c45cc3e420
[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
1752 lines
68 KiB
C++
1752 lines
68 KiB
C++
//===- VectorToSCF.cpp - Convert vector to SCF dialect ----------*- C++ -*-===//
|
|
//
|
|
// 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 lowering of vector transfer operations to SCF.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#include <numeric>
|
|
#include <optional>
|
|
#include <type_traits>
|
|
|
|
#include "mlir/Conversion/VectorToSCF/VectorToSCF.h"
|
|
|
|
#include "mlir/Dialect/Affine/IR/AffineOps.h"
|
|
#include "mlir/Dialect/Arith/IR/Arith.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/Vector/IR/VectorOps.h"
|
|
#include "mlir/Dialect/Vector/Transforms/LoweringPatterns.h"
|
|
#include "mlir/Dialect/Vector/Transforms/VectorTransforms.h"
|
|
#include "mlir/Dialect/Vector/Utils/VectorUtils.h"
|
|
#include "mlir/IR/Builders.h"
|
|
#include "mlir/IR/ImplicitLocOpBuilder.h"
|
|
#include "mlir/Pass/Pass.h"
|
|
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
|
|
#include "mlir/Transforms/Passes.h"
|
|
|
|
namespace mlir {
|
|
#define GEN_PASS_DEF_CONVERTVECTORTOSCF
|
|
#include "mlir/Conversion/Passes.h.inc"
|
|
} // namespace mlir
|
|
|
|
using namespace mlir;
|
|
using vector::TransferReadOp;
|
|
using vector::TransferWriteOp;
|
|
|
|
namespace {
|
|
|
|
/// Attribute name used for labeling transfer ops during progressive lowering.
|
|
static const char kPassLabel[] = "__vector_to_scf_lowering__";
|
|
|
|
/// Return true if this transfer op operates on a source tensor.
|
|
static bool isTensorOp(VectorTransferOpInterface xferOp) {
|
|
if (isa<RankedTensorType>(xferOp.getShapedType())) {
|
|
if (isa<vector::TransferWriteOp>(xferOp)) {
|
|
// TransferWriteOps on tensors have a result.
|
|
assert(xferOp->getNumResults() > 0);
|
|
}
|
|
return true;
|
|
}
|
|
return false;
|
|
}
|
|
|
|
/// Patterns that inherit from this struct have access to
|
|
/// VectorTransferToSCFOptions.
|
|
template <typename OpTy>
|
|
struct VectorToSCFPattern : public OpRewritePattern<OpTy> {
|
|
explicit VectorToSCFPattern(MLIRContext *context,
|
|
VectorTransferToSCFOptions opt)
|
|
: OpRewritePattern<OpTy>(context), options(opt) {}
|
|
|
|
LogicalResult checkLowerTensors(VectorTransferOpInterface xferOp,
|
|
PatternRewriter &rewriter) const {
|
|
if (isTensorOp(xferOp) && !options.lowerTensors) {
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "lowering tensor transfers is disabled");
|
|
}
|
|
return success();
|
|
}
|
|
|
|
VectorTransferToSCFOptions options;
|
|
};
|
|
|
|
/// Given a vector transfer op, calculate which dimension of the `source`
|
|
/// memref should be unpacked in the next application of TransferOpConversion.
|
|
/// A return value of std::nullopt indicates a broadcast.
|
|
template <typename OpTy>
|
|
static std::optional<int64_t> unpackedDim(OpTy xferOp) {
|
|
// TODO: support 0-d corner case.
|
|
assert(xferOp.getTransferRank() > 0 && "unexpected 0-d transfer");
|
|
auto map = xferOp.getPermutationMap();
|
|
if (auto expr = dyn_cast<AffineDimExpr>(map.getResult(0))) {
|
|
return expr.getPosition();
|
|
}
|
|
assert(xferOp.isBroadcastDim(0) &&
|
|
"Expected AffineDimExpr or AffineConstantExpr");
|
|
return std::nullopt;
|
|
}
|
|
|
|
/// Compute the permutation map for the new (N-1)-D vector transfer op. This
|
|
/// map is identical to the current permutation map, but the first result is
|
|
/// omitted.
|
|
template <typename OpTy>
|
|
static AffineMap unpackedPermutationMap(OpBuilder &b, OpTy xferOp) {
|
|
// TODO: support 0-d corner case.
|
|
assert(xferOp.getTransferRank() > 0 && "unexpected 0-d transfer");
|
|
auto map = xferOp.getPermutationMap();
|
|
return AffineMap::get(map.getNumDims(), 0, map.getResults().drop_front(),
|
|
b.getContext());
|
|
}
|
|
|
|
/// Calculate the indices for the new vector transfer op.
|
|
///
|
|
/// E.g.: transfer_read %A[%a, %b, %c, %d] ... : vector<5x4x3xf32> ...
|
|
/// --> transfer_read %A[%a, %b + iv, %c, %d] ... vector<4x3f32>
|
|
/// ^^^^^^
|
|
/// `iv` is the iteration variable of the (new) surrounding loop.
|
|
template <typename OpTy>
|
|
static void getXferIndices(OpBuilder &b, OpTy xferOp, Value iv,
|
|
SmallVector<Value, 8> &indices) {
|
|
typename OpTy::Adaptor adaptor(xferOp);
|
|
// Corresponding memref dim of the vector dim that is unpacked.
|
|
auto dim = unpackedDim(xferOp);
|
|
auto prevIndices = adaptor.getIndices();
|
|
indices.append(prevIndices.begin(), prevIndices.end());
|
|
|
|
Location loc = xferOp.getLoc();
|
|
bool isBroadcast = !dim.has_value();
|
|
if (!isBroadcast) {
|
|
AffineExpr d0, d1;
|
|
bindDims(xferOp.getContext(), d0, d1);
|
|
Value offset = adaptor.getIndices()[*dim];
|
|
indices[*dim] =
|
|
affine::makeComposedAffineApply(b, loc, d0 + d1, {offset, iv});
|
|
}
|
|
}
|
|
|
|
static void maybeYieldValue(OpBuilder &b, Location loc, bool hasRetVal,
|
|
Value value) {
|
|
if (hasRetVal) {
|
|
assert(value && "Expected non-empty value");
|
|
b.create<scf::YieldOp>(loc, value);
|
|
} else {
|
|
b.create<scf::YieldOp>(loc);
|
|
}
|
|
}
|
|
|
|
/// Generates a boolean Value that is true if the iv-th bit in xferOp's mask
|
|
/// is set to true. No such check is generated under following circumstances:
|
|
/// * xferOp does not have a mask.
|
|
/// * xferOp's mask is not 1D. (In case of (N>1)-D, a subvector of the mask is
|
|
/// computed and attached to the new transfer op in the pattern.)
|
|
/// * The to-be-unpacked dim of xferOp is a broadcast.
|
|
template <typename OpTy>
|
|
static Value generateMaskCheck(OpBuilder &b, OpTy xferOp, Value iv) {
|
|
if (!xferOp.getMask())
|
|
return Value();
|
|
if (xferOp.getMaskType().getRank() != 1)
|
|
return Value();
|
|
if (xferOp.isBroadcastDim(0))
|
|
return Value();
|
|
|
|
Location loc = xferOp.getLoc();
|
|
return b.create<vector::ExtractElementOp>(loc, xferOp.getMask(), iv);
|
|
}
|
|
|
|
/// Helper function TransferOpConversion and TransferOp1dConversion.
|
|
/// Generate an in-bounds check if the transfer op may go out-of-bounds on the
|
|
/// specified dimension `dim` with the loop iteration variable `iv`.
|
|
/// E.g., when unpacking dimension 0 from:
|
|
/// ```
|
|
/// %vec = vector.transfer_read %A[%a, %b] %cst
|
|
/// : vector<5x4xf32>, memref<?x?xf32>
|
|
/// ```
|
|
/// An if check similar to this will be generated inside the loop:
|
|
/// ```
|
|
/// %d = memref.dim %A, %c0 : memref<?x?xf32>
|
|
/// if (%a + iv < %d) {
|
|
/// (in-bounds case)
|
|
/// } else {
|
|
/// (out-of-bounds case)
|
|
/// }
|
|
/// ```
|
|
///
|
|
/// If the transfer is 1D and has a mask, this function generates a more complex
|
|
/// check also accounts for potentially masked out elements.
|
|
///
|
|
/// This function variant returns the value returned by `inBoundsCase` or
|
|
/// `outOfBoundsCase`. The MLIR type of the return value must be specified in
|
|
/// `resultTypes`.
|
|
template <typename OpTy>
|
|
static Value generateInBoundsCheck(
|
|
OpBuilder &b, OpTy xferOp, Value iv, std::optional<int64_t> dim,
|
|
TypeRange resultTypes,
|
|
function_ref<Value(OpBuilder &, Location)> inBoundsCase,
|
|
function_ref<Value(OpBuilder &, Location)> outOfBoundsCase = nullptr) {
|
|
bool hasRetVal = !resultTypes.empty();
|
|
Value cond; // Condition to be built...
|
|
|
|
// Condition check 1: Access in-bounds?
|
|
bool isBroadcast = !dim; // No in-bounds check for broadcasts.
|
|
Location loc = xferOp.getLoc();
|
|
ImplicitLocOpBuilder lb(xferOp.getLoc(), b);
|
|
if (!xferOp.isDimInBounds(0) && !isBroadcast) {
|
|
Value memrefDim = vector::createOrFoldDimOp(b, loc, xferOp.getBase(), *dim);
|
|
AffineExpr d0, d1;
|
|
bindDims(xferOp.getContext(), d0, d1);
|
|
Value base = xferOp.getIndices()[*dim];
|
|
Value memrefIdx =
|
|
affine::makeComposedAffineApply(b, loc, d0 + d1, {base, iv});
|
|
cond = lb.create<arith::CmpIOp>(arith::CmpIPredicate::sgt, memrefDim,
|
|
memrefIdx);
|
|
}
|
|
|
|
// Condition check 2: Masked in?
|
|
if (auto maskCond = generateMaskCheck(b, xferOp, iv)) {
|
|
if (cond)
|
|
cond = lb.create<arith::AndIOp>(cond, maskCond);
|
|
else
|
|
cond = maskCond;
|
|
}
|
|
|
|
// If the condition is non-empty, generate an SCF::IfOp.
|
|
if (cond) {
|
|
auto check = lb.create<scf::IfOp>(
|
|
cond,
|
|
/*thenBuilder=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
maybeYieldValue(b, loc, hasRetVal, inBoundsCase(b, loc));
|
|
},
|
|
/*elseBuilder=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
if (outOfBoundsCase) {
|
|
maybeYieldValue(b, loc, hasRetVal, outOfBoundsCase(b, loc));
|
|
} else {
|
|
b.create<scf::YieldOp>(loc);
|
|
}
|
|
});
|
|
|
|
return hasRetVal ? check.getResult(0) : Value();
|
|
}
|
|
|
|
// Condition is empty, no need for an SCF::IfOp.
|
|
return inBoundsCase(b, loc);
|
|
}
|
|
|
|
/// In this function variant, `inBoundsCase` and `outOfBoundsCase` do not have
|
|
/// a return value. Consequently, this function does not have a return value.
|
|
template <typename OpTy>
|
|
static void generateInBoundsCheck(
|
|
OpBuilder &b, OpTy xferOp, Value iv, std::optional<int64_t> dim,
|
|
function_ref<void(OpBuilder &, Location)> inBoundsCase,
|
|
function_ref<void(OpBuilder &, Location)> outOfBoundsCase = nullptr) {
|
|
generateInBoundsCheck(
|
|
b, xferOp, iv, dim, /*resultTypes=*/TypeRange(),
|
|
/*inBoundsCase=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
inBoundsCase(b, loc);
|
|
return Value();
|
|
},
|
|
/*outOfBoundsCase=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
if (outOfBoundsCase)
|
|
outOfBoundsCase(b, loc);
|
|
return Value();
|
|
});
|
|
}
|
|
|
|
/// Given an ArrayAttr, return a copy where the first element is dropped.
|
|
static ArrayAttr dropFirstElem(OpBuilder &b, ArrayAttr attr) {
|
|
if (!attr)
|
|
return attr;
|
|
return ArrayAttr::get(b.getContext(), attr.getValue().drop_front());
|
|
}
|
|
|
|
/// Add the pass label to a vector transfer op if its rank is not the target
|
|
/// rank.
|
|
template <typename OpTy>
|
|
static void maybeApplyPassLabel(OpBuilder &b, OpTy newXferOp,
|
|
unsigned targetRank) {
|
|
if (newXferOp.getVectorType().getRank() > targetRank)
|
|
newXferOp->setAttr(kPassLabel, b.getUnitAttr());
|
|
}
|
|
|
|
namespace lowering_n_d {
|
|
|
|
/// Helper data structure for data and mask buffers.
|
|
struct BufferAllocs {
|
|
Value dataBuffer;
|
|
Value maskBuffer;
|
|
};
|
|
|
|
// TODO: Parallelism and threadlocal considerations with a ParallelScope trait.
|
|
static Operation *getAutomaticAllocationScope(Operation *op) {
|
|
Operation *scope =
|
|
op->getParentWithTrait<OpTrait::AutomaticAllocationScope>();
|
|
assert(scope && "Expected op to be inside automatic allocation scope");
|
|
return scope;
|
|
}
|
|
|
|
/// Allocate temporary buffers for data (vector) and mask (if present).
|
|
template <typename OpTy>
|
|
static BufferAllocs allocBuffers(OpBuilder &b, OpTy xferOp) {
|
|
Location loc = xferOp.getLoc();
|
|
OpBuilder::InsertionGuard guard(b);
|
|
Operation *scope = getAutomaticAllocationScope(xferOp);
|
|
assert(scope->getNumRegions() == 1 &&
|
|
"AutomaticAllocationScope with >1 regions");
|
|
b.setInsertionPointToStart(&scope->getRegion(0).front());
|
|
|
|
BufferAllocs result;
|
|
auto bufferType = MemRefType::get({}, xferOp.getVectorType());
|
|
result.dataBuffer = b.create<memref::AllocaOp>(loc, bufferType);
|
|
|
|
if (xferOp.getMask()) {
|
|
auto maskType = MemRefType::get({}, xferOp.getMask().getType());
|
|
auto maskBuffer = b.create<memref::AllocaOp>(loc, maskType);
|
|
b.setInsertionPoint(xferOp);
|
|
b.create<memref::StoreOp>(loc, xferOp.getMask(), maskBuffer);
|
|
result.maskBuffer = b.create<memref::LoadOp>(loc, maskBuffer, ValueRange());
|
|
}
|
|
|
|
return result;
|
|
}
|
|
|
|
/// Given a MemRefType with VectorType element type, unpack one dimension from
|
|
/// the VectorType into the MemRefType.
|
|
///
|
|
/// E.g.: memref<9xvector<5x6xf32>> --> memref<9x5xvector<6xf32>>
|
|
static FailureOr<MemRefType> unpackOneDim(MemRefType type) {
|
|
auto vectorType = dyn_cast<VectorType>(type.getElementType());
|
|
// Vectors with leading scalable dims are not supported.
|
|
// It may be possible to support these in future by using dynamic memref dims.
|
|
if (vectorType.getScalableDims().front())
|
|
return failure();
|
|
auto memrefShape = type.getShape();
|
|
SmallVector<int64_t, 8> newMemrefShape;
|
|
newMemrefShape.append(memrefShape.begin(), memrefShape.end());
|
|
newMemrefShape.push_back(vectorType.getDimSize(0));
|
|
return MemRefType::get(newMemrefShape,
|
|
VectorType::Builder(vectorType).dropDim(0));
|
|
}
|
|
|
|
/// Given a transfer op, find the memref from which the mask is loaded. This
|
|
/// is similar to Strategy<TransferWriteOp>::getBuffer.
|
|
template <typename OpTy>
|
|
static Value getMaskBuffer(OpTy xferOp) {
|
|
assert(xferOp.getMask() && "Expected that transfer op has mask");
|
|
auto loadOp = xferOp.getMask().template getDefiningOp<memref::LoadOp>();
|
|
assert(loadOp && "Expected transfer op mask produced by LoadOp");
|
|
return loadOp.getMemRef();
|
|
}
|
|
|
|
/// Codegen strategy, depending on the operation.
|
|
template <typename OpTy>
|
|
struct Strategy;
|
|
|
|
/// Code strategy for vector TransferReadOp.
|
|
template <>
|
|
struct Strategy<TransferReadOp> {
|
|
/// Find the StoreOp that is used for writing the current TransferReadOp's
|
|
/// result to the temporary buffer allocation.
|
|
static memref::StoreOp getStoreOp(TransferReadOp xferOp) {
|
|
assert(xferOp->hasOneUse() && "Expected exactly one use of TransferReadOp");
|
|
auto storeOp = dyn_cast<memref::StoreOp>((*xferOp->use_begin()).getOwner());
|
|
assert(storeOp && "Expected TransferReadOp result used by StoreOp");
|
|
return storeOp;
|
|
}
|
|
|
|
/// Find the temporary buffer allocation. All labeled TransferReadOps are
|
|
/// used like this, where %buf is either the buffer allocation or a type cast
|
|
/// of the buffer allocation:
|
|
/// ```
|
|
/// %vec = vector.transfer_read ... { __vector_to_scf_lowering__ } ...
|
|
/// memref.store %vec, %buf[...] ...
|
|
/// ```
|
|
static Value getBuffer(TransferReadOp xferOp) {
|
|
return getStoreOp(xferOp).getMemRef();
|
|
}
|
|
|
|
/// Retrieve the indices of the current StoreOp that stores into the buffer.
|
|
static void getBufferIndices(TransferReadOp xferOp,
|
|
SmallVector<Value, 8> &indices) {
|
|
auto storeOp = getStoreOp(xferOp);
|
|
auto prevIndices = memref::StoreOpAdaptor(storeOp).getIndices();
|
|
indices.append(prevIndices.begin(), prevIndices.end());
|
|
}
|
|
|
|
/// Rewrite the TransferReadOp, assuming that there are no out-of-bounds
|
|
/// accesses on the to-be-unpacked dimension.
|
|
///
|
|
/// 1. Generate a new (N-1)-d TransferReadOp using the loop iteration
|
|
/// variable `iv`.
|
|
/// 2. Store the result into the (already `vector.type_cast`ed) buffer.
|
|
///
|
|
/// E.g.:
|
|
/// ```
|
|
/// %vec = vector.transfer_read %A[%a+%i, %b, %c], %cst
|
|
/// : memref<?x?x?xf32>, vector<4x3xf32>
|
|
/// memref.store %vec, %buf[%i] : memref<5xvector<4x3xf32>>
|
|
/// ```
|
|
/// Is rewritten to:
|
|
/// ```
|
|
/// %casted = vector.type_cast %buf
|
|
/// : memref<5xvector<4x3xf32>> to memref<5x4xvector<3xf32>>
|
|
/// for %j = 0 to 4 {
|
|
/// %vec = vector.transfer_read %A[%a+%i, %b+%j, %c], %cst
|
|
/// : memref<?x?x?xf32>, vector<3xf32>
|
|
/// memref.store %vec, %casted[%i, %j] : memref<5x4xvector<3xf32>>
|
|
/// }
|
|
/// ```
|
|
///
|
|
/// Note: The loop and type cast are generated in TransferOpConversion.
|
|
/// The original TransferReadOp and store op are deleted in `cleanup`.
|
|
/// Note: The `mask` operand is set in TransferOpConversion.
|
|
static TransferReadOp rewriteOp(OpBuilder &b,
|
|
VectorTransferToSCFOptions options,
|
|
TransferReadOp xferOp, Value buffer, Value iv,
|
|
ValueRange /*loopState*/) {
|
|
SmallVector<Value, 8> storeIndices;
|
|
getBufferIndices(xferOp, storeIndices);
|
|
storeIndices.push_back(iv);
|
|
|
|
SmallVector<Value, 8> xferIndices;
|
|
getXferIndices(b, xferOp, iv, xferIndices);
|
|
|
|
Location loc = xferOp.getLoc();
|
|
auto bufferType = dyn_cast<ShapedType>(buffer.getType());
|
|
auto vecType = dyn_cast<VectorType>(bufferType.getElementType());
|
|
auto inBoundsAttr = dropFirstElem(b, xferOp.getInBoundsAttr());
|
|
auto newXferOp = b.create<vector::TransferReadOp>(
|
|
loc, vecType, xferOp.getBase(), xferIndices,
|
|
AffineMapAttr::get(unpackedPermutationMap(b, xferOp)),
|
|
xferOp.getPadding(), Value(), inBoundsAttr);
|
|
|
|
maybeApplyPassLabel(b, newXferOp, options.targetRank);
|
|
|
|
b.create<memref::StoreOp>(loc, newXferOp.getVector(), buffer, storeIndices);
|
|
return newXferOp;
|
|
}
|
|
|
|
/// Handle out-of-bounds accesses on the to-be-unpacked dimension: Write
|
|
/// padding value to the temporary buffer.
|
|
static Value handleOutOfBoundsDim(OpBuilder &b, TransferReadOp xferOp,
|
|
Value buffer, Value iv,
|
|
ValueRange /*loopState*/) {
|
|
SmallVector<Value, 8> storeIndices;
|
|
getBufferIndices(xferOp, storeIndices);
|
|
storeIndices.push_back(iv);
|
|
|
|
Location loc = xferOp.getLoc();
|
|
auto bufferType = dyn_cast<ShapedType>(buffer.getType());
|
|
auto vecType = dyn_cast<VectorType>(bufferType.getElementType());
|
|
auto vec = b.create<vector::SplatOp>(loc, vecType, xferOp.getPadding());
|
|
b.create<memref::StoreOp>(loc, vec, buffer, storeIndices);
|
|
|
|
return Value();
|
|
}
|
|
|
|
/// Cleanup after rewriting the op.
|
|
static void cleanup(PatternRewriter &rewriter, TransferReadOp xferOp,
|
|
scf::ForOp /*forOp*/) {
|
|
rewriter.eraseOp(getStoreOp(xferOp));
|
|
rewriter.eraseOp(xferOp);
|
|
}
|
|
|
|
/// Return the initial loop state for the generated scf.for loop.
|
|
static Value initialLoopState(TransferReadOp xferOp) { return Value(); }
|
|
};
|
|
|
|
/// Codegen strategy for vector TransferWriteOp.
|
|
template <>
|
|
struct Strategy<TransferWriteOp> {
|
|
/// Find the temporary buffer allocation. All labeled TransferWriteOps are
|
|
/// used like this, where %buf is either the buffer allocation or a type cast
|
|
/// of the buffer allocation:
|
|
/// ```
|
|
/// %vec = memref.load %buf[...] ...
|
|
/// vector.transfer_write %vec ... { __vector_to_scf_lowering__ } ...
|
|
/// ```
|
|
static Value getBuffer(TransferWriteOp xferOp) {
|
|
auto loadOp = xferOp.getVector().getDefiningOp<memref::LoadOp>();
|
|
assert(loadOp && "Expected transfer op vector produced by LoadOp");
|
|
return loadOp.getMemRef();
|
|
}
|
|
|
|
/// Retrieve the indices of the current LoadOp that loads from the buffer.
|
|
static void getBufferIndices(TransferWriteOp xferOp,
|
|
SmallVector<Value, 8> &indices) {
|
|
auto loadOp = xferOp.getVector().getDefiningOp<memref::LoadOp>();
|
|
auto prevIndices = memref::LoadOpAdaptor(loadOp).getIndices();
|
|
indices.append(prevIndices.begin(), prevIndices.end());
|
|
}
|
|
|
|
/// Rewrite the TransferWriteOp, assuming that there are no out-of-bounds
|
|
/// accesses on the to-be-unpacked dimension.
|
|
///
|
|
/// 1. Load an (N-1)-d vector from the (already `vector.type_cast`ed) buffer,
|
|
/// using the loop iteration variable `iv`.
|
|
/// 2. Generate a new (N-1)-d TransferWriteOp, writing the loaded vector back
|
|
/// to memory.
|
|
///
|
|
/// Note: For more details, see comments on Strategy<TransferReadOp>.
|
|
static TransferWriteOp rewriteOp(OpBuilder &b,
|
|
VectorTransferToSCFOptions options,
|
|
TransferWriteOp xferOp, Value buffer,
|
|
Value iv, ValueRange loopState) {
|
|
SmallVector<Value, 8> loadIndices;
|
|
getBufferIndices(xferOp, loadIndices);
|
|
loadIndices.push_back(iv);
|
|
|
|
SmallVector<Value, 8> xferIndices;
|
|
getXferIndices(b, xferOp, iv, xferIndices);
|
|
|
|
Location loc = xferOp.getLoc();
|
|
auto vec = b.create<memref::LoadOp>(loc, buffer, loadIndices);
|
|
auto inBoundsAttr = dropFirstElem(b, xferOp.getInBoundsAttr());
|
|
auto source = loopState.empty() ? xferOp.getBase() : loopState[0];
|
|
Type type = isTensorOp(xferOp) ? xferOp.getShapedType() : Type();
|
|
auto newXferOp = b.create<vector::TransferWriteOp>(
|
|
loc, type, vec, source, xferIndices,
|
|
AffineMapAttr::get(unpackedPermutationMap(b, xferOp)), Value(),
|
|
inBoundsAttr);
|
|
|
|
maybeApplyPassLabel(b, newXferOp, options.targetRank);
|
|
|
|
return newXferOp;
|
|
}
|
|
|
|
/// Handle out-of-bounds accesses on the to-be-unpacked dimension.
|
|
static Value handleOutOfBoundsDim(OpBuilder &b, TransferWriteOp xferOp,
|
|
Value buffer, Value iv,
|
|
ValueRange loopState) {
|
|
return isTensorOp(xferOp) ? loopState[0] : Value();
|
|
}
|
|
|
|
/// Cleanup after rewriting the op.
|
|
static void cleanup(PatternRewriter &rewriter, TransferWriteOp xferOp,
|
|
scf::ForOp forOp) {
|
|
if (isTensorOp(xferOp)) {
|
|
assert(forOp->getNumResults() == 1 && "Expected one for loop result");
|
|
rewriter.replaceOp(xferOp, forOp->getResult(0));
|
|
} else {
|
|
rewriter.eraseOp(xferOp);
|
|
}
|
|
}
|
|
|
|
/// Return the initial loop state for the generated scf.for loop.
|
|
static Value initialLoopState(TransferWriteOp xferOp) {
|
|
return isTensorOp(xferOp) ? xferOp.getBase() : Value();
|
|
}
|
|
};
|
|
|
|
template <typename OpTy>
|
|
static LogicalResult checkPrepareXferOp(OpTy xferOp, PatternRewriter &rewriter,
|
|
VectorTransferToSCFOptions options) {
|
|
if (xferOp->hasAttr(kPassLabel))
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "kPassLabel is present (vector-to-scf lowering in progress)");
|
|
if (xferOp.getVectorType().getRank() <= options.targetRank)
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "xferOp vector rank <= transformation target rank");
|
|
if (xferOp.getVectorType().getScalableDims().front())
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "Unpacking of the leading dimension into the memref is not yet "
|
|
"supported for scalable dims");
|
|
if (isTensorOp(xferOp) && !options.lowerTensors)
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "Unpacking for tensors has been disabled.");
|
|
if (xferOp.getVectorType().getElementType() !=
|
|
xferOp.getShapedType().getElementType())
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "Mismatching source and destination element types.");
|
|
|
|
return success();
|
|
}
|
|
|
|
/// Prepare a TransferReadOp for progressive lowering.
|
|
///
|
|
/// 1. Allocate a temporary buffer.
|
|
/// 2. Label the TransferReadOp, marking it eligible for progressive lowering.
|
|
/// 3. Store the result of the TransferReadOp into the temporary buffer.
|
|
/// 4. Load the result from the temporary buffer and replace all uses of the
|
|
/// original TransferReadOp with this load.
|
|
///
|
|
/// E.g.:
|
|
/// ```
|
|
/// %vec = vector.transfer_read %A[%a, %b, %c], %cst
|
|
/// : vector<5x4xf32>, memref<?x?x?xf32>
|
|
/// ```
|
|
/// is rewritten to:
|
|
/// ```
|
|
/// %0 = memref.alloca() : memref<vector<5x4xf32>>
|
|
/// %1 = vector.transfer_read %A[%a, %b, %c], %cst
|
|
/// { __vector_to_scf_lowering__ } : vector<5x4xf32>, memref<?x?x?xf32>
|
|
/// memref.store %1, %0[] : memref<vector<5x4xf32>>
|
|
/// %vec = memref.load %0[] : memref<vector<5x4xf32>>
|
|
/// ```
|
|
///
|
|
/// Note: A second temporary buffer may be allocated for the `mask` operand.
|
|
struct PrepareTransferReadConversion
|
|
: public VectorToSCFPattern<TransferReadOp> {
|
|
using VectorToSCFPattern<TransferReadOp>::VectorToSCFPattern;
|
|
|
|
LogicalResult matchAndRewrite(TransferReadOp xferOp,
|
|
PatternRewriter &rewriter) const override {
|
|
if (checkPrepareXferOp(xferOp, rewriter, options).failed())
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "checkPrepareXferOp conditions not met!");
|
|
|
|
auto buffers = allocBuffers(rewriter, xferOp);
|
|
auto *newXfer = rewriter.clone(*xferOp.getOperation());
|
|
newXfer->setAttr(kPassLabel, rewriter.getUnitAttr());
|
|
if (xferOp.getMask()) {
|
|
dyn_cast<TransferReadOp>(newXfer).getMaskMutable().assign(
|
|
buffers.maskBuffer);
|
|
}
|
|
|
|
Location loc = xferOp.getLoc();
|
|
rewriter.create<memref::StoreOp>(loc, newXfer->getResult(0),
|
|
buffers.dataBuffer);
|
|
rewriter.replaceOpWithNewOp<memref::LoadOp>(xferOp, buffers.dataBuffer);
|
|
|
|
return success();
|
|
}
|
|
};
|
|
|
|
/// Prepare a TransferWriteOp for progressive lowering.
|
|
///
|
|
/// 1. Allocate a temporary buffer.
|
|
/// 2. Store the vector into the buffer.
|
|
/// 3. Load the vector from the buffer again.
|
|
/// 4. Use the loaded vector as a TransferWriteOp operand and label the op,
|
|
/// marking it eligible for progressive lowering via TransferOpConversion.
|
|
///
|
|
/// E.g.:
|
|
/// ```
|
|
/// vector.transfer_write %vec, %A[%a, %b, %c]
|
|
/// : vector<5x4xf32>, memref<?x?x?xf32>
|
|
/// ```
|
|
/// is rewritten to:
|
|
/// ```
|
|
/// %0 = memref.alloca() : memref<vector<5x4xf32>>
|
|
/// memref.store %vec, %0[] : memref<vector<5x4xf32>>
|
|
/// %1 = memref.load %0[] : memref<vector<5x4xf32>>
|
|
/// vector.transfer_write %1, %A[%a, %b, %c] { __vector_to_scf_lowering__ }
|
|
/// : vector<5x4xf32>, memref<?x?x?xf32>
|
|
/// ```
|
|
///
|
|
/// Note: A second temporary buffer may be allocated for the `mask` operand.
|
|
struct PrepareTransferWriteConversion
|
|
: public VectorToSCFPattern<TransferWriteOp> {
|
|
using VectorToSCFPattern<TransferWriteOp>::VectorToSCFPattern;
|
|
|
|
LogicalResult matchAndRewrite(TransferWriteOp xferOp,
|
|
PatternRewriter &rewriter) const override {
|
|
if (checkPrepareXferOp(xferOp, rewriter, options).failed())
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "checkPrepareXferOp conditions not met!");
|
|
|
|
Location loc = xferOp.getLoc();
|
|
auto buffers = allocBuffers(rewriter, xferOp);
|
|
rewriter.create<memref::StoreOp>(loc, xferOp.getVector(),
|
|
buffers.dataBuffer);
|
|
auto loadedVec = rewriter.create<memref::LoadOp>(loc, buffers.dataBuffer);
|
|
rewriter.modifyOpInPlace(xferOp, [&]() {
|
|
xferOp.getValueToStoreMutable().assign(loadedVec);
|
|
xferOp->setAttr(kPassLabel, rewriter.getUnitAttr());
|
|
});
|
|
|
|
if (xferOp.getMask()) {
|
|
rewriter.modifyOpInPlace(xferOp, [&]() {
|
|
xferOp.getMaskMutable().assign(buffers.maskBuffer);
|
|
});
|
|
}
|
|
|
|
return success();
|
|
}
|
|
};
|
|
|
|
/// Decompose a n-D PrintOp into a loop of elementary/scalar prints. This allows
|
|
/// printing both 1D scalable vectors and n-D fixed size vectors.
|
|
///
|
|
/// E.g.:
|
|
/// ```
|
|
/// vector.print %v : vector<[4]xi32>
|
|
/// ```
|
|
/// is rewritten to:
|
|
/// ```
|
|
/// %c0 = arith.constant 0 : index
|
|
/// %c4 = arith.constant 4 : index
|
|
/// %c1 = arith.constant 1 : index
|
|
/// %vscale = vector.vscale
|
|
/// %length = arith.muli %vscale, %c4 : index
|
|
/// %lastIndex = arith.subi %length, %c1 : index
|
|
/// vector.print punctuation <open>
|
|
/// scf.for %i = %c0 to %length step %c1 {
|
|
/// %el = vector.extractelement %v[%i : index] : vector<[4]xi32>
|
|
/// vector.print %el : i32 punctuation <no_punctuation>
|
|
/// %notLastIndex = arith.cmpi ult, %i, %lastIndex : index
|
|
/// scf.if %notLastIndex {
|
|
/// vector.print punctuation <comma>
|
|
/// }
|
|
/// }
|
|
/// vector.print punctuation <close>
|
|
/// vector.print
|
|
/// ```
|
|
struct DecomposePrintOpConversion : public VectorToSCFPattern<vector::PrintOp> {
|
|
using VectorToSCFPattern<vector::PrintOp>::VectorToSCFPattern;
|
|
LogicalResult matchAndRewrite(vector::PrintOp printOp,
|
|
PatternRewriter &rewriter) const override {
|
|
if (!printOp.getSource())
|
|
return failure();
|
|
|
|
VectorType vectorType = dyn_cast<VectorType>(printOp.getPrintType());
|
|
if (!vectorType)
|
|
return failure();
|
|
|
|
// Currently >= 2D scalable vectors are not supported.
|
|
// These can't be lowered to LLVM (as LLVM does not support scalable vectors
|
|
// of scalable vectors), and due to limitations of current ops can't be
|
|
// indexed with SSA values or flattened. This may change after
|
|
// https://reviews.llvm.org/D155034, though there still needs to be a path
|
|
// for lowering to LLVM.
|
|
if (vectorType.getRank() > 1 && vectorType.isScalable())
|
|
return failure();
|
|
|
|
auto loc = printOp.getLoc();
|
|
auto value = printOp.getSource();
|
|
|
|
if (auto intTy = dyn_cast<IntegerType>(vectorType.getElementType())) {
|
|
// Oddly sized integers are (somewhat) buggy on a lot of backends, so to
|
|
// avoid issues extend them to a more standard size.
|
|
// https://github.com/llvm/llvm-project/issues/30613
|
|
auto width = intTy.getWidth();
|
|
auto legalWidth = llvm::NextPowerOf2(std::max(8u, width) - 1);
|
|
auto legalIntTy = IntegerType::get(rewriter.getContext(), legalWidth,
|
|
intTy.getSignedness());
|
|
// arith can only take signless integers, so we must cast back and forth.
|
|
auto signlessSourceVectorType =
|
|
vectorType.cloneWith({}, getIntTypeWithSignlessSemantics(intTy));
|
|
auto signlessTargetVectorType =
|
|
vectorType.cloneWith({}, getIntTypeWithSignlessSemantics(legalIntTy));
|
|
auto targetVectorType = vectorType.cloneWith({}, legalIntTy);
|
|
value = rewriter.create<vector::BitCastOp>(loc, signlessSourceVectorType,
|
|
value);
|
|
if (value.getType() != signlessTargetVectorType) {
|
|
if (width == 1 || intTy.isUnsigned())
|
|
value = rewriter.create<arith::ExtUIOp>(loc, signlessTargetVectorType,
|
|
value);
|
|
else
|
|
value = rewriter.create<arith::ExtSIOp>(loc, signlessTargetVectorType,
|
|
value);
|
|
}
|
|
value = rewriter.create<vector::BitCastOp>(loc, targetVectorType, value);
|
|
vectorType = targetVectorType;
|
|
}
|
|
|
|
auto scalableDimensions = vectorType.getScalableDims();
|
|
auto shape = vectorType.getShape();
|
|
constexpr int64_t singletonShape[] = {1};
|
|
if (vectorType.getRank() == 0)
|
|
shape = singletonShape;
|
|
|
|
if (vectorType.getRank() != 1) {
|
|
// Flatten n-D vectors to 1D. This is done to allow indexing with a
|
|
// non-constant value (which can currently only be done via
|
|
// vector.extractelement for 1D vectors).
|
|
auto flatLength = std::accumulate(shape.begin(), shape.end(), 1,
|
|
std::multiplies<int64_t>());
|
|
auto flatVectorType =
|
|
VectorType::get({flatLength}, vectorType.getElementType());
|
|
value = rewriter.create<vector::ShapeCastOp>(loc, flatVectorType, value);
|
|
}
|
|
|
|
vector::PrintOp firstClose;
|
|
SmallVector<Value, 8> loopIndices;
|
|
for (unsigned d = 0; d < shape.size(); d++) {
|
|
// Setup loop bounds and step.
|
|
Value lowerBound = rewriter.create<arith::ConstantIndexOp>(loc, 0);
|
|
Value upperBound = rewriter.create<arith::ConstantIndexOp>(loc, shape[d]);
|
|
Value step = rewriter.create<arith::ConstantIndexOp>(loc, 1);
|
|
if (!scalableDimensions.empty() && scalableDimensions[d]) {
|
|
auto vscale = rewriter.create<vector::VectorScaleOp>(
|
|
loc, rewriter.getIndexType());
|
|
upperBound = rewriter.create<arith::MulIOp>(loc, upperBound, vscale);
|
|
}
|
|
auto lastIndex = rewriter.create<arith::SubIOp>(loc, upperBound, step);
|
|
|
|
// Create a loop to print the elements surrounded by parentheses.
|
|
rewriter.create<vector::PrintOp>(loc, vector::PrintPunctuation::Open);
|
|
auto loop =
|
|
rewriter.create<scf::ForOp>(loc, lowerBound, upperBound, step);
|
|
auto printClose = rewriter.create<vector::PrintOp>(
|
|
loc, vector::PrintPunctuation::Close);
|
|
if (!firstClose)
|
|
firstClose = printClose;
|
|
|
|
auto loopIdx = loop.getInductionVar();
|
|
loopIndices.push_back(loopIdx);
|
|
|
|
// Print a comma after all but the last element.
|
|
rewriter.setInsertionPointToStart(loop.getBody());
|
|
auto notLastIndex = rewriter.create<arith::CmpIOp>(
|
|
loc, arith::CmpIPredicate::ult, loopIdx, lastIndex);
|
|
rewriter.create<scf::IfOp>(loc, notLastIndex,
|
|
[&](OpBuilder &builder, Location loc) {
|
|
builder.create<vector::PrintOp>(
|
|
loc, vector::PrintPunctuation::Comma);
|
|
builder.create<scf::YieldOp>(loc);
|
|
});
|
|
|
|
rewriter.setInsertionPointToStart(loop.getBody());
|
|
}
|
|
|
|
// Compute the flattened index.
|
|
// Note: For the > rank 1 vectors this assumes non-scalable.
|
|
Value flatIndex;
|
|
auto currentStride = 1;
|
|
for (int d = shape.size() - 1; d >= 0; d--) {
|
|
auto stride = rewriter.create<arith::ConstantIndexOp>(loc, currentStride);
|
|
auto index = rewriter.create<arith::MulIOp>(loc, stride, loopIndices[d]);
|
|
if (flatIndex)
|
|
flatIndex = rewriter.create<arith::AddIOp>(loc, flatIndex, index);
|
|
else
|
|
flatIndex = index;
|
|
currentStride *= shape[d];
|
|
}
|
|
|
|
// Print the scalar elements in the inner most loop.
|
|
auto element =
|
|
rewriter.create<vector::ExtractElementOp>(loc, value, flatIndex);
|
|
rewriter.create<vector::PrintOp>(loc, element,
|
|
vector::PrintPunctuation::NoPunctuation);
|
|
|
|
rewriter.setInsertionPointAfter(firstClose);
|
|
rewriter.create<vector::PrintOp>(loc, printOp.getPunctuation());
|
|
rewriter.eraseOp(printOp);
|
|
return success();
|
|
}
|
|
|
|
static IntegerType getIntTypeWithSignlessSemantics(IntegerType intTy) {
|
|
return IntegerType::get(intTy.getContext(), intTy.getWidth(),
|
|
IntegerType::Signless);
|
|
};
|
|
};
|
|
|
|
/// Progressive lowering of vector transfer ops: Unpack one dimension.
|
|
///
|
|
/// 1. Unpack one dimension from the current buffer type and cast the buffer
|
|
/// to that new type. E.g.:
|
|
/// ```
|
|
/// %vec = memref.load %0[%1] : memref<5xvector<4x3xf32>>
|
|
/// vector.transfer_write %vec ...
|
|
/// ```
|
|
/// The following cast is generated:
|
|
/// ```
|
|
/// %casted = vector.type_cast %0
|
|
/// : memref<5xvector<4x3xf32>> to memref<5x4xvector<3xf32>>
|
|
/// ```
|
|
/// 2. Generate a for loop and rewrite the transfer op according to the
|
|
/// corresponding Strategy<OpTy>. If the to-be-unpacked dimension can be
|
|
/// out-of-bounds, generate an if-check and handle both cases separately.
|
|
/// 3. Clean up according to the corresponding Strategy<OpTy>.
|
|
///
|
|
/// Note: If the transfer op is a TransferWriteOp and operates on a tensor
|
|
/// source (as opposed to a memref source), then each iteration of the generated
|
|
/// scf.for loop yields the new tensor value. E.g.:
|
|
/// ```
|
|
/// %result = scf.for i = 0 to 5 {
|
|
/// %0 = memref.load %buffer[i] : memref<5xvector<4x3xf32>>
|
|
/// %1 = vector.transfer_write %0, %source[...]
|
|
/// : vector<4x3xf32>, tensor<5x4x3xf32>
|
|
/// scf.yield %1 : tensor<5x4x3xf32>
|
|
/// }
|
|
/// ```
|
|
template <typename OpTy>
|
|
struct TransferOpConversion : public VectorToSCFPattern<OpTy> {
|
|
using VectorToSCFPattern<OpTy>::VectorToSCFPattern;
|
|
|
|
void initialize() {
|
|
// This pattern recursively unpacks one dimension at a time. The recursion
|
|
// bounded as the rank is strictly decreasing.
|
|
this->setHasBoundedRewriteRecursion();
|
|
}
|
|
|
|
static void getMaskBufferLoadIndices(OpTy xferOp, Value castedMaskBuffer,
|
|
SmallVectorImpl<Value> &loadIndices,
|
|
Value iv) {
|
|
assert(xferOp.getMask() && "Expected transfer op to have mask");
|
|
|
|
// Add load indices from the previous iteration.
|
|
// The mask buffer depends on the permutation map, which makes determining
|
|
// the indices quite complex, so this is why we need to "look back" to the
|
|
// previous iteration to find the right indices.
|
|
Value maskBuffer = getMaskBuffer(xferOp);
|
|
for (Operation *user : maskBuffer.getUsers()) {
|
|
// If there is no previous load op, then the indices are empty.
|
|
if (auto loadOp = dyn_cast<memref::LoadOp>(user)) {
|
|
Operation::operand_range prevIndices = loadOp.getIndices();
|
|
loadIndices.append(prevIndices.begin(), prevIndices.end());
|
|
break;
|
|
}
|
|
}
|
|
|
|
// In case of broadcast: Use same indices to load from memref
|
|
// as before.
|
|
if (!xferOp.isBroadcastDim(0))
|
|
loadIndices.push_back(iv);
|
|
}
|
|
|
|
LogicalResult matchAndRewrite(OpTy xferOp,
|
|
PatternRewriter &rewriter) const override {
|
|
if (!xferOp->hasAttr(kPassLabel))
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "kPassLabel is present (progressing lowering in progress)");
|
|
|
|
// Find and cast data buffer. How the buffer can be found depends on OpTy.
|
|
ImplicitLocOpBuilder locB(xferOp.getLoc(), rewriter);
|
|
Value dataBuffer = Strategy<OpTy>::getBuffer(xferOp);
|
|
auto dataBufferType = dyn_cast<MemRefType>(dataBuffer.getType());
|
|
FailureOr<MemRefType> castedDataType = unpackOneDim(dataBufferType);
|
|
if (failed(castedDataType))
|
|
return rewriter.notifyMatchFailure(xferOp,
|
|
"Failed to unpack one vector dim.");
|
|
|
|
auto castedDataBuffer =
|
|
locB.create<vector::TypeCastOp>(*castedDataType, dataBuffer);
|
|
|
|
// If the xferOp has a mask: Find and cast mask buffer.
|
|
Value castedMaskBuffer;
|
|
if (xferOp.getMask()) {
|
|
Value maskBuffer = getMaskBuffer(xferOp);
|
|
if (xferOp.isBroadcastDim(0) || xferOp.getMaskType().getRank() == 1) {
|
|
// Do not unpack a dimension of the mask, if:
|
|
// * To-be-unpacked transfer op dimension is a broadcast.
|
|
// * Mask is 1D, i.e., the mask cannot be further unpacked.
|
|
// (That means that all remaining dimensions of the transfer op must
|
|
// be broadcasted.)
|
|
castedMaskBuffer = maskBuffer;
|
|
} else {
|
|
// It's safe to assume the mask buffer can be unpacked if the data
|
|
// buffer was unpacked.
|
|
auto maskBufferType = cast<MemRefType>(maskBuffer.getType());
|
|
MemRefType castedMaskType = *unpackOneDim(maskBufferType);
|
|
castedMaskBuffer =
|
|
locB.create<vector::TypeCastOp>(castedMaskType, maskBuffer);
|
|
}
|
|
}
|
|
|
|
// Loop bounds and step.
|
|
auto lb = locB.create<arith::ConstantIndexOp>(0);
|
|
auto ub = locB.create<arith::ConstantIndexOp>(
|
|
castedDataType->getDimSize(castedDataType->getRank() - 1));
|
|
auto step = locB.create<arith::ConstantIndexOp>(1);
|
|
// TransferWriteOps that operate on tensors return the modified tensor and
|
|
// require a loop state.
|
|
auto loopState = Strategy<OpTy>::initialLoopState(xferOp);
|
|
|
|
// Generate for loop.
|
|
auto result = locB.create<scf::ForOp>(
|
|
lb, ub, step, loopState ? ValueRange(loopState) : ValueRange(),
|
|
[&](OpBuilder &b, Location loc, Value iv, ValueRange loopState) {
|
|
Type stateType = loopState.empty() ? Type() : loopState[0].getType();
|
|
|
|
auto result = generateInBoundsCheck(
|
|
b, xferOp, iv, unpackedDim(xferOp),
|
|
stateType ? TypeRange(stateType) : TypeRange(),
|
|
/*inBoundsCase=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
// Create new transfer op.
|
|
OpTy newXfer = Strategy<OpTy>::rewriteOp(
|
|
b, this->options, xferOp, castedDataBuffer, iv, loopState);
|
|
|
|
// If old transfer op has a mask: Set mask on new transfer op.
|
|
// Special case: If the mask of the old transfer op is 1D and
|
|
// the unpacked dim is not a broadcast, no mask is needed on
|
|
// the new transfer op.
|
|
if (xferOp.getMask() && (xferOp.isBroadcastDim(0) ||
|
|
xferOp.getMaskType().getRank() > 1)) {
|
|
OpBuilder::InsertionGuard guard(b);
|
|
b.setInsertionPoint(newXfer); // Insert load before newXfer.
|
|
|
|
SmallVector<Value, 8> loadIndices;
|
|
getMaskBufferLoadIndices(xferOp, castedMaskBuffer,
|
|
loadIndices, iv);
|
|
auto mask = b.create<memref::LoadOp>(loc, castedMaskBuffer,
|
|
loadIndices);
|
|
rewriter.modifyOpInPlace(newXfer, [&]() {
|
|
newXfer.getMaskMutable().assign(mask);
|
|
});
|
|
}
|
|
|
|
return loopState.empty() ? Value() : newXfer->getResult(0);
|
|
},
|
|
/*outOfBoundsCase=*/
|
|
[&](OpBuilder &b, Location /*loc*/) {
|
|
return Strategy<OpTy>::handleOutOfBoundsDim(
|
|
b, xferOp, castedDataBuffer, iv, loopState);
|
|
});
|
|
|
|
maybeYieldValue(b, loc, !loopState.empty(), result);
|
|
});
|
|
|
|
Strategy<OpTy>::cleanup(rewriter, xferOp, result);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
/// Retrieves the dimensions sizes of a mask. Currently supports CreateMaskOp
|
|
/// and ConstantMaskOp.
|
|
template <typename VscaleConstantBuilder>
|
|
static FailureOr<SmallVector<OpFoldResult>>
|
|
getMaskDimSizes(Value mask, VscaleConstantBuilder &createVscaleMultiple) {
|
|
if (!mask)
|
|
return SmallVector<OpFoldResult>{};
|
|
if (auto createMaskOp = mask.getDefiningOp<vector::CreateMaskOp>()) {
|
|
return llvm::map_to_vector(createMaskOp.getOperands(), [](Value dimSize) {
|
|
return OpFoldResult(dimSize);
|
|
});
|
|
}
|
|
if (auto constantMask = mask.getDefiningOp<vector::ConstantMaskOp>()) {
|
|
int dimIdx = 0;
|
|
VectorType maskType = constantMask.getVectorType();
|
|
auto indexType = IndexType::get(mask.getContext());
|
|
return llvm::map_to_vector(
|
|
constantMask.getMaskDimSizes(), [&](int64_t dimSize) {
|
|
// A scalable dim in a constant_mask means vscale x dimSize.
|
|
if (maskType.getScalableDims()[dimIdx++])
|
|
return OpFoldResult(createVscaleMultiple(dimSize));
|
|
return OpFoldResult(IntegerAttr::get(indexType, dimSize));
|
|
});
|
|
}
|
|
return failure();
|
|
}
|
|
|
|
/// Scalable vector lowering of transfer_write(transpose). This lowering only
|
|
/// supports rank 2 (scalable) vectors, but can be used in conjunction with
|
|
/// `UnrollTransferWriteConversion` to support n-D cases. The unroll conversion
|
|
/// unrolls until the first scalable dimension.
|
|
///
|
|
/// Example:
|
|
///
|
|
/// BEFORE:
|
|
/// ```mlir
|
|
/// %transpose = vector.transpose %vec, [1, 0]
|
|
/// : vector<4x[4]xf32> to vector<[4]x4xf32>
|
|
/// vector.transfer_write %transpose, %dest[%i, %j] {in_bounds = [true, true]}
|
|
/// : vector<[4]x4xf32>, memref<?x?xf32>
|
|
/// ```
|
|
///
|
|
/// AFTER:
|
|
/// ```mlir
|
|
/// %c1 = arith.constant 1 : index
|
|
/// %c4 = arith.constant 4 : index
|
|
/// %c0 = arith.constant 0 : index
|
|
/// %0 = vector.extract %arg0[0] : vector<[4]xf32> from vector<4x[4]xf32>
|
|
/// %1 = vector.extract %arg0[1] : vector<[4]xf32> from vector<4x[4]xf32>
|
|
/// %2 = vector.extract %arg0[2] : vector<[4]xf32> from vector<4x[4]xf32>
|
|
/// %3 = vector.extract %arg0[3] : vector<[4]xf32> from vector<4x[4]xf32>
|
|
/// %vscale = vector.vscale
|
|
/// %c4_vscale = arith.muli %vscale, %c4 : index
|
|
/// scf.for %idx = %c0 to %c4_vscale step %c1 {
|
|
/// %4 = vector.extract %0[%idx] : f32 from vector<[4]xf32>
|
|
/// %5 = vector.extract %1[%idx] : f32 from vector<[4]xf32>
|
|
/// %6 = vector.extract %2[%idx] : f32 from vector<[4]xf32>
|
|
/// %7 = vector.extract %3[%idx] : f32 from vector<[4]xf32>
|
|
/// %slice_i = affine.apply #map(%idx)[%i]
|
|
/// %slice = vector.from_elements %4, %5, %6, %7 : vector<4xf32>
|
|
/// vector.transfer_write %slice, %arg1[%slice_i, %j] {in_bounds = [true]}
|
|
/// : vector<4xf32>, memref<?x?xf32>
|
|
/// }
|
|
/// ```
|
|
struct ScalableTransposeTransferWriteConversion
|
|
: VectorToSCFPattern<vector::TransferWriteOp> {
|
|
using VectorToSCFPattern::VectorToSCFPattern;
|
|
|
|
LogicalResult matchAndRewrite(TransferWriteOp writeOp,
|
|
PatternRewriter &rewriter) const override {
|
|
if (failed(checkLowerTensors(writeOp, rewriter)))
|
|
return failure();
|
|
|
|
VectorType vectorType = writeOp.getVectorType();
|
|
|
|
// Note: By comparing the scalable dims to an ArrayRef of length two this
|
|
// implicitly checks the rank (is also two).
|
|
ArrayRef<bool> scalableFlags = vectorType.getScalableDims();
|
|
if (scalableFlags != ArrayRef<bool>{true, false}) {
|
|
return rewriter.notifyMatchFailure(
|
|
writeOp, "expected vector of the form vector<[N]xMxty>");
|
|
}
|
|
|
|
auto permutationMap = writeOp.getPermutationMap();
|
|
if (!permutationMap.isIdentity()) {
|
|
return rewriter.notifyMatchFailure(
|
|
writeOp, "non-identity permutations are unsupported (lower first)");
|
|
}
|
|
|
|
// Note: This pattern is only lowering the leading dimension (to a loop),
|
|
// so we only check if the leading dimension is in bounds. The in-bounds
|
|
// attribute for the trailing dimension will be propagated.
|
|
if (!writeOp.isDimInBounds(0)) {
|
|
return rewriter.notifyMatchFailure(
|
|
writeOp, "out-of-bounds dims are unsupported (use masking)");
|
|
}
|
|
|
|
Value vector = writeOp.getVector();
|
|
auto transposeOp = vector.getDefiningOp<vector::TransposeOp>();
|
|
if (!transposeOp ||
|
|
transposeOp.getPermutation() != ArrayRef<int64_t>{1, 0}) {
|
|
return rewriter.notifyMatchFailure(writeOp, "source not transpose");
|
|
}
|
|
|
|
auto loc = writeOp.getLoc();
|
|
auto createVscaleMultiple =
|
|
vector::makeVscaleConstantBuilder(rewriter, loc);
|
|
|
|
auto maskDims = getMaskDimSizes(writeOp.getMask(), createVscaleMultiple);
|
|
if (failed(maskDims)) {
|
|
return rewriter.notifyMatchFailure(writeOp,
|
|
"failed to resolve mask dims");
|
|
}
|
|
|
|
int64_t fixedDimSize = vectorType.getDimSize(1);
|
|
auto fixedDimOffsets = llvm::seq(fixedDimSize);
|
|
|
|
// Extract all slices from the source of the transpose.
|
|
auto transposeSource = transposeOp.getVector();
|
|
SmallVector<Value> transposeSourceSlices =
|
|
llvm::map_to_vector(fixedDimOffsets, [&](int64_t idx) -> Value {
|
|
return rewriter.create<vector::ExtractOp>(loc, transposeSource, idx);
|
|
});
|
|
|
|
// Loop bounds and step.
|
|
auto lb = rewriter.create<arith::ConstantIndexOp>(loc, 0);
|
|
auto ub =
|
|
maskDims->empty()
|
|
? Value(createVscaleMultiple(vectorType.getDimSize(0)))
|
|
: vector::getAsValues(rewriter, loc, maskDims->front()).front();
|
|
auto step = rewriter.create<arith::ConstantIndexOp>(loc, 1);
|
|
|
|
// Generate a new mask for the slice.
|
|
VectorType sliceType = VectorType::Builder(vectorType).dropDim(0);
|
|
Value sliceMask = nullptr;
|
|
if (!maskDims->empty()) {
|
|
sliceMask = rewriter.create<vector::CreateMaskOp>(
|
|
loc, sliceType.clone(rewriter.getI1Type()),
|
|
ArrayRef<OpFoldResult>(*maskDims).drop_front());
|
|
}
|
|
|
|
Value initDest = isTensorOp(writeOp) ? writeOp.getBase() : Value{};
|
|
ValueRange initLoopArgs = initDest ? initDest : ValueRange{};
|
|
auto result = rewriter.create<scf::ForOp>(
|
|
loc, lb, ub, step, initLoopArgs,
|
|
[&](OpBuilder &b, Location loc, Value iv, ValueRange loopIterArgs) {
|
|
// Indices for the new transfer op.
|
|
SmallVector<Value, 8> xferIndices;
|
|
getXferIndices(b, writeOp, iv, xferIndices);
|
|
|
|
// Extract a transposed slice from the source vector.
|
|
SmallVector<Value> transposeElements =
|
|
llvm::map_to_vector(fixedDimOffsets, [&](int64_t idx) -> Value {
|
|
return b.create<vector::ExtractOp>(
|
|
loc, transposeSourceSlices[idx], iv);
|
|
});
|
|
auto sliceVec = b.create<vector::FromElementsOp>(loc, sliceType,
|
|
transposeElements);
|
|
|
|
// Create the transfer_write for the slice.
|
|
Value dest =
|
|
loopIterArgs.empty() ? writeOp.getBase() : loopIterArgs.front();
|
|
auto newWriteOp = b.create<vector::TransferWriteOp>(
|
|
loc, sliceVec, dest, xferIndices,
|
|
ArrayRef<bool>(writeOp.getInBoundsValues()).drop_front());
|
|
if (sliceMask)
|
|
newWriteOp.getMaskMutable().assign(sliceMask);
|
|
|
|
// Yield from the loop.
|
|
b.create<scf::YieldOp>(loc, loopIterArgs.empty()
|
|
? ValueRange{}
|
|
: newWriteOp.getResult());
|
|
});
|
|
|
|
if (isTensorOp(writeOp))
|
|
rewriter.replaceOp(writeOp, result);
|
|
else
|
|
rewriter.eraseOp(writeOp);
|
|
|
|
return success();
|
|
}
|
|
};
|
|
|
|
} // namespace lowering_n_d
|
|
|
|
namespace lowering_n_d_unrolled {
|
|
|
|
/// If the original transfer op has a mask, compute the mask of the new transfer
|
|
/// op (for the current iteration `i`) and assign it.
|
|
template <typename OpTy>
|
|
static void maybeAssignMask(OpBuilder &b, OpTy xferOp, OpTy newXferOp,
|
|
int64_t i) {
|
|
if (!xferOp.getMask())
|
|
return;
|
|
|
|
if (xferOp.isBroadcastDim(0)) {
|
|
// To-be-unpacked dimension is a broadcast, which does not have a
|
|
// corresponding mask dimension. Mask attribute remains unchanged.
|
|
newXferOp.getMaskMutable().assign(xferOp.getMask());
|
|
return;
|
|
}
|
|
|
|
if (xferOp.getMaskType().getRank() > 1) {
|
|
// Unpack one dimension of the mask.
|
|
OpBuilder::InsertionGuard guard(b);
|
|
b.setInsertionPoint(newXferOp); // Insert load before newXfer.
|
|
|
|
llvm::SmallVector<int64_t, 1> indices({i});
|
|
Location loc = xferOp.getLoc();
|
|
auto newMask = b.create<vector::ExtractOp>(loc, xferOp.getMask(), indices);
|
|
newXferOp.getMaskMutable().assign(newMask);
|
|
}
|
|
|
|
// If we end up here: The mask of the old transfer op is 1D and the unpacked
|
|
// dim is not a broadcast, so no mask is needed on the new transfer op.
|
|
// `generateInBoundsCheck` will have evaluated the mask already.
|
|
}
|
|
|
|
/// Progressive lowering of vector TransferReadOp with unrolling: Unpack one
|
|
/// dimension. This is similar to TransferOpConversion<TransferReadOp>, but no
|
|
/// memref buffer is allocated and the SCF loop is fully unrolled.
|
|
///
|
|
/// ```
|
|
/// E.g.:
|
|
/// ```
|
|
/// %vec = vector.transfer_read %A[%a, %b, %c], %padding
|
|
/// : memref<?x?x?xf32>, vector<5x4xf32>
|
|
/// ```
|
|
/// is rewritten to IR such as (simplified):
|
|
/// ```
|
|
/// %v_init = splat %padding : vector<5x4xf32>
|
|
/// %tmp0 = vector.transfer_read %A[%a, %b, %c], %padding
|
|
/// : memref<?x?x?xf32>, vector<4xf32>
|
|
/// %v0 = vector.insert %tmp0, %v_init[0] : vector<4xf32> into vector<5x4xf32>
|
|
/// %tmp1 = vector.transfer_read %A[%a, %b + 1, %c], %padding
|
|
/// : memref<?x?x?xf32>, vector<4xf32>
|
|
/// %v1 = vector.insert %tmp1, %v0[1] : vector<4xf32> into vector<5x4xf32>
|
|
/// ...
|
|
/// %tmp4 = vector.transfer_read %A[%a, %b + 4, %c], %padding
|
|
/// : memref<?x?x?xf32>, vector<4xf32>
|
|
/// %vec = vector.insert %tmp1, %v3[4] : vector<4xf32> into vector<5x4xf32>
|
|
/// ```
|
|
///
|
|
/// Note: As an optimization, if the result of the original TransferReadOp
|
|
/// was directly inserted into another vector, no new %v_init vector is created.
|
|
/// Instead, the new TransferReadOp results are inserted into that vector.
|
|
struct UnrollTransferReadConversion
|
|
: public VectorToSCFPattern<TransferReadOp> {
|
|
using VectorToSCFPattern<TransferReadOp>::VectorToSCFPattern;
|
|
|
|
void initialize() {
|
|
// This pattern recursively unpacks one dimension at a time. The recursion
|
|
// bounded as the rank is strictly decreasing.
|
|
setHasBoundedRewriteRecursion();
|
|
}
|
|
|
|
/// Get or build the vector into which the newly created TransferReadOp
|
|
/// results are inserted.
|
|
Value buildResultVector(PatternRewriter &rewriter,
|
|
TransferReadOp xferOp) const {
|
|
if (auto insertOp = getInsertOp(xferOp))
|
|
return insertOp.getDest();
|
|
Location loc = xferOp.getLoc();
|
|
return rewriter.create<vector::SplatOp>(loc, xferOp.getVectorType(),
|
|
xferOp.getPadding());
|
|
}
|
|
|
|
/// If the result of the TransferReadOp has exactly one user, which is a
|
|
/// vector::InsertOp, return that operation.
|
|
vector::InsertOp getInsertOp(TransferReadOp xferOp) const {
|
|
if (xferOp->hasOneUse()) {
|
|
Operation *xferOpUser = *xferOp->getUsers().begin();
|
|
if (auto insertOp = dyn_cast<vector::InsertOp>(xferOpUser))
|
|
return insertOp;
|
|
}
|
|
|
|
return vector::InsertOp();
|
|
}
|
|
|
|
/// If the result of the TransferReadOp has exactly one user, which is a
|
|
/// vector::InsertOp, return that operation's indices.
|
|
void getInsertionIndices(TransferReadOp xferOp,
|
|
SmallVectorImpl<OpFoldResult> &indices) const {
|
|
if (auto insertOp = getInsertOp(xferOp)) {
|
|
auto pos = insertOp.getMixedPosition();
|
|
indices.append(pos.begin(), pos.end());
|
|
}
|
|
}
|
|
|
|
/// Rewrite the op: Unpack one dimension. Can handle masks, out-of-bounds
|
|
/// accesses, and broadcasts and transposes in permutation maps.
|
|
LogicalResult matchAndRewrite(TransferReadOp xferOp,
|
|
PatternRewriter &rewriter) const override {
|
|
if (xferOp.getVectorType().getRank() <= options.targetRank)
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "vector rank is less or equal to target rank");
|
|
if (failed(checkLowerTensors(xferOp, rewriter)))
|
|
return failure();
|
|
if (xferOp.getVectorType().getElementType() !=
|
|
xferOp.getShapedType().getElementType())
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "not yet supported: element type mismatch");
|
|
auto xferVecType = xferOp.getVectorType();
|
|
if (xferVecType.getScalableDims()[0]) {
|
|
return rewriter.notifyMatchFailure(
|
|
xferOp, "scalable dimensions cannot be unrolled at compile time");
|
|
}
|
|
|
|
auto insertOp = getInsertOp(xferOp);
|
|
auto vec = buildResultVector(rewriter, xferOp);
|
|
auto vecType = dyn_cast<VectorType>(vec.getType());
|
|
|
|
VectorType newXferVecType = VectorType::Builder(xferVecType).dropDim(0);
|
|
|
|
int64_t dimSize = xferVecType.getShape()[0];
|
|
|
|
// Generate fully unrolled loop of transfer ops.
|
|
Location loc = xferOp.getLoc();
|
|
for (int64_t i = 0; i < dimSize; ++i) {
|
|
Value iv = rewriter.create<arith::ConstantIndexOp>(loc, i);
|
|
|
|
vec = generateInBoundsCheck(
|
|
rewriter, xferOp, iv, unpackedDim(xferOp), TypeRange(vecType),
|
|
/*inBoundsCase=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
// Indices for the new transfer op.
|
|
SmallVector<Value, 8> xferIndices;
|
|
getXferIndices(b, xferOp, iv, xferIndices);
|
|
|
|
// Indices for the new vector.insert op.
|
|
SmallVector<OpFoldResult, 8> insertionIndices;
|
|
getInsertionIndices(xferOp, insertionIndices);
|
|
insertionIndices.push_back(rewriter.getIndexAttr(i));
|
|
|
|
auto inBoundsAttr = dropFirstElem(b, xferOp.getInBoundsAttr());
|
|
auto newXferOp = b.create<vector::TransferReadOp>(
|
|
loc, newXferVecType, xferOp.getBase(), xferIndices,
|
|
AffineMapAttr::get(unpackedPermutationMap(b, xferOp)),
|
|
xferOp.getPadding(), Value(), inBoundsAttr);
|
|
maybeAssignMask(b, xferOp, newXferOp, i);
|
|
return b.create<vector::InsertOp>(loc, newXferOp, vec,
|
|
insertionIndices);
|
|
},
|
|
/*outOfBoundsCase=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
// Loop through original (unmodified) vector.
|
|
return vec;
|
|
});
|
|
}
|
|
|
|
if (insertOp) {
|
|
// Rewrite single user of the old TransferReadOp, which was an InsertOp.
|
|
rewriter.replaceOp(insertOp, vec);
|
|
rewriter.eraseOp(xferOp);
|
|
} else {
|
|
rewriter.replaceOp(xferOp, vec);
|
|
}
|
|
|
|
return success();
|
|
}
|
|
};
|
|
|
|
/// Progressive lowering of vector TransferWriteOp with unrolling: Unpack one
|
|
/// dimension. This is similar to TransferOpConversion<TransferWriteOp>, but no
|
|
/// memref buffer is allocated and the SCF loop is fully unrolled.
|
|
///
|
|
/// ```
|
|
/// E.g.:
|
|
/// ```
|
|
/// vector.transfer_write %vec, %A[%a, %b, %c]
|
|
/// : vector<5x4xf32>, memref<?x?x?xf32>
|
|
/// ```
|
|
/// is rewritten to IR such as (simplified):
|
|
/// ```
|
|
/// %v0 = vector.extract %vec[0] : vector<4xf32> from vector<5x4xf32>
|
|
/// vector.transfer_write %v0, %A[%a, %b, %c] : vector<4xf32>, memref<...>
|
|
/// %v1 = vector.extract %vec[1] : vector<4xf32> from vector<5x4xf32>
|
|
/// vector.transfer_write %v1, %A[%a, %b + 1, %c] : vector<4xf32>, memref<...>
|
|
/// ...
|
|
/// %v4 = vector.extract %vec[4] : vector<4xf32> from vector<5x4xf32>
|
|
/// vector.transfer_write %v4, %A[%a, %b + 4, %c] : vector<4xf32>, memref<...>
|
|
/// ```
|
|
///
|
|
/// Note: As an optimization, if the vector of the original TransferWriteOp
|
|
/// was directly extracted from another vector via an ExtractOp `a`, extract
|
|
/// the vectors for the newly generated TransferWriteOps from `a`'s input. By
|
|
/// doing so, `a` may become dead, and the number of ExtractOps generated during
|
|
/// recursive application of this pattern will be minimal.
|
|
struct UnrollTransferWriteConversion
|
|
: public VectorToSCFPattern<TransferWriteOp> {
|
|
using VectorToSCFPattern<TransferWriteOp>::VectorToSCFPattern;
|
|
|
|
void initialize() {
|
|
// This pattern recursively unpacks one dimension at a time. The recursion
|
|
// bounded as the rank is strictly decreasing.
|
|
setHasBoundedRewriteRecursion();
|
|
}
|
|
|
|
/// Return the vector from which newly generated ExtracOps will extract.
|
|
Value getDataVector(TransferWriteOp xferOp) const {
|
|
if (auto extractOp = getExtractOp(xferOp))
|
|
return extractOp.getVector();
|
|
return xferOp.getVector();
|
|
}
|
|
|
|
/// If the input of the given TransferWriteOp is an ExtractOp, return it.
|
|
vector::ExtractOp getExtractOp(TransferWriteOp xferOp) const {
|
|
if (auto *op = xferOp.getVector().getDefiningOp())
|
|
return dyn_cast<vector::ExtractOp>(op);
|
|
return vector::ExtractOp();
|
|
}
|
|
|
|
/// If the input of the given TransferWriteOp is an ExtractOp, return its
|
|
/// indices.
|
|
void getExtractionIndices(TransferWriteOp xferOp,
|
|
SmallVectorImpl<OpFoldResult> &indices) const {
|
|
if (auto extractOp = getExtractOp(xferOp)) {
|
|
auto pos = extractOp.getMixedPosition();
|
|
indices.append(pos.begin(), pos.end());
|
|
}
|
|
}
|
|
|
|
/// Rewrite the op: Unpack one dimension. Can handle masks, out-of-bounds
|
|
/// accesses, and broadcasts and transposes in permutation maps.
|
|
LogicalResult matchAndRewrite(TransferWriteOp xferOp,
|
|
PatternRewriter &rewriter) const override {
|
|
VectorType inputVectorTy = xferOp.getVectorType();
|
|
|
|
if (inputVectorTy.getRank() <= options.targetRank)
|
|
return failure();
|
|
|
|
if (failed(checkLowerTensors(xferOp, rewriter)))
|
|
return failure();
|
|
// Transfer ops that modify the element type are not supported atm.
|
|
if (inputVectorTy.getElementType() !=
|
|
xferOp.getShapedType().getElementType())
|
|
return failure();
|
|
|
|
auto vec = getDataVector(xferOp);
|
|
if (inputVectorTy.getScalableDims()[0]) {
|
|
// Cannot unroll a scalable dimension at compile time.
|
|
return failure();
|
|
}
|
|
|
|
int64_t dimSize = inputVectorTy.getShape()[0];
|
|
Value source = xferOp.getBase(); // memref or tensor to be written to.
|
|
auto sourceType = isTensorOp(xferOp) ? xferOp.getShapedType() : Type();
|
|
|
|
// Generate fully unrolled loop of transfer ops.
|
|
Location loc = xferOp.getLoc();
|
|
for (int64_t i = 0; i < dimSize; ++i) {
|
|
Value iv = rewriter.create<arith::ConstantIndexOp>(loc, i);
|
|
|
|
auto updatedSource = generateInBoundsCheck(
|
|
rewriter, xferOp, iv, unpackedDim(xferOp),
|
|
isTensorOp(xferOp) ? TypeRange(sourceType) : TypeRange(),
|
|
/*inBoundsCase=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
// Indices for the new transfer op.
|
|
SmallVector<Value, 8> xferIndices;
|
|
getXferIndices(b, xferOp, iv, xferIndices);
|
|
|
|
// Indices for the new vector.extract op.
|
|
SmallVector<OpFoldResult, 8> extractionIndices;
|
|
getExtractionIndices(xferOp, extractionIndices);
|
|
extractionIndices.push_back(b.getI64IntegerAttr(i));
|
|
|
|
auto extracted =
|
|
b.create<vector::ExtractOp>(loc, vec, extractionIndices);
|
|
auto inBoundsAttr = dropFirstElem(b, xferOp.getInBoundsAttr());
|
|
Value xferVec;
|
|
if (inputVectorTy.getRank() == 1) {
|
|
// When target-rank=0, unrolling would causes the vector input
|
|
// argument into `transfer_write` to become a scalar. We solve
|
|
// this by broadcasting the scalar to a 0D vector.
|
|
xferVec = b.create<vector::BroadcastOp>(
|
|
loc, VectorType::get({}, extracted.getType()), extracted);
|
|
} else {
|
|
xferVec = extracted;
|
|
}
|
|
auto newXferOp = b.create<vector::TransferWriteOp>(
|
|
loc, sourceType, xferVec, source, xferIndices,
|
|
AffineMapAttr::get(unpackedPermutationMap(b, xferOp)), Value(),
|
|
inBoundsAttr);
|
|
|
|
maybeAssignMask(b, xferOp, newXferOp, i);
|
|
|
|
return isTensorOp(xferOp) ? newXferOp->getResult(0) : Value();
|
|
},
|
|
/*outOfBoundsCase=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
return isTensorOp(xferOp) ? source : Value();
|
|
});
|
|
|
|
if (isTensorOp(xferOp))
|
|
source = updatedSource;
|
|
}
|
|
|
|
if (isTensorOp(xferOp))
|
|
rewriter.replaceOp(xferOp, source);
|
|
else
|
|
rewriter.eraseOp(xferOp);
|
|
|
|
return success();
|
|
}
|
|
};
|
|
|
|
} // namespace lowering_n_d_unrolled
|
|
|
|
namespace lowering_1_d {
|
|
|
|
/// Compute the indices into the memref for the LoadOp/StoreOp generated as
|
|
/// part of TransferOp1dConversion. Return the memref dimension on which
|
|
/// the transfer is operating. A return value of std::nullopt indicates a
|
|
/// broadcast.
|
|
template <typename OpTy>
|
|
static std::optional<int64_t>
|
|
get1dMemrefIndices(OpBuilder &b, OpTy xferOp, Value iv,
|
|
SmallVector<Value, 8> &memrefIndices) {
|
|
auto indices = xferOp.getIndices();
|
|
auto map = xferOp.getPermutationMap();
|
|
assert(xferOp.getTransferRank() > 0 && "unexpected 0-d transfer");
|
|
|
|
memrefIndices.append(indices.begin(), indices.end());
|
|
assert(map.getNumResults() == 1 &&
|
|
"Expected 1 permutation map result for 1D transfer");
|
|
if (auto expr = dyn_cast<AffineDimExpr>(map.getResult(0))) {
|
|
Location loc = xferOp.getLoc();
|
|
auto dim = expr.getPosition();
|
|
AffineExpr d0, d1;
|
|
bindDims(xferOp.getContext(), d0, d1);
|
|
Value offset = memrefIndices[dim];
|
|
memrefIndices[dim] =
|
|
affine::makeComposedAffineApply(b, loc, d0 + d1, {offset, iv});
|
|
return dim;
|
|
}
|
|
|
|
assert(xferOp.isBroadcastDim(0) &&
|
|
"Expected AffineDimExpr or AffineConstantExpr");
|
|
return std::nullopt;
|
|
}
|
|
|
|
/// Codegen strategy for TransferOp1dConversion, depending on the
|
|
/// operation.
|
|
template <typename OpTy>
|
|
struct Strategy1d;
|
|
|
|
/// Codegen strategy for TransferReadOp.
|
|
template <>
|
|
struct Strategy1d<TransferReadOp> {
|
|
static void generateForLoopBody(OpBuilder &b, Location loc,
|
|
TransferReadOp xferOp, Value iv,
|
|
ValueRange loopState) {
|
|
SmallVector<Value, 8> indices;
|
|
auto dim = get1dMemrefIndices(b, xferOp, iv, indices);
|
|
auto vec = loopState[0];
|
|
|
|
// In case of out-of-bounds access, leave `vec` as is (was initialized with
|
|
// padding value).
|
|
auto nextVec = generateInBoundsCheck(
|
|
b, xferOp, iv, dim, TypeRange(xferOp.getVectorType()),
|
|
/*inBoundsCase=*/
|
|
[&](OpBuilder &b, Location loc) {
|
|
Value val = b.create<memref::LoadOp>(loc, xferOp.getBase(), indices);
|
|
return b.create<vector::InsertElementOp>(loc, val, vec, iv);
|
|
},
|
|
/*outOfBoundsCase=*/
|
|
[&](OpBuilder & /*b*/, Location loc) { return vec; });
|
|
b.create<scf::YieldOp>(loc, nextVec);
|
|
}
|
|
|
|
static Value initialLoopState(OpBuilder &b, TransferReadOp xferOp) {
|
|
// Inititalize vector with padding value.
|
|
Location loc = xferOp.getLoc();
|
|
return b.create<vector::SplatOp>(loc, xferOp.getVectorType(),
|
|
xferOp.getPadding());
|
|
}
|
|
};
|
|
|
|
/// Codegen strategy for TransferWriteOp.
|
|
template <>
|
|
struct Strategy1d<TransferWriteOp> {
|
|
static void generateForLoopBody(OpBuilder &b, Location loc,
|
|
TransferWriteOp xferOp, Value iv,
|
|
ValueRange /*loopState*/) {
|
|
SmallVector<Value, 8> indices;
|
|
auto dim = get1dMemrefIndices(b, xferOp, iv, indices);
|
|
|
|
// Nothing to do in case of out-of-bounds access.
|
|
generateInBoundsCheck(
|
|
b, xferOp, iv, dim,
|
|
/*inBoundsCase=*/[&](OpBuilder &b, Location loc) {
|
|
auto val =
|
|
b.create<vector::ExtractElementOp>(loc, xferOp.getVector(), iv);
|
|
b.create<memref::StoreOp>(loc, val, xferOp.getBase(), indices);
|
|
});
|
|
b.create<scf::YieldOp>(loc);
|
|
}
|
|
|
|
static Value initialLoopState(OpBuilder &b, TransferWriteOp xferOp) {
|
|
return Value();
|
|
}
|
|
};
|
|
|
|
/// Lower a 1D vector transfer op to SCF using scalar loads/stores. This is
|
|
/// necessary in cases where a 1D vector transfer op cannot be lowered into
|
|
/// vector load/stores due to non-unit strides or broadcasts:
|
|
///
|
|
/// * Transfer dimension is not the last memref dimension
|
|
/// * Transfer dimension is a broadcast (i.e., scalar load + broadcast)
|
|
/// * Memref has a layout map with non-unit stride on the last dimension
|
|
///
|
|
/// This pattern generates IR as follows:
|
|
///
|
|
/// 1. Generate a for loop iterating over each vector element.
|
|
/// 2. Inside the loop, generate a InsertElementOp or ExtractElementOp,
|
|
/// depending on OpTy.
|
|
///
|
|
/// TODO: In some cases (no masking, etc.), LLVM::MatrixColumnMajorLoadOp
|
|
/// can be generated instead of TransferOp1dConversion. Add such a pattern
|
|
/// to ConvertVectorToLLVM.
|
|
///
|
|
/// E.g.:
|
|
/// ```
|
|
/// vector.transfer_write %vec, %A[%a, %b]
|
|
/// {permutation_map = affine_map<(d0, d1) -> (d0)>, in_bounds = [true]}
|
|
/// : vector<9xf32>, memref<?x?xf32>
|
|
/// ```
|
|
/// Is rewritten to approximately the following pseudo-IR:
|
|
/// ```
|
|
/// for i = 0 to 9 {
|
|
/// %t = vector.extractelement %vec[i] : vector<9xf32>
|
|
/// memref.store %t, %arg0[%a + i, %b] : memref<?x?xf32>
|
|
/// }
|
|
/// ```
|
|
template <typename OpTy>
|
|
struct TransferOp1dConversion : public VectorToSCFPattern<OpTy> {
|
|
using VectorToSCFPattern<OpTy>::VectorToSCFPattern;
|
|
|
|
LogicalResult matchAndRewrite(OpTy xferOp,
|
|
PatternRewriter &rewriter) const override {
|
|
// TODO: support 0-d corner case.
|
|
if (xferOp.getTransferRank() == 0)
|
|
return failure();
|
|
auto map = xferOp.getPermutationMap();
|
|
auto memRefType = dyn_cast<MemRefType>(xferOp.getShapedType());
|
|
|
|
if (!memRefType)
|
|
return failure();
|
|
if (xferOp.getVectorType().getRank() != 1)
|
|
return failure();
|
|
if (map.isMinorIdentity() && memRefType.isLastDimUnitStride())
|
|
return failure(); // Handled by ConvertVectorToLLVM
|
|
|
|
// Loop bounds, step, state...
|
|
Location loc = xferOp.getLoc();
|
|
auto vecType = xferOp.getVectorType();
|
|
auto lb = rewriter.create<arith::ConstantIndexOp>(loc, 0);
|
|
Value ub =
|
|
rewriter.create<arith::ConstantIndexOp>(loc, vecType.getDimSize(0));
|
|
if (vecType.isScalable()) {
|
|
Value vscale =
|
|
rewriter.create<vector::VectorScaleOp>(loc, rewriter.getIndexType());
|
|
ub = rewriter.create<arith::MulIOp>(loc, ub, vscale);
|
|
}
|
|
auto step = rewriter.create<arith::ConstantIndexOp>(loc, 1);
|
|
auto loopState = Strategy1d<OpTy>::initialLoopState(rewriter, xferOp);
|
|
|
|
// Generate for loop.
|
|
rewriter.replaceOpWithNewOp<scf::ForOp>(
|
|
xferOp, lb, ub, step, loopState ? ValueRange(loopState) : ValueRange(),
|
|
[&](OpBuilder &b, Location loc, Value iv, ValueRange loopState) {
|
|
Strategy1d<OpTy>::generateForLoopBody(b, loc, xferOp, iv, loopState);
|
|
});
|
|
|
|
return success();
|
|
}
|
|
};
|
|
|
|
} // namespace lowering_1_d
|
|
} // namespace
|
|
|
|
void mlir::populateVectorToSCFConversionPatterns(
|
|
RewritePatternSet &patterns, const VectorTransferToSCFOptions &options) {
|
|
if (options.unroll) {
|
|
patterns.add<lowering_n_d_unrolled::UnrollTransferReadConversion,
|
|
lowering_n_d_unrolled::UnrollTransferWriteConversion>(
|
|
patterns.getContext(), options);
|
|
} else {
|
|
patterns.add<lowering_n_d::PrepareTransferReadConversion,
|
|
lowering_n_d::PrepareTransferWriteConversion,
|
|
lowering_n_d::TransferOpConversion<TransferReadOp>,
|
|
lowering_n_d::TransferOpConversion<TransferWriteOp>>(
|
|
patterns.getContext(), options);
|
|
}
|
|
if (options.lowerScalable) {
|
|
patterns.add<lowering_n_d::ScalableTransposeTransferWriteConversion>(
|
|
patterns.getContext(), options);
|
|
}
|
|
if (options.targetRank == 1) {
|
|
patterns.add<lowering_1_d::TransferOp1dConversion<TransferReadOp>,
|
|
lowering_1_d::TransferOp1dConversion<TransferWriteOp>>(
|
|
patterns.getContext(), options);
|
|
}
|
|
patterns.add<lowering_n_d::DecomposePrintOpConversion>(patterns.getContext(),
|
|
options);
|
|
}
|
|
|
|
namespace {
|
|
|
|
struct ConvertVectorToSCFPass
|
|
: public impl::ConvertVectorToSCFBase<ConvertVectorToSCFPass> {
|
|
ConvertVectorToSCFPass() = default;
|
|
ConvertVectorToSCFPass(const VectorTransferToSCFOptions &options) {
|
|
this->fullUnroll = options.unroll;
|
|
this->targetRank = options.targetRank;
|
|
this->lowerTensors = options.lowerTensors;
|
|
this->lowerScalable = options.lowerScalable;
|
|
}
|
|
|
|
void runOnOperation() override {
|
|
VectorTransferToSCFOptions options;
|
|
options.unroll = fullUnroll;
|
|
options.targetRank = targetRank;
|
|
options.lowerTensors = lowerTensors;
|
|
options.lowerScalable = lowerScalable;
|
|
|
|
// Lower permutation maps first.
|
|
RewritePatternSet lowerTransferPatterns(&getContext());
|
|
mlir::vector::populateVectorTransferPermutationMapLoweringPatterns(
|
|
lowerTransferPatterns);
|
|
(void)applyPatternsGreedily(getOperation(),
|
|
std::move(lowerTransferPatterns));
|
|
|
|
RewritePatternSet patterns(&getContext());
|
|
populateVectorToSCFConversionPatterns(patterns, options);
|
|
(void)applyPatternsGreedily(getOperation(), std::move(patterns));
|
|
}
|
|
};
|
|
|
|
} // namespace
|
|
|
|
std::unique_ptr<Pass>
|
|
mlir::createConvertVectorToSCFPass(const VectorTransferToSCFOptions &options) {
|
|
return std::make_unique<ConvertVectorToSCFPass>(options);
|
|
}
|