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/Dialect/Affine/Analysis/Utils.h" |
15 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
16 | #include "mlir/Dialect/Affine/LoopUtils.h" |
17 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
18 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
19 | #include "mlir/Pass/Pass.h" |
20 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
21 | #include "mlir/Transforms/Passes.h" |
22 | |
23 | #define PASS_NAME "test-affine-data-copy" |
24 | |
25 | using namespace mlir; |
26 | using namespace mlir::affine; |
27 | |
28 | namespace { |
29 | |
30 | struct TestAffineDataCopy |
31 | : public PassWrapper<TestAffineDataCopy, OperationPass<func::FuncOp>> { |
32 | MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(TestAffineDataCopy) |
33 | |
34 | StringRef getArgument() const final { return PASS_NAME; } |
35 | StringRef getDescription() const final { |
36 | return "Tests affine data copy utility functions." ; |
37 | } |
38 | TestAffineDataCopy() = default; |
39 | TestAffineDataCopy(const TestAffineDataCopy &pass) : PassWrapper(pass){}; |
40 | |
41 | void getDependentDialects(DialectRegistry ®istry) const override { |
42 | registry.insert<memref::MemRefDialect>(); |
43 | } |
44 | void runOnOperation() override; |
45 | |
46 | private: |
47 | Option<bool> clMemRefFilter{ |
48 | *this, "memref-filter" , |
49 | llvm::cl::desc( |
50 | "Enable memref filter testing in affine data copy optimization" ), |
51 | llvm::cl::init(Val: false)}; |
52 | Option<bool> clTestGenerateCopyForMemRegion{ |
53 | *this, "for-memref-region" , |
54 | llvm::cl::desc("Test copy generation for a single memref region" ), |
55 | llvm::cl::init(Val: false)}; |
56 | Option<uint64_t> clCapacityLimit{ |
57 | *this, "capacity-kib" , |
58 | llvm::cl::desc("Test copy generation enforcing a limit of capacity " |
59 | "(default: unlimited)" ), |
60 | llvm::cl::init(UINT64_MAX)}; |
61 | }; |
62 | |
63 | } // namespace |
64 | |
65 | void TestAffineDataCopy::runOnOperation() { |
66 | // Gather all AffineForOps by loop depth. |
67 | std::vector<SmallVector<AffineForOp, 2>> depthToLoops; |
68 | gatherLoops(getOperation(), depthToLoops); |
69 | if (depthToLoops.empty()) |
70 | return; |
71 | |
72 | // Only support tests with a single loop nest and a single innermost loop |
73 | // for now. |
74 | unsigned innermostLoopIdx = depthToLoops.size() - 1; |
75 | if (depthToLoops[0].size() != 1 || depthToLoops[innermostLoopIdx].size() != 1) |
76 | return; |
77 | |
78 | auto loopNest = depthToLoops[0][0]; |
79 | auto innermostLoop = depthToLoops[innermostLoopIdx][0]; |
80 | AffineLoadOp load; |
81 | // For simplicity, these options are tested on the first memref being loaded |
82 | // from in the innermost loop. |
83 | if (clMemRefFilter || clTestGenerateCopyForMemRegion) { |
84 | for (auto &op : *innermostLoop.getBody()) { |
85 | if (auto ld = dyn_cast<AffineLoadOp>(op)) { |
86 | load = ld; |
87 | break; |
88 | } |
89 | } |
90 | if (!load) |
91 | return; |
92 | } |
93 | |
94 | AffineCopyOptions copyOptions = {/*generateDma=*/false, |
95 | /*slowMemorySpace=*/0, |
96 | /*fastMemorySpace=*/0, |
97 | /*tagMemorySpace=*/0, |
98 | /*fastMemCapacityBytes=*/clCapacityLimit}; |
99 | DenseSet<Operation *> copyNests; |
100 | if (clMemRefFilter) { |
101 | if (failed(affineDataCopyGenerate(loopNest, copyOptions, load.getMemRef(), |
102 | copyNests))) |
103 | return; |
104 | } else if (clTestGenerateCopyForMemRegion) { |
105 | CopyGenerateResult result; |
106 | MemRefRegion region(loopNest.getLoc()); |
107 | if (failed(region.compute(op: load, /*loopDepth=*/0))) |
108 | return; |
109 | if (failed(generateCopyForMemRegion(region, loopNest, copyOptions, result))) |
110 | return; |
111 | } else if (failed(affineDataCopyGenerate(loopNest, copyOptions, |
112 | /*filterMemref=*/std::nullopt, |
113 | copyNests))) { |
114 | return; |
115 | } |
116 | |
117 | // Promote any single iteration loops in the copy nests and simplify |
118 | // load/stores. |
119 | SmallVector<Operation *, 4> copyOps; |
120 | for (Operation *nest : copyNests) { |
121 | // With a post order walk, the erasure of loops does not affect |
122 | // continuation of the walk or the collection of load/store ops. |
123 | nest->walk(callback: [&](Operation *op) { |
124 | if (auto forOp = dyn_cast<AffineForOp>(op)) |
125 | (void)promoteIfSingleIteration(forOp); |
126 | else if (auto loadOp = dyn_cast<AffineLoadOp>(op)) |
127 | copyOps.push_back(Elt: loadOp); |
128 | else if (auto storeOp = dyn_cast<AffineStoreOp>(op)) |
129 | copyOps.push_back(Elt: storeOp); |
130 | }); |
131 | } |
132 | |
133 | // Promoting single iteration loops could lead to simplification of |
134 | // generated load's/store's, and the latter could anyway also be |
135 | // canonicalized. |
136 | RewritePatternSet patterns(&getContext()); |
137 | for (Operation *op : copyOps) { |
138 | patterns.clear(); |
139 | if (isa<AffineLoadOp>(op)) { |
140 | AffineLoadOp::getCanonicalizationPatterns(patterns, &getContext()); |
141 | } else { |
142 | assert(isa<AffineStoreOp>(op) && "expected affine store op" ); |
143 | AffineStoreOp::getCanonicalizationPatterns(patterns, &getContext()); |
144 | } |
145 | } |
146 | GreedyRewriteConfig config; |
147 | config.setStrictness(GreedyRewriteStrictness::ExistingAndNewOps); |
148 | (void)applyOpPatternsGreedily(copyOps, std::move(patterns), config); |
149 | } |
150 | |
151 | namespace mlir { |
152 | void registerTestAffineDataCopyPass() { |
153 | PassRegistration<TestAffineDataCopy>(); |
154 | } |
155 | } // namespace mlir |
156 | |