| 1 | //===- PromoteShuffleToAMDGPU.cpp - Promote shuffle to AMDGPU -------------===// |
| 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 contains patterns to try to promote `gpu.shuffle`s to specialized |
| 10 | // AMDGPU intrinsics. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "mlir/Dialect/GPU/Transforms/ROCDL/Passes.h" |
| 15 | #include "mlir/Dialect/LLVMIR/ROCDLTargetInfo.h" |
| 16 | |
| 17 | #include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h" |
| 18 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 19 | #include "mlir/Dialect/GPU/IR/GPUDialect.h" |
| 20 | #include "mlir/IR/PatternMatch.h" |
| 21 | #include <optional> |
| 22 | |
| 23 | using namespace mlir; |
| 24 | |
| 25 | namespace { |
| 26 | /// Try to promote `gpu.shuffle` to `amdgpu.swizzle_bitmode`, width must be 64 |
| 27 | /// and offset must be a constant integer in the range [0, 31]. |
| 28 | struct PromoteShuffleToSwizzlePattern |
| 29 | : public OpRewritePattern<gpu::ShuffleOp> { |
| 30 | using Base::Base; |
| 31 | |
| 32 | LogicalResult matchAndRewrite(gpu::ShuffleOp op, |
| 33 | PatternRewriter &rewriter) const override { |
| 34 | if (op.getMode() != gpu::ShuffleMode::XOR) |
| 35 | return rewriter.notifyMatchFailure(arg&: op, |
| 36 | msg: "only xor shuffle mode is supported" ); |
| 37 | |
| 38 | if (!isConstantIntValue(ofr: op.getWidth(), value: 64)) |
| 39 | return rewriter.notifyMatchFailure(arg&: op, |
| 40 | msg: "only 64 width shuffle is supported" ); |
| 41 | |
| 42 | std::optional<int64_t> offset = getConstantIntValue(ofr: op.getOffset()); |
| 43 | if (!offset) |
| 44 | return rewriter.notifyMatchFailure(arg&: op, |
| 45 | msg: "offset must be a constant integer" ); |
| 46 | |
| 47 | int64_t offsetValue = *offset; |
| 48 | if (offsetValue < 0 || offsetValue >= 32) |
| 49 | return rewriter.notifyMatchFailure(arg&: op, |
| 50 | msg: "offset must be in the range [0, 31]" ); |
| 51 | |
| 52 | Location loc = op.getLoc(); |
| 53 | Value res = amdgpu::SwizzleBitModeOp::create( |
| 54 | builder&: rewriter, location: loc, result: op.getResult(i: 0).getType(), src: op.getValue(), |
| 55 | /*and_mask=*/31, |
| 56 | /*orMask=*/or_mask: 0, /*xorMask=*/xor_mask: offsetValue); |
| 57 | Value valid = arith::ConstantIntOp::create(builder&: rewriter, location: loc, value: 1, /*width*/ 1); |
| 58 | rewriter.replaceOp(op, newValues: {res, valid}); |
| 59 | return success(); |
| 60 | } |
| 61 | }; |
| 62 | |
| 63 | /// Try to promote `gpu.shuffle` to `amdgpu.permlane_swap`, width must be 64 |
| 64 | /// and offset must be a constant integer in the set {16, 32}. |
| 65 | struct PromoteShuffleToPermlanePattern |
| 66 | : public OpRewritePattern<gpu::ShuffleOp> { |
| 67 | using Base::Base; |
| 68 | |
| 69 | LogicalResult matchAndRewrite(gpu::ShuffleOp op, |
| 70 | PatternRewriter &rewriter) const override { |
| 71 | if (op.getMode() != gpu::ShuffleMode::XOR) |
| 72 | return rewriter.notifyMatchFailure(arg&: op, |
| 73 | msg: "only xor shuffle mode is supported" ); |
| 74 | |
| 75 | if (!isConstantIntValue(ofr: op.getWidth(), value: 64)) |
| 76 | return rewriter.notifyMatchFailure(arg&: op, |
| 77 | msg: "only 64 width shuffle is supported" ); |
| 78 | |
| 79 | std::optional<int64_t> offset = getConstantIntValue(ofr: op.getOffset()); |
| 80 | if (!offset) |
| 81 | return rewriter.notifyMatchFailure(arg&: op, |
| 82 | msg: "offset must be a constant integer" ); |
| 83 | |
| 84 | int64_t offsetValue = *offset; |
| 85 | if (offsetValue != 16 && offsetValue != 32) |
| 86 | return rewriter.notifyMatchFailure(arg&: op, msg: "offset must be either 15 or 31" ); |
| 87 | |
| 88 | Location loc = op.getLoc(); |
| 89 | Value res = amdgpu::PermlaneSwapOp::create( |
| 90 | builder&: rewriter, location: loc, result: op.getResult(i: 0).getType(), src: op.getValue(), row_length: offsetValue); |
| 91 | Value valid = arith::ConstantIntOp::create(builder&: rewriter, location: loc, value: 1, /*width*/ 1); |
| 92 | rewriter.replaceOp(op, newValues: {res, valid}); |
| 93 | return success(); |
| 94 | } |
| 95 | }; |
| 96 | } // namespace |
| 97 | |
| 98 | void mlir::populateGpuPromoteShuffleToAMDGPUPatterns( |
| 99 | RewritePatternSet &patterns, std::optional<ROCDL::TargetInfo> target) { |
| 100 | patterns.add<PromoteShuffleToSwizzlePattern>(arg: patterns.getContext(), |
| 101 | /*benefit*/ args: 1); |
| 102 | if (target && target->has(feature: llvm::AMDGPU::FEAT_PERMLANE32_SWAP)) |
| 103 | patterns.add<PromoteShuffleToPermlanePattern>(arg: patterns.getContext(), |
| 104 | /*benefit*/ args: 2); |
| 105 | } |
| 106 | |