xref: /llvm-project/mlir/lib/Dialect/Linalg/Transforms/Fusion.cpp (revision 0a81ace0047a2de93e71c82cdf0977fc989660df)
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 #include <optional>
35 
36 #define DEBUG_TYPE "linalg-fusion"
37 
38 using namespace mlir;
39 using namespace mlir::linalg;
40 
41 /// Implements a simple high-level fusion pass on linalg structured operations.
42 ///
43 /// In each block, linalg ops are processed in reverse textual order.
44 /// Given a linalg op `O`, fusion occurs by:
45 ///   1. inspecting the linalg ops that write into the views read by `O`. There
46 ///      are 2 cases:
47 ///      a) buffer case: use the SSA value of the views and a simple alias
48 ///         analysis on subview ops to determine producer-consumer dependences;
49 ///      b) tensor case: use SSA use-def chains on extract_slice ops;
50 ///   2. greedily fuse the linalg ops that produce the subview/extract_slice.
51 ///   3. inspect the fused ops and determine whether they have other remaining
52 ///      LinalgOp uses. If not, then erase the original producing linalg op.
53 ///
54 /// More advanced use cases, analyses as well as profitability heuristics are
55 /// left for future work.
56 
57 struct ShapeDimension {
58   Value shape;
59   unsigned dimension;
60 };
61 
62 // Given an `op`, returns the first (`shape`, `dimension`) pair that identifies
63 // the loop range at `loopDepth`. The semantics of the loopToOperandRangesMaps
64 // guarantees at least one such dimension is found. If multiple candidates exist
65 // they must agree by construction (i.e. have the same size) and we just return
66 // the first one.
67 static ShapeDimension
68 getShapeDefiningLoopRange(LinalgOp op, unsigned loopDepth,
69                           bool fromSubViewOpOnly = false) {
70   // Iterate over the inputs and outputs in order.
71   // Extract the subranges from the linearized ranges.
72   for (OpOperand &opOperand : op->getOpOperands()) {
73     // The method `getRangeFromOperandShape` requires using SubViewOp or
74     // ExtractSliceOps. If the value isn't defined from there continue.
75     // todo: The method should be adapted to get the values from
76     // `ViewInterface`. The interface needs a `getOrCreateRanges` method which
77     // currently returns a `linalg.range`. The fix here is to move this op to
78     // `std` dialect and add the method to `ViewInterface`.
79     if (fromSubViewOpOnly &&
80         !isa_and_nonnull<memref::SubViewOp, tensor::ExtractSliceOp>(
81             opOperand.get().getDefiningOp()))
82       continue;
83 
84     AffineMap map = op.getMatchingIndexingMap(&opOperand);
85     LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange I/O idx: "
86                             << opOperand.getOperandNumber() << "\n");
87     LLVM_DEBUG(llvm::dbgs()
88                << "getShapeDefiningLoopRange map: " << map << "\n");
89     SmallVector<Value, 8> shapeRanges(map.getNumResults(), nullptr);
90     for (const auto &en : llvm::enumerate(map.getResults())) {
91       auto dimExpr = en.value().dyn_cast<AffineDimExpr>();
92       if (!dimExpr)
93         continue;
94       if (loopDepth == en.value().cast<AffineDimExpr>().getPosition()) {
95         LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange loopDepth: "
96                                 << loopDepth << "\n");
97         LLVM_DEBUG(llvm::dbgs() << "getShapeDefiningLoopRange shape: "
98                                 << opOperand.get() << "\n");
99         return ShapeDimension{opOperand.get(),
100                               static_cast<unsigned>(en.index())};
101       }
102     }
103   }
104   llvm_unreachable("Expect to be able to extract a shape defining loop range");
105 }
106 
107 static SmallVector<Value> getTiledOperands(LinalgOp producer) {
108   return producer->getOperands();
109 }
110 
111 /// Fuses the producer by cloning the `producer`. The `fusedLoopsAndRanges`
112 /// provides the loop range information for the fused loops. The rest are
113 /// obtained from the producer itself, since they are not tiled + fused.
114 static LinalgOp fuse(OpBuilder &b, LinalgOp producer,
115                      const DenseMap<unsigned, Range> &fusedLoopsAndRanges) {
116   SmallVector<OpFoldResult> ivs, tileSizes, sizeBounds;
117   SmallVector<Range> loopRanges;
118   Location loc = producer.getLoc();
119 
120   for (unsigned i = 0, e = producer.getNumLoops(); i < e; ++i) {
121     auto shapeDim = getShapeDefiningLoopRange(producer, i);
122     OpFoldResult dim =
123         createFoldedDimOp(b, loc, shapeDim.shape, shapeDim.dimension);
124     sizeBounds.push_back(dim);
125     auto it = fusedLoopsAndRanges.find(i);
126     if (it != fusedLoopsAndRanges.end()) {
127       ivs.push_back(it->second.offset);
128       tileSizes.push_back(it->second.size);
129       loopRanges.push_back(it->second);
130       LLVM_DEBUG(llvm::dbgs() << "tiled loop#" << i << " with LoopRange "
131                               << loopRanges.back() << "\n");
132     } else {
133       tileSizes.push_back(b.getIndexAttr(0));
134       loopRanges.push_back(Range{b.getIndexAttr(0), dim, b.getIndexAttr(1)});
135       LLVM_DEBUG(llvm::dbgs() << "full loop#" << i << " with LoopRange "
136                               << loopRanges.back() << "\n");
137     }
138   }
139 
140   SmallVector<Value, 8> clonedShapes;
141   clonedShapes.reserve(producer->getNumOperands());
142 
143   // Compute subranges for all tensor input/output operands.
144   clonedShapes.append(makeTiledShapes(
145       b, loc, producer, getTiledOperands(producer), ivs, tileSizes, sizeBounds,
146       /**omitPartialTileCheck=*/false));
147 
148   // Iterate over the results in order.
149   // Extract the subtensor type from the linearized range.
150   // Since we do not enforce any canonicalizations on the fly, this is always
151   // fully dynamic at construction time.
152   SmallVector<Type, 4> resultTypes;
153   resultTypes.reserve(producer->getNumResults());
154   for (OpOperand *operand : producer.getDpsInitOperands()) {
155     auto tensorType = operand->get().getType().dyn_cast<RankedTensorType>();
156     if (!tensorType)
157       continue;
158     unsigned rank = tensorType.getRank();
159     SmallVector<int64_t, 4> staticOffsetsVector(
160         rank, ShapedType::kDynamic);
161     SmallVector<int64_t, 4> staticSizesVector(rank, ShapedType::kDynamic);
162     SmallVector<int64_t, 4> staticStridesVector(
163         rank, ShapedType::kDynamic);
164     resultTypes.push_back(tensor::ExtractSliceOp::inferResultType(
165         tensorType, staticOffsetsVector, staticSizesVector,
166         staticStridesVector));
167   }
168 
169   Operation *clonedOp = clone(b, producer, resultTypes, clonedShapes);
170 
171   // Shift all IndexOp results by the tile offset.
172   SmallVector<OpFoldResult> allIvs = llvm::to_vector(
173       llvm::map_range(loopRanges, [&](Range range) { return range.offset; }));
174   offsetIndices(b, clonedOp, allIvs);
175 
176   return clonedOp;
177 }
178 
179 /// Get the loop range for a dimension `dim` based on the `shapedOperand`. It is
180 /// expected to be defined by a subview op or an extract_slice op.
181 static Range getRangeFromOperandShape(OpBuilder &b, Location loc,
182                                       Value shapedOperand, unsigned dim) {
183   Operation *shapeProducingOp = shapedOperand.getDefiningOp();
184   if (auto subViewOp = dyn_cast<memref::SubViewOp>(shapeProducingOp))
185     return subViewOp.getOrCreateRanges(b, loc)[dim];
186   if (auto sliceOp = dyn_cast<tensor::ExtractSliceOp>(shapeProducingOp))
187     return sliceOp.getOrCreateRanges(b, loc)[dim];
188   llvm_unreachable("SubviewOp or ExtractSliceOp expected");
189 }
190 
191 /// Fuses the producer into the loop immediately enclosing the consumer.
192 /// This is achieved by "recomputing" the producer at the time it
193 /// is needed just before the consumer.
194 static LinalgOp fuse(OpBuilder &b, LinalgOp producerOp, AffineMap producerMap,
195                      OpOperand &consumerOpOperand) {
196   LLVM_DEBUG(llvm::dbgs() << "Producer map: " << producerMap << "\n");
197   DenseMap<unsigned, Range> fusedLoopsAndRanges;
198   Value shapedOperand = consumerOpOperand.get();
199   for (const auto &en : llvm::enumerate(producerMap.getResults())) {
200     unsigned posInProducerLoop = en.value().cast<AffineDimExpr>().getPosition();
201     fusedLoopsAndRanges[posInProducerLoop] = getRangeFromOperandShape(
202         b, consumerOpOperand.getOwner()->getLoc(), shapedOperand, en.index());
203   }
204   return fuse(b, producerOp, fusedLoopsAndRanges);
205 }
206 
207 // Encode structural fusion safety preconditions.
208 // Some of these will be lifted in the future with better analysis.
209 static bool isStructurallyFusableProducer(LinalgOp producer, Value consumedView,
210                                           LinalgOp consumer) {
211   assert(producer.hasBufferSemantics() &&
212          "expected linalg op with buffer semantics");
213   assert(consumer.hasBufferSemantics() &&
214          "expected linalg op with buffer semantics");
215   if (producer.getNumDpsInits() != 1) {
216     LLVM_DEBUG(llvm::dbgs() << "\nNot structurally fusable (multi-output)");
217     return false;
218   }
219   // Only fuse when the producer block dominates.
220   DominanceInfo dom(producer.getOperation());
221   if (!dom.dominates(producer->getBlock(), consumer->getBlock())) {
222     LLVM_DEBUG(
223         llvm::dbgs()
224         << "\nNot structurally fusable (producer block does not dominate)");
225     return false;
226   }
227   return true;
228 }
229 
230 bool mlir::linalg::isProducerLastWriteOfView(const LinalgDependenceGraph &graph,
231                                              LinalgOp consumer,
232                                              Value consumedView,
233                                              LinalgOp producer) {
234   assert(producer.hasBufferSemantics() &&
235          "expected linalg op with buffer semantics");
236   assert(consumer.hasBufferSemantics() &&
237          "expected linalg op with buffer semantics");
238   // Make some simple structural checks that alleviate the need for more
239   // complex analyses.
240   if (!isStructurallyFusableProducer(producer, consumedView, consumer)) {
241     LLVM_DEBUG(llvm::dbgs() << "\n***Not static last write due to structure:\t"
242                             << *producer.getOperation());
243     return false;
244   }
245   // Check for any interleaved write to consumedView.
246   if (!graph.findCoveringWrites(producer, consumer, consumedView).empty()) {
247     LLVM_DEBUG(llvm::dbgs() << "\n***Not fusable due to interleaved write:\t"
248                             << *producer.getOperation());
249     return false;
250   }
251   return true;
252 }
253 
254 bool mlir::linalg::isFusableInto(const LinalgDependenceGraph &graph,
255                                  LinalgOp consumer, Value consumedView,
256                                  LinalgOp producer) {
257   assert(producer.hasBufferSemantics() &&
258          "expected linalg op with buffer semantics");
259   assert(consumer.hasBufferSemantics() &&
260          "expected linalg op with buffer semantics");
261   if (!isProducerLastWriteOfView(graph, consumer, consumedView, producer))
262     return false;
263   // Check for any fusion-preventing dependence to any shape read/written that
264   // would violate dependences.
265   if (!graph.findCoveringDependences(producer, consumer).empty()) {
266     LLVM_DEBUG(llvm::dbgs()
267                << "\n***Not fusable due to an interleaved dependence:\t"
268                << *producer.getOperation());
269     return false;
270   }
271   return true;
272 }
273 
274 /// For `consumer` with buffer semantics, find the Linalg operation on buffers
275 /// that is the last writer of `consumerOpOperand`. For now the fusable
276 /// dependence is returned as an instance of the `dependenceGraph`.
277 static FailureOr<LinalgDependenceGraph::LinalgDependenceGraphElem>
278 findFusableProducer(OpOperand &consumerOpOperand,
279                     const LinalgDependenceGraph &dependenceGraph) {
280   LLVM_DEBUG(llvm::dbgs() << "findFusableProducer for: "
281                           << consumerOpOperand.get() << " @"
282                           << consumerOpOperand.getOperandNumber() << " in "
283                           << *consumerOpOperand.getOwner() << "\n");
284   LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner());
285   if (!consumerOp)
286     return failure();
287 
288   // Only consider RAW and WAW atm.
289   for (auto depType : {
290            LinalgDependenceGraph::DependenceType::RAW,
291            LinalgDependenceGraph::DependenceType::WAW,
292        }) {
293     LLVM_DEBUG(llvm::dbgs()
294                << "Dependencies into: " << *consumerOp.getOperation() << "\n");
295     for (auto dependence : llvm::make_filter_range(
296              dependenceGraph.getDependencesInto(consumerOp, depType),
297              [&](LinalgDependenceGraph::LinalgDependenceGraphElem elem) {
298                LLVM_DEBUG(llvm::dbgs() << "Inspect dependence btw: "
299                                        << elem.getIndexingValue() << " and "
300                                        << elem.getDependentValue() << "\n");
301                Value v = elem.getIndexingValue();
302                std::optional<unsigned> operandNum =
303                    elem.getIndexingOpViewOperandNum();
304                return isa<LinalgOp>(elem.getDependentOp()) &&
305                       v == consumerOpOperand.get() && operandNum &&
306                       *operandNum == consumerOpOperand.getOperandNumber();
307              })) {
308       // Consumer consumes this view, `isStructurallyFusableProducer` also
309       // checks whether it is a strict subview of the producer view.
310       auto producer = cast<LinalgOp>(dependence.getDependentOp());
311       LLVM_DEBUG(llvm::dbgs()
312                  << "\n"
313                  << LinalgDependenceGraph::getDependenceTypeStr(depType)
314                  << "producer: " << *dependence.getDependentOp()
315                  << " view: " << dependence.getDependentValue() << "\n");
316 
317       // If the producer and consumer have tensor semantics, the only dependence
318       // between them is through a RAW dependence and they are fusable by
319       // construction. For buffer semantics need additional checks.
320       if (producer.hasBufferSemantics() && consumerOp.hasBufferSemantics() &&
321           isFusableInto(dependenceGraph, consumerOp, consumerOpOperand.get(),
322                         producer))
323         return dependence;
324       if (producer.hasTensorSemantics() && consumerOp.hasTensorSemantics()) {
325         assert(dependence.dependenceType ==
326                LinalgDependenceGraph::DependenceType::RAW);
327         return dependence;
328       }
329     }
330   }
331   return failure();
332 }
333 
334 FailureOr<FusionInfo>
335 mlir::linalg::fuseProducerOfBuffer(OpBuilder &b, OpOperand &consumerOpOperand,
336                                    const LinalgDependenceGraph &graph) {
337   std::optional<LinalgDependenceGraph::LinalgDependenceGraphElem>
338       fusableDependence = findFusableProducer(consumerOpOperand, graph);
339   if (!fusableDependence)
340     return failure();
341 
342   LinalgOp producerOp = dyn_cast<LinalgOp>(fusableDependence->getDependentOp());
343   if (!producerOp)
344     return failure();
345 
346   // If producer is already in the same block as consumer, we are done.
347   if (consumerOpOperand.get().getParentBlock() ==
348       fusableDependence->getDependentValue().getParentBlock())
349     return failure();
350 
351   std::optional<AffineMap> producerMap =
352       fusableDependence->getDependentOpViewIndexingMap();
353   if (!producerMap)
354     return failure();
355 
356   // Must be a subview or an extract_slice to guarantee there are loops we can
357   // fuse into.
358   auto subView = consumerOpOperand.get().getDefiningOp<memref::SubViewOp>();
359   if (!subView) {
360     LLVM_DEBUG(llvm::dbgs() << "\nNot fusable (not a subview)");
361     return failure();
362   }
363 
364   // Fuse `producer` just before `consumer`.
365   OpBuilder::InsertionGuard g(b);
366   b.setInsertionPoint(consumerOpOperand.getOwner());
367   LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: "
368                           << *consumerOpOperand.getOwner() << "\n");
369 
370   auto fusedProducer = fuse(b, producerOp, *producerMap, consumerOpOperand);
371   return FusionInfo{producerOp, fusedProducer};
372 }
373 
374 /// Walk back use-def chain through scf::For yields.
375 /// Sets `producer` and `outputIndex` if it finds a producer LinalgOp
376 
377 // TODO(ravishankarm, ntv): This can be moved into the dependence graphs
378 // dependence tracking since the dependence tracking is similar to what is done
379 // w.r.t to buffers.
380 static void getProducerOfTensor(Value tensor, OpResult &opResult) {
381   if (!tensor.getType().isa<RankedTensorType>())
382     return;
383 
384   while (true) {
385     LLVM_DEBUG(llvm::dbgs() << "\ngetProducerOfTensor: " << tensor);
386     if (auto linalgOp = tensor.getDefiningOp<LinalgOp>()) {
387       opResult = tensor.cast<OpResult>();
388       return;
389     }
390     if (auto sliceOp = tensor.getDefiningOp<tensor::ExtractSliceOp>()) {
391       tensor = sliceOp.getSource();
392       continue;
393     }
394     if (auto blockArg = tensor.dyn_cast<BlockArgument>()) {
395       if (auto forOp = blockArg.getDefiningOp<scf::ForOp>()) {
396         tensor = *(forOp.getIterOperands().begin() + blockArg.getArgNumber());
397         continue;
398       }
399     }
400     return;
401   }
402 }
403 
404 FailureOr<FusionInfo>
405 mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpOperand &consumerOpOperand) {
406   Value inputTensor = consumerOpOperand.get();
407   OpResult producerOpResult;
408   getProducerOfTensor(inputTensor, producerOpResult);
409   if (!producerOpResult) {
410     LLVM_DEBUG(llvm::dbgs() << "\nUnable to find producer");
411     return failure();
412   }
413   return fuseProducerOfTensor(b, producerOpResult, consumerOpOperand);
414 }
415 
416 FailureOr<FusionInfo>
417 mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpResult producerOpResult,
418                                    OpOperand &consumerOpOperand) {
419   auto producerOp = dyn_cast<LinalgOp>(producerOpResult.getOwner());
420   if (!producerOp)
421     return failure();
422 
423   LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner());
424   if (!consumerOp)
425     return failure();
426 
427   Value inputTensor = consumerOpOperand.get();
428 
429   // Must be an extract_slice op to guarantee there are loops we can fuse into.
430   auto sliceOp = inputTensor.getDefiningOp<tensor::ExtractSliceOp>();
431   if (!sliceOp) {
432     LLVM_DEBUG(llvm::dbgs()
433                << "\nNot fusable, not an extract_slice op: " << inputTensor);
434     return failure();
435   }
436 
437   // If producer is already in the same block as consumer, we are done.
438   if (consumerOpOperand.get().getParentBlock() ==
439       producerOpResult.getParentBlock())
440     return failure();
441 
442   // Insert fused `producer` just before `consumer`.
443   OpBuilder::InsertionGuard g(b);
444   b.setInsertionPoint(consumerOp);
445   LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: " << *consumerOp << "\n");
446   OpOperand *opOperand =
447       producerOp.getDpsInitOperand(producerOpResult.getResultNumber());
448   LinalgOp fusedProducer =
449       fuse(b, producerOp, producerOp.getMatchingIndexingMap(opOperand),
450            consumerOpOperand);
451 
452   // Replace use.
453   // Canonicalizations are not guaranteed to have happened before constructing
454   // `fusedProducer`. In the tensor case this can result in temporary type
455   // mismatches. Insert a `tensor.cast` op to propagate the transformation
456   // invariant that types are compatible.
457   Value def = fusedProducer->getResult(producerOpResult.getResultNumber());
458   Type consumerType = consumerOpOperand.get().getType();
459   if (consumerType != def.getType())
460     def = b.create<tensor::CastOp>(fusedProducer.getLoc(), consumerType, def);
461   consumerOpOperand.set(def);
462   return FusionInfo{cast<LinalgOp>(producerOpResult.getOwner()), fusedProducer};
463 }
464