xref: /llvm-project/mlir/lib/Dialect/SCF/Transforms/ParallelLoopCollapsing.cpp (revision 981932bc57aeaa6d4dd02b011d848ed9fa4dd615)
1 //===- ParallelLoopCollapsing.cpp - Pass collapsing parallel loop indices -===//
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 #include "mlir/Dialect/SCF/Transforms/Passes.h"
10 
11 #include "mlir/Dialect/SCF/IR/SCF.h"
12 #include "mlir/Dialect/SCF/Utils/Utils.h"
13 #include "mlir/Transforms/RegionUtils.h"
14 #include "llvm/ADT/SmallSet.h"
15 #include "llvm/Support/CommandLine.h"
16 #include "llvm/Support/Debug.h"
17 
18 namespace mlir {
19 #define GEN_PASS_DEF_TESTSCFPARALLELLOOPCOLLAPSING
20 #include "mlir/Dialect/SCF/Transforms/Passes.h.inc"
21 } // namespace mlir
22 
23 #define DEBUG_TYPE "parallel-loop-collapsing"
24 
25 using namespace mlir;
26 
27 namespace {
28 struct TestSCFParallelLoopCollapsing
29     : public impl::TestSCFParallelLoopCollapsingBase<
30           TestSCFParallelLoopCollapsing> {
31   void runOnOperation() override {
32     Operation *module = getOperation();
33 
34     // The common case for GPU dialect will be simplifying the ParallelOp to 3
35     // arguments, so we do that here to simplify things.
36     llvm::SmallVector<std::vector<unsigned>, 3> combinedLoops;
37 
38     // Gather the input args into the format required by
39     // `collapseParallelLoops`.
40     if (!clCollapsedIndices0.empty())
41       combinedLoops.push_back(clCollapsedIndices0);
42     if (!clCollapsedIndices1.empty()) {
43       if (clCollapsedIndices0.empty()) {
44         llvm::errs()
45             << "collapsed-indices-1 specified but not collapsed-indices-0";
46         signalPassFailure();
47         return;
48       }
49       combinedLoops.push_back(clCollapsedIndices1);
50     }
51     if (!clCollapsedIndices2.empty()) {
52       if (clCollapsedIndices1.empty()) {
53         llvm::errs()
54             << "collapsed-indices-2 specified but not collapsed-indices-1";
55         signalPassFailure();
56         return;
57       }
58       combinedLoops.push_back(clCollapsedIndices2);
59     }
60 
61     if (combinedLoops.empty()) {
62       llvm::errs() << "No collapsed-indices were specified. This pass is only "
63                       "for testing and does not automatically collapse all "
64                       "parallel loops or similar.";
65       signalPassFailure();
66       return;
67     }
68 
69     // Confirm that the specified loops are [0,N) by testing that N values exist
70     // with the maximum value being N-1.
71     llvm::SmallSet<unsigned, 8> flattenedCombinedLoops;
72     unsigned maxCollapsedIndex = 0;
73     for (auto &loops : combinedLoops) {
74       for (auto &loop : loops) {
75         flattenedCombinedLoops.insert(loop);
76         maxCollapsedIndex = std::max(maxCollapsedIndex, loop);
77       }
78     }
79 
80     if (maxCollapsedIndex != flattenedCombinedLoops.size() - 1 ||
81         !flattenedCombinedLoops.contains(maxCollapsedIndex)) {
82       llvm::errs()
83           << "collapsed-indices arguments must include all values [0,N).";
84       signalPassFailure();
85       return;
86     }
87 
88     // Only apply the transformation on parallel loops where the specified
89     // transformation is valid, but do NOT early abort in the case of invalid
90     // loops.
91     module->walk([&](scf::ParallelOp op) {
92       if (flattenedCombinedLoops.size() != op.getNumLoops()) {
93         op.emitOpError("has ")
94             << op.getNumLoops()
95             << " iter args while this limited functionality testing pass was "
96                "configured only for loops with exactly "
97             << flattenedCombinedLoops.size() << " iter args.";
98         return;
99       }
100       collapseParallelLoops(op, combinedLoops);
101     });
102   }
103 };
104 } // namespace
105 
106 std::unique_ptr<Pass> mlir::createTestSCFParallelLoopCollapsingPass() {
107   return std::make_unique<TestSCFParallelLoopCollapsing>();
108 }
109