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 "PassDetail.h"
14 #include "mlir/Dialect/Linalg/IR/LinalgTypes.h"
15 #include "mlir/Dialect/Linalg/Passes.h"
16 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
17 #include "mlir/Dialect/Linalg/Utils/Utils.h"
18 #include "mlir/Dialect/MemRef/IR/MemRef.h"
19 #include "mlir/Dialect/Tensor/IR/Tensor.h"
20 #include "mlir/IR/AffineExpr.h"
21 #include "mlir/IR/AffineMap.h"
22 #include "mlir/Transforms/FoldUtils.h"
23 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
24 
25 #include "llvm/Support/CommandLine.h"
26 
27 using namespace mlir;
28 using namespace mlir::linalg;
29 using namespace mlir::scf;
30 
31 #define DEBUG_TYPE "linalg-tiling"
32 
33 static bool isZero(Value v) {
34   if (auto cst = v.getDefiningOp<ConstantIndexOp>())
35     return cst.getValue() == 0;
36   return false;
37 }
38 
39 using LoopIndexToRangeIndexMap = DenseMap<int, int>;
40 
41 // Creates a number of ranges equal to the number of non-zero in `tileSizes`.
42 // One for each loop of the LinalgOp that is tiled. The `tileSizes` argument has
43 // one entry per surrounding loop. It uses zero as the convention that a
44 // particular loop is not tiled. This convention simplifies implementations by
45 // avoiding affine map manipulations.
46 // The returned ranges correspond to the loop ranges, in the proper order, that
47 // are tiled and for which new loops will be created. Also the function returns
48 // a map from loop indices of the LinalgOp to the corresponding non-empty range
49 // indices of newly created loops.
50 static std::tuple<SmallVector<Range, 4>, LoopIndexToRangeIndexMap>
51 makeTiledLoopRanges(OpBuilder &b, Location loc, AffineMap map,
52                     ValueRange allShapeSizes, ValueRange allTileSizes) {
53   assert(allTileSizes.size() == map.getNumResults());
54   // Apply `map` to get shape sizes in loop order.
55   auto shapeSizes = applyMapToValues(b, loc, map, allShapeSizes);
56   SmallVector<Value, 4> tileSizes(allTileSizes.begin(), allTileSizes.end());
57 
58   // Traverse the tile sizes, which are in loop order, erase zeros everywhere.
59   LoopIndexToRangeIndexMap loopIndexToRangeIndex;
60   for (int idx = 0, e = tileSizes.size(), zerosCount = 0; idx < e; ++idx) {
61     if (isZero(tileSizes[idx - zerosCount])) {
62       shapeSizes.erase(shapeSizes.begin() + idx - zerosCount);
63       tileSizes.erase(tileSizes.begin() + idx - zerosCount);
64       ++zerosCount;
65       continue;
66     }
67     loopIndexToRangeIndex[idx] = idx - zerosCount;
68   }
69 
70   // Create a new range with the applied tile sizes.
71   SmallVector<Range, 4> res;
72   for (unsigned idx = 0, e = tileSizes.size(); idx < e; ++idx)
73     res.push_back(Range{b.create<ConstantIndexOp>(loc, 0), shapeSizes[idx],
74                         tileSizes[idx]});
75   return std::make_tuple(res, loopIndexToRangeIndex);
76 }
77 
78 // All indices returned by IndexOp should be invariant with respect to tiling.
79 // Therefore, if an operation is tiled, we have to transform the indices
80 // accordingly, i.e. offset them by the values of the corresponding induction
81 // variables that are captured implicitly in the body of the op.
82 //
83 // Example. `linalg.generic` before tiling:
84 //
85 // #id_2d = (i, j) -> (i, j)
86 // #pointwise_2d_trait = {
87 //   indexing_maps = [#id_2d, #id_2d],
88 //   iterator_types = ["parallel", "parallel"]
89 // }
90 // linalg.generic #pointwise_2d_trait %operand, %result {
91 //   ^bb0(%operand_in: f32, %result_in: f32):
92 //     %i = linalg.index 0 : index
93 //     %j = linalg.index 1 : index
94 //     <some operations that use %i, %j>
95 // }: memref<50x100xf32>, memref<50x100xf32>
96 //
97 // After tiling pass with tiles sizes 10 and 25:
98 //
99 // #strided = (i, j)[s0, s1, s2] -> (i * s1 + s0 + j * s2)
100 //
101 // %c1 = constant 1 : index
102 // %c0 = constant 0 : index
103 // %c25 = constant 25 : index
104 // %c10 = constant 10 : index
105 // operand_dim_0 = dim %operand, 0 : memref<50x100xf32>
106 // operand_dim_1 = dim %operand, 1 : memref<50x100xf32>
107 // scf.for %k = %c0 to operand_dim_0 step %c10 {
108 //   scf.for %l = %c0 to operand_dim_1 step %c25 {
109 //     %4 = std.subview %operand[%k, %l][%c10, %c25][%c1, %c1]
110 //       : memref<50x100xf32> to memref<?x?xf32, #strided>
111 //     %5 = std.subview %result[%k, %l][%c10, %c25][%c1, %c1]
112 //       : memref<50x100xf32> to memref<?x?xf32, #strided>
113 //     linalg.generic pointwise_2d_trait %4, %5 {
114 //     ^bb0(%operand_in: f32, %result_in: f32):
115 //       %i = linalg.index 0 : index
116 //       %j = linalg.index 1 : index
117 //       // Indices `k` and `l` are implicitly captured in the body.
118 //       %transformed_i = addi %i, %k : index // index `i` is offset by %k
119 //       %transformed_j = addi %j, %l : index // index `j` is offset by %l
120 //       // Every use of %i, %j is replaced with %transformed_i, %transformed_j
121 //       <some operations that use %transformed_i, %transformed_j>
122 //     }: memref<?x?xf32, #strided>, memref<?x?xf32, #strided>
123 //   }
124 // }
125 //
126 // TODO: Investigate whether mixing implicit and explicit indices
127 // does not lead to losing information.
128 static void
129 transformIndexOps(OpBuilder &b, LinalgOp op, SmallVectorImpl<Value> &ivs,
130                   const LoopIndexToRangeIndexMap &loopIndexToRangeIndex) {
131   // Skip operations that have no region attached.
132   if (op->getNumRegions() == 0)
133     return;
134   assert(op->getNumRegions() == 1 && op->getRegion(0).getBlocks().size() == 1 &&
135          "expected linalg operation to have one block.");
136   Block &block = op->getRegion(0).front();
137 
138   for (IndexOp indexOp : block.getOps<linalg::IndexOp>()) {
139     auto rangeIndex = loopIndexToRangeIndex.find(indexOp.dim());
140     if (rangeIndex == loopIndexToRangeIndex.end())
141       continue;
142     // Offset the index by the value of the corresponding induction variable and
143     // replace all uses of the previous value.
144     OpBuilder::InsertionGuard g(b);
145     b.setInsertionPointAfter(indexOp);
146     AffineExpr index, iv;
147     bindDims(b.getContext(), index, iv);
148     AffineApplyOp applyOp = b.create<AffineApplyOp>(
149         indexOp.getLoc(), index + iv,
150         ValueRange{indexOp.getResult(), ivs[rangeIndex->second]});
151     indexOp.getResult().replaceAllUsesExcept(applyOp, applyOp);
152   }
153 }
154 
155 // Insert a tile `source` into the destination tensor `dest`. The position at
156 // which the tile is inserted (as well as size of tile) is taken from a given
157 // ExtractSliceOp `sliceOp`.
158 static Value insertSliceIntoTensor(OpBuilder &b, Location loc,
159                                    tensor::ExtractSliceOp sliceOp, Value source,
160                                    Value dest) {
161   return b.create<tensor::InsertSliceOp>(
162       loc, sliceOp.source().getType(), source, dest, sliceOp.offsets(),
163       sliceOp.sizes(), sliceOp.strides(), sliceOp.static_offsets(),
164       sliceOp.static_sizes(), sliceOp.static_strides());
165 }
166 
167 template <typename LoopTy>
168 static Optional<TiledLinalgOp>
169 tileLinalgOpImpl(OpBuilder &b, LinalgOp op, ValueRange tileSizes,
170                  const LinalgTilingOptions &options) {
171   auto nLoops = op.getNumLoops();
172   // Initial tile sizes may be too big, only take the first nLoops.
173   tileSizes = tileSizes.take_front(nLoops);
174 
175   if (llvm::all_of(tileSizes, isZero))
176     return llvm::None;
177 
178   if (auto convOp = dyn_cast<linalg::ConvOp>(op.getOperation())) {
179     // For conv op only support tiling along batch dimension (which is the first
180     // loop).
181     if (convOp.padding() && !llvm::all_of(tileSizes.drop_front(), isZero))
182       return llvm::None;
183   }
184 
185   // 1. Build the tiled loop ranges.
186   auto allShapeSizes = op.createFlatListOfOperandDims(b, op.getLoc());
187   AffineMap shapeSizesToLoopsMap = op.getShapesToLoopsMap();
188   if (!shapeSizesToLoopsMap)
189     return llvm::None;
190 
191   SmallVector<Range, 4> loopRanges;
192   LoopIndexToRangeIndexMap loopIndexToRangeIndex;
193   std::tie(loopRanges, loopIndexToRangeIndex) = makeTiledLoopRanges(
194       b, op.getLoc(), shapeSizesToLoopsMap, allShapeSizes, tileSizes);
195 
196   SmallVector<Attribute, 4> iteratorTypes;
197   for (auto attr :
198        enumerate(op.iterator_types().cast<ArrayAttr>().getValue())) {
199     if (loopIndexToRangeIndex.count(attr.index()))
200       iteratorTypes.push_back(attr.value());
201   }
202   // If interchangeVector is empty, use the identity. Build the permutation map
203   // otherwise.
204   auto invPermutationMap =
205       AffineMap::getMultiDimIdentityMap(tileSizes.size(), b.getContext());
206   if (!options.interchangeVector.empty()) {
207     // Based on the pruned iterations (due to zero tile size), recompute the
208     // interchange vector.
209     SmallVector<unsigned, 4> interchangeVector;
210     interchangeVector.reserve(options.interchangeVector.size());
211     for (auto pos : options.interchangeVector) {
212       auto it = loopIndexToRangeIndex.find(pos);
213       if (it == loopIndexToRangeIndex.end())
214         continue;
215       interchangeVector.push_back(it->second);
216     }
217     // Interchange vector is guaranteed to be a permutation,
218     // `inversePermutation` must succeed.
219     invPermutationMap = inversePermutation(
220         AffineMap::getPermutationMap(interchangeVector, b.getContext()));
221     assert(invPermutationMap);
222     applyPermutationToVector(loopRanges, interchangeVector);
223     applyPermutationToVector(iteratorTypes, interchangeVector);
224   }
225 
226   // 2. Create the tiled loops.
227   LinalgOp res = op;
228   SmallVector<Value, 4> ivs, tensorResults;
229   auto tiledLoopBodyBuilder = [&](OpBuilder &b, Location loc,
230                                   ValueRange localIvs,
231                                   ValueRange iterArgs) -> scf::ValueVector {
232     ivs.assign(localIvs.begin(), localIvs.end());
233 
234     // When an `interchangeVector` is present, it has been applied to the
235     // loop ranges and the iterator types. Apply its inverse to the
236     // resulting loop `ivs` to match the op definition.
237     SmallVector<Value, 4> interchangedIvs;
238     if (!options.interchangeVector.empty())
239       interchangedIvs = applyMapToValues(b, loc, invPermutationMap, ivs);
240     else
241       interchangedIvs.assign(ivs.begin(), ivs.end());
242 
243     assert(op.getOutputTensorOperands().size() == iterArgs.size() &&
244            "num output tensors must match number of loop iter arguments");
245 
246     SmallVector<Value> operands = op.getInputOperands();
247     SmallVector<Value> outputBuffers = op.getOutputBufferOperands();
248     // TODO: thanks to simplifying assumption we do not need to worry about
249     // order of output buffers and tensors: there is only ever one kind.
250     assert(outputBuffers.empty() || iterArgs.empty());
251     operands.append(outputBuffers.begin(), outputBuffers.end());
252     operands.append(iterArgs.begin(), iterArgs.end());
253     auto sizeBounds =
254         applyMapToValues(b, loc, shapeSizesToLoopsMap, allShapeSizes);
255     SmallVector<Value, 4> tiledOperands = makeTiledShapes(
256         b, loc, op, operands, interchangedIvs, tileSizes, sizeBounds);
257 
258     // TODO: use an interface/adaptor to avoid leaking position in
259     // `tiledOperands`.
260     SmallVector<Type, 4> resultTensorTypes;
261     for (OpOperand *opOperand : op.getOutputTensorOperands())
262       resultTensorTypes.push_back(
263           tiledOperands[opOperand->getOperandNumber()].getType());
264 
265     res = op.clone(b, loc, resultTensorTypes, tiledOperands);
266 
267     // Insert a insert_slice for each output tensor.
268     unsigned resultIdx = 0;
269     for (OpOperand *opOperand : op.getOutputTensorOperands()) {
270       // TODO: use an interface/adaptor to avoid leaking position in
271       // `tiledOperands`.
272       Value outputTensor = tiledOperands[opOperand->getOperandNumber()];
273       if (auto sliceOp = outputTensor.getDefiningOp<tensor::ExtractSliceOp>()) {
274         tensorResults.push_back(insertSliceIntoTensor(
275             b, loc, sliceOp, res->getResult(resultIdx), sliceOp.source()));
276       } else {
277         tensorResults.push_back(res->getResult(resultIdx));
278       }
279       ++resultIdx;
280     }
281     return scf::ValueVector(tensorResults.begin(), tensorResults.end());
282   };
283   GenerateLoopNest<LoopTy>::doit(b, op.getLoc(), loopRanges, op, iteratorTypes,
284                                  tiledLoopBodyBuilder, options.distribution,
285                                  options.distributionTypes);
286 
287   // 3. Transform IndexOp results w.r.t. the tiling.
288   transformIndexOps(b, res, ivs, loopIndexToRangeIndex);
289 
290   // 4. Gather the newly created loops and return them with the new op.
291   SmallVector<Operation *, 8> loops;
292   loops.reserve(ivs.size());
293   for (auto iv : ivs) {
294     if (iv.isa<BlockArgument>()) {
295       loops.push_back(iv.cast<BlockArgument>().getOwner()->getParentOp());
296       assert(loops.back() && "no owner found for induction variable!");
297     } else {
298       // TODO: Instead of doing this, try to recover the ops used instead of the
299       // loop.
300       loops.push_back(nullptr);
301     }
302   }
303 
304   // 5. Get the tensor results from the outermost loop if available. Otherwise
305   // use the previously captured `tensorResults`.
306   Operation *outermostLoop = nullptr;
307   for (Operation *loop : loops)
308     if ((outermostLoop = loop))
309       break;
310 
311   return TiledLinalgOp{
312       res, loops, outermostLoop ? outermostLoop->getResults() : tensorResults};
313 }
314 
315 template <typename LoopTy>
316 Optional<TiledLinalgOp> static tileLinalgOpImpl(
317     OpBuilder &b, LinalgOp op, const LinalgTilingOptions &options) {
318   OpBuilder::InsertionGuard g(b);
319   b.setInsertionPoint(op);
320 
321   if (!options.tileSizeComputationFunction)
322     return llvm::None;
323 
324   // Enforce the convention that "tiling by zero" skips tiling a particular
325   // dimension. This convention is significantly simpler to handle instead of
326   // adjusting affine maps to account for missing dimensions.
327   auto nLoops = op.getNumLoops();
328   SmallVector<Value, 4> tileSizeVector =
329       options.tileSizeComputationFunction(b, op);
330   if (tileSizeVector.size() < nLoops) {
331     auto zero = b.create<ConstantIndexOp>(op.getLoc(), 0);
332     tileSizeVector.append(nLoops - tileSizeVector.size(), zero);
333   }
334 
335   return tileLinalgOpImpl<LoopTy>(b, op, tileSizeVector, options);
336 }
337 
338 Optional<TiledLinalgOp>
339 mlir::linalg::tileLinalgOp(OpBuilder &b, LinalgOp op,
340                            const LinalgTilingOptions &options) {
341   switch (options.loopType) {
342   case LinalgTilingLoopType::Loops:
343     return tileLinalgOpImpl<scf::ForOp>(b, op, options);
344   case LinalgTilingLoopType::ParallelLoops:
345     return tileLinalgOpImpl<scf::ParallelOp>(b, op, options);
346   case LinalgTilingLoopType::TiledLoops:
347     return tileLinalgOpImpl<linalg::TiledLoopOp>(b, op, options);
348   default:;
349   }
350   return llvm::None;
351 }
352 
353 /// Generate a loop nest around a given PadTensorOp (for tiling). `newPadOp`
354 /// and `loopNest` are output parameters that return the new (tiled) PadTensorOp
355 /// and the loop nest.
356 static LogicalResult tilePadTensorOp(OpBuilder &builder, PadTensorOp op,
357                                      PadTensorOp &newPadOp, LoopNest &loopNest,
358                                      const LinalgTilingOptions &options) {
359   // Can tile only PadTensorOp that have an output operand.
360   if (!op.output())
361     return failure();
362 
363   Location loc = op.getLoc();
364   OpBuilder::InsertionGuard g(builder);
365   builder.setInsertionPoint(op);
366 
367   // Clone PadTensorOp so that the existing op can be replaced more easily.
368   newPadOp = cast<PadTensorOp>(builder.clone(*op.getOperation()));
369   // Get rank and tile sizes.
370   int64_t rank = op.getResultType().getRank();
371   SmallVector<Value> tileSizes =
372       options.tileSizeComputationFunction(builder, op);
373   assert(static_cast<int64_t>(tileSizes.size()) == rank);
374   // Compute lower and upper bounds of the loop nest.
375   SmallVector<Value> lbs, dims, steps;
376   for (int64_t i = 0; i < rank; ++i) {
377     if (!isZero(tileSizes[i])) {
378       lbs.push_back(builder.create<ConstantIndexOp>(loc, 0));
379       dims.push_back(builder.create<tensor::DimOp>(loc, op.output(), i));
380       steps.push_back(tileSizes[i]);
381     }
382   }
383   // Generate loop nest: One loop per dimension.
384   loopNest = mlir::scf::buildLoopNest(
385       builder, loc, lbs, /*ubs=*/dims, steps, ValueRange(op.output()),
386       [&](OpBuilder &b, Location loc, ValueRange localIvs,
387           ValueRange iterArgs) -> scf::ValueVector {
388         // Compute offsets and sizes of ExtractSliceOp.
389         SmallVector<Value> offsets =
390             computeTileOffsets(b, loc, localIvs, tileSizes);
391         SmallVector<Value> sizes =
392             computeTileSizes(b, loc, localIvs, tileSizes, dims);
393         // Create ExtractSliceOp: Extract a tile from the PadTensorOp.
394         // Note: The PadTensorOp is located outside of the loop nest. It is
395         // later moved inside by ExtractSliceOfPadTensorSwapPattern.
396         auto map = AffineMap::getMultiDimIdentityMap(rank, b.getContext());
397         Value tiledOutput = makeTiledShape(b, loc, newPadOp->getResult(0),
398                                            tileSizes, map, offsets, sizes);
399         auto sliceOp = tiledOutput.getDefiningOp<tensor::ExtractSliceOp>();
400         assert(sliceOp && "expected ExtractSliceOp");
401         // Insert the tile into the output tensor.
402         Value yieldValue =
403             insertSliceIntoTensor(b, loc, sliceOp, sliceOp, iterArgs[0]);
404         return scf::ValueVector({yieldValue});
405       });
406   return success();
407 }
408 
409 namespace {
410 struct PadTensorOpTilingPattern : public OpRewritePattern<PadTensorOp> {
411   PadTensorOpTilingPattern(MLIRContext *ctx, LinalgTilingOptions opt)
412       : OpRewritePattern<PadTensorOp>(ctx), options(opt) {}
413 
414   LogicalResult matchAndRewrite(PadTensorOp op,
415                                 PatternRewriter &rewriter) const override {
416     if (op->hasAttr(LinalgTransforms::kLinalgTransformMarker))
417       return failure();
418     PadTensorOp newPadOp;
419     LoopNest loopNest;
420     if (failed(tilePadTensorOp(rewriter, op, newPadOp, loopNest, options)))
421       return failure();
422     newPadOp->setAttr(LinalgTransforms::kLinalgTransformMarker,
423                       rewriter.getUnitAttr());
424     // Replace all uses of the original PadTensorOp.
425     rewriter.replaceOp(op, loopNest.getResults()[0]);
426     return success();
427   }
428 
429   LinalgTilingOptions options;
430 };
431 } // namespace
432 
433 namespace {
434 /// Helper classes for type list expansion.
435 template <typename... OpTypes>
436 class CanonicalizationPatternList;
437 
438 template <>
439 class CanonicalizationPatternList<> {
440 public:
441   static void insert(RewritePatternSet &patterns) {}
442 };
443 
444 template <typename OpTy, typename... OpTypes>
445 class CanonicalizationPatternList<OpTy, OpTypes...> {
446 public:
447   static void insert(RewritePatternSet &patterns) {
448     OpTy::getCanonicalizationPatterns(patterns, patterns.getContext());
449     CanonicalizationPatternList<OpTypes...>::insert(patterns);
450   }
451 };
452 
453 /// Helper classes for type list expansion.
454 template <typename... OpTypes>
455 class RewritePatternList;
456 
457 template <>
458 class RewritePatternList<> {
459 public:
460   static void insert(RewritePatternSet &patterns,
461                      const LinalgTilingOptions &options) {}
462 };
463 
464 template <typename OpTy, typename... OpTypes>
465 class RewritePatternList<OpTy, OpTypes...> {
466 public:
467   static void insert(RewritePatternSet &patterns,
468                      const LinalgTilingOptions &options) {
469     auto *ctx = patterns.getContext();
470     patterns.add<LinalgTilingPattern<OpTy>>(
471         ctx, options,
472         LinalgTransformationFilter(ArrayRef<Identifier>{},
473                                    Identifier::get("tiled", ctx)));
474     RewritePatternList<OpTypes...>::insert(patterns, options);
475   }
476 };
477 } // namespace
478 
479 RewritePatternSet
480 mlir::linalg::getLinalgTilingCanonicalizationPatterns(MLIRContext *ctx) {
481   RewritePatternSet patterns(ctx);
482   populateLinalgTilingCanonicalizationPatterns(patterns);
483   return patterns;
484 }
485 
486 void mlir::linalg::populateLinalgTilingCanonicalizationPatterns(
487     RewritePatternSet &patterns) {
488   auto *ctx = patterns.getContext();
489   AffineApplyOp::getCanonicalizationPatterns(patterns, ctx);
490   AffineForOp::getCanonicalizationPatterns(patterns, ctx);
491   AffineMinOp::getCanonicalizationPatterns(patterns, ctx);
492   AffineMaxOp::getCanonicalizationPatterns(patterns, ctx);
493   scf::ForOp::getCanonicalizationPatterns(patterns, ctx);
494   scf::ParallelOp::getCanonicalizationPatterns(patterns, ctx);
495   ConstantIndexOp::getCanonicalizationPatterns(patterns, ctx);
496   tensor::ExtractSliceOp::getCanonicalizationPatterns(patterns, ctx);
497   tensor::InsertSliceOp::getCanonicalizationPatterns(patterns, ctx);
498   memref::SubViewOp::getCanonicalizationPatterns(patterns, ctx);
499   tensor::CastOp::getCanonicalizationPatterns(patterns, ctx);
500   memref::ViewOp::getCanonicalizationPatterns(patterns, ctx);
501   PadTensorOp::getCanonicalizationPatterns(patterns, ctx);
502   ctx->getLoadedDialect<LinalgDialect>()->getCanonicalizationPatterns(patterns);
503   CanonicalizationPatternList<
504 #define GET_OP_LIST
505 #include "mlir/Dialect/Linalg/IR/LinalgStructuredOps.cpp.inc"
506       >::insert(patterns);
507 }
508 
509 /// Populate the given list with patterns that apply Linalg tiling.
510 static void insertTilingPatterns(RewritePatternSet &patterns,
511                                  const LinalgTilingOptions &options) {
512   RewritePatternList<GenericOp,
513 #define GET_OP_LIST
514 #include "mlir/Dialect/Linalg/IR/LinalgStructuredOps.cpp.inc"
515                      >::insert(patterns, options);
516   patterns.add<PadTensorOpTilingPattern>(patterns.getContext(), options);
517 }
518 
519 static void applyExtractSliceOfPadTensorSwapPattern(FuncOp funcOp) {
520   MLIRContext *ctx = funcOp.getContext();
521   RewritePatternSet patterns(ctx);
522   patterns.add<ExtractSliceOfPadTensorSwapPattern>(patterns.getContext());
523   (void)applyPatternsAndFoldGreedily(funcOp, std::move(patterns));
524   (void)applyPatternsAndFoldGreedily(
525       funcOp, getLinalgTilingCanonicalizationPatterns(ctx));
526 }
527 
528 static void
529 applyTilingToLoopPatterns(LinalgTilingLoopType loopType, FuncOp funcOp,
530                           ArrayRef<int64_t> tileSizes,
531                           ArrayRef<StringRef> distributionTypes = {}) {
532   auto options = LinalgTilingOptions()
533                      .setTileSizes(tileSizes)
534                      .setLoopType(loopType)
535                      .setDistributionTypes(distributionTypes);
536   MLIRContext *ctx = funcOp.getContext();
537   RewritePatternSet patterns(ctx);
538   insertTilingPatterns(patterns, options);
539   patterns.add<AffineMinSCFCanonicalizationPattern>(patterns.getContext());
540   (void)applyPatternsAndFoldGreedily(funcOp, std::move(patterns));
541   (void)applyPatternsAndFoldGreedily(
542       funcOp, getLinalgTilingCanonicalizationPatterns(ctx));
543   // Drop the marker.
544   funcOp.walk([](LinalgOp op) {
545     op->removeAttr(LinalgTransforms::kLinalgTransformMarker);
546   });
547 
548   // Apply swap pattern after generating loop nest and running
549   // canonicalizations.
550   applyExtractSliceOfPadTensorSwapPattern(funcOp);
551 }
552 
553 namespace {
554 struct LinalgTilingPass : public LinalgTilingBase<LinalgTilingPass> {
555   LinalgTilingPass() = default;
556   LinalgTilingPass(ArrayRef<int64_t> sizes) { tileSizes = sizes; }
557 
558   void runOnFunction() override {
559     applyTilingToLoopPatterns(LinalgTilingLoopType::Loops, getFunction(),
560                               tileSizes);
561   }
562 };
563 
564 struct LinalgTilingToParallelLoopsPass
565     : public LinalgTilingToParallelLoopsBase<LinalgTilingToParallelLoopsPass> {
566   LinalgTilingToParallelLoopsPass() = default;
567   LinalgTilingToParallelLoopsPass(ArrayRef<int64_t> sizes) {
568     tileSizes = sizes;
569   }
570 
571   void runOnFunction() override {
572     applyTilingToLoopPatterns(LinalgTilingLoopType::ParallelLoops,
573                               getFunction(), tileSizes);
574   }
575 };
576 
577 struct LinalgTilingToTiledLoopsPass
578     : public LinalgTilingToTiledLoopsBase<LinalgTilingToTiledLoopsPass> {
579   LinalgTilingToTiledLoopsPass() = default;
580   LinalgTilingToTiledLoopsPass(ArrayRef<int64_t> sizes,
581                                ArrayRef<StringRef> types) {
582     tileSizes = sizes;
583     distributionTypes = llvm::to_vector<2>(
584         llvm::map_range(types, [](StringRef ref) { return ref.str(); }));
585   }
586 
587   void runOnFunction() override {
588     applyTilingToLoopPatterns(
589         LinalgTilingLoopType::TiledLoops, getFunction(), tileSizes,
590         llvm::to_vector<2>(
591             llvm::map_range(distributionTypes,
592                             [](std::string &str) { return StringRef(str); })));
593   }
594 };
595 
596 } // namespace
597 
598 std::unique_ptr<OperationPass<FuncOp>>
599 mlir::createLinalgTilingPass(ArrayRef<int64_t> tileSizes) {
600   return std::make_unique<LinalgTilingPass>(tileSizes);
601 }
602 
603 std::unique_ptr<OperationPass<FuncOp>>
604 mlir::createLinalgTilingToParallelLoopsPass(ArrayRef<int64_t> tileSizes) {
605   return std::make_unique<LinalgTilingToParallelLoopsPass>(tileSizes);
606 }
607 
608 std::unique_ptr<OperationPass<FuncOp>>
609 mlir::createLinalgTilingToTiledLoopPass(ArrayRef<int64_t> tileSizes,
610                                         ArrayRef<StringRef> distributionTypes) {
611   return std::make_unique<LinalgTilingToTiledLoopsPass>(tileSizes,
612                                                         distributionTypes);
613 }
614