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