1 //===- ShapeToStandard.cpp - conversion from Shape to Standard 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 #include "mlir/Conversion/ShapeToStandard/ShapeToStandard.h"
10 
11 #include "../PassDetail.h"
12 #include "mlir/Dialect/SCF/SCF.h"
13 #include "mlir/Dialect/Shape/IR/Shape.h"
14 #include "mlir/Dialect/StandardOps/IR/Ops.h"
15 #include "mlir/IR/BlockAndValueMapping.h"
16 #include "mlir/Transforms/DialectConversion.h"
17 
18 using namespace mlir;
19 using namespace mlir::shape;
20 using namespace mlir::scf;
21 
22 /// Conversion patterns.
23 namespace {
24 class AnyOpConversion : public OpConversionPattern<AnyOp> {
25 public:
26   using OpConversionPattern<AnyOp>::OpConversionPattern;
27 
28   LogicalResult
29   matchAndRewrite(AnyOp op, ArrayRef<Value> operands,
30                   ConversionPatternRewriter &rewriter) const override;
31 };
32 } // namespace
33 
34 LogicalResult
35 AnyOpConversion::matchAndRewrite(AnyOp op, ArrayRef<Value> operands,
36                                  ConversionPatternRewriter &rewriter) const {
37   AnyOp::Adaptor transformed(operands);
38 
39   // Replace `any` with its first operand.
40   // Any operand would be a valid substitution.
41   rewriter.replaceOp(op, {transformed.inputs().front()});
42   return success();
43 }
44 
45 namespace {
46 template <typename SrcOpTy, typename DstOpTy>
47 class BinaryOpConversion : public OpConversionPattern<SrcOpTy> {
48 public:
49   using OpConversionPattern<SrcOpTy>::OpConversionPattern;
50 
51   LogicalResult
52   matchAndRewrite(SrcOpTy op, ArrayRef<Value> operands,
53                   ConversionPatternRewriter &rewriter) const override {
54     typename SrcOpTy::Adaptor transformed(operands);
55 
56     // For now, only error-free types are supported by this lowering.
57     if (op.getType().template isa<SizeType>())
58       return failure();
59 
60     rewriter.replaceOpWithNewOp<DstOpTy>(op, transformed.lhs(),
61                                          transformed.rhs());
62     return success();
63   }
64 };
65 } // namespace
66 
67 namespace {
68 struct BroadcastOpConverter : public OpConversionPattern<BroadcastOp> {
69   using OpConversionPattern<BroadcastOp>::OpConversionPattern;
70 
71   LogicalResult
72   matchAndRewrite(BroadcastOp op, ArrayRef<Value> operands,
73                   ConversionPatternRewriter &rewriter) const override;
74 };
75 } // namespace
76 
77 LogicalResult BroadcastOpConverter::matchAndRewrite(
78     BroadcastOp op, ArrayRef<Value> operands,
79     ConversionPatternRewriter &rewriter) const {
80   // For now, this lowering is only defined on `tensor<?xindex>` operands, not
81   // on shapes.
82   if (op.getType().isa<ShapeType>())
83     return failure();
84 
85   assert(!op.lhs().getType().isa<ShapeType>() &&
86          !op.rhs().getType().isa<ShapeType>());
87   auto loc = op.getLoc();
88   BroadcastOp::Adaptor transformed(operands);
89   Value zero = rewriter.create<ConstantIndexOp>(loc, 0);
90   Value one = rewriter.create<ConstantIndexOp>(loc, 1);
91 
92   // Find smaller and greater rank and extent tensor.
93   Value lhsRank = rewriter.create<DimOp>(loc, op.lhs(), zero);
94   Value rhsRank = rewriter.create<DimOp>(loc, op.rhs(), zero);
95   Value lhsRankULE =
96       rewriter.create<CmpIOp>(loc, CmpIPredicate::ule, lhsRank, rhsRank);
97   Type indexTy = rewriter.getIndexType();
98   Value lesserRank =
99       rewriter.create<SelectOp>(loc, lhsRankULE, lhsRank, rhsRank);
100   Value greaterRank =
101       rewriter.create<SelectOp>(loc, lhsRankULE, rhsRank, lhsRank);
102   auto erasedRankType =
103       RankedTensorType::get({ShapedType::kDynamicSize}, indexTy);
104   Value rankErasedLhs =
105       rewriter.create<TensorCastOp>(loc, erasedRankType, transformed.lhs());
106   Value rankErasedRhs =
107       rewriter.create<TensorCastOp>(loc, erasedRankType, transformed.rhs());
108   Value lesserRankOperand =
109       rewriter.create<SelectOp>(loc, lhsRankULE, rankErasedLhs, rankErasedRhs);
110   Value greaterRankOperand =
111       rewriter.create<SelectOp>(loc, lhsRankULE, rankErasedRhs, rankErasedLhs);
112 
113   Value rankDiff =
114       rewriter.create<SubIOp>(loc, indexTy, greaterRank, lesserRank);
115   rewriter.replaceOpWithNewOp<DynamicTensorFromElementsOp>(
116       op, getExtentTensorType(op.getContext()), ValueRange{greaterRank},
117       [&](OpBuilder &b, Location loc, ValueRange args) {
118         Value outputDimension = args[0];
119         Value isUnchallengedDimension = b.create<CmpIOp>(
120             loc, CmpIPredicate::ult, outputDimension, rankDiff);
121         Value greaterRankOperandExtent = b.create<ExtractElementOp>(
122             loc, greaterRankOperand, outputDimension);
123         // The initial dimensions of the greater-rank operand are unchallenged,
124         // so we can take them as-is. Otherwise, we need to do a comparison.
125         // We need an actual branch here (instead of a select) because the
126         // lesser-rank operand might be rank 0, so any extract_element would be
127         // invalid.
128         auto ifOp = b.create<IfOp>(
129             loc, TypeRange{indexTy}, isUnchallengedDimension,
130             [&](OpBuilder &b, Location loc) {
131               b.create<scf::YieldOp>(loc, greaterRankOperandExtent);
132             },
133             [&](OpBuilder &b, Location loc) {
134               // The broadcasting logic is:
135               // - if one extent (here we arbitrarily choose the extent from
136               // the greater-rank operand) is equal to 1, then take the extent
137               // from the other operand
138               // - otherwise, take the extent as-is.
139               // Note that this logic remains correct in the presence of
140               // dimensions of zero extent.
141               Value lesserRankOperandDimension =
142                   b.create<SubIOp>(loc, indexTy, outputDimension, rankDiff);
143               Value lesserRankOperandExtent = b.create<ExtractElementOp>(
144                   loc, lesserRankOperand,
145                   ValueRange{lesserRankOperandDimension});
146               Value greaterRankOperandExtentIsOne = b.create<CmpIOp>(
147                   loc, CmpIPredicate::eq, greaterRankOperandExtent, one);
148               Value broadcastedExtent = b.create<SelectOp>(
149                   loc, greaterRankOperandExtentIsOne, lesserRankOperandExtent,
150                   greaterRankOperandExtent);
151               b.create<scf::YieldOp>(loc, broadcastedExtent);
152             });
153         b.create<mlir::YieldOp>(loc, ifOp.getResult(0));
154       });
155   return success();
156 }
157 
158 namespace {
159 class ConstShapeOpConverter : public OpConversionPattern<ConstShapeOp> {
160 public:
161   using OpConversionPattern<ConstShapeOp>::OpConversionPattern;
162 
163   LogicalResult
164   matchAndRewrite(ConstShapeOp op, ArrayRef<Value> operands,
165                   ConversionPatternRewriter &rewriter) const override;
166 };
167 } // namespace
168 
169 LogicalResult ConstShapeOpConverter::matchAndRewrite(
170     ConstShapeOp op, ArrayRef<Value> operands,
171     ConversionPatternRewriter &rewriter) const {
172 
173   // For now, this lowering supports only extent tensors, not `shape.shape`
174   // types.
175   if (op.getType().isa<ShapeType>())
176     return failure();
177 
178   auto loc = op.getLoc();
179   SmallVector<Value, 4> extentOperands;
180   for (auto extent : op.shape()) {
181     extentOperands.push_back(
182         rewriter.create<ConstantIndexOp>(loc, extent.getLimitedValue()));
183   }
184   Type indexTy = rewriter.getIndexType();
185   Value tensor =
186       rewriter.create<TensorFromElementsOp>(loc, indexTy, extentOperands);
187   Type resultTy = RankedTensorType::get({ShapedType::kDynamicSize}, indexTy);
188   rewriter.replaceOpWithNewOp<TensorCastOp>(op, tensor, resultTy);
189   return success();
190 }
191 
192 namespace {
193 class ConstSizeOpConversion : public OpConversionPattern<ConstSizeOp> {
194 public:
195   using OpConversionPattern<ConstSizeOp>::OpConversionPattern;
196 
197   LogicalResult
198   matchAndRewrite(ConstSizeOp op, ArrayRef<Value> operands,
199                   ConversionPatternRewriter &rewriter) const override;
200 };
201 } // namespace
202 
203 LogicalResult ConstSizeOpConversion::matchAndRewrite(
204     ConstSizeOp op, ArrayRef<Value> operands,
205     ConversionPatternRewriter &rewriter) const {
206   rewriter.replaceOpWithNewOp<ConstantIndexOp>(op, op.value().getSExtValue());
207   return success();
208 }
209 
210 namespace {
211 class GetExtentOpConverter : public OpConversionPattern<GetExtentOp> {
212   using OpConversionPattern<GetExtentOp>::OpConversionPattern;
213 
214   LogicalResult
215   matchAndRewrite(GetExtentOp op, ArrayRef<Value> operands,
216                   ConversionPatternRewriter &rewriter) const override;
217 };
218 } // namespace
219 
220 LogicalResult GetExtentOpConverter::matchAndRewrite(
221     GetExtentOp op, ArrayRef<Value> operands,
222     ConversionPatternRewriter &rewriter) const {
223   GetExtentOp::Adaptor transformed(operands);
224 
225   // For now, only error-free types are supported by this lowering.
226   if (op.getType().isa<SizeType>())
227     return failure();
228 
229   // Derive shape extent directly from shape origin if possible. This
230   // circumvents the necessity to materialize the shape in memory.
231   if (auto shapeOfOp = op.shape().getDefiningOp<ShapeOfOp>()) {
232     if (shapeOfOp.arg().getType().isa<ShapedType>()) {
233       rewriter.replaceOpWithNewOp<DimOp>(op, shapeOfOp.arg(),
234                                          transformed.dim());
235       return success();
236     }
237   }
238 
239   rewriter.replaceOpWithNewOp<ExtractElementOp>(op, rewriter.getIndexType(),
240                                                 transformed.shape(),
241                                                 ValueRange{transformed.dim()});
242   return success();
243 }
244 
245 namespace {
246 class RankOpConverter : public OpConversionPattern<shape::RankOp> {
247 public:
248   using OpConversionPattern<shape::RankOp>::OpConversionPattern;
249 
250   LogicalResult
251   matchAndRewrite(shape::RankOp op, ArrayRef<Value> operands,
252                   ConversionPatternRewriter &rewriter) const override;
253 };
254 } // namespace
255 
256 LogicalResult
257 RankOpConverter::matchAndRewrite(shape::RankOp op, ArrayRef<Value> operands,
258                                  ConversionPatternRewriter &rewriter) const {
259   // For now, this lowering supports only error-free types.
260   if (op.getType().isa<SizeType>())
261     return failure();
262 
263   shape::RankOp::Adaptor transformed(operands);
264   rewriter.replaceOpWithNewOp<DimOp>(op, transformed.shape(), 0);
265   return success();
266 }
267 
268 namespace {
269 /// Converts `shape.reduce` to `scf.for`.
270 struct ReduceOpConverter : public OpConversionPattern<shape::ReduceOp> {
271 public:
272   using OpConversionPattern::OpConversionPattern;
273 
274   LogicalResult
275   matchAndRewrite(shape::ReduceOp op, ArrayRef<Value> operands,
276                   ConversionPatternRewriter &rewriter) const final;
277 };
278 } // namespace
279 
280 LogicalResult
281 ReduceOpConverter::matchAndRewrite(shape::ReduceOp op, ArrayRef<Value> operands,
282                                    ConversionPatternRewriter &rewriter) const {
283   // For now, this lowering is only defined on `tensor<?xindex>` operands.
284   if (op.shape().getType().isa<ShapeType>())
285     return failure();
286 
287   auto loc = op.getLoc();
288   shape::ReduceOp::Adaptor transformed(operands);
289 
290   Value zero = rewriter.create<ConstantIndexOp>(loc, 0);
291   Value one = rewriter.create<ConstantIndexOp>(loc, 1);
292   Type indexTy = rewriter.getIndexType();
293   Value rank = rewriter.create<DimOp>(loc, indexTy, transformed.shape(), zero);
294 
295   auto loop = rewriter.create<scf::ForOp>(
296       loc, zero, rank, one, op.initVals(),
297       [&](OpBuilder &b, Location loc, Value iv, ValueRange args) {
298         Value extent = b.create<ExtractElementOp>(loc, transformed.shape(), iv);
299 
300         SmallVector<Value, 2> mappedValues{iv, extent};
301         mappedValues.append(args.begin(), args.end());
302 
303         BlockAndValueMapping mapping;
304         Block *reduceBody = op.getBody();
305         mapping.map(reduceBody->getArguments(), mappedValues);
306         for (auto &nested : reduceBody->without_terminator())
307           b.clone(nested, mapping);
308 
309         SmallVector<Value, 2> mappedResults;
310         for (auto result : reduceBody->getTerminator()->getOperands())
311           mappedResults.push_back(mapping.lookup(result));
312         b.create<scf::YieldOp>(loc, mappedResults);
313       });
314 
315   rewriter.replaceOp(op, loop.getResults());
316   return success();
317 }
318 
319 namespace {
320 /// Converts `shape.shape_eq` to an `scf.for` loop. For now, the lowering is
321 /// only defined on `tensor<?xindex>` operands. The test for equality first
322 /// compares their size and, if equal, checks every extent for equality.
323 ///
324 /// Example:
325 ///
326 /// %result = shape.shape_eq %a, %b : tensor<?xindex>, tensor<?xindex>
327 ///
328 /// becomes
329 ///
330 /// %c0 = constant 0 : index
331 /// %0 = dim %arg0, %c0 : tensor<?xindex>
332 /// %1 = dim %arg1, %c0 : tensor<?xindex>
333 /// %2 = cmpi "eq", %0, %1 : index
334 /// %result = scf.if %2 -> (i1) {
335 ///   %c1 = constant 1 : index
336 ///   %true = constant true
337 ///   %4 = scf.for %arg2 = %c0 to %0 step %c1 iter_args(%arg3 = %true) -> (i1) {
338 ///     %5 = extract_element %arg0[%arg2] : tensor<?xindex>
339 ///     %6 = extract_element %arg1[%arg2] : tensor<?xindex>
340 ///     %7 = cmpi "eq", %5, %6 : index
341 ///     %8 = and %arg3, %7 : i1
342 ///     scf.yield %8 : i1
343 ///   }
344 ///   scf.yield %4 : i1
345 /// } else {
346 ///   %false = constant false
347 ///   scf.yield %false : i1
348 /// }
349 ///
350 struct ShapeEqOpConverter : public OpConversionPattern<ShapeEqOp> {
351   using OpConversionPattern<ShapeEqOp>::OpConversionPattern;
352 
353   LogicalResult
354   matchAndRewrite(ShapeEqOp op, ArrayRef<Value> operands,
355                   ConversionPatternRewriter &rewriter) const override;
356 };
357 } // namespace
358 
359 LogicalResult
360 ShapeEqOpConverter::matchAndRewrite(ShapeEqOp op, ArrayRef<Value> operands,
361                                     ConversionPatternRewriter &rewriter) const {
362   // For now, this lowering is only defined on `tensor<?xindex>` operands, not
363   // on shapes.
364   if (op.lhs().getType().isa<ShapeType>() ||
365       op.rhs().getType().isa<ShapeType>()) {
366     return failure();
367   }
368 
369   ShapeEqOp::Adaptor transformed(operands);
370   auto loc = op.getLoc();
371   Type indexTy = rewriter.getIndexType();
372   Value zero = rewriter.create<ConstantIndexOp>(loc, 0);
373   Value lhsRank = rewriter.create<DimOp>(loc, indexTy, transformed.lhs(), zero);
374   Value rhsRank = rewriter.create<DimOp>(loc, indexTy, transformed.rhs(), zero);
375   Value eqRank =
376       rewriter.create<CmpIOp>(loc, CmpIPredicate::eq, lhsRank, rhsRank);
377   Type i1Ty = rewriter.getI1Type();
378   rewriter.replaceOpWithNewOp<IfOp>(
379       op, i1Ty, eqRank,
380       [&](OpBuilder &b, Location loc) {
381         Value one = b.create<ConstantIndexOp>(loc, 1);
382         Value init = b.create<ConstantOp>(loc, i1Ty, b.getBoolAttr(true));
383         auto loop = b.create<scf::ForOp>(
384             loc, zero, lhsRank, one, ValueRange{init},
385             [&](OpBuilder &b, Location nestedLoc, Value iv, ValueRange args) {
386               Value conj = args[0];
387               Value lhsExtent =
388                   b.create<ExtractElementOp>(loc, transformed.lhs(), iv);
389               Value rhsExtent =
390                   b.create<ExtractElementOp>(loc, transformed.rhs(), iv);
391               Value eqExtent = b.create<CmpIOp>(loc, CmpIPredicate::eq,
392                                                 lhsExtent, rhsExtent);
393               Value conjNext = b.create<AndOp>(loc, conj, eqExtent);
394               b.create<scf::YieldOp>(loc, ValueRange({conjNext}));
395             });
396         b.create<scf::YieldOp>(loc, loop.getResults());
397       },
398       [&](OpBuilder &b, Location loc) {
399         Value result = b.create<ConstantOp>(loc, i1Ty, b.getBoolAttr(false));
400         b.create<scf::YieldOp>(loc, result);
401       });
402   return success();
403 }
404 
405 namespace {
406 class ShapeOfOpConversion : public OpConversionPattern<ShapeOfOp> {
407 public:
408   using OpConversionPattern<ShapeOfOp>::OpConversionPattern;
409 
410   LogicalResult
411   matchAndRewrite(ShapeOfOp op, ArrayRef<Value> operands,
412                   ConversionPatternRewriter &rewriter) const override;
413 };
414 } // namespace
415 
416 LogicalResult ShapeOfOpConversion::matchAndRewrite(
417     ShapeOfOp op, ArrayRef<Value> operands,
418     ConversionPatternRewriter &rewriter) const {
419 
420   // For now, only error-free types are supported by this lowering.
421   if (op.getType().isa<ShapeType>())
422     return failure();
423 
424   // For ranked tensor arguments, lower to `tensor_from_elements`.
425   auto loc = op.getLoc();
426   ShapeOfOp::Adaptor transformed(operands);
427   Value tensor = transformed.arg();
428   Type tensorTy = tensor.getType();
429   if (tensorTy.isa<RankedTensorType>()) {
430 
431     // Build values for individual extents.
432     SmallVector<Value, 8> extentValues;
433     RankedTensorType rankedTensorTy = tensorTy.cast<RankedTensorType>();
434     int64_t rank = rankedTensorTy.getRank();
435     for (int64_t i = 0; i < rank; i++) {
436       if (rankedTensorTy.isDynamicDim(i)) {
437         Value extent = rewriter.create<DimOp>(loc, tensor, i);
438         extentValues.push_back(extent);
439       } else {
440         Value extent =
441             rewriter.create<ConstantIndexOp>(loc, rankedTensorTy.getDimSize(i));
442         extentValues.push_back(extent);
443       }
444     }
445 
446     // Materialize extent tensor.
447     Value staticExtentTensor = rewriter.create<TensorFromElementsOp>(
448         loc, rewriter.getIndexType(), extentValues);
449     rewriter.replaceOpWithNewOp<TensorCastOp>(op, staticExtentTensor,
450                                               op.getType());
451     return success();
452   }
453 
454   // Lower to `dynamic_tensor_from_elements` otherwise.
455   auto *ctx = rewriter.getContext();
456   Value rank = rewriter.create<mlir::RankOp>(loc, tensor);
457   rewriter.replaceOpWithNewOp<DynamicTensorFromElementsOp>(
458       op, getExtentTensorType(ctx), ValueRange{rank},
459       [&](OpBuilder &b, Location loc, ValueRange args) {
460         Value dim = args.front();
461         Value extent = b.create<DimOp>(loc, tensor, dim);
462         b.create<mlir::YieldOp>(loc, extent);
463       });
464 
465   return success();
466 }
467 
468 namespace {
469 class ToExtentTensorOpConversion
470     : public OpConversionPattern<ToExtentTensorOp> {
471 public:
472   using OpConversionPattern<ToExtentTensorOp>::OpConversionPattern;
473 
474   LogicalResult
475   matchAndRewrite(ToExtentTensorOp op, ArrayRef<Value> operands,
476                   ConversionPatternRewriter &rewriter) const override {
477     ToExtentTensorOpAdaptor adaptor(operands);
478 
479     if (!adaptor.input().getType().isa<RankedTensorType>())
480       return rewriter.notifyMatchFailure(op, "input needs to be a tensor");
481 
482     rewriter.replaceOpWithNewOp<TensorCastOp>(op, adaptor.input(),
483                                               op.getType());
484     return success();
485   }
486 };
487 } // namespace
488 
489 namespace {
490 /// Conversion pass.
491 class ConvertShapeToStandardPass
492     : public ConvertShapeToStandardBase<ConvertShapeToStandardPass> {
493 
494   void runOnOperation() override;
495 };
496 } // namespace
497 
498 void ConvertShapeToStandardPass::runOnOperation() {
499   // Setup target legality.
500   MLIRContext &ctx = getContext();
501   ConversionTarget target(ctx);
502   target.addLegalDialect<StandardOpsDialect, SCFDialect>();
503   target.addLegalOp<FuncOp, ModuleOp, ModuleTerminatorOp>();
504 
505   // Setup conversion patterns.
506   OwningRewritePatternList patterns;
507   populateShapeToStandardConversionPatterns(patterns, &ctx);
508 
509   // Apply conversion.
510   auto module = getOperation();
511   if (failed(applyPartialConversion(module, target, std::move(patterns))))
512     signalPassFailure();
513 }
514 
515 void mlir::populateShapeToStandardConversionPatterns(
516     OwningRewritePatternList &patterns, MLIRContext *ctx) {
517   // clang-format off
518   patterns.insert<
519       AnyOpConversion,
520       BinaryOpConversion<AddOp, AddIOp>,
521       BinaryOpConversion<MulOp, MulIOp>,
522       BroadcastOpConverter,
523       ConstShapeOpConverter,
524       ConstSizeOpConversion,
525       GetExtentOpConverter,
526       RankOpConverter,
527       ReduceOpConverter,
528       ShapeEqOpConverter,
529       ShapeOfOpConversion,
530       ToExtentTensorOpConversion>(ctx);
531   // clang-format on
532 }
533 
534 std::unique_ptr<OperationPass<ModuleOp>>
535 mlir::createConvertShapeToStandardPass() {
536   return std::make_unique<ConvertShapeToStandardPass>();
537 }
538