Files
llvm-project/mlir/lib/Dialect/Vector/Transforms/LowerVectorShuffle.cpp

111 lines
4.0 KiB
C++

//===- LowerVectorShuffle.cpp - Lower 'vector.shuffle' operation ----------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements the lowering of complex `vector.shuffle` operation to a
// set of simpler operations supported by LLVM/SPIR-V.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/Dialect/Vector/Transforms/LoweringPatterns.h"
#include "mlir/IR/PatternMatch.h"
#define DEBUG_TYPE "vector-shuffle-lowering"
using namespace mlir;
using namespace mlir::vector;
namespace {
/// Lowers a `vector.shuffle` operation with mixed-size inputs to a new
/// `vector.shuffle` which promotes the smaller input to the larger vector size
/// and an updated version of the original `vector.shuffle`.
///
/// Example:
///
/// %0 = vector.shuffle %v1, %v2 [0, 2, 1, 3] : vector<2xf32>, vector<4xf32>
///
/// is lowered to:
///
/// %0 = vector.shuffle %v1, %v1 [0, 1, -1, -1] :
/// vector<2xf32>, vector<2xf32>
/// %1 = vector.shuffle %0, %v2 [0, 4, 1, 5] :
/// vector<4xf32>, vector<4xf32>
///
/// Note: This transformation helps legalize vector.shuffle ops when lowering
/// to SPIR-V/LLVM, which don't support shuffle operations with mixed-size
/// inputs.
///
struct MixedSizeInputShuffleOpRewrite final
: OpRewritePattern<vector::ShuffleOp> {
using Base::Base;
LogicalResult matchAndRewrite(vector::ShuffleOp shuffleOp,
PatternRewriter &rewriter) const override {
auto v1Type = shuffleOp.getV1VectorType();
auto v2Type = shuffleOp.getV2VectorType();
// Only support 1-D shuffle for now.
if (v1Type.getRank() != 1 || v2Type.getRank() != 1)
return failure();
// Bail out if inputs don't have mixed sizes.
int64_t v1OrigNumElems = v1Type.getNumElements();
int64_t v2OrigNumElems = v2Type.getNumElements();
if (v1OrigNumElems == v2OrigNumElems)
return failure();
// Determine which input needs promotion.
bool promoteV1 = v1OrigNumElems < v2OrigNumElems;
Value inputToPromote = promoteV1 ? shuffleOp.getV1() : shuffleOp.getV2();
VectorType promotedType = promoteV1 ? v2Type : v1Type;
int64_t origNumElems = promoteV1 ? v1OrigNumElems : v2OrigNumElems;
int64_t promotedNumElems = promoteV1 ? v2OrigNumElems : v1OrigNumElems;
// Create a shuffle with a mask that preserves existing elements and fills
// up with poison.
SmallVector<int64_t> promoteMask(promotedNumElems, ShuffleOp::kPoisonIndex);
for (int64_t i = 0; i < origNumElems; ++i)
promoteMask[i] = i;
Value promotedInput =
vector::ShuffleOp::create(rewriter, shuffleOp.getLoc(), promotedType,
inputToPromote, inputToPromote, promoteMask);
// Create the final shuffle with the promoted inputs.
Value promotedV1 = promoteV1 ? promotedInput : shuffleOp.getV1();
Value promotedV2 = promoteV1 ? shuffleOp.getV2() : promotedInput;
SmallVector<int64_t> newMask;
if (!promoteV1) {
newMask = to_vector(shuffleOp.getMask());
} else {
// Adjust V2 indices to account for the new V1 size.
for (auto idx : shuffleOp.getMask()) {
int64_t newIdx = idx;
if (idx >= v1OrigNumElems) {
newIdx += promotedNumElems - v1OrigNumElems;
}
newMask.push_back(newIdx);
}
}
rewriter.replaceOpWithNewOp<vector::ShuffleOp>(
shuffleOp, shuffleOp.getResultVectorType(), promotedV1, promotedV2,
newMask);
return success();
}
};
} // namespace
void mlir::vector::populateVectorShuffleLoweringPatterns(
RewritePatternSet &patterns, PatternBenefit benefit) {
patterns.add<MixedSizeInputShuffleOpRewrite>(patterns.getContext(), benefit);
}