Frank Schlimbach 26b1f61c14
[mlir][shard,mpi] Fixing lowering allgather shard->mpi->llvm (#178870)
`shard.allgather` concatenates along a specified gather-axis. However,
`mpi.allgather` always concatenates along the first dimension and there
is no MPI operation that allows gathering along an arbitrary axis.
Hence, if gather-axis!=0, we need to create a temporary buffer where we
gather along the first dimension and then copy from that buffer to the
final output along the specified gather-axis. This is not ideal by far.

Along the way also
- fixing computation of memref size in mpitollvm
- adding a simple canonicalization pattern for comm_size for easier
debugging
- adding more tests
2026-02-04 18:36:03 +01:00

99 lines
3.4 KiB
C++

//===- MPIOps.cpp - MPI dialect ops implementation ------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/MPI/IR/MPI.h"
#include "mlir/Dialect/MPI/IR/Utils.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinAttributes.h"
using namespace mlir;
using namespace mlir::mpi;
namespace {
// If input memref has dynamic shape and is a cast and if the cast's input has
// static shape, fold the cast's static input into the given operation.
template <typename OpT>
struct FoldCast final : public mlir::OpRewritePattern<OpT> {
using mlir::OpRewritePattern<OpT>::OpRewritePattern;
LogicalResult matchAndRewrite(OpT op,
mlir::PatternRewriter &b) const override {
auto mRef = op.getRef();
if (mRef.getType().hasStaticShape()) {
return mlir::failure();
}
auto defOp = mRef.getDefiningOp();
if (!defOp || !mlir::isa<mlir::memref::CastOp>(defOp)) {
return mlir::failure();
}
auto src = mlir::cast<mlir::memref::CastOp>(defOp).getSource();
if (!src.getType().hasStaticShape()) {
return mlir::failure();
}
op.getRefMutable().assign(src);
return mlir::success();
}
};
struct FoldRank final : public mlir::OpRewritePattern<mlir::mpi::CommRankOp> {
using mlir::OpRewritePattern<mlir::mpi::CommRankOp>::OpRewritePattern;
LogicalResult matchAndRewrite(mlir::mpi::CommRankOp op,
mlir::PatternRewriter &b) const override {
return FoldToDLTIConst(op, "MPI:comm_world_rank", b);
}
};
struct FoldSize final : public mlir::OpRewritePattern<mlir::mpi::CommSizeOp> {
using mlir::OpRewritePattern<mlir::mpi::CommSizeOp>::OpRewritePattern;
LogicalResult matchAndRewrite(mlir::mpi::CommSizeOp op,
mlir::PatternRewriter &b) const override {
return FoldToDLTIConst(op, "MPI:comm_world_size", b);
}
};
} // namespace
void mlir::mpi::SendOp::getCanonicalizationPatterns(
mlir::RewritePatternSet &results, mlir::MLIRContext *context) {
results.add<FoldCast<mlir::mpi::SendOp>>(context);
}
void mlir::mpi::RecvOp::getCanonicalizationPatterns(
mlir::RewritePatternSet &results, mlir::MLIRContext *context) {
results.add<FoldCast<mlir::mpi::RecvOp>>(context);
}
void mlir::mpi::ISendOp::getCanonicalizationPatterns(
mlir::RewritePatternSet &results, mlir::MLIRContext *context) {
results.add<FoldCast<mlir::mpi::ISendOp>>(context);
}
void mlir::mpi::IRecvOp::getCanonicalizationPatterns(
mlir::RewritePatternSet &results, mlir::MLIRContext *context) {
results.add<FoldCast<mlir::mpi::IRecvOp>>(context);
}
void mlir::mpi::CommRankOp::getCanonicalizationPatterns(
mlir::RewritePatternSet &results, mlir::MLIRContext *context) {
results.add<FoldRank>(context);
}
void mlir::mpi::CommSizeOp::getCanonicalizationPatterns(
mlir::RewritePatternSet &results, mlir::MLIRContext *context) {
results.add<FoldSize>(context);
}
//===----------------------------------------------------------------------===//
// TableGen'd op method definitions
//===----------------------------------------------------------------------===//
#define GET_OP_CLASSES
#include "mlir/Dialect/MPI/IR/MPIOps.cpp.inc"