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