#include "mlir/Conversion/SCFToOpenMP/SCFToOpenMP.h"
#include "mlir/Analysis/SliceAnalysis.h"
#include "mlir/Dialect/Affine/Analysis/LoopAnalysis.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/OpenMP/OpenMPDialect.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/IR/ImplicitLocOpBuilder.h"
#include "mlir/IR/SymbolTable.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/DialectConversion.h"
namespace mlir {
#define GEN_PASS_DEF_CONVERTSCFTOOPENMPPASS
#include "mlir/Conversion/Passes.h.inc"
}
using namespace mlir;
template <typename... OpTy>
static bool matchSimpleReduction(Block &block) {
if (block.empty() || llvm::hasSingleElement(block) ||
std::next(block.begin(), 2) != block.end())
return false;
if (block.getNumArguments() != 2)
return false;
SmallVector<Operation *, 4> combinerOps;
Value reducedVal = matchReduction({block.getArguments()[1]},
0, combinerOps);
if (!reducedVal || !isa<BlockArgument>(reducedVal) || combinerOps.size() != 1)
return false;
return isa<OpTy...>(combinerOps[0]) &&
isa<scf::ReduceReturnOp>(block.back()) &&
block.front().getOperands() == block.getArguments();
}
template <
typename CompareOpTy, typename SelectOpTy,
typename Predicate = decltype(std::declval<CompareOpTy>().getPredicate())>
static bool
matchSelectReduction(Block &block, ArrayRef<Predicate> lessThanPredicates,
ArrayRef<Predicate> greaterThanPredicates, bool &isMin) {
static_assert(
llvm::is_one_of<SelectOpTy, arith::SelectOp, LLVM::SelectOp>::value,
"only arithmetic and llvm select ops are supported");
if (block.empty() || llvm::hasSingleElement(block) ||
std::next(block.begin(), 2) == block.end() ||
std::next(block.begin(), 3) != block.end())
return false;
auto compare = dyn_cast<CompareOpTy>(block.front());
auto select = dyn_cast<SelectOpTy>(block.front().getNextNode());
auto terminator = dyn_cast<scf::ReduceReturnOp>(block.back());
if (!compare || !select || !terminator)
return false;
if (compare->getOperands() != block.getArguments())
return false;
bool isLess;
if (llvm::is_contained(lessThanPredicates, compare.getPredicate())) {
isLess = true;
} else if (llvm::is_contained(greaterThanPredicates,
compare.getPredicate())) {
isLess = false;
} else {
return false;
}
if (select.getCondition() != compare.getResult())
return false;
constexpr unsigned kTrueValue = 1;
constexpr unsigned kFalseValue = 2;
bool sameOperands = select.getOperand(kTrueValue) == compare.getLhs() &&
select.getOperand(kFalseValue) == compare.getRhs();
bool swappedOperands = select.getOperand(kTrueValue) == compare.getRhs() &&
select.getOperand(kFalseValue) == compare.getLhs();
if (!sameOperands && !swappedOperands)
return false;
if (select.getResult() != terminator.getResult())
return false;
isMin = (isLess && sameOperands) || (!isLess && swappedOperands);
return isMin || (isLess & swappedOperands) || (!isLess && sameOperands);
}
static const llvm::fltSemantics &fltSemanticsForType(FloatType type) {
if (type.isF16())
return llvm::APFloat::IEEEhalf();
if (type.isF32())
return llvm::APFloat::IEEEsingle();
if (type.isF64())
return llvm::APFloat::IEEEdouble();
if (type.isF128())
return llvm::APFloat::IEEEquad();
if (type.isBF16())
return llvm::APFloat::BFloat();
if (type.isF80())
return llvm::APFloat::x87DoubleExtended();
llvm_unreachable("unknown float type");
}
static Attribute minMaxValueForFloat(Type type, bool min) {
auto fltType = cast<FloatType>(type);
return FloatAttr::get(
type, llvm::APFloat::getLargest(fltSemanticsForType(fltType), min));
}
static Attribute minMaxValueForSignedInt(Type type, bool min) {
auto intType = cast<IntegerType>(type);
unsigned bitwidth = intType.getWidth();
return IntegerAttr::get(type, min ? llvm::APInt::getSignedMinValue(bitwidth)
: llvm::APInt::getSignedMaxValue(bitwidth));
}
static Attribute minMaxValueForUnsignedInt(Type type, bool min) {
auto intType = cast<IntegerType>(type);
unsigned bitwidth = intType.getWidth();
return IntegerAttr::get(type, min ? llvm::APInt::getZero(bitwidth)
: llvm::APInt::getAllOnes(bitwidth));
}
static omp::DeclareReductionOp
createDecl(PatternRewriter &builder, SymbolTable &symbolTable,
scf::ReduceOp reduce, int64_t reductionIndex, Attribute initValue) {
OpBuilder::InsertionGuard guard(builder);
Type type = reduce.getOperands()[reductionIndex].getType();
auto decl = builder.create<omp::DeclareReductionOp>(reduce.getLoc(),
"__scf_reduction", type);
symbolTable.insert(decl);
builder.createBlock(&decl.getInitializerRegion(),
decl.getInitializerRegion().end(), {type},
{reduce.getOperands()[reductionIndex].getLoc()});
builder.setInsertionPointToEnd(&decl.getInitializerRegion().back());
Value init =
builder.create<LLVM::ConstantOp>(reduce.getLoc(), type, initValue);
builder.create<omp::YieldOp>(reduce.getLoc(), init);
Operation *terminator =
&reduce.getReductions()[reductionIndex].front().back();
assert(isa<scf::ReduceReturnOp>(terminator) &&
"expected reduce op to be terminated by redure return");
builder.setInsertionPoint(terminator);
builder.replaceOpWithNewOp<omp::YieldOp>(terminator,
terminator->getOperands());
builder.inlineRegionBefore(reduce.getReductions()[reductionIndex],
decl.getReductionRegion(),
decl.getReductionRegion().end());
return decl;
}
static omp::DeclareReductionOp addAtomicRMW(OpBuilder &builder,
LLVM::AtomicBinOp atomicKind,
omp::DeclareReductionOp decl,
scf::ReduceOp reduce,
int64_t reductionIndex) {
OpBuilder::InsertionGuard guard(builder);
auto ptrType = LLVM::LLVMPointerType::get(builder.getContext());
Location reduceOperandLoc = reduce.getOperands()[reductionIndex].getLoc();
builder.createBlock(&decl.getAtomicReductionRegion(),
decl.getAtomicReductionRegion().end(), {ptrType, ptrType},
{reduceOperandLoc, reduceOperandLoc});
Block *atomicBlock = &decl.getAtomicReductionRegion().back();
builder.setInsertionPointToEnd(atomicBlock);
Value loaded = builder.create<LLVM::LoadOp>(reduce.getLoc(), decl.getType(),
atomicBlock->getArgument(1));
builder.create<LLVM::AtomicRMWOp>(reduce.getLoc(), atomicKind,
atomicBlock->getArgument(0), loaded,
LLVM::AtomicOrdering::monotonic);
builder.create<omp::YieldOp>(reduce.getLoc(), ArrayRef<Value>());
return decl;
}
static omp::DeclareReductionOp declareReduction(PatternRewriter &builder,
scf::ReduceOp reduce,
int64_t reductionIndex) {
Operation *container = SymbolTable::getNearestSymbolTable(reduce);
SymbolTable symbolTable(container);
Operation *insertionPoint = reduce;
while (insertionPoint->getParentOp() != container)
insertionPoint = insertionPoint->getParentOp();
OpBuilder::InsertionGuard guard(builder);
builder.setInsertionPoint(insertionPoint);
assert(llvm::hasSingleElement(reduce.getReductions()[reductionIndex]) &&
"expected reduction region to have a single element");
Type type = reduce.getOperands()[reductionIndex].getType();
Block &reduction = reduce.getReductions()[reductionIndex].front();
if (matchSimpleReduction<arith::AddFOp, LLVM::FAddOp>(reduction)) {
omp::DeclareReductionOp decl =
createDecl(builder, symbolTable, reduce, reductionIndex,
builder.getFloatAttr(type, 0.0));
return addAtomicRMW(builder, LLVM::AtomicBinOp::fadd, decl, reduce,
reductionIndex);
}
if (matchSimpleReduction<arith::AddIOp, LLVM::AddOp>(reduction)) {
omp::DeclareReductionOp decl =
createDecl(builder, symbolTable, reduce, reductionIndex,
builder.getIntegerAttr(type, 0));
return addAtomicRMW(builder, LLVM::AtomicBinOp::add, decl, reduce,
reductionIndex);
}
if (matchSimpleReduction<arith::OrIOp, LLVM::OrOp>(reduction)) {
omp::DeclareReductionOp decl =
createDecl(builder, symbolTable, reduce, reductionIndex,
builder.getIntegerAttr(type, 0));
return addAtomicRMW(builder, LLVM::AtomicBinOp::_or, decl, reduce,
reductionIndex);
}
if (matchSimpleReduction<arith::XOrIOp, LLVM::XOrOp>(reduction)) {
omp::DeclareReductionOp decl =
createDecl(builder, symbolTable, reduce, reductionIndex,
builder.getIntegerAttr(type, 0));
return addAtomicRMW(builder, LLVM::AtomicBinOp::_xor, decl, reduce,
reductionIndex);
}
if (matchSimpleReduction<arith::AndIOp, LLVM::AndOp>(reduction)) {
omp::DeclareReductionOp decl = createDecl(
builder, symbolTable, reduce, reductionIndex,
builder.getIntegerAttr(
type, llvm::APInt::getAllOnes(type.getIntOrFloatBitWidth())));
return addAtomicRMW(builder, LLVM::AtomicBinOp::_and, decl, reduce,
reductionIndex);
}
if (matchSimpleReduction<arith::MulFOp, LLVM::FMulOp>(reduction)) {
return createDecl(builder, symbolTable, reduce, reductionIndex,
builder.getFloatAttr(type, 1.0));
}
if (matchSimpleReduction<arith::MulIOp, LLVM::MulOp>(reduction)) {
return createDecl(builder, symbolTable, reduce, reductionIndex,
builder.getIntegerAttr(type, 1));
}
bool isMin;
if (matchSelectReduction<arith::CmpFOp, arith::SelectOp>(
reduction, {arith::CmpFPredicate::OLT, arith::CmpFPredicate::OLE},
{arith::CmpFPredicate::OGT, arith::CmpFPredicate::OGE}, isMin) ||
matchSelectReduction<LLVM::FCmpOp, LLVM::SelectOp>(
reduction, {LLVM::FCmpPredicate::olt, LLVM::FCmpPredicate::ole},
{LLVM::FCmpPredicate::ogt, LLVM::FCmpPredicate::oge}, isMin)) {
return createDecl(builder, symbolTable, reduce, reductionIndex,
minMaxValueForFloat(type, !isMin));
}
if (matchSelectReduction<arith::CmpIOp, arith::SelectOp>(
reduction, {arith::CmpIPredicate::slt, arith::CmpIPredicate::sle},
{arith::CmpIPredicate::sgt, arith::CmpIPredicate::sge}, isMin) ||
matchSelectReduction<LLVM::ICmpOp, LLVM::SelectOp>(
reduction, {LLVM::ICmpPredicate::slt, LLVM::ICmpPredicate::sle},
{LLVM::ICmpPredicate::sgt, LLVM::ICmpPredicate::sge}, isMin)) {
omp::DeclareReductionOp decl =
createDecl(builder, symbolTable, reduce, reductionIndex,
minMaxValueForSignedInt(type, !isMin));
return addAtomicRMW(builder,
isMin ? LLVM::AtomicBinOp::min : LLVM::AtomicBinOp::max,
decl, reduce, reductionIndex);
}
if (matchSelectReduction<arith::CmpIOp, arith::SelectOp>(
reduction, {arith::CmpIPredicate::ult, arith::CmpIPredicate::ule},
{arith::CmpIPredicate::ugt, arith::CmpIPredicate::uge}, isMin) ||
matchSelectReduction<LLVM::ICmpOp, LLVM::SelectOp>(
reduction, {LLVM::ICmpPredicate::ugt, LLVM::ICmpPredicate::ule},
{LLVM::ICmpPredicate::ugt, LLVM::ICmpPredicate::uge}, isMin)) {
omp::DeclareReductionOp decl =
createDecl(builder, symbolTable, reduce, reductionIndex,
minMaxValueForUnsignedInt(type, !isMin));
return addAtomicRMW(
builder, isMin ? LLVM::AtomicBinOp::umin : LLVM::AtomicBinOp::umax,
decl, reduce, reductionIndex);
}
return nullptr;
}
namespace {
struct ParallelOpLowering : public OpRewritePattern<scf::ParallelOp> {
static constexpr unsigned kUseOpenMPDefaultNumThreads = 0;
unsigned numThreads;
ParallelOpLowering(MLIRContext *context,
unsigned numThreads = kUseOpenMPDefaultNumThreads)
: OpRewritePattern<scf::ParallelOp>(context), numThreads(numThreads) {}
LogicalResult matchAndRewrite(scf::ParallelOp parallelOp,
PatternRewriter &rewriter) const override {
SmallVector<Attribute> reductionDeclSymbols;
SmallVector<omp::DeclareReductionOp> ompReductionDecls;
auto reduce = cast<scf::ReduceOp>(parallelOp.getBody()->getTerminator());
for (int64_t i = 0, e = parallelOp.getNumReductions(); i < e; ++i) {
omp::DeclareReductionOp decl = declareReduction(rewriter, reduce, i);
ompReductionDecls.push_back(decl);
if (!decl)
return failure();
reductionDeclSymbols.push_back(
SymbolRefAttr::get(rewriter.getContext(), decl.getSymName()));
}
Location loc = parallelOp.getLoc();
Value one = rewriter.create<LLVM::ConstantOp>(
loc, rewriter.getIntegerType(64), rewriter.getI64IntegerAttr(1));
SmallVector<Value> reductionVariables;
reductionVariables.reserve(parallelOp.getNumReductions());
auto ptrType = LLVM::LLVMPointerType::get(parallelOp.getContext());
for (Value init : parallelOp.getInitVals()) {
assert((LLVM::isCompatibleType(init.getType()) ||
isa<LLVM::PointerElementTypeInterface>(init.getType())) &&
"cannot create a reduction variable if the type is not an LLVM "
"pointer element");
Value storage =
rewriter.create<LLVM::AllocaOp>(loc, ptrType, init.getType(), one, 0);
rewriter.create<LLVM::StoreOp>(loc, init, storage);
reductionVariables.push_back(storage);
}
for (auto [x, y, rD] : llvm::zip_equal(
reductionVariables, reduce.getOperands(), ompReductionDecls)) {
OpBuilder::InsertionGuard guard(rewriter);
rewriter.setInsertionPoint(reduce);
Region &redRegion = rD.getReductionRegion();
assert(redRegion.hasOneBlock() &&
"expect reduction region to have one block");
Value pvtRedVar = parallelOp.getRegion().addArgument(x.getType(), loc);
Value pvtRedVal = rewriter.create<LLVM::LoadOp>(reduce.getLoc(),
rD.getType(), pvtRedVar);
mlir::OpBuilder builder(rewriter.getContext());
builder.setInsertionPoint(reduce);
mlir::IRMapping mapper;
assert(redRegion.getNumArguments() == 2 &&
"expect reduction region to have two arguments");
mapper.map(redRegion.getArgument(0), pvtRedVal);
mapper.map(redRegion.getArgument(1), y);
for (auto &op : redRegion.getOps()) {
Operation *cloneOp = builder.clone(op, mapper);
if (auto yieldOp = dyn_cast<omp::YieldOp>(*cloneOp)) {
assert(yieldOp && yieldOp.getResults().size() == 1 &&
"expect YieldOp in reduction region to return one result");
Value redVal = yieldOp.getResults()[0];
rewriter.create<LLVM::StoreOp>(loc, redVal, pvtRedVar);
rewriter.eraseOp(yieldOp);
break;
}
}
}
rewriter.eraseOp(reduce);
Value numThreadsVar;
if (numThreads > 0) {
numThreadsVar = rewriter.create<LLVM::ConstantOp>(
loc, rewriter.getI32IntegerAttr(numThreads));
}
auto ompParallel = rewriter.create<omp::ParallelOp>(
loc,
Value{},
numThreadsVar,
llvm::SmallVector<Value>{},
llvm::SmallVector<Value>{},
llvm::SmallVector<Value>{},
DenseBoolArrayAttr{},
ArrayAttr{},
omp::ClauseProcBindKindAttr{},
ValueRange(),
nullptr);
{
OpBuilder::InsertionGuard guard(rewriter);
rewriter.createBlock(&ompParallel.getRegion());
{
OpBuilder::InsertionGuard allocaGuard(rewriter);
auto wsloopOp = rewriter.create<omp::WsloopOp>(parallelOp.getLoc());
if (!reductionVariables.empty()) {
wsloopOp.setReductionsAttr(
ArrayAttr::get(rewriter.getContext(), reductionDeclSymbols));
wsloopOp.getReductionVarsMutable().append(reductionVariables);
llvm::SmallVector<bool> byRefVec;
byRefVec.resize(reductionVariables.size(), false);
wsloopOp.setReductionVarsByref(
DenseBoolArrayAttr::get(rewriter.getContext(), byRefVec));
}
rewriter.create<omp::TerminatorOp>(loc);
llvm::SmallVector<mlir::Type> reductionTypes;
reductionTypes.reserve(reductionVariables.size());
llvm::transform(reductionVariables, std::back_inserter(reductionTypes),
[](mlir::Value v) { return v.getType(); });
rewriter.createBlock(
&wsloopOp.getRegion(), {}, reductionTypes,
llvm::SmallVector<mlir::Location>(reductionVariables.size(),
parallelOp.getLoc()));
rewriter.setInsertionPoint(
rewriter.create<omp::TerminatorOp>(parallelOp.getLoc()));
auto loopOp = rewriter.create<omp::LoopNestOp>(
parallelOp.getLoc(), parallelOp.getLowerBound(),
parallelOp.getUpperBound(), parallelOp.getStep());
rewriter.inlineRegionBefore(parallelOp.getRegion(), loopOp.getRegion(),
loopOp.getRegion().begin());
mlir::Block &loopOpEntryBlock = loopOp.getRegion().front();
unsigned numLoops = parallelOp.getNumLoops();
rewriter.replaceAllUsesWith(
loopOpEntryBlock.getArguments().drop_front(numLoops),
wsloopOp.getRegion().getArguments());
loopOpEntryBlock.eraseArguments(
numLoops, loopOpEntryBlock.getNumArguments() - numLoops);
Block *ops =
rewriter.splitBlock(&loopOpEntryBlock, loopOpEntryBlock.begin());
rewriter.setInsertionPointToStart(&loopOpEntryBlock);
auto scope = rewriter.create<memref::AllocaScopeOp>(parallelOp.getLoc(),
TypeRange());
rewriter.create<omp::YieldOp>(loc, ValueRange());
Block *scopeBlock = rewriter.createBlock(&scope.getBodyRegion());
rewriter.mergeBlocks(ops, scopeBlock);
rewriter.setInsertionPointToEnd(&*scope.getBodyRegion().begin());
rewriter.create<memref::AllocaScopeReturnOp>(loc, ValueRange());
}
}
SmallVector<Value> results;
results.reserve(reductionVariables.size());
for (auto [variable, type] :
llvm::zip(reductionVariables, parallelOp.getResultTypes())) {
Value res = rewriter.create<LLVM::LoadOp>(loc, type, variable);
results.push_back(res);
}
rewriter.replaceOp(parallelOp, results);
return success();
}
};
static LogicalResult applyPatterns(ModuleOp module, unsigned numThreads) {
ConversionTarget target(*module.getContext());
target.addIllegalOp<scf::ReduceOp, scf::ReduceReturnOp, scf::ParallelOp>();
target.addLegalDialect<omp::OpenMPDialect, LLVM::LLVMDialect,
memref::MemRefDialect>();
RewritePatternSet patterns(module.getContext());
patterns.add<ParallelOpLowering>(module.getContext(), numThreads);
FrozenRewritePatternSet frozen(std::move(patterns));
return applyPartialConversion(module, target, frozen);
}
struct SCFToOpenMPPass
: public impl::ConvertSCFToOpenMPPassBase<SCFToOpenMPPass> {
using Base::Base;
void runOnOperation() override {
if (failed(applyPatterns(getOperation(), numThreads)))
signalPassFailure();
}
};
}