| 1 | //===- BufferizableOpInterfaceImpl.cpp - Impl. of BufferizableOpInterface -===// |
| 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/Tensor/Transforms/BufferizableOpInterfaceImpl.h" |
| 10 | |
| 11 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 12 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 13 | #include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h" |
| 14 | #include "mlir/Dialect/Bufferization/IR/Bufferization.h" |
| 15 | #include "mlir/Dialect/Bufferization/IR/DstBufferizableOpInterfaceImpl.h" |
| 16 | #include "mlir/Dialect/Linalg/IR/Linalg.h" |
| 17 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| 18 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 19 | #include "mlir/Dialect/Tensor/Transforms/SubsetOpInterfaceImpl.h" |
| 20 | #include "mlir/Dialect/Utils/StaticValueUtils.h" |
| 21 | #include "mlir/IR/BuiltinTypeInterfaces.h" |
| 22 | #include "mlir/IR/Dialect.h" |
| 23 | #include "mlir/IR/Operation.h" |
| 24 | |
| 25 | using namespace mlir; |
| 26 | using namespace mlir::bufferization; |
| 27 | using namespace mlir::tensor; |
| 28 | |
| 29 | namespace mlir { |
| 30 | namespace tensor { |
| 31 | namespace { |
| 32 | |
| 33 | struct CastOpInterface |
| 34 | : public BufferizableOpInterface::ExternalModel<CastOpInterface, |
| 35 | tensor::CastOp> { |
| 36 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 37 | const AnalysisState &state) const { |
| 38 | return false; |
| 39 | } |
| 40 | |
| 41 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 42 | const AnalysisState &state) const { |
| 43 | return false; |
| 44 | } |
| 45 | |
| 46 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 47 | const AnalysisState &state) const { |
| 48 | return {{op->getResult(idx: 0), BufferRelation::Equivalent}}; |
| 49 | } |
| 50 | |
| 51 | FailureOr<BufferLikeType> |
| 52 | getBufferType(Operation *op, Value value, const BufferizationOptions &options, |
| 53 | const BufferizationState &state, |
| 54 | SmallVector<Value> &invocationStack) const { |
| 55 | auto castOp = cast<tensor::CastOp>(Val: op); |
| 56 | auto maybeSrcBufferType = |
| 57 | bufferization::detail::asMemRefType(bufferType: bufferization::getBufferType( |
| 58 | value: castOp.getSource(), options, state, invocationStack)); |
| 59 | if (failed(Result: maybeSrcBufferType)) |
| 60 | return failure(); |
| 61 | Attribute memorySpace = maybeSrcBufferType->getMemorySpace(); |
| 62 | |
| 63 | // Note: `getMemRefTypeWithFullyDynamicLayout` returns an unranked memref |
| 64 | // type in case the input is an unranked tensor type. |
| 65 | |
| 66 | // Case 1: Casting an unranked tensor |
| 67 | if (isa<UnrankedTensorType>(Val: castOp.getSource().getType())) { |
| 68 | // When casting to a ranked tensor, we cannot infer any static offset or |
| 69 | // strides from the source. Assume fully dynamic. |
| 70 | return cast<BufferLikeType>( |
| 71 | Val: getMemRefTypeWithFullyDynamicLayout(tensorType: castOp.getType(), memorySpace)); |
| 72 | } |
| 73 | |
| 74 | // Case 2: Casting to an unranked tensor type |
| 75 | if (isa<UnrankedTensorType>(Val: castOp.getType())) { |
| 76 | return cast<BufferLikeType>( |
| 77 | Val: getMemRefTypeWithFullyDynamicLayout(tensorType: castOp.getType(), memorySpace)); |
| 78 | } |
| 79 | |
| 80 | // Case 3: Ranked tensor -> ranked tensor. The offsets and strides do not |
| 81 | // change. |
| 82 | auto rankedResultType = cast<RankedTensorType>(Val: castOp.getType()); |
| 83 | return cast<BufferLikeType>(Val: MemRefType::get( |
| 84 | shape: rankedResultType.getShape(), elementType: rankedResultType.getElementType(), |
| 85 | layout: llvm::cast<MemRefType>(Val&: *maybeSrcBufferType).getLayout(), memorySpace)); |
| 86 | } |
| 87 | |
| 88 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 89 | const BufferizationOptions &options, |
| 90 | BufferizationState &state) const { |
| 91 | auto castOp = cast<tensor::CastOp>(Val: op); |
| 92 | |
| 93 | // The result buffer still has the old (pre-cast) type. |
| 94 | FailureOr<Value> resultBuffer = |
| 95 | getBuffer(rewriter, value: castOp.getSource(), options, state); |
| 96 | if (failed(Result: resultBuffer)) |
| 97 | return failure(); |
| 98 | |
| 99 | // Compute the new type. |
| 100 | auto resultMemRefType = |
| 101 | bufferization::getBufferType(value: castOp.getResult(), options, state); |
| 102 | if (failed(Result: resultMemRefType)) |
| 103 | return failure(); |
| 104 | if (resultBuffer->getType() == *resultMemRefType) { |
| 105 | // This cast is a no-op. |
| 106 | replaceOpWithBufferizedValues(rewriter, op, values: *resultBuffer); |
| 107 | return success(); |
| 108 | } |
| 109 | |
| 110 | // Replace the op with a memref.cast. |
| 111 | assert(memref::CastOp::areCastCompatible(resultBuffer->getType(), |
| 112 | *resultMemRefType) && |
| 113 | "CallOp::bufferize: cast incompatible" ); |
| 114 | replaceOpWithNewBufferizedOp<memref::CastOp>( |
| 115 | rewriter, op, args&: *resultMemRefType, args&: *resultBuffer); |
| 116 | |
| 117 | return success(); |
| 118 | } |
| 119 | }; |
| 120 | |
| 121 | /// Bufferization of tensor.collapse_shape. Replace with memref.collapse_shape. |
| 122 | struct CollapseShapeOpInterface |
| 123 | : public BufferizableOpInterface::ExternalModel<CollapseShapeOpInterface, |
| 124 | tensor::CollapseShapeOp> { |
| 125 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 126 | const AnalysisState &state) const { |
| 127 | // tensor.collapse_shape may reallocate, at which point the source buffer is |
| 128 | // copied. I.e., there will be a memory read side effect on the bufferized |
| 129 | // source. This function conservatively returns "true" because whether a |
| 130 | // copy will be created or not is not known at this point. |
| 131 | return true; |
| 132 | } |
| 133 | |
| 134 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 135 | const AnalysisState &state) const { |
| 136 | return false; |
| 137 | } |
| 138 | |
| 139 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 140 | const AnalysisState &state) const { |
| 141 | // TODO: CollapseShapeOp may allocate at runtime. |
| 142 | return {{op->getOpResult(idx: 0), BufferRelation::Equivalent}}; |
| 143 | } |
| 144 | |
| 145 | FailureOr<BufferLikeType> |
| 146 | getBufferType(Operation *op, Value value, const BufferizationOptions &options, |
| 147 | const BufferizationState &state, |
| 148 | SmallVector<Value> &invocationStack) const { |
| 149 | auto collapseShapeOp = cast<tensor::CollapseShapeOp>(Val: op); |
| 150 | auto maybeSrcBufferType = bufferization::getBufferType( |
| 151 | value: collapseShapeOp.getSrc(), options, state, invocationStack); |
| 152 | if (failed(Result: maybeSrcBufferType)) |
| 153 | return failure(); |
| 154 | auto srcBufferType = llvm::cast<MemRefType>(Val&: *maybeSrcBufferType); |
| 155 | bool canBeCollapsed = memref::CollapseShapeOp::isGuaranteedCollapsible( |
| 156 | srcType: srcBufferType, reassociation: collapseShapeOp.getReassociationIndices()); |
| 157 | |
| 158 | if (!canBeCollapsed) { |
| 159 | // If dims cannot be collapsed, this op bufferizes to a new allocation. |
| 160 | RankedTensorType tensorResultType = collapseShapeOp.getResultType(); |
| 161 | return cast<BufferLikeType>( |
| 162 | Val: bufferization::getMemRefTypeWithStaticIdentityLayout( |
| 163 | tensorType: tensorResultType, memorySpace: srcBufferType.getMemorySpace())); |
| 164 | } |
| 165 | |
| 166 | return cast<BufferLikeType>(Val: memref::CollapseShapeOp::computeCollapsedType( |
| 167 | srcType: srcBufferType, reassociation: collapseShapeOp.getReassociationIndices())); |
| 168 | } |
| 169 | |
| 170 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 171 | const BufferizationOptions &options, |
| 172 | BufferizationState &state) const { |
| 173 | auto collapseShapeOp = cast<tensor::CollapseShapeOp>(Val: op); |
| 174 | RankedTensorType tensorResultType = collapseShapeOp.getResultType(); |
| 175 | FailureOr<Value> maybeBuffer = |
| 176 | getBuffer(rewriter, value: collapseShapeOp.getSrc(), options, state); |
| 177 | if (failed(Result: maybeBuffer)) |
| 178 | return failure(); |
| 179 | Value buffer = *maybeBuffer; |
| 180 | auto bufferType = cast<MemRefType>(Val: buffer.getType()); |
| 181 | |
| 182 | if (tensorResultType.getRank() == 0) { |
| 183 | // 0-d collapses must go through a different op builder. |
| 184 | MemRefType resultType; |
| 185 | |
| 186 | if (bufferType.getLayout().isIdentity()) { |
| 187 | // Standard layout: result type has no offset. |
| 188 | MemRefLayoutAttrInterface layout; |
| 189 | resultType = MemRefType::get(shape: {}, elementType: tensorResultType.getElementType(), |
| 190 | layout, memorySpace: bufferType.getMemorySpace()); |
| 191 | } else { |
| 192 | // Source memref has a layout map: result type has the same offset as |
| 193 | // the source type. |
| 194 | SmallVector<int64_t> strides; |
| 195 | int64_t offset; |
| 196 | if (failed(Result: bufferType.getStridesAndOffset(strides, offset))) |
| 197 | return failure(); |
| 198 | resultType = MemRefType::get( |
| 199 | shape: {}, elementType: tensorResultType.getElementType(), |
| 200 | layout: StridedLayoutAttr::get(context: op->getContext(), offset, strides: {}), |
| 201 | memorySpace: bufferType.getMemorySpace()); |
| 202 | } |
| 203 | |
| 204 | replaceOpWithNewBufferizedOp<memref::CollapseShapeOp>( |
| 205 | rewriter, op, args&: resultType, args&: buffer, args: collapseShapeOp.getReassociation()); |
| 206 | return success(); |
| 207 | } |
| 208 | |
| 209 | // If the dims are not collapsible (due to an incompatible source layout |
| 210 | // map), force an out-of-place bufferization, i.e., a buffer copy. This |
| 211 | // newly allocated buffer will have no layout map and thus be collapsible. |
| 212 | bool canBeCollapsed = memref::CollapseShapeOp::isGuaranteedCollapsible( |
| 213 | srcType: bufferType, reassociation: collapseShapeOp.getReassociationIndices()); |
| 214 | if (!canBeCollapsed) { |
| 215 | // TODO: Create alloc_tensor ops during TensorCopyInsertion. |
| 216 | AnalysisState analysisState(options); |
| 217 | FailureOr<Value> tensorAlloc = allocateTensorForShapedValue( |
| 218 | b&: rewriter, loc: op->getLoc(), shapedValue: collapseShapeOp.getSrc(), options, state); |
| 219 | if (failed(Result: tensorAlloc)) |
| 220 | return failure(); |
| 221 | auto memrefType = |
| 222 | MemRefType::get(shape: collapseShapeOp.getSrcType().getShape(), |
| 223 | elementType: collapseShapeOp.getSrcType().getElementType(), |
| 224 | map: AffineMap(), memorySpace: bufferType.getMemorySpace()); |
| 225 | buffer = bufferization::ToBufferOp::create(builder&: rewriter, location: op->getLoc(), |
| 226 | buffer: memrefType, tensor: *tensorAlloc); |
| 227 | } |
| 228 | |
| 229 | // Result type is inferred by the builder. |
| 230 | replaceOpWithNewBufferizedOp<memref::CollapseShapeOp>( |
| 231 | rewriter, op, args&: buffer, args: collapseShapeOp.getReassociationIndices()); |
| 232 | return success(); |
| 233 | } |
| 234 | }; |
| 235 | |
| 236 | /// Bufferization of tensor.dim. Replace with memref.dim. |
| 237 | struct DimOpInterface |
| 238 | : public BufferizableOpInterface::ExternalModel<DimOpInterface, |
| 239 | tensor::DimOp> { |
| 240 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 241 | const AnalysisState &state) const { |
| 242 | // The op reads the tensor's metadata but not its contents. |
| 243 | return false; |
| 244 | } |
| 245 | |
| 246 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 247 | const AnalysisState &state) const { |
| 248 | return false; |
| 249 | } |
| 250 | |
| 251 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 252 | const AnalysisState &state) const { |
| 253 | return {}; |
| 254 | } |
| 255 | |
| 256 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 257 | const BufferizationOptions &options, |
| 258 | BufferizationState &state) const { |
| 259 | auto dimOp = cast<tensor::DimOp>(Val: op); |
| 260 | FailureOr<Value> v = getBuffer(rewriter, value: dimOp.getSource(), options, state); |
| 261 | if (failed(Result: v)) |
| 262 | return failure(); |
| 263 | replaceOpWithNewBufferizedOp<memref::DimOp>(rewriter, op, args&: *v, |
| 264 | args: dimOp.getIndex()); |
| 265 | return success(); |
| 266 | } |
| 267 | }; |
| 268 | |
| 269 | /// Bufferization of "tensor.empty". Replace with "bufferization.alloc_tensor". |
| 270 | struct EmptyOpInterface |
| 271 | : public BufferizableOpInterface::ExternalModel<EmptyOpInterface, |
| 272 | tensor::EmptyOp> { |
| 273 | bool bufferizesToAllocation(Operation *op, Value value) const { return true; } |
| 274 | |
| 275 | bool resultBufferizesToMemoryWrite(Operation *op, OpResult opResult, |
| 276 | const AnalysisState &state) const { |
| 277 | // The returned tensor does not have specified contents. |
| 278 | return false; |
| 279 | } |
| 280 | |
| 281 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 282 | const BufferizationOptions &options, |
| 283 | BufferizationState &state) const { |
| 284 | auto emptyOp = cast<tensor::EmptyOp>(Val: op); |
| 285 | |
| 286 | // Optimization: Fold away the op if it has no uses. |
| 287 | if (op->getUses().empty()) { |
| 288 | rewriter.eraseOp(op); |
| 289 | return success(); |
| 290 | } |
| 291 | |
| 292 | // Allocate a tensor. This emits a "bufferization.alloc_tensor" op. |
| 293 | FailureOr<Value> allocTensor = allocateTensorForShapedValue( |
| 294 | b&: rewriter, loc: op->getLoc(), shapedValue: emptyOp.getResult(), options, state, |
| 295 | /*copy=*/false); |
| 296 | if (failed(Result: allocTensor)) |
| 297 | return failure(); |
| 298 | rewriter.replaceOp(op, newValues: *allocTensor); |
| 299 | return success(); |
| 300 | } |
| 301 | }; |
| 302 | |
| 303 | /// Bufferization of tensor.expand_shape. Replace with memref.expand_shape. |
| 304 | struct ExpandShapeOpInterface |
| 305 | : public BufferizableOpInterface::ExternalModel<ExpandShapeOpInterface, |
| 306 | tensor::ExpandShapeOp> { |
| 307 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 308 | const AnalysisState &state) const { |
| 309 | // In contrast to tensor.collapse_shape, this op can always be bufferized |
| 310 | // without a copy. |
| 311 | return false; |
| 312 | } |
| 313 | |
| 314 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 315 | const AnalysisState &state) const { |
| 316 | return false; |
| 317 | } |
| 318 | |
| 319 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 320 | const AnalysisState &state) const { |
| 321 | return {{op->getOpResult(idx: 0), BufferRelation::Equivalent}}; |
| 322 | } |
| 323 | |
| 324 | FailureOr<BufferLikeType> |
| 325 | getBufferType(Operation *op, Value value, const BufferizationOptions &options, |
| 326 | const BufferizationState &state, |
| 327 | SmallVector<Value> &invocationStack) const { |
| 328 | auto expandShapeOp = cast<tensor::ExpandShapeOp>(Val: op); |
| 329 | auto maybeSrcBufferType = bufferization::getBufferType( |
| 330 | value: expandShapeOp.getSrc(), options, state, invocationStack); |
| 331 | if (failed(Result: maybeSrcBufferType)) |
| 332 | return failure(); |
| 333 | auto srcBufferType = llvm::cast<MemRefType>(Val&: *maybeSrcBufferType); |
| 334 | auto maybeResultType = memref::ExpandShapeOp::computeExpandedType( |
| 335 | srcType: srcBufferType, resultShape: expandShapeOp.getResultType().getShape(), |
| 336 | reassociation: expandShapeOp.getReassociationIndices()); |
| 337 | if (failed(Result: maybeResultType)) |
| 338 | return failure(); |
| 339 | return cast<BufferLikeType>(Val&: *maybeResultType); |
| 340 | } |
| 341 | |
| 342 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 343 | const BufferizationOptions &options, |
| 344 | BufferizationState &state) const { |
| 345 | auto expandShapeOp = cast<tensor::ExpandShapeOp>(Val: op); |
| 346 | FailureOr<BufferLikeType> maybeResultType = |
| 347 | bufferization::getBufferType(value: expandShapeOp.getResult(), options, state); |
| 348 | if (failed(Result: maybeResultType)) |
| 349 | return failure(); |
| 350 | FailureOr<Value> buffer = |
| 351 | getBuffer(rewriter, value: expandShapeOp.getSrc(), options, state); |
| 352 | if (failed(Result: buffer)) |
| 353 | return failure(); |
| 354 | |
| 355 | auto memrefExpandShape = memref::ExpandShapeOp::create( |
| 356 | builder&: rewriter, location: op->getLoc(), resultType: *maybeResultType, src: *buffer, |
| 357 | reassociation: expandShapeOp.getReassociationIndices(), |
| 358 | outputShape: expandShapeOp.getMixedOutputShape()); |
| 359 | replaceOpWithBufferizedValues(rewriter, op, |
| 360 | values: memrefExpandShape->getResults()); |
| 361 | return success(); |
| 362 | } |
| 363 | }; |
| 364 | |
| 365 | /// Bufferization of tensor.extract_slice. Replace with memref.subview. |
| 366 | struct |
| 367 | : public BufferizableOpInterface::ExternalModel<ExtractSliceOpInterface, |
| 368 | tensor::ExtractSliceOp> { |
| 369 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 370 | const AnalysisState &state) const { |
| 371 | return false; |
| 372 | } |
| 373 | |
| 374 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 375 | const AnalysisState &state) const { |
| 376 | return false; |
| 377 | } |
| 378 | |
| 379 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 380 | const AnalysisState &state) const { |
| 381 | return {{op->getOpResult(idx: 0), BufferRelation::Unknown}}; |
| 382 | } |
| 383 | |
| 384 | LogicalResult (Operation *op, RewriterBase &rewriter, |
| 385 | const BufferizationOptions &options, |
| 386 | BufferizationState &state) const { |
| 387 | auto = cast<tensor::ExtractSliceOp>(Val: op); |
| 388 | SmallVector<OpFoldResult> mixedOffsets = extractSliceOp.getMixedOffsets(); |
| 389 | SmallVector<OpFoldResult> mixedSizes = extractSliceOp.getMixedSizes(); |
| 390 | SmallVector<OpFoldResult> mixedStrides = extractSliceOp.getMixedStrides(); |
| 391 | Location loc = extractSliceOp.getLoc(); |
| 392 | |
| 393 | // Get source buffer. |
| 394 | FailureOr<Value> srcMemref = |
| 395 | getBuffer(rewriter, value: extractSliceOp.getSource(), options, state); |
| 396 | if (failed(Result: srcMemref)) |
| 397 | return failure(); |
| 398 | |
| 399 | // Take a subview of the source buffer. |
| 400 | auto resultMemrefType = bufferization::getBufferType( |
| 401 | value: extractSliceOp.getResult(), options, state); |
| 402 | if (failed(Result: resultMemrefType)) |
| 403 | return failure(); |
| 404 | Value subView = memref::SubViewOp::create( |
| 405 | builder&: rewriter, location: loc, resultType: llvm::cast<MemRefType>(Val&: *resultMemrefType), source: *srcMemref, |
| 406 | offsets: mixedOffsets, sizes: mixedSizes, strides: mixedStrides); |
| 407 | |
| 408 | replaceOpWithBufferizedValues(rewriter, op, values: subView); |
| 409 | return success(); |
| 410 | } |
| 411 | |
| 412 | FailureOr<BufferLikeType> |
| 413 | (Operation *op, Value value, const BufferizationOptions &options, |
| 414 | const BufferizationState &state, |
| 415 | SmallVector<Value> &invocationStack) const { |
| 416 | auto = cast<tensor::ExtractSliceOp>(Val: op); |
| 417 | assert(value == extractSliceOp.getResult() && "invalid value" ); |
| 418 | auto srcMemrefType = bufferization::getBufferType( |
| 419 | value: extractSliceOp.getSource(), options, state, invocationStack); |
| 420 | if (failed(Result: srcMemrefType)) |
| 421 | return failure(); |
| 422 | SmallVector<OpFoldResult> mixedOffsets = extractSliceOp.getMixedOffsets(); |
| 423 | SmallVector<OpFoldResult> mixedSizes = extractSliceOp.getMixedSizes(); |
| 424 | SmallVector<OpFoldResult> mixedStrides = extractSliceOp.getMixedStrides(); |
| 425 | return cast<BufferLikeType>(Val: memref::SubViewOp::inferRankReducedResultType( |
| 426 | resultShape: extractSliceOp.getType().getShape(), |
| 427 | sourceMemRefType: llvm::cast<MemRefType>(Val&: *srcMemrefType), staticOffsets: mixedOffsets, staticSizes: mixedSizes, |
| 428 | staticStrides: mixedStrides)); |
| 429 | } |
| 430 | }; |
| 431 | |
| 432 | /// Bufferization of tensor.extract. Replace with memref.load. |
| 433 | struct |
| 434 | : public BufferizableOpInterface::ExternalModel<ExtractOpInterface, |
| 435 | tensor::ExtractOp> { |
| 436 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 437 | const AnalysisState &state) const { |
| 438 | return true; |
| 439 | } |
| 440 | |
| 441 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 442 | const AnalysisState &state) const { |
| 443 | return false; |
| 444 | } |
| 445 | |
| 446 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 447 | const AnalysisState &state) const { |
| 448 | return {}; |
| 449 | } |
| 450 | |
| 451 | LogicalResult (Operation *op, RewriterBase &rewriter, |
| 452 | const BufferizationOptions &options, |
| 453 | BufferizationState &state) const { |
| 454 | auto = cast<tensor::ExtractOp>(Val: op); |
| 455 | FailureOr<Value> srcMemref = |
| 456 | getBuffer(rewriter, value: extractOp.getTensor(), options, state); |
| 457 | if (failed(Result: srcMemref)) |
| 458 | return failure(); |
| 459 | replaceOpWithNewBufferizedOp<memref::LoadOp>(rewriter, op, args&: *srcMemref, |
| 460 | args: extractOp.getIndices()); |
| 461 | return success(); |
| 462 | } |
| 463 | }; |
| 464 | |
| 465 | // Implements backtracking to traverse indices of the output buffer while |
| 466 | // iterating over op.elements(). |
| 467 | static void createStores(RewriterBase &rewriter, Location loc, int dim, |
| 468 | Value buffer, ArrayRef<int64_t> shape, |
| 469 | ArrayRef<Value> constants, |
| 470 | OperandRange::iterator &elementIt, |
| 471 | SmallVectorImpl<Value> &indices) { |
| 472 | if (dim == static_cast<int>(shape.size()) - 1) { |
| 473 | for (int i = 0; i < shape.back(); ++i) { |
| 474 | indices.back() = constants[i]; |
| 475 | memref::StoreOp::create(builder&: rewriter, location: loc, valueToStore: *elementIt, memref: buffer, indices); |
| 476 | ++elementIt; |
| 477 | } |
| 478 | return; |
| 479 | } |
| 480 | for (int i = 0; i < shape[dim]; ++i) { |
| 481 | indices[dim] = constants[i]; |
| 482 | createStores(rewriter, loc, dim: dim + 1, buffer, shape, constants, elementIt, |
| 483 | indices); |
| 484 | } |
| 485 | } |
| 486 | |
| 487 | /// Bufferization of tensor.from_elements. |
| 488 | struct FromElementsOpInterface |
| 489 | : public BufferizableOpInterface::ExternalModel<FromElementsOpInterface, |
| 490 | tensor::FromElementsOp> { |
| 491 | |
| 492 | bool bufferizesToAllocation(Operation *op, Value value) const { return true; } |
| 493 | |
| 494 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 495 | const BufferizationOptions &options, |
| 496 | BufferizationState &state) const { |
| 497 | auto fromElementsOp = cast<tensor::FromElementsOp>(Val: op); |
| 498 | auto tensorType = cast<RankedTensorType>(Val: fromElementsOp.getType()); |
| 499 | |
| 500 | // Allocate a buffer for the result. |
| 501 | Location loc = op->getLoc(); |
| 502 | auto shape = tensorType.getShape(); |
| 503 | // TODO: Create alloc_tensor ops during TensorCopyInsertion. |
| 504 | FailureOr<Value> tensorAlloc = allocateTensorForShapedValue( |
| 505 | b&: rewriter, loc, shapedValue: fromElementsOp.getResult(), options, state, |
| 506 | /*copy=*/false); |
| 507 | if (failed(Result: tensorAlloc)) |
| 508 | return failure(); |
| 509 | FailureOr<BufferLikeType> memrefType = |
| 510 | bufferization::getBufferType(value: *tensorAlloc, options, state); |
| 511 | if (failed(Result: memrefType)) |
| 512 | return failure(); |
| 513 | Value buffer = bufferization::ToBufferOp::create(builder&: rewriter, location: op->getLoc(), |
| 514 | buffer: *memrefType, tensor: *tensorAlloc); |
| 515 | |
| 516 | // Case: tensor<0xelem_type>. |
| 517 | if (fromElementsOp.getElements().empty()) { |
| 518 | replaceOpWithBufferizedValues(rewriter, op, values: buffer); |
| 519 | return success(); |
| 520 | } |
| 521 | |
| 522 | // Case: tensor<elem_type>. |
| 523 | if (shape.empty()) { |
| 524 | memref::StoreOp::create(builder&: rewriter, location: loc, |
| 525 | valueToStore: fromElementsOp.getElements().front(), memref: buffer); |
| 526 | replaceOpWithBufferizedValues(rewriter, op, values: buffer); |
| 527 | return success(); |
| 528 | } |
| 529 | |
| 530 | // Create constants for the range of possible indices [0, max{shape_i}). |
| 531 | auto maxDim = *llvm::max_element(Range&: shape); |
| 532 | SmallVector<Value, 2> constants; |
| 533 | constants.reserve(N: maxDim); |
| 534 | for (int i = 0; i < maxDim; ++i) |
| 535 | constants.push_back(Elt: arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: i)); |
| 536 | |
| 537 | // Traverse all `elements` and create `memref.store` ops. |
| 538 | auto elementIt = fromElementsOp.getElements().begin(); |
| 539 | SmallVector<Value, 2> indices(tensorType.getRank(), constants[0]); |
| 540 | createStores(rewriter, loc, /*dim=*/0, buffer, shape, constants, elementIt, |
| 541 | indices); |
| 542 | |
| 543 | replaceOpWithBufferizedValues(rewriter, op, values: buffer); |
| 544 | |
| 545 | return success(); |
| 546 | } |
| 547 | }; |
| 548 | |
| 549 | /// Lower the body of a tensor.generate like op (one index-typed bbArg per dim). |
| 550 | /// Such ops are lowered to linalg.map with the given tensor as a destination. |
| 551 | /// |
| 552 | /// Example: |
| 553 | /// ``` |
| 554 | /// %r = tensor.generate %x, %y { |
| 555 | /// ^bb0(%arg0: index, %arg1: index): |
| 556 | /// %0 = "some_op"(%arg0, %arg1) : (index, index) -> (index) |
| 557 | /// tensor.yield %0 : index |
| 558 | /// } : tensor<?x?xindex> |
| 559 | /// ``` |
| 560 | /// |
| 561 | /// Is lowered to: |
| 562 | /// ``` |
| 563 | /// linalg.map ins() outs(%dest) { |
| 564 | /// %d0 = linalg.index 0 : index |
| 565 | /// %d1 = linalg.index 1 : index |
| 566 | /// %0 = "some_op"(%d0, %d1) : (index, index) -> (index) |
| 567 | /// linalg.yield %0 : index |
| 568 | /// } |
| 569 | /// ``` |
| 570 | static Value lowerGenerateLikeOpBody(RewriterBase &rewriter, Location loc, |
| 571 | Value tensorDestination, |
| 572 | ValueRange dynamicSizes, |
| 573 | Region &generateBody) { |
| 574 | assert(generateBody.hasOneBlock() && "expected body with single block" ); |
| 575 | auto tensorType = cast<RankedTensorType>(Val: tensorDestination.getType()); |
| 576 | assert(generateBody.getNumArguments() == tensorType.getRank() && |
| 577 | "rank mismatch" ); |
| 578 | |
| 579 | // Create linalg::MapOp. |
| 580 | OpBuilder::InsertionGuard g(rewriter); |
| 581 | auto linalgOp = |
| 582 | linalg::MapOp::create(builder&: rewriter, location: loc, result: tensorType, /*inputs=*/ValueRange(), |
| 583 | /*init=*/tensorDestination); |
| 584 | Block &linalgBody = linalgOp.getMapper().emplaceBlock(); |
| 585 | linalgBody.addArgument(type: tensorType.getElementType(), loc); |
| 586 | |
| 587 | // Create linalg::IndexOps. |
| 588 | rewriter.setInsertionPointToStart(&linalgBody); |
| 589 | SmallVector<Value> indices; |
| 590 | for (int64_t dim = 0; dim < tensorType.getRank(); ++dim) |
| 591 | indices.push_back(Elt: linalg::IndexOp::create(builder&: rewriter, location: loc, dim)); |
| 592 | |
| 593 | // Move over body. |
| 594 | rewriter.mergeBlocks(source: &generateBody.front(), dest: &linalgBody, argValues: indices); |
| 595 | auto yieldOp = cast<tensor::YieldOp>(Val: linalgBody.getTerminator()); |
| 596 | rewriter.replaceOpWithNewOp<linalg::YieldOp>(op: yieldOp, args: yieldOp.getValue()); |
| 597 | |
| 598 | return linalgOp.getResult()[0]; |
| 599 | } |
| 600 | |
| 601 | /// Bufferization of tensor.generate. |
| 602 | struct GenerateOpInterface |
| 603 | : public BufferizableOpInterface::ExternalModel<GenerateOpInterface, |
| 604 | tensor::GenerateOp> { |
| 605 | |
| 606 | bool bufferizesToAllocation(Operation *op, Value value) const { return true; } |
| 607 | |
| 608 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 609 | const BufferizationOptions &options, |
| 610 | BufferizationState &state) const { |
| 611 | auto generateOp = cast<tensor::GenerateOp>(Val: op); |
| 612 | |
| 613 | auto type = generateOp.getResult().getType(); |
| 614 | |
| 615 | // TODO: Implement memory space for this op. |
| 616 | if (options.defaultMemorySpaceFn(cast<TensorLikeType>(Val&: type)) != Attribute()) |
| 617 | return op->emitError(message: "memory space not implemented yet" ); |
| 618 | |
| 619 | // Allocate memory. |
| 620 | Location loc = op->getLoc(); |
| 621 | FailureOr<Value> tensorAlloc = allocateTensorForShapedValue( |
| 622 | b&: rewriter, loc, shapedValue: generateOp.getResult(), options, state, |
| 623 | /*copy=*/false); |
| 624 | if (failed(Result: tensorAlloc)) |
| 625 | return failure(); |
| 626 | |
| 627 | Value result = lowerGenerateLikeOpBody(rewriter, loc, tensorDestination: *tensorAlloc, |
| 628 | dynamicSizes: generateOp.getDynamicExtents(), |
| 629 | generateBody&: generateOp.getBody()); |
| 630 | rewriter.replaceOp(op: generateOp, newValues: result); |
| 631 | |
| 632 | return success(); |
| 633 | } |
| 634 | }; |
| 635 | |
| 636 | /// Bufferization of tensor.insert. Replace with memref.store. |
| 637 | /// |
| 638 | /// Note: DstBufferizableOpInterfaceExternalModel provides many default method |
| 639 | /// implementations for DestinationStyle ops. |
| 640 | struct InsertOpInterface |
| 641 | : public DstBufferizableOpInterfaceExternalModel<InsertOpInterface, |
| 642 | tensor::InsertOp> { |
| 643 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 644 | const BufferizationOptions &options, |
| 645 | BufferizationState &state) const { |
| 646 | auto insertOp = cast<tensor::InsertOp>(Val: op); |
| 647 | FailureOr<Value> destMemref = |
| 648 | getBuffer(rewriter, value: insertOp.getDest(), options, state); |
| 649 | if (failed(Result: destMemref)) |
| 650 | return failure(); |
| 651 | memref::StoreOp::create(builder&: rewriter, location: insertOp.getLoc(), valueToStore: insertOp.getScalar(), |
| 652 | memref: *destMemref, indices: insertOp.getIndices()); |
| 653 | replaceOpWithBufferizedValues(rewriter, op, values: *destMemref); |
| 654 | return success(); |
| 655 | } |
| 656 | }; |
| 657 | |
| 658 | template <typename InsertOpTy> |
| 659 | static bool insertSliceOpRequiresRead(InsertOpTy insertSliceOp, |
| 660 | OpOperand &opOperand) { |
| 661 | // The source is always read. |
| 662 | if (opOperand == insertSliceOp.getSourceMutable()) |
| 663 | return true; |
| 664 | |
| 665 | // For the destination, it depends... |
| 666 | assert(opOperand == insertSliceOp.getDestMutable() && "expected dest" ); |
| 667 | |
| 668 | // Dest is not read if it is entirely overwritten. E.g.: |
| 669 | // tensor.insert_slice %a into %t[0][10][1] : ... into tensor<10xf32> |
| 670 | bool allOffsetsZero = |
| 671 | llvm::all_of(insertSliceOp.getMixedOffsets(), isZeroInteger); |
| 672 | RankedTensorType destType = insertSliceOp.getDestType(); |
| 673 | bool sizesMatchDestSizes = |
| 674 | areConstantIntValues(insertSliceOp.getMixedSizes(), destType.getShape()); |
| 675 | bool allStridesOne = |
| 676 | areAllConstantIntValue(insertSliceOp.getMixedStrides(), 1); |
| 677 | return !(allOffsetsZero && sizesMatchDestSizes && allStridesOne); |
| 678 | } |
| 679 | |
| 680 | /// Bufferization of tensor.insert_slice. Replace with a memory copy. Under |
| 681 | /// certain circumstances, this op can also be a no-op. |
| 682 | /// |
| 683 | /// Note: DstBufferizableOpInterfaceExternalModel provides many default method |
| 684 | /// implementations for DestinationStyle ops. |
| 685 | struct InsertSliceOpInterface |
| 686 | : public DstBufferizableOpInterfaceExternalModel<InsertSliceOpInterface, |
| 687 | tensor::InsertSliceOp> { |
| 688 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 689 | const AnalysisState &state) const { |
| 690 | return insertSliceOpRequiresRead(insertSliceOp: cast<tensor::InsertSliceOp>(Val: op), |
| 691 | opOperand); |
| 692 | } |
| 693 | |
| 694 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 695 | const BufferizationOptions &options, |
| 696 | BufferizationState &state) const { |
| 697 | // insert_slice ops arise from tiling and bufferizing them out-of-place is |
| 698 | // generally a deal breaker. When used with loops, this ends up cloning the |
| 699 | // whole tensor on every single iteration and is a symptom of a |
| 700 | // catastrophically bad scheduling decision. |
| 701 | // TODO: be very loud about it or even consider failing the pass. |
| 702 | auto insertSliceOp = cast<tensor::InsertSliceOp>(Val: op); |
| 703 | SmallVector<OpFoldResult> mixedOffsets = insertSliceOp.getMixedOffsets(); |
| 704 | SmallVector<OpFoldResult> mixedSizes = insertSliceOp.getMixedSizes(); |
| 705 | SmallVector<OpFoldResult> mixedStrides = insertSliceOp.getMixedStrides(); |
| 706 | Location loc = insertSliceOp.getLoc(); |
| 707 | |
| 708 | // Get destination buffer. |
| 709 | FailureOr<Value> dstMemref = |
| 710 | getBuffer(rewriter, value: insertSliceOp.getDest(), options, state); |
| 711 | if (failed(Result: dstMemref)) |
| 712 | return failure(); |
| 713 | |
| 714 | // Take a subview of the destination buffer. |
| 715 | auto dstMemrefType = cast<MemRefType>(Val: dstMemref->getType()); |
| 716 | MemRefType subviewMemRefType = |
| 717 | memref::SubViewOp::inferRankReducedResultType( |
| 718 | resultShape: insertSliceOp.getSourceType().getShape(), sourceMemRefType: dstMemrefType, |
| 719 | staticOffsets: mixedOffsets, staticSizes: mixedSizes, staticStrides: mixedStrides); |
| 720 | Value subView = |
| 721 | memref::SubViewOp::create(builder&: rewriter, location: loc, resultType: subviewMemRefType, source: *dstMemref, |
| 722 | offsets: mixedOffsets, sizes: mixedSizes, strides: mixedStrides); |
| 723 | |
| 724 | // Copy tensor. If this tensor.insert_slice has a matching |
| 725 | // tensor.extract_slice, the copy operation will eventually fold away. |
| 726 | FailureOr<Value> srcMemref = |
| 727 | getBuffer(rewriter, value: insertSliceOp.getSource(), options, state); |
| 728 | if (failed(Result: srcMemref)) |
| 729 | return failure(); |
| 730 | if (failed(Result: options.memCpyFn(rewriter, loc, *srcMemref, subView))) |
| 731 | return failure(); |
| 732 | |
| 733 | replaceOpWithBufferizedValues(rewriter, op, values: *dstMemref); |
| 734 | return success(); |
| 735 | } |
| 736 | }; |
| 737 | |
| 738 | /// Bufferization of tensor.pad. Replace with bufferization.alloc_tensor + |
| 739 | /// linalg.map + insert_slice. |
| 740 | /// For best performance, vectorize before bufferization (better performance in |
| 741 | /// case of padding with a constant). |
| 742 | struct PadOpInterface |
| 743 | : public BufferizableOpInterface::ExternalModel<PadOpInterface, |
| 744 | tensor::PadOp> { |
| 745 | bool bufferizesToAllocation(Operation *op, Value value) const { return true; } |
| 746 | |
| 747 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 748 | const AnalysisState &state) const { |
| 749 | return true; |
| 750 | } |
| 751 | |
| 752 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 753 | const AnalysisState &state) const { |
| 754 | return false; |
| 755 | } |
| 756 | |
| 757 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 758 | const AnalysisState &state) const { |
| 759 | return {}; |
| 760 | } |
| 761 | |
| 762 | FailureOr<BufferLikeType> |
| 763 | getBufferType(Operation *op, Value value, const BufferizationOptions &options, |
| 764 | const BufferizationState &state, |
| 765 | SmallVector<Value> &invocationStack) const { |
| 766 | // Infer memory space from the source tensor. |
| 767 | auto padOp = cast<tensor::PadOp>(Val: op); |
| 768 | auto maybeSrcBufferType = |
| 769 | bufferization::detail::asMemRefType(bufferType: bufferization::getBufferType( |
| 770 | value: padOp.getSource(), options, state, invocationStack)); |
| 771 | if (failed(Result: maybeSrcBufferType)) |
| 772 | return failure(); |
| 773 | MemRefLayoutAttrInterface layout; |
| 774 | return cast<BufferLikeType>( |
| 775 | Val: MemRefType::get(shape: padOp.getResultType().getShape(), |
| 776 | elementType: padOp.getResultType().getElementType(), layout, |
| 777 | memorySpace: maybeSrcBufferType->getMemorySpace())); |
| 778 | } |
| 779 | |
| 780 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 781 | const BufferizationOptions &options, |
| 782 | BufferizationState &state) const { |
| 783 | auto padOp = cast<tensor::PadOp>(Val: op); |
| 784 | Location loc = padOp.getLoc(); |
| 785 | RankedTensorType resultType = padOp.getResultType(); |
| 786 | RankedTensorType srcType = padOp.getSourceType(); |
| 787 | |
| 788 | auto toValue = [&](OpFoldResult ofr) { |
| 789 | if (auto value = dyn_cast<Value>(Val&: ofr)) |
| 790 | return value; |
| 791 | return arith::ConstantIndexOp::create(builder&: rewriter, location: loc, |
| 792 | value: *getConstantIntValue(ofr)) |
| 793 | .getResult(); |
| 794 | }; |
| 795 | |
| 796 | // Compute dynamic result dimensions. |
| 797 | SmallVector<OpFoldResult> mixedLowPad = padOp.getMixedLowPad(); |
| 798 | SmallVector<OpFoldResult> mixedHighPad = padOp.getMixedHighPad(); |
| 799 | SmallVector<Value> dynamicSizes; |
| 800 | for (int64_t i = 0; i < resultType.getRank(); ++i) { |
| 801 | if (!resultType.isDynamicDim(idx: i)) |
| 802 | continue; |
| 803 | Value srcDim = tensor::DimOp::create(builder&: rewriter, location: loc, source: padOp.getSource(), index: i); |
| 804 | Value lowPad = toValue(mixedLowPad[i]); |
| 805 | Value highPad = toValue(mixedHighPad[i]); |
| 806 | AffineExpr s0, s1, s2; |
| 807 | bindSymbols(ctx: op->getContext(), exprs&: s0, exprs&: s1, exprs&: s2); |
| 808 | AffineExpr sumExpr = s0 + s1 + s2; |
| 809 | Value sum = affine::AffineApplyOp::create( |
| 810 | builder&: rewriter, location: loc, exprList: sumExpr, mapOperands: ValueRange{srcDim, lowPad, highPad}); |
| 811 | dynamicSizes.push_back(Elt: sum); |
| 812 | } |
| 813 | |
| 814 | // Allocate a buffer for the padded result. |
| 815 | FailureOr<Value> tensorAlloc = allocateTensorForShapedValue( |
| 816 | b&: rewriter, loc, shapedValue: padOp.getResult(), options, state, |
| 817 | /*copy=*/false); |
| 818 | if (failed(Result: tensorAlloc)) |
| 819 | return failure(); |
| 820 | |
| 821 | // tensor::PadOp is like tensor::GenerateOp: The only difference is that |
| 822 | // only a part of the generated tensor is needed. For simplicity, we reuse |
| 823 | // the same functionality here. |
| 824 | Value filledBuffer = lowerGenerateLikeOpBody( |
| 825 | rewriter, loc, tensorDestination: *tensorAlloc, dynamicSizes, generateBody&: padOp.getBodyRegion()); |
| 826 | |
| 827 | // Create tensor::InsertSliceOp. |
| 828 | SmallVector<OpFoldResult> sliceSizes = |
| 829 | getMixedSizes(builder&: rewriter, loc, value: padOp.getSource()); |
| 830 | SmallVector<OpFoldResult> sliceStrides(srcType.getRank(), |
| 831 | rewriter.getIndexAttr(value: 1)); |
| 832 | rewriter.replaceOpWithNewOp<tensor::InsertSliceOp>( |
| 833 | op: padOp, args: padOp.getSource(), args&: filledBuffer, |
| 834 | /*offsets=*/args: padOp.getMixedLowPad(), args&: sliceSizes, args&: sliceStrides); |
| 835 | |
| 836 | return success(); |
| 837 | } |
| 838 | }; |
| 839 | |
| 840 | /// Bufferization of tensor.rank. Replace with memref.rank. |
| 841 | struct RankOpInterface |
| 842 | : public BufferizableOpInterface::ExternalModel<RankOpInterface, |
| 843 | tensor::RankOp> { |
| 844 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 845 | const AnalysisState &state) const { |
| 846 | // The op reads the tensor's metadata but not its contents. |
| 847 | return false; |
| 848 | } |
| 849 | |
| 850 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 851 | const AnalysisState &state) const { |
| 852 | return false; |
| 853 | } |
| 854 | |
| 855 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 856 | const AnalysisState &state) const { |
| 857 | return {}; |
| 858 | } |
| 859 | |
| 860 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 861 | const BufferizationOptions &options, |
| 862 | BufferizationState &state) const { |
| 863 | auto rankOp = cast<tensor::RankOp>(Val: op); |
| 864 | FailureOr<Value> v = |
| 865 | getBuffer(rewriter, value: rankOp.getTensor(), options, state); |
| 866 | if (failed(Result: v)) |
| 867 | return failure(); |
| 868 | replaceOpWithNewBufferizedOp<memref::RankOp>(rewriter, op, args: rankOp.getType(), |
| 869 | args&: *v); |
| 870 | return success(); |
| 871 | } |
| 872 | }; |
| 873 | |
| 874 | /// Bufferization of tensor.reshape. Replace with memref.reshape. |
| 875 | struct ReshapeOpInterface |
| 876 | : public BufferizableOpInterface::ExternalModel<ReshapeOpInterface, |
| 877 | tensor::ReshapeOp> { |
| 878 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 879 | const AnalysisState &state) const { |
| 880 | // Depending on the layout map, the source buffer may have to be copied. |
| 881 | auto reshapeOp = cast<tensor::ReshapeOp>(Val: op); |
| 882 | return opOperand == reshapeOp.getShapeMutable(); |
| 883 | } |
| 884 | |
| 885 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 886 | const AnalysisState &state) const { |
| 887 | return false; |
| 888 | } |
| 889 | |
| 890 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 891 | const AnalysisState &state) const { |
| 892 | // Only the 'source' operand aliases the result. |
| 893 | auto reshapeOp = cast<tensor::ReshapeOp>(Val: op); |
| 894 | if (reshapeOp.getSourceMutable() != opOperand) |
| 895 | return {}; |
| 896 | return {{op->getOpResult(idx: 0), BufferRelation::Equivalent}}; |
| 897 | } |
| 898 | |
| 899 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 900 | const BufferizationOptions &options, |
| 901 | BufferizationState &state) const { |
| 902 | auto reshapeOp = cast<tensor::ReshapeOp>(Val: op); |
| 903 | FailureOr<Value> srcBuffer = |
| 904 | getBuffer(rewriter, value: reshapeOp.getSource(), options, state); |
| 905 | FailureOr<Value> shapeBuffer = |
| 906 | getBuffer(rewriter, value: reshapeOp.getShape(), options, state); |
| 907 | if (failed(Result: srcBuffer) || failed(Result: shapeBuffer)) |
| 908 | return failure(); |
| 909 | auto maybeResultMemRefType = |
| 910 | bufferization::getBufferType(value: reshapeOp.getResult(), options, state); |
| 911 | if (failed(Result: maybeResultMemRefType)) |
| 912 | return failure(); |
| 913 | |
| 914 | // memref.reshape requires the source buffer to have an identity layout. |
| 915 | // If the source memref does not have an identity layout, copy the source |
| 916 | // into a new buffer with an identity layout. |
| 917 | auto srcType = llvm::dyn_cast<MemRefType>(Val: srcBuffer->getType()); |
| 918 | if (srcType && !srcType.getLayout().isIdentity()) { |
| 919 | FailureOr<Value> tensorAlloc = allocateTensorForShapedValue( |
| 920 | b&: rewriter, loc: op->getLoc(), shapedValue: reshapeOp.getSource(), options, state); |
| 921 | if (failed(Result: tensorAlloc)) |
| 922 | return failure(); |
| 923 | auto memrefType = MemRefType::get( |
| 924 | shape: srcType.getShape(), elementType: srcType.getElementType(), map: AffineMap(), |
| 925 | memorySpace: cast<BaseMemRefType>(Val: srcBuffer->getType()).getMemorySpace()); |
| 926 | srcBuffer = bufferization::ToBufferOp::create(builder&: rewriter, location: op->getLoc(), |
| 927 | buffer: memrefType, tensor: *tensorAlloc) |
| 928 | .getResult(); |
| 929 | } |
| 930 | |
| 931 | replaceOpWithNewBufferizedOp<memref::ReshapeOp>( |
| 932 | rewriter, op, args&: maybeResultMemRefType.value(), args&: *srcBuffer, args&: *shapeBuffer); |
| 933 | return success(); |
| 934 | } |
| 935 | |
| 936 | FailureOr<BufferLikeType> |
| 937 | getBufferType(Operation *op, Value value, const BufferizationOptions &options, |
| 938 | const BufferizationState &state, |
| 939 | SmallVector<Value> &invocationStack) const { |
| 940 | auto reshapeOp = cast<tensor::ReshapeOp>(Val: op); |
| 941 | assert(value == reshapeOp.getResult() && "unexpected value provided" ); |
| 942 | auto maybeSourceBufferType = bufferization::getBufferType( |
| 943 | value: reshapeOp.getSource(), options, state, invocationStack); |
| 944 | if (failed(Result: maybeSourceBufferType)) |
| 945 | return failure(); |
| 946 | return cast<BufferLikeType>(Val: getMemRefTypeWithStaticIdentityLayout( |
| 947 | tensorType: reshapeOp.getResult().getType(), |
| 948 | memorySpace: cast<BaseMemRefType>(Val&: maybeSourceBufferType.value()).getMemorySpace())); |
| 949 | } |
| 950 | }; |
| 951 | |
| 952 | /// Analysis of ParallelInsertSliceOp. |
| 953 | struct ParallelInsertSliceOpInterface |
| 954 | : public BufferizableOpInterface::ExternalModel< |
| 955 | ParallelInsertSliceOpInterface, ParallelInsertSliceOp> { |
| 956 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 957 | const AnalysisState &state) const { |
| 958 | return {}; |
| 959 | } |
| 960 | |
| 961 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 962 | const AnalysisState &state) const { |
| 963 | return opOperand == cast<ParallelInsertSliceOp>(Val: op).getSourceMutable(); |
| 964 | } |
| 965 | |
| 966 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 967 | const AnalysisState &state) const { |
| 968 | auto parallelInsertSliceOp = cast<ParallelInsertSliceOp>(Val: op); |
| 969 | return opOperand == parallelInsertSliceOp.getDestMutable(); |
| 970 | } |
| 971 | |
| 972 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 973 | const BufferizationOptions &options, |
| 974 | BufferizationState &state) const { |
| 975 | OpBuilder::InsertionGuard g(rewriter); |
| 976 | auto parallelInsertSliceOp = cast<ParallelInsertSliceOp>(Val: op); |
| 977 | InParallelOpInterface parallelCombiningParent = |
| 978 | parallelInsertSliceOp.getParallelCombiningParent(); |
| 979 | |
| 980 | // Bufferize the op outside of the in parallel terminator. |
| 981 | rewriter.setInsertionPoint(parallelCombiningParent); |
| 982 | |
| 983 | // Get source and destination buffers. |
| 984 | FailureOr<Value> destBuffer = |
| 985 | getBuffer(rewriter, value: parallelInsertSliceOp.getDest(), options, state); |
| 986 | if (failed(Result: destBuffer)) |
| 987 | return failure(); |
| 988 | FailureOr<Value> srcBuffer = |
| 989 | getBuffer(rewriter, value: parallelInsertSliceOp.getSource(), options, state); |
| 990 | if (failed(Result: srcBuffer)) |
| 991 | return failure(); |
| 992 | |
| 993 | // Take a subview of the destination buffer. |
| 994 | auto destBufferType = cast<MemRefType>(Val: destBuffer->getType()); |
| 995 | MemRefType subviewMemRefType = |
| 996 | memref::SubViewOp::inferRankReducedResultType( |
| 997 | resultShape: parallelInsertSliceOp.getSourceType().getShape(), sourceMemRefType: destBufferType, |
| 998 | staticOffsets: parallelInsertSliceOp.getMixedOffsets(), |
| 999 | staticSizes: parallelInsertSliceOp.getMixedSizes(), |
| 1000 | staticStrides: parallelInsertSliceOp.getMixedStrides()); |
| 1001 | Value subview = memref::SubViewOp::create( |
| 1002 | builder&: rewriter, location: parallelInsertSliceOp.getLoc(), resultType: subviewMemRefType, |
| 1003 | source: *destBuffer, offsets: parallelInsertSliceOp.getMixedOffsets(), |
| 1004 | sizes: parallelInsertSliceOp.getMixedSizes(), |
| 1005 | strides: parallelInsertSliceOp.getMixedStrides()); |
| 1006 | |
| 1007 | // This memcpy will fold away if everything bufferizes in-place. |
| 1008 | if (failed(Result: options.memCpyFn(rewriter, parallelInsertSliceOp.getLoc(), |
| 1009 | *srcBuffer, subview))) |
| 1010 | return failure(); |
| 1011 | |
| 1012 | // In case the source was allocated in the same block, make sure that the |
| 1013 | // deallocation op (if any) appears after the memcpy. By default, deallocs |
| 1014 | // are placed before the terminator, but this does not work for ForallOp |
| 1015 | // because the terminator does more than just yielding a value. |
| 1016 | // |
| 1017 | // Note: This is not a problem for the destination buffer because these are |
| 1018 | // assumed to always bufferize in-place. |
| 1019 | for (Operation *user : srcBuffer->getUsers()) { |
| 1020 | if (hasEffect<MemoryEffects::Free>(op: user)) { |
| 1021 | if (user->getBlock() == parallelCombiningParent->getBlock()) |
| 1022 | rewriter.moveOpBefore(op: user, existingOp: user->getBlock()->getTerminator()); |
| 1023 | break; |
| 1024 | } |
| 1025 | } |
| 1026 | |
| 1027 | // Delete the op. |
| 1028 | rewriter.eraseOp(op); |
| 1029 | return success(); |
| 1030 | } |
| 1031 | |
| 1032 | /// tensor.parallel_insert_slice op has implicit inplace behavior. We |
| 1033 | /// shouldn't create copy to resolve conflict. |
| 1034 | LogicalResult |
| 1035 | resolveConflicts(Operation *op, RewriterBase &rewriter, |
| 1036 | const AnalysisState &analysisState, |
| 1037 | const BufferizationState &bufferizationState) const { |
| 1038 | return success(); |
| 1039 | } |
| 1040 | }; |
| 1041 | |
| 1042 | /// Bufferization of tensor.splat. Bufferizes to a new allocation that is filled |
| 1043 | /// with a linalg.map. Similar to tensor.generate. |
| 1044 | struct SplatOpInterface |
| 1045 | : public BufferizableOpInterface::ExternalModel<SplatOpInterface, |
| 1046 | tensor::SplatOp> { |
| 1047 | |
| 1048 | bool bufferizesToAllocation(Operation *op, Value value) const { return true; } |
| 1049 | |
| 1050 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 1051 | const BufferizationOptions &options, |
| 1052 | BufferizationState &state) const { |
| 1053 | OpBuilder::InsertionGuard g(rewriter); |
| 1054 | auto splatOp = cast<tensor::SplatOp>(Val: op); |
| 1055 | |
| 1056 | // Allocate memory. |
| 1057 | Location loc = op->getLoc(); |
| 1058 | FailureOr<Value> tensorAlloc = allocateTensorForShapedValue( |
| 1059 | b&: rewriter, loc, shapedValue: splatOp.getResult(), options, state, |
| 1060 | /*copy=*/false); |
| 1061 | if (failed(Result: tensorAlloc)) |
| 1062 | return failure(); |
| 1063 | |
| 1064 | // Create linalg::MapOp. |
| 1065 | auto tensorType = cast<RankedTensorType>(Val: tensorAlloc->getType()); |
| 1066 | |
| 1067 | // TODO: Implement memory space for this op. |
| 1068 | if (options.defaultMemorySpaceFn(cast<TensorLikeType>(Val&: tensorType)) != |
| 1069 | Attribute()) |
| 1070 | return op->emitError(message: "memory space not implemented yet" ); |
| 1071 | |
| 1072 | auto linalgOp = linalg::MapOp::create(builder&: rewriter, location: loc, result: tensorType, |
| 1073 | /*inputs=*/ValueRange(), |
| 1074 | /*init=*/*tensorAlloc); |
| 1075 | Block &linalgBody = linalgOp.getMapper().emplaceBlock(); |
| 1076 | linalgBody.addArgument(type: tensorType.getElementType(), loc); |
| 1077 | |
| 1078 | // Create linalg::IndexOps. |
| 1079 | rewriter.setInsertionPointToStart(&linalgBody); |
| 1080 | linalg::YieldOp::create(builder&: rewriter, location: loc, values: splatOp.getInput()); |
| 1081 | rewriter.replaceOp(op: splatOp, newValues: linalgOp.getResult()[0]); |
| 1082 | |
| 1083 | return success(); |
| 1084 | } |
| 1085 | }; |
| 1086 | |
| 1087 | /// Bufferization of tensor.concat. Bufferizes to a new allocation that is |
| 1088 | /// filled with copy ops. Similar to tensor.from_elements, but using memref.copy |
| 1089 | /// on subviews instead of memref.store. |
| 1090 | struct ConcatOpInterface |
| 1091 | : public BufferizableOpInterface::ExternalModel<ConcatOpInterface, |
| 1092 | tensor::ConcatOp> { |
| 1093 | |
| 1094 | bool bufferizesToAllocation(Operation *op, Value value) const { return true; } |
| 1095 | |
| 1096 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 1097 | const AnalysisState &state) const { |
| 1098 | return false; |
| 1099 | } |
| 1100 | |
| 1101 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 1102 | const AnalysisState &state) const { |
| 1103 | return true; |
| 1104 | } |
| 1105 | |
| 1106 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 1107 | const AnalysisState &state) const { |
| 1108 | return {}; |
| 1109 | } |
| 1110 | |
| 1111 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 1112 | const BufferizationOptions &options, |
| 1113 | BufferizationState &state) const { |
| 1114 | OpBuilder::InsertionGuard g(rewriter); |
| 1115 | auto concatOp = cast<tensor::ConcatOp>(Val: op); |
| 1116 | |
| 1117 | // Allocate memory. |
| 1118 | Location loc = op->getLoc(); |
| 1119 | FailureOr<Value> tensorAlloc = allocateTensorForShapedValue( |
| 1120 | b&: rewriter, loc, shapedValue: concatOp.getResult(), options, state, |
| 1121 | /*copy=*/false); |
| 1122 | if (failed(Result: tensorAlloc)) |
| 1123 | return failure(); |
| 1124 | auto tensorType = cast<RankedTensorType>(Val: tensorAlloc->getType()); |
| 1125 | FailureOr<BufferLikeType> memrefType = |
| 1126 | bufferization::getBufferType(value: *tensorAlloc, options, state); |
| 1127 | if (failed(Result: memrefType)) |
| 1128 | return failure(); |
| 1129 | Value dstBuffer = bufferization::ToBufferOp::create( |
| 1130 | builder&: rewriter, location: op->getLoc(), buffer: *memrefType, tensor: *tensorAlloc); |
| 1131 | |
| 1132 | // Extract the dimension for the concat op |
| 1133 | uint64_t concatDim = concatOp.getDim(); |
| 1134 | |
| 1135 | SmallVector<OpFoldResult> offsets(tensorType.getRank(), |
| 1136 | rewriter.getIndexAttr(value: 0)); |
| 1137 | SmallVector<OpFoldResult> strides(tensorType.getRank(), |
| 1138 | rewriter.getIndexAttr(value: 1)); |
| 1139 | SmallVector<OpFoldResult> sizes = |
| 1140 | memref::getMixedSizes(builder&: rewriter, loc, value: dstBuffer); |
| 1141 | |
| 1142 | AffineExpr s0, s1; |
| 1143 | bindSymbols(ctx: rewriter.getContext(), exprs&: s0, exprs&: s1); |
| 1144 | auto sum = [&](OpFoldResult v1, OpFoldResult v2) { |
| 1145 | return affine::makeComposedFoldedAffineApply(b&: rewriter, loc, expr: s0 + s1, |
| 1146 | operands: {v1, v2}); |
| 1147 | }; |
| 1148 | |
| 1149 | OpFoldResult concatDimOffset = rewriter.getIndexAttr(value: 0); |
| 1150 | for (auto operand : concatOp.getInputs()) { |
| 1151 | // Get the buffer for the operand. |
| 1152 | FailureOr<Value> srcBuffer = getBuffer(rewriter, value: operand, options, state); |
| 1153 | if (failed(Result: srcBuffer)) |
| 1154 | return failure(); |
| 1155 | |
| 1156 | // Each operand may have a different size along the concat dimension, |
| 1157 | // so the offset on that axis must accumulate through the loop, and the |
| 1158 | // size must change to the size of the current operand. |
| 1159 | auto operandTensorType = cast<RankedTensorType>(Val: operand.getType()); |
| 1160 | offsets[concatDim] = concatDimOffset; |
| 1161 | OpFoldResult concatDimSize = |
| 1162 | memref::getMixedSize(builder&: rewriter, loc, value: *srcBuffer, dim: concatDim); |
| 1163 | sizes[concatDim] = concatDimSize; |
| 1164 | |
| 1165 | // Create a subview of the destination buffer. |
| 1166 | auto dstMemrefType = cast<MemRefType>(Val&: *memrefType); |
| 1167 | MemRefType subviewMemRefType = |
| 1168 | memref::SubViewOp::inferRankReducedResultType( |
| 1169 | resultShape: operandTensorType.getShape(), sourceMemRefType: dstMemrefType, staticOffsets: offsets, staticSizes: sizes, |
| 1170 | staticStrides: strides); |
| 1171 | Value subview = memref::SubViewOp::create( |
| 1172 | builder&: rewriter, location: loc, resultType: subviewMemRefType, source: dstBuffer, offsets, sizes, strides); |
| 1173 | |
| 1174 | // Copy the source buffer into the destination subview. |
| 1175 | if (failed(Result: options.memCpyFn(rewriter, loc, *srcBuffer, subview))) |
| 1176 | return failure(); |
| 1177 | |
| 1178 | concatDimOffset = sum(concatDimOffset, concatDimSize); |
| 1179 | } |
| 1180 | |
| 1181 | replaceOpWithBufferizedValues(rewriter, op, values: dstBuffer); |
| 1182 | return success(); |
| 1183 | } |
| 1184 | }; |
| 1185 | |
| 1186 | } // namespace |
| 1187 | } // namespace tensor |
| 1188 | } // namespace mlir |
| 1189 | |
| 1190 | void mlir::tensor::registerBufferizableOpInterfaceExternalModels( |
| 1191 | DialectRegistry ®istry) { |
| 1192 | registry.addExtension(extensionFn: +[](MLIRContext *ctx, tensor::TensorDialect *dialect) { |
| 1193 | CastOp::attachInterface<CastOpInterface>(context&: *ctx); |
| 1194 | CollapseShapeOp::attachInterface<CollapseShapeOpInterface>(context&: *ctx); |
| 1195 | ConcatOp::attachInterface<ConcatOpInterface>(context&: *ctx); |
| 1196 | DimOp::attachInterface<DimOpInterface>(context&: *ctx); |
| 1197 | EmptyOp::attachInterface<EmptyOpInterface>(context&: *ctx); |
| 1198 | ExpandShapeOp::attachInterface<ExpandShapeOpInterface>(context&: *ctx); |
| 1199 | ExtractSliceOp::attachInterface<ExtractSliceOpInterface>(context&: *ctx); |
| 1200 | ExtractOp::attachInterface<ExtractOpInterface>(context&: *ctx); |
| 1201 | FromElementsOp::attachInterface<FromElementsOpInterface>(context&: *ctx); |
| 1202 | GenerateOp::attachInterface<GenerateOpInterface>(context&: *ctx); |
| 1203 | InsertOp::attachInterface<InsertOpInterface>(context&: *ctx); |
| 1204 | InsertSliceOp::attachInterface<InsertSliceOpInterface>(context&: *ctx); |
| 1205 | PadOp::attachInterface<PadOpInterface>(context&: *ctx); |
| 1206 | ParallelInsertSliceOp::attachInterface<ParallelInsertSliceOpInterface>( |
| 1207 | context&: *ctx); |
| 1208 | RankOp::attachInterface<RankOpInterface>(context&: *ctx); |
| 1209 | ReshapeOp::attachInterface<ReshapeOpInterface>(context&: *ctx); |
| 1210 | SplatOp::attachInterface<SplatOpInterface>(context&: *ctx); |
| 1211 | |
| 1212 | // Load additional dialects of which ops may get created. |
| 1213 | ctx->loadDialect<arith::ArithDialect, linalg::LinalgDialect>(); |
| 1214 | }); |
| 1215 | |
| 1216 | // Bufferization requires SubsetInsertionOpInterface models. Make sure that |
| 1217 | // they are registered. |
| 1218 | tensor::registerSubsetOpInterfaceExternalModels(registry); |
| 1219 | } |
| 1220 | |