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 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern(
115     LinalgTilingOptions options, LinalgMarker marker, PatternBenefit benefit)
116     : RewritePattern(benefit, MatchAnyOpTypeTag()), marker(marker),
117       options(options) {}
118 
119 LogicalResult mlir::linalg::LinalgBaseTilingPattern::matchAndRewriteBase(
120     Operation *op, PatternRewriter &rewriter, TiledLinalgOp &result) const {
121   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
122   if (!linalgOp)
123     return failure();
124   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
125     return failure();
126 
127   Optional<TiledLinalgOp> res = tileLinalgOp(rewriter, linalgOp, options);
128 
129   if (!res)
130     return failure();
131 
132   // Return relevant information to derived pattern.
133   result = *res;
134 
135   // New marker if specified.
136   marker.replaceLinalgMarker(rewriter, res->op.getOperation());
137   return success();
138 }
139 
140 mlir::linalg::LinalgBaseTileAndFusePattern::LinalgBaseTileAndFusePattern(
141     StringRef opName, MLIRContext *context,
142     const LinalgDependenceGraph &dependenceGraph,
143     LinalgTilingOptions tilingOptions, LinalgFusionOptions fusionOptions,
144     LinalgMarker marker, LinalgMarker fusedOpMarker,
145     LinalgMarker originalOpMarker, PatternBenefit benefit)
146     : RewritePattern(opName, {}, benefit, context),
147       dependenceGraph(dependenceGraph), tilingOptions(tilingOptions),
148       fusionOptions(fusionOptions), marker(marker),
149       fusedOpMarker(fusedOpMarker), originalOpMarker(originalOpMarker) {}
150 
151 LogicalResult mlir::linalg::LinalgBaseTileAndFusePattern::matchAndRewrite(
152     Operation *op, PatternRewriter &rewriter) const {
153   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
154   if (!linalgOp)
155     return failure();
156   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
157     return failure();
158   if (!linalgOp.hasBufferSemantics())
159     return failure();
160 
161   DenseSet<Operation *> producers;
162   producers.insert(linalgOp);
163   for (auto dependence : dependenceGraph.getDependentOperations(linalgOp)) {
164     if (!fusionOptions.indicesToFuse.count(
165             dependence.indexingOpView->getOperandNumber()))
166       continue;
167     if (isa<LinalgOp>(dependence.dependentOpView->getOwner()))
168       producers.insert(dependence.dependentOpView->getOwner());
169   }
170 
171   SmallVector<LinalgOp, 1> fusionOps;
172   for (auto it = op->getBlock()->begin(), ie = Block::iterator(op); it != ie;
173        ++it) {
174     auto producerLinalgOp = dyn_cast<LinalgOp>(&(*it));
175     if (producerLinalgOp && producers.count(producerLinalgOp))
176       fusionOps.push_back(producerLinalgOp);
177   }
178   fusionOps.push_back(linalgOp);
179 
180   SmallVector<Value, 4> tileSizes =
181       tilingOptions.tileSizeComputationFunction(rewriter, op);
182   LinalgTilingOptions instanceTilingOptions = tilingOptions;
183   instanceTilingOptions.setTileSizes(tileSizes);
184   Optional<TiledAndFusedLinalgOps> tiledAndFusedOps = tileAndFuseLinalgOps(
185       rewriter, fusionOps, dependenceGraph, instanceTilingOptions);
186   if (!tiledAndFusedOps)
187     return failure();
188 
189   // Tile the unfused loops;
190   SmallVector<Value, 4> unfusedLoopTileSizes;
191   Value zero = rewriter.create<ConstantIndexOp>(op->getLoc(), 0);
192   for (auto tileSize : enumerate(tileSizes)) {
193     if (tiledAndFusedOps->fusedLoopDims.count(tileSize.index()))
194       unfusedLoopTileSizes.push_back(zero);
195     else
196       unfusedLoopTileSizes.push_back(tileSize.value());
197   }
198   // Tile the loop only if there is a non-zero tile size.
199   if (unfusedLoopTileSizes.size() > linalgOp.getNumLoops())
200     unfusedLoopTileSizes.resize(linalgOp.getNumLoops());
201   if (llvm::any_of(unfusedLoopTileSizes, [](Value val) {
202         if (auto cst = val.getDefiningOp<ConstantIndexOp>())
203           return cst.getValue() != 0;
204         return true;
205       })) {
206     LinalgTilingOptions unfusedTilingOptions = tilingOptions;
207     unfusedTilingOptions.setTileSizes(unfusedLoopTileSizes);
208     Optional<TiledLinalgOp> unfusedTiledOp =
209         tileLinalgOp(rewriter, tiledAndFusedOps->op, unfusedTilingOptions);
210     if (!unfusedTiledOp)
211       return failure();
212     rewriter.eraseOp(tiledAndFusedOps->op);
213     tiledAndFusedOps->op = unfusedTiledOp->op;
214   }
215 
216   marker.replaceLinalgMarker(rewriter, tiledAndFusedOps->op.getOperation());
217   for (auto fusedOp : tiledAndFusedOps->fusedProducers) {
218     fusedOpMarker.replaceLinalgMarker(rewriter, fusedOp.getOperation());
219   }
220   for (auto origProducerOp : ArrayRef<LinalgOp>(fusionOps).drop_back()) {
221     originalOpMarker.replaceLinalgMarker(rewriter,
222                                          origProducerOp.getOperation());
223   }
224   rewriter.updateRootInPlace(
225       op, [&]() { originalOpMarker.replaceLinalgMarker(rewriter, op); });
226   return success();
227 }
228 
229 /// Linalg base interchange pattern.
230 mlir::linalg::LinalgBaseInterchangePattern::LinalgBaseInterchangePattern(
231     StringRef opName, MLIRContext *context,
232     ArrayRef<unsigned> interchangeVector, LinalgMarker marker,
233     PatternBenefit benefit)
234     : RewritePattern(opName, {}, benefit, context), marker(marker),
235       interchangeVector(interchangeVector.begin(), interchangeVector.end()) {}
236 
237 LogicalResult mlir::linalg::LinalgBaseInterchangePattern::matchAndRewrite(
238     Operation *op, PatternRewriter &rewriter) const {
239   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
240   if (!linalgOp)
241     return failure();
242   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
243     return failure();
244   if (failed(interchangeGenericLinalgOpPrecondition(op, interchangeVector)))
245     return failure();
246 
247   // TODO: figure out how this interplays with named ops. In particular this
248   // should break the named op property.
249   rewriter.updateRootInPlace(op, [&]() {
250     interchange(linalgOp, interchangeVector);
251     // New marker if specified.
252     marker.replaceLinalgMarker(rewriter, op);
253   });
254   return success();
255 }
256 
257 mlir::linalg::LinalgBasePromotionPattern::LinalgBasePromotionPattern(
258     StringRef opName, MLIRContext *context, LinalgPromotionOptions options,
259     LinalgMarker marker, PatternBenefit benefit)
260     : RewritePattern(opName, {}, benefit, context), marker(marker),
261       options(options) {}
262 
263 LogicalResult mlir::linalg::LinalgBasePromotionPattern::matchAndRewrite(
264     Operation *op, PatternRewriter &rewriter) const {
265   if (failed(marker.checkAndNotify(rewriter, op)))
266     return failure();
267   if (failed(promoteSubviewsPrecondition(op, options)))
268     return failure();
269 
270   // TODO: We cannot use root update here. This pattern is creating other ops,
271   // so if the promotion fails, those need to be cleaned up, which doesnt seem
272   // to be happening here. So to fail properly, we should be cloning the op and
273   // deleting the previous op. This needs more investigation.
274   rewriter.startRootUpdate(op);
275   Optional<LinalgOp> promotedOp = promoteSubViews(rewriter, op, options);
276   if (!promotedOp) {
277     rewriter.cancelRootUpdate(op);
278     return op->emitError("subview promotion failed");
279   }
280   rewriter.finalizeRootUpdate(op);
281   marker.replaceLinalgMarker(rewriter, op);
282   return success();
283 }
284 
285 mlir::linalg::LinalgBaseVectorizationPattern::LinalgBaseVectorizationPattern(
286     StringRef opName, MLIRContext *context, LinalgMarker marker,
287     PatternBenefit benefit)
288     : RewritePattern(opName, {}, benefit, context), marker(marker) {}
289 
290 LogicalResult mlir::linalg::LinalgBaseVectorizationPattern::matchAndRewrite(
291     Operation *op, PatternRewriter &rewriter) const {
292   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
293   if (!linalgOp)
294     return failure();
295   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
296     return failure();
297   if (failed(vectorizeLinalgOpPrecondition(op)))
298     return failure();
299   vectorizeLinalgOp(rewriter, op);
300   rewriter.eraseOp(op);
301   return success();
302 }
303 
304 LogicalResult mlir::linalg::applyStagedPatterns(
305     Operation *op, ArrayRef<FrozenRewritePatternList> stage1Patterns,
306     const FrozenRewritePatternList &stage2Patterns,
307     function_ref<LogicalResult(Operation *)> stage3Lambda) {
308   unsigned iteration = 0;
309   (void)iteration;
310   for (const auto &patterns : stage1Patterns) {
311     LLVM_DEBUG(DBGS() << "Before 1st stage, iter: " << ++iteration << "\n"
312                       << *op);
313     if (failed(applyPatternsAndFoldGreedily(op, patterns))) {
314       LLVM_DEBUG(DBGS() << "Underlying first stage rewrite did not converge");
315       return failure();
316     }
317     LLVM_DEBUG(DBGS() << "After 1st stage, iter: " << ++iteration << "\n"
318                       << *op);
319     if (failed(applyPatternsAndFoldGreedily(op, stage2Patterns))) {
320       LLVM_DEBUG(DBGS() << "Underlying 2nd stage rewrite did not converge");
321       return failure();
322     }
323     LLVM_DEBUG(DBGS() << "After 2nd stage, iter : " << iteration << "\n"
324                       << *op);
325     if (stage3Lambda) {
326       if (failed(stage3Lambda(op)))
327         return failure();
328       LLVM_DEBUG(DBGS() << "After 3rd stage, iter : " << iteration << "\n"
329                         << *op);
330     }
331   }
332   return success();
333 }
334 
335 /// Traverse `e` and return an AffineExpr where all occurrences of `dim` have
336 /// been replaced by either:
337 ///  - `min` if `positivePath` is true when we reach an occurrence of `dim`
338 ///  - `max` if `positivePath` is true when we reach an occurrence of `dim`
339 /// `positivePath` is negated each time we hit a multiplicative or divisive
340 /// binary op with a constant negative coefficient.
341 static AffineExpr substWithMin(AffineExpr e, AffineExpr dim, AffineExpr min,
342                                AffineExpr max, bool positivePath = true) {
343   if (e == dim)
344     return positivePath ? min : max;
345   if (auto bin = e.dyn_cast<AffineBinaryOpExpr>()) {
346     AffineExpr lhs = bin.getLHS();
347     AffineExpr rhs = bin.getRHS();
348     if (bin.getKind() == mlir::AffineExprKind::Add)
349       return substWithMin(lhs, dim, min, max, positivePath) +
350              substWithMin(rhs, dim, min, max, positivePath);
351 
352     auto c1 = bin.getLHS().dyn_cast<AffineConstantExpr>();
353     auto c2 = bin.getRHS().dyn_cast<AffineConstantExpr>();
354     if (c1 && c1.getValue() < 0)
355       return getAffineBinaryOpExpr(
356           bin.getKind(), c1, substWithMin(rhs, dim, min, max, !positivePath));
357     if (c2 && c2.getValue() < 0)
358       return getAffineBinaryOpExpr(
359           bin.getKind(), substWithMin(lhs, dim, min, max, !positivePath), c2);
360     return getAffineBinaryOpExpr(
361         bin.getKind(), substWithMin(lhs, dim, min, max, positivePath),
362         substWithMin(rhs, dim, min, max, positivePath));
363   }
364   return e;
365 }
366 
367 /// Given the `lbVal`, `ubVal` and `stepVal` of a loop, append `lbVal` and
368 /// `ubVal` to `dims` and `stepVal` to `symbols`.
369 /// Create new AffineDimExpr (`%lb` and `%ub`) and AffineSymbolExpr (`%step`)
370 /// with positions matching the newly appended values. Substitute occurrences of
371 /// `dimExpr` by either the min expression (i.e. `%lb`) or the max expression
372 /// (i.e. `%lb + %step * floordiv(%ub -1 - %lb, %step)`), depending on whether
373 /// the induction variable is used with a positive or negative  coefficient.
374 static AffineExpr substituteLoopInExpr(AffineExpr expr, AffineExpr dimExpr,
375                                        Value lbVal, Value ubVal, Value stepVal,
376                                        SmallVectorImpl<Value> &dims,
377                                        SmallVectorImpl<Value> &symbols) {
378   MLIRContext *ctx = lbVal.getContext();
379   AffineExpr lb = getAffineDimExpr(dims.size(), ctx);
380   dims.push_back(lbVal);
381   AffineExpr ub = getAffineDimExpr(dims.size(), ctx);
382   dims.push_back(ubVal);
383   AffineExpr step = getAffineSymbolExpr(symbols.size(), ctx);
384   symbols.push_back(stepVal);
385   LLVM_DEBUG(DBGS() << "Before: " << expr << "\n");
386   AffineExpr ee = substWithMin(expr, dimExpr, lb,
387                                lb + step * ((ub - 1) - lb).floorDiv(step));
388   LLVM_DEBUG(DBGS() << "After: " << expr << "\n");
389   return ee;
390 }
391 
392 /// Traverse the `dims` and substitute known min or max expressions in place of
393 /// induction variables in `exprs`.
394 static AffineMap substitute(AffineMap map, SmallVectorImpl<Value> &dims,
395                             SmallVectorImpl<Value> &symbols) {
396   auto exprs = llvm::to_vector<4>(map.getResults());
397   for (AffineExpr &expr : exprs) {
398     bool substituted = true;
399     while (substituted) {
400       substituted = false;
401       for (unsigned dimIdx = 0; dimIdx < dims.size(); ++dimIdx) {
402         Value dim = dims[dimIdx];
403         AffineExpr dimExpr = getAffineDimExpr(dimIdx, expr.getContext());
404         LLVM_DEBUG(DBGS() << "Subst: " << dim << " @ " << dimExpr << "\n");
405         AffineExpr substitutedExpr;
406         if (auto forOp = scf::getForInductionVarOwner(dim))
407           substitutedExpr = substituteLoopInExpr(
408               expr, dimExpr, forOp.lowerBound(), forOp.upperBound(),
409               forOp.step(), dims, symbols);
410 
411         if (auto parallelForOp = scf::getParallelForInductionVarOwner(dim))
412           for (unsigned idx = 0, e = parallelForOp.getNumLoops(); idx < e;
413                ++idx)
414             substitutedExpr = substituteLoopInExpr(
415                 expr, dimExpr, parallelForOp.lowerBound()[idx],
416                 parallelForOp.upperBound()[idx], parallelForOp.step()[idx],
417                 dims, symbols);
418 
419         if (!substitutedExpr)
420           continue;
421 
422         substituted = (substitutedExpr != expr);
423         expr = substitutedExpr;
424       }
425     }
426 
427     // Cleanup and simplify the results.
428     // This needs to happen outside of the loop iterating on dims.size() since
429     // it modifies dims.
430     SmallVector<Value, 4> operands(dims.begin(), dims.end());
431     operands.append(symbols.begin(), symbols.end());
432     auto map = AffineMap::get(dims.size(), symbols.size(), exprs,
433                               exprs.front().getContext());
434 
435     LLVM_DEBUG(DBGS() << "Map to simplify: " << map << "\n");
436 
437     // Pull in affine.apply operations and compose them fully into the
438     // result.
439     fullyComposeAffineMapAndOperands(&map, &operands);
440     canonicalizeMapAndOperands(&map, &operands);
441     map = simplifyAffineMap(map);
442     // Assign the results.
443     exprs.assign(map.getResults().begin(), map.getResults().end());
444     dims.assign(operands.begin(), operands.begin() + map.getNumDims());
445     symbols.assign(operands.begin() + map.getNumDims(), operands.end());
446 
447     LLVM_DEBUG(DBGS() << "Map simplified: " << map << "\n");
448   }
449 
450   assert(!exprs.empty() && "Unexpected empty exprs");
451   return AffineMap::get(dims.size(), symbols.size(), exprs, map.getContext());
452 }
453 
454 LogicalResult AffineMinSCFCanonicalizationPattern::matchAndRewrite(
455     AffineMinOp minOp, PatternRewriter &rewriter) const {
456   LLVM_DEBUG(DBGS() << "Canonicalize AffineMinSCF: " << *minOp.getOperation()
457                     << "\n");
458 
459   SmallVector<Value, 4> dims(minOp.getDimOperands()),
460       symbols(minOp.getSymbolOperands());
461   AffineMap map = substitute(minOp.getAffineMap(), dims, symbols);
462 
463   LLVM_DEBUG(DBGS() << "Resulting map: " << map << "\n");
464 
465   // Check whether any of the expressions, when subtracted from all other
466   // expressions, produces only >= 0 constants. If so, it is the min.
467   for (auto e : minOp.getAffineMap().getResults()) {
468     LLVM_DEBUG(DBGS() << "Candidate min: " << e << "\n");
469     if (!e.isSymbolicOrConstant())
470       continue;
471 
472     auto isNonPositive = [](AffineExpr e) {
473       if (auto cst = e.dyn_cast<AffineConstantExpr>())
474         return cst.getValue() < 0;
475       return true;
476     };
477 
478     // Build the subMap and check everything is statically known to be
479     // positive.
480     SmallVector<AffineExpr, 4> subExprs;
481     subExprs.reserve(map.getNumResults());
482     for (auto ee : map.getResults())
483       subExprs.push_back(ee - e);
484     MLIRContext *ctx = minOp.getContext();
485     AffineMap subMap = simplifyAffineMap(
486         AffineMap::get(map.getNumDims(), map.getNumSymbols(), subExprs, ctx));
487     LLVM_DEBUG(DBGS() << "simplified subMap: " << subMap << "\n");
488     if (llvm::any_of(subMap.getResults(), isNonPositive))
489       continue;
490 
491     // Static min found.
492     if (auto cst = e.dyn_cast<AffineConstantExpr>()) {
493       rewriter.replaceOpWithNewOp<ConstantIndexOp>(minOp, cst.getValue());
494     } else {
495       auto resultMap = AffineMap::get(0, map.getNumSymbols(), {e}, ctx);
496       SmallVector<Value, 4> resultOperands = dims;
497       resultOperands.append(symbols.begin(), symbols.end());
498       canonicalizeMapAndOperands(&resultMap, &resultOperands);
499       resultMap = simplifyAffineMap(resultMap);
500       rewriter.replaceOpWithNewOp<AffineApplyOp>(minOp, resultMap,
501                                                  resultOperands);
502     }
503     return success();
504   }
505 
506   return failure();
507 }
508