Skip to content

[mlir][vector] add support for linearizing vector.bitcast in VectorLinearize #123110

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 8 commits into from
Jan 27, 2025
Merged

[mlir][vector] add support for linearizing vector.bitcast in VectorLinearize #123110

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
d067a5a
add linearize pattern for bitcast
chencha3 Jan 15, 2025
5f35883
code format
chencha3 Jan 15, 2025
f0b3ae0
update tests
chencha3 Jan 16, 2025
25a8f39
add test for scalable vector
chencha3 Jan 16, 2025
ea2e518
fix naming in tests
chencha3 Jan 27, 2025
7a3109d
Merge branch 'main' into linearize_bitcast
chencha3 Jan 27, 2025
054b700
fix naming in tests
chencha3 Jan 27, 2025
d112120
Merge branch 'main' into linearize_bitcast
chencha3 Jan 27, 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
51 changes: 46 additions & 5 deletions mlir/lib/Dialect/Vector/Transforms/VectorLinearize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,13 +72,14 @@ struct LinearizeConstant final : OpConversionPattern<arith::ConstantOp> {
auto resType =
getTypeConverter()->convertType<VectorType>(constOp.getType());

if (!resType)
return rewriter.notifyMatchFailure(loc, "can't convert return type");

if (resType.isScalable() && !isa<SplatElementsAttr>(constOp.getValue()))
return rewriter.notifyMatchFailure(
loc,
"Cannot linearize a constant scalable vector that's not a splat");

if (!resType)
return rewriter.notifyMatchFailure(loc, "can't convert return type");
if (!isLessThanTargetBitWidth(constOp, targetVectorBitWidth))
return rewriter.notifyMatchFailure(
loc, "Can't flatten since targetBitWidth <= OpSize");
Expand Down Expand Up @@ -459,6 +460,45 @@ struct LinearizeVectorInsert final
private:
unsigned targetVectorBitWidth;
};

/// This pattern converts the BitCastOp that works on nD (n > 1)
/// vectors to a BitCastOp that works on linearized vectors.
/// Following,
/// vector.bitcast %v1: vector<4x2xf32> to vector<4x4xf16>
/// is converted to :
/// %v1_1d = vector.shape_cast %v1: vector<4x2xf32> to vector<8xf32>
/// %out_1d = vector.bitcast %v1_1d: vector<8xf32> to vector<16xf16>
/// %out_nd = vector.shape_cast %out_1d: vector<16xf16> to vector<4x4xf16>
struct LinearizeVectorBitCast final
Copy link
Contributor

Choose a reason for hiding this comment

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

Please add documentation and tests for scalable vectors (bailing out for scalable vectors would be fine, though this should just work).

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Thanks. I added comments and tests for scalable vectors.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

@banach-space Do you have any more concerns?

Copy link
Contributor

Choose a reason for hiding this comment

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

Apologies for the delay—this slipped off my radar.

Everything looks good, but I have a small request regarding the tests.

: public OpConversionPattern<vector::BitCastOp> {
using OpConversionPattern::OpConversionPattern;
LinearizeVectorBitCast(
const TypeConverter &typeConverter, MLIRContext *context,
unsigned targetVectBitWidth = std::numeric_limits<unsigned>::max(),
PatternBenefit benefit = 1)
: OpConversionPattern(typeConverter, context, benefit),
targetVectorBitWidth(targetVectBitWidth) {}
LogicalResult
matchAndRewrite(vector::BitCastOp castOp, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = castOp.getLoc();
auto resType = getTypeConverter()->convertType(castOp.getType());
if (!resType)
return rewriter.notifyMatchFailure(loc, "can't convert return type.");

if (!isLessThanTargetBitWidth(castOp, targetVectorBitWidth))
return rewriter.notifyMatchFailure(
loc, "Can't flatten since targetBitWidth <= OpSize");

rewriter.replaceOpWithNewOp<vector::BitCastOp>(castOp, resType,
adaptor.getSource());
return mlir::success();
}

private:
unsigned targetVectorBitWidth;
};

} // namespace

void mlir::vector::populateVectorLinearizeTypeConversionsAndLegality(
Expand All @@ -485,7 +525,7 @@ void mlir::vector::populateVectorLinearizeTypeConversionsAndLegality(
typeConverter.addTargetMaterialization(materializeCast);
target.markUnknownOpDynamicallyLegal(
[=](Operation *op) -> std::optional<bool> {
if ((isa<arith::ConstantOp>(op) ||
if ((isa<arith::ConstantOp>(op) || isa<vector::BitCastOp>(op) ||
op->hasTrait<OpTrait::Vectorizable>())) {
return (isLessThanTargetBitWidth(op, targetBitWidth)
? typeConverter.isLegal(op)
Expand All @@ -494,8 +534,9 @@ void mlir::vector::populateVectorLinearizeTypeConversionsAndLegality(
return std::nullopt;
});

patterns.add<LinearizeConstant, LinearizeVectorizable>(
typeConverter, patterns.getContext(), targetBitWidth);
patterns
.add<LinearizeConstant, LinearizeVectorizable, LinearizeVectorBitCast>(
typeConverter, patterns.getContext(), targetBitWidth);
}

void mlir::vector::populateVectorLinearizeShuffleLikeOpsPatterns(
Expand Down
67 changes: 66 additions & 1 deletion mlir/test/Dialect/Vector/linearize.mlir
View file
Open in desktop
Copy link
Contributor

Choose a reason for hiding this comment

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

Would you mind using more descriptive LIT variables ? (e.g. RES or DOWNCAST/UPCAST for vector.shape_cast) This would be more consistent with the existing convention within the file.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Yes, fixed them and thanks.

Original file line number Diff line number Diff line change
Expand Up @@ -179,7 +179,7 @@ func.func @test_extract_strided_slice_1(%arg0 : vector<4x8xf32>) -> vector<2x2xf

// ALL-LABEL: func.func @test_extract_strided_slice_1_scalable(
// ALL-SAME: %[[VAL_0:.*]]: vector<4x[8]xf32>) -> vector<2x[8]xf32> {
func.func @test_extract_strided_slice_1_scalable(%arg0: vector<4x[8]xf32>) -> vector<2x[8]xf32> {
func.func @test_extract_strided_slice_1_scalable(%arg0: vector<4x[8]xf32>) -> vector<2x[8]xf32> {
// ALL-NOT: vector.shuffle
// ALL-NOT: vector.shape_cast
// ALL: %[[RES:.*]] = vector.extract_strided_slice %[[VAL_0]] {offsets = [1, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x[8]xf32> to vector<2x[8]xf32>
Expand Down Expand Up @@ -318,3 +318,68 @@ func.func @test_vector_extract_scalar() {
%0 = vector.extract %cst[0] : i32 from vector<4xi32>
return
}

// -----

// ALL-LABEL: test_vector_bitcast
// ALL-SAME: %[[ARG_0:.*]]: vector<4x4xf32>
func.func @test_vector_bitcast(%arg0: vector<4x4xf32>) -> vector<4x8xf16> {
// DEFAULT: %[[DOWNCAST:.*]] = vector.shape_cast %[[ARG_0]] : vector<4x4xf32> to vector<16xf32>
// DEFAULT: %[[BITCAST:.*]] = vector.bitcast %[[DOWNCAST]] : vector<16xf32> to vector<32xf16>
// DEFAULT: %[[UPCAST:.*]] = vector.shape_cast %[[BITCAST]] : vector<32xf16> to vector<4x8xf16>

// BW-128: %[[UPCAST:.*]] = vector.bitcast %[[ARG_0]] : vector<4x4xf32> to vector<4x8xf16>
// BW-0: %[[BITCAST:.*]] = vector.bitcast %[[ARG_0]] : vector<4x4xf32> to vector<4x8xf16>
%1 = vector.bitcast %arg0 : vector<4x4xf32> to vector<4x8xf16>
return %1 : vector<4x8xf16>
}

// -----

// ALL-LABEL: test_vector_bitcast
// ALL-SAME: %[[ARG_0:.*]]: vector<4x2xf32>
func.func @test_vector_bitcast(%arg0: vector<4x2xf32>) -> vector<4x4xf16> {
// DEFAULT: %[[DOWNCAST:.*]] = vector.shape_cast %[[ARG_0]] : vector<4x2xf32> to vector<8xf32>
// DEFAULT: %[[BITCAST:.*]] = vector.bitcast %[[DOWNCAST]] : vector<8xf32> to vector<16xf16>
// DEFAULT: %[[UPCAST:.*]] = vector.shape_cast %[[BITCAST]] : vector<16xf16> to vector<4x4xf16>
// BW-128: %[[DOWNCAST:.*]] = vector.shape_cast %[[ARG_0]] : vector<4x2xf32> to vector<8xf32>
// BW-128: %[[BITCAST:.*]] = vector.bitcast %[[DOWNCAST]] : vector<8xf32> to vector<16xf16>
// BW-128: %[[UPCAST:.*]] = vector.shape_cast %[[BITCAST]] : vector<16xf16> to vector<4x4xf16>

// BW-0: %[[BITCAST:.*]] = vector.bitcast %[[ARG_0]] : vector<4x2xf32> to vector<4x4xf16>
%1 = vector.bitcast %arg0 : vector<4x2xf32> to vector<4x4xf16>
return %1 : vector<4x4xf16>
}

// -----

// ALL-LABEL: test_vector_bitcast
// ALL-SAME: %[[ARG_0:.*]]: vector<4x[2]xf32>
func.func @test_vector_bitcast(%arg0: vector<4x[2]xf32>) -> vector<4x[4]xf16> {
// DEFAULT: %[[DOWNCAST:.*]] = vector.shape_cast %[[ARG_0]] : vector<4x[2]xf32> to vector<[8]xf32>
// DEFAULT: %[[BITCAST:.*]] = vector.bitcast %[[DOWNCAST]] : vector<[8]xf32> to vector<[16]xf16>
// DEFAULT: %[[UPCAST:.*]] = vector.shape_cast %[[BITCAST]] : vector<[16]xf16> to vector<4x[4]xf16>
// BW-128: %[[DOWNCAST:.*]] = vector.shape_cast %[[ARG_0]] : vector<4x[2]xf32> to vector<[8]xf32>
// BW-128: %[[BITCAST:.*]] = vector.bitcast %[[DOWNCAST]] : vector<[8]xf32> to vector<[16]xf16>
// BW-128: %[[UPCAST:.*]] = vector.shape_cast %[[BITCAST]] : vector<[16]xf16> to vector<4x[4]xf16>

// BW-0: %[[BITCAST:.*]] = vector.bitcast %[[ARG_0]] : vector<4x[2]xf32> to vector<4x[4]xf16>
%1 = vector.bitcast %arg0 : vector<4x[2]xf32> to vector<4x[4]xf16>
return %1 : vector<4x[4]xf16>
}

// -----
// ALL-LABEL: test_vector_bitcast
// ALL-SAME: %[[ARG_0:.*]]: vector<[4]x2xf32>
func.func @test_vector_bitcast(%arg0: vector<[4]x2xf32>) -> vector<[4]x4xf16> {
// DEFAULT: %[[DOWNCAST:.*]] = vector.shape_cast %[[ARG_0]] : vector<[4]x2xf32> to vector<[8]xf32>
// DEFAULT: %[[BITCAST:.*]] = vector.bitcast %[[DOWNCAST]] : vector<[8]xf32> to vector<[16]xf16>
// DEFAULT: %[[UPCAST:.*]] = vector.shape_cast %[[BITCAST]] : vector<[16]xf16> to vector<[4]x4xf16>
// BW-128: %[[DOWNCAST:.*]] = vector.shape_cast %[[ARG_0]] : vector<[4]x2xf32> to vector<[8]xf32>
// BW-128: %[[BITCAST:.*]] = vector.bitcast %[[DOWNCAST]] : vector<[8]xf32> to vector<[16]xf16>
// BW-128: %[[UPCAST:.*]] = vector.shape_cast %[[BITCAST]] : vector<[16]xf16> to vector<[4]x4xf16>

// BW-0: %[[BITCAST:.*]] = vector.bitcast %[[ARG_0]] : vector<[4]x2xf32> to vector<[4]x4xf16>
%1 = vector.bitcast %arg0 : vector<[4]x2xf32> to vector<[4]x4xf16>
return %1 : vector<[4]x4xf16>
}
Loading