4887
4888 //===----------------------------------------------------------------------===//
4889 // SplatOp
4890 //===----------------------------------------------------------------------===//
4891
fold(ArrayRef<Attribute> operands)4892 OpFoldResult SplatOp::fold(ArrayRef<Attribute> operands) {
4893 auto constOperand = operands.front();
4894 if (!constOperand.isa_and_nonnull<IntegerAttr, FloatAttr>())
4895 return {};
4896
4897 // SplatElementsAttr::get treats single value for second arg as being a splat.
4898 return SplatElementsAttr::get(getType(), {constOperand});
4899 }
4900
4901 //===----------------------------------------------------------------------===//
4902 // WarpExecuteOnLane0Op
4903 //===----------------------------------------------------------------------===//
4904
print(OpAsmPrinter & p)4905 void WarpExecuteOnLane0Op::print(OpAsmPrinter &p) {
4906 p << "(" << getLaneid() << ")";
4907
4908 SmallVector<StringRef> coreAttr = {getWarpSizeAttrName()};
4909 auto warpSizeAttr = getOperation()->getAttr(getWarpSizeAttrName());
4910 p << "[" << warpSizeAttr.cast<IntegerAttr>().getInt() << "]";
4911
4912 if (!getArgs().empty())
4913 p << " args(" << getArgs() << " : " << getArgs().getTypes() << ")";
4914 if (!getResults().empty())
4915 p << " -> (" << getResults().getTypes() << ')';
4916 p << " ";
4917 p.printRegion(getRegion(),
4918 /*printEntryBlockArgs=*/true,
4919 /*printBlockTerminators=*/!getResults().empty());
4920 p.printOptionalAttrDict(getOperation()->getAttrs(), coreAttr);
4921 }
4922
parse(OpAsmParser & parser,OperationState & result)4923 ParseResult WarpExecuteOnLane0Op::parse(OpAsmParser &parser,
4924 OperationState &result) {
4925 // Create the region.
4926 result.regions.reserve(1);
4927 Region *warpRegion = result.addRegion();
4928
4929 auto &builder = parser.getBuilder();
4930 OpAsmParser::UnresolvedOperand laneId;
4931
4932 // Parse predicate operand.
4933 if (parser.parseLParen() ||
4934 parser.parseOperand(laneId, /*allowResultNumber=*/false) ||
4935 parser.parseRParen())
4936 return failure();
4937
4938 int64_t warpSize;
4939 if (parser.parseLSquare() || parser.parseInteger(warpSize) ||
4940 parser.parseRSquare())
4941 return failure();
4942 result.addAttribute(getWarpSizeAttrName(OperationName(getOperationName(),
4943 builder.getContext())),
4944 builder.getI64IntegerAttr(warpSize));
4945
4946 if (parser.resolveOperand(laneId, builder.getIndexType(), result.operands))
4947 return failure();
4948
4949 llvm::SMLoc inputsOperandsLoc;
4950 SmallVector<OpAsmParser::UnresolvedOperand> inputsOperands;
4951 SmallVector<Type> inputTypes;
4952 if (succeeded(parser.parseOptionalKeyword("args"))) {
4953 if (parser.parseLParen())
4954 return failure();
4955
4956 inputsOperandsLoc = parser.getCurrentLocation();
4957 if (parser.parseOperandList(inputsOperands) ||
4958 parser.parseColonTypeList(inputTypes) || parser.parseRParen())
4959 return failure();
4960 }
4961 if (parser.resolveOperands(inputsOperands, inputTypes, inputsOperandsLoc,
4962 result.operands))
4963 return failure();
4964
4965 // Parse optional results type list.
4966 if (parser.parseOptionalArrowTypeList(result.types))
4967 return failure();
4968 // Parse the region.
4969 if (parser.parseRegion(*warpRegion, /*arguments=*/{},
4970 /*argTypes=*/{}))
4971 return failure();
4972 WarpExecuteOnLane0Op::ensureTerminator(*warpRegion, builder, result.location);
4973
4974 // Parse the optional attribute list.
4975 if (parser.parseOptionalAttrDict(result.attributes))
4976 return failure();
4977 return success();
4978 }
4979
getSuccessorRegions(Optional<unsigned> index,ArrayRef<Attribute> operands,SmallVectorImpl<RegionSuccessor> & regions)4980 void WarpExecuteOnLane0Op::getSuccessorRegions(
4981 Optional<unsigned> index, ArrayRef<Attribute> operands,
4982 SmallVectorImpl<RegionSuccessor> ®ions) {
4983 if (index) {
4984 regions.push_back(RegionSuccessor(getResults()));
4985 return;
4986 }
4987
4988 // The warp region is always executed
4989 regions.push_back(RegionSuccessor(&getWarpRegion()));
4990 }
4991
build(OpBuilder & builder,OperationState & result,TypeRange resultTypes,Value laneId,int64_t warpSize)4992 void WarpExecuteOnLane0Op::build(OpBuilder &builder, OperationState &result,
4993 TypeRange resultTypes, Value laneId,
4994 int64_t warpSize) {
4995 build(builder, result, resultTypes, laneId, warpSize,
4996 /*operands=*/llvm::None, /*argTypes=*/llvm::None);
4997 }
4998
build(OpBuilder & builder,OperationState & result,TypeRange resultTypes,Value laneId,int64_t warpSize,ValueRange args,TypeRange blockArgTypes)4999 void WarpExecuteOnLane0Op::build(OpBuilder &builder, OperationState &result,
5000 TypeRange resultTypes, Value laneId,
5001 int64_t warpSize, ValueRange args,
5002 TypeRange blockArgTypes) {
5003 result.addOperands(laneId);
5004 result.addAttribute(getAttributeNames()[0],
5005 builder.getI64IntegerAttr(warpSize));
5006 result.addTypes(resultTypes);
5007 result.addOperands(args);
5008 assert(args.size() == blockArgTypes.size());
5009 OpBuilder::InsertionGuard guard(builder);
5010 Region *warpRegion = result.addRegion();
5011 Block *block = builder.createBlock(warpRegion);
5012 for (auto it : llvm::zip(blockArgTypes, args))
5013 block->addArgument(std::get<0>(it), std::get<1>(it).getLoc());
5014 }
5015
5016 /// Helper check if the distributed vector type is consistent with the expanded
5017 /// type and distributed size.
verifyDistributedType(Type expanded,Type distributed,int64_t warpSize,Operation * op)5018 static LogicalResult verifyDistributedType(Type expanded, Type distributed,
5019 int64_t warpSize, Operation *op) {
5020 // If the types matches there is no distribution.
5021 if (expanded == distributed)
5022 return success();
5023 auto expandedVecType = expanded.dyn_cast<VectorType>();
5024 auto distributedVecType = distributed.dyn_cast<VectorType>();
5025 if (!expandedVecType || !distributedVecType)
5026 return op->emitOpError("expected vector type for distributed operands.");
5027 if (expandedVecType.getRank() != distributedVecType.getRank() ||
5028 expandedVecType.getElementType() != distributedVecType.getElementType())
5029 return op->emitOpError(
5030 "expected distributed vectors to have same rank and element type.");
5031 bool foundDistributedDim = false;
5032 for (int64_t i = 0, e = expandedVecType.getRank(); i < e; i++) {
5033 if (expandedVecType.getDimSize(i) == distributedVecType.getDimSize(i))
5034 continue;
5035 if (expandedVecType.getDimSize(i) ==
5036 distributedVecType.getDimSize(i) * warpSize) {
5037 if (foundDistributedDim)
5038 return op->emitOpError()
5039 << "expected only one dimension to be distributed from "
5040 << expandedVecType << " to " << distributedVecType;
5041 foundDistributedDim = true;
5042 continue;
5043 }
5044 return op->emitOpError() << "incompatible distribution dimensions from "
5045 << expandedVecType << " to " << distributedVecType;
5046 }
5047 return success();
5048 }
5049
verify()5050 LogicalResult WarpExecuteOnLane0Op::verify() {
5051 if (getArgs().size() != getWarpRegion().getNumArguments())
5052 return emitOpError(
5053 "expected same number op arguments and block arguments.");
5054 auto yield =
5055 cast<YieldOp>(getWarpRegion().getBlocks().begin()->getTerminator());
5056 if (yield.getNumOperands() != getNumResults())
5057 return emitOpError(
5058 "expected same number of yield operands and return values.");
5059 int64_t warpSize = getWarpSize();
5060 for (auto it : llvm::zip(getWarpRegion().getArguments(), getArgs())) {
5061 if (failed(verifyDistributedType(std::get<0>(it).getType(),
5062 std::get<1>(it).getType(), warpSize,
5063 getOperation())))
5064 return failure();
5065 }
5066 for (auto it : llvm::zip(yield.getOperands(), getResults())) {
5067 if (failed(verifyDistributedType(std::get<0>(it).getType(),
5068 std::get<1>(it).getType(), warpSize,
5069 getOperation())))
5070 return failure();
5071 }
5072 return success();
5073 }
5074
areTypesCompatible(Type lhs,Type rhs)5075 bool WarpExecuteOnLane0Op::areTypesCompatible(Type lhs, Type rhs) {
5076 return succeeded(
5077 verifyDistributedType(lhs, rhs, getWarpSize(), getOperation()));
5078 }
5079
makeArithReduction(OpBuilder & b,Location loc,CombiningKind kind,Value v1,Value v2)5080 Value mlir::vector::makeArithReduction(OpBuilder &b, Location loc,
5081 CombiningKind kind, Value v1, Value v2) {
5082 Type t1 = getElementTypeOrSelf(v1.getType());
5083 Type t2 = getElementTypeOrSelf(v2.getType());
5084 switch (kind) {
5085 case CombiningKind::ADD:
5086 if (t1.isIntOrIndex() && t2.isIntOrIndex())
5087 return b.createOrFold<arith::AddIOp>(loc, v1, v2);
5088 else if (t1.isa<FloatType>() && t2.isa<FloatType>())
5089 return b.createOrFold<arith::AddFOp>(loc, v1, v2);
5090 llvm_unreachable("invalid value types for ADD reduction");
5091 case CombiningKind::AND:
5092 assert(t1.isIntOrIndex() && t2.isIntOrIndex() && "expected int values");
5093 return b.createOrFold<arith::AndIOp>(loc, v1, v2);
5094 case CombiningKind::MAXF:
5095 assert(t1.isa<FloatType>() && t2.isa<FloatType>() &&
5096 "expected float values");
5097 return b.createOrFold<arith::MaxFOp>(loc, v1, v2);
5098 case CombiningKind::MINF:
5099 assert(t1.isa<FloatType>() && t2.isa<FloatType>() &&
5100 "expected float values");
5101 return b.createOrFold<arith::MinFOp>(loc, v1, v2);
5102 case CombiningKind::MAXSI:
5103 assert(t1.isIntOrIndex() && t2.isIntOrIndex() && "expected int values");
5104 return b.createOrFold<arith::MaxSIOp>(loc, v1, v2);
5105 case CombiningKind::MINSI:
5106 assert(t1.isIntOrIndex() && t2.isIntOrIndex() && "expected int values");
5107 return b.createOrFold<arith::MinSIOp>(loc, v1, v2);
5108 case CombiningKind::MAXUI:
5109 assert(t1.isIntOrIndex() && t2.isIntOrIndex() && "expected int values");
5110 return b.createOrFold<arith::MaxUIOp>(loc, v1, v2);
5111 case CombiningKind::MINUI:
5112 assert(t1.isIntOrIndex() && t2.isIntOrIndex() && "expected int values");
5113 return b.createOrFold<arith::MinUIOp>(loc, v1, v2);
5114 case CombiningKind::MUL:
5115 if (t1.isIntOrIndex() && t2.isIntOrIndex())
5116 return b.createOrFold<arith::MulIOp>(loc, v1, v2);
5117 else if (t1.isa<FloatType>() && t2.isa<FloatType>())
5118 return b.createOrFold<arith::MulFOp>(loc, v1, v2);
5119 llvm_unreachable("invalid value types for MUL reduction");
5120 case CombiningKind::OR:
5121 assert(t1.isIntOrIndex() && t2.isIntOrIndex() && "expected int values");
5122 return b.createOrFold<arith::OrIOp>(loc, v1, v2);
5123 case CombiningKind::XOR:
5124 assert(t1.isIntOrIndex() && t2.isIntOrIndex() && "expected int values");
5125 return b.createOrFold<arith::XOrIOp>(loc, v1, v2);
5126 };
5127 llvm_unreachable("unknown CombiningKind");
5128 }
5129
5130 //===----------------------------------------------------------------------===//
5131 // TableGen'd op method definitions
5132 //===----------------------------------------------------------------------===//
5133
5134 #define GET_OP_CLASSES
5135 #include "mlir/Dialect/Vector/IR/VectorOps.cpp.inc"
5136