| 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 "mlir/Dialect/Arith/IR/Arith.h" |
| 12 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
| 13 | #include "mlir/Dialect/SCF/IR/SCF.h" |
| 14 | #include "mlir/Dialect/Shape/IR/Shape.h" |
| 15 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 16 | #include "mlir/IR/IRMapping.h" |
| 17 | #include "mlir/Pass/Pass.h" |
| 18 | #include "mlir/Transforms/DialectConversion.h" |
| 19 | #include "llvm/ADT/STLExtras.h" |
| 20 | |
| 21 | namespace mlir { |
| 22 | #define GEN_PASS_DEF_CONVERTSHAPETOSTANDARDPASS |
| 23 | #include "mlir/Conversion/Passes.h.inc" |
| 24 | } // namespace mlir |
| 25 | |
| 26 | using namespace mlir; |
| 27 | using namespace mlir::shape; |
| 28 | using namespace mlir::scf; |
| 29 | |
| 30 | /// Conversion patterns. |
| 31 | namespace { |
| 32 | class AnyOpConversion : public OpConversionPattern<AnyOp> { |
| 33 | public: |
| 34 | using OpConversionPattern<AnyOp>::OpConversionPattern; |
| 35 | |
| 36 | LogicalResult |
| 37 | matchAndRewrite(AnyOp op, OpAdaptor adaptor, |
| 38 | ConversionPatternRewriter &rewriter) const override; |
| 39 | }; |
| 40 | } // namespace |
| 41 | |
| 42 | LogicalResult |
| 43 | AnyOpConversion::matchAndRewrite(AnyOp op, OpAdaptor adaptor, |
| 44 | ConversionPatternRewriter &rewriter) const { |
| 45 | // Replace `any` with its first operand. |
| 46 | // Any operand would be a valid substitution. |
| 47 | rewriter.replaceOp(op, newValues: {adaptor.getInputs().front()}); |
| 48 | return success(); |
| 49 | } |
| 50 | |
| 51 | namespace { |
| 52 | template <typename SrcOpTy, typename DstOpTy> |
| 53 | class BinaryOpConversion : public OpConversionPattern<SrcOpTy> { |
| 54 | public: |
| 55 | using OpConversionPattern<SrcOpTy>::OpConversionPattern; |
| 56 | |
| 57 | LogicalResult |
| 58 | matchAndRewrite(SrcOpTy op, typename SrcOpTy::Adaptor adaptor, |
| 59 | ConversionPatternRewriter &rewriter) const override { |
| 60 | // For now, only error-free types are supported by this lowering. |
| 61 | if (isa<SizeType>(op.getType())) |
| 62 | return failure(); |
| 63 | |
| 64 | rewriter.replaceOpWithNewOp<DstOpTy>(op, adaptor.getLhs(), |
| 65 | adaptor.getRhs()); |
| 66 | return success(); |
| 67 | } |
| 68 | }; |
| 69 | } // namespace |
| 70 | |
| 71 | namespace { |
| 72 | struct BroadcastOpConverter : public OpConversionPattern<BroadcastOp> { |
| 73 | using OpConversionPattern<BroadcastOp>::OpConversionPattern; |
| 74 | |
| 75 | LogicalResult |
| 76 | matchAndRewrite(BroadcastOp op, OpAdaptor adaptor, |
| 77 | ConversionPatternRewriter &rewriter) const override; |
| 78 | }; |
| 79 | |
| 80 | // Get the resulting extent in a given dimension. This is computed with any |
| 81 | // number of extent tensors and shifted offsets into them. |
| 82 | Value getBroadcastedDim(ImplicitLocOpBuilder lb, ValueRange extentTensors, |
| 83 | ValueRange rankDiffs, Value outputDimension) { |
| 84 | Value one = arith::ConstantIndexOp::create(builder&: lb, value: 1); |
| 85 | Value broadcastedDim = one; |
| 86 | for (auto tup : llvm::zip(t&: extentTensors, u&: rankDiffs)) { |
| 87 | Value shape = std::get<0>(t&: tup); |
| 88 | Value rankDiff = std::get<1>(t&: tup); |
| 89 | Value outOfBounds = arith::CmpIOp::create(builder&: lb, predicate: arith::CmpIPredicate::ult, |
| 90 | lhs: outputDimension, rhs: rankDiff); |
| 91 | Type indexTy = lb.getIndexType(); |
| 92 | broadcastedDim = |
| 93 | IfOp::create( |
| 94 | builder&: lb, cond: outOfBounds, |
| 95 | thenBuilder: [&](OpBuilder &b, Location loc) { |
| 96 | scf::YieldOp::create(builder&: b, location: loc, results: broadcastedDim); |
| 97 | }, |
| 98 | elseBuilder: [&](OpBuilder &b, Location loc) { |
| 99 | // The broadcasting logic is: |
| 100 | // - if one extent (here we arbitrarily choose the |
| 101 | // extent from the greater-rank operand) is equal to 1, |
| 102 | // then take the extent from the other operand |
| 103 | // - otherwise, take the extent as-is. |
| 104 | // Note that this logic remains correct in the presence |
| 105 | // of dimensions of zero extent. |
| 106 | Value lesserRankOperandDimension = arith::SubIOp::create( |
| 107 | builder&: b, location: loc, result: indexTy, lhs: outputDimension, rhs: rankDiff); |
| 108 | Value lesserRankOperandExtent = tensor::ExtractOp::create( |
| 109 | builder&: b, location: loc, tensor: shape, indices: ValueRange{lesserRankOperandDimension}); |
| 110 | |
| 111 | Value dimIsOne = |
| 112 | arith::CmpIOp::create(builder&: b, location: loc, predicate: arith::CmpIPredicate::eq, |
| 113 | lhs: lesserRankOperandExtent, rhs: one); |
| 114 | Value dim = arith::SelectOp::create( |
| 115 | builder&: b, location: loc, condition: dimIsOne, true_value: broadcastedDim, false_value: lesserRankOperandExtent); |
| 116 | scf::YieldOp::create(builder&: b, location: loc, results: dim); |
| 117 | }) |
| 118 | .getResult(i: 0); |
| 119 | } |
| 120 | return broadcastedDim; |
| 121 | } |
| 122 | } // namespace |
| 123 | |
| 124 | LogicalResult BroadcastOpConverter::matchAndRewrite( |
| 125 | BroadcastOp op, OpAdaptor adaptor, |
| 126 | ConversionPatternRewriter &rewriter) const { |
| 127 | // For now, this lowering is only defined on `tensor<?xindex>` operands, not |
| 128 | // on shapes. |
| 129 | if (isa<ShapeType>(Val: op.getType())) |
| 130 | return failure(); |
| 131 | |
| 132 | auto loc = op.getLoc(); |
| 133 | ImplicitLocOpBuilder lb(loc, rewriter); |
| 134 | |
| 135 | Value zero = arith::ConstantIndexOp::create(builder&: lb, value: 0); |
| 136 | Type indexTy = lb.getIndexType(); |
| 137 | |
| 138 | // Save all the ranks for bounds checking. Because this is a tensor |
| 139 | // representing the shape extents, the rank is the extent of the only |
| 140 | // dimension in the tensor. |
| 141 | SmallVector<Value> ranks, rankDiffs; |
| 142 | llvm::append_range(C&: ranks, R: llvm::map_range(C: adaptor.getShapes(), F: [&](Value v) { |
| 143 | return tensor::DimOp::create(builder&: lb, source: v, index: zero); |
| 144 | })); |
| 145 | |
| 146 | // Find the maximum rank |
| 147 | Value maxRank = ranks.front(); |
| 148 | for (Value v : llvm::drop_begin(RangeOrContainer&: ranks, N: 1)) { |
| 149 | maxRank = arith::MaxUIOp::create(builder&: lb, lhs: v, rhs: maxRank); |
| 150 | } |
| 151 | |
| 152 | // Calculate the difference of ranks and the maximum rank for later offsets. |
| 153 | llvm::append_range(C&: rankDiffs, R: llvm::map_range(C&: ranks, F: [&](Value v) { |
| 154 | return arith::SubIOp::create(builder&: lb, result: indexTy, lhs: maxRank, rhs: v); |
| 155 | })); |
| 156 | |
| 157 | Value replacement = tensor::GenerateOp::create( |
| 158 | builder&: lb, resultTy: getExtentTensorType(ctx: lb.getContext()), dynamicExtents: ValueRange{maxRank}, |
| 159 | odsArg2: [&](OpBuilder &b, Location loc, ValueRange args) { |
| 160 | Value broadcastedDim = |
| 161 | getBroadcastedDim(lb: ImplicitLocOpBuilder(loc, b), extentTensors: adaptor.getShapes(), |
| 162 | rankDiffs, outputDimension: args[0]); |
| 163 | |
| 164 | tensor::YieldOp::create(builder&: b, location: loc, value: broadcastedDim); |
| 165 | }); |
| 166 | if (replacement.getType() != op.getType()) |
| 167 | replacement = tensor::CastOp::create(builder&: lb, dest: op.getType(), source: replacement); |
| 168 | rewriter.replaceOp(op, newValues: replacement); |
| 169 | return success(); |
| 170 | } |
| 171 | |
| 172 | namespace { |
| 173 | class ConstShapeOpConverter : public OpConversionPattern<ConstShapeOp> { |
| 174 | public: |
| 175 | using OpConversionPattern<ConstShapeOp>::OpConversionPattern; |
| 176 | |
| 177 | LogicalResult |
| 178 | matchAndRewrite(ConstShapeOp op, OpAdaptor adaptor, |
| 179 | ConversionPatternRewriter &rewriter) const override; |
| 180 | }; |
| 181 | } // namespace |
| 182 | |
| 183 | LogicalResult ConstShapeOpConverter::matchAndRewrite( |
| 184 | ConstShapeOp op, OpAdaptor adaptor, |
| 185 | ConversionPatternRewriter &rewriter) const { |
| 186 | |
| 187 | // For now, this lowering supports only extent tensors, not `shape.shape` |
| 188 | // types. |
| 189 | if (isa<ShapeType>(Val: op.getType())) |
| 190 | return failure(); |
| 191 | |
| 192 | auto loc = op.getLoc(); |
| 193 | SmallVector<Value, 4> extentOperands; |
| 194 | for (auto extent : op.getShape()) { |
| 195 | extentOperands.push_back(Elt: arith::ConstantIndexOp::create( |
| 196 | builder&: rewriter, location: loc, value: extent.getLimitedValue())); |
| 197 | } |
| 198 | Type resultTy = |
| 199 | RankedTensorType::get(shape: {op.getShape().size()}, elementType: rewriter.getIndexType()); |
| 200 | Value tensor = |
| 201 | tensor::FromElementsOp::create(builder&: rewriter, location: loc, result: resultTy, elements: extentOperands); |
| 202 | rewriter.replaceOpWithNewOp<tensor::CastOp>(op, args&: resultTy, args&: tensor); |
| 203 | return success(); |
| 204 | } |
| 205 | |
| 206 | namespace { |
| 207 | class ConstSizeOpConversion : public OpConversionPattern<ConstSizeOp> { |
| 208 | public: |
| 209 | using OpConversionPattern<ConstSizeOp>::OpConversionPattern; |
| 210 | |
| 211 | LogicalResult |
| 212 | matchAndRewrite(ConstSizeOp op, OpAdaptor adaptor, |
| 213 | ConversionPatternRewriter &rewriter) const override; |
| 214 | }; |
| 215 | } // namespace |
| 216 | |
| 217 | LogicalResult ConstSizeOpConversion::matchAndRewrite( |
| 218 | ConstSizeOp op, OpAdaptor adaptor, |
| 219 | ConversionPatternRewriter &rewriter) const { |
| 220 | rewriter.replaceOpWithNewOp<arith::ConstantIndexOp>( |
| 221 | op, args: op.getValue().getSExtValue()); |
| 222 | return success(); |
| 223 | } |
| 224 | |
| 225 | namespace { |
| 226 | struct IsBroadcastableOpConverter |
| 227 | : public OpConversionPattern<IsBroadcastableOp> { |
| 228 | using OpConversionPattern<IsBroadcastableOp>::OpConversionPattern; |
| 229 | |
| 230 | LogicalResult |
| 231 | matchAndRewrite(IsBroadcastableOp op, OpAdaptor adaptor, |
| 232 | ConversionPatternRewriter &rewriter) const override; |
| 233 | }; |
| 234 | } // namespace |
| 235 | |
| 236 | LogicalResult IsBroadcastableOpConverter::matchAndRewrite( |
| 237 | IsBroadcastableOp op, OpAdaptor adaptor, |
| 238 | ConversionPatternRewriter &rewriter) const { |
| 239 | // For now, this lowering is only defined on `tensor<?xindex>` operands, not |
| 240 | // on shapes. |
| 241 | if (!llvm::all_of(Range: op.getShapes(), |
| 242 | P: [](Value v) { return !isa<ShapeType>(Val: v.getType()); })) |
| 243 | return failure(); |
| 244 | |
| 245 | auto loc = op.getLoc(); |
| 246 | ImplicitLocOpBuilder lb(loc, rewriter); |
| 247 | Value zero = arith::ConstantIndexOp::create(builder&: lb, value: 0); |
| 248 | Value one = arith::ConstantIndexOp::create(builder&: lb, value: 1); |
| 249 | Type indexTy = lb.getIndexType(); |
| 250 | |
| 251 | // Save all the ranks for bounds checking. Because this is a tensor |
| 252 | // representing the shape extents, the rank is the extent of the only |
| 253 | // dimension in the tensor. |
| 254 | SmallVector<Value> ranks, rankDiffs; |
| 255 | llvm::append_range(C&: ranks, R: llvm::map_range(C: adaptor.getShapes(), F: [&](Value v) { |
| 256 | return tensor::DimOp::create(builder&: lb, source: v, index: zero); |
| 257 | })); |
| 258 | |
| 259 | // Find the maximum rank |
| 260 | Value maxRank = ranks.front(); |
| 261 | for (Value v : llvm::drop_begin(RangeOrContainer&: ranks, N: 1)) { |
| 262 | maxRank = arith::MaxUIOp::create(builder&: lb, lhs: v, rhs: maxRank); |
| 263 | } |
| 264 | |
| 265 | // Calculate the difference of ranks and the maximum rank for later offsets. |
| 266 | llvm::append_range(C&: rankDiffs, R: llvm::map_range(C&: ranks, F: [&](Value v) { |
| 267 | return arith::SubIOp::create(builder&: lb, result: indexTy, lhs: maxRank, rhs: v); |
| 268 | })); |
| 269 | |
| 270 | Type i1Ty = rewriter.getI1Type(); |
| 271 | Value trueVal = arith::ConstantOp::create(builder&: rewriter, location: loc, result: i1Ty, |
| 272 | value: rewriter.getBoolAttr(value: true)); |
| 273 | |
| 274 | auto reduceResult = ForOp::create( |
| 275 | builder&: lb, location: loc, lowerBound: zero, upperBound: maxRank, step: one, initArgs: ValueRange{trueVal}, |
| 276 | odsArg4: [&](OpBuilder &b, Location loc, Value iv, ValueRange iterArgs) { |
| 277 | // Find a non-1 dim, if it exists. Note that the first part of this |
| 278 | // could reuse the Broadcast lowering entirely, but we redo the work |
| 279 | // here to make optimizations easier between the two loops. |
| 280 | Value broadcastedDim = getBroadcastedDim( |
| 281 | lb: ImplicitLocOpBuilder(loc, b), extentTensors: adaptor.getShapes(), rankDiffs, outputDimension: iv); |
| 282 | |
| 283 | Value broadcastable = iterArgs[0]; |
| 284 | for (auto tup : llvm::zip(t: adaptor.getShapes(), u&: rankDiffs)) { |
| 285 | Value shape, rankDiff; |
| 286 | std::tie(args&: shape, args&: rankDiff) = tup; |
| 287 | Value outOfBounds = arith::CmpIOp::create( |
| 288 | builder&: b, location: loc, predicate: arith::CmpIPredicate::ult, lhs: iv, rhs: rankDiff); |
| 289 | broadcastable = |
| 290 | IfOp::create( |
| 291 | builder&: b, location: loc, cond: outOfBounds, |
| 292 | thenBuilder: [&](OpBuilder &b, Location loc) { |
| 293 | // Non existent dimensions are always broadcastable |
| 294 | scf::YieldOp::create(builder&: b, location: loc, results: broadcastable); |
| 295 | }, |
| 296 | elseBuilder: [&](OpBuilder &b, Location loc) { |
| 297 | // Every value needs to be either 1, or the same non-1 |
| 298 | // value to be broadcastable in this dim. |
| 299 | Value operandDimension = |
| 300 | arith::SubIOp::create(builder&: b, location: loc, result: indexTy, lhs: iv, rhs: rankDiff); |
| 301 | Value dimensionExtent = tensor::ExtractOp::create( |
| 302 | builder&: b, location: loc, tensor: shape, indices: ValueRange{operandDimension}); |
| 303 | |
| 304 | Value equalOne = arith::CmpIOp::create( |
| 305 | builder&: b, location: loc, predicate: arith::CmpIPredicate::eq, lhs: dimensionExtent, rhs: one); |
| 306 | Value equalBroadcasted = |
| 307 | arith::CmpIOp::create(builder&: b, location: loc, predicate: arith::CmpIPredicate::eq, |
| 308 | lhs: dimensionExtent, rhs: broadcastedDim); |
| 309 | Value result = arith::AndIOp::create( |
| 310 | builder&: b, location: loc, lhs: broadcastable, |
| 311 | rhs: arith::OrIOp::create(builder&: b, location: loc, lhs: equalOne, |
| 312 | rhs: equalBroadcasted)); |
| 313 | scf::YieldOp::create(builder&: b, location: loc, results: result); |
| 314 | }) |
| 315 | .getResult(i: 0); |
| 316 | } |
| 317 | |
| 318 | scf::YieldOp::create(builder&: b, location: loc, results: broadcastable); |
| 319 | }); |
| 320 | |
| 321 | rewriter.replaceOp(op, newValues: reduceResult.getResults().front()); |
| 322 | return success(); |
| 323 | } |
| 324 | |
| 325 | namespace { |
| 326 | class DimOpConverter : public OpConversionPattern<DimOp> { |
| 327 | using OpConversionPattern<DimOp>::OpConversionPattern; |
| 328 | |
| 329 | LogicalResult |
| 330 | matchAndRewrite(DimOp op, OpAdaptor adaptor, |
| 331 | ConversionPatternRewriter &rewriter) const override; |
| 332 | }; |
| 333 | } // namespace |
| 334 | |
| 335 | LogicalResult |
| 336 | DimOpConverter::matchAndRewrite(DimOp op, OpAdaptor adaptor, |
| 337 | ConversionPatternRewriter &rewriter) const { |
| 338 | // Lower to dim(X, i) to get_extent(shape_of(X), i) and rely on further |
| 339 | // lowerings. This can be further optimized if needed to avoid intermediate |
| 340 | // steps. |
| 341 | auto shapeOf = shape::ShapeOfOp::create(builder&: rewriter, location: op.getLoc(), arg: op.getValue()); |
| 342 | rewriter.replaceOpWithNewOp<shape::GetExtentOp>(op, args: op.getType(), args&: shapeOf, |
| 343 | args: op.getIndex()); |
| 344 | return success(); |
| 345 | } |
| 346 | |
| 347 | namespace { |
| 348 | class GetExtentOpConverter : public OpConversionPattern<GetExtentOp> { |
| 349 | using OpConversionPattern<GetExtentOp>::OpConversionPattern; |
| 350 | |
| 351 | LogicalResult |
| 352 | matchAndRewrite(GetExtentOp op, OpAdaptor adaptor, |
| 353 | ConversionPatternRewriter &rewriter) const override; |
| 354 | }; |
| 355 | } // namespace |
| 356 | |
| 357 | LogicalResult GetExtentOpConverter::matchAndRewrite( |
| 358 | GetExtentOp op, OpAdaptor adaptor, |
| 359 | ConversionPatternRewriter &rewriter) const { |
| 360 | // For now, only error-free types are supported by this lowering. |
| 361 | if (isa<SizeType>(Val: op.getType())) |
| 362 | return failure(); |
| 363 | |
| 364 | // Derive shape extent directly from shape origin if possible. This |
| 365 | // circumvents the necessity to materialize the shape in memory. |
| 366 | if (auto shapeOfOp = op.getShape().getDefiningOp<ShapeOfOp>()) { |
| 367 | if (isa<ShapedType>(Val: shapeOfOp.getArg().getType())) { |
| 368 | rewriter.replaceOpWithNewOp<tensor::DimOp>(op, args: shapeOfOp.getArg(), |
| 369 | args: adaptor.getDim()); |
| 370 | return success(); |
| 371 | } |
| 372 | } |
| 373 | |
| 374 | rewriter.replaceOpWithNewOp<tensor::ExtractOp>(op, args: rewriter.getIndexType(), |
| 375 | args: adaptor.getShape(), |
| 376 | args: ValueRange{adaptor.getDim()}); |
| 377 | return success(); |
| 378 | } |
| 379 | |
| 380 | namespace { |
| 381 | class RankOpConverter : public OpConversionPattern<shape::RankOp> { |
| 382 | public: |
| 383 | using OpConversionPattern<shape::RankOp>::OpConversionPattern; |
| 384 | |
| 385 | LogicalResult |
| 386 | matchAndRewrite(shape::RankOp op, OpAdaptor adaptor, |
| 387 | ConversionPatternRewriter &rewriter) const override; |
| 388 | }; |
| 389 | } // namespace |
| 390 | |
| 391 | LogicalResult |
| 392 | RankOpConverter::matchAndRewrite(shape::RankOp op, OpAdaptor adaptor, |
| 393 | ConversionPatternRewriter &rewriter) const { |
| 394 | // For now, this lowering supports only error-free types. |
| 395 | if (isa<SizeType>(Val: op.getType())) |
| 396 | return failure(); |
| 397 | |
| 398 | rewriter.replaceOpWithNewOp<tensor::DimOp>(op, args: adaptor.getShape(), args: 0); |
| 399 | return success(); |
| 400 | } |
| 401 | |
| 402 | namespace { |
| 403 | /// Converts `shape.reduce` to `scf.for`. |
| 404 | struct ReduceOpConverter : public OpConversionPattern<shape::ReduceOp> { |
| 405 | public: |
| 406 | using Base::Base; |
| 407 | |
| 408 | LogicalResult |
| 409 | matchAndRewrite(shape::ReduceOp op, OpAdaptor adaptor, |
| 410 | ConversionPatternRewriter &rewriter) const final; |
| 411 | }; |
| 412 | } // namespace |
| 413 | |
| 414 | LogicalResult |
| 415 | ReduceOpConverter::matchAndRewrite(shape::ReduceOp op, OpAdaptor adaptor, |
| 416 | ConversionPatternRewriter &rewriter) const { |
| 417 | // For now, this lowering is only defined on `tensor<?xindex>` operands. |
| 418 | if (isa<ShapeType>(Val: op.getShape().getType())) |
| 419 | return failure(); |
| 420 | |
| 421 | auto loc = op.getLoc(); |
| 422 | |
| 423 | Value zero = arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: 0); |
| 424 | Value one = arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: 1); |
| 425 | Type indexTy = rewriter.getIndexType(); |
| 426 | Value rank = |
| 427 | tensor::DimOp::create(builder&: rewriter, location: loc, result: indexTy, source: adaptor.getShape(), index: zero); |
| 428 | |
| 429 | auto loop = scf::ForOp::create( |
| 430 | builder&: rewriter, location: loc, lowerBound: zero, upperBound: rank, step: one, initArgs: op.getInitVals(), |
| 431 | odsArg4: [&](OpBuilder &b, Location loc, Value iv, ValueRange args) { |
| 432 | Value extent = |
| 433 | tensor::ExtractOp::create(builder&: b, location: loc, tensor: adaptor.getShape(), indices: iv); |
| 434 | |
| 435 | SmallVector<Value, 2> mappedValues{iv, extent}; |
| 436 | mappedValues.append(in_start: args.begin(), in_end: args.end()); |
| 437 | |
| 438 | IRMapping mapping; |
| 439 | Block *reduceBody = op.getBody(); |
| 440 | mapping.map(from: reduceBody->getArguments(), to&: mappedValues); |
| 441 | for (auto &nested : reduceBody->without_terminator()) |
| 442 | b.clone(op&: nested, mapper&: mapping); |
| 443 | |
| 444 | SmallVector<Value, 2> mappedResults; |
| 445 | for (auto result : reduceBody->getTerminator()->getOperands()) |
| 446 | mappedResults.push_back(Elt: mapping.lookup(from: result)); |
| 447 | scf::YieldOp::create(builder&: b, location: loc, results: mappedResults); |
| 448 | }); |
| 449 | |
| 450 | rewriter.replaceOp(op, newValues: loop.getResults()); |
| 451 | return success(); |
| 452 | } |
| 453 | |
| 454 | namespace { |
| 455 | /// Converts `shape.shape_eq` to an `scf.for` loop. For now, the lowering is |
| 456 | /// only defined on `tensor<?xindex>` operands. The test for equality first |
| 457 | /// compares their size and, if equal, checks every extent for equality. |
| 458 | /// |
| 459 | /// Example: |
| 460 | /// |
| 461 | /// %result = shape.shape_eq %a, %b : tensor<?xindex>, tensor<?xindex> |
| 462 | /// |
| 463 | /// becomes |
| 464 | /// |
| 465 | /// %c0 = arith.constant 0 : index |
| 466 | /// %0 = dim %arg0, %c0 : tensor<?xindex> |
| 467 | /// %1 = dim %arg1, %c0 : tensor<?xindex> |
| 468 | /// %2 = arith.cmpi "eq", %0, %1 : index |
| 469 | /// %result = scf.if %2 -> (i1) { |
| 470 | /// %c1 = arith.constant 1 : index |
| 471 | /// %true = arith.constant true |
| 472 | /// %4 = scf.for %arg2 = %c0 to %0 step %c1 iter_args(%arg3 = %true) -> (i1) { |
| 473 | /// %5 = tensor.extract %arg0[%arg2] : tensor<?xindex> |
| 474 | /// %6 = tensor.extract %arg1[%arg2] : tensor<?xindex> |
| 475 | /// %7 = arith.cmpi "eq", %5, %6 : index |
| 476 | /// %8 = arith.andi %arg3, %7 : i1 |
| 477 | /// scf.yield %8 : i1 |
| 478 | /// } |
| 479 | /// scf.yield %4 : i1 |
| 480 | /// } else { |
| 481 | /// %false = arith.constant false |
| 482 | /// scf.yield %false : i1 |
| 483 | /// } |
| 484 | /// |
| 485 | struct ShapeEqOpConverter : public OpConversionPattern<ShapeEqOp> { |
| 486 | using OpConversionPattern<ShapeEqOp>::OpConversionPattern; |
| 487 | |
| 488 | LogicalResult |
| 489 | matchAndRewrite(ShapeEqOp op, OpAdaptor adaptor, |
| 490 | ConversionPatternRewriter &rewriter) const override; |
| 491 | }; |
| 492 | } // namespace |
| 493 | |
| 494 | LogicalResult |
| 495 | ShapeEqOpConverter::matchAndRewrite(ShapeEqOp op, OpAdaptor adaptor, |
| 496 | ConversionPatternRewriter &rewriter) const { |
| 497 | if (!llvm::all_of(Range: op.getShapes(), |
| 498 | P: [](Value v) { return !isa<ShapeType>(Val: v.getType()); })) |
| 499 | return failure(); |
| 500 | |
| 501 | Type i1Ty = rewriter.getI1Type(); |
| 502 | if (op.getShapes().size() <= 1) { |
| 503 | rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, args&: i1Ty, |
| 504 | args: rewriter.getBoolAttr(value: true)); |
| 505 | return success(); |
| 506 | } |
| 507 | |
| 508 | auto loc = op.getLoc(); |
| 509 | Type indexTy = rewriter.getIndexType(); |
| 510 | Value zero = arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: 0); |
| 511 | Value firstShape = adaptor.getShapes().front(); |
| 512 | Value firstRank = |
| 513 | tensor::DimOp::create(builder&: rewriter, location: loc, result: indexTy, source: firstShape, index: zero); |
| 514 | Value result = nullptr; |
| 515 | // Generate a linear sequence of compares, all with firstShape as lhs. |
| 516 | for (Value shape : adaptor.getShapes().drop_front(n: 1)) { |
| 517 | Value rank = tensor::DimOp::create(builder&: rewriter, location: loc, result: indexTy, source: shape, index: zero); |
| 518 | Value eqRank = arith::CmpIOp::create( |
| 519 | builder&: rewriter, location: loc, predicate: arith::CmpIPredicate::eq, lhs: firstRank, rhs: rank); |
| 520 | auto same = IfOp::create( |
| 521 | builder&: rewriter, location: loc, cond: eqRank, |
| 522 | thenBuilder: [&](OpBuilder &b, Location loc) { |
| 523 | Value one = arith::ConstantIndexOp::create(builder&: b, location: loc, value: 1); |
| 524 | Value init = |
| 525 | arith::ConstantOp::create(builder&: b, location: loc, result: i1Ty, value: b.getBoolAttr(value: true)); |
| 526 | auto loop = scf::ForOp::create( |
| 527 | builder&: b, location: loc, lowerBound: zero, upperBound: firstRank, step: one, initArgs: ValueRange{init}, |
| 528 | odsArg4: [&](OpBuilder &b, Location nestedLoc, Value iv, ValueRange args) { |
| 529 | Value conj = args[0]; |
| 530 | Value lhsExtent = |
| 531 | tensor::ExtractOp::create(builder&: b, location: loc, tensor: firstShape, indices: iv); |
| 532 | Value rhsExtent = tensor::ExtractOp::create(builder&: b, location: loc, tensor: shape, indices: iv); |
| 533 | Value eqExtent = arith::CmpIOp::create( |
| 534 | builder&: b, location: loc, predicate: arith::CmpIPredicate::eq, lhs: lhsExtent, rhs: rhsExtent); |
| 535 | Value conjNext = arith::AndIOp::create(builder&: b, location: loc, lhs: conj, rhs: eqExtent); |
| 536 | scf::YieldOp::create(builder&: b, location: loc, results: ValueRange({conjNext})); |
| 537 | }); |
| 538 | scf::YieldOp::create(builder&: b, location: loc, results: loop.getResults()); |
| 539 | }, |
| 540 | elseBuilder: [&](OpBuilder &b, Location loc) { |
| 541 | Value result = |
| 542 | arith::ConstantOp::create(builder&: b, location: loc, result: i1Ty, value: b.getBoolAttr(value: false)); |
| 543 | scf::YieldOp::create(builder&: b, location: loc, results: result); |
| 544 | }); |
| 545 | result = !result ? same.getResult(i: 0) |
| 546 | : arith::AndIOp::create(builder&: rewriter, location: loc, lhs: result, |
| 547 | rhs: same.getResult(i: 0)); |
| 548 | } |
| 549 | rewriter.replaceOp(op, newValues: result); |
| 550 | return success(); |
| 551 | } |
| 552 | |
| 553 | namespace { |
| 554 | class ShapeOfOpConversion : public OpConversionPattern<ShapeOfOp> { |
| 555 | public: |
| 556 | using OpConversionPattern<ShapeOfOp>::OpConversionPattern; |
| 557 | |
| 558 | LogicalResult |
| 559 | matchAndRewrite(ShapeOfOp op, OpAdaptor adaptor, |
| 560 | ConversionPatternRewriter &rewriter) const override; |
| 561 | }; |
| 562 | } // namespace |
| 563 | |
| 564 | LogicalResult ShapeOfOpConversion::matchAndRewrite( |
| 565 | ShapeOfOp op, OpAdaptor adaptor, |
| 566 | ConversionPatternRewriter &rewriter) const { |
| 567 | |
| 568 | // For now, only error-free types are supported by this lowering. |
| 569 | if (isa<ShapeType>(Val: op.getType())) |
| 570 | return failure(); |
| 571 | |
| 572 | // For ranked tensor arguments, lower to `tensor.from_elements`. |
| 573 | auto loc = op.getLoc(); |
| 574 | Value tensor = adaptor.getArg(); |
| 575 | Type tensorTy = tensor.getType(); |
| 576 | if (isa<RankedTensorType>(Val: tensorTy)) { |
| 577 | |
| 578 | // Build values for individual extents. |
| 579 | SmallVector<Value, 8> extentValues; |
| 580 | RankedTensorType rankedTensorTy = cast<RankedTensorType>(Val&: tensorTy); |
| 581 | int64_t rank = rankedTensorTy.getRank(); |
| 582 | for (int64_t i = 0; i < rank; i++) { |
| 583 | if (rankedTensorTy.isDynamicDim(idx: i)) { |
| 584 | Value extent = tensor::DimOp::create(builder&: rewriter, location: loc, source: tensor, index: i); |
| 585 | extentValues.push_back(Elt: extent); |
| 586 | } else { |
| 587 | Value extent = arith::ConstantIndexOp::create( |
| 588 | builder&: rewriter, location: loc, value: rankedTensorTy.getDimSize(idx: i)); |
| 589 | extentValues.push_back(Elt: extent); |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | // Materialize extent tensor. |
| 594 | Value staticExtentTensor = tensor::FromElementsOp::create( |
| 595 | builder&: rewriter, location: loc, result: RankedTensorType::get(shape: {rank}, elementType: rewriter.getIndexType()), |
| 596 | elements: extentValues); |
| 597 | rewriter.replaceOpWithNewOp<tensor::CastOp>(op, args: op.getType(), |
| 598 | args&: staticExtentTensor); |
| 599 | return success(); |
| 600 | } |
| 601 | |
| 602 | // Lower to `tensor.generate` otherwise. |
| 603 | auto *ctx = rewriter.getContext(); |
| 604 | Value rank = tensor::RankOp::create(builder&: rewriter, location: loc, tensor); |
| 605 | rewriter.replaceOpWithNewOp<tensor::GenerateOp>( |
| 606 | op, args: getExtentTensorType(ctx), args: ValueRange{rank}, |
| 607 | args: [&](OpBuilder &b, Location loc, ValueRange args) { |
| 608 | Value dim = args.front(); |
| 609 | Value extent = tensor::DimOp::create(builder&: b, location: loc, source: tensor, index: dim); |
| 610 | tensor::YieldOp::create(builder&: b, location: loc, value: extent); |
| 611 | }); |
| 612 | |
| 613 | return success(); |
| 614 | } |
| 615 | |
| 616 | namespace { |
| 617 | class SplitAtOpConversion : public OpConversionPattern<SplitAtOp> { |
| 618 | public: |
| 619 | using OpConversionPattern<SplitAtOp>::OpConversionPattern; |
| 620 | |
| 621 | LogicalResult |
| 622 | matchAndRewrite(SplitAtOp op, OpAdaptor adaptor, |
| 623 | ConversionPatternRewriter &rewriter) const override; |
| 624 | }; |
| 625 | } // namespace |
| 626 | |
| 627 | LogicalResult SplitAtOpConversion::matchAndRewrite( |
| 628 | SplitAtOp op, OpAdaptor adaptor, |
| 629 | ConversionPatternRewriter &rewriter) const { |
| 630 | // Error conditions are not implemented, only lower if all operands and |
| 631 | // results are extent tensors. |
| 632 | if (llvm::any_of(Range: ValueRange{op.getOperand(), op.getHead(), op.getTail()}, |
| 633 | P: [](Value v) { return isa<ShapeType>(Val: v.getType()); })) |
| 634 | return failure(); |
| 635 | |
| 636 | ImplicitLocOpBuilder b(op.getLoc(), rewriter); |
| 637 | Value zero = arith::ConstantIndexOp::create(builder&: b, value: 0); |
| 638 | Value rank = tensor::DimOp::create(builder&: b, source: adaptor.getOperand(), index: zero); |
| 639 | |
| 640 | // index < 0 ? index + rank : index |
| 641 | Value originalIndex = adaptor.getIndex(); |
| 642 | Value add = arith::AddIOp::create(builder&: b, lhs: originalIndex, rhs: rank); |
| 643 | Value indexIsNegative = |
| 644 | arith::CmpIOp::create(builder&: b, predicate: arith::CmpIPredicate::slt, lhs: originalIndex, rhs: zero); |
| 645 | Value index = arith::SelectOp::create(builder&: b, condition: indexIsNegative, true_value: add, false_value: originalIndex); |
| 646 | |
| 647 | Value one = arith::ConstantIndexOp::create(builder&: b, value: 1); |
| 648 | Value head = |
| 649 | tensor::ExtractSliceOp::create(builder&: b, source: adaptor.getOperand(), offsets: zero, sizes: index, strides: one); |
| 650 | Value tailSize = arith::SubIOp::create(builder&: b, lhs: rank, rhs: index); |
| 651 | Value tail = tensor::ExtractSliceOp::create(builder&: b, source: adaptor.getOperand(), offsets: index, |
| 652 | sizes: tailSize, strides: one); |
| 653 | rewriter.replaceOp(op, newValues: {head, tail}); |
| 654 | return success(); |
| 655 | } |
| 656 | |
| 657 | namespace { |
| 658 | class ToExtentTensorOpConversion |
| 659 | : public OpConversionPattern<ToExtentTensorOp> { |
| 660 | public: |
| 661 | using OpConversionPattern<ToExtentTensorOp>::OpConversionPattern; |
| 662 | |
| 663 | LogicalResult |
| 664 | matchAndRewrite(ToExtentTensorOp op, OpAdaptor adaptor, |
| 665 | ConversionPatternRewriter &rewriter) const override { |
| 666 | if (!isa<RankedTensorType>(Val: adaptor.getInput().getType())) |
| 667 | return rewriter.notifyMatchFailure(arg&: op, msg: "input needs to be a tensor" ); |
| 668 | |
| 669 | rewriter.replaceOpWithNewOp<tensor::CastOp>(op, args: op.getType(), |
| 670 | args: adaptor.getInput()); |
| 671 | return success(); |
| 672 | } |
| 673 | }; |
| 674 | } // namespace |
| 675 | |
| 676 | namespace { |
| 677 | /// Import the Shape Ops to Std Patterns. |
| 678 | #include "ShapeToStandard.cpp.inc" |
| 679 | } // namespace |
| 680 | |
| 681 | namespace { |
| 682 | /// Conversion pass. |
| 683 | class ConvertShapeToStandardPass |
| 684 | : public impl::ConvertShapeToStandardPassBase<ConvertShapeToStandardPass> { |
| 685 | |
| 686 | void runOnOperation() override; |
| 687 | }; |
| 688 | } // namespace |
| 689 | |
| 690 | void ConvertShapeToStandardPass::runOnOperation() { |
| 691 | // Setup target legality. |
| 692 | MLIRContext &ctx = getContext(); |
| 693 | ConversionTarget target(ctx); |
| 694 | target.addLegalDialect<arith::ArithDialect, SCFDialect, |
| 695 | tensor::TensorDialect>(); |
| 696 | target.addLegalOp<CstrRequireOp, func::FuncOp, ModuleOp>(); |
| 697 | |
| 698 | // Setup conversion patterns. |
| 699 | RewritePatternSet patterns(&ctx); |
| 700 | populateShapeToStandardConversionPatterns(patterns); |
| 701 | |
| 702 | // Apply conversion. |
| 703 | auto module = getOperation(); |
| 704 | if (failed(Result: applyPartialConversion(op: module, target, patterns: std::move(patterns)))) |
| 705 | signalPassFailure(); |
| 706 | } |
| 707 | |
| 708 | void mlir::populateShapeToStandardConversionPatterns( |
| 709 | RewritePatternSet &patterns) { |
| 710 | // clang-format off |
| 711 | populateWithGenerated(patterns); |
| 712 | patterns.add< |
| 713 | AnyOpConversion, |
| 714 | BinaryOpConversion<AddOp, arith::AddIOp>, |
| 715 | BinaryOpConversion<MulOp, arith::MulIOp>, |
| 716 | BroadcastOpConverter, |
| 717 | ConstShapeOpConverter, |
| 718 | ConstSizeOpConversion, |
| 719 | DimOpConverter, |
| 720 | IsBroadcastableOpConverter, |
| 721 | GetExtentOpConverter, |
| 722 | RankOpConverter, |
| 723 | ReduceOpConverter, |
| 724 | ShapeEqOpConverter, |
| 725 | ShapeOfOpConversion, |
| 726 | SplitAtOpConversion, |
| 727 | ToExtentTensorOpConversion>(arg: patterns.getContext()); |
| 728 | // clang-format on |
| 729 | } |
| 730 | |