1 //===- FoldSubViewOps.cpp - Fold memref.subview ops -----------------------===//
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 transformation pass folds loading/storing from/to subview ops into
10 // loading/storing from/to the original memref.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "mlir/Dialect/Affine/IR/AffineOps.h"
15 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
16 #include "mlir/Dialect/MemRef/IR/MemRef.h"
17 #include "mlir/Dialect/MemRef/Transforms/Passes.h"
18 #include "mlir/Dialect/StandardOps/IR/Ops.h"
19 #include "mlir/Dialect/Vector/VectorOps.h"
20 #include "mlir/IR/BuiltinTypes.h"
21 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
22 
23 using namespace mlir;
24 
25 //===----------------------------------------------------------------------===//
26 // Utility functions
27 //===----------------------------------------------------------------------===//
28 
29 /// Given the 'indices' of an load/store operation where the memref is a result
30 /// of a subview op, returns the indices w.r.t to the source memref of the
31 /// subview op. For example
32 ///
33 /// %0 = ... : memref<12x42xf32>
34 /// %1 = subview %0[%arg0, %arg1][][%stride1, %stride2] : memref<12x42xf32> to
35 ///          memref<4x4xf32, offset=?, strides=[?, ?]>
36 /// %2 = load %1[%i1, %i2] : memref<4x4xf32, offset=?, strides=[?, ?]>
37 ///
38 /// could be folded into
39 ///
40 /// %2 = load %0[%arg0 + %i1 * %stride1][%arg1 + %i2 * %stride2] :
41 ///          memref<12x42xf32>
42 static LogicalResult
43 resolveSourceIndices(Location loc, PatternRewriter &rewriter,
44                      memref::SubViewOp subViewOp, ValueRange indices,
45                      SmallVectorImpl<Value> &sourceIndices) {
46   SmallVector<OpFoldResult> mixedOffsets = subViewOp.getMixedOffsets();
47   SmallVector<OpFoldResult> mixedSizes = subViewOp.getMixedSizes();
48   SmallVector<OpFoldResult> mixedStrides = subViewOp.getMixedStrides();
49 
50   SmallVector<Value> useIndices;
51   // Check if this is rank-reducing case. Then for every unit-dim size add a
52   // zero to the indices.
53   unsigned resultDim = 0;
54   llvm::SmallDenseSet<unsigned> unusedDims = subViewOp.getDroppedDims();
55   for (auto dim : llvm::seq<unsigned>(0, subViewOp.getSourceType().getRank())) {
56     if (unusedDims.count(dim))
57       useIndices.push_back(rewriter.create<arith::ConstantIndexOp>(loc, 0));
58     else
59       useIndices.push_back(indices[resultDim++]);
60   }
61   if (useIndices.size() != mixedOffsets.size())
62     return failure();
63   sourceIndices.resize(useIndices.size());
64   for (auto index : llvm::seq<size_t>(0, mixedOffsets.size())) {
65     SmallVector<Value> dynamicOperands;
66     AffineExpr expr = rewriter.getAffineDimExpr(0);
67     unsigned numSymbols = 0;
68     dynamicOperands.push_back(useIndices[index]);
69 
70     // Multiply the stride;
71     if (auto attr = mixedStrides[index].dyn_cast<Attribute>()) {
72       expr = expr * attr.cast<IntegerAttr>().getInt();
73     } else {
74       dynamicOperands.push_back(mixedStrides[index].get<Value>());
75       expr = expr * rewriter.getAffineSymbolExpr(numSymbols++);
76     }
77 
78     // Add the offset.
79     if (auto attr = mixedOffsets[index].dyn_cast<Attribute>()) {
80       expr = expr + attr.cast<IntegerAttr>().getInt();
81     } else {
82       dynamicOperands.push_back(mixedOffsets[index].get<Value>());
83       expr = expr + rewriter.getAffineSymbolExpr(numSymbols++);
84     }
85     Location loc = subViewOp.getLoc();
86     sourceIndices[index] = rewriter.create<AffineApplyOp>(
87         loc, AffineMap::get(1, numSymbols, expr), dynamicOperands);
88   }
89   return success();
90 }
91 
92 /// Helpers to access the memref operand for each op.
93 static Value getMemRefOperand(memref::LoadOp op) { return op.memref(); }
94 
95 static Value getMemRefOperand(vector::TransferReadOp op) { return op.source(); }
96 
97 static Value getMemRefOperand(memref::StoreOp op) { return op.memref(); }
98 
99 static Value getMemRefOperand(vector::TransferWriteOp op) {
100   return op.source();
101 }
102 
103 /// Given the permutation map of the original
104 /// `vector.transfer_read`/`vector.transfer_write` operations compute the
105 /// permutation map to use after the subview is folded with it.
106 static AffineMap getPermutationMap(MLIRContext *context,
107                                    memref::SubViewOp subViewOp,
108                                    AffineMap currPermutationMap) {
109   llvm::SmallDenseSet<unsigned> unusedDims = subViewOp.getDroppedDims();
110   SmallVector<AffineExpr> exprs;
111   int64_t sourceRank = subViewOp.getSourceType().getRank();
112   for (auto dim : llvm::seq<int64_t>(0, sourceRank)) {
113     if (unusedDims.count(dim))
114       continue;
115     exprs.push_back(getAffineDimExpr(dim, context));
116   }
117   auto resultDimToSourceDimMap = AffineMap::get(sourceRank, 0, exprs, context);
118   return currPermutationMap.compose(resultDimToSourceDimMap);
119 }
120 
121 //===----------------------------------------------------------------------===//
122 // Patterns
123 //===----------------------------------------------------------------------===//
124 
125 namespace {
126 /// Merges subview operation with load/transferRead operation.
127 template <typename OpTy>
128 class LoadOpOfSubViewFolder final : public OpRewritePattern<OpTy> {
129 public:
130   using OpRewritePattern<OpTy>::OpRewritePattern;
131 
132   LogicalResult matchAndRewrite(OpTy loadOp,
133                                 PatternRewriter &rewriter) const override;
134 
135 private:
136   void replaceOp(OpTy loadOp, memref::SubViewOp subViewOp,
137                  ArrayRef<Value> sourceIndices,
138                  PatternRewriter &rewriter) const;
139 };
140 
141 /// Merges subview operation with store/transferWriteOp operation.
142 template <typename OpTy>
143 class StoreOpOfSubViewFolder final : public OpRewritePattern<OpTy> {
144 public:
145   using OpRewritePattern<OpTy>::OpRewritePattern;
146 
147   LogicalResult matchAndRewrite(OpTy storeOp,
148                                 PatternRewriter &rewriter) const override;
149 
150 private:
151   void replaceOp(OpTy storeOp, memref::SubViewOp subViewOp,
152                  ArrayRef<Value> sourceIndices,
153                  PatternRewriter &rewriter) const;
154 };
155 
156 template <>
157 void LoadOpOfSubViewFolder<memref::LoadOp>::replaceOp(
158     memref::LoadOp loadOp, memref::SubViewOp subViewOp,
159     ArrayRef<Value> sourceIndices, PatternRewriter &rewriter) const {
160   rewriter.replaceOpWithNewOp<memref::LoadOp>(loadOp, subViewOp.source(),
161                                               sourceIndices);
162 }
163 
164 template <>
165 void LoadOpOfSubViewFolder<vector::TransferReadOp>::replaceOp(
166     vector::TransferReadOp loadOp, memref::SubViewOp subViewOp,
167     ArrayRef<Value> sourceIndices, PatternRewriter &rewriter) const {
168   rewriter.replaceOpWithNewOp<vector::TransferReadOp>(
169       loadOp, loadOp.getVectorType(), subViewOp.source(), sourceIndices,
170       getPermutationMap(rewriter.getContext(), subViewOp,
171                         loadOp.permutation_map()),
172       loadOp.padding(), loadOp.in_boundsAttr());
173 }
174 
175 template <>
176 void StoreOpOfSubViewFolder<memref::StoreOp>::replaceOp(
177     memref::StoreOp storeOp, memref::SubViewOp subViewOp,
178     ArrayRef<Value> sourceIndices, PatternRewriter &rewriter) const {
179   rewriter.replaceOpWithNewOp<memref::StoreOp>(
180       storeOp, storeOp.value(), subViewOp.source(), sourceIndices);
181 }
182 
183 template <>
184 void StoreOpOfSubViewFolder<vector::TransferWriteOp>::replaceOp(
185     vector::TransferWriteOp transferWriteOp, memref::SubViewOp subViewOp,
186     ArrayRef<Value> sourceIndices, PatternRewriter &rewriter) const {
187   rewriter.replaceOpWithNewOp<vector::TransferWriteOp>(
188       transferWriteOp, transferWriteOp.vector(), subViewOp.source(),
189       sourceIndices,
190       getPermutationMap(rewriter.getContext(), subViewOp,
191                         transferWriteOp.permutation_map()),
192       transferWriteOp.in_boundsAttr());
193 }
194 } // namespace
195 
196 template <typename OpTy>
197 LogicalResult
198 LoadOpOfSubViewFolder<OpTy>::matchAndRewrite(OpTy loadOp,
199                                              PatternRewriter &rewriter) const {
200   auto subViewOp =
201       getMemRefOperand(loadOp).template getDefiningOp<memref::SubViewOp>();
202   if (!subViewOp)
203     return failure();
204 
205   SmallVector<Value, 4> sourceIndices;
206   if (failed(resolveSourceIndices(loadOp.getLoc(), rewriter, subViewOp,
207                                   loadOp.indices(), sourceIndices)))
208     return failure();
209 
210   replaceOp(loadOp, subViewOp, sourceIndices, rewriter);
211   return success();
212 }
213 
214 template <typename OpTy>
215 LogicalResult
216 StoreOpOfSubViewFolder<OpTy>::matchAndRewrite(OpTy storeOp,
217                                               PatternRewriter &rewriter) const {
218   auto subViewOp =
219       getMemRefOperand(storeOp).template getDefiningOp<memref::SubViewOp>();
220   if (!subViewOp)
221     return failure();
222 
223   SmallVector<Value, 4> sourceIndices;
224   if (failed(resolveSourceIndices(storeOp.getLoc(), rewriter, subViewOp,
225                                   storeOp.indices(), sourceIndices)))
226     return failure();
227 
228   replaceOp(storeOp, subViewOp, sourceIndices, rewriter);
229   return success();
230 }
231 
232 void memref::populateFoldSubViewOpPatterns(RewritePatternSet &patterns) {
233   patterns.add<LoadOpOfSubViewFolder<memref::LoadOp>,
234                LoadOpOfSubViewFolder<vector::TransferReadOp>,
235                StoreOpOfSubViewFolder<memref::StoreOp>,
236                StoreOpOfSubViewFolder<vector::TransferWriteOp>>(
237       patterns.getContext());
238 }
239 
240 //===----------------------------------------------------------------------===//
241 // Pass registration
242 //===----------------------------------------------------------------------===//
243 
244 namespace {
245 
246 #define GEN_PASS_CLASSES
247 #include "mlir/Dialect/MemRef/Transforms/Passes.h.inc"
248 
249 struct FoldSubViewOpsPass final
250     : public FoldSubViewOpsBase<FoldSubViewOpsPass> {
251   void runOnOperation() override;
252 };
253 
254 } // namespace
255 
256 void FoldSubViewOpsPass::runOnOperation() {
257   RewritePatternSet patterns(&getContext());
258   memref::populateFoldSubViewOpPatterns(patterns);
259   (void)applyPatternsAndFoldGreedily(getOperation()->getRegions(),
260                                      std::move(patterns));
261 }
262 
263 std::unique_ptr<Pass> memref::createFoldSubViewOpsPass() {
264   return std::make_unique<FoldSubViewOpsPass>();
265 }
266