560564f51c
This is to avoid confusion when dealing with reduction/combining kinds. For example, see a recent PR comment: https://github.com/llvm/llvm-project/pull/75846#discussion_r1430722175. Previously, they were picked to mostly mirror the names of the llvm vector reduction intrinsics: https://llvm.org/docs/LangRef.html#llvm-vector-reduce-fmin-intrinsic. In isolation, it was not clear if `<maxf>` has `arith.maxnumf` or `arith.maximumf` semantics. The new reduction kind names map 1:1 to arith ops, which makes it easier to tell/look up their semantics. Because both the vector and the gpu dialect depend on the arith dialect, it's more natural to align names with those in arith than with the lowering to llvm intrinsics. Issue: https://github.com/llvm/llvm-project/issues/72354
1346 lines
52 KiB
C++
1346 lines
52 KiB
C++
//===- LowerVectorContract.cpp - Lower 'vector.contract' operation --------===//
|
|
//
|
|
// 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 target-independent rewrites and utilities to lower the
|
|
// 'vector.contract' operation.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#include "mlir/Dialect/Affine/IR/AffineOps.h"
|
|
#include "mlir/Dialect/Arith/IR/Arith.h"
|
|
#include "mlir/Dialect/Arith/Utils/Utils.h"
|
|
#include "mlir/Dialect/Linalg/IR/Linalg.h"
|
|
#include "mlir/Dialect/MemRef/IR/MemRef.h"
|
|
#include "mlir/Dialect/SCF/IR/SCF.h"
|
|
#include "mlir/Dialect/Tensor/IR/Tensor.h"
|
|
#include "mlir/Dialect/Utils/IndexingUtils.h"
|
|
#include "mlir/Dialect/Utils/StructuredOpsUtils.h"
|
|
#include "mlir/Dialect/Vector/IR/VectorOps.h"
|
|
#include "mlir/Dialect/Vector/Transforms/LoweringPatterns.h"
|
|
#include "mlir/Dialect/Vector/Utils/VectorUtils.h"
|
|
#include "mlir/IR/BuiltinAttributeInterfaces.h"
|
|
#include "mlir/IR/BuiltinTypes.h"
|
|
#include "mlir/IR/ImplicitLocOpBuilder.h"
|
|
#include "mlir/IR/Location.h"
|
|
#include "mlir/IR/Matchers.h"
|
|
#include "mlir/IR/PatternMatch.h"
|
|
#include "mlir/IR/TypeUtilities.h"
|
|
#include "mlir/Interfaces/VectorInterfaces.h"
|
|
#include "mlir/Support/LogicalResult.h"
|
|
|
|
#define DEBUG_TYPE "vector-contract-lowering"
|
|
|
|
using namespace mlir;
|
|
using namespace mlir::vector;
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Helper functions
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Helper to find an index in an affine map.
|
|
static std::optional<int64_t> getResultIndex(AffineMap map, int64_t index) {
|
|
for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
|
|
int64_t idx = map.getDimPosition(i);
|
|
if (idx == index)
|
|
return i;
|
|
}
|
|
return std::nullopt;
|
|
}
|
|
|
|
// Helper to construct iterator types with one index removed.
|
|
static SmallVector<Attribute> adjustIter(ArrayAttr iteratorTypes,
|
|
int64_t index) {
|
|
SmallVector<Attribute> results;
|
|
for (const auto &it : llvm::enumerate(iteratorTypes)) {
|
|
int64_t idx = it.index();
|
|
if (idx == index)
|
|
continue;
|
|
results.push_back(it.value());
|
|
}
|
|
return results;
|
|
}
|
|
|
|
// Helper to construct an affine map with one index removed.
|
|
static AffineMap adjustMap(AffineMap map, int64_t index,
|
|
PatternRewriter &rewriter) {
|
|
auto *ctx = rewriter.getContext();
|
|
SmallVector<AffineExpr> results;
|
|
for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
|
|
int64_t idx = map.getDimPosition(i);
|
|
if (idx == index)
|
|
continue;
|
|
// Re-insert remaining indices, but renamed when occurring
|
|
// after the removed index.
|
|
auto targetExpr = getAffineDimExpr(idx < index ? idx : idx - 1, ctx);
|
|
results.push_back(targetExpr);
|
|
}
|
|
return AffineMap::get(map.getNumDims() - 1, 0, results, ctx);
|
|
}
|
|
|
|
// Helper method to possibly drop a dimension in a load.
|
|
// TODO
|
|
static Value reshapeLoad(Location loc, Value val, VectorType type,
|
|
int64_t index, int64_t pos,
|
|
PatternRewriter &rewriter) {
|
|
if (index == -1)
|
|
return val;
|
|
|
|
// At extraction dimension?
|
|
if (index == 0)
|
|
return rewriter.create<vector::ExtractOp>(loc, val, pos);
|
|
|
|
// Unroll leading dimensions.
|
|
VectorType vType = VectorType::Builder(type).dropDim(0);
|
|
VectorType resType = VectorType::Builder(type).dropDim(index);
|
|
Value result = rewriter.create<arith::ConstantOp>(
|
|
loc, resType, rewriter.getZeroAttr(resType));
|
|
for (int64_t d = 0, e = resType.getDimSize(0); d < e; d++) {
|
|
Value ext = rewriter.create<vector::ExtractOp>(loc, val, d);
|
|
Value load = reshapeLoad(loc, ext, vType, index - 1, pos, rewriter);
|
|
result = rewriter.create<vector::InsertOp>(loc, load, result, d);
|
|
}
|
|
return result;
|
|
}
|
|
|
|
// Helper method to possibly drop a dimension in a store.
|
|
// TODO
|
|
static Value reshapeStore(Location loc, Value val, Value result,
|
|
VectorType type, int64_t index, int64_t pos,
|
|
PatternRewriter &rewriter) {
|
|
// Unmodified?
|
|
if (index == -1)
|
|
return val;
|
|
// At insertion dimension?
|
|
if (index == 0)
|
|
return rewriter.create<vector::InsertOp>(loc, val, result, pos);
|
|
|
|
// Unroll leading dimensions.
|
|
VectorType vType = VectorType::Builder(type).dropDim(0);
|
|
for (int64_t d = 0, e = type.getDimSize(0); d < e; d++) {
|
|
Value ext = rewriter.create<vector::ExtractOp>(loc, result, d);
|
|
Value ins = rewriter.create<vector::ExtractOp>(loc, val, d);
|
|
Value sto = reshapeStore(loc, ins, ext, vType, index - 1, pos, rewriter);
|
|
result = rewriter.create<vector::InsertOp>(loc, sto, result, d);
|
|
}
|
|
return result;
|
|
}
|
|
|
|
/// Helper to create arithmetic operation associated with a kind of contraction.
|
|
static std::optional<Value>
|
|
createContractArithOp(Location loc, Value x, Value y, Value acc,
|
|
vector::CombiningKind kind, PatternRewriter &rewriter,
|
|
bool isInt, Value mask = Value()) {
|
|
using vector::CombiningKind;
|
|
Value mul;
|
|
|
|
if (isInt) {
|
|
if (kind == CombiningKind::MINNUMF || kind == CombiningKind::MAXNUMF ||
|
|
kind == CombiningKind::MINIMUMF || kind == CombiningKind::MAXIMUMF)
|
|
// Only valid for floating point types.
|
|
return std::nullopt;
|
|
mul = rewriter.create<arith::MulIOp>(loc, x, y);
|
|
} else {
|
|
// Float case.
|
|
if (kind == CombiningKind::AND || kind == CombiningKind::MINUI ||
|
|
kind == CombiningKind::MINSI || kind == CombiningKind::MAXUI ||
|
|
kind == CombiningKind::MAXSI || kind == CombiningKind::OR ||
|
|
kind == CombiningKind::XOR)
|
|
// Only valid for integer types.
|
|
return std::nullopt;
|
|
// Special case for fused multiply-add.
|
|
if (acc && isa<VectorType>(acc.getType()) && kind == CombiningKind::ADD) {
|
|
Value fma = rewriter.create<vector::FMAOp>(loc, x, y, acc);
|
|
if (mask)
|
|
// The fma op doesn't need explicit masking. However, fma ops used in
|
|
// reductions must preserve previous 'acc' values for masked-out lanes.
|
|
fma = selectPassthru(rewriter, mask, fma, acc);
|
|
return fma;
|
|
}
|
|
mul = rewriter.create<arith::MulFOp>(loc, x, y);
|
|
}
|
|
|
|
if (!acc)
|
|
return std::optional<Value>(mul);
|
|
|
|
return makeArithReduction(rewriter, loc, kind, mul, acc,
|
|
/*fastmath=*/nullptr, mask);
|
|
}
|
|
|
|
/// Return the positions of the reductions in the given map.
|
|
static SmallVector<int64_t> getReductionIndex(AffineMap map,
|
|
ArrayAttr iteratorTypes) {
|
|
SmallVector<int64_t> dimsIdx;
|
|
for (unsigned i = 0, e = map.getNumResults(); i < e; i++) {
|
|
if (isReductionIterator(iteratorTypes[map.getDimPosition(i)]))
|
|
dimsIdx.push_back(i);
|
|
}
|
|
return dimsIdx;
|
|
}
|
|
|
|
/// Look for a given dimension in an affine map and return its position. Return
|
|
/// std::nullopt if the dimension is not in the map results.
|
|
static std::optional<unsigned> getDimPosition(AffineMap map, unsigned dim) {
|
|
for (unsigned i = 0, e = map.getNumResults(); i < e; i++) {
|
|
if (map.getDimPosition(i) == dim)
|
|
return i;
|
|
}
|
|
return std::nullopt;
|
|
}
|
|
|
|
/// Creates an AddIOp if `isInt` is true otherwise create an arith::AddFOp using
|
|
/// operands `x` and `y`.
|
|
static Value createAdd(Location loc, Value x, Value y, bool isInt,
|
|
PatternRewriter &rewriter) {
|
|
if (isInt)
|
|
return rewriter.create<arith::AddIOp>(loc, x, y);
|
|
return rewriter.create<arith::AddFOp>(loc, x, y);
|
|
}
|
|
|
|
/// Creates a MulIOp if `isInt` is true otherwise create an MulFOp using
|
|
/// operands `x and `y`.
|
|
static Value createMul(Location loc, Value x, Value y, bool isInt,
|
|
PatternRewriter &rewriter) {
|
|
if (isInt)
|
|
return rewriter.create<arith::MulIOp>(loc, x, y);
|
|
return rewriter.create<arith::MulFOp>(loc, x, y);
|
|
}
|
|
|
|
namespace {
|
|
|
|
/// Progressive lowering of a `vector.contract %a, %b, %c` with row-major matmul
|
|
/// semantics to:
|
|
/// ```
|
|
/// %flattened_a = vector.shape_cast %a
|
|
/// %flattened_b = vector.shape_cast %b
|
|
/// %flattened_d = vector.matmul %flattened_a, %flattened_b
|
|
/// %d = vector.shape_cast %%flattened_d
|
|
/// %e = add %c, %d
|
|
/// ```
|
|
/// `vector.matmul` later lowers to `llvm.matrix.multiply`.
|
|
//
|
|
/// This only kicks in when VectorTransformsOptions is set to OuterProduct and
|
|
/// the vector.contract op is a row-major matrix multiply.
|
|
class ContractionOpToMatmulOpLowering
|
|
: public OpRewritePattern<vector::ContractionOp> {
|
|
public:
|
|
using OpRewritePattern::OpRewritePattern;
|
|
|
|
using FilterConstraintType =
|
|
std::function<LogicalResult(vector::ContractionOp op)>;
|
|
|
|
static LogicalResult defaultFilter(vector::ContractionOp op) {
|
|
return success();
|
|
}
|
|
|
|
ContractionOpToMatmulOpLowering(
|
|
vector::VectorTransformsOptions vectorTransformOptions,
|
|
MLIRContext *context, PatternBenefit benefit = 1,
|
|
FilterConstraintType constraint = defaultFilter)
|
|
: OpRewritePattern<vector::ContractionOp>(context, benefit),
|
|
vectorTransformOptions(vectorTransformOptions),
|
|
filter(std::move(constraint)) {}
|
|
|
|
LogicalResult matchAndRewrite(vector::ContractionOp op,
|
|
PatternRewriter &rewriter) const override;
|
|
|
|
private:
|
|
/// Options to control the vector patterns.
|
|
vector::VectorTransformsOptions vectorTransformOptions;
|
|
FilterConstraintType filter;
|
|
};
|
|
|
|
/// Progressive lowering of a `vector.contract %a, %b, %c` with row-major matmul
|
|
/// semantics to a reduction_size-unrolled sequence:
|
|
/// ```
|
|
/// %at = vector.transpose %a, [1, 0]
|
|
/// %bRow0 = vector.extract %b[0]
|
|
/// %atRow0 = vector.extract %at[0]
|
|
/// %c0 = vector.outerproduct %atRow0, %bRow0, %c
|
|
/// ...
|
|
/// %bRowK = vector.extract %b[K]
|
|
/// %atRowK = vector.extract %at[K]
|
|
/// %cK = vector.outerproduct %atRowK, %bRowK, %cK-1
|
|
/// ```
|
|
///
|
|
/// This only kicks in when VectorTransformsOptions is set to OuterProduct and
|
|
/// the vector.contract op is a row-major matrix multiply.
|
|
class ContractionOpToOuterProductOpLowering
|
|
: public OpRewritePattern<vector::ContractionOp> {
|
|
public:
|
|
using OpRewritePattern::OpRewritePattern;
|
|
|
|
using FilterConstraintType =
|
|
std::function<LogicalResult(vector::ContractionOp op)>;
|
|
|
|
static LogicalResult defaultFilter(vector::ContractionOp op) {
|
|
return success();
|
|
}
|
|
|
|
ContractionOpToOuterProductOpLowering(
|
|
vector::VectorTransformsOptions vectorTransformOptions,
|
|
MLIRContext *context, PatternBenefit benefit = 1,
|
|
FilterConstraintType constraint = defaultFilter)
|
|
: OpRewritePattern<vector::ContractionOp>(context, benefit),
|
|
vectorTransformOptions(vectorTransformOptions),
|
|
filter(std::move(constraint)) {}
|
|
|
|
LogicalResult matchAndRewrite(vector::ContractionOp op,
|
|
PatternRewriter &rewriter) const override;
|
|
|
|
private:
|
|
/// Options to control the vector patterns.
|
|
vector::VectorTransformsOptions vectorTransformOptions;
|
|
FilterConstraintType filter;
|
|
};
|
|
|
|
/// Progressive lowering of a `vector.contract %a, %b, %c` with row-major matmul
|
|
/// semantics to an output-size-unrolled sequence:
|
|
/// ```
|
|
/// %out = arith.constant ... : vector<MxNxelt_type>
|
|
/// %bt = vector.transpose %b, [1, 0]
|
|
/// %aRow0 = vector.extract %a[0]
|
|
/// %btRow0 = vector.extract %bt[0]
|
|
/// %c00 = vector.reduce %atRow0, %bRow0
|
|
/// %out00 = vector.insert %c00, %out[0, 0]
|
|
/// ...
|
|
/// %aRowLast = vector.extract %at[M-1]
|
|
/// %btRowLast = vector.extract %b[N-1]
|
|
/// %cLastLast = vector.reduce %atRowLast, %bRowLast
|
|
/// %outcLastLast = vector.insert %cLastLast, %out[M-1, N-1]
|
|
/// ```
|
|
///
|
|
/// This only kicks in when VectorTransformsOptions is set to Dot and
|
|
/// the vector.contract op is a row-major matmul or matvec.
|
|
class ContractionOpToDotLowering
|
|
: public OpRewritePattern<vector::ContractionOp> {
|
|
public:
|
|
using OpRewritePattern::OpRewritePattern;
|
|
|
|
using FilterConstraintType =
|
|
std::function<LogicalResult(vector::ContractionOp op)>;
|
|
|
|
static LogicalResult defaultFilter(vector::ContractionOp op) {
|
|
return success();
|
|
}
|
|
|
|
ContractionOpToDotLowering(
|
|
vector::VectorTransformsOptions vectorTransformOptions,
|
|
MLIRContext *context, PatternBenefit benefit = 1,
|
|
const FilterConstraintType &constraint = defaultFilter)
|
|
: OpRewritePattern<vector::ContractionOp>(context, benefit),
|
|
vectorTransformOptions(vectorTransformOptions), filter(defaultFilter) {}
|
|
|
|
LogicalResult matchAndRewrite(vector::ContractionOp op,
|
|
PatternRewriter &rewriter) const override;
|
|
|
|
private:
|
|
/// Options to control the vector patterns.
|
|
vector::VectorTransformsOptions vectorTransformOptions;
|
|
FilterConstraintType filter;
|
|
};
|
|
|
|
/// Progressive lowering of ContractionOp.
|
|
///
|
|
/// One:
|
|
/// %x = vector.contract with at least one free/batch dimension
|
|
/// is replaced by:
|
|
/// %a = vector.contract with one less free/batch dimension
|
|
/// %b = vector.contract with one less free/batch dimension
|
|
/// ..
|
|
/// %x = combine %a %b ..
|
|
/// until a pure contraction is reached (no free/batch dimensions),
|
|
/// which is replaced by a dot-product.
|
|
///
|
|
/// This only kicks in when either VectorTransformsOptions is set
|
|
/// to Dot or when other contraction patterns fail.
|
|
class ContractionOpLowering : public OpRewritePattern<vector::ContractionOp> {
|
|
public:
|
|
using OpRewritePattern::OpRewritePattern;
|
|
using FilterConstraintType =
|
|
std::function<LogicalResult(vector::ContractionOp op)>;
|
|
|
|
static LogicalResult defaultFilter(vector::ContractionOp op) {
|
|
return success();
|
|
}
|
|
|
|
ContractionOpLowering(vector::VectorTransformsOptions vectorTransformOptions,
|
|
MLIRContext *context, PatternBenefit benefit = 1,
|
|
FilterConstraintType constraint = defaultFilter)
|
|
: OpRewritePattern<vector::ContractionOp>(context, benefit),
|
|
vectorTransformOptions(vectorTransformOptions),
|
|
filter(std::move(constraint)) {}
|
|
|
|
LogicalResult matchAndRewrite(vector::ContractionOp op,
|
|
PatternRewriter &rewriter) const override;
|
|
|
|
private:
|
|
/// Options to control the vector patterns.
|
|
vector::VectorTransformsOptions vectorTransformOptions;
|
|
FilterConstraintType filter;
|
|
// Lower one parallel dimension.
|
|
FailureOr<Value> lowerParallel(PatternRewriter &rewriter,
|
|
vector::ContractionOp op, int64_t lhsIndex,
|
|
int64_t rhsIndex, Value mask) const;
|
|
// Lower one reduction dimension.
|
|
FailureOr<Value> lowerReduction(PatternRewriter &rewriter,
|
|
vector::ContractionOp op, Value mask) const;
|
|
};
|
|
|
|
/// Generate a vector implementation for matmat, matvec and tmatvec.
|
|
/// This unrolls outer-products along the reduction dimension.
|
|
struct UnrolledOuterProductGenerator
|
|
: public StructuredGenerator<vector::ContractionOp, vector::IteratorType> {
|
|
UnrolledOuterProductGenerator(RewriterBase &b, vector::ContractionOp op)
|
|
: StructuredGenerator<vector::ContractionOp, vector::IteratorType>(b, op),
|
|
kind(op.getKind()), lhs(op.getLhs()), rhs(op.getRhs()),
|
|
res(op.getAcc()), lhsType(op.getLhsType()) {
|
|
auto maskableOp = cast<MaskableOpInterface>(op.getOperation());
|
|
if (maskableOp.isMasked())
|
|
mask = maskableOp.getMaskingOp().getMask();
|
|
}
|
|
|
|
Value t(Value v, ArrayRef<int64_t> perm = {1, 0}) {
|
|
if (!v)
|
|
return v;
|
|
return rewriter.create<vector::TransposeOp>(loc, v, perm);
|
|
}
|
|
|
|
Value promote(Value v, Type dstElementType) {
|
|
Type elementType = v.getType();
|
|
auto vecType = dyn_cast<VectorType>(elementType);
|
|
if (vecType)
|
|
elementType = vecType.getElementType();
|
|
if (elementType == dstElementType)
|
|
return v;
|
|
Type promotedType = dstElementType;
|
|
if (vecType)
|
|
promotedType = vecType.clone(promotedType);
|
|
if (isa<FloatType>(dstElementType))
|
|
return rewriter.create<arith::ExtFOp>(loc, promotedType, v);
|
|
return rewriter.create<arith::ExtSIOp>(loc, promotedType, v);
|
|
}
|
|
|
|
FailureOr<Value> outerProd(Value lhs, Value rhs, Value res,
|
|
VectorType lhsType, int reductionDim,
|
|
std::optional<Value> maybeMask = std::nullopt) {
|
|
// Unrolling a scalable dimension would be incorrect - bail out.
|
|
if (lhsType.getScalableDims()[reductionDim])
|
|
return failure();
|
|
|
|
int reductionSize = lhsType.getDimSize(reductionDim);
|
|
assert(reductionSize > 0 &&
|
|
"Reduction dim must be a known static size to allow unrolling");
|
|
|
|
// Incremental support for masking.
|
|
if (mask && !maybeMask.has_value())
|
|
return failure();
|
|
|
|
Type resElementType = cast<VectorType>(res.getType()).getElementType();
|
|
for (int64_t k = 0; k < reductionSize; ++k) {
|
|
Value extractA = rewriter.create<vector::ExtractOp>(loc, lhs, k);
|
|
Value extractB = rewriter.create<vector::ExtractOp>(loc, rhs, k);
|
|
extractA = promote(extractA, resElementType);
|
|
extractB = promote(extractB, resElementType);
|
|
Value extractMask;
|
|
if (maybeMask.has_value() && maybeMask.value())
|
|
extractMask =
|
|
rewriter.create<vector::ExtractOp>(loc, maybeMask.value(), k);
|
|
|
|
Operation *outerProdOp = rewriter.create<vector::OuterProductOp>(
|
|
loc, res.getType(), extractA, extractB, res, kind);
|
|
res = maskOperation(rewriter, outerProdOp, extractMask)->getResult(0);
|
|
}
|
|
return res;
|
|
}
|
|
|
|
/// Two outer parallel, one inner reduction (matmat flavor).
|
|
FailureOr<Value> matmat() {
|
|
if (!iters({Par(), Par(), Red()}))
|
|
return failure();
|
|
// Set up the parallel/reduction structure in the right form.
|
|
AffineExpr m, n, k;
|
|
bindDims(rewriter.getContext(), m, n, k);
|
|
Value transposedMask = t(mask, {2, 0, 1});
|
|
// Classical row-major matmul: Just permute the lhs.
|
|
if (layout({{m, k}, {k, n}, {m, n}}))
|
|
return outerProd(t(lhs), rhs, res, lhsType, /*reductionDim=*/1,
|
|
transposedMask);
|
|
// TODO: may be better to fail and use some vector<k> -> scalar reduction.
|
|
if (layout({{m, k}, {n, k}, {m, n}})) {
|
|
Value tlhs = t(lhs);
|
|
return outerProd(tlhs, t(rhs), res, lhsType, /*reductionDim=*/1,
|
|
transposedMask);
|
|
}
|
|
// No need to permute anything.
|
|
if (layout({{k, m}, {k, n}, {m, n}}))
|
|
return outerProd(lhs, rhs, res, lhsType, /*reductionDim=*/0,
|
|
transposedMask);
|
|
// Just permute the rhs.
|
|
if (layout({{k, m}, {n, k}, {m, n}}))
|
|
return outerProd(lhs, t(rhs), res, lhsType, /*reductionDim=*/0,
|
|
transposedMask);
|
|
// Transposed output: swap RHS and LHS.
|
|
// Classical row-major matmul: permute the lhs.
|
|
if (layout({{m, k}, {k, n}, {n, m}}))
|
|
return outerProd(rhs, t(lhs), res, lhsType, /*reductionDim=*/1,
|
|
transposedMask);
|
|
// TODO: may be better to fail and use some vector<k> -> scalar reduction.
|
|
if (layout({{m, k}, {n, k}, {n, m}})) {
|
|
Value trhs = t(rhs);
|
|
return outerProd(trhs, t(lhs), res, lhsType, /*reductionDim=*/1,
|
|
transposedMask);
|
|
}
|
|
if (layout({{k, m}, {k, n}, {n, m}}))
|
|
return outerProd(rhs, lhs, res, lhsType, /*reductionDim=*/0,
|
|
transposedMask);
|
|
if (layout({{k, m}, {n, k}, {n, m}}))
|
|
return outerProd(t(rhs), lhs, res, lhsType, /*reductionDim=*/0,
|
|
transposedMask);
|
|
return failure();
|
|
}
|
|
|
|
//
|
|
// One outer parallel, one inner reduction (matvec flavor).
|
|
// Mask needs to be transposed everywhere to turn the reduction dimension
|
|
// outermost as required by outerproduct.
|
|
//
|
|
FailureOr<Value> matvec() {
|
|
if (!iters({Par(), Red()}))
|
|
return failure();
|
|
AffineExpr m, k;
|
|
bindDims(rewriter.getContext(), m, k);
|
|
Value transposedMask = t(mask);
|
|
|
|
// Case mat-vec: transpose.
|
|
if (layout({{m, k}, {k}, {m}}))
|
|
return outerProd(t(lhs), rhs, res, lhsType, /*reductionDim=*/1,
|
|
transposedMask);
|
|
// Case mat-trans-vec: ready to go.
|
|
if (layout({{k, m}, {k}, {m}}))
|
|
return outerProd(lhs, rhs, res, lhsType, /*reductionDim=*/0,
|
|
transposedMask);
|
|
// Case vec-mat: swap and transpose.
|
|
if (layout({{k}, {m, k}, {m}}))
|
|
return outerProd(t(rhs), lhs, res, lhsType, /*reductionDim=*/0,
|
|
transposedMask);
|
|
// Case vec-mat-trans: swap and ready to go.
|
|
if (layout({{k}, {k, m}, {m}}))
|
|
return outerProd(rhs, lhs, res, lhsType, /*reductionDim=*/0,
|
|
transposedMask);
|
|
return failure();
|
|
}
|
|
|
|
//
|
|
// One outer reduction, one inner parallel (tmatvec flavor).
|
|
// Mask already has the shape of the outer product.
|
|
//
|
|
FailureOr<Value> tmatvec() {
|
|
if (!iters({Red(), Par()}))
|
|
return failure();
|
|
AffineExpr k, m;
|
|
bindDims(rewriter.getContext(), k, m);
|
|
|
|
// Case mat-vec: transpose.
|
|
if (layout({{m, k}, {k}, {m}}))
|
|
return outerProd(t(lhs), rhs, res, lhsType, /*reductionDim=*/1, mask);
|
|
// Case mat-trans-vec: ready to go.
|
|
if (layout({{k, m}, {k}, {m}}))
|
|
return outerProd(lhs, rhs, res, lhsType, /*reductionDim=*/0, mask);
|
|
// Case vec-mat: swap and transpose.
|
|
if (layout({{k}, {m, k}, {m}}))
|
|
return outerProd(t(rhs), lhs, res, lhsType, /*reductionDim=*/0, mask);
|
|
// Case vec-mat-trans: swap and ready to go.
|
|
if (layout({{k}, {k, m}, {m}}))
|
|
return outerProd(rhs, lhs, res, lhsType, /*reductionDim=*/0, mask);
|
|
return failure();
|
|
}
|
|
|
|
private:
|
|
vector::CombiningKind kind;
|
|
Value lhs, rhs, res, mask;
|
|
VectorType lhsType;
|
|
};
|
|
|
|
/// Progressively lower a `vector.contract %a, %b, %c` with row-major matmul
|
|
/// semantics to a reduction_size-unrolled sequence:
|
|
/// ```
|
|
/// %at = vector.transpose %a, [1, 0]
|
|
/// %bRow0 = vector.extract %b[0]
|
|
/// %atRow0 = vector.extract %at[0]
|
|
/// %c0 = vector.outerproduct %atRow0, %bRow0, %c
|
|
/// ...
|
|
/// %bRowK = vector.extract %b[K]
|
|
/// %atRowK = vector.extract %at[K]
|
|
/// %cK = vector.outerproduct %atRowK, %bRowK, %cK-1
|
|
/// ```
|
|
///
|
|
/// This only kicks in when VectorTransformsOptions is set to OuterProduct but
|
|
/// otherwise supports any layout permutation of the matrix-multiply.
|
|
LogicalResult ContractionOpToOuterProductOpLowering::matchAndRewrite(
|
|
vector::ContractionOp op, PatternRewriter &rewriter) const {
|
|
if (vectorTransformOptions.vectorContractLowering !=
|
|
vector::VectorContractLowering::OuterProduct)
|
|
return failure();
|
|
|
|
if (failed(filter(op)))
|
|
return failure();
|
|
|
|
// Vector mask setup.
|
|
OpBuilder::InsertionGuard guard(rewriter);
|
|
auto maskableOp = cast<vector::MaskableOpInterface>(op.getOperation());
|
|
Operation *rootOp;
|
|
if (maskableOp.isMasked()) {
|
|
rewriter.setInsertionPoint(maskableOp.getMaskingOp());
|
|
rootOp = maskableOp.getMaskingOp();
|
|
} else {
|
|
rootOp = op;
|
|
}
|
|
|
|
UnrolledOuterProductGenerator e(rewriter, op);
|
|
FailureOr<Value> matmatRes = e.matmat();
|
|
if (succeeded(matmatRes)) {
|
|
rewriter.replaceOp(rootOp, *matmatRes);
|
|
return success();
|
|
}
|
|
FailureOr<Value> matvecRes = e.matvec();
|
|
if (succeeded(matvecRes)) {
|
|
rewriter.replaceOp(rootOp, *matvecRes);
|
|
return success();
|
|
}
|
|
FailureOr<Value> tmatvecRes = e.tmatvec();
|
|
if (succeeded(tmatvecRes)) {
|
|
rewriter.replaceOp(rootOp, *tmatvecRes);
|
|
return success();
|
|
}
|
|
|
|
return failure();
|
|
}
|
|
|
|
LogicalResult
|
|
ContractionOpToDotLowering::matchAndRewrite(vector::ContractionOp op,
|
|
PatternRewriter &rewriter) const {
|
|
// TODO: Support vector.mask.
|
|
auto maskableOp = cast<MaskableOpInterface>(op.getOperation());
|
|
if (maskableOp.isMasked())
|
|
return failure();
|
|
|
|
if (failed(filter(op)))
|
|
return failure();
|
|
|
|
if (vectorTransformOptions.vectorContractLowering !=
|
|
vector::VectorContractLowering::Dot)
|
|
return failure();
|
|
|
|
auto iteratorTypes = op.getIteratorTypes().getValue();
|
|
static constexpr std::array<int64_t, 2> perm = {1, 0};
|
|
Location loc = op.getLoc();
|
|
Value lhs = op.getLhs(), rhs = op.getRhs();
|
|
|
|
using MapList = ArrayRef<ArrayRef<AffineExpr>>;
|
|
auto infer = [](MapList m) { return AffineMap::inferFromExprList(m); };
|
|
AffineExpr m, n, k;
|
|
bindDims(rewriter.getContext(), m, n, k);
|
|
SmallVector<AffineMap> maps = op.getIndexingMapsArray();
|
|
//
|
|
// In the following we wish to make the reduction dimension innermost so we
|
|
// can load vectors and just fmul + reduce into a scalar.
|
|
//
|
|
if (isParallelIterator(iteratorTypes[0]) &&
|
|
isParallelIterator(iteratorTypes[1]) &&
|
|
isReductionIterator(iteratorTypes[2])) {
|
|
//
|
|
// Two outer parallel, one inner reduction (matmat flavor).
|
|
//
|
|
if (maps == infer({{m, k}, {k, n}, {m, n}})) {
|
|
rhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm);
|
|
} else if (maps == infer({{m, k}, {n, k}, {m, n}})) {
|
|
// No need to permute anything.
|
|
} else if (maps == infer({{k, m}, {k, n}, {m, n}})) {
|
|
lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
|
|
rhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm);
|
|
} else if (maps == infer({{k, m}, {n, k}, {m, n}})) {
|
|
lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
|
|
} else if (maps == infer({{m, k}, {k, n}, {n, m}})) {
|
|
// This is the classical row-major matmul. Just permute the lhs.
|
|
Value tmp = lhs;
|
|
lhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm);
|
|
rhs = tmp;
|
|
} else if (maps == infer({{m, k}, {n, k}, {n, m}})) {
|
|
std::swap(lhs, rhs);
|
|
} else if (maps == infer({{k, m}, {k, n}, {n, m}})) {
|
|
Value tmp = lhs;
|
|
lhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm);
|
|
rhs = rewriter.create<vector::TransposeOp>(loc, tmp, perm);
|
|
} else if (maps == infer({{k, m}, {n, k}, {n, m}})) {
|
|
Value tmp = rhs;
|
|
rhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
|
|
lhs = tmp;
|
|
} else {
|
|
return failure();
|
|
}
|
|
} else if (isParallelIterator(iteratorTypes[0]) &&
|
|
isReductionIterator(iteratorTypes[1])) {
|
|
//
|
|
// One outer parallel, one inner reduction (matvec flavor)
|
|
//
|
|
if (maps == infer({{m, n}, {n}, {m}})) {
|
|
// No need to permute anything.
|
|
} else if (maps == infer({{n, m}, {n}, {m}})) {
|
|
lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
|
|
} else if (maps == infer({{n}, {m, n}, {m}})) {
|
|
std::swap(lhs, rhs);
|
|
} else if (maps == infer({{n}, {n, m}, {m}})) {
|
|
std::swap(lhs, rhs);
|
|
lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
|
|
} else {
|
|
return failure();
|
|
}
|
|
} else {
|
|
return failure();
|
|
}
|
|
|
|
VectorType dstType = cast<VectorType>(op.getResultType());
|
|
assert(dstType.getRank() >= 1 && dstType.getRank() <= 2 &&
|
|
"Expected dst type of rank 1 or 2");
|
|
|
|
unsigned rank = dstType.getRank();
|
|
unsigned dstRows = dstType.getShape()[0];
|
|
unsigned dstColumns = rank == 1 ? 1 : dstType.getShape()[1];
|
|
|
|
// ExtractOp does not allow dynamic indexing, we must unroll explicitly.
|
|
Value res = rewriter.create<arith::ConstantOp>(loc, dstType,
|
|
rewriter.getZeroAttr(dstType));
|
|
bool isInt = isa<IntegerType>(dstType.getElementType());
|
|
for (unsigned r = 0; r < dstRows; ++r) {
|
|
Value a = rewriter.create<vector::ExtractOp>(op.getLoc(), lhs, r);
|
|
for (unsigned c = 0; c < dstColumns; ++c) {
|
|
Value b = rank == 1
|
|
? rhs
|
|
: rewriter.create<vector::ExtractOp>(op.getLoc(), rhs, c);
|
|
Value m = createMul(op.getLoc(), a, b, isInt, rewriter);
|
|
Value reduced = rewriter.create<vector::ReductionOp>(
|
|
op.getLoc(), vector::CombiningKind::ADD, m);
|
|
|
|
SmallVector<int64_t, 2> pos = rank == 1 ? SmallVector<int64_t, 2>{r}
|
|
: SmallVector<int64_t, 2>{r, c};
|
|
res = rewriter.create<vector::InsertOp>(op.getLoc(), reduced, res, pos);
|
|
}
|
|
}
|
|
if (auto acc = op.getAcc())
|
|
res = createAdd(op.getLoc(), res, acc, isInt, rewriter);
|
|
rewriter.replaceOp(op, res);
|
|
return success();
|
|
}
|
|
|
|
/// Lower vector.contract with all size one reduction dimensions to
|
|
/// elementwise ops when possible.
|
|
struct ContractOpToElementwise
|
|
: public OpRewritePattern<vector::ContractionOp> {
|
|
using OpRewritePattern::OpRewritePattern;
|
|
using FilterConstraintType =
|
|
std::function<LogicalResult(vector::ContractionOp op)>;
|
|
static LogicalResult defaultFilter(vector::ContractionOp op) {
|
|
return success();
|
|
}
|
|
ContractOpToElementwise(
|
|
vector::VectorTransformsOptions vectorTransformOptions,
|
|
MLIRContext *context, PatternBenefit benefit = 1,
|
|
const FilterConstraintType &constraint = defaultFilter)
|
|
: OpRewritePattern<vector::ContractionOp>(context, benefit),
|
|
vectorTransformOptions(vectorTransformOptions), filter(defaultFilter) {}
|
|
|
|
LogicalResult matchAndRewrite(vector::ContractionOp contractOp,
|
|
PatternRewriter &rewriter) const override {
|
|
// TODO: Support vector.mask.
|
|
auto maskableOp = cast<MaskableOpInterface>(contractOp.getOperation());
|
|
if (maskableOp.isMasked())
|
|
return failure();
|
|
|
|
if (failed(filter(contractOp)))
|
|
return failure();
|
|
|
|
if (vectorTransformOptions.vectorContractLowering !=
|
|
vector::VectorContractLowering::ParallelArith)
|
|
return failure();
|
|
|
|
ArrayRef<int64_t> lhsShape = contractOp.getLhsType().getShape();
|
|
ArrayRef<int64_t> rhsShape = contractOp.getRhsType().getShape();
|
|
AffineMap lhsMap = contractOp.getIndexingMapsArray()[0];
|
|
AffineMap rhsMap = contractOp.getIndexingMapsArray()[1];
|
|
SmallVector<int64_t> lhsReductionDims =
|
|
getReductionIndex(lhsMap, contractOp.getIteratorTypes());
|
|
SmallVector<int64_t> rhsReductionDims =
|
|
getReductionIndex(rhsMap, contractOp.getIteratorTypes());
|
|
// All the reduction dimensions must be a size 1.
|
|
for (int64_t dim : lhsReductionDims) {
|
|
if (lhsShape[dim] != 1)
|
|
return failure();
|
|
}
|
|
for (int64_t dim : rhsReductionDims) {
|
|
if (rhsShape[dim] != 1)
|
|
return failure();
|
|
}
|
|
AffineMap accMap = contractOp.getIndexingMapsArray()[2];
|
|
unsigned numParallelDims = accMap.getNumResults();
|
|
unsigned numLhsDimToBroadcast =
|
|
numParallelDims - (lhsMap.getNumResults() - lhsReductionDims.size());
|
|
unsigned numRhsDimToBroadcast =
|
|
numParallelDims - (rhsMap.getNumResults() - rhsReductionDims.size());
|
|
SmallVector<int64_t> lhsDims;
|
|
SmallVector<int64_t> lhsTranspose;
|
|
SmallVector<int64_t> rhsDims;
|
|
SmallVector<int64_t> rhsTranspose;
|
|
for (int64_t dim : lhsReductionDims)
|
|
lhsTranspose.push_back(numLhsDimToBroadcast + dim);
|
|
for (int64_t dim : rhsReductionDims)
|
|
rhsTranspose.push_back(numRhsDimToBroadcast + dim);
|
|
// Loop through the parallel dimensions to calculate the dimensions to
|
|
// broadcast and to permute in order to extract only parallel dimensions.
|
|
for (unsigned i = 0; i < numParallelDims; i++) {
|
|
std::optional<unsigned> lhsDim =
|
|
getDimPosition(lhsMap, accMap.getDimPosition(i));
|
|
if (lhsDim) {
|
|
lhsTranspose.push_back(numLhsDimToBroadcast + *lhsDim);
|
|
} else {
|
|
// If the parallel dimension doesn't exist we will have to broadcast it.
|
|
lhsDims.push_back(
|
|
cast<VectorType>(contractOp.getResultType()).getDimSize(i));
|
|
lhsTranspose.push_back(lhsDims.size() - 1);
|
|
}
|
|
std::optional<unsigned> rhsDim =
|
|
getDimPosition(rhsMap, accMap.getDimPosition(i));
|
|
if (rhsDim) {
|
|
rhsTranspose.push_back(numRhsDimToBroadcast + *rhsDim);
|
|
} else {
|
|
// If the parallel dimension doesn't exist we will have to broadcast it.
|
|
rhsDims.push_back(
|
|
cast<VectorType>(contractOp.getResultType()).getDimSize(i));
|
|
rhsTranspose.push_back(rhsDims.size() - 1);
|
|
}
|
|
}
|
|
Value newLhs = contractOp.getLhs();
|
|
Value newRhs = contractOp.getRhs();
|
|
Location loc = contractOp.getLoc();
|
|
if (!lhsDims.empty()) {
|
|
lhsDims.append(lhsShape.begin(), lhsShape.end());
|
|
auto expandedType =
|
|
VectorType::get(lhsDims, contractOp.getLhsType().getElementType());
|
|
newLhs = rewriter.create<vector::BroadcastOp>(loc, expandedType, newLhs);
|
|
}
|
|
if (!rhsDims.empty()) {
|
|
rhsDims.append(rhsShape.begin(), rhsShape.end());
|
|
auto expandedType =
|
|
VectorType::get(rhsDims, contractOp.getRhsType().getElementType());
|
|
newRhs = rewriter.create<vector::BroadcastOp>(loc, expandedType, newRhs);
|
|
}
|
|
bool isInt = contractOp.getLhsType().getElementType().isIntOrIndex();
|
|
newLhs = rewriter.create<vector::TransposeOp>(loc, newLhs, lhsTranspose);
|
|
newRhs = rewriter.create<vector::TransposeOp>(loc, newRhs, rhsTranspose);
|
|
SmallVector<int64_t> lhsOffsets(lhsReductionDims.size(), 0);
|
|
SmallVector<int64_t> rhsOffsets(rhsReductionDims.size(), 0);
|
|
newLhs = rewriter.create<vector::ExtractOp>(loc, newLhs, lhsOffsets);
|
|
newRhs = rewriter.create<vector::ExtractOp>(loc, newRhs, rhsOffsets);
|
|
std::optional<Value> result =
|
|
createContractArithOp(loc, newLhs, newRhs, contractOp.getAcc(),
|
|
contractOp.getKind(), rewriter, isInt);
|
|
rewriter.replaceOp(contractOp, {*result});
|
|
return success();
|
|
}
|
|
|
|
private:
|
|
/// Options to control the vector patterns.
|
|
vector::VectorTransformsOptions vectorTransformOptions;
|
|
FilterConstraintType filter;
|
|
};
|
|
|
|
/// Progressive lowering of ContractionOp.
|
|
/// One:
|
|
/// %x = vector.contract with at least one free/batch dimension
|
|
/// is replaced by:
|
|
/// %a = vector.contract with one less free/batch dimension
|
|
/// %b = vector.contract with one less free/batch dimension
|
|
/// ..
|
|
/// %x = combine %a %b ..
|
|
/// until a pure contraction is reached (no free/batch dimensions),
|
|
/// which is replaced by a dot-product.
|
|
///
|
|
/// This only kicks in when either VectorTransformsOptions is set
|
|
/// to DOT or when other contraction patterns fail.
|
|
//
|
|
// TODO: break down into transpose/reshape/cast ops
|
|
// when they become available to avoid code dup
|
|
// TODO: investigate lowering order impact on performance
|
|
LogicalResult
|
|
ContractionOpLowering::matchAndRewrite(vector::ContractionOp op,
|
|
PatternRewriter &rewriter) const {
|
|
if (failed(filter(op)))
|
|
return failure();
|
|
|
|
// TODO: support mixed mode contract lowering.
|
|
if (op.getLhsType().getElementType() !=
|
|
getElementTypeOrSelf(op.getAccType()) ||
|
|
op.getRhsType().getElementType() != getElementTypeOrSelf(op.getAccType()))
|
|
return failure();
|
|
|
|
// TODO: the code below assumes the default contraction, make sure it supports
|
|
// other kinds before enabling this lowering.
|
|
if (op.getKind() != vector::CombiningKind::ADD) {
|
|
return rewriter.notifyMatchFailure(
|
|
op, "contractions other than 'add' not supported");
|
|
}
|
|
|
|
// TODO: implement benefits, cost models.
|
|
MLIRContext *ctx = op.getContext();
|
|
ContractionOpToMatmulOpLowering pat1(vectorTransformOptions, ctx);
|
|
if (succeeded(pat1.matchAndRewrite(op, rewriter)))
|
|
return success();
|
|
ContractionOpToOuterProductOpLowering pat2(vectorTransformOptions, ctx);
|
|
if (succeeded(pat2.matchAndRewrite(op, rewriter)))
|
|
return success();
|
|
ContractionOpToDotLowering pat3(vectorTransformOptions, ctx);
|
|
if (succeeded(pat3.matchAndRewrite(op, rewriter)))
|
|
return success();
|
|
ContractOpToElementwise pat4(vectorTransformOptions, ctx);
|
|
if (succeeded(pat4.matchAndRewrite(op, rewriter)))
|
|
return success();
|
|
|
|
// Vector mask setup.
|
|
OpBuilder::InsertionGuard guard(rewriter);
|
|
Operation *rootOp = op;
|
|
Value mask;
|
|
if (op.isMasked()) {
|
|
rewriter.setInsertionPoint(op.getMaskingOp());
|
|
rootOp = op.getMaskingOp();
|
|
mask = op.getMaskingOp().getMask();
|
|
}
|
|
|
|
// Find first batch dimension in LHS/RHS, and lower when found.
|
|
std::vector<std::pair<int64_t, int64_t>> batchDimMap = op.getBatchDimMap();
|
|
if (!batchDimMap.empty()) {
|
|
int64_t lhsIndex = batchDimMap[0].first;
|
|
int64_t rhsIndex = batchDimMap[0].second;
|
|
auto newOp = lowerParallel(rewriter, op, lhsIndex, rhsIndex, mask);
|
|
if (failed(newOp))
|
|
return failure();
|
|
rewriter.replaceOp(rootOp, *newOp);
|
|
return success();
|
|
}
|
|
|
|
// Collect contracting dimensions.
|
|
std::vector<std::pair<int64_t, int64_t>> contractingDimMap =
|
|
op.getContractingDimMap();
|
|
DenseSet<int64_t> lhsContractingDimSet;
|
|
DenseSet<int64_t> rhsContractingDimSet;
|
|
for (auto &dimPair : contractingDimMap) {
|
|
lhsContractingDimSet.insert(dimPair.first);
|
|
rhsContractingDimSet.insert(dimPair.second);
|
|
}
|
|
|
|
// Find first free dimension in LHS, and lower when found.
|
|
VectorType lhsType = op.getLhsType();
|
|
for (int64_t lhsIndex = 0, e = lhsType.getRank(); lhsIndex < e; ++lhsIndex) {
|
|
if (lhsContractingDimSet.count(lhsIndex) == 0) {
|
|
auto newOp = lowerParallel(rewriter, op, lhsIndex, /*rhsIndex=*/-1, mask);
|
|
if (failed(newOp))
|
|
return failure();
|
|
rewriter.replaceOp(rootOp, *newOp);
|
|
return success();
|
|
}
|
|
}
|
|
|
|
// Find first free dimension in RHS, and lower when found.
|
|
VectorType rhsType = op.getRhsType();
|
|
for (int64_t rhsIndex = 0, e = rhsType.getRank(); rhsIndex < e; ++rhsIndex) {
|
|
if (rhsContractingDimSet.count(rhsIndex) == 0) {
|
|
auto newOp = lowerParallel(rewriter, op, /*lhsIndex=*/-1, rhsIndex, mask);
|
|
if (failed(newOp))
|
|
return failure();
|
|
rewriter.replaceOp(rootOp, *newOp);
|
|
return success();
|
|
}
|
|
}
|
|
|
|
// Lower the first remaining reduction dimension.
|
|
if (!contractingDimMap.empty()) {
|
|
auto newOp = lowerReduction(rewriter, op, mask);
|
|
if (failed(newOp))
|
|
return failure();
|
|
rewriter.replaceOp(rootOp, *newOp);
|
|
return success();
|
|
}
|
|
|
|
return failure();
|
|
}
|
|
|
|
// Lower one parallel dimension.
|
|
// Incidentally also tolerates unit-size (hence trivial) reduction dimensions.
|
|
// TODO: consider reusing existing contract unrolling
|
|
FailureOr<Value> ContractionOpLowering::lowerParallel(PatternRewriter &rewriter,
|
|
vector::ContractionOp op,
|
|
int64_t lhsIndex,
|
|
int64_t rhsIndex,
|
|
Value mask) const {
|
|
VectorType lhsType = op.getLhsType();
|
|
VectorType rhsType = op.getRhsType();
|
|
VectorType resType = cast<VectorType>(op.getResultType());
|
|
// Find the iterator type index and result index.
|
|
SmallVector<AffineMap> iMap = op.getIndexingMapsArray();
|
|
int64_t iterIndex = -1;
|
|
int64_t dimSize = -1;
|
|
if (lhsIndex >= 0) {
|
|
iterIndex = iMap[0].getDimPosition(lhsIndex);
|
|
if (rhsIndex >= 0 && iterIndex != iMap[1].getDimPosition(rhsIndex))
|
|
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
|
|
diag << "expected lhsIndex=" << lhsIndex << " and rhsIndex=" << rhsIndex
|
|
<< " to map to the same dimension";
|
|
});
|
|
if (lhsType.getScalableDims()[lhsIndex])
|
|
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
|
|
diag << "Unrolling scalable dimension (lhsIndex=" << lhsIndex
|
|
<< ") is not supported yet";
|
|
});
|
|
dimSize = lhsType.getDimSize(lhsIndex);
|
|
} else if (rhsIndex >= 0) {
|
|
iterIndex = iMap[1].getDimPosition(rhsIndex);
|
|
if (rhsType.getScalableDims()[rhsIndex])
|
|
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
|
|
diag << "Unrolling scalable dimension (rhsIndex=" << rhsIndex
|
|
<< ") is not supported yet";
|
|
});
|
|
dimSize = rhsType.getDimSize(rhsIndex);
|
|
}
|
|
if (iterIndex < 0)
|
|
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
|
|
diag << "expected either lhsIndex=" << lhsIndex
|
|
<< " or rhsIndex=" << rhsIndex << " to be nonnegative";
|
|
});
|
|
// value_or(-1) means that we tolerate a dimension not appearing
|
|
// in the result map. That can't happen for actual parallel iterators, but
|
|
// the caller ContractionOpLowering::matchAndRewrite is currently calling
|
|
// lowerParallel also for the case of unit-size reduction dims appearing only
|
|
// on one of LHS or RHS, not both. At the moment, such cases are created by
|
|
// CastAwayContractionLeadingOneDim, so we need to either support that or
|
|
// modify that pattern.
|
|
int64_t resIndex = getResultIndex(iMap[2], iterIndex).value_or(-1);
|
|
if (resIndex == -1 && dimSize != 1)
|
|
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
|
|
diag << "expected the dimension for iterIndex=" << iterIndex
|
|
<< " to either appear in the result map, or to be a unit dimension";
|
|
});
|
|
|
|
// Construct new iterator types and affine map array attribute.
|
|
std::array<AffineMap, 3> lowIndexingMaps = {
|
|
adjustMap(iMap[0], iterIndex, rewriter),
|
|
adjustMap(iMap[1], iterIndex, rewriter),
|
|
adjustMap(iMap[2], iterIndex, rewriter)};
|
|
auto lowAffine = rewriter.getAffineMapArrayAttr(lowIndexingMaps);
|
|
auto lowIter =
|
|
rewriter.getArrayAttr(adjustIter(op.getIteratorTypes(), iterIndex));
|
|
// Unroll into a series of lower dimensional vector.contract ops.
|
|
Location loc = op.getLoc();
|
|
Value result = rewriter.create<arith::ConstantOp>(
|
|
loc, resType, rewriter.getZeroAttr(resType));
|
|
|
|
for (int64_t d = 0; d < dimSize; ++d) {
|
|
auto lhs = reshapeLoad(loc, op.getLhs(), lhsType, lhsIndex, d, rewriter);
|
|
auto rhs = reshapeLoad(loc, op.getRhs(), rhsType, rhsIndex, d, rewriter);
|
|
auto acc = reshapeLoad(loc, op.getAcc(), resType, resIndex, d, rewriter);
|
|
|
|
Value lowMask;
|
|
if (mask)
|
|
lowMask = reshapeLoad(loc, mask, cast<VectorType>(mask.getType()),
|
|
iterIndex, d, rewriter);
|
|
|
|
Operation *lowContract = rewriter.create<vector::ContractionOp>(
|
|
loc, lhs, rhs, acc, lowAffine, lowIter);
|
|
lowContract = maskOperation(rewriter, lowContract, lowMask);
|
|
result = reshapeStore(loc, lowContract->getResult(0), result, resType,
|
|
resIndex, d, rewriter);
|
|
}
|
|
return result;
|
|
}
|
|
|
|
// Lower one reduction dimension.
|
|
FailureOr<Value> ContractionOpLowering::lowerReduction(
|
|
PatternRewriter &rewriter, vector::ContractionOp op, Value mask) const {
|
|
auto loc = op.getLoc();
|
|
VectorType lhsType = op.getLhsType();
|
|
VectorType rhsType = op.getRhsType();
|
|
Type resType = op.getResultType();
|
|
if (isa<VectorType>(resType))
|
|
return rewriter.notifyMatchFailure(op,
|
|
"did not expect a VectorType result");
|
|
bool isInt = isa<IntegerType>(resType);
|
|
// Use iterator index 0.
|
|
int64_t iterIndex = 0;
|
|
SmallVector<AffineMap> iMap = op.getIndexingMapsArray();
|
|
std::optional<int64_t> lookupLhs = getResultIndex(iMap[0], iterIndex);
|
|
std::optional<int64_t> lookupRhs = getResultIndex(iMap[1], iterIndex);
|
|
if (!lookupLhs.has_value())
|
|
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
|
|
diag << "expected iterIndex=" << iterIndex << "to map to a LHS dimension";
|
|
});
|
|
if (!lookupRhs.has_value())
|
|
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
|
|
diag << "expected iterIndex=" << iterIndex << "to map to a RHS dimension";
|
|
});
|
|
int64_t lhsIndex = *lookupLhs;
|
|
int64_t rhsIndex = *lookupRhs;
|
|
int64_t dimSize = lhsType.getDimSize(lhsIndex);
|
|
if (dimSize != rhsType.getDimSize(rhsIndex))
|
|
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
|
|
diag << "expect LHS dimension " << lhsIndex
|
|
<< " to have the same size as RHS dimension " << rhsIndex;
|
|
});
|
|
// Base case.
|
|
if (lhsType.getRank() == 1) {
|
|
if (rhsType.getRank() != 1)
|
|
return rewriter.notifyMatchFailure(
|
|
op, "When LHS has rank 1, expected also RHS to have rank 1");
|
|
Value m = createMul(loc, op.getLhs(), op.getRhs(), isInt, rewriter);
|
|
auto kind = vector::CombiningKind::ADD;
|
|
|
|
Value acc = op.getAcc();
|
|
Operation *reductionOp =
|
|
acc ? rewriter.create<vector::ReductionOp>(loc, kind, m, acc)
|
|
: rewriter.create<vector::ReductionOp>(loc, kind, m);
|
|
return maskOperation(rewriter, reductionOp, mask)->getResult(0);
|
|
}
|
|
// Construct new iterator types and affine map array attribute.
|
|
std::array<AffineMap, 3> lowIndexingMaps = {
|
|
adjustMap(iMap[0], iterIndex, rewriter),
|
|
adjustMap(iMap[1], iterIndex, rewriter),
|
|
adjustMap(iMap[2], iterIndex, rewriter)};
|
|
auto lowAffine = rewriter.getAffineMapArrayAttr(lowIndexingMaps);
|
|
auto lowIter =
|
|
rewriter.getArrayAttr(adjustIter(op.getIteratorTypes(), iterIndex));
|
|
// Unroll into a series of lower dimensional vector.contract ops.
|
|
// By feeding the initial accumulator into the first contraction,
|
|
// and the result of each contraction into the next, eventually
|
|
// the sum of all reductions is computed.
|
|
Value result = op.getAcc();
|
|
for (int64_t d = 0; d < dimSize; ++d) {
|
|
auto lhs = reshapeLoad(loc, op.getLhs(), lhsType, lhsIndex, d, rewriter);
|
|
auto rhs = reshapeLoad(loc, op.getRhs(), rhsType, rhsIndex, d, rewriter);
|
|
Value newMask;
|
|
if (mask)
|
|
newMask = reshapeLoad(loc, mask, cast<VectorType>(mask.getType()),
|
|
iterIndex, d, rewriter);
|
|
|
|
Operation *newContract = rewriter.create<vector::ContractionOp>(
|
|
loc, lhs, rhs, result, lowAffine, lowIter);
|
|
result = maskOperation(rewriter, newContract, newMask)->getResult(0);
|
|
}
|
|
return result;
|
|
}
|
|
|
|
/// Progressive lowering of OuterProductOp.
|
|
/// One:
|
|
/// %x = vector.outerproduct %lhs, %rhs, %acc
|
|
/// is replaced by:
|
|
/// %z = zero-result
|
|
/// %0 = vector.extract %lhs[0]
|
|
/// %1 = vector.broadcast %0
|
|
/// %2 = vector.extract %acc[0]
|
|
/// %3 = vector.fma %1, %rhs, %2
|
|
/// %4 = vector.insert %3, %z[0]
|
|
/// ..
|
|
/// %x = vector.insert %.., %..[N-1]
|
|
///
|
|
class OuterProductOpLowering : public OpRewritePattern<vector::OuterProductOp> {
|
|
public:
|
|
using OpRewritePattern::OpRewritePattern;
|
|
|
|
LogicalResult matchAndRewrite(vector::OuterProductOp op,
|
|
PatternRewriter &rewriter) const override {
|
|
VectorType resType = op.getResultVectorType();
|
|
if ((resType.getShape().size() >= 2) && resType.allDimsScalable())
|
|
return failure();
|
|
|
|
auto loc = op.getLoc();
|
|
|
|
VectorType lhsType = op.getOperandVectorTypeLHS();
|
|
VectorType rhsType = dyn_cast<VectorType>(op.getOperandTypeRHS());
|
|
Type eltType = resType.getElementType();
|
|
bool isInt = isa<IntegerType, IndexType>(eltType);
|
|
Value acc = op.getAcc();
|
|
vector::CombiningKind kind = op.getKind();
|
|
|
|
// Vector mask setup.
|
|
OpBuilder::InsertionGuard guard(rewriter);
|
|
auto maskableOp = cast<vector::MaskableOpInterface>(op.getOperation());
|
|
Operation *rootOp;
|
|
Value mask;
|
|
if (maskableOp.isMasked()) {
|
|
rewriter.setInsertionPoint(maskableOp.getMaskingOp());
|
|
rootOp = maskableOp.getMaskingOp();
|
|
mask = maskableOp.getMaskingOp().getMask();
|
|
} else {
|
|
rootOp = op;
|
|
}
|
|
|
|
if (!rhsType) {
|
|
// Special case: AXPY operation.
|
|
Value b = rewriter.create<vector::BroadcastOp>(loc, lhsType, op.getRhs());
|
|
std::optional<Value> mult = createContractArithOp(
|
|
loc, op.getLhs(), b, acc, kind, rewriter, isInt, mask);
|
|
if (!mult.has_value())
|
|
return failure();
|
|
rewriter.replaceOp(rootOp, *mult);
|
|
return success();
|
|
}
|
|
|
|
Value result = rewriter.create<arith::ConstantOp>(
|
|
loc, resType, rewriter.getZeroAttr(resType));
|
|
for (int64_t d = 0, e = resType.getDimSize(0); d < e; ++d) {
|
|
Value x = rewriter.create<vector::ExtractOp>(loc, op.getLhs(), d);
|
|
Value a = rewriter.create<vector::BroadcastOp>(loc, rhsType, x);
|
|
Value r = nullptr;
|
|
if (acc)
|
|
r = rewriter.create<vector::ExtractOp>(loc, acc, d);
|
|
Value extrMask;
|
|
if (mask)
|
|
extrMask = rewriter.create<vector::ExtractOp>(loc, mask, d);
|
|
|
|
std::optional<Value> m = createContractArithOp(
|
|
loc, a, op.getRhs(), r, kind, rewriter, isInt, extrMask);
|
|
if (!m.has_value())
|
|
return failure();
|
|
result = rewriter.create<vector::InsertOp>(loc, *m, result, d);
|
|
}
|
|
|
|
rewriter.replaceOp(rootOp, result);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
/// Progressively lower a `vector.contract %a, %b, %c` with row-major matmul
|
|
/// semantics to:
|
|
/// ```
|
|
/// %mta = maybe_transpose
|
|
/// %mtb = maybe_transpose
|
|
/// %flattened_a = vector.shape_cast %mta
|
|
/// %flattened_b = vector.shape_cast %mtb
|
|
/// %flattened_d = vector.matmul %flattened_a, %flattened_b
|
|
/// %mtd = vector.shape_cast %flattened_d
|
|
/// %d = maybe_untranspose %mtd
|
|
/// %e = add %c, %d
|
|
/// ```
|
|
/// `vector.matmul` later lowers to `llvm.matrix.multiply`.
|
|
//
|
|
/// This only kicks in when VectorTransformsOptions is set to `Matmul`.
|
|
/// vector.transpose operations are inserted if the vector.contract op is not a
|
|
/// row-major matrix multiply.
|
|
LogicalResult
|
|
ContractionOpToMatmulOpLowering::matchAndRewrite(vector::ContractionOp op,
|
|
PatternRewriter &rew) const {
|
|
// TODO: Support vector.mask.
|
|
auto maskableOp = cast<MaskableOpInterface>(op.getOperation());
|
|
if (maskableOp.isMasked())
|
|
return failure();
|
|
|
|
if (vectorTransformOptions.vectorContractLowering !=
|
|
vector::VectorContractLowering::Matmul)
|
|
return failure();
|
|
if (failed(filter(op)))
|
|
return failure();
|
|
|
|
auto iteratorTypes = op.getIteratorTypes().getValue();
|
|
if (!isParallelIterator(iteratorTypes[0]) ||
|
|
!isParallelIterator(iteratorTypes[1]) ||
|
|
!isReductionIterator(iteratorTypes[2]))
|
|
return failure();
|
|
|
|
Type elementType = op.getLhsType().getElementType();
|
|
if (!elementType.isIntOrFloat())
|
|
return failure();
|
|
|
|
Type dstElementType = op.getType();
|
|
if (auto vecType = dyn_cast<VectorType>(dstElementType))
|
|
dstElementType = vecType.getElementType();
|
|
if (elementType != dstElementType)
|
|
return failure();
|
|
|
|
// Perform lhs + rhs transpositions to conform to matmul row-major semantics.
|
|
// Bail out if the contraction cannot be put in this form.
|
|
MLIRContext *ctx = op.getContext();
|
|
Location loc = op.getLoc();
|
|
AffineExpr m, n, k;
|
|
bindDims(rew.getContext(), m, n, k);
|
|
// LHS must be A(m, k) or A(k, m).
|
|
Value lhs = op.getLhs();
|
|
auto lhsMap = op.getIndexingMapsArray()[0];
|
|
if (lhsMap == AffineMap::get(3, 0, {k, m}, ctx))
|
|
lhs = rew.create<vector::TransposeOp>(loc, lhs, ArrayRef<int64_t>{1, 0});
|
|
else if (lhsMap != AffineMap::get(3, 0, {m, k}, ctx))
|
|
return failure();
|
|
|
|
// RHS must be B(k, n) or B(n, k).
|
|
Value rhs = op.getRhs();
|
|
auto rhsMap = op.getIndexingMapsArray()[1];
|
|
if (rhsMap == AffineMap::get(3, 0, {n, k}, ctx))
|
|
rhs = rew.create<vector::TransposeOp>(loc, rhs, ArrayRef<int64_t>{1, 0});
|
|
else if (rhsMap != AffineMap::get(3, 0, {k, n}, ctx))
|
|
return failure();
|
|
|
|
// At this point lhs and rhs are in row-major.
|
|
VectorType lhsType = cast<VectorType>(lhs.getType());
|
|
VectorType rhsType = cast<VectorType>(rhs.getType());
|
|
int64_t lhsRows = lhsType.getDimSize(0);
|
|
int64_t lhsColumns = lhsType.getDimSize(1);
|
|
int64_t rhsColumns = rhsType.getDimSize(1);
|
|
|
|
Type flattenedLHSType =
|
|
VectorType::get(lhsType.getNumElements(), lhsType.getElementType());
|
|
lhs = rew.create<vector::ShapeCastOp>(loc, flattenedLHSType, lhs);
|
|
|
|
Type flattenedRHSType =
|
|
VectorType::get(rhsType.getNumElements(), rhsType.getElementType());
|
|
rhs = rew.create<vector::ShapeCastOp>(loc, flattenedRHSType, rhs);
|
|
|
|
Value mul = rew.create<vector::MatmulOp>(loc, lhs, rhs, lhsRows, lhsColumns,
|
|
rhsColumns);
|
|
mul = rew.create<vector::ShapeCastOp>(
|
|
loc,
|
|
VectorType::get({lhsRows, rhsColumns},
|
|
getElementTypeOrSelf(op.getAcc().getType())),
|
|
mul);
|
|
|
|
// ACC must be C(m, n) or C(n, m).
|
|
auto accMap = op.getIndexingMapsArray()[2];
|
|
if (accMap == AffineMap::get(3, 0, {n, m}, ctx))
|
|
mul = rew.create<vector::TransposeOp>(loc, mul, ArrayRef<int64_t>{1, 0});
|
|
else if (accMap != AffineMap::get(3, 0, {m, n}, ctx))
|
|
llvm_unreachable("invalid contraction semantics");
|
|
|
|
Value res =
|
|
isa<IntegerType>(elementType)
|
|
? static_cast<Value>(rew.create<arith::AddIOp>(loc, op.getAcc(), mul))
|
|
: static_cast<Value>(
|
|
rew.create<arith::AddFOp>(loc, op.getAcc(), mul));
|
|
|
|
rew.replaceOp(op, res);
|
|
return success();
|
|
}
|
|
} // namespace
|
|
|
|
void mlir::vector::populateVectorContractLoweringPatterns(
|
|
RewritePatternSet &patterns, VectorTransformsOptions options,
|
|
PatternBenefit benefit, bool disableOuterProductLowering) {
|
|
if (!disableOuterProductLowering)
|
|
patterns.add<OuterProductOpLowering>(patterns.getContext(), benefit);
|
|
patterns.add<ContractionOpLowering, ContractionOpToMatmulOpLowering,
|
|
ContractionOpToOuterProductOpLowering>(
|
|
options, patterns.getContext(), benefit);
|
|
}
|
|
|
|
void mlir::vector::populateVectorOuterProductLoweringPatterns(
|
|
RewritePatternSet &patterns, PatternBenefit benefit) {
|
|
patterns.add<OuterProductOpLowering>(patterns.getContext(), benefit);
|
|
}
|