| 1 | //===- ParallelLoopFusion.cpp - Code to perform loop fusion ---------------===// |
| 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 loop fusion on parallel loops. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "mlir/Dialect/SCF/Transforms/Passes.h" |
| 14 | |
| 15 | #include "mlir/Analysis/AliasAnalysis.h" |
| 16 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 17 | #include "mlir/Dialect/Index/IR/IndexOps.h" |
| 18 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| 19 | #include "mlir/Dialect/MemRef/Utils/MemRefUtils.h" |
| 20 | #include "mlir/Dialect/SCF/IR/SCF.h" |
| 21 | #include "mlir/Dialect/SCF/Transforms/Transforms.h" |
| 22 | #include "mlir/Dialect/SCF/Utils/Utils.h" |
| 23 | #include "mlir/Dialect/Utils/IndexingUtils.h" |
| 24 | #include "mlir/Dialect/Utils/StaticValueUtils.h" |
| 25 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 26 | #include "mlir/IR/Builders.h" |
| 27 | #include "mlir/IR/BuiltinTypes.h" |
| 28 | #include "mlir/IR/IRMapping.h" |
| 29 | #include "mlir/IR/Matchers.h" |
| 30 | #include "mlir/IR/OpDefinition.h" |
| 31 | #include "mlir/IR/OperationSupport.h" |
| 32 | #include "mlir/IR/PatternMatch.h" |
| 33 | #include "mlir/IR/Value.h" |
| 34 | #include "mlir/Interfaces/SideEffectInterfaces.h" |
| 35 | |
| 36 | #include "llvm/ADT/STLExtras.h" |
| 37 | #include "llvm/ADT/SetVector.h" |
| 38 | #include "llvm/ADT/SmallBitVector.h" |
| 39 | #include "llvm/ADT/TypeSwitch.h" |
| 40 | #include "llvm/Support/InterleavedRange.h" |
| 41 | |
| 42 | #include "llvm/Support/DebugLog.h" |
| 43 | #include <numeric> |
| 44 | #include <optional> |
| 45 | #include <tuple> |
| 46 | #define DEBUG_TYPE "parallel-loop-fusion" |
| 47 | |
| 48 | namespace mlir { |
| 49 | #define GEN_PASS_DEF_SCFPARALLELLOOPFUSION |
| 50 | #include "mlir/Dialect/SCF/Transforms/Passes.h.inc" |
| 51 | } // namespace mlir |
| 52 | |
| 53 | using namespace mlir; |
| 54 | using namespace mlir::scf; |
| 55 | |
| 56 | /// Verify there are no nested ParallelOps. |
| 57 | static bool hasNestedParallelOp(ParallelOp ploop) { |
| 58 | auto walkResult = |
| 59 | ploop.getBody()->walk(callback: [](ParallelOp) { return WalkResult::interrupt(); }); |
| 60 | return walkResult.wasInterrupted(); |
| 61 | } |
| 62 | |
| 63 | /// Verify equal iteration spaces. |
| 64 | static bool equalIterationSpaces(ParallelOp firstPloop, |
| 65 | ParallelOp secondPloop) { |
| 66 | if (firstPloop.getNumLoops() != secondPloop.getNumLoops() || |
| 67 | firstPloop.getUnsignedCmp() != secondPloop.getUnsignedCmp()) |
| 68 | return false; |
| 69 | |
| 70 | // Two bounds match if they are the same value, or if both are constants |
| 71 | // holding the same value. The latter matters because equivalent bounds are |
| 72 | // often materialized by distinct `arith.constant` ops, which leaves the |
| 73 | // iteration spaces equal even though the SSA values differ. |
| 74 | auto matchOperands = [&](const OperandRange &lhs, |
| 75 | const OperandRange &rhs) -> bool { |
| 76 | // TODO: Extend this to support aliases. |
| 77 | return std::equal(first1: lhs.begin(), last1: lhs.end(), first2: rhs.begin(), |
| 78 | binary_pred: [](Value lhsValue, Value rhsValue) { |
| 79 | if (lhsValue == rhsValue) |
| 80 | return true; |
| 81 | std::optional<int64_t> lhsConst = |
| 82 | getConstantIntValue(ofr: lhsValue); |
| 83 | std::optional<int64_t> rhsConst = |
| 84 | getConstantIntValue(ofr: rhsValue); |
| 85 | return lhsConst && rhsConst && *lhsConst == *rhsConst; |
| 86 | }); |
| 87 | }; |
| 88 | return matchOperands(firstPloop.getLowerBound(), |
| 89 | secondPloop.getLowerBound()) && |
| 90 | matchOperands(firstPloop.getUpperBound(), |
| 91 | secondPloop.getUpperBound()) && |
| 92 | matchOperands(firstPloop.getStep(), secondPloop.getStep()); |
| 93 | } |
| 94 | |
| 95 | /// Check if both operations are the same type of memory write op and |
| 96 | /// write to the same memory location (same buffer and same indices). |
| 97 | static bool opsWriteSameMemLocation(Operation *op1, Operation *op2) { |
| 98 | if (!op1 || !op2 || op1->getName() != op2->getName()) |
| 99 | return false; |
| 100 | if (op1 == op2) |
| 101 | return true; |
| 102 | // support only these memory-writing ops for now |
| 103 | if (!isa<memref::StoreOp, vector::TransferWriteOp, vector::StoreOp>(Val: op1)) |
| 104 | return false; |
| 105 | bool opsAreIdentical = |
| 106 | llvm::TypeSwitch<Operation *, bool>(op1) |
| 107 | .Case(caseFn: [&](memref::StoreOp storeOp1) { |
| 108 | auto storeOp2 = cast<memref::StoreOp>(Val: op2); |
| 109 | return (storeOp1.getMemRef() == storeOp2.getMemRef()) && |
| 110 | (storeOp1.getIndices() == storeOp2.getIndices()); |
| 111 | }) |
| 112 | .Case(caseFn: [&](vector::TransferWriteOp writeOp1) { |
| 113 | auto writeOp2 = cast<vector::TransferWriteOp>(Val: op2); |
| 114 | return (writeOp1.getBase() == writeOp2.getBase()) && |
| 115 | (writeOp1.getIndices() == writeOp2.getIndices()) && |
| 116 | (writeOp1.getMask() == writeOp2.getMask()) && |
| 117 | (writeOp1.getValueToStore().getType() == |
| 118 | writeOp2.getValueToStore().getType()) && |
| 119 | (writeOp1.getInBounds() == writeOp2.getInBounds()); |
| 120 | }) |
| 121 | .Case(caseFn: [&](vector::StoreOp vecStoreOp1) { |
| 122 | auto vecStoreOp2 = cast<vector::StoreOp>(Val: op2); |
| 123 | return (vecStoreOp1.getBase() == vecStoreOp2.getBase()) && |
| 124 | (vecStoreOp1.getIndices() == vecStoreOp2.getIndices()) && |
| 125 | (vecStoreOp1.getValueToStore().getType() == |
| 126 | vecStoreOp2.getValueToStore().getType()) && |
| 127 | (vecStoreOp1.getAlignment() == vecStoreOp2.getAlignment()) && |
| 128 | (vecStoreOp1.getNontemporal() == |
| 129 | vecStoreOp2.getNontemporal()); |
| 130 | }) |
| 131 | .Default(defaultFn: [](Operation *) { return false; }); |
| 132 | return opsAreIdentical; |
| 133 | } |
| 134 | |
| 135 | /// Check if val1 (from the first parallel loop) and val2 (from the |
| 136 | /// second) are equivalent, considering the mapping of induction variables from |
| 137 | /// the first to the second parallel loop. |
| 138 | static bool valsAreEquivalent(Value val1, Value val2, |
| 139 | const IRMapping &loopsIVsMap) { |
| 140 | if (val1 == val2 || loopsIVsMap.lookupOrDefault(from: val1) == val2 || |
| 141 | loopsIVsMap.lookupOrDefault(from: val2) == val1) |
| 142 | return true; |
| 143 | Operation *val1DefOp = val1.getDefiningOp(); |
| 144 | Operation *val2DefOp = val2.getDefiningOp(); |
| 145 | if (!val1DefOp || !val2DefOp) |
| 146 | return false; |
| 147 | if (!isMemoryEffectFree(op: val1DefOp) || !isMemoryEffectFree(op: val2DefOp)) |
| 148 | return false; |
| 149 | return OperationEquivalence::isEquivalentTo( |
| 150 | lhs: val1DefOp, rhs: val2DefOp, |
| 151 | checkEquivalent: [&](Value v1, Value v2) { |
| 152 | return success(IsSuccess: loopsIVsMap.lookupOrDefault(from: v1) == v2 || |
| 153 | loopsIVsMap.lookupOrDefault(from: v2) == v1); |
| 154 | }, |
| 155 | /*markEquivalent=*/nullptr, flags: OperationEquivalence::Flags::IgnoreLocations); |
| 156 | } |
| 157 | |
| 158 | /// If the `expr` value is the result of an integer addition of `base` and a |
| 159 | /// constant, return the constant. |
| 160 | static std::optional<int64_t> getAddConstant(Value expr, Value base, |
| 161 | const IRMapping &loopsIVsMap) { |
| 162 | if (auto addOp = expr.getDefiningOp<arith::AddIOp>()) { |
| 163 | if (auto constOp = getConstantIntValue(ofr: addOp.getLhs()); |
| 164 | constOp && valsAreEquivalent(val1: addOp.getRhs(), val2: base, loopsIVsMap)) |
| 165 | return constOp.value(); |
| 166 | if (auto constOp = getConstantIntValue(ofr: addOp.getRhs()); |
| 167 | constOp && valsAreEquivalent(val1: addOp.getLhs(), val2: base, loopsIVsMap)) |
| 168 | return constOp.value(); |
| 169 | return std::nullopt; |
| 170 | } |
| 171 | |
| 172 | if (auto addOp = expr.getDefiningOp<index::AddOp>()) { |
| 173 | if (auto constOp = getConstantIntValue(ofr: addOp.getLhs()); |
| 174 | constOp && valsAreEquivalent(val1: addOp.getRhs(), val2: base, loopsIVsMap)) |
| 175 | return constOp.value(); |
| 176 | if (auto constOp = getConstantIntValue(ofr: addOp.getRhs()); |
| 177 | constOp && valsAreEquivalent(val1: addOp.getLhs(), val2: base, loopsIVsMap)) |
| 178 | return constOp.value(); |
| 179 | return std::nullopt; |
| 180 | } |
| 181 | |
| 182 | if (auto applyOp = expr.getDefiningOp<affine::AffineApplyOp>()) { |
| 183 | AffineMap map = applyOp.getAffineMap(); |
| 184 | if (map.getNumResults() != 1 || map.getNumDims() != 1 || |
| 185 | map.getNumSymbols() != 0) |
| 186 | return std::nullopt; |
| 187 | if (!valsAreEquivalent(val1: applyOp.getOperand(i: 0), val2: base, loopsIVsMap)) |
| 188 | return std::nullopt; |
| 189 | AffineExpr result = map.getResult(idx: 0); |
| 190 | auto bin = dyn_cast<AffineBinaryOpExpr>(Val&: result); |
| 191 | if (!bin || bin.getKind() != AffineExprKind::Add) |
| 192 | return std::nullopt; |
| 193 | auto lhsDim = dyn_cast<AffineDimExpr>(Val: bin.getLHS()); |
| 194 | auto rhsDim = dyn_cast<AffineDimExpr>(Val: bin.getRHS()); |
| 195 | auto lhsConst = dyn_cast<AffineConstantExpr>(Val: bin.getLHS()); |
| 196 | auto rhsConst = dyn_cast<AffineConstantExpr>(Val: bin.getRHS()); |
| 197 | if (lhsConst && rhsDim) |
| 198 | return lhsConst.getValue(); |
| 199 | if (rhsConst && lhsDim) |
| 200 | return rhsConst.getValue(); |
| 201 | } |
| 202 | return std::nullopt; |
| 203 | } |
| 204 | |
| 205 | // Return true if the scalar load index may hit any element covered by a |
| 206 | // vector.store/transfer_write along a single memref dimension. Supported cases: |
| 207 | // |
| 208 | // 1) Direct index match (with optional offset): |
| 209 | // vector.transfer_write %v, %A[%i] : vector<4xf32>, memref<...> |
| 210 | // %x = memref.load %A[%i] : memref<...> |
| 211 | // |
| 212 | // 2) Loop IV range intersects the write range: |
| 213 | // vector.transfer_write %v, %A[%c0] : vector<4xf32>, memref<...> |
| 214 | // scf.for %k = %c0 to %c4 step %c1 { %x = memref.load %A[%k] } |
| 215 | // |
| 216 | // 3) Constant index (or IV + constant) within the write range: |
| 217 | // vector.transfer_write %v, %A[%c0] : vector<4xf32>, memref<...> |
| 218 | // %x = memref.load %A[%c2] : memref<...> |
| 219 | // %y = memref.load %A[%i + %c1] : memref<...> |
| 220 | // |
| 221 | // Args: |
| 222 | // - loadIndex: index used by the scalar load for this dimension. |
| 223 | // - offset: subview offset for the base memref dimension (if any). |
| 224 | // - writeIndex: index used by the transfer_write for this dimension. Can be |
| 225 | // null if the dim was dropped by a rank reducing subview, whose result is |
| 226 | // written by the vector.write. |
| 227 | // - extent: vector size along this dimension (number of elements written). |
| 228 | // - loopsIVsMap: IV equivalence map between fused loops. |
| 229 | static bool loadIndexWithinWriteRange(Value loadIndex, OpFoldResult offset, |
| 230 | Value writeIndex, int64_t extent, |
| 231 | const IRMapping &loopsIVsMap) { |
| 232 | if (extent <= 0) |
| 233 | return false; |
| 234 | |
| 235 | // Extract constant loop bounds for loop IVs (e.g. from scf.for). |
| 236 | auto getConstLoopBoundsForIV = |
| 237 | [](Value index) -> std::optional<std::tuple<int64_t, int64_t, int64_t>> { |
| 238 | auto blockArg = dyn_cast<BlockArgument>(Val&: index); |
| 239 | if (!blockArg) |
| 240 | return std::nullopt; |
| 241 | auto *parentOp = blockArg.getOwner()->getParentOp(); |
| 242 | auto loopLike = dyn_cast<LoopLikeOpInterface>(Val: parentOp); |
| 243 | if (!loopLike) |
| 244 | return std::nullopt; |
| 245 | auto ranges = getConstLoopBounds(loopOp: loopLike); |
| 246 | if (ranges.empty()) |
| 247 | return std::nullopt; |
| 248 | |
| 249 | auto ivs = loopLike.getLoopInductionVars(); |
| 250 | if (!ivs) |
| 251 | return std::nullopt; |
| 252 | auto it = llvm::find(Range&: *ivs, Val: blockArg); |
| 253 | if (it == ivs->end()) |
| 254 | return std::nullopt; |
| 255 | unsigned pos = std::distance(first: ivs->begin(), last: it); |
| 256 | if (pos >= ranges.size()) |
| 257 | return std::nullopt; |
| 258 | auto [lb, ub, step] = ranges[pos]; |
| 259 | return std::make_tuple(args&: lb, args&: ub, args&: step); |
| 260 | }; |
| 261 | |
| 262 | std::optional<int64_t> offsetConst = getConstantIntValue(ofr: offset); |
| 263 | std::optional<int64_t> writeConst = |
| 264 | writeIndex ? getConstantIntValue(ofr: writeIndex) : std::optional<int64_t>(0); |
| 265 | if (!writeConst && writeIndex) { |
| 266 | // Treat single-iteration IVs as constants for matching. |
| 267 | if (auto bounds = getConstLoopBoundsForIV(writeIndex)) { |
| 268 | auto [lb, ub, step] = *bounds; |
| 269 | if (step > 0 && ub == lb + step) |
| 270 | writeConst = lb; |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | // Check whether a loop IV is fully contained in a constant write range. |
| 275 | auto loopIVWithinRange = [](int64_t lb, int64_t ub, int64_t step, |
| 276 | int64_t rangeStart, int64_t rangeExtent) -> bool { |
| 277 | if (rangeExtent <= 0 || step <= 0) |
| 278 | return false; |
| 279 | if (ub <= lb) |
| 280 | return false; |
| 281 | int64_t rangeEnd = rangeStart + rangeExtent; |
| 282 | return lb >= rangeStart && ub <= rangeEnd; |
| 283 | }; |
| 284 | |
| 285 | if (offsetConst && writeConst) { |
| 286 | // Constant start of the write range; check constant load or loop IV range. |
| 287 | int64_t start = *offsetConst + *writeConst; |
| 288 | if (auto loadConst = getConstantIntValue(ofr: loadIndex)) |
| 289 | return (*loadConst >= start && *loadConst < start + extent); |
| 290 | if (auto bounds = getConstLoopBoundsForIV(loadIndex)) { |
| 291 | auto [lb, ub, step] = *bounds; |
| 292 | return loopIVWithinRange(lb, ub, step, start, extent); |
| 293 | } |
| 294 | } |
| 295 | |
| 296 | if (writeIndex) { |
| 297 | // Direct IV match (or IV + constant) against the write index. |
| 298 | if (offsetConst && *offsetConst == 0 && |
| 299 | valsAreEquivalent(val1: loadIndex, val2: writeIndex, loopsIVsMap)) |
| 300 | return true; |
| 301 | if (auto addConst = getAddConstant(expr: loadIndex, base: writeIndex, loopsIVsMap)) { |
| 302 | // Match load index of the form writeIndex + C within the write extent. |
| 303 | if (offsetConst) { |
| 304 | int64_t start = *offsetConst; |
| 305 | return (*addConst >= start && *addConst < start + extent); |
| 306 | } |
| 307 | } |
| 308 | return false; |
| 309 | } |
| 310 | |
| 311 | if (auto offsetVal = dyn_cast<Value>(Val&: offset)) { |
| 312 | // Exact match when extent is 1 and the load hits the offset value. |
| 313 | if (extent == 1 && valsAreEquivalent(val1: loadIndex, val2: offsetVal, loopsIVsMap)) |
| 314 | return true; |
| 315 | } |
| 316 | |
| 317 | return false; |
| 318 | } |
| 319 | |
| 320 | /// Return the base memref value used by the given memory op. |
| 321 | static Value getBaseMemref(Operation *op) { |
| 322 | // TODO: use the common interface for memory ops once available. |
| 323 | return llvm::TypeSwitch<Operation *, Value>(op) |
| 324 | .Case(caseFn: [&](memref::LoadOp load) { return load.getMemRef(); }) |
| 325 | .Case(caseFn: [&](memref::StoreOp store) { return store.getMemRef(); }) |
| 326 | .Case(caseFn: [&](vector::TransferReadOp read) { return read.getBase(); }) |
| 327 | .Case(caseFn: [&](vector::TransferWriteOp write) { return write.getBase(); }) |
| 328 | .Case(caseFn: [&](vector::LoadOp load) { return load.getBase(); }) |
| 329 | .Case(caseFn: [&](vector::StoreOp store) { return store.getBase(); }) |
| 330 | .Default(defaultFn: [](Operation *) { return Value(); }); |
| 331 | } |
| 332 | |
| 333 | /// Recognize scalar memref.load of an element produced by a vector write |
| 334 | /// (vector.transfer_write or vector.store, optionally through a rank-reducing |
| 335 | /// unit-stride subview) of the same buffer. This covers the pattern where a |
| 336 | /// vector write stores a full lane pack and a subsequent scalar load reads an |
| 337 | /// element from that lane pack. EXAMPLE: |
| 338 | /// vector.transfer_write %V, %arg[%x, %y, ..., 0] {in_bounds = [true]} : |
| 339 | /// vector<4xf32>, memref<4xf32, strided<[1], offset: ?>> |
| 340 | /// scf.for %iter = %c0 to %c4 step %c1 iter_args(...) -> (f32) { |
| 341 | /// %0 = memref.load %arg[%x, %y, ..., %iter] : memref<1x128x16x4xf32> |
| 342 | /// ... |
| 343 | /// } |
| 344 | /// |
| 345 | static bool isLoadOnWrittenVector(memref::LoadOp loadOp, Value writeBase, |
| 346 | ValueRange writeIndices, VectorType vecTy, |
| 347 | ArrayRef<int64_t> vectorDimForWriteDim, |
| 348 | const IRMapping &ivsMap) { |
| 349 | if (!vecTy) |
| 350 | return false; |
| 351 | |
| 352 | Value base = writeBase; |
| 353 | // The write base if there is no subview, or the subview source otherwise. |
| 354 | MemrefValue baseMemref = nullptr; |
| 355 | SmallVector<OpFoldResult> offsets; |
| 356 | llvm::SmallBitVector droppedDims; |
| 357 | bool hasSubview = false; |
| 358 | auto *ctx = loadOp.getContext(); |
| 359 | if (auto subView = base.getDefiningOp<memref::SubViewOp>()) { |
| 360 | if (!subView.hasUnitStride()) |
| 361 | return false; |
| 362 | baseMemref = cast<MemrefValue>(Val: subView.getSource()); |
| 363 | offsets = llvm::to_vector(Range: subView.getMixedOffsets()); |
| 364 | droppedDims = subView.getDroppedDims(); |
| 365 | hasSubview = true; |
| 366 | } else { |
| 367 | baseMemref = dyn_cast<MemrefValue>(Val&: base); |
| 368 | if (!baseMemref) |
| 369 | return false; |
| 370 | } |
| 371 | |
| 372 | auto loadIndices = loadOp.getIndices(); |
| 373 | unsigned baseRank = baseMemref.getType().getRank(); |
| 374 | if ((loadOp.getMemref() != baseMemref) || (loadIndices.size() != baseRank)) |
| 375 | return false; |
| 376 | |
| 377 | unsigned writeRank = writeIndices.size(); |
| 378 | if ((!hasSubview && writeRank != baseRank) || |
| 379 | (hasSubview && offsets.size() != baseRank) || |
| 380 | (vectorDimForWriteDim.size() != writeRank)) |
| 381 | return false; |
| 382 | |
| 383 | auto zeroAttr = IntegerAttr::get(type: IndexType::get(context: ctx), value: 0); |
| 384 | unsigned writeMemrefDim = 0; |
| 385 | for (unsigned baseDim : llvm::seq(Size: baseRank)) { |
| 386 | bool wasDropped = (hasSubview && droppedDims.test(Idx: baseDim)); |
| 387 | int64_t vectorDim = !wasDropped ? vectorDimForWriteDim[writeMemrefDim] : -1; |
| 388 | int64_t extent = 1; |
| 389 | if (vectorDim >= 0) { |
| 390 | int64_t dimSize = vecTy.getDimSize(idx: vectorDim); |
| 391 | if (dimSize == ShapedType::kDynamic) |
| 392 | return false; |
| 393 | extent = dimSize; |
| 394 | } |
| 395 | Value writeIndex = !wasDropped ? writeIndices[writeMemrefDim] : Value(); |
| 396 | OpFoldResult offset = |
| 397 | hasSubview ? offsets[baseDim] : OpFoldResult(zeroAttr); |
| 398 | if (!loadIndexWithinWriteRange(loadIndex: loadIndices[baseDim], offset, writeIndex, |
| 399 | extent, loopsIVsMap: ivsMap)) |
| 400 | return false; |
| 401 | if (!wasDropped) |
| 402 | ++writeMemrefDim; |
| 403 | } |
| 404 | |
| 405 | return true; |
| 406 | } |
| 407 | |
| 408 | /// Recognize scalar memref.load of an element produced by a |
| 409 | /// vector.transfer_write |
| 410 | static bool loadMatchesVectorWrite(memref::LoadOp loadOp, |
| 411 | vector::TransferWriteOp writeOp, |
| 412 | const IRMapping &ivsMap) { |
| 413 | auto vecTy = dyn_cast<VectorType>(Val: writeOp.getVector().getType()); |
| 414 | if (!vecTy) |
| 415 | return false; |
| 416 | |
| 417 | // A masked write may not write the element that the load reads. |
| 418 | if (writeOp.getMask() || writeOp.isMasked()) |
| 419 | return false; |
| 420 | |
| 421 | unsigned writeRank = writeOp.getIndices().size(); |
| 422 | AffineMap permutationMap = writeOp.getPermutationMap(); |
| 423 | if (!permutationMap.isProjectedPermutation() || |
| 424 | permutationMap.getNumResults() != vecTy.getRank() || |
| 425 | permutationMap.getNumDims() != writeRank) |
| 426 | return false; |
| 427 | |
| 428 | SmallVector<int64_t> vectorDimForWriteDim(writeRank, -1); |
| 429 | for (unsigned vecDim = 0; vecDim < permutationMap.getNumResults(); ++vecDim) { |
| 430 | auto dimExpr = dyn_cast<AffineDimExpr>(Val: permutationMap.getResult(idx: vecDim)); |
| 431 | if (!dimExpr) |
| 432 | return false; |
| 433 | unsigned writeDim = dimExpr.getPosition(); |
| 434 | if (writeDim >= writeRank || vectorDimForWriteDim[writeDim] != -1) |
| 435 | return false; |
| 436 | vectorDimForWriteDim[writeDim] = vecDim; |
| 437 | } |
| 438 | |
| 439 | return isLoadOnWrittenVector(loadOp, writeBase: writeOp.getBase(), writeIndices: writeOp.getIndices(), |
| 440 | vecTy, vectorDimForWriteDim, ivsMap); |
| 441 | } |
| 442 | |
| 443 | /// Recognize scalar memref.load of an element produced by a vector.store |
| 444 | static bool loadMatchesVectorStore(memref::LoadOp loadOp, |
| 445 | vector::StoreOp storeOp, |
| 446 | const IRMapping &ivsMap) { |
| 447 | auto vecTy = dyn_cast<VectorType>(Val: storeOp.getValueToStore().getType()); |
| 448 | if (!vecTy) |
| 449 | return false; |
| 450 | |
| 451 | unsigned writeRank = storeOp.getIndices().size(); |
| 452 | if (vecTy.getRank() > writeRank) |
| 453 | return false; |
| 454 | |
| 455 | SmallVector<int64_t> vectorDimForWriteDim(writeRank, -1); |
| 456 | unsigned vecRank = vecTy.getRank(); |
| 457 | for (unsigned i = 0; i < vecRank; ++i) { |
| 458 | unsigned writeDim = writeRank - vecRank + i; |
| 459 | vectorDimForWriteDim[writeDim] = i; |
| 460 | } |
| 461 | |
| 462 | return isLoadOnWrittenVector(loadOp, writeBase: storeOp.getBase(), writeIndices: storeOp.getIndices(), |
| 463 | vecTy, vectorDimForWriteDim, ivsMap); |
| 464 | } |
| 465 | |
| 466 | /// Check if both operations access the same positions of the same |
| 467 | /// buffer, but one of the two does it through a rank-reducing full subview of |
| 468 | /// the buffer (the other's base). EXAMPLE: |
| 469 | /// memref.store %a, %buf[%c0, %i, %j] : memref<1x2x2xf32> |
| 470 | /// %alias = memref.subview %buf[0, 0, 0][1, 2, 2][1, 1, 1]: memref<1x2x2xf32> |
| 471 | /// to memref<2x2xf32> |
| 472 | /// %val = memref.load %alias[%i, %j] : memref<2x2xf32> |
| 473 | template <typename OpTy1, typename OpTy2> |
| 474 | static bool opsAccessSameIndicesViaRankReducingSubview( |
| 475 | OpTy1 op1, OpTy2 op2, const IRMapping &firstToSecondPloopIVsMap, |
| 476 | OpBuilder &b) { |
| 477 | auto base1 = cast<MemrefValue>(getBaseMemref(op1)); |
| 478 | auto base2 = cast<MemrefValue>(getBaseMemref(op2)); |
| 479 | if (!base1 || !base2) |
| 480 | return false; |
| 481 | |
| 482 | auto accessThroughTrivialSubviewIsSame = |
| 483 | [&b](memref::SubViewOp subView, ValueRange subViewAccess, |
| 484 | ValueRange sourceAccess, const IRMapping &ivsMap) -> bool { |
| 485 | SmallVector<Value> resolvedSubviewAccess; |
| 486 | LogicalResult resolved = resolveSourceIndicesRankReducingSubview( |
| 487 | loc: subView.getLoc(), b, subViewOp: subView, indices: subViewAccess, sourceIndices&: resolvedSubviewAccess); |
| 488 | if (failed(Result: resolved) || |
| 489 | (resolvedSubviewAccess.size() != sourceAccess.size())) |
| 490 | return false; |
| 491 | for (auto [dimIdx, resolvedIndex] : |
| 492 | llvm::enumerate(First&: resolvedSubviewAccess)) { |
| 493 | if (!matchPattern(value: resolvedIndex, pattern: m_Zero()) && |
| 494 | !valsAreEquivalent(val1: resolvedIndex, val2: sourceAccess[dimIdx], loopsIVsMap: ivsMap)) |
| 495 | return false; |
| 496 | } |
| 497 | return true; |
| 498 | }; |
| 499 | |
| 500 | // Case 1: op1 uses a subview of op2's base. |
| 501 | if (auto subView = base1.template getDefiningOp<memref::SubViewOp>(); |
| 502 | subView && |
| 503 | memref::isSameViewOrTrivialAlias( |
| 504 | a: base2, b: cast<MemrefValue>(subView.getSource())) && |
| 505 | accessThroughTrivialSubviewIsSame(subView, op1.getIndices(), |
| 506 | op2.getIndices(), |
| 507 | firstToSecondPloopIVsMap)) |
| 508 | return true; |
| 509 | |
| 510 | // Case 2: op2 uses a subview of op1's base. |
| 511 | if (auto subView = base2.template getDefiningOp<memref::SubViewOp>(); |
| 512 | subView && |
| 513 | memref::isSameViewOrTrivialAlias( |
| 514 | a: base1, b: cast<MemrefValue>(subView.getSource())) && |
| 515 | accessThroughTrivialSubviewIsSame(subView, op2.getIndices(), |
| 516 | op1.getIndices(), |
| 517 | firstToSecondPloopIVsMap)) |
| 518 | return true; |
| 519 | |
| 520 | return false; |
| 521 | } |
| 522 | |
| 523 | /// Check if both memory read/write operations access the same indices |
| 524 | /// (considering also the mapping of induction variables from the first to the |
| 525 | /// second parallel loop). |
| 526 | template <typename OpTy1, typename OpTy2> |
| 527 | static bool opsAccessSameIndices(OpTy1 op1, OpTy2 op2, |
| 528 | const IRMapping &loopsIVsMap, OpBuilder &b) { |
| 529 | auto indices1 = op1.getIndices(); |
| 530 | auto indices2 = op2.getIndices(); |
| 531 | if (indices1.size() != indices2.size()) |
| 532 | return opsAccessSameIndicesViaRankReducingSubview(op1, op2, loopsIVsMap, b); |
| 533 | for (auto [idx1, idx2] : llvm::zip(indices1, indices2)) { |
| 534 | if (!valsAreEquivalent(idx1, idx2, loopsIVsMap)) |
| 535 | return false; |
| 536 | } |
| 537 | return true; |
| 538 | } |
| 539 | |
| 540 | /// Check if the loadOp reads from the same memory location (same buffer, |
| 541 | /// same indices and same properties) as written by the storeOp. |
| 542 | static bool |
| 543 | loadsFromSameMemoryLocationWrittenBy(Operation *loadOp, Operation *storeOp, |
| 544 | const IRMapping &firstToSecondPloopIVsMap, |
| 545 | OpBuilder &b) { |
| 546 | if (!loadOp || !storeOp) |
| 547 | return false; |
| 548 | // Support only these memory-reading ops for now |
| 549 | if (!isa<memref::LoadOp, vector::TransferReadOp, vector::LoadOp>(Val: loadOp)) |
| 550 | return false; |
| 551 | bool accessSameMemory = |
| 552 | llvm::TypeSwitch<Operation *, bool>(loadOp) |
| 553 | .Case(caseFn: [&](memref::LoadOp memLoadOp) { |
| 554 | if (auto memStoreOp = dyn_cast<memref::StoreOp>(Val: storeOp)) |
| 555 | return opsAccessSameIndices(op1: memLoadOp, op2: memStoreOp, |
| 556 | loopsIVsMap: firstToSecondPloopIVsMap, b); |
| 557 | if (auto vecWriteOp = dyn_cast<vector::TransferWriteOp>(Val: storeOp)) |
| 558 | return loadMatchesVectorWrite(loadOp: memLoadOp, writeOp: vecWriteOp, |
| 559 | ivsMap: firstToSecondPloopIVsMap); |
| 560 | if (auto vecStoreOp = dyn_cast<vector::StoreOp>(Val: storeOp)) |
| 561 | return loadMatchesVectorStore(loadOp: memLoadOp, storeOp: vecStoreOp, |
| 562 | ivsMap: firstToSecondPloopIVsMap); |
| 563 | return false; |
| 564 | }) |
| 565 | .Case(caseFn: [&](vector::TransferReadOp vecReadOp) { |
| 566 | auto vecWriteOp = dyn_cast<vector::TransferWriteOp>(Val: storeOp); |
| 567 | if (!vecWriteOp) |
| 568 | return false; |
| 569 | // A write masked by `vector.mask` may write fewer elements than |
| 570 | // the read reads. |
| 571 | return opsAccessSameIndices(op1: vecReadOp, op2: vecWriteOp, |
| 572 | loopsIVsMap: firstToSecondPloopIVsMap, b) && |
| 573 | (vecReadOp.getVectorType() == vecWriteOp.getVectorType()) && |
| 574 | (vecReadOp.getPermutationMap() == |
| 575 | vecWriteOp.getPermutationMap()) && |
| 576 | (vecReadOp.getMask() == vecWriteOp.getMask()) && |
| 577 | !vecWriteOp.isMasked() && |
| 578 | (vecReadOp.getInBounds() == vecWriteOp.getInBounds()); |
| 579 | }) |
| 580 | .Case(caseFn: [&](vector::LoadOp vecLoadOp) { |
| 581 | auto vecStoreOp = dyn_cast<vector::StoreOp>(Val: storeOp); |
| 582 | if (!vecStoreOp) |
| 583 | return false; |
| 584 | return opsAccessSameIndices(op1: vecLoadOp, op2: vecStoreOp, |
| 585 | loopsIVsMap: firstToSecondPloopIVsMap, b) && |
| 586 | (vecLoadOp.getVectorType() == vecStoreOp.getVectorType()) && |
| 587 | (vecLoadOp.getAlignment() == vecStoreOp.getAlignment()); |
| 588 | }) |
| 589 | .Default(defaultFn: [](Operation *) { return false; }); |
| 590 | return accessSameMemory; |
| 591 | } |
| 592 | |
| 593 | static Value getStoreOpTargetBuffer(Operation *op) { |
| 594 | return llvm::TypeSwitch<Operation *, Value>(op) |
| 595 | .Case(caseFn: [&](memref::StoreOp storeOp) { return storeOp.getMemRef(); }) |
| 596 | .Case(caseFn: [&](vector::TransferWriteOp writeOp) { return writeOp.getBase(); }) |
| 597 | .Case(caseFn: [&](vector::StoreOp vecStoreOp) { return vecStoreOp.getBase(); }) |
| 598 | .Default(defaultFn: [](Operation *) { return Value(); }); |
| 599 | } |
| 600 | |
| 601 | /// To be called when `mayAlias(val1, val2)` is true. Check if the potential |
| 602 | /// aliasing between the loadOp and storeOp can be resolved by analyzing their |
| 603 | /// access patterns. |
| 604 | static bool canResolveAlias(Operation *loadOp, Operation *storeOp, |
| 605 | const IRMapping &loopsIVsMap) { |
| 606 | if (auto transfWriteOp = dyn_cast<vector::TransferWriteOp>(Val: storeOp); |
| 607 | transfWriteOp && isa<memref::LoadOp>(Val: loadOp)) |
| 608 | return loadMatchesVectorWrite(loadOp: cast<memref::LoadOp>(Val: loadOp), writeOp: transfWriteOp, |
| 609 | ivsMap: loopsIVsMap); |
| 610 | if (auto vecStoreOp = dyn_cast<vector::StoreOp>(Val: storeOp); |
| 611 | vecStoreOp && isa<memref::LoadOp>(Val: loadOp)) |
| 612 | return loadMatchesVectorStore(loadOp: cast<memref::LoadOp>(Val: loadOp), storeOp: vecStoreOp, |
| 613 | ivsMap: loopsIVsMap); |
| 614 | return false; |
| 615 | } |
| 616 | |
| 617 | /// Check that the parallel loops have no mixed access to the same buffers. |
| 618 | /// Return `true` if the second parallel loop does not read or write the buffers |
| 619 | /// written by the first loop using different indices. |
| 620 | static bool haveNoDataDependenciesExceptSameIndex( |
| 621 | ParallelOp firstPloop, ParallelOp secondPloop, |
| 622 | const IRMapping &firstToSecondPloopIndices, |
| 623 | llvm::function_ref<bool(Value, Value)> mayAlias, OpBuilder &b) { |
| 624 | // Map buffers to their store/write ops in the firstPloop |
| 625 | DenseMap<Value, SmallVector<Operation *>> bufferStoresInFirstPloop; |
| 626 | // Record all the memory buffers used in store/write ops found in firstPloop |
| 627 | llvm::SmallSetVector<Value, 4> buffersWrittenInFirstPloop; |
| 628 | |
| 629 | auto collectStoreOpsInWalk = [&](Operation *op) { |
| 630 | auto memOpInterf = dyn_cast_if_present<MemoryEffectOpInterface>(Val: op); |
| 631 | // Ignore ops that don't write to memory |
| 632 | if (!memOpInterf || (!memOpInterf.hasEffect<MemoryEffects::Write>() && |
| 633 | !memOpInterf.hasEffect<MemoryEffects::Free>())) |
| 634 | return WalkResult::advance(); |
| 635 | |
| 636 | // Only these memory-writing ops are supported for now: |
| 637 | // memref.store, vector.transfer_write, vector.store |
| 638 | Value storeOpBase = getStoreOpTargetBuffer(op); |
| 639 | if (!storeOpBase) |
| 640 | return WalkResult::interrupt(); |
| 641 | |
| 642 | // Expect the base operand to be a Memref |
| 643 | MemrefValue storeOpBaseMemref = dyn_cast<MemrefValue>(Val&: storeOpBase); |
| 644 | if (!storeOpBaseMemref) |
| 645 | return WalkResult::interrupt(); |
| 646 | // Get the original memref buffer, skipping full view-like ops |
| 647 | Value buffer = memref::skipFullyAliasingOperations(source: storeOpBaseMemref); |
| 648 | bufferStoresInFirstPloop[buffer].push_back(Elt: op); |
| 649 | buffersWrittenInFirstPloop.insert(X: buffer); |
| 650 | return WalkResult::advance(); |
| 651 | }; |
| 652 | |
| 653 | // Walk the first parallel loop to collect all store/write ops and their |
| 654 | // target buffers |
| 655 | if (firstPloop.getBody()->walk(callback&: collectStoreOpsInWalk).wasInterrupted()) |
| 656 | return false; |
| 657 | |
| 658 | // Check that this load/read op encountered while walking the second parallel |
| 659 | // loop does not have incompatible data dependencies with the store/write ops |
| 660 | // collected from the first parallel loop: the loops can be fused only if in |
| 661 | // the 2nd loop there are no loads/stores from/to the buffers written in the |
| 662 | // 1st loop, except when on the same exact memory location (same indices) as |
| 663 | // written in the 1st loop. |
| 664 | auto checkLoadInWalkHasNoIncompatibleDataDeps = [&](Operation *loadOp) { |
| 665 | auto memOpInterf = dyn_cast_if_present<MemoryEffectOpInterface>(Val: loadOp); |
| 666 | // To be conservative, we should stop on ops that don't advertise their |
| 667 | // memory effects. However, many ops don't implement MemoryEffectOpInterface |
| 668 | // yet, so for now we just skip them. |
| 669 | // TODO: once more ops add MemoryEffectOpInterface, interrupt the walk here. |
| 670 | if (!memOpInterf && |
| 671 | !loadOp->hasTrait<mlir::OpTrait::HasRecursiveMemoryEffects>()) |
| 672 | return WalkResult::advance(); |
| 673 | // Ignore ops that don't read from memory, and wrapping ops that have nested |
| 674 | // memory effects (e.g. loops, conditionals) as they will be analyzed when |
| 675 | // visiting their nested ops. |
| 676 | if ((!memOpInterf && |
| 677 | loadOp->hasTrait<mlir::OpTrait::HasRecursiveMemoryEffects>()) || |
| 678 | (memOpInterf && !memOpInterf.hasEffect<MemoryEffects::Read>())) |
| 679 | return WalkResult::advance(); |
| 680 | // Support only these memory-reading ops for now |
| 681 | if (!isa<memref::LoadOp, vector::TransferReadOp, vector::LoadOp>(Val: loadOp) || |
| 682 | !isa<MemrefValue>(Val: loadOp->getOperand(idx: 0))) |
| 683 | return WalkResult::interrupt(); |
| 684 | |
| 685 | MemrefValue loadOpBase = cast<MemrefValue>(Val: loadOp->getOperand(idx: 0)); |
| 686 | MemrefValue loadedOrigBuf = memref::skipFullyAliasingOperations(source: loadOpBase); |
| 687 | |
| 688 | for (Value storedMem : buffersWrittenInFirstPloop) |
| 689 | if ((storedMem != loadedOrigBuf) && mayAlias(storedMem, loadedOrigBuf) && |
| 690 | !llvm::all_of(Range&: bufferStoresInFirstPloop[storedMem], |
| 691 | P: [&](Operation *storeOp) { |
| 692 | return canResolveAlias(loadOp, storeOp, |
| 693 | loopsIVsMap: firstToSecondPloopIndices); |
| 694 | })) { |
| 695 | return WalkResult::interrupt(); |
| 696 | } |
| 697 | |
| 698 | auto writeOpsIt = bufferStoresInFirstPloop.find(Val: loadedOrigBuf); |
| 699 | if (writeOpsIt == bufferStoresInFirstPloop.end()) |
| 700 | return WalkResult::advance(); |
| 701 | // Store/write ops to this buffer in the firstPloop |
| 702 | SmallVector<mlir::Operation *> &writeOps = writeOpsIt->second; |
| 703 | |
| 704 | // If the first loop has no writes to this buffer, continue |
| 705 | if (writeOps.empty()) |
| 706 | return WalkResult::advance(); |
| 707 | |
| 708 | Operation *writeOp = writeOps.front(); |
| 709 | |
| 710 | // In the first parallel loop, multiple writes to the same memref are |
| 711 | // allowed only on the same memory location |
| 712 | if (!llvm::all_of(Range&: writeOps, P: [&](Operation *otherWriteOp) { |
| 713 | return opsWriteSameMemLocation(op1: writeOp, op2: otherWriteOp); |
| 714 | })) { |
| 715 | return WalkResult::interrupt(); |
| 716 | } |
| 717 | |
| 718 | // Check that the load in secondPloop reads from the same memory location as |
| 719 | // written by the corresponding store in firstPloop |
| 720 | if (!loadsFromSameMemoryLocationWrittenBy(loadOp, storeOp: writeOp, |
| 721 | firstToSecondPloopIVsMap: firstToSecondPloopIndices, b)) { |
| 722 | return WalkResult::interrupt(); |
| 723 | } |
| 724 | |
| 725 | return WalkResult::advance(); |
| 726 | }; |
| 727 | |
| 728 | // Walk the second parallel loop to check load/read ops against the stores |
| 729 | // collected from the first parallel loop. |
| 730 | return !secondPloop.getBody() |
| 731 | ->walk(callback&: checkLoadInWalkHasNoIncompatibleDataDeps) |
| 732 | .wasInterrupted(); |
| 733 | } |
| 734 | |
| 735 | /// Check that in each loop there are no read ops on the buffers written |
| 736 | /// by the other loop, except when reading from the same exact memory location |
| 737 | /// (same indices) as written in the other loop. |
| 738 | static bool |
| 739 | noIncompatibleDataDependencies(ParallelOp firstPloop, ParallelOp secondPloop, |
| 740 | const IRMapping &firstToSecondPloopIndices, |
| 741 | llvm::function_ref<bool(Value, Value)> mayAlias, |
| 742 | OpBuilder &b) { |
| 743 | if (!haveNoDataDependenciesExceptSameIndex( |
| 744 | firstPloop, secondPloop, firstToSecondPloopIndices, mayAlias, b)) |
| 745 | return false; |
| 746 | |
| 747 | IRMapping secondToFirstPloopIndices; |
| 748 | secondToFirstPloopIndices.map(from: secondPloop.getBody()->getArguments(), |
| 749 | to: firstPloop.getBody()->getArguments()); |
| 750 | return haveNoDataDependenciesExceptSameIndex( |
| 751 | firstPloop: secondPloop, secondPloop: firstPloop, firstToSecondPloopIndices: secondToFirstPloopIndices, mayAlias, b); |
| 752 | } |
| 753 | |
| 754 | /// Check if fusion of the two parallel loops is legal: |
| 755 | /// i.e. no nested parallel loops, equal iteration spaces, |
| 756 | /// and no incompatible data dependencies between the loops. |
| 757 | static bool isFusionLegal(ParallelOp firstPloop, ParallelOp secondPloop, |
| 758 | const IRMapping &firstToSecondPloopIndices, |
| 759 | llvm::function_ref<bool(Value, Value)> mayAlias, |
| 760 | OpBuilder &b) { |
| 761 | if (hasNestedParallelOp(ploop: firstPloop) || hasNestedParallelOp(ploop: secondPloop) || |
| 762 | !equalIterationSpaces(firstPloop, secondPloop) || |
| 763 | !noIncompatibleDataDependencies(firstPloop, secondPloop, |
| 764 | firstToSecondPloopIndices, mayAlias, b)) |
| 765 | return false; |
| 766 | |
| 767 | // We are fusing first loop into second, make sure there are no users of the |
| 768 | // first loop results between loops. |
| 769 | DominanceInfo dom; |
| 770 | for (Operation *user : firstPloop->getUsers()) { |
| 771 | if (!dom.properlyDominates(a: secondPloop, b: user, /*enclosingOpOk*/ false)) |
| 772 | return false; |
| 773 | } |
| 774 | return true; |
| 775 | } |
| 776 | |
| 777 | // Returns new parallel loop where two loops matching indices param are |
| 778 | // interchanged |
| 779 | static std::optional<ParallelOp> |
| 780 | interchangeLoops(OpBuilder &builder, ParallelOp &loop, |
| 781 | const ArrayRef<int64_t> &indices) { |
| 782 | assert(loop.getNumLoops() == indices.size()); |
| 783 | if (loop.getNumLoops() < 2) |
| 784 | return std::nullopt; |
| 785 | |
| 786 | // Replace the parallel loop with the same parallel loop. |
| 787 | builder.setInsertionPoint(loop); |
| 788 | SmallVector<Value> newLB = |
| 789 | applyPermutation(input: SmallVector<Value>(loop.getLowerBound()), permutation: indices); |
| 790 | SmallVector<Value> newUB = |
| 791 | applyPermutation(input: SmallVector<Value>(loop.getUpperBound()), permutation: indices); |
| 792 | SmallVector<Value> newStep = |
| 793 | applyPermutation(input: SmallVector<Value>(loop.getStep()), permutation: indices); |
| 794 | auto newOp = |
| 795 | ParallelOp::create(builder, location: loop.getLoc(), lowerBounds: newLB, upperBounds: newUB, steps: newStep, |
| 796 | initVals: loop.getInitVals(), bodyBuilderFn: nullptr, unsignedCmp: loop.getUnsignedCmp()); |
| 797 | auto ivs = loop.getInductionVars(); |
| 798 | SmallVector<Value> newIvs = applyPermutation( |
| 799 | input: newOp.getInductionVars(), permutation: invertPermutationVector(permutation: indices)); |
| 800 | IRMapping mapping; |
| 801 | for (auto [iv, riv] : llvm::zip(t&: ivs, u&: newIvs)) { |
| 802 | mapping.map(from: iv, to: riv); |
| 803 | } |
| 804 | |
| 805 | // Copy parallel loop body |
| 806 | auto b = OpBuilder::atBlockBegin(block: newOp.getBody()); |
| 807 | for (auto &o : loop.getNumReductions() |
| 808 | ? loop.getBodyRegion().front() |
| 809 | : loop.getBodyRegion().front().without_terminator()) { |
| 810 | b.clone(op&: o, mapper&: mapping); |
| 811 | } |
| 812 | return newOp; |
| 813 | } |
| 814 | |
| 815 | struct LoopIV { |
| 816 | Value lBound, uBound, step; |
| 817 | bool operator!=(LoopIV const &other) const { return !(*this == other); } |
| 818 | bool operator==(LoopIV const &other) const { |
| 819 | return lBound == other.lBound && uBound == other.uBound && |
| 820 | step == other.step; |
| 821 | } |
| 822 | }; |
| 823 | |
| 824 | template <> |
| 825 | struct llvm::DenseMapInfo<LoopIV> { |
| 826 | static inline bool isEqual(const LoopIV &lhs, const LoopIV &rhs) { |
| 827 | return (lhs == rhs); |
| 828 | } |
| 829 | |
| 830 | static inline unsigned getHashValue(const LoopIV &val) { |
| 831 | return llvm::hash_combine( |
| 832 | args: DenseMapInfo<mlir::Value>::getHashValue(val: val.lBound), |
| 833 | args: DenseMapInfo<mlir::Value>::getHashValue(val: val.uBound), |
| 834 | args: DenseMapInfo<mlir::Value>::getHashValue(val: val.step)); |
| 835 | } |
| 836 | }; |
| 837 | |
| 838 | // Returns vector of candidate permutation indices vectors, |
| 839 | // can be empty. Caps the number of extra candidate permutations |
| 840 | // explored to avoid combinatorial explosion. This makes the search |
| 841 | // intentionally incomplete. |
| 842 | static SmallVector<SmallVector<int64_t>> |
| 843 | computeCandidateInterchangePermutations(ParallelOp &firstPloop, |
| 844 | ParallelOp &secondPloop, |
| 845 | int permBudget = 120) { |
| 846 | // Check preconditions |
| 847 | if (firstPloop.getNumLoops() < 2 || |
| 848 | firstPloop.getNumLoops() != secondPloop.getNumLoops()) |
| 849 | return {}; |
| 850 | |
| 851 | SmallVector<LoopIV> firstIVs(firstPloop.getNumLoops()); |
| 852 | SmallVector<LoopIV> secondIVs(secondPloop.getNumLoops()); |
| 853 | llvm::SmallSetVector<LoopIV, 6> unique; |
| 854 | for (unsigned index : llvm::seq(Size: firstPloop.getNumLoops())) { |
| 855 | firstIVs[index].lBound = firstPloop.getLowerBound()[index]; |
| 856 | firstIVs[index].uBound = firstPloop.getUpperBound()[index]; |
| 857 | firstIVs[index].step = firstPloop.getStep()[index]; |
| 858 | secondIVs[index].lBound = secondPloop.getLowerBound()[index]; |
| 859 | secondIVs[index].uBound = secondPloop.getUpperBound()[index]; |
| 860 | secondIVs[index].step = secondPloop.getStep()[index]; |
| 861 | unique.insert(X: firstIVs[index]); |
| 862 | } |
| 863 | |
| 864 | SmallVector<bool> diffIVs(firstPloop.getNumLoops()); |
| 865 | llvm::transform( |
| 866 | Range: llvm::zip(t&: firstIVs, u&: secondIVs), d_first: diffIVs.begin(), |
| 867 | F: [](auto const &pair) { return std::get<0>(pair) != std::get<1>(pair); }); |
| 868 | |
| 869 | SmallVector<int64_t> indices; |
| 870 | for (auto [idx, val] : enumerate(First&: diffIVs)) |
| 871 | if (val) |
| 872 | indices.push_back(Elt: idx); |
| 873 | |
| 874 | // Not a permutation shortcut |
| 875 | if (indices.size() == 1) |
| 876 | return {}; |
| 877 | |
| 878 | // Initialize with identity permutations |
| 879 | SmallVector<int64_t> basic(firstIVs.size()); |
| 880 | std::iota(first: basic.begin(), last: basic.end(), value: 0); |
| 881 | |
| 882 | if (indices.empty() && unique.size() == firstIVs.size()) |
| 883 | return {}; |
| 884 | |
| 885 | if (indices.size() > 1) { |
| 886 | // Determine whether the iteration space of the first loop is a permutation |
| 887 | // of the second and collect remaps. |
| 888 | SmallVector<int64_t> remaps; |
| 889 | for (auto fIdx : indices) { |
| 890 | for (auto sIdx : indices) { |
| 891 | // can be remapped |
| 892 | if (fIdx != sIdx && firstIVs[fIdx] == secondIVs[sIdx] && |
| 893 | remaps.end() == std::find(first: remaps.begin(), last: remaps.end(), val: sIdx)) { |
| 894 | remaps.push_back(Elt: sIdx); |
| 895 | break; |
| 896 | } |
| 897 | } |
| 898 | } |
| 899 | |
| 900 | // Not a permutation |
| 901 | if (indices.size() != remaps.size()) |
| 902 | return {}; |
| 903 | |
| 904 | // compose permutation indices |
| 905 | for (auto [from, to] : zip(t&: indices, u&: remaps)) { |
| 906 | basic[from] = to; |
| 907 | } |
| 908 | |
| 909 | LDBG() << "Collected basic permutations: " |
| 910 | << llvm::interleaved_array(R: basic); |
| 911 | |
| 912 | // All axes are unique, no further permutatons needed |
| 913 | if (unique.size() == firstIVs.size()) { |
| 914 | return {basic}; |
| 915 | } |
| 916 | } |
| 917 | |
| 918 | // |
| 919 | // Permute equal axes |
| 920 | assert(unique.size() != firstIVs.size() && |
| 921 | "Expected at least two equal axes" ); |
| 922 | |
| 923 | // Collect equal axes to groups |
| 924 | SmallVector<SmallVector<int64_t>> {basic}; |
| 925 | SmallVector<SmallVector<int64_t>> groups; |
| 926 | for (auto iv : unique) { |
| 927 | SmallVector<int64_t> group; |
| 928 | for (unsigned index : llvm::seq(Size: firstIVs.size())) { |
| 929 | if (firstIVs[index] == iv) |
| 930 | group.push_back(Elt: index); |
| 931 | } |
| 932 | if (group.size() > 1) |
| 933 | groups.push_back(Elt: std::move(group)); |
| 934 | } |
| 935 | |
| 936 | // Permute axes groups |
| 937 | SmallVector<SmallVector<int64_t>> rmpdGroups(groups); |
| 938 | bool repeat = true; |
| 939 | while (repeat && permBudget) { |
| 940 | repeat = false; |
| 941 | for (auto const &[group, groupRemaps] : zip(t&: groups, u&: rmpdGroups)) { |
| 942 | repeat |= std::next_permutation(first: groupRemaps.begin(), last: groupRemaps.end()); |
| 943 | if (repeat) |
| 944 | break; |
| 945 | } |
| 946 | |
| 947 | if (repeat) { |
| 948 | SmallVector<int64_t> (basic); |
| 949 | for (auto const &[group, groupRemaps] : zip(t&: groups, u&: rmpdGroups)) { |
| 950 | for (auto [from, to] : zip(t&: group, u&: groupRemaps)) |
| 951 | extra[from] = basic[to]; |
| 952 | } |
| 953 | if (basic != extra) { |
| 954 | LDBG() << "Collected extra permutations: " |
| 955 | << llvm::interleaved_array(R: extra); |
| 956 | |
| 957 | extraResults.push_back(Elt: std::move(extra)); |
| 958 | permBudget--; |
| 959 | } |
| 960 | } |
| 961 | } |
| 962 | |
| 963 | return extraResults; |
| 964 | } |
| 965 | |
| 966 | /// Prepend operations of firstPloop's body into secondPloop's body. |
| 967 | /// Update secondPloop with new loop. |
| 968 | static void applyLoopFusion(ParallelOp &firstPloop, ParallelOp &secondPloop, |
| 969 | OpBuilder &builder) { |
| 970 | Block *block1 = firstPloop.getBody(); |
| 971 | Block *block2 = secondPloop.getBody(); |
| 972 | ValueRange inits1 = firstPloop.getInitVals(); |
| 973 | ValueRange inits2 = secondPloop.getInitVals(); |
| 974 | |
| 975 | SmallVector<Value> newInitVars(inits1.begin(), inits1.end()); |
| 976 | newInitVars.append(in_start: inits2.begin(), in_end: inits2.end()); |
| 977 | |
| 978 | IRRewriter b(builder); |
| 979 | b.setInsertionPoint(secondPloop); |
| 980 | auto newSecondPloop = |
| 981 | ParallelOp::create(builder&: b, location: secondPloop.getLoc(), lowerBounds: secondPloop.getLowerBound(), |
| 982 | upperBounds: secondPloop.getUpperBound(), steps: secondPloop.getStep(), |
| 983 | initVals: newInitVars, bodyBuilderFn: nullptr, unsignedCmp: secondPloop.getUnsignedCmp()); |
| 984 | |
| 985 | Block *newBlock = newSecondPloop.getBody(); |
| 986 | auto term1 = cast<ReduceOp>(Val: block1->getTerminator()); |
| 987 | auto term2 = cast<ReduceOp>(Val: block2->getTerminator()); |
| 988 | |
| 989 | b.inlineBlockBefore(source: block2, dest: newBlock, before: newBlock->begin(), |
| 990 | argValues: newBlock->getArguments()); |
| 991 | b.inlineBlockBefore(source: block1, dest: newBlock, before: newBlock->begin(), |
| 992 | argValues: newBlock->getArguments()); |
| 993 | |
| 994 | ValueRange results = newSecondPloop.getResults(); |
| 995 | if (!results.empty()) { |
| 996 | b.setInsertionPointToEnd(newBlock); |
| 997 | |
| 998 | ValueRange reduceArgs1 = term1.getOperands(); |
| 999 | ValueRange reduceArgs2 = term2.getOperands(); |
| 1000 | SmallVector<Value> newReduceArgs(reduceArgs1.begin(), reduceArgs1.end()); |
| 1001 | newReduceArgs.append(in_start: reduceArgs2.begin(), in_end: reduceArgs2.end()); |
| 1002 | |
| 1003 | auto newReduceOp = scf::ReduceOp::create(builder&: b, location: term2.getLoc(), operands: newReduceArgs); |
| 1004 | |
| 1005 | for (auto &&[i, reg] : llvm::enumerate(First: llvm::concat<Region>( |
| 1006 | Ranges: term1.getReductions(), Ranges: term2.getReductions()))) { |
| 1007 | Block &oldRedBlock = reg.front(); |
| 1008 | Block &newRedBlock = newReduceOp.getReductions()[i].front(); |
| 1009 | b.inlineBlockBefore(source: &oldRedBlock, dest: &newRedBlock, before: newRedBlock.begin(), |
| 1010 | argValues: newRedBlock.getArguments()); |
| 1011 | } |
| 1012 | |
| 1013 | firstPloop.replaceAllUsesWith(values: results.take_front(n: inits1.size())); |
| 1014 | secondPloop.replaceAllUsesWith(values: results.take_back(n: inits2.size())); |
| 1015 | } |
| 1016 | term1->erase(); |
| 1017 | term2->erase(); |
| 1018 | firstPloop.erase(); |
| 1019 | secondPloop.erase(); |
| 1020 | secondPloop = newSecondPloop; |
| 1021 | } |
| 1022 | |
| 1023 | /// Check fusion pre-conditions and call fusion if it is possible |
| 1024 | static void fuseIfLegal(ParallelOp firstPloop, ParallelOp &secondPloop, |
| 1025 | OpBuilder builder, |
| 1026 | llvm::function_ref<bool(Value, Value)> mayAlias) { |
| 1027 | Block *block1 = firstPloop.getBody(); |
| 1028 | Block *block2 = secondPloop.getBody(); |
| 1029 | IRMapping firstToSecondPloopIndices; |
| 1030 | firstToSecondPloopIndices.map(from: block1->getArguments(), to: block2->getArguments()); |
| 1031 | |
| 1032 | if (isFusionLegal(firstPloop, secondPloop, firstToSecondPloopIndices, |
| 1033 | mayAlias, b&: builder)) { |
| 1034 | applyLoopFusion(firstPloop, secondPloop, builder); |
| 1035 | return; |
| 1036 | } |
| 1037 | |
| 1038 | // If iteration space of the second parallel loop is a permutation of the |
| 1039 | // first one then interchange iteration space of the second parallel loop |
| 1040 | // and re-asses possibility of fusion. |
| 1041 | for (auto &perms : |
| 1042 | computeCandidateInterchangePermutations(firstPloop, secondPloop)) { |
| 1043 | OpBuilder::InsertionGuard guard(builder); |
| 1044 | LDBG() << "Applied permutation: " << llvm::interleaved_array(R: perms); |
| 1045 | |
| 1046 | auto newLoop = interchangeLoops(builder, loop&: secondPloop, indices: perms); |
| 1047 | firstToSecondPloopIndices.clear(); |
| 1048 | firstToSecondPloopIndices.map(from: block1->getArguments(), |
| 1049 | to: newLoop->getBody()->getArguments()); |
| 1050 | if (!isFusionLegal(firstPloop, secondPloop: *newLoop, firstToSecondPloopIndices, |
| 1051 | mayAlias, b&: builder)) { |
| 1052 | LDBG() << "Rejected: " << newLoop; |
| 1053 | |
| 1054 | newLoop->erase(); |
| 1055 | continue; |
| 1056 | } |
| 1057 | |
| 1058 | secondPloop.replaceAllUsesWith(values: newLoop->getResults()); |
| 1059 | secondPloop->erase(); |
| 1060 | secondPloop = *newLoop; |
| 1061 | applyLoopFusion(firstPloop, secondPloop, builder); |
| 1062 | break; |
| 1063 | } |
| 1064 | } |
| 1065 | |
| 1066 | void mlir::scf::naivelyFuseParallelOps( |
| 1067 | Region ®ion, llvm::function_ref<bool(Value, Value)> mayAlias) { |
| 1068 | OpBuilder b(region); |
| 1069 | // Consider every single block and attempt to fuse adjacent loops. |
| 1070 | SmallVector<SmallVector<ParallelOp>, 1> ploopChains; |
| 1071 | for (auto &block : region) { |
| 1072 | ploopChains.clear(); |
| 1073 | ploopChains.push_back(Elt: {}); |
| 1074 | |
| 1075 | // Not using `walk()` to traverse only top-level parallel loops and also |
| 1076 | // make sure that there are no side-effecting ops between the parallel |
| 1077 | // loops. |
| 1078 | bool noSideEffects = true; |
| 1079 | for (auto &op : block) { |
| 1080 | if (auto ploop = dyn_cast<ParallelOp>(Val&: op)) { |
| 1081 | if (noSideEffects) { |
| 1082 | ploopChains.back().push_back(Elt: ploop); |
| 1083 | } else { |
| 1084 | ploopChains.push_back(Elt: {ploop}); |
| 1085 | noSideEffects = true; |
| 1086 | } |
| 1087 | continue; |
| 1088 | } |
| 1089 | // TODO: Handle region side effects properly. |
| 1090 | noSideEffects &= isMemoryEffectFree(op: &op) && op.getNumRegions() == 0; |
| 1091 | } |
| 1092 | for (MutableArrayRef<ParallelOp> ploops : ploopChains) { |
| 1093 | for (int i = 0, e = ploops.size(); i + 1 < e; ++i) |
| 1094 | fuseIfLegal(firstPloop: ploops[i], secondPloop&: ploops[i + 1], builder: b, mayAlias); |
| 1095 | } |
| 1096 | } |
| 1097 | } |
| 1098 | |
| 1099 | namespace { |
| 1100 | struct ParallelLoopFusion |
| 1101 | : public impl::SCFParallelLoopFusionBase<ParallelLoopFusion> { |
| 1102 | void runOnOperation() override { |
| 1103 | auto &aa = getAnalysis<AliasAnalysis>(); |
| 1104 | |
| 1105 | auto mayAlias = [&](Value val1, Value val2) -> bool { |
| 1106 | // If the memref is defined in one of the parallel loops body, careful |
| 1107 | // alias analysis is needed. |
| 1108 | // TODO: check if this is still needed as a separate check. |
| 1109 | auto val1Def = val1.getDefiningOp(); |
| 1110 | auto val2Def = val2.getDefiningOp(); |
| 1111 | auto val1Loop = |
| 1112 | val1Def ? val1Def->getParentOfType<ParallelOp>() : nullptr; |
| 1113 | auto val2Loop = |
| 1114 | val2Def ? val2Def->getParentOfType<ParallelOp>() : nullptr; |
| 1115 | if (val1Loop != val2Loop) |
| 1116 | return true; |
| 1117 | |
| 1118 | return !aa.alias(lhs: val1, rhs: val2).isNo(); |
| 1119 | }; |
| 1120 | |
| 1121 | getOperation()->walk(callback: [&](Operation *child) { |
| 1122 | for (Region ®ion : child->getRegions()) |
| 1123 | naivelyFuseParallelOps(region, mayAlias); |
| 1124 | }); |
| 1125 | } |
| 1126 | }; |
| 1127 | } // namespace |
| 1128 | |
| 1129 | std::unique_ptr<Pass> mlir::createParallelLoopFusionPass() { |
| 1130 | return std::make_unique<ParallelLoopFusion>(); |
| 1131 | } |
| 1132 | |