| 1 | //===- VectorTransferOpTransforms.cpp - transfer op transforms ------------===// |
| 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 functions concerned with optimizing transfer_read and |
| 10 | // transfer_write ops. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 15 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 16 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| 17 | #include "mlir/Dialect/MemRef/Utils/MemRefUtils.h" |
| 18 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 19 | #include "mlir/Dialect/Utils/IndexingUtils.h" |
| 20 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 21 | #include "mlir/Dialect/Vector/Transforms/LoweringPatterns.h" |
| 22 | #include "mlir/Dialect/Vector/Transforms/VectorTransforms.h" |
| 23 | #include "mlir/Dialect/Vector/Utils/VectorUtils.h" |
| 24 | #include "mlir/IR/BuiltinTypes.h" |
| 25 | #include "mlir/IR/Dominance.h" |
| 26 | #include "mlir/IR/Matchers.h" |
| 27 | #include "mlir/IR/Operation.h" |
| 28 | #include "mlir/Interfaces/SideEffectInterfaces.h" |
| 29 | #include "llvm/ADT/STLExtras.h" |
| 30 | #include "llvm/ADT/StringRef.h" |
| 31 | #include "llvm/Support/Casting.h" |
| 32 | #include "llvm/Support/DebugLog.h" |
| 33 | #include "llvm/Support/LogicalResult.h" |
| 34 | |
| 35 | #define DEBUG_TYPE "vector-transfer-opt" |
| 36 | |
| 37 | using namespace mlir; |
| 38 | |
| 39 | /// Return the ancestor op in the region or nullptr if the region is not |
| 40 | /// an ancestor of the op. |
| 41 | static Operation *findAncestorOpInRegion(Region *region, Operation *op) { |
| 42 | LDBG() << " Finding ancestor of " << *op << " in region" ; |
| 43 | for (; op != nullptr && op->getParentRegion() != region; |
| 44 | op = op->getParentOp()) |
| 45 | ; |
| 46 | if (op) { |
| 47 | LDBG() << " -> Ancestor: " << *op; |
| 48 | } else { |
| 49 | LDBG() << " -> Ancestor: nullptr" ; |
| 50 | } |
| 51 | return op; |
| 52 | } |
| 53 | |
| 54 | namespace { |
| 55 | |
| 56 | class TransferOptimization { |
| 57 | public: |
| 58 | TransferOptimization(RewriterBase &rewriter, Operation *op) |
| 59 | : rewriter(rewriter), dominators(op), postDominators(op) {} |
| 60 | void deadStoreOp(vector::TransferWriteOp); |
| 61 | void storeToLoadForwarding(vector::TransferReadOp); |
| 62 | void removeDeadOp() { |
| 63 | LDBG() << "Removing " << opToErase.size() << " dead operations" ; |
| 64 | for (Operation *op : opToErase) { |
| 65 | LDBG() << " -> Erasing: " << *op; |
| 66 | rewriter.eraseOp(op); |
| 67 | } |
| 68 | opToErase.clear(); |
| 69 | } |
| 70 | |
| 71 | private: |
| 72 | RewriterBase &rewriter; |
| 73 | bool isReachable(Operation *start, Operation *dest); |
| 74 | DominanceInfo dominators; |
| 75 | PostDominanceInfo postDominators; |
| 76 | std::vector<Operation *> opToErase; |
| 77 | }; |
| 78 | |
| 79 | } // namespace |
| 80 | /// Return true if there is a path from start operation to dest operation, |
| 81 | /// otherwise return false. The operations have to be in the same region. |
| 82 | bool TransferOptimization::isReachable(Operation *start, Operation *dest) { |
| 83 | LDBG() << " Checking reachability from " << *start << " to " << *dest; |
| 84 | assert(start->getParentRegion() == dest->getParentRegion() && |
| 85 | "This function only works for ops i the same region" ); |
| 86 | // Simple case where the start op dominate the destination. |
| 87 | if (dominators.dominates(a: start, b: dest)) { |
| 88 | LDBG() << " -> Start dominates dest, reachable" ; |
| 89 | return true; |
| 90 | } |
| 91 | bool blockReachable = start->getBlock()->isReachable(other: dest->getBlock()); |
| 92 | LDBG() << " -> Block reachable: " << blockReachable; |
| 93 | return blockReachable; |
| 94 | } |
| 95 | |
| 96 | /// For transfer_write to overwrite fully another transfer_write must: |
| 97 | /// 1. Access the same memref with the same indices and vector type. |
| 98 | /// 2. Post-dominate the other transfer_write operation. |
| 99 | /// If several candidates are available, one must be post-dominated by all the |
| 100 | /// others since they are all post-dominating the same transfer_write. We only |
| 101 | /// consider the transfer_write post-dominated by all the other candidates as |
| 102 | /// this will be the first transfer_write executed after the potentially dead |
| 103 | /// transfer_write. |
| 104 | /// If we found such an overwriting transfer_write we know that the original |
| 105 | /// transfer_write is dead if all reads that can be reached from the potentially |
| 106 | /// dead transfer_write are dominated by the overwriting transfer_write. |
| 107 | void TransferOptimization::deadStoreOp(vector::TransferWriteOp write) { |
| 108 | LDBG() << "=== Starting deadStoreOp analysis for: " << *write.getOperation(); |
| 109 | llvm::SmallVector<Operation *, 8> blockingAccesses; |
| 110 | Operation *firstOverwriteCandidate = nullptr; |
| 111 | Value source = memref::skipViewLikeOps(source: cast<MemrefValue>(Val: write.getBase())); |
| 112 | LDBG() << "Source memref (after skipping view-like ops): " << source; |
| 113 | llvm::SmallVector<Operation *, 32> users(source.getUsers().begin(), |
| 114 | source.getUsers().end()); |
| 115 | LDBG() << "Found " << users.size() << " users of source memref" ; |
| 116 | llvm::SmallDenseSet<Operation *, 32> processed; |
| 117 | while (!users.empty()) { |
| 118 | Operation *user = users.pop_back_val(); |
| 119 | LDBG() << "Processing user: " << *user; |
| 120 | // If the user has already been processed skip. |
| 121 | if (!processed.insert(V: user).second) { |
| 122 | LDBG() << " -> Already processed, skipping" ; |
| 123 | continue; |
| 124 | } |
| 125 | if (auto viewLike = dyn_cast<ViewLikeOpInterface>(Val: user)) { |
| 126 | LDBG() << " -> View-like operation, following to destination" ; |
| 127 | Value viewDest = viewLike.getViewDest(); |
| 128 | users.append(in_start: viewDest.getUsers().begin(), in_end: viewDest.getUsers().end()); |
| 129 | continue; |
| 130 | } |
| 131 | if (isMemoryEffectFree(op: user)) { |
| 132 | LDBG() << " -> Memory effect free, skipping" ; |
| 133 | continue; |
| 134 | } |
| 135 | if (user == write.getOperation()) { |
| 136 | LDBG() << " -> Same as write operation, skipping" ; |
| 137 | continue; |
| 138 | } |
| 139 | if (auto nextWrite = dyn_cast<vector::TransferWriteOp>(Val: user)) { |
| 140 | LDBG() << " -> Found transfer_write candidate: " << *nextWrite; |
| 141 | // Check candidate that can override the store. |
| 142 | bool sameView = memref::isSameViewOrTrivialAlias( |
| 143 | a: cast<MemrefValue>(Val: nextWrite.getBase()), |
| 144 | b: cast<MemrefValue>(Val: write.getBase())); |
| 145 | bool sameValue = checkSameValueWAW(write: nextWrite, priorWrite: write); |
| 146 | bool postDominates = postDominators.postDominates(a: nextWrite, b: write); |
| 147 | LDBG() << " -> Same view: " << sameView |
| 148 | << ", Same value: " << sameValue |
| 149 | << ", Post-dominates: " << postDominates; |
| 150 | |
| 151 | if (sameView && sameValue && postDominates) { |
| 152 | LDBG() << " -> Valid overwrite candidate found" ; |
| 153 | if (firstOverwriteCandidate == nullptr || |
| 154 | postDominators.postDominates(a: firstOverwriteCandidate, b: nextWrite)) { |
| 155 | LDBG() << " -> New first overwrite candidate: " << *nextWrite; |
| 156 | firstOverwriteCandidate = nextWrite; |
| 157 | } else { |
| 158 | LDBG() << " -> Keeping existing first overwrite candidate" ; |
| 159 | assert( |
| 160 | postDominators.postDominates(nextWrite, firstOverwriteCandidate)); |
| 161 | } |
| 162 | continue; |
| 163 | } |
| 164 | LDBG() << " -> Not a valid overwrite candidate" ; |
| 165 | } |
| 166 | if (auto transferOp = dyn_cast<VectorTransferOpInterface>(Val: user)) { |
| 167 | LDBG() << " -> Found vector transfer operation: " << *transferOp; |
| 168 | // Don't need to consider disjoint accesses. |
| 169 | bool isDisjoint = vector::isDisjointTransferSet( |
| 170 | transferA: cast<VectorTransferOpInterface>(Val: write.getOperation()), |
| 171 | transferB: cast<VectorTransferOpInterface>(Val: transferOp.getOperation()), |
| 172 | /*testDynamicValueUsingBounds=*/true); |
| 173 | LDBG() << " -> Is disjoint: " << isDisjoint; |
| 174 | if (isDisjoint) { |
| 175 | LDBG() << " -> Skipping disjoint access" ; |
| 176 | continue; |
| 177 | } |
| 178 | } |
| 179 | LDBG() << " -> Adding to blocking accesses: " << *user; |
| 180 | blockingAccesses.push_back(Elt: user); |
| 181 | } |
| 182 | LDBG() << "Finished processing users. Found " << blockingAccesses.size() |
| 183 | << " blocking accesses" ; |
| 184 | |
| 185 | if (firstOverwriteCandidate == nullptr) { |
| 186 | LDBG() << "No overwrite candidate found, store is not dead" ; |
| 187 | return; |
| 188 | } |
| 189 | |
| 190 | LDBG() << "First overwrite candidate: " << *firstOverwriteCandidate; |
| 191 | Region *topRegion = firstOverwriteCandidate->getParentRegion(); |
| 192 | Operation *writeAncestor = findAncestorOpInRegion(region: topRegion, op: write); |
| 193 | assert(writeAncestor && |
| 194 | "write op should be recursively part of the top region" ); |
| 195 | LDBG() << "Write ancestor in top region: " << *writeAncestor; |
| 196 | |
| 197 | LDBG() << "Checking " << blockingAccesses.size() |
| 198 | << " blocking accesses for reachability" ; |
| 199 | for (Operation *access : blockingAccesses) { |
| 200 | LDBG() << "Checking blocking access: " << *access; |
| 201 | Operation *accessAncestor = findAncestorOpInRegion(region: topRegion, op: access); |
| 202 | // TODO: if the access and write have the same ancestor we could recurse in |
| 203 | // the region to know if the access is reachable with more precision. |
| 204 | if (accessAncestor == nullptr) { |
| 205 | LDBG() << " -> No ancestor in top region, skipping" ; |
| 206 | continue; |
| 207 | } |
| 208 | |
| 209 | bool isReachableFromWrite = isReachable(start: writeAncestor, dest: accessAncestor); |
| 210 | LDBG() << " -> Is reachable from write: " << isReachableFromWrite; |
| 211 | if (!isReachableFromWrite) { |
| 212 | LDBG() << " -> Not reachable, skipping" ; |
| 213 | continue; |
| 214 | } |
| 215 | |
| 216 | bool overwriteDominatesAccess = |
| 217 | dominators.dominates(a: firstOverwriteCandidate, b: accessAncestor); |
| 218 | LDBG() << " -> Overwrite dominates access: " << overwriteDominatesAccess; |
| 219 | if (!overwriteDominatesAccess) { |
| 220 | LDBG() << "Store may not be dead due to op: " << *accessAncestor; |
| 221 | return; |
| 222 | } |
| 223 | LDBG() << " -> Access is dominated by overwrite, continuing" ; |
| 224 | } |
| 225 | LDBG() << "Found dead store: " << *write.getOperation() |
| 226 | << " overwritten by: " << *firstOverwriteCandidate; |
| 227 | opToErase.push_back(x: write.getOperation()); |
| 228 | } |
| 229 | |
| 230 | /// A transfer_write candidate to storeToLoad forwarding must: |
| 231 | /// 1. Access the same memref with the same indices and vector type as the |
| 232 | /// transfer_read. |
| 233 | /// 2. Dominate the transfer_read operation. |
| 234 | /// If several candidates are available, one must be dominated by all the others |
| 235 | /// since they are all dominating the same transfer_read. We only consider the |
| 236 | /// transfer_write dominated by all the other candidates as this will be the |
| 237 | /// last transfer_write executed before the transfer_read. |
| 238 | /// If we found such a candidate we can do the forwarding if all the other |
| 239 | /// potentially aliasing ops that may reach the transfer_read are post-dominated |
| 240 | /// by the transfer_write. |
| 241 | void TransferOptimization::storeToLoadForwarding(vector::TransferReadOp read) { |
| 242 | LDBG() << "=== Starting storeToLoadForwarding analysis for: " |
| 243 | << *read.getOperation(); |
| 244 | if (read.hasOutOfBoundsDim()) { |
| 245 | LDBG() << "Read has out-of-bounds dimensions, skipping" ; |
| 246 | return; |
| 247 | } |
| 248 | SmallVector<Operation *, 8> blockingWrites; |
| 249 | vector::TransferWriteOp lastwrite = nullptr; |
| 250 | Value source = memref::skipViewLikeOps(source: cast<MemrefValue>(Val: read.getBase())); |
| 251 | LDBG() << "Source memref (after skipping view-like ops): " << source; |
| 252 | llvm::SmallVector<Operation *, 32> users(source.getUsers().begin(), |
| 253 | source.getUsers().end()); |
| 254 | LDBG() << "Found " << users.size() << " users of source memref" ; |
| 255 | llvm::SmallDenseSet<Operation *, 32> processed; |
| 256 | while (!users.empty()) { |
| 257 | Operation *user = users.pop_back_val(); |
| 258 | LDBG() << "Processing user: " << *user; |
| 259 | // If the user has already been processed skip. |
| 260 | if (!processed.insert(V: user).second) { |
| 261 | LDBG() << " -> Already processed, skipping" ; |
| 262 | continue; |
| 263 | } |
| 264 | if (auto viewLike = dyn_cast<ViewLikeOpInterface>(Val: user)) { |
| 265 | LDBG() << " -> View-like operation, following to destination" ; |
| 266 | Value viewDest = viewLike.getViewDest(); |
| 267 | users.append(in_start: viewDest.getUsers().begin(), in_end: viewDest.getUsers().end()); |
| 268 | continue; |
| 269 | } |
| 270 | if (isMemoryEffectFree(op: user) || isa<vector::TransferReadOp>(Val: user)) { |
| 271 | LDBG() << " -> Memory effect free or transfer_read, skipping" ; |
| 272 | continue; |
| 273 | } |
| 274 | if (auto write = dyn_cast<vector::TransferWriteOp>(Val: user)) { |
| 275 | LDBG() << " -> Found transfer_write candidate: " << *write; |
| 276 | // If there is a write, but we can prove that it is disjoint we can ignore |
| 277 | // the write. |
| 278 | bool isDisjoint = vector::isDisjointTransferSet( |
| 279 | transferA: cast<VectorTransferOpInterface>(Val: write.getOperation()), |
| 280 | transferB: cast<VectorTransferOpInterface>(Val: read.getOperation()), |
| 281 | /*testDynamicValueUsingBounds=*/true); |
| 282 | LDBG() << " -> Is disjoint: " << isDisjoint; |
| 283 | if (isDisjoint) { |
| 284 | LDBG() << " -> Skipping disjoint write" ; |
| 285 | continue; |
| 286 | } |
| 287 | |
| 288 | bool sameView = |
| 289 | memref::isSameViewOrTrivialAlias(a: cast<MemrefValue>(Val: read.getBase()), |
| 290 | b: cast<MemrefValue>(Val: write.getBase())); |
| 291 | bool dominates = dominators.dominates(a: write, b: read); |
| 292 | bool sameValue = checkSameValueRAW(defWrite: write, read); |
| 293 | LDBG() << " -> Same view: " << sameView << ", Dominates: " << dominates |
| 294 | << ", Same value: " << sameValue; |
| 295 | |
| 296 | if (sameView && dominates && sameValue) { |
| 297 | LDBG() << " -> Valid forwarding candidate found" ; |
| 298 | if (lastwrite == nullptr || dominators.dominates(a: lastwrite, b: write)) { |
| 299 | LDBG() << " -> New last write candidate: " << *write; |
| 300 | lastwrite = write; |
| 301 | } else { |
| 302 | LDBG() << " -> Keeping existing last write candidate" ; |
| 303 | assert(dominators.dominates(write, lastwrite)); |
| 304 | } |
| 305 | continue; |
| 306 | } |
| 307 | LDBG() << " -> Not a valid forwarding candidate" ; |
| 308 | } |
| 309 | LDBG() << " -> Adding to blocking writes: " << *user; |
| 310 | blockingWrites.push_back(Elt: user); |
| 311 | } |
| 312 | LDBG() << "Finished processing users. Found " << blockingWrites.size() |
| 313 | << " blocking writes" ; |
| 314 | |
| 315 | if (lastwrite == nullptr) { |
| 316 | LDBG() << "No last write candidate found, cannot forward" ; |
| 317 | return; |
| 318 | } |
| 319 | |
| 320 | LDBG() << "Last write candidate: " << *lastwrite; |
| 321 | Region *topRegion = lastwrite->getParentRegion(); |
| 322 | Operation *readAncestor = findAncestorOpInRegion(region: topRegion, op: read); |
| 323 | assert(readAncestor && |
| 324 | "read op should be recursively part of the top region" ); |
| 325 | LDBG() << "Read ancestor in top region: " << *readAncestor; |
| 326 | |
| 327 | LDBG() << "Checking " << blockingWrites.size() |
| 328 | << " blocking writes for post-dominance" ; |
| 329 | for (Operation *write : blockingWrites) { |
| 330 | LDBG() << "Checking blocking write: " << *write; |
| 331 | Operation *writeAncestor = findAncestorOpInRegion(region: topRegion, op: write); |
| 332 | if (writeAncestor) { |
| 333 | LDBG() << " -> Write ancestor: " << *writeAncestor; |
| 334 | } else { |
| 335 | LDBG() << " -> Write ancestor: nullptr" ; |
| 336 | } |
| 337 | |
| 338 | // TODO: if the store and read have the same ancestor we could recurse in |
| 339 | // the region to know if the read is reachable with more precision. |
| 340 | if (writeAncestor == nullptr) { |
| 341 | LDBG() << " -> No ancestor in top region, skipping" ; |
| 342 | continue; |
| 343 | } |
| 344 | |
| 345 | bool isReachableToRead = isReachable(start: writeAncestor, dest: readAncestor); |
| 346 | LDBG() << " -> Is reachable to read: " << isReachableToRead; |
| 347 | if (!isReachableToRead) { |
| 348 | LDBG() << " -> Not reachable, skipping" ; |
| 349 | continue; |
| 350 | } |
| 351 | |
| 352 | bool lastWritePostDominates = |
| 353 | postDominators.postDominates(a: lastwrite, b: write); |
| 354 | LDBG() << " -> Last write post-dominates blocking write: " |
| 355 | << lastWritePostDominates; |
| 356 | if (!lastWritePostDominates) { |
| 357 | LDBG() << "Fail to do write to read forwarding due to op: " << *write; |
| 358 | return; |
| 359 | } |
| 360 | LDBG() << " -> Blocking write is post-dominated, continuing" ; |
| 361 | } |
| 362 | |
| 363 | LDBG() << "Forward value from " << *lastwrite.getOperation() |
| 364 | << " to: " << *read.getOperation(); |
| 365 | read.replaceAllUsesWith(newValue: lastwrite.getVector()); |
| 366 | opToErase.push_back(x: read.getOperation()); |
| 367 | } |
| 368 | |
| 369 | /// Converts OpFoldResults to int64_t shape without unit dims. |
| 370 | static SmallVector<int64_t> getReducedShape(ArrayRef<OpFoldResult> mixedSizes) { |
| 371 | SmallVector<int64_t> reducedShape; |
| 372 | for (const auto size : mixedSizes) { |
| 373 | if (llvm::dyn_cast_if_present<Value>(Val: size)) { |
| 374 | reducedShape.push_back(Elt: ShapedType::kDynamic); |
| 375 | continue; |
| 376 | } |
| 377 | |
| 378 | auto value = cast<IntegerAttr>(Val: cast<Attribute>(Val: size)).getValue(); |
| 379 | if (value == 1) |
| 380 | continue; |
| 381 | reducedShape.push_back(Elt: value.getSExtValue()); |
| 382 | } |
| 383 | return reducedShape; |
| 384 | } |
| 385 | |
| 386 | /// Drops unit dimensions from the input MemRefType. |
| 387 | static MemRefType dropUnitDims(MemRefType inputType, |
| 388 | ArrayRef<OpFoldResult> offsets, |
| 389 | ArrayRef<OpFoldResult> sizes, |
| 390 | ArrayRef<OpFoldResult> strides) { |
| 391 | auto targetShape = getReducedShape(mixedSizes: sizes); |
| 392 | MemRefType rankReducedType = memref::SubViewOp::inferRankReducedResultType( |
| 393 | resultShape: targetShape, sourceMemRefType: inputType, staticOffsets: offsets, staticSizes: sizes, staticStrides: strides); |
| 394 | return rankReducedType.canonicalizeStridedLayout(); |
| 395 | } |
| 396 | |
| 397 | /// Creates a rank-reducing memref.subview op that drops unit dims from its |
| 398 | /// input. Or just returns the input if it was already without unit dims. |
| 399 | static Value rankReducingSubviewDroppingUnitDims(PatternRewriter &rewriter, |
| 400 | mlir::Location loc, |
| 401 | Value input) { |
| 402 | MemRefType inputType = cast<MemRefType>(Val: input.getType()); |
| 403 | SmallVector<OpFoldResult> offsets(inputType.getRank(), |
| 404 | rewriter.getIndexAttr(value: 0)); |
| 405 | SmallVector<OpFoldResult> sizes = memref::getMixedSizes(builder&: rewriter, loc, value: input); |
| 406 | SmallVector<OpFoldResult> strides(inputType.getRank(), |
| 407 | rewriter.getIndexAttr(value: 1)); |
| 408 | MemRefType resultType = dropUnitDims(inputType, offsets, sizes, strides); |
| 409 | |
| 410 | if (resultType.canonicalizeStridedLayout() == |
| 411 | inputType.canonicalizeStridedLayout()) |
| 412 | return input; |
| 413 | return memref::SubViewOp::create(builder&: rewriter, location: loc, resultType, source: input, offsets, |
| 414 | sizes, strides); |
| 415 | } |
| 416 | |
| 417 | /// Returns the number of dims that aren't unit dims. |
| 418 | static int getReducedRank(ArrayRef<int64_t> shape) { |
| 419 | return llvm::count_if(Range&: shape, P: [](int64_t dimSize) { return dimSize != 1; }); |
| 420 | } |
| 421 | |
| 422 | /// Trims non-scalable one dimensions from `oldType` and returns the result |
| 423 | /// type. |
| 424 | static VectorType trimNonScalableUnitDims(VectorType oldType) { |
| 425 | SmallVector<int64_t> newShape; |
| 426 | SmallVector<bool> newScalableDims; |
| 427 | for (auto [dimIdx, dimSize] : llvm::enumerate(First: oldType.getShape())) { |
| 428 | if (dimSize == 1 && !oldType.getScalableDims()[dimIdx]) |
| 429 | continue; |
| 430 | newShape.push_back(Elt: dimSize); |
| 431 | newScalableDims.push_back(Elt: oldType.getScalableDims()[dimIdx]); |
| 432 | } |
| 433 | return VectorType::get(shape: newShape, elementType: oldType.getElementType(), scalableDims: newScalableDims); |
| 434 | } |
| 435 | |
| 436 | static bool isUnitDimMask(Value maskDimSize) { |
| 437 | return matchPattern(value: maskDimSize, pattern: m_One()); |
| 438 | } |
| 439 | static bool isUnitDimMask(int64_t maskDimSize) { return maskDimSize == 1; } |
| 440 | |
| 441 | /// Rewrites a mask op to drop non-scalable unit dimensions. |
| 442 | /// Supports vector.create_mask and vector.constant_mask. |
| 443 | template <typename MaskOp> |
| 444 | static FailureOr<Value> maskDropNonScalableUnitDims(PatternRewriter &rewriter, |
| 445 | Location loc, MaskOp op) { |
| 446 | auto type = op.getType(); |
| 447 | VectorType reducedType = trimNonScalableUnitDims(type); |
| 448 | if (reducedType.getRank() == type.getRank()) |
| 449 | return failure(); |
| 450 | |
| 451 | using ElemType = std::decay_t<decltype(*op.getMaskDimSizes().begin())>; |
| 452 | SmallVector<ElemType> reduced; |
| 453 | for (auto [dim, dimIsScalable, elem] : llvm::zip_equal( |
| 454 | type.getShape(), type.getScalableDims(), op.getMaskDimSizes())) { |
| 455 | if (dim == 1 && !dimIsScalable) { |
| 456 | if (!isUnitDimMask(elem)) |
| 457 | return failure(); |
| 458 | continue; |
| 459 | } |
| 460 | reduced.push_back(elem); |
| 461 | } |
| 462 | return MaskOp::create(rewriter, loc, reducedType, reduced).getResult(); |
| 463 | } |
| 464 | |
| 465 | namespace { |
| 466 | |
| 467 | /// Rewrites `vector.transfer_read` ops where the source has unit dims, by |
| 468 | /// inserting a memref.subview dropping those unit dims. The vector shapes are |
| 469 | /// also reduced accordingly. |
| 470 | class TransferReadDropUnitDimsPattern |
| 471 | : public vector::MaskableOpRewritePattern<vector::TransferReadOp> { |
| 472 | using MaskableOpRewritePattern::MaskableOpRewritePattern; |
| 473 | |
| 474 | FailureOr<Value> |
| 475 | matchAndRewriteMaskableOp(vector::TransferReadOp transferReadOp, |
| 476 | vector::MaskingOpInterface maskingOp, |
| 477 | PatternRewriter &rewriter) const override { |
| 478 | LDBG() << "=== TransferReadDropUnitDimsPattern: Analyzing " |
| 479 | << *transferReadOp; |
| 480 | auto loc = transferReadOp.getLoc(); |
| 481 | Value vector = transferReadOp.getVector(); |
| 482 | VectorType vectorType = cast<VectorType>(Val: vector.getType()); |
| 483 | Value source = transferReadOp.getBase(); |
| 484 | MemRefType sourceType = dyn_cast<MemRefType>(Val: source.getType()); |
| 485 | // TODO: support tensor types. |
| 486 | if (!sourceType) { |
| 487 | LDBG() << " -> Not a MemRefType, skipping" ; |
| 488 | return failure(); |
| 489 | } |
| 490 | // TODO: generalize this pattern, relax the requirements here. |
| 491 | if (transferReadOp.hasOutOfBoundsDim()) { |
| 492 | LDBG() << " -> Has out-of-bounds dimensions, skipping" ; |
| 493 | return failure(); |
| 494 | } |
| 495 | if (!transferReadOp.getPermutationMap().isMinorIdentity()) { |
| 496 | LDBG() << " -> Not minor identity permutation map, skipping" ; |
| 497 | return failure(); |
| 498 | } |
| 499 | // Check if the source shape can be further reduced. |
| 500 | int reducedRank = getReducedRank(shape: sourceType.getShape()); |
| 501 | LDBG() << " -> Source rank: " << sourceType.getRank() |
| 502 | << ", Reduced rank: " << reducedRank; |
| 503 | if (reducedRank == sourceType.getRank()) { |
| 504 | LDBG() << " -> No unit dimensions to drop, skipping" ; |
| 505 | return failure(); |
| 506 | } |
| 507 | // TODO: Extend vector.mask to support 0-d vectors. In the meantime, bail |
| 508 | // out. |
| 509 | if (reducedRank == 0 && maskingOp) { |
| 510 | LDBG() << " -> 0-d vector with masking not supported, skipping" ; |
| 511 | return failure(); |
| 512 | } |
| 513 | // Check if the reduced vector shape matches the reduced source shape. |
| 514 | // Otherwise, this case is not supported yet. |
| 515 | VectorType reducedVectorType = trimNonScalableUnitDims(oldType: vectorType); |
| 516 | LDBG() << " -> Vector type: " << vectorType |
| 517 | << ", Reduced vector type: " << reducedVectorType; |
| 518 | if (reducedRank != reducedVectorType.getRank()) { |
| 519 | LDBG() << " -> Reduced ranks don't match, skipping" ; |
| 520 | return failure(); |
| 521 | } |
| 522 | if (llvm::any_of(Range: transferReadOp.getIndices(), P: [](Value v) { |
| 523 | return getConstantIntValue(ofr: v) != static_cast<int64_t>(0); |
| 524 | })) { |
| 525 | LDBG() << " -> Non-zero indices found, skipping" ; |
| 526 | return failure(); |
| 527 | } |
| 528 | |
| 529 | Value maskOp = transferReadOp.getMask(); |
| 530 | if (maskOp) { |
| 531 | LDBG() << " -> Processing mask operation" ; |
| 532 | auto maskVectorType = cast<VectorType>(Val: maskOp.getType()); |
| 533 | FailureOr<Value> rankReducedMaskOp = failure(); |
| 534 | if (auto createMaskOp = maskOp.getDefiningOp<vector::CreateMaskOp>()) |
| 535 | rankReducedMaskOp = |
| 536 | maskDropNonScalableUnitDims(rewriter, loc, op: createMaskOp); |
| 537 | else if (auto constantMaskOp = |
| 538 | maskOp.getDefiningOp<vector::ConstantMaskOp>()) |
| 539 | rankReducedMaskOp = |
| 540 | maskDropNonScalableUnitDims(rewriter, loc, op: constantMaskOp); |
| 541 | else |
| 542 | return rewriter.notifyMatchFailure( |
| 543 | arg&: transferReadOp, |
| 544 | msg: "unsupported mask op, only 'vector.create_mask' and " |
| 545 | "'vector.constant_mask' are currently supported" ); |
| 546 | |
| 547 | if (succeeded(Result: rankReducedMaskOp)) { |
| 548 | maskOp = *rankReducedMaskOp; |
| 549 | LDBG() << " -> Successfully reduced mask dimensions" ; |
| 550 | } else if (maskVectorType.getRank() != reducedVectorType.getRank()) { |
| 551 | return rewriter.notifyMatchFailure( |
| 552 | arg&: transferReadOp, msg: "Mask reduction required, but failed" ); |
| 553 | } |
| 554 | } |
| 555 | |
| 556 | LDBG() << " -> Creating rank-reduced subview and new transfer_read" ; |
| 557 | Value reducedShapeSource = |
| 558 | rankReducingSubviewDroppingUnitDims(rewriter, loc, input: source); |
| 559 | Value c0 = arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: 0); |
| 560 | Repeated<Value> zeros(reducedRank, c0); |
| 561 | auto identityMap = rewriter.getMultiDimIdentityMap(rank: reducedRank); |
| 562 | SmallVector<bool> inBounds(reducedVectorType.getRank(), true); |
| 563 | Operation *newTransferReadOp = vector::TransferReadOp::create( |
| 564 | builder&: rewriter, location: loc, result: reducedVectorType, base: reducedShapeSource, indices: zeros, |
| 565 | permutation_map: identityMap, padding: transferReadOp.getPadding(), mask: maskOp, |
| 566 | in_bounds: rewriter.getBoolArrayAttr(values: inBounds)); |
| 567 | LDBG() << " -> Created new transfer_read: " << *newTransferReadOp; |
| 568 | |
| 569 | if (maskingOp) { |
| 570 | LDBG() << " -> Applying masking operation" ; |
| 571 | auto shapeCastMask = rewriter.createOrFold<vector::ShapeCastOp>( |
| 572 | location: loc, args: reducedVectorType.cloneWith(shape: std::nullopt, elementType: rewriter.getI1Type()), |
| 573 | args: maskingOp.getMask()); |
| 574 | newTransferReadOp = mlir::vector::maskOperation( |
| 575 | builder&: rewriter, maskableOp: newTransferReadOp, mask: shapeCastMask); |
| 576 | } |
| 577 | |
| 578 | auto shapeCast = rewriter.createOrFold<vector::ShapeCastOp>( |
| 579 | location: loc, args&: vectorType, args: newTransferReadOp->getResults()[0]); |
| 580 | LDBG() << " -> Created shape cast: " << *shapeCast.getDefiningOp(); |
| 581 | LDBG() << " -> Pattern match successful, returning result" ; |
| 582 | |
| 583 | return shapeCast; |
| 584 | } |
| 585 | }; |
| 586 | |
| 587 | /// Rewrites `vector.transfer_write` ops where the "source" (i.e. destination) |
| 588 | /// has unit dims, by inserting a `memref.subview` dropping those unit dims. The |
| 589 | /// vector shapes are also reduced accordingly. |
| 590 | class TransferWriteDropUnitDimsPattern |
| 591 | : public vector::MaskableOpRewritePattern<vector::TransferWriteOp> { |
| 592 | using MaskableOpRewritePattern::MaskableOpRewritePattern; |
| 593 | |
| 594 | FailureOr<Value> |
| 595 | matchAndRewriteMaskableOp(vector::TransferWriteOp transferWriteOp, |
| 596 | vector::MaskingOpInterface maskingOp, |
| 597 | PatternRewriter &rewriter) const override { |
| 598 | LDBG() << "=== TransferWriteDropUnitDimsPattern: Analyzing " |
| 599 | << *transferWriteOp; |
| 600 | auto loc = transferWriteOp.getLoc(); |
| 601 | Value vector = transferWriteOp.getVector(); |
| 602 | VectorType vectorType = cast<VectorType>(Val: vector.getType()); |
| 603 | Value source = transferWriteOp.getBase(); |
| 604 | MemRefType sourceType = dyn_cast<MemRefType>(Val: source.getType()); |
| 605 | // TODO: support tensor type. |
| 606 | if (!sourceType) { |
| 607 | LDBG() << " -> Not a MemRefType, skipping" ; |
| 608 | return failure(); |
| 609 | } |
| 610 | // TODO: generalize this pattern, relax the requirements here. |
| 611 | if (transferWriteOp.hasOutOfBoundsDim()) { |
| 612 | LDBG() << " -> Has out-of-bounds dimensions, skipping" ; |
| 613 | return failure(); |
| 614 | } |
| 615 | if (!transferWriteOp.getPermutationMap().isMinorIdentity()) { |
| 616 | LDBG() << " -> Not minor identity permutation map, skipping" ; |
| 617 | return failure(); |
| 618 | } |
| 619 | // Check if the destination shape can be further reduced. |
| 620 | int reducedRank = getReducedRank(shape: sourceType.getShape()); |
| 621 | LDBG() << " -> Source rank: " << sourceType.getRank() |
| 622 | << ", Reduced rank: " << reducedRank; |
| 623 | if (reducedRank == sourceType.getRank()) { |
| 624 | LDBG() << " -> No unit dimensions to drop, skipping" ; |
| 625 | return failure(); |
| 626 | } |
| 627 | // TODO: Extend vector.mask to support 0-d vectors. In the meantime, bail |
| 628 | // out. |
| 629 | if (reducedRank == 0 && maskingOp) { |
| 630 | LDBG() << " -> 0-d vector with masking not supported, skipping" ; |
| 631 | return failure(); |
| 632 | } |
| 633 | // Check if the reduced vector shape matches the reduced destination shape. |
| 634 | // Otherwise, this case is not supported yet. |
| 635 | VectorType reducedVectorType = trimNonScalableUnitDims(oldType: vectorType); |
| 636 | LDBG() << " -> Vector type: " << vectorType |
| 637 | << ", Reduced vector type: " << reducedVectorType; |
| 638 | if (reducedRank != reducedVectorType.getRank()) { |
| 639 | LDBG() << " -> Reduced ranks don't match, skipping" ; |
| 640 | return failure(); |
| 641 | } |
| 642 | if (llvm::any_of(Range: transferWriteOp.getIndices(), P: [](Value v) { |
| 643 | return getConstantIntValue(ofr: v) != static_cast<int64_t>(0); |
| 644 | })) { |
| 645 | LDBG() << " -> Non-zero indices found, skipping" ; |
| 646 | return failure(); |
| 647 | } |
| 648 | |
| 649 | Value maskOp = transferWriteOp.getMask(); |
| 650 | if (maskOp) { |
| 651 | LDBG() << " -> Processing mask operation" ; |
| 652 | auto maskVectorType = cast<VectorType>(Val: maskOp.getType()); |
| 653 | FailureOr<Value> rankReducedMask = failure(); |
| 654 | if (auto createMaskOp = maskOp.getDefiningOp<vector::CreateMaskOp>()) |
| 655 | rankReducedMask = |
| 656 | maskDropNonScalableUnitDims(rewriter, loc, op: createMaskOp); |
| 657 | else if (auto constantMaskOp = |
| 658 | maskOp.getDefiningOp<vector::ConstantMaskOp>()) |
| 659 | rankReducedMask = |
| 660 | maskDropNonScalableUnitDims(rewriter, loc, op: constantMaskOp); |
| 661 | else |
| 662 | return rewriter.notifyMatchFailure( |
| 663 | arg&: transferWriteOp, |
| 664 | msg: "unsupported mask op, only 'vector.create_mask' and " |
| 665 | "'vector.constant_mask' are currently supported" ); |
| 666 | |
| 667 | if (succeeded(Result: rankReducedMask)) { |
| 668 | maskOp = *rankReducedMask; |
| 669 | LDBG() << " -> Successfully reduced mask dimensions" ; |
| 670 | } else if (maskVectorType.getRank() != reducedVectorType.getRank()) { |
| 671 | return rewriter.notifyMatchFailure( |
| 672 | arg&: transferWriteOp, msg: "Mask reduction required, but failed" ); |
| 673 | } |
| 674 | } |
| 675 | LDBG() << " -> Creating rank-reduced subview and new transfer_write" ; |
| 676 | Value reducedShapeSource = |
| 677 | rankReducingSubviewDroppingUnitDims(rewriter, loc, input: source); |
| 678 | Value c0 = arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: 0); |
| 679 | Repeated<Value> zeros(reducedRank, c0); |
| 680 | auto identityMap = rewriter.getMultiDimIdentityMap(rank: reducedRank); |
| 681 | SmallVector<bool> inBounds(reducedVectorType.getRank(), true); |
| 682 | auto shapeCastSrc = rewriter.createOrFold<vector::ShapeCastOp>( |
| 683 | location: loc, args&: reducedVectorType, args&: vector); |
| 684 | Operation *newXferWrite = vector::TransferWriteOp::create( |
| 685 | builder&: rewriter, location: loc, result: Type(), valueToStore: shapeCastSrc, base: reducedShapeSource, indices: zeros, |
| 686 | permutation_map: identityMap, mask: maskOp, in_bounds: rewriter.getBoolArrayAttr(values: inBounds)); |
| 687 | LDBG() << " -> Created new transfer_write: " << *newXferWrite; |
| 688 | |
| 689 | if (maskingOp) { |
| 690 | LDBG() << " -> Applying masking operation" ; |
| 691 | auto shapeCastMask = rewriter.createOrFold<vector::ShapeCastOp>( |
| 692 | location: loc, args: reducedVectorType.cloneWith(shape: std::nullopt, elementType: rewriter.getI1Type()), |
| 693 | args: maskingOp.getMask()); |
| 694 | newXferWrite = |
| 695 | mlir::vector::maskOperation(builder&: rewriter, maskableOp: newXferWrite, mask: shapeCastMask); |
| 696 | } |
| 697 | |
| 698 | if (transferWriteOp.hasPureTensorSemantics()) { |
| 699 | LDBG() << " -> Pattern match successful (tensor semantics), returning " |
| 700 | "result" ; |
| 701 | return newXferWrite->getResults()[0]; |
| 702 | } |
| 703 | |
| 704 | // With Memref semantics, there's no return value. Use empty value to signal |
| 705 | // success. |
| 706 | LDBG() << " -> Pattern match successful (memref semantics)" ; |
| 707 | return Value(); |
| 708 | } |
| 709 | }; |
| 710 | |
| 711 | } // namespace |
| 712 | |
| 713 | /// Creates a memref.collapse_shape collapsing all inner dimensions of the |
| 714 | /// input starting at `firstDimToCollapse`. |
| 715 | static Value collapseInnerDims(PatternRewriter &rewriter, mlir::Location loc, |
| 716 | Value input, int64_t firstDimToCollapse) { |
| 717 | ShapedType inputType = cast<ShapedType>(Val: input.getType()); |
| 718 | if (inputType.getRank() == 1) |
| 719 | return input; |
| 720 | SmallVector<ReassociationIndices> reassociation; |
| 721 | for (int64_t i = 0; i < firstDimToCollapse; ++i) |
| 722 | reassociation.push_back(Elt: ReassociationIndices{i}); |
| 723 | ReassociationIndices collapsedIndices; |
| 724 | for (int64_t i = firstDimToCollapse; i < inputType.getRank(); ++i) |
| 725 | collapsedIndices.push_back(Elt: i); |
| 726 | reassociation.push_back(Elt: collapsedIndices); |
| 727 | return memref::CollapseShapeOp::create(builder&: rewriter, location: loc, src: input, reassociation); |
| 728 | } |
| 729 | |
| 730 | /// Returns the new indices that collapses the inner dimensions starting from |
| 731 | /// the `firstDimToCollapse` dimension. |
| 732 | static SmallVector<Value> getCollapsedIndices(RewriterBase &rewriter, |
| 733 | Location loc, |
| 734 | ArrayRef<int64_t> shape, |
| 735 | ValueRange indices, |
| 736 | int64_t firstDimToCollapse) { |
| 737 | assert(firstDimToCollapse < static_cast<int64_t>(indices.size())); |
| 738 | |
| 739 | // If all the collapsed indices are zero then no extra logic is needed. |
| 740 | // Otherwise, a new offset/index has to be computed. |
| 741 | SmallVector<Value> indicesAfterCollapsing( |
| 742 | indices.begin(), indices.begin() + firstDimToCollapse); |
| 743 | SmallVector<Value> indicesToCollapse(indices.begin() + firstDimToCollapse, |
| 744 | indices.end()); |
| 745 | if (llvm::all_of(Range&: indicesToCollapse, P: isZeroInteger)) { |
| 746 | indicesAfterCollapsing.push_back(Elt: indicesToCollapse[0]); |
| 747 | return indicesAfterCollapsing; |
| 748 | } |
| 749 | |
| 750 | // Compute the remaining trailing index/offset required for reading from |
| 751 | // the collapsed memref: |
| 752 | // |
| 753 | // offset = 0 |
| 754 | // for (i = firstDimToCollapse; i < outputRank; ++i) |
| 755 | // offset += sourceType.getDimSize(i) * transferReadOp.indices[i] |
| 756 | // |
| 757 | // For this example: |
| 758 | // %2 = vector.transfer_read/write %arg4[%c0, %arg0, %c0] (...) : |
| 759 | // memref<1x43x2xi32>, vector<1x2xi32> |
| 760 | // which would be collapsed to: |
| 761 | // %1 = vector.transfer_read/write %collapse_shape[%c0, %offset] (...) : |
| 762 | // memref<1x86xi32>, vector<2xi32> |
| 763 | // one would get the following offset: |
| 764 | // %offset = %arg0 * 43 |
| 765 | OpFoldResult collapsedOffset = |
| 766 | arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: 0).getResult(); |
| 767 | |
| 768 | auto collapsedStrides = computeSuffixProduct( |
| 769 | sizes: ArrayRef<int64_t>(shape.begin() + firstDimToCollapse, shape.end())); |
| 770 | |
| 771 | // Compute the collapsed offset. |
| 772 | auto &&[collapsedExpr, collapsedVals] = |
| 773 | computeLinearIndex(sourceOffset: collapsedOffset, strides: collapsedStrides, indices: indicesToCollapse); |
| 774 | collapsedOffset = affine::makeComposedFoldedAffineApply( |
| 775 | b&: rewriter, loc, expr: collapsedExpr, operands: collapsedVals); |
| 776 | |
| 777 | if (auto value = dyn_cast<Value>(Val&: collapsedOffset)) { |
| 778 | indicesAfterCollapsing.push_back(Elt: value); |
| 779 | } else { |
| 780 | indicesAfterCollapsing.push_back(Elt: arith::ConstantIndexOp::create( |
| 781 | builder&: rewriter, location: loc, value: *getConstantIntValue(ofr: collapsedOffset))); |
| 782 | } |
| 783 | |
| 784 | return indicesAfterCollapsing; |
| 785 | } |
| 786 | |
| 787 | namespace { |
| 788 | /// Rewrites contiguous row-major vector.transfer_read ops by inserting |
| 789 | /// memref.collapse_shape on the source so that the resulting |
| 790 | /// vector.transfer_read has a 1D source. Requires the source shape to be |
| 791 | /// already reduced i.e. without unit dims. |
| 792 | /// |
| 793 | /// If `targetVectorBitwidth` is provided, the flattening will only happen if |
| 794 | /// the trailing dimension of the vector read is smaller than the provided |
| 795 | /// bitwidth. |
| 796 | class FlattenContiguousRowMajorTransferReadPattern |
| 797 | : public OpRewritePattern<vector::TransferReadOp> { |
| 798 | public: |
| 799 | FlattenContiguousRowMajorTransferReadPattern(MLIRContext *context, |
| 800 | unsigned vectorBitwidth, |
| 801 | PatternBenefit benefit) |
| 802 | : OpRewritePattern<vector::TransferReadOp>(context, benefit), |
| 803 | targetVectorBitwidth(vectorBitwidth) {} |
| 804 | |
| 805 | LogicalResult matchAndRewrite(vector::TransferReadOp transferReadOp, |
| 806 | PatternRewriter &rewriter) const override { |
| 807 | LDBG() << "=== FlattenContiguousRowMajorTransferReadPattern: Analyzing " |
| 808 | << *transferReadOp; |
| 809 | auto loc = transferReadOp.getLoc(); |
| 810 | Value vector = transferReadOp.getVector(); |
| 811 | VectorType vectorType = cast<VectorType>(Val: vector.getType()); |
| 812 | auto source = transferReadOp.getBase(); |
| 813 | MemRefType sourceType = dyn_cast<MemRefType>(Val: source.getType()); |
| 814 | |
| 815 | // 0. Check pre-conditions |
| 816 | // Contiguity check is valid on tensors only. |
| 817 | if (!sourceType) { |
| 818 | LDBG() << " -> Not a MemRefType, skipping" ; |
| 819 | return failure(); |
| 820 | } |
| 821 | // If this is already 0D/1D, there's nothing to do. |
| 822 | if (vectorType.getRank() <= 1) { |
| 823 | LDBG() << " -> Already 0D/1D, skipping" ; |
| 824 | return failure(); |
| 825 | } |
| 826 | if (!vectorType.getElementType().isSignlessIntOrFloat()) { |
| 827 | LDBG() << " -> Not signless int or float, skipping" ; |
| 828 | return failure(); |
| 829 | } |
| 830 | unsigned trailingVectorDimBitwidth = |
| 831 | vectorType.getShape().back() * vectorType.getElementTypeBitWidth(); |
| 832 | LDBG() << " -> Trailing vector dim bitwidth: " << trailingVectorDimBitwidth |
| 833 | << ", target: " << targetVectorBitwidth; |
| 834 | if (trailingVectorDimBitwidth >= targetVectorBitwidth) { |
| 835 | LDBG() << " -> Trailing dim bitwidth >= target, skipping" ; |
| 836 | return failure(); |
| 837 | } |
| 838 | if (!vector::isContiguousSlice(memrefType: sourceType, vectorType)) { |
| 839 | LDBG() << " -> Not contiguous slice, skipping" ; |
| 840 | return failure(); |
| 841 | } |
| 842 | // TODO: generalize this pattern, relax the requirements here. |
| 843 | if (transferReadOp.hasOutOfBoundsDim()) { |
| 844 | LDBG() << " -> Has out-of-bounds dimensions, skipping" ; |
| 845 | return failure(); |
| 846 | } |
| 847 | if (!transferReadOp.getPermutationMap().isMinorIdentity()) { |
| 848 | LDBG() << " -> Not minor identity permutation map, skipping" ; |
| 849 | return failure(); |
| 850 | } |
| 851 | if (transferReadOp.getMask() || transferReadOp.isMasked()) { |
| 852 | LDBG() << " -> Has mask, skipping" ; |
| 853 | return failure(); |
| 854 | } |
| 855 | |
| 856 | // Determine vector dimensions to collapse. |
| 857 | // Ignore a leading sequence of adjacent unit dimensions in the vector. |
| 858 | ArrayRef<int64_t> collapsedVectorShape = |
| 859 | vectorType.getShape().drop_while(Pred: [](auto v) { return v == 1; }); |
| 860 | size_t collapsedVecRank = collapsedVectorShape.size(); |
| 861 | // Limit the collapse of multi-dimensional unit vectors (e.g. <1x1x1xf32>) |
| 862 | // to a 1D single-element vector. |
| 863 | if (collapsedVecRank == 0) |
| 864 | collapsedVecRank = 1; |
| 865 | |
| 866 | // Determine the first memref dimension to collapse - just enough so we can |
| 867 | // read a flattened vector. |
| 868 | int64_t firstDimToCollapse = sourceType.getRank() - collapsedVecRank; |
| 869 | LDBG() << " -> First dimension to collapse: " << firstDimToCollapse; |
| 870 | |
| 871 | // 1. Collapse the source memref |
| 872 | LDBG() << " -> Collapsing source memref" ; |
| 873 | Value collapsedSource = |
| 874 | collapseInnerDims(rewriter, loc, input: source, firstDimToCollapse); |
| 875 | MemRefType collapsedSourceType = |
| 876 | cast<MemRefType>(Val: collapsedSource.getType()); |
| 877 | int64_t collapsedRank = collapsedSourceType.getRank(); |
| 878 | assert(collapsedRank == firstDimToCollapse + 1); |
| 879 | LDBG() << " -> Collapsed source type: " << collapsedSourceType; |
| 880 | |
| 881 | // 2. Generate input args for a new vector.transfer_read that will read |
| 882 | // from the collapsed memref. |
| 883 | // 2.1. New dim exprs + affine map |
| 884 | SmallVector<AffineExpr, 1> dimExprs{ |
| 885 | getAffineDimExpr(position: firstDimToCollapse, context: rewriter.getContext())}; |
| 886 | auto collapsedMap = |
| 887 | AffineMap::get(dimCount: collapsedRank, symbolCount: 0, results: dimExprs, context: rewriter.getContext()); |
| 888 | |
| 889 | // 2.2 New indices |
| 890 | SmallVector<Value> collapsedIndices = |
| 891 | getCollapsedIndices(rewriter, loc, shape: sourceType.getShape(), |
| 892 | indices: transferReadOp.getIndices(), firstDimToCollapse); |
| 893 | |
| 894 | // 3. Create new vector.transfer_read that reads from the collapsed memref |
| 895 | VectorType flatVectorType = VectorType::get(shape: {vectorType.getNumElements()}, |
| 896 | elementType: vectorType.getElementType()); |
| 897 | LDBG() << " -> Creating flattened vector type: " << flatVectorType; |
| 898 | vector::TransferReadOp flatRead = vector::TransferReadOp::create( |
| 899 | builder&: rewriter, location: loc, vectorType: flatVectorType, source: collapsedSource, indices: collapsedIndices, |
| 900 | padding: transferReadOp.getPadding(), permutationMap: collapsedMap); |
| 901 | flatRead.setInBoundsAttr(rewriter.getBoolArrayAttr(values: {true})); |
| 902 | LDBG() << " -> Created flat transfer_read: " << *flatRead; |
| 903 | |
| 904 | // 4. Replace the old transfer_read with the new one reading from the |
| 905 | // collapsed shape |
| 906 | LDBG() << " -> Replacing with shape cast" ; |
| 907 | rewriter.replaceOpWithNewOp<vector::ShapeCastOp>( |
| 908 | op: transferReadOp, args: cast<VectorType>(Val: vector.getType()), args&: flatRead); |
| 909 | LDBG() << " -> Pattern match successful" ; |
| 910 | return success(); |
| 911 | } |
| 912 | |
| 913 | private: |
| 914 | // Minimum bitwidth that the trailing vector dimension should have after |
| 915 | // flattening. |
| 916 | unsigned targetVectorBitwidth; |
| 917 | }; |
| 918 | |
| 919 | /// Rewrites contiguous row-major vector.transfer_write ops by inserting |
| 920 | /// memref.collapse_shape on the source so that the resulting |
| 921 | /// vector.transfer_write has a 1D source. Requires the source shape to be |
| 922 | /// already reduced i.e. without unit dims. |
| 923 | /// |
| 924 | /// If `targetVectorBitwidth` is provided, the flattening will only happen if |
| 925 | /// the trailing dimension of the vector read is smaller than the provided |
| 926 | /// bitwidth. |
| 927 | class FlattenContiguousRowMajorTransferWritePattern |
| 928 | : public OpRewritePattern<vector::TransferWriteOp> { |
| 929 | public: |
| 930 | FlattenContiguousRowMajorTransferWritePattern(MLIRContext *context, |
| 931 | unsigned vectorBitwidth, |
| 932 | PatternBenefit benefit) |
| 933 | : OpRewritePattern<vector::TransferWriteOp>(context, benefit), |
| 934 | targetVectorBitwidth(vectorBitwidth) {} |
| 935 | |
| 936 | LogicalResult matchAndRewrite(vector::TransferWriteOp transferWriteOp, |
| 937 | PatternRewriter &rewriter) const override { |
| 938 | auto loc = transferWriteOp.getLoc(); |
| 939 | Value vector = transferWriteOp.getVector(); |
| 940 | VectorType vectorType = cast<VectorType>(Val: vector.getType()); |
| 941 | Value source = transferWriteOp.getBase(); |
| 942 | MemRefType sourceType = dyn_cast<MemRefType>(Val: source.getType()); |
| 943 | |
| 944 | // 0. Check pre-conditions |
| 945 | // Contiguity check is valid on tensors only. |
| 946 | if (!sourceType) |
| 947 | return failure(); |
| 948 | // If this is already 0D/1D, there's nothing to do. |
| 949 | if (vectorType.getRank() <= 1) |
| 950 | // Already 0D/1D, nothing to do. |
| 951 | return failure(); |
| 952 | if (!vectorType.getElementType().isSignlessIntOrFloat()) |
| 953 | return failure(); |
| 954 | unsigned trailingVectorDimBitwidth = |
| 955 | vectorType.getShape().back() * vectorType.getElementTypeBitWidth(); |
| 956 | if (trailingVectorDimBitwidth >= targetVectorBitwidth) |
| 957 | return failure(); |
| 958 | if (!vector::isContiguousSlice(memrefType: sourceType, vectorType)) |
| 959 | return failure(); |
| 960 | // TODO: generalize this pattern, relax the requirements here. |
| 961 | if (transferWriteOp.hasOutOfBoundsDim()) |
| 962 | return failure(); |
| 963 | if (!transferWriteOp.getPermutationMap().isMinorIdentity()) |
| 964 | return failure(); |
| 965 | if (transferWriteOp.getMask() || transferWriteOp.isMasked()) |
| 966 | return failure(); |
| 967 | |
| 968 | // Determine vector dimensions to collapse. |
| 969 | // Ignore a leading sequence of adjacent unit dimensions in the vector. |
| 970 | ArrayRef<int64_t> collapsedVectorShape = |
| 971 | vectorType.getShape().drop_while(Pred: [](auto v) { return v == 1; }); |
| 972 | size_t collapsedVecRank = collapsedVectorShape.size(); |
| 973 | // Limit the collapse of multi-dimensional unit vectors (e.g. <1x1x1xf32>) |
| 974 | // to a 1D single-element vector. |
| 975 | if (collapsedVecRank == 0) |
| 976 | collapsedVecRank = 1; |
| 977 | |
| 978 | // Determine the first memref dimension to collapse - just enough so we can |
| 979 | // read a flattened vector. |
| 980 | int64_t firstDimToCollapse = sourceType.getRank() - collapsedVecRank; |
| 981 | |
| 982 | // 1. Collapse the source memref |
| 983 | Value collapsedSource = |
| 984 | collapseInnerDims(rewriter, loc, input: source, firstDimToCollapse); |
| 985 | MemRefType collapsedSourceType = |
| 986 | cast<MemRefType>(Val: collapsedSource.getType()); |
| 987 | int64_t collapsedRank = collapsedSourceType.getRank(); |
| 988 | assert(collapsedRank == firstDimToCollapse + 1); |
| 989 | |
| 990 | // 2. Generate input args for a new vector.transfer_read that will read |
| 991 | // from the collapsed memref. |
| 992 | // 2.1. New dim exprs + affine map |
| 993 | SmallVector<AffineExpr, 1> dimExprs{ |
| 994 | getAffineDimExpr(position: firstDimToCollapse, context: rewriter.getContext())}; |
| 995 | auto collapsedMap = |
| 996 | AffineMap::get(dimCount: collapsedRank, symbolCount: 0, results: dimExprs, context: rewriter.getContext()); |
| 997 | |
| 998 | // 2.2 New indices |
| 999 | SmallVector<Value> collapsedIndices = |
| 1000 | getCollapsedIndices(rewriter, loc, shape: sourceType.getShape(), |
| 1001 | indices: transferWriteOp.getIndices(), firstDimToCollapse); |
| 1002 | |
| 1003 | // 3. Create new vector.transfer_write that writes to the collapsed memref |
| 1004 | VectorType flatVectorType = VectorType::get(shape: {vectorType.getNumElements()}, |
| 1005 | elementType: vectorType.getElementType()); |
| 1006 | Value flatVector = |
| 1007 | vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: flatVectorType, source: vector); |
| 1008 | vector::TransferWriteOp flatWrite = vector::TransferWriteOp::create( |
| 1009 | builder&: rewriter, location: loc, vector: flatVector, dest: collapsedSource, indices: collapsedIndices, |
| 1010 | permutationMap: collapsedMap); |
| 1011 | flatWrite.setInBoundsAttr(rewriter.getBoolArrayAttr(values: {true})); |
| 1012 | |
| 1013 | // 4. Replace the old transfer_write with the new one writing the |
| 1014 | // collapsed shape |
| 1015 | rewriter.eraseOp(op: transferWriteOp); |
| 1016 | return success(); |
| 1017 | } |
| 1018 | |
| 1019 | private: |
| 1020 | // Minimum bitwidth that the trailing vector dimension should have after |
| 1021 | // flattening. |
| 1022 | unsigned targetVectorBitwidth; |
| 1023 | }; |
| 1024 | |
| 1025 | /// Rewrite `vector.extract(vector.transfer_read)` to `memref.load`. |
| 1026 | /// |
| 1027 | /// All the users of the transfer op must be `vector.extract` ops. If |
| 1028 | /// `allowMultipleUses` is set to true, rewrite transfer ops with any number of |
| 1029 | /// users. Otherwise, rewrite only if the extract op is the single user of the |
| 1030 | /// transfer op. Rewriting a single vector load with multiple scalar loads may |
| 1031 | /// negatively affect performance. |
| 1032 | class |
| 1033 | : public OpRewritePattern<vector::ExtractOp> { |
| 1034 | public: |
| 1035 | (MLIRContext *context, |
| 1036 | PatternBenefit benefit, |
| 1037 | bool allowMultipleUses) |
| 1038 | : OpRewritePattern(context, benefit), |
| 1039 | allowMultipleUses(allowMultipleUses) {} |
| 1040 | |
| 1041 | LogicalResult matchAndRewrite(vector::ExtractOp , |
| 1042 | PatternRewriter &rewriter) const override { |
| 1043 | // Match phase. |
| 1044 | auto xferOp = extractOp.getSource().getDefiningOp<vector::TransferReadOp>(); |
| 1045 | if (!xferOp) |
| 1046 | return failure(); |
| 1047 | // Check that we are extracting a scalar and not a sub-vector. |
| 1048 | if (isa<VectorType>(Val: extractOp.getResult().getType())) |
| 1049 | return failure(); |
| 1050 | // If multiple uses are not allowed, check if xfer has a single use. |
| 1051 | if (!allowMultipleUses && !xferOp.getResult().hasOneUse()) |
| 1052 | return failure(); |
| 1053 | // If multiple uses are allowed, check if all the xfer uses are extract ops. |
| 1054 | if (allowMultipleUses && |
| 1055 | !llvm::all_of(Range: xferOp->getUses(), P: [](OpOperand &use) { |
| 1056 | return isa<vector::ExtractOp>(Val: use.getOwner()); |
| 1057 | })) |
| 1058 | return failure(); |
| 1059 | // Mask not supported. |
| 1060 | if (xferOp.getMask()) |
| 1061 | return failure(); |
| 1062 | // Map not supported. |
| 1063 | if (!xferOp.getPermutationMap().isMinorIdentity()) |
| 1064 | return failure(); |
| 1065 | // Cannot rewrite if the indices may be out of bounds. |
| 1066 | if (xferOp.hasOutOfBoundsDim()) |
| 1067 | return failure(); |
| 1068 | |
| 1069 | // Rewrite phase: construct scalar load. |
| 1070 | SmallVector<Value> newIndices(xferOp.getIndices().begin(), |
| 1071 | xferOp.getIndices().end()); |
| 1072 | for (auto [i, pos] : llvm::enumerate(First: extractOp.getMixedPosition())) { |
| 1073 | int64_t idx = newIndices.size() - extractOp.getNumIndices() + i; |
| 1074 | |
| 1075 | // Compute affine expression `newIndices[idx] + pos` where `pos` can be |
| 1076 | // either a constant or a value. |
| 1077 | OpFoldResult composedIdx; |
| 1078 | if (auto attr = dyn_cast<Attribute>(Val&: pos)) { |
| 1079 | int64_t offset = cast<IntegerAttr>(Val&: attr).getInt(); |
| 1080 | composedIdx = affine::makeComposedFoldedAffineApply( |
| 1081 | b&: rewriter, loc: extractOp.getLoc(), |
| 1082 | expr: rewriter.getAffineSymbolExpr(position: 0) + offset, operands: {newIndices[idx]}); |
| 1083 | } else { |
| 1084 | Value dynamicOffset = cast<Value>(Val&: pos); |
| 1085 | AffineExpr sym0, sym1; |
| 1086 | bindSymbols(ctx: rewriter.getContext(), exprs&: sym0, exprs&: sym1); |
| 1087 | composedIdx = affine::makeComposedFoldedAffineApply( |
| 1088 | b&: rewriter, loc: extractOp.getLoc(), expr: sym0 + sym1, |
| 1089 | operands: {newIndices[idx], dynamicOffset}); |
| 1090 | } |
| 1091 | |
| 1092 | // Update the corresponding index with the folded result. |
| 1093 | if (auto value = dyn_cast<Value>(Val&: composedIdx)) { |
| 1094 | newIndices[idx] = value; |
| 1095 | } else { |
| 1096 | newIndices[idx] = arith::ConstantIndexOp::create( |
| 1097 | builder&: rewriter, location: extractOp.getLoc(), value: *getConstantIntValue(ofr: composedIdx)); |
| 1098 | } |
| 1099 | } |
| 1100 | if (isa<MemRefType>(Val: xferOp.getBase().getType())) { |
| 1101 | rewriter.replaceOpWithNewOp<memref::LoadOp>(op: extractOp, args: xferOp.getBase(), |
| 1102 | args&: newIndices); |
| 1103 | } else { |
| 1104 | rewriter.replaceOpWithNewOp<tensor::ExtractOp>( |
| 1105 | op: extractOp, args: xferOp.getBase(), args&: newIndices); |
| 1106 | } |
| 1107 | |
| 1108 | return success(); |
| 1109 | } |
| 1110 | |
| 1111 | private: |
| 1112 | bool ; |
| 1113 | }; |
| 1114 | |
| 1115 | /// Rewrite transfer_writes of vectors of size 1 (e.g., vector<1x1xf32>) |
| 1116 | /// to memref.store. |
| 1117 | class RewriteScalarWrite : public OpRewritePattern<vector::TransferWriteOp> { |
| 1118 | using Base::Base; |
| 1119 | |
| 1120 | LogicalResult matchAndRewrite(vector::TransferWriteOp xferOp, |
| 1121 | PatternRewriter &rewriter) const override { |
| 1122 | // Must be a scalar write. |
| 1123 | auto vecType = xferOp.getVectorType(); |
| 1124 | if (!llvm::all_of(Range: vecType.getShape(), P: [](int64_t sz) { return sz == 1; })) |
| 1125 | return failure(); |
| 1126 | // Mask not supported. |
| 1127 | if (xferOp.getMask() || xferOp.isMasked()) |
| 1128 | return failure(); |
| 1129 | // Map not supported. |
| 1130 | if (!xferOp.getPermutationMap().isMinorIdentity()) |
| 1131 | return failure(); |
| 1132 | // Only float and integer element types are supported. |
| 1133 | Value scalar = vector::ExtractOp::create(builder&: rewriter, location: xferOp.getLoc(), |
| 1134 | source: xferOp.getVector()); |
| 1135 | // Construct a scalar store. |
| 1136 | if (isa<MemRefType>(Val: xferOp.getBase().getType())) { |
| 1137 | rewriter.replaceOpWithNewOp<memref::StoreOp>( |
| 1138 | op: xferOp, args&: scalar, args: xferOp.getBase(), args: xferOp.getIndices()); |
| 1139 | } else { |
| 1140 | rewriter.replaceOpWithNewOp<tensor::InsertOp>( |
| 1141 | op: xferOp, args&: scalar, args: xferOp.getBase(), args: xferOp.getIndices()); |
| 1142 | } |
| 1143 | return success(); |
| 1144 | } |
| 1145 | }; |
| 1146 | |
| 1147 | } // namespace |
| 1148 | |
| 1149 | void mlir::vector::transferOpflowOpt(RewriterBase &rewriter, |
| 1150 | Operation *rootOp) { |
| 1151 | LDBG() << "=== Starting transferOpflowOpt on root operation: " |
| 1152 | << OpWithFlags(rootOp, OpPrintingFlags().skipRegions()); |
| 1153 | TransferOptimization opt(rewriter, rootOp); |
| 1154 | |
| 1155 | // Run store to load forwarding first since it can expose more dead store |
| 1156 | // opportunity. |
| 1157 | LDBG() << "Phase 1: Store-to-load forwarding" ; |
| 1158 | int readCount = 0; |
| 1159 | rootOp->walk(callback: [&](vector::TransferReadOp read) { |
| 1160 | if (isa<MemRefType>(Val: read.getShapedType())) { |
| 1161 | LDBG() << "Processing transfer_read #" << ++readCount << ": " << *read; |
| 1162 | opt.storeToLoadForwarding(read); |
| 1163 | } |
| 1164 | }); |
| 1165 | LDBG() << "Phase 1 complete. Removing dead operations from forwarding" ; |
| 1166 | opt.removeDeadOp(); |
| 1167 | |
| 1168 | LDBG() << "Phase 2: Dead store elimination" ; |
| 1169 | int writeCount = 0; |
| 1170 | rootOp->walk(callback: [&](vector::TransferWriteOp write) { |
| 1171 | if (isa<MemRefType>(Val: write.getShapedType())) { |
| 1172 | LDBG() << "Processing transfer_write #" << ++writeCount << ": " << *write; |
| 1173 | opt.deadStoreOp(write); |
| 1174 | } |
| 1175 | }); |
| 1176 | LDBG() << "Phase 2 complete. Removing dead operations from dead store " |
| 1177 | "elimination" ; |
| 1178 | opt.removeDeadOp(); |
| 1179 | LDBG() << "=== transferOpflowOpt complete" ; |
| 1180 | } |
| 1181 | |
| 1182 | void mlir::vector::populateScalarVectorTransferLoweringPatterns( |
| 1183 | RewritePatternSet &patterns, PatternBenefit benefit, |
| 1184 | bool allowMultipleUses) { |
| 1185 | patterns.add<RewriteScalarExtractOfTransferRead>(arg: patterns.getContext(), |
| 1186 | args&: benefit, args&: allowMultipleUses); |
| 1187 | patterns.add<RewriteScalarWrite>(arg: patterns.getContext(), args&: benefit); |
| 1188 | } |
| 1189 | |
| 1190 | void mlir::vector::populateVectorTransferDropUnitDimsPatterns( |
| 1191 | RewritePatternSet &patterns, PatternBenefit benefit) { |
| 1192 | patterns |
| 1193 | .add<TransferReadDropUnitDimsPattern, TransferWriteDropUnitDimsPattern>( |
| 1194 | arg: patterns.getContext(), args&: benefit); |
| 1195 | } |
| 1196 | |
| 1197 | void mlir::vector::populateFlattenVectorTransferPatterns( |
| 1198 | RewritePatternSet &patterns, unsigned targetVectorBitwidth, |
| 1199 | PatternBenefit benefit) { |
| 1200 | patterns.add<FlattenContiguousRowMajorTransferReadPattern, |
| 1201 | FlattenContiguousRowMajorTransferWritePattern>( |
| 1202 | arg: patterns.getContext(), args&: targetVectorBitwidth, args&: benefit); |
| 1203 | populateDropUnitDimWithShapeCastPatterns(patterns, benefit); |
| 1204 | } |
| 1205 | |