1 | //===- Fusion.cpp - Implementation of linalg Fusion -----------------------===// |
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 the linalg dialect Fusion pass. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
14 | #include "mlir/Dialect/Arith/IR/Arith.h" |
15 | #include "mlir/Dialect/Linalg/IR/Linalg.h" |
16 | #include "mlir/Dialect/Linalg/Passes.h" |
17 | #include "mlir/Dialect/Linalg/Transforms/Transforms.h" |
18 | #include "mlir/Dialect/Linalg/Utils/Utils.h" |
19 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
20 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
21 | #include "mlir/Dialect/Tensor/Utils/Utils.h" |
22 | #include "mlir/IR/AffineExpr.h" |
23 | #include "mlir/IR/AffineMap.h" |
24 | #include "mlir/IR/Dominance.h" |
25 | #include "mlir/Support/LLVM.h" |
26 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
27 | #include "mlir/Transforms/RegionUtils.h" |
28 | #include "llvm/ADT/MapVector.h" |
29 | #include "llvm/ADT/ScopeExit.h" |
30 | #include "llvm/ADT/SmallBitVector.h" |
31 | #include "llvm/Support/CommandLine.h" |
32 | #include "llvm/Support/Debug.h" |
33 | |
34 | #include <optional> |
35 | #include <set> |
36 | |
37 | #define DEBUG_TYPE "linalg-fusion" |
38 | |
39 | using namespace mlir; |
40 | using namespace mlir::linalg; |
41 | |
42 | /// Implements a simple high-level fusion pass on linalg structured operations. |
43 | /// |
44 | /// In each block, linalg ops are processed in reverse textual order. |
45 | /// Given a linalg op `O`, fusion occurs by: |
46 | /// 1. inspecting the linalg ops that write into the views read by `O`. There |
47 | /// are 2 cases: |
48 | /// a) buffer case: use the SSA value of the views and a simple alias |
49 | /// analysis on subview ops to determine producer-consumer dependences; |
50 | /// b) tensor case: use SSA use-def chains on extract_slice ops; |
51 | /// 2. greedily fuse the linalg ops that produce the subview/extract_slice. |
52 | /// 3. inspect the fused ops and determine whether they have other remaining |
53 | /// LinalgOp uses. If not, then erase the original producing linalg op. |
54 | /// |
55 | /// More advanced use cases, analyses as well as profitability heuristics are |
56 | /// left for future work. |
57 | |
58 | struct ShapeDimension { |
59 | Value shape; |
60 | unsigned dimension; |
61 | }; |
62 | |
63 | // Given an `op`, returns the first (`shape`, `dimension`) pair that identifies |
64 | // the loop range at `loopDepth`. The semantics of the loopToOperandRangesMaps |
65 | // guarantees at least one such dimension is found. If multiple candidates exist |
66 | // they must agree by construction (i.e. have the same size) and we just return |
67 | // the first one. |
68 | static ShapeDimension |
69 | getShapeDefiningLoopRange(LinalgOp op, unsigned loopDepth, |
70 | bool fromSubViewOpOnly = false) { |
71 | // Iterate over the inputs and outputs in order. |
72 | // Extract the subranges from the linearized ranges. |
73 | for (OpOperand &opOperand : op->getOpOperands()) { |
74 | // The method `getRangeFromOperandShape` requires using SubViewOp or |
75 | // ExtractSliceOps. If the value isn't defined from there continue. |
76 | // todo: The method should be adapted to get the values from |
77 | // `ViewInterface`. The interface needs a `getOrCreateRanges` method which |
78 | // currently returns a `linalg.range`. The fix here is to move this op to |
79 | // `std` dialect and add the method to `ViewInterface`. |
80 | if (fromSubViewOpOnly && |
81 | !isa_and_nonnull<memref::SubViewOp, tensor::ExtractSliceOp>( |
82 | opOperand.get().getDefiningOp())) |
83 | continue; |
84 | |
85 | AffineMap map = op.getMatchingIndexingMap(&opOperand); |
86 | LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange I/O idx: " |
87 | << opOperand.getOperandNumber() << "\n" ); |
88 | LLVM_DEBUG(llvm::dbgs() |
89 | << "getShapeDefiningLoopRange map: " << map << "\n" ); |
90 | for (const auto &en : llvm::enumerate(map.getResults())) { |
91 | auto dimExpr = dyn_cast<AffineDimExpr>(en.value()); |
92 | if (!dimExpr) |
93 | continue; |
94 | if (loopDepth == cast<AffineDimExpr>(en.value()).getPosition()) { |
95 | LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange loopDepth: " |
96 | << loopDepth << "\n" ); |
97 | LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange shape: " |
98 | << opOperand.get() << "\n" ); |
99 | return ShapeDimension{opOperand.get(), |
100 | static_cast<unsigned>(en.index())}; |
101 | } |
102 | } |
103 | } |
104 | llvm_unreachable("Expect to be able to extract a shape defining loop range" ); |
105 | } |
106 | |
107 | static SmallVector<Value> getTiledOperands(LinalgOp producer) { |
108 | return producer->getOperands(); |
109 | } |
110 | |
111 | /// Fuses the producer by cloning the `producer`. The `fusedLoopsAndRanges` |
112 | /// provides the loop range information for the fused loops. The rest are |
113 | /// obtained from the producer itself, since they are not tiled + fused. |
114 | static LinalgOp fuse(OpBuilder &b, LinalgOp producer, |
115 | const DenseMap<unsigned, Range> &fusedLoopsAndRanges) { |
116 | SmallVector<OpFoldResult> ivs, tileSizes, sizeBounds; |
117 | SmallVector<Range> loopRanges; |
118 | Location loc = producer.getLoc(); |
119 | |
120 | for (unsigned i = 0, e = producer.getNumLoops(); i < e; ++i) { |
121 | auto shapeDim = getShapeDefiningLoopRange(producer, i); |
122 | OpFoldResult dim = |
123 | createFoldedDimOp(b, loc, shapeDim.shape, shapeDim.dimension); |
124 | sizeBounds.push_back(Elt: dim); |
125 | auto it = fusedLoopsAndRanges.find(Val: i); |
126 | if (it != fusedLoopsAndRanges.end()) { |
127 | ivs.push_back(Elt: it->second.offset); |
128 | tileSizes.push_back(Elt: it->second.size); |
129 | loopRanges.push_back(Elt: it->second); |
130 | LLVM_DEBUG(llvm::dbgs() << "tiled loop#" << i << " with LoopRange " |
131 | << loopRanges.back() << "\n" ); |
132 | } else { |
133 | tileSizes.push_back(b.getIndexAttr(0)); |
134 | loopRanges.push_back(Elt: Range{b.getIndexAttr(0), .size: dim, b.getIndexAttr(1)}); |
135 | LLVM_DEBUG(llvm::dbgs() << "full loop#" << i << " with LoopRange " |
136 | << loopRanges.back() << "\n" ); |
137 | } |
138 | } |
139 | |
140 | SmallVector<Value, 8> clonedShapes; |
141 | clonedShapes.reserve(N: producer->getNumOperands()); |
142 | |
143 | // Compute subranges for all tensor input/output operands. |
144 | clonedShapes.append(makeTiledShapes( |
145 | b, loc, producer, getTiledOperands(producer), ivs, tileSizes, sizeBounds, |
146 | /**omitPartialTileCheck=*/false)); |
147 | |
148 | // Take result types from the tiled init operands. |
149 | MutableOperandRange producerDpsInits = producer.getDpsInitsMutable(); |
150 | SmallVector<Type, 4> resultTypes; |
151 | resultTypes.reserve(N: producer->getNumResults()); |
152 | int64_t firstInitOperandIdx = |
153 | producerDpsInits.getAsOperandRange().getBeginOperandIndex(); |
154 | for (int64_t i = 0, e = producer->getNumResults(); i < e; ++i) { |
155 | resultTypes.push_back(Elt: clonedShapes[firstInitOperandIdx + i].getType()); |
156 | } |
157 | |
158 | // Clone the producer with new operands and result types. |
159 | LinalgOp clonedOp = clone(b, producer, resultTypes, clonedShapes); |
160 | |
161 | // Shift all IndexOp results by the tile offset. |
162 | SmallVector<OpFoldResult> allIvs = llvm::to_vector( |
163 | Range: llvm::map_range(C&: loopRanges, F: [&](Range range) { return range.offset; })); |
164 | offsetIndices(b, clonedOp, allIvs); |
165 | |
166 | return clonedOp; |
167 | } |
168 | |
169 | /// Get the loop range for a dimension `dim` based on the `shapedOperand`. It is |
170 | /// expected to be defined by a subview op or an extract_slice op. |
171 | static Range getRangeFromOperandShape(OpBuilder &b, Location loc, |
172 | Value shapedOperand, unsigned dim) { |
173 | Operation *shapeProducingOp = shapedOperand.getDefiningOp(); |
174 | if (auto subViewOp = dyn_cast<memref::SubViewOp>(shapeProducingOp)) |
175 | return subViewOp.getOrCreateRanges(b, loc)[dim]; |
176 | if (auto sliceOp = dyn_cast<tensor::ExtractSliceOp>(shapeProducingOp)) |
177 | return sliceOp.getOrCreateRanges(b, loc)[dim]; |
178 | llvm_unreachable("SubviewOp or ExtractSliceOp expected" ); |
179 | } |
180 | |
181 | /// Fuses the producer into the loop immediately enclosing the consumer. |
182 | /// This is achieved by "recomputing" the producer at the time it |
183 | /// is needed just before the consumer. |
184 | static LinalgOp fuse(OpBuilder &b, LinalgOp producerOp, AffineMap producerMap, |
185 | OpOperand &consumerOpOperand) { |
186 | LLVM_DEBUG(llvm::dbgs() << "Producer map: " << producerMap << "\n" ); |
187 | DenseMap<unsigned, Range> fusedLoopsAndRanges; |
188 | Value shapedOperand = consumerOpOperand.get(); |
189 | for (const auto &en : llvm::enumerate(First: producerMap.getResults())) { |
190 | unsigned posInProducerLoop = cast<AffineDimExpr>(Val: en.value()).getPosition(); |
191 | fusedLoopsAndRanges[posInProducerLoop] = getRangeFromOperandShape( |
192 | b, loc: consumerOpOperand.getOwner()->getLoc(), shapedOperand, dim: en.index()); |
193 | } |
194 | return fuse(b, producerOp, fusedLoopsAndRanges); |
195 | } |
196 | |
197 | /// Walk back use-def chain through scf::For yields. |
198 | /// Sets `producer` and `outputIndex` if it finds a producer LinalgOp |
199 | |
200 | // TODO(ravishankarm, ntv): This can be moved into the dependence graphs |
201 | // dependence tracking since the dependence tracking is similar to what is done |
202 | // w.r.t to buffers. |
203 | static void getProducerOfTensor(Value tensor, OpResult &opResult) { |
204 | if (!isa<RankedTensorType>(Val: tensor.getType())) |
205 | return; |
206 | |
207 | while (true) { |
208 | LLVM_DEBUG(llvm::dbgs() << "\ngetProducerOfTensor: " << tensor); |
209 | if (auto linalgOp = tensor.getDefiningOp<LinalgOp>()) { |
210 | opResult = cast<OpResult>(Val&: tensor); |
211 | return; |
212 | } |
213 | if (auto sliceOp = tensor.getDefiningOp<tensor::ExtractSliceOp>()) { |
214 | tensor = sliceOp.getSource(); |
215 | continue; |
216 | } |
217 | if (auto blockArg = dyn_cast<BlockArgument>(Val&: tensor)) { |
218 | if (auto forOp = blockArg.getDefiningOp<scf::ForOp>()) { |
219 | tensor = forOp.getInitArgs()[blockArg.getArgNumber()]; |
220 | continue; |
221 | } |
222 | } |
223 | return; |
224 | } |
225 | } |
226 | |
227 | FailureOr<FusionInfo> |
228 | mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpOperand &consumerOpOperand) { |
229 | Value inputTensor = consumerOpOperand.get(); |
230 | OpResult producerOpResult; |
231 | getProducerOfTensor(tensor: inputTensor, opResult&: producerOpResult); |
232 | if (!producerOpResult) { |
233 | LLVM_DEBUG(llvm::dbgs() << "\nUnable to find producer" ); |
234 | return failure(); |
235 | } |
236 | return fuseProducerOfTensor(b, producerOpResult, consumerOpOperand); |
237 | } |
238 | |
239 | FailureOr<FusionInfo> |
240 | mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpResult producerOpResult, |
241 | OpOperand &consumerOpOperand) { |
242 | auto producerOp = dyn_cast<LinalgOp>(producerOpResult.getOwner()); |
243 | if (!producerOp) |
244 | return failure(); |
245 | |
246 | LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner()); |
247 | if (!consumerOp) |
248 | return failure(); |
249 | |
250 | Value inputTensor = consumerOpOperand.get(); |
251 | |
252 | // Must be an extract_slice op to guarantee there are loops we can fuse into. |
253 | auto sliceOp = inputTensor.getDefiningOp<tensor::ExtractSliceOp>(); |
254 | if (!sliceOp) { |
255 | LLVM_DEBUG(llvm::dbgs() |
256 | << "\nNot fusable, not an extract_slice op: " << inputTensor); |
257 | return failure(); |
258 | } |
259 | |
260 | // If producer is already in the same block as consumer, we are done. |
261 | if (consumerOpOperand.get().getParentBlock() == |
262 | producerOpResult.getParentBlock()) |
263 | return failure(); |
264 | |
265 | // Insert fused `producer` just before `consumer`. |
266 | OpBuilder::InsertionGuard g(b); |
267 | b.setInsertionPoint(consumerOp); |
268 | LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: " << *consumerOp << "\n" ); |
269 | OpOperand *opOperand = |
270 | producerOp.getDpsInitOperand(producerOpResult.getResultNumber()); |
271 | LinalgOp fusedProducer = |
272 | fuse(b, producerOp, producerOp.getMatchingIndexingMap(opOperand), |
273 | consumerOpOperand); |
274 | |
275 | // Replace use. |
276 | Value def = fusedProducer->getResult(producerOpResult.getResultNumber()); |
277 | Type consumerType = consumerOpOperand.get().getType(); |
278 | // Check if rank-reduction occurred as part of the extract_slice. If yes, |
279 | // collapse the dropped dimensions. |
280 | if (cast<ShapedType>(consumerType).getRank() != |
281 | cast<ShapedType>(def.getType()).getRank()) { |
282 | llvm::SmallBitVector droppedDims = sliceOp.getDroppedDims(); |
283 | def = |
284 | tensor::dropGivenUnitDims(b, fusedProducer.getLoc(), def, droppedDims); |
285 | } |
286 | // Canonicalizations are not guaranteed to have happened before constructing |
287 | // `fusedProducer`. In the tensor case this can result in temporary type |
288 | // mismatches. Insert a `tensor.cast` op to propagate the transformation |
289 | // invariant that types are compatible. |
290 | if (consumerType != def.getType()) |
291 | def = b.create<tensor::CastOp>(fusedProducer.getLoc(), consumerType, def); |
292 | consumerOpOperand.set(def); |
293 | return FusionInfo{cast<LinalgOp>(producerOpResult.getOwner()), fusedProducer}; |
294 | } |
295 | |