1 //===- Utils.cpp - Utilities to support the Linalg dialect ----------------===//
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 utilities for the Linalg dialect.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "mlir/Dialect/Linalg/Utils/Utils.h"
14 
15 #include "mlir/Analysis/AffineStructures.h"
16 #include "mlir/Analysis/SliceAnalysis.h"
17 #include "mlir/Dialect/Affine/IR/AffineOps.h"
18 #include "mlir/Dialect/Affine/IR/AffineValueMap.h"
19 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
20 #include "mlir/Dialect/Linalg/IR/LinalgOps.h"
21 #include "mlir/Dialect/Linalg/IR/LinalgTypes.h"
22 #include "mlir/Dialect/MemRef/IR/MemRef.h"
23 #include "mlir/Dialect/SCF/SCF.h"
24 #include "mlir/Dialect/StandardOps/IR/Ops.h"
25 #include "mlir/Dialect/StandardOps/Utils/Utils.h"
26 #include "mlir/Dialect/Tensor/IR/Tensor.h"
27 #include "mlir/Dialect/Utils/StaticValueUtils.h"
28 #include "mlir/IR/AffineExpr.h"
29 #include "mlir/IR/AffineExprVisitor.h"
30 #include "mlir/IR/AffineMap.h"
31 #include "mlir/IR/Matchers.h"
32 #include "mlir/IR/OpImplementation.h"
33 #include "mlir/Pass/Pass.h"
34 #include "mlir/Transforms/LoopUtils.h"
35 #include "llvm/ADT/TypeSwitch.h"
36 #include "llvm/Support/Debug.h"
37 
38 #define DEBUG_TYPE "linalg-utils"
39 
40 using namespace mlir;
41 using namespace mlir::linalg;
42 using namespace mlir::scf;
43 
44 static bool isZero(Value v) {
45   if (auto cst = v.getDefiningOp<arith::ConstantIndexOp>())
46     return cst.value() == 0;
47   return false;
48 }
49 
50 namespace {
51 
52 // Helper visitor to determine whether an AffineExpr is tiled.
53 // This is achieved by traversing every AffineDimExpr with position `pos` and
54 // checking whether the corresponding `tileSizes[pos]` is non-zero.
55 // This also enforces only positive coefficients occur in multiplications.
56 //
57 // Example:
58 //   `d0 + 2 * d1 + d3` is tiled by [0, 0, 0, 2] but not by [0, 0, 2, 0]
59 //
60 struct TileCheck : public AffineExprVisitor<TileCheck> {
61   TileCheck(ValueRange tileSizes) : isTiled(false), tileSizes(tileSizes) {}
62 
63   void visitDimExpr(AffineDimExpr expr) {
64     isTiled |= !isZero(tileSizes[expr.getPosition()]);
65   }
66   void visitAffineBinaryOpExpr(AffineBinaryOpExpr expr) {
67     visit(expr.getLHS());
68     visit(expr.getRHS());
69     if (expr.getKind() == mlir::AffineExprKind::Mul)
70       assert(expr.getRHS().cast<AffineConstantExpr>().getValue() > 0 &&
71              "nonpositive multiplying coefficient");
72   }
73   bool isTiled;
74   ValueRange tileSizes;
75 };
76 
77 } // namespace
78 
79 static bool isTiled(AffineExpr expr, ValueRange tileSizes) {
80   if (!expr)
81     return false;
82   TileCheck t(tileSizes);
83   t.visit(expr);
84   return t.isTiled;
85 }
86 
87 // Checks whether the `map  varies with respect to a non-zero `tileSize`.
88 static bool isTiled(AffineMap map, ValueRange tileSizes) {
89   if (!map)
90     return false;
91   for (unsigned r = 0; r < map.getNumResults(); ++r)
92     if (isTiled(map.getResult(r), tileSizes))
93       return true;
94   return false;
95 }
96 
97 Optional<RegionMatcher::BinaryOpKind>
98 RegionMatcher::matchAsScalarBinaryOp(GenericOp op) {
99   auto &region = op.region();
100   if (!llvm::hasSingleElement(region))
101     return llvm::None;
102 
103   Block &block = region.front();
104   if (block.getNumArguments() != 2 ||
105       !block.getArgument(0).getType().isSignlessIntOrFloat() ||
106       !block.getArgument(1).getType().isSignlessIntOrFloat())
107     return llvm::None;
108 
109   auto &ops = block.getOperations();
110   if (!llvm::hasSingleElement(block.without_terminator()))
111     return llvm::None;
112 
113   using mlir::matchers::m_Val;
114   auto a = m_Val(block.getArgument(0));
115   auto b = m_Val(block.getArgument(1));
116 
117   auto addPattern = m_Op<linalg::YieldOp>(m_Op<arith::AddIOp>(a, b));
118   if (addPattern.match(&ops.back()))
119     return BinaryOpKind::IAdd;
120 
121   return llvm::None;
122 }
123 
124 /// Explicit instantiation of loop nest generator for different loop types.
125 template struct mlir::linalg::GenerateLoopNest<scf::ForOp>;
126 template struct mlir::linalg::GenerateLoopNest<scf::ParallelOp>;
127 template struct mlir::linalg::GenerateLoopNest<AffineForOp>;
128 template struct mlir::linalg::GenerateLoopNest<TiledLoopOp>;
129 
130 /// Given a list of subview ranges, extract individual values for lower, upper
131 /// bounds and steps and put them into the corresponding vectors.
132 static void unpackRanges(ArrayRef<Range> ranges, SmallVectorImpl<Value> &lbs,
133                          SmallVectorImpl<Value> &ubs,
134                          SmallVectorImpl<Value> &steps) {
135   for (Range range : ranges) {
136     lbs.emplace_back(range.offset);
137     ubs.emplace_back(range.size);
138     steps.emplace_back(range.stride);
139   }
140 }
141 
142 namespace mlir {
143 namespace linalg {
144 
145 bool isPermutation(ArrayRef<int64_t> permutation) {
146   // Count the number of appearances for all indices.
147   SmallVector<int64_t> indexCounts(permutation.size(), 0);
148   for (auto index : permutation) {
149     // Exit if the index is out-of-range.
150     if (index < 0 || index >= static_cast<int64_t>(permutation.size()))
151       return false;
152     indexCounts[index]++;
153   }
154   // Return true if all indices appear once.
155   return count(indexCounts, 1) == static_cast<int64_t>(permutation.size());
156 }
157 
158 /// Helper function that creates a memref::DimOp or tensor::DimOp depending on
159 /// the type of `source`.
160 Value createOrFoldDimOp(OpBuilder &b, Location loc, Value source, int64_t dim) {
161   if (source.getType().isa<UnrankedMemRefType, MemRefType>())
162     return b.createOrFold<memref::DimOp>(loc, source, dim);
163   if (source.getType().isa<UnrankedTensorType, RankedTensorType>())
164     return b.createOrFold<tensor::DimOp>(loc, source, dim);
165   llvm_unreachable("Expected MemRefType or TensorType");
166 }
167 
168 /// Given an operation, retrieves the value of each dynamic dimension through
169 /// constructing the necessary DimOp operators.
170 SmallVector<Value, 4> getDynOperands(Location loc, Value val, OpBuilder &b) {
171   SmallVector<Value, 4> dynOperands;
172   auto shapedType = val.getType().cast<ShapedType>();
173   for (auto dim : llvm::enumerate(shapedType.getShape())) {
174     if (dim.value() == ShapedType::kDynamicSize)
175       dynOperands.push_back(createOrFoldDimOp(b, loc, val, dim.index()));
176   }
177   return dynOperands;
178 }
179 
180 void getUpperBoundForIndex(Value value, AffineMap &boundMap,
181                            SmallVectorImpl<Value> &boundOperands) {
182   // Initialize `boundMap` and `boundOperands` to the identity returning
183   // `value`. This combination is the default result of the method if no
184   // simplification is possible.
185   assert(value.getType().isIndex() && "expect value to have index type");
186   boundMap = AffineMap::getMultiDimIdentityMap(1, value.getContext());
187   boundOperands.assign({value});
188   canonicalizeMapAndOperands(&boundMap, &boundOperands);
189 
190   // Continue only if there is an affine index computation to simplify.
191   Operation *definingOp = value.getDefiningOp();
192   if (!definingOp || !isa<AffineApplyOp, AffineMinOp>(definingOp))
193     return;
194 
195   // Get the backward slice containing the affine index computation.
196   SetVector<Operation *> backwardSlice;
197   getBackwardSlice(definingOp, &backwardSlice, [](Operation *op) {
198     return isa<AffineApplyOp, AffineMinOp>(op);
199   });
200   backwardSlice.insert(definingOp);
201 
202   // Setup a system of affine constraints that describe the index computation.
203   FlatAffineValueConstraints constraints;
204 
205   // Helper to find or create an identifier for the given value.
206   auto findOrCreateId = [&](Value value) {
207     if (!constraints.containsId(value)) {
208       constraints.appendDimId(value);
209       return true;
210     }
211     unsigned pos;
212     constraints.findId(value, &pos);
213     return pos < constraints.getNumDimIds();
214   };
215   // Helper to get the position for the given value.
216   auto getPosition = [&](Value value) {
217     unsigned pos;
218     bool exists = constraints.findId(value, &pos);
219     (void)exists;
220     assert(exists && "expect to find the identifier");
221     return pos;
222   };
223 
224   // Add the affine operations in `backwardSlice` to the constraints.
225   for (Operation *op : llvm::reverse(backwardSlice)) {
226     // Add an identifier for all op results and operands.
227     if (!(llvm::all_of(op->getResults(), findOrCreateId) &&
228           llvm::all_of(op->getOperands(), findOrCreateId)))
229       return;
230     // Add AffineApplyOps to the constraints.
231     if (auto applyOp = dyn_cast<AffineApplyOp>(op)) {
232       AffineValueMap valueMap(applyOp.getAffineMap(), applyOp.getOperands(),
233                               applyOp.getResult());
234       if (failed(constraints.composeMap(&valueMap)))
235         return;
236       continue;
237     }
238     // Add AffineMinOps to the constraints.
239     auto minOp = cast<AffineMinOp>(op);
240     AffineMap map = constraints.computeAlignedMap(minOp.getAffineMap(),
241                                                   minOp.getOperands());
242     if (failed(constraints.addBound(FlatAffineConstraints::UB,
243                                     getPosition(minOp.getResult()), map)))
244       return;
245   }
246 
247   // Obtain an upper bound for the affine index computation by projecting out
248   // all temporary results and expressing the upper bound for `value` in terms
249   // of the terminals of the index computation.
250   SmallVector<AffineMap> lowerBounds(1), upperBounds(1);
251   constraints.getSliceBounds(getPosition(value), 1, value.getContext(),
252                              &lowerBounds, &upperBounds);
253 
254   // Verify `upperBounds[0]` is valid and has at least one result.
255   if (!upperBounds[0] || upperBounds[0].getNumResults() == 0)
256     return;
257 
258   // Set `boundMap` and `boundOperands` to the computed upper bound.
259   boundMap = upperBounds[0];
260   constraints.getAllValues(&boundOperands);
261   erase_value(boundOperands, value);
262   canonicalizeMapAndOperands(&boundMap, &boundOperands);
263 }
264 
265 FailureOr<int64_t> getConstantUpperBoundForIndex(Value value) {
266   // Compute an upper bound for `value`.
267   AffineMap boundMap;
268   SmallVector<Value> boundOperands;
269   getUpperBoundForIndex(value, boundMap, boundOperands);
270 
271   // Search the results of `boundMap` for constant upper bounds.
272   SmallVector<int64_t> constantBounds;
273   for (AffineExpr result : boundMap.getResults())
274     if (auto constExpr = result.dyn_cast<AffineConstantExpr>())
275       constantBounds.push_back(constExpr.getValue());
276 
277   // Return the minimal upper bound or failure if none is found.
278   if (constantBounds.empty())
279     return failure();
280   return *std::min_element(constantBounds.begin(), constantBounds.end());
281 }
282 
283 tensor::ExtractSliceOp makeComposedExtractSliceOp(
284     OpBuilder &b, Location loc, Value source, ArrayRef<OpFoldResult> offsets,
285     ArrayRef<OpFoldResult> sizes, ArrayRef<OpFoldResult> strides) {
286   assert(source && "expect source to be nonzero");
287 
288   // Do not fold if the producer is not an ExtractSliceOp.
289   auto producerOp = source.getDefiningOp<tensor::ExtractSliceOp>();
290   if (!producerOp)
291     return b.create<tensor::ExtractSliceOp>(loc, source, offsets, sizes,
292                                             strides);
293 
294   // Do not fold if the producer is rank reducing or if there are any non-unit
295   // strides. Supporting non-unit strides complicates the offset computation
296   // since the consumer offsets need to be multiplied by the producer strides.
297   // TODO: support non-unit strides once there are use cases.
298   SmallVector<OpFoldResult> allStrides = producerOp.getMixedStrides();
299   allStrides.append(strides.begin(), strides.end());
300   bool hasNonUnitStride = any_of(allStrides, [](OpFoldResult ofr) {
301     return getConstantIntValue(ofr) != static_cast<int64_t>(1);
302   });
303   if (hasNonUnitStride ||
304       producerOp.getSourceType().getRank() !=
305           producerOp.getResult().getType().cast<ShapedType>().getRank())
306     return b.create<tensor::ExtractSliceOp>(loc, source, offsets, sizes,
307                                             strides);
308 
309   // Fold the producer by adding the offests and extracting the slice directly
310   // from the producer source tensor.
311   SmallVector<OpFoldResult> foldedOffsets(offsets.begin(), offsets.end());
312   AffineExpr dim1, dim2;
313   bindDims(b.getContext(), dim1, dim2);
314   for (auto en : enumerate(producerOp.getMixedOffsets())) {
315     SmallVector<Value> offsetValues = {
316         getValueOrCreateConstantIndexOp(b, loc, foldedOffsets[en.index()]),
317         getValueOrCreateConstantIndexOp(b, loc, en.value())};
318     foldedOffsets[en.index()] =
319         makeComposedAffineApply(b, loc, dim1 + dim2, offsetValues).getResult();
320   }
321   return b.create<tensor::ExtractSliceOp>(loc, producerOp.source(),
322                                           foldedOffsets, sizes, strides);
323 }
324 
325 /// Specialization to build an scf "for" nest.
326 template <>
327 void GenerateLoopNest<scf::ForOp>::doit(
328     OpBuilder &b, Location loc, ArrayRef<Range> loopRanges, LinalgOp linalgOp,
329     ArrayRef<Attribute> iteratorTypes,
330     function_ref<scf::ValueVector(OpBuilder &, Location, ValueRange,
331                                   ValueRange)>
332         bodyBuilderFn,
333     Optional<LinalgLoopDistributionOptions> distributionOptions,
334     ArrayRef<StringRef> distributionTypes) {
335   SmallVector<Value> iterArgInitValues = linalgOp.getOutputTensorOperands();
336   // Create procInfo so it dominates loops, if appropriate.
337   SmallVector<ProcInfo, 4> procInfo;
338   SmallVector<DistributionMethod, 0> distributionMethod;
339   if (distributionOptions.hasValue()) {
340     // Collect loop ranges for parallel dimensions.
341     SmallVector<Range, 2> parallelLoopRanges;
342     for (auto iteratorType : enumerate(iteratorTypes))
343       if (isParallelIterator(iteratorType.value()))
344         parallelLoopRanges.push_back(loopRanges[iteratorType.index()]);
345 
346     // Get their distribution schemes.
347     distributionMethod = distributionOptions->distributionMethod;
348     if (distributionMethod.size() < parallelLoopRanges.size())
349       parallelLoopRanges.resize(distributionMethod.size());
350     procInfo = distributionOptions->procInfo(b, loc, parallelLoopRanges);
351   }
352 
353   SmallVector<Value, 4> lbs, ubs, steps;
354   unpackRanges(loopRanges, lbs, ubs, steps);
355   LoopNest loopNest = mlir::scf::buildLoopNest(
356       b, loc, lbs, ubs, steps, iterArgInitValues,
357       [&](OpBuilder &b, Location loc, ValueRange ivs, ValueRange iterArgs) {
358         assert(iterArgs.size() == linalgOp.getOutputTensorOperands().size() &&
359                "expect the number of output tensors and iter args to match");
360         SmallVector<Value> operandValuesToUse =
361             linalgOp.getInputAndOutputOperands();
362         if (!iterArgs.empty()) {
363           operandValuesToUse = linalgOp.getInputOperands();
364           operandValuesToUse.append(iterArgs.begin(), iterArgs.end());
365         }
366         return bodyBuilderFn(b, loc, ivs, operandValuesToUse);
367       });
368 
369   if (!distributionOptions || loopNest.loops.empty())
370     return;
371 
372   // Filter out scf.for loops that were created out of parallel dimensions.
373   SmallVector<scf::ForOp, 4> loops;
374   for (auto iteratorType : enumerate(iteratorTypes))
375     if (isParallelIterator(iteratorType.value()))
376       loops.push_back(loopNest.loops[iteratorType.index()]);
377 
378   // Distribute - only supports cyclic distribution for now.
379   for (auto it : llvm::zip(loops, procInfo, distributionMethod))
380     if (std::get<2>(it) == DistributionMethod::Cyclic)
381       mapLoopToProcessorIds(std::get<0>(it), std::get<1>(it).procId,
382                             std::get<1>(it).nprocs);
383 }
384 
385 /// Specialization to build affine "for" nest.
386 template <>
387 void GenerateLoopNest<AffineForOp>::doit(
388     OpBuilder &b, Location loc, ArrayRef<Range> loopRanges, LinalgOp linalgOp,
389     ArrayRef<Attribute> iteratorTypes,
390     function_ref<scf::ValueVector(OpBuilder &, Location, ValueRange,
391                                   ValueRange)>
392         bodyBuilderFn,
393     Optional<LinalgLoopDistributionOptions>, ArrayRef<StringRef>) {
394   SmallVector<Value> iterArgInitValues = linalgOp.getOutputTensorOperands();
395   assert(iterArgInitValues.empty() && "unexpected AffineForOp init values");
396   SmallVector<Value, 4> lbs, ubs, steps;
397   unpackRanges(loopRanges, lbs, ubs, steps);
398 
399   // Affine loops require constant steps.
400   SmallVector<int64_t, 4> constantSteps;
401   constantSteps.reserve(steps.size());
402   for (Value v : steps) {
403     auto op = v.getDefiningOp<arith::ConstantIndexOp>();
404     assert(op && "Affine loops require constant steps");
405     constantSteps.push_back(op.value());
406   }
407 
408   mlir::buildAffineLoopNest(b, loc, lbs, ubs, constantSteps,
409                             [&](OpBuilder &b, Location loc, ValueRange ivs) {
410                               SmallVector<Value> operandValuesToUse =
411                                   linalgOp.getInputAndOutputOperands();
412                               bodyBuilderFn(b, loc, ivs, operandValuesToUse);
413                             });
414 }
415 
416 /// Specialization to build an linalg.tiled_loop
417 template <>
418 void GenerateLoopNest<TiledLoopOp>::doit(
419     OpBuilder &b, Location loc, ArrayRef<Range> loopRanges, LinalgOp linalgOp,
420     ArrayRef<Attribute> iteratorTypes,
421     function_ref<scf::ValueVector(OpBuilder &, Location, ValueRange,
422                                   ValueRange)>
423         bodyBuilderFn,
424     Optional<LinalgLoopDistributionOptions> distributionOptions,
425     ArrayRef<StringRef> distributionTypes) {
426   SmallVector<ProcInfo, 2> procInfo;
427   SmallVector<Value, 4> lbs, ubs, steps;
428   unpackRanges(loopRanges, lbs, ubs, steps);
429 
430   auto wrappedBuilderFn = [&](OpBuilder &nestedBuilder, Location nestedLoc,
431                               ValueRange ivs, ValueRange inputs,
432                               ValueRange outputs) {
433     SmallVector<Value> operandValuesToUse = inputs;
434     operandValuesToUse.append(outputs.begin(), outputs.end());
435     scf::ValueVector results =
436         bodyBuilderFn(nestedBuilder, nestedLoc, ivs, operandValuesToUse);
437     nestedBuilder.create<linalg::YieldOp>(nestedLoc, results);
438   };
439 
440   SmallVector<Value> inputOperands = linalgOp.getInputOperands();
441   SmallVector<Value> outputOperands = linalgOp.getOutputOperands();
442   auto tiledLoop =
443       b.create<TiledLoopOp>(loc, lbs, ubs, steps, inputOperands, outputOperands,
444                             b.getArrayAttr(iteratorTypes), wrappedBuilderFn);
445   if (!distributionTypes.empty())
446     tiledLoop.setDistributionTypes(b, distributionTypes);
447 }
448 
449 /// Update the `lb`, `ub` and `step` to get per processor `lb`, `ub` and `step`.
450 void updateBoundsForCyclicDistribution(OpBuilder &b, Location loc, Value procId,
451                                        Value nprocs, Value &lb, Value &ub,
452                                        Value &step) {
453   AffineExpr d0, d1;
454   bindDims(b.getContext(), d0, d1);
455   AffineExpr s0 = getAffineSymbolExpr(0, b.getContext());
456   lb = makeComposedAffineApply(b, loc, d0 + d1 * s0, {lb, procId, step});
457   step = makeComposedAffineApply(b, loc, d0 * s0, {nprocs, step});
458 }
459 
460 /// Generates a loop nest consisting of scf.parallel and scf.for, depending
461 /// on the `iteratorTypes.` Consecutive parallel loops create a single
462 /// scf.parallel operation; each sequential loop creates a new scf.for
463 /// operation. The body of the innermost loop is populated by
464 /// `bodyBuilderFn` that accepts a range of induction variables for all
465 /// loops. `ivStorage` is used to store the partial list of induction
466 /// variables.
467 // TODO: this function can be made iterative instead. However, it
468 // will have at most as many recursive calls as nested loops, which rarely
469 // exceeds 10.
470 static void generateParallelLoopNest(
471     OpBuilder &b, Location loc, ValueRange lbs, ValueRange ubs,
472     ValueRange steps, ArrayRef<Attribute> iteratorTypes,
473     function_ref<void(OpBuilder &, Location, ValueRange)> bodyBuilderFn,
474     SmallVectorImpl<Value> &ivStorage,
475     ArrayRef<DistributionMethod> distributionMethod = {}) {
476   assert(lbs.size() == ubs.size());
477   assert(lbs.size() == steps.size());
478   assert(lbs.size() == iteratorTypes.size());
479 
480   // If there are no (more) loops to be generated, generate the body and be
481   // done with it.
482   if (iteratorTypes.empty()) {
483     bodyBuilderFn(b, loc, ivStorage);
484     return;
485   }
486 
487   // Find the outermost parallel loops and drop their types from the list.
488   unsigned nLoops = iteratorTypes.size();
489   unsigned nOuterPar =
490       nLoops - iteratorTypes.drop_while(isParallelIterator).size();
491 
492   // If there are no outer parallel loops, generate one sequential loop and
493   // recurse. Note that we wouldn't have dropped anything from `iteratorTypes`
494   // in this case.
495   if (nOuterPar == 0) {
496     LoopNest singleLoop = buildLoopNest(
497         b, loc, lbs.take_front(), ubs.take_front(), steps.take_front(),
498         [&](OpBuilder &b, Location loc, ValueRange ivs) {
499           ivStorage.append(ivs.begin(), ivs.end());
500           generateParallelLoopNest(b, loc, lbs.drop_front(), ubs.drop_front(),
501                                    steps.drop_front(),
502                                    iteratorTypes.drop_front(), bodyBuilderFn,
503                                    ivStorage, distributionMethod);
504         });
505     return;
506   }
507   if (distributionMethod.empty()) {
508     // Generate a single parallel loop-nest operation for all outermost
509     // parallel loops and recurse.
510     b.create<scf::ParallelOp>(
511         loc, lbs.take_front(nOuterPar), ubs.take_front(nOuterPar),
512         steps.take_front(nOuterPar),
513         [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange localIvs) {
514           ivStorage.append(localIvs.begin(), localIvs.end());
515           generateParallelLoopNest(
516               nestedBuilder, nestedLoc, lbs.drop_front(nOuterPar),
517               ubs.drop_front(nOuterPar), steps.drop_front(nOuterPar),
518               iteratorTypes.drop_front(nOuterPar), bodyBuilderFn, ivStorage,
519               (distributionMethod.size() < nOuterPar)
520                   ? ArrayRef<DistributionMethod>()
521                   : distributionMethod.drop_front(nOuterPar));
522         });
523     return;
524   }
525 
526   // Process all consecutive similarly distributed loops simultaneously.
527   DistributionMethod methodToUse = distributionMethod[0];
528   unsigned numProcessed = 1;
529   for (unsigned i = 1; i < nOuterPar && i < distributionMethod.size(); ++i) {
530     if (distributionMethod[i] != methodToUse)
531       break;
532     numProcessed++;
533   }
534 
535   switch (methodToUse) {
536   case DistributionMethod::Cyclic: {
537     // Generate a single parallel loop-nest operation for all outermost
538     // parallel loops and recurse.
539     b.create<scf::ParallelOp>(
540         loc, lbs.take_front(numProcessed), ubs.take_front(numProcessed),
541         steps.take_front(numProcessed),
542         [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange localIvs) {
543           ivStorage.append(localIvs.begin(), localIvs.end());
544           generateParallelLoopNest(
545               nestedBuilder, nestedLoc, lbs.drop_front(numProcessed),
546               ubs.drop_front(numProcessed), steps.drop_front(numProcessed),
547               iteratorTypes.drop_front(numProcessed), bodyBuilderFn, ivStorage,
548               (distributionMethod.size() < numProcessed)
549                   ? ArrayRef<DistributionMethod>()
550                   : distributionMethod.drop_front(numProcessed));
551         });
552     return;
553   }
554   case DistributionMethod::CyclicNumProcsGeNumIters: {
555     // Check (for the processed loops) that the iteration is in-bounds.
556     ArithBuilder ab(b, loc);
557     Value cond = ab.slt(lbs[0], ubs[0]);
558     for (unsigned i = 1; i < numProcessed; ++i)
559       cond = ab._and(cond, ab.slt(lbs[i], ubs[i]));
560     ivStorage.append(lbs.begin(), std::next(lbs.begin(), numProcessed));
561     b.create<scf::IfOp>(loc, cond, [&](OpBuilder &b, Location loc) {
562       generateParallelLoopNest(
563           b, loc, lbs.drop_front(numProcessed), ubs.drop_front(numProcessed),
564           steps.drop_front(numProcessed),
565           iteratorTypes.drop_front(numProcessed), bodyBuilderFn, ivStorage,
566           distributionMethod.drop_front(numProcessed));
567       b.create<scf::YieldOp>(loc, ValueRange{});
568     });
569     return;
570   }
571   case DistributionMethod::CyclicNumProcsEqNumIters:
572     // No check/loops needed here. Set the `%iv` to be the `%lb` and proceed
573     // with inner loop generation.
574     ivStorage.append(lbs.begin(), std::next(lbs.begin(), numProcessed));
575     generateParallelLoopNest(
576         b, loc, lbs.drop_front(numProcessed), ubs.drop_front(numProcessed),
577         steps.drop_front(numProcessed), iteratorTypes.drop_front(numProcessed),
578         bodyBuilderFn, ivStorage, distributionMethod.drop_front(numProcessed));
579     return;
580   }
581 }
582 
583 /// Specialization for generating a mix of parallel and sequential scf loops.
584 template <>
585 void GenerateLoopNest<scf::ParallelOp>::doit(
586     OpBuilder &b, Location loc, ArrayRef<Range> loopRanges, LinalgOp linalgOp,
587     ArrayRef<Attribute> iteratorTypes,
588     function_ref<scf::ValueVector(OpBuilder &, Location, ValueRange,
589                                   ValueRange)>
590         bodyBuilderFn,
591     Optional<LinalgLoopDistributionOptions> distributionOptions,
592     ArrayRef<StringRef> distributionTypes) {
593   SmallVector<Value> iterArgInitValues = linalgOp.getOutputTensorOperands();
594   assert(iterArgInitValues.empty() && "unexpected ParallelOp init values");
595   // This function may be passed more iterator types than ranges.
596   assert(iteratorTypes.size() >= loopRanges.size() &&
597          "expected iterator type for all ranges");
598   iteratorTypes = iteratorTypes.take_front(loopRanges.size());
599   SmallVector<Value, 8> lbsStorage, ubsStorage, stepsStorage, ivs;
600   unsigned numLoops = iteratorTypes.size();
601   ivs.reserve(numLoops);
602   lbsStorage.reserve(numLoops);
603   ubsStorage.reserve(numLoops);
604   stepsStorage.reserve(numLoops);
605 
606   // Get the loop lb, ub, and step.
607   unpackRanges(loopRanges, lbsStorage, ubsStorage, stepsStorage);
608 
609   // Modify the lb, ub, and step based on the distribution options.
610   SmallVector<DistributionMethod, 0> distributionMethod;
611   if (distributionOptions) {
612     auto &options = distributionOptions.getValue();
613     distributionMethod.assign(distributionOptions->distributionMethod.begin(),
614                               distributionOptions->distributionMethod.end());
615     SmallVector<Range, 2> parallelLoopRanges;
616     for (auto iteratorType : enumerate(iteratorTypes)) {
617       if (isParallelIterator(iteratorType.value()))
618         parallelLoopRanges.push_back(loopRanges[iteratorType.index()]);
619     }
620     if (distributionMethod.size() < parallelLoopRanges.size())
621       parallelLoopRanges.resize(distributionMethod.size());
622     SmallVector<ProcInfo, 2> procInfo =
623         options.procInfo(b, loc, parallelLoopRanges);
624     unsigned index = 0;
625     for (auto iteratorType : enumerate(iteratorTypes)) {
626       if (index >= procInfo.size())
627         break;
628       if (isParallelIterator(iteratorType.value())) {
629         unsigned i = iteratorType.index();
630         updateBoundsForCyclicDistribution(b, loc, procInfo[index].procId,
631                                           procInfo[index].nprocs, lbsStorage[i],
632                                           ubsStorage[i], stepsStorage[i]);
633         index++;
634       }
635     }
636   }
637   ValueRange lbs(lbsStorage), ubs(ubsStorage), steps(stepsStorage);
638   generateParallelLoopNest(
639       b, loc, lbs, ubs, steps, iteratorTypes,
640       [&](OpBuilder &b, Location loc, ValueRange ivs) {
641         SmallVector<Value> operandValuesToUse =
642             linalgOp.getInputAndOutputOperands();
643         bodyBuilderFn(b, loc, ivs, operandValuesToUse);
644       },
645       ivs, distributionMethod);
646 
647   assert(ivs.size() == iteratorTypes.size() && "did not generate enough loops");
648 }
649 
650 static Value fullyComposeAndAffineApply(OpBuilder &b, Location loc,
651                                         AffineExpr expr, ValueRange operands) {
652   AffineMap map = AffineMap::inferFromExprList({expr}).front();
653   SmallVector<Value> normalizedOperands(operands.begin(), operands.end());
654   mlir::fullyComposeAffineMapAndOperands(&map, &normalizedOperands);
655   canonicalizeMapAndOperands(&map, &normalizedOperands);
656   return b.createOrFold<AffineApplyOp>(loc, map, normalizedOperands);
657 }
658 
659 Value makeTiledShape(OpBuilder &builder, Location loc, Value valueToTile,
660                      ValueRange tileSizes, AffineMap map, ValueRange lbs,
661                      ValueRange ubs, ValueRange subShapeSizes) {
662   auto shapedType = valueToTile.getType().dyn_cast<ShapedType>();
663   assert(shapedType && "only shaped types can be tiled");
664   ArrayRef<int64_t> shape = shapedType.getShape();
665   int64_t rank = shapedType.getRank();
666 
667   // Construct a new subview / extract_slice for the tile.
668   SmallVector<OpFoldResult, 4> offsets, sizes, strides;
669   offsets.reserve(rank);
670   sizes.reserve(rank);
671   strides.reserve(rank);
672   for (unsigned r = 0; r < rank; ++r) {
673     LLVM_DEBUG(llvm::dbgs() << "makeTiledShape: for dim#" << r);
674     if (!isTiled(map.getSubMap({r}), tileSizes)) {
675       offsets.push_back(builder.getIndexAttr(0));
676       Value dim = createOrFoldDimOp(builder, loc, valueToTile, r);
677       sizes.push_back(getAsOpFoldResult(dim));
678       strides.push_back(builder.getIndexAttr(1));
679       LLVM_DEBUG(llvm::dbgs() << ": not tiled: use size: " << dim << "\n");
680       continue;
681     }
682     LLVM_DEBUG(llvm::dbgs() << ": tiled: figure out subsize...\n");
683 
684     // Tiling creates a new slice at the proper index, the slice step is 1
685     // (i.e. the op does not subsample, stepping occurs in the loop).
686     auto m = map.getSubMap({r});
687     LLVM_DEBUG(llvm::dbgs() << "makeTiledShape: submap: " << m << "\n");
688     auto offset = applyMapToValues(builder, loc, m, lbs).front();
689     offsets.push_back(offset);
690     auto closedIntSize =
691         applyMapToValues(builder, loc, m, subShapeSizes).front();
692     // Resulting size needs to be made half open interval again.
693     AffineExpr s0 = getAffineSymbolExpr(0, builder.getContext());
694     Value size =
695         fullyComposeAndAffineApply(builder, loc, s0 + 1, closedIntSize);
696     LLVM_DEBUG(llvm::dbgs() << "makeTiledShape: raw size: " << size << "\n");
697 
698     // The size of the subview / extract_slice should be trimmed to avoid
699     // out-of-bounds accesses, unless:
700     // a. We statically know the subshape size divides the shape size evenly.
701     // b. The subshape size is 1. According to the way the loops are set up,
702     //    tensors with "0" dimensions would never be constructed.
703     int64_t shapeSize = shape[r];
704     auto sizeCst = size.getDefiningOp<arith::ConstantIndexOp>();
705     auto hasTileSizeOne = sizeCst && sizeCst.value() == 1;
706     auto dividesEvenly = sizeCst && !ShapedType::isDynamic(shapeSize) &&
707                          ((shapeSize % sizeCst.value()) == 0);
708     if (!hasTileSizeOne && !dividesEvenly) {
709       LLVM_DEBUG(llvm::dbgs() << "makeTiledShape: shapeSize=" << shapeSize
710                               << ", size: " << size
711                               << ": make sure in bound with affine.min\n");
712 
713       AffineExpr dim0, dim1, dim2;
714       bindDims(builder.getContext(), dim0, dim1, dim2);
715 
716       // Get the dimension size for this dimension. We need to first calculate
717       // the max index and then plus one. This is important because for
718       // convolution ops, we have its input window dimension's affine map of the
719       // form `(d0 * s0 + d1)`, where `d0`/`d1 is an output/filter window
720       // dimension and `s0` is stride. Directly use the dimension size of
721       // output/filer window dimensions will cause incorrect calculation.
722       AffineMap minusOneMap =
723           AffineMap::inferFromExprList({ArrayRef<AffineExpr>{dim0 - 1}})
724               .front();
725       AffineMap plusOneMap =
726           AffineMap::inferFromExprList({ArrayRef<AffineExpr>{dim0 + 1}})
727               .front();
728       auto maxIndices = llvm::to_vector<8>(llvm::map_range(ubs, [&](Value ub) {
729         return makeComposedAffineApply(builder, loc, minusOneMap, {ub})
730             .getResult();
731       }));
732       Value maxIndex = applyMapToValues(builder, loc, m, maxIndices).front();
733       Value d = makeComposedAffineApply(builder, loc, plusOneMap, {maxIndex});
734 
735       // Compute min(size, dim - offset) to avoid out-of-bounds accesses.
736       AffineMap minMap = AffineMap::inferFromExprList(
737                              {ArrayRef<AffineExpr>{dim0, dim1 - dim2}})
738                              .front();
739       SmallVector<Value, 4> operands{size, d, offset};
740       fullyComposeAffineMapAndOperands(&minMap, &operands);
741       canonicalizeMapAndOperands(&minMap, &operands);
742       size = builder.create<AffineMinOp>(loc, builder.getIndexType(), minMap,
743                                          operands);
744     }
745 
746     sizes.push_back(size);
747     LLVM_DEBUG(llvm::dbgs()
748                << "makeTiledShape: new offset: " << offset << "\n");
749     LLVM_DEBUG(llvm::dbgs() << "makeTiledShape: new size: " << size << "\n");
750     strides.push_back(builder.getIndexAttr(1));
751   }
752 
753   auto *sliceOp = TypeSwitch<ShapedType, Operation *>(shapedType)
754                       .Case([&](MemRefType) {
755                         return builder.create<memref::SubViewOp>(
756                             loc, valueToTile, offsets, sizes, strides);
757                       })
758                       .Case([&](RankedTensorType) {
759                         return makeComposedExtractSliceOp(
760                             builder, loc, valueToTile, offsets, sizes, strides);
761                       })
762                       .Default([](ShapedType) -> Operation * {
763                         llvm_unreachable("Unexpected shaped type");
764                       });
765   return sliceOp->getResult(0);
766 }
767 
768 SmallVector<Value> computeTileOffsets(OpBuilder &b, Location loc,
769                                       ValueRange ivs, ValueRange tileSizes) {
770   SmallVector<Value> offsets;
771   for (unsigned idx = 0, idxIvs = 0, e = tileSizes.size(); idx < e; ++idx) {
772     LLVM_DEBUG(llvm::dbgs() << "makeTiledShapes: for loop#" << idx << "\n");
773     bool isTiled = !isZero(tileSizes[idx]);
774     offsets.push_back(
775         isTiled ? ivs[idxIvs++]
776                 : b.create<arith::ConstantIndexOp>(loc, 0).getResult());
777     LLVM_DEBUG(llvm::dbgs()
778                << "computeTileOffsets: " << offsets.back() << "\n");
779   }
780   return offsets;
781 }
782 
783 SmallVector<Value> computeTileSizes(OpBuilder &b, Location loc, ValueRange ivs,
784                                     ValueRange tileSizes,
785                                     ArrayRef<Value> sizeBounds) {
786   SmallVector<Value> sizes;
787   for (unsigned idx = 0, e = tileSizes.size(); idx < e; ++idx) {
788     bool isTiled = !isZero(tileSizes[idx]);
789     // Before composing, we need to make range a closed interval.
790     Value size = isTiled ? tileSizes[idx] : sizeBounds[idx];
791     AffineExpr d0 = getAffineDimExpr(0, b.getContext());
792     sizes.push_back(fullyComposeAndAffineApply(b, loc, d0 - 1, size));
793     LLVM_DEBUG(llvm::dbgs() << "computeTileSizes: " << sizes.back() << "\n");
794   }
795   return sizes;
796 }
797 
798 SmallVector<Value, 4> makeTiledShapes(OpBuilder &b, Location loc,
799                                       LinalgOp linalgOp,
800                                       ArrayRef<Value> valuesToTile,
801                                       ValueRange ivs, ValueRange tileSizes,
802                                       ArrayRef<Value> sizeBounds) {
803   assert(ivs.size() == static_cast<size_t>(llvm::count_if(
804                            llvm::make_range(tileSizes.begin(), tileSizes.end()),
805                            [](Value v) { return !isZero(v); })) &&
806          "expected as many ivs as non-zero sizes");
807 
808   // Construct (potentially temporary) mins and maxes on which to apply maps
809   // that define tile subshapes.
810   SmallVector<Value> lbs = computeTileOffsets(b, loc, ivs, tileSizes);
811   SmallVector<Value> subShapeSizes =
812       computeTileSizes(b, loc, ivs, tileSizes, sizeBounds);
813 
814   assert(static_cast<int64_t>(valuesToTile.size()) ==
815              linalgOp.getNumInputsAndOutputs() &&
816          "expected one value to tile for every operand");
817   SmallVector<Value, 4> tiledShapes;
818   tiledShapes.reserve(valuesToTile.size());
819   for (OpOperand *opOperand : linalgOp.getInputAndOutputOperands()) {
820     Value shapedOp = valuesToTile[opOperand->getOperandNumber()];
821     LLVM_DEBUG(llvm::dbgs() << "makeTiledShapes: for operand " << shapedOp);
822     AffineMap map = linalgOp.getTiedIndexingMap(opOperand);
823     // Use `opOperand` as is if it is not tiled and not an output tensor. Having
824     // an extract/insert slice pair for all output tensors simplifies follow up
825     // transformations such as padding and bufferization since the
826     // extract/insert slice pairs make the accessed iteration argument
827     // subdomains explicit.
828     if (!isTiled(map, tileSizes) && !linalgOp.isOutputTensor(opOperand)) {
829       tiledShapes.push_back(shapedOp);
830       LLVM_DEBUG(llvm::dbgs() << ": not tiled: use shape: "
831                               << opOperand->get().getType() << "\n");
832       continue;
833     }
834     LLVM_DEBUG(llvm::dbgs() << ": tiled: figure out subshape...\n");
835 
836     tiledShapes.push_back(makeTiledShape(b, loc, shapedOp, tileSizes, map, lbs,
837                                          sizeBounds, subShapeSizes));
838   }
839 
840   return tiledShapes;
841 }
842 
843 void addTileLoopIvsToIndexOpResults(OpBuilder &b, LinalgOp tiledOp,
844                                     ArrayRef<Value> ivs) {
845   if (tiledOp.hasIndexSemantics()) {
846     for (IndexOp indexOp : tiledOp.getBlock()->getOps<IndexOp>()) {
847       if (ivs[indexOp.dim()] == nullptr)
848         continue;
849       OpBuilder::InsertionGuard guard(b);
850       b.setInsertionPointAfter(indexOp);
851       AffineExpr index, offset;
852       bindDims(b.getContext(), index, offset);
853       AffineApplyOp applyOp = makeComposedAffineApply(
854           b, indexOp.getLoc(), index + offset,
855           ValueRange{indexOp.getResult(), ivs[indexOp.dim()]});
856       indexOp.getResult().replaceAllUsesExcept(applyOp, applyOp);
857     }
858   }
859 }
860 
861 } // namespace linalg
862 } // namespace mlir
863