| //===- Fusion.cpp - Implementation of linalg Fusion -----------------------===// |
| // |
| // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| // See https://llvm.org/LICENSE.txt for license information. |
| // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| // |
| //===----------------------------------------------------------------------===// |
| // |
| // This file implements the linalg dialect Fusion pass. |
| // |
| //===----------------------------------------------------------------------===// |
| |
| #include "PassDetail.h" |
| #include "mlir/Dialect/Affine/IR/AffineOps.h" |
| #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h" |
| #include "mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h" |
| #include "mlir/Dialect/Linalg/IR/LinalgOps.h" |
| #include "mlir/Dialect/Linalg/IR/LinalgTypes.h" |
| #include "mlir/Dialect/Linalg/Passes.h" |
| #include "mlir/Dialect/Linalg/Transforms/Transforms.h" |
| #include "mlir/Dialect/Linalg/Utils/Utils.h" |
| #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| #include "mlir/IR/AffineExpr.h" |
| #include "mlir/IR/AffineMap.h" |
| #include "mlir/IR/Dominance.h" |
| #include "mlir/Support/LLVM.h" |
| #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
| #include "mlir/Transforms/RegionUtils.h" |
| #include "llvm/ADT/MapVector.h" |
| #include "llvm/ADT/ScopeExit.h" |
| #include "llvm/Support/CommandLine.h" |
| #include "llvm/Support/Debug.h" |
| |
| #include <set> |
| |
| #define DEBUG_TYPE "linalg-fusion" |
| |
| using namespace mlir; |
| using namespace mlir::linalg; |
| |
| using llvm::dbgs; |
| |
| /// Implements a simple high-level fusion pass on linalg structured operations. |
| /// |
| /// In each block, linalg ops are processed in reverse textual order. |
| /// Given a linalg op `O`, fusion occurs by: |
| /// 1. inspecting the linalg ops that write into the views read by `O`. There |
| /// are 2 cases: |
| /// a) buffer case: use the SSA value of the views and a simple alias |
| /// analysis on subview ops to determine producer-consumer dependences; |
| /// b) tensor case: use SSA use-def chains on extract_slice ops; |
| /// 2. greedily fuse the linalg ops that produce the subview/extract_slice. |
| /// 3. inspect the fused ops and determine whether they have other remaining |
| /// LinalgOp uses. If not, then erase the original producing linalg op. |
| /// |
| /// More advanced use cases, analyses as well as profitability heuristics are |
| /// left for future work. |
| |
| struct ShapeDimension { |
| Value shape; |
| unsigned dimension; |
| }; |
| |
| // Given an `op`, returns the first (`shape`, `dimension`) pair that identifies |
| // the loop range at `loopDepth`. The semantics of the loopToOperandRangesMaps |
| // guarantees at least one such dimension is found. If multiple candidates exist |
| // they must agree by construction (i.e. have the same size) and we just return |
| // the first one. |
| static ShapeDimension |
| getShapeDefiningLoopRange(LinalgOp op, unsigned loopDepth, |
| bool fromSubViewOpOnly = false) { |
| // Iterate over the inputs and outputs in order. |
| // Extract the subranges from the linearized ranges. |
| for (OpOperand *opOperand : op.getInputAndOutputOperands()) { |
| // The method `getRangeFromOperandShape` requires using SubViewOp or |
| // ExtractSliceOps. If the value isn't defined from there continue. |
| // todo: The method should be adapted to get the values from |
| // `ViewInterface`. The interface needs a `getOrCreateRanges` method which |
| // currently returns a `linalg.range`. The fix here is to move this op to |
| // `std` dialect and add the method to `ViewInterface`. |
| if (fromSubViewOpOnly && |
| !isa_and_nonnull<memref::SubViewOp, tensor::ExtractSliceOp>( |
| opOperand->get().getDefiningOp())) |
| continue; |
| |
| AffineMap map = op.getTiedIndexingMap(opOperand); |
| LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange I/O idx: " |
| << opOperand->getOperandNumber() << "\n"); |
| LLVM_DEBUG(llvm::dbgs() |
| << "getShapeDefiningLoopRange map: " << map << "\n"); |
| SmallVector<Value, 8> shapeRanges(map.getNumResults(), nullptr); |
| for (auto en : llvm::enumerate(map.getResults())) { |
| auto dimExpr = en.value().dyn_cast<AffineDimExpr>(); |
| if (!dimExpr) |
| continue; |
| if (loopDepth == en.value().cast<AffineDimExpr>().getPosition()) { |
| LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange loopDepth: " |
| << loopDepth << "\n"); |
| LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange shape: " |
| << opOperand->get() << "\n"); |
| return ShapeDimension{opOperand->get(), |
| static_cast<unsigned>(en.index())}; |
| } |
| } |
| } |
| llvm_unreachable("Expect to be able to extract a shape defining loop range"); |
| } |
| |
| // Return tiled operands for the fused producer op. When fusing into |
| // `linalg.tiled_loop` one has to update `input` and `output` arguments of the |
| // loop correspondingly. |
| // Each input tensor of the producer op has to be added to `inputs` of the |
| // `tiled_loop` if it is not present there already. Each output tensor has to |
| // be added either to `inputs` or to `outputs` of `linalg.tiled_loop` depending |
| // on whether the correponding result is an input or an output to the loop. |
| // |
| // NOTE: This way of updating the arguments of the `tiled_loop` assumes that the |
| // intermediate result is not used by any other operation but the consumer. A |
| // more generic way is to append all missing output tensors of the producer to |
| // the tiled loop outputs and hence modify the number of the results, since we |
| // would need to add the intermediate results to `linalg.yield`. After that a |
| // canonicalization pass would move the unused output args of the `tiled_loop` |
| // to the `input` section. |
| static SmallVector<Value> getTiledOperands(OpBuilder &b, LinalgOp producer) { |
| auto tiledLoop = dyn_cast<TiledLoopOp>(b.getBlock()->getParentOp()); |
| if (!tiledLoop) |
| return producer.getInputAndOutputOperands(); |
| |
| SmallVector<Value> tiledOperands; |
| assert(producer.hasTensorSemantics() && |
| "only fusion on tensors is currently supported for TiledLinalgOp"); |
| |
| for (OpOperand *producerInput : producer.getInputOperands()) { |
| OpOperand *addedInput = tiledLoop.findInputOperand(producerInput->get()); |
| if (addedInput == nullptr) |
| addedInput = &tiledLoop.appendInputOperand(b, producerInput->get()); |
| BlockArgument addedBlockArg = tiledLoop.getTiedBlockArgument(*addedInput); |
| tiledOperands.push_back(addedBlockArg); |
| } |
| for (OpOperand *producerOutput : producer.getOutputOperands()) { |
| OpResult result = producer.getTiedOpResult(producerOutput); |
| OpOperand *resultInputOperand = tiledLoop.findInputOperand(result); |
| OpOperand *resultOutputOperand = tiledLoop.findOutputOperand(result); |
| assert((resultInputOperand != nullptr) ^ (resultOutputOperand != nullptr) && |
| "The result should be present in `input` or `output` args of " |
| "`tiled_loop"); |
| |
| bool isInput = resultInputOperand; |
| int opNumber = isInput ? resultInputOperand->getOperandNumber() |
| : resultOutputOperand->getOperandNumber(); |
| |
| OpOperand *addedOutput = tiledLoop.findOutputOperand(producerOutput->get()); |
| if (addedOutput == nullptr) |
| addedOutput = |
| isInput ? &tiledLoop.appendInputOperand(b, producerOutput->get()) |
| : &tiledLoop.appendOutputOperand(b, producerOutput->get()); |
| |
| OpOperand &resultOperand = tiledLoop->getOpOperand(opNumber); |
| auto addedBlockArg = tiledLoop.getTiedBlockArgument(*addedOutput); |
| auto resultOperandBlockArg = tiledLoop.getTiedBlockArgument(resultOperand); |
| resultOperandBlockArg.replaceAllUsesWith(addedBlockArg); |
| tiledLoop.eraseOperand(b, resultOperand); |
| tiledOperands.push_back(addedBlockArg); |
| } |
| return tiledOperands; |
| } |
| |
| /// Fuses the producer by cloning the `producer`. The `fusedLoopsAndRanges` |
| /// provides the loop range information for the fused loops. The rest are |
| /// obtained from the producer itself, since they are not tiled + fused. |
| static LinalgOp fuse(OpBuilder &b, LinalgOp producer, |
| const DenseMap<unsigned, Range> &fusedLoopsAndRanges) { |
| SmallVector<Value, 8> ivs, tileSizes, sizeBounds; |
| SmallVector<Range, 8> loopRanges; |
| Location loc = producer.getLoc(); |
| auto zero = b.create<arith::ConstantIndexOp>(loc, 0); |
| auto one = b.create<arith::ConstantIndexOp>(loc, 1); |
| |
| for (unsigned i = 0, e = producer.getNumLoops(); i < e; ++i) { |
| auto shapeDim = getShapeDefiningLoopRange(producer, i); |
| Value dim = createOrFoldDimOp(b, loc, shapeDim.shape, shapeDim.dimension); |
| sizeBounds.push_back(dim); |
| auto it = fusedLoopsAndRanges.find(i); |
| if (it != fusedLoopsAndRanges.end()) { |
| ivs.push_back(it->second.offset); |
| tileSizes.push_back(it->second.size); |
| loopRanges.push_back(it->second); |
| LLVM_DEBUG(llvm::dbgs() << "tiled loop#" << i << " with LoopRange " |
| << loopRanges.back() << "\n"); |
| } else { |
| tileSizes.push_back(zero); |
| loopRanges.push_back(Range{zero, dim, one}); |
| LLVM_DEBUG(llvm::dbgs() << "full loop#" << i << " with LoopRange " |
| << loopRanges.back() << "\n"); |
| } |
| } |
| |
| SmallVector<Value, 8> clonedShapes; |
| clonedShapes.reserve(producer.getNumInputsAndOutputs()); |
| |
| // Compute subranges for all tensor input/output operands. |
| clonedShapes.append(makeTiledShapes(b, loc, producer, |
| getTiledOperands(b, producer), ivs, |
| tileSizes, sizeBounds)); |
| |
| // Iterate over the results in order. |
| // Extract the subtensor type from the linearized range. |
| // Since we do not enforce any canonicalizations on the fly, this is always |
| // fully dynamic at construction time. |
| SmallVector<Type, 4> resultTypes; |
| resultTypes.reserve(producer->getNumResults()); |
| for (RankedTensorType t : producer.getOutputTensorTypes()) { |
| unsigned rank = t.getRank(); |
| SmallVector<int64_t, 4> staticOffsetsVector( |
| rank, ShapedType::kDynamicStrideOrOffset); |
| SmallVector<int64_t, 4> staticSizesVector(rank, ShapedType::kDynamicSize); |
| SmallVector<int64_t, 4> staticStridesVector( |
| rank, ShapedType::kDynamicStrideOrOffset); |
| resultTypes.push_back(tensor::ExtractSliceOp::inferResultType( |
| t.cast<RankedTensorType>(), staticOffsetsVector, staticSizesVector, |
| staticStridesVector)); |
| } |
| |
| Operation *clonedOp = producer.clone(b, loc, resultTypes, clonedShapes); |
| |
| // Shift all IndexOp results by the tile offset. |
| SmallVector<Value> allIvs; |
| transform(loopRanges, std::back_inserter(allIvs), |
| [](Range range) { return range.offset; }); |
| addTileLoopIvsToIndexOpResults(b, clonedOp, allIvs); |
| |
| return clonedOp; |
| } |
| |
| /// Get the loop range for a dimension `dim` based on the `shapedOperand`. It is |
| /// expected to be defined by a subview op or an extract_slice op. |
| static Range getRangeFromOperandShape(OpBuilder &b, Location loc, |
| Value shapedOperand, unsigned dim) { |
| Operation *shapeProducingOp = shapedOperand.getDefiningOp(); |
| if (auto subViewOp = dyn_cast<memref::SubViewOp>(shapeProducingOp)) |
| return subViewOp.getOrCreateRanges(b, loc)[dim]; |
| if (auto sliceOp = dyn_cast<tensor::ExtractSliceOp>(shapeProducingOp)) |
| return sliceOp.getOrCreateRanges(b, loc)[dim]; |
| llvm_unreachable("SubviewOp or ExtractSliceOp expected"); |
| } |
| |
| /// Fuses the producer into the loop immediately enclosing the consumer. |
| /// This is achieved by "recomputing" the producer at the time it |
| /// is needed just before the consumer. |
| static LinalgOp fuse(OpBuilder &b, LinalgOp producerOp, AffineMap producerMap, |
| OpOperand &consumerOpOperand) { |
| LLVM_DEBUG(llvm::dbgs() << "Producer map: " << producerMap << "\n"); |
| DenseMap<unsigned, Range> fusedLoopsAndRanges; |
| Value shapedOperand = consumerOpOperand.get(); |
| for (auto en : llvm::enumerate(producerMap.getResults())) { |
| unsigned posInProducerLoop = en.value().cast<AffineDimExpr>().getPosition(); |
| fusedLoopsAndRanges[posInProducerLoop] = getRangeFromOperandShape( |
| b, consumerOpOperand.getOwner()->getLoc(), shapedOperand, en.index()); |
| } |
| return fuse(b, producerOp, fusedLoopsAndRanges); |
| } |
| |
| // Encode structural fusion safety preconditions. |
| // Some of these will be lifted in the future with better analysis. |
| static bool isStructurallyFusableProducer(LinalgOp producer, Value consumedView, |
| LinalgOp consumer) { |
| assert(producer.hasBufferSemantics() && |
| "expected linalg op with buffer semantics"); |
| assert(consumer.hasBufferSemantics() && |
| "expected linalg op with buffer semantics"); |
| if (producer.getNumOutputs() != 1) { |
| LLVM_DEBUG(llvm::dbgs() << "\nNot structurally fusable (multi-output)"); |
| return false; |
| } |
| // Only fuse when the producer block dominates. |
| DominanceInfo dom(producer.getOperation()); |
| if (!dom.dominates(producer->getBlock(), consumer->getBlock())) { |
| LLVM_DEBUG( |
| llvm::dbgs() |
| << "\nNot structurally fusable (producer block does not dominate)"); |
| return false; |
| } |
| return true; |
| } |
| |
| bool mlir::linalg::isProducerLastWriteOfView(const LinalgDependenceGraph &graph, |
| LinalgOp consumer, |
| Value consumedView, |
| LinalgOp producer) { |
| assert(producer.hasBufferSemantics() && |
| "expected linalg op with buffer semantics"); |
| assert(consumer.hasBufferSemantics() && |
| "expected linalg op with buffer semantics"); |
| // Make some simple structural checks that alleviate the need for more |
| // complex analyses. |
| if (!isStructurallyFusableProducer(producer, consumedView, consumer)) { |
| LLVM_DEBUG(llvm::dbgs() << "\n***Not static last write due to structure:\t" |
| << *producer.getOperation()); |
| return false; |
| } |
| // Check for any interleaved write to consumedView. |
| if (!graph.findCoveringWrites(producer, consumer, consumedView).empty()) { |
| LLVM_DEBUG(llvm::dbgs() << "\n***Not fusable due to interleaved write:\t" |
| << *producer.getOperation()); |
| return false; |
| } |
| return true; |
| } |
| |
| bool mlir::linalg::isFusableInto(const LinalgDependenceGraph &graph, |
| LinalgOp consumer, Value consumedView, |
| LinalgOp producer) { |
| assert(producer.hasBufferSemantics() && |
| "expected linalg op with buffer semantics"); |
| assert(consumer.hasBufferSemantics() && |
| "expected linalg op with buffer semantics"); |
| if (!isProducerLastWriteOfView(graph, consumer, consumedView, producer)) |
| return false; |
| // Check for any fusion-preventing dependence to any shape read/written that |
| // would violate dependences. |
| if (!graph.findCoveringDependences(producer, consumer).empty()) { |
| LLVM_DEBUG(llvm::dbgs() |
| << "\n***Not fusable due to an interleaved dependence:\t" |
| << *producer.getOperation()); |
| return false; |
| } |
| return true; |
| } |
| |
| /// For `consumer` with buffer semantics, find the Linalg operation on buffers |
| /// that is the last writer of `consumerOpOperand`. For now the fusable |
| /// dependence is returned as an instance of the `dependenceGraph`. |
| static FailureOr<LinalgDependenceGraph::LinalgDependenceGraphElem> |
| findFusableProducer(OpOperand &consumerOpOperand, |
| const LinalgDependenceGraph &dependenceGraph) { |
| LLVM_DEBUG(llvm::dbgs() << "findFusableProducer for: " |
| << consumerOpOperand.get() << " @" |
| << consumerOpOperand.getOperandNumber() << " in " |
| << *consumerOpOperand.getOwner() << "\n"); |
| LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner()); |
| if (!consumerOp) |
| return failure(); |
| |
| // Only consider RAW and WAW atm. |
| for (auto depType : { |
| LinalgDependenceGraph::DependenceType::RAW, |
| LinalgDependenceGraph::DependenceType::WAW, |
| }) { |
| LLVM_DEBUG(llvm::dbgs() |
| << "Dependencies into: " << *consumerOp.getOperation() << "\n"); |
| for (auto dependence : llvm::make_filter_range( |
| dependenceGraph.getDependencesInto(consumerOp, depType), |
| [&](LinalgDependenceGraph::LinalgDependenceGraphElem elem) { |
| LLVM_DEBUG(llvm::dbgs() << "Inspect dependence btw: " |
| << elem.getIndexingValue() << " and " |
| << elem.getDependentValue() << "\n"); |
| Value v = elem.getIndexingValue(); |
| Optional<unsigned> operandNum = |
| elem.getIndexingOpViewOperandNum(); |
| return isa<LinalgOp>(elem.getDependentOp()) && |
| v == consumerOpOperand.get() && operandNum && |
| operandNum.getValue() == |
| consumerOpOperand.getOperandNumber(); |
| })) { |
| // Consumer consumes this view, `isStructurallyFusableProducer` also |
| // checks whether it is a strict subview of the producer view. |
| auto producer = cast<LinalgOp>(dependence.getDependentOp()); |
| LLVM_DEBUG(llvm::dbgs() |
| << "\n" |
| << LinalgDependenceGraph::getDependenceTypeStr(depType) |
| << "producer: " << *dependence.getDependentOp() |
| << " view: " << dependence.getDependentValue() << "\n"); |
| |
| // If the producer and consumer have tensor semantics, the only dependence |
| // between them is through a RAW dependence and they are fusable by |
| // construction. For buffer semantics need additional checks. |
| if (producer.hasBufferSemantics() && consumerOp.hasBufferSemantics() && |
| isFusableInto(dependenceGraph, consumerOp, consumerOpOperand.get(), |
| producer)) |
| return dependence; |
| if (producer.hasTensorSemantics() && consumerOp.hasTensorSemantics()) { |
| assert(dependence.dependenceType == |
| LinalgDependenceGraph::DependenceType::RAW); |
| return dependence; |
| } |
| } |
| } |
| return failure(); |
| } |
| |
| FailureOr<FusionInfo> |
| mlir::linalg::fuseProducerOfBuffer(OpBuilder &b, OpOperand &consumerOpOperand, |
| const LinalgDependenceGraph &graph) { |
| Optional<LinalgDependenceGraph::LinalgDependenceGraphElem> fusableDependence = |
| findFusableProducer(consumerOpOperand, graph); |
| if (!fusableDependence) |
| return failure(); |
| |
| LinalgOp producerOp = dyn_cast<LinalgOp>(fusableDependence->getDependentOp()); |
| if (!producerOp) |
| return failure(); |
| |
| // If producer is already in the same block as consumer, we are done. |
| if (consumerOpOperand.get().getParentBlock() == |
| fusableDependence->getDependentValue().getParentBlock()) |
| return failure(); |
| |
| Optional<AffineMap> producerMap = |
| fusableDependence->getDependentOpViewIndexingMap(); |
| if (!producerMap) |
| return failure(); |
| |
| // Must be a subview or an extract_slice to guarantee there are loops we can |
| // fuse into. |
| auto subView = consumerOpOperand.get().getDefiningOp<memref::SubViewOp>(); |
| if (!subView) { |
| LLVM_DEBUG(llvm::dbgs() << "\nNot fusable (not a subview)"); |
| return failure(); |
| } |
| |
| // Fuse `producer` just before `consumer`. |
| OpBuilder::InsertionGuard g(b); |
| b.setInsertionPoint(consumerOpOperand.getOwner()); |
| LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: " |
| << *consumerOpOperand.getOwner() << "\n"); |
| |
| auto fusedProducer = fuse(b, producerOp, *producerMap, consumerOpOperand); |
| return FusionInfo{producerOp, fusedProducer}; |
| } |
| |
| /// Walk back use-def chain through scf::For yields. |
| /// Sets `producer` and `outputIndex` if it finds a producer LinalgOp |
| |
| // TODO(ravishankarm, ntv): This can be moved into the dependence graphs |
| // dependence tracking since the dependence tracking is similar to what is done |
| // w.r.t to buffers. |
| static void getProducerOfTensor(Value tensor, OpResult &opResult) { |
| if (!tensor.getType().isa<RankedTensorType>()) |
| return; |
| |
| while (true) { |
| LLVM_DEBUG(llvm::dbgs() << "\ngetProducerOfTensor: " << tensor); |
| if (auto linalgOp = tensor.getDefiningOp<LinalgOp>()) { |
| opResult = tensor.cast<OpResult>(); |
| return; |
| } |
| if (auto sliceOp = tensor.getDefiningOp<tensor::ExtractSliceOp>()) { |
| tensor = sliceOp.source(); |
| continue; |
| } |
| if (auto blockArg = tensor.dyn_cast<BlockArgument>()) { |
| if (auto forOp = blockArg.getDefiningOp<scf::ForOp>()) { |
| tensor = *(forOp.getIterOperands().begin() + blockArg.getArgNumber()); |
| continue; |
| } |
| } |
| return; |
| } |
| } |
| |
| FailureOr<FusionInfo> |
| mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpOperand &consumerOpOperand) { |
| Value inputTensor = consumerOpOperand.get(); |
| OpResult producerOpResult; |
| getProducerOfTensor(inputTensor, producerOpResult); |
| if (!producerOpResult) { |
| LLVM_DEBUG(llvm::dbgs() << "\nUnable to find producer"); |
| return failure(); |
| } |
| return fuseProducerOfTensor(b, producerOpResult, consumerOpOperand); |
| } |
| |
| FailureOr<FusionInfo> |
| mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpResult producerOpResult, |
| OpOperand &consumerOpOperand) { |
| auto producerOp = dyn_cast<LinalgOp>(producerOpResult.getOwner()); |
| if (!producerOp) |
| return failure(); |
| |
| LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner()); |
| if (!consumerOp) |
| return failure(); |
| |
| Value inputTensor = consumerOpOperand.get(); |
| |
| // Must be an extract_slice op to guarantee there are loops we can fuse into. |
| auto sliceOp = inputTensor.getDefiningOp<tensor::ExtractSliceOp>(); |
| if (!sliceOp) { |
| LLVM_DEBUG(llvm::dbgs() |
| << "\nNot fusable, not an extract_slice op: " << inputTensor); |
| return failure(); |
| } |
| |
| // If producer is already in the same block as consumer, we are done. |
| if (consumerOpOperand.get().getParentBlock() == |
| producerOpResult.getParentBlock()) |
| return failure(); |
| |
| // Insert fused `producer` just before `consumer`. |
| OpBuilder::InsertionGuard g(b); |
| b.setInsertionPoint(consumerOp); |
| LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: " << *consumerOp << "\n"); |
| OpOperand *opOperand = |
| producerOp.getOutputOperand(producerOpResult.getResultNumber()); |
| LinalgOp fusedProducer = |
| fuse(b, producerOp, producerOp.getTiedIndexingMap(opOperand), |
| consumerOpOperand); |
| |
| // Replace use. |
| // Canonicalizations are not guaranteed to have happened before constructing |
| // `fusedProducer`. In the tensor case this can result in temporary type |
| // mismatches. Insert a `tensor.cast` op to propagate the transformation |
| // invariant that types are compatible. |
| Value def = fusedProducer->getResult(producerOpResult.getResultNumber()); |
| Type consumerType = consumerOpOperand.get().getType(); |
| if (consumerType != def.getType()) |
| def = b.create<tensor::CastOp>(fusedProducer.getLoc(), consumerType, def); |
| consumerOpOperand.set(def); |
| return FusionInfo{cast<LinalgOp>(producerOpResult.getOwner()), fusedProducer}; |
| } |
| |
| /// Prune all dimensions that are of reduction iterator type from `map`. |
| static AffineMap pruneReductionDimsFromMap(ArrayRef<Attribute> iteratorTypes, |
| AffineMap map) { |
| llvm::SmallDenseSet<unsigned> projectedDims; |
| for (auto attr : llvm::enumerate(iteratorTypes)) { |
| if (!isParallelIterator(attr.value())) |
| projectedDims.insert(attr.index()); |
| } |
| return getProjectedMap(map, projectedDims); |
| } |
| |
| /// Returns the mapping from iterations in the consumer that write to the same |
| /// location as the iterations in the producer. To do so use |
| /// - indexing map of the fused view in the consumer : consumerIndexMap |
| /// - indexing map of the fused view in the producer : producerIndexMap |
| /// consumerLoopToProducerLoop = |
| /// inverse(producerIndexMap).compose(consumerIndexMap) |
| static FailureOr<AffineMap> getConsumerLoopToProducerLoopMap( |
| LinalgDependenceGraph::LinalgDependenceGraphElem dependence) { |
| auto producer = dyn_cast<LinalgOp>(dependence.getDependentOp()); |
| if (!producer) |
| return failure(); |
| |
| Optional<AffineMap> producerIndexingMap = |
| dependence.getDependentOpViewIndexingMap(); |
| Optional<AffineMap> consumerIndexingMap = |
| dependence.getIndexingOpViewIndexingMap(); |
| if (!producerIndexingMap || !consumerIndexingMap) |
| return failure(); |
| |
| AffineMap prunedProducerIndexingMap = pruneReductionDimsFromMap( |
| producer.iterator_types().getValue(), *producerIndexingMap); |
| if (!prunedProducerIndexingMap.isPermutation()) |
| return failure(); |
| |
| if (consumerIndexingMap->getNumResults() != |
| prunedProducerIndexingMap.getNumResults()) |
| return failure(); |
| |
| LLVM_DEBUG({ |
| llvm::dbgs() << "\t producerMap : "; |
| producerIndexingMap->print(llvm::dbgs()); |
| llvm::dbgs() << " pruned : "; |
| prunedProducerIndexingMap.print(llvm::dbgs()); |
| llvm::dbgs() << "\n"; |
| llvm::dbgs() << "\t consumerMap : "; |
| consumerIndexingMap->print(llvm::dbgs()); |
| llvm::dbgs() << "\n"; |
| }); |
| |
| AffineMap invProducerIndexMap = inversePermutation(prunedProducerIndexingMap); |
| if (!invProducerIndexMap) |
| return failure(); |
| |
| return invProducerIndexMap.compose(*consumerIndexingMap); |
| } |
| |
| /// Given a projected permutation `map`, returns true if the map changes the |
| /// order in which the fused loop dimension appear. |
| static bool doesTransposeAccess(AffineMap map, |
| const std::set<unsigned> &fusableLoops) { |
| Optional<unsigned> lastFusableLoop; |
| for (unsigned pos : llvm::map_range(map.getResults(), [](AffineExpr expr) { |
| return expr.cast<AffineDimExpr>().getPosition(); |
| })) { |
| if (!fusableLoops.count(pos)) |
| continue; |
| if (!lastFusableLoop) { |
| lastFusableLoop = pos; |
| continue; |
| } |
| if (pos <= lastFusableLoop.getValue()) |
| return true; |
| lastFusableLoop = pos; |
| } |
| return false; |
| } |
| |
| /// Returns the positions of the loop in `op` that can be tiled based on the |
| /// operations that are to be fused with it. For example, in a |
| /// |
| /// linalg.matmul ins(%a, %b : ...) outs(%c : ...) |
| /// |
| /// if the producer of %a needs to be fused with this op, only the `i` loop of |
| /// the matmul can be tiled while fusing. If producer of %a, and %b are to be |
| /// fused, then no loops can be tiled while fusing. The conditions used are: |
| /// 1. Only parallel loops can be used for tile + fuse. Find the number of |
| /// common outer parallel loops between the op and its producers being fused. |
| /// 2. Of the parallel loops only some can be fused. Only those loops can be |
| /// fused such where the fusable loops iteration space only touches one tile |
| /// of the fused operation. This is because the producer (which is writing |
| /// the fused subview) has update semantics. |
| /// |
| /// Since an inverse computation is needed, we need to consider the projection |
| /// of the producerIndexMap w.r.t the parallel loops. The actual fusable loops |
| /// are the dimensions of the consumerLoopToProducerLoop map that correspond to |
| /// parallel loops and appear in the result of the map |
| /// |
| /// Example 1: |
| /// linalg.fill(%cst, %c) |
| /// linalg.matmul ins(%a, %b) outs(%c) |
| /// Number of parallel loops : 2 |
| /// producerIndexMap = affine_map<(i, j) ->(i , j)> |
| /// consumerIndexMap = affine_map<(i, j, k) -> (i, j)> |
| /// consumerLoopToProducerLoop = affine_map<(i, j, k) -> (i, j)> |
| /// Fused dimensions : i, j |
| /// |
| /// Example 2: |
| /// linalg.matmul ins(%a, %b) outs(%c) |
| /// linalg.generic {indexing_maps = [affine_map<(i, j) -> (j, i)>, ... |
| /// iterator_types = ["parallel", "parallel"]} |
| /// ins(%c) ... |
| /// |
| /// Number of parallel loops = 2: |
| /// producerIndexMap (projected to parallel loops) = |
| /// affine_map<(i, j) -> (i, j)> |
| /// consumerLoopToProducerLoop2 = affine_map<(i, j) -> (j, i)> |
| /// Fused dimensions : i, j |
| /// |
| /// Example 3: |
| /// linalg.copy(%s, %b) |
| /// linalg.matmul ins(%a, %b) outs(%c) |
| /// |
| /// Number of parallel loops = 2 |
| /// produceIndexMap : affine_map<(i, j) -> (i, j)> |
| /// consumerLoopToProduceLoops = affine_map<(i, j, k) -> (k, j)> |
| /// submap with only parallel loops = affine_map<(i, j) -> (j)> |
| /// Fused dimensions : j |
| static std::set<unsigned> |
| collectFusableLoops(ArrayRef<LinalgOp> ops, |
| const FusableOpDependencesTy &fusableDependences) { |
| assert(!ops.empty()); |
| auto getNumOuterParallelLoops = [](LinalgOp linalgOp) { |
| return linalgOp.iterator_types() |
| .getValue() |
| .take_while([](Attribute attr) -> bool { |
| return attr.cast<StringAttr>().getValue() == |
| getParallelIteratorTypeName(); |
| }) |
| .size(); |
| }; |
| |
| size_t numOuterParallelLoops = getNumOuterParallelLoops(ops.back()); |
| for (auto op : ops.drop_back()) { |
| numOuterParallelLoops = |
| std::min(numOuterParallelLoops, getNumOuterParallelLoops(op)); |
| } |
| |
| std::set<unsigned> fusableLoops; |
| auto range = llvm::seq<unsigned>(0, numOuterParallelLoops); |
| fusableLoops.insert(range.begin(), range.end()); |
| |
| for (auto op : reverse(ops)) { |
| for (auto dependence : fusableDependences.lookup(op)) { |
| LLVM_DEBUG({ |
| llvm::dbgs() << "\t fusable :"; |
| for (unsigned i : fusableLoops) |
| llvm::dbgs() << " " << i; |
| llvm::dbgs() << "\n"; |
| }); |
| |
| Optional<AffineMap> consumerLoopToProducerLoop = |
| getConsumerLoopToProducerLoopMap(dependence); |
| if (!consumerLoopToProducerLoop) { |
| op.emitRemark("failed to get map from consumer loop to producer loop"); |
| return {}; |
| } |
| // todo: This condition is only an implementation limitation. When fusing |
| // the operation, if the accesses in the producer/consumer are transposes |
| // of each other, the loop bounds for the tiled producer can be |
| // manipulated accordingly. This requires some additional bookkeeping in |
| // the implementation of tile+fuse that is deferred to later. |
| if (doesTransposeAccess(*consumerLoopToProducerLoop, fusableLoops)) { |
| op.emitRemark("unhandled fusion when fusion requires permutation"); |
| return {}; |
| } |
| |
| std::set<unsigned> candidates; |
| for (AffineExpr expr : consumerLoopToProducerLoop->getResults()) { |
| unsigned position = expr.cast<AffineDimExpr>().getPosition(); |
| if (fusableLoops.count(position)) |
| candidates.insert(position); |
| } |
| LLVM_DEBUG({ |
| llvm::dbgs() << "\t candidates :"; |
| for (unsigned i : candidates) |
| llvm::dbgs() << " " << i; |
| llvm::dbgs() << "\n"; |
| }); |
| if (candidates.empty()) |
| return {}; |
| std::swap(candidates, fusableLoops); |
| } |
| } |
| |
| return fusableLoops; |
| } |
| |
| /// Find all dependences that are fusable. |
| FusableOpDependencesTy mlir::linalg::findAllFusableDependences( |
| ArrayRef<LinalgOp> ops, const LinalgDependenceGraph &dependenceGraph) { |
| FusableOpDependencesTy fusableDependences; |
| DenseMap<Operation *, SmallVector<AffineMap, 1>> fusedProducerIndexingMap; |
| for (LinalgOp op : reverse(ops)) { |
| for (OpOperand *opOperand : op.getInputAndOutputOperands()) { |
| Optional<LinalgDependenceGraph::LinalgDependenceGraphElem> |
| fusableDependence = findFusableProducer(*opOperand, dependenceGraph); |
| if (!fusableDependence) |
| continue; |
| LinalgOp producerOp = |
| dyn_cast<LinalgOp>(fusableDependence->getDependentOp()); |
| if (!producerOp) |
| continue; |
| // Do not fuse dependences that are to operations not in the same basic |
| // block. This avoid moving fused operations across loops that might |
| // themselves carry dependency making the fusion illegal. |
| if (producerOp->getBlock() != op->getBlock()) |
| continue; |
| |
| // Make sure that the indexing map of the view used for fusion in the |
| // producer is a projected permutation. |
| Optional<AffineMap> producerMap = |
| fusableDependence->getDependentOpViewIndexingMap(); |
| Optional<AffineMap> consumerMap = |
| fusableDependence->getIndexingOpViewIndexingMap(); |
| assert( |
| consumerMap && |
| "unable to find indexing map of operand/result of indexing OpView"); |
| fusedProducerIndexingMap[producerOp.getOperation()].push_back( |
| *consumerMap); |
| if (!producerMap || !producerMap->isProjectedPermutation() || |
| !consumerMap->isProjectedPermutation()) |
| continue; |
| |
| fusableDependences[producerOp.getOperation()].push_back( |
| *fusableDependence); |
| } |
| } |
| // TODO: Currently fusion would not be legal if the fusable dependence is to |
| // the same producer but different indexing map in the consumer. Fix this, but |
| // in the meanwhile disallow such a fusion. |
| for (auto useIndexingMapsList : fusedProducerIndexingMap) { |
| AffineMap map1 = useIndexingMapsList.second.front(); |
| for (AffineMap map2 : |
| ArrayRef<AffineMap>(useIndexingMapsList.second).drop_front()) { |
| if (map1 != map2) { |
| fusableDependences.erase(useIndexingMapsList.first); |
| break; |
| } |
| } |
| } |
| return fusableDependences; |
| } |
| |
| /// Tile the fused loops in the root operation, by setting the tile sizes for |
| /// all other loops to zero (those will be tiled later). |
| static FailureOr<TiledLinalgOp> |
| tileRootOperation(OpBuilder &b, LinalgOp op, ArrayRef<Value> tileSizeVector, |
| const LinalgTilingOptions &options, |
| const std::set<unsigned> &fusedLoops) { |
| SmallVector<Value, 4> tileSizes(tileSizeVector.begin(), tileSizeVector.end()); |
| auto zero = b.create<arith::ConstantIndexOp>(op.getLoc(), 0); |
| for (unsigned i = 0, e = tileSizes.size(); i != e; ++i) |
| if (!fusedLoops.count(i)) |
| tileSizes[i] = zero; |
| LinalgTilingOptions tileFusedLoopsOptions = options; |
| tileFusedLoopsOptions.setTileSizes(tileSizes); |
| return tileLinalgOp(b, op, tileFusedLoopsOptions); |
| } |
| |
| /// Fuse the operations in `fusionCandidates` with `tiledOp`. Latter is expected |
| /// to be a tiled operation such that it is valid to fuse all operations in |
| /// `fusionCandidates`, i.e. move the operation within the inter-tile loops of |
| /// `tiledOp`. |
| static SmallVector<LinalgOp, 1> |
| fuseOperations(OpBuilder &b, LinalgOp rootOp, TiledLinalgOp tiledLinalgOp, |
| ArrayRef<LinalgOp> fusionCandidates, |
| const FusableOpDependencesTy &fusableDependences, |
| const std::set<unsigned> &fusedLoops) { |
| LinalgOp tiledOp = tiledLinalgOp.op; |
| OpBuilder::InsertionGuard guard(b); |
| b.setInsertionPoint(tiledOp); |
| |
| DenseMap<unsigned, Range> fusedLoopsAndRanges; |
| for (unsigned loop : fusedLoops) { |
| ShapeDimension shapeDim = getShapeDefiningLoopRange(tiledOp, loop, true); |
| fusedLoopsAndRanges[loop] = getRangeFromOperandShape( |
| b, tiledOp.getLoc(), shapeDim.shape, shapeDim.dimension); |
| } |
| |
| SmallVector<LinalgOp, 1> fusedOps(fusionCandidates.size()); |
| DenseMap<Operation *, LinalgOp> origOpToFusedOp; |
| origOpToFusedOp[rootOp.getOperation()] = tiledOp; |
| for (auto candidate : enumerate(llvm::reverse(fusionCandidates))) { |
| LinalgOp origOp = candidate.value(); |
| LinalgOp fusedOp = fuse(b, origOp, fusedLoopsAndRanges); |
| origOpToFusedOp[origOp.getOperation()] = fusedOp; |
| fusedOps[fusionCandidates.size() - candidate.index() - 1] = fusedOp; |
| |
| // Prepare the builder for the next insertion point. |
| auto guard = llvm::make_scope_exit([&]() { b.setInsertionPoint(fusedOp); }); |
| if (!origOp.hasTensorSemantics()) |
| continue; |
| |
| // If the producer consumer operations are linalg operations on tensors, the |
| // dependence is due to value produced (as a return tensor) by the producer |
| // and used in the consumer. The returned value of the fused op needs to be |
| // made the operand of the tiled/fused consumer operation. By construction |
| // the value returned by the producer is the value used by the consumer. |
| for (auto &dependence : fusableDependences.lookup(origOp.getOperation())) { |
| if (dependence.dependenceType != |
| LinalgDependenceGraph::DependenceType::RAW) |
| continue; |
| |
| unsigned resultIndex = |
| dependence.getDependentOpViewResultNum().getValue(); |
| LinalgOp consumer = origOpToFusedOp.lookup(dependence.getIndexingOp()); |
| if (!consumer) |
| continue; |
| |
| Value replacementValue = fusedOp.getOperation()->getResult(resultIndex); |
| consumer.getOperation()->setOperand( |
| dependence.getIndexingOpViewOperandNum().getValue(), |
| replacementValue); |
| } |
| |
| // At this point, all Linalg uses of the tensors produced by `origOp` have |
| // been replaced. However, there may still be "output tensor"-like uses |
| // coming from WAW dependencies. |
| // All these uses are iter_args of the outermost loop (TODO: add a check). |
| // Such iter_args uses serve 2 purposes: |
| // 1. give a shape to the output |
| // 2. encode destructive updates that may be inplaceable by bufferization. |
| // To keep the second type of information while letting the unfused op die |
| // unused, we need to forward the producer output operand. |
| if (auto forOp = dyn_cast<scf::ForOp>(tiledLinalgOp.loops.front())) { |
| for (auto &operand : forOp.getIterOpOperands()) { |
| if (auto opResult = operand.get().dyn_cast<OpResult>()) { |
| if (opResult.getOwner() == origOp) { |
| Value output = |
| origOp.getOutputOperand(opResult.getResultNumber())->get(); |
| assert(output.getType().isa<RankedTensorType>()); |
| operand.set(output); |
| } |
| } |
| } |
| } |
| } |
| return fusedOps; |
| } |
| |
| static FailureOr<TiledAndFusedLinalgOps> |
| tileAndFuseLinalgOpsImpl(OpBuilder &b, ArrayRef<LinalgOp> ops, |
| const LinalgDependenceGraph &dependenceGraph, |
| const LinalgTilingOptions &tilingOptions) { |
| if (ops.size() < 2) |
| return failure(); |
| LinalgOp rootOp = ops.back(); |
| if (!llvm::all_of( |
| ops, |
| [](LinalgOp linalgOp) { return linalgOp.hasBufferSemantics(); }) && |
| !llvm::all_of(ops, [](LinalgOp linalgOp) { |
| return linalgOp.hasTensorSemantics(); |
| })) { |
| rootOp.emitError( |
| "unable to fuse operations that have tensor semantics with operations " |
| "that have buffer semantics and viceversa."); |
| return failure(); |
| } |
| // TODO: Support interchange with tile + fuse. This might actually help do |
| // better fusion. |
| if (!tilingOptions.interchangeVector.empty()) { |
| rootOp.emitRemark("unable to handle tile and fuse with interchange"); |
| return failure(); |
| } |
| |
| OpBuilder::InsertionGuard guard(b); |
| b.setInsertionPoint(rootOp); |
| |
| // Find all the producers. |
| LLVM_DEBUG(llvm::dbgs() << "findAllFusableDependences\n"); |
| FusableOpDependencesTy fusableDependences = |
| findAllFusableDependences(ops, dependenceGraph); |
| if (fusableDependences.empty()) { |
| LLVM_DEBUG(llvm::dbgs() << "no fusable dependencies found\n"); |
| return failure(); |
| } |
| |
| TiledAndFusedLinalgOps ret; |
| // Find the loops that can be tiled and fused. |
| LLVM_DEBUG(llvm::dbgs() << "collectFusableLoops\n"); |
| ret.fusedLoopDims = collectFusableLoops(ops, fusableDependences); |
| |
| // If there are no fusable dependences or there are no tile+fusable loops, |
| // just return. |
| if (ret.fusedLoopDims.empty()) { |
| LLVM_DEBUG(llvm::dbgs() << "no fusable loops found\n"); |
| return failure(); |
| } |
| |
| // Tile the fused loops in the last operation in the list. |
| SmallVector<Value, 4> tileSizeVector = |
| tilingOptions.tileSizeComputationFunction(b, rootOp); |
| FailureOr<TiledLinalgOp> tiledRootOp = tileRootOperation( |
| b, rootOp, tileSizeVector, tilingOptions, ret.fusedLoopDims); |
| if (failed(tiledRootOp)) { |
| rootOp.emitRemark("failed to tile the fused loops"); |
| return failure(); |
| } |
| ret.op = tiledRootOp->op; |
| ret.fusedLoops.assign(tiledRootOp->loops.begin(), tiledRootOp->loops.end()); |
| |
| // Fuse the other operations into the fused inter-tile loops produced above. |
| ret.fusedProducers = fuseOperations(b, rootOp, *tiledRootOp, ops.drop_back(), |
| fusableDependences, ret.fusedLoopDims); |
| |
| return ret; |
| } |
| |
| FailureOr<TiledAndFusedLinalgOps> |
| mlir::linalg::tileAndFuseLinalgOps(OpBuilder &b, ArrayRef<LinalgOp> ops, |
| const LinalgDependenceGraph &dependenceGraph, |
| const LinalgTilingOptions &tilingOptions) { |
| switch (tilingOptions.loopType) { |
| case LinalgTilingLoopType::Loops: |
| case LinalgTilingLoopType::ParallelLoops: |
| case LinalgTilingLoopType::TiledLoops: |
| return tileAndFuseLinalgOpsImpl(b, ops, dependenceGraph, tilingOptions); |
| default:; |
| } |
| return failure(); |
| } |