xref: /llvm-project/mlir/lib/Dialect/Linalg/Transforms/Fusion.cpp (revision b4db15a949646f45011f31c58133adab59f8ddb0)
1 //===- Fusion.cpp - Implementation of linalg Fusion -----------------------===//
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 the linalg dialect Fusion pass.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "mlir/Dialect/Affine/IR/AffineOps.h"
14 #include "mlir/Dialect/Arith/IR/Arith.h"
15 #include "mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h"
16 #include "mlir/Dialect/Linalg/IR/Linalg.h"
17 #include "mlir/Dialect/Linalg/Passes.h"
18 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
19 #include "mlir/Dialect/Linalg/Utils/Utils.h"
20 #include "mlir/Dialect/MemRef/IR/MemRef.h"
21 #include "mlir/Dialect/Tensor/IR/Tensor.h"
22 #include "mlir/IR/AffineExpr.h"
23 #include "mlir/IR/AffineMap.h"
24 #include "mlir/IR/Dominance.h"
25 #include "mlir/Support/LLVM.h"
26 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
27 #include "mlir/Transforms/RegionUtils.h"
28 #include "llvm/ADT/MapVector.h"
29 #include "llvm/ADT/ScopeExit.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 
33 #include <set>
34 
35 #define DEBUG_TYPE "linalg-fusion"
36 
37 using namespace mlir;
38 using namespace mlir::linalg;
39 
40 /// Implements a simple high-level fusion pass on linalg structured operations.
41 ///
42 /// In each block, linalg ops are processed in reverse textual order.
43 /// Given a linalg op `O`, fusion occurs by:
44 ///   1. inspecting the linalg ops that write into the views read by `O`. There
45 ///      are 2 cases:
46 ///      a) buffer case: use the SSA value of the views and a simple alias
47 ///         analysis on subview ops to determine producer-consumer dependences;
48 ///      b) tensor case: use SSA use-def chains on extract_slice ops;
49 ///   2. greedily fuse the linalg ops that produce the subview/extract_slice.
50 ///   3. inspect the fused ops and determine whether they have other remaining
51 ///      LinalgOp uses. If not, then erase the original producing linalg op.
52 ///
53 /// More advanced use cases, analyses as well as profitability heuristics are
54 /// left for future work.
55 
56 struct ShapeDimension {
57   Value shape;
58   unsigned dimension;
59 };
60 
61 // Given an `op`, returns the first (`shape`, `dimension`) pair that identifies
62 // the loop range at `loopDepth`. The semantics of the loopToOperandRangesMaps
63 // guarantees at least one such dimension is found. If multiple candidates exist
64 // they must agree by construction (i.e. have the same size) and we just return
65 // the first one.
66 static ShapeDimension
67 getShapeDefiningLoopRange(LinalgOp op, unsigned loopDepth,
68                           bool fromSubViewOpOnly = false) {
69   // Iterate over the inputs and outputs in order.
70   // Extract the subranges from the linearized ranges.
71   for (OpOperand &opOperand : op->getOpOperands()) {
72     // The method `getRangeFromOperandShape` requires using SubViewOp or
73     // ExtractSliceOps. If the value isn't defined from there continue.
74     // todo: The method should be adapted to get the values from
75     // `ViewInterface`. The interface needs a `getOrCreateRanges` method which
76     // currently returns a `linalg.range`. The fix here is to move this op to
77     // `std` dialect and add the method to `ViewInterface`.
78     if (fromSubViewOpOnly &&
79         !isa_and_nonnull<memref::SubViewOp, tensor::ExtractSliceOp>(
80             opOperand.get().getDefiningOp()))
81       continue;
82 
83     AffineMap map = op.getMatchingIndexingMap(&opOperand);
84     LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange I/O idx: "
85                             << opOperand.getOperandNumber() << "\n");
86     LLVM_DEBUG(llvm::dbgs()
87                << "getShapeDefiningLoopRange map: " << map << "\n");
88     SmallVector<Value, 8> shapeRanges(map.getNumResults(), nullptr);
89     for (const auto &en : llvm::enumerate(map.getResults())) {
90       auto dimExpr = en.value().dyn_cast<AffineDimExpr>();
91       if (!dimExpr)
92         continue;
93       if (loopDepth == en.value().cast<AffineDimExpr>().getPosition()) {
94         LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange loopDepth: "
95                                 << loopDepth << "\n");
96         LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange shape: "
97                                 << opOperand.get() << "\n");
98         return ShapeDimension{opOperand.get(),
99                               static_cast<unsigned>(en.index())};
100       }
101     }
102   }
103   llvm_unreachable("Expect to be able to extract a shape defining loop range");
104 }
105 
106 static SmallVector<Value> getTiledOperands(LinalgOp producer) {
107   return producer->getOperands();
108 }
109 
110 /// Fuses the producer by cloning the `producer`. The `fusedLoopsAndRanges`
111 /// provides the loop range information for the fused loops. The rest are
112 /// obtained from the producer itself, since they are not tiled + fused.
113 static LinalgOp fuse(OpBuilder &b, LinalgOp producer,
114                      const DenseMap<unsigned, Range> &fusedLoopsAndRanges) {
115   SmallVector<OpFoldResult> ivs, tileSizes, sizeBounds;
116   SmallVector<Range> loopRanges;
117   Location loc = producer.getLoc();
118 
119   for (unsigned i = 0, e = producer.getNumLoops(); i < e; ++i) {
120     auto shapeDim = getShapeDefiningLoopRange(producer, i);
121     OpFoldResult dim =
122         createFoldedDimOp(b, loc, shapeDim.shape, shapeDim.dimension);
123     sizeBounds.push_back(dim);
124     auto it = fusedLoopsAndRanges.find(i);
125     if (it != fusedLoopsAndRanges.end()) {
126       ivs.push_back(it->second.offset);
127       tileSizes.push_back(it->second.size);
128       loopRanges.push_back(it->second);
129       LLVM_DEBUG(llvm::dbgs() << "tiled loop#" << i << " with LoopRange "
130                               << loopRanges.back() << "\n");
131     } else {
132       tileSizes.push_back(b.getIndexAttr(0));
133       loopRanges.push_back(Range{b.getIndexAttr(0), dim, b.getIndexAttr(1)});
134       LLVM_DEBUG(llvm::dbgs() << "full loop#" << i << " with LoopRange "
135                               << loopRanges.back() << "\n");
136     }
137   }
138 
139   SmallVector<Value, 8> clonedShapes;
140   clonedShapes.reserve(producer->getNumOperands());
141 
142   // Compute subranges for all tensor input/output operands.
143   clonedShapes.append(makeTiledShapes(
144       b, loc, producer, getTiledOperands(producer), ivs, tileSizes, sizeBounds,
145       /**omitPartialTileCheck=*/false));
146 
147   // Iterate over the results in order.
148   // Extract the subtensor type from the linearized range.
149   // Since we do not enforce any canonicalizations on the fly, this is always
150   // fully dynamic at construction time.
151   SmallVector<Type, 4> resultTypes;
152   resultTypes.reserve(producer->getNumResults());
153   for (OpOperand *operand : producer.getDpsInitOperands()) {
154     auto tensorType = operand->get().getType().dyn_cast<RankedTensorType>();
155     if (!tensorType)
156       continue;
157     unsigned rank = tensorType.getRank();
158     SmallVector<int64_t, 4> staticOffsetsVector(
159         rank, ShapedType::kDynamicStrideOrOffset);
160     SmallVector<int64_t, 4> staticSizesVector(rank, ShapedType::kDynamicSize);
161     SmallVector<int64_t, 4> staticStridesVector(
162         rank, ShapedType::kDynamicStrideOrOffset);
163     resultTypes.push_back(tensor::ExtractSliceOp::inferResultType(
164         tensorType, staticOffsetsVector, staticSizesVector,
165         staticStridesVector));
166   }
167 
168   Operation *clonedOp = producer.clone(b, loc, resultTypes, clonedShapes);
169 
170   // Shift all IndexOp results by the tile offset.
171   SmallVector<OpFoldResult> allIvs = llvm::to_vector(
172       llvm::map_range(loopRanges, [&](Range range) { return range.offset; }));
173   offsetIndices(b, clonedOp, allIvs);
174 
175   return clonedOp;
176 }
177 
178 /// Get the loop range for a dimension `dim` based on the `shapedOperand`. It is
179 /// expected to be defined by a subview op or an extract_slice op.
180 static Range getRangeFromOperandShape(OpBuilder &b, Location loc,
181                                       Value shapedOperand, unsigned dim) {
182   Operation *shapeProducingOp = shapedOperand.getDefiningOp();
183   if (auto subViewOp = dyn_cast<memref::SubViewOp>(shapeProducingOp))
184     return subViewOp.getOrCreateRanges(b, loc)[dim];
185   if (auto sliceOp = dyn_cast<tensor::ExtractSliceOp>(shapeProducingOp))
186     return sliceOp.getOrCreateRanges(b, loc)[dim];
187   llvm_unreachable("SubviewOp or ExtractSliceOp expected");
188 }
189 
190 /// Fuses the producer into the loop immediately enclosing the consumer.
191 /// This is achieved by "recomputing" the producer at the time it
192 /// is needed just before the consumer.
193 static LinalgOp fuse(OpBuilder &b, LinalgOp producerOp, AffineMap producerMap,
194                      OpOperand &consumerOpOperand) {
195   LLVM_DEBUG(llvm::dbgs() << "Producer map: " << producerMap << "\n");
196   DenseMap<unsigned, Range> fusedLoopsAndRanges;
197   Value shapedOperand = consumerOpOperand.get();
198   for (const auto &en : llvm::enumerate(producerMap.getResults())) {
199     unsigned posInProducerLoop = en.value().cast<AffineDimExpr>().getPosition();
200     fusedLoopsAndRanges[posInProducerLoop] = getRangeFromOperandShape(
201         b, consumerOpOperand.getOwner()->getLoc(), shapedOperand, en.index());
202   }
203   return fuse(b, producerOp, fusedLoopsAndRanges);
204 }
205 
206 // Encode structural fusion safety preconditions.
207 // Some of these will be lifted in the future with better analysis.
208 static bool isStructurallyFusableProducer(LinalgOp producer, Value consumedView,
209                                           LinalgOp consumer) {
210   assert(producer.hasBufferSemantics() &&
211          "expected linalg op with buffer semantics");
212   assert(consumer.hasBufferSemantics() &&
213          "expected linalg op with buffer semantics");
214   if (producer.getNumDpsInits() != 1) {
215     LLVM_DEBUG(llvm::dbgs() << "\nNot structurally fusable (multi-output)");
216     return false;
217   }
218   // Only fuse when the producer block dominates.
219   DominanceInfo dom(producer.getOperation());
220   if (!dom.dominates(producer->getBlock(), consumer->getBlock())) {
221     LLVM_DEBUG(
222         llvm::dbgs()
223         << "\nNot structurally fusable (producer block does not dominate)");
224     return false;
225   }
226   return true;
227 }
228 
229 bool mlir::linalg::isProducerLastWriteOfView(const LinalgDependenceGraph &graph,
230                                              LinalgOp consumer,
231                                              Value consumedView,
232                                              LinalgOp producer) {
233   assert(producer.hasBufferSemantics() &&
234          "expected linalg op with buffer semantics");
235   assert(consumer.hasBufferSemantics() &&
236          "expected linalg op with buffer semantics");
237   // Make some simple structural checks that alleviate the need for more
238   // complex analyses.
239   if (!isStructurallyFusableProducer(producer, consumedView, consumer)) {
240     LLVM_DEBUG(llvm::dbgs() << "\n***Not static last write due to structure:\t"
241                             << *producer.getOperation());
242     return false;
243   }
244   // Check for any interleaved write to consumedView.
245   if (!graph.findCoveringWrites(producer, consumer, consumedView).empty()) {
246     LLVM_DEBUG(llvm::dbgs() << "\n***Not fusable due to interleaved write:\t"
247                             << *producer.getOperation());
248     return false;
249   }
250   return true;
251 }
252 
253 bool mlir::linalg::isFusableInto(const LinalgDependenceGraph &graph,
254                                  LinalgOp consumer, Value consumedView,
255                                  LinalgOp producer) {
256   assert(producer.hasBufferSemantics() &&
257          "expected linalg op with buffer semantics");
258   assert(consumer.hasBufferSemantics() &&
259          "expected linalg op with buffer semantics");
260   if (!isProducerLastWriteOfView(graph, consumer, consumedView, producer))
261     return false;
262   // Check for any fusion-preventing dependence to any shape read/written that
263   // would violate dependences.
264   if (!graph.findCoveringDependences(producer, consumer).empty()) {
265     LLVM_DEBUG(llvm::dbgs()
266                << "\n***Not fusable due to an interleaved dependence:\t"
267                << *producer.getOperation());
268     return false;
269   }
270   return true;
271 }
272 
273 /// For `consumer` with buffer semantics, find the Linalg operation on buffers
274 /// that is the last writer of `consumerOpOperand`. For now the fusable
275 /// dependence is returned as an instance of the `dependenceGraph`.
276 static FailureOr<LinalgDependenceGraph::LinalgDependenceGraphElem>
277 findFusableProducer(OpOperand &consumerOpOperand,
278                     const LinalgDependenceGraph &dependenceGraph) {
279   LLVM_DEBUG(llvm::dbgs() << "findFusableProducer for: "
280                           << consumerOpOperand.get() << " @"
281                           << consumerOpOperand.getOperandNumber() << " in "
282                           << *consumerOpOperand.getOwner() << "\n");
283   LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner());
284   if (!consumerOp)
285     return failure();
286 
287   // Only consider RAW and WAW atm.
288   for (auto depType : {
289            LinalgDependenceGraph::DependenceType::RAW,
290            LinalgDependenceGraph::DependenceType::WAW,
291        }) {
292     LLVM_DEBUG(llvm::dbgs()
293                << "Dependencies into: " << *consumerOp.getOperation() << "\n");
294     for (auto dependence : llvm::make_filter_range(
295              dependenceGraph.getDependencesInto(consumerOp, depType),
296              [&](LinalgDependenceGraph::LinalgDependenceGraphElem elem) {
297                LLVM_DEBUG(llvm::dbgs() << "Inspect dependence btw: "
298                                        << elem.getIndexingValue() << " and "
299                                        << elem.getDependentValue() << "\n");
300                Value v = elem.getIndexingValue();
301                Optional<unsigned> operandNum =
302                    elem.getIndexingOpViewOperandNum();
303                return isa<LinalgOp>(elem.getDependentOp()) &&
304                       v == consumerOpOperand.get() && operandNum &&
305                       *operandNum == consumerOpOperand.getOperandNumber();
306              })) {
307       // Consumer consumes this view, `isStructurallyFusableProducer` also
308       // checks whether it is a strict subview of the producer view.
309       auto producer = cast<LinalgOp>(dependence.getDependentOp());
310       LLVM_DEBUG(llvm::dbgs()
311                  << "\n"
312                  << LinalgDependenceGraph::getDependenceTypeStr(depType)
313                  << "producer: " << *dependence.getDependentOp()
314                  << " view: " << dependence.getDependentValue() << "\n");
315 
316       // If the producer and consumer have tensor semantics, the only dependence
317       // between them is through a RAW dependence and they are fusable by
318       // construction. For buffer semantics need additional checks.
319       if (producer.hasBufferSemantics() && consumerOp.hasBufferSemantics() &&
320           isFusableInto(dependenceGraph, consumerOp, consumerOpOperand.get(),
321                         producer))
322         return dependence;
323       if (producer.hasTensorSemantics() && consumerOp.hasTensorSemantics()) {
324         assert(dependence.dependenceType ==
325                LinalgDependenceGraph::DependenceType::RAW);
326         return dependence;
327       }
328     }
329   }
330   return failure();
331 }
332 
333 FailureOr<FusionInfo>
334 mlir::linalg::fuseProducerOfBuffer(OpBuilder &b, OpOperand &consumerOpOperand,
335                                    const LinalgDependenceGraph &graph) {
336   Optional<LinalgDependenceGraph::LinalgDependenceGraphElem> fusableDependence =
337       findFusableProducer(consumerOpOperand, graph);
338   if (!fusableDependence)
339     return failure();
340 
341   LinalgOp producerOp = dyn_cast<LinalgOp>(fusableDependence->getDependentOp());
342   if (!producerOp)
343     return failure();
344 
345   // If producer is already in the same block as consumer, we are done.
346   if (consumerOpOperand.get().getParentBlock() ==
347       fusableDependence->getDependentValue().getParentBlock())
348     return failure();
349 
350   Optional<AffineMap> producerMap =
351       fusableDependence->getDependentOpViewIndexingMap();
352   if (!producerMap)
353     return failure();
354 
355   // Must be a subview or an extract_slice to guarantee there are loops we can
356   // fuse into.
357   auto subView = consumerOpOperand.get().getDefiningOp<memref::SubViewOp>();
358   if (!subView) {
359     LLVM_DEBUG(llvm::dbgs() << "\nNot fusable (not a subview)");
360     return failure();
361   }
362 
363   // Fuse `producer` just before `consumer`.
364   OpBuilder::InsertionGuard g(b);
365   b.setInsertionPoint(consumerOpOperand.getOwner());
366   LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: "
367                           << *consumerOpOperand.getOwner() << "\n");
368 
369   auto fusedProducer = fuse(b, producerOp, *producerMap, consumerOpOperand);
370   return FusionInfo{producerOp, fusedProducer};
371 }
372 
373 /// Walk back use-def chain through scf::For yields.
374 /// Sets `producer` and `outputIndex` if it finds a producer LinalgOp
375 
376 // TODO(ravishankarm, ntv): This can be moved into the dependence graphs
377 // dependence tracking since the dependence tracking is similar to what is done
378 // w.r.t to buffers.
379 static void getProducerOfTensor(Value tensor, OpResult &opResult) {
380   if (!tensor.getType().isa<RankedTensorType>())
381     return;
382 
383   while (true) {
384     LLVM_DEBUG(llvm::dbgs() << "\ngetProducerOfTensor: " << tensor);
385     if (auto linalgOp = tensor.getDefiningOp<LinalgOp>()) {
386       opResult = tensor.cast<OpResult>();
387       return;
388     }
389     if (auto sliceOp = tensor.getDefiningOp<tensor::ExtractSliceOp>()) {
390       tensor = sliceOp.getSource();
391       continue;
392     }
393     if (auto blockArg = tensor.dyn_cast<BlockArgument>()) {
394       if (auto forOp = blockArg.getDefiningOp<scf::ForOp>()) {
395         tensor = *(forOp.getIterOperands().begin() + blockArg.getArgNumber());
396         continue;
397       }
398     }
399     return;
400   }
401 }
402 
403 FailureOr<FusionInfo>
404 mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpOperand &consumerOpOperand) {
405   Value inputTensor = consumerOpOperand.get();
406   OpResult producerOpResult;
407   getProducerOfTensor(inputTensor, producerOpResult);
408   if (!producerOpResult) {
409     LLVM_DEBUG(llvm::dbgs() << "\nUnable to find producer");
410     return failure();
411   }
412   return fuseProducerOfTensor(b, producerOpResult, consumerOpOperand);
413 }
414 
415 FailureOr<FusionInfo>
416 mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpResult producerOpResult,
417                                    OpOperand &consumerOpOperand) {
418   auto producerOp = dyn_cast<LinalgOp>(producerOpResult.getOwner());
419   if (!producerOp)
420     return failure();
421 
422   LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner());
423   if (!consumerOp)
424     return failure();
425 
426   Value inputTensor = consumerOpOperand.get();
427 
428   // Must be an extract_slice op to guarantee there are loops we can fuse into.
429   auto sliceOp = inputTensor.getDefiningOp<tensor::ExtractSliceOp>();
430   if (!sliceOp) {
431     LLVM_DEBUG(llvm::dbgs()
432                << "\nNot fusable, not an extract_slice op: " << inputTensor);
433     return failure();
434   }
435 
436   // If producer is already in the same block as consumer, we are done.
437   if (consumerOpOperand.get().getParentBlock() ==
438       producerOpResult.getParentBlock())
439     return failure();
440 
441   // Insert fused `producer` just before `consumer`.
442   OpBuilder::InsertionGuard g(b);
443   b.setInsertionPoint(consumerOp);
444   LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: " << *consumerOp << "\n");
445   OpOperand *opOperand =
446       producerOp.getDpsInitOperand(producerOpResult.getResultNumber());
447   LinalgOp fusedProducer =
448       fuse(b, producerOp, producerOp.getMatchingIndexingMap(opOperand),
449            consumerOpOperand);
450 
451   // Replace use.
452   // Canonicalizations are not guaranteed to have happened before constructing
453   // `fusedProducer`. In the tensor case this can result in temporary type
454   // mismatches. Insert a `tensor.cast` op to propagate the transformation
455   // invariant that types are compatible.
456   Value def = fusedProducer->getResult(producerOpResult.getResultNumber());
457   Type consumerType = consumerOpOperand.get().getType();
458   if (consumerType != def.getType())
459     def = b.create<tensor::CastOp>(fusedProducer.getLoc(), consumerType, def);
460   consumerOpOperand.set(def);
461   return FusionInfo{cast<LinalgOp>(producerOpResult.getOwner()), fusedProducer};
462 }
463