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/Linalg/Analysis/DependenceAnalysis.h" 16 #include "mlir/Dialect/Linalg/IR/LinalgOps.h" 17 #include "mlir/Dialect/Linalg/Utils/Utils.h" 18 #include "mlir/Dialect/StandardOps/EDSC/Intrinsics.h" 19 #include "mlir/Dialect/Utils/StructuredOpsUtils.h" 20 #include "mlir/Dialect/Vector/EDSC/Intrinsics.h" 21 #include "mlir/Dialect/Vector/VectorOps.h" 22 #include "mlir/IR/AffineExpr.h" 23 #include "mlir/IR/Matchers.h" 24 #include "mlir/IR/PatternMatch.h" 25 #include "mlir/Pass/Pass.h" 26 #include "mlir/Support/LLVM.h" 27 #include "llvm/Support/Debug.h" 28 #include "llvm/Support/raw_ostream.h" 29 #include <type_traits> 30 31 #define DEBUG_TYPE "linalg-transforms" 32 33 using namespace mlir; 34 using namespace mlir::edsc; 35 using namespace mlir::edsc::intrinsics; 36 using namespace mlir::linalg; 37 38 #define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE << "]: ") 39 40 //===----------------------------------------------------------------------===// 41 // Transformations exposed as rewrite patterns. 42 //===----------------------------------------------------------------------===// 43 // Marker used as attribute name in generated Linalg rewriting transformations. 44 const StringLiteral mlir::linalg::LinalgTransforms::kLinalgTransformMarker = 45 "__internal_linalg_transform__"; 46 47 mlir::linalg::LinalgMarker::LinalgMarker(ArrayRef<Identifier> matchDisjunction, 48 Optional<Identifier> replacement) 49 : matchDisjunction(matchDisjunction.begin(), matchDisjunction.end()), 50 replacement(replacement) {} 51 52 LogicalResult 53 mlir::linalg::LinalgMarker::checkAndNotify(PatternRewriter &rewriter, 54 Operation *op) const { 55 auto attr = op->template getAttrOfType<StringAttr>( 56 LinalgTransforms::kLinalgTransformMarker); 57 58 if (!attr) { 59 // 1. Has no marker case and matchDisjunction is empty. 60 if (matchDisjunction.empty()) 61 return success(); 62 63 // 2. Has no marker but was expecting a marker. 64 return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) { 65 diag << " does not have any marker from list: "; 66 interleaveComma(matchDisjunction, diag); 67 }); 68 } 69 70 // 4. Match explicit marker. 71 for (auto marker : matchDisjunction) 72 if (attr.getValue() == marker) 73 return success(); 74 75 // 5. Fail to match. 76 return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) { 77 diag << " does not have any marker from list: "; 78 interleaveComma(matchDisjunction, diag); 79 }); 80 } 81 82 void mlir::linalg::LinalgMarker::replaceLinalgMarker(PatternRewriter &rewriter, 83 Operation *op) const { 84 if (replacement.hasValue()) 85 op->setAttr(LinalgTransforms::kLinalgTransformMarker, 86 rewriter.getStringAttr(replacement.getValue())); 87 else 88 op->removeAttr(Identifier::get(LinalgTransforms::kLinalgTransformMarker, 89 rewriter.getContext())); 90 } 91 92 LinalgTilingOptions & 93 mlir::linalg::LinalgTilingOptions::setTileSizes(ArrayRef<int64_t> ts) { 94 SmallVector<int64_t, 4> tileSizes(ts.begin(), ts.end()); 95 tileSizeComputationFunction = [tileSizes](OpBuilder &b, Operation *op) { 96 OpBuilder::InsertionGuard guard(b); 97 b.setInsertionPointToStart( 98 &op->getParentOfType<FuncOp>().getBody().front()); 99 return llvm::to_vector<4>(map_range(tileSizes, [&](int64_t s) { 100 Value v = b.create<ConstantIndexOp>(op->getLoc(), s); 101 return v; 102 })); 103 }; 104 return *this; 105 } 106 107 /// Linalg base tiling pattern. 108 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern( 109 StringRef opName, MLIRContext *context, LinalgTilingOptions options, 110 LinalgMarker marker, PatternBenefit benefit) 111 : RewritePattern(opName, {}, benefit, context), marker(marker), 112 options(options) {} 113 114 LogicalResult mlir::linalg::LinalgBaseTilingPattern::matchAndRewrite( 115 Operation *op, PatternRewriter &rewriter) const { 116 LinalgOp linalgOp = dyn_cast<LinalgOp>(op); 117 if (!linalgOp) 118 return failure(); 119 if (failed(marker.checkAndNotify(rewriter, linalgOp))) 120 return failure(); 121 122 Optional<TiledLinalgOp> res = tileLinalgOp(rewriter, linalgOp, options); 123 124 if (!res) 125 return failure(); 126 127 // New marker if specified. 128 marker.replaceLinalgMarker(rewriter, res->op.getOperation()); 129 return success(); 130 } 131 132 /// Linalg base interchange pattern. 133 mlir::linalg::LinalgBaseInterchangePattern::LinalgBaseInterchangePattern( 134 StringRef opName, MLIRContext *context, 135 ArrayRef<unsigned> interchangeVector, LinalgMarker marker, 136 PatternBenefit benefit) 137 : RewritePattern(opName, {}, benefit, context), marker(marker), 138 interchangeVector(interchangeVector.begin(), interchangeVector.end()) {} 139 140 LogicalResult mlir::linalg::LinalgBaseInterchangePattern::matchAndRewrite( 141 Operation *op, PatternRewriter &rewriter) const { 142 LinalgOp linalgOp = dyn_cast<LinalgOp>(op); 143 if (!linalgOp) 144 return failure(); 145 if (failed(marker.checkAndNotify(rewriter, linalgOp))) 146 return failure(); 147 if (failed(interchangeGenericLinalgOpPrecondition(op, interchangeVector))) 148 return failure(); 149 150 // TODO: figure out how this interplays with named ops. In particular this 151 // should break the named op property. 152 rewriter.updateRootInPlace(op, [&]() { 153 interchange(linalgOp, interchangeVector); 154 // New marker if specified. 155 marker.replaceLinalgMarker(rewriter, op); 156 }); 157 return success(); 158 } 159 160 mlir::linalg::LinalgBasePromotionPattern::LinalgBasePromotionPattern( 161 StringRef opName, MLIRContext *context, LinalgPromotionOptions options, 162 LinalgMarker marker, PatternBenefit benefit) 163 : RewritePattern(opName, {}, benefit, context), marker(marker), 164 options(options) {} 165 166 LogicalResult mlir::linalg::LinalgBasePromotionPattern::matchAndRewrite( 167 Operation *op, PatternRewriter &rewriter) const { 168 if (failed(marker.checkAndNotify(rewriter, op))) 169 return failure(); 170 if (failed(promoteSubviewsPrecondition(op, options))) 171 return failure(); 172 173 // TODO: We cannot use root update here. This pattern is creating other ops, 174 // so if the promotion fails, those need to be cleaned up, which doesnt seem 175 // to be happening here. So to fail properly, we should be cloning the op and 176 // deleting the previous op. This needs more investigation. 177 rewriter.startRootUpdate(op); 178 Optional<LinalgOp> promotedOp = promoteSubViews(rewriter, op, options); 179 if (!promotedOp) { 180 rewriter.cancelRootUpdate(op); 181 return op->emitError("subview promotion failed"); 182 } 183 rewriter.finalizeRootUpdate(op); 184 marker.replaceLinalgMarker(rewriter, op); 185 return success(); 186 } 187 188 mlir::linalg::LinalgBaseVectorizationPattern::LinalgBaseVectorizationPattern( 189 StringRef opName, MLIRContext *context, LinalgMarker marker, 190 PatternBenefit benefit) 191 : RewritePattern(opName, {}, benefit, context), marker(marker) {} 192 193 LogicalResult mlir::linalg::LinalgBaseVectorizationPattern::matchAndRewrite( 194 Operation *op, PatternRewriter &rewriter) const { 195 LinalgOp linalgOp = dyn_cast<LinalgOp>(op); 196 if (!linalgOp) 197 return failure(); 198 if (failed(marker.checkAndNotify(rewriter, linalgOp))) 199 return failure(); 200 if (failed(vectorizeLinalgOpPrecondition(op))) 201 return failure(); 202 vectorizeLinalgOp(rewriter, op); 203 rewriter.eraseOp(op); 204 return success(); 205 } 206 207 LogicalResult mlir::linalg::applyStagedPatterns( 208 Operation *op, ArrayRef<OwningRewritePatternList> stage1Patterns, 209 const OwningRewritePatternList &stage2Patterns, 210 function_ref<LogicalResult(Operation *)> stage3Lambda) { 211 unsigned iteration = 0; 212 (void)iteration; 213 for (const auto &patterns : stage1Patterns) { 214 LLVM_DEBUG(DBGS() << "Before 1st stage, iter: " << ++iteration << "\n" 215 << *op); 216 if (failed(applyPatternsAndFoldGreedily(op, patterns))) { 217 LLVM_DEBUG(DBGS() << "Underlying first stage rewrite did not converge"); 218 return failure(); 219 } 220 LLVM_DEBUG(DBGS() << "After 1st stage, iter: " << ++iteration << "\n" 221 << *op); 222 if (failed(applyPatternsAndFoldGreedily(op, stage2Patterns))) { 223 LLVM_DEBUG(DBGS() << "Underlying 2nd stage rewrite did not converge"); 224 return failure(); 225 } 226 LLVM_DEBUG(DBGS() << "After 2nd stage, iter : " << iteration << "\n" 227 << *op); 228 if (stage3Lambda) { 229 if (failed(stage3Lambda(op))) 230 return failure(); 231 LLVM_DEBUG(DBGS() << "After 3rd stage, iter : " << iteration << "\n" 232 << *op); 233 } 234 } 235 return success(); 236 } 237 238 /// Traverse `e` and return an AffineExpr where all occurrences of `dim` have 239 /// been replaced by either: 240 /// - `min` if `positivePath` is true when we reach an occurrence of `dim` 241 /// - `max` if `positivePath` is true when we reach an occurrence of `dim` 242 /// `positivePath` is negated each time we hit a multiplicative or divisive 243 /// binary op with a constant negative coefficient. 244 static AffineExpr substWithMin(AffineExpr e, AffineExpr dim, AffineExpr min, 245 AffineExpr max, bool positivePath = true) { 246 if (e == dim) 247 return positivePath ? min : max; 248 if (auto bin = e.dyn_cast<AffineBinaryOpExpr>()) { 249 AffineExpr lhs = bin.getLHS(); 250 AffineExpr rhs = bin.getRHS(); 251 if (bin.getKind() == mlir::AffineExprKind::Add) 252 return substWithMin(lhs, dim, min, max, positivePath) + 253 substWithMin(rhs, dim, min, max, positivePath); 254 255 auto c1 = bin.getLHS().dyn_cast<AffineConstantExpr>(); 256 auto c2 = bin.getRHS().dyn_cast<AffineConstantExpr>(); 257 if (c1 && c1.getValue() < 0) 258 return getAffineBinaryOpExpr( 259 bin.getKind(), c1, substWithMin(rhs, dim, min, max, !positivePath)); 260 if (c2 && c2.getValue() < 0) 261 return getAffineBinaryOpExpr( 262 bin.getKind(), substWithMin(lhs, dim, min, max, !positivePath), c2); 263 return getAffineBinaryOpExpr( 264 bin.getKind(), substWithMin(lhs, dim, min, max, positivePath), 265 substWithMin(rhs, dim, min, max, positivePath)); 266 } 267 return e; 268 } 269 270 /// Given the `lbVal`, `ubVal` and `stepVal` of a loop, append `lbVal` and 271 /// `ubVal` to `dims` and `stepVal` to `symbols`. 272 /// Create new AffineDimExpr (`%lb` and `%ub`) and AffineSymbolExpr (`%step`) 273 /// with positions matching the newly appended values. Substitute occurrences of 274 /// `dimExpr` by either the min expression (i.e. `%lb`) or the max expression 275 /// (i.e. `%lb + %step * floordiv(%ub -1 - %lb, %step)`), depending on whether 276 /// the induction variable is used with a positive or negative coefficient. 277 static AffineExpr substituteLoopInExpr(AffineExpr expr, AffineExpr dimExpr, 278 Value lbVal, Value ubVal, Value stepVal, 279 SmallVectorImpl<Value> &dims, 280 SmallVectorImpl<Value> &symbols) { 281 MLIRContext *ctx = lbVal.getContext(); 282 AffineExpr lb = getAffineDimExpr(dims.size(), ctx); 283 dims.push_back(lbVal); 284 AffineExpr ub = getAffineDimExpr(dims.size(), ctx); 285 dims.push_back(ubVal); 286 AffineExpr step = getAffineSymbolExpr(symbols.size(), ctx); 287 symbols.push_back(stepVal); 288 LLVM_DEBUG(DBGS() << "Before: " << expr << "\n"); 289 AffineExpr ee = substWithMin(expr, dimExpr, lb, 290 lb + step * ((ub - 1) - lb).floorDiv(step)); 291 LLVM_DEBUG(DBGS() << "After: " << expr << "\n"); 292 return ee; 293 } 294 295 /// Traverse the `dims` and substitute known min or max expressions in place of 296 /// induction variables in `exprs`. 297 static AffineMap substitute(AffineMap map, SmallVectorImpl<Value> &dims, 298 SmallVectorImpl<Value> &symbols) { 299 auto exprs = llvm::to_vector<4>(map.getResults()); 300 for (AffineExpr &expr : exprs) { 301 bool substituted = true; 302 while (substituted) { 303 substituted = false; 304 for (unsigned dimIdx = 0; dimIdx < dims.size(); ++dimIdx) { 305 Value dim = dims[dimIdx]; 306 AffineExpr dimExpr = getAffineDimExpr(dimIdx, expr.getContext()); 307 LLVM_DEBUG(DBGS() << "Subst: " << dim << " @ " << dimExpr << "\n"); 308 AffineExpr substitutedExpr; 309 if (auto forOp = scf::getForInductionVarOwner(dim)) 310 substitutedExpr = substituteLoopInExpr( 311 expr, dimExpr, forOp.lowerBound(), forOp.upperBound(), 312 forOp.step(), dims, symbols); 313 314 if (auto parallelForOp = scf::getParallelForInductionVarOwner(dim)) 315 for (unsigned idx = 0, e = parallelForOp.getNumLoops(); idx < e; 316 ++idx) 317 substitutedExpr = substituteLoopInExpr( 318 expr, dimExpr, parallelForOp.lowerBound()[idx], 319 parallelForOp.upperBound()[idx], parallelForOp.step()[idx], 320 dims, symbols); 321 322 if (!substitutedExpr) 323 continue; 324 325 substituted = (substitutedExpr != expr); 326 expr = substitutedExpr; 327 } 328 } 329 330 // Cleanup and simplify the results. 331 // This needs to happen outside of the loop iterating on dims.size() since 332 // it modifies dims. 333 SmallVector<Value, 4> operands(dims.begin(), dims.end()); 334 operands.append(symbols.begin(), symbols.end()); 335 auto map = AffineMap::get(dims.size(), symbols.size(), exprs, 336 exprs.front().getContext()); 337 338 LLVM_DEBUG(DBGS() << "Map to simplify: " << map << "\n"); 339 340 // Pull in affine.apply operations and compose them fully into the 341 // result. 342 fullyComposeAffineMapAndOperands(&map, &operands); 343 canonicalizeMapAndOperands(&map, &operands); 344 map = simplifyAffineMap(map); 345 // Assign the results. 346 exprs.assign(map.getResults().begin(), map.getResults().end()); 347 dims.assign(operands.begin(), operands.begin() + map.getNumDims()); 348 symbols.assign(operands.begin() + map.getNumDims(), operands.end()); 349 350 LLVM_DEBUG(DBGS() << "Map simplified: " << map << "\n"); 351 } 352 353 assert(!exprs.empty() && "Unexpected empty exprs"); 354 return AffineMap::get(dims.size(), symbols.size(), exprs, map.getContext()); 355 } 356 357 LogicalResult AffineMinSCFCanonicalizationPattern::matchAndRewrite( 358 AffineMinOp minOp, PatternRewriter &rewriter) const { 359 LLVM_DEBUG(DBGS() << "Canonicalize AffineMinSCF: " << *minOp.getOperation() 360 << "\n"); 361 362 SmallVector<Value, 4> dims(minOp.getDimOperands()), 363 symbols(minOp.getSymbolOperands()); 364 AffineMap map = substitute(minOp.getAffineMap(), dims, symbols); 365 366 LLVM_DEBUG(DBGS() << "Resulting map: " << map << "\n"); 367 368 // Check whether any of the expressions, when subtracted from all other 369 // expressions, produces only >= 0 constants. If so, it is the min. 370 for (auto e : minOp.getAffineMap().getResults()) { 371 LLVM_DEBUG(DBGS() << "Candidate min: " << e << "\n"); 372 if (!e.isSymbolicOrConstant()) 373 continue; 374 375 auto isNonPositive = [](AffineExpr e) { 376 if (auto cst = e.dyn_cast<AffineConstantExpr>()) 377 return cst.getValue() < 0; 378 return true; 379 }; 380 381 // Build the subMap and check everything is statically known to be 382 // positive. 383 SmallVector<AffineExpr, 4> subExprs; 384 subExprs.reserve(map.getNumResults()); 385 for (auto ee : map.getResults()) 386 subExprs.push_back(ee - e); 387 MLIRContext *ctx = minOp.getContext(); 388 AffineMap subMap = simplifyAffineMap( 389 AffineMap::get(map.getNumDims(), map.getNumSymbols(), subExprs, ctx)); 390 LLVM_DEBUG(DBGS() << "simplified subMap: " << subMap << "\n"); 391 if (llvm::any_of(subMap.getResults(), isNonPositive)) 392 continue; 393 394 // Static min found. 395 if (auto cst = e.dyn_cast<AffineConstantExpr>()) { 396 rewriter.replaceOpWithNewOp<ConstantIndexOp>(minOp, cst.getValue()); 397 } else { 398 auto resultMap = AffineMap::get(0, map.getNumSymbols(), {e}, ctx); 399 SmallVector<Value, 4> resultOperands = dims; 400 resultOperands.append(symbols.begin(), symbols.end()); 401 canonicalizeMapAndOperands(&resultMap, &resultOperands); 402 resultMap = simplifyAffineMap(resultMap); 403 rewriter.replaceOpWithNewOp<AffineApplyOp>(minOp, resultMap, 404 resultOperands); 405 } 406 return success(); 407 } 408 409 return failure(); 410 } 411