c56bd7ab79
This patch adds support for masked vectorisation of depthwise 1D WC
convolutions,`linalg.depthwise_conv_1d_nwc_wc`. This is implemented by
adding support for masking.
Two major assumptions are made:
* only the channel dimension can be dynamic/scalable (i.e. the
trailing dim),
* when specifying vector sizes to use in the vectoriser, only the size
corresponding to the channel dim is effectively used (other dims are
inferred from the context).
In terms of scalable vectorisation, this should be sufficient to cover
all practical cases (i.e. making arbitrary dim scalable wouldn't make
much sense). As for more generic cases with dynamic shapes (e.g. W or N
dims being dynamic), more work would be needed. In particular, one would
have to consider the filter and input/output tensors separately.
320 lines
13 KiB
C++
320 lines
13 KiB
C++
//===- VectorUtils.cpp - MLIR Utilities for VectorOps ------------------===//
|
|
//
|
|
// Part of the MLIR 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 utility methods for working with the Vector dialect.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#include "mlir/Dialect/Vector/Utils/VectorUtils.h"
|
|
|
|
#include "mlir/Dialect/Affine/Analysis/LoopAnalysis.h"
|
|
#include "mlir/Dialect/Affine/IR/AffineOps.h"
|
|
#include "mlir/Dialect/Arith/IR/Arith.h"
|
|
#include "mlir/Dialect/Func/IR/FuncOps.h"
|
|
#include "mlir/Dialect/MemRef/IR/MemRef.h"
|
|
#include "mlir/Dialect/Tensor/IR/Tensor.h"
|
|
#include "mlir/Dialect/Utils/IndexingUtils.h"
|
|
#include "mlir/Dialect/Vector/IR/VectorOps.h"
|
|
#include "mlir/IR/Builders.h"
|
|
#include "mlir/IR/IntegerSet.h"
|
|
#include "mlir/IR/Operation.h"
|
|
#include "mlir/IR/TypeUtilities.h"
|
|
#include "mlir/Support/LLVM.h"
|
|
#include "mlir/Support/MathExtras.h"
|
|
|
|
#include "llvm/ADT/DenseSet.h"
|
|
#include "llvm/ADT/SetVector.h"
|
|
|
|
using namespace mlir;
|
|
|
|
/// Helper function that creates a memref::DimOp or tensor::DimOp depending on
|
|
/// the type of `source`.
|
|
Value mlir::vector::createOrFoldDimOp(OpBuilder &b, Location loc, Value source,
|
|
int64_t dim) {
|
|
if (isa<UnrankedMemRefType, MemRefType>(source.getType()))
|
|
return b.createOrFold<memref::DimOp>(loc, source, dim);
|
|
if (isa<UnrankedTensorType, RankedTensorType>(source.getType()))
|
|
return b.createOrFold<tensor::DimOp>(loc, source, dim);
|
|
llvm_unreachable("Expected MemRefType or TensorType");
|
|
}
|
|
|
|
/// Given the n-D transpose pattern 'transp', return true if 'dim0' and 'dim1'
|
|
/// should be transposed with each other within the context of their 2D
|
|
/// transposition slice.
|
|
///
|
|
/// Example 1: dim0 = 0, dim1 = 2, transp = [2, 1, 0]
|
|
/// Return true: dim0 and dim1 are transposed within the context of their 2D
|
|
/// transposition slice ([1, 0]).
|
|
///
|
|
/// Example 2: dim0 = 0, dim1 = 1, transp = [2, 1, 0]
|
|
/// Return true: dim0 and dim1 are transposed within the context of their 2D
|
|
/// transposition slice ([1, 0]). Paradoxically, note how dim1 (1) is *not*
|
|
/// transposed within the full context of the transposition.
|
|
///
|
|
/// Example 3: dim0 = 0, dim1 = 1, transp = [2, 0, 1]
|
|
/// Return false: dim0 and dim1 are *not* transposed within the context of
|
|
/// their 2D transposition slice ([0, 1]). Paradoxically, note how dim0 (0)
|
|
/// and dim1 (1) are transposed within the full context of the of the
|
|
/// transposition.
|
|
static bool areDimsTransposedIn2DSlice(int64_t dim0, int64_t dim1,
|
|
ArrayRef<int64_t> transp) {
|
|
// Perform a linear scan along the dimensions of the transposed pattern. If
|
|
// dim0 is found first, dim0 and dim1 are not transposed within the context of
|
|
// their 2D slice. Otherwise, 'dim1' is found first and they are transposed.
|
|
for (int64_t permDim : transp) {
|
|
if (permDim == dim0)
|
|
return false;
|
|
if (permDim == dim1)
|
|
return true;
|
|
}
|
|
|
|
llvm_unreachable("Ill-formed transpose pattern");
|
|
}
|
|
|
|
FailureOr<std::pair<int, int>>
|
|
mlir::vector::isTranspose2DSlice(vector::TransposeOp op) {
|
|
VectorType srcType = op.getSourceVectorType();
|
|
SmallVector<int64_t> srcGtOneDims;
|
|
for (auto [index, size] : llvm::enumerate(srcType.getShape()))
|
|
if (size > 1)
|
|
srcGtOneDims.push_back(index);
|
|
|
|
if (srcGtOneDims.size() != 2)
|
|
return failure();
|
|
|
|
// Check whether the two source vector dimensions that are greater than one
|
|
// must be transposed with each other so that we can apply one of the 2-D
|
|
// transpose pattens. Otherwise, these patterns are not applicable.
|
|
if (!areDimsTransposedIn2DSlice(srcGtOneDims[0], srcGtOneDims[1],
|
|
op.getPermutation()))
|
|
return failure();
|
|
|
|
return std::pair<int, int>(srcGtOneDims[0], srcGtOneDims[1]);
|
|
}
|
|
|
|
/// Constructs a permutation map from memref indices to vector dimension.
|
|
///
|
|
/// The implementation uses the knowledge of the mapping of enclosing loop to
|
|
/// vector dimension. `enclosingLoopToVectorDim` carries this information as a
|
|
/// map with:
|
|
/// - keys representing "vectorized enclosing loops";
|
|
/// - values representing the corresponding vector dimension.
|
|
/// The algorithm traverses "vectorized enclosing loops" and extracts the
|
|
/// at-most-one MemRef index that is invariant along said loop. This index is
|
|
/// guaranteed to be at most one by construction: otherwise the MemRef is not
|
|
/// vectorizable.
|
|
/// If this invariant index is found, it is added to the permutation_map at the
|
|
/// proper vector dimension.
|
|
/// If no index is found to be invariant, 0 is added to the permutation_map and
|
|
/// corresponds to a vector broadcast along that dimension.
|
|
///
|
|
/// Returns an empty AffineMap if `enclosingLoopToVectorDim` is empty,
|
|
/// signalling that no permutation map can be constructed given
|
|
/// `enclosingLoopToVectorDim`.
|
|
///
|
|
/// Examples can be found in the documentation of `makePermutationMap`, in the
|
|
/// header file.
|
|
static AffineMap makePermutationMap(
|
|
ArrayRef<Value> indices,
|
|
const DenseMap<Operation *, unsigned> &enclosingLoopToVectorDim) {
|
|
if (enclosingLoopToVectorDim.empty())
|
|
return AffineMap();
|
|
MLIRContext *context =
|
|
enclosingLoopToVectorDim.begin()->getFirst()->getContext();
|
|
SmallVector<AffineExpr> perm(enclosingLoopToVectorDim.size(),
|
|
getAffineConstantExpr(0, context));
|
|
|
|
for (auto kvp : enclosingLoopToVectorDim) {
|
|
assert(kvp.second < perm.size());
|
|
auto invariants = affine::getInvariantAccesses(
|
|
cast<affine::AffineForOp>(kvp.first).getInductionVar(), indices);
|
|
unsigned numIndices = indices.size();
|
|
unsigned countInvariantIndices = 0;
|
|
for (unsigned dim = 0; dim < numIndices; ++dim) {
|
|
if (!invariants.count(indices[dim])) {
|
|
assert(perm[kvp.second] == getAffineConstantExpr(0, context) &&
|
|
"permutationMap already has an entry along dim");
|
|
perm[kvp.second] = getAffineDimExpr(dim, context);
|
|
} else {
|
|
++countInvariantIndices;
|
|
}
|
|
}
|
|
assert((countInvariantIndices == numIndices ||
|
|
countInvariantIndices == numIndices - 1) &&
|
|
"Vectorization prerequisite violated: at most 1 index may be "
|
|
"invariant wrt a vectorized loop");
|
|
(void)countInvariantIndices;
|
|
}
|
|
return AffineMap::get(indices.size(), 0, perm, context);
|
|
}
|
|
|
|
/// Implementation detail that walks up the parents and records the ones with
|
|
/// the specified type.
|
|
/// TODO: could also be implemented as a collect parents followed by a
|
|
/// filter and made available outside this file.
|
|
template <typename T>
|
|
static SetVector<Operation *> getParentsOfType(Block *block) {
|
|
SetVector<Operation *> res;
|
|
auto *current = block->getParentOp();
|
|
while (current) {
|
|
if ([[maybe_unused]] auto typedParent = dyn_cast<T>(current)) {
|
|
assert(res.count(current) == 0 && "Already inserted");
|
|
res.insert(current);
|
|
}
|
|
current = current->getParentOp();
|
|
}
|
|
return res;
|
|
}
|
|
|
|
/// Returns the enclosing AffineForOp, from closest to farthest.
|
|
static SetVector<Operation *> getEnclosingforOps(Block *block) {
|
|
return getParentsOfType<affine::AffineForOp>(block);
|
|
}
|
|
|
|
AffineMap mlir::makePermutationMap(
|
|
Block *insertPoint, ArrayRef<Value> indices,
|
|
const DenseMap<Operation *, unsigned> &loopToVectorDim) {
|
|
DenseMap<Operation *, unsigned> enclosingLoopToVectorDim;
|
|
auto enclosingLoops = getEnclosingforOps(insertPoint);
|
|
for (auto *forInst : enclosingLoops) {
|
|
auto it = loopToVectorDim.find(forInst);
|
|
if (it != loopToVectorDim.end()) {
|
|
enclosingLoopToVectorDim.insert(*it);
|
|
}
|
|
}
|
|
return ::makePermutationMap(indices, enclosingLoopToVectorDim);
|
|
}
|
|
|
|
AffineMap mlir::makePermutationMap(
|
|
Operation *op, ArrayRef<Value> indices,
|
|
const DenseMap<Operation *, unsigned> &loopToVectorDim) {
|
|
return makePermutationMap(op->getBlock(), indices, loopToVectorDim);
|
|
}
|
|
|
|
bool matcher::operatesOnSuperVectorsOf(Operation &op,
|
|
VectorType subVectorType) {
|
|
// First, extract the vector type and distinguish between:
|
|
// a. ops that *must* lower a super-vector (i.e. vector.transfer_read,
|
|
// vector.transfer_write); and
|
|
// b. ops that *may* lower a super-vector (all other ops).
|
|
// The ops that *may* lower a super-vector only do so if the super-vector to
|
|
// sub-vector ratio exists. The ops that *must* lower a super-vector are
|
|
// explicitly checked for this property.
|
|
/// TODO: there should be a single function for all ops to do this so we
|
|
/// do not have to special case. Maybe a trait, or just a method, unclear atm.
|
|
bool mustDivide = false;
|
|
(void)mustDivide;
|
|
VectorType superVectorType;
|
|
if (auto transfer = dyn_cast<VectorTransferOpInterface>(op)) {
|
|
superVectorType = transfer.getVectorType();
|
|
mustDivide = true;
|
|
} else if (op.getNumResults() == 0) {
|
|
if (!isa<func::ReturnOp>(op)) {
|
|
op.emitError("NYI: assuming only return operations can have 0 "
|
|
" results at this point");
|
|
}
|
|
return false;
|
|
} else if (op.getNumResults() == 1) {
|
|
if (auto v = dyn_cast<VectorType>(op.getResult(0).getType())) {
|
|
superVectorType = v;
|
|
} else {
|
|
// Not a vector type.
|
|
return false;
|
|
}
|
|
} else {
|
|
// Not a vector.transfer and has more than 1 result, fail hard for now to
|
|
// wake us up when something changes.
|
|
op.emitError("NYI: operation has more than 1 result");
|
|
return false;
|
|
}
|
|
|
|
// Get the ratio.
|
|
auto ratio =
|
|
computeShapeRatio(superVectorType.getShape(), subVectorType.getShape());
|
|
|
|
// Sanity check.
|
|
assert((ratio || !mustDivide) &&
|
|
"vector.transfer operation in which super-vector size is not an"
|
|
" integer multiple of sub-vector size");
|
|
|
|
// This catches cases that are not strictly necessary to have multiplicity but
|
|
// still aren't divisible by the sub-vector shape.
|
|
// This could be useful information if we wanted to reshape at the level of
|
|
// the vector type (but we would have to look at the compute and distinguish
|
|
// between parallel, reduction and possibly other cases.
|
|
return ratio.has_value();
|
|
}
|
|
|
|
bool vector::isContiguousSlice(MemRefType memrefType, VectorType vectorType) {
|
|
if (vectorType.isScalable())
|
|
return false;
|
|
|
|
ArrayRef<int64_t> vectorShape = vectorType.getShape();
|
|
auto vecRank = vectorType.getRank();
|
|
|
|
if (!trailingNDimsContiguous(memrefType, vecRank))
|
|
return false;
|
|
|
|
// Extract the trailing dims and strides of the input memref
|
|
auto memrefShape = memrefType.getShape().take_back(vecRank);
|
|
|
|
// Compare the dims of `vectorType` against `memrefType` (in reverse).
|
|
// In the most basic case, all dims will match.
|
|
auto firstNonMatchingDim =
|
|
std::mismatch(vectorShape.rbegin(), vectorShape.rend(),
|
|
memrefShape.rbegin(), memrefShape.rend());
|
|
if (firstNonMatchingDim.first == vectorShape.rend())
|
|
return true;
|
|
|
|
// One non-matching dim is still fine, however the remaining leading dims of
|
|
// `vectorType` need to be 1.
|
|
SmallVector<int64_t> leadingDims(++firstNonMatchingDim.first,
|
|
vectorShape.rend());
|
|
|
|
return llvm::all_of(leadingDims, [](auto x) { return x == 1; });
|
|
}
|
|
|
|
std::optional<StaticTileOffsetRange>
|
|
vector::createUnrollIterator(VectorType vType, int64_t targetRank) {
|
|
if (vType.getRank() <= targetRank)
|
|
return {};
|
|
// Attempt to unroll until targetRank or the first scalable dimension (which
|
|
// cannot be unrolled).
|
|
auto shapeToUnroll = vType.getShape().drop_back(targetRank);
|
|
auto scalableDimsToUnroll = vType.getScalableDims().drop_back(targetRank);
|
|
auto it =
|
|
std::find(scalableDimsToUnroll.begin(), scalableDimsToUnroll.end(), true);
|
|
auto firstScalableDim = it - scalableDimsToUnroll.begin();
|
|
if (firstScalableDim == 0)
|
|
return {};
|
|
// All scalable dimensions should be removed now.
|
|
scalableDimsToUnroll = scalableDimsToUnroll.slice(0, firstScalableDim);
|
|
assert(!llvm::is_contained(scalableDimsToUnroll, true) &&
|
|
"unexpected leading scalable dimension");
|
|
// Create an unroll iterator for leading dimensions.
|
|
shapeToUnroll = shapeToUnroll.slice(0, firstScalableDim);
|
|
return StaticTileOffsetRange(shapeToUnroll, /*unrollStep=*/1);
|
|
}
|
|
|
|
SmallVector<OpFoldResult> vector::getMixedSizesXfer(bool hasTensorSemantics,
|
|
Operation *xfer,
|
|
RewriterBase &rewriter) {
|
|
auto loc = xfer->getLoc();
|
|
|
|
Value base = TypeSwitch<Operation *, Value>(xfer)
|
|
.Case<vector::TransferReadOp>(
|
|
[&](auto readOp) { return readOp.getSource(); })
|
|
.Case<vector::TransferWriteOp>(
|
|
[&](auto writeOp) { return writeOp.getOperand(1); });
|
|
|
|
SmallVector<OpFoldResult> mixedSourceDims =
|
|
hasTensorSemantics ? tensor::getMixedSizes(rewriter, loc, base)
|
|
: memref::getMixedSizes(rewriter, loc, base);
|
|
return mixedSourceDims;
|
|
}
|