1 //===- ParallelLoopTiling.cpp - Tiles loop.parallel ---------------===// 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 loop tiling on parallel loops. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "PassDetail.h" 14 #include "mlir/Dialect/Affine/IR/AffineOps.h" 15 #include "mlir/Dialect/SCF/Passes.h" 16 #include "mlir/Dialect/SCF/SCF.h" 17 #include "mlir/Dialect/SCF/Transforms.h" 18 #include "mlir/Dialect/StandardOps/IR/Ops.h" 19 #include "mlir/Transforms/RegionUtils.h" 20 #include "llvm/Support/CommandLine.h" 21 22 using namespace mlir; 23 using namespace mlir::scf; 24 25 /// Tile a parallel loop of the form 26 /// loop.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3) 27 /// step (%arg4, %arg5) 28 /// 29 /// into 30 /// loop.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3) 31 /// step (%arg4*tileSize[0], 32 /// %arg5*tileSize[1]) 33 /// loop.parallel (%j0, %j1) = (0, 0) to (min(tileSize[0], %arg2-%j0) 34 /// min(tileSize[1], %arg3-%j1)) 35 /// step (%arg4, %arg5) 36 /// The old loop is replaced with the new one. 37 void mlir::scf::tileParallelLoop(ParallelOp op, ArrayRef<int64_t> tileSizes) { 38 OpBuilder b(op); 39 auto zero = b.create<ConstantIndexOp>(op.getLoc(), 0); 40 SmallVector<Value, 2> tileSizeConstants; 41 tileSizeConstants.reserve(op.upperBound().size()); 42 for (size_t i = 0, end = op.upperBound().size(); i != end; ++i) { 43 if (i < tileSizes.size()) 44 tileSizeConstants.push_back( 45 b.create<ConstantIndexOp>(op.getLoc(), tileSizes[i])); 46 else 47 // Just pick 1 for the remaining dimensions. 48 tileSizeConstants.push_back(b.create<ConstantIndexOp>(op.getLoc(), 1)); 49 } 50 51 // Create the outer loop with adjusted steps. 52 SmallVector<Value, 2> newSteps; 53 newSteps.reserve(op.step().size()); 54 for (auto step : llvm::zip(op.step(), tileSizeConstants)) { 55 newSteps.push_back( 56 b.create<MulIOp>(op.getLoc(), std::get<0>(step), std::get<1>(step))); 57 } 58 auto outerLoop = b.create<ParallelOp>(op.getLoc(), op.lowerBound(), 59 op.upperBound(), newSteps); 60 b.setInsertionPointToStart(outerLoop.getBody()); 61 62 // Compute min(size, dim - offset) to avoid out-of-bounds accesses. 63 // FIXME: Instead of using min, we want to replicate the tail. This would give 64 // the inner loop constant bounds for easy vectorization. 65 auto minMap = AffineMap::get( 66 /*dimCount=*/3, /*symbolCount=*/0, 67 {getAffineDimExpr(/*position=*/0, b.getContext()), 68 getAffineDimExpr(/*position=*/1, b.getContext()) - 69 getAffineDimExpr(/*position=*/2, b.getContext())}, 70 b.getContext()); 71 72 // Create the inner loop with adjusted bounds. 73 SmallVector<Value, 2> newBounds; 74 newBounds.reserve(op.upperBound().size()); 75 for (auto bounds : llvm::zip(tileSizeConstants, outerLoop.upperBound(), 76 outerLoop.getInductionVars())) { 77 newBounds.push_back(b.create<AffineMinOp>( 78 op.getLoc(), b.getIndexType(), minMap, 79 ValueRange{std::get<0>(bounds), std::get<1>(bounds), 80 std::get<2>(bounds)})); 81 } 82 auto innerLoop = b.create<ParallelOp>( 83 op.getLoc(), SmallVector<Value, 2>(newBounds.size(), zero), newBounds, 84 op.step()); 85 86 // Steal the body of the old parallel loop and erase it. 87 innerLoop.region().takeBody(op.region()); 88 op.erase(); 89 } 90 91 /// Get a list of most nested parallel loops. Assumes that ParallelOps are only 92 /// directly nested. 93 static bool getInnermostNestedLoops(Block *block, 94 SmallVectorImpl<ParallelOp> &loops) { 95 bool hasInnerLoop = false; 96 for (auto parallelOp : block->getOps<ParallelOp>()) { 97 hasInnerLoop = true; 98 if (!getInnermostNestedLoops(parallelOp.getBody(), loops)) 99 loops.push_back(parallelOp); 100 } 101 return hasInnerLoop; 102 } 103 104 namespace { 105 struct ParallelLoopTiling 106 : public LoopParallelLoopTilingBase<ParallelLoopTiling> { 107 ParallelLoopTiling() = default; 108 explicit ParallelLoopTiling(ArrayRef<int64_t> tileSizes) { 109 this->tileSizes = tileSizes; 110 } 111 112 void runOnFunction() override { 113 SmallVector<ParallelOp, 2> mostNestedParallelOps; 114 for (Block &block : getFunction()) { 115 getInnermostNestedLoops(&block, mostNestedParallelOps); 116 } 117 for (ParallelOp pLoop : mostNestedParallelOps) { 118 tileParallelLoop(pLoop, tileSizes); 119 } 120 } 121 }; 122 } // namespace 123 124 std::unique_ptr<Pass> 125 mlir::createParallelLoopTilingPass(ArrayRef<int64_t> tileSizes) { 126 return std::make_unique<ParallelLoopTiling>(tileSizes); 127 } 128