1 | //===- EraseUnusedOperandsAndResults.cpp ----------------------------------===// |
---|---|
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 | #include "mlir/Dialect/Linalg/Transforms/Transforms.h" |
10 | |
11 | #include "mlir/Dialect/Linalg/IR/Linalg.h" |
12 | |
13 | using namespace mlir; |
14 | using namespace mlir::linalg; |
15 | |
16 | /// Return `true` if the `result` of an operation `genericOp` is dead. |
17 | static bool isResultValueDead(linalg::GenericOp genericOp, OpResult result) { |
18 | if (!result.use_empty()) |
19 | return false; |
20 | // If out operand not used in payload, we can drop it. |
21 | OpOperand *outputOpOperand = |
22 | genericOp.getDpsInitOperand(result.getResultNumber()); |
23 | if (!genericOp.payloadUsesValueFromOperand(outputOpOperand)) |
24 | return true; |
25 | |
26 | // The out operand that is part of a payload can be dropped if |
27 | // these conditions are met: |
28 | // - Result from out operand is dead. |
29 | // - User of arg is yield. |
30 | // - outArg data is not being used by other outArgs. |
31 | |
32 | // Check block arg and cycle from out operand has a single use. |
33 | BlockArgument outputArg = |
34 | genericOp.getRegionOutputArgs()[result.getResultNumber()]; |
35 | if (!outputArg.hasOneUse()) |
36 | return false; |
37 | Operation *argUserOp = *outputArg.user_begin(); |
38 | |
39 | // Check argUser has no other use. |
40 | if (!argUserOp->use_empty()) |
41 | return false; |
42 | |
43 | // Check that argUser is a yield. |
44 | auto yieldOp = dyn_cast<linalg::YieldOp>(argUserOp); |
45 | if (!yieldOp) |
46 | return false; |
47 | |
48 | // Check outArg data is not being used by other outArgs. |
49 | if (yieldOp.getOperand(result.getResultNumber()) != outputArg) |
50 | return false; |
51 | |
52 | return true; |
53 | } |
54 | |
55 | //===---------------------------------------------------------------------===// |
56 | // Helper methods for operand deduplication and dead results elimination |
57 | //===---------------------------------------------------------------------===// |
58 | |
59 | // Deduplicate input operands, and return the |
60 | // - Mapping from operand position in the original op, to operand position in |
61 | // the canonicalized op. |
62 | // - The preserved input operands list (by reference). |
63 | llvm::SmallDenseMap<unsigned, unsigned> static deduplicateInputOperands( |
64 | GenericOp genericOp, SmallVector<OpOperand *> &droppedOpOperands, |
65 | SmallVector<Value> &newInputOperands, |
66 | SmallVector<AffineMap> &newIndexingMaps) { |
67 | llvm::SmallDenseMap<unsigned, unsigned> origToNewPos; |
68 | llvm::SmallDenseMap<std::pair<Value, AffineMap>, unsigned> dedupedInputs; |
69 | for (const auto &en : llvm::enumerate(genericOp.getDpsInputOperands())) { |
70 | OpOperand *inputOpOperand = en.value(); |
71 | // Check if operand is dead and if dropping the indexing map makes the |
72 | // loops to shape computation invalid. |
73 | if (!genericOp.payloadUsesValueFromOperand(inputOpOperand)) { |
74 | // Add the current operands to the list of potentially droppable |
75 | // operands. If it cannot be dropped, this needs to be popped back. |
76 | droppedOpOperands.push_back(inputOpOperand); |
77 | if (genericOp.canOpOperandsBeDropped(droppedOpOperands)) |
78 | continue; |
79 | droppedOpOperands.pop_back(); |
80 | } |
81 | |
82 | // Check if this operand is a duplicate. |
83 | AffineMap indexingMap = genericOp.getMatchingIndexingMap(inputOpOperand); |
84 | auto it = |
85 | dedupedInputs.find(std::make_pair(inputOpOperand->get(), indexingMap)); |
86 | if (it != dedupedInputs.end()) { |
87 | origToNewPos[en.index()] = it->second; |
88 | droppedOpOperands.push_back(inputOpOperand); |
89 | continue; |
90 | } |
91 | |
92 | // This is a preserved argument. |
93 | origToNewPos[en.index()] = newInputOperands.size(); |
94 | dedupedInputs[{inputOpOperand->get(), indexingMap}] = |
95 | newInputOperands.size(); |
96 | newInputOperands.push_back(inputOpOperand->get()); |
97 | newIndexingMaps.push_back(indexingMap); |
98 | } |
99 | return origToNewPos; |
100 | } |
101 | |
102 | // Deduplicate output operands, and return the |
103 | // - Mapping from operand position in the original op, to operand position in |
104 | // the canonicalized op. |
105 | // - The preserved output operands list (by reference). |
106 | llvm::SmallDenseMap<unsigned, unsigned> static deduplicateOutputOperands( |
107 | GenericOp genericOp, SmallVector<OpOperand *> &droppedOpOperands, |
108 | SmallVector<Value> &newOutputOperands, |
109 | SmallVector<AffineMap> &newIndexingMaps, bool removeOutputs) { |
110 | llvm::SmallDenseMap<unsigned, unsigned> origToNewPos; |
111 | llvm::SmallDenseMap<std::tuple<Value, AffineMap, Value>, unsigned> |
112 | dedupedOutpts; |
113 | // If the op doesn't have tensor semantics or outputs should not be removed, |
114 | // keep all the outputs as preserved. |
115 | if (!genericOp.hasPureTensorSemantics() || !removeOutputs) { |
116 | for (const auto &en : llvm::enumerate(genericOp.getDpsInitsMutable())) { |
117 | origToNewPos[en.index()] = newOutputOperands.size(); |
118 | newOutputOperands.push_back(en.value().get()); |
119 | newIndexingMaps.push_back(genericOp.getMatchingIndexingMap(&en.value())); |
120 | } |
121 | return origToNewPos; |
122 | } |
123 | // Output argument can be dropped if the result has |
124 | // - no users, and |
125 | // - it is not used in the payload, and |
126 | // - the corresponding indexing maps are not needed for loop bound |
127 | // computation. |
128 | auto yieldOp = cast<YieldOp>(genericOp.getBody()->getTerminator()); |
129 | for (const auto &outputOpOperand : |
130 | llvm::enumerate(genericOp.getDpsInitsMutable())) { |
131 | OpResult result = genericOp.getTiedOpResult(&outputOpOperand.value()); |
132 | AffineMap indexingMap = |
133 | genericOp.getMatchingIndexingMap(&outputOpOperand.value()); |
134 | auto key = std::make_tuple(outputOpOperand.value().get(), indexingMap, |
135 | yieldOp->getOperand(outputOpOperand.index())); |
136 | if (isResultValueDead(genericOp, result)) { |
137 | // Check if the opoperand can be dropped without affecting loop |
138 | // bound computation. Add the operand to the list of dropped op |
139 | // operand for checking. If it cannot be dropped, need to pop the |
140 | // value back. |
141 | droppedOpOperands.push_back(&outputOpOperand.value()); |
142 | if (genericOp.canOpOperandsBeDropped(droppedOpOperands)) { |
143 | continue; |
144 | } |
145 | droppedOpOperands.pop_back(); |
146 | } |
147 | |
148 | if (!genericOp.payloadUsesValueFromOperand(&outputOpOperand.value())) { |
149 | // The out operand can also be dropped if it is computed redundantly |
150 | // by another result, the conditions for that are |
151 | // - The same operand is used as the out operand |
152 | // - The same indexing map is used |
153 | // - The same yield value is used. |
154 | auto it = dedupedOutpts.find(key); |
155 | if (it != dedupedOutpts.end()) { |
156 | origToNewPos[outputOpOperand.index()] = it->second; |
157 | droppedOpOperands.push_back(&outputOpOperand.value()); |
158 | continue; |
159 | } |
160 | } |
161 | |
162 | origToNewPos[outputOpOperand.index()] = newOutputOperands.size(); |
163 | dedupedOutpts[key] = newOutputOperands.size(); |
164 | newOutputOperands.push_back(outputOpOperand.value().get()); |
165 | newIndexingMaps.push_back( |
166 | genericOp.getMatchingIndexingMap(&outputOpOperand.value())); |
167 | } |
168 | return origToNewPos; |
169 | } |
170 | |
171 | // Populate the body of the canonicalized operation. |
172 | static void populateOpPayload( |
173 | GenericOp genericOp, GenericOp newOp, |
174 | const llvm::SmallDenseMap<unsigned, unsigned> &origInsToNewInsPos, |
175 | const llvm::SmallDenseMap<unsigned, unsigned> &origOutsToNewOutsPos, |
176 | RewriterBase &rewriter) { |
177 | // Merge the body of the original op with the new op. |
178 | Block *newOpBlock = &newOp.getRegion().front(); |
179 | assert(newOpBlock->empty() && "expected new op to have an empty payload"); |
180 | Block *origOpBlock = &genericOp.getRegion().front(); |
181 | SmallVector<Value> replacements(origOpBlock->getNumArguments(), nullptr); |
182 | |
183 | // Replace all arguments in the original op, with arguments from the |
184 | // canonicalized op. |
185 | auto updateReplacements = |
186 | [&](SmallVector<OpOperand *> &origOperands, |
187 | SmallVector<OpOperand *> &newOperands, |
188 | const llvm::SmallDenseMap<unsigned, unsigned> &map) { |
189 | for (const auto &origOperand : llvm::enumerate(First&: origOperands)) { |
190 | auto it = map.find(Val: origOperand.index()); |
191 | if (it == map.end()) |
192 | continue; |
193 | OpOperand *newOperand = newOperands[it->second]; |
194 | replacements[origOperand.value()->getOperandNumber()] = |
195 | newOpBlock->getArgument(i: newOperand->getOperandNumber()); |
196 | } |
197 | }; |
198 | |
199 | SmallVector<OpOperand *> origInputOperands = genericOp.getDpsInputOperands(); |
200 | SmallVector<OpOperand *> newInputOperands = newOp.getDpsInputOperands(); |
201 | updateReplacements(origInputOperands, newInputOperands, origInsToNewInsPos); |
202 | |
203 | SmallVector<OpOperand *> origOutputOperands = llvm::to_vector(llvm::map_range( |
204 | genericOp.getDpsInitsMutable(), [](OpOperand &o) { return &o; })); |
205 | SmallVector<OpOperand *> newOutputOperands = llvm::to_vector(llvm::map_range( |
206 | newOp.getDpsInitsMutable(), [](OpOperand &o) { return &o; })); |
207 | updateReplacements(origOutputOperands, newOutputOperands, |
208 | origOutsToNewOutsPos); |
209 | |
210 | // Drop the unused yield args. |
211 | if (newOp.getNumDpsInits() != genericOp.getNumDpsInits()) { |
212 | OpBuilder::InsertionGuard g(rewriter); |
213 | YieldOp origYieldOp = cast<YieldOp>(origOpBlock->getTerminator()); |
214 | rewriter.setInsertionPoint(origYieldOp); |
215 | |
216 | SmallVector<Value> newYieldVals(newOp.getNumDpsInits(), nullptr); |
217 | for (const auto &yieldOpOperands : |
218 | llvm::enumerate(origYieldOp.getValues())) { |
219 | auto it = origOutsToNewOutsPos.find(yieldOpOperands.index()); |
220 | if (it == origOutsToNewOutsPos.end()) |
221 | continue; |
222 | newYieldVals[it->second] = yieldOpOperands.value(); |
223 | } |
224 | rewriter.replaceOpWithNewOp<YieldOp>(origYieldOp, newYieldVals); |
225 | } |
226 | |
227 | rewriter.mergeBlocks(source: origOpBlock, dest: newOpBlock, argValues: replacements); |
228 | } |
229 | |
230 | FailureOr<linalg::GenericOp> |
231 | mlir::linalg::deduplicateOperandsAndRemoveDeadResults( |
232 | RewriterBase &rewriter, linalg::GenericOp genericOp, bool removeOutputs) { |
233 | // Create a map from argument position in the original op to the argument |
234 | // position in the new op. If the argument is dropped it wont have an entry. |
235 | SmallVector<OpOperand *> droppedOpOperands; |
236 | |
237 | // Information needed to build the new op. |
238 | SmallVector<Value> newInputOperands, newOutputOperands; |
239 | SmallVector<AffineMap> newIndexingMaps; |
240 | |
241 | // Gather information about duplicate input operands. |
242 | llvm::SmallDenseMap<unsigned, unsigned> origInsToNewInsPos = |
243 | deduplicateInputOperands(genericOp, droppedOpOperands, newInputOperands, |
244 | newIndexingMaps); |
245 | |
246 | // Gather information about the dropped outputs. |
247 | llvm::SmallDenseMap<unsigned, unsigned> origOutsToNewOutsPos = |
248 | deduplicateOutputOperands(genericOp, droppedOpOperands, newOutputOperands, |
249 | newIndexingMaps, removeOutputs); |
250 | |
251 | // Check if there is any change to operands. |
252 | if (newInputOperands.size() + newOutputOperands.size() == |
253 | genericOp->getNumOperands()) |
254 | return genericOp; |
255 | |
256 | // Create the new op with the body being empty. |
257 | Location loc = genericOp.getLoc(); |
258 | SmallVector<Type> newResultTypes; |
259 | for (Value v : newOutputOperands) |
260 | if (isa<TensorType>(Val: v.getType())) |
261 | newResultTypes.push_back(Elt: v.getType()); |
262 | auto newOp = rewriter.create<GenericOp>( |
263 | loc, newResultTypes, newInputOperands, newOutputOperands, |
264 | rewriter.getAffineMapArrayAttr(newIndexingMaps), |
265 | genericOp.getIteratorTypes(), genericOp.getDocAttr(), |
266 | genericOp.getLibraryCallAttr(), |
267 | [](OpBuilder & /*builder*/, Location /*loc*/, ValueRange /*args*/) { |
268 | return; |
269 | }); |
270 | // Copy over unknown attributes. They might be load bearing for some flow. |
271 | ArrayRef<StringRef> odsAttrs = genericOp.getAttributeNames(); |
272 | for (NamedAttribute kv : genericOp->getAttrs()) |
273 | if (!llvm::is_contained(odsAttrs, kv.getName().getValue())) |
274 | newOp->setAttr(kv.getName(), kv.getValue()); |
275 | |
276 | // Fix up the payload of the canonicalized operation. |
277 | populateOpPayload(genericOp, newOp, origInsToNewInsPos, origOutsToNewOutsPos, |
278 | rewriter); |
279 | |
280 | // Replace all live uses of the op. |
281 | SmallVector<Value> replacementsVals(genericOp->getNumResults(), nullptr); |
282 | for (const auto &result : llvm::enumerate(genericOp.getResults())) { |
283 | auto it = origOutsToNewOutsPos.find(result.index()); |
284 | if (it == origOutsToNewOutsPos.end()) |
285 | continue; |
286 | replacementsVals[result.index()] = newOp.getResult(it->second); |
287 | } |
288 | rewriter.replaceOp(genericOp, replacementsVals); |
289 | return newOp; |
290 | } |
291 | |
292 | namespace { |
293 | |
294 | struct DeduplicateAndRemoveDeadOperandsAndResults |
295 | : public OpRewritePattern<GenericOp> { |
296 | DeduplicateAndRemoveDeadOperandsAndResults(MLIRContext *ctx, |
297 | bool removeOutputs) |
298 | : OpRewritePattern<GenericOp>(ctx), removeOutputs(removeOutputs) {} |
299 | |
300 | LogicalResult matchAndRewrite(GenericOp genericOp, |
301 | PatternRewriter &rewriter) const override { |
302 | FailureOr<GenericOp> newOp = deduplicateOperandsAndRemoveDeadResults( |
303 | rewriter, genericOp, removeOutputs); |
304 | if (failed(Result: newOp) || newOp.value() == genericOp) { |
305 | return rewriter.notifyMatchFailure( |
306 | genericOp, "failed to dedup operands/remove dead results"); |
307 | } |
308 | return success(); |
309 | } |
310 | |
311 | private: |
312 | /// If unset, outputs are not modified by this pattern. |
313 | bool removeOutputs; |
314 | }; |
315 | |
316 | /// Remove unused cycles. |
317 | /// We can remove unused cycle within a payload of generic region |
318 | /// if these conditions are met: |
319 | /// - Result from out operand is dead. |
320 | /// - Block arg from out operand has a single use in the %cycle |
321 | /// instruction. |
322 | /// - Cycle has a single use and it is in yield. |
323 | struct RemoveUnusedCycleInGenericOp : public OpRewritePattern<GenericOp> { |
324 | using OpRewritePattern<GenericOp>::OpRewritePattern; |
325 | |
326 | LogicalResult matchAndRewrite(GenericOp genericOp, |
327 | PatternRewriter &rewriter) const override { |
328 | |
329 | // If the op doesnt have tensor semantics, preserve the outputs as is. |
330 | if (!genericOp.hasPureTensorSemantics()) |
331 | return failure(); |
332 | |
333 | bool hasRemovedCycles = false; |
334 | // Iterate over output operands and remove any unused cycles. |
335 | for (const auto &outputOpOperand : |
336 | llvm::enumerate(genericOp.getDpsInits())) { |
337 | |
338 | // Check that result from out operand is dead. |
339 | Value result = genericOp.getResult(outputOpOperand.index()); |
340 | if (!result.use_empty()) |
341 | continue; |
342 | |
343 | // Check that outputArg has one use in cycle. |
344 | BlockArgument outputArg = |
345 | genericOp.getRegionOutputArgs()[outputOpOperand.index()]; |
346 | if (!outputArg.hasOneUse()) |
347 | continue; |
348 | |
349 | // Check cycle has at most one use. |
350 | Operation *cycleOp = *outputArg.user_begin(); |
351 | if (!cycleOp->hasOneUse()) |
352 | continue; |
353 | |
354 | // Check that the cycleUser is a yield. |
355 | Operation *cycleUserOp = *cycleOp->user_begin(); |
356 | if (!isa<linalg::YieldOp>(cycleUserOp)) |
357 | continue; |
358 | |
359 | // Check that argIndex matches yieldIndex, else data is being used. |
360 | if (cycleUserOp->getOperand(outputOpOperand.index()) != |
361 | cycleOp->getResult(0)) |
362 | continue; |
363 | |
364 | // Directly replace the cycle with the blockArg such that |
365 | // Deduplicate pattern can eliminate it along with unused yield. |
366 | rewriter.replaceOp(cycleOp, outputArg); |
367 | rewriter.modifyOpInPlace(genericOp, [] {}); |
368 | hasRemovedCycles = true; |
369 | } |
370 | |
371 | if (hasRemovedCycles) { |
372 | return success(); |
373 | } |
374 | |
375 | return failure(); |
376 | } |
377 | }; |
378 | |
379 | /// Fold uses of duplicate inputs in the body of a linalg.generic. E.g.: |
380 | /// ``` |
381 | /// linalg.generic ins(%a, %b, %a, %b) outs(%a) |
382 | /// ^bb0(%in0, %in1, %in2, %in3, %out1) |
383 | /// ``` |
384 | /// Assuming that all %a and %b have the same index map: |
385 | /// * All uses of %in0 and %in2 are replaced with %out1 |
386 | /// * All uses of %in1 are replaced with %in3 |
387 | /// This pattern can enable additional canonicalizations: In the above example, |
388 | /// %in0, %in1 and %in3 have no uses anymore and their corresponding operands |
389 | /// can be folded away. This pattern does not modify uses of output block args. |
390 | struct FoldDuplicateInputBbArgs : public OpRewritePattern<GenericOp> { |
391 | using OpRewritePattern<GenericOp>::OpRewritePattern; |
392 | |
393 | LogicalResult matchAndRewrite(GenericOp genericOp, |
394 | PatternRewriter &rewriter) const override { |
395 | // Find replacement bbArgs for all input bbArg. |
396 | DenseMap<int, int> replacements; |
397 | for (int i = 0; i < genericOp.getNumDpsInputs(); ++i) { |
398 | // Skip bbArgs that have no uses. |
399 | if (genericOp.getBody()->getArgument(i).getUses().empty()) |
400 | continue; |
401 | // Find replacement bbArg. This can be an input or an output bbArg. |
402 | for (int j = genericOp->getNumOperands() - 1; j > i; --j) { |
403 | if (genericOp->getOperand(i) == genericOp->getOperand(j) && |
404 | genericOp.getIndexingMapsArray()[i] == |
405 | genericOp.getIndexingMapsArray()[j]) { |
406 | replacements[i] = j; |
407 | break; |
408 | } |
409 | } |
410 | } |
411 | |
412 | // Stop here if no replacements were found. |
413 | if (replacements.empty()) |
414 | return failure(); |
415 | |
416 | // Rewrite the op. |
417 | rewriter.modifyOpInPlace(genericOp, [&]() { |
418 | for (auto [before, after] : replacements) { |
419 | BlockArgument bbArg = genericOp.getBody()->getArgument(before); |
420 | BlockArgument replacement = genericOp.getBody()->getArgument(after); |
421 | rewriter.replaceAllUsesWith(from: bbArg, to: replacement); |
422 | } |
423 | }); |
424 | |
425 | return success(); |
426 | } |
427 | }; |
428 | |
429 | } // namespace |
430 | |
431 | void mlir::linalg::populateEraseUnusedOperandsAndResultsPatterns( |
432 | RewritePatternSet &patterns) { |
433 | patterns.insert<DeduplicateAndRemoveDeadOperandsAndResults>( |
434 | arg: patterns.getContext(), /*removeOutputs=*/args: true); |
435 | patterns.insert<RemoveUnusedCycleInGenericOp>(arg: patterns.getContext()); |
436 | } |
437 | |
438 | void mlir::linalg::populateEraseUnnecessaryInputsPatterns( |
439 | RewritePatternSet &patterns) { |
440 | patterns.insert<DeduplicateAndRemoveDeadOperandsAndResults>( |
441 | arg: patterns.getContext(), /*removeOutputs=*/args: false); |
442 | patterns.insert<FoldDuplicateInputBbArgs>(arg: patterns.getContext()); |
443 | } |
444 |
Definitions
- isResultValueDead
- deduplicateInputOperands
- deduplicateOutputOperands
- populateOpPayload
- deduplicateOperandsAndRemoveDeadResults
- DeduplicateAndRemoveDeadOperandsAndResults
- DeduplicateAndRemoveDeadOperandsAndResults
- matchAndRewrite
- RemoveUnusedCycleInGenericOp
- matchAndRewrite
- FoldDuplicateInputBbArgs
- matchAndRewrite
- populateEraseUnusedOperandsAndResultsPatterns
Update your C++ knowledge – Modern C++11/14/17 Training
Find out more