1 //===- TestAffineDataCopy.cpp - Test affine data copy utility -------------===// 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 a pass to test affine data copy utility functions and 10 // options. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "mlir/Analysis/Utils.h" 15 #include "mlir/Dialect/Affine/IR/AffineOps.h" 16 #include "mlir/Dialect/MemRef/IR/MemRef.h" 17 #include "mlir/Pass/Pass.h" 18 #include "mlir/Transforms/GreedyPatternRewriteDriver.h" 19 #include "mlir/Transforms/LoopUtils.h" 20 #include "mlir/Transforms/Passes.h" 21 22 #define PASS_NAME "test-affine-data-copy" 23 24 using namespace mlir; 25 26 static llvm::cl::OptionCategory clOptionsCategory(PASS_NAME " options"); 27 28 namespace { 29 30 struct TestAffineDataCopy 31 : public PassWrapper<TestAffineDataCopy, FunctionPass> { 32 StringRef getArgument() const final { return PASS_NAME; } 33 StringRef getDescription() const final { 34 return "Tests affine data copy utility functions."; 35 } 36 TestAffineDataCopy() = default; 37 TestAffineDataCopy(const TestAffineDataCopy &pass){}; 38 39 void getDependentDialects(DialectRegistry ®istry) const override { 40 registry.insert<memref::MemRefDialect>(); 41 } 42 void runOnFunction() override; 43 44 private: 45 Option<bool> clMemRefFilter{ 46 *this, "memref-filter", 47 llvm::cl::desc( 48 "Enable memref filter testing in affine data copy optimization"), 49 llvm::cl::init(false)}; 50 Option<bool> clTestGenerateCopyForMemRegion{ 51 *this, "for-memref-region", 52 llvm::cl::desc("Test copy generation for a single memref region"), 53 llvm::cl::init(false)}; 54 }; 55 56 } // end anonymous namespace 57 58 void TestAffineDataCopy::runOnFunction() { 59 // Gather all AffineForOps by loop depth. 60 std::vector<SmallVector<AffineForOp, 2>> depthToLoops; 61 gatherLoops(getFunction(), depthToLoops); 62 assert(depthToLoops.size() && "Loop nest not found"); 63 64 // Only support tests with a single loop nest and a single innermost loop 65 // for now. 66 unsigned innermostLoopIdx = depthToLoops.size() - 1; 67 if (depthToLoops[0].size() != 1 || depthToLoops[innermostLoopIdx].size() != 1) 68 return; 69 70 auto loopNest = depthToLoops[0][0]; 71 auto innermostLoop = depthToLoops[innermostLoopIdx][0]; 72 AffineLoadOp load; 73 if (clMemRefFilter || clTestGenerateCopyForMemRegion) { 74 // Gather MemRef filter. For simplicity, we use the first loaded memref 75 // found in the innermost loop. 76 for (auto &op : *innermostLoop.getBody()) { 77 if (auto ld = dyn_cast<AffineLoadOp>(op)) { 78 load = ld; 79 break; 80 } 81 } 82 } 83 if (!load) 84 return; 85 86 AffineCopyOptions copyOptions = {/*generateDma=*/false, 87 /*slowMemorySpace=*/0, 88 /*fastMemorySpace=*/0, 89 /*tagMemorySpace=*/0, 90 /*fastMemCapacityBytes=*/32 * 1024 * 1024UL}; 91 DenseSet<Operation *> copyNests; 92 if (clMemRefFilter) { 93 affineDataCopyGenerate(loopNest, copyOptions, load.getMemRef(), copyNests); 94 } else if (clTestGenerateCopyForMemRegion) { 95 CopyGenerateResult result; 96 MemRefRegion region(loopNest.getLoc()); 97 (void)region.compute(load, /*loopDepth=*/0); 98 (void)generateCopyForMemRegion(region, loopNest, copyOptions, result); 99 } 100 101 // Promote any single iteration loops in the copy nests and simplify 102 // load/stores. 103 SmallVector<Operation *, 4> copyOps; 104 for (Operation *nest : copyNests) { 105 // With a post order walk, the erasure of loops does not affect 106 // continuation of the walk or the collection of load/store ops. 107 nest->walk([&](Operation *op) { 108 if (auto forOp = dyn_cast<AffineForOp>(op)) 109 (void)promoteIfSingleIteration(forOp); 110 else if (auto loadOp = dyn_cast<AffineLoadOp>(op)) 111 copyOps.push_back(loadOp); 112 else if (auto storeOp = dyn_cast<AffineStoreOp>(op)) 113 copyOps.push_back(storeOp); 114 }); 115 } 116 117 // Promoting single iteration loops could lead to simplification of 118 // generated load's/store's, and the latter could anyway also be 119 // canonicalized. 120 RewritePatternSet patterns(&getContext()); 121 for (Operation *op : copyOps) { 122 patterns.clear(); 123 if (isa<AffineLoadOp>(op)) { 124 AffineLoadOp::getCanonicalizationPatterns(patterns, &getContext()); 125 } else { 126 assert(isa<AffineStoreOp>(op) && "expected affine store op"); 127 AffineStoreOp::getCanonicalizationPatterns(patterns, &getContext()); 128 } 129 } 130 (void)applyOpPatternsAndFold(copyOps, std::move(patterns), /*strict=*/true); 131 } 132 133 namespace mlir { 134 void registerTestAffineDataCopyPass() { 135 PassRegistration<TestAffineDataCopy>(); 136 } 137 } // namespace mlir 138