1 //===- VectorToGPU.cpp - Convert vector to GPU dialect ----------*- C++ -*-===// 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 lowering of vector operations to GPU dialect ops. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include <type_traits> 14 15 #include "mlir/Conversion/VectorToGPU/VectorToGPU.h" 16 17 #include "../PassDetail.h" 18 #include "mlir/Analysis/SliceAnalysis.h" 19 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h" 20 #include "mlir/Dialect/GPU/GPUDialect.h" 21 #include "mlir/Dialect/MemRef/IR/MemRef.h" 22 #include "mlir/Dialect/SCF/SCF.h" 23 #include "mlir/Dialect/Utils/StructuredOpsUtils.h" 24 #include "mlir/Dialect/Vector/VectorOps.h" 25 #include "mlir/Dialect/Vector/VectorUtils.h" 26 #include "mlir/IR/Builders.h" 27 #include "mlir/Pass/Pass.h" 28 #include "mlir/Transforms/GreedyPatternRewriteDriver.h" 29 #include "mlir/Transforms/Passes.h" 30 31 using namespace mlir; 32 33 // Return true if the contract op can be convert to MMA matmul. 34 static bool contractSupportsMMAMatrixType(vector::ContractionOp contract) { 35 if (llvm::size(contract.masks()) != 0) 36 return false; 37 38 using MapList = ArrayRef<ArrayRef<AffineExpr>>; 39 auto infer = [](MapList m) { return AffineMap::inferFromExprList(m); }; 40 AffineExpr m, n, k; 41 bindDims(contract.getContext(), m, n, k); 42 auto iteratorTypes = contract.iterator_types().getValue(); 43 if (!(isParallelIterator(iteratorTypes[0]) && 44 isParallelIterator(iteratorTypes[1]) && 45 isReductionIterator(iteratorTypes[2]))) 46 return false; 47 48 // The contract needs to represent a matmul to be able to convert to 49 // MMAMatrix matmul. 50 if (contract.getIndexingMaps() != infer({{m, k}, {k, n}, {m, n}})) 51 return false; 52 53 return true; 54 } 55 56 // Return the stide for the dimension 0 of |type| if it is a memref and has a 57 // constant stride. 58 static llvm::Optional<int64_t> 59 getMemrefConstantHorizontalStride(ShapedType type) { 60 auto memrefType = type.dyn_cast<MemRefType>(); 61 if (!memrefType) 62 return false; 63 int64_t offset = 0; 64 SmallVector<int64_t, 2> strides; 65 if (failed(getStridesAndOffset(memrefType, strides, offset))) 66 return llvm::None; 67 if (strides[0] == ShapedType::kDynamicStrideOrOffset) 68 return llvm::None; 69 return strides[0]; 70 } 71 72 // Return true if the transfer op can be converted to a MMA matrix load. 73 static bool transferReadSupportsMMAMatrixType(vector::TransferReadOp readOp) { 74 if (readOp.mask() || readOp.hasOutOfBoundsDim() || 75 readOp.getVectorType().getRank() != 2) 76 return false; 77 if (!getMemrefConstantHorizontalStride(readOp.getShapedType())) 78 return false; 79 AffineMap map = readOp.permutation_map(); 80 OpBuilder b(readOp.getContext()); 81 AffineExpr innerDim = b.getAffineDimExpr(map.getNumDims() - 1); 82 AffineExpr zero = b.getAffineConstantExpr(0); 83 auto broadcastInnerDim = AffineMap::get(map.getNumDims(), 0, {zero, innerDim}, 84 readOp.getContext()); 85 // TODO: Support transpose once it is added to GPU dialect ops. 86 // For now we only support (d0, d1) -> (d0, d1) and (d0, d1) -> (0, d1). 87 return !(!map.isMinorIdentity() && map != broadcastInnerDim); 88 } 89 90 // Return true if the transfer op can be converted to a MMA matrix store. 91 static bool 92 transferWriteSupportsMMAMatrixType(vector::TransferWriteOp writeOp) { 93 // TODO: support 0-d corner case. 94 if (writeOp.getTransferRank() == 0) 95 return false; 96 97 if (writeOp.mask() || writeOp.hasOutOfBoundsDim() || 98 writeOp.getVectorType().getRank() != 2) 99 return false; 100 if (!getMemrefConstantHorizontalStride(writeOp.getShapedType())) 101 return false; 102 // TODO: Support transpose once it is added to GPU dialect ops. 103 if (!writeOp.permutation_map().isMinorIdentity()) 104 return false; 105 return true; 106 } 107 108 /// Return true if the constant is a splat to a 2D vector so that it can be 109 /// converted to a MMA constant matrix op. 110 static bool constantSupportsMMAMatrixType(arith::ConstantOp constantOp) { 111 auto vecType = constantOp.getType().dyn_cast<VectorType>(); 112 if (!vecType || vecType.getRank() != 2) 113 return false; 114 return constantOp.getValue().isa<SplatElementsAttr>(); 115 } 116 117 /// Return true if this is a broadcast from scalar to a 2D vector. 118 static bool broadcastSupportsMMAMatrixType(vector::BroadcastOp broadcastOp) { 119 return broadcastOp.getVectorType().getRank() == 2 && 120 broadcastOp.source().getType().isa<FloatType>(); 121 } 122 123 /// Return the MMA elementwise enum associated with `op` if it is supported. 124 /// Return `llvm::None` otherwise. 125 static llvm::Optional<gpu::MMAElementwiseOp> 126 convertElementwiseOpToMMA(Operation *op) { 127 if (isa<arith::AddFOp>(op)) 128 return gpu::MMAElementwiseOp::ADDF; 129 if (isa<arith::MulFOp>(op)) 130 return gpu::MMAElementwiseOp::MULF; 131 if (isa<arith::MaxFOp>(op)) 132 return gpu::MMAElementwiseOp::MAXF; 133 if (isa<arith::MinFOp>(op)) 134 return gpu::MMAElementwiseOp::MINF; 135 if (isa<arith::DivFOp>(op)) 136 return gpu::MMAElementwiseOp::DIVF; 137 return llvm::None; 138 } 139 140 /// Return true if the op is supported as elementwise op on MMAMatrix type. 141 static bool elementwiseSupportsMMAMatrixType(Operation *op) { 142 return convertElementwiseOpToMMA(op).hasValue(); 143 } 144 145 static bool supportsMMaMatrixType(Operation *op) { 146 if (isa<scf::ForOp, scf::YieldOp>(op)) 147 return true; 148 if (auto transferRead = dyn_cast<vector::TransferReadOp>(op)) 149 return transferReadSupportsMMAMatrixType(transferRead); 150 if (auto transferWrite = dyn_cast<vector::TransferWriteOp>(op)) 151 return transferWriteSupportsMMAMatrixType(transferWrite); 152 if (auto contract = dyn_cast<vector::ContractionOp>(op)) 153 return contractSupportsMMAMatrixType(contract); 154 if (auto constant = dyn_cast<arith::ConstantOp>(op)) 155 return constantSupportsMMAMatrixType(constant); 156 if (auto broadcast = dyn_cast<vector::BroadcastOp>(op)) 157 return broadcastSupportsMMAMatrixType(broadcast); 158 return elementwiseSupportsMMAMatrixType(op); 159 } 160 161 /// Return an unsorted slice handling scf.for region differently than 162 /// `getSlice`. In scf.for we only want to include as part of the slice elements 163 /// that are part of the use/def chain. 164 static SetVector<Operation *> getSliceContract(Operation *op, 165 TransitiveFilter backwardFilter, 166 TransitiveFilter forwardFilter) { 167 SetVector<Operation *> slice; 168 slice.insert(op); 169 unsigned currentIndex = 0; 170 SetVector<Operation *> backwardSlice; 171 SetVector<Operation *> forwardSlice; 172 while (currentIndex != slice.size()) { 173 auto *currentOp = (slice)[currentIndex]; 174 // Compute and insert the backwardSlice starting from currentOp. 175 backwardSlice.clear(); 176 getBackwardSlice(currentOp, &backwardSlice, backwardFilter); 177 slice.insert(backwardSlice.begin(), backwardSlice.end()); 178 179 // Compute and insert the forwardSlice starting from currentOp. 180 forwardSlice.clear(); 181 // Special case for ForOp, we don't want to include the whole region but 182 // only the value using the region arguments. 183 // TODO: We should refine this to only care about the region arguments being 184 // converted to matrix type. 185 if (auto forOp = dyn_cast<scf::ForOp>(currentOp)) { 186 for (Value forOpResult : forOp.getResults()) 187 getForwardSlice(forOpResult, &forwardSlice, forwardFilter); 188 for (BlockArgument &arg : forOp.getRegionIterArgs()) 189 getForwardSlice(arg, &forwardSlice, forwardFilter); 190 } else { 191 getForwardSlice(currentOp, &forwardSlice, forwardFilter); 192 } 193 slice.insert(forwardSlice.begin(), forwardSlice.end()); 194 ++currentIndex; 195 } 196 return slice; 197 } 198 199 // Analyze slice of operations based on convert op to figure out if the whole 200 // slice can be converted to MMA operations. 201 static SetVector<Operation *> getOpToConvert(mlir::Operation *op) { 202 auto hasVectorDest = [](Operation *op) { 203 return llvm::any_of(op->getResultTypes(), 204 [](Type t) { return t.isa<VectorType>(); }); 205 }; 206 auto hasVectorSrc = [](Operation *op) { 207 return llvm::any_of(op->getOperandTypes(), 208 [](Type t) { return t.isa<VectorType>(); }); 209 }; 210 SetVector<Operation *> opToConvert; 211 op->walk([&](vector::ContractionOp contract) { 212 if (opToConvert.contains(contract.getOperation())) 213 return; 214 SetVector<Operation *> dependentOps = 215 getSliceContract(contract, hasVectorDest, hasVectorSrc); 216 // If any instruction cannot use MMA matrix type drop the whole 217 // chain. MMA matrix are stored in an opaque type so they cannot be used 218 // by all operations. 219 if (llvm::any_of(dependentOps, 220 [](Operation *op) { return !supportsMMaMatrixType(op); })) 221 return; 222 opToConvert.insert(dependentOps.begin(), dependentOps.end()); 223 }); 224 // Sort the operations so that we can convert them in topological order. 225 return topologicalSort(opToConvert); 226 } 227 228 namespace { 229 // Transform contract into (m, k)x(k, n)x(m, n) form so that it can be converted 230 // to MMA matmul. 231 struct PrepareContractToGPUMMA 232 : public OpRewritePattern<vector::ContractionOp> { 233 using OpRewritePattern<vector::ContractionOp>::OpRewritePattern; 234 235 LogicalResult matchAndRewrite(vector::ContractionOp op, 236 PatternRewriter &rewriter) const override { 237 Location loc = op.getLoc(); 238 Value lhs = op.lhs(), rhs = op.rhs(), res = op.acc(); 239 240 // Set up the parallel/reduction structure in right form. 241 using MapList = ArrayRef<ArrayRef<AffineExpr>>; 242 auto infer = [](MapList m) { return AffineMap::inferFromExprList(m); }; 243 AffineExpr m, n, k; 244 bindDims(rewriter.getContext(), m, n, k); 245 static constexpr std::array<int64_t, 2> perm = {1, 0}; 246 auto iteratorTypes = op.iterator_types().getValue(); 247 SmallVector<AffineMap, 4> maps = op.getIndexingMaps(); 248 if (!(isParallelIterator(iteratorTypes[0]) && 249 isParallelIterator(iteratorTypes[1]) && 250 isReductionIterator(iteratorTypes[2]))) 251 return failure(); 252 // 253 // Two outer parallel, one inner reduction (matmat flavor). 254 // 255 if (maps == infer({{m, k}, {k, n}, {m, n}})) { 256 // This is the classical row-major matmul, nothing to do. 257 return failure(); 258 } 259 if (maps == infer({{m, k}, {n, k}, {m, n}})) { 260 rhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm); 261 } else if (maps == infer({{k, m}, {k, n}, {m, n}})) { 262 lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm); 263 } else if (maps == infer({{k, m}, {n, k}, {m, n}})) { 264 rhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm); 265 lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm); 266 } else if (maps == infer({{m, k}, {k, n}, {n, m}})) { 267 std::swap(rhs, lhs); 268 rhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm); 269 lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm); 270 } else if (maps == infer({{m, k}, {n, k}, {n, m}})) { 271 std::swap(rhs, lhs); 272 rhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm); 273 } else if (maps == infer({{k, m}, {k, n}, {n, m}})) { 274 std::swap(lhs, rhs); 275 lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm); 276 } else if (maps == infer({{k, m}, {n, k}, {n, m}})) { 277 std::swap(lhs, rhs); 278 } else { 279 return failure(); 280 } 281 rewriter.replaceOpWithNewOp<vector::ContractionOp>( 282 op, lhs, rhs, res, 283 rewriter.getAffineMapArrayAttr(infer({{m, k}, {k, n}, {m, n}})), 284 op.iterator_types()); 285 return success(); 286 } 287 }; 288 289 // Merge transpose op into the transfer read op. Transpose are not supported on 290 // MMA types but MMA load can transpose the matrix when loading. 291 struct CombineTransferReadOpTranspose final 292 : public OpRewritePattern<vector::TransposeOp> { 293 using OpRewritePattern<vector::TransposeOp>::OpRewritePattern; 294 295 LogicalResult matchAndRewrite(vector::TransposeOp op, 296 PatternRewriter &rewriter) const override { 297 auto transferReadOp = op.vector().getDefiningOp<vector::TransferReadOp>(); 298 if (!transferReadOp) 299 return failure(); 300 301 // TODO: support 0-d corner case. 302 if (transferReadOp.getTransferRank() == 0) 303 return failure(); 304 305 if (transferReadOp.mask() || transferReadOp.hasOutOfBoundsDim()) 306 return failure(); 307 SmallVector<int64_t, 2> perm; 308 op.getTransp(perm); 309 SmallVector<unsigned, 2> permU; 310 for (int64_t o : perm) 311 permU.push_back(unsigned(o)); 312 AffineMap permutationMap = 313 AffineMap::getPermutationMap(permU, op.getContext()); 314 AffineMap newMap = permutationMap.compose(transferReadOp.permutation_map()); 315 rewriter.replaceOpWithNewOp<vector::TransferReadOp>( 316 op, op.getType(), transferReadOp.source(), transferReadOp.indices(), 317 AffineMapAttr::get(newMap), transferReadOp.padding(), 318 transferReadOp.mask(), transferReadOp.in_boundsAttr()); 319 return success(); 320 } 321 }; 322 323 } // namespace 324 325 // MMA types have different layout based on how they are used in matmul ops. 326 // Figure the right layout to use by looking at op uses. 327 // TODO: Change the GPU dialect to abstract the layout at the this level and 328 // only care about it during lowering to NVVM. 329 template <typename OpTy> 330 static const char *inferFragType(OpTy op) { 331 for (Operation *users : op->getUsers()) { 332 auto contract = dyn_cast<vector::ContractionOp>(users); 333 if (!contract) 334 continue; 335 if (contract.lhs() == op.getResult()) 336 return "AOp"; 337 if (contract.rhs() == op.getResult()) 338 return "BOp"; 339 } 340 return "COp"; 341 } 342 343 static void convertTransferReadOp(vector::TransferReadOp op, 344 llvm::DenseMap<Value, Value> &valueMapping) { 345 assert(op.getTransferRank() > 0 && "unexpected 0-d transfer"); 346 assert(transferReadSupportsMMAMatrixType(op)); 347 Optional<int64_t> stride = 348 getMemrefConstantHorizontalStride(op.getShapedType()); 349 AffineMap map = op.permutation_map(); 350 // Handle broadcast by setting the stride to 0. 351 if (map.getResult(0).isa<AffineConstantExpr>()) { 352 assert(map.getResult(0).cast<AffineConstantExpr>().getValue() == 0); 353 stride = 0; 354 } 355 assert(stride); 356 const char *fragType = inferFragType(op); 357 gpu::MMAMatrixType type = 358 gpu::MMAMatrixType::get(op.getVectorType().getShape(), 359 op.getVectorType().getElementType(), fragType); 360 OpBuilder b(op); 361 Value load = b.create<gpu::SubgroupMmaLoadMatrixOp>( 362 op.getLoc(), type, op.source(), op.indices(), b.getIndexAttr(*stride)); 363 valueMapping[op.getResult()] = load; 364 } 365 366 static void convertTransferWriteOp(vector::TransferWriteOp op, 367 llvm::DenseMap<Value, Value> &valueMapping) { 368 assert(transferWriteSupportsMMAMatrixType(op)); 369 Optional<int64_t> stride = 370 getMemrefConstantHorizontalStride(op.getShapedType()); 371 assert(stride); 372 OpBuilder b(op); 373 Value matrix = valueMapping.find(op.vector())->second; 374 b.create<gpu::SubgroupMmaStoreMatrixOp>( 375 op.getLoc(), matrix, op.source(), op.indices(), b.getIndexAttr(*stride)); 376 op.erase(); 377 } 378 379 static void convertContractOp(vector::ContractionOp op, 380 llvm::DenseMap<Value, Value> &valueMapping) { 381 OpBuilder b(op); 382 Value opA = valueMapping.find(op.lhs())->second; 383 Value opB = valueMapping.find(op.rhs())->second; 384 Value opC = valueMapping.find(op.acc())->second; 385 Value matmul = b.create<gpu::SubgroupMmaComputeOp>(op.getLoc(), opC.getType(), 386 opA, opB, opC); 387 valueMapping[op.getResult()] = matmul; 388 } 389 390 /// Convert a 2D splat ConstantOp to a SubgroupMmaConstantMatrix op. 391 static void convertConstantOp(arith::ConstantOp op, 392 llvm::DenseMap<Value, Value> &valueMapping) { 393 assert(constantSupportsMMAMatrixType(op)); 394 OpBuilder b(op); 395 Attribute splat = 396 op.getValue().cast<SplatElementsAttr>().getSplatValue<Attribute>(); 397 auto scalarConstant = 398 b.create<arith::ConstantOp>(op.getLoc(), splat.getType(), splat); 399 const char *fragType = inferFragType(op); 400 auto vecType = op.getType().cast<VectorType>(); 401 gpu::MMAMatrixType type = gpu::MMAMatrixType::get( 402 vecType.getShape(), vecType.getElementType(), llvm::StringRef(fragType)); 403 auto matrix = b.create<gpu::SubgroupMmaConstantMatrixOp>(op.getLoc(), type, 404 scalarConstant); 405 valueMapping[op.getResult()] = matrix; 406 } 407 408 /// Convert a vector.broadcast from scalar to a SubgroupMmaConstantMatrix op. 409 static void convertBroadcastOp(vector::BroadcastOp op, 410 llvm::DenseMap<Value, Value> &valueMapping) { 411 assert(broadcastSupportsMMAMatrixType(op)); 412 OpBuilder b(op); 413 const char *fragType = inferFragType(op); 414 auto vecType = op.getVectorType(); 415 gpu::MMAMatrixType type = gpu::MMAMatrixType::get( 416 vecType.getShape(), vecType.getElementType(), llvm::StringRef(fragType)); 417 auto matrix = b.create<gpu::SubgroupMmaConstantMatrixOp>(op.getLoc(), type, 418 op.source()); 419 valueMapping[op.getResult()] = matrix; 420 } 421 422 // Replace ForOp with a new ForOp with extra operands. The YieldOp is not 423 // updated and needs to be updated separatly for the loop to be correct. 424 static scf::ForOp replaceForOpWithNewSignature(OpBuilder &b, scf::ForOp loop, 425 ValueRange newIterOperands) { 426 // Create a new loop before the existing one, with the extra operands. 427 OpBuilder::InsertionGuard g(b); 428 b.setInsertionPoint(loop); 429 auto operands = llvm::to_vector<4>(loop.getIterOperands()); 430 operands.append(newIterOperands.begin(), newIterOperands.end()); 431 scf::ForOp newLoop = 432 b.create<scf::ForOp>(loop.getLoc(), loop.getLowerBound(), 433 loop.getUpperBound(), loop.getStep(), operands); 434 newLoop.getBody()->erase(); 435 newLoop.getLoopBody().getBlocks().splice( 436 newLoop.getLoopBody().getBlocks().begin(), 437 loop.getLoopBody().getBlocks()); 438 for (auto operand : newIterOperands) 439 newLoop.getBody()->addArgument(operand.getType()); 440 441 for (auto it : llvm::zip(loop.getResults(), newLoop.getResults().take_front( 442 loop.getNumResults()))) 443 std::get<0>(it).replaceAllUsesWith(std::get<1>(it)); 444 loop.erase(); 445 return newLoop; 446 } 447 448 static void convertForOp(scf::ForOp op, 449 llvm::DenseMap<Value, Value> &valueMapping) { 450 SmallVector<Value> newOperands; 451 SmallVector<std::pair<size_t, size_t>> argMapping; 452 for (const auto &operand : llvm::enumerate(op.getIterOperands())) { 453 auto it = valueMapping.find(operand.value()); 454 if (it == valueMapping.end()) 455 continue; 456 argMapping.push_back(std::make_pair( 457 operand.index(), op.getNumIterOperands() + newOperands.size())); 458 newOperands.push_back(it->second); 459 } 460 OpBuilder b(op); 461 scf::ForOp newForOp = replaceForOpWithNewSignature(b, op, newOperands); 462 Block &loopBody = *newForOp.getBody(); 463 for (auto mapping : argMapping) { 464 valueMapping[newForOp.getResult(mapping.first)] = 465 newForOp.getResult(mapping.second); 466 valueMapping[loopBody.getArgument(mapping.first + 467 newForOp.getNumInductionVars())] = 468 loopBody.getArgument(mapping.second + newForOp.getNumInductionVars()); 469 } 470 } 471 472 static void convertYieldOp(scf::YieldOp op, 473 llvm::DenseMap<Value, Value> &valueMapping) { 474 OpBuilder b(op); 475 auto loop = cast<scf::ForOp>(op->getParentOp()); 476 auto yieldOperands = llvm::to_vector<4>(op.getOperands()); 477 for (const auto &operand : llvm::enumerate(op.getOperands())) { 478 auto it = valueMapping.find(operand.value()); 479 if (it == valueMapping.end()) 480 continue; 481 // Replace the yield of old value with the for op argument to make it easier 482 // to remove the dead code. 483 yieldOperands[operand.index()] = loop.getIterOperands()[operand.index()]; 484 yieldOperands.push_back(it->second); 485 } 486 b.create<scf::YieldOp>(op.getLoc(), yieldOperands); 487 op.erase(); 488 } 489 490 /// Convert an elementwise op to the equivalent elementwise op on MMA matrix. 491 static void convertElementwiseOp(Operation *op, gpu::MMAElementwiseOp opType, 492 llvm::DenseMap<Value, Value> &valueMapping) { 493 OpBuilder b(op); 494 SmallVector<Value> matrixOperands; 495 for (Value operand : op->getOperands()) 496 matrixOperands.push_back(valueMapping.find(operand)->second); 497 Value newOp = b.create<gpu::SubgroupMmaElementwiseOp>( 498 op->getLoc(), matrixOperands[0].getType(), matrixOperands, opType); 499 valueMapping[op->getResult(0)] = newOp; 500 } 501 502 namespace mlir { 503 504 void populatePrepareVectorToMMAPatterns(RewritePatternSet &patterns) { 505 patterns.add<PrepareContractToGPUMMA, CombineTransferReadOpTranspose>( 506 patterns.getContext()); 507 } 508 509 void convertVectorToMMAOps(FuncOp funcOp) { 510 SetVector<Operation *> ops = getOpToConvert(funcOp); 511 llvm::DenseMap<Value, Value> valueMapping; 512 for (Operation *op : ops) { 513 if (auto transferRead = dyn_cast<vector::TransferReadOp>(op)) { 514 convertTransferReadOp(transferRead, valueMapping); 515 } else if (auto transferWrite = dyn_cast<vector::TransferWriteOp>(op)) { 516 convertTransferWriteOp(transferWrite, valueMapping); 517 } else if (auto contractOp = dyn_cast<vector::ContractionOp>(op)) { 518 convertContractOp(contractOp, valueMapping); 519 } else if (auto constantOp = dyn_cast<arith::ConstantOp>(op)) { 520 convertConstantOp(constantOp, valueMapping); 521 } else if (auto broadcastOp = dyn_cast<vector::BroadcastOp>(op)) { 522 convertBroadcastOp(broadcastOp, valueMapping); 523 } else if (auto forOp = dyn_cast<scf::ForOp>(op)) { 524 convertForOp(forOp, valueMapping); 525 } else if (auto yiledOp = dyn_cast<scf::YieldOp>(op)) { 526 convertYieldOp(yiledOp, valueMapping); 527 } else if (auto elementwiseType = convertElementwiseOpToMMA(op)) { 528 convertElementwiseOp(op, *elementwiseType, valueMapping); 529 } 530 } 531 } 532 533 } // namespace mlir 534 namespace { 535 536 struct ConvertVectorToGPUPass 537 : public ConvertVectorToGPUBase<ConvertVectorToGPUPass> { 538 void runOnFunction() override { 539 RewritePatternSet patterns(getFunction().getContext()); 540 populatePrepareVectorToMMAPatterns(patterns); 541 (void)applyPatternsAndFoldGreedily(getFunction(), std::move(patterns)); 542 543 convertVectorToMMAOps(getFunction()); 544 } 545 }; 546 547 } // namespace 548 549 std::unique_ptr<Pass> mlir::createConvertVectorToGPUPass() { 550 return std::make_unique<ConvertVectorToGPUPass>(); 551 } 552