| 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 | |
| 16 | #include "mlir/Analysis/SliceAnalysis.h" |
| 17 | #include "mlir/Dialect/Affine/Analysis/LoopAnalysis.h" |
| 18 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 19 | #include "mlir/Dialect/LLVMIR/LLVMArithOps.h" |
| 20 | #include "mlir/Dialect/LLVMIR/LLVMConstantOps.h" |
| 21 | #include "mlir/Dialect/LLVMIR/LLVMMemoryOps.h" |
| 22 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| 23 | #include "mlir/Dialect/OpenMP/OpenMPDialect.h" |
| 24 | #include "mlir/Dialect/SCF/IR/SCF.h" |
| 25 | #include "mlir/IR/SymbolTable.h" |
| 26 | #include "mlir/IR/TypeUtilities.h" |
| 27 | #include "mlir/Pass/Pass.h" |
| 28 | #include "mlir/Transforms/WalkPatternRewriteDriver.h" |
| 29 | |
| 30 | namespace mlir { |
| 31 | #define GEN_PASS_DEF_CONVERTSCFTOOPENMPPASS |
| 32 | #include "mlir/Conversion/Passes.h.inc" |
| 33 | } // namespace mlir |
| 34 | |
| 35 | using namespace mlir; |
| 36 | |
| 37 | /// Matches a block containing a "simple" reduction. The expected shape of the |
| 38 | /// block is as follows. |
| 39 | /// |
| 40 | /// ^bb(%arg0, %arg1): |
| 41 | /// %0 = OpTy(%arg0, %arg1) |
| 42 | /// scf.reduce.return %0 |
| 43 | template <typename... OpTy> |
| 44 | static bool matchSimpleReduction(Block &block) { |
| 45 | if (block.empty() || llvm::hasSingleElement(C&: block) || |
| 46 | std::next(x: block.begin(), n: 2) != block.end()) |
| 47 | return false; |
| 48 | |
| 49 | if (block.getNumArguments() != 2) |
| 50 | return false; |
| 51 | |
| 52 | SmallVector<Operation *, 4> combinerOps; |
| 53 | Value reducedVal = matchReduction(iterCarriedArgs: {block.getArguments()[1]}, |
| 54 | /*redPos=*/0, combinerOps); |
| 55 | |
| 56 | if (!reducedVal || !isa<BlockArgument>(Val: reducedVal) || combinerOps.size() != 1) |
| 57 | return false; |
| 58 | |
| 59 | return isa<OpTy...>(combinerOps[0]) && |
| 60 | isa<scf::ReduceReturnOp>(Val: block.back()) && |
| 61 | block.front().getOperands() == block.getArguments(); |
| 62 | } |
| 63 | |
| 64 | /// Matches a block containing a select-based min/max reduction. The types of |
| 65 | /// select and compare operations are provided as template arguments. The |
| 66 | /// comparison predicates suitable for min and max are provided as function |
| 67 | /// arguments. If a reduction is matched, `ifMin` will be set if the reduction |
| 68 | /// compute the minimum and unset if it computes the maximum, otherwise it |
| 69 | /// remains unmodified. The expected shape of the block is as follows. |
| 70 | /// |
| 71 | /// ^bb(%arg0, %arg1): |
| 72 | /// %0 = CompareOpTy(<one-of-predicates>, %arg0, %arg1) |
| 73 | /// %1 = SelectOpTy(%0, %arg0, %arg1) // %arg0, %arg1 may be swapped here. |
| 74 | /// scf.reduce.return %1 |
| 75 | template < |
| 76 | typename CompareOpTy, typename SelectOpTy, |
| 77 | typename Predicate = decltype(std::declval<CompareOpTy>().getPredicate())> |
| 78 | static bool |
| 79 | matchSelectReduction(Block &block, ArrayRef<Predicate> lessThanPredicates, |
| 80 | ArrayRef<Predicate> greaterThanPredicates, bool &isMin) { |
| 81 | static_assert( |
| 82 | llvm::is_one_of<SelectOpTy, arith::SelectOp, LLVM::SelectOp>::value, |
| 83 | "only arithmetic and llvm select ops are supported" ); |
| 84 | |
| 85 | // Expect exactly three operations in the block. |
| 86 | if (block.empty() || llvm::hasSingleElement(C&: block) || |
| 87 | std::next(x: block.begin(), n: 2) == block.end() || |
| 88 | std::next(x: block.begin(), n: 3) != block.end()) |
| 89 | return false; |
| 90 | |
| 91 | // Check op kinds. |
| 92 | auto compare = dyn_cast<CompareOpTy>(block.front()); |
| 93 | auto select = dyn_cast<SelectOpTy>(block.front().getNextNode()); |
| 94 | auto terminator = dyn_cast<scf::ReduceReturnOp>(Val&: block.back()); |
| 95 | if (!compare || !select || !terminator) |
| 96 | return false; |
| 97 | |
| 98 | // Block arguments must be compared. |
| 99 | if (compare->getOperands() != block.getArguments()) |
| 100 | return false; |
| 101 | |
| 102 | // Detect whether the comparison is less-than or greater-than, otherwise bail. |
| 103 | bool isLess; |
| 104 | if (llvm::is_contained(lessThanPredicates, compare.getPredicate())) { |
| 105 | isLess = true; |
| 106 | } else if (llvm::is_contained(greaterThanPredicates, |
| 107 | compare.getPredicate())) { |
| 108 | isLess = false; |
| 109 | } else { |
| 110 | return false; |
| 111 | } |
| 112 | |
| 113 | if (select.getCondition() != compare.getResult()) |
| 114 | return false; |
| 115 | |
| 116 | // Detect if the operands are swapped between cmpf and select. Match the |
| 117 | // comparison type with the requested type or with the opposite of the |
| 118 | // requested type if the operands are swapped. Use generic accessors because |
| 119 | // std and LLVM versions of select have different operand names but identical |
| 120 | // positions. |
| 121 | constexpr unsigned kTrueValue = 1; |
| 122 | constexpr unsigned kFalseValue = 2; |
| 123 | bool sameOperands = select.getOperand(kTrueValue) == compare.getLhs() && |
| 124 | select.getOperand(kFalseValue) == compare.getRhs(); |
| 125 | bool swappedOperands = select.getOperand(kTrueValue) == compare.getRhs() && |
| 126 | select.getOperand(kFalseValue) == compare.getLhs(); |
| 127 | if (!sameOperands && !swappedOperands) |
| 128 | return false; |
| 129 | |
| 130 | if (select.getResult() != terminator.getResult()) |
| 131 | return false; |
| 132 | |
| 133 | // The reduction is a min if it uses less-than predicates with same operands |
| 134 | // or greather-than predicates with swapped operands. Similarly for max. |
| 135 | isMin = (isLess && sameOperands) || (!isLess && swappedOperands); |
| 136 | return isMin || (isLess & swappedOperands) || (!isLess && sameOperands); |
| 137 | } |
| 138 | |
| 139 | /// Returns the float semantics for the given float type. |
| 140 | static const llvm::fltSemantics &fltSemanticsForType(FloatType type) { |
| 141 | if (type.isF16()) |
| 142 | return llvm::APFloat::IEEEhalf(); |
| 143 | if (type.isF32()) |
| 144 | return llvm::APFloat::IEEEsingle(); |
| 145 | if (type.isF64()) |
| 146 | return llvm::APFloat::IEEEdouble(); |
| 147 | if (type.isF128()) |
| 148 | return llvm::APFloat::IEEEquad(); |
| 149 | if (type.isBF16()) |
| 150 | return llvm::APFloat::BFloat(); |
| 151 | if (type.isF80()) |
| 152 | return llvm::APFloat::x87DoubleExtended(); |
| 153 | llvm_unreachable("unknown float type" ); |
| 154 | } |
| 155 | |
| 156 | /// Helper to create a splat attribute for vector types, or return the scalar |
| 157 | /// attribute for scalar types. |
| 158 | static Attribute getSplatOrScalarAttr(Type type, Attribute val) { |
| 159 | if (auto vecType = dyn_cast<VectorType>(Val&: type)) |
| 160 | return DenseElementsAttr::get(type: vecType, values: val); |
| 161 | return val; |
| 162 | } |
| 163 | |
| 164 | /// Returns an attribute with the minimum (if `min` is set) or the maximum value |
| 165 | /// (otherwise) for the given float type. |
| 166 | static Attribute minMaxValueForFloat(Type type, bool min) { |
| 167 | Type elType = getElementTypeOrSelf(type); |
| 168 | auto fltType = cast<FloatType>(Val&: elType); |
| 169 | auto val = llvm::APFloat::getLargest(Sem: fltSemanticsForType(type: fltType), Negative: min); |
| 170 | |
| 171 | return getSplatOrScalarAttr(type, val: FloatAttr::get(type: elType, value: val)); |
| 172 | } |
| 173 | |
| 174 | /// Returns an attribute with the signed integer minimum (if `min` is set) or |
| 175 | /// the maximum value (otherwise) for the given integer type, regardless of its |
| 176 | /// signedness semantics (only the width is considered). |
| 177 | static Attribute minMaxValueForSignedInt(Type type, bool min) { |
| 178 | Type elType = getElementTypeOrSelf(type); |
| 179 | auto intType = cast<IntegerType>(Val&: elType); |
| 180 | unsigned bitwidth = intType.getWidth(); |
| 181 | auto val = min ? llvm::APInt::getSignedMinValue(numBits: bitwidth) |
| 182 | : llvm::APInt::getSignedMaxValue(numBits: bitwidth); |
| 183 | |
| 184 | return getSplatOrScalarAttr(type, val: IntegerAttr::get(type: elType, value: val)); |
| 185 | } |
| 186 | |
| 187 | /// Returns an attribute with the unsigned integer minimum (if `min` is set) or |
| 188 | /// the maximum value (otherwise) for the given integer type, regardless of its |
| 189 | /// signedness semantics (only the width is considered). |
| 190 | static Attribute minMaxValueForUnsignedInt(Type type, bool min) { |
| 191 | Type elType = getElementTypeOrSelf(type); |
| 192 | auto intType = cast<IntegerType>(Val&: elType); |
| 193 | unsigned bitwidth = intType.getWidth(); |
| 194 | auto val = |
| 195 | min ? llvm::APInt::getZero(numBits: bitwidth) : llvm::APInt::getAllOnes(numBits: bitwidth); |
| 196 | |
| 197 | return getSplatOrScalarAttr(type, val: IntegerAttr::get(type: elType, value: val)); |
| 198 | } |
| 199 | |
| 200 | /// Creates an OpenMP reduction declaration and inserts it into the provided |
| 201 | /// symbol table. The declaration has a constant initializer with the neutral |
| 202 | /// value `initValue`, and the `reductionIndex`-th reduction combiner carried |
| 203 | /// over from `reduce`. |
| 204 | static omp::DeclareReductionOp |
| 205 | createDecl(PatternRewriter &builder, SymbolTable &symbolTable, |
| 206 | scf::ReduceOp reduce, int64_t reductionIndex, Attribute initValue) { |
| 207 | OpBuilder::InsertionGuard guard(builder); |
| 208 | Type type = reduce.getOperands()[reductionIndex].getType(); |
| 209 | auto decl = omp::DeclareReductionOp::create(builder, location: reduce.getLoc(), |
| 210 | sym_name: "__scf_reduction" , |
| 211 | /*sym_visibility=*/nullptr, type, |
| 212 | /*byref_element_type=*/{}); |
| 213 | symbolTable.insert(symbol: decl); |
| 214 | |
| 215 | builder.createBlock(parent: &decl.getInitializerRegion(), |
| 216 | insertPt: decl.getInitializerRegion().end(), argTypes: {type}, |
| 217 | locs: {reduce.getOperands()[reductionIndex].getLoc()}); |
| 218 | builder.setInsertionPointToEnd(&decl.getInitializerRegion().back()); |
| 219 | Value init = |
| 220 | LLVM::ConstantOp::create(builder, location: reduce.getLoc(), res: type, value: initValue); |
| 221 | omp::YieldOp::create(builder, location: reduce.getLoc(), results: init); |
| 222 | |
| 223 | Operation *terminator = |
| 224 | &reduce.getReductions()[reductionIndex].front().back(); |
| 225 | assert(isa<scf::ReduceReturnOp>(terminator) && |
| 226 | "expected reduce op to be terminated by reduce return" ); |
| 227 | builder.setInsertionPoint(terminator); |
| 228 | builder.replaceOpWithNewOp<omp::YieldOp>(op: terminator, |
| 229 | args: terminator->getOperands()); |
| 230 | builder.inlineRegionBefore(region&: reduce.getReductions()[reductionIndex], |
| 231 | parent&: decl.getReductionRegion(), |
| 232 | before: decl.getReductionRegion().end()); |
| 233 | return decl; |
| 234 | } |
| 235 | |
| 236 | /// Adds an atomic reduction combiner to the given OpenMP reduction declaration |
| 237 | /// using llvm.atomicrmw of the given kind. |
| 238 | static omp::DeclareReductionOp addAtomicRMW(OpBuilder &builder, |
| 239 | LLVM::AtomicBinOp atomicKind, |
| 240 | omp::DeclareReductionOp decl, |
| 241 | scf::ReduceOp reduce, |
| 242 | int64_t reductionIndex) { |
| 243 | OpBuilder::InsertionGuard guard(builder); |
| 244 | auto ptrType = LLVM::LLVMPointerType::get(context: builder.getContext()); |
| 245 | Location reduceOperandLoc = reduce.getOperands()[reductionIndex].getLoc(); |
| 246 | builder.createBlock(parent: &decl.getAtomicReductionRegion(), |
| 247 | insertPt: decl.getAtomicReductionRegion().end(), argTypes: {ptrType, ptrType}, |
| 248 | locs: {reduceOperandLoc, reduceOperandLoc}); |
| 249 | Block *atomicBlock = &decl.getAtomicReductionRegion().back(); |
| 250 | builder.setInsertionPointToEnd(atomicBlock); |
| 251 | Value loaded = LLVM::LoadOp::create(builder, location: reduce.getLoc(), type: decl.getType(), |
| 252 | addr: atomicBlock->getArgument(i: 1)); |
| 253 | LLVM::AtomicRMWOp::create(builder, location: reduce.getLoc(), binOp: atomicKind, |
| 254 | ptr: atomicBlock->getArgument(i: 0), val: loaded, |
| 255 | ordering: LLVM::AtomicOrdering::monotonic); |
| 256 | omp::YieldOp::create(builder, location: reduce.getLoc(), results: ArrayRef<Value>()); |
| 257 | return decl; |
| 258 | } |
| 259 | |
| 260 | /// Returns true if the type is supported by llvm.atomicrmw. |
| 261 | /// LLVM IR currently does not support atomic operations on vector types. |
| 262 | /// See LLVM Language Reference Manual on 'atomicrmw'. |
| 263 | static bool supportsAtomic(Type type) { return !isa<VectorType>(Val: type); } |
| 264 | |
| 265 | /// Creates an OpenMP reduction declaration that corresponds to the given SCF |
| 266 | /// reduction and returns it. Recognizes common reductions in order to identify |
| 267 | /// the neutral value, necessary for the OpenMP declaration. If the reduction |
| 268 | /// cannot be recognized, returns null. |
| 269 | static omp::DeclareReductionOp declareReduction(PatternRewriter &builder, |
| 270 | scf::ReduceOp reduce, |
| 271 | int64_t reductionIndex) { |
| 272 | Operation *container = SymbolTable::getNearestSymbolTable(from: reduce); |
| 273 | SymbolTable symbolTable(container); |
| 274 | |
| 275 | // Insert reduction declarations in the symbol-table ancestor before the |
| 276 | // ancestor of the current insertion point. |
| 277 | Operation *insertionPoint = reduce; |
| 278 | while (insertionPoint->getParentOp() != container) |
| 279 | insertionPoint = insertionPoint->getParentOp(); |
| 280 | OpBuilder::InsertionGuard guard(builder); |
| 281 | builder.setInsertionPoint(insertionPoint); |
| 282 | |
| 283 | assert(llvm::hasSingleElement(reduce.getReductions()[reductionIndex]) && |
| 284 | "expected reduction region to have a single element" ); |
| 285 | |
| 286 | // Match simple binary reductions that can be expressed with atomicrmw. |
| 287 | Type type = reduce.getOperands()[reductionIndex].getType(); |
| 288 | Block &reduction = reduce.getReductions()[reductionIndex].front(); |
| 289 | |
| 290 | // Handle scalar element type extraction for vector bitwidth safety. |
| 291 | Type elType = getElementTypeOrSelf(type); |
| 292 | |
| 293 | // Arithmetic Reductions |
| 294 | if (matchSimpleReduction<arith::AddFOp, LLVM::FAddOp>(block&: reduction)) { |
| 295 | omp::DeclareReductionOp decl = createDecl( |
| 296 | builder, symbolTable, reduce, reductionIndex, |
| 297 | initValue: getSplatOrScalarAttr(type, val: builder.getFloatAttr(type: elType, value: 0.0))); |
| 298 | return supportsAtomic(type) ? addAtomicRMW(builder, atomicKind: LLVM::AtomicBinOp::fadd, |
| 299 | decl, reduce, reductionIndex) |
| 300 | : decl; |
| 301 | } |
| 302 | if (matchSimpleReduction<arith::AddIOp, LLVM::AddOp>(block&: reduction)) { |
| 303 | omp::DeclareReductionOp decl = createDecl( |
| 304 | builder, symbolTable, reduce, reductionIndex, |
| 305 | initValue: getSplatOrScalarAttr(type, val: builder.getIntegerAttr(type: elType, value: 0))); |
| 306 | return supportsAtomic(type) ? addAtomicRMW(builder, atomicKind: LLVM::AtomicBinOp::add, |
| 307 | decl, reduce, reductionIndex) |
| 308 | : decl; |
| 309 | } |
| 310 | if (matchSimpleReduction<arith::OrIOp, LLVM::OrOp>(block&: reduction)) { |
| 311 | omp::DeclareReductionOp decl = createDecl( |
| 312 | builder, symbolTable, reduce, reductionIndex, |
| 313 | initValue: getSplatOrScalarAttr(type, val: builder.getIntegerAttr(type: elType, value: 0))); |
| 314 | return supportsAtomic(type) ? addAtomicRMW(builder, atomicKind: LLVM::AtomicBinOp::_or, |
| 315 | decl, reduce, reductionIndex) |
| 316 | : decl; |
| 317 | } |
| 318 | if (matchSimpleReduction<arith::XOrIOp, LLVM::XOrOp>(block&: reduction)) { |
| 319 | omp::DeclareReductionOp decl = createDecl( |
| 320 | builder, symbolTable, reduce, reductionIndex, |
| 321 | initValue: getSplatOrScalarAttr(type, val: builder.getIntegerAttr(type: elType, value: 0))); |
| 322 | return supportsAtomic(type) ? addAtomicRMW(builder, atomicKind: LLVM::AtomicBinOp::_xor, |
| 323 | decl, reduce, reductionIndex) |
| 324 | : decl; |
| 325 | } |
| 326 | if (matchSimpleReduction<arith::AndIOp, LLVM::AndOp>(block&: reduction)) { |
| 327 | APInt allOnes = llvm::APInt::getAllOnes(numBits: elType.getIntOrFloatBitWidth()); |
| 328 | omp::DeclareReductionOp decl = createDecl( |
| 329 | builder, symbolTable, reduce, reductionIndex, |
| 330 | initValue: getSplatOrScalarAttr(type, val: builder.getIntegerAttr(type: elType, value: allOnes))); |
| 331 | return supportsAtomic(type) ? addAtomicRMW(builder, atomicKind: LLVM::AtomicBinOp::_and, |
| 332 | decl, reduce, reductionIndex) |
| 333 | : decl; |
| 334 | } |
| 335 | |
| 336 | // Match simple binary reductions that cannot be expressed with atomicrmw. |
| 337 | // TODO: add atomic region using cmpxchg (which needs atomic load to be |
| 338 | // available as an op). |
| 339 | if (matchSimpleReduction<arith::MulFOp, LLVM::FMulOp>(block&: reduction)) { |
| 340 | return createDecl( |
| 341 | builder, symbolTable, reduce, reductionIndex, |
| 342 | initValue: getSplatOrScalarAttr(type, val: builder.getFloatAttr(type: elType, value: 1.0))); |
| 343 | } |
| 344 | |
| 345 | if (matchSimpleReduction<arith::MulIOp, LLVM::MulOp>(block&: reduction)) { |
| 346 | return createDecl( |
| 347 | builder, symbolTable, reduce, reductionIndex, |
| 348 | initValue: getSplatOrScalarAttr(type, val: builder.getIntegerAttr(type: elType, value: 1))); |
| 349 | } |
| 350 | |
| 351 | // Match select-based min/max reductions. |
| 352 | bool isMin; |
| 353 | // Floating Point Min/Max |
| 354 | if (matchSelectReduction<arith::CmpFOp, arith::SelectOp, |
| 355 | arith::CmpFPredicate>( |
| 356 | block&: reduction, lessThanPredicates: {arith::CmpFPredicate::OLT, arith::CmpFPredicate::OLE}, |
| 357 | greaterThanPredicates: {arith::CmpFPredicate::OGT, arith::CmpFPredicate::OGE}, isMin) || |
| 358 | matchSelectReduction<arith::CmpFOp, arith::SelectOp, |
| 359 | arith::CmpFPredicate>( |
| 360 | block&: reduction, lessThanPredicates: {arith::CmpFPredicate::OGT, arith::CmpFPredicate::OGE}, |
| 361 | greaterThanPredicates: {arith::CmpFPredicate::OLT, arith::CmpFPredicate::OLE}, isMin)) { |
| 362 | return createDecl(builder, symbolTable, reduce, reductionIndex, |
| 363 | initValue: minMaxValueForFloat(type, min: !isMin)); |
| 364 | } |
| 365 | |
| 366 | // Integer Min/Max |
| 367 | if (matchSelectReduction<arith::CmpIOp, arith::SelectOp, |
| 368 | arith::CmpIPredicate>( |
| 369 | block&: reduction, lessThanPredicates: {arith::CmpIPredicate::slt, arith::CmpIPredicate::sle}, |
| 370 | greaterThanPredicates: {arith::CmpIPredicate::sgt, arith::CmpIPredicate::sge}, isMin) || |
| 371 | matchSelectReduction<arith::CmpIOp, arith::SelectOp, |
| 372 | arith::CmpIPredicate>( |
| 373 | block&: reduction, lessThanPredicates: {arith::CmpIPredicate::sgt, arith::CmpIPredicate::sge}, |
| 374 | greaterThanPredicates: {arith::CmpIPredicate::slt, arith::CmpIPredicate::sle}, isMin)) { |
| 375 | omp::DeclareReductionOp decl = |
| 376 | createDecl(builder, symbolTable, reduce, reductionIndex, |
| 377 | initValue: minMaxValueForSignedInt(type, min: !isMin)); |
| 378 | return supportsAtomic(type) ? addAtomicRMW(builder, |
| 379 | atomicKind: isMin ? LLVM::AtomicBinOp::min |
| 380 | : LLVM::AtomicBinOp::max, |
| 381 | decl, reduce, reductionIndex) |
| 382 | : decl; |
| 383 | } |
| 384 | |
| 385 | // Unsigned Integer Min/Max |
| 386 | if (matchSelectReduction<arith::CmpIOp, arith::SelectOp, |
| 387 | arith::CmpIPredicate>( |
| 388 | block&: reduction, lessThanPredicates: {arith::CmpIPredicate::ult, arith::CmpIPredicate::ule}, |
| 389 | greaterThanPredicates: {arith::CmpIPredicate::ugt, arith::CmpIPredicate::uge}, isMin) || |
| 390 | matchSelectReduction<arith::CmpIOp, arith::SelectOp, |
| 391 | arith::CmpIPredicate>( |
| 392 | block&: reduction, lessThanPredicates: {arith::CmpIPredicate::ugt, arith::CmpIPredicate::uge}, |
| 393 | greaterThanPredicates: {arith::CmpIPredicate::ult, arith::CmpIPredicate::ule}, isMin)) { |
| 394 | omp::DeclareReductionOp decl = |
| 395 | createDecl(builder, symbolTable, reduce, reductionIndex, |
| 396 | initValue: minMaxValueForUnsignedInt(type, min: !isMin)); |
| 397 | return supportsAtomic(type) ? addAtomicRMW(builder, |
| 398 | atomicKind: isMin ? LLVM::AtomicBinOp::umin |
| 399 | : LLVM::AtomicBinOp::umax, |
| 400 | decl, reduce, reductionIndex) |
| 401 | : decl; |
| 402 | } |
| 403 | |
| 404 | return nullptr; |
| 405 | } |
| 406 | |
| 407 | namespace { |
| 408 | |
| 409 | struct ParallelOpLowering : public OpRewritePattern<scf::ParallelOp> { |
| 410 | static constexpr unsigned kUseOpenMPDefaultNumThreads = 0; |
| 411 | unsigned numThreads; |
| 412 | |
| 413 | ParallelOpLowering(MLIRContext *context, |
| 414 | unsigned numThreads = kUseOpenMPDefaultNumThreads) |
| 415 | : OpRewritePattern<scf::ParallelOp>(context), numThreads(numThreads) {} |
| 416 | |
| 417 | LogicalResult matchAndRewrite(scf::ParallelOp parallelOp, |
| 418 | PatternRewriter &rewriter) const override { |
| 419 | if (parallelOp.getUnsignedCmp()) |
| 420 | return rewriter.notifyMatchFailure( |
| 421 | arg&: parallelOp, msg: "unsigned loop bounds are not supported" ); |
| 422 | // Bail out early if any reduction init value has a type that is not |
| 423 | // compatible with LLVM (e.g. index), since we cannot allocate a reduction |
| 424 | // variable for such types. |
| 425 | for (Value init : parallelOp.getInitVals()) { |
| 426 | if (!LLVM::isCompatibleType(type: init.getType()) && |
| 427 | !isa<LLVM::PointerElementTypeInterface>(Val: init.getType())) |
| 428 | return rewriter.notifyMatchFailure( |
| 429 | arg&: parallelOp, msg: "reduction init type is not an LLVM-compatible type" ); |
| 430 | } |
| 431 | |
| 432 | // Declare reductions. |
| 433 | // TODO: consider checking it here is already a compatible reduction |
| 434 | // declaration and use it instead of redeclaring. |
| 435 | SmallVector<Attribute> reductionSyms; |
| 436 | SmallVector<omp::DeclareReductionOp> ompReductionDecls; |
| 437 | auto reduce = cast<scf::ReduceOp>(Val: parallelOp.getBody()->getTerminator()); |
| 438 | for (int64_t i = 0, e = parallelOp.getNumReductions(); i < e; ++i) { |
| 439 | omp::DeclareReductionOp decl = declareReduction(builder&: rewriter, reduce, reductionIndex: i); |
| 440 | ompReductionDecls.push_back(Elt: decl); |
| 441 | if (!decl) |
| 442 | return failure(); |
| 443 | reductionSyms.push_back( |
| 444 | Elt: SymbolRefAttr::get(ctx: rewriter.getContext(), value: decl.getSymName())); |
| 445 | } |
| 446 | |
| 447 | // Allocate reduction variables. Make sure the we don't overflow the stack |
| 448 | // with local `alloca`s by saving and restoring the stack pointer. |
| 449 | Location loc = parallelOp.getLoc(); |
| 450 | Value one = |
| 451 | LLVM::ConstantOp::create(builder&: rewriter, location: loc, res: rewriter.getIntegerType(width: 64), |
| 452 | value: rewriter.getI64IntegerAttr(value: 1)); |
| 453 | SmallVector<Value> reductionVariables; |
| 454 | reductionVariables.reserve(N: parallelOp.getNumReductions()); |
| 455 | auto ptrType = LLVM::LLVMPointerType::get(context: parallelOp.getContext()); |
| 456 | for (Value init : parallelOp.getInitVals()) { |
| 457 | Value storage = LLVM::AllocaOp::create(builder&: rewriter, location: loc, resultType: ptrType, |
| 458 | elementType: init.getType(), arraySize: one, alignment: 0); |
| 459 | LLVM::StoreOp::create(builder&: rewriter, location: loc, value: init, addr: storage); |
| 460 | reductionVariables.push_back(Elt: storage); |
| 461 | } |
| 462 | |
| 463 | // Replace the reduction operations contained in this loop. Must be done |
| 464 | // here rather than in a separate pattern to have access to the list of |
| 465 | // reduction variables. |
| 466 | for (auto [x, y, rD] : llvm::zip_equal( |
| 467 | t&: reductionVariables, u: reduce.getOperands(), args&: ompReductionDecls)) { |
| 468 | OpBuilder::InsertionGuard guard(rewriter); |
| 469 | rewriter.setInsertionPoint(reduce); |
| 470 | Region &redRegion = rD.getReductionRegion(); |
| 471 | // The SCF dialect by definition contains only structured operations |
| 472 | // and hence the SCF reduction region will contain a single block. |
| 473 | // The ompReductionDecls region is a copy of the SCF reduction region |
| 474 | // and hence has the same property. |
| 475 | assert(redRegion.hasOneBlock() && |
| 476 | "expect reduction region to have one block" ); |
| 477 | Value pvtRedVar = parallelOp.getRegion().addArgument(type: x.getType(), loc); |
| 478 | Value pvtRedVal = LLVM::LoadOp::create(builder&: rewriter, location: reduce.getLoc(), |
| 479 | type: rD.getType(), addr: pvtRedVar); |
| 480 | // Make a copy of the reduction combiner region in the body |
| 481 | mlir::OpBuilder builder(rewriter.getContext()); |
| 482 | builder.setInsertionPoint(reduce); |
| 483 | mlir::IRMapping mapper; |
| 484 | assert(redRegion.getNumArguments() == 2 && |
| 485 | "expect reduction region to have two arguments" ); |
| 486 | mapper.map(from: redRegion.getArgument(i: 0), to: pvtRedVal); |
| 487 | mapper.map(from: redRegion.getArgument(i: 1), to: y); |
| 488 | for (auto &op : redRegion.getOps()) { |
| 489 | Operation *cloneOp = builder.clone(op, mapper); |
| 490 | if (auto yieldOp = dyn_cast<omp::YieldOp>(Val&: *cloneOp)) { |
| 491 | assert(yieldOp && yieldOp.getResults().size() == 1 && |
| 492 | "expect YieldOp in reduction region to return one result" ); |
| 493 | Value redVal = yieldOp.getResults()[0]; |
| 494 | LLVM::StoreOp::create(builder&: rewriter, location: loc, value: redVal, addr: pvtRedVar); |
| 495 | rewriter.eraseOp(op: yieldOp); |
| 496 | break; |
| 497 | } |
| 498 | } |
| 499 | } |
| 500 | rewriter.eraseOp(op: reduce); |
| 501 | |
| 502 | SmallVector<Value> numThreadsVars; |
| 503 | if (numThreads > 0) { |
| 504 | Value numThreadsVar = LLVM::ConstantOp::create( |
| 505 | builder&: rewriter, location: loc, value: rewriter.getI32IntegerAttr(value: numThreads)); |
| 506 | numThreadsVars.push_back(Elt: numThreadsVar); |
| 507 | } |
| 508 | // Create the parallel wrapper. |
| 509 | auto ompParallel = omp::ParallelOp::create( |
| 510 | builder&: rewriter, location: loc, |
| 511 | /* allocate_vars = */ llvm::SmallVector<Value>{}, |
| 512 | /* allocator_vars = */ llvm::SmallVector<Value>{}, |
| 513 | /* allocate_alignments = */ nullptr, |
| 514 | /* allocate_private_indices = */ nullptr, |
| 515 | /* if_expr = */ Value{}, |
| 516 | /* num_threads_vars = */ numThreadsVars, |
| 517 | /* private_vars = */ ValueRange(), |
| 518 | /* private_syms = */ nullptr, |
| 519 | /* private_needs_barrier = */ false, |
| 520 | /* proc_bind_kind = */ omp::ClauseProcBindKindAttr{}, |
| 521 | /* reduction_mods = */ nullptr, |
| 522 | /* reduction_vars = */ llvm::SmallVector<Value>{}, |
| 523 | /* reduction_byref = */ DenseBoolArrayAttr{}, |
| 524 | /* reduction_syms = */ ArrayAttr{}); |
| 525 | { |
| 526 | |
| 527 | OpBuilder::InsertionGuard guard(rewriter); |
| 528 | rewriter.createBlock(parent: &ompParallel.getRegion()); |
| 529 | |
| 530 | // Replace the loop. |
| 531 | { |
| 532 | OpBuilder::InsertionGuard allocaGuard(rewriter); |
| 533 | // Create worksharing loop wrapper. |
| 534 | auto wsloopOp = omp::WsloopOp::create(builder&: rewriter, location: parallelOp.getLoc()); |
| 535 | if (!reductionVariables.empty()) { |
| 536 | wsloopOp.setReductionSymsAttr( |
| 537 | ArrayAttr::get(context: rewriter.getContext(), value: reductionSyms)); |
| 538 | wsloopOp.getReductionVarsMutable().append(values: reductionVariables); |
| 539 | llvm::SmallVector<bool> reductionByRef; |
| 540 | // false because these reductions always reduce scalars and so do |
| 541 | // not need to pass by reference |
| 542 | reductionByRef.resize(N: reductionVariables.size(), NV: false); |
| 543 | wsloopOp.setReductionByref( |
| 544 | DenseBoolArrayAttr::get(context: rewriter.getContext(), content: reductionByRef)); |
| 545 | } |
| 546 | omp::TerminatorOp::create(builder&: rewriter, location: loc); // omp.parallel terminator. |
| 547 | |
| 548 | // The wrapper's entry block arguments will define the reduction |
| 549 | // variables. |
| 550 | llvm::SmallVector<mlir::Type> reductionTypes; |
| 551 | reductionTypes.reserve(N: reductionVariables.size()); |
| 552 | llvm::transform(Range&: reductionVariables, d_first: std::back_inserter(x&: reductionTypes), |
| 553 | F: [](mlir::Value v) { return v.getType(); }); |
| 554 | rewriter.createBlock( |
| 555 | parent: &wsloopOp.getRegion(), insertPt: {}, argTypes: reductionTypes, |
| 556 | locs: llvm::SmallVector<mlir::Location>(reductionVariables.size(), |
| 557 | parallelOp.getLoc())); |
| 558 | |
| 559 | // Create loop nest and populate region with contents of scf.parallel. |
| 560 | auto loopOp = omp::LoopNestOp::create( |
| 561 | builder&: rewriter, location: parallelOp.getLoc(), collapse_num_loops: parallelOp.getLowerBound().size(), |
| 562 | loop_lower_bounds: parallelOp.getLowerBound(), loop_upper_bounds: parallelOp.getUpperBound(), |
| 563 | loop_steps: parallelOp.getStep(), /*loop_inclusive=*/false, |
| 564 | /*tile_sizes=*/nullptr); |
| 565 | |
| 566 | rewriter.inlineRegionBefore(region&: parallelOp.getRegion(), parent&: loopOp.getRegion(), |
| 567 | before: loopOp.getRegion().begin()); |
| 568 | |
| 569 | // Remove reduction-related block arguments from omp.loop_nest and |
| 570 | // redirect uses to the corresponding omp.wsloop block argument. |
| 571 | mlir::Block &loopOpEntryBlock = loopOp.getRegion().front(); |
| 572 | unsigned numLoops = parallelOp.getNumLoops(); |
| 573 | rewriter.replaceAllUsesWith( |
| 574 | from: loopOpEntryBlock.getArguments().drop_front(N: numLoops), |
| 575 | to: wsloopOp.getRegion().getArguments()); |
| 576 | loopOpEntryBlock.eraseArguments( |
| 577 | start: numLoops, num: loopOpEntryBlock.getNumArguments() - numLoops); |
| 578 | |
| 579 | Block *ops = |
| 580 | rewriter.splitBlock(block: &loopOpEntryBlock, before: loopOpEntryBlock.begin()); |
| 581 | rewriter.setInsertionPointToStart(&loopOpEntryBlock); |
| 582 | |
| 583 | auto scope = memref::AllocaScopeOp::create( |
| 584 | builder&: rewriter, location: parallelOp.getLoc(), results: TypeRange()); |
| 585 | omp::YieldOp::create(builder&: rewriter, location: loc, results: ValueRange()); |
| 586 | Block *scopeBlock = rewriter.createBlock(parent: &scope.getBodyRegion()); |
| 587 | rewriter.mergeBlocks(source: ops, dest: scopeBlock); |
| 588 | rewriter.setInsertionPointToEnd(&*scope.getBodyRegion().begin()); |
| 589 | memref::AllocaScopeReturnOp::create(builder&: rewriter, location: loc, results: ValueRange()); |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | // Load loop results. |
| 594 | SmallVector<Value> results; |
| 595 | results.reserve(N: reductionVariables.size()); |
| 596 | for (auto [variable, type] : |
| 597 | llvm::zip(t&: reductionVariables, u: parallelOp.getResultTypes())) { |
| 598 | Value res = LLVM::LoadOp::create(builder&: rewriter, location: loc, type, addr: variable); |
| 599 | results.push_back(Elt: res); |
| 600 | } |
| 601 | rewriter.replaceOp(op: parallelOp, newValues: results); |
| 602 | |
| 603 | return success(); |
| 604 | } |
| 605 | }; |
| 606 | |
| 607 | /// Applies the conversion patterns in the given function. |
| 608 | static LogicalResult applyPatterns(ModuleOp module, unsigned numThreads) { |
| 609 | RewritePatternSet patterns(module.getContext()); |
| 610 | patterns.add<ParallelOpLowering>(arg: module.getContext(), args&: numThreads); |
| 611 | FrozenRewritePatternSet frozen(std::move(patterns)); |
| 612 | walkAndApplyPatterns(op: module, patterns: frozen); |
| 613 | auto status = module.walk(callback: [](Operation *op) { |
| 614 | if (isa<scf::ReduceOp, scf::ReduceReturnOp, scf::ParallelOp>(Val: op)) { |
| 615 | op->emitError(message: "unconverted operation found" ); |
| 616 | return WalkResult::interrupt(); |
| 617 | } |
| 618 | return WalkResult::advance(); |
| 619 | }); |
| 620 | return failure(IsFailure: status.wasInterrupted()); |
| 621 | } |
| 622 | |
| 623 | /// A pass converting SCF operations to OpenMP operations. |
| 624 | struct SCFToOpenMPPass |
| 625 | : public impl::ConvertSCFToOpenMPPassBase<SCFToOpenMPPass> { |
| 626 | |
| 627 | using Base::Base; |
| 628 | |
| 629 | /// Pass entry point. |
| 630 | void runOnOperation() override { |
| 631 | if (failed(Result: applyPatterns(module: getOperation(), numThreads))) |
| 632 | signalPassFailure(); |
| 633 | } |
| 634 | }; |
| 635 | |
| 636 | } // namespace |
| 637 | |