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/IR/PatternMatch.h"
25 #include "mlir/Pass/Pass.h"
26 #include "mlir/Support/LLVM.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::matchAndRewrite(
115     Operation *op, PatternRewriter &rewriter) const {
116   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
117   if (!linalgOp)
118     return failure();
119   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
120     return failure();
121 
122   Optional<TiledLinalgOp> res = tileLinalgOp(rewriter, linalgOp, options);
123 
124   if (!res)
125     return failure();
126 
127   // New marker if specified.
128   marker.replaceLinalgMarker(rewriter, res->op.getOperation());
129 
130   rewriter.eraseOp(op);
131   return success();
132 }
133 
134 /// Linalg base interchange pattern.
135 mlir::linalg::LinalgBaseInterchangePattern::LinalgBaseInterchangePattern(
136     StringRef opName, MLIRContext *context,
137     ArrayRef<unsigned> interchangeVector, LinalgMarker marker,
138     PatternBenefit benefit)
139     : RewritePattern(opName, {}, benefit, context), marker(marker),
140       interchangeVector(interchangeVector.begin(), interchangeVector.end()) {}
141 
142 LogicalResult mlir::linalg::LinalgBaseInterchangePattern::matchAndRewrite(
143     Operation *op, PatternRewriter &rewriter) const {
144   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
145   if (!linalgOp)
146     return failure();
147   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
148     return failure();
149   if (failed(interchangeGenericLinalgOpPrecondition(op, interchangeVector)))
150     return failure();
151 
152   // TODO: figure out how this interplays with named ops. In particular this
153   // should break the named op property.
154   rewriter.updateRootInPlace(op, [&]() {
155     interchange(linalgOp, interchangeVector);
156     // New marker if specified.
157     marker.replaceLinalgMarker(rewriter, op);
158   });
159   return success();
160 }
161 
162 mlir::linalg::LinalgBasePromotionPattern::LinalgBasePromotionPattern(
163     StringRef opName, MLIRContext *context, LinalgPromotionOptions options,
164     LinalgMarker marker, PatternBenefit benefit)
165     : RewritePattern(opName, {}, benefit, context), marker(marker),
166       options(options) {}
167 
168 LogicalResult mlir::linalg::LinalgBasePromotionPattern::matchAndRewrite(
169     Operation *op, PatternRewriter &rewriter) const {
170   if (failed(marker.checkAndNotify(rewriter, op)))
171     return failure();
172   if (failed(promoteSubviewsPrecondition(op, options)))
173     return failure();
174 
175   // TODO: We cannot use root update here. This pattern is creating other ops,
176   // so if the promotion fails, those need to be cleaned up, which doesnt seem
177   // to be happening here. So to fail properly, we should be cloning the op and
178   // deleting the previous op. This needs more investigation.
179   rewriter.startRootUpdate(op);
180   Optional<LinalgOp> promotedOp = promoteSubViews(rewriter, op, options);
181   if (!promotedOp) {
182     rewriter.cancelRootUpdate(op);
183     return op->emitError("subview promotion failed");
184   }
185   rewriter.finalizeRootUpdate(op);
186   marker.replaceLinalgMarker(rewriter, op);
187   return success();
188 }
189 
190 mlir::linalg::LinalgBaseVectorizationPattern::LinalgBaseVectorizationPattern(
191     StringRef opName, MLIRContext *context, LinalgMarker marker,
192     PatternBenefit benefit)
193     : RewritePattern(opName, {}, benefit, context), marker(marker) {}
194 
195 LogicalResult mlir::linalg::LinalgBaseVectorizationPattern::matchAndRewrite(
196     Operation *op, PatternRewriter &rewriter) const {
197   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
198   if (!linalgOp)
199     return failure();
200   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
201     return failure();
202   if (failed(vectorizeLinalgOpPrecondition(op)))
203     return failure();
204   vectorizeLinalgOp(rewriter, op);
205   rewriter.eraseOp(op);
206   return success();
207 }
208 
209 LogicalResult mlir::linalg::applyStagedPatterns(
210     Operation *op, ArrayRef<OwningRewritePatternList> stage1Patterns,
211     const OwningRewritePatternList &stage2Patterns,
212     function_ref<LogicalResult(Operation *)> stage3Lambda) {
213   unsigned iteration = 0;
214   (void)iteration;
215   for (const auto &patterns : stage1Patterns) {
216     LLVM_DEBUG(DBGS() << "Before 1st stage, iter: " << ++iteration << "\n"
217                       << *op);
218     if (failed(applyPatternsAndFoldGreedily(op, patterns))) {
219       LLVM_DEBUG(DBGS() << "Underlying first stage rewrite did not converge");
220       return failure();
221     }
222     LLVM_DEBUG(DBGS() << "After 1st stage, iter: " << ++iteration << "\n"
223                       << *op);
224     if (failed(applyPatternsAndFoldGreedily(op, stage2Patterns))) {
225       LLVM_DEBUG(DBGS() << "Underlying 2nd stage rewrite did not converge");
226       return failure();
227     }
228     LLVM_DEBUG(DBGS() << "After 2nd stage, iter : " << iteration << "\n"
229                       << *op);
230     if (stage3Lambda) {
231       if (failed(stage3Lambda(op)))
232         return failure();
233       LLVM_DEBUG(DBGS() << "After 3rd stage, iter : " << iteration << "\n"
234                         << *op);
235     }
236   }
237   return success();
238 }
239 
240 /// Traverse `e` and return an AffineExpr where all occurrences of `dim` have
241 /// been replaced by either:
242 ///  - `min` if `positivePath` is true when we reach an occurrence of `dim`
243 ///  - `max` if `positivePath` is true when we reach an occurrence of `dim`
244 /// `positivePath` is negated each time we hit a multiplicative or divisive
245 /// binary op with a constant negative coefficient.
246 static AffineExpr substWithMin(AffineExpr e, AffineExpr dim, AffineExpr min,
247                                AffineExpr max, bool positivePath = true) {
248   if (e == dim)
249     return positivePath ? min : max;
250   if (auto bin = e.dyn_cast<AffineBinaryOpExpr>()) {
251     AffineExpr lhs = bin.getLHS();
252     AffineExpr rhs = bin.getRHS();
253     if (bin.getKind() == mlir::AffineExprKind::Add)
254       return substWithMin(lhs, dim, min, max, positivePath) +
255              substWithMin(rhs, dim, min, max, positivePath);
256 
257     auto c1 = bin.getLHS().dyn_cast<AffineConstantExpr>();
258     auto c2 = bin.getRHS().dyn_cast<AffineConstantExpr>();
259     if (c1 && c1.getValue() < 0)
260       return getAffineBinaryOpExpr(
261           bin.getKind(), c1, substWithMin(rhs, dim, min, max, !positivePath));
262     if (c2 && c2.getValue() < 0)
263       return getAffineBinaryOpExpr(
264           bin.getKind(), substWithMin(lhs, dim, min, max, !positivePath), c2);
265     return getAffineBinaryOpExpr(
266         bin.getKind(), substWithMin(lhs, dim, min, max, positivePath),
267         substWithMin(rhs, dim, min, max, positivePath));
268   }
269   return e;
270 }
271 
272 /// Given the `lbVal`, `ubVal` and `stepVal` of a loop, append `lbVal` and
273 /// `ubVal` to `dims` and `stepVal` to `symbols`.
274 /// Create new AffineDimExpr (`%lb` and `%ub`) and AffineSymbolExpr (`%step`)
275 /// with positions matching the newly appended values. Substitute occurrences of
276 /// `dimExpr` by either the min expression (i.e. `%lb`) or the max expression
277 /// (i.e. `%lb + %step * floordiv(%ub -1 - %lb, %step)`), depending on whether
278 /// the induction variable is used with a positive or negative  coefficient.
279 static AffineExpr substituteLoopInExpr(AffineExpr expr, AffineExpr dimExpr,
280                                        Value lbVal, Value ubVal, Value stepVal,
281                                        SmallVectorImpl<Value> &dims,
282                                        SmallVectorImpl<Value> &symbols) {
283   MLIRContext *ctx = lbVal.getContext();
284   AffineExpr lb = getAffineDimExpr(dims.size(), ctx);
285   dims.push_back(lbVal);
286   AffineExpr ub = getAffineDimExpr(dims.size(), ctx);
287   dims.push_back(ubVal);
288   AffineExpr step = getAffineSymbolExpr(symbols.size(), ctx);
289   symbols.push_back(stepVal);
290   LLVM_DEBUG(DBGS() << "Before: " << expr << "\n");
291   AffineExpr ee = substWithMin(expr, dimExpr, lb,
292                                lb + step * ((ub - 1) - lb).floorDiv(step));
293   LLVM_DEBUG(DBGS() << "After: " << expr << "\n");
294   return ee;
295 }
296 
297 /// Traverse the `dims` and substitute known min or max expressions in place of
298 /// induction variables in `exprs`.
299 static AffineMap substitute(AffineMap map, SmallVectorImpl<Value> &dims,
300                             SmallVectorImpl<Value> &symbols) {
301   auto exprs = llvm::to_vector<4>(map.getResults());
302   for (AffineExpr &expr : exprs) {
303     bool substituted = true;
304     while (substituted) {
305       substituted = false;
306       for (unsigned dimIdx = 0; dimIdx < dims.size(); ++dimIdx) {
307         Value dim = dims[dimIdx];
308         AffineExpr dimExpr = getAffineDimExpr(dimIdx, expr.getContext());
309         LLVM_DEBUG(DBGS() << "Subst: " << dim << " @ " << dimExpr << "\n");
310         AffineExpr substitutedExpr;
311         if (auto forOp = scf::getForInductionVarOwner(dim))
312           substitutedExpr = substituteLoopInExpr(
313               expr, dimExpr, forOp.lowerBound(), forOp.upperBound(),
314               forOp.step(), dims, symbols);
315 
316         if (auto parallelForOp = scf::getParallelForInductionVarOwner(dim))
317           for (unsigned idx = 0, e = parallelForOp.getNumLoops(); idx < e;
318                ++idx)
319             substitutedExpr = substituteLoopInExpr(
320                 expr, dimExpr, parallelForOp.lowerBound()[idx],
321                 parallelForOp.upperBound()[idx], parallelForOp.step()[idx],
322                 dims, symbols);
323 
324         if (!substitutedExpr)
325           continue;
326 
327         substituted = (substitutedExpr != expr);
328         expr = substitutedExpr;
329       }
330     }
331 
332     // Cleanup and simplify the results.
333     // This needs to happen outside of the loop iterating on dims.size() since
334     // it modifies dims.
335     SmallVector<Value, 4> operands(dims.begin(), dims.end());
336     operands.append(symbols.begin(), symbols.end());
337     auto map = AffineMap::get(dims.size(), symbols.size(), exprs,
338                               exprs.front().getContext());
339 
340     LLVM_DEBUG(DBGS() << "Map to simplify: " << map << "\n");
341 
342     // Pull in affine.apply operations and compose them fully into the
343     // result.
344     fullyComposeAffineMapAndOperands(&map, &operands);
345     canonicalizeMapAndOperands(&map, &operands);
346     map = simplifyAffineMap(map);
347     // Assign the results.
348     exprs.assign(map.getResults().begin(), map.getResults().end());
349     dims.assign(operands.begin(), operands.begin() + map.getNumDims());
350     symbols.assign(operands.begin() + map.getNumDims(), operands.end());
351 
352     LLVM_DEBUG(DBGS() << "Map simplified: " << map << "\n");
353   }
354 
355   assert(!exprs.empty() && "Unexpected empty exprs");
356   return AffineMap::get(dims.size(), symbols.size(), exprs, map.getContext());
357 }
358 
359 LogicalResult AffineMinSCFCanonicalizationPattern::matchAndRewrite(
360     AffineMinOp minOp, PatternRewriter &rewriter) const {
361   LLVM_DEBUG(DBGS() << "Canonicalize AffineMinSCF: " << *minOp.getOperation()
362                     << "\n");
363 
364   SmallVector<Value, 4> dims(minOp.getDimOperands()),
365       symbols(minOp.getSymbolOperands());
366   AffineMap map = substitute(minOp.getAffineMap(), dims, symbols);
367 
368   LLVM_DEBUG(DBGS() << "Resulting map: " << map << "\n");
369 
370   // Check whether any of the expressions, when subtracted from all other
371   // expressions, produces only >= 0 constants. If so, it is the min.
372   for (auto e : minOp.getAffineMap().getResults()) {
373     LLVM_DEBUG(DBGS() << "Candidate min: " << e << "\n");
374     if (!e.isSymbolicOrConstant())
375       continue;
376 
377     auto isNonPositive = [](AffineExpr e) {
378       if (auto cst = e.dyn_cast<AffineConstantExpr>())
379         return cst.getValue() < 0;
380       return true;
381     };
382 
383     // Build the subMap and check everything is statically known to be
384     // positive.
385     SmallVector<AffineExpr, 4> subExprs;
386     subExprs.reserve(map.getNumResults());
387     for (auto ee : map.getResults())
388       subExprs.push_back(ee - e);
389     MLIRContext *ctx = minOp.getContext();
390     AffineMap subMap = simplifyAffineMap(
391         AffineMap::get(map.getNumDims(), map.getNumSymbols(), subExprs, ctx));
392     LLVM_DEBUG(DBGS() << "simplified subMap: " << subMap << "\n");
393     if (llvm::any_of(subMap.getResults(), isNonPositive))
394       continue;
395 
396     // Static min found.
397     if (auto cst = e.dyn_cast<AffineConstantExpr>()) {
398       rewriter.replaceOpWithNewOp<ConstantIndexOp>(minOp, cst.getValue());
399     } else {
400       auto resultMap = AffineMap::get(0, map.getNumSymbols(), {e}, ctx);
401       SmallVector<Value, 4> resultOperands = dims;
402       resultOperands.append(symbols.begin(), symbols.end());
403       canonicalizeMapAndOperands(&resultMap, &resultOperands);
404       resultMap = simplifyAffineMap(resultMap);
405       rewriter.replaceOpWithNewOp<AffineApplyOp>(minOp, resultMap,
406                                                  resultOperands);
407     }
408     return success();
409   }
410 
411   return failure();
412 }
413