xref: /llvm-project/mlir/lib/Dialect/SCF/Transforms/ForallToParallel.cpp (revision 6b4c12284795a3030e37b17047271a47a69bb587)
1 //===- ForallToParallel.cpp - scf.forall to scf.parallel loop conversion --===//
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 // Transforms SCF.ForallOp's into SCF.ParallelOps's.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "mlir/Dialect/SCF/IR/SCF.h"
14 #include "mlir/Dialect/SCF/Transforms/Passes.h"
15 #include "mlir/Dialect/SCF/Transforms/Transforms.h"
16 #include "mlir/IR/PatternMatch.h"
17 
18 namespace mlir {
19 #define GEN_PASS_DEF_SCFFORALLTOPARALLELLOOP
20 #include "mlir/Dialect/SCF/Transforms/Passes.h.inc"
21 } // namespace mlir
22 
23 using namespace mlir;
24 
forallToParallelLoop(RewriterBase & rewriter,scf::ForallOp forallOp,scf::ParallelOp * result)25 LogicalResult mlir::scf::forallToParallelLoop(RewriterBase &rewriter,
26                                               scf::ForallOp forallOp,
27                                               scf::ParallelOp *result) {
28   OpBuilder::InsertionGuard guard(rewriter);
29   rewriter.setInsertionPoint(forallOp);
30 
31   Location loc = forallOp.getLoc();
32   if (!forallOp.getOutputs().empty())
33     return rewriter.notifyMatchFailure(
34         forallOp,
35         "only fully bufferized scf.forall ops can be lowered to scf.parallel");
36 
37   // Convert mixed bounds and steps to SSA values.
38   SmallVector<Value> lbs = forallOp.getLowerBound(rewriter);
39   SmallVector<Value> ubs = forallOp.getUpperBound(rewriter);
40   SmallVector<Value> steps = forallOp.getStep(rewriter);
41 
42   // Create empty scf.parallel op.
43   auto parallelOp = rewriter.create<scf::ParallelOp>(loc, lbs, ubs, steps);
44   rewriter.eraseBlock(&parallelOp.getRegion().front());
45   rewriter.inlineRegionBefore(forallOp.getRegion(), parallelOp.getRegion(),
46                               parallelOp.getRegion().begin());
47   // Replace the terminator.
48   rewriter.setInsertionPointToEnd(&parallelOp.getRegion().front());
49   rewriter.replaceOpWithNewOp<scf::ReduceOp>(
50       parallelOp.getRegion().front().getTerminator());
51 
52   // If the mapping attribute is present, propagate to the new parallelOp.
53   if (forallOp.getMapping())
54     parallelOp->setAttr("mapping", *forallOp.getMapping());
55 
56   // Erase the scf.forall op.
57   rewriter.replaceOp(forallOp, parallelOp);
58 
59   if (result)
60     *result = parallelOp;
61 
62   return success();
63 }
64 
65 namespace {
66 struct ForallToParallelLoop final
67     : public impl::SCFForallToParallelLoopBase<ForallToParallelLoop> {
runOnOperation__anonc6d64d9a0111::ForallToParallelLoop68   void runOnOperation() override {
69     Operation *parentOp = getOperation();
70     IRRewriter rewriter(parentOp->getContext());
71 
72     parentOp->walk([&](scf::ForallOp forallOp) {
73       if (failed(scf::forallToParallelLoop(rewriter, forallOp))) {
74         return signalPassFailure();
75       }
76     });
77   }
78 };
79 } // namespace
80 
createForallToParallelLoopPass()81 std::unique_ptr<Pass> mlir::createForallToParallelLoopPass() {
82   return std::make_unique<ForallToParallelLoop>();
83 }
84