Skip to content

[MLIR][XeGPU] Extend unrolling support for scatter ops with chunk_size #144447

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 20 commits into from
Jun 17, 2025
Merged

[MLIR][XeGPU] Extend unrolling support for scatter ops with chunk_size #144447

Show file tree
Hide file tree
Changes from 18 commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
bac8bc6
add unroll supportfor reduce and broadcast
Jianhui-Li Jun 9, 2025
30b099e
add create_desc unrolling and test
Jianhui-Li Jun 9, 2025
8a0e145
add unrolling support scatter operations
Jianhui-Li Jun 10, 2025
c91156c
clang format
Jianhui-Li Jun 10, 2025
e532cbe
Merge branch 'main' into unroll-scatter-ops
Jianhui-Li Jun 10, 2025
30cb8d8
Update mlir/test/Dialect/XeGPU/xegpu-unroll-patterns.mlir
Jianhui-Li Jun 13, 2025
f493e52
Update xegpu-unroll-patterns.mlir
Jianhui-Li Jun 13, 2025
2606a4b
reject support for chunk_size
Jianhui-Li Jun 13, 2025
a3e064d
small fixes
Jianhui-Li Jun 13, 2025
b383194
Merge branch 'unroll-scatter-ops' of https://github.com/Jianhui-Li/ll…
Jianhui-Li Jun 13, 2025
b3d5937
smal fix
Jianhui-Li Jun 14, 2025
248981b
add support for create_tdesc with chunk_size
Jianhui-Li Jun 15, 2025
66ab4aa
add unrolling support for load/store/prefetch/update with chunk_size
Jianhui-Li Jun 15, 2025
7cbeebb
small bug fixes
Jianhui-Li Jun 16, 2025
47fe143
add tests
Jianhui-Li Jun 16, 2025
a569bc8
clang format fix
Jianhui-Li Jun 16, 2025
2ad1947
Merge branch 'main' into unroll-scatter-ops-chunksize
Jianhui-Li Jun 16, 2025
41b839d
remove 1 complex test
Jianhui-Li Jun 16, 2025
3c3023c
address review feedback
Jianhui-Li Jun 17, 2025
19d8d13
clang-format
Jianhui-Li Jun 17, 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
178 changes: 136 additions & 42 deletions mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -402,30 +402,58 @@ struct UnrollCreateDescOp : public UnrollPattern<xegpu::CreateDescOp> {
PatternRewriter &rewriter) const override {
Location loc = op.getLoc();
xegpu::TensorDescType tdescTy = op.getType();
TypedValue<::mlir::VectorType> indiceVec = op.getOffsets();
VectorType indiceVecTy = indiceVec.getType();

// check if the tensor descriptor type is a 1d vector type
if (tdescTy.getRank() > 1)
if (!tdescTy.isScattered())
return failure();

std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
if (!targetShape)
return failure();

auto newTdescTy = getUnrolledTypes(tdescTy, *targetShape)[0];

TypedValue<::mlir::VectorType> indiceVec = op.getOffsets();
VectorType indiceVecTy = indiceVec.getType();
SmallVector<int64_t> targetIndiceShape(*targetShape);
int64_t originalChunkSize = tdescTy.getChunkSize();
// IndiceVec is 1 dim lower than tdescTy when chunkSize is larger than 1.
if (originalChunkSize > 1)
targetIndiceShape.pop_back();

auto newTdescTy = getUnrolledTypes(tdescTy, *targetShape)[0];
SmallVector<Type> convertedIndiceTypes =
getUnrolledTypes(indiceVecTy, *targetShape);
getUnrolledTypes(indiceVecTy, targetIndiceShape);
SmallVector<Value> convertedIndiceVec =
pack(indiceVec, convertedIndiceTypes, *targetShape, loc, rewriter);
pack(indiceVec, convertedIndiceTypes, targetIndiceShape, loc, rewriter);

SmallVector<Value> newOps;
for (auto indice : convertedIndiceVec) {
auto newOp = rewriter.create<xegpu::CreateDescOp>(loc, newTdescTy,
op.getSource(), indice);
newOps.push_back(newOp);

// more indices is need when chunkSize > 1. Since a big load from one
// address could be break into multiple small loads.
if (originalChunkSize > 1) {
int64_t blockedChunkSize = targetShape->back();
int64_t numNewChunks = originalChunkSize / blockedChunkSize;

for (auto [indice, indiceType] :
llvm::zip(convertedIndiceVec, convertedIndiceTypes)) {
for (int64_t i = 0; i < numNewChunks; ++i) {
// Compute the offset
Value inc = rewriter.create<arith::ConstantIndexOp>(
loc, i * blockedChunkSize);
Value incVec = rewriter.create<vector::SplatOp>(loc, indiceType, inc);
Value offsetIndice =
rewriter.create<arith::AddIOp>(loc, indice, incVec);

auto newOp = rewriter.create<xegpu::CreateDescOp>(
loc, newTdescTy, op.getSource(), offsetIndice);

newOps.push_back(newOp);
}
}
} else {
for (auto indice : convertedIndiceVec) {
auto newOp = rewriter.create<xegpu::CreateDescOp>(
loc, newTdescTy, op.getSource(), indice);
newOps.push_back(newOp);
}
}

Value castOp = unpack(newOps, tdescTy, *targetShape, loc, rewriter);
Expand All @@ -444,16 +472,18 @@ struct UnrollLoadGatherOp : public UnrollPattern<xegpu::LoadGatherOp> {
VectorType valueTy = llvm::dyn_cast<VectorType>(op.getValue().getType());
xegpu::TensorDescType tdescTy = op.getTensorDescType();

// check if the tensor descriptor type is a 1d vector type
if (tdescTy.getRank() > 1)
if (!tdescTy.isScattered())
return failure();

VectorType maskTy = llvm::dyn_cast<VectorType>(op.getMask().getType());

std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
if (!targetShape)
return failure();

SmallVector<int64_t> targetMaskShape(*targetShape);
int64_t originalChunkSize = tdescTy.getChunkSize();

VectorType maskTy = llvm::dyn_cast<VectorType>(op.getMask().getType());

Type elemTy = tdescTy.getElementType();
VectorType newValueTy = valueTy.cloneWith(*targetShape, elemTy);

Expand All @@ -462,10 +492,32 @@ struct UnrollLoadGatherOp : public UnrollPattern<xegpu::LoadGatherOp> {
SmallVector<Value> convertedTdescs = pack(
op.getTensorDesc(), convertedTdescTypes, *targetShape, loc, rewriter);

SmallVector<Type> convertedMaskTypes =
getUnrolledTypes(maskTy, *targetShape);
SmallVector<Value> convertedMasks =
pack(op.getMask(), convertedMaskTypes, *targetShape, loc, rewriter);
SmallVector<Type> convertedMaskTypes;
SmallVector<Value> convertedMasks;

if (originalChunkSize > 1) {
targetMaskShape.pop_back();
convertedMaskTypes = getUnrolledTypes(maskTy, targetMaskShape);
SmallVector<Value> convertedMasks1D = pack(
op.getMask(), convertedMaskTypes, targetMaskShape, loc, rewriter);
int64_t blockedChunkSize = targetShape->back();
int64_t numNewChunks = originalChunkSize / blockedChunkSize;

for (auto mask : convertedMasks1D) {
for (int64_t i = 0; i < numNewChunks; ++i) {
convertedMasks.push_back(mask);
}
}
// This is to handle the transpose effect when chunkSize > 1.
if (targetShape && targetShape->size() > 1) {
std::swap((*targetShape)[0], (*targetShape)[1]);
newValueTy = valueTy.cloneWith(*targetShape, elemTy);
}
} else {
convertedMaskTypes = getUnrolledTypes(maskTy, targetMaskShape);
convertedMasks = pack(op.getMask(), convertedMaskTypes, targetMaskShape,
loc, rewriter);
}

SmallVector<Value> newOps;
for (auto [t, m] : llvm::zip(convertedTdescs, convertedMasks)) {
Expand All @@ -476,7 +528,6 @@ struct UnrollLoadGatherOp : public UnrollPattern<xegpu::LoadGatherOp> {
}

Value castOp = unpack(newOps, op.getType(), *targetShape, loc, rewriter);

rewriter.replaceOp(op, castOp);
return success();
}
Expand All @@ -490,7 +541,7 @@ struct UnrollPrefetchOp : public UnrollPattern<xegpu::PrefetchOp> {
xegpu::TensorDescType tdescTy = op.getTensorDescType();

// check if the tensor descriptor type is a 1d vector type
if (tdescTy.getRank() > 1)
if (tdescTy.getRank() > 2)
return failure();

std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
Expand Down Expand Up @@ -519,30 +570,51 @@ struct UnrollStoreScatterOp : public UnrollPattern<xegpu::StoreScatterOp> {
VectorType valueTy = llvm::dyn_cast<VectorType>(op.getValue().getType());
xegpu::TensorDescType tdescTy = op.getTensorDescType();

// check if the tensor descriptor type is a 1d vector type
if (tdescTy.getRank() > 1)
if (!tdescTy.isScattered())
return failure();

VectorType maskTy = llvm::dyn_cast<VectorType>(op.getMask().getType());

std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
if (!targetShape)
return failure();

SmallVector<Type> convertedValTypes =
getUnrolledTypes(valueTy, *targetShape);
SmallVector<int64_t> targetIndiceShape(*targetShape);
int64_t originalChunkSize = tdescTy.getChunkSize();

VectorType maskTy = llvm::dyn_cast<VectorType>(op.getMask().getType());

SmallVector<Type> convertedTdescTypes =
getUnrolledTypes(tdescTy, *targetShape);

SmallVector<Value> convertedValues =
pack(op.getValue(), convertedValTypes, *targetShape, loc, rewriter);
SmallVector<Value> convertedTdescs = pack(
op.getTensorDesc(), convertedTdescTypes, *targetShape, loc, rewriter);

SmallVector<Type> convertedMaskTypes =
getUnrolledTypes(maskTy, *targetShape);
SmallVector<Value> convertedMasks =
pack(op.getMask(), convertedMaskTypes, *targetShape, loc, rewriter);
SmallVector<Type> convertedMaskTypes;
SmallVector<Value> convertedMasks;

if (originalChunkSize > 1) {
int64_t blockedChunkSize = targetShape->back();
int64_t numNewChunks = originalChunkSize / blockedChunkSize;
convertedMaskTypes = getUnrolledTypes(maskTy, (*targetShape)[0]);
SmallVector<Value> convertedMasks1D = pack(
op.getMask(), convertedMaskTypes, (*targetShape)[0], loc, rewriter);

for (auto mask : convertedMasks1D) {
for (int64_t i = 0; i < numNewChunks; ++i) {
convertedMasks.push_back(mask);
}
}
// This is to handle the transpose effect when chunkSize > 1.
std::swap((*targetShape)[0], (*targetShape)[1]);

} else {
convertedMaskTypes = getUnrolledTypes(maskTy, *targetShape);
convertedMasks =
pack(op.getMask(), convertedMaskTypes, *targetShape, loc, rewriter);
}

SmallVector<Type> convertedValTypes =
getUnrolledTypes(valueTy, *targetShape);
SmallVector<Value> convertedValues =
pack(op.getValue(), convertedValTypes, *targetShape, loc, rewriter);

for (size_t i = 0; i < convertedValues.size(); ++i) {
Value v = convertedValues[i];
Expand All @@ -565,8 +637,10 @@ struct UnrollUpdateOffsetOp : public UnrollPattern<xegpu::UpdateOffsetOp> {
Location loc = op.getLoc();
xegpu::TensorDescType tdescTy = op.getTensorDescType();

// check if the tensor descriptor type is a 1d vector type
if (tdescTy.getRank() > 1)
if (tdescTy.getRank() > 2)
return failure();

if (!tdescTy.isScattered())
return failure();

std::optional<SmallVector<int64_t>> targetShape = getTargetShape(op);
Expand All @@ -580,12 +654,32 @@ struct UnrollUpdateOffsetOp : public UnrollPattern<xegpu::UpdateOffsetOp> {

TypedValue<::mlir::VectorType> offsetVec = op.getOffsets();
VectorType offsetVecTy = offsetVec.getType();
SmallVector<Type> convertedOffsetTypes =
getUnrolledTypes(offsetVecTy, *targetShape);
SmallVector<Value> convertedOffsetVec =
pack(offsetVec, convertedOffsetTypes, *targetShape, loc, rewriter);

SmallVector<Type> convertedOffsetTypes;
SmallVector<Value> convertedOffsetVec;
SmallVector<Value> newOps;
int64_t originalChunkSize = tdescTy.getChunkSize();
if (originalChunkSize > 1) {
SmallVector<int64_t> shape1D(targetShape->begin(),
targetShape->end() - 1);
convertedOffsetTypes = getUnrolledTypes(offsetVecTy, shape1D);
SmallVector<Value> convertedOffsetVec1D =
pack(offsetVec, convertedOffsetTypes, shape1D, loc, rewriter);

int64_t blockedChunkSize = targetShape->back();
int64_t numNewChunks = originalChunkSize / blockedChunkSize;

for (auto offset : convertedOffsetVec1D) {
for (int64_t i = 0; i < numNewChunks; ++i) {
convertedOffsetVec.push_back(offset);
}
}

} else {
convertedOffsetTypes = getUnrolledTypes(offsetVecTy, *targetShape);
convertedOffsetVec =
pack(offsetVec, convertedOffsetTypes, *targetShape, loc, rewriter);
}

for (auto [t, o] : llvm::zip(convertedTdesc, convertedOffsetVec)) {
auto newOp =
rewriter.create<xegpu::UpdateOffsetOp>(loc, t.getType(), t, o);
Expand Down
Loading
Loading