shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
llvm-project/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
Andrzej Warzyński c45cc3e420
[mlir][vector] Standardize base Naming Across Vector Ops (NFC) (#137859) 
[mlir][vector] Standardize base Naming Across Vector Ops (NFC)

This change standardizes the naming convention for the argument
representing the value to read from or write to in Vector ops that
interface with Tensors or MemRefs. Specifically, it ensures that all
such ops use the name `base` (i.e., the base address or location to
which offsets are applied).

Updated operations:

* `vector.transfer_read`,
* `vector.transfer_write`.

For reference, these ops already use `base`:

* `vector.load`, `vector.store`, `vector.scatter`, `vector.gather`,
  `vector.expandload`, `vector.compressstore`, `vector.maskedstore`,
  `vector.maskedload`.

This is a non-functional change (NFC) and does not alter the semantics of these
operations. However, it does require users of the XFer ops to switch from
`op.getSource()` to `op.getBase()`.

To ease the transition, this PR temporarily adds a `getSource()` interface
method for compatibility. This is intended for downstream use only and should
not be relied on upstream. The method will be removed prior to the LLVM 21
release.

Implements #131602
2025-05-12 09:44:50 +01:00

6904 lines
270 KiB
C++
Raw Blame History

//===- VectorOps.cpp - MLIR Vector 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
//
//===----------------------------------------------------------------------===//
//
// This file implements convenience types for working with super-vectorization
// operations, in particular super-vector loads and stores.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/Conversion/ConvertToLLVM/ToLLVMInterface.h"
#include "mlir/Dialect/Affine/IR/ValueBoundsOpInterfaceImpl.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Arith/Utils/Utils.h"
#include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/UB/IR/UBOps.h"
#include "mlir/Dialect/Utils/IndexingUtils.h"
#include "mlir/Dialect/Utils/StructuredOpsUtils.h"
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/AffineMap.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinAttributes.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/DialectImplementation.h"
#include "mlir/IR/IRMapping.h"
#include "mlir/IR/OpImplementation.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/IR/TypeUtilities.h"
#include "mlir/Interfaces/SubsetOpInterface.h"
#include "mlir/Interfaces/ValueBoundsOpInterface.h"
#include "mlir/Support/LLVM.h"
#include "mlir/Transforms/InliningUtils.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringSet.h"
#include "llvm/ADT/TypeSwitch.h"
#include "llvm/Support/Casting.h"
#include <cassert>
#include <cstdint>
#include <numeric>
#include "mlir/Dialect/Vector/IR/VectorDialect.cpp.inc"
// Pull in all enum type and utility function definitions.
#include "mlir/Dialect/Vector/IR/VectorEnums.cpp.inc"
using namespace mlir;
using namespace mlir::vector;
/// Helper enum to classify mask value.
enum class MaskFormat {
AllTrue = 0,
AllFalse = 1,
Unknown = 2,
};
/// Helper method to classify a mask value. Currently, the method
/// looks "under the hood" of a constant value with dense attributes
/// and a constant mask operation (since the client may be called at
/// various stages during progressive lowering).
static MaskFormat getMaskFormat(Value mask) {
if (auto c = mask.getDefiningOp<arith::ConstantOp>()) {
// Inspect constant dense values. We count up for bits that
// are set, count down for bits that are cleared, and bail
// when a mix is detected.
if (auto denseElts = llvm::dyn_cast<DenseIntElementsAttr>(c.getValue())) {
int64_t val = 0;
for (bool b : denseElts.getValues<bool>())
if (b && val >= 0)
val++;
else if (!b && val <= 0)
val--;
else
return MaskFormat::Unknown;
if (val > 0)
return MaskFormat::AllTrue;
if (val < 0)
return MaskFormat::AllFalse;
}
} else if (auto m = mask.getDefiningOp<ConstantMaskOp>()) {
// Inspect constant mask index. If the index exceeds the
// dimension size, all bits are set. If the index is zero
// or less, no bits are set.
ArrayRef<int64_t> masks = m.getMaskDimSizes();
auto shape = m.getType().getShape();
bool allTrue = true;
bool allFalse = true;
for (auto [maskIdx, dimSize] : llvm::zip_equal(masks, shape)) {
if (maskIdx < dimSize)
allTrue = false;
if (maskIdx > 0)
allFalse = false;
}
if (allTrue)
return MaskFormat::AllTrue;
if (allFalse)
return MaskFormat::AllFalse;
} else if (auto m = mask.getDefiningOp<CreateMaskOp>()) {
// Finds all-false create_masks. An all-true create_mask requires all
// dims to be constants, so that'll be folded to a constant_mask, then
// detected in the constant_mask case.
auto maskOperands = m.getOperands();
for (Value operand : maskOperands) {
if (auto constantOp = operand.getDefiningOp<arith::ConstantOp>()) {
int64_t dimSize =
llvm::cast<IntegerAttr>(constantOp.getValue()).getInt();
if (dimSize <= 0)
return MaskFormat::AllFalse;
}
}
return MaskFormat::Unknown;
}
return MaskFormat::Unknown;
}
/// Default callback to build a region with a 'vector.yield' terminator with no
/// arguments.
void mlir::vector::buildTerminatedBody(OpBuilder &builder, Location loc) {
builder.create<vector::YieldOp>(loc);
}
// Helper for verifying combining kinds in contractions and reductions.
static bool isSupportedCombiningKind(CombiningKind combiningKind,
Type elementType) {
switch (combiningKind) {
case CombiningKind::ADD:
case CombiningKind::MUL:
return elementType.isIntOrIndexOrFloat();
case CombiningKind::MINUI:
case CombiningKind::MINSI:
case CombiningKind::MAXUI:
case CombiningKind::MAXSI:
case CombiningKind::AND:
case CombiningKind::OR:
case CombiningKind::XOR:
return elementType.isIntOrIndex();
case CombiningKind::MINNUMF:
case CombiningKind::MAXNUMF:
case CombiningKind::MINIMUMF:
case CombiningKind::MAXIMUMF:
return llvm::isa<FloatType>(elementType);
}
return false;
}
/// Returns the effective rank of the vector to read/write for Xfer Ops
///
/// When the element type of the shaped type is _a scalar_, this will simply
/// return the rank of the vector ( the result for xfer_read or the value to
/// store for xfer_write).
///
/// When the element type of the base shaped type is _a vector_, returns the
/// difference between the original vector type and the element type of the
/// shaped type.
///
/// EXAMPLE 1 (element type is _a scalar_):
/// - shapedType = tensor<10x20xf32>, vectorType = vector<2x4xf32>
/// - shapedType.getElementType() = f32 (rank 0)
/// - vectorType.getRank() = 2
/// - Result = 2 - 0 = 2
///
/// EXAMPLE 2 (element type is _a vector_):
/// - shapedType = tensor<10xvector<20xf32>>, vectorType = vector<20xf32>
/// - shapedType.getElementType() = vector<20xf32> (rank 1)
/// - vectorType.getRank() = 1
/// - Result = 1 - 1 = 0
///
/// This is used to determine the number of minor dimensions for identity maps
/// in vector transfer Ops.
static unsigned getEffectiveVectorRankForXferOp(ShapedType shapedType,
VectorType vectorType) {
unsigned elementVectorRank = 0;
VectorType elementVectorType =
llvm::dyn_cast<VectorType>(shapedType.getElementType());
if (elementVectorType)
elementVectorRank += elementVectorType.getRank();
return vectorType.getRank() - elementVectorRank;
}
AffineMap mlir::vector::getTransferMinorIdentityMap(ShapedType shapedType,
VectorType vectorType) {
// 0-d transfers are to/from tensor<t>/memref<t> and vector<1xt>.
// TODO: replace once we have 0-d vectors.
if (shapedType.getRank() == 0 &&
vectorType.getShape() == ArrayRef<int64_t>{1})
return AffineMap::get(
/*numDims=*/0, /*numSymbols=*/0,
getAffineConstantExpr(0, shapedType.getContext()));
return AffineMap::getMinorIdentityMap(
shapedType.getRank(),
getEffectiveVectorRankForXferOp(shapedType, vectorType),
shapedType.getContext());
}
/// Check if `write` is of a constant splat and the masked `read` is padded with
/// the same splat value -- meaning it could be the same value as the initial
/// constant splat.
static bool isSplatWriteConsistentWithMaskedRead(vector::TransferWriteOp write,
vector::TransferReadOp read) {
auto readMask = read.getMask();
auto writeMask = write.getMask();
// Check if the masks are consistent. The splat value could be the same if the
// read is masked (and padded with the splat value), and the write is unmasked
// or has the same mask. Note this does not allow the case where the write is
// masked and the read is unmasked, as then the read could be of more elements
// than the write (which may not be the same value).
bool couldBeSameSplat = readMask && (!writeMask || writeMask == readMask);
if (!couldBeSameSplat)
return false;
// Check for constant splat (as the source of the write).
DenseElementsAttr splatAttr;
if (!matchPattern(write.getVector(),
m_Constant<DenseElementsAttr>(&splatAttr)) ||
!splatAttr.isSplat()) {
return false;
}
// The padding of the read and the constant splat value must be the same.
Attribute padAttr;
if (!matchPattern(read.getPadding(), m_Constant(&padAttr)))
return false;
return padAttr == splatAttr.getSplatValue<Attribute>();
}
bool mlir::vector::checkSameValueRAW(vector::TransferWriteOp defWrite,
vector::TransferReadOp read) {
return !defWrite.hasOutOfBoundsDim() &&
defWrite.getIndices() == read.getIndices() &&
defWrite.getVectorType() == read.getVectorType() &&
defWrite.getPermutationMap() == read.getPermutationMap() &&
((!defWrite.getMask() && !read.getMask()) ||
isSplatWriteConsistentWithMaskedRead(defWrite, read));
}
bool mlir::vector::checkSameValueWAW(vector::TransferWriteOp write,
vector::TransferWriteOp priorWrite) {
return priorWrite.getIndices() == write.getIndices() &&
priorWrite.getMask() == write.getMask() &&
priorWrite.getVectorType() == write.getVectorType() &&
priorWrite.getPermutationMap() == write.getPermutationMap();
}
bool mlir::vector::isDisjointTransferIndices(
VectorTransferOpInterface transferA, VectorTransferOpInterface transferB,
bool testDynamicValueUsingBounds) {
// For simplicity only look at transfer of same type.
if (transferA.getVectorType() != transferB.getVectorType())
return false;
unsigned rankOffset = transferA.getLeadingShapedRank();
for (unsigned i = 0, e = transferA.getIndices().size(); i < e; i++) {
Value indexA = transferA.getIndices()[i];
Value indexB = transferB.getIndices()[i];
std::optional<int64_t> cstIndexA = getConstantIntValue(indexA);
std::optional<int64_t> cstIndexB = getConstantIntValue(indexB);
if (i < rankOffset) {
// For leading dimensions, if we can prove that index are different we
// know we are accessing disjoint slices.
if (cstIndexA.has_value() && cstIndexB.has_value()) {
if (*cstIndexA != *cstIndexB)
return true;
continue;
}
if (testDynamicValueUsingBounds) {
// First try to see if we can fully compose and simplify the affine
// expression as a fast track.
FailureOr<uint64_t> delta =
affine::fullyComposeAndComputeConstantDelta(indexA, indexB);
if (succeeded(delta) && *delta != 0)
return true;
FailureOr<bool> testEqual =
ValueBoundsConstraintSet::areEqual(indexA, indexB);
if (succeeded(testEqual) && !testEqual.value())
return true;
}
} else {
// For this dimension, we slice a part of the memref we need to make sure
// the intervals accessed don't overlap.
int64_t vectorDim = transferA.getVectorType().getDimSize(i - rankOffset);
if (cstIndexA.has_value() && cstIndexB.has_value()) {
int64_t distance = std::abs(*cstIndexA - *cstIndexB);
if (distance >= vectorDim)
return true;
continue;
}
if (testDynamicValueUsingBounds) {
// First try to see if we can fully compose and simplify the affine
// expression as a fast track.
FailureOr<int64_t> delta =
affine::fullyComposeAndComputeConstantDelta(indexA, indexB);
if (succeeded(delta) && std::abs(*delta) >= vectorDim)
return true;
FailureOr<int64_t> computeDelta =
ValueBoundsConstraintSet::computeConstantDelta(indexA, indexB);
if (succeeded(computeDelta)) {
if (std::abs(computeDelta.value()) >= vectorDim)
return true;
}
}
}
}
return false;
}
bool mlir::vector::isDisjointTransferSet(VectorTransferOpInterface transferA,
VectorTransferOpInterface transferB,
bool testDynamicValueUsingBounds) {
if (transferA.getBase() != transferB.getBase())
return false;
return isDisjointTransferIndices(transferA, transferB,
testDynamicValueUsingBounds);
}
// Helper to iterate over n-D vector slice elements. Calculate the next
// `position` in the n-D vector of size `shape`, applying an offset `offsets`.
// Modifies the `position` in place. Returns a failure when `position` becomes
// the end position.
static LogicalResult incSlicePosition(MutableArrayRef<int64_t> position,
ArrayRef<int64_t> shape,
ArrayRef<int64_t> offsets) {
for (auto [posInDim, dimSize, offsetInDim] :
llvm::reverse(llvm::zip_equal(position, shape, offsets))) {
++posInDim;
if (posInDim < dimSize + offsetInDim)
return success();
// Carry the overflow to the next loop iteration.
posInDim = offsetInDim;
}
return failure();
}
/// Returns the integer numbers in `values`. `values` are expected to be
/// constant operations.
SmallVector<int64_t> vector::getAsIntegers(ArrayRef<Value> values) {
SmallVector<int64_t> ints;
llvm::transform(values, std::back_inserter(ints), [](Value value) {
auto constOp = value.getDefiningOp<arith::ConstantIndexOp>();
assert(constOp && "Unexpected non-constant index");
return constOp.value();
});
return ints;
}
/// Returns the integer numbers in `foldResults`. `foldResults` are expected to
/// be constant operations.
SmallVector<int64_t> vector::getAsIntegers(ArrayRef<OpFoldResult> foldResults) {
SmallVector<int64_t> ints;
llvm::transform(
foldResults, std::back_inserter(ints), [](OpFoldResult foldResult) {
assert(isa<Attribute>(foldResult) && "Unexpected non-constant index");
return cast<IntegerAttr>(cast<Attribute>(foldResult)).getInt();
});
return ints;
}
/// Convert `foldResults` into Values. Integer attributes are converted to
/// constant op.
SmallVector<Value> vector::getAsValues(OpBuilder &builder, Location loc,
ArrayRef<OpFoldResult> foldResults) {
SmallVector<Value> values;
llvm::transform(foldResults, std::back_inserter(values),
[&](OpFoldResult foldResult) {
if (auto attr = dyn_cast<Attribute>(foldResult))
return builder
.create<arith::ConstantIndexOp>(
loc, cast<IntegerAttr>(attr).getInt())
.getResult();
return cast<Value>(foldResult);
});
return values;
}
std::optional<int64_t> vector::getConstantVscaleMultiplier(Value value) {
if (value.getDefiningOp<vector::VectorScaleOp>())
return 1;
auto mul = value.getDefiningOp<arith::MulIOp>();
if (!mul)
return {};
auto lhs = mul.getLhs();
auto rhs = mul.getRhs();
if (lhs.getDefiningOp<vector::VectorScaleOp>())
return getConstantIntValue(rhs);
if (rhs.getDefiningOp<vector::VectorScaleOp>())
return getConstantIntValue(lhs);
return {};
}
//===----------------------------------------------------------------------===//
// CombiningKindAttr
//===----------------------------------------------------------------------===//
namespace mlir {
namespace vector {
namespace detail {
struct BitmaskEnumStorage : public AttributeStorage {
using KeyTy = uint64_t;
BitmaskEnumStorage(KeyTy val) : value(val) {}
bool operator==(const KeyTy &key) const { return value == key; }
static BitmaskEnumStorage *construct(AttributeStorageAllocator &allocator,
const KeyTy &key) {
return new (allocator.allocate<BitmaskEnumStorage>())
BitmaskEnumStorage(key);
}
KeyTy value = 0;
};
} // namespace detail
} // namespace vector
} // namespace mlir
//===----------------------------------------------------------------------===//
// VectorDialect
//===----------------------------------------------------------------------===//
namespace {
/// This class defines the interface for handling inlining with vector dialect
/// operations.
struct VectorInlinerInterface : public DialectInlinerInterface {
using DialectInlinerInterface::DialectInlinerInterface;
/// All vector dialect ops can be inlined.
bool isLegalToInline(Operation *, Region *, bool, IRMapping &) const final {
return true;
}
};
} // namespace
void VectorDialect::initialize() {
addAttributes<
#define GET_ATTRDEF_LIST
#include "mlir/Dialect/Vector/IR/VectorAttributes.cpp.inc"
>();
addOperations<
#define GET_OP_LIST
#include "mlir/Dialect/Vector/IR/VectorOps.cpp.inc"
>();
addInterfaces<VectorInlinerInterface>();
declarePromisedInterfaces<bufferization::BufferizableOpInterface,
TransferReadOp, TransferWriteOp, GatherOp, MaskOp,
YieldOp>();
declarePromisedInterfaces<SubsetOpInterface, TransferReadOp,
TransferWriteOp>();
declarePromisedInterface<SubsetExtractionOpInterface, TransferReadOp>();
declarePromisedInterface<SubsetInsertionOpInterface, TransferWriteOp>();
declarePromisedInterface<ConvertToLLVMPatternInterface, VectorDialect>();
}
/// Materialize a single constant operation from a given attribute value with
/// the desired resultant type.
Operation *VectorDialect::materializeConstant(OpBuilder &builder,
Attribute value, Type type,
Location loc) {
if (isa<ub::PoisonAttrInterface>(value))
return value.getDialect().materializeConstant(builder, value, type, loc);
return arith::ConstantOp::materialize(builder, value, type, loc);
}
IntegerType vector::getVectorSubscriptType(Builder &builder) {
return builder.getIntegerType(64);
}
ArrayAttr vector::getVectorSubscriptAttr(Builder &builder,
ArrayRef<int64_t> values) {
return builder.getI64ArrayAttr(values);
}
//===----------------------------------------------------------------------===//
// MultiDimReductionOp
//===----------------------------------------------------------------------===//
void vector::MultiDimReductionOp::build(OpBuilder &builder,
OperationState &result, Value source,
Value acc, ArrayRef<bool> reductionMask,
CombiningKind kind) {
SmallVector<int64_t> reductionDims;
for (const auto &en : llvm::enumerate(reductionMask))
if (en.value())
reductionDims.push_back(en.index());
build(builder, result, kind, source, acc, reductionDims);
}
OpFoldResult MultiDimReductionOp::fold(FoldAdaptor adaptor) {
// Single parallel dim, this is a noop.
if (getSourceVectorType().getRank() == 1 && !isReducedDim(0))
return getSource();
return {};
}
std::optional<SmallVector<int64_t, 4>>
MultiDimReductionOp::getShapeForUnroll() {
return llvm::to_vector<4>(getSourceVectorType().getShape());
}
LogicalResult MultiDimReductionOp::verify() {
SmallVector<int64_t> targetShape;
SmallVector<bool> scalableDims;
Type inferredReturnType;
auto sourceScalableDims = getSourceVectorType().getScalableDims();
for (auto [dimIdx, dimSize] :
llvm::enumerate(getSourceVectorType().getShape()))
if (!llvm::any_of(getReductionDims(),
[dimIdx = dimIdx](int64_t reductionDimIdx) {
return reductionDimIdx == static_cast<int64_t>(dimIdx);
})) {
targetShape.push_back(dimSize);
scalableDims.push_back(sourceScalableDims[dimIdx]);
}
// TODO: update to also allow 0-d vectors when available.
if (targetShape.empty())
inferredReturnType = getSourceVectorType().getElementType();
else
inferredReturnType = VectorType::get(
targetShape, getSourceVectorType().getElementType(), scalableDims);
if (getType() != inferredReturnType)
return emitOpError() << "destination type " << getType()
<< " is incompatible with source type "
<< getSourceVectorType();
return success();
}
/// Returns the mask type expected by this operation.
Type MultiDimReductionOp::getExpectedMaskType() {
auto vecType = getSourceVectorType();
return VectorType::get(vecType.getShape(),
IntegerType::get(vecType.getContext(), /*width=*/1),
vecType.getScalableDims());
}
namespace {
// Only unit dimensions that are being reduced are folded. If the dimension is
// unit, but not reduced, it is not folded, thereby keeping the output type the
// same. If not all dimensions which are reduced are of unit dimension, this
// transformation does nothing. This is just a generalization of
// ElideSingleElementReduction for ReduceOp.
struct ElideUnitDimsInMultiDimReduction
: public OpRewritePattern<MultiDimReductionOp> {
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(MultiDimReductionOp reductionOp,
PatternRewriter &rewriter) const override {
ArrayRef<int64_t> shape = reductionOp.getSourceVectorType().getShape();
for (const auto &dim : enumerate(shape)) {
if (reductionOp.isReducedDim(dim.index()) && dim.value() != 1)
return failure();
}
// Vector mask setup.
OpBuilder::InsertionGuard guard(rewriter);
Operation *rootOp;
Value mask;
if (reductionOp.isMasked()) {
rewriter.setInsertionPoint(reductionOp.getMaskingOp());
rootOp = reductionOp.getMaskingOp();
mask = reductionOp.getMaskingOp().getMask();
} else {
rootOp = reductionOp;
}
Location loc = reductionOp.getLoc();
Value acc = reductionOp.getAcc();
Value cast;
if (auto dstVecType = dyn_cast<VectorType>(reductionOp.getDestType())) {
if (mask) {
VectorType newMaskType =
VectorType::get(dstVecType.getShape(), rewriter.getI1Type(),
dstVecType.getScalableDims());
mask = rewriter.create<vector::ShapeCastOp>(loc, newMaskType, mask);
}
cast = rewriter.create<vector::ShapeCastOp>(
loc, reductionOp.getDestType(), reductionOp.getSource());
} else {
// This means we are reducing all the dimensions, and all reduction
// dimensions are of size 1. So a simple extraction would do.
if (mask)
mask = rewriter.create<vector::ExtractOp>(loc, mask);
cast = rewriter.create<vector::ExtractOp>(loc, reductionOp.getSource());
}
Value result =
vector::makeArithReduction(rewriter, loc, reductionOp.getKind(), acc,
cast, /*fastmath=*/nullptr, mask);
rewriter.replaceOp(rootOp, result);
return success();
}
};
} // namespace
void MultiDimReductionOp::getCanonicalizationPatterns(
RewritePatternSet &results, MLIRContext *context) {
results.add<ElideUnitDimsInMultiDimReduction>(context);
}
//===----------------------------------------------------------------------===//
// ReductionOp
//===----------------------------------------------------------------------===//
void vector::ReductionOp::build(OpBuilder &builder, OperationState &result,
CombiningKind kind, Value vector,
arith::FastMathFlags fastMathFlags) {
build(builder, result, kind, vector, /*acc=*/Value(), fastMathFlags);
}
void vector::ReductionOp::build(OpBuilder &builder, OperationState &result,
CombiningKind kind, Value vector, Value acc,
arith::FastMathFlags fastMathFlags) {
build(builder, result,
llvm::cast<VectorType>(vector.getType()).getElementType(), kind, vector,
acc, fastMathFlags);
}
LogicalResult ReductionOp::verify() {
// Verify for 0-D and 1-D vector.
int64_t rank = getSourceVectorType().getRank();
if (rank > 1)
return emitOpError("unsupported reduction rank: ") << rank;
// Verify supported reduction kind.
Type eltType = getDest().getType();
if (!isSupportedCombiningKind(getKind(), eltType))
return emitOpError("unsupported reduction type '")
<< eltType << "' for kind '" << stringifyCombiningKind(getKind())
<< "'";
return success();
}
// MaskableOpInterface methods.
/// Returns the mask type expected by this operation.
Type ReductionOp::getExpectedMaskType() {
auto vecType = getSourceVectorType();
return VectorType::get(vecType.getShape(),
IntegerType::get(vecType.getContext(), /*width=*/1),
vecType.getScalableDims());
}
Value mlir::vector::getVectorReductionOp(arith::AtomicRMWKind op,
OpBuilder &builder, Location loc,
Value vector) {
switch (op) {
case arith::AtomicRMWKind::addf:
case arith::AtomicRMWKind::addi:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::ADD, vector);
case arith::AtomicRMWKind::mulf:
case arith::AtomicRMWKind::muli:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::MUL, vector);
case arith::AtomicRMWKind::minimumf:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::MINIMUMF, vector);
case arith::AtomicRMWKind::mins:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::MINSI, vector);
case arith::AtomicRMWKind::minu:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::MINUI, vector);
case arith::AtomicRMWKind::maximumf:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::MAXIMUMF, vector);
case arith::AtomicRMWKind::maxs:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::MAXSI, vector);
case arith::AtomicRMWKind::maxu:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::MAXUI, vector);
case arith::AtomicRMWKind::andi:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::AND, vector);
case arith::AtomicRMWKind::ori:
return builder.create<vector::ReductionOp>(vector.getLoc(),
CombiningKind::OR, vector);
// TODO: Add remaining reduction operations.
default:
(void)emitOptionalError(loc, "Reduction operation type not supported");
break;
}
return nullptr;
}
std::optional<SmallVector<int64_t, 4>> ReductionOp::getShapeForUnroll() {
return llvm::to_vector<4>(getSourceVectorType().getShape());
}
namespace {
struct ElideSingleElementReduction : public OpRewritePattern<ReductionOp> {
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ReductionOp reductionOp,
PatternRewriter &rewriter) const override {
// Vector mask setup.
OpBuilder::InsertionGuard guard(rewriter);
auto maskableOp =
cast<vector::MaskableOpInterface>(reductionOp.getOperation());
Operation *rootOp;
Value mask;
if (maskableOp.isMasked()) {
rewriter.setInsertionPoint(maskableOp.getMaskingOp());
rootOp = maskableOp.getMaskingOp();
mask = maskableOp.getMaskingOp().getMask();
} else {
rootOp = reductionOp;
}
auto vectorType = reductionOp.getSourceVectorType();
if (vectorType.getRank() != 0 && vectorType.getDimSize(0) != 1)
return failure();
Location loc = reductionOp.getLoc();
if (mask)
mask = rewriter.create<ExtractOp>(loc, mask);
Value result = rewriter.create<ExtractOp>(loc, reductionOp.getVector());
if (Value acc = reductionOp.getAcc())
result = vector::makeArithReduction(rewriter, loc, reductionOp.getKind(),
result, acc,
reductionOp.getFastmathAttr(), mask);
rewriter.replaceOp(rootOp, result);
return success();
}
};
} // namespace
void ReductionOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<ElideSingleElementReduction>(context);
}
//===----------------------------------------------------------------------===//
// ContractionOp
//===----------------------------------------------------------------------===//
void vector::ContractionOp::build(OpBuilder &builder, OperationState &result,
Value lhs, Value rhs, Value acc,
ArrayRef<ArrayRef<AffineExpr>> indexingExprs,
ArrayRef<IteratorType> iteratorTypes) {
result.addOperands({lhs, rhs, acc});
result.addTypes(acc.getType());
result.addAttribute(
getIndexingMapsAttrName(result.name),
builder.getAffineMapArrayAttr(
AffineMap::inferFromExprList(indexingExprs, builder.getContext())));
result.addAttribute(
getIteratorTypesAttrName(result.name),
builder.getArrayAttr(llvm::to_vector(llvm::map_range(
iteratorTypes, [&](IteratorType t) -> mlir::Attribute {
return IteratorTypeAttr::get(builder.getContext(), t);
}))));
}
void vector::ContractionOp::build(OpBuilder &builder, OperationState &result,
Value lhs, Value rhs, Value acc,
ArrayAttr indexingMaps,
ArrayAttr iteratorTypes) {
build(builder, result, lhs, rhs, acc, indexingMaps, iteratorTypes,
ContractionOp::getDefaultKind());
}
void vector::ContractionOp::build(OpBuilder &builder, OperationState &result,
Value lhs, Value rhs, Value acc,
ArrayAttr indexingMaps,
ArrayAttr iteratorTypes, CombiningKind kind) {
result.addOperands({lhs, rhs, acc});
result.addTypes(acc.getType());
result.addAttribute(getIndexingMapsAttrName(result.name), indexingMaps);
result.addAttribute(getIteratorTypesAttrName(result.name), iteratorTypes);
result.addAttribute(getKindAttrName(result.name),
CombiningKindAttr::get(builder.getContext(), kind));
}
ParseResult ContractionOp::parse(OpAsmParser &parser, OperationState &result) {
OpAsmParser::UnresolvedOperand lhsInfo;
OpAsmParser::UnresolvedOperand rhsInfo;
OpAsmParser::UnresolvedOperand accInfo;
SmallVector<OpAsmParser::UnresolvedOperand, 2> masksInfo;
SmallVector<Type, 2> types;
Type resultType;
auto loc = parser.getCurrentLocation();
DictionaryAttr dictAttr;
// TODO: Unify linalg op attribute parsing.
if (parser.parseAttribute(dictAttr) || parser.parseOperand(lhsInfo) ||
parser.parseComma() || parser.parseOperand(rhsInfo) ||
parser.parseComma() || parser.parseOperand(accInfo) ||
parser.parseTrailingOperandList(masksInfo) ||
parser.parseOptionalAttrDict(result.attributes) ||
parser.parseColonTypeList(types) ||
parser.parseKeywordType("into", resultType) ||
parser.resolveOperand(lhsInfo, types[0], result.operands) ||
parser.resolveOperand(rhsInfo, types[1], result.operands) ||
parser.resolveOperand(accInfo, resultType, result.operands) ||
parser.addTypeToList(resultType, result.types))
return failure();
result.attributes.append(dictAttr.getValue().begin(),
dictAttr.getValue().end());
// Convert array of string into an array of IteratyType enums. This is needed,
// because tests still use the old format when 'iterator_types' attribute is
// represented as an array of strings.
// TODO: Remove this conversion once tests are fixed.
auto iteratorTypes = dyn_cast_or_null<ArrayAttr>(
result.attributes.get(getIteratorTypesAttrName(result.name)));
if (!iteratorTypes) {
return parser.emitError(loc)
<< "expected " << getIteratorTypesAttrName(result.name)
<< " array attribute";
}
SmallVector<Attribute> iteratorTypeAttrs;
for (StringRef s : iteratorTypes.getAsValueRange<StringAttr>()) {
auto maybeIteratorType = symbolizeIteratorType(s);
if (!maybeIteratorType.has_value())
return parser.emitError(loc) << "unexpected iterator_type (" << s << ")";
iteratorTypeAttrs.push_back(
IteratorTypeAttr::get(parser.getContext(), maybeIteratorType.value()));
}
result.attributes.set(getIteratorTypesAttrName(result.name),
parser.getBuilder().getArrayAttr(iteratorTypeAttrs));
if (!result.attributes.get(getKindAttrName(result.name))) {
result.addAttribute(
getKindAttrName(result.name),
CombiningKindAttr::get(result.getContext(),
ContractionOp::getDefaultKind()));
}
if (masksInfo.empty())
return success();
if (masksInfo.size() != 2)
return parser.emitError(parser.getNameLoc(),
"expected zero or exactly 2 vector mask operands");
auto lhsType = llvm::cast<VectorType>(types[0]);
auto rhsType = llvm::cast<VectorType>(types[1]);
auto maskElementType = parser.getBuilder().getI1Type();
std::array<VectorType, 2> maskTypes = {
VectorType::Builder(lhsType).setElementType(maskElementType),
VectorType::Builder(rhsType).setElementType(maskElementType)};
if (parser.resolveOperands(masksInfo, maskTypes, loc, result.operands))
return failure();
return success();
}
void ContractionOp::print(OpAsmPrinter &p) {
// TODO: Unify printing code with linalg ops.
auto attrNames = getTraitAttrNames();
llvm::StringSet<> traitAttrsSet;
traitAttrsSet.insert_range(attrNames);
SmallVector<NamedAttribute, 8> attrs;
for (auto attr : (*this)->getAttrs()) {
if (attr.getName() == getIteratorTypesAttrName()) {
auto iteratorTypes =
llvm::cast<ArrayAttr>(attr.getValue())
.getAsValueRange<IteratorTypeAttr, IteratorType>();
// Convert IteratorType enums into the string representation. This is
// needed, because tests still use the old format when 'iterator_types'
// attribute is represented as an array of strings.
// TODO: Remove this conversion once tests are fixed.
SmallVector<Attribute> iteratorTypeNames = llvm::to_vector(
llvm::map_range(iteratorTypes, [&](IteratorType t) -> Attribute {
return StringAttr::get(getContext(), stringifyIteratorType(t));
}));
attrs.emplace_back(getIteratorTypesAttrName(),
ArrayAttr::get(getContext(), iteratorTypeNames));
} else if (traitAttrsSet.count(attr.getName().strref()) > 0)
attrs.push_back(attr);
}
auto dictAttr = DictionaryAttr::get(getContext(), attrs);
p << " " << dictAttr << " " << getLhs() << ", ";
p << getRhs() << ", " << getAcc();
p.printOptionalAttrDict((*this)->getAttrs(), attrNames);
p << " : " << getLhs().getType() << ", " << getRhs().getType() << " into "
<< getResultType();
}
static bool verifyDimMap(VectorType lhsType, VectorType rhsType,
const std::vector<std::pair<int64_t, int64_t>> &map) {
for (auto &dimPair : map) {
if (dimPair.first < 0 || dimPair.first >= lhsType.getRank() ||
dimPair.second < 0 || dimPair.second >= rhsType.getRank() ||
lhsType.getDimSize(dimPair.first) != rhsType.getDimSize(dimPair.second))
return false;
}
return true;
}
static LogicalResult verifyOutputShape(
ContractionOp op, VectorType lhsType, VectorType rhsType, Type accType,
Type resType,
const std::vector<std::pair<int64_t, int64_t>> &contractingDimMap,
const std::vector<std::pair<int64_t, int64_t>> &batchDimMap) {
DenseSet<int64_t> lhsContractingDimSet;
DenseSet<int64_t> rhsContractingDimSet;
for (auto &dimPair : contractingDimMap) {
lhsContractingDimSet.insert(dimPair.first);
rhsContractingDimSet.insert(dimPair.second);
}
DenseSet<int64_t> rhsBatchDimSet(llvm::from_range,
llvm::make_second_range(batchDimMap));
// Add free and batch dimensions from 'lhsType' to 'expectedResultDims'.
SmallVector<int64_t, 4> expectedResultDims;
for (int64_t i = 0, e = lhsType.getRank(); i < e; ++i) {
if (lhsContractingDimSet.count(i) > 0)
continue;
expectedResultDims.push_back(lhsType.getDimSize(i));
}
// Add free dimensions from 'rhsType' to 'expectedResultDims'.
for (int64_t i = 0, e = rhsType.getRank(); i < e; ++i) {
if (rhsContractingDimSet.count(i) > 0 || rhsBatchDimSet.count(i) > 0)
continue;
expectedResultDims.push_back(rhsType.getDimSize(i));
}
// Verify 'expectedResultDims'.
if (expectedResultDims.empty()) {
// No batch or free dimension implies a scalar result.
if (llvm::isa<VectorType>(resType) || llvm::isa<VectorType>(accType))
return op.emitOpError("invalid accumulator/result vector shape");
} else {
// At least one batch or free dimension implies a vector result.
auto resVectorType = llvm::dyn_cast<VectorType>(resType);
auto accVectorType = llvm::dyn_cast<VectorType>(accType);
if (!resVectorType || !accVectorType)
return op.emitOpError("invalid accumulator/result vector shape");
// Infer expected result vector type. Lhs + rhs map and lhs + rhs vector
// types fully define the result vector type. This assumes the affine maps
// are well-formed, which must have been verified already.
MLIRContext *ctx = op.getContext();
AffineMap lhsMap = op.getIndexingMapsArray()[0];
AffineMap rhsMap = op.getIndexingMapsArray()[1];
if (getUnusedDimsBitVector({lhsMap, rhsMap}).any())
return op.emitOpError(
"expected all dimensions to be either a LHS or a RHS dimension");
SmallVector<AffineExpr, 4> extents(lhsMap.getNumInputs());
for (auto pair :
{std::make_pair(lhsType, lhsMap), std::make_pair(rhsType, rhsMap)}) {
VectorType v = pair.first;
auto map = pair.second;
for (unsigned idx = 0, e = v.getRank(); idx < e; ++idx) {
unsigned pos = map.getDimPosition(idx);
if (!extents[pos])
extents[pos] = getAffineConstantExpr(v.getShape()[idx], ctx);
}
}
if (!llvm::all_of(extents, [](AffineExpr e) { return e; }))
return op.emitOpError("expected all dimensions to get an extent as "
"either a LHS or a RHS dimension");
AffineMap resMap = op.getIndexingMapsArray()[2];
auto extentsMap = AffineMap::get(/*dimCount=*/extents.size(),
/*symbolCount=*/0, extents, ctx);
// Compose the resMap with the extentsMap, which is a constant map.
AffineMap expectedMap = simplifyAffineMap(resMap.compose(extentsMap));
assert(llvm::all_of(expectedMap.getResults(),
llvm::IsaPred<AffineConstantExpr>) &&
"expected constant extent along all dimensions.");
// Extract the expected shape and build the type.
auto expectedShape = llvm::to_vector<4>(
llvm::map_range(expectedMap.getResults(), [](AffineExpr e) {
return cast<AffineConstantExpr>(e).getValue();
}));
auto expected =
VectorType::get(expectedShape, resVectorType.getElementType(),
resVectorType.getScalableDims());
if (resVectorType != expected || accVectorType != expected)
return op.emitOpError(
"invalid accumulator/result vector shape, expected: ")
<< expected;
}
return success();
}
LogicalResult ContractionOp::verify() {
VectorType lhsType = getLhsType();
VectorType rhsType = getRhsType();
Type accType = getAccType();
Type resType = getResultType();
if (llvm::isa<IntegerType>(lhsType.getElementType())) {
if (!lhsType.getElementType().isSignlessInteger())
return emitOpError("only supports signless integer types");
}
// Verify that an indexing map was specified for each vector operand.
if (getIndexingMapsArray().size() != 3)
return emitOpError("expected an indexing map for each vector operand");
// Verify that each index map has 'numIterators' inputs, no symbols, and
// that the number of map outputs equals the rank of its associated
// vector operand.
unsigned numIterators = getIteratorTypes().getValue().size();
for (const auto &it : llvm::enumerate(getIndexingMapsArray())) {
auto index = it.index();
auto map = it.value();
if (map.getNumSymbols() != 0)
return emitOpError("expected indexing map ")
<< index << " to have no symbols";
auto vectorType = llvm::dyn_cast<VectorType>(getOperand(index).getType());
unsigned rank = vectorType ? vectorType.getShape().size() : 0;
// Verify that the map has the right number of inputs, outputs, and indices.
// This also correctly accounts for (..) -> () for rank-0 results.
if (map.getNumDims() != numIterators)
return emitOpError("expected indexing map ")
<< index << " to have " << numIterators << " number of inputs";
if (map.getNumResults() != rank)
return emitOpError("expected indexing map ")
<< index << " to have " << rank << " number of outputs";
if (!map.isProjectedPermutation())
return emitOpError("expected indexing map ")
<< index << " to be a projected permutation of its inputs";
}
auto contractingDimMap = getContractingDimMap();
auto batchDimMap = getBatchDimMap();
// Verify at least one contracting dimension pair was specified.
if (contractingDimMap.empty())
return emitOpError("expected at least one contracting dimension pair");
// Verify contracting dimension map was properly constructed.
if (!verifyDimMap(lhsType, rhsType, contractingDimMap))
return emitOpError("invalid contracting dimension map");
// Verify batch dimension map was properly constructed.
if (!verifyDimMap(lhsType, rhsType, batchDimMap))
return emitOpError("invalid batch dimension map");
// Verify 'accType' and 'resType' shape.
if (failed(verifyOutputShape(*this, lhsType, rhsType, accType, resType,
contractingDimMap, batchDimMap)))
return failure();
// Verify supported combining kind.
auto vectorType = llvm::dyn_cast<VectorType>(resType);
auto elementType = vectorType ? vectorType.getElementType() : resType;
if (!isSupportedCombiningKind(getKind(), elementType))
return emitOpError("unsupported contraction type");
return success();
}
// MaskableOpInterface methods.
/// Returns the mask type expected by this operation. Mostly used for
/// verification purposes. It requires the operation to be vectorized."
Type ContractionOp::getExpectedMaskType() {
auto indexingMaps = this->getIndexingMapsArray();
AffineMap lhsIdxMap = indexingMaps[0];
AffineMap rhsIdxMap = indexingMaps[1];
VectorType lhsType = this->getLhsType();
VectorType rhsType = this->getRhsType();
unsigned numVecDims = lhsIdxMap.getNumDims();
SmallVector<int64_t> maskShape(numVecDims, ShapedType::kDynamic);
SmallVector<bool> maskShapeScalableDims(numVecDims, false);
// Using the information in the indexing maps, extract the size of each
// dimension in the vector.contract operation from the two input operands.
for (auto [dimIdx, dimSize] : llvm::enumerate(lhsType.getShape())) {
maskShape[lhsIdxMap.getDimPosition(dimIdx)] = dimSize;
maskShapeScalableDims[lhsIdxMap.getDimPosition(dimIdx)] =
lhsType.getScalableDims()[dimIdx];
}
for (auto [dimIdx, dimSize] : llvm::enumerate(rhsType.getShape())) {
maskShape[rhsIdxMap.getDimPosition(dimIdx)] = dimSize;
maskShapeScalableDims[rhsIdxMap.getDimPosition(dimIdx)] =
rhsType.getScalableDims()[dimIdx];
}
assert(!ShapedType::isDynamicShape(maskShape) &&
"Mask shape couldn't be computed");
return VectorType::get(maskShape,
IntegerType::get(lhsType.getContext(), /*width=*/1),
maskShapeScalableDims);
}
SmallVector<StringRef> ContractionOp::getTraitAttrNames() {
return SmallVector<StringRef>{getIndexingMapsAttrName(),
getIteratorTypesAttrName(), getKindAttrName()};
}
static int64_t getResultIndex(AffineMap map, AffineExpr targetExpr) {
for (int64_t i = 0, e = map.getNumResults(); i < e; ++i)
if (targetExpr == map.getResult(i))
return i;
return -1;
}
static std::vector<std::pair<int64_t, int64_t>>
getDimMap(ArrayRef<AffineMap> indexingMaps, ArrayAttr iteratorTypes,
IteratorType targetIteratorType, MLIRContext *context) {
std::vector<std::pair<int64_t, int64_t>> dimMap;
for (const auto &it : llvm::enumerate(iteratorTypes)) {
auto iteratorType = llvm::cast<IteratorTypeAttr>(it.value()).getValue();
if (iteratorType != targetIteratorType)
continue;
// Search lhs/rhs map results for 'targetExpr'.
auto targetExpr = getAffineDimExpr(it.index(), context);
int64_t lhsDim = getResultIndex(indexingMaps[0], targetExpr);
int64_t rhsDim = getResultIndex(indexingMaps[1], targetExpr);
if (lhsDim >= 0 && rhsDim >= 0)
dimMap.emplace_back(lhsDim, rhsDim);
}
return dimMap;
}
void ContractionOp::getIterationBounds(
SmallVectorImpl<int64_t> &iterationBounds) {
auto lhsShape = getLhsType().getShape();
auto resVectorType = llvm::dyn_cast<VectorType>(getResultType());
SmallVector<AffineMap, 4> indexingMaps(getIndexingMapsArray());
for (const auto &it : llvm::enumerate(getIteratorTypes())) {
// Search lhs/rhs map results for 'targetExpr'.
auto targetExpr = getAffineDimExpr(it.index(), getContext());
auto iteratorType = llvm::cast<IteratorTypeAttr>(it.value()).getValue();
if (iteratorType == IteratorType::reduction) {
// Get reduction dim size from lhs shape (same size in rhsShape).
int64_t lhsDimIndex = getResultIndex(indexingMaps[0], targetExpr);
assert(lhsDimIndex >= 0);
iterationBounds.push_back(lhsShape[lhsDimIndex]);
continue;
}
// Get parallel dimension size from result shape.
int64_t resDimIndex = getResultIndex(indexingMaps[2], targetExpr);
assert(resDimIndex >= 0);
assert(resVectorType != nullptr);
iterationBounds.push_back(resVectorType.getShape()[resDimIndex]);
}
}
void ContractionOp::getIterationIndexMap(
std::vector<DenseMap<int64_t, int64_t>> &iterationIndexMap) {
unsigned numMaps = getIndexingMapsArray().size();
iterationIndexMap.resize(numMaps);
for (const auto &it : llvm::enumerate(getIndexingMapsArray())) {
auto index = it.index();
auto map = it.value();
for (unsigned i = 0, e = map.getNumResults(); i < e; ++i) {
auto dim = cast<AffineDimExpr>(map.getResult(i));
iterationIndexMap[index][dim.getPosition()] = i;
}
}
}
std::vector<std::pair<int64_t, int64_t>> ContractionOp::getContractingDimMap() {
SmallVector<AffineMap, 4> indexingMaps(getIndexingMapsArray());
return getDimMap(indexingMaps, getIteratorTypes(), IteratorType::reduction,
getContext());
}
std::vector<std::pair<int64_t, int64_t>> ContractionOp::getBatchDimMap() {
SmallVector<AffineMap, 4> indexingMaps(getIndexingMapsArray());
return getDimMap(indexingMaps, getIteratorTypes(), IteratorType::parallel,
getContext());
}
std::optional<SmallVector<int64_t, 4>> ContractionOp::getShapeForUnroll() {
SmallVector<int64_t, 4> shape;
getIterationBounds(shape);
return shape;
}
/// Return a fused vector::ContractionOp which represents a patterns such as:
///
/// ```mlir
/// %c0 = vector.constant 0: ...
/// %c = vector.contract %a, %b, %c0: ...
/// %e = add %c, %d: ...
/// ```
///
/// by:
///
/// ```mlir
/// %e = vector.contract %a, %b, %d: ...
/// ```
///
/// Return null if the canonicalization does not apply.
// TODO: This should be a folding of Add into Contract in core but while they
// live in different dialects, it is not possible without unnatural
// dependencies.
template <typename AddOpType>
struct CanonicalizeContractAdd : public OpRewritePattern<AddOpType> {
using OpRewritePattern<AddOpType>::OpRewritePattern;
LogicalResult matchAndRewrite(AddOpType addOp,
PatternRewriter &rewriter) const override {
auto canonicalize = [&](Value maybeContraction,
Value otherOperand) -> vector::ContractionOp {
vector::ContractionOp contractionOp =
dyn_cast_or_null<vector::ContractionOp>(
maybeContraction.getDefiningOp());
if (!contractionOp)
return vector::ContractionOp();
if (auto maybeZero = dyn_cast_or_null<arith::ConstantOp>(
contractionOp.getAcc().getDefiningOp())) {
if (maybeZero.getValue() ==
rewriter.getZeroAttr(contractionOp.getAcc().getType())) {
IRMapping bvm;
bvm.map(contractionOp.getAcc(), otherOperand);
auto newContraction =
cast<vector::ContractionOp>(rewriter.clone(*contractionOp, bvm));
rewriter.replaceOp(addOp, newContraction.getResult());
return newContraction;
}
}
return vector::ContractionOp();
};
Value a = addOp->getOperand(0), b = addOp->getOperand(1);
vector::ContractionOp contract = canonicalize(a, b);
contract = contract ? contract : canonicalize(b, a);
return contract ? success() : failure();
}
};
void ContractionOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<CanonicalizeContractAdd<arith::AddIOp>,
CanonicalizeContractAdd<arith::AddFOp>>(context);
}
//===----------------------------------------------------------------------===//
// ExtractElementOp
//===----------------------------------------------------------------------===//
void ExtractElementOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
SetIntRangeFn setResultRanges) {
setResultRanges(getResult(), argRanges.front());
}
void vector::ExtractElementOp::build(OpBuilder &builder, OperationState &result,
Value source) {
result.addOperands({source});
result.addTypes(llvm::cast<VectorType>(source.getType()).getElementType());
}
LogicalResult vector::ExtractElementOp::verify() {
VectorType vectorType = getSourceVectorType();
if (vectorType.getRank() == 0) {
if (getPosition())
return emitOpError("expected position to be empty with 0-D vector");
return success();
}
if (vectorType.getRank() != 1)
return emitOpError("unexpected >1 vector rank");
if (!getPosition())
return emitOpError("expected position for 1-D vector");
return success();
}
OpFoldResult vector::ExtractElementOp::fold(FoldAdaptor adaptor) {
// Skip the 0-D vector here now.
if (!adaptor.getPosition())
return {};
// Fold extractelement (splat X) -> X.
if (auto splat = getVector().getDefiningOp<vector::SplatOp>())
return splat.getInput();
// Fold extractelement(broadcast(X)) -> X.
if (auto broadcast = getVector().getDefiningOp<vector::BroadcastOp>())
if (!llvm::isa<VectorType>(broadcast.getSource().getType()))
return broadcast.getSource();
auto src = dyn_cast_or_null<DenseElementsAttr>(adaptor.getVector());
auto pos = dyn_cast_or_null<IntegerAttr>(adaptor.getPosition());
if (!pos || !src)
return {};
auto srcElements = src.getValues<Attribute>();
uint64_t posIdx = pos.getInt();
if (posIdx >= srcElements.size())
return {};
return srcElements[posIdx];
}
// Returns `true` if `index` is either within [0, maxIndex) or equal to
// `poisonValue`.
static bool isValidPositiveIndexOrPoison(int64_t index, int64_t poisonValue,
int64_t maxIndex) {
return index == poisonValue || (index >= 0 && index < maxIndex);
}
//===----------------------------------------------------------------------===//
// ExtractOp
//===----------------------------------------------------------------------===//
void ExtractOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
SetIntRangeFn setResultRanges) {
setResultRanges(getResult(), argRanges.front());
}
void vector::ExtractOp::build(OpBuilder &builder, OperationState &result,
Value source) {
auto vectorTy = cast<VectorType>(source.getType());
build(builder, result, source, SmallVector<int64_t>(vectorTy.getRank(), 0));
}
void vector::ExtractOp::build(OpBuilder &builder, OperationState &result,
Value source, int64_t position) {
build(builder, result, source, ArrayRef<int64_t>{position});
}
void vector::ExtractOp::build(OpBuilder &builder, OperationState &result,
Value source, OpFoldResult position) {
build(builder, result, source, ArrayRef<OpFoldResult>{position});
}
void vector::ExtractOp::build(OpBuilder &builder, OperationState &result,
Value source, ArrayRef<int64_t> position) {
build(builder, result, source, /*dynamic_position=*/ArrayRef<Value>(),
builder.getDenseI64ArrayAttr(position));
}
void vector::ExtractOp::build(OpBuilder &builder, OperationState &result,
Value source, ArrayRef<OpFoldResult> position) {
SmallVector<int64_t> staticPos;
SmallVector<Value> dynamicPos;
dispatchIndexOpFoldResults(position, dynamicPos, staticPos);
build(builder, result, source, dynamicPos,
builder.getDenseI64ArrayAttr(staticPos));
}
LogicalResult
ExtractOp::inferReturnTypes(MLIRContext *, std::optional<Location>,
ExtractOp::Adaptor adaptor,
SmallVectorImpl<Type> &inferredReturnTypes) {
auto vectorType = llvm::cast<VectorType>(adaptor.getVector().getType());
if (static_cast<int64_t>(adaptor.getStaticPosition().size()) ==
vectorType.getRank()) {
inferredReturnTypes.push_back(vectorType.getElementType());
} else {
auto n = std::min<size_t>(adaptor.getStaticPosition().size(),
vectorType.getRank());
inferredReturnTypes.push_back(VectorType::get(
vectorType.getShape().drop_front(n), vectorType.getElementType(),
vectorType.getScalableDims().drop_front(n)));
}
return success();
}
bool ExtractOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
// Allow extracting 1-element vectors instead of scalars.
auto isCompatible = [](TypeRange l, TypeRange r) {
auto vectorType = llvm::dyn_cast<VectorType>(l.front());
return vectorType && vectorType.getShape().equals({1}) &&
vectorType.getElementType() == r.front();
};
if (l.size() == 1 && r.size() == 1 &&
(isCompatible(l, r) || isCompatible(r, l)))
return true;
return l == r;
}
LogicalResult vector::ExtractOp::verify() {
// Note: This check must come before getMixedPosition() to prevent a crash.
auto dynamicMarkersCount =
llvm::count_if(getStaticPosition(), ShapedType::isDynamic);
if (static_cast<size_t>(dynamicMarkersCount) != getDynamicPosition().size())
return emitOpError(
"mismatch between dynamic and static positions (kDynamic marker but no "
"corresponding dynamic position) -- this can only happen due to an "
"incorrect fold/rewrite");
auto position = getMixedPosition();
if (position.size() > static_cast<unsigned>(getSourceVectorType().getRank()))
return emitOpError(
"expected position attribute of rank no greater than vector rank");
for (auto [idx, pos] : llvm::enumerate(position)) {
if (auto attr = dyn_cast<Attribute>(pos)) {
int64_t constIdx = cast<IntegerAttr>(attr).getInt();
if (!isValidPositiveIndexOrPoison(
constIdx, kPoisonIndex, getSourceVectorType().getDimSize(idx))) {
return emitOpError("expected position attribute #")
<< (idx + 1)
<< " to be a non-negative integer smaller than the "
"corresponding vector dimension or poison (-1)";
}
}
}
return success();
}
template <typename IntType>
static SmallVector<IntType> extractVector(ArrayAttr arrayAttr) {
return llvm::to_vector<4>(llvm::map_range(
arrayAttr.getAsRange<IntegerAttr>(),
[](IntegerAttr attr) { return static_cast<IntType>(attr.getInt()); }));
}
/// Fold the result of chains of ExtractOp in place by simply concatenating the
/// positions.
static LogicalResult foldExtractOpFromExtractChain(ExtractOp extractOp) {
if (!extractOp.getVector().getDefiningOp<ExtractOp>())
return failure();
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition())
return failure();
SmallVector<int64_t> globalPosition;
ExtractOp currentOp = extractOp;
ArrayRef<int64_t> extrPos = currentOp.getStaticPosition();
globalPosition.append(extrPos.rbegin(), extrPos.rend());
while (ExtractOp nextOp = currentOp.getVector().getDefiningOp<ExtractOp>()) {
currentOp = nextOp;
// TODO: Canonicalization for dynamic position not implemented yet.
if (currentOp.hasDynamicPosition())
return failure();
ArrayRef<int64_t> extrPos = currentOp.getStaticPosition();
globalPosition.append(extrPos.rbegin(), extrPos.rend());
}
extractOp.setOperand(0, currentOp.getVector());
// OpBuilder is only used as a helper to build an I64ArrayAttr.
OpBuilder b(extractOp.getContext());
std::reverse(globalPosition.begin(), globalPosition.end());
extractOp.setStaticPosition(globalPosition);
return success();
}
namespace {
/// Fold an ExtractOp that is fed by a chain of InsertOps and TransposeOps.
/// Walk back a chain of InsertOp/TransposeOp until we hit a match.
/// Compose TransposeOp permutations as we walk back.
/// This helper class keeps an updated extraction position `extractPosition`
/// with extra trailing sentinels.
/// The sentinels encode the internal transposition status of the result vector.
/// As we iterate, extractPosition is permuted and updated.
class ExtractFromInsertTransposeChainState {
public:
ExtractFromInsertTransposeChainState(ExtractOp e);
/// Iterate over producing insert and transpose ops until we find a fold.
Value fold();
private:
/// Return true if the vector at position `a` is contained within the vector
/// at position `b`. Under insert/extract semantics, this is the same as `a`
/// is a prefix of `b`.
template <typename ContainerA, typename ContainerB>
bool isContainedWithin(const ContainerA &a, const ContainerB &b) {
return a.size() <= b.size() &&
std::equal(a.begin(), a.begin() + a.size(), b.begin());
}
/// Return true if the vector at position `a` intersects the vector at
/// position `b`. Under insert/extract semantics, this is the same as equality
/// of all entries of `a` that are >=0 with the corresponding entries of b.
/// Comparison is on the common prefix (i.e. zip).
template <typename ContainerA, typename ContainerB>
bool intersectsWhereNonNegative(const ContainerA &a, const ContainerB &b) {
for (auto [elemA, elemB] : llvm::zip(a, b)) {
if (elemA < 0 || elemB < 0)
continue;
if (elemA != elemB)
return false;
}
return true;
}
/// Folding is only possible in the absence of an internal permutation in the
/// result vector.
bool canFold() {
return (sentinels == ArrayRef(extractPosition).drop_front(extractedRank));
}
// Helper to get the next defining op of interest.
void updateStateForNextIteration(Value v) {
nextInsertOp = v.getDefiningOp<vector::InsertOp>();
nextTransposeOp = v.getDefiningOp<vector::TransposeOp>();
};
// Case 1. If we hit a transpose, just compose the map and iterate.
// Invariant: insert + transpose do not change rank, we can always compose.
LogicalResult handleTransposeOp();
// Case 2: the insert position matches extractPosition exactly, early return.
LogicalResult handleInsertOpWithMatchingPos(Value &res);
/// Case 3: if the insert position is a prefix of extractPosition, extract a
/// portion of the source of the insert.
/// Example:
/// ```
/// %ins = vector.insert %source, %vest[1]: vector<3x4> into vector<2x3x4x5>
/// // extractPosition == [1, 2, 3]
/// %ext = vector.extract %ins[1, 0]: vector<5> from vector<3x4x5>
/// // can fold to vector.extract %source[0, 3]
/// %ext = vector.extract %source[3]: vector<6> from vector<5x6>
/// ```
/// To traverse through %source, we need to set the leading dims to 0 and
/// drop the extra leading dims.
/// This method updates the internal state.
LogicalResult handleInsertOpWithPrefixPos(Value &res);
/// Try to fold in place to extract(source, extractPosition) and return the
/// folded result. Return null if folding is not possible (e.g. due to an
/// internal transposition in the result).
Value tryToFoldExtractOpInPlace(Value source);
ExtractOp extractOp;
int64_t vectorRank;
int64_t extractedRank;
InsertOp nextInsertOp;
TransposeOp nextTransposeOp;
/// Sentinel values that encode the internal permutation status of the result.
/// They are set to (-1, ... , -k) at the beginning and appended to
/// `extractPosition`.
/// In the end, the tail of `extractPosition` must be exactly `sentinels` to
/// ensure that there is no internal transposition.
/// Internal transposition cannot be accounted for with a folding pattern.
// TODO: We could relax the internal transposition with an extra transposition
// operation in a future canonicalizer.
SmallVector<int64_t> sentinels;
SmallVector<int64_t> extractPosition;
};
} // namespace
ExtractFromInsertTransposeChainState::ExtractFromInsertTransposeChainState(
ExtractOp e)
: extractOp(e), vectorRank(extractOp.getSourceVectorType().getRank()),
extractedRank(extractOp.getNumIndices()) {
assert(vectorRank >= extractedRank && "Extracted position overflow");
sentinels.reserve(vectorRank - extractedRank);
for (int64_t i = 0, e = vectorRank - extractedRank; i < e; ++i)
sentinels.push_back(-(i + 1));
extractPosition.assign(extractOp.getStaticPosition().begin(),
extractOp.getStaticPosition().end());
llvm::append_range(extractPosition, sentinels);
}
// Case 1. If we hit a transpose, just compose the map and iterate.
// Invariant: insert + transpose do not change rank, we can always compose.
LogicalResult ExtractFromInsertTransposeChainState::handleTransposeOp() {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition())
return failure();
if (!nextTransposeOp)
return failure();
AffineMap m = inversePermutation(AffineMap::getPermutationMap(
nextTransposeOp.getPermutation(), extractOp.getContext()));
extractPosition = applyPermutationMap(m, ArrayRef(extractPosition));
return success();
}
// Case 2: the insert position matches extractPosition exactly, early return.
LogicalResult
ExtractFromInsertTransposeChainState::handleInsertOpWithMatchingPos(
Value &res) {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition() || nextInsertOp.hasDynamicPosition())
return failure();
ArrayRef<int64_t> insertedPos = nextInsertOp.getStaticPosition();
if (insertedPos != llvm::ArrayRef(extractPosition).take_front(extractedRank))
return failure();
// Case 2.a. early-exit fold.
res = nextInsertOp.getValueToStore();
// Case 2.b. if internal transposition is present, canFold will be false.
return success(canFold());
}
/// Case 3: if inserted position is a prefix of extractPosition,
/// extract a portion of the source of the insertion.
/// This method updates the internal state.
LogicalResult
ExtractFromInsertTransposeChainState::handleInsertOpWithPrefixPos(Value &res) {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition() || nextInsertOp.hasDynamicPosition())
return failure();
ArrayRef<int64_t> insertedPos = nextInsertOp.getStaticPosition();
if (!isContainedWithin(insertedPos, extractPosition))
return failure();
// Set leading dims to zero.
std::fill_n(extractPosition.begin(), insertedPos.size(), 0);
// Drop extra leading dims.
extractPosition.erase(extractPosition.begin(),
extractPosition.begin() + insertedPos.size());
extractedRank = extractPosition.size() - sentinels.size();
// Case 3.a. early-exit fold (break and delegate to post-while path).
res = nextInsertOp.getValueToStore();
// Case 3.b. if internal transposition is present, canFold will be false.
return success();
}
/// Try to fold in place to extract(source, extractPosition) and return the
/// folded result. Return null if folding is not possible (e.g. due to an
/// internal transposition in the result).
Value ExtractFromInsertTransposeChainState::tryToFoldExtractOpInPlace(
Value source) {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition())
return Value();
// If we can't fold (either internal transposition, or nothing to fold), bail.
bool nothingToFold = (source == extractOp.getVector());
if (nothingToFold || !canFold())
return Value();
// Otherwise, fold by updating the op inplace and return its result.
OpBuilder b(extractOp.getContext());
extractOp.setStaticPosition(
ArrayRef(extractPosition).take_front(extractedRank));
extractOp.getVectorMutable().assign(source);
return extractOp.getResult();
}
/// Iterate over producing insert and transpose ops until we find a fold.
Value ExtractFromInsertTransposeChainState::fold() {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition())
return Value();
Value valueToExtractFrom = extractOp.getVector();
updateStateForNextIteration(valueToExtractFrom);
while (nextInsertOp || nextTransposeOp) {
// Case 1. If we hit a transpose, just compose the map and iterate.
// Invariant: insert + transpose do not change rank, we can always compose.
if (succeeded(handleTransposeOp())) {
valueToExtractFrom = nextTransposeOp.getVector();
updateStateForNextIteration(valueToExtractFrom);
continue;
}
Value result;
// Case 2: the position match exactly.
if (succeeded(handleInsertOpWithMatchingPos(result)))
return result;
// Case 3: if the inserted position is a prefix of extractPosition, we can
// just extract a portion of the source of the insert.
if (succeeded(handleInsertOpWithPrefixPos(result)))
return tryToFoldExtractOpInPlace(result);
// Case 4: extractPositionRef intersects insertedPosRef on non-sentinel
// values. This is a more difficult case and we bail.
ArrayRef<int64_t> insertedPos = nextInsertOp.getStaticPosition();
if (isContainedWithin(extractPosition, insertedPos) ||
intersectsWhereNonNegative(extractPosition, insertedPos))
return Value();
// Case 5: No intersection, we forward the extract to insertOp.dest().
valueToExtractFrom = nextInsertOp.getDest();
updateStateForNextIteration(valueToExtractFrom);
}
// If after all this we can fold, go for it.
return tryToFoldExtractOpInPlace(valueToExtractFrom);
}
/// Returns true if the operation has a 0-D vector type operand or result.
static bool hasZeroDimVectors(Operation *op) {
auto hasZeroDimVectorType = [](Type type) -> bool {
auto vecType = dyn_cast<VectorType>(type);
return vecType && vecType.getRank() == 0;
};
return llvm::any_of(op->getOperandTypes(), hasZeroDimVectorType) ||
llvm::any_of(op->getResultTypes(), hasZeroDimVectorType);
}
/// Fold extractOp with scalar result coming from BroadcastOp or SplatOp.
static Value foldExtractFromBroadcast(ExtractOp extractOp) {
Operation *defOp = extractOp.getVector().getDefiningOp();
if (!defOp || !isa<vector::BroadcastOp, SplatOp>(defOp))
return Value();
Value source = defOp->getOperand(0);
if (extractOp.getType() == source.getType())
return source;
auto getRank = [](Type type) {
return llvm::isa<VectorType>(type) ? llvm::cast<VectorType>(type).getRank()
: 0;
};
// If splat or broadcast from a scalar, just return the source scalar.
unsigned broadcastSrcRank = getRank(source.getType());
if (broadcastSrcRank == 0 && source.getType() == extractOp.getType())
return source;
unsigned extractResultRank = getRank(extractOp.getType());
if (extractResultRank > broadcastSrcRank)
return Value();
// Check that the dimension of the result haven't been broadcasted.
auto extractVecType = llvm::dyn_cast<VectorType>(extractOp.getType());
auto broadcastVecType = llvm::dyn_cast<VectorType>(source.getType());
if (extractVecType && broadcastVecType &&
extractVecType.getShape() !=
broadcastVecType.getShape().take_back(extractResultRank))
return Value();
auto broadcastOp = cast<vector::BroadcastOp>(defOp);
int64_t broadcastDstRank = broadcastOp.getResultVectorType().getRank();
// Detect all the positions that come from "dim-1" broadcasting.
// These dimensions correspond to "dim-1" broadcasted dims; set the mathching
// extract position to `0` when extracting from the source operand.
llvm::SetVector<int64_t> broadcastedUnitDims =
broadcastOp.computeBroadcastedUnitDims();
SmallVector<OpFoldResult> extractPos(extractOp.getMixedPosition());
OpBuilder b(extractOp.getContext());
int64_t broadcastRankDiff = broadcastDstRank - broadcastSrcRank;
for (int64_t i = broadcastRankDiff, e = extractPos.size(); i < e; ++i)
if (broadcastedUnitDims.contains(i))
extractPos[i] = b.getIndexAttr(0);
// `rankDiff` leading dimensions correspond to new broadcasted dims, drop the
// matching extract position when extracting from the source operand.
int64_t rankDiff = broadcastSrcRank - extractResultRank;
extractPos.erase(extractPos.begin(),
std::next(extractPos.begin(), extractPos.size() - rankDiff));
// OpBuilder is only used as a helper to build an I64ArrayAttr.
auto [staticPos, dynPos] = decomposeMixedValues(extractPos);
extractOp->setOperands(
llvm::to_vector(llvm::concat<Value>(ValueRange(source), dynPos)));
extractOp.setStaticPosition(staticPos);
return extractOp.getResult();
}
/// Fold extractOp coming from ShuffleOp.
///
/// Example:
///
/// %shuffle = vector.shuffle %a, %b [0, 8, 7, 15]
/// : vector<8xf32>, vector<8xf32>
/// %extract = vector.extract %shuffle[3] : f32 from vector<4xf32>
/// ->
/// %extract = vector.extract %b[7] : f32 from vector<8xf32>
///
static Value foldExtractFromShuffle(ExtractOp extractOp) {
// Dynamic positions are not folded as the resulting code would be more
// complex than the input code.
if (extractOp.hasDynamicPosition())
return Value();
auto shuffleOp = extractOp.getVector().getDefiningOp<ShuffleOp>();
if (!shuffleOp)
return Value();
// TODO: 0-D or multi-dimensional vectors not supported yet.
if (shuffleOp.getResultVectorType().getRank() != 1)
return Value();
int64_t inputVecSize = shuffleOp.getV1().getType().getShape()[0];
auto shuffleMask = shuffleOp.getMask();
int64_t extractIdx = extractOp.getStaticPosition()[0];
int64_t shuffleIdx = shuffleMask[extractIdx];
// Find the shuffled vector to extract from based on the shuffle index.
if (shuffleIdx < inputVecSize) {
extractOp.setOperand(0, shuffleOp.getV1());
extractOp.setStaticPosition({shuffleIdx});
} else {
extractOp.setOperand(0, shuffleOp.getV2());
extractOp.setStaticPosition({shuffleIdx - inputVecSize});
}
return extractOp.getResult();
}
// Fold extractOp with source coming from ShapeCast op.
static Value foldExtractFromShapeCast(ExtractOp extractOp) {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition())
return Value();
auto shapeCastOp = extractOp.getVector().getDefiningOp<vector::ShapeCastOp>();
if (!shapeCastOp)
return Value();
// Get the nth dimension size starting from lowest dimension.
auto getDimReverse = [](VectorType type, int64_t n) {
return type.getShape().take_back(n + 1).front();
};
int64_t destinationRank =
llvm::isa<VectorType>(extractOp.getType())
? llvm::cast<VectorType>(extractOp.getType()).getRank()
: 0;
if (destinationRank > shapeCastOp.getSourceVectorType().getRank())
return Value();
if (destinationRank > 0) {
auto destinationType =
llvm::cast<VectorType>(extractOp.getResult().getType());
for (int64_t i = 0; i < destinationRank; i++) {
// The lowest dimension of the destination must match the lowest
// dimension of the shapecast op source.
// TODO: This case could be support in a canonicalization pattern.
if (getDimReverse(shapeCastOp.getSourceVectorType(), i) !=
getDimReverse(destinationType, i))
return Value();
}
}
// Extract the strides associated with the extract op vector source. Then use
// this to calculate a linearized position for the extract.
SmallVector<int64_t> extractedPos(extractOp.getStaticPosition());
std::reverse(extractedPos.begin(), extractedPos.end());
SmallVector<int64_t, 4> strides;
int64_t stride = 1;
for (int64_t i = 0, e = extractedPos.size(); i < e; i++) {
strides.push_back(stride);
stride *=
getDimReverse(extractOp.getSourceVectorType(), i + destinationRank);
}
int64_t position = linearize(extractedPos, strides);
// Then extract the strides associated to the shapeCast op vector source and
// delinearize the position using those strides.
SmallVector<int64_t, 4> newStrides;
int64_t numDimension =
shapeCastOp.getSourceVectorType().getRank() - destinationRank;
stride = 1;
for (int64_t i = 0; i < numDimension; i++) {
newStrides.push_back(stride);
stride *=
getDimReverse(shapeCastOp.getSourceVectorType(), i + destinationRank);
}
std::reverse(newStrides.begin(), newStrides.end());
SmallVector<int64_t, 4> newPosition = delinearize(position, newStrides);
// OpBuilder is only used as a helper to build an I64ArrayAttr.
OpBuilder b(extractOp.getContext());
extractOp.setStaticPosition(newPosition);
extractOp.setOperand(0, shapeCastOp.getSource());
return extractOp.getResult();
}
/// Fold an ExtractOp from ExtractStridedSliceOp.
static Value foldExtractFromExtractStrided(ExtractOp extractOp) {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition())
return Value();
auto extractStridedSliceOp =
extractOp.getVector().getDefiningOp<vector::ExtractStridedSliceOp>();
if (!extractStridedSliceOp)
return Value();
// 0-D vectors not supported.
assert(!hasZeroDimVectors(extractOp) && "0-D vectors not supported");
if (hasZeroDimVectors(extractStridedSliceOp))
return Value();
// Return if 'extractStridedSliceOp' has non-unit strides.
if (extractStridedSliceOp.hasNonUnitStrides())
return Value();
// Trim offsets for dimensions fully extracted.
auto sliceOffsets =
extractVector<int64_t>(extractStridedSliceOp.getOffsets());
while (!sliceOffsets.empty()) {
size_t lastOffset = sliceOffsets.size() - 1;
if (sliceOffsets.back() != 0 ||
extractStridedSliceOp.getType().getDimSize(lastOffset) !=
extractStridedSliceOp.getSourceVectorType().getDimSize(lastOffset))
break;
sliceOffsets.pop_back();
}
unsigned destinationRank = 0;
if (auto vecType = llvm::dyn_cast<VectorType>(extractOp.getType()))
destinationRank = vecType.getRank();
// The dimensions of the result need to be untouched by the
// extractStridedSlice op.
if (destinationRank > extractStridedSliceOp.getSourceVectorType().getRank() -
sliceOffsets.size())
return Value();
SmallVector<int64_t> extractedPos(extractOp.getStaticPosition());
assert(extractedPos.size() >= sliceOffsets.size());
for (size_t i = 0, e = sliceOffsets.size(); i < e; i++)
extractedPos[i] = extractedPos[i] + sliceOffsets[i];
extractOp.getVectorMutable().assign(extractStridedSliceOp.getVector());
// OpBuilder is only used as a helper to build an I64ArrayAttr.
OpBuilder b(extractOp.getContext());
extractOp.setStaticPosition(extractedPos);
return extractOp.getResult();
}
/// Fold extract_op fed from a chain of insertStridedSlice ops.
static Value foldExtractStridedOpFromInsertChain(ExtractOp extractOp) {
// TODO: Canonicalization for dynamic position not implemented yet.
if (extractOp.hasDynamicPosition())
return Value();
int64_t destinationRank =
llvm::isa<VectorType>(extractOp.getType())
? llvm::cast<VectorType>(extractOp.getType()).getRank()
: 0;
auto insertOp = extractOp.getVector().getDefiningOp<InsertStridedSliceOp>();
if (!insertOp)
return Value();
// 0-D vectors not supported.
assert(!hasZeroDimVectors(extractOp) && "0-D vectors not supported");
if (hasZeroDimVectors(insertOp))
return Value();
while (insertOp) {
int64_t insertRankDiff = insertOp.getDestVectorType().getRank() -
insertOp.getSourceVectorType().getRank();
if (destinationRank > insertOp.getSourceVectorType().getRank())
return Value();
auto insertOffsets = extractVector<int64_t>(insertOp.getOffsets());
ArrayRef<int64_t> extractOffsets = extractOp.getStaticPosition();
if (llvm::any_of(insertOp.getStrides(), [](Attribute attr) {
return llvm::cast<IntegerAttr>(attr).getInt() != 1;
}))
return Value();
bool disjoint = false;
SmallVector<int64_t, 4> offsetDiffs;
for (unsigned dim = 0, e = extractOffsets.size(); dim < e; ++dim) {
int64_t start = insertOffsets[dim];
int64_t size =
(dim < insertRankDiff)
? 1
: insertOp.getSourceVectorType().getDimSize(dim - insertRankDiff);
int64_t end = start + size;
int64_t offset = extractOffsets[dim];
// Check if the start of the extract offset is in the interval inserted.
if (start <= offset && offset < end) {
if (dim >= insertRankDiff)
offsetDiffs.push_back(offset - start);
continue;
}
disjoint = true;
break;
}
// The extract element chunk overlap with the vector inserted.
if (!disjoint) {
// If any of the inner dimensions are only partially inserted we have a
// partial overlap.
int64_t srcRankDiff =
insertOp.getSourceVectorType().getRank() - destinationRank;
for (int64_t i = 0; i < destinationRank; i++) {
if (insertOp.getSourceVectorType().getDimSize(i + srcRankDiff) !=
insertOp.getDestVectorType().getDimSize(i + srcRankDiff +
insertRankDiff))
return Value();
}
extractOp.getVectorMutable().assign(insertOp.getValueToStore());
// OpBuilder is only used as a helper to build an I64ArrayAttr.
OpBuilder b(extractOp.getContext());
extractOp.setStaticPosition(offsetDiffs);
return extractOp.getResult();
}
// If the chunk extracted is disjoint from the chunk inserted, keep
// looking in the insert chain.
insertOp = insertOp.getDest().getDefiningOp<InsertStridedSliceOp>();
}
return Value();
}
/// Try to fold the extraction of a scalar from a vector defined by
/// vector.from_elements. E.g.:
///
/// %0 = vector.from_elements %a, %b : vector<2xf32>
/// %1 = vector.extract %0[0] : f32 from vector<2xf32>
/// ==> fold to %a
static Value foldScalarExtractFromFromElements(ExtractOp extractOp) {
// Dynamic extractions cannot be folded.
if (extractOp.hasDynamicPosition())
return {};
// Look for extract(from_elements).
auto fromElementsOp = extractOp.getVector().getDefiningOp<FromElementsOp>();
if (!fromElementsOp)
return {};
// Scalable vectors are not supported.
auto vecType = llvm::cast<VectorType>(fromElementsOp.getType());
if (vecType.isScalable())
return {};
// Only extractions of scalars are supported.
int64_t rank = vecType.getRank();
ArrayRef<int64_t> indices = extractOp.getStaticPosition();
if (extractOp.getType() != vecType.getElementType())
return {};
assert(static_cast<int64_t>(indices.size()) == rank &&
"unexpected number of indices");
// Compute flattened/linearized index and fold to operand.
int flatIndex = 0;
int stride = 1;
for (int i = rank - 1; i >= 0; --i) {
flatIndex += indices[i] * stride;
stride *= vecType.getDimSize(i);
}
return fromElementsOp.getElements()[flatIndex];
}
/// If the dynamic indices of `extractOp` or `insertOp` are in fact constants,
/// then fold it.
template <typename OpType, typename AdaptorType>
static Value extractInsertFoldConstantOp(OpType op, AdaptorType adaptor,
SmallVectorImpl<Value> &operands) {
std::vector<int64_t> staticPosition = op.getStaticPosition().vec();
OperandRange dynamicPosition = op.getDynamicPosition();
ArrayRef<Attribute> dynamicPositionAttr = adaptor.getDynamicPosition();
ArrayRef<int64_t> vectorShape;
if constexpr (std::is_same_v<OpType, ExtractOp>)
vectorShape = op.getSourceVectorType().getShape();
else
vectorShape = op.getDestVectorType().getShape();
// If the dynamic operands is empty, it is returned directly.
if (!dynamicPosition.size())
return {};
// `index` is used to iterate over the `dynamicPosition`.
unsigned index = 0;
// `opChange` is a flag. If it is true, it means to update `op` in place.
bool opChange = false;
for (unsigned i = 0, e = staticPosition.size(); i < e; ++i) {
if (!ShapedType::isDynamic(staticPosition[i]))
continue;
Attribute positionAttr = dynamicPositionAttr[index];
Value position = dynamicPosition[index++];
if (auto attr = mlir::dyn_cast_if_present<IntegerAttr>(positionAttr)) {
int64_t value = attr.getInt();
// Do not fold if the value is out of bounds (-1 signifies a poison
// value rather than OOB index).
if (value >= -1 && value < vectorShape[i]) {
staticPosition[i] = attr.getInt();
opChange = true;
continue;
}
}
operands.push_back(position);
}
if (opChange) {
op.setStaticPosition(staticPosition);
op.getOperation()->setOperands(operands);
return op.getResult();
}
return {};
}
/// Fold an insert or extract operation into an poison value when a poison index
/// is found at any dimension of the static position.
static Attribute foldPoisonIndexInsertExtractOp(MLIRContext *context,
ArrayRef<int64_t> staticPos,
int64_t poisonVal) {
if (!is_contained(staticPos, poisonVal))
return {};
return ub::PoisonAttr::get(context);
}
/// Fold a vector extract from is a poison source.
static Attribute foldPoisonSrcExtractOp(Attribute srcAttr) {
if (isa_and_nonnull<ub::PoisonAttr>(srcAttr))
return srcAttr;
return {};
}
/// Fold a vector extract extracting from a DenseElementsAttr.
static Attribute foldDenseElementsAttrSrcExtractOp(ExtractOp extractOp,
Attribute srcAttr) {
auto denseAttr = dyn_cast_if_present<DenseElementsAttr>(srcAttr);
if (!denseAttr) {
return {};
}
if (denseAttr.isSplat()) {
Attribute newAttr = denseAttr.getSplatValue<Attribute>();
if (auto vecDstType = dyn_cast<VectorType>(extractOp.getType()))
newAttr = DenseElementsAttr::get(vecDstType, newAttr);
return newAttr;
}
auto vecTy = cast<VectorType>(extractOp.getSourceVectorType());
if (vecTy.isScalable())
return {};
if (extractOp.hasDynamicPosition()) {
return {};
}
// Materializing subsets of a large constant array can generally lead to
// explosion in IR size because of different combination of subsets that
// can exist. However, vector.extract is a restricted form of subset
// extract where you can only extract non-overlapping (or the same) subset for
// a given rank of the subset. Because of this property, the IR size can only
// increase at most by `rank * size(array)` from a single constant array being
// extracted by multiple extracts.
// Calculate the linearized position of the continuous chunk of elements to
// extract.
SmallVector<int64_t> completePositions(vecTy.getRank(), 0);
copy(extractOp.getStaticPosition(), completePositions.begin());
int64_t startPos =
linearize(completePositions, computeStrides(vecTy.getShape()));
auto denseValuesBegin = denseAttr.value_begin<TypedAttr>() + startPos;
TypedAttr newAttr;
if (auto resVecTy = dyn_cast<VectorType>(extractOp.getType())) {
SmallVector<Attribute> elementValues(
denseValuesBegin, denseValuesBegin + resVecTy.getNumElements());
newAttr = DenseElementsAttr::get(resVecTy, elementValues);
} else {
newAttr = *denseValuesBegin;
}
return newAttr;
}
OpFoldResult ExtractOp::fold(FoldAdaptor adaptor) {
// Fold "vector.extract %v[] : vector<2x2xf32> from vector<2x2xf32>" to %v.
// Note: Do not fold "vector.extract %v[] : f32 from vector<f32>" (type
// mismatch).
if (getNumIndices() == 0 && getVector().getType() == getResult().getType())
return getVector();
if (auto res = foldPoisonIndexInsertExtractOp(
getContext(), adaptor.getStaticPosition(), kPoisonIndex))
return res;
if (auto res = foldPoisonSrcExtractOp(adaptor.getVector()))
return res;
if (auto res = foldDenseElementsAttrSrcExtractOp(*this, adaptor.getVector()))
return res;
if (succeeded(foldExtractOpFromExtractChain(*this)))
return getResult();
if (auto res = ExtractFromInsertTransposeChainState(*this).fold())
return res;
if (auto res = foldExtractFromBroadcast(*this))
return res;
if (auto res = foldExtractFromShuffle(*this))
return res;
if (auto res = foldExtractFromShapeCast(*this))
return res;
if (auto val = foldExtractFromExtractStrided(*this))
return val;
if (auto val = foldExtractStridedOpFromInsertChain(*this))
return val;
if (auto val = foldScalarExtractFromFromElements(*this))
return val;
SmallVector<Value> operands = {getVector()};
if (auto val = extractInsertFoldConstantOp(*this, adaptor, operands))
return val;
return OpFoldResult();
}
namespace {
// Pattern to rewrite a ExtractOp(Broadcast) -> Broadcast.
class ExtractOpFromBroadcast final : public OpRewritePattern<ExtractOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ExtractOp extractOp,
PatternRewriter &rewriter) const override {
Operation *defOp = extractOp.getVector().getDefiningOp();
if (!defOp || !isa<vector::BroadcastOp, SplatOp>(defOp))
return failure();
Value source = defOp->getOperand(0);
if (extractOp.getType() == source.getType())
return failure();
auto getRank = [](Type type) {
return llvm::isa<VectorType>(type)
? llvm::cast<VectorType>(type).getRank()
: 0;
};
unsigned broadcastSrcRank = getRank(source.getType());
unsigned extractResultRank = getRank(extractOp.getType());
// We only consider the case where the rank of the source is less than or
// equal to the rank of the extract dst. The other cases are handled in the
// folding patterns.
if (extractResultRank < broadcastSrcRank)
return failure();
// For scalar result, the input can only be a rank-0 vector, which will
// be handled by the folder.
if (extractResultRank == 0)
return failure();
rewriter.replaceOpWithNewOp<vector::BroadcastOp>(
extractOp, extractOp.getType(), source);
return success();
}
};
// Pattern to rewrite a ExtractOp(CreateMask) -> CreateMask.
class ExtractOpFromCreateMask final : public OpRewritePattern<ExtractOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ExtractOp extractOp,
PatternRewriter &rewriter) const override {
auto createMaskOp =
extractOp.getVector().getDefiningOp<vector::CreateMaskOp>();
if (!createMaskOp)
return failure();
VectorType extractedMaskType =
llvm::dyn_cast<VectorType>(extractOp.getResult().getType());
if (!extractedMaskType)
return failure();
auto maskOperands = createMaskOp.getOperands();
ArrayRef<int64_t> extractOpPos = extractOp.getStaticPosition();
VectorType maskType = createMaskOp.getVectorType();
bool containsUnknownDims = false;
bool allFalse = getMaskFormat(createMaskOp) == MaskFormat::AllFalse;
for (size_t dimIdx = 0; !allFalse && dimIdx < extractOpPos.size();
dimIdx++) {
int64_t pos = extractOpPos[dimIdx];
Value operand = maskOperands[dimIdx];
auto constantOp = operand.getDefiningOp<arith::ConstantOp>();
if (!constantOp) {
// Bounds of this dim unknown.
containsUnknownDims = true;
continue;
}
int64_t createMaskBound =
llvm::cast<IntegerAttr>(constantOp.getValue()).getInt();
if (pos != ShapedType::kDynamic) {
// If any position is outside the range from the `create_mask`, then the
// extracted mask will be all-false.
allFalse |= pos >= createMaskBound;
} else if (createMaskBound < maskType.getDimSize(dimIdx)) {
// This dim is not all-true and since this is a dynamic index we don't
// know if the extraction is within the true or false region.
// Note: Zero dims have already handled via getMaskFormat().
containsUnknownDims = true;
}
}
if (allFalse) {
rewriter.replaceOpWithNewOp<arith::ConstantOp>(
extractOp, DenseElementsAttr::get(extractedMaskType, false));
} else if (!containsUnknownDims) {
rewriter.replaceOpWithNewOp<vector::CreateMaskOp>(
extractOp, extractedMaskType,
maskOperands.drop_front(extractOpPos.size()));
} else {
return failure();
}
return success();
}
};
// Folds extract(shape_cast(..)) into shape_cast when the total element count
// does not change.
LogicalResult foldExtractFromShapeCastToShapeCast(ExtractOp extractOp,
PatternRewriter &rewriter) {
auto castOp = extractOp.getVector().getDefiningOp<ShapeCastOp>();
if (!castOp)
return failure();
VectorType sourceType = castOp.getSourceVectorType();
auto targetType = dyn_cast<VectorType>(extractOp.getResult().getType());
if (!targetType)
return failure();
if (sourceType.getNumElements() != targetType.getNumElements())
return failure();
rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(extractOp, targetType,
castOp.getSource());
return success();
}
/// Try to canonicalize the extraction of a subvector from a vector defined by
/// vector.from_elements. E.g.:
///
/// %0 = vector.from_elements %a, %b, %a, %a : vector<2x2xf32>
/// %1 = vector.extract %0[0] : vector<2xf32> from vector<2x2xf32>
/// ==> canonicalize to vector.from_elements %a, %b : vector<2xf32>
LogicalResult foldExtractFromFromElements(ExtractOp extractOp,
PatternRewriter &rewriter) {
// Dynamic positions are not supported.
if (extractOp.hasDynamicPosition())
return failure();
// Scalar extracts are handled by the folder.
auto resultType = dyn_cast<VectorType>(extractOp.getType());
if (!resultType)
return failure();
// Look for extracts from a from_elements op.
auto fromElementsOp = extractOp.getVector().getDefiningOp<FromElementsOp>();
if (!fromElementsOp)
return failure();
VectorType inputType = fromElementsOp.getType();
// Scalable vectors are not supported.
if (resultType.isScalable() || inputType.isScalable())
return failure();
// Compute the position of first extracted element and flatten/linearize the
// position.
SmallVector<int64_t> firstElementPos =
llvm::to_vector(extractOp.getStaticPosition());
firstElementPos.append(/*NumInputs=*/resultType.getRank(), /*Elt=*/0);
int flatIndex = 0;
int stride = 1;
for (int64_t i = inputType.getRank() - 1; i >= 0; --i) {
flatIndex += firstElementPos[i] * stride;
stride *= inputType.getDimSize(i);
}
// Replace the op with a smaller from_elements op.
rewriter.replaceOpWithNewOp<FromElementsOp>(
extractOp, resultType,
fromElementsOp.getElements().slice(flatIndex,
resultType.getNumElements()));
return success();
}
} // namespace
void ExtractOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<ExtractOpFromBroadcast, ExtractOpFromCreateMask>(context);
results.add(foldExtractFromShapeCastToShapeCast);
results.add(foldExtractFromFromElements);
}
static void populateFromInt64AttrArray(ArrayAttr arrayAttr,
SmallVectorImpl<int64_t> &results) {
for (auto attr : arrayAttr)
results.push_back(llvm::cast<IntegerAttr>(attr).getInt());
}
//===----------------------------------------------------------------------===//
// FmaOp
//===----------------------------------------------------------------------===//
std::optional<SmallVector<int64_t, 4>> FMAOp::getShapeForUnroll() {
return llvm::to_vector<4>(getVectorType().getShape());
}
//===----------------------------------------------------------------------===//
// FromElementsOp
//===----------------------------------------------------------------------===//
/// Rewrite a vector.from_elements into a vector.splat if all elements are the
/// same SSA value. E.g.:
///
/// %0 = vector.from_elements %a, %a, %a : vector<3xf32>
/// ==> rewrite to vector.splat %a : vector<3xf32>
static LogicalResult rewriteFromElementsAsSplat(FromElementsOp fromElementsOp,
PatternRewriter &rewriter) {
if (!llvm::all_equal(fromElementsOp.getElements()))
return failure();
rewriter.replaceOpWithNewOp<SplatOp>(fromElementsOp, fromElementsOp.getType(),
fromElementsOp.getElements().front());
return success();
}
void FromElementsOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add(rewriteFromElementsAsSplat);
}
//===----------------------------------------------------------------------===//
// BroadcastOp
//===----------------------------------------------------------------------===//
void BroadcastOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
SetIntRangeFn setResultRanges) {
setResultRanges(getResult(), argRanges.front());
}
/// Return the dimensions of the result vector that were formerly ones in the
/// source tensor and thus correspond to "dim-1" broadcasting.
static llvm::SetVector<int64_t>
computeBroadcastedUnitDims(ArrayRef<int64_t> srcShape,
ArrayRef<int64_t> dstShape) {
int64_t rankDiff = dstShape.size() - srcShape.size();
int64_t dstDim = rankDiff;
llvm::SetVector<int64_t> res;
for (auto [s1, s2] :
llvm::zip_equal(srcShape, dstShape.drop_front(rankDiff))) {
if (s1 != s2) {
assert(s1 == 1 && "expected \"dim-1\" broadcasting");
res.insert(dstDim);
}
++dstDim;
}
return res;
}
llvm::SetVector<int64_t> BroadcastOp::computeBroadcastedUnitDims() {
// Scalar broadcast is without any unit dim broadcast.
auto srcVectorType = llvm::dyn_cast<VectorType>(getSourceType());
if (!srcVectorType)
return {};
return ::computeBroadcastedUnitDims(srcVectorType.getShape(),
getResultVectorType().getShape());
}
/// Broadcast `value` to a vector of `dstShape`, knowing that exactly the
/// `broadcastedDims` dimensions in the dstShape are broadcasted.
/// This requires (and asserts) that the broadcast is free of "dim-1"
/// broadcasting.
/// Since vector.broadcast only allows expanding leading dimensions, an extra
/// vector.transpose may be inserted to make the broadcast possible.
/// `value`, `dstShape` and `broadcastedDims` must be properly specified or
/// the helper will assert. This means:
/// 1. `dstShape` must not be empty.
/// 2. `broadcastedDims` must be confined to [0 .. rank(value.getVectorType)]
/// 2. `dstShape` trimmed of the dimensions specified in `broadcastedDims`
// must match the `value` shape.
Value BroadcastOp::createOrFoldBroadcastOp(
OpBuilder &b, Value value, ArrayRef<int64_t> dstShape,
const llvm::SetVector<int64_t> &broadcastedDims) {
assert(!dstShape.empty() && "unexpected empty dst shape");
// Well-formedness check.
SmallVector<int64_t> checkShape;
for (int i = 0, e = dstShape.size(); i < e; ++i) {
if (broadcastedDims.contains(i))
continue;
checkShape.push_back(dstShape[i]);
}
assert(broadcastedDims.size() == dstShape.size() - checkShape.size() &&
"ill-formed broadcastedDims contains values not confined to "
"destVectorShape");
Location loc = value.getLoc();
Type elementType = getElementTypeOrSelf(value.getType());
VectorType srcVectorType = llvm::dyn_cast<VectorType>(value.getType());
VectorType dstVectorType = VectorType::get(dstShape, elementType);
// Step 2. If scalar -> dstShape broadcast, just do it.
if (!srcVectorType) {
assert(checkShape.empty() &&
"ill-formed createOrFoldBroadcastOp arguments");
return b.createOrFold<vector::BroadcastOp>(loc, dstVectorType, value);
}
assert(srcVectorType.getShape().equals(checkShape) &&
"ill-formed createOrFoldBroadcastOp arguments");
// Step 3. Since vector.broadcast only allows creating leading dims,
// vector -> dstShape broadcast may require a transpose.
// Traverse the dims in order and construct:
// 1. The leading entries of the broadcastShape that is guaranteed to be
// achievable by a simple broadcast.
// 2. The induced permutation for the subsequent vector.transpose that will
// bring us from `broadcastShape` back to he desired `dstShape`.
// If the induced permutation is not the identity, create a vector.transpose.
SmallVector<int64_t> broadcastShape, permutation(dstShape.size(), -1);
broadcastShape.reserve(dstShape.size());
// Consider the example:
// srcShape = 2x4
// dstShape = 1x2x3x4x5
// broadcastedDims = [0, 2, 4]
//
// We want to build:
// broadcastShape = 1x3x5x2x4
// permutation = [0, 2, 4, 1, 3]
// ---V--- -----V-----
// leading broadcast part src shape part
//
// Note that the trailing dims of broadcastShape are exactly the srcShape
// by construction.
// nextSrcShapeDim is used to keep track of where in the permutation the
// "src shape part" occurs.
int64_t nextSrcShapeDim = broadcastedDims.size();
for (int64_t i = 0, e = dstShape.size(); i < e; ++i) {
if (broadcastedDims.contains(i)) {
// 3.a. For each dim in the dst shape, if it is a broadcasted dim,
// bring it to the head of the broadcastShape.
// It will need to be permuted back from `broadcastShape.size() - 1` into
// position `i`.
broadcastShape.push_back(dstShape[i]);
permutation[i] = broadcastShape.size() - 1;
} else {
// 3.b. Otherwise, the dim is not broadcasted, it comes from the src
// shape and needs to be permuted into position `i`.
// Don't touch `broadcastShape` here, the whole srcShape will be
// appended after.
permutation[i] = nextSrcShapeDim++;
}
}
// 3.c. Append the srcShape.
llvm::append_range(broadcastShape, srcVectorType.getShape());
// Ensure there are no "dim-1" broadcasts.
assert(::computeBroadcastedUnitDims(srcVectorType.getShape(), broadcastShape)
.empty() &&
"unexpected \"dim-1\" broadcast");
VectorType broadcastType = VectorType::get(broadcastShape, elementType);
assert(vector::isBroadcastableTo(value.getType(), broadcastType) ==
vector::BroadcastableToResult::Success &&
"must be broadcastable");
Value res = b.createOrFold<vector::BroadcastOp>(loc, broadcastType, value);
// Step 4. If we find any dimension that indeed needs to be permuted,
// immediately return a new vector.transpose.
for (int64_t i = 0, e = permutation.size(); i < e; ++i)
if (permutation[i] != i)
return b.createOrFold<vector::TransposeOp>(loc, res, permutation);
// Otherwise return res.
return res;
}
BroadcastableToResult mlir::vector::isBroadcastableTo(
Type srcType, VectorType dstVectorType,
std::pair<VectorDim, VectorDim> *mismatchingDims) {
// Broadcast scalar to vector of the same element type.
if (srcType.isIntOrIndexOrFloat() && dstVectorType &&
getElementTypeOrSelf(srcType) == getElementTypeOrSelf(dstVectorType))
return BroadcastableToResult::Success;
// From now on, only vectors broadcast.
VectorType srcVectorType = llvm::dyn_cast<VectorType>(srcType);
if (!srcVectorType)
return BroadcastableToResult::SourceTypeNotAVector;
int64_t srcRank = srcVectorType.getRank();
int64_t dstRank = dstVectorType.getRank();
if (srcRank > dstRank)
return BroadcastableToResult::SourceRankHigher;
// Source has an exact match or singleton value for all trailing dimensions
// (all leading dimensions are simply duplicated).
int64_t lead = dstRank - srcRank;
for (int64_t dimIdx = 0; dimIdx < srcRank; ++dimIdx) {
// Have mismatching dims (in the sense of vector.broadcast semantics) been
// encountered?
bool foundMismatchingDims = false;
// Check fixed-width dims.
int64_t srcDim = srcVectorType.getDimSize(dimIdx);
int64_t dstDim = dstVectorType.getDimSize(lead + dimIdx);
if (srcDim != 1 && srcDim != dstDim)
foundMismatchingDims = true;
// Check scalable flags.
bool srcDimScalableFlag = srcVectorType.getScalableDims()[dimIdx];
bool dstDimScalableFlag = dstVectorType.getScalableDims()[lead + dimIdx];
if ((srcDim == 1 && srcDimScalableFlag && dstDim != 1) ||
// 1 -> [N] is fine, everything else should be rejected when mixing
// fixed-width and scalable dims
(srcDimScalableFlag != dstDimScalableFlag &&
(srcDim != 1 || srcDimScalableFlag)))
foundMismatchingDims = true;
if (foundMismatchingDims) {
if (mismatchingDims != nullptr) {
mismatchingDims->first.dim = srcDim;
mismatchingDims->first.isScalable = srcDimScalableFlag;
mismatchingDims->second.dim = dstDim;
mismatchingDims->second.isScalable = dstDimScalableFlag;
}
return BroadcastableToResult::DimensionMismatch;
}
}
return BroadcastableToResult::Success;
}
LogicalResult BroadcastOp::verify() {
std::pair<VectorDim, VectorDim> mismatchingDims;
BroadcastableToResult res = isBroadcastableTo(
getSourceType(), getResultVectorType(), &mismatchingDims);
if (res == BroadcastableToResult::Success)
return success();
if (res == BroadcastableToResult::SourceRankHigher)
return emitOpError("source rank higher than destination rank");
if (res == BroadcastableToResult::DimensionMismatch) {
return emitOpError("dimension mismatch (")
<< (mismatchingDims.first.isScalable ? "[" : "")
<< mismatchingDims.first.dim
<< (mismatchingDims.first.isScalable ? "]" : "") << " vs. "
<< (mismatchingDims.second.isScalable ? "[" : "")
<< mismatchingDims.second.dim
<< (mismatchingDims.second.isScalable ? "]" : "") << ")";
}
if (res == BroadcastableToResult::SourceTypeNotAVector)
return emitOpError("source type is not a vector");
llvm_unreachable("unexpected vector.broadcast op error");
}
OpFoldResult BroadcastOp::fold(FoldAdaptor adaptor) {
if (getSourceType() == getResultVectorType())
return getSource();
if (!adaptor.getSource())
return {};
auto vectorType = getResultVectorType();
if (auto attr = llvm::dyn_cast<IntegerAttr>(adaptor.getSource())) {
if (vectorType.getElementType() != attr.getType())
return {};
return DenseElementsAttr::get(vectorType, attr);
}
if (auto attr = llvm::dyn_cast<FloatAttr>(adaptor.getSource())) {
if (vectorType.getElementType() != attr.getType())
return {};
return DenseElementsAttr::get(vectorType, attr);
}
if (auto attr = llvm::dyn_cast<SplatElementsAttr>(adaptor.getSource()))
return DenseElementsAttr::get(vectorType, attr.getSplatValue<Attribute>());
if (llvm::dyn_cast<ub::PoisonAttr>(adaptor.getSource()))
return ub::PoisonAttr::get(getContext());
return {};
}
namespace {
// Fold broadcast1(broadcast2(x)) into broadcast1(x).
struct BroadcastFolder : public OpRewritePattern<BroadcastOp> {
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(BroadcastOp broadcastOp,
PatternRewriter &rewriter) const override {
auto srcBroadcast = broadcastOp.getSource().getDefiningOp<BroadcastOp>();
if (!srcBroadcast)
return failure();
rewriter.replaceOpWithNewOp<BroadcastOp>(broadcastOp,
broadcastOp.getResultVectorType(),
srcBroadcast.getSource());
return success();
}
};
} // namespace
void BroadcastOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
// BroadcastToShapeCast is not a default canonicalization, it is opt-in by
// calling `populateCastAwayVectorLeadingOneDimPatterns`
results.add<BroadcastFolder>(context);
}
//===----------------------------------------------------------------------===//
// ShuffleOp
//===----------------------------------------------------------------------===//
LogicalResult ShuffleOp::verify() {
VectorType resultType = getResultVectorType();
VectorType v1Type = getV1VectorType();
VectorType v2Type = getV2VectorType();
// Verify ranks.
int64_t resRank = resultType.getRank();
int64_t v1Rank = v1Type.getRank();
int64_t v2Rank = v2Type.getRank();
bool wellFormed0DCase = v1Rank == 0 && v2Rank == 0 && resRank == 1;
bool wellFormedNDCase = v1Rank == resRank && v2Rank == resRank;
if (!wellFormed0DCase && !wellFormedNDCase)
return emitOpError("rank mismatch");
// Verify all but leading dimension sizes.
for (int64_t r = 1; r < v1Rank; ++r) {
int64_t resDim = resultType.getDimSize(r);
int64_t v1Dim = v1Type.getDimSize(r);
int64_t v2Dim = v2Type.getDimSize(r);
if (resDim != v1Dim || v1Dim != v2Dim)
return emitOpError("dimension mismatch");
}
// Verify mask length.
ArrayRef<int64_t> mask = getMask();
int64_t maskLength = mask.size();
if (maskLength <= 0)
return emitOpError("invalid mask length");
if (maskLength != resultType.getDimSize(0))
return emitOpError("mask length mismatch");
// Verify all indices.
int64_t indexSize = (v1Type.getRank() == 0 ? 1 : v1Type.getDimSize(0)) +
(v2Type.getRank() == 0 ? 1 : v2Type.getDimSize(0));
for (auto [idx, maskPos] : llvm::enumerate(mask)) {
if (!isValidPositiveIndexOrPoison(maskPos, kPoisonIndex, indexSize))
return emitOpError("mask index #") << (idx + 1) << " out of range";
}
return success();
}
LogicalResult
ShuffleOp::inferReturnTypes(MLIRContext *, std::optional<Location>,
ShuffleOp::Adaptor adaptor,
SmallVectorImpl<Type> &inferredReturnTypes) {
auto v1Type = llvm::cast<VectorType>(adaptor.getV1().getType());
auto v1Rank = v1Type.getRank();
// Construct resulting type: leading dimension matches mask
// length, all trailing dimensions match the operands.
SmallVector<int64_t, 4> shape;
shape.reserve(v1Rank);
shape.push_back(std::max<size_t>(1, adaptor.getMask().size()));
// In the 0-D case there is no trailing shape to append.
if (v1Rank > 0)
llvm::append_range(shape, v1Type.getShape().drop_front());
inferredReturnTypes.push_back(
VectorType::get(shape, v1Type.getElementType()));
return success();
}
template <typename T>
static bool isStepIndexArray(ArrayRef<T> idxArr, uint64_t begin, size_t width) {
T expected = begin;
return idxArr.size() == width && llvm::all_of(idxArr, [&expected](T value) {
return value == expected++;
});
}
OpFoldResult vector::ShuffleOp::fold(FoldAdaptor adaptor) {
auto v1Type = getV1VectorType();
auto v2Type = getV2VectorType();
assert(!v1Type.isScalable() && !v2Type.isScalable() &&
"Vector shuffle does not support scalable vectors");
// For consistency: 0-D shuffle return type is 1-D, this cannot be a folding
// but must be a canonicalization into a vector.broadcast.
if (v1Type.getRank() == 0)
return {};
// Fold shuffle V1, V2, [0, 1, 2, 3] : <4xi32>, <2xi32> -> V1.
auto mask = getMask();
if (isStepIndexArray(mask, 0, v1Type.getDimSize(0)))
return getV1();
// Fold shuffle V1, V2, [4, 5] : <4xi32>, <2xi32> -> V2.
if (isStepIndexArray(mask, v1Type.getDimSize(0), v2Type.getDimSize(0)))
return getV2();
Attribute v1Attr = adaptor.getV1(), v2Attr = adaptor.getV2();
if (!v1Attr || !v2Attr)
return {};
// Fold shuffle poison, poison -> poison.
bool isV1Poison = isa<ub::PoisonAttr>(v1Attr);
bool isV2Poison = isa<ub::PoisonAttr>(v2Attr);
if (isV1Poison && isV2Poison)
return ub::PoisonAttr::get(getContext());
// Only support 1-D for now to avoid complicated n-D DenseElementsAttr
// manipulation.
if (v1Type.getRank() != 1)
return {};
// Poison input attributes need special handling as they are not
// DenseElementsAttr. If an index is poison, we select the first element of
// the first non-poison input.
SmallVector<Attribute> v1Elements, v2Elements;
Attribute poisonElement;
if (!isV2Poison) {
v2Elements =
to_vector(cast<DenseElementsAttr>(v2Attr).getValues<Attribute>());
poisonElement = v2Elements[0];
}
if (!isV1Poison) {
v1Elements =
to_vector(cast<DenseElementsAttr>(v1Attr).getValues<Attribute>());
poisonElement = v1Elements[0];
}
SmallVector<Attribute> results;
int64_t v1Size = v1Type.getDimSize(0);
for (int64_t maskIdx : mask) {
Attribute indexedElm;
// TODO: Return a partial poison vector when supported by the UB dialect.
if (maskIdx == ShuffleOp::kPoisonIndex) {
indexedElm = poisonElement;
} else {
if (maskIdx < v1Size)
indexedElm = isV1Poison ? poisonElement : v1Elements[maskIdx];
else
indexedElm = isV2Poison ? poisonElement : v2Elements[maskIdx - v1Size];
}
results.push_back(indexedElm);
}
return DenseElementsAttr::get(getResultVectorType(), results);
}
namespace {
// Pattern to rewrite a 0-D shuffle with [0] or [1] mask returning a 1-D vector
// to a broadcast.
struct Canonicalize0DShuffleOp : public OpRewritePattern<ShuffleOp> {
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ShuffleOp shuffleOp,
PatternRewriter &rewriter) const override {
VectorType v1VectorType = shuffleOp.getV1VectorType();
ArrayRef<int64_t> mask = shuffleOp.getMask();
if (v1VectorType.getRank() > 0)
return failure();
if (mask.size() != 1)
return failure();
VectorType resType = VectorType::Builder(v1VectorType).setShape({1});
if (mask[0] == 0)
rewriter.replaceOpWithNewOp<vector::BroadcastOp>(shuffleOp, resType,
shuffleOp.getV1());
else
rewriter.replaceOpWithNewOp<vector::BroadcastOp>(shuffleOp, resType,
shuffleOp.getV2());
return success();
}
};
/// Pattern to rewrite a ShuffleOp(SplatOp, SplatOp) to SplatOp.
class ShuffleSplat final : public OpRewritePattern<ShuffleOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ShuffleOp op,
PatternRewriter &rewriter) const override {
auto v1Splat = op.getV1().getDefiningOp<SplatOp>();
auto v2Splat = op.getV2().getDefiningOp<SplatOp>();
if (!v1Splat || !v2Splat)
return failure();
if (v1Splat.getInput() != v2Splat.getInput())
return failure();
rewriter.replaceOpWithNewOp<SplatOp>(op, op.getType(), v1Splat.getInput());
return success();
}
};
/// Pattern to rewrite a fixed-size interleave via vector.shuffle to
/// vector.interleave.
class ShuffleInterleave : public OpRewritePattern<ShuffleOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ShuffleOp op,
PatternRewriter &rewriter) const override {
VectorType resultType = op.getResultVectorType();
if (resultType.isScalable())
return rewriter.notifyMatchFailure(
op, "ShuffleOp can't represent a scalable interleave");
if (resultType.getRank() != 1)
return rewriter.notifyMatchFailure(
op, "ShuffleOp can't represent an n-D interleave");
VectorType sourceType = op.getV1VectorType();
if (sourceType != op.getV2VectorType() ||
sourceType.getNumElements() * 2 != resultType.getNumElements()) {
return rewriter.notifyMatchFailure(
op, "ShuffleOp types don't match an interleave");
}
ArrayRef<int64_t> shuffleMask = op.getMask();
int64_t resultVectorSize = resultType.getNumElements();
for (int i = 0, e = resultVectorSize / 2; i < e; ++i) {
int64_t maskValueA = shuffleMask[i * 2];
int64_t maskValueB = shuffleMask[(i * 2) + 1];
if (maskValueA != i || maskValueB != (resultVectorSize / 2) + i)
return rewriter.notifyMatchFailure(op,
"ShuffleOp mask not interleaving");
}
rewriter.replaceOpWithNewOp<InterleaveOp>(op, op.getV1(), op.getV2());
return success();
}
};
} // namespace
void ShuffleOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<ShuffleSplat, ShuffleInterleave, Canonicalize0DShuffleOp>(
context);
}
//===----------------------------------------------------------------------===//
// InsertElementOp
//===----------------------------------------------------------------------===//
void InsertElementOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
SetIntRangeFn setResultRanges) {
setResultRanges(getResult(), argRanges[0].rangeUnion(argRanges[1]));
}
void InsertElementOp::build(OpBuilder &builder, OperationState &result,
Value source, Value dest) {
build(builder, result, source, dest, {});
}
LogicalResult InsertElementOp::verify() {
auto dstVectorType = getDestVectorType();
if (dstVectorType.getRank() == 0) {
if (getPosition())
return emitOpError("expected position to be empty with 0-D vector");
return success();
}
if (dstVectorType.getRank() != 1)
return emitOpError("unexpected >1 vector rank");
if (!getPosition())
return emitOpError("expected position for 1-D vector");
return success();
}
OpFoldResult vector::InsertElementOp::fold(FoldAdaptor adaptor) {
// Skip the 0-D vector here.
if (!adaptor.getPosition())
return {};
auto src = dyn_cast_or_null<TypedAttr>(adaptor.getSource());
auto dst = dyn_cast_or_null<DenseElementsAttr>(adaptor.getDest());
auto pos = dyn_cast_or_null<IntegerAttr>(adaptor.getPosition());
if (!src || !dst || !pos)
return {};
if (src.getType() != getDestVectorType().getElementType())
return {};
auto dstElements = dst.getValues<Attribute>();
SmallVector<Attribute> results(dstElements);
uint64_t posIdx = pos.getInt();
if (posIdx >= results.size())
return {};
results[posIdx] = src;
return DenseElementsAttr::get(getDestVectorType(), results);
}
//===----------------------------------------------------------------------===//
// InsertOp
//===----------------------------------------------------------------------===//
void vector::InsertOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
SetIntRangeFn setResultRanges) {
setResultRanges(getResult(), argRanges[0].rangeUnion(argRanges[1]));
}
void vector::InsertOp::build(OpBuilder &builder, OperationState &result,
Value source, Value dest) {
auto vectorTy = cast<VectorType>(dest.getType());
build(builder, result, source, dest,
SmallVector<int64_t>(vectorTy.getRank(), 0));
}
void vector::InsertOp::build(OpBuilder &builder, OperationState &result,
Value source, Value dest, int64_t position) {
build(builder, result, source, dest, ArrayRef<int64_t>{position});
}
void vector::InsertOp::build(OpBuilder &builder, OperationState &result,
Value source, Value dest, OpFoldResult position) {
build(builder, result, source, dest, ArrayRef<OpFoldResult>{position});
}
void vector::InsertOp::build(OpBuilder &builder, OperationState &result,
Value source, Value dest,
ArrayRef<int64_t> position) {
SmallVector<OpFoldResult> posVals;
posVals.reserve(position.size());
llvm::transform(position, std::back_inserter(posVals),
[&](int64_t pos) { return builder.getI64IntegerAttr(pos); });
build(builder, result, source, dest, posVals);
}
void vector::InsertOp::build(OpBuilder &builder, OperationState &result,
Value source, Value dest,
ArrayRef<OpFoldResult> position) {
SmallVector<int64_t> staticPos;
SmallVector<Value> dynamicPos;
dispatchIndexOpFoldResults(position, dynamicPos, staticPos);
build(builder, result, source, dest, dynamicPos,
builder.getDenseI64ArrayAttr(staticPos));
}
LogicalResult InsertOp::verify() {
SmallVector<OpFoldResult> position = getMixedPosition();
auto destVectorType = getDestVectorType();
if (position.size() > static_cast<unsigned>(destVectorType.getRank()))
return emitOpError(
"expected position attribute of rank no greater than dest vector rank");
auto srcVectorType = llvm::dyn_cast<VectorType>(getValueToStoreType());
if (srcVectorType &&
(static_cast<unsigned>(srcVectorType.getRank()) + position.size() !=
static_cast<unsigned>(destVectorType.getRank())))
return emitOpError("expected position attribute rank + source rank to "
"match dest vector rank");
if (!srcVectorType &&
(position.size() != static_cast<unsigned>(destVectorType.getRank())))
return emitOpError(
"expected position attribute rank to match the dest vector rank");
for (auto [idx, pos] : llvm::enumerate(position)) {
if (auto attr = dyn_cast<Attribute>(pos)) {
int64_t constIdx = cast<IntegerAttr>(attr).getInt();
if (!isValidPositiveIndexOrPoison(constIdx, kPoisonIndex,
destVectorType.getDimSize(idx))) {
return emitOpError("expected position attribute #")
<< (idx + 1)
<< " to be a non-negative integer smaller than the "
"corresponding "
"dest vector dimension";
}
}
}
return success();
}
namespace {
// If insertOp is only inserting unit dimensions it can be transformed to a
// broadcast.
class InsertToBroadcast final : public OpRewritePattern<InsertOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(InsertOp insertOp,
PatternRewriter &rewriter) const override {
auto srcVecType =
llvm::dyn_cast<VectorType>(insertOp.getValueToStoreType());
if (!srcVecType || insertOp.getDestVectorType().getNumElements() !=
srcVecType.getNumElements())
return failure();
rewriter.replaceOpWithNewOp<BroadcastOp>(
insertOp, insertOp.getDestVectorType(), insertOp.getValueToStore());
return success();
}
};
/// Pattern to rewrite a InsertOp(SplatOp, SplatOp) to SplatOp.
class InsertSplatToSplat final : public OpRewritePattern<InsertOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(InsertOp op,
PatternRewriter &rewriter) const override {
auto srcSplat = op.getValueToStore().getDefiningOp<SplatOp>();
auto dstSplat = op.getDest().getDefiningOp<SplatOp>();
if (!srcSplat || !dstSplat)
return failure();
if (srcSplat.getInput() != dstSplat.getInput())
return failure();
rewriter.replaceOpWithNewOp<SplatOp>(op, op.getType(), srcSplat.getInput());
return success();
}
};
} // namespace
static Attribute
foldDenseElementsAttrDestInsertOp(InsertOp insertOp, Attribute srcAttr,
Attribute dstAttr,
int64_t maxVectorSizeFoldThreshold) {
if (insertOp.hasDynamicPosition())
return {};
auto denseDst = llvm::dyn_cast_if_present<DenseElementsAttr>(dstAttr);
if (!denseDst)
return {};
if (!srcAttr) {
return {};
}
VectorType destTy = insertOp.getDestVectorType();
if (destTy.isScalable())
return {};
// Make sure we do not create too many large constants.
if (destTy.getNumElements() > maxVectorSizeFoldThreshold &&
!insertOp->hasOneUse())
return {};
// Calculate the linearized position of the continuous chunk of elements to
// insert.
llvm::SmallVector<int64_t> completePositions(destTy.getRank(), 0);
copy(insertOp.getStaticPosition(), completePositions.begin());
int64_t insertBeginPosition =
linearize(completePositions, computeStrides(destTy.getShape()));
SmallVector<Attribute> insertedValues;
Type destEltType = destTy.getElementType();
/// Converts the expected type to an IntegerAttr if there's
/// a mismatch.
auto convertIntegerAttr = [](Attribute attr, Type expectedType) -> Attribute {
if (auto intAttr = mlir::dyn_cast<IntegerAttr>(attr)) {
if (intAttr.getType() != expectedType)
return IntegerAttr::get(expectedType, intAttr.getInt());
}
return attr;
};
// The `convertIntegerAttr` method specifically handles the case
// for `llvm.mlir.constant` which can hold an attribute with a
// different type than the return type.
if (auto denseSource = llvm::dyn_cast<DenseElementsAttr>(srcAttr)) {
for (auto value : denseSource.getValues<Attribute>())
insertedValues.push_back(convertIntegerAttr(value, destEltType));
} else {
insertedValues.push_back(convertIntegerAttr(srcAttr, destEltType));
}
auto allValues = llvm::to_vector(denseDst.getValues<Attribute>());
copy(insertedValues, allValues.begin() + insertBeginPosition);
auto newAttr = DenseElementsAttr::get(destTy, allValues);
return newAttr;
}
void InsertOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<InsertToBroadcast, BroadcastFolder, InsertSplatToSplat>(context);
}
OpFoldResult vector::InsertOp::fold(FoldAdaptor adaptor) {
// Do not create constants with more than `vectorSizeFoldThreashold` elements,
// unless the source vector constant has a single use.
constexpr int64_t vectorSizeFoldThreshold = 256;
// Fold "vector.insert %v, %dest [] : vector<2x2xf32> from vector<2x2xf32>" to
// %v. Note: Do not fold "vector.insert %v, %dest [] : f32 into vector<f32>"
// (type mismatch).
if (getNumIndices() == 0 && getValueToStoreType() == getType())
return getValueToStore();
SmallVector<Value> operands = {getValueToStore(), getDest()};
if (auto val = extractInsertFoldConstantOp(*this, adaptor, operands))
return val;
if (auto res = foldPoisonIndexInsertExtractOp(
getContext(), adaptor.getStaticPosition(), kPoisonIndex))
return res;
if (auto res = foldDenseElementsAttrDestInsertOp(
*this, adaptor.getValueToStore(), adaptor.getDest(),
vectorSizeFoldThreshold)) {
return res;
}
return {};
}
//===----------------------------------------------------------------------===//
// InsertStridedSliceOp
//===----------------------------------------------------------------------===//
void InsertStridedSliceOp::build(OpBuilder &builder, OperationState &result,
Value source, Value dest,
ArrayRef<int64_t> offsets,
ArrayRef<int64_t> strides) {
result.addOperands({source, dest});
auto offsetsAttr = getVectorSubscriptAttr(builder, offsets);
auto stridesAttr = getVectorSubscriptAttr(builder, strides);
result.addTypes(dest.getType());
result.addAttribute(InsertStridedSliceOp::getOffsetsAttrName(result.name),
offsetsAttr);
result.addAttribute(InsertStridedSliceOp::getStridesAttrName(result.name),
stridesAttr);
}
// TODO: Should be moved to Tablegen ConfinedAttr attributes.
template <typename OpType>
static LogicalResult isIntegerArrayAttrSmallerThanShape(OpType op,
ArrayAttr arrayAttr,
ArrayRef<int64_t> shape,
StringRef attrName) {
if (arrayAttr.size() > shape.size())
return op.emitOpError("expected ")
<< attrName << " attribute of rank no greater than vector rank";
return success();
}
// Returns true if all integers in `arrayAttr` are in the half-open [min, max}
// interval. If `halfOpen` is true then the admissible interval is [min, max).
// Otherwise, the admissible interval is [min, max].
template <typename OpType>
static LogicalResult
isIntegerArrayAttrConfinedToRange(OpType op, ArrayAttr arrayAttr, int64_t min,
int64_t max, StringRef attrName,
bool halfOpen = true) {
for (auto attr : arrayAttr) {
auto val = llvm::cast<IntegerAttr>(attr).getInt();
auto upper = max;
if (!halfOpen)
upper += 1;
if (val < min || val >= upper)
return op.emitOpError("expected ") << attrName << " to be confined to ["
<< min << ", " << upper << ")";
}
return success();
}
// Returns true if all integers in `arrayAttr` are in the half-open [min, max}
// interval. If `halfOpen` is true then the admissible interval is [min, max).
// Otherwise, the admissible interval is [min, max].
template <typename OpType>
static LogicalResult
isIntegerArrayAttrConfinedToShape(OpType op, ArrayAttr arrayAttr,
ArrayRef<int64_t> shape, StringRef attrName,
bool halfOpen = true, int64_t min = 0) {
for (auto [index, attrDimPair] :
llvm::enumerate(llvm::zip_first(arrayAttr, shape))) {
int64_t val = llvm::cast<IntegerAttr>(std::get<0>(attrDimPair)).getInt();
int64_t max = std::get<1>(attrDimPair);
if (!halfOpen)
max += 1;
if (val < min || val >= max)
return op.emitOpError("expected ")
<< attrName << " dimension " << index << " to be confined to ["
<< min << ", " << max << ")";
}
return success();
}
// Returns true if, for all indices i = 0..shape.size()-1, val is in the
// [min, max} interval:
// val = `arrayAttr1[i]` + `arrayAttr2[i]`,
// If `halfOpen` is true then the admissible interval is [min, max). Otherwise,
// the admissible interval is [min, max].
template <typename OpType>
static LogicalResult isSumOfIntegerArrayAttrConfinedToShape(
OpType op, ArrayAttr arrayAttr1, ArrayAttr arrayAttr2,
ArrayRef<int64_t> shape, StringRef attrName1, StringRef attrName2,
bool halfOpen = true, int64_t min = 1) {
assert(arrayAttr1.size() <= shape.size());
assert(arrayAttr2.size() <= shape.size());
for (auto [index, it] :
llvm::enumerate(llvm::zip(arrayAttr1, arrayAttr2, shape))) {
auto val1 = llvm::cast<IntegerAttr>(std::get<0>(it)).getInt();
auto val2 = llvm::cast<IntegerAttr>(std::get<1>(it)).getInt();
int64_t max = std::get<2>(it);
if (!halfOpen)
max += 1;
if (val1 + val2 < 0 || val1 + val2 >= max)
return op.emitOpError("expected sum(")
<< attrName1 << ", " << attrName2 << ") dimension " << index
<< " to be confined to [" << min << ", " << max << ")";
}
return success();
}
static ArrayAttr makeI64ArrayAttr(ArrayRef<int64_t> values,
MLIRContext *context) {
auto attrs = llvm::map_range(values, [context](int64_t v) -> Attribute {
return IntegerAttr::get(IntegerType::get(context, 64), APInt(64, v));
});
return ArrayAttr::get(context, llvm::to_vector<8>(attrs));
}
LogicalResult InsertStridedSliceOp::verify() {
auto sourceVectorType = getSourceVectorType();
auto destVectorType = getDestVectorType();
auto offsets = getOffsetsAttr();
auto strides = getStridesAttr();
if (offsets.size() != static_cast<unsigned>(destVectorType.getRank()))
return emitOpError(
"expected offsets of same size as destination vector rank");
if (strides.size() != static_cast<unsigned>(sourceVectorType.getRank()))
return emitOpError("expected strides of same size as source vector rank");
if (sourceVectorType.getRank() > destVectorType.getRank())
return emitOpError(
"expected source rank to be no greater than destination rank");
auto sourceShape = sourceVectorType.getShape();
auto destShape = destVectorType.getShape();
SmallVector<int64_t, 4> sourceShapeAsDestShape(
destShape.size() - sourceShape.size(), 0);
sourceShapeAsDestShape.append(sourceShape.begin(), sourceShape.end());
auto offName = InsertStridedSliceOp::getOffsetsAttrName();
auto stridesName = InsertStridedSliceOp::getStridesAttrName();
if (failed(isIntegerArrayAttrConfinedToShape(*this, offsets, destShape,
offName)) ||
failed(isIntegerArrayAttrConfinedToRange(*this, strides, /*min=*/1,
/*max=*/1, stridesName,
/*halfOpen=*/false)) ||
failed(isSumOfIntegerArrayAttrConfinedToShape(
*this, offsets,
makeI64ArrayAttr(sourceShapeAsDestShape, getContext()), destShape,
offName, "source vector shape",
/*halfOpen=*/false, /*min=*/1)))
return failure();
unsigned rankDiff = destShape.size() - sourceShape.size();
for (unsigned idx = 0; idx < sourceShape.size(); ++idx) {
if (sourceVectorType.getScalableDims()[idx] !=
destVectorType.getScalableDims()[idx + rankDiff]) {
return emitOpError("mismatching scalable flags (at source vector idx=")
<< idx << ")";
}
if (sourceVectorType.getScalableDims()[idx]) {
auto sourceSize = sourceShape[idx];
auto destSize = destShape[idx + rankDiff];
if (sourceSize != destSize) {
return emitOpError("expected size at idx=")
<< idx
<< (" to match the corresponding base size from the input "
"vector (")
<< sourceSize << (" vs ") << destSize << (")");
}
}
}
return success();
}
namespace {
/// Pattern to rewrite an InsertStridedSliceOp(SplatOp(X):src_type,
/// SplatOp(X):dst_type) to SplatOp(X):dst_type.
class FoldInsertStridedSliceSplat final
: public OpRewritePattern<InsertStridedSliceOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(InsertStridedSliceOp insertStridedSliceOp,
PatternRewriter &rewriter) const override {
auto srcSplatOp =
insertStridedSliceOp.getValueToStore().getDefiningOp<vector::SplatOp>();
auto destSplatOp =
insertStridedSliceOp.getDest().getDefiningOp<vector::SplatOp>();
if (!srcSplatOp || !destSplatOp)
return failure();
if (srcSplatOp.getInput() != destSplatOp.getInput())
return failure();
rewriter.replaceOp(insertStridedSliceOp, insertStridedSliceOp.getDest());
return success();
}
};
/// Pattern to rewrite an InsertStridedSliceOp(ExtractStridedSliceOp(dst), dst)
/// to dst.
class FoldInsertStridedSliceOfExtract final
: public OpRewritePattern<InsertStridedSliceOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(InsertStridedSliceOp insertStridedSliceOp,
PatternRewriter &rewriter) const override {
auto extractStridedSliceOp =
insertStridedSliceOp.getValueToStore()
.getDefiningOp<vector::ExtractStridedSliceOp>();
if (!extractStridedSliceOp)
return failure();
if (extractStridedSliceOp.getOperand() != insertStridedSliceOp.getDest())
return failure();
// Check if have the same strides and offsets.
if (extractStridedSliceOp.getStrides() !=
insertStridedSliceOp.getStrides() ||
extractStridedSliceOp.getOffsets() != insertStridedSliceOp.getOffsets())
return failure();
rewriter.replaceOp(insertStridedSliceOp, insertStridedSliceOp.getDest());
return success();
}
};
// Pattern to rewrite an InsertStridedSliceOp(ConstantOp into ConstantOp) ->
// ConstantOp.
class InsertStridedSliceConstantFolder final
: public OpRewritePattern<InsertStridedSliceOp> {
public:
using OpRewritePattern::OpRewritePattern;
// Do not create constants with more than `vectorSizeFoldThreashold` elements,
// unless the source vector constant has a single use.
static constexpr int64_t vectorSizeFoldThreshold = 256;
LogicalResult matchAndRewrite(InsertStridedSliceOp op,
PatternRewriter &rewriter) const override {
// Return if 'InsertOp' operand is not defined by a compatible vector
// ConstantOp.
TypedValue<VectorType> destVector = op.getDest();
Attribute vectorDestCst;
if (!matchPattern(destVector, m_Constant(&vectorDestCst)))
return failure();
VectorType destTy = destVector.getType();
if (destTy.isScalable())
return failure();
// Make sure we do not create too many large constants.
if (destTy.getNumElements() > vectorSizeFoldThreshold &&
!destVector.hasOneUse())
return failure();
TypedValue<VectorType> sourceValue = op.getValueToStore();
Attribute sourceCst;
if (!matchPattern(sourceValue, m_Constant(&sourceCst)))
return failure();
// TODO: Support poison.
if (isa<ub::PoisonAttr>(vectorDestCst) || isa<ub::PoisonAttr>(sourceCst))
return failure();
// TODO: Handle non-unit strides when they become available.
if (op.hasNonUnitStrides())
return failure();
VectorType sliceVecTy = sourceValue.getType();
ArrayRef<int64_t> sliceShape = sliceVecTy.getShape();
int64_t rankDifference = destTy.getRank() - sliceVecTy.getRank();
SmallVector<int64_t, 4> offsets = getI64SubArray(op.getOffsets());
SmallVector<int64_t, 4> destStrides = computeStrides(destTy.getShape());
// Calcualte the destination element indices by enumerating all slice
// positions within the destination and linearizing them. The enumeration
// order is lexicographic which yields a sequence of monotonically
// increasing linearized position indices.
// Because the destination may have higher dimensionality then the slice,
// we keep track of two overlapping sets of positions and offsets.
auto denseDest = llvm::cast<DenseElementsAttr>(vectorDestCst);
auto denseSlice = llvm::cast<DenseElementsAttr>(sourceCst);
auto sliceValuesIt = denseSlice.value_begin<Attribute>();
auto newValues = llvm::to_vector(denseDest.getValues<Attribute>());
SmallVector<int64_t> currDestPosition(offsets.begin(), offsets.end());
MutableArrayRef<int64_t> currSlicePosition(
currDestPosition.begin() + rankDifference, currDestPosition.end());
ArrayRef<int64_t> sliceOffsets(offsets.begin() + rankDifference,
offsets.end());
do {
int64_t linearizedPosition = linearize(currDestPosition, destStrides);
assert(linearizedPosition < destTy.getNumElements() && "Invalid index");
assert(sliceValuesIt != denseSlice.value_end<Attribute>() &&
"Invalid slice element");
newValues[linearizedPosition] = *sliceValuesIt;
++sliceValuesIt;
} while (succeeded(
incSlicePosition(currSlicePosition, sliceShape, sliceOffsets)));
auto newAttr = DenseElementsAttr::get(destTy, newValues);
rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, newAttr);
return success();
}
};
} // namespace
void vector::InsertStridedSliceOp::getCanonicalizationPatterns(
RewritePatternSet &results, MLIRContext *context) {
results.add<FoldInsertStridedSliceSplat, FoldInsertStridedSliceOfExtract,
InsertStridedSliceConstantFolder>(context);
}
OpFoldResult InsertStridedSliceOp::fold(FoldAdaptor adaptor) {
if (getSourceVectorType() == getDestVectorType())
return getValueToStore();
return {};
}
//===----------------------------------------------------------------------===//
// OuterProductOp
//===----------------------------------------------------------------------===//
/// Build an op without mask, use the type of `acc` as the return type.
void OuterProductOp::build(OpBuilder &builder, OperationState &result,
Value lhs, Value rhs, Value acc) {
result.addOperands({lhs, rhs, acc});
result.addTypes(acc.getType());
}
void OuterProductOp::print(OpAsmPrinter &p) {
p << " " << getLhs() << ", " << getRhs();
if (getAcc()) {
p << ", " << getAcc();
p.printOptionalAttrDict((*this)->getAttrs());
}
p << " : " << getLhs().getType() << ", " << getRhs().getType();
}
ParseResult OuterProductOp::parse(OpAsmParser &parser, OperationState &result) {
SmallVector<OpAsmParser::UnresolvedOperand, 3> operandsInfo;
Type tLHS, tRHS;
if (parser.parseOperandList(operandsInfo) ||
parser.parseOptionalAttrDict(result.attributes) ||
parser.parseColonType(tLHS) || parser.parseComma() ||
parser.parseType(tRHS))
return failure();
if (operandsInfo.size() < 2)
return parser.emitError(parser.getNameLoc(),
"expected at least 2 operands");
VectorType vLHS = llvm::dyn_cast<VectorType>(tLHS);
VectorType vRHS = llvm::dyn_cast<VectorType>(tRHS);
if (!vLHS)
return parser.emitError(parser.getNameLoc(),
"expected vector type for operand #1");
VectorType resType;
if (vRHS) {
SmallVector<bool> scalableDimsRes{vLHS.getScalableDims()[0],
vRHS.getScalableDims()[0]};
resType = VectorType::get({vLHS.getDimSize(0), vRHS.getDimSize(0)},
vLHS.getElementType(), scalableDimsRes);
} else {
// Scalar RHS operand
SmallVector<bool> scalableDimsRes{vLHS.getScalableDims()[0]};
resType = VectorType::get({vLHS.getDimSize(0)}, vLHS.getElementType(),
scalableDimsRes);
}
if (!result.attributes.get(OuterProductOp::getKindAttrName(result.name))) {
result.attributes.append(
OuterProductOp::getKindAttrName(result.name),
CombiningKindAttr::get(result.getContext(),
OuterProductOp::getDefaultKind()));
}
return failure(
parser.resolveOperand(operandsInfo[0], tLHS, result.operands) ||
parser.resolveOperand(operandsInfo[1], tRHS, result.operands) ||
(operandsInfo.size() > 2 &&
parser.resolveOperand(operandsInfo[2], resType, result.operands)) ||
parser.addTypeToList(resType, result.types));
}
LogicalResult OuterProductOp::verify() {
Type tRHS = getOperandTypeRHS();
VectorType vLHS = getOperandVectorTypeLHS(),
vRHS = llvm::dyn_cast<VectorType>(tRHS),
vACC = getOperandVectorTypeACC(), vRES = getResultVectorType();
if (vLHS.getRank() != 1)
return emitOpError("expected 1-d vector for operand #1");
if (vRHS) {
// Proper OUTER operation.
if (vRHS.getRank() != 1)
return emitOpError("expected 1-d vector for operand #2");
if (vRES.getRank() != 2)
return emitOpError("expected 2-d vector result");
if (vLHS.getDimSize(0) != vRES.getDimSize(0))
return emitOpError("expected #1 operand dim to match result dim #1");
if (vRHS.getDimSize(0) != vRES.getDimSize(1))
return emitOpError("expected #2 operand dim to match result dim #2");
if (vLHS.isScalable() && !vRHS.isScalable()) {
// This restriction reflects what's currently supported in terms of
// scalable vectors. However, we could relax this if there's a use case.
return emitOpError(
"expected either both or only #2 operand dim to be scalable");
}
} else {
// An AXPY operation.
if (vRES.getRank() != 1)
return emitOpError("expected 1-d vector result");
if (vLHS.getDimSize(0) != vRES.getDimSize(0))
return emitOpError("expected #1 operand dim to match result dim #1");
}
if (vACC && vACC != vRES)
return emitOpError("expected operand #3 of same type as result type");
// Verify supported combining kind.
if (!isSupportedCombiningKind(getKind(), vRES.getElementType()))
return emitOpError("unsupported outerproduct type");
return success();
}
// MaskableOpInterface methods.
/// Returns the mask type expected by this operation. Mostly used for
/// verification purposes. It requires the operation to be vectorized."
Type OuterProductOp::getExpectedMaskType() {
auto vecType = this->getResultVectorType();
return VectorType::get(vecType.getShape(),
IntegerType::get(vecType.getContext(), /*width=*/1),
vecType.getScalableDims());
}
//===----------------------------------------------------------------------===//
// ExtractStridedSliceOp
//===----------------------------------------------------------------------===//
// Inference works as follows:
// 1. Add 'sizes' from prefix of dims in 'offsets'.
// 2. Add sizes from 'vectorType' for remaining dims.
// Scalable flags are inherited from 'vectorType'.
static Type inferStridedSliceOpResultType(VectorType vectorType,
ArrayAttr offsets, ArrayAttr sizes,
ArrayAttr strides) {
assert(offsets.size() == sizes.size() && offsets.size() == strides.size());
SmallVector<int64_t, 4> shape;
shape.reserve(vectorType.getRank());
unsigned idx = 0;
for (unsigned e = offsets.size(); idx < e; ++idx)
shape.push_back(llvm::cast<IntegerAttr>(sizes[idx]).getInt());
for (unsigned e = vectorType.getShape().size(); idx < e; ++idx)
shape.push_back(vectorType.getShape()[idx]);
return VectorType::get(shape, vectorType.getElementType(),
vectorType.getScalableDims());
}
void ExtractStridedSliceOp::build(OpBuilder &builder, OperationState &result,
Value source, ArrayRef<int64_t> offsets,
ArrayRef<int64_t> sizes,
ArrayRef<int64_t> strides) {
result.addOperands(source);
auto offsetsAttr = getVectorSubscriptAttr(builder, offsets);
auto sizesAttr = getVectorSubscriptAttr(builder, sizes);
auto stridesAttr = getVectorSubscriptAttr(builder, strides);
result.addTypes(
inferStridedSliceOpResultType(llvm::cast<VectorType>(source.getType()),
offsetsAttr, sizesAttr, stridesAttr));
result.addAttribute(ExtractStridedSliceOp::getOffsetsAttrName(result.name),
offsetsAttr);
result.addAttribute(ExtractStridedSliceOp::getSizesAttrName(result.name),
sizesAttr);
result.addAttribute(ExtractStridedSliceOp::getStridesAttrName(result.name),
stridesAttr);
}
LogicalResult ExtractStridedSliceOp::verify() {
auto type = getSourceVectorType();
auto offsets = getOffsetsAttr();
auto sizes = getSizesAttr();
auto strides = getStridesAttr();
if (offsets.size() != sizes.size() || offsets.size() != strides.size())
return emitOpError(
"expected offsets, sizes and strides attributes of same size");
auto shape = type.getShape();
auto offName = getOffsetsAttrName();
auto sizesName = getSizesAttrName();
auto stridesName = getStridesAttrName();
if (failed(
isIntegerArrayAttrSmallerThanShape(*this, offsets, shape, offName)) ||
failed(
isIntegerArrayAttrSmallerThanShape(*this, sizes, shape, sizesName)) ||
failed(isIntegerArrayAttrSmallerThanShape(*this, strides, shape,
stridesName)) ||
failed(
isIntegerArrayAttrConfinedToShape(*this, offsets, shape, offName)) ||
failed(isIntegerArrayAttrConfinedToShape(*this, sizes, shape, sizesName,
/*halfOpen=*/false,
/*min=*/1)) ||
failed(isIntegerArrayAttrConfinedToRange(*this, strides, /*min=*/1,
/*max=*/1, stridesName,
/*halfOpen=*/false)) ||
failed(isSumOfIntegerArrayAttrConfinedToShape(*this, offsets, sizes,
shape, offName, sizesName,
/*halfOpen=*/false)))
return failure();
auto resultType = inferStridedSliceOpResultType(getSourceVectorType(),
offsets, sizes, strides);
if (getResult().getType() != resultType)
return emitOpError("expected result type to be ") << resultType;
for (unsigned idx = 0; idx < sizes.size(); ++idx) {
if (type.getScalableDims()[idx]) {
auto inputDim = type.getShape()[idx];
auto inputSize = llvm::cast<IntegerAttr>(sizes[idx]).getInt();
if (inputDim != inputSize)
return emitOpError("expected size at idx=")
<< idx
<< (" to match the corresponding base size from the input "
"vector (")
<< inputSize << (" vs ") << inputDim << (")");
}
}
return success();
}
// When the source of ExtractStrided comes from a chain of InsertStrided ops try
// to use the source of the InsertStrided ops if we can detect that the
// extracted vector is a subset of one of the vector inserted.
static LogicalResult
foldExtractStridedOpFromInsertChain(ExtractStridedSliceOp op) {
// Helper to extract integer out of ArrayAttr.
auto getElement = [](ArrayAttr array, int idx) {
return llvm::cast<IntegerAttr>(array[idx]).getInt();
};
ArrayAttr extractOffsets = op.getOffsets();
ArrayAttr extractStrides = op.getStrides();
ArrayAttr extractSizes = op.getSizes();
auto insertOp = op.getVector().getDefiningOp<InsertStridedSliceOp>();
while (insertOp) {
if (op.getSourceVectorType().getRank() !=
insertOp.getSourceVectorType().getRank())
return failure();
ArrayAttr insertOffsets = insertOp.getOffsets();
ArrayAttr insertStrides = insertOp.getStrides();
// If the rank of extract is greater than the rank of insert, we are likely
// extracting a partial chunk of the vector inserted.
if (extractOffsets.size() > insertOffsets.size())
return failure();
bool patialoverlap = false;
bool disjoint = false;
SmallVector<int64_t, 4> offsetDiffs;
for (unsigned dim = 0, e = extractOffsets.size(); dim < e; ++dim) {
if (getElement(extractStrides, dim) != getElement(insertStrides, dim))
return failure();
int64_t start = getElement(insertOffsets, dim);
int64_t end = start + insertOp.getSourceVectorType().getDimSize(dim);
int64_t offset = getElement(extractOffsets, dim);
int64_t size = getElement(extractSizes, dim);
// Check if the start of the extract offset is in the interval inserted.
if (start <= offset && offset < end) {
// If the extract interval overlaps but is not fully included we may
// have a partial overlap that will prevent any folding.
if (offset + size > end)
patialoverlap = true;
offsetDiffs.push_back(offset - start);
continue;
}
disjoint = true;
break;
}
// The extract element chunk is a subset of the insert element.
if (!disjoint && !patialoverlap) {
op.setOperand(insertOp.getValueToStore());
// OpBuilder is only used as a helper to build an I64ArrayAttr.
OpBuilder b(op.getContext());
op.setOffsetsAttr(b.getI64ArrayAttr(offsetDiffs));
return success();
}
// If the chunk extracted is disjoint from the chunk inserted, keep looking
// in the insert chain.
if (disjoint)
insertOp = insertOp.getDest().getDefiningOp<InsertStridedSliceOp>();
else {
// The extracted vector partially overlap the inserted vector, we cannot
// fold.
return failure();
}
}
return failure();
}
// ExtractStridedSliceOp(non-splat ConstantOp) -> ConstantOp.
static OpFoldResult
foldExtractStridedSliceNonSplatConstant(ExtractStridedSliceOp op,
Attribute foldInput) {
auto dense = llvm::dyn_cast_if_present<DenseElementsAttr>(foldInput);
if (!dense)
return {};
// TODO: Handle non-unit strides when they become available.
if (op.hasNonUnitStrides())
return {};
VectorType sourceVecTy = op.getSourceVectorType();
ArrayRef<int64_t> sourceShape = sourceVecTy.getShape();
SmallVector<int64_t, 4> sourceStrides = computeStrides(sourceShape);
VectorType sliceVecTy = op.getType();
ArrayRef<int64_t> sliceShape = sliceVecTy.getShape();
int64_t rank = sliceVecTy.getRank();
// Expand offsets and sizes to match the vector rank.
SmallVector<int64_t, 4> offsets(rank, 0);
copy(getI64SubArray(op.getOffsets()), offsets.begin());
SmallVector<int64_t, 4> sizes(sourceShape);
copy(getI64SubArray(op.getSizes()), sizes.begin());
// Calculate the slice elements by enumerating all slice positions and
// linearizing them. The enumeration order is lexicographic which yields a
// sequence of monotonically increasing linearized position indices.
const auto denseValuesBegin = dense.value_begin<Attribute>();
SmallVector<Attribute> sliceValues;
sliceValues.reserve(sliceVecTy.getNumElements());
SmallVector<int64_t> currSlicePosition(offsets.begin(), offsets.end());
do {
int64_t linearizedPosition = linearize(currSlicePosition, sourceStrides);
assert(linearizedPosition < sourceVecTy.getNumElements() &&
"Invalid index");
sliceValues.push_back(*(denseValuesBegin + linearizedPosition));
} while (succeeded(incSlicePosition(currSlicePosition, sliceShape, offsets)));
assert(static_cast<int64_t>(sliceValues.size()) ==
sliceVecTy.getNumElements() &&
"Invalid number of slice elements");
return DenseElementsAttr::get(sliceVecTy, sliceValues);
}
OpFoldResult ExtractStridedSliceOp::fold(FoldAdaptor adaptor) {
if (getSourceVectorType() == getResult().getType())
return getVector();
if (succeeded(foldExtractStridedOpFromInsertChain(*this)))
return getResult();
// ExtractStridedSliceOp(splat ConstantOp) -> ConstantOp.
if (auto splat =
llvm::dyn_cast_if_present<SplatElementsAttr>(adaptor.getVector()))
DenseElementsAttr::get(getType(), splat.getSplatValue<Attribute>());
// ExtractStridedSliceOp(non-splat ConstantOp) -> ConstantOp.
return foldExtractStridedSliceNonSplatConstant(*this, adaptor.getVector());
}
void ExtractStridedSliceOp::getOffsets(SmallVectorImpl<int64_t> &results) {
populateFromInt64AttrArray(getOffsets(), results);
}
namespace {
// Pattern to rewrite an ExtractStridedSliceOp(ConstantMaskOp) to
// ConstantMaskOp.
class StridedSliceConstantMaskFolder final
: public OpRewritePattern<ExtractStridedSliceOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ExtractStridedSliceOp extractStridedSliceOp,
PatternRewriter &rewriter) const override {
// Return if 'extractStridedSliceOp' operand is not defined by a
// ConstantMaskOp.
auto *defOp = extractStridedSliceOp.getVector().getDefiningOp();
auto constantMaskOp = dyn_cast_or_null<ConstantMaskOp>(defOp);
if (!constantMaskOp)
return failure();
// Return if 'extractStridedSliceOp' has non-unit strides.
if (extractStridedSliceOp.hasNonUnitStrides())
return failure();
// Gather constant mask dimension sizes.
ArrayRef<int64_t> maskDimSizes = constantMaskOp.getMaskDimSizes();
// Gather strided slice offsets and sizes.
SmallVector<int64_t, 4> sliceOffsets;
populateFromInt64AttrArray(extractStridedSliceOp.getOffsets(),
sliceOffsets);
SmallVector<int64_t, 4> sliceSizes;
populateFromInt64AttrArray(extractStridedSliceOp.getSizes(), sliceSizes);
// Compute slice of vector mask region.
SmallVector<int64_t, 4> sliceMaskDimSizes;
sliceMaskDimSizes.reserve(maskDimSizes.size());
for (auto [maskDimSize, sliceOffset, sliceSize] :
llvm::zip(maskDimSizes, sliceOffsets, sliceSizes)) {
int64_t sliceMaskDimSize = std::max(
static_cast<int64_t>(0),
std::min(sliceOffset + sliceSize, maskDimSize) - sliceOffset);
sliceMaskDimSizes.push_back(sliceMaskDimSize);
}
// Add unchanged dimensions.
if (sliceMaskDimSizes.size() < maskDimSizes.size())
for (size_t i = sliceMaskDimSizes.size(); i < maskDimSizes.size(); ++i)
sliceMaskDimSizes.push_back(maskDimSizes[i]);
// If any of 'sliceMaskDimSizes' are zero, then set all to zero (masked
// region is a conjunction of mask dim intervals).
if (llvm::is_contained(sliceMaskDimSizes, 0))
sliceMaskDimSizes.assign(maskDimSizes.size(), 0);
// Replace 'extractStridedSliceOp' with ConstantMaskOp with sliced mask
// region.
rewriter.replaceOpWithNewOp<ConstantMaskOp>(
extractStridedSliceOp, extractStridedSliceOp.getResult().getType(),
sliceMaskDimSizes);
return success();
}
};
// Pattern to rewrite an ExtractStridedSliceOp(BroadcastOp) to
// BroadcastOp(ExtractStrideSliceOp).
class StridedSliceBroadcast final
: public OpRewritePattern<ExtractStridedSliceOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ExtractStridedSliceOp op,
PatternRewriter &rewriter) const override {
auto broadcast = op.getVector().getDefiningOp<BroadcastOp>();
if (!broadcast)
return failure();
auto srcVecType =
llvm::dyn_cast<VectorType>(broadcast.getSource().getType());
unsigned srcRank = srcVecType ? srcVecType.getRank() : 0;
auto dstVecType = llvm::cast<VectorType>(op.getType());
unsigned dstRank = dstVecType.getRank();
unsigned rankDiff = dstRank - srcRank;
// Check if the most inner dimensions of the source of the broadcast are the
// same as the destination of the extract. If this is the case we can just
// use a broadcast as the original dimensions are untouched.
bool lowerDimMatch = true;
for (unsigned i = 0; i < srcRank; i++) {
if (srcVecType.getDimSize(i) != dstVecType.getDimSize(i + rankDiff)) {
lowerDimMatch = false;
break;
}
}
Value source = broadcast.getSource();
// If the inner dimensions don't match, it means we need to extract from the
// source of the orignal broadcast and then broadcast the extracted value.
// We also need to handle degenerated cases where the source is effectively
// just a single scalar.
bool isScalarSrc = (srcRank == 0 || srcVecType.getNumElements() == 1);
if (!lowerDimMatch && !isScalarSrc) {
source = rewriter.create<ExtractStridedSliceOp>(
op->getLoc(), source,
getI64SubArray(op.getOffsets(), /* dropFront=*/rankDiff),
getI64SubArray(op.getSizes(), /* dropFront=*/rankDiff),
getI64SubArray(op.getStrides(), /* dropFront=*/rankDiff));
}
rewriter.replaceOpWithNewOp<BroadcastOp>(op, op.getType(), source);
return success();
}
};
/// Pattern to rewrite an ExtractStridedSliceOp(SplatOp) to SplatOp.
class StridedSliceSplat final : public OpRewritePattern<ExtractStridedSliceOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ExtractStridedSliceOp op,
PatternRewriter &rewriter) const override {
auto splat = op.getVector().getDefiningOp<SplatOp>();
if (!splat)
return failure();
rewriter.replaceOpWithNewOp<SplatOp>(op, op.getType(), splat.getInput());
return success();
}
};
/// Pattern to rewrite simple cases of N-D extract_strided_slice, where the
/// slice is contiguous, into extract and shape_cast.
///
/// Example:
/// Before:
/// %1 = vector.extract_strided_slice %arg0 {
/// offsets = [0, 0, 0, 0, 0],
/// sizes = [1, 1, 1, 1, 8],
/// strides = [1, 1, 1, 1, 1]
/// } : vector<8x1x1x2x8xi8> to vector<1x1x1x1x8xi8>
/// After:
/// %0 = vector.extract %arg0[0, 0, 0, 0]
/// : vector<8xi8> from vector<8x1x1x2x8xi8>
/// %1 = vector.shape_cast %0
/// : vector<8xi8> to vector<1x1x1x1x8xi8>
///
class ContiguousExtractStridedSliceToExtract final
: public OpRewritePattern<ExtractStridedSliceOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ExtractStridedSliceOp op,
PatternRewriter &rewriter) const override {
if (op.hasNonUnitStrides())
return failure();
Value source = op.getOperand();
auto sourceType = cast<VectorType>(source.getType());
if (sourceType.isScalable() || sourceType.getRank() == 0)
return failure();
// Compute the number of offsets to pass to ExtractOp::build. That is the
// difference between the source rank and the desired slice rank. We walk
// the dimensions from innermost out, and stop when the next slice dimension
// is not full-size.
SmallVector<int64_t> sizes = getI64SubArray(op.getSizes());
int numOffsets;
for (numOffsets = sizes.size(); numOffsets > 0; --numOffsets) {
if (sizes[numOffsets - 1] != sourceType.getDimSize(numOffsets - 1))
break;
}
// If the created extract op would have no offsets, then this whole
// extract_strided_slice is the identity and should have been handled by
// other canonicalizations.
if (numOffsets == 0)
return failure();
// If not even the inner-most dimension is full-size, this op can't be
// rewritten as an ExtractOp.
if (numOffsets == sourceType.getRank() &&
static_cast<int>(sizes.size()) == sourceType.getRank())
return failure();
// The outer dimensions must have unit size.
for (int i = 0; i < numOffsets; ++i) {
if (sizes[i] != 1)
return failure();
}
// Avoid generating slices that have leading unit dimensions. The shape_cast
// op that we create below would take bad generic fallback patterns
// (ShapeCastOpRewritePattern).
while (numOffsets < static_cast<int>(sizes.size()) - 1 &&
sizes[numOffsets] == 1) {
++numOffsets;
}
SmallVector<int64_t> offsets = getI64SubArray(op.getOffsets());
auto extractOffsets = ArrayRef(offsets).take_front(numOffsets);
Value extract = rewriter.create<vector::ExtractOp>(op->getLoc(), source,
extractOffsets);
rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(op, op.getType(), extract);
return success();
}
};
} // namespace
void ExtractStridedSliceOp::getCanonicalizationPatterns(
RewritePatternSet &results, MLIRContext *context) {
// Pattern to rewrite a ExtractStridedSliceOp(ConstantMaskOp) ->
// ConstantMaskOp and ExtractStridedSliceOp(ConstantOp) -> ConstantOp.
results.add<StridedSliceConstantMaskFolder, StridedSliceBroadcast,
StridedSliceSplat, ContiguousExtractStridedSliceToExtract>(
context);
}
//===----------------------------------------------------------------------===//
// TransferReadOp
//===----------------------------------------------------------------------===//
/// 1. Builder that sets padding to zero and an empty mask (variant with attrs).
void TransferReadOp::build(OpBuilder &builder, OperationState &result,
VectorType vectorType, Value source,
ValueRange indices, AffineMapAttr permutationMapAttr,
/*optional*/ ArrayAttr inBoundsAttr) {
Type elemType = llvm::cast<ShapedType>(source.getType()).getElementType();
Value padding = builder.create<arith::ConstantOp>(
result.location, elemType, builder.getZeroAttr(elemType));
build(builder, result, vectorType, source, indices, permutationMapAttr,
padding, /*mask=*/Value(), inBoundsAttr);
}
/// 2. Builder that sets padding to zero an empty mask (variant without attrs).
void TransferReadOp::build(OpBuilder &builder, OperationState &result,
VectorType vectorType, Value source,
ValueRange indices, AffineMap permutationMap,
std::optional<ArrayRef<bool>> inBounds) {
auto permutationMapAttr = AffineMapAttr::get(permutationMap);
auto inBoundsAttr = (inBounds && !inBounds.value().empty())
? builder.getBoolArrayAttr(inBounds.value())
: builder.getBoolArrayAttr(
SmallVector<bool>(vectorType.getRank(), false));
build(builder, result, vectorType, source, indices, permutationMapAttr,
inBoundsAttr);
}
/// 3. Builder that sets permutation map to 'getMinorIdentityMap'.
void TransferReadOp::build(OpBuilder &builder, OperationState &result,
VectorType vectorType, Value source,
ValueRange indices, Value padding,
std::optional<ArrayRef<bool>> inBounds) {
AffineMap permutationMap = getTransferMinorIdentityMap(
llvm::cast<ShapedType>(source.getType()), vectorType);
auto permutationMapAttr = AffineMapAttr::get(permutationMap);
auto inBoundsAttr = (inBounds && !inBounds.value().empty())
? builder.getBoolArrayAttr(inBounds.value())
: builder.getBoolArrayAttr(
SmallVector<bool>(vectorType.getRank(), false));
build(builder, result, vectorType, source, indices, permutationMapAttr,
padding,
/*mask=*/Value(), inBoundsAttr);
}
/// 4. Builder that sets padding to zero and permutation map to
/// 'getMinorIdentityMap'.
void TransferReadOp::build(OpBuilder &builder, OperationState &result,
VectorType vectorType, Value source,
ValueRange indices,
std::optional<ArrayRef<bool>> inBounds) {
Type elemType = llvm::cast<ShapedType>(source.getType()).getElementType();
Value padding = builder.create<arith::ConstantOp>(
result.location, elemType, builder.getZeroAttr(elemType));
build(builder, result, vectorType, source, indices, padding, inBounds);
}
template <typename EmitFun>
static LogicalResult verifyPermutationMap(AffineMap permutationMap,
EmitFun emitOpError) {
SmallVector<bool, 8> seen(permutationMap.getNumInputs(), false);
for (auto expr : permutationMap.getResults()) {
auto dim = dyn_cast<AffineDimExpr>(expr);
auto zero = dyn_cast<AffineConstantExpr>(expr);
if (zero) {
if (zero.getValue() != 0) {
return emitOpError(
"requires a projected permutation_map (at most one dim or the zero "
"constant can appear in each result)");
}
continue;
}
if (!dim) {
return emitOpError("requires a projected permutation_map (at most one "
"dim or the zero constant can appear in each result)");
}
if (seen[dim.getPosition()]) {
return emitOpError(
"requires a permutation_map that is a permutation (found one dim "
"used more than once)");
}
seen[dim.getPosition()] = true;
}
return success();
}
static LogicalResult
verifyTransferOp(VectorTransferOpInterface op, ShapedType shapedType,
VectorType vectorType, VectorType maskType,
VectorType inferredMaskType, AffineMap permutationMap,
ArrayAttr inBounds) {
if (op->hasAttr("masked")) {
return op->emitOpError("masked attribute has been removed. "
"Use in_bounds instead.");
}
if (!llvm::isa<MemRefType, RankedTensorType>(shapedType))
return op->emitOpError(
"requires source to be a memref or ranked tensor type");
auto elementType = shapedType.getElementType();
DataLayout dataLayout = DataLayout::closest(op);
if (auto vectorElementType = llvm::dyn_cast<VectorType>(elementType)) {
// Memref or tensor has vector element type.
unsigned sourceVecSize =
dataLayout.getTypeSizeInBits(vectorElementType.getElementType()) *
vectorElementType.getShape().back();
unsigned resultVecSize =
dataLayout.getTypeSizeInBits(vectorType.getElementType()) *
vectorType.getShape().back();
if (resultVecSize % sourceVecSize != 0)
return op->emitOpError(
"requires the bitwidth of the minor 1-D vector to be an integral "
"multiple of the bitwidth of the minor 1-D vector of the source");
unsigned sourceVecEltRank = vectorElementType.getRank();
unsigned resultVecRank = vectorType.getRank();
if (sourceVecEltRank > resultVecRank)
return op->emitOpError(
"requires source vector element and vector result ranks to match.");
unsigned rankOffset = resultVecRank - sourceVecEltRank;
// Check that permutation map results match 'rankOffset' of vector type.
if (permutationMap.getNumResults() != rankOffset)
return op->emitOpError("requires a permutation_map with result dims of "
"the same rank as the vector type");
if (maskType)
return op->emitOpError("does not support masks with vector element type");
} else {
// Memref or tensor has scalar element type.
unsigned minorSize =
vectorType.getRank() == 0 ? 1 : vectorType.getShape().back();
unsigned resultVecSize =
dataLayout.getTypeSizeInBits(vectorType.getElementType()) * minorSize;
if (resultVecSize % dataLayout.getTypeSizeInBits(elementType) != 0)
return op->emitOpError(
"requires the bitwidth of the minor 1-D vector to be an integral "
"multiple of the bitwidth of the source element type");
// Check that permutation map results match rank of vector type.
if (permutationMap.getNumResults() != vectorType.getRank())
return op->emitOpError("requires a permutation_map with result dims of "
"the same rank as the vector type");
}
if (permutationMap.getNumSymbols() != 0)
return op->emitOpError("requires permutation_map without symbols");
if (permutationMap.getNumInputs() != shapedType.getRank())
return op->emitOpError("requires a permutation_map with input dims of the "
"same rank as the source type");
if (maskType && maskType != inferredMaskType)
return op->emitOpError("inferred mask type (")
<< inferredMaskType << ") and mask operand type (" << maskType
<< ") don't match";
if (permutationMap.getNumResults() != static_cast<int64_t>(inBounds.size()))
return op->emitOpError("expects the in_bounds attr of same rank "
"as permutation_map results: ")
<< AffineMapAttr::get(permutationMap)
<< " vs inBounds of size: " << inBounds.size();
return success();
}
static void printTransferAttrs(OpAsmPrinter &p, VectorTransferOpInterface op) {
SmallVector<StringRef, 3> elidedAttrs;
elidedAttrs.push_back(TransferReadOp::getOperandSegmentSizeAttr());
if (op.getPermutationMap().isMinorIdentity())
elidedAttrs.push_back(op.getPermutationMapAttrName());
// Elide in_bounds attribute if all dims are out-of-bounds.
if (llvm::none_of(op.getInBoundsValues(), [](bool b) { return b; }))
elidedAttrs.push_back(op.getInBoundsAttrName());
p.printOptionalAttrDict(op->getAttrs(), elidedAttrs);
}
void TransferReadOp::print(OpAsmPrinter &p) {
p << " " << getBase() << "[" << getIndices() << "], " << getPadding();
if (getMask())
p << ", " << getMask();
printTransferAttrs(p, *this);
p << " : " << getShapedType() << ", " << getVectorType();
}
VectorType mlir::vector::inferTransferOpMaskType(VectorType vecType,
AffineMap permMap) {
auto i1Type = IntegerType::get(permMap.getContext(), 1);
AffineMap invPermMap = inversePermutation(compressUnusedDims(permMap));
assert(invPermMap && "Inversed permutation map couldn't be computed");
SmallVector<int64_t, 8> maskShape = invPermMap.compose(vecType.getShape());
// The MaskOp specification doesn't support 0-D vectors at the moment. Turn a
// 0-D mask into a single-element 1-D mask.
if (maskShape.empty())
maskShape.push_back(1);
SmallVector<bool> scalableDims =
applyPermutationMap(invPermMap, vecType.getScalableDims());
return VectorType::get(maskShape, i1Type, scalableDims);
}
ParseResult TransferReadOp::parse(OpAsmParser &parser, OperationState &result) {
auto &builder = parser.getBuilder();
SMLoc typesLoc;
OpAsmParser::UnresolvedOperand sourceInfo;
SmallVector<OpAsmParser::UnresolvedOperand, 8> indexInfo;
OpAsmParser::UnresolvedOperand paddingInfo;
SmallVector<Type, 2> types;
OpAsmParser::UnresolvedOperand maskInfo;
// Parsing with support for paddingValue.
if (parser.parseOperand(sourceInfo) ||
parser.parseOperandList(indexInfo, OpAsmParser::Delimiter::Square) ||
parser.parseComma() || parser.parseOperand(paddingInfo))
return failure();
ParseResult hasMask = parser.parseOptionalComma();
if (hasMask.succeeded()) {
if (parser.parseOperand(maskInfo))
return failure();
}
if (parser.parseOptionalAttrDict(result.attributes) ||
parser.getCurrentLocation(&typesLoc) || parser.parseColonTypeList(types))
return failure();
if (types.size() != 2)
return parser.emitError(typesLoc, "requires two types");
auto indexType = builder.getIndexType();
auto shapedType = llvm::dyn_cast<ShapedType>(types[0]);
if (!shapedType || !llvm::isa<MemRefType, RankedTensorType>(shapedType))
return parser.emitError(typesLoc, "requires memref or ranked tensor type");
VectorType vectorType = llvm::dyn_cast<VectorType>(types[1]);
if (!vectorType)
return parser.emitError(typesLoc, "requires vector type");
auto permMapAttrName = TransferReadOp::getPermutationMapAttrName(result.name);
Attribute permMapAttr = result.attributes.get(permMapAttrName);
AffineMap permMap;
if (!permMapAttr) {
if (shapedType.getRank() <
getEffectiveVectorRankForXferOp(shapedType, vectorType))
return parser.emitError(typesLoc,
"expected a custom permutation_map when "
"rank(source) != rank(destination)");
permMap = getTransferMinorIdentityMap(shapedType, vectorType);
result.attributes.set(permMapAttrName, AffineMapAttr::get(permMap));
} else {
permMap = llvm::cast<AffineMapAttr>(permMapAttr).getValue();
}
auto inBoundsAttrName = TransferReadOp::getInBoundsAttrName(result.name);
Attribute inBoundsAttr = result.attributes.get(inBoundsAttrName);
if (!inBoundsAttr) {
result.addAttribute(inBoundsAttrName,
builder.getBoolArrayAttr(
SmallVector<bool>(permMap.getNumResults(), false)));
}
if (parser.resolveOperand(sourceInfo, shapedType, result.operands) ||
parser.resolveOperands(indexInfo, indexType, result.operands) ||
parser.resolveOperand(paddingInfo, shapedType.getElementType(),
result.operands))
return failure();
if (hasMask.succeeded()) {
if (llvm::dyn_cast<VectorType>(shapedType.getElementType()))
return parser.emitError(
maskInfo.location, "does not support masks with vector element type");
if (vectorType.getRank() != permMap.getNumResults()) {
return parser.emitError(typesLoc,
"expected the same rank for the vector and the "
"results of the permutation map");
}
// Instead of adding the mask type as an op type, compute it based on the
// vector type and the permutation map (to keep the type signature small).
auto maskType = inferTransferOpMaskType(vectorType, permMap);
if (parser.resolveOperand(maskInfo, maskType, result.operands))
return failure();
}
result.addAttribute(TransferReadOp::getOperandSegmentSizeAttr(),
builder.getDenseI32ArrayAttr(
{1, static_cast<int32_t>(indexInfo.size()), 1,
static_cast<int32_t>(hasMask.succeeded())}));
return parser.addTypeToList(vectorType, result.types);
}
LogicalResult TransferReadOp::verify() {
// Consistency of elemental types in source and vector.
ShapedType shapedType = getShapedType();
VectorType vectorType = getVectorType();
VectorType maskType = getMaskType();
auto paddingType = getPadding().getType();
auto permutationMap = getPermutationMap();
VectorType inferredMaskType =
maskType ? inferTransferOpMaskType(vectorType, permutationMap)
: VectorType();
auto sourceElementType = shapedType.getElementType();
if (static_cast<int64_t>(getIndices().size()) != shapedType.getRank())
return emitOpError("requires ") << shapedType.getRank() << " indices";
if (failed(verifyTransferOp(cast<VectorTransferOpInterface>(getOperation()),
shapedType, vectorType, maskType,
inferredMaskType, permutationMap, getInBounds())))
return failure();
if (auto sourceVectorElementType =
llvm::dyn_cast<VectorType>(sourceElementType)) {
// Source has vector element type.
// Check that 'sourceVectorElementType' and 'paddingType' types match.
if (sourceVectorElementType != paddingType)
return emitOpError(
"requires source element type and padding type to match.");
} else {
// Check that 'paddingType' is valid to store in a vector type.
if (!VectorType::isValidElementType(paddingType))
return emitOpError("requires valid padding vector elemental type");
// Check that padding type and vector element types match.
if (paddingType != sourceElementType)
return emitOpError(
"requires formal padding and source of the same elemental type");
}
return verifyPermutationMap(permutationMap,
[&](Twine t) { return emitOpError(t); });
}
// MaskableOpInterface methods.
/// Returns the mask type expected by this operation. Mostly used for
/// verification purposes. It requires the operation to be vectorized."
Type TransferReadOp::getExpectedMaskType() {
return inferTransferOpMaskType(getVectorType(), getPermutationMap());
}
//===----------------------------------------------------------------------===//
// TransferReadOp: VectorTransferOpInterface methods.
//===----------------------------------------------------------------------===//
VectorType TransferReadOp::getVectorType() {
return cast<VectorType>(getVector().getType());
}
template <typename TransferOp>
static bool isInBounds(TransferOp op, int64_t resultIdx, int64_t indicesIdx) {
// TODO: support more aggressive createOrFold on:
// op.getIndices()[indicesIdx] + vectorType < dim(op.getSource(), indicesIdx)
if (op.getShapedType().isDynamicDim(indicesIdx))
return false;
Value index = op.getIndices()[indicesIdx];
std::optional<int64_t> cstOp = getConstantIntValue(index);
if (!cstOp.has_value())
return false;
int64_t sourceSize = op.getShapedType().getDimSize(indicesIdx);
int64_t vectorSize = op.getVectorType().getDimSize(resultIdx);
return cstOp.value() + vectorSize <= sourceSize;
}
template <typename TransferOp>
static LogicalResult foldTransferInBoundsAttribute(TransferOp op) {
// TODO: support 0-d corner case.
// TODO: Be less conservative.
if (op.getTransferRank() == 0)
return failure();
AffineMap permutationMap = op.getPermutationMap();
bool changed = false;
SmallVector<bool, 4> newInBounds;
newInBounds.reserve(op.getTransferRank());
// Idxs of non-bcast dims - used when analysing bcast dims.
SmallVector<unsigned> nonBcastDims;
// 1. Process non-broadcast dims
for (unsigned i = 0; i < op.getTransferRank(); ++i) {
// 1.1. Already marked as in-bounds, nothing to see here.
if (op.isDimInBounds(i)) {
newInBounds.push_back(true);
continue;
}
// 1.2. Currently out-of-bounds, check whether we can statically determine
// it is inBounds.
bool inBounds = false;
auto dimExpr = dyn_cast<AffineDimExpr>(permutationMap.getResult(i));
if (dimExpr) {
inBounds = isInBounds(op, /*resultIdx=*/i,
/*indicesIdx=*/dimExpr.getPosition());
nonBcastDims.push_back(i);
}
newInBounds.push_back(inBounds);
// We commit the pattern if it is "more inbounds".
changed |= inBounds;
}
// 2. Handle broadcast dims
// If all non-broadcast dims are "in bounds", then all bcast dims should be
// "in bounds" as well.
bool allNonBcastDimsInBounds = llvm::all_of(
nonBcastDims, [&newInBounds](unsigned idx) { return newInBounds[idx]; });
if (allNonBcastDimsInBounds) {
for (size_t idx : permutationMap.getBroadcastDims()) {
changed |= !newInBounds[idx];
newInBounds[idx] = true;
}
}
if (!changed)
return failure();
// OpBuilder is only used as a helper to build an I64ArrayAttr.
OpBuilder b(op.getContext());
op.setInBoundsAttr(b.getBoolArrayAttr(newInBounds));
return success();
}
template <typename TransferOp>
static LogicalResult foldTransferFullMask(TransferOp op) {
auto mask = op.getMask();
if (!mask)
return failure();
if (getMaskFormat(mask) != MaskFormat::AllTrue)
return failure();
op.getMaskMutable().clear();
return success();
}
/// ```
/// %w0 = vector.transfer_write %v0, %arg0[%c1, %c0] {in_bounds = [true, true]}
/// : vector<1x4xf32>, tensor<4x4xf32>
/// %0 = vector.transfer_read %w0[%c1, %c0], %cf0 {in_bounds = [true, true]}
/// : tensor<4x4xf32>, vector<1x4xf32>
/// ```
/// -> Folds into
/// ```
/// %v0
/// ```
static Value foldRAW(TransferReadOp readOp) {
if (!llvm::isa<RankedTensorType>(readOp.getShapedType()))
return {};
auto defWrite = readOp.getBase().getDefiningOp<vector::TransferWriteOp>();
while (defWrite) {
if (checkSameValueRAW(defWrite, readOp))
return defWrite.getVector();
if (!isDisjointTransferIndices(
cast<VectorTransferOpInterface>(defWrite.getOperation()),
cast<VectorTransferOpInterface>(readOp.getOperation())))
break;
defWrite = defWrite.getBase().getDefiningOp<vector::TransferWriteOp>();
}
return {};
}
OpFoldResult TransferReadOp::fold(FoldAdaptor) {
if (Value vec = foldRAW(*this))
return vec;
/// transfer_read(memrefcast) -> transfer_read
if (succeeded(foldTransferInBoundsAttribute(*this)))
return getResult();
if (succeeded(foldTransferFullMask(*this)))
return getResult();
if (succeeded(memref::foldMemRefCast(*this)))
return getResult();
if (succeeded(tensor::foldTensorCast(*this)))
return getResult();
return OpFoldResult();
}
std::optional<SmallVector<int64_t, 4>> TransferReadOp::getShapeForUnroll() {
return llvm::to_vector<4>(getVectorType().getShape());
}
void TransferReadOp::getEffects(
SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>>
&effects) {
if (llvm::isa<MemRefType>(getShapedType()))
effects.emplace_back(MemoryEffects::Read::get(), &getBaseMutable(),
SideEffects::DefaultResource::get());
}
Speculation::Speculatability TransferReadOp::getSpeculatability() {
if (hasPureTensorSemantics())
return Speculation::Speculatable;
return Speculation::NotSpeculatable;
}
namespace {
/// Store to load forwarding for transfer operations with permuation maps.
/// Even if the permutation maps are different we can still propagate the store
/// into the load if the size of the dimensions read and written match. Then we
/// can replace the transfer_read + transfer_write by vector.broadcast and
/// vector.transpose.
/// Example:
/// ```
/// %w0 = vector.transfer_write %v0, %arg0[%c0, %c0, %c0]
/// {in_bounds = [true, true],
/// permutation_map = affine_map<(d0, d1, d2) -> (d2, d1)>} :
/// vector<4x1xf32>, tensor<4x4x4xf32>
/// %r = vector.transfer_read %w0[%c0, %c0, %c0], %cf0
/// {in_bounds = [true, true, true, true],
/// permutation_map = affine_map<(d0, d1, d2) -> (d1, 0, d2, 0)>} :
/// tensor<4x4x4xf32>, vector<1x100x4x5xf32>
/// ```
/// To:
/// ```
/// %0 = vector.broadcast %arg1 : vector<4x1xf32> to vector<100x5x4x1xf32>
/// %r = vector.transpose %0, [3, 0, 2, 1] :
/// vector<100x5x4x1xf32> to vector<1x100x4x5xf32>
/// ```
struct TransferReadAfterWriteToBroadcast
: public OpRewritePattern<TransferReadOp> {
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(TransferReadOp readOp,
PatternRewriter &rewriter) const override {
if (readOp.hasOutOfBoundsDim() ||
!llvm::isa<RankedTensorType>(readOp.getShapedType()))
return failure();
auto defWrite = readOp.getBase().getDefiningOp<vector::TransferWriteOp>();
if (!defWrite)
return failure();
// TODO: If the written transfer chunk is a superset of the read transfer
// chunk we could do an extract_strided_slice.
if (readOp.getTransferChunkAccessed() !=
defWrite.getTransferChunkAccessed())
return failure();
// TODO: Support cases where a dim is explicitly written but implicitly
// read (i.e., a unit dim that is rank reduced).
if (getUnusedDimsBitVector({readOp.getPermutationMap()}) !=
getUnusedDimsBitVector({defWrite.getPermutationMap()}))
return failure();
if (readOp.getIndices() != defWrite.getIndices() ||
readOp.getMask() != defWrite.getMask())
return failure();
Value vec = defWrite.getVector();
// TODO: loop through the chain of transfer_write if we can prove that they
// don't overlap with the transfer_read. This requires improving
// `isDisjointTransferIndices` helper.
AffineMap readMap = compressUnusedDims(readOp.getPermutationMap());
AffineMap writeMap = compressUnusedDims(defWrite.getPermutationMap());
AffineMap map = readMap.compose(writeMap);
if (map.getNumResults() == 0)
return failure();
// Calculate the permutation to apply to go from the vector stored to the
// vector read.
SmallVector<unsigned> permutation;
if (!map.isPermutationOfMinorIdentityWithBroadcasting(permutation))
return failure();
Location loc = readOp.getLoc();
// Calculate the broadcast shape by applying the reverse permutation to the
// final shape we want.
ArrayRef<int64_t> destShape = readOp.getVectorType().getShape();
SmallVector<int64_t> broadcastShape(destShape.size());
SmallVector<bool> broadcastScalableFlags(destShape.size());
for (const auto &pos : llvm::enumerate(permutation)) {
broadcastShape[pos.value()] = destShape[pos.index()];
broadcastScalableFlags[pos.value()] =
readOp.getVectorType().getScalableDims()[pos.index()];
}
VectorType broadcastedType = VectorType::get(
broadcastShape, defWrite.getVectorType().getElementType(),
broadcastScalableFlags);
vec = rewriter.create<vector::BroadcastOp>(loc, broadcastedType, vec);
SmallVector<int64_t> transposePerm(permutation.begin(), permutation.end());
rewriter.replaceOpWithNewOp<vector::TransposeOp>(readOp, vec,
transposePerm);
return success();
}
};
} // namespace
void TransferReadOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<TransferReadAfterWriteToBroadcast>(context);
}
//===----------------------------------------------------------------------===//
// TransferWriteOp
//===----------------------------------------------------------------------===//
/// 1. Builder with type inference.
void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
Value vector, Value dest, ValueRange indices,
AffineMapAttr permutationMapAttr,
/*optional*/ Value mask,
/*optional*/ ArrayAttr inBoundsAttr) {
Type resultType = llvm::dyn_cast<RankedTensorType>(dest.getType());
build(builder, result, resultType, vector, dest, indices, permutationMapAttr,
mask, inBoundsAttr);
}
/// 2. Builder with type inference that sets an empty mask (variant with attrs).
void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
Value vector, Value dest, ValueRange indices,
AffineMapAttr permutationMapAttr,
/*optional*/ ArrayAttr inBoundsAttr) {
build(builder, result, vector, dest, indices, permutationMapAttr,
/*mask=*/Value(), inBoundsAttr);
}
/// 3. Builder with type inference that sets an empty mask (variant without
/// attrs)
void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
Value vector, Value dest, ValueRange indices,
AffineMap permutationMap,
std::optional<ArrayRef<bool>> inBounds) {
auto permutationMapAttr = AffineMapAttr::get(permutationMap);
auto inBoundsAttr =
(inBounds && !inBounds.value().empty())
? builder.getBoolArrayAttr(inBounds.value())
: builder.getBoolArrayAttr(SmallVector<bool>(
llvm::cast<VectorType>(vector.getType()).getRank(), false));
build(builder, result, vector, dest, indices, permutationMapAttr,
/*mask=*/Value(), inBoundsAttr);
}
/// 4. Builder with type inference that sets an empty mask and sets permutation
/// map to 'getMinorIdentityMap'.
void TransferWriteOp::build(OpBuilder &builder, OperationState &result,
Value vector, Value dest, ValueRange indices,
std::optional<ArrayRef<bool>> inBounds) {
auto vectorType = llvm::cast<VectorType>(vector.getType());
AffineMap permutationMap = getTransferMinorIdentityMap(
llvm::cast<ShapedType>(dest.getType()), vectorType);
build(builder, result, vector, dest, indices, permutationMap, inBounds);
}
ParseResult TransferWriteOp::parse(OpAsmParser &parser,
OperationState &result) {
auto &builder = parser.getBuilder();
SMLoc typesLoc;
OpAsmParser::UnresolvedOperand vectorInfo, sourceInfo;
SmallVector<OpAsmParser::UnresolvedOperand, 8> indexInfo;
SmallVector<Type, 2> types;
OpAsmParser::UnresolvedOperand maskInfo;
if (parser.parseOperand(vectorInfo) || parser.parseComma() ||
parser.parseOperand(sourceInfo) ||
parser.parseOperandList(indexInfo, OpAsmParser::Delimiter::Square))
return failure();
ParseResult hasMask = parser.parseOptionalComma();
if (hasMask.succeeded() && parser.parseOperand(maskInfo))
return failure();
if (parser.parseOptionalAttrDict(result.attributes) ||
parser.getCurrentLocation(&typesLoc) || parser.parseColonTypeList(types))
return failure();
if (types.size() != 2)
return parser.emitError(typesLoc, "requires two types");
auto indexType = builder.getIndexType();
VectorType vectorType = llvm::dyn_cast<VectorType>(types[0]);
if (!vectorType)
return parser.emitError(typesLoc, "requires vector type");
ShapedType shapedType = llvm::dyn_cast<ShapedType>(types[1]);
if (!shapedType || !llvm::isa<MemRefType, RankedTensorType>(shapedType))
return parser.emitError(typesLoc, "requires memref or ranked tensor type");
auto permMapAttrName =
TransferWriteOp::getPermutationMapAttrName(result.name);
auto permMapAttr = result.attributes.get(permMapAttrName);
AffineMap permMap;
if (!permMapAttr) {
if (shapedType.getRank() <
getEffectiveVectorRankForXferOp(shapedType, vectorType))
return parser.emitError(typesLoc,
"expected a custom permutation_map when "
"rank(source) != rank(destination)");
permMap = getTransferMinorIdentityMap(shapedType, vectorType);
result.attributes.set(permMapAttrName, AffineMapAttr::get(permMap));
} else {
permMap = llvm::cast<AffineMapAttr>(permMapAttr).getValue();
}
auto inBoundsAttrName = TransferWriteOp::getInBoundsAttrName(result.name);
Attribute inBoundsAttr = result.attributes.get(inBoundsAttrName);
if (!inBoundsAttr) {
result.addAttribute(inBoundsAttrName,
builder.getBoolArrayAttr(
SmallVector<bool>(permMap.getNumResults(), false)));
}
if (parser.resolveOperand(vectorInfo, vectorType, result.operands) ||
parser.resolveOperand(sourceInfo, shapedType, result.operands) ||
parser.resolveOperands(indexInfo, indexType, result.operands))
return failure();
if (hasMask.succeeded()) {
if (llvm::dyn_cast<VectorType>(shapedType.getElementType()))
return parser.emitError(
maskInfo.location, "does not support masks with vector element type");
if (vectorType.getRank() != permMap.getNumResults()) {
return parser.emitError(typesLoc,
"expected the same rank for the vector and the "
"results of the permutation map");
}
auto maskType = inferTransferOpMaskType(vectorType, permMap);
if (parser.resolveOperand(maskInfo, maskType, result.operands))
return failure();
}
result.addAttribute(TransferWriteOp::getOperandSegmentSizeAttr(),
builder.getDenseI32ArrayAttr(
{1, 1, static_cast<int32_t>(indexInfo.size()),
static_cast<int32_t>(hasMask.succeeded())}));
return failure(llvm::isa<RankedTensorType>(shapedType) &&
parser.addTypeToList(shapedType, result.types));
}
void TransferWriteOp::print(OpAsmPrinter &p) {
p << " " << getVector() << ", " << getBase() << "[" << getIndices() << "]";
if (getMask())
p << ", " << getMask();
printTransferAttrs(p, *this);
p << " : " << getVectorType() << ", " << getShapedType();
}
LogicalResult TransferWriteOp::verify() {
// Consistency of elemental types in shape and vector.
ShapedType shapedType = getShapedType();
VectorType vectorType = getVectorType();
VectorType maskType = getMaskType();
auto permutationMap = getPermutationMap();
VectorType inferredMaskType =
maskType ? inferTransferOpMaskType(vectorType, permutationMap)
: VectorType();
if (llvm::size(getIndices()) != shapedType.getRank())
return emitOpError("requires ") << shapedType.getRank() << " indices";
// We do not allow broadcast dimensions on TransferWriteOps for the moment,
// as the semantics is unclear. This can be revisited later if necessary.
if (hasBroadcastDim())
return emitOpError("should not have broadcast dimensions");
if (failed(verifyTransferOp(cast<VectorTransferOpInterface>(getOperation()),
shapedType, vectorType, maskType,
inferredMaskType, permutationMap, getInBounds())))
return failure();
return verifyPermutationMap(permutationMap,
[&](Twine t) { return emitOpError(t); });
}
//===----------------------------------------------------------------------===//
// TransferWriteOp: MaskableOpInterface methods.
//===----------------------------------------------------------------------===//
/// Returns the mask type expected by this operation. Mostly used for
/// verification purposes.
Type TransferWriteOp::getExpectedMaskType() {
return inferTransferOpMaskType(getVectorType(), getPermutationMap());
}
//===----------------------------------------------------------------------===//
// TransferWriteOp: VectorTransferOpInterface methods.
//===----------------------------------------------------------------------===//
Value TransferWriteOp::getVector() { return getOperand(0); }
VectorType TransferWriteOp::getVectorType() {
return cast<VectorType>(getValueToStore().getType());
}
//===----------------------------------------------------------------------===//
// TransferWriteOp: fold methods.
//===----------------------------------------------------------------------===//
/// Fold:
/// ```
/// %t1 = ...
/// %v = vector.transfer_read %t0[%c0...], {in_bounds = [true...]} :
/// tensor<static_sizesxf32>, vector<static_sizesxf32>
/// %t2 = vector.transfer_write %v, %t1[%c0...] {in_bounds = [true...]} :
/// vector<static_sizesxf32>, tensor<static_sizesxf32>
/// ```
///
/// into:
///
/// ```
/// %t0
/// ```
///
/// The producer of t1 may or may not be DCE'd depending on whether it is a
/// block argument or has side effects.
static LogicalResult foldReadInitWrite(TransferWriteOp write,
ArrayRef<Attribute>,
SmallVectorImpl<OpFoldResult> &results) {
// TODO: support 0-d corner case.
if (write.getTransferRank() == 0)
return failure();
auto rankedTensorType =
llvm::dyn_cast<RankedTensorType>(write.getBase().getType());
// If not operating on tensors, bail.
if (!rankedTensorType)
return failure();
// If no read, bail.
auto read = write.getVector().getDefiningOp<vector::TransferReadOp>();
if (!read)
return failure();
// TODO: support 0-d corner case.
if (read.getTransferRank() == 0)
return failure();
// For now, only accept minor identity. Future: composition is minor identity.
if (!read.getPermutationMap().isMinorIdentity() ||
!write.getPermutationMap().isMinorIdentity())
return failure();
// Bail on mismatching ranks.
if (read.getTransferRank() != write.getTransferRank())
return failure();
// Bail on potential out-of-bounds accesses.
if (read.hasOutOfBoundsDim() || write.hasOutOfBoundsDim())
return failure();
// Tensor types must be the same.
if (read.getBase().getType() != rankedTensorType)
return failure();
// Vector types must be the same.
if (read.getVectorType() != write.getVectorType())
return failure();
// Vector and Tensor shapes must match.
if (read.getVectorType().getShape() != rankedTensorType.getShape())
return failure();
// If any index is nonzero.
auto isNotConstantZero = [](Value v) {
auto cstOp = getConstantIntValue(v);
return !cstOp.has_value() || cstOp.value() != 0;
};
if (llvm::any_of(read.getIndices(), isNotConstantZero) ||
llvm::any_of(write.getIndices(), isNotConstantZero))
return failure();
// Success.
results.push_back(read.getBase());
return success();
}
static bool checkSameValueWAR(vector::TransferReadOp read,
vector::TransferWriteOp write) {
return read.getBase() == write.getBase() &&
read.getIndices() == write.getIndices() &&
read.getPermutationMap() == write.getPermutationMap() &&
read.getVectorType() == write.getVectorType() && !read.getMask() &&
!write.getMask();
}
/// Fold transfer_write write after read:
/// ```
/// %t0 = ...
/// %v = vector.transfer_read %t0[%c0...] :
/// tensor<static_sizesxf32>, vector<static_sizesxf32>
/// %t1 = vector.transfer_write %v, %t0[%c0...] :
/// vector<static_sizesxf32>, tensor<static_sizesxf32>
/// ```
///
/// into:
///
/// ```
/// %t0
/// ```
static LogicalResult foldWAR(TransferWriteOp write,
SmallVectorImpl<OpFoldResult> &results) {
if (!llvm::isa<RankedTensorType>(write.getBase().getType()))
return failure();
auto read = write.getVector().getDefiningOp<vector::TransferReadOp>();
if (!read)
return failure();
if (!checkSameValueWAR(read, write))
return failure();
results.push_back(read.getBase());
return success();
}
LogicalResult TransferWriteOp::fold(FoldAdaptor adaptor,
SmallVectorImpl<OpFoldResult> &results) {
if (succeeded(foldReadInitWrite(*this, adaptor.getOperands(), results)))
return success();
if (succeeded(foldWAR(*this, results)))
return success();
if (succeeded(foldTransferInBoundsAttribute(*this)))
return success();
if (succeeded(foldTransferFullMask(*this)))
return success();
return memref::foldMemRefCast(*this);
}
//===----------------------------------------------------------------------===//
// TransferWriteOp: other methods.
//===----------------------------------------------------------------------===//
std::optional<SmallVector<int64_t, 4>> TransferWriteOp::getShapeForUnroll() {
return llvm::to_vector<4>(getVectorType().getShape());
}
void TransferWriteOp::getEffects(
SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>>
&effects) {
if (llvm::isa<MemRefType>(getShapedType()))
effects.emplace_back(MemoryEffects::Write::get(), &getValueToStoreMutable(),
SideEffects::DefaultResource::get());
}
Speculation::Speculatability TransferWriteOp::getSpeculatability() {
if (hasPureTensorSemantics())
return Speculation::Speculatable;
return Speculation::NotSpeculatable;
}
namespace {
/// Remove dead transfer write from the SSA chain so that it an be eliminated by
/// DCE
/// ```
/// %w0 = vector.transfer_write %v0, %arg0[%c1, %c0] {in_bounds = [true, true]}
/// : vector<1x4xf32>, tensor<4x4xf32>
/// %w1 = vector.transfer_write %v0, %w0[%c2, %c0] {in_bounds = [true, true]}
/// : vector<1x4xf32>, tensor<4x4xf32>
/// %w2 = vector.transfer_write %v1, %w1[%c1, %c0] {in_bounds = [true, true]}
/// : vector<1x4xf32>, tensor<4x4xf32>
/// ```
///
/// into:
///
/// ```
/// %w0 = vector.transfer_write %v0, %arg0[%c1, %c0] {in_bounds = [true, true]}
/// : vector<1x4xf32>, tensor<4x4xf32>
/// %w1 = vector.transfer_write %v0, %arg0[%c2, %c0] {in_bounds = [true, true]}
/// : vector<1x4xf32>, tensor<4x4xf32>
/// %w2 = vector.transfer_write %v1, %w1[%c1, %c0] {in_bounds = [true, true]}
/// : vector<1x4xf32>, tensor<4x4xf32>
/// ```
///
/// `%w0 = vector.transfer_write` op will be removed by DCE if it doesn't have
/// any other uses.
class FoldWaw final : public OpRewritePattern<TransferWriteOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(TransferWriteOp writeOp,
PatternRewriter &rewriter) const override {
if (!llvm::isa<RankedTensorType>(writeOp.getShapedType()))
return failure();
vector::TransferWriteOp writeToModify = writeOp;
auto defWrite = writeOp.getBase().getDefiningOp<vector::TransferWriteOp>();
while (defWrite) {
if (checkSameValueWAW(writeOp, defWrite)) {
rewriter.modifyOpInPlace(writeToModify, [&]() {
writeToModify.getBaseMutable().assign(defWrite.getBase());
});
return success();
}
if (!isDisjointTransferIndices(
cast<VectorTransferOpInterface>(defWrite.getOperation()),
cast<VectorTransferOpInterface>(writeOp.getOperation())))
break;
// If the previous write op doesn't have any other use we an safely look
// at the previous store to see if it can be removed.
if (!defWrite->hasOneUse())
break;
writeToModify = defWrite;
defWrite = defWrite.getBase().getDefiningOp<vector::TransferWriteOp>();
}
return failure();
}
};
/// Rewrite tensor::ExtractSliceOp(vector::TransferWriteOp) to
/// vector::TransferWriteOp(tensor::ExtractSliceOp) if the full slice is
/// overwritten and inserted into another tensor. After this rewrite, the
/// operations bufferize in-place since all of them work on the same slice.
///
/// For example:
/// ```mlir
/// %0 = vector.transfer_write %vec, %init_tensor[%c0, %c0]
/// : vector<8x16xf32>, tensor<8x16xf32>
/// %1 = tensor.extract_slice %0[0, 0] [%sz0, %sz1] [1, 1]
/// : tensor<8x16xf32> to tensor<?x?xf32>
/// %r = tensor.insert_slice %1 into %iter_arg[%iv0, %iv1] [%sz0, %sz1] [1, 1]
/// : tensor<?x?xf32> into tensor<27x37xf32>
/// ```
/// folds to
/// ```mlir
/// %0 = tensor.extract_slice %iter_arg[%iv0, %iv1] [%sz0, %sz1] [1, 1]
/// : tensor<27x37xf32> to tensor<?x?xf32>
/// %1 = vector.transfer_write %vec, %0[%c0, %c0]
/// : vector<8x16xf32>, tensor<?x?xf32>
/// %r = tensor.insert_slice %1 into %iter_arg[%iv0, %iv1] [%sz0, %sz1] [1, 1]
/// : tensor<?x?xf32> into tensor<27x37xf32>
/// ```
struct SwapExtractSliceOfTransferWrite
: public OpRewritePattern<tensor::InsertSliceOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(tensor::InsertSliceOp insertOp,
PatternRewriter &rewriter) const override {
if (!insertOp.hasUnitStride())
return failure();
auto extractOp =
insertOp.getSource().getDefiningOp<tensor::ExtractSliceOp>();
if (!extractOp || !extractOp.hasUnitStride() || !extractOp->hasOneUse())
return failure();
auto transferOp = extractOp.getSource().getDefiningOp<TransferWriteOp>();
if (!transferOp || !transferOp->hasOneUse())
return failure();
// Fail if vector::TransferWriteOp or tensor::ExtractSliceOp is
// rank-reducing.
if (insertOp.getSourceType().getRank() != transferOp.getTransferRank()) {
return rewriter.notifyMatchFailure(insertOp,
"use-def chain is rank-reducing");
}
// Fail if tensor::ExtractSliceOp has non-zero offset.
if (!extractOp.hasZeroOffset()) {
return rewriter.notifyMatchFailure(insertOp,
"ExtractSliceOp has non-zero offset");
}
// Fail if tensor::TransferWriteOp has non-zero offset.
if (!llvm::all_of(transferOp.getIndices(), [](Value value) {
return getConstantIntValue(value) == static_cast<int64_t>(0);
})) {
return rewriter.notifyMatchFailure(insertOp,
"TranferWriteOp has non-zero offset");
}
// Fail if tensor::ExtractSliceOp and tensor::InsertSliceOp sizes differ.
if (insertOp.getMixedSizes().size() != extractOp.getMixedSizes().size()) {
return rewriter.notifyMatchFailure(
insertOp, "InsertSliceOp and ExtractSliceOp ranks differ");
}
for (auto [insertSize, extractSize] :
llvm::zip_equal(insertOp.getMixedSizes(), extractOp.getMixedSizes())) {
if (!isEqualConstantIntOrValue(insertSize, extractSize)) {
return rewriter.notifyMatchFailure(
insertOp, "InsertSliceOp and ExtractSliceOp sizes differ");
}
}
// Fail if the vector::TransferWriteOp may not overwrite the full tensor.
assert(transferOp.getVectorType().hasStaticShape() &&
"expected vector to have a static shape");
ArrayRef<int64_t> vectorShape = transferOp.getVectorType().getShape();
SmallVector<int64_t> resultShape = applyPermutationMap(
transferOp.getPermutationMap(), transferOp.getShapedType().getShape());
if (transferOp.getMask() || !vectorShape.equals(resultShape)) {
return rewriter.notifyMatchFailure(
insertOp, "TransferWriteOp may not write the full tensor.");
}
// Swap the tensor::ExtractSliceOp in front of the vector::TransferWriteOp.
// Set all in_bounds to false and let the folder infer them.
SmallVector<bool> newInBounds(vectorShape.size(), false);
auto newExtractOp = rewriter.create<tensor::ExtractSliceOp>(
extractOp.getLoc(), insertOp.getSourceType(), insertOp.getDest(),
insertOp.getMixedOffsets(), insertOp.getMixedSizes(),
insertOp.getMixedStrides());
auto newTransferWriteOp = rewriter.create<TransferWriteOp>(
transferOp.getLoc(), transferOp.getVector(), newExtractOp.getResult(),
transferOp.getIndices(), transferOp.getPermutationMapAttr(),
rewriter.getBoolArrayAttr(newInBounds));
rewriter.modifyOpInPlace(insertOp, [&]() {
insertOp.getSourceMutable().assign(newTransferWriteOp.getResult());
});
return success();
}
};
} // namespace
void TransferWriteOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<FoldWaw, SwapExtractSliceOfTransferWrite>(context);
}
//===----------------------------------------------------------------------===//
// LoadOp
//===----------------------------------------------------------------------===//
static LogicalResult verifyLoadStoreMemRefLayout(Operation *op,
VectorType vecTy,
MemRefType memRefTy) {
// If rank==0 or size==1 it's equivalent to scalar load/store, so we don't
// need any strides limitations.
if (!vecTy.isScalable() &&
(vecTy.getRank() == 0 || vecTy.getNumElements() == 1))
return success();
if (!memRefTy.isLastDimUnitStride())
return op->emitOpError("most minor memref dim must have unit stride");
return success();
}
LogicalResult vector::LoadOp::verify() {
VectorType resVecTy = getVectorType();
MemRefType memRefTy = getMemRefType();
if (failed(verifyLoadStoreMemRefLayout(*this, resVecTy, memRefTy)))
return failure();
if (memRefTy.getRank() < resVecTy.getRank())
return emitOpError(
"destination memref has lower rank than the result vector");
// Checks for vector memrefs.
Type memElemTy = memRefTy.getElementType();
if (auto memVecTy = llvm::dyn_cast<VectorType>(memElemTy)) {
if (memVecTy != resVecTy)
return emitOpError("base memref and result vector types should match");
memElemTy = memVecTy.getElementType();
}
if (resVecTy.getElementType() != memElemTy)
return emitOpError("base and result element types should match");
if (llvm::size(getIndices()) != memRefTy.getRank())
return emitOpError("requires ") << memRefTy.getRank() << " indices";
return success();
}
OpFoldResult LoadOp::fold(FoldAdaptor) {
if (succeeded(memref::foldMemRefCast(*this)))
return getResult();
return OpFoldResult();
}
//===----------------------------------------------------------------------===//
// StoreOp
//===----------------------------------------------------------------------===//
LogicalResult vector::StoreOp::verify() {
VectorType valueVecTy = getVectorType();
MemRefType memRefTy = getMemRefType();
if (failed(verifyLoadStoreMemRefLayout(*this, valueVecTy, memRefTy)))
return failure();
if (memRefTy.getRank() < valueVecTy.getRank())
return emitOpError("source memref has lower rank than the vector to store");
// Checks for vector memrefs.
Type memElemTy = memRefTy.getElementType();
if (auto memVecTy = llvm::dyn_cast<VectorType>(memElemTy)) {
if (memVecTy != valueVecTy)
return emitOpError(
"base memref and valueToStore vector types should match");
memElemTy = memVecTy.getElementType();
}
if (valueVecTy.getElementType() != memElemTy)
return emitOpError("base and valueToStore element type should match");
if (llvm::size(getIndices()) != memRefTy.getRank())
return emitOpError("requires ") << memRefTy.getRank() << " indices";
return success();
}
LogicalResult StoreOp::fold(FoldAdaptor adaptor,
SmallVectorImpl<OpFoldResult> &results) {
return memref::foldMemRefCast(*this);
}
//===----------------------------------------------------------------------===//
// MaskedLoadOp
//===----------------------------------------------------------------------===//
LogicalResult MaskedLoadOp::verify() {
VectorType maskVType = getMaskVectorType();
VectorType passVType = getPassThruVectorType();
VectorType resVType = getVectorType();
MemRefType memType = getMemRefType();
if (resVType.getElementType() != memType.getElementType())
return emitOpError("base and result element type should match");
if (llvm::size(getIndices()) != memType.getRank())
return emitOpError("requires ") << memType.getRank() << " indices";
if (resVType.getShape() != maskVType.getShape())
return emitOpError("expected result shape to match mask shape");
if (resVType != passVType)
return emitOpError("expected pass_thru of same type as result type");
return success();
}
namespace {
class MaskedLoadFolder final : public OpRewritePattern<MaskedLoadOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(MaskedLoadOp load,
PatternRewriter &rewriter) const override {
switch (getMaskFormat(load.getMask())) {
case MaskFormat::AllTrue:
rewriter.replaceOpWithNewOp<vector::LoadOp>(
load, load.getType(), load.getBase(), load.getIndices());
return success();
case MaskFormat::AllFalse:
rewriter.replaceOp(load, load.getPassThru());
return success();
case MaskFormat::Unknown:
return failure();
}
llvm_unreachable("Unexpected 1DMaskFormat on MaskedLoad");
}
};
} // namespace
void MaskedLoadOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<MaskedLoadFolder>(context);
}
OpFoldResult MaskedLoadOp::fold(FoldAdaptor) {
if (succeeded(memref::foldMemRefCast(*this)))
return getResult();
return OpFoldResult();
}
//===----------------------------------------------------------------------===//
// MaskedStoreOp
//===----------------------------------------------------------------------===//
LogicalResult MaskedStoreOp::verify() {
VectorType maskVType = getMaskVectorType();
VectorType valueVType = getVectorType();
MemRefType memType = getMemRefType();
if (valueVType.getElementType() != memType.getElementType())
return emitOpError("base and valueToStore element type should match");
if (llvm::size(getIndices()) != memType.getRank())
return emitOpError("requires ") << memType.getRank() << " indices";
if (valueVType.getShape() != maskVType.getShape())
return emitOpError("expected valueToStore shape to match mask shape");
return success();
}
namespace {
class MaskedStoreFolder final : public OpRewritePattern<MaskedStoreOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(MaskedStoreOp store,
PatternRewriter &rewriter) const override {
switch (getMaskFormat(store.getMask())) {
case MaskFormat::AllTrue:
rewriter.replaceOpWithNewOp<vector::StoreOp>(
store, store.getValueToStore(), store.getBase(), store.getIndices());
return success();
case MaskFormat::AllFalse:
rewriter.eraseOp(store);
return success();
case MaskFormat::Unknown:
return failure();
}
llvm_unreachable("Unexpected 1DMaskFormat on MaskedStore");
}
};
} // namespace
void MaskedStoreOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<MaskedStoreFolder>(context);
}
LogicalResult MaskedStoreOp::fold(FoldAdaptor adaptor,
SmallVectorImpl<OpFoldResult> &results) {
return memref::foldMemRefCast(*this);
}
//===----------------------------------------------------------------------===//
// GatherOp
//===----------------------------------------------------------------------===//
LogicalResult GatherOp::verify() {
VectorType indVType = getIndexVectorType();
VectorType maskVType = getMaskVectorType();
VectorType resVType = getVectorType();
ShapedType baseType = getBaseType();
if (!llvm::isa<MemRefType, RankedTensorType>(baseType))
return emitOpError("requires base to be a memref or ranked tensor type");
if (resVType.getElementType() != baseType.getElementType())
return emitOpError("base and result element type should match");
if (llvm::size(getIndices()) != baseType.getRank())
return emitOpError("requires ") << baseType.getRank() << " indices";
if (resVType.getShape() != indVType.getShape())
return emitOpError("expected result dim to match indices dim");
if (resVType.getShape() != maskVType.getShape())
return emitOpError("expected result dim to match mask dim");
if (resVType != getPassThruVectorType())
return emitOpError("expected pass_thru of same type as result type");
return success();
}
// MaskableOpInterface methods.
/// Returns the mask type expected by this operation. Mostly used for
/// verification purposes. It requires the operation to be vectorized."
Type GatherOp::getExpectedMaskType() {
auto vecType = this->getIndexVectorType();
return VectorType::get(vecType.getShape(),
IntegerType::get(vecType.getContext(), /*width=*/1),
vecType.getScalableDims());
}
std::optional<SmallVector<int64_t, 4>> GatherOp::getShapeForUnroll() {
return llvm::to_vector<4>(getVectorType().getShape());
}
/// Cheeck if `indexVec` is constant 1D vec of consecutive values [0, 1, 2, ...]
static LogicalResult isZeroBasedContiguousSeq(Value indexVec) {
auto vecType = dyn_cast<VectorType>(indexVec.getType());
if (!vecType || vecType.getRank() != 1 || vecType.isScalable())
return failure();
if (indexVec.getDefiningOp<StepOp>())
return success();
DenseIntElementsAttr elements;
if (!matchPattern(indexVec, m_Constant(&elements)))
return failure();
return success(
llvm::equal(elements, llvm::seq<int64_t>(0, vecType.getNumElements())));
}
namespace {
class GatherFolder final : public OpRewritePattern<GatherOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(GatherOp gather,
PatternRewriter &rewriter) const override {
switch (getMaskFormat(gather.getMask())) {
case MaskFormat::AllTrue:
return failure(); // no unmasked equivalent
case MaskFormat::AllFalse:
rewriter.replaceOp(gather, gather.getPassThru());
return success();
case MaskFormat::Unknown:
return failure();
}
llvm_unreachable("Unexpected 1DMaskFormat on GatherFolder");
}
};
/// Fold gathers with consecutive offsets [0, 1, 2, ...] into contiguous
/// maskedload. Only 1D fixed vectors are supported for now.
class FoldContiguousGather final : public OpRewritePattern<GatherOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(GatherOp op,
PatternRewriter &rewriter) const override {
if (!isa<MemRefType>(op.getBase().getType()))
return rewriter.notifyMatchFailure(op, "base must be of memref type");
if (failed(isZeroBasedContiguousSeq(op.getIndexVec())))
return failure();
rewriter.replaceOpWithNewOp<MaskedLoadOp>(op, op.getType(), op.getBase(),
op.getIndices(), op.getMask(),
op.getPassThru());
return success();
}
};
} // namespace
void GatherOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<GatherFolder, FoldContiguousGather>(context);
}
//===----------------------------------------------------------------------===//
// ScatterOp
//===----------------------------------------------------------------------===//
LogicalResult ScatterOp::verify() {
VectorType indVType = getIndexVectorType();
VectorType maskVType = getMaskVectorType();
VectorType valueVType = getVectorType();
MemRefType memType = getMemRefType();
if (valueVType.getElementType() != memType.getElementType())
return emitOpError("base and valueToStore element type should match");
if (llvm::size(getIndices()) != memType.getRank())
return emitOpError("requires ") << memType.getRank() << " indices";
if (valueVType.getShape() != indVType.getShape())
return emitOpError("expected valueToStore dim to match indices dim");
if (valueVType.getShape() != maskVType.getShape())
return emitOpError("expected valueToStore dim to match mask dim");
return success();
}
namespace {
class ScatterFolder final : public OpRewritePattern<ScatterOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ScatterOp scatter,
PatternRewriter &rewriter) const override {
switch (getMaskFormat(scatter.getMask())) {
case MaskFormat::AllTrue:
return failure(); // no unmasked equivalent
case MaskFormat::AllFalse:
rewriter.eraseOp(scatter);
return success();
case MaskFormat::Unknown:
return failure();
}
llvm_unreachable("Unexpected 1DMaskFormat on ScatterFolder");
}
};
/// Fold scatters with consecutive offsets [0, 1, 2, ...] into contiguous
/// maskedstore. Only 1D fixed vectors are supported for now.
class FoldContiguousScatter final : public OpRewritePattern<ScatterOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ScatterOp op,
PatternRewriter &rewriter) const override {
if (failed(isZeroBasedContiguousSeq(op.getIndexVec())))
return failure();
rewriter.replaceOpWithNewOp<MaskedStoreOp>(
op, op.getBase(), op.getIndices(), op.getMask(), op.getValueToStore());
return success();
}
};
} // namespace
void ScatterOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<ScatterFolder, FoldContiguousScatter>(context);
}
//===----------------------------------------------------------------------===//
// ExpandLoadOp
//===----------------------------------------------------------------------===//
LogicalResult ExpandLoadOp::verify() {
VectorType maskVType = getMaskVectorType();
VectorType passVType = getPassThruVectorType();
VectorType resVType = getVectorType();
MemRefType memType = getMemRefType();
if (resVType.getElementType() != memType.getElementType())
return emitOpError("base and result element type should match");
if (llvm::size(getIndices()) != memType.getRank())
return emitOpError("requires ") << memType.getRank() << " indices";
if (resVType.getDimSize(0) != maskVType.getDimSize(0))
return emitOpError("expected result dim to match mask dim");
if (resVType != passVType)
return emitOpError("expected pass_thru of same type as result type");
return success();
}
namespace {
class ExpandLoadFolder final : public OpRewritePattern<ExpandLoadOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ExpandLoadOp expand,
PatternRewriter &rewriter) const override {
switch (getMaskFormat(expand.getMask())) {
case MaskFormat::AllTrue:
rewriter.replaceOpWithNewOp<vector::LoadOp>(
expand, expand.getType(), expand.getBase(), expand.getIndices());
return success();
case MaskFormat::AllFalse:
rewriter.replaceOp(expand, expand.getPassThru());
return success();
case MaskFormat::Unknown:
return failure();
}
llvm_unreachable("Unexpected 1DMaskFormat on ExpandLoadFolder");
}
};
} // namespace
void ExpandLoadOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<ExpandLoadFolder>(context);
}
//===----------------------------------------------------------------------===//
// CompressStoreOp
//===----------------------------------------------------------------------===//
LogicalResult CompressStoreOp::verify() {
VectorType maskVType = getMaskVectorType();
VectorType valueVType = getVectorType();
MemRefType memType = getMemRefType();
if (valueVType.getElementType() != memType.getElementType())
return emitOpError("base and valueToStore element type should match");
if (llvm::size(getIndices()) != memType.getRank())
return emitOpError("requires ") << memType.getRank() << " indices";
if (valueVType.getDimSize(0) != maskVType.getDimSize(0))
return emitOpError("expected valueToStore dim to match mask dim");
return success();
}
namespace {
class CompressStoreFolder final : public OpRewritePattern<CompressStoreOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(CompressStoreOp compress,
PatternRewriter &rewriter) const override {
switch (getMaskFormat(compress.getMask())) {
case MaskFormat::AllTrue:
rewriter.replaceOpWithNewOp<vector::StoreOp>(
compress, compress.getValueToStore(), compress.getBase(),
compress.getIndices());
return success();
case MaskFormat::AllFalse:
rewriter.eraseOp(compress);
return success();
case MaskFormat::Unknown:
return failure();
}
llvm_unreachable("Unexpected 1DMaskFormat on CompressStoreFolder");
}
};
} // namespace
void CompressStoreOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<CompressStoreFolder>(context);
}
//===----------------------------------------------------------------------===//
// ShapeCastOp
//===----------------------------------------------------------------------===//
void ShapeCastOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
SetIntRangeFn setResultRanges) {
setResultRanges(getResult(), argRanges.front());
}
LogicalResult ShapeCastOp::verify() {
VectorType sourceType = getSourceVectorType();
VectorType resultType = getResultVectorType();
// Check that element type is preserved
if (sourceType.getElementType() != resultType.getElementType())
return emitOpError("has different source and result element types");
// Check that number of elements is preserved
int64_t sourceNElms = sourceType.getNumElements();
int64_t resultNElms = resultType.getNumElements();
if (sourceNElms != resultNElms) {
return emitOpError() << "has different number of elements at source ("
<< sourceNElms << ") and result (" << resultNElms
<< ")";
}
// Check that (non-)scalability is preserved
int64_t sourceNScalableDims = sourceType.getNumScalableDims();
int64_t resultNScalableDims = resultType.getNumScalableDims();
if (sourceNScalableDims != resultNScalableDims)
return emitOpError() << "has different number of scalable dims at source ("
<< sourceNScalableDims << ") and result ("
<< resultNScalableDims << ")";
return success();
}
namespace {
/// Return true if `transpose` does not permute a pair of non-unit dims.
/// By `order preserving` we mean that the flattened versions of the input and
/// output vectors are (numerically) identical. In other words `transpose` is
/// effectively a shape cast.
bool isOrderPreserving(TransposeOp transpose) {
ArrayRef<int64_t> permutation = transpose.getPermutation();
VectorType sourceType = transpose.getSourceVectorType();
ArrayRef<int64_t> inShape = sourceType.getShape();
ArrayRef<bool> inDimIsScalable = sourceType.getScalableDims();
auto isNonScalableUnitDim = [&](int64_t dim) {
return inShape[dim] == 1 && !inDimIsScalable[dim];
};
int64_t current = 0;
for (auto p : permutation) {
if (!isNonScalableUnitDim(p)) {
if (p < current) {
return false;
}
current = p;
}
}
return true;
}
} // namespace
OpFoldResult ShapeCastOp::fold(FoldAdaptor adaptor) {
VectorType resultType = getType();
// No-op shape cast.
if (getSource().getType() == resultType)
return getSource();
// shape_cast(shape_cast(x)) -> shape_cast(x)
if (auto precedingShapeCast = getSource().getDefiningOp<ShapeCastOp>()) {
setOperand(precedingShapeCast.getSource());
return getResult();
}
// shape_cast(transpose(x)) -> shape_cast(x)
if (auto transpose = getSource().getDefiningOp<TransposeOp>()) {
// This folder does
// shape_cast(transpose) -> shape_cast
// But another pattern, ConvertIllegalShapeCastOpsToTransposes, does
// shape_cast -> shape_cast(transpose)
// i.e. the complete opposite. When paired, these 2 patterns can cause
// infinite cycles in pattern rewriting.
// ConvertIllegalShapeCastOpsToTransposes only matches on scalable
// vectors, so by disabling this folder for scalable vectors the
// cycle is avoided.
// TODO: Check if ConvertIllegalShapeCastOpsToTransposes is
// still needed. If it's not, then we can fold here.
if (!transpose.getType().isScalable() && isOrderPreserving(transpose)) {
setOperand(transpose.getVector());
return getResult();
}
return {};
}
// Y = shape_cast(broadcast(X))
// -> X, if X and Y have same type
if (auto bcastOp = getSource().getDefiningOp<BroadcastOp>()) {
if (bcastOp.getSourceType() == resultType)
return bcastOp.getSource();
}
// shape_cast(constant) -> constant
if (auto splatAttr =
llvm::dyn_cast_if_present<SplatElementsAttr>(adaptor.getSource()))
return splatAttr.reshape(getType());
// shape_cast(poison) -> poison
if (llvm::dyn_cast_if_present<ub::PoisonAttr>(adaptor.getSource())) {
return ub::PoisonAttr::get(getContext());
}
return {};
}
namespace {
/// Helper function that computes a new vector type based on the input vector
/// type by removing the trailing one dims:
///
/// vector<4x1x1xi1> --> vector<4x1xi1>
///
static VectorType trimTrailingOneDims(VectorType oldType) {
ArrayRef<int64_t> oldShape = oldType.getShape();
ArrayRef<int64_t> newShape = oldShape;
ArrayRef<bool> oldScalableDims = oldType.getScalableDims();
ArrayRef<bool> newScalableDims = oldScalableDims;
while (!newShape.empty() && newShape.back() == 1 && !newScalableDims.back()) {
newShape = newShape.drop_back(1);
newScalableDims = newScalableDims.drop_back(1);
}
// Make sure we have at least 1 dimension.
// TODO: Add support for 0-D vectors.
if (newShape.empty()) {
newShape = oldShape.take_back();
newScalableDims = oldScalableDims.take_back();
}
return VectorType::get(newShape, oldType.getElementType(), newScalableDims);
}
/// Folds qualifying shape_cast(create_mask) into a new create_mask
///
/// Looks at `vector.shape_cast` Ops that simply "drop" the trailing unit
/// dimension. If the input vector comes from `vector.create_mask` for which
/// the corresponding mask input value is 1 (e.g. `%c1` below), then it is safe
/// to fold shape_cast into create_mask.
///
/// BEFORE:
/// %1 = vector.create_mask %c1, %dim, %c1, %c1 : vector<1x[4]x1x1xi1>
/// %2 = vector.shape_cast %1 : vector<1x[4]x1x1xi1> to vector<1x[4]xi1>
/// AFTER:
/// %0 = vector.create_mask %c1, %dim : vector<1x[4]xi1>
class ShapeCastCreateMaskFolderTrailingOneDim final
: public OpRewritePattern<ShapeCastOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ShapeCastOp shapeOp,
PatternRewriter &rewriter) const override {
Value shapeOpSrc = shapeOp->getOperand(0);
auto createMaskOp = shapeOpSrc.getDefiningOp<vector::CreateMaskOp>();
auto constantMaskOp = shapeOpSrc.getDefiningOp<vector::ConstantMaskOp>();
if (!createMaskOp && !constantMaskOp)
return failure();
VectorType shapeOpResTy = shapeOp.getResultVectorType();
VectorType shapeOpSrcTy = shapeOp.getSourceVectorType();
VectorType newVecType = trimTrailingOneDims(shapeOpSrcTy);
if (newVecType != shapeOpResTy)
return failure();
auto numDimsToDrop =
shapeOpSrcTy.getShape().size() - shapeOpResTy.getShape().size();
// No unit dims to drop
if (!numDimsToDrop)
return failure();
if (createMaskOp) {
auto maskOperands = createMaskOp.getOperands();
auto numMaskOperands = maskOperands.size();
// Check every mask dim size to see whether it can be dropped
for (size_t i = numMaskOperands - 1; i >= numMaskOperands - numDimsToDrop;
--i) {
auto constant = maskOperands[i].getDefiningOp<arith::ConstantIndexOp>();
if (!constant || (constant.value() != 1))
return failure();
}
SmallVector<Value> newMaskOperands =
maskOperands.drop_back(numDimsToDrop);
rewriter.replaceOpWithNewOp<vector::CreateMaskOp>(shapeOp, shapeOpResTy,
newMaskOperands);
return success();
}
if (constantMaskOp) {
auto maskDimSizes = constantMaskOp.getMaskDimSizes();
auto numMaskOperands = maskDimSizes.size();
// Check every mask dim size to see whether it can be dropped
for (size_t i = numMaskOperands - 1; i >= numMaskOperands - numDimsToDrop;
--i) {
if (maskDimSizes[i] != 1)
return failure();
}
auto newMaskOperands = maskDimSizes.drop_back(numDimsToDrop);
rewriter.replaceOpWithNewOp<vector::ConstantMaskOp>(shapeOp, shapeOpResTy,
newMaskOperands);
return success();
}
return failure();
}
};
/// Pattern to rewrite Y = ShapeCast(Broadcast(X)) as either
/// i) Y = ShapeCast(X), or
/// ii) Y = Broadcast(X)
/// If both (i) and (ii) are possible, (i) is chosen.
class ShapeCastBroadcastFolder final : public OpRewritePattern<ShapeCastOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(ShapeCastOp shapeCastOp,
PatternRewriter &rewriter) const override {
auto broadcastOp =
shapeCastOp.getSource().getDefiningOp<vector::BroadcastOp>();
if (!broadcastOp)
return failure();
auto srcVectorType = dyn_cast<VectorType>(broadcastOp.getSourceType());
bool srcIsScalar = !srcVectorType;
// Replace Y = ShapeCast(Broadcast(X)) with Y = ShapeCast(X).
// Example:
// %0 = vector.broadcast %in : vector<3x4xf32> to vector<1x3x4xf32>
// %1 = vector.shape_cast %0 : vector<1x3x4xf32> to vector<12xf32>
// to
// %1 = vector.shape_cast %in : vector<3x4xf32> to vector<12xf32>
if (srcVectorType) {
if (srcVectorType.getNumElements() ==
shapeCastOp.getResultVectorType().getNumElements()) {
rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(
shapeCastOp, shapeCastOp.getResultVectorType(),
broadcastOp.getSource());
return success();
}
}
// Replace Y = ShapeCast(Broadcast(X)) with Y = Broadcast(X)
// Example
// %0 = vector.broadcast %in : vector<3xf32> to vector<2x4x3xf32>
// %1 = vector.shape_cast %0 : vector<2x4x3xf32> to vector<8x3xf32>
// to
// %1 = vector.broadcast %in : vector<3xf32> to vector<8x3xf32>
VectorType dstVectorType = shapeCastOp.getResultVectorType();
if (srcIsScalar || isBroadcastableTo(srcVectorType, dstVectorType) ==
BroadcastableToResult::Success) {
rewriter.replaceOpWithNewOp<vector::BroadcastOp>(
shapeCastOp, dstVectorType, broadcastOp.getSource());
return success();
}
return failure();
}
};
} // namespace
void ShapeCastOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results
.add<ShapeCastCreateMaskFolderTrailingOneDim, ShapeCastBroadcastFolder>(
context);
}
//===----------------------------------------------------------------------===//
// VectorBitCastOp
//===----------------------------------------------------------------------===//
LogicalResult BitCastOp::verify() {
auto sourceVectorType = getSourceVectorType();
auto resultVectorType = getResultVectorType();
for (int64_t i = 0, e = sourceVectorType.getRank() - 1; i < e; i++) {
if (sourceVectorType.getDimSize(i) != resultVectorType.getDimSize(i))
return emitOpError("dimension size mismatch at: ") << i;
}
DataLayout dataLayout = DataLayout::closest(*this);
auto sourceElementBits =
dataLayout.getTypeSizeInBits(sourceVectorType.getElementType());
auto resultElementBits =
dataLayout.getTypeSizeInBits(resultVectorType.getElementType());
if (sourceVectorType.getRank() == 0) {
if (sourceElementBits != resultElementBits)
return emitOpError("source/result bitwidth of the 0-D vector element "
"types must be equal");
} else if (sourceElementBits * sourceVectorType.getShape().back() !=
resultElementBits * resultVectorType.getShape().back()) {
return emitOpError(
"source/result bitwidth of the minor 1-D vectors must be equal");
}
return success();
}
OpFoldResult BitCastOp::fold(FoldAdaptor adaptor) {
// Nop cast.
if (getSource().getType() == getResult().getType())
return getSource();
// Canceling bitcasts.
if (auto otherOp = getSource().getDefiningOp<BitCastOp>()) {
if (getResult().getType() == otherOp.getSource().getType())
return otherOp.getSource();
setOperand(otherOp.getSource());
return getResult();
}
Attribute sourceConstant = adaptor.getSource();
if (!sourceConstant)
return {};
Type srcElemType = getSourceVectorType().getElementType();
Type dstElemType = getResultVectorType().getElementType();
if (auto floatPack = llvm::dyn_cast<DenseFPElementsAttr>(sourceConstant)) {
if (floatPack.isSplat()) {
auto splat = floatPack.getSplatValue<FloatAttr>();
// Casting fp16 into fp32.
if (srcElemType.isF16() && dstElemType.isF32()) {
uint32_t bits = static_cast<uint32_t>(
splat.getValue().bitcastToAPInt().getZExtValue());
// Duplicate the 16-bit pattern.
bits = (bits << 16) | (bits & 0xffff);
APInt intBits(32, bits);
APFloat floatBits(llvm::APFloat::IEEEsingle(), intBits);
return DenseElementsAttr::get(getResultVectorType(), floatBits);
}
}
}
if (auto intPack = llvm::dyn_cast<DenseIntElementsAttr>(sourceConstant)) {
if (intPack.isSplat()) {
auto splat = intPack.getSplatValue<IntegerAttr>();
if (llvm::isa<IntegerType>(dstElemType)) {
uint64_t srcBitWidth = srcElemType.getIntOrFloatBitWidth();
uint64_t dstBitWidth = dstElemType.getIntOrFloatBitWidth();
// Casting to a larger integer bit width.
if (dstBitWidth > srcBitWidth && dstBitWidth % srcBitWidth == 0) {
APInt intBits = splat.getValue().zext(dstBitWidth);
// Duplicate the lower width element.
for (uint64_t i = 0; i < dstBitWidth / srcBitWidth - 1; i++)
intBits = (intBits << srcBitWidth) | intBits;
return DenseElementsAttr::get(getResultVectorType(), intBits);
}
}
}
}
return {};
}
//===----------------------------------------------------------------------===//
// TypeCastOp
//===----------------------------------------------------------------------===//
static SmallVector<int64_t, 8> extractShape(MemRefType memRefType) {
auto vectorType = llvm::dyn_cast<VectorType>(memRefType.getElementType());
SmallVector<int64_t, 8> res(memRefType.getShape());
if (vectorType)
res.append(vectorType.getShape().begin(), vectorType.getShape().end());
return res;
}
/// Build the canonical memRefType with a single vector.
/// E.g. memref<4 x 5 x vector<6 x f32>> -> memref<vector<4 x 5 x 6 x f32>>.
void TypeCastOp::build(OpBuilder &builder, OperationState &result,
Value source) {
result.addOperands(source);
MemRefType memRefType = llvm::cast<MemRefType>(source.getType());
VectorType vectorType =
VectorType::get(extractShape(memRefType),
getElementTypeOrSelf(getElementTypeOrSelf(memRefType)));
result.addTypes(MemRefType::get({}, vectorType, MemRefLayoutAttrInterface(),
memRefType.getMemorySpace()));
}
LogicalResult TypeCastOp::verify() {
MemRefType canonicalType = getMemRefType().canonicalizeStridedLayout();
if (!canonicalType.getLayout().isIdentity())
return emitOpError("expects operand to be a memref with identity layout");
if (!getResultMemRefType().getLayout().isIdentity())
return emitOpError("expects result to be a memref with identity layout");
if (getResultMemRefType().getMemorySpace() !=
getMemRefType().getMemorySpace())
return emitOpError("expects result in same memory space");
auto sourceType = getMemRefType();
auto resultType = getResultMemRefType();
if (getElementTypeOrSelf(getElementTypeOrSelf(sourceType)) !=
getElementTypeOrSelf(getElementTypeOrSelf(resultType)))
return emitOpError(
"expects result and operand with same underlying scalar type: ")
<< resultType;
if (extractShape(sourceType) != extractShape(resultType))
return emitOpError(
"expects concatenated result and operand shapes to be equal: ")
<< resultType;
return success();
}
//===----------------------------------------------------------------------===//
// TransposeOp
//===----------------------------------------------------------------------===//
void vector::TransposeOp::build(OpBuilder &builder, OperationState &result,
Value vector, ArrayRef<int64_t> permutation) {
VectorType vt = llvm::cast<VectorType>(vector.getType());
SmallVector<int64_t, 4> transposedShape(vt.getRank());
SmallVector<bool, 4> transposedScalableDims(vt.getRank());
for (unsigned i = 0; i < permutation.size(); ++i) {
transposedShape[i] = vt.getShape()[permutation[i]];
transposedScalableDims[i] = vt.getScalableDims()[permutation[i]];
}
result.addOperands(vector);
result.addTypes(VectorType::get(transposedShape, vt.getElementType(),
transposedScalableDims));
result.addAttribute(TransposeOp::getPermutationAttrName(result.name),
builder.getDenseI64ArrayAttr(permutation));
}
OpFoldResult vector::TransposeOp::fold(FoldAdaptor adaptor) {
// Eliminate splat constant transpose ops.
if (auto splat =
llvm::dyn_cast_if_present<SplatElementsAttr>(adaptor.getVector()))
return splat.reshape(getResultVectorType());
// Eliminate poison transpose ops.
if (llvm::dyn_cast_if_present<ub::PoisonAttr>(adaptor.getVector()))
return ub::PoisonAttr::get(getContext());
// Eliminate identity transpose ops. This happens when the dimensions of the
// input vector remain in their original order after the transpose operation.
ArrayRef<int64_t> perm = getPermutation();
// Check if the permutation of the dimensions contains sequential values:
// {0, 1, 2, ...}.
for (int64_t i = 0, e = perm.size(); i < e; i++) {
if (perm[i] != i)
return {};
}
return getVector();
}
LogicalResult vector::TransposeOp::verify() {
VectorType vectorType = getSourceVectorType();
VectorType resultType = getResultVectorType();
int64_t rank = resultType.getRank();
if (vectorType.getRank() != rank)
return emitOpError("vector result rank mismatch: ") << rank;
// Verify transposition array.
ArrayRef<int64_t> perm = getPermutation();
int64_t size = perm.size();
if (rank != size)
return emitOpError("transposition length mismatch: ") << size;
SmallVector<bool, 8> seen(rank, false);
for (const auto &ta : llvm::enumerate(perm)) {
if (ta.value() < 0 || ta.value() >= rank)
return emitOpError("transposition index out of range: ") << ta.value();
if (seen[ta.value()])
return emitOpError("duplicate position index: ") << ta.value();
seen[ta.value()] = true;
if (resultType.getDimSize(ta.index()) != vectorType.getDimSize(ta.value()))
return emitOpError("dimension size mismatch at: ") << ta.value();
}
return success();
}
std::optional<SmallVector<int64_t, 4>> TransposeOp::getShapeForUnroll() {
return llvm::to_vector<4>(getResultVectorType().getShape());
}
namespace {
// Rewrites two back-to-back TransposeOp operations into a single TransposeOp.
class TransposeFolder final : public OpRewritePattern<vector::TransposeOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(vector::TransposeOp transposeOp,
PatternRewriter &rewriter) const override {
// Composes two permutations: result[i] = permutation1[permutation2[i]].
auto composePermutations = [](ArrayRef<int64_t> permutation1,
ArrayRef<int64_t> permutation2) {
SmallVector<int64_t, 4> result;
for (auto index : permutation2)
result.push_back(permutation1[index]);
return result;
};
// Return if the input of 'transposeOp' is not defined by another transpose.
vector::TransposeOp parentTransposeOp =
transposeOp.getVector().getDefiningOp<vector::TransposeOp>();
if (!parentTransposeOp)
return failure();
SmallVector<int64_t, 4> permutation = composePermutations(
parentTransposeOp.getPermutation(), transposeOp.getPermutation());
// Replace 'transposeOp' with a new transpose operation.
rewriter.replaceOpWithNewOp<vector::TransposeOp>(
transposeOp, transposeOp.getResult().getType(),
parentTransposeOp.getVector(), permutation);
return success();
}
};
// Folds transpose(splat x : src_type) : res_type into splat x : res_type.
class FoldTransposeSplat final : public OpRewritePattern<TransposeOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(TransposeOp transposeOp,
PatternRewriter &rewriter) const override {
auto splatOp = transposeOp.getVector().getDefiningOp<vector::SplatOp>();
if (!splatOp)
return failure();
rewriter.replaceOpWithNewOp<vector::SplatOp>(
transposeOp, transposeOp.getResultVectorType(), splatOp.getInput());
return success();
}
};
/// Folds transpose(create_mask) into a new transposed create_mask.
class FoldTransposeCreateMask final : public OpRewritePattern<TransposeOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(TransposeOp transpOp,
PatternRewriter &rewriter) const override {
Value transposeSrc = transpOp.getVector();
auto createMaskOp = transposeSrc.getDefiningOp<vector::CreateMaskOp>();
auto constantMaskOp = transposeSrc.getDefiningOp<vector::ConstantMaskOp>();
if (!createMaskOp && !constantMaskOp)
return failure();
// Get the transpose permutation and apply it to the vector.create_mask or
// vector.constant_mask operands.
ArrayRef<int64_t> permutation = transpOp.getPermutation();
if (createMaskOp) {
auto maskOperands = createMaskOp.getOperands();
SmallVector<Value> newOperands(maskOperands.begin(), maskOperands.end());
applyPermutationToVector(newOperands, permutation);
rewriter.replaceOpWithNewOp<vector::CreateMaskOp>(
transpOp, transpOp.getResultVectorType(), newOperands);
return success();
}
// ConstantMaskOp case.
auto maskDimSizes = constantMaskOp.getMaskDimSizes();
auto newMaskDimSizes = applyPermutation(maskDimSizes, permutation);
rewriter.replaceOpWithNewOp<vector::ConstantMaskOp>(
transpOp, transpOp.getResultVectorType(), newMaskDimSizes);
return success();
}
};
/// Folds transpose(shape_cast) into a new shape_cast.
class FoldTransposeShapeCast final : public OpRewritePattern<TransposeOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(TransposeOp transposeOp,
PatternRewriter &rewriter) const override {
auto shapeCastOp =
transposeOp.getVector().getDefiningOp<vector::ShapeCastOp>();
if (!shapeCastOp)
return failure();
if (!isOrderPreserving(transposeOp))
return failure();
VectorType resultType = transposeOp.getType();
// We don't need to check isValidShapeCast at this point, because it is
// guaranteed that merging the transpose into the the shape_cast is a valid
// shape_cast, because the transpose just inserts/removes ones.
rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(transposeOp, resultType,
shapeCastOp.getSource());
return success();
}
};
/// Folds transpose(broadcast(x)) to broadcast(x) if the transpose is
/// 'order preserving', where 'order preserving' means the flattened
/// inputs and outputs of the transpose have identical (numerical) values.
///
/// Example:
/// ```
/// %0 = vector.broadcast %input : vector<1x1xi32> to vector<1x8xi32>
/// %1 = vector.transpose %0, [1, 0] : vector<1x8xi32>
/// to vector<8x1xi32>
/// ```
/// can be rewritten as the equivalent
/// ```
/// %0 = vector.broadcast %input : vector<1x1xi32> to vector<8x1xi32>.
/// ```
/// The algorithm works by partitioning dimensions into groups that can be
/// locally permuted while preserving order, and checks that the transpose
/// only permutes within these groups.
///
/// Groups are either contiguous sequences of 1s, or non-1s (1-element groups).
/// Consider broadcasting 4x1x1x7 to 2x3x4x5x6x7. This is equivalent to
/// broadcasting from 1x1x4x1x1x7.
/// ^^^ ^ ^^^ ^
/// groups: 0 1 2 3
/// Order preserving permutations for this example are ones that only permute
/// within the groups [0,1] and [3,4], like (1 0 2 4 3 5 6).
class FoldTransposeBroadcast : public OpRewritePattern<vector::TransposeOp> {
public:
using OpRewritePattern::OpRewritePattern;
FoldTransposeBroadcast(MLIRContext *context, PatternBenefit benefit = 1)
: OpRewritePattern<vector::TransposeOp>(context, benefit) {}
LogicalResult matchAndRewrite(vector::TransposeOp transpose,
PatternRewriter &rewriter) const override {
vector::BroadcastOp broadcast =
transpose.getVector().getDefiningOp<vector::BroadcastOp>();
if (!broadcast) {
return rewriter.notifyMatchFailure(transpose,
"not preceded by a broadcast");
}
auto inputType = dyn_cast<VectorType>(broadcast.getSourceType());
VectorType outputType = transpose.getResultVectorType();
// transpose(broadcast(scalar)) -> broadcast(scalar) is always valid
bool inputIsScalar = !inputType;
if (inputIsScalar) {
rewriter.replaceOpWithNewOp<vector::BroadcastOp>(transpose, outputType,
transpose.getVector());
return success();
}
ArrayRef<int64_t> permutation = transpose.getPermutation();
ArrayRef<int64_t> inputShape = inputType.getShape();
int64_t inputRank = inputType.getRank();
int64_t outputRank = transpose.getType().getRank();
int64_t deltaRank = outputRank - inputRank;
int low = 0;
for (int inputIndex = 0; inputIndex < inputRank; ++inputIndex) {
bool notOne = inputShape[inputIndex] != 1;
bool prevNotOne = (inputIndex != 0 && inputShape[inputIndex - 1] != 1);
bool groupEndFound = notOne || prevNotOne;
if (groupEndFound) {
int high = inputIndex + deltaRank;
// Return failure if not all permutation destinations for indices in
// [low, high) are in [low, high), i.e. the permutation is not local to
// the group.
for (int i = low; i < high; ++i) {
if (permutation[i] < low || permutation[i] >= high) {
return rewriter.notifyMatchFailure(
transpose, "permutation not local to group");
}
}
}
}
// We don't need to check the final group [low, outputRank) because if it is
// not locally bound, there must be a preceding group that already failed
// the check (impossible to have just 1 non-locally bound group).
// The preceding logic also ensures that at this point, the output of the
// transpose is definitely broadcastable from the input shape, assert so:
assert(vector::isBroadcastableTo(inputType, outputType) ==
vector::BroadcastableToResult::Success &&
"not broadcastable directly to transpose output");
rewriter.replaceOpWithNewOp<vector::BroadcastOp>(transpose, outputType,
transpose.getVector());
return success();
}
};
} // namespace
void vector::TransposeOp::getCanonicalizationPatterns(
RewritePatternSet &results, MLIRContext *context) {
results.add<FoldTransposeCreateMask, FoldTransposeShapeCast, TransposeFolder,
FoldTransposeSplat, FoldTransposeBroadcast>(context);
}
//===----------------------------------------------------------------------===//
// ConstantMaskOp
//===----------------------------------------------------------------------===//
void ConstantMaskOp::build(OpBuilder &builder, OperationState &result,
VectorType type, ConstantMaskKind kind) {
assert(kind == ConstantMaskKind::AllTrue ||
kind == ConstantMaskKind::AllFalse);
build(builder, result, type,
kind == ConstantMaskKind::AllTrue
? type.getShape()
: SmallVector<int64_t>(type.getRank(), 0));
}
LogicalResult ConstantMaskOp::verify() {
auto resultType = llvm::cast<VectorType>(getResult().getType());
// Check the corner case of 0-D vectors first.
if (resultType.getRank() == 0) {
if (getMaskDimSizes().size() != 1)
return emitError("array attr must have length 1 for 0-D vectors");
auto dim = getMaskDimSizes()[0];
if (dim != 0 && dim != 1)
return emitError("mask dim size must be either 0 or 1 for 0-D vectors");
return success();
}
// Verify that array attr size matches the rank of the vector result.
if (static_cast<int64_t>(getMaskDimSizes().size()) != resultType.getRank())
return emitOpError(
"must specify array attr of size equal vector result rank");
// Verify that each array attr element is in bounds of corresponding vector
// result dimension size.
auto resultShape = resultType.getShape();
auto resultScalableDims = resultType.getScalableDims();
ArrayRef<int64_t> maskDimSizes = getMaskDimSizes();
for (const auto [index, maskDimSize] : llvm::enumerate(maskDimSizes)) {
if (maskDimSize < 0 || maskDimSize > resultShape[index])
return emitOpError(
"array attr of size out of bounds of vector result dimension size");
if (resultScalableDims[index] && maskDimSize != 0 &&
maskDimSize != resultShape[index])
return emitOpError(
"only supports 'none set' or 'all set' scalable dimensions");
}
// Verify that if one mask dim size is zero, they all should be zero (because
// the mask region is a conjunction of each mask dimension interval).
bool anyZeros = llvm::is_contained(maskDimSizes, 0);
bool allZeros = llvm::all_of(maskDimSizes, [](int64_t s) { return s == 0; });
if (anyZeros && !allZeros)
return emitOpError("expected all mask dim sizes to be zeros, "
"as a result of conjunction with zero mask dim");
return success();
}
bool ConstantMaskOp::isAllOnesMask() {
auto resultType = getVectorType();
// Check the corner case of 0-D vectors first.
if (resultType.getRank() == 0) {
assert(getMaskDimSizes().size() == 1 && "invalid sizes for zero rank mask");
return getMaskDimSizes()[0] == 1;
}
for (const auto [resultSize, maskDimSize] :
llvm::zip_equal(resultType.getShape(), getMaskDimSizes())) {
if (maskDimSize < resultSize)
return false;
}
return true;
}
//===----------------------------------------------------------------------===//
// CreateMaskOp
//===----------------------------------------------------------------------===//
void CreateMaskOp::build(OpBuilder &builder, OperationState &result,
VectorType type,
ArrayRef<OpFoldResult> mixedOperands) {
SmallVector<Value> operands =
getValueOrCreateConstantIndexOp(builder, result.location, mixedOperands);
build(builder, result, type, operands);
}
LogicalResult CreateMaskOp::verify() {
auto vectorType = llvm::cast<VectorType>(getResult().getType());
// Verify that an operand was specified for each result vector each dimension.
if (vectorType.getRank() == 0) {
if (getNumOperands() != 1)
return emitOpError(
"must specify exactly one operand for 0-D create_mask");
} else if (getNumOperands() !=
llvm::cast<VectorType>(getResult().getType()).getRank()) {
return emitOpError(
"must specify an operand for each result vector dimension");
}
return success();
}
namespace {
/// Pattern to rewrite a CreateMaskOp with a ConstantMaskOp.
///
/// Ex 1:
/// %c2 = arith.constant 2 : index
/// %c3 = arith.constant 3 : index
/// %0 = vector.create_mask %c3, %c2 : vector<4x3xi1>
/// Becomes:
/// vector.constant_mask [3, 2] : vector<4x3xi1>
///
/// Ex 2:
/// %c_neg_1 = arith.constant -1 : index
/// %0 = vector.create_mask %c_neg_1 : vector<[8]xi1>
/// becomes:
/// vector.constant_mask [0] : vector<[8]xi1>
///
/// Ex 3:
/// %c8 = arith.constant 8 : index
/// %c16 = arith.constant 16 : index
/// %0 = vector.vscale
/// %1 = arith.muli %0, %c16 : index
/// %10 = vector.create_mask %c8, %1 : vector<8x[16]xi1>
/// becomes:
/// %0 = vector.constant_mask [8, 16] : vector<8x[16]xi1>
class CreateMaskFolder final : public OpRewritePattern<CreateMaskOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(CreateMaskOp createMaskOp,
PatternRewriter &rewriter) const override {
VectorType maskType = createMaskOp.getVectorType();
ArrayRef<int64_t> maskTypeDimSizes = maskType.getShape();
ArrayRef<bool> maskTypeDimScalableFlags = maskType.getScalableDims();
// Special case: Rank zero shape.
constexpr std::array<int64_t, 1> rankZeroShape{1};
constexpr std::array<bool, 1> rankZeroScalableDims{false};
if (maskType.getRank() == 0) {
maskTypeDimSizes = rankZeroShape;
maskTypeDimScalableFlags = rankZeroScalableDims;
}
// Determine if this CreateMaskOp can be folded to a ConstantMaskOp and
// collect the `constantDims` (for the ConstantMaskOp).
SmallVector<int64_t, 4> constantDims;
for (auto [i, dimSize] : llvm::enumerate(createMaskOp.getOperands())) {
if (auto intSize = getConstantIntValue(dimSize)) {
// Constant value.
// If the mask dim is non-scalable this can be any value.
// If the mask dim is scalable only zero (all-false) is supported.
if (maskTypeDimScalableFlags[i] && intSize >= 0)
return failure();
constantDims.push_back(*intSize);
} else if (auto vscaleMultiplier = getConstantVscaleMultiplier(dimSize)) {
// Constant vscale multiple (e.g. 4 x vscale).
// Must be all-true to fold to a ConstantMask.
if (vscaleMultiplier < maskTypeDimSizes[i])
return failure();
constantDims.push_back(*vscaleMultiplier);
} else {
return failure();
}
}
// Clamp values to constant_mask bounds.
for (auto [value, maskDimSize] : llvm::zip(constantDims, maskTypeDimSizes))
value = std::clamp<int64_t>(value, 0, maskDimSize);
// If one of dim sizes is zero, set all dims to zero.
if (llvm::is_contained(constantDims, 0))
constantDims.assign(constantDims.size(), 0);
// Replace 'createMaskOp' with ConstantMaskOp.
rewriter.replaceOpWithNewOp<ConstantMaskOp>(createMaskOp, maskType,
constantDims);
return success();
}
};
} // namespace
void CreateMaskOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<CreateMaskFolder>(context);
}
//===----------------------------------------------------------------------===//
// MaskOp
//===----------------------------------------------------------------------===//
void MaskOp::build(
OpBuilder &builder, OperationState &result, Value mask,
Operation *maskableOp,
function_ref<void(OpBuilder &, Operation *)> maskRegionBuilder) {
assert(maskRegionBuilder &&
"builder callback for 'maskRegion' must be present");
result.addOperands(mask);
OpBuilder::InsertionGuard guard(builder);
Region *maskRegion = result.addRegion();
builder.createBlock(maskRegion);
maskRegionBuilder(builder, maskableOp);
}
void MaskOp::build(
OpBuilder &builder, OperationState &result, TypeRange resultTypes,
Value mask, Operation *maskableOp,
function_ref<void(OpBuilder &, Operation *)> maskRegionBuilder) {
build(builder, result, resultTypes, mask, /*passthru=*/Value(), maskableOp,
maskRegionBuilder);
}
void MaskOp::build(
OpBuilder &builder, OperationState &result, TypeRange resultTypes,
Value mask, Value passthru, Operation *maskableOp,
function_ref<void(OpBuilder &, Operation *)> maskRegionBuilder) {
build(builder, result, mask, maskableOp, maskRegionBuilder);
if (passthru)
result.addOperands(passthru);
result.addTypes(resultTypes);
}
ParseResult MaskOp::parse(OpAsmParser &parser, OperationState &result) {
// Create the op region.
result.regions.reserve(1);
Region &maskRegion = *result.addRegion();
auto &builder = parser.getBuilder();
// Parse all the operands.
OpAsmParser::UnresolvedOperand mask;
if (parser.parseOperand(mask))
return failure();
// Optional passthru operand.
OpAsmParser::UnresolvedOperand passthru;
ParseResult parsePassthru = parser.parseOptionalComma();
if (parsePassthru.succeeded() && parser.parseOperand(passthru))
return failure();
// Parse op region.
if (parser.parseRegion(maskRegion, /*arguments=*/{}, /*argTypes=*/{}))
return failure();
MaskOp::ensureTerminator(maskRegion, builder, result.location);
// Parse the optional attribute list.
if (parser.parseOptionalAttrDict(result.attributes))
return failure();
// Parse all the types.
Type maskType;
if (parser.parseColonType(maskType))
return failure();
SmallVector<Type> resultTypes;
if (parser.parseOptionalArrowTypeList(resultTypes))
return failure();
result.types.append(resultTypes);
// Resolve operands.
if (parser.resolveOperand(mask, maskType, result.operands))
return failure();
if (parsePassthru.succeeded())
if (parser.resolveOperand(passthru, resultTypes[0], result.operands))
return failure();
return success();
}
void mlir::vector::MaskOp::print(OpAsmPrinter &p) {
p << " " << getMask();
if (getPassthru())
p << ", " << getPassthru();
// Print single masked operation and skip terminator.
p << " { ";
Block *singleBlock = &getMaskRegion().getBlocks().front();
if (singleBlock && !singleBlock->getOperations().empty())
p.printCustomOrGenericOp(&singleBlock->front());
p << " }";
p.printOptionalAttrDict(getOperation()->getAttrs());
p << " : " << getMask().getType();
if (getNumResults() > 0)
p << " -> " << getResultTypes();
}
void MaskOp::ensureTerminator(Region &region, Builder &builder, Location loc) {
OpTrait::SingleBlockImplicitTerminator<vector::YieldOp>::Impl<
MaskOp>::ensureTerminator(region, builder, loc);
// Keep the default yield terminator if the number of masked operations is not
// the expected. This case will trigger a verification failure.
Block &block = region.front();
if (block.getOperations().size() != 2)
return;
// Replace default yield terminator with a new one that returns the results
// from the masked operation.
OpBuilder opBuilder(builder.getContext());
Operation *maskedOp = &block.front();
Operation *oldYieldOp = &block.back();
assert(isa<vector::YieldOp>(oldYieldOp) && "Expected vector::YieldOp");
// Empty vector.mask op.
if (maskedOp == oldYieldOp)
return;
opBuilder.setInsertionPoint(oldYieldOp);
opBuilder.create<vector::YieldOp>(loc, maskedOp->getResults());
oldYieldOp->dropAllReferences();
oldYieldOp->erase();
}
LogicalResult MaskOp::verify() {
// Structural checks.
Block &block = getMaskRegion().getBlocks().front();
if (block.getOperations().empty())
return emitOpError("expects a terminator within the mask region");
unsigned numMaskRegionOps = block.getOperations().size();
if (numMaskRegionOps > 2)
return emitOpError("expects only one operation to mask");
// Terminator checks.
auto terminator = dyn_cast<vector::YieldOp>(block.back());
if (!terminator)
return emitOpError("expects a terminator within the mask region");
if (terminator->getNumOperands() != getNumResults())
return emitOpError(
"expects number of results to match mask region yielded values");
// Empty vector.mask. Nothing else to check.
if (numMaskRegionOps == 1)
return success();
auto maskableOp = dyn_cast<MaskableOpInterface>(block.front());
if (!maskableOp)
return emitOpError("expects a MaskableOpInterface within the mask region");
// Result checks.
if (maskableOp->getNumResults() != getNumResults())
return emitOpError("expects number of results to match maskable operation "
"number of results");
if (!llvm::equal(maskableOp->getResultTypes(), getResultTypes()))
return emitOpError(
"expects result type to match maskable operation result type");
if (llvm::count_if(maskableOp->getResultTypes(),
[](Type t) { return llvm::isa<VectorType>(t); }) > 1)
return emitOpError("multiple vector results not supported");
// Mask checks.
Type expectedMaskType = maskableOp.getExpectedMaskType();
if (getMask().getType() != expectedMaskType)
return emitOpError("expects a ")
<< expectedMaskType << " mask for the maskable operation";
// Passthru checks.
Value passthru = getPassthru();
if (passthru) {
if (!maskableOp.supportsPassthru())
return emitOpError(
"doesn't expect a passthru argument for this maskable operation");
if (maskableOp->getNumResults() != 1)
return emitOpError("expects result when passthru argument is provided");
if (passthru.getType() != maskableOp->getResultTypes()[0])
return emitOpError("expects passthru type to match result type");
}
return success();
}
/// Folds vector.mask ops with an all-true mask.
LogicalResult MaskOp::fold(FoldAdaptor adaptor,
SmallVectorImpl<OpFoldResult> &results) {
MaskFormat maskFormat = getMaskFormat(getMask());
if (isEmpty())
return failure();
if (maskFormat != MaskFormat::AllTrue)
return failure();
// Move maskable operation outside of the `vector.mask` region.
Operation *maskableOp = getMaskableOp();
maskableOp->dropAllUses();
maskableOp->moveBefore(getOperation());
llvm::append_range(results, maskableOp->getResults());
return success();
}
// Elides empty vector.mask operations with or without return values. Propagates
// the yielded values by the vector.yield terminator, if any, or erases the op,
// otherwise.
class ElideEmptyMaskOp : public OpRewritePattern<MaskOp> {
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(MaskOp maskOp,
PatternRewriter &rewriter) const override {
auto maskingOp = cast<MaskingOpInterface>(maskOp.getOperation());
if (maskingOp.getMaskableOp())
return failure();
if (!maskOp.isEmpty())
return failure();
Block *block = maskOp.getMaskBlock();
auto terminator = cast<vector::YieldOp>(block->front());
if (terminator.getNumOperands() == 0)
rewriter.eraseOp(maskOp);
else
rewriter.replaceOp(maskOp, terminator.getOperands());
return success();
}
};
void MaskOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<ElideEmptyMaskOp>(context);
}
// MaskingOpInterface definitions.
/// Returns the operation masked by this 'vector.mask'.
Operation *MaskOp::getMaskableOp() {
Block *block = getMaskBlock();
if (block->getOperations().size() < 2)
return nullptr;
return &block->front();
}
/// Returns true if 'vector.mask' has a passthru value.
bool MaskOp::hasPassthru() { return getPassthru() != Value(); }
//===----------------------------------------------------------------------===//
// ScanOp
//===----------------------------------------------------------------------===//
LogicalResult ScanOp::verify() {
VectorType srcType = getSourceType();
VectorType initialType = getInitialValueType();
// Check reduction dimension < rank.
int64_t srcRank = srcType.getRank();
int64_t reductionDim = getReductionDim();
if (reductionDim >= srcRank)
return emitOpError("reduction dimension ")
<< reductionDim << " has to be less than " << srcRank;
// Check that rank(initial_value) = rank(src) - 1.
int64_t initialValueRank = initialType.getRank();
if (initialValueRank != srcRank - 1)
return emitOpError("initial value rank ")
<< initialValueRank << " has to be equal to " << srcRank - 1;
// Check shapes of initial value and src.
ArrayRef<int64_t> srcShape = srcType.getShape();
ArrayRef<int64_t> initialValueShapes = initialType.getShape();
SmallVector<int64_t> expectedShape;
for (int i = 0; i < srcRank; i++) {
if (i != reductionDim)
expectedShape.push_back(srcShape[i]);
}
if (!llvm::equal(initialValueShapes, expectedShape)) {
return emitOpError("incompatible input/initial value shapes");
}
// Verify supported reduction kind.
Type eltType = getDestType().getElementType();
if (!isSupportedCombiningKind(getKind(), eltType))
return emitOpError("unsupported reduction type ")
<< eltType << " for kind '" << stringifyCombiningKind(getKind())
<< "'";
return success();
}
void mlir::vector::populateVectorToVectorCanonicalizationPatterns(
RewritePatternSet &patterns, PatternBenefit benefit) {
patterns
.add<CreateMaskFolder, MaskedLoadFolder, MaskedStoreFolder, GatherFolder,
ScatterFolder, ExpandLoadFolder, CompressStoreFolder,
StridedSliceConstantMaskFolder, TransposeFolder>(
patterns.getContext(), benefit);
}
//===----------------------------------------------------------------------===//
// SplatOp
//===----------------------------------------------------------------------===//
OpFoldResult SplatOp::fold(FoldAdaptor adaptor) {
auto constOperand = adaptor.getInput();
if (!isa_and_nonnull<IntegerAttr, FloatAttr>(constOperand))
return {};
// SplatElementsAttr::get treats single value for second arg as being a splat.
return SplatElementsAttr::get(getType(), {constOperand});
}
void SplatOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
SetIntRangeFn setResultRanges) {
setResultRanges(getResult(), argRanges.front());
}
Value mlir::vector::makeArithReduction(OpBuilder &b, Location loc,
CombiningKind kind, Value v1, Value acc,
arith::FastMathFlagsAttr fastmath,
Value mask) {
Type t1 = getElementTypeOrSelf(v1.getType());
Type tAcc = getElementTypeOrSelf(acc.getType());
Value result;
switch (kind) {
case CombiningKind::ADD:
if (t1.isIntOrIndex() && tAcc.isIntOrIndex())
result = b.createOrFold<arith::AddIOp>(loc, v1, acc);
else if (llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc))
result = b.createOrFold<arith::AddFOp>(loc, v1, acc, fastmath);
else
llvm_unreachable("invalid value types for ADD reduction");
break;
case CombiningKind::AND:
assert(t1.isIntOrIndex() && tAcc.isIntOrIndex() && "expected int values");
result = b.createOrFold<arith::AndIOp>(loc, v1, acc);
break;
case CombiningKind::MAXNUMF:
assert(llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc) &&
"expected float values");
result = b.createOrFold<arith::MaxNumFOp>(loc, v1, acc, fastmath);
break;
case CombiningKind::MAXIMUMF:
assert(llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc) &&
"expected float values");
result = b.createOrFold<arith::MaximumFOp>(loc, v1, acc, fastmath);
break;
case CombiningKind::MINNUMF:
assert(llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc) &&
"expected float values");
result = b.createOrFold<arith::MinNumFOp>(loc, v1, acc, fastmath);
break;
case CombiningKind::MINIMUMF:
assert(llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc) &&
"expected float values");
result = b.createOrFold<arith::MinimumFOp>(loc, v1, acc, fastmath);
break;
case CombiningKind::MAXSI:
assert(t1.isIntOrIndex() && tAcc.isIntOrIndex() && "expected int values");
result = b.createOrFold<arith::MaxSIOp>(loc, v1, acc);
break;
case CombiningKind::MINSI:
assert(t1.isIntOrIndex() && tAcc.isIntOrIndex() && "expected int values");
result = b.createOrFold<arith::MinSIOp>(loc, v1, acc);
break;
case CombiningKind::MAXUI:
assert(t1.isIntOrIndex() && tAcc.isIntOrIndex() && "expected int values");
result = b.createOrFold<arith::MaxUIOp>(loc, v1, acc);
break;
case CombiningKind::MINUI:
assert(t1.isIntOrIndex() && tAcc.isIntOrIndex() && "expected int values");
result = b.createOrFold<arith::MinUIOp>(loc, v1, acc);
break;
case CombiningKind::MUL:
if (t1.isIntOrIndex() && tAcc.isIntOrIndex())
result = b.createOrFold<arith::MulIOp>(loc, v1, acc);
else if (llvm::isa<FloatType>(t1) && llvm::isa<FloatType>(tAcc))
result = b.createOrFold<arith::MulFOp>(loc, v1, acc, fastmath);
else
llvm_unreachable("invalid value types for MUL reduction");
break;
case CombiningKind::OR:
assert(t1.isIntOrIndex() && tAcc.isIntOrIndex() && "expected int values");
result = b.createOrFold<arith::OrIOp>(loc, v1, acc);
break;
case CombiningKind::XOR:
assert(t1.isIntOrIndex() && tAcc.isIntOrIndex() && "expected int values");
result = b.createOrFold<arith::XOrIOp>(loc, v1, acc);
break;
};
assert(result && "unknown CombiningKind");
return selectPassthru(b, mask, result, acc);
}
//===----------------------------------------------------------------------===//
// Vector Masking Utilities
//===----------------------------------------------------------------------===//
/// Create the vector.yield-ended region of a vector.mask op with `maskableOp`
/// as masked operation.
void mlir::vector::createMaskOpRegion(OpBuilder &builder,
Operation *maskableOp) {
assert(maskableOp->getBlock() && "MaskableOp must be inserted into a block");
Block *insBlock = builder.getInsertionBlock();
// Create a block and move the op to that block.
insBlock->getOperations().splice(
insBlock->begin(), maskableOp->getBlock()->getOperations(), maskableOp);
builder.create<YieldOp>(maskableOp->getLoc(), maskableOp->getResults());
}
/// Creates a vector.mask operation around a maskable operation. Returns the
/// vector.mask operation if the mask provided is valid. Otherwise, returns
/// the maskable operation itself.
Operation *mlir::vector::maskOperation(OpBuilder &builder,
Operation *maskableOp, Value mask,
Value passthru) {
if (!mask)
return maskableOp;
if (passthru)
return builder.create<MaskOp>(maskableOp->getLoc(),
maskableOp->getResultTypes(), mask, passthru,
maskableOp, createMaskOpRegion);
return builder.create<MaskOp>(maskableOp->getLoc(),
maskableOp->getResultTypes(), mask, maskableOp,
createMaskOpRegion);
}
/// Creates a vector select operation that picks values from `newValue` or
/// `passthru` for each result vector lane based on `mask`. This utility is used
/// to propagate the pass-thru value of vector.mask or for cases where only the
/// pass-thru value propagation is needed. VP intrinsics do not support
/// pass-thru values and every mask-out lane is set to poison. LLVM backends are
/// usually able to match op + select patterns and fold them into a native
/// target instructions.
Value mlir::vector::selectPassthru(OpBuilder &builder, Value mask,
Value newValue, Value passthru) {
if (!mask)
return newValue;
return builder.create<arith::SelectOp>(newValue.getLoc(), newValue.getType(),
mask, newValue, passthru);
}
//===----------------------------------------------------------------------===//
// TableGen'd op method definitions
//===----------------------------------------------------------------------===//
#define GET_ATTRDEF_CLASSES
#include "mlir/Dialect/Vector/IR/VectorAttributes.cpp.inc"
#define GET_OP_CLASSES
#include "mlir/Dialect/Vector/IR/VectorOps.cpp.inc"
Reference in New Issue View Git Blame Copy Permalink