Skip to content

Lower allreduce #144716

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Jun 23, 2025
Merged

Lower allreduce #144716

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
24bc1ca
Lowering mesh.all_reduce to mpi calls.
fschlimb Mar 27, 2025
7588408
fixing in-place mesh.all_reduce
fschlimb Apr 2, 2025
d114eb8
addressing review nits
fschlimb Jun 16, 2025
58cc509
Apply suggestions from code review
fschlimb Jun 16, 2025
5f7929a
OpClass -> ReductionOp
fschlimb Jun 18, 2025
e901c39
separate conversion and rewerite patterns
fschlimb Jun 18, 2025
27c8373
formatting
fschlimb Jun 23, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions mlir/include/mlir/Conversion/Passes.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -905,6 +905,8 @@ def ConvertMeshToMPIPass : Pass<"convert-mesh-to-mpi"> {
shard/partition sizes depend on the rank.
}];
let dependentDialects = [
"affine::AffineDialect",
"arith::ArithDialect",
"memref::MemRefDialect",
"mpi::MPIDialect",
"scf::SCFDialect",
Expand Down
1 change: 1 addition & 0 deletions mlir/include/mlir/Dialect/MPI/IR/MPI.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,7 @@
#include "mlir/IR/Dialect.h"
#include "mlir/IR/OpDefinition.h"
#include "mlir/IR/OpImplementation.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"

//===----------------------------------------------------------------------===//
// MPIDialect
Expand Down
2 changes: 1 addition & 1 deletion mlir/include/mlir/Dialect/MPI/IR/MPI.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,7 +230,7 @@ def MPI_OpMinloc : I32EnumAttrCase<"MPI_MINLOC", 11, "MPI_MINLOC">;
def MPI_OpMaxloc : I32EnumAttrCase<"MPI_MAXLOC", 12, "MPI_MAXLOC">;
def MPI_OpReplace : I32EnumAttrCase<"MPI_REPLACE", 13, "MPI_REPLACE">;

def MPI_OpClassEnum : I32EnumAttr<"MPI_OpClassEnum", "MPI operation class", [
def MPI_ReductionOpEnum : I32EnumAttr<"MPI_ReductionOpEnum", "MPI operation class", [
MPI_OpNull,
MPI_OpMax,
MPI_OpMin,
Expand Down
12 changes: 7 additions & 5 deletions mlir/include/mlir/Dialect/MPI/IR/MPIOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@

include "mlir/Dialect/MPI/IR/MPI.td"
include "mlir/Dialect/MPI/IR/MPITypes.td"
include "mlir/Interfaces/SideEffectInterfaces.td"

class MPI_Op<string mnemonic, list<Trait> traits = []>
: Op<MPI_Dialect, mnemonic, traits>;
Expand Down Expand Up @@ -41,7 +42,7 @@ def MPI_InitOp : MPI_Op<"init", []> {
// CommWorldOp
//===----------------------------------------------------------------------===//

def MPI_CommWorldOp : MPI_Op<"comm_world", []> {
def MPI_CommWorldOp : MPI_Op<"comm_world", [Pure]> {
let summary = "Get the World communicator, equivalent to `MPI_COMM_WORLD`";
let description = [{
This operation returns the predefined MPI_COMM_WORLD communicator.
Expand All @@ -56,7 +57,7 @@ def MPI_CommWorldOp : MPI_Op<"comm_world", []> {
// CommRankOp
//===----------------------------------------------------------------------===//

def MPI_CommRankOp : MPI_Op<"comm_rank", []> {
def MPI_CommRankOp : MPI_Op<"comm_rank", [Pure]> {
let summary = "Get the current rank, equivalent to "
"`MPI_Comm_rank(comm, &rank)`";
let description = [{
Expand All @@ -72,13 +73,14 @@ def MPI_CommRankOp : MPI_Op<"comm_rank", []> {
);

let assemblyFormat = "`(` $comm `)` attr-dict `:` type(results)";
let hasCanonicalizer = 1;
}

//===----------------------------------------------------------------------===//
// CommSizeOp
//===----------------------------------------------------------------------===//

def MPI_CommSizeOp : MPI_Op<"comm_size", []> {
def MPI_CommSizeOp : MPI_Op<"comm_size", [Pure]> {
let summary = "Get the size of the group associated to the communicator, "
"equivalent to `MPI_Comm_size(comm, &size)`";
let description = [{
Expand All @@ -100,7 +102,7 @@ def MPI_CommSizeOp : MPI_Op<"comm_size", []> {
// CommSplitOp
//===----------------------------------------------------------------------===//

def MPI_CommSplitOp : MPI_Op<"comm_split", []> {
def MPI_CommSplitOp : MPI_Op<"comm_split", [Pure]> {
let summary = "Partition the group associated with the given communicator into "
"disjoint subgroups";
let description = [{
Expand Down Expand Up @@ -281,7 +283,7 @@ def MPI_AllReduceOp : MPI_Op<"allreduce", []> {
let arguments = (
ins AnyMemRef : $sendbuf,
AnyMemRef : $recvbuf,
MPI_OpClassEnum : $op,
MPI_ReductionOpEnum : $op,
MPI_Comm : $comm
);

Expand Down
5 changes: 5 additions & 0 deletions mlir/include/mlir/Dialect/Mesh/IR/MeshOps.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -212,6 +212,11 @@ void maybeInsertSourceShardingAnnotation(MeshSharding sharding,
OpOperand &operand,
OpBuilder &builder);

/// Converts a vector of OpFoldResults (ints) into vector of Values of the
/// provided type.
SmallVector<Value> getMixedAsValues(OpBuilder b, const Location &loc,
llvm::ArrayRef<int64_t> statics,
ValueRange dynamics, Type type = Type());
} // namespace mesh
} // namespace mlir

Expand Down
4 changes: 2 additions & 2 deletions mlir/include/mlir/Dialect/Mesh/IR/MeshOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -584,11 +584,11 @@ def Mesh_AllReduceOp : Mesh_CollectiveCommunicationOpBase<"all_reduce", [
```
}];
let arguments = !con(commonArgs, (ins
AnyRankedTensor:$input,
AnyTypeOf<[AnyMemRef, AnyRankedTensor]>:$input,
Copy link
Contributor

@AntonLydike AntonLydike Jun 18, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I wonder if this is the way to go, isn't the sharding dialect supposed to be tensor only? I would expect MPI to insert some bufferization dialect ops to get to the memref. Although admittedly, my knowledge here is very flaky.

Copy link
Contributor Author

@fschlimb fschlimb Jun 23, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The communication primitives might be added during bufferization, like through implementing the BufferizableOpInterface. Hence they must support tensors and memrefs.

Copy link
Contributor Author

@fschlimb fschlimb Jun 23, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

MPI dialect is strictly buffer semantics. There is no room for tensors there.

DefaultValuedAttr<Mesh_ReductionKindAttr, "::mlir::mesh::ReductionKind::Sum">:$reduction
));
let results = (outs
AnyRankedTensor:$result
AnyTypeOf<[AnyMemRef, AnyRankedTensor]>:$result
);
let assemblyFormat = [{
$input `on` $mesh (`mesh_axes` `=` $mesh_axes^)? (`reduction` `=` $reduction^)?
Expand Down
10 changes: 5 additions & 5 deletions mlir/include/mlir/Dialect/Mesh/Transforms/Simplifications.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,9 +62,9 @@ void populateAllReduceEndomorphismSimplificationPatterns(
auto isEndomorphismOp = [reduction](Operation *op,
std::optional<Operation *> referenceOp) {
auto allReduceOp = llvm::dyn_cast<AllReduceOp>(op);
if (!allReduceOp ||
allReduceOp.getInput().getType().getElementType() !=
allReduceOp.getResult().getType().getElementType() ||
auto inType = cast<ShapedType>(allReduceOp.getInput().getType());
auto outType = cast<ShapedType>(allReduceOp.getResult().getType());
if (!allReduceOp || inType.getElementType() != outType.getElementType() ||
allReduceOp.getReduction() != reduction) {
return false;
}
Expand All @@ -83,9 +83,9 @@ void populateAllReduceEndomorphismSimplificationPatterns(
}

auto refAllReduceOp = llvm::dyn_cast<AllReduceOp>(referenceOp.value());
auto refType = cast<ShapedType>(refAllReduceOp.getResult().getType());
return refAllReduceOp->getAttrs() == allReduceOp->getAttrs() &&
allReduceOp.getInput().getType().getElementType() ==
refAllReduceOp.getInput().getType().getElementType();
inType.getElementType() == refType.getElementType();
};
auto isAlgebraicOp = [](Operation *op) {
return static_cast<bool>(llvm::dyn_cast<AlgebraicOp>(op));
Expand Down
5 changes: 5 additions & 0 deletions mlir/include/mlir/Dialect/Mesh/Transforms/Transforms.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,11 @@ createCollectiveProcessGroupSize(MeshOp mesh, ArrayRef<MeshAxis> axes,
TypedValue<IndexType> createProcessLinearIndex(StringRef mesh,
ArrayRef<MeshAxis> meshAxes,
ImplicitLocOpBuilder &builder);
// Get process linear index from a multi-index along the given mesh axes .
TypedValue<IndexType>
createProcessLinearIndex(StringRef mesh, ValueRange processInGroupMultiIndex,
ArrayRef<MeshAxis> meshAxes,
ImplicitLocOpBuilder &builder);

} // namespace mesh
} // namespace mlir
Expand Down
62 changes: 31 additions & 31 deletions mlir/lib/Conversion/MPIToLLVM/MPIToLLVM.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -116,7 +116,7 @@ class MPIImplTraits {
/// enum value.
virtual Value getMPIOp(const Location loc,
ConversionPatternRewriter &rewriter,
mpi::MPI_OpClassEnum opAttr) = 0;
mpi::MPI_ReductionOpEnum opAttr) = 0;
};

//===----------------------------------------------------------------------===//
Expand Down Expand Up @@ -199,49 +199,49 @@ class MPICHImplTraits : public MPIImplTraits {
}

Value getMPIOp(const Location loc, ConversionPatternRewriter &rewriter,
mpi::MPI_OpClassEnum opAttr) override {
mpi::MPI_ReductionOpEnum opAttr) override {
int32_t op = MPI_NO_OP;
switch (opAttr) {
case mpi::MPI_OpClassEnum::MPI_OP_NULL:
case mpi::MPI_ReductionOpEnum::MPI_OP_NULL:
op = MPI_NO_OP;
break;
case mpi::MPI_OpClassEnum::MPI_MAX:
case mpi::MPI_ReductionOpEnum::MPI_MAX:
op = MPI_MAX;
break;
case mpi::MPI_OpClassEnum::MPI_MIN:
case mpi::MPI_ReductionOpEnum::MPI_MIN:
op = MPI_MIN;
break;
case mpi::MPI_OpClassEnum::MPI_SUM:
case mpi::MPI_ReductionOpEnum::MPI_SUM:
op = MPI_SUM;
break;
case mpi::MPI_OpClassEnum::MPI_PROD:
case mpi::MPI_ReductionOpEnum::MPI_PROD:
op = MPI_PROD;
break;
case mpi::MPI_OpClassEnum::MPI_LAND:
case mpi::MPI_ReductionOpEnum::MPI_LAND:
op = MPI_LAND;
break;
case mpi::MPI_OpClassEnum::MPI_BAND:
case mpi::MPI_ReductionOpEnum::MPI_BAND:
op = MPI_BAND;
break;
case mpi::MPI_OpClassEnum::MPI_LOR:
case mpi::MPI_ReductionOpEnum::MPI_LOR:
op = MPI_LOR;
break;
case mpi::MPI_OpClassEnum::MPI_BOR:
case mpi::MPI_ReductionOpEnum::MPI_BOR:
op = MPI_BOR;
break;
case mpi::MPI_OpClassEnum::MPI_LXOR:
case mpi::MPI_ReductionOpEnum::MPI_LXOR:
op = MPI_LXOR;
break;
case mpi::MPI_OpClassEnum::MPI_BXOR:
case mpi::MPI_ReductionOpEnum::MPI_BXOR:
op = MPI_BXOR;
break;
case mpi::MPI_OpClassEnum::MPI_MINLOC:
case mpi::MPI_ReductionOpEnum::MPI_MINLOC:
op = MPI_MINLOC;
break;
case mpi::MPI_OpClassEnum::MPI_MAXLOC:
case mpi::MPI_ReductionOpEnum::MPI_MAXLOC:
op = MPI_MAXLOC;
break;
case mpi::MPI_OpClassEnum::MPI_REPLACE:
case mpi::MPI_ReductionOpEnum::MPI_REPLACE:
op = MPI_REPLACE;
break;
}
Expand Down Expand Up @@ -336,49 +336,49 @@ class OMPIImplTraits : public MPIImplTraits {
}

Value getMPIOp(const Location loc, ConversionPatternRewriter &rewriter,
mpi::MPI_OpClassEnum opAttr) override {
mpi::MPI_ReductionOpEnum opAttr) override {
StringRef op;
switch (opAttr) {
case mpi::MPI_OpClassEnum::MPI_OP_NULL:
case mpi::MPI_ReductionOpEnum::MPI_OP_NULL:
op = "ompi_mpi_no_op";
break;
case mpi::MPI_OpClassEnum::MPI_MAX:
case mpi::MPI_ReductionOpEnum::MPI_MAX:
op = "ompi_mpi_max";
break;
case mpi::MPI_OpClassEnum::MPI_MIN:
case mpi::MPI_ReductionOpEnum::MPI_MIN:
op = "ompi_mpi_min";
break;
case mpi::MPI_OpClassEnum::MPI_SUM:
case mpi::MPI_ReductionOpEnum::MPI_SUM:
op = "ompi_mpi_sum";
break;
case mpi::MPI_OpClassEnum::MPI_PROD:
case mpi::MPI_ReductionOpEnum::MPI_PROD:
op = "ompi_mpi_prod";
break;
case mpi::MPI_OpClassEnum::MPI_LAND:
case mpi::MPI_ReductionOpEnum::MPI_LAND:
op = "ompi_mpi_land";
break;
case mpi::MPI_OpClassEnum::MPI_BAND:
case mpi::MPI_ReductionOpEnum::MPI_BAND:
op = "ompi_mpi_band";
break;
case mpi::MPI_OpClassEnum::MPI_LOR:
case mpi::MPI_ReductionOpEnum::MPI_LOR:
op = "ompi_mpi_lor";
break;
case mpi::MPI_OpClassEnum::MPI_BOR:
case mpi::MPI_ReductionOpEnum::MPI_BOR:
op = "ompi_mpi_bor";
break;
case mpi::MPI_OpClassEnum::MPI_LXOR:
case mpi::MPI_ReductionOpEnum::MPI_LXOR:
op = "ompi_mpi_lxor";
break;
case mpi::MPI_OpClassEnum::MPI_BXOR:
case mpi::MPI_ReductionOpEnum::MPI_BXOR:
op = "ompi_mpi_bxor";
break;
case mpi::MPI_OpClassEnum::MPI_MINLOC:
case mpi::MPI_ReductionOpEnum::MPI_MINLOC:
op = "ompi_mpi_minloc";
break;
case mpi::MPI_OpClassEnum::MPI_MAXLOC:
case mpi::MPI_ReductionOpEnum::MPI_MAXLOC:
op = "ompi_mpi_maxloc";
break;
case mpi::MPI_OpClassEnum::MPI_REPLACE:
case mpi::MPI_ReductionOpEnum::MPI_REPLACE:
op = "ompi_mpi_replace";
break;
}
Expand Down
Loading