Skip to content

[mlir][spirv] Split codegen for float min/max reductions and others (NFC) #69023

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

Closed
wants to merge 1 commit into from
Closed

[mlir][spirv] Split codegen for float min/max reductions and others (NFC) #69023

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
130 changes: 101 additions & 29 deletions mlir/lib/Conversion/VectorToSPIRV/VectorToSPIRV.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/IR/BuiltinAttributes.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/Location.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/IR/TypeUtilities.h"
Expand Down Expand Up @@ -351,15 +352,13 @@ struct VectorInsertStridedSliceOpConvert final
}
};

template <class SPIRVFMaxOp, class SPIRVFMinOp, class SPIRVUMaxOp,
class SPIRVUMinOp, class SPIRVSMaxOp, class SPIRVSMinOp>
struct VectorReductionPattern final
: public OpConversionPattern<vector::ReductionOp> {
template <typename Derived>
struct VectorReductionPatternBase : OpConversionPattern<vector::ReductionOp> {
using OpConversionPattern::OpConversionPattern;

LogicalResult
matchAndRewrite(vector::ReductionOp reduceOp, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
ConversionPatternRewriter &rewriter) const final {
Type resultType = typeConverter->convertType(reduceOp.getType());
if (!resultType)
return failure();
Expand All @@ -368,9 +367,22 @@ struct VectorReductionPattern final
if (!srcVectorType || srcVectorType.getRank() != 1)
return rewriter.notifyMatchFailure(reduceOp, "not 1-D vector source");

// Extract all elements.
SmallVector<Value> extractedElements =
extractAllElements(reduceOp, adaptor, srcVectorType, rewriter);

const auto &self = static_cast<const Derived &>(*this);

return self.reduceExtracted(reduceOp, extractedElements, resultType,
rewriter);
}

private:
SmallVector<Value>
extractAllElements(vector::ReductionOp reduceOp, OpAdaptor adaptor,
VectorType srcVectorType,
ConversionPatternRewriter &rewriter) const {
int numElements = srcVectorType.getDimSize(0);
SmallVector<Value, 4> values;
SmallVector<Value> values;
values.reserve(numElements + (adaptor.getAcc() != nullptr));
Location loc = reduceOp.getLoc();
for (int i = 0; i < numElements; ++i) {
Expand All @@ -381,9 +393,26 @@ struct VectorReductionPattern final
if (Value acc = adaptor.getAcc())
values.push_back(acc);

// Reduce them.
Value result = values.front();
for (Value next : llvm::ArrayRef(values).drop_front()) {
return values;
}
};

#define VECTOR_REDUCTION_BASE \
VectorReductionPatternBase<VectorReductionPattern<SPIRVUMaxOp, SPIRVUMinOp, \
SPIRVSMaxOp, SPIRVSMinOp>>
Comment on lines +400 to +402
Copy link
Member

Choose a reason for hiding this comment

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

Can we make this a using declaration instead?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Unfortunately, no.
The macro is needed to avoid repetition between the base class and explicitly inheriting the constructor of the base class.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

There is other option though but it might look weird:

template <typename SPIRVUMaxOp, typename SPIRVUMinOp, typename SPIRVSMaxOp,
          typename SPIRVSMinOp, typename..., 
          typename Base = VectorReductionPatternBase<VectorReductionPattern<SPIRVUMaxOp, SPIRVUMinOp, SPIRVSMaxOp, SPIRVSMinOp>>>
struct VectorReductionPattern final : Base {
  using Base::Base;
  ...
};

typename... is needed to protect from possible wrongful overrides of the Base and typename Base is used here as an alias.

Copy link
Member

@kuhar kuhar Oct 14, 2023 •
edited
Loading

Choose a reason for hiding this comment

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

How about something like this:

template <typename... Args>
struct VectorReductionPattern;

typename <typename... Args>
using CrtpBase = VectorReductionPatternBase<VectorReductionPattern, Args...>;

template <typename... Args>
struct VectorReductionPattern final : CrtpBase<Args...> {
  using Base = CrtpBase<Args...>; 
   using Base::Base;
   // define SPIRVUMaxOp et al. in VectorReductionPatternBase
  ...
};

just an idea, I haven't tried compiling this

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Actually, it should work, I'll definitely give it a try! My idea wouldn't work because VectorReductionPattern is not declared before using it.

template <typename SPIRVUMaxOp, typename SPIRVUMinOp, typename SPIRVSMaxOp,
typename SPIRVSMinOp>
struct VectorReductionPattern final : VECTOR_REDUCTION_BASE {
using Base = VECTOR_REDUCTION_BASE;
using Base::Base;

LogicalResult reduceExtracted(vector::ReductionOp reduceOp,
ArrayRef<Value> extractedElements,
Type resultType,
ConversionPatternRewriter &rewriter) const {
mlir::Location loc = reduceOp->getLoc();
Value result = extractedElements.front();
for (Value next : llvm::ArrayRef(extractedElements).drop_front()) {
switch (reduceOp.getKind()) {

#define INT_AND_FLOAT_CASE(kind, iop, fop) \
Expand All @@ -403,10 +432,6 @@ struct VectorReductionPattern final

INT_AND_FLOAT_CASE(ADD, IAddOp, FAddOp);
INT_AND_FLOAT_CASE(MUL, IMulOp, FMulOp);
INT_OR_FLOAT_CASE(MAXIMUMF, SPIRVFMaxOp);
INT_OR_FLOAT_CASE(MINIMUMF, SPIRVFMinOp);
INT_OR_FLOAT_CASE(MAXF, SPIRVFMaxOp);
INT_OR_FLOAT_CASE(MINF, SPIRVFMinOp);
INT_OR_FLOAT_CASE(MINUI, SPIRVUMinOp);
INT_OR_FLOAT_CASE(MINSI, SPIRVSMinOp);
INT_OR_FLOAT_CASE(MAXUI, SPIRVUMaxOp);
Expand All @@ -416,13 +441,57 @@ struct VectorReductionPattern final
case vector::CombiningKind::OR:
case vector::CombiningKind::XOR:
return rewriter.notifyMatchFailure(reduceOp, "unimplemented");
default:
return rewriter.notifyMatchFailure(reduceOp, "not handled here");
}
}

rewriter.replaceOp(reduceOp, result);
return success();
}
};
#undef VECTOR_REDUCTION_BASE
#undef INT_AND_FLOAT_CASE
#undef INT_OR_FLOAT_CASE

#define MIN_MAX_PATTERN_BASE \
VectorReductionPatternBase< \
VectorReductionFloatMinMax<SPIRVFMaxOp, SPIRVFMinOp>>
template <class SPIRVFMaxOp, class SPIRVFMinOp>
struct VectorReductionFloatMinMax final : MIN_MAX_PATTERN_BASE {
using Base = MIN_MAX_PATTERN_BASE;
using Base::Base;

LogicalResult reduceExtracted(vector::ReductionOp reduceOp,
ArrayRef<Value> extractedElements,
Type resultType,
ConversionPatternRewriter &rewriter) const {
mlir::Location loc = reduceOp->getLoc();
Value result = extractedElements.front();
for (Value next : llvm::ArrayRef(extractedElements).drop_front()) {
switch (reduceOp.getKind()) {

#define INT_OR_FLOAT_CASE(kind, fop) \
case vector::CombiningKind::kind: \
result = rewriter.create<fop>(loc, resultType, result, next); \
break

INT_OR_FLOAT_CASE(MAXIMUMF, SPIRVFMaxOp);
INT_OR_FLOAT_CASE(MINIMUMF, SPIRVFMinOp);
INT_OR_FLOAT_CASE(MAXF, SPIRVFMaxOp);
INT_OR_FLOAT_CASE(MINF, SPIRVFMinOp);

default:
return rewriter.notifyMatchFailure(reduceOp, "not handled here");
}
}

rewriter.replaceOp(reduceOp, result);
return success();
}
};
#undef MIN_MAX_PATTERN_BASE
#undef INT_OR_FLOAT_CASE

class VectorSplatPattern final : public OpConversionPattern<vector::SplatOp> {
public:
Expand Down Expand Up @@ -604,25 +673,28 @@ struct VectorReductionToDotProd final : OpRewritePattern<vector::ReductionOp> {
};

} // namespace
#define CL_MAX_MIN_OPS \
spirv::CLFMaxOp, spirv::CLFMinOp, spirv::CLUMaxOp, spirv::CLUMinOp, \
spirv::CLSMaxOp, spirv::CLSMinOp
#define CL_INT_MAX_MIN_OPS \
spirv::CLUMaxOp, spirv::CLUMinOp, spirv::CLSMaxOp, spirv::CLSMinOp

#define GL_INT_MAX_MIN_OPS \
spirv::GLUMaxOp, spirv::GLUMinOp, spirv::GLSMaxOp, spirv::GLSMinOp

#define GL_MAX_MIN_OPS \
spirv::GLFMaxOp, spirv::GLFMinOp, spirv::GLUMaxOp, spirv::GLUMinOp, \
spirv::GLSMaxOp, spirv::GLSMinOp
#define CL_FLOAT_MAX_MIN_OPS spirv::CLFMaxOp, spirv::CLFMinOp
#define GL_FLOAT_MAX_MIN_OPS spirv::GLFMaxOp, spirv::GLFMinOp

void mlir::populateVectorToSPIRVPatterns(SPIRVTypeConverter &typeConverter,
RewritePatternSet &patterns) {
patterns.add<VectorBitcastConvert, VectorBroadcastConvert,
VectorExtractElementOpConvert, VectorExtractOpConvert,
VectorExtractStridedSliceOpConvert,
VectorFmaOpConvert<spirv::GLFmaOp>,
VectorFmaOpConvert<spirv::CLFmaOp>, VectorInsertElementOpConvert,
VectorInsertOpConvert, VectorReductionPattern<GL_MAX_MIN_OPS>,
VectorReductionPattern<CL_MAX_MIN_OPS>, VectorShapeCast,
VectorInsertStridedSliceOpConvert, VectorShuffleOpConvert,
VectorSplatPattern>(typeConverter, patterns.getContext());
patterns.add<
VectorBitcastConvert, VectorBroadcastConvert,
VectorExtractElementOpConvert, VectorExtractOpConvert,
VectorExtractStridedSliceOpConvert, VectorFmaOpConvert<spirv::GLFmaOp>,
VectorFmaOpConvert<spirv::CLFmaOp>, VectorInsertElementOpConvert,
VectorInsertOpConvert, VectorReductionPattern<GL_INT_MAX_MIN_OPS>,
VectorReductionPattern<CL_INT_MAX_MIN_OPS>,
VectorReductionFloatMinMax<CL_FLOAT_MAX_MIN_OPS>,
VectorReductionFloatMinMax<GL_FLOAT_MAX_MIN_OPS>, VectorShapeCast,
VectorInsertStridedSliceOpConvert, VectorShuffleOpConvert,
VectorSplatPattern>(typeConverter, patterns.getContext());
}

void mlir::populateVectorReductionToSPIRVDotProductPatterns(
Expand Down