1 //===- DropUnitDims.cpp - Pass to drop use of unit-extent for broadcasting ===//
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 patterns/pass to remove usage of unit-extent dimensions
10 // to specify broadcasting in favor of more canonical representation of the
11 // computation
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "PassDetail.h"
16 #include "mlir/Dialect/Linalg/IR/LinalgOps.h"
17 #include "mlir/Dialect/Linalg/IR/LinalgTypes.h"
18 #include "mlir/Dialect/Linalg/Passes.h"
19 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
20 #include "mlir/Dialect/Linalg/Utils/Utils.h"
21 #include "mlir/Dialect/Tensor/IR/Tensor.h"
22 #include "mlir/IR/AffineExpr.h"
23 #include "mlir/IR/AffineMap.h"
24 #include "mlir/IR/BuiltinTypes.h"
25 #include "mlir/Transforms/FoldUtils.h"
26 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Debug.h"
29 
30 #define DEBUG_TYPE "linalg-drop-unit-dims"
31 
32 using namespace mlir;
33 using namespace mlir::linalg;
34 
35 /// Implements a pass that canonicalizes the uses of unit-extent dimensions for
36 /// broadcasting. For example,
37 ///
38 /// ```mlir
39 /// #accesses = [
40 ///   affine_map<(d0, d1) -> (0, d1)>,
41 ///   affine_map<(d0, d1) -> (d0, 0)>,
42 ///   affine_map<(d0, d1) -> (d0, d1)>
43 /// ]
44 ///
45 /// #trait = {
46 ///   args_in = 2,
47 ///   args_out = 1,
48 ///   indexing_maps = #accesses,
49 ///   iterator_types = ["parallel", "parallel"],
50 ///   library_call = "some_external_fn"
51 /// }
52 ///
53 /// func @broadcast_test(%arg0 : tensor<5xf32>, %arg1 : tensor<5xf32>) ->
54 /// tensor<5x5xf32>
55 /// {
56 ///   %0 = linalg.tensor_reshape %arg0 [affine_map<(d0, d1) -> (d0, d1)>] :
57 ///        tensor<5xf32> into tensor<1x5xf32>
58 ///   %1 = linalg.tensor_reshape %arg1 [affine_map<(d0, d1) -> (d0, d1)>] :
59 ///        tensor<5xf32> into tensor<5x1xf32>
60 ///   %2 = linalg.generic #trait %0, %1 {
61 ///        ^bb0(%arg2: f32, %arg3: f32):
62 ///          %3 = addf %arg2, %arg3 : f32
63 ///          linalg.yield %3 : f32
64 ///        } : tensor<1x5xf32>, tensor<5x1xf32> -> tensor<5x5xf32>
65 ///   return %2 : tensor<5x5xf32>
66 /// }
67 ///
68 /// would canonicalize to
69 ///
70 /// ```mlir
71 /// #accesses = [
72 ///   affine_map<(d0, d1) -> (d1)>,
73 ///   affine_map<(d0, d1) -> (d0)>,
74 ///   affine_map<(d0, d1) -> (d0, d1)>
75 /// ]
76 ///
77 /// #trait = {
78 ///   args_in = 2,
79 ///   args_out = 1,
80 ///   indexing_maps = #accesses,
81 ///   iterator_types = ["parallel", "parallel"],
82 ///   library_call = "some_external_fn"
83 /// }
84 ///
85 /// func @broadcast_test(%arg0 : tensor<5xf32>, %arg1 : tensor<5xf32>) ->
86 /// tensor<5x5xf32>
87 /// {
88 ///   %0 = linalg.generic #trait %arg0, %arg1 {
89 ///        ^bb0(%arg2: f32, %arg3: f32):
90 ///          %3 = addf %arg2, %arg3 : f32
91 ///          linalg.yield %3 : f32
92 ///        } : tensor<5xf32>, tensor<5xf32> -> tensor<5x5xf32>
93 ///   return %0 : tensor<5x5xf32>
94 /// }
95 
96 /// Given dims of the iteration space of a structured op that are known to be
97 /// single trip count (`unitDims`), return the indexing maps to use in the
98 /// canonicalized op with these dims removed, given the original `indexingMaps`.
replaceUnitDims(DenseSet<unsigned> & unitDims,ArrayRef<AffineMap> indexingMaps,MLIRContext * context)99 static ArrayAttr replaceUnitDims(DenseSet<unsigned> &unitDims,
100                                  ArrayRef<AffineMap> indexingMaps,
101                                  MLIRContext *context) {
102   if (indexingMaps.empty())
103     return nullptr;
104   unsigned numIterationDims = indexingMaps.front().getNumDims();
105   unsigned numSymbols = indexingMaps.front().getNumSymbols();
106 
107   // Compute the replacement for each dim expr.
108   SmallVector<AffineExpr, 4> dimReplacements;
109   dimReplacements.reserve(numIterationDims);
110   unsigned numKeptDims = 0;
111   for (unsigned dim : llvm::seq<unsigned>(0, numIterationDims)) {
112     if (unitDims.count(dim))
113       dimReplacements.push_back(getAffineConstantExpr(0, context));
114     else
115       dimReplacements.push_back(getAffineDimExpr(numKeptDims++, context));
116   }
117 
118   // Symbols remain the same.
119   SmallVector<AffineExpr, 4> symReplacements;
120   symReplacements.reserve(numSymbols);
121   for (unsigned symbol : llvm::seq<unsigned>(0, numSymbols))
122     symReplacements.push_back(getAffineSymbolExpr(symbol, context));
123 
124   SmallVector<AffineMap, 4> newIndexingMaps;
125   newIndexingMaps.reserve(indexingMaps.size());
126   for (AffineMap operandMap : indexingMaps) {
127     // Expected indexing maps to have no symbols.
128     if (operandMap.getNumSymbols())
129       return nullptr;
130     newIndexingMaps.push_back(simplifyAffineMap(
131         operandMap.replaceDimsAndSymbols(dimReplacements, symReplacements,
132                                          numIterationDims - unitDims.size(),
133                                          numSymbols)));
134   }
135 
136   // Check that the new index maps are invertible. If not, something went
137   // wrong, so abort.
138   if (!inversePermutation(concatAffineMaps(newIndexingMaps)))
139     return nullptr;
140   return ArrayAttr::get(context,
141                         llvm::to_vector<4>(llvm::map_range(
142                             newIndexingMaps, [](AffineMap map) -> Attribute {
143                               return AffineMapAttr::get(map);
144                             })));
145 }
146 
147 /// Update the index accesses of linalg operations having index semantics.
replaceUnitDimIndexOps(GenericOp genericOp,const DenseSet<unsigned> & unitDims,PatternRewriter & rewriter)148 static void replaceUnitDimIndexOps(GenericOp genericOp,
149                                    const DenseSet<unsigned> &unitDims,
150                                    PatternRewriter &rewriter) {
151   assert(genericOp->getNumRegions() == 1 &&
152          genericOp->getRegion(0).getBlocks().size() == 1 &&
153          "expected generic operation to have one block.");
154   Block &block = genericOp->getRegion(0).front();
155 
156   for (IndexOp indexOp : llvm::make_early_inc_range(block.getOps<IndexOp>())) {
157     OpBuilder::InsertionGuard guard(rewriter);
158     rewriter.setInsertionPoint(indexOp);
159     if (unitDims.count(indexOp.dim()) != 0) {
160       rewriter.replaceOpWithNewOp<ConstantIndexOp>(indexOp, 0);
161     } else {
162       // Update the dimension of the index operation if needed.
163       unsigned droppedDims = llvm::count_if(
164           unitDims, [&](unsigned dim) { return dim < indexOp.dim(); });
165       if (droppedDims != 0)
166         rewriter.replaceOpWithNewOp<IndexOp>(indexOp,
167                                              indexOp.dim() - droppedDims);
168     }
169   }
170 }
171 
172 namespace {
173 /// Pattern to fold unit-trip count loops in GenericOps.
174 struct FoldUnitDimLoops : public OpRewritePattern<GenericOp> {
175   using OpRewritePattern<GenericOp>::OpRewritePattern;
matchAndRewrite__anon6cffb90d0311::FoldUnitDimLoops176   LogicalResult matchAndRewrite(GenericOp genericOp,
177                                 PatternRewriter &rewriter) const override {
178     SmallVector<AffineMap, 4> indexingMaps = genericOp.getIndexingMaps();
179     if (indexingMaps.empty())
180       return failure();
181 
182     // Check if any of the iteration dimensions are unit-trip count. They will
183     // end up being unit-trip count if they are used to index into a unit-dim
184     // tensor/memref.
185     AffineMap invertedMap = inversePermutation(concatAffineMaps(indexingMaps));
186     if (!invertedMap)
187       return failure();
188     SmallVector<int64_t> dims = genericOp.getStaticShape();
189 
190     DenseSet<unsigned> unitDims;
191     SmallVector<unsigned, 4> unitDimsReductionLoops;
192     ArrayAttr iteratorTypes = genericOp.iterator_types();
193     for (auto expr : enumerate(invertedMap.getResults())) {
194       if (AffineDimExpr dimExpr = expr.value().dyn_cast<AffineDimExpr>())
195         if (dims[dimExpr.getPosition()] == 1)
196           unitDims.insert(expr.index());
197     }
198 
199     if (unitDims.empty())
200       return failure();
201 
202     // Compute the modified indexing maps.
203     MLIRContext *context = rewriter.getContext();
204     ArrayAttr newIndexingMapAttr =
205         replaceUnitDims(unitDims, indexingMaps, context);
206     if (!newIndexingMapAttr)
207       return genericOp.emitError("unable to compute modified indexing_maps");
208 
209     // Compute the iterator types of the modified op by dropping the one-trip
210     // count loops.
211     SmallVector<Attribute, 4> newIteratorTypes;
212     for (auto attr : llvm::enumerate(iteratorTypes)) {
213       if (!unitDims.count(attr.index()))
214         newIteratorTypes.push_back(attr.value());
215     }
216 
217     rewriter.startRootUpdate(genericOp);
218     genericOp.indexing_mapsAttr(newIndexingMapAttr);
219     genericOp.iterator_typesAttr(ArrayAttr::get(context, newIteratorTypes));
220     replaceUnitDimIndexOps(genericOp, unitDims, rewriter);
221     rewriter.finalizeRootUpdate(genericOp);
222     return success();
223   }
224 };
225 
226 struct UnitExtentReplacementInfo {
227   Type type;
228   AffineMap indexMap;
229   ArrayAttr reassociation;
230 };
231 } // namespace
232 
233 /// Utility function for replacing operands/results to a linalg generic
234 /// operation with unit-extent dimensions. These can be replaced with
235 /// an operand/result with the unit-extent dimension removed. This is only done
236 /// if the indexing map used to access that didimensionmension has a
237 /// AffineConstantExpr of value 0. Given the `type` of an result/operand of a
238 /// Linalg op, and its `indexMap` the utility function returns:
239 /// - the new type with dimensions of size 1 removed.
240 /// - modified index map that can be used to access the replaced result/operand
241 /// - the reassociation that converts from the original tensor type to the
242 ///   modified tensor type.
243 static llvm::Optional<UnitExtentReplacementInfo>
replaceUnitExtents(GenericOp genericOp,OpOperand * opOperand,MLIRContext * context)244 replaceUnitExtents(GenericOp genericOp, OpOperand *opOperand,
245                    MLIRContext *context) {
246   AffineMap indexingMap = genericOp.getTiedIndexingMap(opOperand);
247   ArrayRef<int64_t> shape = genericOp.getShape(opOperand);
248   ArrayRef<AffineExpr> exprs = indexingMap.getResults();
249   SmallVector<AffineExpr> reassociations;
250   SmallVector<Attribute> reassociationMaps;
251   SmallVector<AffineExpr> newIndexExprs;
252   SmallVector<int64_t> newShape;
253 
254   int64_t origRank = genericOp.getRank(opOperand);
255   AffineExpr zeroExpr = getAffineConstantExpr(0, context);
256   auto isUnitExtent = [&](int64_t dim) -> bool {
257     return shape[dim] == 1 && exprs[dim] == zeroExpr;
258   };
259 
260   // Early return for memrefs with affine maps to represent that we will always
261   // leave them unchanged.
262   Type actualType = opOperand->get().getType();
263   if (auto memref = actualType.dyn_cast<MemRefType>()) {
264     if (!memref.getAffineMaps().empty())
265       return llvm::None;
266   }
267 
268   int64_t dim = 0;
269   // Fold dimensions that are unit-extent at the beginning of the tensor.
270   while (dim < origRank && isUnitExtent(dim))
271     reassociations.push_back(getAffineDimExpr(dim++, context));
272   while (dim < origRank) {
273     reassociations.push_back(getAffineDimExpr(dim, context));
274     newIndexExprs.push_back(exprs[dim]);
275     newShape.push_back(shape[dim]);
276     // Fold all following dimensions that are unit-extent.
277     while (dim + 1 < origRank && isUnitExtent(dim + 1)) {
278       ++dim;
279       reassociations.push_back(getAffineDimExpr(dim, context));
280     }
281     reassociationMaps.push_back(AffineMapAttr::get(AffineMap::get(
282         origRank, /*symbolCount = */ 0, reassociations, context)));
283     reassociations.clear();
284     ++dim;
285   }
286 
287   // Compute the tensor or scalar replacement type.
288   Type elementType = getElementTypeOrSelf(opOperand->get());
289   Type replacementType;
290   if (elementType == opOperand->get().getType()) {
291     replacementType = elementType;
292   } else if (actualType.isa<RankedTensorType>()) {
293     replacementType = RankedTensorType::get(newShape, elementType);
294   } else if (actualType.isa<MemRefType>()) {
295     replacementType = MemRefType::get(newShape, elementType);
296   }
297   assert(replacementType && "unsupported shaped type");
298   UnitExtentReplacementInfo info = {replacementType,
299                                     AffineMap::get(indexingMap.getNumDims(),
300                                                    indexingMap.getNumSymbols(),
301                                                    newIndexExprs, context),
302                                     ArrayAttr::get(context, reassociationMaps)};
303   return info;
304 }
305 
306 namespace {
307 
308 SmallVector<ReassociationExprs, 2>
convertAffineMapArrayToExprs(ArrayAttr affineMapArrayAttr)309 convertAffineMapArrayToExprs(ArrayAttr affineMapArrayAttr) {
310   SmallVector<ReassociationExprs, 2> reassociationExprs;
311   for (auto attr : affineMapArrayAttr)
312     reassociationExprs.push_back(
313         llvm::to_vector<4>(attr.cast<AffineMapAttr>().getValue().getResults()));
314   return reassociationExprs;
315 }
316 
317 /// Pattern to replace tensor/buffer operands/results that are unit extents.
318 struct ReplaceUnitExtents : public OpRewritePattern<GenericOp> {
319   using OpRewritePattern<GenericOp>::OpRewritePattern;
320 
321   // Return the original value if the type is unchanged, or reshape it. Return a
322   // nullptr if this is an unsupported type.
maybeExpand__anon6cffb90d0511::ReplaceUnitExtents323   Value maybeExpand(Value result, Type origResultType,
324                     ArrayAttr reassociationMap, Location loc,
325                     PatternRewriter &rewriter) const {
326     if (origResultType == result.getType())
327       return result;
328     if (origResultType.isa<RankedTensorType>()) {
329       return rewriter.create<linalg::TensorExpandShapeOp>(
330           loc, origResultType, result,
331           convertAffineMapArrayToExprs(reassociationMap));
332     }
333     if (origResultType.isa<MemRefType>()) {
334       return rewriter.create<memref::ExpandShapeOp>(
335           loc, origResultType, result,
336           convertAffineMapArrayToExprs(reassociationMap));
337     }
338     return nullptr;
339   };
340 
341   // Return the original value if the type is unchanged, or reshape it. Return a
342   // nullptr if this is an unsupported type.
maybeCollapse__anon6cffb90d0511::ReplaceUnitExtents343   Value maybeCollapse(Value operand, Type newInputOutputType,
344                       ArrayAttr reassociationMap, Location loc,
345                       PatternRewriter &rewriter) const {
346     auto operandType = operand.getType();
347     if (operandType == newInputOutputType)
348       return operand;
349     if (operandType.isa<MemRefType>()) {
350       return rewriter.create<memref::CollapseShapeOp>(
351           loc, newInputOutputType, operand,
352           convertAffineMapArrayToExprs(reassociationMap));
353     }
354     if (operandType.isa<RankedTensorType>()) {
355       return rewriter.create<linalg::TensorCollapseShapeOp>(
356           loc, newInputOutputType, operand,
357           convertAffineMapArrayToExprs(reassociationMap));
358     }
359     return nullptr;
360   };
361 
matchAndRewrite__anon6cffb90d0511::ReplaceUnitExtents362   LogicalResult matchAndRewrite(GenericOp genericOp,
363                                 PatternRewriter &rewriter) const override {
364     // Skip the pattern if the op has any tensor with special encoding.
365     if (llvm::any_of(genericOp->getOperandTypes(), [](Type type) {
366           auto tensorType = type.dyn_cast<RankedTensorType>();
367           return tensorType && tensorType.getEncoding() != nullptr;
368         }))
369       return failure();
370     MLIRContext *context = rewriter.getContext();
371     Location loc = genericOp.getLoc();
372 
373     SmallVector<AffineMap> newIndexingMaps;
374     SmallVector<ArrayAttr> reassociationMaps;
375     SmallVector<Type> newInputOutputTypes;
376     bool doCanonicalization = false;
377     for (OpOperand *opOperand : genericOp.getInputAndOutputOperands()) {
378       auto replacementInfo = replaceUnitExtents(genericOp, opOperand, context);
379       if (replacementInfo) {
380         reassociationMaps.push_back(replacementInfo->reassociation);
381         newIndexingMaps.push_back(replacementInfo->indexMap);
382         newInputOutputTypes.push_back(replacementInfo->type);
383         doCanonicalization |=
384             replacementInfo->type != opOperand->get().getType();
385       } else {
386         // If replaceUnitExtents cannot handle this case, maintain the same
387         // type, indexing map, and create a set of mappings representing an
388         // identity matrix.
389         newInputOutputTypes.push_back(opOperand->get().getType());
390         newIndexingMaps.push_back(genericOp.getTiedIndexingMap(opOperand));
391         int64_t origRank = genericOp.getRank(opOperand);
392         auto maps = llvm::to_vector<8>(llvm::map_range(
393             llvm::seq<int64_t>(0, origRank), [&](int64_t dim) -> Attribute {
394               return AffineMapAttr::get(
395                   AffineMap::get(origRank, /*symbolCount = */ 0,
396                                  getAffineDimExpr(dim, context), context));
397             }));
398         reassociationMaps.push_back(ArrayAttr::get(context, maps));
399       }
400     }
401 
402     // If the indexing maps of the result operation are not invertible (i.e. not
403     // legal), abort.
404     if (!doCanonicalization ||
405         !inversePermutation(concatAffineMaps(newIndexingMaps)))
406       return failure();
407 
408     // If any operand type change, insert a reshape to convert from the original
409     // type to the new type.
410     // TODO: get rid of flattenedIdx which assumes operand order and contiguity.
411     unsigned flattenedIdx = 0;
412     auto insertReshapes = [&](ValueRange values) {
413       SmallVector<Value, 4> res;
414       res.reserve(values.size());
415       for (auto operand : values) {
416         auto reshapedValue =
417             maybeCollapse(operand, newInputOutputTypes[flattenedIdx],
418                           reassociationMaps[flattenedIdx], loc, rewriter);
419         assert(reshapedValue &&
420                "expected ranked MemRef or Tensor operand type");
421         res.push_back(reshapedValue);
422         ++flattenedIdx;
423       }
424       return res;
425     };
426 
427     SmallVector<Value, 4> newInputs = insertReshapes(genericOp.inputs());
428     SmallVector<Value, 4> newOutputs = insertReshapes(genericOp.outputs());
429 
430     // If any result type changes, insert a reshape to convert from the original
431     // type to the new type.
432     SmallVector<Type, 4> resultTypes;
433     resultTypes.reserve(genericOp.getNumResults());
434     for (unsigned i : llvm::seq<unsigned>(0, genericOp.getNumResults()))
435       resultTypes.push_back(newInputOutputTypes[i + genericOp.getNumInputs()]);
436     GenericOp replacementOp = rewriter.create<GenericOp>(
437         loc, resultTypes, newInputs, newOutputs, newIndexingMaps,
438         llvm::to_vector<4>(
439             genericOp.iterator_types().template getAsValueRange<StringAttr>()));
440     rewriter.inlineRegionBefore(genericOp.region(), replacementOp.region(),
441                                 replacementOp.region().begin());
442 
443     // If any result tensor has a modified shape, then add reshape to recover
444     // the original shape.
445     SmallVector<Value, 4> resultReplacements;
446     for (auto result : llvm::enumerate(replacementOp.getResults())) {
447       unsigned index = result.index() + replacementOp.getNumInputs();
448       auto origResultType = genericOp.getResult(result.index()).getType();
449 
450       auto newResult = maybeExpand(result.value(), origResultType,
451                                    reassociationMaps[index], loc, rewriter);
452       assert(newResult &&
453              "unexpected output type other than ranked MemRef or Tensor");
454       resultReplacements.push_back(newResult);
455     }
456     rewriter.replaceOp(genericOp, resultReplacements);
457     return success();
458   }
459 };
460 } // namespace
461 
462 /// Get the reassociation maps to fold the result of a extract_slice (or source
463 /// of a insert_slice) operation with given offsets, and sizes to its
464 /// rank-reduced version. This is only done for the cases where the size is 1
465 /// and offset is 0. Strictly speaking the offset 0 is not required in general,
466 /// but non-zero offsets are not handled by SPIR-V backend at this point (and
467 /// potentially cannot be handled).
468 static Optional<SmallVector<ReassociationIndices>>
getReassociationMapForFoldingUnitDims(ArrayRef<OpFoldResult> mixedSizes)469 getReassociationMapForFoldingUnitDims(ArrayRef<OpFoldResult> mixedSizes) {
470   SmallVector<ReassociationIndices> reassociation;
471   ReassociationIndices curr;
472   for (auto it : llvm::enumerate(mixedSizes)) {
473     auto dim = it.index();
474     auto size = it.value();
475     curr.push_back(dim);
476     auto attr = size.dyn_cast<Attribute>();
477     if (attr && attr.cast<IntegerAttr>().getInt() == 1)
478       continue;
479     reassociation.emplace_back(ReassociationIndices{});
480     std::swap(reassociation.back(), curr);
481   }
482   // When the reassociations are not empty, then fold the remaining
483   // unit-dimensions into the last dimension.  If the reassociations so far is
484   // empty, then leave it emtpy. This will fold everything to a rank-0 tensor.
485   if (!curr.empty() && !reassociation.empty())
486     reassociation.back().append(curr.begin(), curr.end());
487   return reassociation;
488 }
489 
490 namespace {
491 /// Convert `extract_slice` operations to rank-reduced versions.
492 struct UseRankReducedExtractSliceOp
493     : public OpRewritePattern<tensor::ExtractSliceOp> {
494   using OpRewritePattern<tensor::ExtractSliceOp>::OpRewritePattern;
495 
matchAndRewrite__anon6cffb90d0911::UseRankReducedExtractSliceOp496   LogicalResult matchAndRewrite(tensor::ExtractSliceOp sliceOp,
497                                 PatternRewriter &rewriter) const override {
498     RankedTensorType resultType = sliceOp.getType();
499     SmallVector<OpFoldResult> offsets = sliceOp.getMixedOffsets();
500     SmallVector<OpFoldResult> sizes = sliceOp.getMixedSizes();
501     SmallVector<OpFoldResult> strides = sliceOp.getMixedStrides();
502     auto reassociation = getReassociationMapForFoldingUnitDims(sizes);
503     if (!reassociation ||
504         reassociation->size() == static_cast<size_t>(resultType.getRank()))
505       return failure();
506     auto rankReducedType = tensor::ExtractSliceOp::inferRankReducedResultType(
507                                reassociation->size(), sliceOp.getSourceType(),
508                                offsets, sizes, strides)
509                                .cast<RankedTensorType>();
510 
511     Location loc = sliceOp.getLoc();
512     Value newSlice = rewriter.create<tensor::ExtractSliceOp>(
513         loc, rankReducedType, sliceOp.source(), offsets, sizes, strides);
514     rewriter.replaceOpWithNewOp<TensorExpandShapeOp>(sliceOp, resultType,
515                                                      newSlice, *reassociation);
516     return success();
517   }
518 };
519 
520 /// Convert `insert_slice` operations to rank-reduced versions.
521 struct UseRankReducedInsertSliceOp
522     : public OpRewritePattern<tensor::InsertSliceOp> {
523   using OpRewritePattern<tensor::InsertSliceOp>::OpRewritePattern;
524 
matchAndRewrite__anon6cffb90d0911::UseRankReducedInsertSliceOp525   LogicalResult matchAndRewrite(tensor::InsertSliceOp insertOp,
526                                 PatternRewriter &rewriter) const override {
527     RankedTensorType sourceType = insertOp.getSourceType();
528     SmallVector<OpFoldResult> offsets = insertOp.getMixedOffsets();
529     SmallVector<OpFoldResult> sizes = insertOp.getMixedSizes();
530     SmallVector<OpFoldResult> strides = insertOp.getMixedStrides();
531     auto reassociation = getReassociationMapForFoldingUnitDims(sizes);
532     if (!reassociation ||
533         reassociation->size() == static_cast<size_t>(sourceType.getRank()))
534       return failure();
535     Location loc = insertOp.getLoc();
536     auto reshapedSource = rewriter.create<TensorCollapseShapeOp>(
537         loc, insertOp.source(), *reassociation);
538     rewriter.replaceOpWithNewOp<tensor::InsertSliceOp>(
539         insertOp, reshapedSource, insertOp.dest(), insertOp.getMixedOffsets(),
540         insertOp.getMixedSizes(), insertOp.getMixedStrides());
541     return success();
542   }
543 };
544 } // namespace
545 
546 /// Patterns that are used to canonicalize the use of unit-extent dims for
547 /// broadcasting.
populateFoldUnitExtentDimsPatterns(RewritePatternSet & patterns)548 void mlir::linalg::populateFoldUnitExtentDimsPatterns(
549     RewritePatternSet &patterns) {
550   auto *context = patterns.getContext();
551   patterns.add<FoldUnitDimLoops, ReplaceUnitExtents,
552                UseRankReducedExtractSliceOp, UseRankReducedInsertSliceOp>(
553       context);
554   TensorCollapseShapeOp::getCanonicalizationPatterns(patterns, context);
555   TensorExpandShapeOp::getCanonicalizationPatterns(patterns, context);
556 }
557 
558 namespace {
559 /// Pass that removes unit-extent dims within generic ops.
560 struct LinalgFoldUnitExtentDimsPass
561     : public LinalgFoldUnitExtentDimsBase<LinalgFoldUnitExtentDimsPass> {
runOnFunction__anon6cffb90d0a11::LinalgFoldUnitExtentDimsPass562   void runOnFunction() override {
563     FuncOp funcOp = getFunction();
564     MLIRContext *context = funcOp.getContext();
565     RewritePatternSet patterns(context);
566     if (foldOneTripLoopsOnly)
567       patterns.add<FoldUnitDimLoops>(context);
568     else
569       populateFoldUnitExtentDimsPatterns(patterns);
570     (void)applyPatternsAndFoldGreedily(funcOp.getBody(), std::move(patterns));
571   }
572 };
573 } // namespace
574 
575 std::unique_ptr<OperationPass<FuncOp>>
createLinalgFoldUnitExtentDimsPass()576 mlir::createLinalgFoldUnitExtentDimsPass() {
577   return std::make_unique<LinalgFoldUnitExtentDimsPass>();
578 }
579