xref: /llvm-project/mlir/lib/Dialect/Linalg/Transforms/Fusion.cpp (revision 039b969b32b64b64123dce30dd28ec4e343d893f)
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 "PassDetail.h"
14 #include "mlir/Dialect/Affine/IR/AffineOps.h"
15 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
16 #include "mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h"
17 #include "mlir/Dialect/Linalg/IR/Linalg.h"
18 #include "mlir/Dialect/Linalg/Passes.h"
19 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
20 #include "mlir/Dialect/Linalg/Utils/Utils.h"
21 #include "mlir/Dialect/MemRef/IR/MemRef.h"
22 #include "mlir/Dialect/Tensor/IR/Tensor.h"
23 #include "mlir/IR/AffineExpr.h"
24 #include "mlir/IR/AffineMap.h"
25 #include "mlir/IR/Dominance.h"
26 #include "mlir/Support/LLVM.h"
27 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
28 #include "mlir/Transforms/RegionUtils.h"
29 #include "llvm/ADT/MapVector.h"
30 #include "llvm/ADT/ScopeExit.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 
34 #include <set>
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.getInputAndOutputOperands()) {
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.getTiedIndexingMap(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.getInputAndOutputOperands();
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.getNumInputsAndOutputs());
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 (RankedTensorType t : producer.getOutputTensorTypes()) {
155     unsigned rank = t.getRank();
156     SmallVector<int64_t, 4> staticOffsetsVector(
157         rank, ShapedType::kDynamicStrideOrOffset);
158     SmallVector<int64_t, 4> staticSizesVector(rank, ShapedType::kDynamicSize);
159     SmallVector<int64_t, 4> staticStridesVector(
160         rank, ShapedType::kDynamicStrideOrOffset);
161     resultTypes.push_back(tensor::ExtractSliceOp::inferResultType(
162         t.cast<RankedTensorType>(), staticOffsetsVector, staticSizesVector,
163         staticStridesVector));
164   }
165 
166   Operation *clonedOp = producer.clone(b, loc, resultTypes, clonedShapes);
167 
168   // Shift all IndexOp results by the tile offset.
169   SmallVector<OpFoldResult> allIvs = llvm::to_vector(
170       llvm::map_range(loopRanges, [&](Range range) { return range.offset; }));
171   offsetIndices(b, clonedOp, allIvs);
172 
173   return clonedOp;
174 }
175 
176 /// Get the loop range for a dimension `dim` based on the `shapedOperand`. It is
177 /// expected to be defined by a subview op or an extract_slice op.
178 static Range getRangeFromOperandShape(OpBuilder &b, Location loc,
179                                       Value shapedOperand, unsigned dim) {
180   Operation *shapeProducingOp = shapedOperand.getDefiningOp();
181   if (auto subViewOp = dyn_cast<memref::SubViewOp>(shapeProducingOp))
182     return subViewOp.getOrCreateRanges(b, loc)[dim];
183   if (auto sliceOp = dyn_cast<tensor::ExtractSliceOp>(shapeProducingOp))
184     return sliceOp.getOrCreateRanges(b, loc)[dim];
185   llvm_unreachable("SubviewOp or ExtractSliceOp expected");
186 }
187 
188 /// Fuses the producer into the loop immediately enclosing the consumer.
189 /// This is achieved by "recomputing" the producer at the time it
190 /// is needed just before the consumer.
191 static LinalgOp fuse(OpBuilder &b, LinalgOp producerOp, AffineMap producerMap,
192                      OpOperand &consumerOpOperand) {
193   LLVM_DEBUG(llvm::dbgs() << "Producer map: " << producerMap << "\n");
194   DenseMap<unsigned, Range> fusedLoopsAndRanges;
195   Value shapedOperand = consumerOpOperand.get();
196   for (const auto &en : llvm::enumerate(producerMap.getResults())) {
197     unsigned posInProducerLoop = en.value().cast<AffineDimExpr>().getPosition();
198     fusedLoopsAndRanges[posInProducerLoop] = getRangeFromOperandShape(
199         b, consumerOpOperand.getOwner()->getLoc(), shapedOperand, en.index());
200   }
201   return fuse(b, producerOp, fusedLoopsAndRanges);
202 }
203 
204 // Encode structural fusion safety preconditions.
205 // Some of these will be lifted in the future with better analysis.
206 static bool isStructurallyFusableProducer(LinalgOp producer, Value consumedView,
207                                           LinalgOp consumer) {
208   assert(producer.hasBufferSemantics() &&
209          "expected linalg op with buffer semantics");
210   assert(consumer.hasBufferSemantics() &&
211          "expected linalg op with buffer semantics");
212   if (producer.getNumOutputs() != 1) {
213     LLVM_DEBUG(llvm::dbgs() << "\nNot structurally fusable (multi-output)");
214     return false;
215   }
216   // Only fuse when the producer block dominates.
217   DominanceInfo dom(producer.getOperation());
218   if (!dom.dominates(producer->getBlock(), consumer->getBlock())) {
219     LLVM_DEBUG(
220         llvm::dbgs()
221         << "\nNot structurally fusable (producer block does not dominate)");
222     return false;
223   }
224   return true;
225 }
226 
227 bool mlir::linalg::isProducerLastWriteOfView(const LinalgDependenceGraph &graph,
228                                              LinalgOp consumer,
229                                              Value consumedView,
230                                              LinalgOp producer) {
231   assert(producer.hasBufferSemantics() &&
232          "expected linalg op with buffer semantics");
233   assert(consumer.hasBufferSemantics() &&
234          "expected linalg op with buffer semantics");
235   // Make some simple structural checks that alleviate the need for more
236   // complex analyses.
237   if (!isStructurallyFusableProducer(producer, consumedView, consumer)) {
238     LLVM_DEBUG(llvm::dbgs() << "\n***Not static last write due to structure:\t"
239                             << *producer.getOperation());
240     return false;
241   }
242   // Check for any interleaved write to consumedView.
243   if (!graph.findCoveringWrites(producer, consumer, consumedView).empty()) {
244     LLVM_DEBUG(llvm::dbgs() << "\n***Not fusable due to interleaved write:\t"
245                             << *producer.getOperation());
246     return false;
247   }
248   return true;
249 }
250 
251 bool mlir::linalg::isFusableInto(const LinalgDependenceGraph &graph,
252                                  LinalgOp consumer, Value consumedView,
253                                  LinalgOp producer) {
254   assert(producer.hasBufferSemantics() &&
255          "expected linalg op with buffer semantics");
256   assert(consumer.hasBufferSemantics() &&
257          "expected linalg op with buffer semantics");
258   if (!isProducerLastWriteOfView(graph, consumer, consumedView, producer))
259     return false;
260   // Check for any fusion-preventing dependence to any shape read/written that
261   // would violate dependences.
262   if (!graph.findCoveringDependences(producer, consumer).empty()) {
263     LLVM_DEBUG(llvm::dbgs()
264                << "\n***Not fusable due to an interleaved dependence:\t"
265                << *producer.getOperation());
266     return false;
267   }
268   return true;
269 }
270 
271 /// For `consumer` with buffer semantics, find the Linalg operation on buffers
272 /// that is the last writer of `consumerOpOperand`. For now the fusable
273 /// dependence is returned as an instance of the `dependenceGraph`.
274 static FailureOr<LinalgDependenceGraph::LinalgDependenceGraphElem>
275 findFusableProducer(OpOperand &consumerOpOperand,
276                     const LinalgDependenceGraph &dependenceGraph) {
277   LLVM_DEBUG(llvm::dbgs() << "findFusableProducer for: "
278                           << consumerOpOperand.get() << " @"
279                           << consumerOpOperand.getOperandNumber() << " in "
280                           << *consumerOpOperand.getOwner() << "\n");
281   LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner());
282   if (!consumerOp)
283     return failure();
284 
285   // Only consider RAW and WAW atm.
286   for (auto depType : {
287            LinalgDependenceGraph::DependenceType::RAW,
288            LinalgDependenceGraph::DependenceType::WAW,
289        }) {
290     LLVM_DEBUG(llvm::dbgs()
291                << "Dependencies into: " << *consumerOp.getOperation() << "\n");
292     for (auto dependence : llvm::make_filter_range(
293              dependenceGraph.getDependencesInto(consumerOp, depType),
294              [&](LinalgDependenceGraph::LinalgDependenceGraphElem elem) {
295                LLVM_DEBUG(llvm::dbgs() << "Inspect dependence btw: "
296                                        << elem.getIndexingValue() << " and "
297                                        << elem.getDependentValue() << "\n");
298                Value v = elem.getIndexingValue();
299                Optional<unsigned> operandNum =
300                    elem.getIndexingOpViewOperandNum();
301                return isa<LinalgOp>(elem.getDependentOp()) &&
302                       v == consumerOpOperand.get() && operandNum &&
303                       *operandNum == consumerOpOperand.getOperandNumber();
304              })) {
305       // Consumer consumes this view, `isStructurallyFusableProducer` also
306       // checks whether it is a strict subview of the producer view.
307       auto producer = cast<LinalgOp>(dependence.getDependentOp());
308       LLVM_DEBUG(llvm::dbgs()
309                  << "\n"
310                  << LinalgDependenceGraph::getDependenceTypeStr(depType)
311                  << "producer: " << *dependence.getDependentOp()
312                  << " view: " << dependence.getDependentValue() << "\n");
313 
314       // If the producer and consumer have tensor semantics, the only dependence
315       // between them is through a RAW dependence and they are fusable by
316       // construction. For buffer semantics need additional checks.
317       if (producer.hasBufferSemantics() && consumerOp.hasBufferSemantics() &&
318           isFusableInto(dependenceGraph, consumerOp, consumerOpOperand.get(),
319                         producer))
320         return dependence;
321       if (producer.hasTensorSemantics() && consumerOp.hasTensorSemantics()) {
322         assert(dependence.dependenceType ==
323                LinalgDependenceGraph::DependenceType::RAW);
324         return dependence;
325       }
326     }
327   }
328   return failure();
329 }
330 
331 FailureOr<FusionInfo>
332 mlir::linalg::fuseProducerOfBuffer(OpBuilder &b, OpOperand &consumerOpOperand,
333                                    const LinalgDependenceGraph &graph) {
334   Optional<LinalgDependenceGraph::LinalgDependenceGraphElem> fusableDependence =
335       findFusableProducer(consumerOpOperand, graph);
336   if (!fusableDependence)
337     return failure();
338 
339   LinalgOp producerOp = dyn_cast<LinalgOp>(fusableDependence->getDependentOp());
340   if (!producerOp)
341     return failure();
342 
343   // If producer is already in the same block as consumer, we are done.
344   if (consumerOpOperand.get().getParentBlock() ==
345       fusableDependence->getDependentValue().getParentBlock())
346     return failure();
347 
348   Optional<AffineMap> producerMap =
349       fusableDependence->getDependentOpViewIndexingMap();
350   if (!producerMap)
351     return failure();
352 
353   // Must be a subview or an extract_slice to guarantee there are loops we can
354   // fuse into.
355   auto subView = consumerOpOperand.get().getDefiningOp<memref::SubViewOp>();
356   if (!subView) {
357     LLVM_DEBUG(llvm::dbgs() << "\nNot fusable (not a subview)");
358     return failure();
359   }
360 
361   // Fuse `producer` just before `consumer`.
362   OpBuilder::InsertionGuard g(b);
363   b.setInsertionPoint(consumerOpOperand.getOwner());
364   LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: "
365                           << *consumerOpOperand.getOwner() << "\n");
366 
367   auto fusedProducer = fuse(b, producerOp, *producerMap, consumerOpOperand);
368   return FusionInfo{producerOp, fusedProducer};
369 }
370 
371 /// Walk back use-def chain through scf::For yields.
372 /// Sets `producer` and `outputIndex` if it finds a producer LinalgOp
373 
374 // TODO(ravishankarm, ntv): This can be moved into the dependence graphs
375 // dependence tracking since the dependence tracking is similar to what is done
376 // w.r.t to buffers.
377 static void getProducerOfTensor(Value tensor, OpResult &opResult) {
378   if (!tensor.getType().isa<RankedTensorType>())
379     return;
380 
381   while (true) {
382     LLVM_DEBUG(llvm::dbgs() << "\ngetProducerOfTensor: " << tensor);
383     if (auto linalgOp = tensor.getDefiningOp<LinalgOp>()) {
384       opResult = tensor.cast<OpResult>();
385       return;
386     }
387     if (auto sliceOp = tensor.getDefiningOp<tensor::ExtractSliceOp>()) {
388       tensor = sliceOp.getSource();
389       continue;
390     }
391     if (auto blockArg = tensor.dyn_cast<BlockArgument>()) {
392       if (auto forOp = blockArg.getDefiningOp<scf::ForOp>()) {
393         tensor = *(forOp.getIterOperands().begin() + blockArg.getArgNumber());
394         continue;
395       }
396     }
397     return;
398   }
399 }
400 
401 FailureOr<FusionInfo>
402 mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpOperand &consumerOpOperand) {
403   Value inputTensor = consumerOpOperand.get();
404   OpResult producerOpResult;
405   getProducerOfTensor(inputTensor, producerOpResult);
406   if (!producerOpResult) {
407     LLVM_DEBUG(llvm::dbgs() << "\nUnable to find producer");
408     return failure();
409   }
410   return fuseProducerOfTensor(b, producerOpResult, consumerOpOperand);
411 }
412 
413 FailureOr<FusionInfo>
414 mlir::linalg::fuseProducerOfTensor(OpBuilder &b, OpResult producerOpResult,
415                                    OpOperand &consumerOpOperand) {
416   auto producerOp = dyn_cast<LinalgOp>(producerOpResult.getOwner());
417   if (!producerOp)
418     return failure();
419 
420   LinalgOp consumerOp = dyn_cast<LinalgOp>(consumerOpOperand.getOwner());
421   if (!consumerOp)
422     return failure();
423 
424   Value inputTensor = consumerOpOperand.get();
425 
426   // Must be an extract_slice op to guarantee there are loops we can fuse into.
427   auto sliceOp = inputTensor.getDefiningOp<tensor::ExtractSliceOp>();
428   if (!sliceOp) {
429     LLVM_DEBUG(llvm::dbgs()
430                << "\nNot fusable, not an extract_slice op: " << inputTensor);
431     return failure();
432   }
433 
434   // If producer is already in the same block as consumer, we are done.
435   if (consumerOpOperand.get().getParentBlock() ==
436       producerOpResult.getParentBlock())
437     return failure();
438 
439   // Insert fused `producer` just before `consumer`.
440   OpBuilder::InsertionGuard g(b);
441   b.setInsertionPoint(consumerOp);
442   LLVM_DEBUG(llvm::dbgs() << "Fuse into consumer: " << *consumerOp << "\n");
443   OpOperand *opOperand =
444       producerOp.getOutputOperand(producerOpResult.getResultNumber());
445   LinalgOp fusedProducer =
446       fuse(b, producerOp, producerOp.getTiedIndexingMap(opOperand),
447            consumerOpOperand);
448 
449   // Replace use.
450   // Canonicalizations are not guaranteed to have happened before constructing
451   // `fusedProducer`. In the tensor case this can result in temporary type
452   // mismatches. Insert a `tensor.cast` op to propagate the transformation
453   // invariant that types are compatible.
454   Value def = fusedProducer->getResult(producerOpResult.getResultNumber());
455   Type consumerType = consumerOpOperand.get().getType();
456   if (consumerType != def.getType())
457     def = b.create<tensor::CastOp>(fusedProducer.getLoc(), consumerType, def);
458   consumerOpOperand.set(def);
459   return FusionInfo{cast<LinalgOp>(producerOpResult.getOwner()), fusedProducer};
460 }
461