Google Git
Sign in
llvm / llvm-project / mlir / ba4e03e4144915ddf10f0b5f7fec65b1674bd6b8 / . / lib / Dialect / Vector / Transforms / VectorDistribute.cpp
blob: c84eb2c9f8857d514cf8c258b008c044d67e79e4 [file] [log] [blame]
//===- VectorDistribute.cpp - patterns to do vector distribution ----------===//
//
// 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
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/GPU/IR/GPUDialect.h"
#include "mlir/Dialect/GPU/Utils/DistributionUtils.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/Dialect/Vector/Transforms/VectorDistribution.h"
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/Attributes.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"
#include "mlir/Transforms/RegionUtils.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallVectorExtras.h"
#include "llvm/Support/FormatVariadic.h"
#include <utility>
using namespace mlir;
using namespace mlir::vector;
using namespace mlir::gpu;
/// Currently the distribution map is implicit based on the vector shape. In the
/// future it will be part of the op.
/// Example:
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<1x16x2xf32>) {
/// ...
/// gpu.yield %3 : vector<32x16x64xf32>
/// }
/// ```
/// Would have an implicit map of:
/// `(d0, d1, d2) -> (d0, d2)`
static AffineMap calculateImplicitMap(VectorType sequentialType,
VectorType distributedType) {
SmallVector<AffineExpr> perm;
perm.reserve(1);
// Check which dimensions of the sequential type are different than the
// dimensions of the distributed type to know the distributed dimensions. Then
// associate each distributed dimension to an ID in order.
for (unsigned i = 0, e = sequentialType.getRank(); i < e; i++) {
if (sequentialType.getDimSize(i) != distributedType.getDimSize(i))
perm.push_back(getAffineDimExpr(i, distributedType.getContext()));
}
auto map = AffineMap::get(sequentialType.getRank(), 0, perm,
distributedType.getContext());
return map;
}
/// Given a sequential and distributed vector type, returns the distributed
/// dimension. This function expects that only a single dimension is
/// distributed.
static int getDistributedDim(VectorType sequentialType,
VectorType distributedType) {
assert(sequentialType.getRank() == distributedType.getRank() &&
"sequential and distributed vector types must have the same rank");
int64_t distributedDim = -1;
for (int64_t i = 0; i < sequentialType.getRank(); ++i) {
if (distributedType.getDimSize(i) != sequentialType.getDimSize(i)) {
// Keep this assert here in case WarpExecuteOnLane0Op gets extended to
// support distributing multiple dimensions in the future.
assert(distributedDim == -1 && "found multiple distributed dims");
distributedDim = i;
}
}
return distributedDim;
}
namespace {
/// Helper struct to create the load / store operations that permit transit
/// through the parallel / sequential and the sequential / parallel boundaries
/// when performing `rewriteWarpOpToScfFor`.
///
/// The vector distribution dimension is inferred from the vector types.
struct DistributedLoadStoreHelper {
DistributedLoadStoreHelper(Value sequentialVal, Value distributedVal,
Value laneId, Value zero)
: sequentialVal(sequentialVal), distributedVal(distributedVal),
laneId(laneId), zero(zero) {
sequentialVectorType = dyn_cast<VectorType>(sequentialVal.getType());
distributedVectorType = dyn_cast<VectorType>(distributedVal.getType());
if (sequentialVectorType && distributedVectorType)
distributionMap =
calculateImplicitMap(sequentialVectorType, distributedVectorType);
}
Value buildDistributedOffset(RewriterBase &b, Location loc, int64_t index) {
int64_t distributedSize = distributedVectorType.getDimSize(index);
AffineExpr tid = getAffineSymbolExpr(0, b.getContext());
return b.createOrFold<affine::AffineApplyOp>(loc, tid * distributedSize,
ArrayRef<Value>{laneId});
}
/// Create a store during the process of distributing the
/// `vector.warp_execute_on_thread_0` op.
/// Vector distribution assumes the following convention regarding the
/// temporary buffers that are created to transition values. This **must**
/// be properly specified in the `options.warpAllocationFn`:
/// 1. scalars of type T transit through a memref<1xT>.
/// 2. vectors of type V<shapexT> transit through a memref<shapexT>
Operation *buildStore(RewriterBase &b, Location loc, Value val,
Value buffer) {
assert((val == distributedVal || val == sequentialVal) &&
"Must store either the preregistered distributed or the "
"preregistered sequential value.");
// Scalar case can directly use memref.store.
if (!isa<VectorType>(val.getType()))
return memref::StoreOp::create(b, loc, val, buffer, zero);
// Vector case must use vector::TransferWriteOp which will later lower to
// vector.store of memref.store depending on further lowerings.
int64_t rank = sequentialVectorType.getRank();
SmallVector<Value> indices(rank, zero);
if (val == distributedVal) {
for (auto dimExpr : distributionMap.getResults()) {
int64_t index = cast<AffineDimExpr>(dimExpr).getPosition();
indices[index] = buildDistributedOffset(b, loc, index);
}
}
SmallVector<bool> inBounds(indices.size(), true);
return vector::TransferWriteOp::create(
b, loc, val, buffer, indices,
ArrayRef<bool>(inBounds.begin(), inBounds.end()));
}
/// Create a load during the process of distributing the
/// `vector.warp_execute_on_thread_0` op.
/// Vector distribution assumes the following convention regarding the
/// temporary buffers that are created to transition values. This **must**
/// be properly specified in the `options.warpAllocationFn`:
/// 1. scalars of type T transit through a memref<1xT>.
/// 2. vectors of type V<shapexT> transit through a memref<shapexT>
///
/// When broadcastMode is true, the load is not distributed to account for
/// the broadcast semantics of the `gpu.warp_execute_on_lane_0` op.
///
/// Example:
///
/// ```
/// %r = gpu.warp_execute_on_lane_0(...) -> (f32) {
/// gpu.yield %cst : f32
/// }
/// // Both types are f32. The constant %cst is broadcasted to all lanes.
/// ```
/// This behavior described in more detail in the documentation of the op.
Value buildLoad(RewriterBase &b, Location loc, Type type, Value buffer) {
// Scalar case can directly use memref.store.
if (!isa<VectorType>(type))
return memref::LoadOp::create(b, loc, buffer, zero);
// Other cases must be vector atm.
// Vector case must use vector::TransferReadOp which will later lower to
// vector.read of memref.read depending on further lowerings.
assert((type == distributedVectorType || type == sequentialVectorType) &&
"Must store either the preregistered distributed or the "
"preregistered sequential type.");
SmallVector<Value> indices(sequentialVectorType.getRank(), zero);
if (type == distributedVectorType) {
for (auto dimExpr : distributionMap.getResults()) {
int64_t index = cast<AffineDimExpr>(dimExpr).getPosition();
indices[index] = buildDistributedOffset(b, loc, index);
}
}
SmallVector<bool> inBounds(indices.size(), true);
return vector::TransferReadOp::create(
b, loc, cast<VectorType>(type), buffer, indices,
/*padding=*/std::nullopt,
ArrayRef<bool>(inBounds.begin(), inBounds.end()));
}
Value sequentialVal, distributedVal, laneId, zero;
VectorType sequentialVectorType, distributedVectorType;
AffineMap distributionMap;
};
} // namespace
// Clones `op` into a new operation that takes `operands` and returns
// `resultTypes`.
static Operation *cloneOpWithOperandsAndTypes(RewriterBase &rewriter,
Location loc, Operation *op,
ArrayRef<Value> operands,
ArrayRef<Type> resultTypes) {
OperationState res(loc, op->getName().getStringRef(), operands, resultTypes,
op->getAttrs());
return rewriter.create(res);
}
namespace {
/// Rewrite a WarpExecuteOnLane0Op into a predicated scf.if op where the single
/// thread `laneId` executes the entirety of the computation.
///
/// After the transformation:
/// - the IR within the scf.if op can be thought of as executing sequentially
/// (from the point of view of threads along `laneId`).
/// - the IR outside of the scf.if op can be thought of as executing in
/// parallel (from the point of view of threads along `laneId`).
///
/// Values that need to transit through the parallel / sequential and the
/// sequential / parallel boundaries do so via reads and writes to a temporary
/// memory location.
///
/// The transformation proceeds in multiple steps:
/// 1. Create the scf.if op.
/// 2. Insert appropriate (alloc, write)-pairs before the scf.if and reads
/// within the scf.if to transit the values captured from above.
/// 3. Synchronize before the scf.if to ensure all writes inserted in 2. are
/// consistent within the scf.if.
/// 4. Move the body of the WarpExecuteOnLane0Op inside the scf.if.
/// 5. Insert appropriate writes within scf.if and reads after the scf.if to
/// transit the values returned by the op.
/// 6. Synchronize after the scf.if to ensure all writes inserted in 5. are
/// consistent after the scf.if.
/// 7. Perform late cleanups.
///
/// All this assumes the vector distribution occurs along the most minor
/// distributed vector dimension.
struct WarpOpToScfIfPattern : public WarpDistributionPattern {
WarpOpToScfIfPattern(MLIRContext *context,
const WarpExecuteOnLane0LoweringOptions &options,
PatternBenefit benefit = 1)
: WarpDistributionPattern(context, benefit), options(options) {}
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
assert(warpOp.getBodyRegion().hasOneBlock() &&
"expected WarpOp with single block");
Block *warpOpBody = &warpOp.getBodyRegion().front();
Location loc = warpOp.getLoc();
// Passed all checks. Start rewriting.
OpBuilder::InsertionGuard g(rewriter);
rewriter.setInsertionPoint(warpOp);
// Step 1: Create scf.if op.
Value c0 = arith::ConstantIndexOp::create(rewriter, loc, 0);
Value isLane0 = arith::CmpIOp::create(
rewriter, loc, arith::CmpIPredicate::eq, warpOp.getLaneid(), c0);
auto ifOp = scf::IfOp::create(rewriter, loc, isLane0,
/*withElseRegion=*/false);
rewriter.eraseOp(ifOp.thenBlock()->getTerminator());
// Step 2: insert appropriate (alloc, write)-pairs before the scf.if and
// reads within the scf.if to transit the values captured from above.
SmallVector<Value> bbArgReplacements;
for (const auto &it : llvm::enumerate(warpOp.getArgs())) {
Value sequentialVal = warpOpBody->getArgument(it.index());
Value distributedVal = it.value();
DistributedLoadStoreHelper helper(sequentialVal, distributedVal,
warpOp.getLaneid(), c0);
// Create buffer before the ifOp.
rewriter.setInsertionPoint(ifOp);
Value buffer = options.warpAllocationFn(loc, rewriter, warpOp,
sequentialVal.getType());
// Store distributed vector into buffer, before the ifOp.
helper.buildStore(rewriter, loc, distributedVal, buffer);
// Load sequential vector from buffer, inside the ifOp.
rewriter.setInsertionPointToStart(ifOp.thenBlock());
bbArgReplacements.push_back(
helper.buildLoad(rewriter, loc, sequentialVal.getType(), buffer));
}
// Step 3. Insert sync after all the stores and before all the loads.
if (!warpOp.getArgs().empty()) {
rewriter.setInsertionPoint(ifOp);
options.warpSyncronizationFn(loc, rewriter, warpOp);
}
// Step 4. Move body of warpOp to ifOp.
rewriter.mergeBlocks(warpOpBody, ifOp.thenBlock(), bbArgReplacements);
// Step 5. Insert appropriate writes within scf.if and reads after the
// scf.if to transit the values returned by the op.
// TODO: at this point, we can reuse the shared memory from previous
// buffers.
SmallVector<Value> replacements;
auto yieldOp = cast<gpu::YieldOp>(ifOp.thenBlock()->getTerminator());
Location yieldLoc = yieldOp.getLoc();
for (const auto &it : llvm::enumerate(yieldOp.getOperands())) {
Value sequentialVal = it.value();
Value distributedVal = warpOp->getResult(it.index());
DistributedLoadStoreHelper helper(sequentialVal, distributedVal,
warpOp.getLaneid(), c0);
// Create buffer before the ifOp.
rewriter.setInsertionPoint(ifOp);
Value buffer = options.warpAllocationFn(loc, rewriter, warpOp,
sequentialVal.getType());
// Store yielded value into buffer, inside the ifOp, before the
// terminator.
rewriter.setInsertionPoint(yieldOp);
helper.buildStore(rewriter, loc, sequentialVal, buffer);
// Load distributed value from buffer, after the warpOp.
rewriter.setInsertionPointAfter(ifOp);
// Result type and yielded value type are the same. This is a broadcast.
// E.g.:
// %r = gpu.warp_execute_on_lane_0(...) -> (f32) {
// gpu.yield %cst : f32
// }
// Both types are f32. The constant %cst is broadcasted to all lanes.
// This is described in more detail in the documentation of the op.
replacements.push_back(
helper.buildLoad(rewriter, loc, distributedVal.getType(), buffer));
}
// Step 6. Insert sync after all the stores and before all the loads.
if (!yieldOp.getOperands().empty()) {
rewriter.setInsertionPointAfter(ifOp);
options.warpSyncronizationFn(loc, rewriter, warpOp);
}
// Step 7. Delete terminator and add empty scf.yield.
rewriter.eraseOp(yieldOp);
rewriter.setInsertionPointToEnd(ifOp.thenBlock());
scf::YieldOp::create(rewriter, yieldLoc);
// Compute replacements for WarpOp results.
rewriter.replaceOp(warpOp, replacements);
return success();
}
private:
const WarpExecuteOnLane0LoweringOptions &options;
};
/// Return the distributed vector type based on the original type and the
/// distribution map. The map is expected to have a dimension equal to the
/// original type rank and should be a projection where the results are the
/// distributed dimensions. The number of results should be equal to the number
/// of warp sizes which is currently limited to 1.
/// Example: For a vector<16x32x64> distributed with a map(d0, d1, d2) -> (d1)
/// and a warp size of 16 would distribute the second dimension (associated to
/// d1) and return vector<16x2x64>
static VectorType getDistributedType(VectorType originalType, AffineMap map,
int64_t warpSize) {
SmallVector<int64_t> targetShape(originalType.getShape());
for (unsigned i = 0, e = map.getNumResults(); i < e; i++) {
unsigned position = map.getDimPosition(i);
if (targetShape[position] % warpSize != 0) {
if (warpSize % targetShape[position] != 0) {
return VectorType();
}
warpSize /= targetShape[position];
targetShape[position] = 1;
continue;
}
targetShape[position] = targetShape[position] / warpSize;
warpSize = 1;
break;
}
if (warpSize != 1) {
return VectorType();
}
VectorType targetType =
VectorType::get(targetShape, originalType.getElementType());
return targetType;
}
/// Distribute transfer_write ops based on the affine map returned by
/// `distributionMapFn`. Writes of size more than `maxNumElementToExtract`
/// will not be distributed (it should be less than the warp size).
///
/// Example:
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%id){
/// ...
/// vector.transfer_write %v, %A[%c0] : vector<32xf32>, memref<128xf32>
/// gpu.yield
/// }
/// ```
/// To
/// ```
/// %r:3 = gpu.warp_execute_on_lane_0(%id) -> (vector<1xf32>) {
/// ...
/// gpu.yield %v : vector<32xf32>
/// }
/// vector.transfer_write %v, %A[%id] : vector<1xf32>, memref<128xf32>
struct WarpOpTransferWrite : public WarpDistributionPattern {
WarpOpTransferWrite(MLIRContext *ctx, DistributionMapFn fn,
unsigned maxNumElementsToExtract, PatternBenefit b = 1)
: WarpDistributionPattern(ctx, b), distributionMapFn(std::move(fn)),
maxNumElementsToExtract(maxNumElementsToExtract) {}
/// Distribute the TransferWriteOp. Only 1D distributions and vector dims that
/// are multiples of the distribution ratio are supported at the moment.
LogicalResult tryDistributeOp(RewriterBase &rewriter,
vector::TransferWriteOp writeOp,
WarpExecuteOnLane0Op warpOp) const {
VectorType writtenVectorType = writeOp.getVectorType();
// 1. If the write is 0-D, we just clone it into a new WarpExecuteOnLane0Op
// to separate it from the rest.
if (writtenVectorType.getRank() == 0)
return failure();
// 2. Compute the distributed type.
AffineMap map = distributionMapFn(writeOp.getVector());
VectorType targetType =
getDistributedType(writtenVectorType, map, warpOp.getWarpSize());
if (!targetType)
return failure();
// 2.5 Compute the distributed type for the new mask;
VectorType maskType;
if (writeOp.getMask()) {
// TODO: Distribution of masked writes with non-trivial permutation maps
// requires the distribution of the mask to elementwise match the
// distribution of the permuted written vector. Currently the details
// of which lane is responsible for which element is captured strictly
// by shape information on the warp op, and thus requires materializing
// the permutation in IR.
if (!writeOp.getPermutationMap().isMinorIdentity())
return failure();
maskType =
getDistributedType(writeOp.getMaskType(), map, warpOp.getWarpSize());
}
// 3. clone the write into a new WarpExecuteOnLane0Op to separate it from
// the rest.
vector::TransferWriteOp newWriteOp =
cloneWriteOp(rewriter, warpOp, writeOp, targetType, maskType);
// 4. Reindex the write using the distribution map.
auto newWarpOp =
newWriteOp.getVector().getDefiningOp<WarpExecuteOnLane0Op>();
// Delinearize the lane id based on the way threads are divided across the
// vector. To get the number of threads per vector dimension, divide the
// sequential size by the distributed size along each dim.
rewriter.setInsertionPoint(newWriteOp);
SmallVector<OpFoldResult> delinearizedIdSizes;
for (auto [seqSize, distSize] :
llvm::zip_equal(writtenVectorType.getShape(), targetType.getShape())) {
assert(seqSize % distSize == 0 && "Invalid distributed vector shape");
delinearizedIdSizes.push_back(rewriter.getIndexAttr(seqSize / distSize));
}
SmallVector<Value> delinearized;
if (map.getNumResults() > 1) {
delinearized = mlir::affine::AffineDelinearizeIndexOp::create(
rewriter, newWarpOp.getLoc(), newWarpOp.getLaneid(),
delinearizedIdSizes)
.getResults();
} else {
// If there is only one map result, we can elide the delinearization
// op and use the lane id directly.
delinearized.append(targetType.getRank(), newWarpOp.getLaneid());
}
AffineMap indexMap = map.compose(newWriteOp.getPermutationMap());
Location loc = newWriteOp.getLoc();
SmallVector<Value> indices(newWriteOp.getIndices().begin(),
newWriteOp.getIndices().end());
for (auto it : llvm::zip(indexMap.getResults(), map.getResults())) {
AffineExpr d0, d1;
bindDims(newWarpOp.getContext(), d0, d1);
auto indexExpr = dyn_cast<AffineDimExpr>(std::get<0>(it));
if (!indexExpr)
continue;
unsigned indexPos = indexExpr.getPosition();
unsigned vectorPos = cast<AffineDimExpr>(std::get<1>(it)).getPosition();
Value laneId = delinearized[vectorPos];
auto scale =
rewriter.getAffineConstantExpr(targetType.getDimSize(vectorPos));
indices[indexPos] = affine::makeComposedAffineApply(
rewriter, loc, d0 + scale * d1, {indices[indexPos], laneId});
}
newWriteOp.getIndicesMutable().assign(indices);
return success();
}
/// Extract TransferWriteOps of vector<1x> into a separate warp op.
LogicalResult tryExtractOp(RewriterBase &rewriter,
vector::TransferWriteOp writeOp,
WarpExecuteOnLane0Op warpOp) const {
Location loc = writeOp.getLoc();
VectorType vecType = writeOp.getVectorType();
if (vecType.getNumElements() > maxNumElementsToExtract) {
return rewriter.notifyMatchFailure(
warpOp,
llvm::formatv(
"writes more elements ({0}) than allowed to extract ({1})",
vecType.getNumElements(), maxNumElementsToExtract));
}
// Do not process warp ops that contain only TransferWriteOps.
if (llvm::all_of(warpOp.getOps(),
llvm::IsaPred<vector::TransferWriteOp, gpu::YieldOp>))
return failure();
SmallVector<Value> yieldValues = {writeOp.getVector()};
SmallVector<Type> retTypes = {vecType};
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, yieldValues, retTypes, newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
// Create a second warp op that contains only writeOp.
auto secondWarpOp = WarpExecuteOnLane0Op::create(rewriter, loc, TypeRange(),
newWarpOp.getLaneid(),
newWarpOp.getWarpSize());
Block &body = secondWarpOp.getBodyRegion().front();
rewriter.setInsertionPointToStart(&body);
auto newWriteOp =
cast<vector::TransferWriteOp>(rewriter.clone(*writeOp.getOperation()));
newWriteOp.getValueToStoreMutable().assign(
newWarpOp.getResult(newRetIndices[0]));
rewriter.eraseOp(writeOp);
gpu::YieldOp::create(rewriter, newWarpOp.getLoc());
return success();
}
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
gpu::YieldOp yield = warpOp.getTerminator();
Operation *lastNode = yield->getPrevNode();
auto writeOp = dyn_cast_or_null<vector::TransferWriteOp>(lastNode);
if (!writeOp)
return failure();
Value maybeMask = writeOp.getMask();
if (!llvm::all_of(writeOp->getOperands(), [&](Value value) {
return writeOp.getVector() == value ||
(maybeMask && maybeMask == value) ||
warpOp.isDefinedOutsideOfRegion(value);
}))
return failure();
if (succeeded(tryDistributeOp(rewriter, writeOp, warpOp)))
return success();
// Masked writes not supported for extraction.
if (writeOp.getMask())
return failure();
if (succeeded(tryExtractOp(rewriter, writeOp, warpOp)))
return success();
return failure();
}
private:
/// Clone `writeOp` assumed to be nested under `warpOp` into a new warp
/// execute op with the proper return type. The new write op is updated to
/// write the result of the new warp execute op. The old `writeOp` is deleted.
vector::TransferWriteOp cloneWriteOp(RewriterBase &rewriter,
WarpExecuteOnLane0Op warpOp,
vector::TransferWriteOp writeOp,
VectorType targetType,
VectorType maybeMaskType) const {
assert(writeOp->getParentOp() == warpOp &&
"write must be nested immediately under warp");
OpBuilder::InsertionGuard g(rewriter);
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp;
if (maybeMaskType) {
newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, ValueRange{writeOp.getVector(), writeOp.getMask()},
TypeRange{targetType, maybeMaskType}, newRetIndices);
} else {
newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, ValueRange{{writeOp.getVector()}},
TypeRange{targetType}, newRetIndices);
}
rewriter.setInsertionPointAfter(newWarpOp);
auto newWriteOp =
cast<vector::TransferWriteOp>(rewriter.clone(*writeOp.getOperation()));
rewriter.eraseOp(writeOp);
newWriteOp.getValueToStoreMutable().assign(
newWarpOp.getResult(newRetIndices[0]));
if (maybeMaskType)
newWriteOp.getMaskMutable().assign(newWarpOp.getResult(newRetIndices[1]));
return newWriteOp;
}
DistributionMapFn distributionMapFn;
unsigned maxNumElementsToExtract = 1;
};
/// Sink out elementwise op feeding into a warp op yield.
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<1xf32>) {
/// ...
/// %3 = arith.addf %1, %2 : vector<32xf32>
/// gpu.yield %3 : vector<32xf32>
/// }
/// ```
/// To
/// ```
/// %r:3 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<1xf32>,
/// vector<1xf32>, vector<1xf32>) {
/// ...
/// %4 = arith.addf %2, %3 : vector<32xf32>
/// gpu.yield %4, %2, %3 : vector<32xf32>, vector<32xf32>,
/// vector<32xf32>
/// }
/// %0 = arith.addf %r#1, %r#2 : vector<1xf32>
struct WarpOpElementwise : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *yieldOperand = getWarpResult(warpOp, [](Operation *op) {
return OpTrait::hasElementwiseMappableTraits(op);
});
if (!yieldOperand)
return failure();
Operation *elementWise = yieldOperand->get().getDefiningOp();
unsigned operandIndex = yieldOperand->getOperandNumber();
Value distributedVal = warpOp.getResult(operandIndex);
SmallVector<Value> yieldValues;
SmallVector<Type> retTypes;
Location loc = warpOp.getLoc();
for (OpOperand &operand : elementWise->getOpOperands()) {
Type targetType;
if (auto vecType = dyn_cast<VectorType>(distributedVal.getType())) {
// If the result type is a vector, the operands must also be vectors.
auto operandType = cast<VectorType>(operand.get().getType());
targetType =
VectorType::get(vecType.getShape(), operandType.getElementType());
} else {
auto operandType = operand.get().getType();
assert(!isa<VectorType>(operandType) &&
"unexpected yield of vector from op with scalar result type");
targetType = operandType;
}
retTypes.push_back(targetType);
yieldValues.push_back(operand.get());
}
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, yieldValues, retTypes, newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
SmallVector<Value> newOperands(elementWise->getOperands().begin(),
elementWise->getOperands().end());
for (unsigned i : llvm::seq(unsigned(0), elementWise->getNumOperands())) {
newOperands[i] = newWarpOp.getResult(newRetIndices[i]);
}
OpBuilder::InsertionGuard g(rewriter);
rewriter.setInsertionPointAfter(newWarpOp);
Operation *newOp = cloneOpWithOperandsAndTypes(
rewriter, loc, elementWise, newOperands,
{newWarpOp.getResult(operandIndex).getType()});
rewriter.replaceAllUsesWith(newWarpOp.getResult(operandIndex),
newOp->getResult(0));
return success();
}
};
/// Sink out splat constant op feeding into a warp op yield.
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<1xf32>) {
/// ...
/// %cst = arith.constant dense<2.0> : vector<32xf32>
/// gpu.yield %cst : vector<32xf32>
/// }
/// ```
/// To
/// ```
/// gpu.warp_execute_on_lane_0(%arg0 {
/// ...
/// }
/// %0 = arith.constant dense<2.0> : vector<1xf32>
struct WarpOpConstant : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *yieldOperand =
getWarpResult(warpOp, llvm::IsaPred<arith::ConstantOp>);
if (!yieldOperand)
return failure();
auto constantOp = yieldOperand->get().getDefiningOp<arith::ConstantOp>();
auto dense = dyn_cast<SplatElementsAttr>(constantOp.getValue());
if (!dense)
return failure();
// Notify the rewriter that the warp op is changing (see the comment on
// the WarpOpTransferRead pattern).
rewriter.startOpModification(warpOp);
unsigned operandIndex = yieldOperand->getOperandNumber();
Attribute scalarAttr = dense.getSplatValue<Attribute>();
auto newAttr = DenseElementsAttr::get(
cast<ShapedType>(warpOp.getResult(operandIndex).getType()), scalarAttr);
Location loc = warpOp.getLoc();
rewriter.setInsertionPointAfter(warpOp);
Value distConstant = arith::ConstantOp::create(rewriter, loc, newAttr);
rewriter.replaceAllUsesWith(warpOp.getResult(operandIndex), distConstant);
rewriter.finalizeOpModification(warpOp);
return success();
}
};
/// Sink out step op feeding into a warp op yield.
/// Vector step op is treated similar to arith.constant, apart from
/// the result that represents a sequence [0, vec_size).
/// Due to the to vec_size == warp_size limitation,
/// we can simply wrap the lane id into a vector (i.e., broadcast).
/// Supporting vec_size != warp_size may involve preserving the step
/// result and using additional arith ops (the exact details are TBD).
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<1xindex>) {
/// ...
/// %cst = vector.step : vector<32xindex>
/// gpu.yield %cst : vector<1xindex>
/// }
/// ```
/// To
/// ```
/// gpu.warp_execute_on_lane_0(%arg0) {
/// ...
/// }
/// %lane_id_vec = vector.broadcast %arg0 : index to vector<1xindex>
struct WarpOpStep final : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *yieldOperand =
getWarpResult(warpOp, llvm::IsaPred<vector::StepOp>);
if (!yieldOperand)
return failure();
const unsigned operandIdx = yieldOperand->getOperandNumber();
auto stepOp = yieldOperand->get().getDefiningOp<vector::StepOp>();
VectorType resTy = stepOp.getResult().getType();
if (resTy.getNumElements() != static_cast<int64_t>(warpOp.getWarpSize()))
return rewriter.notifyMatchFailure(
warpOp,
llvm::formatv("Expected result size ({0}) to be of warp size ({1})",
resTy.getNumElements(), warpOp.getWarpSize()));
VectorType newVecTy =
cast<VectorType>(warpOp.getResult(operandIdx).getType());
rewriter.setInsertionPointAfter(warpOp);
Value laneIdVec = vector::BroadcastOp::create(rewriter, warpOp.getLoc(),
newVecTy, warpOp.getLaneid());
rewriter.replaceAllUsesWith(warpOp.getResult(operandIdx), laneIdVec);
return success();
}
};
/// Sink out transfer_read op feeding into a warp op yield.
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<1xf32>) {
/// ...
// %2 = vector.transfer_read %src[%c0], %cst : memref<1024xf32>,
// vector<32xf32>
/// gpu.yield %2 : vector<32xf32>
/// }
/// ```
/// To
/// ```
/// %dead = gpu.warp_execute_on_lane_0(%arg0) -> (vector<1xf32>,
/// vector<1xf32>, vector<1xf32>) {
/// ...
/// %2 = vector.transfer_read %src[%c0], %cst : memref<1024xf32>,
/// vector<32xf32> gpu.yield %2 : vector<32xf32>
/// }
/// %0 = vector.transfer_read %src[%c0], %cst : memref<1024xf32>, vector<1xf32>
struct WarpOpTransferRead : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
// Try to find a distributable yielded read. Note that this pattern can
// still fail at the end after distribution, in which case this might have
// missed another distributable read.
OpOperand *operand = getWarpResult(warpOp, [](Operation *op) {
// Don't duplicate transfer_read ops when distributing.
return isa<vector::TransferReadOp>(op) && op->hasOneUse();
});
if (!operand)
return rewriter.notifyMatchFailure(
warpOp, "warp result is not a vector.transfer_read op");
auto read = operand->get().getDefiningOp<vector::TransferReadOp>();
// Source must be defined outside of the region.
if (!warpOp.isDefinedOutsideOfRegion(read.getBase()))
return rewriter.notifyMatchFailure(
read, "source must be defined outside of the region");
unsigned operandIndex = operand->getOperandNumber();
Value distributedVal = warpOp.getResult(operandIndex);
SmallVector<Value, 4> indices(read.getIndices().begin(),
read.getIndices().end());
auto sequentialType = cast<VectorType>(read.getResult().getType());
auto distributedType = cast<VectorType>(distributedVal.getType());
AffineMap map = calculateImplicitMap(sequentialType, distributedType);
AffineMap indexMap = map.compose(read.getPermutationMap());
// Try to delinearize the lane ID to match the rank expected for
// distribution.
SmallVector<Value> delinearizedIds;
if (!delinearizeLaneId(rewriter, read.getLoc(), sequentialType.getShape(),
distributedType.getShape(), warpOp.getWarpSize(),
warpOp.getLaneid(), delinearizedIds)) {
return rewriter.notifyMatchFailure(
read, "cannot delinearize lane ID for distribution");
}
assert(!delinearizedIds.empty() || map.getNumResults() == 0);
// Distribute indices and the mask (if present).
OpBuilder::InsertionGuard g(rewriter);
SmallVector<Value> additionalResults(indices.begin(), indices.end());
SmallVector<Type> additionalResultTypes(indices.size(),
rewriter.getIndexType());
additionalResults.push_back(read.getPadding());
additionalResultTypes.push_back(read.getPadding().getType());
bool hasMask = false;
if (read.getMask()) {
hasMask = true;
// TODO: Distribution of masked reads with non-trivial permutation maps
// requires the distribution of the mask to elementwise match the
// distribution of the permuted written vector. Currently the details
// of which lane is responsible for which element is captured strictly
// by shape information on the warp op, and thus requires materializing
// the permutation in IR.
if (!mlir::compressUnusedDims(read.getPermutationMap()).isIdentity())
return rewriter.notifyMatchFailure(
read, "non-trivial permutation maps not supported");
VectorType maskType =
getDistributedType(read.getMaskType(), map, warpOp.getWarpSize());
additionalResults.push_back(read.getMask());
additionalResultTypes.push_back(maskType);
}
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, additionalResults, additionalResultTypes,
newRetIndices);
distributedVal = newWarpOp.getResult(operandIndex);
// Distributed indices were appended first.
SmallVector<Value> newIndices;
for (int64_t i = 0, e = indices.size(); i < e; ++i)
newIndices.push_back(newWarpOp.getResult(newRetIndices[i]));
rewriter.setInsertionPointAfter(newWarpOp);
for (auto it : llvm::zip_equal(indexMap.getResults(), map.getResults())) {
AffineExpr d0, d1;
bindDims(read.getContext(), d0, d1);
auto indexExpr = dyn_cast<AffineDimExpr>(std::get<0>(it));
if (!indexExpr)
continue;
unsigned indexPos = indexExpr.getPosition();
unsigned vectorPos = cast<AffineDimExpr>(std::get<1>(it)).getPosition();
int64_t scale = distributedType.getDimSize(vectorPos);
newIndices[indexPos] = affine::makeComposedAffineApply(
rewriter, read.getLoc(), d0 + scale * d1,
{newIndices[indexPos], delinearizedIds[vectorPos]});
}
// Distributed padding value was appended right after the indices.
Value newPadding = newWarpOp.getResult(newRetIndices[indices.size()]);
// Distributed mask value was added at the end (if the op has a mask).
Value newMask =
hasMask ? newWarpOp.getResult(newRetIndices[newRetIndices.size() - 1])
: Value();
auto newRead = vector::TransferReadOp::create(
rewriter, read.getLoc(), distributedVal.getType(), read.getBase(),
newIndices, read.getPermutationMapAttr(), newPadding, newMask,
read.getInBoundsAttr());
rewriter.replaceAllUsesWith(distributedVal, newRead);
return success();
}
};
/// Remove any result that has no use along with the matching yieldOp operand.
// TODO: Move this in WarpExecuteOnLane0Op canonicalization.
struct WarpOpDeadResult : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
SmallVector<Type> newResultTypes;
newResultTypes.reserve(warpOp->getNumResults());
SmallVector<Value> newYieldValues;
newYieldValues.reserve(warpOp->getNumResults());
DenseMap<Value, int64_t> dedupYieldOperandPositionMap;
DenseMap<OpResult, int64_t> dedupResultPositionMap;
gpu::YieldOp yield = warpOp.getTerminator();
// Some values may be yielded multiple times and correspond to multiple
// results. Deduplicating occurs by taking each result with its matching
// yielded value, and:
// 1. recording the unique first position at which the value is yielded.
// 2. recording for the result, the first position at which the dedup'ed
// value is yielded.
// 3. skipping from the new result types / new yielded values any result
// that has no use or whose yielded value has already been seen.
for (OpResult result : warpOp.getResults()) {
Value yieldOperand = yield.getOperand(result.getResultNumber());
auto it = dedupYieldOperandPositionMap.insert(
std::make_pair(yieldOperand, newResultTypes.size()));
dedupResultPositionMap.insert(std::make_pair(result, it.first->second));
if (result.use_empty() || !it.second)
continue;
newResultTypes.push_back(result.getType());
newYieldValues.push_back(yieldOperand);
}
// No modification, exit early.
if (yield.getNumOperands() == newYieldValues.size())
return failure();
// Move the body of the old warpOp to a new warpOp.
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndReplaceReturns(
rewriter, warpOp, newYieldValues, newResultTypes);
// Simplify the new warp op after dropping dead results.
newWarpOp.getBody()->walk([&](Operation *op) {
if (isOpTriviallyDead(op))
rewriter.eraseOp(op);
});
// Replace results of the old warpOp by the new, deduplicated results.
SmallVector<Value> newValues;
newValues.reserve(warpOp->getNumResults());
for (OpResult result : warpOp.getResults()) {
if (result.use_empty())
newValues.push_back(Value());
else
newValues.push_back(
newWarpOp.getResult(dedupResultPositionMap.lookup(result)));
}
rewriter.replaceOp(warpOp, newValues);
return success();
}
};
// If an operand is directly yielded out of the region we can forward it
// directly and it doesn't need to go through the region.
struct WarpOpForwardOperand : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
gpu::YieldOp yield = warpOp.getTerminator();
Value valForwarded;
unsigned resultIndex;
for (OpOperand &operand : yield->getOpOperands()) {
Value result = warpOp.getResult(operand.getOperandNumber());
if (result.use_empty())
continue;
// Assume all the values coming from above are uniform.
if (!warpOp.getBodyRegion().isAncestor(operand.get().getParentRegion())) {
if (result.getType() != operand.get().getType())
continue;
valForwarded = operand.get();
resultIndex = operand.getOperandNumber();
break;
}
auto arg = dyn_cast<BlockArgument>(operand.get());
if (!arg || arg.getOwner()->getParentOp() != warpOp.getOperation())
continue;
Value warpOperand = warpOp.getArgs()[arg.getArgNumber()];
if (result.getType() != warpOperand.getType())
continue;
valForwarded = warpOperand;
resultIndex = operand.getOperandNumber();
break;
}
if (!valForwarded)
return failure();
// Notify the rewriter that the warp op is changing (see the comment on
// the WarpOpTransferRead pattern).
rewriter.startOpModification(warpOp);
rewriter.replaceAllUsesWith(warpOp.getResult(resultIndex), valForwarded);
rewriter.finalizeOpModification(warpOp);
return success();
}
};
struct WarpOpBroadcast : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *operand =
getWarpResult(warpOp, llvm::IsaPred<vector::BroadcastOp>);
if (!operand)
return failure();
unsigned int operandNumber = operand->getOperandNumber();
auto broadcastOp = operand->get().getDefiningOp<vector::BroadcastOp>();
Location loc = broadcastOp.getLoc();
auto destVecType =
cast<VectorType>(warpOp->getResultTypes()[operandNumber]);
Value broadcastSrc = broadcastOp.getSource();
Type broadcastSrcType = broadcastSrc.getType();
// Check that the broadcast actually spans a set of values uniformly across
// all threads. In other words, check that each thread can reconstruct
// their own broadcast.
// For that we simply check that the broadcast we want to build makes sense.
if (vector::isBroadcastableTo(broadcastSrcType, destVecType) !=
vector::BroadcastableToResult::Success)
return failure();
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {broadcastSrc}, {broadcastSrcType}, newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value broadcasted = vector::BroadcastOp::create(
rewriter, loc, destVecType, newWarpOp->getResult(newRetIndices[0]));
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber),
broadcasted);
return success();
}
};
/// Pattern to move shape cast out of the warp op. shape cast is basically a
/// no-op for warp distribution; we need to handle the shape though.
struct WarpOpShapeCast : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *operand =
getWarpResult(warpOp, llvm::IsaPred<vector::ShapeCastOp>);
if (!operand)
return failure();
auto oldCastOp = operand->get().getDefiningOp<vector::ShapeCastOp>();
unsigned int operandNumber = operand->getOperandNumber();
auto castDistributedType =
cast<VectorType>(warpOp->getResultTypes()[operandNumber]);
VectorType castOriginalType = oldCastOp.getSourceVectorType();
VectorType castResultType = castDistributedType;
// We expect the distributed type to have a smaller rank than the original
// type. Prepend with size-one dimensions to make them the same.
unsigned castDistributedRank = castDistributedType.getRank();
unsigned castOriginalRank = castOriginalType.getRank();
if (castDistributedRank < castOriginalRank) {
SmallVector<int64_t> shape(castOriginalRank - castDistributedRank, 1);
llvm::append_range(shape, castDistributedType.getShape());
castDistributedType =
VectorType::get(shape, castDistributedType.getElementType());
}
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {oldCastOp.getSource()}, {castDistributedType},
newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value newCast = vector::ShapeCastOp::create(
rewriter, oldCastOp.getLoc(), castResultType,
newWarpOp->getResult(newRetIndices[0]));
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber), newCast);
return success();
}
};
/// Sink out vector.create_mask op feeding into a warp op yield.
/// ```
/// %0 = ...
/// %1 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<1xf32>) {
/// ...
/// %mask = vector.create_mask %0 : vector<32xi1>
/// gpu.yield %mask : vector<32xi1>
/// }
/// ```
/// To
/// ```
/// %0 = ...
/// gpu.warp_execute_on_lane_0(%arg0) {
/// ...
/// }
/// %cmp = arith.cmpi ult, %laneid, %0
/// %ub = arith.select %cmp, %c0, %c1
/// %1 = vector.create_mask %ub : vector<1xi1>
struct WarpOpCreateMask : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *yieldOperand =
getWarpResult(warpOp, llvm::IsaPred<vector::CreateMaskOp>);
if (!yieldOperand)
return failure();
auto mask = yieldOperand->get().getDefiningOp<vector::CreateMaskOp>();
// Early exit if any values needed for calculating the new mask indices
// are defined inside the warp op.
if (!llvm::all_of(mask->getOperands(), [&](Value value) {
return warpOp.isDefinedOutsideOfRegion(value);
}))
return failure();
Location loc = mask.getLoc();
unsigned operandIndex = yieldOperand->getOperandNumber();
auto distType = cast<VectorType>(warpOp.getResult(operandIndex).getType());
VectorType seqType = mask.getVectorType();
ArrayRef<int64_t> seqShape = seqType.getShape();
ArrayRef<int64_t> distShape = distType.getShape();
rewriter.setInsertionPointAfter(warpOp);
// Delinearize the lane ID for constructing the distributed mask sizes.
SmallVector<Value> delinearizedIds;
if (!delinearizeLaneId(rewriter, loc, seqShape, distShape,
warpOp.getWarpSize(), warpOp.getLaneid(),
delinearizedIds))
return rewriter.notifyMatchFailure(
mask, "cannot delinearize lane ID for distribution");
assert(!delinearizedIds.empty());
// Notify the rewriter that the warp op is changing (see the comment on
// the WarpOpTransferRead pattern).
rewriter.startOpModification(warpOp);
AffineExpr s0, s1;
bindSymbols(rewriter.getContext(), s0, s1);
SmallVector<Value> newOperands;
for (int i = 0, e = distShape.size(); i < e; ++i) {
// Get `mask_dim_range_upper_limit[i] - lane_id[i] * dist_sizes[i]` to
// find the distance from the largest mask index owned by this lane to the
// original mask size. `vector.create_mask` implicitly clamps mask
// operands to the range [0, mask_vector_size[i]], or in other words, the
// mask sizes are always in the range [0, mask_vector_size[i]).
Value maskDimIdx = affine::makeComposedAffineApply(
rewriter, loc, s1 - s0 * distShape[i],
{delinearizedIds[i], mask.getOperand(i)});
newOperands.push_back(maskDimIdx);
}
auto newMask =
vector::CreateMaskOp::create(rewriter, loc, distType, newOperands);
rewriter.replaceAllUsesWith(warpOp.getResult(operandIndex), newMask);
rewriter.finalizeOpModification(warpOp);
return success();
}
};
/// Sink out insert_strided_slice op feeding into a warp op yield.
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<8x1xf32>) {
/// ...
/// %src = ... : vector<4x32xf32>
/// %dest = ... : vector<8x32xf32>
/// %insert = vector.insert_strided_slice %src, %dest, offsets = [0, 0],
/// strides = [1, 1] : vector<4x32xf32> into vector<8x32xf32>
/// gpu.yield %insert : vector<8x32xf32>
/// }
/// ```
/// To
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<4x1xf32>,
/// vector<8x1xf32>) {
/// ...
/// %src = ... : vector<4x32xf32>
/// %dest = ... : vector<8x32xf32>
/// gpu.yield %src, %dest : vector<4x16xf32>, vector<8x16xf32>
/// }
/// %insert = vector.insert_strided_slice %0#0, %0#1,
/// offsets = [0, 0], strides = [1, 1] : vector<4x1xf32> into vector<8x1xf32>
/// ```
/// NOTE: Current support assumes that both src and dest vectors are distributed
/// to lanes and sinking the insert op does not require any cross lane
/// communication.
struct WarpOpInsertStridedSlice : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *operand =
getWarpResult(warpOp, llvm::IsaPred<vector::InsertStridedSliceOp>);
if (!operand)
return failure();
unsigned int operandNumber = operand->getOperandNumber();
auto insertOp =
operand->get().getDefiningOp<vector::InsertStridedSliceOp>();
auto distributedType =
cast<VectorType>(warpOp.getResult(operandNumber).getType());
// Distributed type must be 2D or higher.
// TODO: Support 1D distributed types.
if (distributedType.getRank() < 2)
return rewriter.notifyMatchFailure(
insertOp, "result vector type must be 2D or higher");
// Find the distributed dimension of the dest vector. There should be
// exactly one.
auto yieldedType = cast<VectorType>(operand->get().getType());
int64_t destDistributedDim =
getDistributedDim(yieldedType, distributedType);
assert(destDistributedDim != -1 && "could not find distributed dimension");
VectorType srcType = insertOp.getSourceVectorType();
VectorType destType = insertOp.getDestVectorType();
// Currently we require that both source (kD) and dest (nD) vectors are
// distributed. This requires that distributedDim (d) is contained in the
// last k dims of the dest vector (d >= n - k).
// TODO: Add support for case where source vector is not distributed.
int64_t sourceDistributedDim =
destDistributedDim - (destType.getRank() - srcType.getRank());
if (sourceDistributedDim < 0)
return rewriter.notifyMatchFailure(
insertOp,
"distributed dimension must be in the last k dims of dest vector");
// Distributed dimension must be fully inserted.
if (srcType.getDimSize(sourceDistributedDim) !=
destType.getDimSize(destDistributedDim))
return rewriter.notifyMatchFailure(
insertOp, "distributed dimension must be fully inserted");
SmallVector<int64_t> newSourceDistShape(
insertOp.getSourceVectorType().getShape());
newSourceDistShape[sourceDistributedDim] =
distributedType.getDimSize(destDistributedDim);
auto newSourceTy =
VectorType::get(newSourceDistShape, distributedType.getElementType());
VectorType newDestTy = distributedType;
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {insertOp.getValueToStore(), insertOp.getDest()},
{newSourceTy, newDestTy}, newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value distributedSource = newWarpOp->getResult(newRetIndices[0]);
Value distributedDest = newWarpOp->getResult(newRetIndices[1]);
// Create a new insert strided slice op that inserts distributed source into
// distributed dest.
Value newInsert = vector::InsertStridedSliceOp::create(
rewriter, insertOp.getLoc(), distributedDest.getType(),
distributedSource, distributedDest, insertOp.getOffsets(),
insertOp.getStrides());
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber), newInsert);
return success();
}
};
/// Sink out extract_strided_slice op feeding into a warp op yield.
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<16x1xf32>) {
/// ...
/// %src = ... : vector<64x32xf32>
/// %extract = vector.extract_strided_slice %src, offsets = [0], sizes = [16],
/// strides = [1] : vector<64x32xf32> to vector<16x32xf32>
/// gpu.yield %extract : vector<16x32xf32>
/// }
/// ```
/// To
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<64x1xf32>) {
/// ...
/// %src = ... : vector<64x32xf32>
/// gpu.yield %src : vector<64x32xf32>
/// }
/// %extract = vector.extract_strided_slice %0, offsets = [0], sizes = [16],
/// strides = [1] : vector<64x1xf32> to vector<16x1xf32>
/// ```
/// NOTE: Current support assumes that the extraction happens only on non
/// distributed dimensions (does not require cross lane communication).
struct WarpOpExtractStridedSlice : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *operand =
getWarpResult(warpOp, llvm::IsaPred<vector::ExtractStridedSliceOp>);
if (!operand)
return failure();
unsigned int operandNumber = operand->getOperandNumber();
auto extractOp =
operand->get().getDefiningOp<vector::ExtractStridedSliceOp>();
auto distributedType =
cast<VectorType>(warpOp.getResult(operandNumber).getType());
// Distributed type must be 2D or higher.
// TODO: Support 1D distributed types.
if (distributedType.getRank() < 2)
return rewriter.notifyMatchFailure(
extractOp, "result vector type must be 2D or higher");
// Find the distributed dimension. There should be exactly one.
auto yieldedType = cast<VectorType>(operand->get().getType());
int64_t distributedDim = getDistributedDim(yieldedType, distributedType);
assert(distributedDim != -1 && "could not find distributed dimension");
int64_t numOfExtractedDims =
static_cast<int64_t>(extractOp.getSizes().size());
// If the distributed dim is included in the extracted dims, then we make
// sure distributed dim is fully extracted. If distributed dim is not
// included in extracted dims, it is guaranteed to be fully extracted (i.e.
// distributed dim comes after all the extracted dims)
// TODO: Partial extraction from distributed dimension require cross lane
// communication.
if (distributedDim < numOfExtractedDims) {
int64_t distributedDimOffset =
llvm::cast<IntegerAttr>(extractOp.getOffsets()[distributedDim])
.getInt();
int64_t distributedDimSize =
llvm::cast<IntegerAttr>(extractOp.getSizes()[distributedDim])
.getInt();
if (distributedDimOffset != 0 ||
distributedDimSize != yieldedType.getDimSize(distributedDim))
return rewriter.notifyMatchFailure(
extractOp, "distributed dimension must be fully extracted");
}
SmallVector<int64_t> newDistributedShape(
extractOp.getSourceVectorType().getShape());
newDistributedShape[distributedDim] =
distributedType.getDimSize(distributedDim);
auto newDistributedType =
VectorType::get(newDistributedShape, distributedType.getElementType());
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {extractOp.getVector()}, {newDistributedType},
newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
SmallVector<Attribute> distributedSizes = llvm::map_to_vector(
extractOp.getSizes(), [](Attribute attr) { return attr; });
// Update the distributed sizes to match the distributed type.
if (distributedDim < static_cast<int64_t>(distributedSizes.size()))
distributedSizes[distributedDim] = rewriter.getI64IntegerAttr(
distributedType.getDimSize(distributedDim));
// Create a new extract strided slice op that extracts from the
// distributed vector.
Value distributedVec = newWarpOp->getResult(newRetIndices[0]);
Value newExtract = vector::ExtractStridedSliceOp::create(
rewriter, extractOp.getLoc(), distributedType, distributedVec,
extractOp.getOffsets(),
ArrayAttr::get(rewriter.getContext(), distributedSizes),
extractOp.getStrides());
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber),
newExtract);
return success();
}
};
/// Pattern to move out vector.extract of single element vector. Those don't
/// need to be distributed and can just be propagated outside of the region.
struct WarpOpExtract : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *operand =
getWarpResult(warpOp, llvm::IsaPred<vector::ExtractOp>);
if (!operand)
return failure();
unsigned int operandNumber = operand->getOperandNumber();
auto extractOp = operand->get().getDefiningOp<vector::ExtractOp>();
VectorType extractSrcType = extractOp.getSourceVectorType();
Location loc = extractOp.getLoc();
// For 1-d or 0-d source cases, we rely on WarpOpExtractScalar pattern.
if (extractSrcType.getRank() <= 1) {
return failure();
}
// All following cases are 2d or higher dimensional source vectors.
if (warpOp.getResult(operandNumber).getType() == operand->get().getType()) {
// There is no distribution, this is a broadcast. Simply move the extract
// out of the warp op.
// TODO: This could be optimized. E.g., in case of a scalar result, let
// one lane extract and shuffle the result to all other lanes (same as
// the 1d case).
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {extractOp.getVector()},
{extractOp.getSourceVectorType()}, newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value distributedVec = newWarpOp->getResult(newRetIndices[0]);
// Extract from distributed vector.
Value newExtract = vector::ExtractOp::create(
rewriter, loc, distributedVec, extractOp.getMixedPosition());
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber),
newExtract);
return success();
}
// Find the distributed dimension. There should be exactly one.
auto distributedType =
cast<VectorType>(warpOp.getResult(operandNumber).getType());
auto yieldedType = cast<VectorType>(operand->get().getType());
int64_t distributedDim = getDistributedDim(yieldedType, distributedType);
assert(distributedDim != -1 && "could not find distributed dimension");
(void)distributedDim;
// Yield source vector from warp op.
SmallVector<int64_t> newDistributedShape(extractSrcType.getShape());
for (int i = 0; i < distributedType.getRank(); ++i)
newDistributedShape[i + extractOp.getNumIndices()] =
distributedType.getDimSize(i);
auto newDistributedType =
VectorType::get(newDistributedShape, distributedType.getElementType());
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {extractOp.getVector()}, {newDistributedType},
newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value distributedVec = newWarpOp->getResult(newRetIndices[0]);
// Extract from distributed vector.
Value newExtract = vector::ExtractOp::create(rewriter, loc, distributedVec,
extractOp.getMixedPosition());
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber),
newExtract);
return success();
}
};
/// Pattern to move out vector.extract with a scalar result.
/// Only supports 1-D and 0-D sources for now.
struct WarpOpExtractScalar : public WarpDistributionPattern {
WarpOpExtractScalar(MLIRContext *ctx, WarpShuffleFromIdxFn fn,
PatternBenefit b = 1)
: WarpDistributionPattern(ctx, b), warpShuffleFromIdxFn(std::move(fn)) {}
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *operand =
getWarpResult(warpOp, llvm::IsaPred<vector::ExtractOp>);
if (!operand)
return failure();
unsigned int operandNumber = operand->getOperandNumber();
auto extractOp = operand->get().getDefiningOp<vector::ExtractOp>();
VectorType extractSrcType = extractOp.getSourceVectorType();
// Only supports 1-D or 0-D sources for now.
if (extractSrcType.getRank() > 1) {
return rewriter.notifyMatchFailure(
extractOp, "only 0-D or 1-D source supported for now");
}
// TODO: Supported shuffle types should be parameterizable, similar to
// `WarpShuffleFromIdxFn`.
if (!extractSrcType.getElementType().isF32() &&
!extractSrcType.getElementType().isInteger(32))
return rewriter.notifyMatchFailure(
extractOp, "only f32/i32 element types are supported");
bool is0dOrVec1Extract = extractSrcType.getNumElements() == 1;
Type elType = extractSrcType.getElementType();
VectorType distributedVecType;
if (!is0dOrVec1Extract) {
assert(extractSrcType.getRank() == 1 &&
"expected that extract src rank is 0 or 1");
if (extractSrcType.getShape()[0] % warpOp.getWarpSize() != 0)
return failure();
int64_t elementsPerLane =
extractSrcType.getShape()[0] / warpOp.getWarpSize();
distributedVecType = VectorType::get({elementsPerLane}, elType);
} else {
distributedVecType = extractSrcType;
}
// Yield source vector and position (if present) from warp op.
SmallVector<Value> additionalResults{extractOp.getVector()};
SmallVector<Type> additionalResultTypes{distributedVecType};
additionalResults.append(
SmallVector<Value>(extractOp.getDynamicPosition()));
additionalResultTypes.append(
SmallVector<Type>(extractOp.getDynamicPosition().getTypes()));
Location loc = extractOp.getLoc();
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, additionalResults, additionalResultTypes,
newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value distributedVec = newWarpOp->getResult(newRetIndices[0]);
// 0d extract: The new warp op broadcasts the source vector to all lanes.
// All lanes extract the scalar.
if (is0dOrVec1Extract) {
Value newExtract;
SmallVector<int64_t> indices(extractSrcType.getRank(), 0);
newExtract =
vector::ExtractOp::create(rewriter, loc, distributedVec, indices);
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber),
newExtract);
return success();
}
int64_t staticPos = extractOp.getStaticPosition()[0];
OpFoldResult pos = ShapedType::isDynamic(staticPos)
? (newWarpOp->getResult(newRetIndices[1]))
: OpFoldResult(rewriter.getIndexAttr(staticPos));
// 1d extract: Distribute the source vector. One lane extracts and shuffles
// the value to all other lanes.
int64_t elementsPerLane = distributedVecType.getShape()[0];
AffineExpr sym0 = getAffineSymbolExpr(0, rewriter.getContext());
// tid of extracting thread: pos / elementsPerLane
Value broadcastFromTid = affine::makeComposedAffineApply(
rewriter, loc, sym0.ceilDiv(elementsPerLane), pos);
// Extract at position: pos % elementsPerLane
Value newPos =
elementsPerLane == 1
? arith::ConstantIndexOp::create(rewriter, loc, 0).getResult()
: affine::makeComposedAffineApply(rewriter, loc,
sym0 % elementsPerLane, pos);
Value extracted =
vector::ExtractOp::create(rewriter, loc, distributedVec, newPos);
// Shuffle the extracted value to all lanes.
Value shuffled = warpShuffleFromIdxFn(
loc, rewriter, extracted, broadcastFromTid, newWarpOp.getWarpSize());
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber), shuffled);
return success();
}
private:
WarpShuffleFromIdxFn warpShuffleFromIdxFn;
};
/// Pattern to move out vector.insert with a scalar input.
/// Only supports 1-D and 0-D destinations for now.
struct WarpOpInsertScalar : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *operand = getWarpResult(warpOp, llvm::IsaPred<vector::InsertOp>);
if (!operand)
return failure();
unsigned int operandNumber = operand->getOperandNumber();
auto insertOp = operand->get().getDefiningOp<vector::InsertOp>();
VectorType vecType = insertOp.getDestVectorType();
VectorType distrType =
cast<VectorType>(warpOp.getResult(operandNumber).getType());
// Only supports 1-D or 0-D destinations for now.
if (vecType.getRank() > 1) {
return rewriter.notifyMatchFailure(
insertOp, "only 0-D or 1-D source supported for now");
}
// Yield destination vector, source scalar and position from warp op.
SmallVector<Value> additionalResults{insertOp.getDest(),
insertOp.getValueToStore()};
SmallVector<Type> additionalResultTypes{
distrType, insertOp.getValueToStore().getType()};
additionalResults.append(SmallVector<Value>(insertOp.getDynamicPosition()));
additionalResultTypes.append(
SmallVector<Type>(insertOp.getDynamicPosition().getTypes()));
Location loc = insertOp.getLoc();
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, additionalResults, additionalResultTypes,
newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value distributedVec = newWarpOp->getResult(newRetIndices[0]);
Value newSource = newWarpOp->getResult(newRetIndices[1]);
rewriter.setInsertionPointAfter(newWarpOp);
OpFoldResult pos;
if (vecType.getRank() != 0) {
int64_t staticPos = insertOp.getStaticPosition()[0];
pos = ShapedType::isDynamic(staticPos)
? (newWarpOp->getResult(newRetIndices[2]))
: OpFoldResult(rewriter.getIndexAttr(staticPos));
}
// This condition is always true for 0-d vectors.
if (vecType == distrType) {
Value newInsert;
SmallVector<OpFoldResult> indices;
if (pos) {
indices.push_back(pos);
}
newInsert = vector::InsertOp::create(rewriter, loc, newSource,
distributedVec, indices);
// Broadcast: Simply move the vector.insert op out.
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber),
newInsert);
return success();
}
// This is a distribution. Only one lane should insert.
int64_t elementsPerLane = distrType.getShape()[0];
AffineExpr sym0 = getAffineSymbolExpr(0, rewriter.getContext());
// tid of extracting thread: pos / elementsPerLane
Value insertingLane = affine::makeComposedAffineApply(
rewriter, loc, sym0.ceilDiv(elementsPerLane), pos);
// Insert position: pos % elementsPerLane
OpFoldResult newPos = affine::makeComposedFoldedAffineApply(
rewriter, loc, sym0 % elementsPerLane, pos);
Value isInsertingLane =
arith::CmpIOp::create(rewriter, loc, arith::CmpIPredicate::eq,
newWarpOp.getLaneid(), insertingLane);
Value newResult =
scf::IfOp::create(
rewriter, loc, isInsertingLane,
/*thenBuilder=*/
[&](OpBuilder &builder, Location loc) {
Value newInsert = vector::InsertOp::create(
builder, loc, newSource, distributedVec, newPos);
scf::YieldOp::create(builder, loc, newInsert);
},
/*elseBuilder=*/
[&](OpBuilder &builder, Location loc) {
scf::YieldOp::create(builder, loc, distributedVec);
})
.getResult(0);
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber), newResult);
return success();
}
};
struct WarpOpInsert : public WarpDistributionPattern {
using Base::Base;
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *operand = getWarpResult(warpOp, llvm::IsaPred<vector::InsertOp>);
if (!operand)
return failure();
unsigned int operandNumber = operand->getOperandNumber();
auto insertOp = operand->get().getDefiningOp<vector::InsertOp>();
Location loc = insertOp.getLoc();
// For 1-d or 0-d destination cases, we rely on WarpOpInsertScalar pattern.
if (insertOp.getDestVectorType().getRank() <= 1) {
return failure();
}
// All following cases are 2d or higher dimensional source vectors.
if (warpOp.getResult(operandNumber).getType() == operand->get().getType()) {
// There is no distribution, this is a broadcast. Simply move the insert
// out of the warp op.
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {insertOp.getValueToStore(), insertOp.getDest()},
{insertOp.getValueToStoreType(), insertOp.getDestVectorType()},
newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value distributedSrc = newWarpOp->getResult(newRetIndices[0]);
Value distributedDest = newWarpOp->getResult(newRetIndices[1]);
Value newResult = vector::InsertOp::create(rewriter, loc, distributedSrc,
distributedDest,
insertOp.getMixedPosition());
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber),
newResult);
return success();
}
// Find the distributed dimension. There should be exactly one.
auto distrDestType =
cast<VectorType>(warpOp.getResult(operandNumber).getType());
auto yieldedType = cast<VectorType>(operand->get().getType());
int64_t distrDestDim = -1;
for (int64_t i = 0; i < yieldedType.getRank(); ++i) {
if (distrDestType.getDimSize(i) != yieldedType.getDimSize(i)) {
// Keep this assert here in case WarpExecuteOnLane0Op gets extended to
// support distributing multiple dimensions in the future.
assert(distrDestDim == -1 && "found multiple distributed dims");
distrDestDim = i;
}
}
assert(distrDestDim != -1 && "could not find distributed dimension");
// Compute the distributed source vector type.
VectorType srcVecType = cast<VectorType>(insertOp.getValueToStoreType());
SmallVector<int64_t> distrSrcShape(srcVecType.getShape());
// E.g.: vector.insert %s, %d [2] : vector<96xf32> into vector<128x96xf32>
// Case 1: distrDestDim = 1 (dim of size 96). In that case, each lane will
// insert a smaller vector<3xf32>.
// Case 2: distrDestDim = 0 (dim of size 128) => distrSrcDim = -1. In that
// case, one lane will insert the source vector<96xf32>. The other
// lanes will not do anything.
int64_t distrSrcDim = distrDestDim - insertOp.getNumIndices();
if (distrSrcDim >= 0)
distrSrcShape[distrSrcDim] = distrDestType.getDimSize(distrDestDim);
auto distrSrcType =
VectorType::get(distrSrcShape, distrDestType.getElementType());
// Yield source and dest vectors from warp op.
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {insertOp.getValueToStore(), insertOp.getDest()},
{distrSrcType, distrDestType}, newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
Value distributedSrc = newWarpOp->getResult(newRetIndices[0]);
Value distributedDest = newWarpOp->getResult(newRetIndices[1]);
// Insert into the distributed vector.
Value newResult;
if (distrSrcDim >= 0) {
// Every lane inserts a small piece.
newResult = vector::InsertOp::create(rewriter, loc, distributedSrc,
distributedDest,
insertOp.getMixedPosition());
} else {
// One lane inserts the entire source vector.
int64_t elementsPerLane = distrDestType.getDimSize(distrDestDim);
SmallVector<OpFoldResult> pos = insertOp.getMixedPosition();
SmallVector<int64_t> newPos = getAsIntegers(pos);
// tid of inserting lane: pos / elementsPerLane
Value insertingLane = arith::ConstantIndexOp::create(
rewriter, loc, newPos[distrDestDim] / elementsPerLane);
Value isInsertingLane =
arith::CmpIOp::create(rewriter, loc, arith::CmpIPredicate::eq,
newWarpOp.getLaneid(), insertingLane);
// Insert position: pos % elementsPerLane
newPos[distrDestDim] %= elementsPerLane;
auto insertingBuilder = [&](OpBuilder &builder, Location loc) {
Value newInsert = vector::InsertOp::create(builder, loc, distributedSrc,
distributedDest, newPos);
scf::YieldOp::create(builder, loc, newInsert);
};
auto nonInsertingBuilder = [&](OpBuilder &builder, Location loc) {
scf::YieldOp::create(builder, loc, distributedDest);
};
newResult = scf::IfOp::create(rewriter, loc, isInsertingLane,
/*thenBuilder=*/insertingBuilder,
/*elseBuilder=*/nonInsertingBuilder)
.getResult(0);
}
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber), newResult);
return success();
}
};
/// Sink scf.for region out of WarpExecuteOnLane0Op. This can be done only if
/// the scf.ForOp is the last operation in the region so that it doesn't
/// change the order of execution. This creates a new scf.for region after the
/// WarpExecuteOnLane0Op. The new scf.for region will contain a new
/// WarpExecuteOnLane0Op region. Example:
/// ```
/// %w = gpu.warp_execute_on_lane_0(%laneid) -> (vector<4xf32>) {
/// ...
/// %v1 = scf.for %arg3 = %c0 to %c128 step %c1 iter_args(%arg4 = %v)
/// -> (vector<128xf32>) {
/// ...
/// scf.yield %r : vector<128xf32>
/// }
/// gpu.yield %v1 : vector<128xf32>
/// }
/// ```
/// To:
/// %w0 = gpu.warp_execute_on_lane_0(%arg0) -> (vector<4xf32>) {
/// ...
/// gpu.yield %v : vector<128xf32>
/// }
/// %w = scf.for %arg3 = %c0 to %c128 step %c1 iter_args(%varg = %q0)
/// -> (vector<4xf32>) {
/// %iw = gpu.warp_execute_on_lane_0(%laneid)
/// args(%varg : vector<4xf32>) -> (vector<4xf32>) {
/// ^bb0(%arg: vector<128xf32>):
/// ...
/// gpu.yield %ir : vector<128xf32>
/// }
/// scf.yield %iw : vector<4xf32>
/// }
/// ```
struct WarpOpScfForOp : public WarpDistributionPattern {
WarpOpScfForOp(MLIRContext *ctx, DistributionMapFn fn, PatternBenefit b = 1)
: WarpDistributionPattern(ctx, b), distributionMapFn(std::move(fn)) {}
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
gpu::YieldOp warpOpYield = warpOp.getTerminator();
// Only pick up `ForOp` if it is the last op in the region.
Operation *lastNode = warpOpYield->getPrevNode();
auto forOp = dyn_cast_or_null<scf::ForOp>(lastNode);
if (!forOp)
return failure();
// Collect Values that come from the `WarpOp` but are outside the `ForOp`.
// Those Values need to be returned by the new warp op.
llvm::SmallSetVector<Value, 32> escapingValues;
SmallVector<Type> escapingValueInputTypes;
SmallVector<Type> escapingValueDistTypes;
mlir::visitUsedValuesDefinedAbove(
forOp.getBodyRegion(), [&](OpOperand *operand) {
Operation *parent = operand->get().getParentRegion()->getParentOp();
if (warpOp->isAncestor(parent)) {
if (!escapingValues.insert(operand->get()))
return;
Type distType = operand->get().getType();
if (auto vecType = dyn_cast<VectorType>(distType)) {
AffineMap map = distributionMapFn(operand->get());
distType = getDistributedType(vecType, map, warpOp.getWarpSize());
}
escapingValueInputTypes.push_back(operand->get().getType());
escapingValueDistTypes.push_back(distType);
}
});
if (llvm::is_contained(escapingValueDistTypes, Type{}))
return failure();
// `WarpOp` can yield two types of values:
// 1. Values that are not results of the `ForOp`:
// These values must also be yielded by the new `WarpOp`. Also, we need
// to record the index mapping for these values to replace them later.
// 2. Values that are results of the `ForOp`:
// In this case, we record the index mapping between the `WarpOp` result
// index and matching `ForOp` result index.
// Additionally, we keep track of the distributed types for all `ForOp`
// vector results.
SmallVector<Value> nonForYieldedValues;
SmallVector<unsigned> nonForResultIndices;
llvm::SmallDenseMap<unsigned, unsigned> forResultMapping;
llvm::SmallDenseMap<unsigned, VectorType> forResultDistTypes;
for (OpOperand &yieldOperand : warpOpYield->getOpOperands()) {
// Yielded value is not a result of the forOp.
if (yieldOperand.get().getDefiningOp() != forOp.getOperation()) {
nonForYieldedValues.push_back(yieldOperand.get());
nonForResultIndices.push_back(yieldOperand.getOperandNumber());
continue;
}
OpResult forResult = cast<OpResult>(yieldOperand.get());
unsigned int forResultNumber = forResult.getResultNumber();
forResultMapping[yieldOperand.getOperandNumber()] = forResultNumber;
// If this `ForOp` result is vector type and it is yielded by the
// `WarpOp`, we keep track the distributed type for this result.
if (!isa<VectorType>(forResult.getType()))
continue;
VectorType distType = cast<VectorType>(
warpOp.getResult(yieldOperand.getOperandNumber()).getType());
forResultDistTypes[forResultNumber] = distType;
}
// Newly created `WarpOp` will yield values in following order:
// 1. All init args of the `ForOp`.
// 2. All escaping values.
// 3. All non-`ForOp` yielded values.
SmallVector<Value> newWarpOpYieldValues;
SmallVector<Type> newWarpOpDistTypes;
for (auto [i, initArg] : llvm::enumerate(forOp.getInitArgs())) {
newWarpOpYieldValues.push_back(initArg);
// Compute the distributed type for this init arg.
Type distType = initArg.getType();
if (auto vecType = dyn_cast<VectorType>(distType)) {
// If the `ForOp` result corresponds to this init arg is already yielded
// we can get the distributed type from `forResultDistTypes` map.
// Otherwise, we compute it using distributionMapFn.
AffineMap map = distributionMapFn(initArg);
distType = forResultDistTypes.count(i)
? forResultDistTypes[i]
: getDistributedType(vecType, map, warpOp.getWarpSize());
}
newWarpOpDistTypes.push_back(distType);
}
// Insert escaping values and their distributed types.
newWarpOpYieldValues.insert(newWarpOpYieldValues.end(),
escapingValues.begin(), escapingValues.end());
newWarpOpDistTypes.insert(newWarpOpDistTypes.end(),
escapingValueDistTypes.begin(),
escapingValueDistTypes.end());
// Next, we insert all non-`ForOp` yielded values and their distributed
// types. We also create a mapping between the non-`ForOp` yielded value
// index and the corresponding new `WarpOp` yield value index (needed to
// update users later).
llvm::SmallDenseMap<unsigned, unsigned> nonForResultMapping;
for (auto [i, v] :
llvm::zip_equal(nonForResultIndices, nonForYieldedValues)) {
nonForResultMapping[i] = newWarpOpYieldValues.size();
newWarpOpYieldValues.push_back(v);
newWarpOpDistTypes.push_back(warpOp.getResult(i).getType());
}
// Create the new `WarpOp` with the updated yield values and types.
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndReplaceReturns(
rewriter, warpOp, newWarpOpYieldValues, newWarpOpDistTypes);
// Next, we create a new `ForOp` with the init args yielded by the new
// `WarpOp`.
const unsigned escapingValuesStartIdx =
forOp.getInitArgs().size(); // `ForOp` init args are positioned before
// escaping values in the new `WarpOp`.
SmallVector<Value> newForOpOperands;
for (size_t i = 0; i < escapingValuesStartIdx; ++i)
newForOpOperands.push_back(newWarpOp.getResult(i));
// Create a new `ForOp` outside the new `WarpOp` region.
OpBuilder::InsertionGuard g(rewriter);
rewriter.setInsertionPointAfter(newWarpOp);
auto newForOp = scf::ForOp::create(
rewriter, forOp.getLoc(), forOp.getLowerBound(), forOp.getUpperBound(),
forOp.getStep(), newForOpOperands, /*bodyBuilder=*/nullptr,
forOp.getUnsignedCmp());
// Next, we insert a new `WarpOp` (called inner `WarpOp`) inside the
// newly created `ForOp`. This `WarpOp` will contain all ops that were
// contained within the original `ForOp` body.
rewriter.setInsertionPointToStart(newForOp.getBody());
SmallVector<Value> innerWarpInput(newForOp.getRegionIterArgs().begin(),
newForOp.getRegionIterArgs().end());
SmallVector<Type> innerWarpInputType(forOp.getResultTypes().begin(),
forOp.getResultTypes().end());
// Escaping values are forwarded to the inner `WarpOp` as its (additional)
// arguments. We keep track of the mapping between these values and their
// argument index in the inner `WarpOp` (to replace users later).
llvm::SmallDenseMap<Value, int64_t> argIndexMapping;
for (size_t i = escapingValuesStartIdx;
i < escapingValuesStartIdx + escapingValues.size(); ++i) {
innerWarpInput.push_back(newWarpOp.getResult(i));
argIndexMapping[escapingValues[i - escapingValuesStartIdx]] =
innerWarpInputType.size();
innerWarpInputType.push_back(
escapingValueInputTypes[i - escapingValuesStartIdx]);
}
// Create the inner `WarpOp` with the new input values and types.
auto innerWarp = WarpExecuteOnLane0Op::create(
rewriter, newWarpOp.getLoc(), newForOp.getResultTypes(),
newWarpOp.getLaneid(), newWarpOp.getWarpSize(), innerWarpInput,
innerWarpInputType);
// Inline the `ForOp` body into the inner `WarpOp` body.
SmallVector<Value> argMapping;
argMapping.push_back(newForOp.getInductionVar());
for (Value args : innerWarp.getBody()->getArguments())
argMapping.push_back(args);
argMapping.resize(forOp.getBody()->getNumArguments());
SmallVector<Value> yieldOperands;
for (Value operand : forOp.getBody()->getTerminator()->getOperands())
yieldOperands.push_back(operand);
rewriter.eraseOp(forOp.getBody()->getTerminator());
rewriter.mergeBlocks(forOp.getBody(), innerWarp.getBody(), argMapping);
// Insert a gpu `YieldOp` at the end of the inner `WarpOp` body that yields
// original `ForOp` results.
rewriter.setInsertionPointToEnd(innerWarp.getBody());
gpu::YieldOp::create(rewriter, innerWarp.getLoc(), yieldOperands);
rewriter.setInsertionPointAfter(innerWarp);
// Insert a scf.yield op at the end of the new `ForOp` body that yields
// the inner `WarpOp` results.
if (!innerWarp.getResults().empty())
scf::YieldOp::create(rewriter, forOp.getLoc(), innerWarp.getResults());
// Update the users of original `WarpOp` results that were coming from the
// original `ForOp` to the corresponding new `ForOp` result.
for (auto [origIdx, newIdx] : forResultMapping)
rewriter.replaceAllUsesExcept(warpOp.getResult(origIdx),
newForOp.getResult(newIdx), newForOp);
// Similarly, update any users of the `WarpOp` results that were not
// results of the `ForOp`.
for (auto [origIdx, newIdx] : nonForResultMapping)
rewriter.replaceAllUsesWith(warpOp.getResult(origIdx),
newWarpOp.getResult(newIdx));
// Remove the original `WarpOp` and `ForOp`, they should not have any uses
// at this point.
rewriter.eraseOp(forOp);
rewriter.eraseOp(warpOp);
// Update any users of escaping values that were forwarded to the
// inner `WarpOp`. These values are now arguments of the inner `WarpOp`.
newForOp.walk([&](Operation *op) {
for (OpOperand &operand : op->getOpOperands()) {
auto it = argIndexMapping.find(operand.get());
if (it == argIndexMapping.end())
continue;
operand.set(innerWarp.getBodyRegion().getArgument(it->second));
}
});
// Finally, hoist out any now uniform code from the inner `WarpOp`.
mlir::vector::moveScalarUniformCode(innerWarp);
return success();
}
private:
DistributionMapFn distributionMapFn;
};
/// A pattern that extracts vector.reduction ops from a WarpExecuteOnLane0Op.
/// The vector is reduced in parallel. Currently limited to vector size
/// matching the warpOp size. E.g.:
/// ```
/// %r = gpu.warp_execute_on_lane_0(%laneid)[32] -> (f32) {
/// %0 = "some_def"() : () -> (vector<32xf32>)
/// %1 = vector.reduction "add", %0 : vector<32xf32> into f32
/// gpu.yield %1 : f32
/// }
/// ```
/// is lowered to:
/// ```
/// %0 = gpu.warp_execute_on_lane_0(%laneid)[32] -> (vector<1xf32>) {
/// %1 = "some_def"() : () -> (vector<32xf32>)
/// gpu.yield %1 : vector<32xf32>
/// }
/// %a = vector.extract %0[0] : f32 from vector<1xf32>
/// %r = ("warp.reduction %a")
/// ```
struct WarpOpReduction : public WarpDistributionPattern {
WarpOpReduction(MLIRContext *context,
DistributedReductionFn distributedReductionFn,
PatternBenefit benefit = 1)
: WarpDistributionPattern(context, benefit),
distributedReductionFn(std::move(distributedReductionFn)) {}
LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
PatternRewriter &rewriter) const override {
OpOperand *yieldOperand =
getWarpResult(warpOp, llvm::IsaPred<vector::ReductionOp>);
if (!yieldOperand)
return failure();
auto reductionOp =
cast<vector::ReductionOp>(yieldOperand->get().getDefiningOp());
auto vectorType = cast<VectorType>(reductionOp.getVector().getType());
// Only rank 1 vectors supported.
if (vectorType.getRank() != 1)
return rewriter.notifyMatchFailure(
warpOp, "Only rank 1 reductions can be distributed.");
// Only warp_size-sized vectors supported.
if (vectorType.getShape()[0] % warpOp.getWarpSize() != 0)
return rewriter.notifyMatchFailure(
warpOp, "Reduction vector dimension must match was size.");
if (!reductionOp.getType().isIntOrFloat())
return rewriter.notifyMatchFailure(
warpOp, "Reduction distribution currently only supports floats and "
"integer types.");
int64_t numElements = vectorType.getShape()[0] / warpOp.getWarpSize();
// Return vector that will be reduced from the WarpExecuteOnLane0Op.
unsigned operandIndex = yieldOperand->getOperandNumber();
SmallVector<Value> yieldValues = {reductionOp.getVector()};
SmallVector<Type> retTypes = {
VectorType::get({numElements}, reductionOp.getType())};
if (reductionOp.getAcc()) {
yieldValues.push_back(reductionOp.getAcc());
retTypes.push_back(reductionOp.getAcc().getType());
}
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, yieldValues, retTypes, newRetIndices);
rewriter.setInsertionPointAfter(newWarpOp);
// Obtain data to reduce for a single lane.
Value laneValVec = newWarpOp.getResult(newRetIndices[0]);
// Distribute and reduce across threads.
Value fullReduce =
distributedReductionFn(reductionOp.getLoc(), rewriter, laneValVec,
reductionOp.getKind(), newWarpOp.getWarpSize());
if (reductionOp.getAcc()) {
fullReduce = vector::makeArithReduction(
rewriter, reductionOp.getLoc(), reductionOp.getKind(), fullReduce,
newWarpOp.getResult(newRetIndices[1]));
}
rewriter.replaceAllUsesWith(newWarpOp.getResult(operandIndex), fullReduce);
return success();
}
private:
DistributedReductionFn distributedReductionFn;
};
} // namespace
void mlir::vector::populateWarpExecuteOnLane0OpToScfForPattern(
RewritePatternSet &patterns,
const WarpExecuteOnLane0LoweringOptions &options, PatternBenefit benefit) {
patterns.add<WarpOpToScfIfPattern>(patterns.getContext(), options, benefit);
}
void mlir::vector::populateDistributeTransferWriteOpPatterns(
RewritePatternSet &patterns, const DistributionMapFn &distributionMapFn,
unsigned maxNumElementsToExtract, PatternBenefit benefit) {
patterns.add<WarpOpTransferWrite>(patterns.getContext(), distributionMapFn,
maxNumElementsToExtract, benefit);
}
void mlir::vector::populatePropagateWarpVectorDistributionPatterns(
RewritePatternSet &patterns, const DistributionMapFn &distributionMapFn,
const WarpShuffleFromIdxFn &warpShuffleFromIdxFn, PatternBenefit benefit,
PatternBenefit readBenefit) {
patterns.add<WarpOpTransferRead>(patterns.getContext(), readBenefit);
patterns
.add<WarpOpElementwise, WarpOpDeadResult, WarpOpBroadcast,
WarpOpShapeCast, WarpOpExtract, WarpOpForwardOperand, WarpOpConstant,
WarpOpInsertScalar, WarpOpInsert, WarpOpCreateMask,
WarpOpExtractStridedSlice, WarpOpInsertStridedSlice, WarpOpStep>(
patterns.getContext(), benefit);
patterns.add<WarpOpExtractScalar>(patterns.getContext(), warpShuffleFromIdxFn,
benefit);
patterns.add<WarpOpScfForOp>(patterns.getContext(), distributionMapFn,
benefit);
}
void mlir::vector::populateDistributeReduction(
RewritePatternSet &patterns,
const DistributedReductionFn &distributedReductionFn,
PatternBenefit benefit) {
patterns.add<WarpOpReduction>(patterns.getContext(), distributedReductionFn,
benefit);
}
/// Helper to know if an op can be hoisted out of the region.
static bool canBeHoisted(Operation *op,
function_ref<bool(Value)> definedOutside) {
return llvm::all_of(op->getOperands(), definedOutside) &&
isMemoryEffectFree(op) && op->getNumRegions() == 0;
}
void mlir::vector::moveScalarUniformCode(WarpExecuteOnLane0Op warpOp) {
Block *body = warpOp.getBody();
// Keep track of the ops we want to hoist.
llvm::SmallSetVector<Operation *, 8> opsToMove;
// Helper to check if a value is or will be defined outside of the region.
auto isDefinedOutsideOfBody = [&](Value value) {
auto *definingOp = value.getDefiningOp();
return (definingOp && opsToMove.count(definingOp)) ||
warpOp.isDefinedOutsideOfRegion(value);
};
// Do not use walk here, as we do not want to go into nested regions and hoist
// operations from there.
for (auto &op : body->without_terminator()) {
bool hasVectorResult = llvm::any_of(op.getResults(), [](Value result) {
return isa<VectorType>(result.getType());
});
if (!hasVectorResult && canBeHoisted(&op, isDefinedOutsideOfBody))
opsToMove.insert(&op);
}
// Move all the ops marked as uniform outside of the region.
for (Operation *op : opsToMove)
op->moveBefore(warpOp);
}