1 //===- LinalgTransforms.cpp - Linalg transformations as patterns ----------===// 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 logic and helpers to expose Linalg transforms as rewrite 10 // patterns. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "mlir/Dialect/Linalg/Transforms/Transforms.h" 15 #include "mlir/Dialect/Affine/Utils.h" 16 #include "mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h" 17 #include "mlir/Dialect/Linalg/IR/LinalgOps.h" 18 #include "mlir/Dialect/Linalg/Utils/Utils.h" 19 #include "mlir/Dialect/StandardOps/EDSC/Intrinsics.h" 20 #include "mlir/Dialect/Utils/StructuredOpsUtils.h" 21 #include "mlir/Dialect/Vector/EDSC/Intrinsics.h" 22 #include "mlir/Dialect/Vector/VectorOps.h" 23 #include "mlir/IR/AffineExpr.h" 24 #include "mlir/IR/Matchers.h" 25 #include "mlir/Pass/Pass.h" 26 #include "mlir/Support/LLVM.h" 27 #include "mlir/Transforms/GreedyPatternRewriteDriver.h" 28 #include "llvm/Support/Debug.h" 29 #include "llvm/Support/raw_ostream.h" 30 #include <type_traits> 31 32 #define DEBUG_TYPE "linalg-transforms" 33 34 using namespace mlir; 35 using namespace mlir::edsc; 36 using namespace mlir::edsc::intrinsics; 37 using namespace mlir::linalg; 38 39 #define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE << "]: ") 40 41 //===----------------------------------------------------------------------===// 42 // Transformations exposed as rewrite patterns. 43 //===----------------------------------------------------------------------===// 44 // Marker used as attribute name in generated Linalg rewriting transformations. 45 const StringLiteral mlir::linalg::LinalgTransforms::kLinalgTransformMarker = 46 "__internal_linalg_transform__"; 47 48 mlir::linalg::LinalgMarker::LinalgMarker(ArrayRef<Identifier> matchDisjunction, 49 Optional<Identifier> replacement) 50 : matchDisjunction(matchDisjunction.begin(), matchDisjunction.end()), 51 replacement(replacement) {} 52 53 LogicalResult 54 mlir::linalg::LinalgMarker::checkAndNotify(PatternRewriter &rewriter, 55 Operation *op) const { 56 auto attr = op->template getAttrOfType<StringAttr>( 57 LinalgTransforms::kLinalgTransformMarker); 58 59 if (!attr) { 60 // 1. Has no marker case and matchDisjunction is empty. 61 if (matchDisjunction.empty()) 62 return success(); 63 64 // 2. Has no marker but was expecting a marker. 65 return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) { 66 diag << " does not have any marker from list: "; 67 interleaveComma(matchDisjunction, diag); 68 }); 69 } 70 71 // 4. Match explicit marker. 72 for (auto marker : matchDisjunction) 73 if (attr.getValue() == marker) 74 return success(); 75 76 // 5. Fail to match. 77 return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) { 78 diag << " does not have any marker from list: "; 79 interleaveComma(matchDisjunction, diag); 80 }); 81 } 82 83 void mlir::linalg::LinalgMarker::replaceLinalgMarker(PatternRewriter &rewriter, 84 Operation *op) const { 85 if (replacement.hasValue()) 86 op->setAttr(LinalgTransforms::kLinalgTransformMarker, 87 rewriter.getStringAttr(replacement.getValue())); 88 else 89 op->removeAttr(Identifier::get(LinalgTransforms::kLinalgTransformMarker, 90 rewriter.getContext())); 91 } 92 93 LinalgTilingOptions & 94 mlir::linalg::LinalgTilingOptions::setTileSizes(ArrayRef<int64_t> ts) { 95 SmallVector<int64_t, 4> tileSizes(ts.begin(), ts.end()); 96 tileSizeComputationFunction = [tileSizes](OpBuilder &b, Operation *op) { 97 OpBuilder::InsertionGuard guard(b); 98 b.setInsertionPointToStart( 99 &op->getParentOfType<FuncOp>().getBody().front()); 100 return llvm::to_vector<4>(map_range(tileSizes, [&](int64_t s) { 101 Value v = b.create<ConstantIndexOp>(op->getLoc(), s); 102 return v; 103 })); 104 }; 105 return *this; 106 } 107 108 /// Linalg base tiling pattern. 109 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern( 110 StringRef opName, MLIRContext *context, LinalgTilingOptions options, 111 LinalgMarker marker, PatternBenefit benefit) 112 : RewritePattern(opName, {}, benefit, context), marker(marker), 113 options(options) {} 114 115 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern( 116 LinalgTilingOptions options, LinalgMarker marker, PatternBenefit benefit) 117 : RewritePattern(benefit, MatchAnyOpTypeTag()), marker(marker), 118 options(options) {} 119 120 LogicalResult mlir::linalg::LinalgBaseTilingPattern::matchAndRewriteBase( 121 Operation *op, PatternRewriter &rewriter, TiledLinalgOp &result) const { 122 LinalgOp linalgOp = dyn_cast<LinalgOp>(op); 123 if (!linalgOp) 124 return failure(); 125 if (failed(marker.checkAndNotify(rewriter, linalgOp))) 126 return failure(); 127 128 Optional<TiledLinalgOp> res = tileLinalgOp(rewriter, linalgOp, options); 129 130 if (!res) 131 return failure(); 132 133 // Return relevant information to derived pattern. 134 result = *res; 135 136 // New marker if specified. 137 marker.replaceLinalgMarker(rewriter, res->op.getOperation()); 138 return success(); 139 } 140 141 mlir::linalg::LinalgBaseTileAndFusePattern::LinalgBaseTileAndFusePattern( 142 StringRef opName, MLIRContext *context, 143 const LinalgDependenceGraph &dependenceGraph, 144 LinalgTilingOptions tilingOptions, LinalgFusionOptions fusionOptions, 145 LinalgMarker marker, LinalgMarker fusedOpMarker, 146 LinalgMarker originalOpMarker, PatternBenefit benefit) 147 : RewritePattern(opName, {}, benefit, context), 148 dependenceGraph(dependenceGraph), tilingOptions(tilingOptions), 149 fusionOptions(fusionOptions), marker(marker), 150 fusedOpMarker(fusedOpMarker), originalOpMarker(originalOpMarker) {} 151 152 LogicalResult mlir::linalg::LinalgBaseTileAndFusePattern::matchAndRewrite( 153 Operation *op, PatternRewriter &rewriter) const { 154 LinalgOp linalgOp = dyn_cast<LinalgOp>(op); 155 if (!linalgOp) 156 return failure(); 157 if (failed(marker.checkAndNotify(rewriter, linalgOp))) 158 return failure(); 159 if (!linalgOp.hasBufferSemantics()) 160 return failure(); 161 162 DenseSet<Operation *> producers; 163 producers.insert(linalgOp); 164 for (auto dependence : dependenceGraph.getDependentOperations(linalgOp)) { 165 if (!fusionOptions.indicesToFuse.count( 166 dependence.indexingOpView->getOperandNumber())) 167 continue; 168 if (isa<LinalgOp>(dependence.dependentOpView->getOwner())) 169 producers.insert(dependence.dependentOpView->getOwner()); 170 } 171 172 SmallVector<LinalgOp, 1> fusionOps; 173 for (auto it = op->getBlock()->begin(), ie = Block::iterator(op); it != ie; 174 ++it) { 175 auto producerLinalgOp = dyn_cast<LinalgOp>(&(*it)); 176 if (producerLinalgOp && producers.count(producerLinalgOp)) 177 fusionOps.push_back(producerLinalgOp); 178 } 179 fusionOps.push_back(linalgOp); 180 181 SmallVector<Value, 4> tileSizes = 182 tilingOptions.tileSizeComputationFunction(rewriter, op); 183 LinalgTilingOptions instanceTilingOptions = tilingOptions; 184 instanceTilingOptions.setTileSizes(tileSizes); 185 Optional<TiledAndFusedLinalgOps> tiledAndFusedOps = tileAndFuseLinalgOps( 186 rewriter, fusionOps, dependenceGraph, instanceTilingOptions); 187 if (!tiledAndFusedOps) 188 return failure(); 189 190 // Tile the unfused loops; 191 SmallVector<Value, 4> unfusedLoopTileSizes; 192 Value zero = rewriter.create<ConstantIndexOp>(op->getLoc(), 0); 193 for (auto tileSize : enumerate(tileSizes)) { 194 if (tiledAndFusedOps->fusedLoopDims.count(tileSize.index())) 195 unfusedLoopTileSizes.push_back(zero); 196 else 197 unfusedLoopTileSizes.push_back(tileSize.value()); 198 } 199 // Tile the loop only if there is a non-zero tile size. 200 if (unfusedLoopTileSizes.size() > linalgOp.getNumLoops()) 201 unfusedLoopTileSizes.resize(linalgOp.getNumLoops()); 202 if (llvm::any_of(unfusedLoopTileSizes, [](Value val) { 203 if (auto cst = val.getDefiningOp<ConstantIndexOp>()) 204 return cst.getValue() != 0; 205 return true; 206 })) { 207 LinalgTilingOptions unfusedTilingOptions = tilingOptions; 208 unfusedTilingOptions.setTileSizes(unfusedLoopTileSizes); 209 Optional<TiledLinalgOp> unfusedTiledOp = 210 tileLinalgOp(rewriter, tiledAndFusedOps->op, unfusedTilingOptions); 211 if (!unfusedTiledOp) 212 return failure(); 213 rewriter.eraseOp(tiledAndFusedOps->op); 214 tiledAndFusedOps->op = unfusedTiledOp->op; 215 } 216 217 marker.replaceLinalgMarker(rewriter, tiledAndFusedOps->op.getOperation()); 218 for (auto fusedOp : tiledAndFusedOps->fusedProducers) { 219 fusedOpMarker.replaceLinalgMarker(rewriter, fusedOp.getOperation()); 220 } 221 for (auto origProducerOp : ArrayRef<LinalgOp>(fusionOps).drop_back()) { 222 originalOpMarker.replaceLinalgMarker(rewriter, 223 origProducerOp.getOperation()); 224 } 225 rewriter.updateRootInPlace( 226 op, [&]() { originalOpMarker.replaceLinalgMarker(rewriter, op); }); 227 return success(); 228 } 229 230 /// Linalg base interchange pattern. 231 mlir::linalg::LinalgBaseInterchangePattern::LinalgBaseInterchangePattern( 232 StringRef opName, MLIRContext *context, 233 ArrayRef<unsigned> interchangeVector, LinalgMarker marker, 234 PatternBenefit benefit) 235 : RewritePattern(opName, {}, benefit, context), marker(marker), 236 interchangeVector(interchangeVector.begin(), interchangeVector.end()) {} 237 238 LogicalResult mlir::linalg::LinalgBaseInterchangePattern::matchAndRewrite( 239 Operation *op, PatternRewriter &rewriter) const { 240 LinalgOp linalgOp = dyn_cast<LinalgOp>(op); 241 if (!linalgOp) 242 return failure(); 243 if (failed(marker.checkAndNotify(rewriter, linalgOp))) 244 return failure(); 245 if (failed(interchangeGenericLinalgOpPrecondition(op, interchangeVector))) 246 return failure(); 247 248 // TODO: figure out how this interplays with named ops. In particular this 249 // should break the named op property. 250 rewriter.updateRootInPlace(op, [&]() { 251 interchange(linalgOp, interchangeVector); 252 // New marker if specified. 253 marker.replaceLinalgMarker(rewriter, op); 254 }); 255 return success(); 256 } 257 258 mlir::linalg::LinalgBasePromotionPattern::LinalgBasePromotionPattern( 259 StringRef opName, MLIRContext *context, LinalgPromotionOptions options, 260 LinalgMarker marker, PatternBenefit benefit) 261 : RewritePattern(opName, {}, benefit, context), marker(marker), 262 options(options) {} 263 264 LogicalResult mlir::linalg::LinalgBasePromotionPattern::matchAndRewrite( 265 Operation *op, PatternRewriter &rewriter) const { 266 if (failed(marker.checkAndNotify(rewriter, op))) 267 return failure(); 268 if (failed(promoteSubviewsPrecondition(op, options))) 269 return failure(); 270 271 // TODO: We cannot use root update here. This pattern is creating other ops, 272 // so if the promotion fails, those need to be cleaned up, which doesnt seem 273 // to be happening here. So to fail properly, we should be cloning the op and 274 // deleting the previous op. This needs more investigation. 275 rewriter.startRootUpdate(op); 276 Optional<LinalgOp> promotedOp = promoteSubViews(rewriter, op, options); 277 if (!promotedOp) { 278 rewriter.cancelRootUpdate(op); 279 return op->emitError("subview promotion failed"); 280 } 281 rewriter.finalizeRootUpdate(op); 282 marker.replaceLinalgMarker(rewriter, op); 283 return success(); 284 } 285 286 mlir::linalg::LinalgBaseVectorizationPattern::LinalgBaseVectorizationPattern( 287 StringRef opName, MLIRContext *context, LinalgMarker marker, 288 PatternBenefit benefit) 289 : RewritePattern(opName, {}, benefit, context), marker(marker) {} 290 291 LogicalResult mlir::linalg::LinalgBaseVectorizationPattern::matchAndRewrite( 292 Operation *op, PatternRewriter &rewriter) const { 293 LinalgOp linalgOp = dyn_cast<LinalgOp>(op); 294 if (!linalgOp) 295 return failure(); 296 if (failed(marker.checkAndNotify(rewriter, linalgOp))) 297 return failure(); 298 if (failed(vectorizeLinalgOpPrecondition(op))) 299 return failure(); 300 vectorizeLinalgOp(rewriter, op); 301 rewriter.eraseOp(op); 302 return success(); 303 } 304 305 LogicalResult mlir::linalg::applyStagedPatterns( 306 Operation *op, ArrayRef<FrozenRewritePatternList> stage1Patterns, 307 const FrozenRewritePatternList &stage2Patterns, 308 function_ref<LogicalResult(Operation *)> stage3Lambda) { 309 unsigned iteration = 0; 310 (void)iteration; 311 for (const auto &patterns : stage1Patterns) { 312 LLVM_DEBUG(DBGS() << "Before 1st stage, iter: " << ++iteration << "\n" 313 << *op); 314 if (failed(applyPatternsAndFoldGreedily(op, patterns))) { 315 LLVM_DEBUG(DBGS() << "Underlying first stage rewrite did not converge"); 316 return failure(); 317 } 318 LLVM_DEBUG(DBGS() << "After 1st stage, iter: " << ++iteration << "\n" 319 << *op); 320 if (failed(applyPatternsAndFoldGreedily(op, stage2Patterns))) { 321 LLVM_DEBUG(DBGS() << "Underlying 2nd stage rewrite did not converge"); 322 return failure(); 323 } 324 LLVM_DEBUG(DBGS() << "After 2nd stage, iter : " << iteration << "\n" 325 << *op); 326 if (stage3Lambda) { 327 if (failed(stage3Lambda(op))) 328 return failure(); 329 LLVM_DEBUG(DBGS() << "After 3rd stage, iter : " << iteration << "\n" 330 << *op); 331 } 332 } 333 return success(); 334 } 335 336 /// Given the `lbVal`, `ubVal` and `stepVal` of a loop, append `lbVal` and 337 /// `ubVal` to `dims` and `stepVal` to `symbols`. 338 /// Create new AffineDimExpr (`%lb` and `%ub`) and AffineSymbolExpr (`%step`) 339 /// with positions matching the newly appended values. Substitute occurrences of 340 /// `dimExpr` by either the min expression (i.e. `%lb`) or the max expression 341 /// (i.e. `%lb + %step * floordiv(%ub -1 - %lb, %step)`), depending on whether 342 /// the induction variable is used with a positive or negative coefficient. 343 static AffineExpr substituteLoopInExpr(AffineExpr expr, AffineExpr dimExpr, 344 Value lbVal, Value ubVal, Value stepVal, 345 SmallVectorImpl<Value> &dims, 346 SmallVectorImpl<Value> &symbols) { 347 MLIRContext *ctx = lbVal.getContext(); 348 AffineExpr lb = getAffineDimExpr(dims.size(), ctx); 349 dims.push_back(lbVal); 350 AffineExpr ub = getAffineDimExpr(dims.size(), ctx); 351 dims.push_back(ubVal); 352 AffineExpr step = getAffineSymbolExpr(symbols.size(), ctx); 353 symbols.push_back(stepVal); 354 LLVM_DEBUG(DBGS() << "Before: " << expr << "\n"); 355 AffineExpr ee = substWithMin(expr, dimExpr, lb, 356 lb + step * ((ub - 1) - lb).floorDiv(step)); 357 LLVM_DEBUG(DBGS() << "After: " << expr << "\n"); 358 return ee; 359 } 360 361 /// Traverse the `dims` and substitute known min or max expressions in place of 362 /// induction variables in `exprs`. 363 static AffineMap substitute(AffineMap map, SmallVectorImpl<Value> &dims, 364 SmallVectorImpl<Value> &symbols) { 365 auto exprs = llvm::to_vector<4>(map.getResults()); 366 for (AffineExpr &expr : exprs) { 367 bool substituted = true; 368 while (substituted) { 369 substituted = false; 370 for (unsigned dimIdx = 0; dimIdx < dims.size(); ++dimIdx) { 371 Value dim = dims[dimIdx]; 372 AffineExpr dimExpr = getAffineDimExpr(dimIdx, expr.getContext()); 373 LLVM_DEBUG(DBGS() << "Subst: " << dim << " @ " << dimExpr << "\n"); 374 AffineExpr substitutedExpr; 375 if (auto forOp = scf::getForInductionVarOwner(dim)) 376 substitutedExpr = substituteLoopInExpr( 377 expr, dimExpr, forOp.lowerBound(), forOp.upperBound(), 378 forOp.step(), dims, symbols); 379 380 if (auto parallelForOp = scf::getParallelForInductionVarOwner(dim)) 381 for (unsigned idx = 0, e = parallelForOp.getNumLoops(); idx < e; 382 ++idx) 383 substitutedExpr = substituteLoopInExpr( 384 expr, dimExpr, parallelForOp.lowerBound()[idx], 385 parallelForOp.upperBound()[idx], parallelForOp.step()[idx], 386 dims, symbols); 387 388 if (!substitutedExpr) 389 continue; 390 391 substituted = (substitutedExpr != expr); 392 expr = substitutedExpr; 393 } 394 } 395 396 // Cleanup and simplify the results. 397 // This needs to happen outside of the loop iterating on dims.size() since 398 // it modifies dims. 399 SmallVector<Value, 4> operands(dims.begin(), dims.end()); 400 operands.append(symbols.begin(), symbols.end()); 401 auto map = AffineMap::get(dims.size(), symbols.size(), exprs, 402 exprs.front().getContext()); 403 404 LLVM_DEBUG(DBGS() << "Map to simplify: " << map << "\n"); 405 406 // Pull in affine.apply operations and compose them fully into the 407 // result. 408 fullyComposeAffineMapAndOperands(&map, &operands); 409 canonicalizeMapAndOperands(&map, &operands); 410 map = simplifyAffineMap(map); 411 // Assign the results. 412 exprs.assign(map.getResults().begin(), map.getResults().end()); 413 dims.assign(operands.begin(), operands.begin() + map.getNumDims()); 414 symbols.assign(operands.begin() + map.getNumDims(), operands.end()); 415 416 LLVM_DEBUG(DBGS() << "Map simplified: " << map << "\n"); 417 } 418 419 assert(!exprs.empty() && "Unexpected empty exprs"); 420 return AffineMap::get(dims.size(), symbols.size(), exprs, map.getContext()); 421 } 422 423 LogicalResult AffineMinSCFCanonicalizationPattern::matchAndRewrite( 424 AffineMinOp minOp, PatternRewriter &rewriter) const { 425 LLVM_DEBUG(DBGS() << "Canonicalize AffineMinSCF: " << *minOp.getOperation() 426 << "\n"); 427 428 SmallVector<Value, 4> dims(minOp.getDimOperands()), 429 symbols(minOp.getSymbolOperands()); 430 AffineMap map = substitute(minOp.getAffineMap(), dims, symbols); 431 432 LLVM_DEBUG(DBGS() << "Resulting map: " << map << "\n"); 433 434 // Check whether any of the expressions, when subtracted from all other 435 // expressions, produces only >= 0 constants. If so, it is the min. 436 for (auto e : minOp.getAffineMap().getResults()) { 437 LLVM_DEBUG(DBGS() << "Candidate min: " << e << "\n"); 438 if (!e.isSymbolicOrConstant()) 439 continue; 440 441 auto isNonPositive = [](AffineExpr e) { 442 if (auto cst = e.dyn_cast<AffineConstantExpr>()) 443 return cst.getValue() < 0; 444 return true; 445 }; 446 447 // Build the subMap and check everything is statically known to be 448 // positive. 449 SmallVector<AffineExpr, 4> subExprs; 450 subExprs.reserve(map.getNumResults()); 451 for (auto ee : map.getResults()) 452 subExprs.push_back(ee - e); 453 MLIRContext *ctx = minOp.getContext(); 454 AffineMap subMap = simplifyAffineMap( 455 AffineMap::get(map.getNumDims(), map.getNumSymbols(), subExprs, ctx)); 456 LLVM_DEBUG(DBGS() << "simplified subMap: " << subMap << "\n"); 457 if (llvm::any_of(subMap.getResults(), isNonPositive)) 458 continue; 459 460 // Static min found. 461 if (auto cst = e.dyn_cast<AffineConstantExpr>()) { 462 rewriter.replaceOpWithNewOp<ConstantIndexOp>(minOp, cst.getValue()); 463 } else { 464 auto resultMap = AffineMap::get(0, map.getNumSymbols(), {e}, ctx); 465 SmallVector<Value, 4> resultOperands = dims; 466 resultOperands.append(symbols.begin(), symbols.end()); 467 canonicalizeMapAndOperands(&resultMap, &resultOperands); 468 resultMap = simplifyAffineMap(resultMap); 469 rewriter.replaceOpWithNewOp<AffineApplyOp>(minOp, resultMap, 470 resultOperands); 471 } 472 return success(); 473 } 474 475 return failure(); 476 } 477