| 1 | //===- ResolveStridedMetadata.cpp - AMDGPU expand_strided_metadata ------===// |
|---|
| 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/AMDGPU/Transforms/Passes.h" |
|---|
| 10 | |
|---|
| 11 | #include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h" |
|---|
| 12 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
|---|
| 13 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
|---|
| 14 | |
|---|
| 15 | namespace mlir::amdgpu { |
|---|
| 16 | #define GEN_PASS_DEF_AMDGPURESOLVESTRIDEDMETADATAPASS |
|---|
| 17 | #include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc" |
|---|
| 18 | } // namespace mlir::amdgpu |
|---|
| 19 | |
|---|
| 20 | using namespace mlir; |
|---|
| 21 | using namespace mlir::amdgpu; |
|---|
| 22 | |
|---|
| 23 | namespace { |
|---|
| 24 | struct AmdgpuResolveStridedMetadataPass |
|---|
| 25 | : public amdgpu::impl::AmdgpuResolveStridedMetadataPassBase< |
|---|
| 26 | AmdgpuResolveStridedMetadataPass> { |
|---|
| 27 | void runOnOperation() override; |
|---|
| 28 | }; |
|---|
| 29 | |
|---|
| 30 | struct final |
|---|
| 31 | : public OpRewritePattern<memref::ExtractStridedMetadataOp> { |
|---|
| 32 | using OpRewritePattern::OpRewritePattern; |
|---|
| 33 | LogicalResult matchAndRewrite(memref::ExtractStridedMetadataOp metadataOp, |
|---|
| 34 | PatternRewriter &rewriter) const override { |
|---|
| 35 | auto castOp = metadataOp.getSource().getDefiningOp<FatRawBufferCastOp>(); |
|---|
| 36 | if (!castOp) |
|---|
| 37 | return rewriter.notifyMatchFailure(metadataOp, |
|---|
| 38 | "not a fat raw buffer cast" ); |
|---|
| 39 | Location loc = castOp.getLoc(); |
|---|
| 40 | auto sourceMetadata = rewriter.create<memref::ExtractStridedMetadataOp>( |
|---|
| 41 | loc, castOp.getSource()); |
|---|
| 42 | SmallVector<Value> results; |
|---|
| 43 | if (metadataOp.getBaseBuffer().use_empty()) { |
|---|
| 44 | results.push_back(Elt: nullptr); |
|---|
| 45 | } else { |
|---|
| 46 | auto baseBufferType = |
|---|
| 47 | cast<MemRefType>(metadataOp.getBaseBuffer().getType()); |
|---|
| 48 | if (baseBufferType == castOp.getResult().getType()) { |
|---|
| 49 | results.push_back(Elt: castOp.getResult()); |
|---|
| 50 | } else { |
|---|
| 51 | results.push_back(rewriter.create<memref::ReinterpretCastOp>( |
|---|
| 52 | loc, baseBufferType, castOp.getResult(), /*offset=*/0, |
|---|
| 53 | /*sizes=*/ArrayRef<int64_t>{}, /*strides=*/ArrayRef<int64_t>{})); |
|---|
| 54 | } |
|---|
| 55 | } |
|---|
| 56 | if (castOp.getResetOffset()) |
|---|
| 57 | results.push_back(rewriter.create<arith::ConstantIndexOp>(location: loc, args: 0)); |
|---|
| 58 | else |
|---|
| 59 | results.push_back(Elt: sourceMetadata.getOffset()); |
|---|
| 60 | llvm::append_range(results, sourceMetadata.getSizes()); |
|---|
| 61 | llvm::append_range(results, sourceMetadata.getStrides()); |
|---|
| 62 | rewriter.replaceOp(metadataOp, results); |
|---|
| 63 | return success(); |
|---|
| 64 | } |
|---|
| 65 | }; |
|---|
| 66 | } // namespace |
|---|
| 67 | |
|---|
| 68 | void mlir::amdgpu::populateAmdgpuResolveStridedMetadataPatterns( |
|---|
| 69 | RewritePatternSet &patterns) { |
|---|
| 70 | patterns.add<ExtractStridedMetadataOnFatRawBufferCastFolder>( |
|---|
| 71 | arg: patterns.getContext()); |
|---|
| 72 | } |
|---|
| 73 | |
|---|
| 74 | void AmdgpuResolveStridedMetadataPass::runOnOperation() { |
|---|
| 75 | RewritePatternSet patterns(&getContext()); |
|---|
| 76 | populateAmdgpuResolveStridedMetadataPatterns(patterns); |
|---|
| 77 | if (failed(applyPatternsGreedily(getOperation(), std::move(patterns)))) |
|---|
| 78 | signalPassFailure(); |
|---|
| 79 | } |
|---|
| 80 | |
|---|
