6512ca7ddb
The motivation is to avoid having to negate `isDynamic*` checks, avoid double negations, and allow for `ShapedType::isStaticDim` to be used in ADT functions without having to wrap it in a lambda performing the negation. Also add the new functions to C and Python bindings.
1579 lines
57 KiB
C++
1579 lines
57 KiB
C++
//===- MeshOps.cpp - Mesh Dialect Operations ------------------------------===//
|
|
//
|
|
// 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/Mesh/IR/MeshOps.h"
|
|
|
|
#include "mlir/Dialect/Arith/IR/Arith.h"
|
|
#include "mlir/Dialect/Mesh/IR/MeshDialect.h"
|
|
#include "mlir/Dialect/Utils/StaticValueUtils.h"
|
|
#include "mlir/IR/Attributes.h"
|
|
#include "mlir/IR/BuiltinAttributes.h"
|
|
#include "mlir/IR/BuiltinTypeInterfaces.h"
|
|
#include "mlir/IR/BuiltinTypes.h"
|
|
#include "mlir/IR/Diagnostics.h"
|
|
#include "mlir/IR/DialectImplementation.h"
|
|
#include "mlir/IR/IRMapping.h"
|
|
#include "mlir/IR/Location.h"
|
|
#include "mlir/IR/PatternMatch.h"
|
|
#include "mlir/IR/TypeUtilities.h"
|
|
#include "mlir/IR/Value.h"
|
|
#include "mlir/Interfaces/ViewLikeInterface.h"
|
|
#include "mlir/Support/LLVM.h"
|
|
#include "mlir/Transforms/InliningUtils.h"
|
|
#include "llvm/ADT/ArrayRef.h"
|
|
#include "llvm/ADT/STLExtras.h"
|
|
#include "llvm/ADT/SmallSet.h"
|
|
#include "llvm/ADT/SmallVector.h"
|
|
#include "llvm/ADT/TypeSwitch.h"
|
|
#include "llvm/Support/Casting.h"
|
|
#include <algorithm>
|
|
#include <functional>
|
|
#include <iterator>
|
|
#include <numeric>
|
|
#include <optional>
|
|
#include <utility>
|
|
|
|
#define DEBUG_TYPE "mesh-ops"
|
|
#define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE << "]: ")
|
|
|
|
using namespace mlir;
|
|
using namespace mlir::mesh;
|
|
|
|
#include "mlir/Dialect/Mesh/IR/MeshDialect.cpp.inc"
|
|
|
|
namespace {
|
|
|
|
struct DimensionSize {
|
|
static DimensionSize dynamic() { return DimensionSize(ShapedType::kDynamic); }
|
|
DimensionSize(int64_t val) : val(val) {}
|
|
int64_t value() const { return val; }
|
|
operator int64_t() const { return val; }
|
|
bool isDynamic() const { return ShapedType::isDynamic(val); }
|
|
|
|
private:
|
|
int64_t val;
|
|
};
|
|
|
|
} // namespace
|
|
|
|
static DimensionSize operator/(DimensionSize lhs, DimensionSize rhs) {
|
|
if (lhs.isDynamic() || rhs.isDynamic()) {
|
|
return DimensionSize::dynamic();
|
|
}
|
|
return lhs.value() / rhs.value();
|
|
}
|
|
|
|
static DimensionSize operator*(DimensionSize lhs, DimensionSize rhs) {
|
|
if (lhs.isDynamic() || rhs.isDynamic()) {
|
|
return DimensionSize::dynamic();
|
|
}
|
|
return lhs.value() * rhs.value();
|
|
}
|
|
|
|
SmallVector<Value> mlir::mesh::getMixedAsValues(OpBuilder b,
|
|
const Location &loc,
|
|
llvm::ArrayRef<int64_t> statics,
|
|
ValueRange dynamics,
|
|
Type type) {
|
|
SmallVector<Value> values;
|
|
auto dyn = dynamics.begin();
|
|
Type i64 = b.getI64Type();
|
|
if (!type)
|
|
type = i64;
|
|
assert((i64 == type || b.getIndexType() == type) &&
|
|
"expected an i64 or an intex type");
|
|
for (auto s : statics) {
|
|
if (s == ShapedType::kDynamic) {
|
|
values.emplace_back(*(dyn++));
|
|
} else {
|
|
TypedAttr val = type == i64 ? b.getI64IntegerAttr(s) : b.getIndexAttr(s);
|
|
values.emplace_back(b.create<arith::ConstantOp>(loc, type, val));
|
|
}
|
|
}
|
|
return values;
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Inliner
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
namespace {
|
|
struct MeshInlinerInterface : public DialectInlinerInterface {
|
|
using DialectInlinerInterface::DialectInlinerInterface;
|
|
// Currently no restrictions are encoded for inlining.
|
|
bool isLegalToInline(Operation *, Operation *, bool) const final {
|
|
return true;
|
|
}
|
|
bool isLegalToInline(Region *, Region *, bool, IRMapping &) const final {
|
|
return true;
|
|
}
|
|
bool isLegalToInline(Operation *, Region *, bool, IRMapping &) const final {
|
|
return true;
|
|
}
|
|
};
|
|
} // namespace
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Mesh dialect
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
void MeshDialect::initialize() {
|
|
addOperations<
|
|
#define GET_OP_LIST
|
|
#include "mlir/Dialect/Mesh/IR/MeshOps.cpp.inc"
|
|
>();
|
|
addAttributes<
|
|
#define GET_ATTRDEF_LIST
|
|
#include "mlir/Dialect/Mesh/IR/MeshAttributes.cpp.inc"
|
|
>();
|
|
addTypes<
|
|
#define GET_TYPEDEF_LIST
|
|
#include "mlir/Dialect/Mesh/IR/MeshTypes.cpp.inc"
|
|
>();
|
|
addInterface<MeshInlinerInterface>();
|
|
}
|
|
|
|
Operation *MeshDialect::materializeConstant(OpBuilder &builder, Attribute value,
|
|
Type type, Location loc) {
|
|
return arith::ConstantOp::materialize(builder, value, type, loc);
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Mesh utilities
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
static FailureOr<MeshOp> getMeshAndVerify(Operation *op,
|
|
FlatSymbolRefAttr meshSymbol,
|
|
SymbolTableCollection &symbolTable) {
|
|
mesh::MeshOp mesh = getMeshOrNull(op, meshSymbol, symbolTable);
|
|
if (!mesh) {
|
|
return op->emitError() << "Undefined required mesh symbol \""
|
|
<< meshSymbol.getValue() << "\".";
|
|
}
|
|
|
|
return mesh;
|
|
}
|
|
|
|
template <typename It>
|
|
bool isUnique(It begin, It end) {
|
|
if (begin == end) {
|
|
return true;
|
|
}
|
|
It next = std::next(begin);
|
|
if (next == end) {
|
|
return true;
|
|
}
|
|
for (; next != end; ++next, ++begin) {
|
|
if (*begin == *next) {
|
|
return false;
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
|
|
static LogicalResult verifyMeshAxes(Location loc, ArrayRef<MeshAxis> axes,
|
|
MeshOp mesh) {
|
|
SmallVector<MeshAxis> sorted = llvm::to_vector(axes);
|
|
llvm::sort(sorted);
|
|
if (!isUnique(sorted.begin(), sorted.end())) {
|
|
return emitError(loc) << "Mesh axes contains duplicate elements.";
|
|
}
|
|
|
|
MeshAxis rank = mesh.getRank();
|
|
for (auto axis : axes) {
|
|
if (axis >= rank || axis < 0) {
|
|
return emitError(loc)
|
|
<< "0-based mesh axis index " << axis
|
|
<< " is out of bounds. The referenced mesh \"" << mesh.getSymName()
|
|
<< "\" is of rank " << rank << ".";
|
|
}
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
template <typename Op>
|
|
static FailureOr<MeshOp>
|
|
getMeshAndVerifyAxes(Op op, SymbolTableCollection &symbolTable) {
|
|
auto mesh =
|
|
::getMeshAndVerify(op.getOperation(), op.getMeshAttr(), symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyMeshAxes(op.getLoc(), op.getMeshAxes(), mesh.value()))) {
|
|
return failure();
|
|
}
|
|
return mesh;
|
|
}
|
|
|
|
template <typename InShape, typename MeshShape, typename SplitAxes,
|
|
typename OutShape>
|
|
static void shardShape(const InShape &inShape, const MeshShape &meshShape,
|
|
const SplitAxes &splitAxes, OutShape &outShape,
|
|
ArrayRef<int64_t> shardedDimsOffsets = {},
|
|
ArrayRef<int64_t> haloSizes = {}) {
|
|
// 0d tensors cannot be sharded and must get replicated
|
|
if (inShape.empty()) {
|
|
assert(outShape.empty());
|
|
return;
|
|
}
|
|
|
|
std::copy(llvm::adl_begin(inShape), llvm::adl_end(inShape),
|
|
llvm::adl_begin(outShape));
|
|
|
|
if (!shardedDimsOffsets.empty()) {
|
|
auto isDynShape = ShapedType::isDynamicShape(meshShape);
|
|
uint64_t pos = 1;
|
|
for (auto [tensorAxis, innerSplitAxes] : llvm::enumerate(splitAxes)) {
|
|
if (!innerSplitAxes.empty()) {
|
|
auto sz = shardedDimsOffsets[pos];
|
|
bool same = !isDynShape;
|
|
if (same) {
|
|
// Find sharded dims in shardedDimsOffsets with same static size on
|
|
// all devices. Use kDynamic for dimensions with dynamic or
|
|
// non-uniform offs in shardedDimsOffsets.
|
|
uint64_t numShards = 0;
|
|
for (auto i : innerSplitAxes.asArrayRef()) {
|
|
numShards += meshShape[i];
|
|
}
|
|
for (size_t i = 1; i < numShards; ++i) {
|
|
if (shardedDimsOffsets[pos + i] - shardedDimsOffsets[pos + i - 1] !=
|
|
sz) {
|
|
same = false;
|
|
break;
|
|
}
|
|
}
|
|
pos += numShards + 1;
|
|
}
|
|
outShape[tensorAxis] = same ? sz : ShapedType::kDynamic;
|
|
}
|
|
}
|
|
} else {
|
|
for (auto [tensorAxis, innerSplitAxes] : llvm::enumerate(splitAxes)) {
|
|
outShape[tensorAxis] = shardDimension(
|
|
inShape[tensorAxis],
|
|
collectiveProcessGroupSize(innerSplitAxes.asArrayRef(), meshShape));
|
|
}
|
|
|
|
if (!haloSizes.empty()) {
|
|
// add halo sizes if requested
|
|
int haloAxis = 0;
|
|
for (auto [tensorAxis, innerSplitAxes] : llvm::enumerate(splitAxes)) {
|
|
if (ShapedType::isStatic(outShape[tensorAxis]) &&
|
|
!innerSplitAxes.empty()) {
|
|
if (haloSizes[haloAxis * 2] >= 0 &&
|
|
haloSizes[haloAxis * 2 + 1] >= 0) {
|
|
outShape[tensorAxis] +=
|
|
haloSizes[haloAxis * 2] + haloSizes[haloAxis * 2 + 1];
|
|
++haloAxis;
|
|
} else {
|
|
outShape[tensorAxis] = ShapedType::kDynamic;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
ShapedType mesh::shardShapedType(ShapedType shape, MeshOp mesh,
|
|
MeshSharding sharding) {
|
|
using Dim = std::decay_t<decltype(shape.getDimSize(0))>;
|
|
SmallVector<Dim> resShapeArr(shape.getShape().size());
|
|
shardShape(shape.getShape(), mesh.getShape(), sharding.getSplitAxes(),
|
|
resShapeArr, sharding.getStaticShardedDimsOffsets(),
|
|
sharding.getStaticHaloSizes());
|
|
return shape.clone(resShapeArr);
|
|
}
|
|
|
|
Type mesh::shardType(Type type, MeshOp mesh, MeshSharding sharding) {
|
|
RankedTensorType rankedTensorType = dyn_cast<RankedTensorType>(type);
|
|
if (rankedTensorType && !rankedTensorType.getShape().empty()) {
|
|
return shardShapedType(rankedTensorType, mesh, sharding);
|
|
}
|
|
return type;
|
|
}
|
|
|
|
static void maybeInsertTargetShardingAnnotationImpl(MeshSharding sharding,
|
|
Value &operandValue,
|
|
Operation *operandOp,
|
|
OpBuilder &builder,
|
|
ShardOp &newShardOp) {
|
|
OpBuilder::InsertionGuard insertionGuard(builder);
|
|
builder.setInsertionPointAfterValue(operandValue);
|
|
ShardOp shardOp = dyn_cast<ShardOp>(operandOp);
|
|
if (shardOp && sharding == shardOp.getSharding() &&
|
|
!shardOp.getAnnotateForUsers()) {
|
|
// No need for anything if the correct sharding is already set.
|
|
if (!newShardOp) {
|
|
newShardOp = shardOp;
|
|
}
|
|
return;
|
|
}
|
|
|
|
if (!newShardOp) {
|
|
auto shardingOp =
|
|
builder.create<ShardingOp>(operandValue.getLoc(), sharding);
|
|
newShardOp =
|
|
builder.create<ShardOp>(operandValue.getLoc(), operandValue, shardingOp,
|
|
/*annotate_for_users*/ false);
|
|
}
|
|
operandValue.replaceUsesWithIf(
|
|
newShardOp, [operandOp, operandValue](OpOperand &use) {
|
|
return use.getOwner() == operandOp && use.get() == operandValue;
|
|
});
|
|
|
|
if (!shardOp || shardOp.getAnnotateForUsers()) {
|
|
return;
|
|
}
|
|
|
|
auto newShardOp2 = builder.create<ShardOp>(operandValue.getLoc(), newShardOp,
|
|
newShardOp.getSharding(),
|
|
/*annotate_for_users*/ true);
|
|
newShardOp.getResult().replaceAllUsesExcept(newShardOp2, newShardOp2);
|
|
}
|
|
|
|
void mlir::mesh::maybeInsertTargetShardingAnnotation(MeshSharding sharding,
|
|
OpResult result,
|
|
OpBuilder &builder) {
|
|
ShardOp newShardOp;
|
|
SmallVector<std::pair<Value, Operation *>> uses;
|
|
for (auto &use : result.getUses()) {
|
|
uses.emplace_back(use.get(), use.getOwner());
|
|
}
|
|
for (auto &[operandValue, operandOp] : uses) {
|
|
maybeInsertTargetShardingAnnotationImpl(sharding, operandValue, operandOp,
|
|
builder, newShardOp);
|
|
}
|
|
}
|
|
|
|
void mlir::mesh::maybeInsertSourceShardingAnnotation(MeshSharding sharding,
|
|
OpOperand &operand,
|
|
OpBuilder &builder) {
|
|
OpBuilder::InsertionGuard insertionGuard(builder);
|
|
Value operandValue = operand.get();
|
|
Operation *operandSrcOp = operandValue.getDefiningOp();
|
|
bool isBlockArg = !operandSrcOp;
|
|
{
|
|
[[maybe_unused]] auto opType =
|
|
dyn_cast<mlir::RankedTensorType>(operandValue.getType());
|
|
assert(!opType || opType.getRank() > 0 || isFullReplication(sharding));
|
|
}
|
|
if (!isa<RankedTensorType>(operandValue.getType()) && operandSrcOp &&
|
|
operandSrcOp->hasTrait<OpTrait::ConstantLike>()) {
|
|
return;
|
|
}
|
|
|
|
Operation *operandOp = operand.getOwner();
|
|
ShardOp shardOp = dyn_cast_or_null<ShardOp>(operandSrcOp);
|
|
|
|
if (shardOp && sharding == shardOp.getSharding() &&
|
|
shardOp.getAnnotateForUsers()) {
|
|
// No need for anything the correct sharding is already set.
|
|
return;
|
|
}
|
|
|
|
builder.setInsertionPoint(operandOp);
|
|
auto shardingOp =
|
|
builder.create<ShardingOp>(operand.get().getLoc(), sharding);
|
|
auto newShardOp =
|
|
builder.create<ShardOp>(operandValue.getLoc(), operandValue, shardingOp,
|
|
/*annotate_for_users*/ true);
|
|
IRRewriter rewriter(builder);
|
|
rewriter.replaceUsesWithIf(
|
|
operandValue, newShardOp, [operandOp, operandValue](OpOperand &use) {
|
|
return use.getOwner() == operandOp && use.get() == operandValue;
|
|
});
|
|
|
|
if (isBlockArg || !shardOp || !shardOp.getAnnotateForUsers()) {
|
|
// No need for resharding.
|
|
return;
|
|
}
|
|
|
|
builder.setInsertionPoint(newShardOp);
|
|
auto newPreceedingShardOp =
|
|
builder.create<ShardOp>(operandValue.getLoc(), operandValue, shardingOp,
|
|
/*annotate_for_users*/ false);
|
|
rewriter.replaceUsesWithIf(
|
|
newShardOp.getSrc(), newPreceedingShardOp, [&newShardOp](OpOperand &use) {
|
|
return use.getOwner() == newShardOp.getOperation();
|
|
});
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.mesh op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult MeshOp::verify() {
|
|
int64_t rank = getRank();
|
|
|
|
if (rank <= 0)
|
|
return emitOpError("rank of mesh is expected to be a positive integer");
|
|
|
|
for (int64_t dimSize : getShape()) {
|
|
if (dimSize < 0 && ShapedType::isStatic(dimSize))
|
|
return emitOpError("dimension size of a mesh is expected to be "
|
|
"non-negative or dynamic");
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.mesh_shape op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
MeshShapeOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = ::getMeshAndVerify(getOperation(), getMeshAttr(), symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyMeshAxes(getLoc(), getAxes(), mesh.value()))) {
|
|
return failure();
|
|
}
|
|
|
|
size_t expectedResultsCount =
|
|
getAxes().empty() ? mesh->getRank() : getAxes().size();
|
|
if (getResult().size() != expectedResultsCount) {
|
|
return emitError() << "Unexpected number of results " << getResult().size()
|
|
<< ". Expected " << expectedResultsCount << ".";
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
void MeshShapeOp::build(OpBuilder &odsBuilder, OperationState &odsState,
|
|
MeshOp mesh) {
|
|
build(odsBuilder, odsState, mesh, SmallVector<MeshAxis>());
|
|
}
|
|
|
|
void MeshShapeOp::build(OpBuilder &odsBuilder, OperationState &odsState,
|
|
MeshOp mesh, ArrayRef<MeshAxis> axes) {
|
|
build(odsBuilder, odsState,
|
|
SmallVector<Type>(axes.empty() ? mesh.getRank() : axes.size(),
|
|
odsBuilder.getIndexType()),
|
|
mesh.getSymName(), MeshAxesAttr::get(odsBuilder.getContext(), axes));
|
|
}
|
|
|
|
void MeshShapeOp::build(OpBuilder &odsBuilder, OperationState &odsState,
|
|
StringRef mesh, ArrayRef<MeshAxis> axes) {
|
|
assert(!axes.empty());
|
|
build(odsBuilder, odsState,
|
|
SmallVector<Type>(axes.size(), odsBuilder.getIndexType()), mesh,
|
|
MeshAxesAttr::get(odsBuilder.getContext(), axes));
|
|
}
|
|
|
|
void MeshShapeOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResults()[0], "mesh_shape");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.sharding
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
void ShardingOp::build(::mlir::OpBuilder &b, ::mlir::OperationState &odsState,
|
|
FlatSymbolRefAttr mesh,
|
|
ArrayRef<MeshAxesAttr> split_axes,
|
|
ArrayRef<MeshAxis> partial_axes,
|
|
mesh::ReductionKind partial_type,
|
|
ArrayRef<int64_t> static_halos,
|
|
ArrayRef<int64_t> static_offsets) {
|
|
return build(
|
|
b, odsState, mesh, MeshAxesArrayAttr::get(b.getContext(), split_axes),
|
|
::mlir::DenseI16ArrayAttr::get(b.getContext(), partial_axes),
|
|
::mlir::mesh::ReductionKindAttr::get(b.getContext(), partial_type),
|
|
::mlir::DenseI64ArrayAttr::get(b.getContext(), static_halos), {},
|
|
::mlir::DenseI64ArrayAttr::get(b.getContext(), static_offsets), {});
|
|
}
|
|
|
|
void ShardingOp::build(::mlir::OpBuilder &b, ::mlir::OperationState &odsState,
|
|
FlatSymbolRefAttr mesh,
|
|
ArrayRef<MeshAxesAttr> split_axes) {
|
|
return build(
|
|
b, odsState, mesh, MeshAxesArrayAttr::get(b.getContext(), split_axes), {},
|
|
::mlir::mesh::ReductionKindAttr::get(b.getContext(), ReductionKind::Sum),
|
|
{}, {}, {}, {});
|
|
}
|
|
|
|
void ShardingOp::build(::mlir::OpBuilder &b, ::mlir::OperationState &odsState,
|
|
llvm::StringRef mesh, ArrayRef<MeshAxesAttr> split_axes,
|
|
ArrayRef<int64_t> static_halos,
|
|
ArrayRef<int64_t> static_offsets) {
|
|
return build(
|
|
b, odsState, FlatSymbolRefAttr::get(b.getContext(), mesh),
|
|
MeshAxesArrayAttr::get(b.getContext(), split_axes), {},
|
|
::mlir::mesh::ReductionKindAttr::get(b.getContext(), ReductionKind::Sum),
|
|
::mlir::DenseI64ArrayAttr::get(b.getContext(), static_halos), {},
|
|
::mlir::DenseI64ArrayAttr::get(b.getContext(), static_offsets), {});
|
|
}
|
|
|
|
void ShardingOp::build(
|
|
::mlir::OpBuilder &b, ::mlir::OperationState &odsState,
|
|
FlatSymbolRefAttr mesh, ArrayRef<MeshAxesAttr> split_axes,
|
|
::mlir::ArrayRef<::mlir::OpFoldResult> halo_sizes,
|
|
::mlir::ArrayRef<::mlir::OpFoldResult> sharded_dims_offsets) {
|
|
mlir::SmallVector<int64_t> staticHalos, staticDims;
|
|
mlir::SmallVector<mlir::Value> dynamicHalos, dynamicDims;
|
|
dispatchIndexOpFoldResults(halo_sizes, dynamicHalos, staticHalos);
|
|
dispatchIndexOpFoldResults(sharded_dims_offsets, dynamicDims, staticDims);
|
|
return build(
|
|
b, odsState, mesh, MeshAxesArrayAttr::get(b.getContext(), split_axes), {},
|
|
::mlir::mesh::ReductionKindAttr::get(b.getContext(), ReductionKind::Sum),
|
|
::mlir::DenseI64ArrayAttr::get(b.getContext(), staticHalos), dynamicHalos,
|
|
::mlir::DenseI64ArrayAttr::get(b.getContext(), staticDims), dynamicDims);
|
|
}
|
|
|
|
void ShardingOp::build(::mlir::OpBuilder &b, ::mlir::OperationState &odsState,
|
|
mlir::mesh::MeshSharding from) {
|
|
|
|
build(b, odsState, ShardingType::get(b.getContext()), from.getMeshAttr(),
|
|
MeshAxesArrayAttr::get(b.getContext(), from.getSplitAxes()),
|
|
from.getPartialAxes().empty()
|
|
? DenseI16ArrayAttr()
|
|
: b.getDenseI16ArrayAttr(from.getPartialAxes()),
|
|
::mlir::mesh::ReductionKindAttr::get(b.getContext(),
|
|
from.getPartialType()),
|
|
from.getStaticShardedDimsOffsets().empty()
|
|
? DenseI64ArrayAttr()
|
|
: b.getDenseI64ArrayAttr(from.getStaticShardedDimsOffsets()),
|
|
from.getDynamicShardedDimsOffsets(),
|
|
from.getStaticHaloSizes().empty()
|
|
? DenseI64ArrayAttr()
|
|
: b.getDenseI64ArrayAttr(from.getStaticHaloSizes()),
|
|
from.getDynamicHaloSizes());
|
|
}
|
|
|
|
LogicalResult ShardingOp::verify() {
|
|
llvm::SmallSet<MeshAxis, 4> visitedAxes;
|
|
|
|
auto checkMeshAxis = [&](ArrayRef<MeshAxis> axesArray) -> LogicalResult {
|
|
for (MeshAxis axis : axesArray) {
|
|
if (axis < 0)
|
|
return emitError() << "mesh axis is expected to be non-negative";
|
|
if (!visitedAxes.insert(axis).second)
|
|
return emitError() << "mesh axis duplicated";
|
|
}
|
|
return success();
|
|
};
|
|
|
|
for (auto subAxes : getSplitAxes().getAxes()) {
|
|
ArrayRef<MeshAxis> subAxesArray = subAxes.asArrayRef();
|
|
if (failed(checkMeshAxis(subAxesArray)))
|
|
return failure();
|
|
}
|
|
if (getPartialAxes().has_value() &&
|
|
failed(checkMeshAxis(getPartialAxes().value())))
|
|
return failure();
|
|
|
|
if (!getStaticHaloSizes().empty() && !getStaticShardedDimsOffsets().empty()) {
|
|
return emitOpError("halo sizes and shard offsets are mutually exclusive");
|
|
}
|
|
|
|
if (!getStaticHaloSizes().empty()) {
|
|
auto numSplitAxes = getSplitAxes().getAxes().size();
|
|
for (auto splitAxis : getSplitAxes().getAxes()) {
|
|
if (splitAxis.empty()) {
|
|
--numSplitAxes;
|
|
}
|
|
}
|
|
if (getStaticHaloSizes().size() != numSplitAxes * 2) {
|
|
return emitError() << "halo sizes must be specified for all split axes.";
|
|
}
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
void ShardingOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "sharding");
|
|
}
|
|
|
|
LogicalResult ShardingOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = ::getMeshAndVerify(getOperation(), getMeshAttr(), symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (mlir::ShapedType::isDynamicShape(mesh->getShape()) &&
|
|
getStaticShardedDimsOffsets().size() > 0) {
|
|
return emitError() << "sharded dims offsets are not allowed for "
|
|
"devices meshes with dynamic shape.";
|
|
}
|
|
|
|
auto shardedDimsOffsets = getStaticShardedDimsOffsets();
|
|
if (!shardedDimsOffsets.empty()) {
|
|
auto meshShape = mesh.value().getShape();
|
|
assert(ShapedType::isStaticShape(meshShape));
|
|
uint64_t pos = 0;
|
|
for (auto [tensorAxis, innerSplitAxes] : llvm::enumerate(getSplitAxes())) {
|
|
if (!innerSplitAxes.empty()) {
|
|
int64_t numShards = 0, off = 0;
|
|
for (auto i : innerSplitAxes.asArrayRef()) {
|
|
numShards += meshShape[i];
|
|
}
|
|
for (int64_t i = 0; i <= numShards; ++i) {
|
|
if (shardedDimsOffsets.size() <= pos + i) {
|
|
return emitError() << "sharded dims offsets has wrong size.";
|
|
}
|
|
if (ShapedType::isStatic(shardedDimsOffsets[pos + i])) {
|
|
if (shardedDimsOffsets[pos + i] < off) {
|
|
return emitError()
|
|
<< "sharded dims offsets must be non-decreasing.";
|
|
}
|
|
off = shardedDimsOffsets[pos + i];
|
|
}
|
|
}
|
|
pos += numShards + 1;
|
|
}
|
|
}
|
|
}
|
|
return success();
|
|
}
|
|
|
|
namespace {
|
|
// Sharding annotations "halo sizes" and "sharded dims offsets"
|
|
// are a mix of attributes and dynamic values. This canonicalization moves
|
|
// constant values to the respective attribute lists, minimizing the number
|
|
// of values.
|
|
// It also removes sharded_dims_sizes and halos if they are effectively "empty".
|
|
class NormalizeSharding final : public OpRewritePattern<ShardingOp> {
|
|
public:
|
|
using OpRewritePattern<ShardingOp>::OpRewritePattern;
|
|
|
|
LogicalResult matchAndRewrite(ShardingOp op,
|
|
PatternRewriter &b) const override {
|
|
auto mixedHalos =
|
|
getMixedValues(op.getStaticHaloSizes(), op.getDynamicHaloSizes(), b);
|
|
auto mixedOffs = getMixedValues(op.getStaticShardedDimsOffsets(),
|
|
op.getDynamicShardedDimsOffsets(), b);
|
|
|
|
// No constant operands were folded, just return;
|
|
bool modified = succeeded(foldDynamicIndexList(mixedHalos, true)) ||
|
|
succeeded(foldDynamicIndexList(mixedOffs, true));
|
|
|
|
auto [staticHalos, dynamicHalos] = decomposeMixedValues(mixedHalos);
|
|
auto [staticOffs, dynamicOffs] = decomposeMixedValues(mixedOffs);
|
|
|
|
if (dynamicHalos.empty() && !staticHalos.empty()) {
|
|
if (staticHalos[0] == 0 && llvm::all_equal(staticHalos)) {
|
|
staticHalos.clear();
|
|
modified = true;
|
|
}
|
|
}
|
|
|
|
// Remove sharded dims offsets if they are effectively the default values,
|
|
// e.g. if they define equi-distance between all neighboring shards.
|
|
// Requires static-only offsets. Compares the first distance as the
|
|
// difference between the first two offsets. Only if all consecutive
|
|
// distances are the same, the offsets are removed.
|
|
if (dynamicOffs.empty() && !staticOffs.empty()) {
|
|
assert(staticOffs.size() >= 2);
|
|
auto diff = staticOffs[1] - staticOffs[0];
|
|
bool all_same = staticOffs.size() > 2;
|
|
for (auto i = 2u; i < staticOffs.size(); ++i) {
|
|
if (staticOffs[i] - staticOffs[i - 1] != diff) {
|
|
all_same = false;
|
|
break;
|
|
}
|
|
}
|
|
if (all_same) {
|
|
staticOffs.clear();
|
|
modified = true;
|
|
}
|
|
}
|
|
|
|
if (!modified) {
|
|
return failure();
|
|
}
|
|
|
|
op.setStaticHaloSizes(staticHalos);
|
|
op.getDynamicHaloSizesMutable().assign(dynamicHalos);
|
|
op.setStaticShardedDimsOffsets(staticOffs);
|
|
op.getDynamicShardedDimsOffsetsMutable().assign(dynamicOffs);
|
|
|
|
return success();
|
|
}
|
|
};
|
|
} // namespace
|
|
|
|
void ShardingOp::getCanonicalizationPatterns(mlir::RewritePatternSet &results,
|
|
mlir::MLIRContext *context) {
|
|
results.add<NormalizeSharding>(context);
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// MeshSharding
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
bool MeshSharding::equalSplitAndPartialAxes(const MeshSharding &rhs) const {
|
|
if (getMesh() != rhs.getMesh()) {
|
|
return false;
|
|
}
|
|
|
|
if (getPartialAxes().size() != rhs.getPartialAxes().size() ||
|
|
(!getPartialAxes().empty() && getPartialType() != rhs.getPartialType()) ||
|
|
!llvm::equal(getPartialAxes(), rhs.getPartialAxes())) {
|
|
return false;
|
|
}
|
|
|
|
auto minSize = std::min(getSplitAxes().size(), rhs.getSplitAxes().size());
|
|
if (!llvm::equal(llvm::make_range(getSplitAxes().begin(),
|
|
getSplitAxes().begin() + minSize),
|
|
llvm::make_range(rhs.getSplitAxes().begin(),
|
|
rhs.getSplitAxes().begin() + minSize))) {
|
|
return false;
|
|
}
|
|
|
|
return llvm::all_of(llvm::drop_begin(getSplitAxes(), minSize),
|
|
std::mem_fn(&MeshAxesAttr::empty)) &&
|
|
llvm::all_of(llvm::drop_begin(rhs.getSplitAxes(), minSize),
|
|
std::mem_fn(&MeshAxesAttr::empty));
|
|
}
|
|
|
|
bool MeshSharding::equalHaloAndShardSizes(const MeshSharding &rhs) const {
|
|
return equalShardSizes(rhs) && equalHaloSizes(rhs);
|
|
}
|
|
|
|
bool MeshSharding::equalShardSizes(const MeshSharding &rhs) const {
|
|
if (rhs.getStaticShardedDimsOffsets().size() !=
|
|
getStaticShardedDimsOffsets().size() ||
|
|
!llvm::equal(getStaticShardedDimsOffsets(),
|
|
rhs.getStaticShardedDimsOffsets())) {
|
|
return false;
|
|
}
|
|
if (rhs.getDynamicShardedDimsOffsets().size() !=
|
|
getDynamicShardedDimsOffsets().size() ||
|
|
!llvm::equal(getDynamicShardedDimsOffsets(),
|
|
rhs.getDynamicShardedDimsOffsets())) {
|
|
return false;
|
|
}
|
|
return true;
|
|
}
|
|
|
|
bool MeshSharding::equalHaloSizes(const MeshSharding &rhs) const {
|
|
if (rhs.getStaticHaloSizes().size() != getStaticHaloSizes().size() ||
|
|
!llvm::equal(getStaticHaloSizes(), rhs.getStaticHaloSizes())) {
|
|
return false;
|
|
}
|
|
if (rhs.getDynamicHaloSizes().size() != getDynamicHaloSizes().size() ||
|
|
!llvm::equal(getDynamicHaloSizes(), rhs.getDynamicHaloSizes())) {
|
|
return false;
|
|
}
|
|
return true;
|
|
}
|
|
|
|
bool MeshSharding::operator==(Value rhs) const {
|
|
return equalSplitAndPartialAxes(rhs) && equalHaloAndShardSizes(rhs);
|
|
}
|
|
|
|
bool MeshSharding::operator!=(Value rhs) const { return !(*this == rhs); }
|
|
|
|
bool MeshSharding::operator==(const MeshSharding &rhs) const {
|
|
return equalSplitAndPartialAxes(rhs) && equalHaloAndShardSizes(rhs);
|
|
}
|
|
|
|
bool MeshSharding::operator!=(const MeshSharding &rhs) const {
|
|
return !(*this == rhs);
|
|
}
|
|
|
|
MeshSharding::MeshSharding(::mlir::FlatSymbolRefAttr mesh_) : mesh(mesh_) {}
|
|
|
|
MeshSharding::MeshSharding(Value rhs) {
|
|
auto shardingOp = mlir::dyn_cast<ShardingOp>(rhs.getDefiningOp());
|
|
assert(shardingOp && "expected sharding op");
|
|
auto splitAxes = shardingOp.getSplitAxes().getAxes();
|
|
auto partialAxes = shardingOp.getPartialAxes().value_or(ArrayRef<MeshAxis>());
|
|
// If splitAxes and partialAxes are empty, use "empty" constructor.
|
|
if (splitAxes.empty() && partialAxes.empty()) {
|
|
*this = MeshSharding(shardingOp.getMeshAttr());
|
|
return;
|
|
}
|
|
*this = get(shardingOp.getMeshAttr(), splitAxes, partialAxes,
|
|
shardingOp.getPartialType().value_or(ReductionKind::Sum),
|
|
shardingOp.getStaticHaloSizes(),
|
|
shardingOp.getStaticShardedDimsOffsets(),
|
|
SmallVector<Value>(shardingOp.getDynamicHaloSizes()),
|
|
SmallVector<Value>(shardingOp.getDynamicShardedDimsOffsets()));
|
|
}
|
|
|
|
MeshSharding MeshSharding::get(::mlir::FlatSymbolRefAttr mesh_,
|
|
ArrayRef<MeshAxesAttr> split_axes_,
|
|
ArrayRef<MeshAxis> partial_axes_,
|
|
ReductionKind partial_type_,
|
|
ArrayRef<int64_t> static_halo_sizes_,
|
|
ArrayRef<int64_t> static_sharded_dims_offsets_,
|
|
ArrayRef<Value> dynamic_halo_sizes_,
|
|
ArrayRef<Value> dynamic_sharded_dims_offsets_) {
|
|
MeshSharding res(mesh_);
|
|
if (split_axes_.empty() && partial_axes_.empty()) {
|
|
return res;
|
|
}
|
|
|
|
res.split_axes.resize(split_axes_.size());
|
|
for (auto [i, axis] : llvm::enumerate(split_axes_)) {
|
|
res.split_axes[i] =
|
|
MeshAxesAttr::get(mesh_.getContext(), axis.asArrayRef());
|
|
}
|
|
|
|
auto clone = [](const auto src, auto &dst) {
|
|
dst.resize(src.size());
|
|
llvm::copy(src, dst.begin());
|
|
};
|
|
|
|
clone(partial_axes_, res.partial_axes);
|
|
res.partial_type = partial_type_;
|
|
clone(static_halo_sizes_, res.static_halo_sizes);
|
|
clone(static_sharded_dims_offsets_, res.static_sharded_dims_offsets);
|
|
clone(dynamic_halo_sizes_, res.dynamic_halo_sizes);
|
|
clone(dynamic_sharded_dims_offsets_, res.dynamic_sharded_dims_offsets);
|
|
|
|
return res;
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.shard_shape
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
void ShardShapeOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult()[0], "shard_shape");
|
|
}
|
|
|
|
void ShardShapeOp::build(::mlir::OpBuilder &odsBuilder,
|
|
::mlir::OperationState &odsState,
|
|
::llvm::ArrayRef<int64_t> dims,
|
|
ArrayRef<Value> dims_dyn, ::mlir::Value sharding,
|
|
::mlir::ValueRange device) {
|
|
SmallVector<mlir::Type> resType(dims.size(), odsBuilder.getIndexType());
|
|
build(odsBuilder, odsState, resType, dims, dims_dyn, sharding,
|
|
SmallVector<int64_t>(device.size(), ShapedType::kDynamic), device);
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.shard op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
void ShardOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "sharding_annotated");
|
|
}
|
|
|
|
namespace {
|
|
// Determine if the given ShardOp is a duplicate of another ShardOp
|
|
// on the same value. This can happen if constant values are sharded.
|
|
class FoldDuplicateShardOp final : public OpRewritePattern<ShardOp> {
|
|
public:
|
|
using OpRewritePattern<ShardOp>::OpRewritePattern;
|
|
|
|
LogicalResult matchAndRewrite(ShardOp op, PatternRewriter &b) const override {
|
|
// Get the use-list of the value being sharded and check if it has more than
|
|
// one use.
|
|
Value value = op.getSrc();
|
|
if (value.hasOneUse() || value.getDefiningOp<ShardOp>()) {
|
|
return failure();
|
|
}
|
|
|
|
// Iterate through the uses of the value to find a duplicate ShardOp.
|
|
for (auto &use : value.getUses()) {
|
|
if (use.getOwner() != op.getOperation()) {
|
|
auto otherOp = dyn_cast<ShardOp>(use.getOwner());
|
|
if (!otherOp || !otherOp->isBeforeInBlock(op)) {
|
|
return failure();
|
|
}
|
|
// Create a MeshSharding object for the current and the other ShardOp
|
|
// If the two are equal replace current op with the other op.
|
|
MeshSharding currentSharding(op.getSharding());
|
|
MeshSharding otherSharding(otherOp.getSharding());
|
|
if (currentSharding == otherSharding) {
|
|
b.replaceAllUsesWith(op.getResult(), otherOp.getResult());
|
|
b.eraseOp(op.getOperation());
|
|
} else {
|
|
// use the other sharding as input for op
|
|
op.getSrcMutable().assign(otherOp.getResult());
|
|
}
|
|
return success();
|
|
}
|
|
}
|
|
|
|
return failure();
|
|
}
|
|
};
|
|
} // namespace
|
|
|
|
void ShardOp::getCanonicalizationPatterns(mlir::RewritePatternSet &results,
|
|
mlir::MLIRContext *context) {
|
|
results.add<FoldDuplicateShardOp>(context);
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.process_multi_index op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
ProcessMultiIndexOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = ::getMeshAndVerify(getOperation(), getMeshAttr(), symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyMeshAxes(getLoc(), getAxes(), mesh.value()))) {
|
|
return failure();
|
|
}
|
|
|
|
size_t expectedResultsCount =
|
|
getAxes().empty() ? mesh->getRank() : getAxes().size();
|
|
if (getResult().size() != expectedResultsCount) {
|
|
return emitError() << "Unexpected number of results " << getResult().size()
|
|
<< ". Expected " << expectedResultsCount << ".";
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
void ProcessMultiIndexOp::build(OpBuilder &odsBuilder, OperationState &odsState,
|
|
MeshOp mesh) {
|
|
build(odsBuilder, odsState,
|
|
SmallVector<Type>(mesh.getRank(), odsBuilder.getIndexType()),
|
|
mesh.getSymName(), ArrayRef<MeshAxis>());
|
|
}
|
|
|
|
void ProcessMultiIndexOp::build(OpBuilder &odsBuilder, OperationState &odsState,
|
|
StringRef mesh, ArrayRef<MeshAxis> axes) {
|
|
build(odsBuilder, odsState,
|
|
SmallVector<Type>(axes.size(), odsBuilder.getIndexType()), mesh,
|
|
MeshAxesAttr::get(odsBuilder.getContext(), axes));
|
|
}
|
|
|
|
void ProcessMultiIndexOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResults()[0], "proc_linear_idx");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.process_linear_index op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
ProcessLinearIndexOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = ::getMeshAndVerify(getOperation(), getMeshAttr(), symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
return success();
|
|
}
|
|
|
|
void ProcessLinearIndexOp::build(OpBuilder &odsBuilder,
|
|
OperationState &odsState, MeshOp mesh) {
|
|
build(odsBuilder, odsState, mesh.getSymName());
|
|
}
|
|
|
|
void ProcessLinearIndexOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "proc_linear_idx");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.neighbors_linear_indices op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
NeighborsLinearIndicesOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = ::getMeshAndVerify(getOperation(), getMeshAttr(), symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
return success();
|
|
}
|
|
|
|
void NeighborsLinearIndicesOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getNeighborDown(), "down_linear_idx");
|
|
setNameFn(getNeighborUp(), "up_linear_idx");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// collective communication ops
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
namespace {
|
|
|
|
template <typename Op>
|
|
struct EmptyMeshAxesCanonicalizationPattern : OpRewritePattern<Op> {
|
|
using OpRewritePattern<Op>::OpRewritePattern;
|
|
LogicalResult matchAndRewrite(Op op,
|
|
PatternRewriter &rewriter) const override {
|
|
auto meshAxes = op.getMeshAxes();
|
|
if (!meshAxes.empty()) {
|
|
return failure();
|
|
}
|
|
if (op.getInput().getType() != op.getResult().getType()) {
|
|
return failure();
|
|
}
|
|
|
|
rewriter.replaceAllUsesWith(op.getResult(), op.getInput());
|
|
rewriter.eraseOp(op.getOperation());
|
|
return success();
|
|
}
|
|
};
|
|
|
|
} // namespace
|
|
|
|
static LogicalResult verifyInGroupDevice(Location loc, StringRef deviceName,
|
|
ArrayRef<int64_t> device,
|
|
Operation::operand_range deviceDynamic,
|
|
ArrayRef<MeshAxis> meshAxes,
|
|
ArrayRef<int64_t> meshShape) {
|
|
if (device.size() != meshAxes.size()) {
|
|
return emitError(loc) << "In-group device \"" << deviceName
|
|
<< "\" has unexpected multi-index size "
|
|
<< device.size() << ". Expected " << meshAxes.size()
|
|
<< ".";
|
|
}
|
|
|
|
for (size_t i = 0; i < device.size(); ++i) {
|
|
if (ShapedType::isStatic(device[i]) &&
|
|
ShapedType::isStatic(meshShape[meshAxes[i]]) &&
|
|
meshShape[meshAxes[i]] <= device[i]) {
|
|
return emitError(loc)
|
|
<< "Out of bounds coordinate " << i << " for in-group device \""
|
|
<< deviceName << "\"."
|
|
<< " Got " << device[i] << ", but expected value in the range [0, "
|
|
<< (meshShape[meshAxes[i]] - 1) << "].";
|
|
}
|
|
}
|
|
return success();
|
|
}
|
|
|
|
template <typename It>
|
|
static auto product(It begin, It end) {
|
|
using ElementType = std::decay_t<decltype(*begin)>;
|
|
return std::accumulate(begin, end, static_cast<ElementType>(1),
|
|
std::multiplies<ElementType>());
|
|
}
|
|
|
|
template <typename R>
|
|
static auto product(R &&range) {
|
|
return product(adl_begin(range), adl_end(range));
|
|
}
|
|
|
|
static LogicalResult verifyDimensionCompatibility(Location loc,
|
|
int64_t expectedDimSize,
|
|
int64_t resultDimSize,
|
|
int64_t resultAxis) {
|
|
if (ShapedType::isStatic(resultDimSize) && expectedDimSize != resultDimSize) {
|
|
return emitError(loc) << "Dimension size mismatch for result axis "
|
|
<< resultAxis << ". Expected "
|
|
<< (ShapedType::isDynamic(expectedDimSize)
|
|
? Twine("dynamic")
|
|
: Twine(expectedDimSize))
|
|
<< ", but got " << resultDimSize << ".";
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
static LogicalResult verifyGatherOperandAndResultShape(
|
|
Value operand, Value result, int64_t gatherAxis,
|
|
ArrayRef<MeshAxis> meshAxes, ArrayRef<int64_t> meshShape) {
|
|
auto resultRank = cast<ShapedType>(result.getType()).getRank();
|
|
if (gatherAxis < 0 || gatherAxis >= resultRank) {
|
|
return emitError(result.getLoc())
|
|
<< "Gather axis " << gatherAxis << " is out of bounds [0, "
|
|
<< resultRank << ").";
|
|
}
|
|
|
|
ShapedType operandType = cast<ShapedType>(operand.getType());
|
|
ShapedType resultType = cast<ShapedType>(result.getType());
|
|
auto deviceGroupSize =
|
|
DimensionSize(collectiveProcessGroupSize(meshAxes, meshShape));
|
|
for (int64_t axis = 0; axis < operandType.getRank(); ++axis) {
|
|
auto operandDimSize = DimensionSize(operandType.getDimSize(axis));
|
|
auto resultDimSize = DimensionSize(resultType.getDimSize(axis));
|
|
auto expectedResultDimSize =
|
|
axis == gatherAxis ? deviceGroupSize * operandDimSize : operandDimSize;
|
|
if (failed(verifyDimensionCompatibility(
|
|
result.getLoc(), expectedResultDimSize, resultDimSize, axis))) {
|
|
return failure();
|
|
}
|
|
}
|
|
return success();
|
|
}
|
|
|
|
static LogicalResult verifyAllToAllOperandAndResultShape(
|
|
Value operand, Value result, int64_t splitAxis, int64_t concatAxis,
|
|
ArrayRef<MeshAxis> meshAxes, ArrayRef<int64_t> meshShape) {
|
|
ShapedType operandType = cast<ShapedType>(operand.getType());
|
|
ShapedType resultType = cast<ShapedType>(result.getType());
|
|
for (int64_t axis = 0; axis < operandType.getRank(); ++axis) {
|
|
if ((axis != splitAxis && axis != concatAxis) || splitAxis == concatAxis) {
|
|
if (failed(verifyDimensionCompatibility(
|
|
result.getLoc(), operandType.getDimSize(axis),
|
|
resultType.getDimSize(axis), axis))) {
|
|
return failure();
|
|
}
|
|
}
|
|
}
|
|
|
|
if (splitAxis == concatAxis) {
|
|
return success();
|
|
}
|
|
|
|
auto deviceGroupSize =
|
|
DimensionSize(collectiveProcessGroupSize(meshAxes, meshShape));
|
|
auto operandConcatDimSize = DimensionSize(operandType.getDimSize(concatAxis));
|
|
auto operandSplitDimSize = DimensionSize(operandType.getDimSize(splitAxis));
|
|
DimensionSize expectedResultConcatDimSize =
|
|
operandConcatDimSize * deviceGroupSize;
|
|
DimensionSize expectedResultSplitDimSize =
|
|
operandSplitDimSize / deviceGroupSize;
|
|
if (!expectedResultSplitDimSize.isDynamic() &&
|
|
int64_t(operandSplitDimSize) % int64_t(deviceGroupSize) != 0) {
|
|
expectedResultSplitDimSize = DimensionSize::dynamic();
|
|
}
|
|
if (failed(verifyDimensionCompatibility(
|
|
result.getLoc(), expectedResultConcatDimSize.value(),
|
|
resultType.getDimSize(concatAxis), concatAxis))) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyDimensionCompatibility(
|
|
result.getLoc(), expectedResultSplitDimSize.value(),
|
|
resultType.getDimSize(splitAxis), splitAxis))) {
|
|
return failure();
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
static LogicalResult verifyScatterOrSliceOperandAndResultShape(
|
|
Value operand, Value result, int64_t tensorAxis,
|
|
ArrayRef<MeshAxis> meshAxes, ArrayRef<int64_t> meshShape) {
|
|
ShapedType operandType = cast<ShapedType>(operand.getType());
|
|
ShapedType resultType = cast<ShapedType>(result.getType());
|
|
for (int64_t axis = 0; axis < operandType.getRank(); ++axis) {
|
|
if (axis != tensorAxis) {
|
|
if (failed(verifyDimensionCompatibility(
|
|
result.getLoc(), operandType.getDimSize(axis),
|
|
resultType.getDimSize(axis), axis))) {
|
|
return failure();
|
|
}
|
|
}
|
|
}
|
|
|
|
auto deviceGroupSize =
|
|
DimensionSize(collectiveProcessGroupSize(meshAxes, meshShape));
|
|
auto operandScatterDimSize =
|
|
DimensionSize(operandType.getDimSize(tensorAxis));
|
|
if (!operandScatterDimSize.isDynamic() && !deviceGroupSize.isDynamic() &&
|
|
int64_t(operandScatterDimSize) % int64_t(deviceGroupSize) != 0) {
|
|
return emitError(result.getLoc())
|
|
<< "Operand dimension size " << int64_t(operandScatterDimSize)
|
|
<< " is not divisible by collective device group size "
|
|
<< int64_t(deviceGroupSize) << " for tensor axis " << tensorAxis
|
|
<< ".";
|
|
}
|
|
DimensionSize expectedResultTensorDimSize =
|
|
operandScatterDimSize / deviceGroupSize;
|
|
if (failed(verifyDimensionCompatibility(
|
|
result.getLoc(), expectedResultTensorDimSize.value(),
|
|
resultType.getDimSize(tensorAxis), tensorAxis))) {
|
|
return failure();
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
static RankedTensorType sliceResultType(Type operandType, MeshOp mesh,
|
|
ArrayRef<MeshAxis> meshAxes,
|
|
int64_t sliceAxis) {
|
|
RankedTensorType operandRankedTensorType =
|
|
cast<RankedTensorType>(operandType);
|
|
DimensionSize operandSliceAxisSize =
|
|
operandRankedTensorType.getShape()[sliceAxis];
|
|
SmallVector<int64_t> resultShape =
|
|
llvm::to_vector(operandRankedTensorType.getShape());
|
|
|
|
resultShape[sliceAxis] =
|
|
operandSliceAxisSize /
|
|
DimensionSize(collectiveProcessGroupSize(meshAxes, mesh));
|
|
return operandRankedTensorType.clone(resultShape);
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.all_gather op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
AllGatherOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
auto gatherAxis = getGatherAxis().getSExtValue();
|
|
return verifyGatherOperandAndResultShape(getOperand(), getResult(),
|
|
gatherAxis, getMeshAxes(),
|
|
mesh.value().getShape());
|
|
}
|
|
|
|
void AllGatherOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<AllGatherOp>>(context);
|
|
}
|
|
|
|
void AllGatherOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "all_gather");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.all_reduce op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
AllReduceOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
return getMeshAndVerifyAxes(*this, symbolTable);
|
|
}
|
|
|
|
void AllReduceOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<AllReduceOp>>(context);
|
|
}
|
|
|
|
void AllReduceOp::build(OpBuilder &odsBuilder, OperationState &odsState,
|
|
Value input, StringRef mesh,
|
|
ArrayRef<MeshAxis> meshAxes, ReductionKind reduction) {
|
|
build(odsBuilder, odsState, input.getType(), mesh, meshAxes, input,
|
|
reduction);
|
|
}
|
|
|
|
void AllReduceOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "all_reduce");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.all_slice op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult AllSliceOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
return verifyScatterOrSliceOperandAndResultShape(
|
|
getOperand(), getResult(), getSliceAxis().getSExtValue(), getMeshAxes(),
|
|
mesh.value().getShape());
|
|
}
|
|
|
|
void AllSliceOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<AllSliceOp>>(context);
|
|
}
|
|
|
|
void AllSliceOp::build(OpBuilder &odsBuilder, OperationState &odsState,
|
|
Value input, MeshOp mesh, ArrayRef<MeshAxis> meshAxes,
|
|
int64_t sliceAxis) {
|
|
Type resultType = sliceResultType(input.getType(), mesh, meshAxes, sliceAxis);
|
|
build(odsBuilder, odsState, resultType, input, mesh.getSymName(), meshAxes,
|
|
sliceAxis);
|
|
}
|
|
|
|
void AllSliceOp::build(OpBuilder &odsBuilder, OperationState &odsState,
|
|
Type resultType, Value input, StringRef mesh,
|
|
ArrayRef<MeshAxis> meshAxes, int64_t sliceAxis) {
|
|
build(odsBuilder, odsState, resultType, mesh, meshAxes, input,
|
|
APInt(sizeof(sliceAxis) * CHAR_BIT, sliceAxis));
|
|
}
|
|
|
|
void AllSliceOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "all_slice");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.all_to_all op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult AllToAllOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
|
|
return verifyAllToAllOperandAndResultShape(
|
|
getOperand(), getResult(), getSplitAxis().getSExtValue(),
|
|
getConcatAxis().getSExtValue(), getMeshAxes(), mesh.value().getShape());
|
|
}
|
|
|
|
void AllToAllOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<AllToAllOp>>(context);
|
|
}
|
|
|
|
void AllToAllOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "all_to_all");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.broadcast op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
BroadcastOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyInGroupDevice(getLoc(), getRootAttrName(), getRoot(),
|
|
getRootDynamic(), getMeshAxes(),
|
|
mesh.value().getShape()))) {
|
|
return failure();
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
void BroadcastOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<BroadcastOp>>(context);
|
|
}
|
|
|
|
void BroadcastOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "broadcast");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.gather op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult GatherOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyInGroupDevice(getLoc(), getRootAttrName(), getRoot(),
|
|
getRootDynamic(), getMeshAxes(),
|
|
mesh.value().getShape()))) {
|
|
return failure();
|
|
}
|
|
|
|
auto gatherAxis = getGatherAxis().getSExtValue();
|
|
return verifyGatherOperandAndResultShape(getInput(), getResult(), gatherAxis,
|
|
getMeshAxes(),
|
|
mesh.value().getShape());
|
|
}
|
|
|
|
void GatherOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<GatherOp>>(context);
|
|
}
|
|
|
|
void GatherOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "gather");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.recv op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult RecvOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (getSource() &&
|
|
failed(verifyInGroupDevice(getLoc(), getSourceAttrName(),
|
|
getSource().value(), getSourceDynamic(),
|
|
getMeshAxes(), mesh.value().getShape()))) {
|
|
return failure();
|
|
}
|
|
return success();
|
|
}
|
|
|
|
void RecvOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<RecvOp>>(context);
|
|
}
|
|
|
|
void RecvOp::getAsmResultNames(function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "recv");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.reduce op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult ReduceOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyInGroupDevice(getLoc(), getRootAttrName(), getRoot(),
|
|
getRootDynamic(), getMeshAxes(),
|
|
mesh.value().getShape()))) {
|
|
return failure();
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
void ReduceOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<ReduceOp>>(context);
|
|
}
|
|
|
|
void ReduceOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "reduce");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.reduce_scatter op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
ReduceScatterOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
|
|
return verifyScatterOrSliceOperandAndResultShape(
|
|
getOperand(), getResult(), getScatterAxis().getSExtValue(), getMeshAxes(),
|
|
mesh.value().getShape());
|
|
}
|
|
|
|
void ReduceScatterOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<ReduceScatterOp>>(context);
|
|
}
|
|
|
|
void ReduceScatterOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "reduce_scatter");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.scatter op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult ScatterOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyInGroupDevice(getLoc(), getRootAttrName(), getRoot(),
|
|
getRootDynamic(), getMeshAxes(),
|
|
mesh.value().getShape()))) {
|
|
return failure();
|
|
}
|
|
|
|
auto scatterAxis = getScatterAxis().getSExtValue();
|
|
return verifyScatterOrSliceOperandAndResultShape(getInput(), getResult(),
|
|
scatterAxis, getMeshAxes(),
|
|
mesh.value().getShape());
|
|
}
|
|
|
|
void ScatterOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<ScatterOp>>(context);
|
|
}
|
|
|
|
void ScatterOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "scatter");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.send op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult SendOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
if (failed(verifyInGroupDevice(getLoc(), getDestinationAttrName(),
|
|
getDestination(), getDestinationDynamic(),
|
|
getMeshAxes(), mesh.value().getShape()))) {
|
|
return failure();
|
|
}
|
|
return success();
|
|
}
|
|
|
|
void SendOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
patterns.add<EmptyMeshAxesCanonicalizationPattern<SendOp>>(context);
|
|
}
|
|
|
|
void SendOp::getAsmResultNames(function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "send");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.shift op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult ShiftOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerifyAxes(*this, symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
|
|
auto meshAxes = getMeshAxes();
|
|
auto shiftAxis = getShiftAxis().getZExtValue();
|
|
if (!llvm::is_contained(meshAxes, shiftAxis)) {
|
|
return emitError() << "Invalid shift axis " << shiftAxis
|
|
<< ". It must be one of the grouping mesh axes.";
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
void ShiftOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
|
|
MLIRContext *context) {
|
|
// TODO: remove op when offset is 0 or if it is a rotate with and
|
|
// offset % shift_axis_mesh_dim_size == 0.
|
|
}
|
|
|
|
void ShiftOp::getAsmResultNames(
|
|
function_ref<void(Value, StringRef)> setNameFn) {
|
|
setNameFn(getResult(), "shift");
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// mesh.update_halo op
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult
|
|
UpdateHaloOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
|
|
auto mesh = getMeshAndVerify(getOperation(), getMeshAttr(), symbolTable);
|
|
if (failed(mesh)) {
|
|
return failure();
|
|
}
|
|
|
|
return success();
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// TableGen'd op method definitions
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#define GET_OP_CLASSES
|
|
#include "mlir/Dialect/Mesh/IR/MeshOps.cpp.inc"
|
|
|
|
#define GET_ATTRDEF_CLASSES
|
|
#include "mlir/Dialect/Mesh/IR/MeshAttributes.cpp.inc"
|
|
|
|
#define GET_TYPEDEF_CLASSES
|
|
#include "mlir/Dialect/Mesh/IR/MeshTypes.cpp.inc"
|
|
|
|
#include "mlir/Dialect/Mesh/IR/MeshEnums.cpp.inc"
|