1 //===- Tiling.cpp - Implementation of linalg Tiling -----------------------===//
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 Tiling pass.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include <utility>
14 
15 #include "PassDetail.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/SCF/Transforms.h"
22 #include "mlir/Dialect/Tensor/IR/Tensor.h"
23 #include "mlir/IR/AffineExpr.h"
24 #include "mlir/IR/AffineMap.h"
25 #include "mlir/Transforms/FoldUtils.h"
26 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
27 
28 #include "llvm/Support/CommandLine.h"
29 
30 using namespace mlir;
31 using namespace mlir::linalg;
32 using namespace mlir::scf;
33 
34 #define DEBUG_TYPE "linalg-tiling"
35 
36 static bool isZero(Value v) {
37   if (auto cst = v.getDefiningOp<arith::ConstantIndexOp>())
38     return cst.value() == 0;
39   return false;
40 }
41 
42 std::tuple<SmallVector<Range, 4>, LoopIndexToRangeIndexMap>
43 mlir::linalg::makeTiledLoopRanges(RewriterBase &b, Location loc, AffineMap map,
44                                   ValueRange allShapeSizes,
45                                   ValueRange allTileSizes) {
46   assert(allTileSizes.size() == map.getNumResults());
47   // Apply `map` to get shape sizes in loop order.
48   auto shapeSizes = applyMapToValues(b, loc, map, allShapeSizes);
49   SmallVector<Value, 4> tileSizes(allTileSizes.begin(), allTileSizes.end());
50 
51   // Traverse the tile sizes, which are in loop order, erase zeros everywhere.
52   LoopIndexToRangeIndexMap loopIndexToRangeIndex;
53   for (int idx = 0, e = tileSizes.size(), zerosCount = 0; idx < e; ++idx) {
54     if (isZero(tileSizes[idx - zerosCount])) {
55       shapeSizes.erase(shapeSizes.begin() + idx - zerosCount);
56       tileSizes.erase(tileSizes.begin() + idx - zerosCount);
57       ++zerosCount;
58       continue;
59     }
60     loopIndexToRangeIndex[idx] = idx - zerosCount;
61   }
62 
63   // Create a new range with the applied tile sizes.
64   SmallVector<Range, 4> res;
65   for (unsigned idx = 0, e = tileSizes.size(); idx < e; ++idx)
66     res.push_back(Range{b.create<arith::ConstantIndexOp>(loc, 0),
67                         shapeSizes[idx], tileSizes[idx]});
68   return std::make_tuple(res, loopIndexToRangeIndex);
69 }
70 
71 void mlir::linalg::transformIndexOps(
72     RewriterBase &b, LinalgOp op, SmallVectorImpl<Value> &ivs,
73     const LoopIndexToRangeIndexMap &loopIndexToRangeIndex) {
74   SmallVector<Value> allIvs(op.getNumLoops(), nullptr);
75   for (auto &en : enumerate(allIvs)) {
76     auto rangeIndex = loopIndexToRangeIndex.find(en.index());
77     if (rangeIndex == loopIndexToRangeIndex.end())
78       continue;
79     en.value() = ivs[rangeIndex->second];
80   }
81   addTileLoopIvsToIndexOpResults(b, op, allIvs);
82 }
83 
84 // Insert a tile `source` into the destination tensor `dest`. The position at
85 // which the tile is inserted (as well as size of tile) is taken from a given
86 // ExtractSliceOp `sliceOp`.
87 static Value insertSliceIntoTensor(RewriterBase &b, Location loc,
88                                    tensor::ExtractSliceOp sliceOp, Value source,
89                                    Value dest) {
90   return b.create<tensor::InsertSliceOp>(
91       loc, sliceOp.source().getType(), source, dest, sliceOp.offsets(),
92       sliceOp.sizes(), sliceOp.strides(), sliceOp.static_offsets(),
93       sliceOp.static_sizes(), sliceOp.static_strides());
94 }
95 
96 template <typename LoopTy>
97 static FailureOr<TiledLinalgOp>
98 tileLinalgOpImpl(RewriterBase &b, LinalgOp op, ValueRange tileSizes,
99                  const LinalgTilingOptions &options) {
100   auto nLoops = op.getNumLoops();
101   // Initial tile sizes may be too big, only take the first nLoops.
102   tileSizes = tileSizes.take_front(nLoops);
103 
104   if (llvm::all_of(tileSizes, isZero)) {
105     TiledLinalgOp tiledOp;
106     tiledOp.op = cast<LinalgOp>(b.clone(*op.getOperation()));
107     tiledOp.tensorResults.assign(tiledOp.op->result_begin(),
108                                  tiledOp.op->result_end());
109     return tiledOp;
110   }
111 
112   // 1. Build the tiled loop ranges.
113   auto allShapeSizes = op.createFlatListOfOperandDims(b, op.getLoc());
114   AffineMap shapeSizesToLoopsMap = op.getShapesToLoopsMap();
115   if (!shapeSizesToLoopsMap)
116     return failure();
117 
118   SmallVector<Range, 4> loopRanges;
119   LoopIndexToRangeIndexMap loopIndexToRangeIndex;
120   std::tie(loopRanges, loopIndexToRangeIndex) = makeTiledLoopRanges(
121       b, op.getLoc(), shapeSizesToLoopsMap, allShapeSizes, tileSizes);
122 
123   SmallVector<Attribute, 4> iteratorTypes;
124   for (const auto &attr :
125        enumerate(op.iterator_types().cast<ArrayAttr>().getValue())) {
126     if (loopIndexToRangeIndex.count(attr.index()))
127       iteratorTypes.push_back(attr.value());
128   }
129   // If interchangeVector is empty, use the identity. Build the permutation map
130   // otherwise.
131   auto invPermutationMap =
132       AffineMap::getMultiDimIdentityMap(tileSizes.size(), b.getContext());
133   if (!options.interchangeVector.empty()) {
134     // Based on the pruned iterations (due to zero tile size), recompute the
135     // interchange vector.
136     SmallVector<unsigned, 4> interchangeVector;
137     interchangeVector.reserve(options.interchangeVector.size());
138     for (auto pos : options.interchangeVector) {
139       auto it = loopIndexToRangeIndex.find(pos);
140       if (it == loopIndexToRangeIndex.end())
141         continue;
142       interchangeVector.push_back(it->second);
143     }
144     // Interchange vector is guaranteed to be a permutation,
145     // `inversePermutation` must succeed.
146     invPermutationMap = inversePermutation(
147         AffineMap::getPermutationMap(interchangeVector, b.getContext()));
148     assert(invPermutationMap);
149     SmallVector<int64_t> permutation(interchangeVector.begin(),
150                                      interchangeVector.end());
151     applyPermutationToVector(loopRanges, permutation);
152     applyPermutationToVector(iteratorTypes, permutation);
153   }
154 
155   // 2. Create the tiled loops.
156   LinalgOp res = op;
157   SmallVector<Value, 4> ivs, tensorResults;
158   auto tiledLoopBodyBuilder =
159       [&](OpBuilder &builder, Location loc, ValueRange localIvs,
160           ValueRange operandValuesToUse) -> scf::ValueVector {
161     ivs.assign(localIvs.begin(), localIvs.end());
162 
163     // When an `interchangeVector` is present, it has been applied to the
164     // loop ranges and the iterator types. Apply its inverse to the
165     // resulting loop `ivs` to match the op definition.
166     SmallVector<Value, 4> interchangedIvs;
167     if (!options.interchangeVector.empty())
168       interchangedIvs = applyMapToValues(b, loc, invPermutationMap, ivs);
169     else
170       interchangedIvs.assign(ivs.begin(), ivs.end());
171 
172     // Tile the `operandValuesToUse` that either match the `op` operands
173     // themselves or the tile loop arguments forwarding them.
174     assert(operandValuesToUse.size() ==
175                static_cast<size_t>(op.getNumInputsAndOutputs()) &&
176            "expect the number of operands and inputs and outputs to match");
177     SmallVector<Value> valuesToTile = operandValuesToUse;
178     auto sizeBounds =
179         applyMapToValues(b, loc, shapeSizesToLoopsMap, allShapeSizes);
180     SmallVector<Value, 4> tiledOperands = makeTiledShapes(
181         b, loc, op, valuesToTile, interchangedIvs, tileSizes, sizeBounds);
182 
183     // TODO: use an interface/adaptor to avoid leaking position in
184     // `tiledOperands`.
185     SmallVector<Type, 4> resultTensorTypes;
186     for (OpOperand *opOperand : op.getOutputTensorOperands())
187       resultTensorTypes.push_back(
188           tiledOperands[opOperand->getOperandNumber()].getType());
189 
190     res = op.clone(b, loc, resultTensorTypes, tiledOperands);
191 
192     // Insert a insert_slice for each output tensor.
193     unsigned resultIdx = 0;
194     for (OpOperand *opOperand : op.getOutputTensorOperands()) {
195       // TODO: use an interface/adaptor to avoid leaking position in
196       // `tiledOperands`.
197       Value outputTensor = tiledOperands[opOperand->getOperandNumber()];
198       // TODO: Propagate RewriterBase everywhere.
199       IRRewriter rewriter(b);
200       if (auto sliceOp = outputTensor.getDefiningOp<tensor::ExtractSliceOp>()) {
201         tensorResults.push_back(insertSliceIntoTensor(rewriter, loc, sliceOp,
202                                                       res->getResult(resultIdx),
203                                                       sliceOp.source()));
204       } else {
205         tensorResults.push_back(res->getResult(resultIdx));
206       }
207       ++resultIdx;
208     }
209     return scf::ValueVector(tensorResults.begin(), tensorResults.end());
210   };
211   GenerateLoopNest<LoopTy>::doit(b, op.getLoc(), loopRanges, op, iteratorTypes,
212                                  tiledLoopBodyBuilder, options.distribution,
213                                  options.distributionTypes);
214 
215   // 3. Transform IndexOp results w.r.t. the tiling.
216   transformIndexOps(b, res, ivs, loopIndexToRangeIndex);
217 
218   // 4. Gather the newly created loops and return them with the new op.
219   SmallVector<Operation *, 8> loops;
220   loops.reserve(ivs.size());
221   for (auto iv : ivs) {
222     if (iv.isa<BlockArgument>()) {
223       loops.push_back(iv.cast<BlockArgument>().getOwner()->getParentOp());
224       assert(loops.back() && "no owner found for induction variable!");
225     } else {
226       // TODO: Instead of doing this, try to recover the ops used instead of the
227       // loop.
228       loops.push_back(nullptr);
229     }
230   }
231 
232   // 5. Get the tensor results from the outermost loop if available. Otherwise
233   // use the previously captured `tensorResults`.
234   Operation *outermostLoop = nullptr;
235   for (Operation *loop : loops)
236     if ((outermostLoop = loop))
237       break;
238 
239   return TiledLinalgOp{
240       res, loops, outermostLoop ? outermostLoop->getResults() : tensorResults};
241 }
242 
243 template <typename LoopTy>
244 FailureOr<TiledLinalgOp> static tileLinalgOpImpl(
245     RewriterBase &b, LinalgOp op, const LinalgTilingOptions &options) {
246   OpBuilder::InsertionGuard g(b);
247   b.setInsertionPoint(op);
248 
249   if (!options.tileSizeComputationFunction)
250     return failure();
251 
252   // Enforce the convention that "tiling by zero" skips tiling a particular
253   // dimension. This convention is significantly simpler to handle instead of
254   // adjusting affine maps to account for missing dimensions.
255   auto nLoops = op.getNumLoops();
256   SmallVector<Value, 4> tileSizeVector =
257       options.tileSizeComputationFunction(b, op);
258   if (tileSizeVector.size() < nLoops) {
259     auto zero = b.create<arith::ConstantIndexOp>(op.getLoc(), 0);
260     tileSizeVector.append(nLoops - tileSizeVector.size(), zero);
261   }
262 
263   return tileLinalgOpImpl<LoopTy>(b, op, tileSizeVector, options);
264 }
265 
266 FailureOr<TiledLinalgOp>
267 mlir::linalg::tileLinalgOp(RewriterBase &b, LinalgOp op,
268                            const LinalgTilingOptions &options) {
269   switch (options.loopType) {
270   case LinalgTilingLoopType::Loops:
271     return tileLinalgOpImpl<scf::ForOp>(b, op, options);
272   case LinalgTilingLoopType::ParallelLoops:
273     return tileLinalgOpImpl<scf::ParallelOp>(b, op, options);
274   default:;
275   }
276   return failure();
277 }
278 
279 /// Generate a loop nest around a given tensor::PadOp (for tiling). `newPadOp`
280 /// and `loopNest` are output parameters that return the new (tiled)
281 /// tensor::PadOp and the loop nest.
282 static LogicalResult tilePadOp(RewriterBase &builder, tensor::PadOp op,
283                                tensor::PadOp &newPadOp, LoopNest &loopNest,
284                                const LinalgTilingOptions &options) {
285   Location loc = op.getLoc();
286   OpBuilder::InsertionGuard g(builder);
287   builder.setInsertionPoint(op);
288 
289   // Clone tensor::PadOp so that the existing op can be replaced more easily.
290   newPadOp = cast<tensor::PadOp>(builder.clone(*op.getOperation()));
291   // Get rank and tile sizes.
292   int64_t rank = op.getResultType().getRank();
293   SmallVector<Value> tileSizes =
294       options.tileSizeComputationFunction(builder, op);
295   // Normalize untiled padding dimensions to 0.
296   Value zero = builder.create<arith::ConstantIndexOp>(loc, 0);
297   tileSizes.append(rank - tileSizes.size(), zero);
298   // Compute lower and upper bounds of the loop nest.
299   TilingInterface tilingInterface =
300       dyn_cast<TilingInterface>(op.getOperation());
301   SmallVector<Range> ranges = tilingInterface.getIterationDomain(builder);
302   SmallVector<Value> lbs, dims, allDims, steps;
303   for (int64_t i = 0; i < rank; ++i) {
304     allDims.push_back(ranges[i].size);
305     if (!isZero(tileSizes[i])) {
306       lbs.push_back(ranges[i].offset);
307       dims.push_back(ranges[i].size);
308       steps.push_back(tileSizes[i]);
309     }
310   }
311   // Generate loop nest: One loop per dimension.
312   SmallVector<Value> destOperand =
313       tilingInterface.getDestinationOperands(builder);
314   loopNest = mlir::scf::buildLoopNest(
315       builder, loc, lbs, /*ubs=*/dims, steps, ValueRange(destOperand),
316       [&](OpBuilder &b, Location loc, ValueRange localIvs,
317           ValueRange iterArgs) -> scf::ValueVector {
318         // Compute offsets and sizes of ExtractSliceOp.
319         SmallVector<Value> offsets =
320             computeTileOffsets(b, loc, localIvs, tileSizes);
321         SmallVector<Value> sizes =
322             computeTileSizes(b, loc, localIvs, tileSizes, allDims);
323         // Create ExtractSliceOp: Extract a tile from the tensor::PadOp.
324         // Note: The tensor::PadOp is located outside of the loop nest. It is
325         // later moved inside by ExtractSliceOfPadTensorSwapPattern.
326         auto map = AffineMap::getMultiDimIdentityMap(rank, b.getContext());
327         Value tiledOutput =
328             makeTiledShape(b, loc, newPadOp->getResult(0), tileSizes, map,
329                            offsets, allDims, sizes);
330         auto sliceOp = tiledOutput.getDefiningOp<tensor::ExtractSliceOp>();
331         assert(sliceOp && "expected ExtractSliceOp");
332         // Insert the tile into the output tensor.
333         // TODO: Propagate RewriterBase everywhere.
334         IRRewriter rewriter(b);
335         Value yieldValue =
336             insertSliceIntoTensor(rewriter, loc, sliceOp, sliceOp, iterArgs[0]);
337         return scf::ValueVector({yieldValue});
338       });
339   return success();
340 }
341 
342 namespace {
343 struct PadOpTilingPattern : public OpRewritePattern<tensor::PadOp> {
344   PadOpTilingPattern(MLIRContext *ctx, LinalgTilingOptions opt)
345       : OpRewritePattern<tensor::PadOp>(ctx), options(std::move(opt)) {}
346 
347   LogicalResult matchAndRewrite(tensor::PadOp op,
348                                 PatternRewriter &rewriter) const override {
349     if (op->hasAttr(LinalgTransforms::kLinalgTransformMarker))
350       return failure();
351     tensor::PadOp newPadOp;
352     LoopNest loopNest;
353     if (failed(tilePadOp(rewriter, op, newPadOp, loopNest, options)))
354       return failure();
355     newPadOp->setAttr(LinalgTransforms::kLinalgTransformMarker,
356                       rewriter.getUnitAttr());
357     // Replace all uses of the original tensor::PadOp.
358     rewriter.replaceOp(op, loopNest.getResults()[0]);
359     return success();
360   }
361 
362   LinalgTilingOptions options;
363 };
364 } // namespace
365 
366 namespace {
367 /// Helper classes for type list expansion.
368 template <typename... OpTypes>
369 class CanonicalizationPatternList;
370 
371 template <>
372 class CanonicalizationPatternList<> {
373 public:
374   static void insert(RewritePatternSet &patterns) {}
375 };
376 
377 template <typename OpTy, typename... OpTypes>
378 class CanonicalizationPatternList<OpTy, OpTypes...> {
379 public:
380   static void insert(RewritePatternSet &patterns) {
381     OpTy::getCanonicalizationPatterns(patterns, patterns.getContext());
382     CanonicalizationPatternList<OpTypes...>::insert(patterns);
383   }
384 };
385 } // namespace
386 
387 RewritePatternSet
388 mlir::linalg::getLinalgTilingCanonicalizationPatterns(MLIRContext *ctx) {
389   RewritePatternSet patterns(ctx);
390   populateLinalgTilingCanonicalizationPatterns(patterns);
391   return patterns;
392 }
393 
394 void mlir::linalg::populateLinalgTilingCanonicalizationPatterns(
395     RewritePatternSet &patterns) {
396   auto *ctx = patterns.getContext();
397   AffineApplyOp::getCanonicalizationPatterns(patterns, ctx);
398   AffineForOp::getCanonicalizationPatterns(patterns, ctx);
399   AffineMinOp::getCanonicalizationPatterns(patterns, ctx);
400   AffineMaxOp::getCanonicalizationPatterns(patterns, ctx);
401   arith::ConstantIndexOp::getCanonicalizationPatterns(patterns, ctx);
402 
403   memref::SubViewOp::getCanonicalizationPatterns(patterns, ctx);
404   memref::ViewOp::getCanonicalizationPatterns(patterns, ctx);
405 
406   scf::ForOp::getCanonicalizationPatterns(patterns, ctx);
407   scf::ParallelOp::getCanonicalizationPatterns(patterns, ctx);
408 
409   tensor::CastOp::getCanonicalizationPatterns(patterns, ctx);
410   tensor::ExtractSliceOp::getCanonicalizationPatterns(patterns, ctx);
411   tensor::InsertSliceOp::getCanonicalizationPatterns(patterns, ctx);
412 
413   InitTensorOp::getCanonicalizationPatterns(patterns, ctx);
414   tensor::PadOp::getCanonicalizationPatterns(patterns, ctx);
415   ctx->getLoadedDialect<LinalgDialect>()->getCanonicalizationPatterns(patterns);
416 
417   CanonicalizationPatternList<
418 #define GET_OP_LIST
419 #include "mlir/Dialect/Linalg/IR/LinalgStructuredOps.cpp.inc"
420       >::insert(patterns);
421 }
422 
423 /// Populate the given list with patterns that apply Linalg tiling.
424 static void insertTilingPatterns(RewritePatternSet &patterns,
425                                  const LinalgTilingOptions &options) {
426   auto *ctx = patterns.getContext();
427   LinalgTransformationFilter f(ArrayRef<StringAttr>{},
428                                StringAttr::get(ctx, "tiled"));
429   TilingPatterns<GenericOp,
430 #define GET_OP_LIST
431 #include "mlir/Dialect/Linalg/IR/LinalgStructuredOps.cpp.inc"
432                  >::insert(patterns, options, f);
433   patterns.add<PadOpTilingPattern>(ctx, options);
434 }
435 
436 void mlir::linalg::populatePadTensorTilingPatterns(
437     RewritePatternSet &patterns, const LinalgTilingOptions &options) {
438   auto *ctx = patterns.getContext();
439   patterns.add<PadOpTilingPattern>(ctx, options);
440 }
441 
442 static void applyExtractSliceOfPadTensorSwapPattern(FuncOp funcOp) {
443   MLIRContext *ctx = funcOp.getContext();
444   RewritePatternSet patterns(ctx);
445   patterns.add<ExtractSliceOfPadTensorSwapPattern>(patterns.getContext());
446   (void)applyPatternsAndFoldGreedily(funcOp, std::move(patterns));
447   (void)applyPatternsAndFoldGreedily(
448       funcOp, getLinalgTilingCanonicalizationPatterns(ctx));
449 }
450 
451 namespace {
452 struct LinalgTilingPass : public LinalgTilingBase<LinalgTilingPass> {
453   LinalgTilingPass() = default;
454   LinalgTilingPass(ArrayRef<int64_t> tileSizes, LinalgTilingLoopType loopType) {
455     this->tileSizes = tileSizes;
456     this->loopType = "";
457     this->loopTypeEnum = loopType;
458   }
459 
460   void runOnOperation() override {
461     FuncOp funcOp = getOperation();
462     LinalgTilingLoopType type =
463         llvm::StringSwitch<LinalgTilingLoopType>(loopType)
464             .Case("for", LinalgTilingLoopType::Loops)
465             .Case("affine", LinalgTilingLoopType::AffineLoops)
466             .Case("parallel", LinalgTilingLoopType::ParallelLoops)
467             .Default(loopTypeEnum);
468     auto options =
469         LinalgTilingOptions().setTileSizes(tileSizes).setLoopType(type);
470     MLIRContext *ctx = funcOp.getContext();
471     RewritePatternSet patterns(ctx);
472     insertTilingPatterns(patterns, options);
473     scf::populateSCFForLoopCanonicalizationPatterns(patterns);
474     (void)applyPatternsAndFoldGreedily(funcOp, std::move(patterns));
475     (void)applyPatternsAndFoldGreedily(
476         funcOp, getLinalgTilingCanonicalizationPatterns(ctx));
477     // Drop the marker.
478     funcOp.walk([](LinalgOp op) {
479       op->removeAttr(LinalgTransforms::kLinalgTransformMarker);
480     });
481 
482     // Apply swap pattern after generating loop nest and running
483     // canonicalizations.
484     applyExtractSliceOfPadTensorSwapPattern(funcOp);
485   }
486 
487   LinalgTilingLoopType loopTypeEnum;
488 };
489 
490 } // namespace
491 
492 std::unique_ptr<OperationPass<FuncOp>>
493 mlir::createLinalgTilingPass(ArrayRef<int64_t> tileSizes,
494                              linalg::LinalgTilingLoopType loopType) {
495   return std::make_unique<LinalgTilingPass>(tileSizes, loopType);
496 }
497