//===- 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/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/StandardOps/EDSC/Intrinsics.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 "llvm/ADT/MapVector.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include #define DEBUG_TYPE "linalg-fusion" using namespace mlir; using namespace mlir::edsc; using namespace mlir::edsc::intrinsics; 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 subtensor ops; /// 2. greedily fuse the linalg ops that produce the subview/subtensor. /// 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. // Fill `offset`, `sizes` and `strides` used to iterate over the shape indexed // by `permutationMap`. static void inferShapeComponents(AffineMap permutationMap, ArrayRef loopRanges, SmallVectorImpl &offsets, SmallVectorImpl &sizes, SmallVectorImpl &strides) { assert(permutationMap.isProjectedPermutation() && "expected some subset of a permutation map"); SmallVector shapeRanges(permutationMap.getNumResults()); unsigned idx = 0; for (AffineExpr e : permutationMap.getResults()) { // loopToOperandRangesMaps are permutations-only, just swap indices. unsigned loopPos = e.cast().getPosition(); shapeRanges[idx++] = loopRanges[loopPos]; } // Construct a new subshape for the tile. unsigned rank = shapeRanges.size(); offsets.reserve(rank); sizes.reserve(rank); strides.reserve(rank); for (auto r : shapeRanges) { offsets.push_back(r.offset); sizes.push_back(r.size); strides.push_back(r.stride); } } // Return a cloned version of `op` that operates on `loopRanges`, assumed to be // a subset of the original loop ranges of `op`. // This is achieved by applying the `loopToOperandRangesMaps` permutation maps // to the `loopRanges` in order to obtain view ranges. static LinalgOp cloneWithLoopRanges(OpBuilder &b, Location loc, LinalgOp op, ArrayRef loopRanges) { SmallVector clonedShapes; clonedShapes.reserve(op.getNumShapedOperands()); // Iterate over the shape operands in order. // Extract the subranges from the linearized ranges. for (auto en : llvm::enumerate(op.getShapedOperands())) { unsigned shapedOperandIdx = en.index(); AffineMap map = op.getIndexingMap(shapedOperandIdx); LLVM_DEBUG(llvm::dbgs() << "shapedOperandIdx: " << shapedOperandIdx << " with indexingMap: " << map << "\n"); SmallVector offsets, sizes, strides; inferShapeComponents(map, loopRanges, offsets, sizes, strides); Value shape = en.value(); Value sub = shape.getType().isa() ? b.create(loc, shape, offsets, sizes, strides) .getResult() : b.create(loc, shape, offsets, sizes, strides) .getResult(); clonedShapes.push_back(sub); } // Append the other operands. auto operands = op.getAssumedNonShapedOperands(); clonedShapes.append(operands.begin(), operands.end()); // 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 resultTypes; resultTypes.reserve(op->getNumResults()); for (RankedTensorType t : op.getOutputTensorTypes()) { unsigned rank = t.getRank(); SmallVector staticOffsetsVector( rank, ShapedType::kDynamicStrideOrOffset); SmallVector staticSizesVector(rank, ShapedType::kDynamicSize); SmallVector staticStridesVector( rank, ShapedType::kDynamicStrideOrOffset); resultTypes.push_back(SubTensorOp::inferResultType( t.cast(), staticOffsetsVector, staticSizesVector, staticStridesVector)); } Operation *clonedOp = op.clone(b, loc, resultTypes, clonedShapes); // When the producer is an IndexedGenericOp, we have to transform its block // IV arguments according to the tiling of the consumer, i.e. offset them by // the values computed in `loopRanges`. if (auto indexedGenericOp = dyn_cast(clonedOp)) { auto &block = indexedGenericOp.region().front(); OpBuilder::InsertionGuard g(b); b.setInsertionPointToStart(&block); for (unsigned i = 0, e = indexedGenericOp.getNumLoops(); i < e; ++i) { Value oldIndex = block.getArgument(i); // TODO: replace by an affine_apply. AddIOp newIndex = b.create(indexedGenericOp.getLoc(), oldIndex, loopRanges[i].offset); oldIndex.replaceAllUsesExcept(newIndex, SmallPtrSet{newIndex}); } } return clonedOp; } 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) { auto maps = op.indexing_maps(); // Iterate over the inputs and outputs in order. // Extract the subranges from the linearized ranges. SmallVector ios(op.getInputsAndOutputBuffers()); for (auto en : llvm::enumerate(ios)) { // The method `getRangeFromOperandShape` requires using SubViewOp or // SubTensorOps. If the value isnt 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(en.value().getDefiningOp())) continue; unsigned idx = en.index(); auto map = maps[idx].cast().getValue(); LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange I/O idx: " << idx << "\n"); LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange map: " << map << "\n"); Value shape = en.value(); SmallVector shapeRanges(map.getNumResults(), nullptr); for (auto en2 : llvm::enumerate(map.getResults())) { auto dimExpr = en2.value().dyn_cast(); if (!dimExpr) continue; if (loopDepth == en2.value().cast().getPosition()) { LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange loopDepth: " << loopDepth << "\n"); LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange shape: " << shape << "\n"); return ShapeDimension{shape, static_cast(en2.index())}; } } } llvm_unreachable("Expect to be able to extract a shape defining loop range"); } /// Fuse 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 &fusedLoopsAndRanges) { unsigned nPar = producer.getNumParallelLoops(); unsigned nRed = producer.getNumReductionLoops(); unsigned nWin = producer.getNumWindowLoops(); SmallVector loopRanges(nPar + nRed + nWin); for (auto fusedLoops : fusedLoopsAndRanges) loopRanges[fusedLoops.first] = fusedLoops.second; // Iterate over all dimensions. For the dimensions not identified by the // producer map for `producerIdx`, we need to explicitly compute the shape // that defines the loop ranges using the `producer`. for (unsigned i = 0, nLoops = loopRanges.size(); i < nLoops; ++i) { if (loopRanges[i].offset) LLVM_DEBUG(llvm::dbgs() << "existing LoopRange: " << loopRanges[i] << "\n"); else { auto shapeDim = getShapeDefiningLoopRange(producer, i); loopRanges[i] = Range{std_constant_index(0), std_dim(shapeDim.shape, shapeDim.dimension), std_constant_index(1)}; LLVM_DEBUG(llvm::dbgs() << "new LoopRange: " << loopRanges[i] << "\n"); } } return cloneWithLoopRanges(b, producer.getLoc(), producer, loopRanges); } /// Get the loop range for a dimension `dim` based on the `shapedOperand`. It is /// expected to be defined by a subview op or a subtensor op. static Range getRangeFromOperandShape(OpBuilder &b, Location loc, Value shapedOperand, unsigned dim) { Operation *shapeProducingOp = shapedOperand.getDefiningOp(); if (auto subViewOp = dyn_cast(shapeProducingOp)) return subViewOp.getOrCreateRanges(b, loc)[dim]; if (auto subTensorOp = dyn_cast(shapeProducingOp)) return subTensorOp.getOrCreateRanges(b, loc)[dim]; llvm_unreachable("SubviewOp or SubTensorOp expected"); } /// Fuses the producer of `producerIdx` into the loop immediately enclosing /// `consumer`. This is achieved by "recomputing" the `producer` at the time it /// is needed just before the `consumer. /// /// Depending on the type of `consumer.getShapedOperand(consumerIdx)`, there are /// 2 cases: /// 1. Buffer case: `producerIdx` is the index of the buffer in /// `producer.getOutputBuffers()`. /// 2. Tensor case: `producerIdx` is the index of the tensor in /// `producer.getResults()`. static LinalgOp fuse(OpBuilder &b, LinalgOp producer, unsigned producerIdx, LinalgOp consumer, unsigned consumerIdx) { AffineMap producerMap = producer.getOutputIndexingMap(producerIdx); LLVM_DEBUG(llvm::dbgs() << "Producer Idx: " << producerIdx << ", producer map: " << producerMap << "\n"); DenseMap fusedLoopsAndRanges; Location loc = consumer.getLoc(); Value shapedOperand = consumer.getShapedOperand(consumerIdx); for (auto en : llvm::enumerate(producerMap.getResults())) { unsigned posInProducerLoop = en.value().cast().getPosition(); fusedLoopsAndRanges[posInProducerLoop] = getRangeFromOperandShape(b, loc, shapedOperand, en.index()); } return fuse(b, producer, 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; } if (auto convOp = dyn_cast(producer.getOperation())) { // TODO: add a level of indirection to linalg.generic. if (convOp.padding()) return false; } if (auto convOp = dyn_cast(consumer.getOperation())) { // TODO: add a level of indirection to linalg.generic. if (convOp.padding()) return false; } return true; } static bool isSameSubView(Value a, Value b) { if (a == b) return true; auto sva = a.getDefiningOp(); auto svb = b.getDefiningOp(); if (!sva || !svb) return false; if (!isSameSubView(sva.getViewSource(), svb.getViewSource())) return false; if (sva.getType() != svb.getType()) return false; if (sva.getNumOperands() != svb.getNumOperands()) return false; if (sva.static_offsets() != svb.static_offsets()) return false; if (sva.static_sizes() != svb.static_sizes()) return false; if (sva.static_strides() != svb.static_strides()) return false; /// Skip the "source" operand. for (unsigned idx = 1, e = sva.getNumOperands(); idx != e; ++idx) if (sva.getOperand(idx) != svb.getOperand(idx)) return false; return true; } static Optional findFusableProducer(LinalgOp consumer, unsigned consumerIdx, const LinalgDependenceGraph &dependenceGraph) { // Only consider RAW and WAW atm. for (auto depType : { LinalgDependenceGraph::DependenceType::RAW, LinalgDependenceGraph::DependenceType::WAW, }) { for (auto dependence : llvm::make_filter_range( dependenceGraph.getDependencesInto(consumer, depType), [consumerIdx]( LinalgDependenceGraph::LinalgDependenceGraphElem elem) { return elem.indexingOpView.operandIndex == consumerIdx; })) { auto producer = cast(dependence.dependentOpView.op); // Check that the dependence is indeed on the input `consumerIdx` view. auto consumedView = consumer.getBuffer(dependence.indexingOpView.operandIndex); if (!isSameSubView(consumer.getBuffer(consumerIdx), consumedView)) continue; // Consumer consumes this view, `isStructurallyFusableProducer` also // checks whether it is a strict subview of the producer view. auto producedView = producer.getBuffer(dependence.dependentOpView.operandIndex); LLVM_DEBUG(llvm::dbgs() << "\n" << LinalgDependenceGraph::getDependenceTypeStr(depType) << "producer: " << *producer.getOperation() << " view: " << producedView << " output index: " << dependence.dependentOpView.operandIndex - producer.getNumInputs() << "\n"); (void)producedView; // Simple fusability checks. if (!isFusableInto(dependenceGraph, consumer, consumedView, producer)) continue; return dependence; } } return {}; } Optional mlir::linalg::fuseProducerOfBuffer(OpBuilder &b, LinalgOp consumer, unsigned consumerIdx, const LinalgDependenceGraph &graph) { Optional fusableDependence = findFusableProducer(consumer, consumerIdx, graph); if (!fusableDependence) return {}; LinalgOp producerOp = cast(fusableDependence->dependentOpView.op); // If producer is already in the same block as consumer, we are done. if (consumer->getBlock() == producerOp->getBlock()) return {}; unsigned producerIdx = fusableDependence->dependentOpView.operandIndex - producerOp.getNumInputs(); Value consumerView = consumer.getShapedOperand(consumerIdx); // Must be a subview or a slice to guarantee there are loops we can fuse // into. auto subView = consumerView.getDefiningOp(); auto slice = consumerView.getDefiningOp(); if (!subView && !slice) { LLVM_DEBUG(llvm::dbgs() << "\nNot fusable (not a subview or slice)"); return {}; } // Fuse `producer` just before `consumer`. OpBuilder::InsertionGuard g(b); b.setInsertionPoint(consumer.getOperation()); ScopedContext scope(b, consumer.getLoc()); LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: " << *consumer << "\n"); auto fusedProducer = fuse(b, producerOp, producerIdx, consumer, consumerIdx); return FusionInfo{producerOp, fusedProducer}; } /// Walk back use-def chain through scf::For yields. /// Sets `producer` and `outputIndex` if it finds a producer LinalgOp static void getProducerOfTensor(Value tensor, LinalgOp &producer, unsigned &outputIndex) { if (!tensor.getType().isa()) return; while (true) { if (auto linalgOp = tensor.getDefiningOp()) { producer = linalgOp; outputIndex = tensor.cast().getResultNumber(); return; } if (auto subTensorOp = tensor.getDefiningOp()) { tensor = subTensorOp.source(); continue; } if (auto blockArg = tensor.dyn_cast()) { if (auto forOp = blockArg.getDefiningOp()) { tensor = forOp.getResult(blockArg.getArgNumber()); continue; } } return; } } Optional mlir::linalg::fuseProducerOfTensor(OpBuilder &b, LinalgOp consumer, unsigned consumerIdx) { Value inputTensor = consumer.getInput(consumerIdx); LinalgOp producerOp; unsigned producerIdx; getProducerOfTensor(inputTensor, producerOp, producerIdx); // Must be a subtensor to guarantee there are loops we can fuse into. auto subTensor = inputTensor.getDefiningOp(); if (!subTensor || !producerOp) { LLVM_DEBUG(llvm::dbgs() << "\nNot fusable (not a subtensor)"); return {}; } // If producer is already in the same block as consumer, we are done. if (consumer->getBlock() == producerOp->getBlock()) return {}; // Insert fused `producer` just before `consumer`. OpBuilder::InsertionGuard g(b); b.setInsertionPoint(consumer.getOperation()); ScopedContext scope(b, consumer.getLoc()); LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: " << *consumer << "\n"); LinalgOp fusedProducer = fuse(b, producerOp, producerIdx, consumer, consumerIdx); // 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(producerIdx); OpOperand &use = consumer->getOpOperand(consumerIdx); Type consumerType = use.get().getType(); if (consumerType != def.getType()) def = b.create(fusedProducer.getLoc(), consumerType, def); use.set(def); return FusionInfo{producerOp, fusedProducer}; } /// Prune all dimensions that are of reduction iterator type from `map`. static AffineMap pruneReductionDimsFromMap(ArrayRef iteratorTypes, AffineMap map) { SmallVector projectedDims; for (auto attr : llvm::enumerate(iteratorTypes)) { if (!isParallelIterator(attr.value())) projectedDims.push_back(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 Optional getConsumerLoopToProducerLoopMap( LinalgDependenceGraph::LinalgDependenceGraphElem dependence) { auto producer = cast(dependence.dependentOpView.op); AffineMap producerIndexingMap = producer.getIndexingMap(dependence.dependentOpView.operandIndex); auto consumer = cast(dependence.indexingOpView.op); AffineMap consumerIndexingMap = consumer.getIndexingMap(dependence.indexingOpView.operandIndex); AffineMap prunedProducerIndexingMap = pruneReductionDimsFromMap( producer.iterator_types().getValue(), producerIndexingMap); if (!prunedProducerIndexingMap.isPermutation()) return None; if (consumerIndexingMap.getNumResults() != prunedProducerIndexingMap.getNumResults()) return None; 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 None; 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 &fusableLoops) { Optional lastFusableLoop; for (unsigned pos : llvm::map_range(map.getResults(), [](AffineExpr expr) { return expr.cast().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(%c, %cst) /// 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 collectFusableLoops(ArrayRef ops, const FusableOpDependencesTy &fusableDependences) { assert(!ops.empty()); auto getNumOuterParallelLoops = [](LinalgOp linalgOp) { return linalgOp.iterator_types() .getValue() .take_while([](Attribute attr) -> bool { return attr.cast().getValue() == getParallelIteratorTypeName(); }) .size(); }; size_t numOuterParallelLoops = getNumOuterParallelLoops(ops.back()); for (auto op : ops.drop_back()) { numOuterParallelLoops = std::min(numOuterParallelLoops, getNumOuterParallelLoops(op)); } std::set fusableLoops; auto range = llvm::seq(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 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 defered to later. if (doesTransposeAccess(*consumerLoopToProducerLoop, fusableLoops)) { op.emitRemark("unhandled fusion when fusion requires permutation"); return {}; } std::set candidates; for (AffineExpr expr : consumerLoopToProducerLoop->getResults()) { unsigned position = expr.cast().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 ops, const LinalgDependenceGraph &dependenceGraph) { FusableOpDependencesTy fusableDependences; // 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. DenseMap fusedProducerIndexingMap; for (LinalgOp op : reverse(ops)) { for (auto operandIndex : llvm::seq(0, op.getNumInputsAndOutputBuffers())) { Optional fusableDependence = findFusableProducer(op, operandIndex, dependenceGraph); if (!fusableDependence) continue; LinalgOp producerOp = cast(fusableDependence->dependentOpView.op); // 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()) { op.emitRemark("unhandled fusion of ops in different basic blocks"); return FusableOpDependencesTy{}; } // Make sure that the indexing map of the view used for fusion in the // producer is a projected permutation. unsigned producerIdx = fusableDependence->dependentOpView.operandIndex; AffineMap producerMap = producerOp.getIndexingMap(producerIdx); if (!producerMap.isProjectedPermutation()) { op.emitRemark( "unhandled non permutation indexing map for fused view in " "producer for operand at index ") << operandIndex; return FusableOpDependencesTy{}; } unsigned consumerIdx = fusableDependence->indexingOpView.operandIndex; AffineMap consumerMap = op.getIndexingMap(consumerIdx); if (!consumerMap.isProjectedPermutation()) { op.emitRemark( "unhandled case where indexing map for fused view in the consumer " "is " "not a projected permuration while fusing at index ") << operandIndex; return FusableOpDependencesTy{}; } // Check if the producer is already a fusion candidate. Cannot fuse this // dependence if it has a different indexing map when used in the // consumer. if (fusedProducerIndexingMap.count(producerOp.getOperation()) && fusedProducerIndexingMap[producerOp.getOperation()] != consumerMap) { op.emitRemark( "unhandled fusion to the same producer but with different " "indexing maps"); return FusableOpDependencesTy{}; } fusedProducerIndexingMap[producerOp.getOperation()] = consumerMap; fusableDependences[producerOp.getOperation()].push_back( *fusableDependence); } } 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 Optional tileRootOperation( OpBuilder &builder, LinalgOp op, ArrayRef tileSizeVector, const LinalgTilingOptions &options, const std::set &fusedLoops) { SmallVector tileSizes(tileSizeVector.begin(), tileSizeVector.end()); auto zero = std_constant_index(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(builder, 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 fuseOperations(OpBuilder &builder, LinalgOp tiledOp, ArrayRef fusionCandidates, const FusableOpDependencesTy &fusableDependences, const std::set &fusedLoops) { OpBuilder::InsertionGuard guard(builder); builder.setInsertionPoint(tiledOp); DenseMap fusedLoopsAndRanges; for (unsigned loop : fusedLoops) { ShapeDimension shapeDim = getShapeDefiningLoopRange(tiledOp, loop, true); fusedLoopsAndRanges[loop] = getRangeFromOperandShape( builder, tiledOp.getLoc(), shapeDim.shape, shapeDim.dimension); } SmallVector fusedOps(fusionCandidates.size()); for (auto candidate : enumerate(llvm::reverse(fusionCandidates))) { LinalgOp fusedOp = fuse(builder, candidate.value(), fusedLoopsAndRanges); fusedOps[fusionCandidates.size() - candidate.index() - 1] = fusedOp; builder.setInsertionPoint(fusedOp); } return fusedOps; } template static Optional tileAndFuseLinalgOpsImpl(OpBuilder &builder, ArrayRef ops, const LinalgDependenceGraph &dependenceGraph, const LinalgTilingOptions &tilingOptions) { if (ops.empty()) return llvm::None; LinalgOp rootOp = ops.back(); for (auto op : enumerate(ops)) { // TODO: Nothing in the fusion of sequence of ops is specific to // buffers. This check can be removed after it is tested on tensors. LinalgOp linalgOp = op.value(); if (!linalgOp.hasBufferSemantics()) { linalgOp.emitError("tile and fuse only tested for buffer operation"); return llvm::None; } } // TODO: Support interchange with tile + fuse. This might actually help do // better fusion. if (!tilingOptions.interchangeVector.empty()) { rootOp.emitError("unable to handle tile and fuse with interchange"); return llvm::None; } OpBuilder::InsertionGuard guard(builder); builder.setInsertionPoint(rootOp); ScopedContext scope(builder, rootOp.getLoc()); // Find all the producers. FusableOpDependencesTy fusableDependences = findAllFusableDependences(ops, dependenceGraph); if (fusableDependences.empty()) return llvm::None; TiledAndFusedLinalgOps ret; // Find the loops that can be tiled and fused. ret.fusedLoopDims = collectFusableLoops(ops, fusableDependences); // If there are no fusable dependences or there are no tile+fusable loops, // just return. if (ret.fusedLoopDims.empty()) { return llvm::None; } // Tile the fused loops in the last operation in the list. SmallVector tileSizeVector = tilingOptions.tileSizeComputationFunction(builder, rootOp); Optional tiledRootOp = tileRootOperation( builder, rootOp, tileSizeVector, tilingOptions, ret.fusedLoopDims); if (!tiledRootOp) { rootOp.emitError("failed to tile the fused loops"); return llvm::None; } 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(builder, ret.op, ops.drop_back(), fusableDependences, ret.fusedLoopDims); return ret; } Optional mlir::linalg::tileAndFuseLinalgOps(OpBuilder &builder, ArrayRef ops, const LinalgDependenceGraph &dependenceGraph, const LinalgTilingOptions &tilingOptions) { switch (tilingOptions.loopType) { case LinalgTilingLoopType::Loops: return tileAndFuseLinalgOpsImpl(builder, ops, dependenceGraph, tilingOptions); case LinalgTilingLoopType::ParallelLoops: return tileAndFuseLinalgOpsImpl( builder, ops, dependenceGraph, tilingOptions); default:; } return llvm::None; }