xref: /llvm-project/mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp (revision be0a7e9f27083ada6072fcc0711ffa5630daa5ec)
1 //===- SCFToOpenMP.cpp - Structured Control Flow to OpenMP conversion -----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements a pass to convert scf.parallel operations into OpenMP
10 // parallel loops.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "mlir/Conversion/SCFToOpenMP/SCFToOpenMP.h"
15 #include "../PassDetail.h"
16 #include "mlir/Analysis/LoopAnalysis.h"
17 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
18 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
19 #include "mlir/Dialect/OpenMP/OpenMPDialect.h"
20 #include "mlir/Dialect/SCF/SCF.h"
21 #include "mlir/Dialect/StandardOps/IR/Ops.h"
22 #include "mlir/IR/ImplicitLocOpBuilder.h"
23 #include "mlir/IR/SymbolTable.h"
24 #include "mlir/Transforms/DialectConversion.h"
25 
26 using namespace mlir;
27 
28 /// Matches a block containing a "simple" reduction. The expected shape of the
29 /// block is as follows.
30 ///
31 ///   ^bb(%arg0, %arg1):
32 ///     %0 = OpTy(%arg0, %arg1)
33 ///     scf.reduce.return %0
34 template <typename... OpTy>
35 static bool matchSimpleReduction(Block &block) {
36   if (block.empty() || llvm::hasSingleElement(block) ||
37       std::next(block.begin(), 2) != block.end())
38     return false;
39 
40   if (block.getNumArguments() != 2)
41     return false;
42 
43   SmallVector<Operation *, 4> combinerOps;
44   Value reducedVal = matchReduction({block.getArguments()[1]},
45                                     /*redPos=*/0, combinerOps);
46 
47   if (!reducedVal || !reducedVal.isa<BlockArgument>() ||
48       combinerOps.size() != 1)
49     return false;
50 
51   return isa<OpTy...>(combinerOps[0]) &&
52          isa<scf::ReduceReturnOp>(block.back()) &&
53          block.front().getOperands() == block.getArguments();
54 }
55 
56 /// Matches a block containing a select-based min/max reduction. The types of
57 /// select and compare operations are provided as template arguments. The
58 /// comparison predicates suitable for min and max are provided as function
59 /// arguments. If a reduction is matched, `ifMin` will be set if the reduction
60 /// compute the minimum and unset if it computes the maximum, otherwise it
61 /// remains unmodified. The expected shape of the block is as follows.
62 ///
63 ///   ^bb(%arg0, %arg1):
64 ///     %0 = CompareOpTy(<one-of-predicates>, %arg0, %arg1)
65 ///     %1 = SelectOpTy(%0, %arg0, %arg1)  // %arg0, %arg1 may be swapped here.
66 ///     scf.reduce.return %1
67 template <
68     typename CompareOpTy, typename SelectOpTy,
69     typename Predicate = decltype(std::declval<CompareOpTy>().getPredicate())>
70 static bool
71 matchSelectReduction(Block &block, ArrayRef<Predicate> lessThanPredicates,
72                      ArrayRef<Predicate> greaterThanPredicates, bool &isMin) {
73   static_assert(llvm::is_one_of<SelectOpTy, SelectOp, LLVM::SelectOp>::value,
74                 "only std and llvm select ops are supported");
75 
76   // Expect exactly three operations in the block.
77   if (block.empty() || llvm::hasSingleElement(block) ||
78       std::next(block.begin(), 2) == block.end() ||
79       std::next(block.begin(), 3) != block.end())
80     return false;
81 
82   // Check op kinds.
83   auto compare = dyn_cast<CompareOpTy>(block.front());
84   auto select = dyn_cast<SelectOpTy>(block.front().getNextNode());
85   auto terminator = dyn_cast<scf::ReduceReturnOp>(block.back());
86   if (!compare || !select || !terminator)
87     return false;
88 
89   // Block arguments must be compared.
90   if (compare->getOperands() != block.getArguments())
91     return false;
92 
93   // Detect whether the comparison is less-than or greater-than, otherwise bail.
94   bool isLess;
95   if (llvm::find(lessThanPredicates, compare.getPredicate()) !=
96       lessThanPredicates.end()) {
97     isLess = true;
98   } else if (llvm::find(greaterThanPredicates, compare.getPredicate()) !=
99              greaterThanPredicates.end()) {
100     isLess = false;
101   } else {
102     return false;
103   }
104 
105   if (select.getCondition() != compare.getResult())
106     return false;
107 
108   // Detect if the operands are swapped between cmpf and select. Match the
109   // comparison type with the requested type or with the opposite of the
110   // requested type if the operands are swapped. Use generic accessors because
111   // std and LLVM versions of select have different operand names but identical
112   // positions.
113   constexpr unsigned kTrueValue = 1;
114   constexpr unsigned kFalseValue = 2;
115   bool sameOperands = select.getOperand(kTrueValue) == compare.getLhs() &&
116                       select.getOperand(kFalseValue) == compare.getRhs();
117   bool swappedOperands = select.getOperand(kTrueValue) == compare.getRhs() &&
118                          select.getOperand(kFalseValue) == compare.getLhs();
119   if (!sameOperands && !swappedOperands)
120     return false;
121 
122   if (select.getResult() != terminator.result())
123     return false;
124 
125   // The reduction is a min if it uses less-than predicates with same operands
126   // or greather-than predicates with swapped operands. Similarly for max.
127   isMin = (isLess && sameOperands) || (!isLess && swappedOperands);
128   return isMin || (isLess & swappedOperands) || (!isLess && sameOperands);
129 }
130 
131 /// Returns the float semantics for the given float type.
132 static const llvm::fltSemantics &fltSemanticsForType(FloatType type) {
133   if (type.isF16())
134     return llvm::APFloat::IEEEhalf();
135   if (type.isF32())
136     return llvm::APFloat::IEEEsingle();
137   if (type.isF64())
138     return llvm::APFloat::IEEEdouble();
139   if (type.isF128())
140     return llvm::APFloat::IEEEquad();
141   if (type.isBF16())
142     return llvm::APFloat::BFloat();
143   if (type.isF80())
144     return llvm::APFloat::x87DoubleExtended();
145   llvm_unreachable("unknown float type");
146 }
147 
148 /// Returns an attribute with the minimum (if `min` is set) or the maximum value
149 /// (otherwise) for the given float type.
150 static Attribute minMaxValueForFloat(Type type, bool min) {
151   auto fltType = type.cast<FloatType>();
152   return FloatAttr::get(
153       type, llvm::APFloat::getLargest(fltSemanticsForType(fltType), min));
154 }
155 
156 /// Returns an attribute with the signed integer minimum (if `min` is set) or
157 /// the maximum value (otherwise) for the given integer type, regardless of its
158 /// signedness semantics (only the width is considered).
159 static Attribute minMaxValueForSignedInt(Type type, bool min) {
160   auto intType = type.cast<IntegerType>();
161   unsigned bitwidth = intType.getWidth();
162   return IntegerAttr::get(type, min ? llvm::APInt::getSignedMinValue(bitwidth)
163                                     : llvm::APInt::getSignedMaxValue(bitwidth));
164 }
165 
166 /// Returns an attribute with the unsigned integer minimum (if `min` is set) or
167 /// the maximum value (otherwise) for the given integer type, regardless of its
168 /// signedness semantics (only the width is considered).
169 static Attribute minMaxValueForUnsignedInt(Type type, bool min) {
170   auto intType = type.cast<IntegerType>();
171   unsigned bitwidth = intType.getWidth();
172   return IntegerAttr::get(type, min ? llvm::APInt::getNullValue(bitwidth)
173                                     : llvm::APInt::getAllOnesValue(bitwidth));
174 }
175 
176 /// Creates an OpenMP reduction declaration and inserts it into the provided
177 /// symbol table. The declaration has a constant initializer with the neutral
178 /// value `initValue`, and the reduction combiner carried over from `reduce`.
179 static omp::ReductionDeclareOp createDecl(PatternRewriter &builder,
180                                           SymbolTable &symbolTable,
181                                           scf::ReduceOp reduce,
182                                           Attribute initValue) {
183   OpBuilder::InsertionGuard guard(builder);
184   auto decl = builder.create<omp::ReductionDeclareOp>(
185       reduce.getLoc(), "__scf_reduction", reduce.operand().getType());
186   symbolTable.insert(decl);
187 
188   Type type = reduce.operand().getType();
189   builder.createBlock(&decl.initializerRegion(), decl.initializerRegion().end(),
190                       {type});
191   builder.setInsertionPointToEnd(&decl.initializerRegion().back());
192   Value init =
193       builder.create<LLVM::ConstantOp>(reduce.getLoc(), type, initValue);
194   builder.create<omp::YieldOp>(reduce.getLoc(), init);
195 
196   Operation *terminator = &reduce.getRegion().front().back();
197   assert(isa<scf::ReduceReturnOp>(terminator) &&
198          "expected reduce op to be terminated by redure return");
199   builder.setInsertionPoint(terminator);
200   builder.replaceOpWithNewOp<omp::YieldOp>(terminator,
201                                            terminator->getOperands());
202   builder.inlineRegionBefore(reduce.getRegion(), decl.reductionRegion(),
203                              decl.reductionRegion().end());
204   return decl;
205 }
206 
207 /// Adds an atomic reduction combiner to the given OpenMP reduction declaration
208 /// using llvm.atomicrmw of the given kind.
209 static omp::ReductionDeclareOp addAtomicRMW(OpBuilder &builder,
210                                             LLVM::AtomicBinOp atomicKind,
211                                             omp::ReductionDeclareOp decl,
212                                             scf::ReduceOp reduce) {
213   OpBuilder::InsertionGuard guard(builder);
214   Type type = reduce.operand().getType();
215   Type ptrType = LLVM::LLVMPointerType::get(type);
216   builder.createBlock(&decl.atomicReductionRegion(),
217                       decl.atomicReductionRegion().end(), {ptrType, ptrType});
218   Block *atomicBlock = &decl.atomicReductionRegion().back();
219   builder.setInsertionPointToEnd(atomicBlock);
220   Value loaded = builder.create<LLVM::LoadOp>(reduce.getLoc(),
221                                               atomicBlock->getArgument(1));
222   builder.create<LLVM::AtomicRMWOp>(reduce.getLoc(), type, atomicKind,
223                                     atomicBlock->getArgument(0), loaded,
224                                     LLVM::AtomicOrdering::monotonic);
225   builder.create<omp::YieldOp>(reduce.getLoc(), ArrayRef<Value>());
226   return decl;
227 }
228 
229 /// Creates an OpenMP reduction declaration that corresponds to the given SCF
230 /// reduction and returns it. Recognizes common reductions in order to identify
231 /// the neutral value, necessary for the OpenMP declaration. If the reduction
232 /// cannot be recognized, returns null.
233 static omp::ReductionDeclareOp declareReduction(PatternRewriter &builder,
234                                                 scf::ReduceOp reduce) {
235   Operation *container = SymbolTable::getNearestSymbolTable(reduce);
236   SymbolTable symbolTable(container);
237 
238   // Insert reduction declarations in the symbol-table ancestor before the
239   // ancestor of the current insertion point.
240   Operation *insertionPoint = reduce;
241   while (insertionPoint->getParentOp() != container)
242     insertionPoint = insertionPoint->getParentOp();
243   OpBuilder::InsertionGuard guard(builder);
244   builder.setInsertionPoint(insertionPoint);
245 
246   assert(llvm::hasSingleElement(reduce.getRegion()) &&
247          "expected reduction region to have a single element");
248 
249   // Match simple binary reductions that can be expressed with atomicrmw.
250   Type type = reduce.operand().getType();
251   Block &reduction = reduce.getRegion().front();
252   if (matchSimpleReduction<arith::AddFOp, LLVM::FAddOp>(reduction)) {
253     omp::ReductionDeclareOp decl = createDecl(builder, symbolTable, reduce,
254                                               builder.getFloatAttr(type, 0.0));
255     return addAtomicRMW(builder, LLVM::AtomicBinOp::fadd, decl, reduce);
256   }
257   if (matchSimpleReduction<arith::AddIOp, LLVM::AddOp>(reduction)) {
258     omp::ReductionDeclareOp decl = createDecl(builder, symbolTable, reduce,
259                                               builder.getIntegerAttr(type, 0));
260     return addAtomicRMW(builder, LLVM::AtomicBinOp::add, decl, reduce);
261   }
262   if (matchSimpleReduction<arith::OrIOp, LLVM::OrOp>(reduction)) {
263     omp::ReductionDeclareOp decl = createDecl(builder, symbolTable, reduce,
264                                               builder.getIntegerAttr(type, 0));
265     return addAtomicRMW(builder, LLVM::AtomicBinOp::_or, decl, reduce);
266   }
267   if (matchSimpleReduction<arith::XOrIOp, LLVM::XOrOp>(reduction)) {
268     omp::ReductionDeclareOp decl = createDecl(builder, symbolTable, reduce,
269                                               builder.getIntegerAttr(type, 0));
270     return addAtomicRMW(builder, LLVM::AtomicBinOp::_xor, decl, reduce);
271   }
272   if (matchSimpleReduction<arith::AndIOp, LLVM::AndOp>(reduction)) {
273     omp::ReductionDeclareOp decl = createDecl(
274         builder, symbolTable, reduce,
275         builder.getIntegerAttr(
276             type, llvm::APInt::getAllOnesValue(type.getIntOrFloatBitWidth())));
277     return addAtomicRMW(builder, LLVM::AtomicBinOp::_and, decl, reduce);
278   }
279 
280   // Match simple binary reductions that cannot be expressed with atomicrmw.
281   // TODO: add atomic region using cmpxchg (which needs atomic load to be
282   // available as an op).
283   if (matchSimpleReduction<arith::MulFOp, LLVM::FMulOp>(reduction)) {
284     return createDecl(builder, symbolTable, reduce,
285                       builder.getFloatAttr(type, 1.0));
286   }
287 
288   // Match select-based min/max reductions.
289   bool isMin;
290   if (matchSelectReduction<arith::CmpFOp, SelectOp>(
291           reduction, {arith::CmpFPredicate::OLT, arith::CmpFPredicate::OLE},
292           {arith::CmpFPredicate::OGT, arith::CmpFPredicate::OGE}, isMin) ||
293       matchSelectReduction<LLVM::FCmpOp, LLVM::SelectOp>(
294           reduction, {LLVM::FCmpPredicate::olt, LLVM::FCmpPredicate::ole},
295           {LLVM::FCmpPredicate::ogt, LLVM::FCmpPredicate::oge}, isMin)) {
296     return createDecl(builder, symbolTable, reduce,
297                       minMaxValueForFloat(type, !isMin));
298   }
299   if (matchSelectReduction<arith::CmpIOp, SelectOp>(
300           reduction, {arith::CmpIPredicate::slt, arith::CmpIPredicate::sle},
301           {arith::CmpIPredicate::sgt, arith::CmpIPredicate::sge}, isMin) ||
302       matchSelectReduction<LLVM::ICmpOp, LLVM::SelectOp>(
303           reduction, {LLVM::ICmpPredicate::slt, LLVM::ICmpPredicate::sle},
304           {LLVM::ICmpPredicate::sgt, LLVM::ICmpPredicate::sge}, isMin)) {
305     omp::ReductionDeclareOp decl = createDecl(
306         builder, symbolTable, reduce, minMaxValueForSignedInt(type, !isMin));
307     return addAtomicRMW(builder,
308                         isMin ? LLVM::AtomicBinOp::min : LLVM::AtomicBinOp::max,
309                         decl, reduce);
310   }
311   if (matchSelectReduction<arith::CmpIOp, SelectOp>(
312           reduction, {arith::CmpIPredicate::ult, arith::CmpIPredicate::ule},
313           {arith::CmpIPredicate::ugt, arith::CmpIPredicate::uge}, isMin) ||
314       matchSelectReduction<LLVM::ICmpOp, LLVM::SelectOp>(
315           reduction, {LLVM::ICmpPredicate::ugt, LLVM::ICmpPredicate::ule},
316           {LLVM::ICmpPredicate::ugt, LLVM::ICmpPredicate::uge}, isMin)) {
317     omp::ReductionDeclareOp decl = createDecl(
318         builder, symbolTable, reduce, minMaxValueForUnsignedInt(type, !isMin));
319     return addAtomicRMW(
320         builder, isMin ? LLVM::AtomicBinOp::umin : LLVM::AtomicBinOp::umax,
321         decl, reduce);
322   }
323 
324   return nullptr;
325 }
326 
327 namespace {
328 
329 struct ParallelOpLowering : public OpRewritePattern<scf::ParallelOp> {
330   using OpRewritePattern<scf::ParallelOp>::OpRewritePattern;
331 
332   LogicalResult matchAndRewrite(scf::ParallelOp parallelOp,
333                                 PatternRewriter &rewriter) const override {
334     // Replace SCF yield with OpenMP yield.
335     {
336       OpBuilder::InsertionGuard guard(rewriter);
337       rewriter.setInsertionPointToEnd(parallelOp.getBody());
338       assert(llvm::hasSingleElement(parallelOp.region()) &&
339              "expected scf.parallel to have one block");
340       rewriter.replaceOpWithNewOp<omp::YieldOp>(
341           parallelOp.getBody()->getTerminator(), ValueRange());
342     }
343 
344     // Declare reductions.
345     // TODO: consider checking it here is already a compatible reduction
346     // declaration and use it instead of redeclaring.
347     SmallVector<Attribute> reductionDeclSymbols;
348     for (auto reduce : parallelOp.getOps<scf::ReduceOp>()) {
349       omp::ReductionDeclareOp decl = declareReduction(rewriter, reduce);
350       if (!decl)
351         return failure();
352       reductionDeclSymbols.push_back(
353           SymbolRefAttr::get(rewriter.getContext(), decl.sym_name()));
354     }
355 
356     // Allocate reduction variables. Make sure the we don't overflow the stack
357     // with local `alloca`s by saving and restoring the stack pointer.
358     Location loc = parallelOp.getLoc();
359     Value one = rewriter.create<LLVM::ConstantOp>(
360         loc, rewriter.getIntegerType(64), rewriter.getI64IntegerAttr(1));
361     SmallVector<Value> reductionVariables;
362     reductionVariables.reserve(parallelOp.getNumReductions());
363     Value token = rewriter.create<LLVM::StackSaveOp>(
364         loc, LLVM::LLVMPointerType::get(rewriter.getIntegerType(8)));
365     for (Value init : parallelOp.initVals()) {
366       assert((LLVM::isCompatibleType(init.getType()) ||
367               init.getType().isa<LLVM::PointerElementTypeInterface>()) &&
368              "cannot create a reduction variable if the type is not an LLVM "
369              "pointer element");
370       Value storage = rewriter.create<LLVM::AllocaOp>(
371           loc, LLVM::LLVMPointerType::get(init.getType()), one, 0);
372       rewriter.create<LLVM::StoreOp>(loc, init, storage);
373       reductionVariables.push_back(storage);
374     }
375 
376     // Replace the reduction operations contained in this loop. Must be done
377     // here rather than in a separate pattern to have access to the list of
378     // reduction variables.
379     for (auto pair :
380          llvm::zip(parallelOp.getOps<scf::ReduceOp>(), reductionVariables)) {
381       OpBuilder::InsertionGuard guard(rewriter);
382       scf::ReduceOp reduceOp = std::get<0>(pair);
383       rewriter.setInsertionPoint(reduceOp);
384       rewriter.replaceOpWithNewOp<omp::ReductionOp>(
385           reduceOp, reduceOp.operand(), std::get<1>(pair));
386     }
387 
388     // Create the parallel wrapper.
389     auto ompParallel = rewriter.create<omp::ParallelOp>(loc);
390     {
391       OpBuilder::InsertionGuard guard(rewriter);
392       rewriter.createBlock(&ompParallel.region());
393 
394       // Replace SCF yield with OpenMP yield.
395       {
396         OpBuilder::InsertionGuard innerGuard(rewriter);
397         rewriter.setInsertionPointToEnd(parallelOp.getBody());
398         assert(llvm::hasSingleElement(parallelOp.region()) &&
399                "expected scf.parallel to have one block");
400         rewriter.replaceOpWithNewOp<omp::YieldOp>(
401             parallelOp.getBody()->getTerminator(), ValueRange());
402       }
403 
404       // Replace the loop.
405       auto loop = rewriter.create<omp::WsLoopOp>(
406           parallelOp.getLoc(), parallelOp.lowerBound(), parallelOp.upperBound(),
407           parallelOp.step());
408       rewriter.create<omp::TerminatorOp>(loc);
409 
410       rewriter.inlineRegionBefore(parallelOp.region(), loop.region(),
411                                   loop.region().begin());
412       if (!reductionVariables.empty()) {
413         loop.reductionsAttr(
414             ArrayAttr::get(rewriter.getContext(), reductionDeclSymbols));
415         loop.reduction_varsMutable().append(reductionVariables);
416       }
417     }
418 
419     // Load loop results.
420     SmallVector<Value> results;
421     results.reserve(reductionVariables.size());
422     for (Value variable : reductionVariables) {
423       Value res = rewriter.create<LLVM::LoadOp>(loc, variable);
424       results.push_back(res);
425     }
426     rewriter.replaceOp(parallelOp, results);
427 
428     rewriter.create<LLVM::StackRestoreOp>(loc, token);
429     return success();
430   }
431 };
432 
433 /// Applies the conversion patterns in the given function.
434 static LogicalResult applyPatterns(ModuleOp module) {
435   ConversionTarget target(*module.getContext());
436   target.addIllegalOp<scf::ReduceOp, scf::ReduceReturnOp, scf::ParallelOp>();
437   target.addLegalDialect<omp::OpenMPDialect, LLVM::LLVMDialect>();
438 
439   RewritePatternSet patterns(module.getContext());
440   patterns.add<ParallelOpLowering>(module.getContext());
441   FrozenRewritePatternSet frozen(std::move(patterns));
442   return applyPartialConversion(module, target, frozen);
443 }
444 
445 /// A pass converting SCF operations to OpenMP operations.
446 struct SCFToOpenMPPass : public ConvertSCFToOpenMPBase<SCFToOpenMPPass> {
447   /// Pass entry point.
448   void runOnOperation() override {
449     if (failed(applyPatterns(getOperation())))
450       signalPassFailure();
451   }
452 };
453 
454 } // namespace
455 
456 std::unique_ptr<OperationPass<ModuleOp>> mlir::createConvertSCFToOpenMPPass() {
457   return std::make_unique<SCFToOpenMPPass>();
458 }
459