Skip to content

[mlir] [vector] Add linearization pattern for vector.create_mask #138214

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.

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 12 commits into from
May 14, 2025
Merged
Show file tree
Hide file tree
Changes from 5 commits
Commits
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
65 changes: 62 additions & 3 deletions mlir/lib/Dialect/Vector/Transforms/VectorLinearize.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -445,6 +445,64 @@ struct LinearizeVectorSplat final
}
};

/// This pattern converts the CreateMaskOp to work on a
/// linearized vector. The pattern currently
/// supports only 2D masks with a unit outer dimension.
/// Following,
/// vector.create_mask %dims : vector<1x4xi1>
/// is converted to:
/// %out_1d = vector.create_mask %dims : vector<4xi1>
/// %out_nd = vector.shape_cast %out_1d : vector<4xi1> to vector<1x4xi1>
struct LinearizeVectorCreateMask final
: OpConversionPattern<vector::CreateMaskOp> {
using OpConversionPattern::OpConversionPattern;

LinearizeVectorCreateMask(const TypeConverter &typeConverter,
MLIRContext *context, PatternBenefit benefit = 1)
: OpConversionPattern(typeConverter, context, benefit) {}

LogicalResult
matchAndRewrite(vector::CreateMaskOp createMaskOp, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
auto srcTy = createMaskOp.getType();
auto srcShape = srcTy.getShape();
if (srcShape.size() != 2)
Copy link
Contributor

Choose a reason for hiding this comment

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

Consider adding
// FIXME: add support for any vector with at most 1 non-unit dimension (like vector<1x4x1xi1>)
here

return rewriter.notifyMatchFailure(createMaskOp,
"only 2D mask is supported.");

if (srcShape[0] != 1)
return rewriter.notifyMatchFailure(
createMaskOp, "only unit outer dimension is supported.");

auto dstTy = getTypeConverter()->convertType(srcTy);
if (!dstTy)
return rewriter.notifyMatchFailure(createMaskOp, "cannot convert type.");

// Compare the first operand with 0. If it's less than or equal to 0,
// create a zero mask, else strip the first operand and create a mask
// using the second operand.
auto firstOperand = adaptor.getOperands().front();
auto zero =
rewriter.create<mlir::arith::ConstantIndexOp>(createMaskOp.getLoc(), 0);
auto isZeroOrNegative = rewriter.create<mlir::arith::CmpIOp>(
createMaskOp.getLoc(), mlir::arith::CmpIPredicate::sle, firstOperand,
zero);
auto isZeroOrNegativeSplat = rewriter.create<mlir::vector::SplatOp>(
createMaskOp.getLoc(), dstTy, isZeroOrNegative);

// Use a select operation to choose between the masks.
auto zeroMask = rewriter.create<mlir::arith::ConstantOp>(
createMaskOp.getLoc(), dstTy, rewriter.getZeroAttr(dstTy));
auto newMask = rewriter.create<mlir::vector::CreateMaskOp>(
createMaskOp.getLoc(), dstTy, adaptor.getOperands().back());
auto result = rewriter.create<mlir::arith::SelectOp>(
Copy link
Contributor

Choose a reason for hiding this comment

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

We may want to use some createAndFold here to get rid of redundant IR...

Copy link
Contributor

@newling newling May 12, 2025

Choose a reason for hiding this comment

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

Thanks for updating the test @nbpatel to use non-constant operands.

nit: my guess is that most of the time, the unit dimension will have extent which is the constant 1, and so createOrFold will still be advisable.

createMaskOp.getLoc(), isZeroOrNegativeSplat, zeroMask, newMask);

rewriter.replaceOp(createMaskOp, result.getResult());
return success();
}
};

} // namespace

/// Return true if the operation `op` does not support scalable vectors and
Expand Down Expand Up @@ -530,9 +588,10 @@ void mlir::vector::populateForVectorLinearize(TypeConverter &typeConverter,
void mlir::vector::populateVectorLinearizeBasePatterns(
const TypeConverter &typeConverter, const ConversionTarget &target,
RewritePatternSet &patterns) {
patterns.add<LinearizeConstantLike, LinearizeVectorizable,
LinearizeVectorBitCast, LinearizeVectorSplat>(
typeConverter, patterns.getContext());
patterns
.add<LinearizeConstantLike, LinearizeVectorizable, LinearizeVectorBitCast,
LinearizeVectorSplat, LinearizeVectorCreateMask>(
typeConverter, patterns.getContext());
}

void mlir::vector::populateVectorLinearizeShuffleLikeOpsPatterns(
Expand Down
38 changes: 38 additions & 0 deletions mlir/test/Dialect/Vector/linearize.mlir
Original file line number Diff line number Diff line change
Expand Up @@ -345,3 +345,41 @@ func.func @linearize_scalable_vector_splat(%arg0: i32) -> vector<4x[2]xi32> {
%0 = vector.splat %arg0 : vector<4x[2]xi32>
return %0 : vector<4x[2]xi32>
}

// -----
// ALL-LABEL: linearize_create_mask
func.func @linearize_create_mask() -> vector<1x16xi1> {
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[C20:.*]] = arith.constant 20 : index
// CHECK: %[[C0_0:.*]] = arith.constant 0 : index
// CHECK: %[[CMP:.*]] = arith.cmpi sle, %[[C0]], %[[C0_0]] : index
// CHECK: %[[SPLAT:.*]] = vector.splat %[[CMP]] : vector<16xi1>
// CHECK: %[[CST:.*]] = arith.constant dense<false> : vector<16xi1>
// CHECK: %[[MASK_1D:.*]] = vector.create_mask %[[C20]] : vector<16xi1>
// CHECK: %[[SELECT:.*]] = arith.select %[[SPLAT]], %[[CST]], %[[MASK_1D]] : vector<16xi1>, vector<16xi1>
// CHECK: %[[CAST:.*]] = vector.shape_cast %[[SELECT]] : vector<16xi1> to vector<1x16xi1>
// CHECK: return %[[CAST]] : vector<1x16xi1>
%c0 = arith.constant 0 : index
%c20 = arith.constant 20 : index
%0 = vector.create_mask %c0, %c20 : vector<1x16xi1>
return %0 : vector<1x16xi1>
}

// -----
// ALL-LABEL: linearize_scalable_create_mask
func.func @linearize_scalable_create_mask() -> vector<1x[16]xi1> {
// CHECK: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[C20:.*]] = arith.constant 20 : index
// CHECK: %[[C0_0:.*]] = arith.constant 0 : index
// CHECK: %[[CMP:.*]] = arith.cmpi sle, %[[C0]], %[[C0_0]] : index
// CHECK: %[[SPLAT:.*]] = vector.splat %[[CMP]] : vector<[16]xi1>
// CHECK: %[[CST:.*]] = arith.constant dense<false> : vector<[16]xi1>
Copy link
Contributor

Choose a reason for hiding this comment

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

most of these checks should be folded with createAndFold

Copy link
Contributor

Choose a reason for hiding this comment

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

Alternatively, use an %arg0: index instead of %c0

// CHECK: %[[MASK_1D:.*]] = vector.create_mask %[[C20]] : vector<[16]xi1>
// CHECK: %[[SELECT:.*]] = arith.select %[[SPLAT]], %[[CST]], %[[MASK_1D]] : vector<[16]xi1>, vector<[16]xi1>
// CHECK: %[[CAST:.*]] = vector.shape_cast %[[SELECT]] : vector<[16]xi1> to vector<1x[16]xi1>
// CHECK: return %[[CAST]] : vector<1x[16]xi1>
%c0 = arith.constant 0 : index
%c20 = arith.constant 20 : index
%0 = vector.create_mask %c0, %c20 : vector<1x[16]xi1>
return %0 : vector<1x[16]xi1>
}
2 changes: 1 addition & 1 deletion mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -973,7 +973,7 @@ struct TestVectorLinearize final
return "Linearizes ND vectors for N >= 2 into 1D vectors";
}
void getDependentDialects(DialectRegistry &registry) const override {
registry.insert<vector::VectorDialect>();
registry.insert<vector::VectorDialect, arith::ArithDialect>();
}

void runOnOperation() override {
Expand Down