1 //===- LinalgTransforms.cpp - Linalg transformations as patterns ----------===//
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 logic and helpers to expose Linalg transforms as rewrite
10 // patterns.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
15 #include "mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h"
16 #include "mlir/Dialect/Linalg/IR/LinalgOps.h"
17 #include "mlir/Dialect/Linalg/Utils/Utils.h"
18 #include "mlir/Dialect/StandardOps/EDSC/Intrinsics.h"
19 #include "mlir/Dialect/Utils/StructuredOpsUtils.h"
20 #include "mlir/Dialect/Vector/EDSC/Intrinsics.h"
21 #include "mlir/Dialect/Vector/VectorOps.h"
22 #include "mlir/IR/AffineExpr.h"
23 #include "mlir/IR/Matchers.h"
24 #include "mlir/Pass/Pass.h"
25 #include "mlir/Support/LLVM.h"
26 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include <type_traits>
30 
31 #define DEBUG_TYPE "linalg-transforms"
32 
33 using namespace mlir;
34 using namespace mlir::edsc;
35 using namespace mlir::edsc::intrinsics;
36 using namespace mlir::linalg;
37 
38 #define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE << "]: ")
39 
40 //===----------------------------------------------------------------------===//
41 // Transformations exposed as rewrite patterns.
42 //===----------------------------------------------------------------------===//
43 // Marker used as attribute name in generated Linalg rewriting transformations.
44 const StringLiteral mlir::linalg::LinalgTransforms::kLinalgTransformMarker =
45     "__internal_linalg_transform__";
46 
47 mlir::linalg::LinalgMarker::LinalgMarker(ArrayRef<Identifier> matchDisjunction,
48                                          Optional<Identifier> replacement)
49     : matchDisjunction(matchDisjunction.begin(), matchDisjunction.end()),
50       replacement(replacement) {}
51 
52 LogicalResult
53 mlir::linalg::LinalgMarker::checkAndNotify(PatternRewriter &rewriter,
54                                            Operation *op) const {
55   auto attr = op->template getAttrOfType<StringAttr>(
56       LinalgTransforms::kLinalgTransformMarker);
57 
58   if (!attr) {
59     // 1. Has no marker case and matchDisjunction is empty.
60     if (matchDisjunction.empty())
61       return success();
62 
63     // 2. Has no marker but was expecting a marker.
64     return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
65       diag << " does not have any marker from list: ";
66       interleaveComma(matchDisjunction, diag);
67     });
68   }
69 
70   // 4. Match explicit marker.
71   for (auto marker : matchDisjunction)
72     if (attr.getValue() == marker)
73       return success();
74 
75   // 5. Fail to match.
76   return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
77     diag << " does not have any marker from list: ";
78     interleaveComma(matchDisjunction, diag);
79   });
80 }
81 
82 void mlir::linalg::LinalgMarker::replaceLinalgMarker(PatternRewriter &rewriter,
83                                                      Operation *op) const {
84   if (replacement.hasValue())
85     op->setAttr(LinalgTransforms::kLinalgTransformMarker,
86                 rewriter.getStringAttr(replacement.getValue()));
87   else
88     op->removeAttr(Identifier::get(LinalgTransforms::kLinalgTransformMarker,
89                                    rewriter.getContext()));
90 }
91 
92 LinalgTilingOptions &
93 mlir::linalg::LinalgTilingOptions::setTileSizes(ArrayRef<int64_t> ts) {
94   SmallVector<int64_t, 4> tileSizes(ts.begin(), ts.end());
95   tileSizeComputationFunction = [tileSizes](OpBuilder &b, Operation *op) {
96     OpBuilder::InsertionGuard guard(b);
97     b.setInsertionPointToStart(
98         &op->getParentOfType<FuncOp>().getBody().front());
99     return llvm::to_vector<4>(map_range(tileSizes, [&](int64_t s) {
100       Value v = b.create<ConstantIndexOp>(op->getLoc(), s);
101       return v;
102     }));
103   };
104   return *this;
105 }
106 
107 /// Linalg base tiling pattern.
108 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern(
109     StringRef opName, MLIRContext *context, LinalgTilingOptions options,
110     LinalgMarker marker, PatternBenefit benefit)
111     : RewritePattern(opName, {}, benefit, context), marker(marker),
112       options(options) {}
113 
114 LogicalResult mlir::linalg::LinalgBaseTilingPattern::matchAndRewriteBase(
115     Operation *op, PatternRewriter &rewriter,
116     SmallVectorImpl<Value> &tensorResults) const {
117   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
118   if (!linalgOp)
119     return failure();
120   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
121     return failure();
122 
123   // If LinalgOp has results, they must all be tied to init tensors.
124   // We enforce this to ensure all tiled ops have been rewritten in
125   // "init tensor" form. This ensures tiling has anchor values into which to
126   // subtensor / subtensor_insert. Otherwise tiling would need to allocate which
127   // is not acceptable.
128   // This would not be the case with a special terminator op that generates the
129   // whole tensor (instead of inserting a subtensor). But the generator-based
130   // abstraction has other issues.
131   if (linalgOp.getNumInitTensors() != linalgOp.getOperation()->getNumResults())
132     return failure();
133 
134   Optional<TiledLinalgOp> res = tileLinalgOp(rewriter, linalgOp, options);
135 
136   if (!res)
137     return failure();
138 
139   // Return relevant information to derived pattern.
140   tensorResults = res->tensorResults;
141 
142   // New marker if specified.
143   marker.replaceLinalgMarker(rewriter, res->op.getOperation());
144   return success();
145 }
146 
147 mlir::linalg::LinalgBaseTileAndFusePattern::LinalgBaseTileAndFusePattern(
148     StringRef opName, MLIRContext *context,
149     const LinalgDependenceGraph &dependenceGraph,
150     LinalgTilingOptions tilingOptions, LinalgFusionOptions fusionOptions,
151     LinalgMarker marker, LinalgMarker fusedOpMarker,
152     LinalgMarker originalOpMarker, PatternBenefit benefit)
153     : RewritePattern(opName, {}, benefit, context),
154       dependenceGraph(dependenceGraph), tilingOptions(tilingOptions),
155       fusionOptions(fusionOptions), marker(marker),
156       fusedOpMarker(fusedOpMarker), originalOpMarker(originalOpMarker) {}
157 
158 LogicalResult mlir::linalg::LinalgBaseTileAndFusePattern::matchAndRewrite(
159     Operation *op, PatternRewriter &rewriter) const {
160   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
161   if (!linalgOp)
162     return failure();
163   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
164     return failure();
165   if (!linalgOp.hasBufferSemantics())
166     return failure();
167 
168   Optional<TiledAndFusedLinalgOps> tiledAndFusedOps = tileAndFuseLinalgOps(
169       rewriter, op, dependenceGraph, tilingOptions, fusionOptions);
170   if (!tiledAndFusedOps)
171     return failure();
172   marker.replaceLinalgMarker(rewriter, tiledAndFusedOps->op.getOperation());
173   for (auto fusedOp : tiledAndFusedOps->fusedProducers) {
174     fusedOpMarker.replaceLinalgMarker(rewriter, fusedOp.getOperation());
175   }
176   for (auto origProducerOp : tiledAndFusedOps->originalProducers)
177     originalOpMarker.replaceLinalgMarker(rewriter,
178                                          origProducerOp.getOperation());
179   rewriter.updateRootInPlace(
180       op, [&]() { originalOpMarker.replaceLinalgMarker(rewriter, op); });
181   return success();
182 }
183 
184 /// Linalg base interchange pattern.
185 mlir::linalg::LinalgBaseInterchangePattern::LinalgBaseInterchangePattern(
186     StringRef opName, MLIRContext *context,
187     ArrayRef<unsigned> interchangeVector, LinalgMarker marker,
188     PatternBenefit benefit)
189     : RewritePattern(opName, {}, benefit, context), marker(marker),
190       interchangeVector(interchangeVector.begin(), interchangeVector.end()) {}
191 
192 LogicalResult mlir::linalg::LinalgBaseInterchangePattern::matchAndRewrite(
193     Operation *op, PatternRewriter &rewriter) const {
194   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
195   if (!linalgOp)
196     return failure();
197   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
198     return failure();
199   if (failed(interchangeGenericLinalgOpPrecondition(op, interchangeVector)))
200     return failure();
201 
202   // TODO: figure out how this interplays with named ops. In particular this
203   // should break the named op property.
204   rewriter.updateRootInPlace(op, [&]() {
205     interchange(linalgOp, interchangeVector);
206     // New marker if specified.
207     marker.replaceLinalgMarker(rewriter, op);
208   });
209   return success();
210 }
211 
212 mlir::linalg::LinalgBasePromotionPattern::LinalgBasePromotionPattern(
213     StringRef opName, MLIRContext *context, LinalgPromotionOptions options,
214     LinalgMarker marker, PatternBenefit benefit)
215     : RewritePattern(opName, {}, benefit, context), marker(marker),
216       options(options) {}
217 
218 LogicalResult mlir::linalg::LinalgBasePromotionPattern::matchAndRewrite(
219     Operation *op, PatternRewriter &rewriter) const {
220   if (failed(marker.checkAndNotify(rewriter, op)))
221     return failure();
222   if (failed(promoteSubviewsPrecondition(op, options)))
223     return failure();
224 
225   // TODO: We cannot use root update here. This pattern is creating other ops,
226   // so if the promotion fails, those need to be cleaned up, which doesnt seem
227   // to be happening here. So to fail properly, we should be cloning the op and
228   // deleting the previous op. This needs more investigation.
229   rewriter.startRootUpdate(op);
230   Optional<LinalgOp> promotedOp = promoteSubViews(rewriter, op, options);
231   if (!promotedOp) {
232     rewriter.cancelRootUpdate(op);
233     return op->emitError("subview promotion failed");
234   }
235   rewriter.finalizeRootUpdate(op);
236   marker.replaceLinalgMarker(rewriter, op);
237   return success();
238 }
239 
240 mlir::linalg::LinalgBaseVectorizationPattern::LinalgBaseVectorizationPattern(
241     StringRef opName, MLIRContext *context, LinalgMarker marker,
242     PatternBenefit benefit)
243     : RewritePattern(opName, {}, benefit, context), marker(marker) {}
244 
245 LogicalResult mlir::linalg::LinalgBaseVectorizationPattern::matchAndRewrite(
246     Operation *op, PatternRewriter &rewriter) const {
247   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
248   if (!linalgOp)
249     return failure();
250   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
251     return failure();
252   if (failed(vectorizeLinalgOpPrecondition(op)))
253     return failure();
254   vectorizeLinalgOp(rewriter, op);
255   rewriter.eraseOp(op);
256   return success();
257 }
258 
259 LogicalResult mlir::linalg::applyStagedPatterns(
260     Operation *op, ArrayRef<FrozenRewritePatternList> stage1Patterns,
261     const FrozenRewritePatternList &stage2Patterns,
262     function_ref<LogicalResult(Operation *)> stage3Lambda) {
263   unsigned iteration = 0;
264   (void)iteration;
265   for (const auto &patterns : stage1Patterns) {
266     LLVM_DEBUG(DBGS() << "Before 1st stage, iter: " << ++iteration << "\n"
267                       << *op);
268     if (failed(applyPatternsAndFoldGreedily(op, patterns))) {
269       LLVM_DEBUG(DBGS() << "Underlying first stage rewrite did not converge");
270       return failure();
271     }
272     LLVM_DEBUG(DBGS() << "After 1st stage, iter: " << ++iteration << "\n"
273                       << *op);
274     if (failed(applyPatternsAndFoldGreedily(op, stage2Patterns))) {
275       LLVM_DEBUG(DBGS() << "Underlying 2nd stage rewrite did not converge");
276       return failure();
277     }
278     LLVM_DEBUG(DBGS() << "After 2nd stage, iter : " << iteration << "\n"
279                       << *op);
280     if (stage3Lambda) {
281       if (failed(stage3Lambda(op)))
282         return failure();
283       LLVM_DEBUG(DBGS() << "After 3rd stage, iter : " << iteration << "\n"
284                         << *op);
285     }
286   }
287   return success();
288 }
289 
290 /// Traverse `e` and return an AffineExpr where all occurrences of `dim` have
291 /// been replaced by either:
292 ///  - `min` if `positivePath` is true when we reach an occurrence of `dim`
293 ///  - `max` if `positivePath` is true when we reach an occurrence of `dim`
294 /// `positivePath` is negated each time we hit a multiplicative or divisive
295 /// binary op with a constant negative coefficient.
296 static AffineExpr substWithMin(AffineExpr e, AffineExpr dim, AffineExpr min,
297                                AffineExpr max, bool positivePath = true) {
298   if (e == dim)
299     return positivePath ? min : max;
300   if (auto bin = e.dyn_cast<AffineBinaryOpExpr>()) {
301     AffineExpr lhs = bin.getLHS();
302     AffineExpr rhs = bin.getRHS();
303     if (bin.getKind() == mlir::AffineExprKind::Add)
304       return substWithMin(lhs, dim, min, max, positivePath) +
305              substWithMin(rhs, dim, min, max, positivePath);
306 
307     auto c1 = bin.getLHS().dyn_cast<AffineConstantExpr>();
308     auto c2 = bin.getRHS().dyn_cast<AffineConstantExpr>();
309     if (c1 && c1.getValue() < 0)
310       return getAffineBinaryOpExpr(
311           bin.getKind(), c1, substWithMin(rhs, dim, min, max, !positivePath));
312     if (c2 && c2.getValue() < 0)
313       return getAffineBinaryOpExpr(
314           bin.getKind(), substWithMin(lhs, dim, min, max, !positivePath), c2);
315     return getAffineBinaryOpExpr(
316         bin.getKind(), substWithMin(lhs, dim, min, max, positivePath),
317         substWithMin(rhs, dim, min, max, positivePath));
318   }
319   return e;
320 }
321 
322 /// Given the `lbVal`, `ubVal` and `stepVal` of a loop, append `lbVal` and
323 /// `ubVal` to `dims` and `stepVal` to `symbols`.
324 /// Create new AffineDimExpr (`%lb` and `%ub`) and AffineSymbolExpr (`%step`)
325 /// with positions matching the newly appended values. Substitute occurrences of
326 /// `dimExpr` by either the min expression (i.e. `%lb`) or the max expression
327 /// (i.e. `%lb + %step * floordiv(%ub -1 - %lb, %step)`), depending on whether
328 /// the induction variable is used with a positive or negative  coefficient.
329 static AffineExpr substituteLoopInExpr(AffineExpr expr, AffineExpr dimExpr,
330                                        Value lbVal, Value ubVal, Value stepVal,
331                                        SmallVectorImpl<Value> &dims,
332                                        SmallVectorImpl<Value> &symbols) {
333   MLIRContext *ctx = lbVal.getContext();
334   AffineExpr lb = getAffineDimExpr(dims.size(), ctx);
335   dims.push_back(lbVal);
336   AffineExpr ub = getAffineDimExpr(dims.size(), ctx);
337   dims.push_back(ubVal);
338   AffineExpr step = getAffineSymbolExpr(symbols.size(), ctx);
339   symbols.push_back(stepVal);
340   LLVM_DEBUG(DBGS() << "Before: " << expr << "\n");
341   AffineExpr ee = substWithMin(expr, dimExpr, lb,
342                                lb + step * ((ub - 1) - lb).floorDiv(step));
343   LLVM_DEBUG(DBGS() << "After: " << expr << "\n");
344   return ee;
345 }
346 
347 /// Traverse the `dims` and substitute known min or max expressions in place of
348 /// induction variables in `exprs`.
349 static AffineMap substitute(AffineMap map, SmallVectorImpl<Value> &dims,
350                             SmallVectorImpl<Value> &symbols) {
351   auto exprs = llvm::to_vector<4>(map.getResults());
352   for (AffineExpr &expr : exprs) {
353     bool substituted = true;
354     while (substituted) {
355       substituted = false;
356       for (unsigned dimIdx = 0; dimIdx < dims.size(); ++dimIdx) {
357         Value dim = dims[dimIdx];
358         AffineExpr dimExpr = getAffineDimExpr(dimIdx, expr.getContext());
359         LLVM_DEBUG(DBGS() << "Subst: " << dim << " @ " << dimExpr << "\n");
360         AffineExpr substitutedExpr;
361         if (auto forOp = scf::getForInductionVarOwner(dim))
362           substitutedExpr = substituteLoopInExpr(
363               expr, dimExpr, forOp.lowerBound(), forOp.upperBound(),
364               forOp.step(), dims, symbols);
365 
366         if (auto parallelForOp = scf::getParallelForInductionVarOwner(dim))
367           for (unsigned idx = 0, e = parallelForOp.getNumLoops(); idx < e;
368                ++idx)
369             substitutedExpr = substituteLoopInExpr(
370                 expr, dimExpr, parallelForOp.lowerBound()[idx],
371                 parallelForOp.upperBound()[idx], parallelForOp.step()[idx],
372                 dims, symbols);
373 
374         if (!substitutedExpr)
375           continue;
376 
377         substituted = (substitutedExpr != expr);
378         expr = substitutedExpr;
379       }
380     }
381 
382     // Cleanup and simplify the results.
383     // This needs to happen outside of the loop iterating on dims.size() since
384     // it modifies dims.
385     SmallVector<Value, 4> operands(dims.begin(), dims.end());
386     operands.append(symbols.begin(), symbols.end());
387     auto map = AffineMap::get(dims.size(), symbols.size(), exprs,
388                               exprs.front().getContext());
389 
390     LLVM_DEBUG(DBGS() << "Map to simplify: " << map << "\n");
391 
392     // Pull in affine.apply operations and compose them fully into the
393     // result.
394     fullyComposeAffineMapAndOperands(&map, &operands);
395     canonicalizeMapAndOperands(&map, &operands);
396     map = simplifyAffineMap(map);
397     // Assign the results.
398     exprs.assign(map.getResults().begin(), map.getResults().end());
399     dims.assign(operands.begin(), operands.begin() + map.getNumDims());
400     symbols.assign(operands.begin() + map.getNumDims(), operands.end());
401 
402     LLVM_DEBUG(DBGS() << "Map simplified: " << map << "\n");
403   }
404 
405   assert(!exprs.empty() && "Unexpected empty exprs");
406   return AffineMap::get(dims.size(), symbols.size(), exprs, map.getContext());
407 }
408 
409 LogicalResult AffineMinSCFCanonicalizationPattern::matchAndRewrite(
410     AffineMinOp minOp, PatternRewriter &rewriter) const {
411   LLVM_DEBUG(DBGS() << "Canonicalize AffineMinSCF: " << *minOp.getOperation()
412                     << "\n");
413 
414   SmallVector<Value, 4> dims(minOp.getDimOperands()),
415       symbols(minOp.getSymbolOperands());
416   AffineMap map = substitute(minOp.getAffineMap(), dims, symbols);
417 
418   LLVM_DEBUG(DBGS() << "Resulting map: " << map << "\n");
419 
420   // Check whether any of the expressions, when subtracted from all other
421   // expressions, produces only >= 0 constants. If so, it is the min.
422   for (auto e : minOp.getAffineMap().getResults()) {
423     LLVM_DEBUG(DBGS() << "Candidate min: " << e << "\n");
424     if (!e.isSymbolicOrConstant())
425       continue;
426 
427     auto isNonPositive = [](AffineExpr e) {
428       if (auto cst = e.dyn_cast<AffineConstantExpr>())
429         return cst.getValue() < 0;
430       return true;
431     };
432 
433     // Build the subMap and check everything is statically known to be
434     // positive.
435     SmallVector<AffineExpr, 4> subExprs;
436     subExprs.reserve(map.getNumResults());
437     for (auto ee : map.getResults())
438       subExprs.push_back(ee - e);
439     MLIRContext *ctx = minOp.getContext();
440     AffineMap subMap = simplifyAffineMap(
441         AffineMap::get(map.getNumDims(), map.getNumSymbols(), subExprs, ctx));
442     LLVM_DEBUG(DBGS() << "simplified subMap: " << subMap << "\n");
443     if (llvm::any_of(subMap.getResults(), isNonPositive))
444       continue;
445 
446     // Static min found.
447     if (auto cst = e.dyn_cast<AffineConstantExpr>()) {
448       rewriter.replaceOpWithNewOp<ConstantIndexOp>(minOp, cst.getValue());
449     } else {
450       auto resultMap = AffineMap::get(0, map.getNumSymbols(), {e}, ctx);
451       SmallVector<Value, 4> resultOperands = dims;
452       resultOperands.append(symbols.begin(), symbols.end());
453       canonicalizeMapAndOperands(&resultMap, &resultOperands);
454       resultMap = simplifyAffineMap(resultMap);
455       rewriter.replaceOpWithNewOp<AffineApplyOp>(minOp, resultMap,
456                                                  resultOperands);
457     }
458     return success();
459   }
460 
461   return failure();
462 }
463