| 1 | //===- ParallelForToNestedFors.cpp - scf.parallel to nested scf.for ops --===// |
| 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.ParallelOp to nested scf.for ops. |
| 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 | #include "llvm/Support/Debug.h" |
| 18 | |
| 19 | namespace mlir { |
| 20 | #define GEN_PASS_DEF_SCFPARALLELFORTONESTEDFORS |
| 21 | #include "mlir/Dialect/SCF/Transforms/Passes.h.inc" |
| 22 | } // namespace mlir |
| 23 | |
| 24 | #define DEBUG_TYPE "parallel-for-to-nested-fors" |
| 25 | using namespace mlir; |
| 26 | |
| 27 | FailureOr<scf::LoopNest> |
| 28 | mlir::scf::parallelForToNestedFors(RewriterBase &rewriter, |
| 29 | scf::ParallelOp parallelOp) { |
| 30 | |
| 31 | if (!parallelOp.getResults().empty()) |
| 32 | return rewriter.notifyMatchFailure( |
| 33 | arg&: parallelOp, msg: "Currently scf.parallel to scf.for conversion doesn't " |
| 34 | "support scf.parallel with results." ); |
| 35 | |
| 36 | rewriter.setInsertionPoint(parallelOp); |
| 37 | |
| 38 | Location loc = parallelOp.getLoc(); |
| 39 | SmallVector<Value> lowerBounds = parallelOp.getLowerBound(); |
| 40 | SmallVector<Value> upperBounds = parallelOp.getUpperBound(); |
| 41 | SmallVector<Value> steps = parallelOp.getStep(); |
| 42 | |
| 43 | assert(lowerBounds.size() == upperBounds.size() && |
| 44 | lowerBounds.size() == steps.size() && |
| 45 | "Mismatched parallel loop bounds" ); |
| 46 | |
| 47 | scf::LoopNest loopNest = |
| 48 | scf::buildLoopNest(builder&: rewriter, loc, lbs: lowerBounds, ubs: upperBounds, steps); |
| 49 | for (scf::ForOp forOp : loopNest.loops) |
| 50 | forOp.setUnsignedCmp(parallelOp.getUnsignedCmp()); |
| 51 | |
| 52 | SmallVector<Value> newInductionVars = llvm::map_to_vector( |
| 53 | C&: loopNest.loops, F: [](scf::ForOp forOp) { return forOp.getInductionVar(); }); |
| 54 | Block *linearizedBody = loopNest.loops.back().getBody(); |
| 55 | Block *parallelBody = parallelOp.getBody(); |
| 56 | rewriter.eraseOp(op: parallelBody->getTerminator()); |
| 57 | rewriter.inlineBlockBefore(source: parallelBody, op: linearizedBody->getTerminator(), |
| 58 | argValues: newInductionVars); |
| 59 | rewriter.eraseOp(op: parallelOp); |
| 60 | return loopNest; |
| 61 | } |
| 62 | |
| 63 | namespace { |
| 64 | struct ParallelForToNestedFors final |
| 65 | : public impl::SCFParallelForToNestedForsBase<ParallelForToNestedFors> { |
| 66 | void runOnOperation() override { |
| 67 | Operation *parentOp = getOperation(); |
| 68 | IRRewriter rewriter(parentOp->getContext()); |
| 69 | |
| 70 | parentOp->walk( |
| 71 | callback: [&](scf::ParallelOp parallelOp) { |
| 72 | if (failed(Result: scf::parallelForToNestedFors(rewriter, parallelOp))) { |
| 73 | LLVM_DEBUG( |
| 74 | llvm::dbgs() |
| 75 | << "Failed to convert scf.parallel to nested scf.for ops for:\n" |
| 76 | << parallelOp << "\n" ); |
| 77 | return WalkResult::advance(); |
| 78 | } |
| 79 | return WalkResult::advance(); |
| 80 | }); |
| 81 | } |
| 82 | }; |
| 83 | } // namespace |
| 84 | |
| 85 | std::unique_ptr<Pass> mlir::createParallelForToNestedForsPass() { |
| 86 | return std::make_unique<ParallelForToNestedFors>(); |
| 87 | } |
| 88 | |