| 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/Vector/Transforms/BufferizableOpInterfaceImpl.h" |
| 10 | |
| 11 | #include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h" |
| 12 | #include "mlir/Dialect/Bufferization/IR/Bufferization.h" |
| 13 | #include "mlir/Dialect/Bufferization/IR/DstBufferizableOpInterfaceImpl.h" |
| 14 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 15 | #include "mlir/IR/Dialect.h" |
| 16 | #include "mlir/IR/Operation.h" |
| 17 | |
| 18 | using namespace mlir; |
| 19 | using namespace mlir::bufferization; |
| 20 | using namespace mlir::vector; |
| 21 | |
| 22 | namespace mlir { |
| 23 | namespace vector { |
| 24 | namespace { |
| 25 | |
| 26 | /// Bufferization of vector.transfer_read. Replaced with a new |
| 27 | /// vector.transfer_read that operates on a memref. |
| 28 | struct TransferReadOpInterface |
| 29 | : public BufferizableOpInterface::ExternalModel<TransferReadOpInterface, |
| 30 | vector::TransferReadOp> { |
| 31 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 32 | const AnalysisState &state) const { |
| 33 | assert(isa<RankedTensorType>(opOperand.get().getType()) && |
| 34 | "only tensor types expected" ); |
| 35 | return true; |
| 36 | } |
| 37 | |
| 38 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 39 | const AnalysisState &state) const { |
| 40 | assert(isa<RankedTensorType>(opOperand.get().getType()) && |
| 41 | "only tensor types expected" ); |
| 42 | return false; |
| 43 | } |
| 44 | |
| 45 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 46 | const AnalysisState &state) const { |
| 47 | return {}; |
| 48 | } |
| 49 | |
| 50 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 51 | const BufferizationOptions &options, |
| 52 | BufferizationState &state) const { |
| 53 | auto readOp = cast<vector::TransferReadOp>(op); |
| 54 | assert(isa<TensorType>(readOp.getShapedType()) && |
| 55 | "only tensor types expected" ); |
| 56 | FailureOr<Value> buffer = |
| 57 | getBuffer(rewriter, readOp.getBase(), options, state); |
| 58 | if (failed(Result: buffer)) |
| 59 | return failure(); |
| 60 | replaceOpWithNewBufferizedOp<vector::TransferReadOp>( |
| 61 | rewriter, readOp, readOp.getVectorType(), *buffer, readOp.getIndices(), |
| 62 | readOp.getPermutationMap(), readOp.getPadding(), readOp.getMask(), |
| 63 | readOp.getInBoundsAttr()); |
| 64 | return success(); |
| 65 | } |
| 66 | }; |
| 67 | |
| 68 | /// Bufferization of vector.transfer_write. Replace with a new |
| 69 | /// vector.transfer_write that operates on a memref. |
| 70 | /// |
| 71 | /// Note: DstBufferizableOpInterfaceExternalModel provides many default method |
| 72 | /// implementations for DestinationStyle ops. |
| 73 | struct TransferWriteOpInterface |
| 74 | : public DstBufferizableOpInterfaceExternalModel<TransferWriteOpInterface, |
| 75 | vector::TransferWriteOp> { |
| 76 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 77 | const AnalysisState &state) const { |
| 78 | auto writeOp = cast<vector::TransferWriteOp>(op); |
| 79 | |
| 80 | // Does not bufferize to a memory read if the vector completely overwrites |
| 81 | // the buffer. |
| 82 | |
| 83 | // Destination must have static shape. |
| 84 | if (!writeOp.getShapedType().hasStaticShape()) |
| 85 | return true; |
| 86 | |
| 87 | // All offsets must be 0. |
| 88 | for (Value offset : writeOp.getIndices()) { |
| 89 | if (getConstantIntValue(offset) != 0) |
| 90 | return true; |
| 91 | } |
| 92 | |
| 93 | // There is no mask. |
| 94 | if (writeOp.isMasked()) |
| 95 | return true; |
| 96 | |
| 97 | // Must write at least the full dimension size. |
| 98 | for (auto [d0, d1] : llvm::zip(writeOp.getShapedType().getShape(), |
| 99 | writeOp.getVectorType().getShape())) { |
| 100 | if (d0 > d1) |
| 101 | return true; |
| 102 | } |
| 103 | |
| 104 | return false; |
| 105 | } |
| 106 | |
| 107 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 108 | const BufferizationOptions &options, |
| 109 | BufferizationState &state) const { |
| 110 | auto writeOp = cast<vector::TransferWriteOp>(op); |
| 111 | assert(isa<TensorType>(writeOp.getShapedType()) && |
| 112 | "only tensor types expected" ); |
| 113 | |
| 114 | // Create a new transfer_write on buffer that doesn't have a return value. |
| 115 | FailureOr<Value> resultBuffer = |
| 116 | getBuffer(rewriter, writeOp.getBase(), options, state); |
| 117 | if (failed(Result: resultBuffer)) |
| 118 | return failure(); |
| 119 | rewriter.create<vector::TransferWriteOp>( |
| 120 | writeOp.getLoc(), writeOp.getVector(), *resultBuffer, |
| 121 | writeOp.getIndices(), writeOp.getPermutationMapAttr(), |
| 122 | writeOp.getMask(), writeOp.getInBoundsAttr()); |
| 123 | replaceOpWithBufferizedValues(rewriter, op, values: *resultBuffer); |
| 124 | |
| 125 | return success(); |
| 126 | } |
| 127 | }; |
| 128 | |
| 129 | /// Bufferization of vector.gather. Replaced with a new vector.gather that |
| 130 | /// operates on a memref. |
| 131 | struct GatherOpInterface |
| 132 | : public BufferizableOpInterface::ExternalModel<GatherOpInterface, |
| 133 | vector::GatherOp> { |
| 134 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 135 | const AnalysisState &state) const { |
| 136 | assert(isa<RankedTensorType>(opOperand.get().getType()) && |
| 137 | "only tensor types expected" ); |
| 138 | return true; |
| 139 | } |
| 140 | |
| 141 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 142 | const AnalysisState &state) const { |
| 143 | assert(isa<RankedTensorType>(opOperand.get().getType()) && |
| 144 | "only tensor types expected" ); |
| 145 | return false; |
| 146 | } |
| 147 | |
| 148 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 149 | const AnalysisState &state) const { |
| 150 | return {}; |
| 151 | } |
| 152 | |
| 153 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 154 | const BufferizationOptions &options, |
| 155 | BufferizationState &state) const { |
| 156 | auto gatherOp = cast<vector::GatherOp>(op); |
| 157 | assert(isa<TensorType>(gatherOp.getBaseType()) && |
| 158 | "only tensor types expected" ); |
| 159 | FailureOr<Value> buffer = |
| 160 | getBuffer(rewriter, gatherOp.getBase(), options, state); |
| 161 | if (failed(Result: buffer)) |
| 162 | return failure(); |
| 163 | replaceOpWithNewBufferizedOp<vector::GatherOp>( |
| 164 | rewriter, gatherOp, gatherOp.getVectorType(), *buffer, |
| 165 | gatherOp.getIndices(), gatherOp.getIndexVec(), gatherOp.getMask(), |
| 166 | gatherOp.getPassThru()); |
| 167 | return success(); |
| 168 | } |
| 169 | }; |
| 170 | |
| 171 | /// Bufferization of vector.mask. Replaced with a new vector.mask that |
| 172 | /// operates on a memref. |
| 173 | struct MaskOpInterface |
| 174 | : public BufferizableOpInterface::ExternalModel<MaskOpInterface, |
| 175 | vector::MaskOp> { |
| 176 | AliasingOpOperandList |
| 177 | getAliasingOpOperands(Operation *op, Value value, |
| 178 | const AnalysisState &state) const { |
| 179 | // MaskOps do not have tensor OpOperands. The yielded values are the result |
| 180 | // of the wrapped op. |
| 181 | auto maskOp = cast<vector::MaskOp>(op); |
| 182 | size_t resultNum = std::distance(first: op->getOpResults().begin(), |
| 183 | last: llvm::find(Range: op->getOpResults(), Val: value)); |
| 184 | auto yieldOp = |
| 185 | cast<vector::YieldOp>(maskOp.getMaskRegion().front().getTerminator()); |
| 186 | return {{&yieldOp->getOpOperand(resultNum), BufferRelation::Equivalent}}; |
| 187 | } |
| 188 | |
| 189 | LogicalResult |
| 190 | resolveConflicts(Operation *op, RewriterBase &rewriter, |
| 191 | const AnalysisState &analysisState, |
| 192 | const BufferizationState &bufferizationState) const { |
| 193 | auto bufferizableOp = cast<BufferizableOpInterface>(op); |
| 194 | if (failed(bufferizableOp.resolveTensorOpOperandConflicts( |
| 195 | rewriter, analysisState, bufferizationState))) |
| 196 | return failure(); |
| 197 | |
| 198 | // TODO: Remove this function when vector.mask bodies can bufferize |
| 199 | // out-of-place. This is currently not supported because yielding allocs |
| 200 | // from a block leads to a memory leak and because vector.mask supports only |
| 201 | // a single op in its body. |
| 202 | auto maskOp = cast<vector::MaskOp>(op); |
| 203 | if (!maskOp.getMaskRegion() |
| 204 | .front() |
| 205 | .getOps<bufferization::AllocTensorOp>() |
| 206 | .empty()) |
| 207 | return op->emitOpError(message: "body must bufferize in-place" ); |
| 208 | |
| 209 | return success(); |
| 210 | } |
| 211 | |
| 212 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 213 | const BufferizationOptions &options, |
| 214 | BufferizationState &state) const { |
| 215 | auto maskOp = cast<vector::MaskOp>(op); |
| 216 | |
| 217 | // Do not bufferize if the masked op is not bufferizable. |
| 218 | Operation *maskedOp = maskOp.getMaskableOp(); |
| 219 | if (!options.dynCastBufferizableOp(maskedOp)) |
| 220 | return success(); |
| 221 | |
| 222 | // Update the terminator: Drop all operands that are not results of the |
| 223 | // masked op. |
| 224 | auto yieldOp = |
| 225 | cast<vector::YieldOp>(maskOp.getMaskRegion().front().getTerminator()); |
| 226 | SmallVector<Value> newReturnValues(maskOp->getNumResults(), Value()); |
| 227 | SmallVector<Value> newYieldedValues; |
| 228 | for (const auto &it : llvm::enumerate(yieldOp.getOperands())) { |
| 229 | if (llvm::is_contained(maskedOp->getOpResults(), it.value())) { |
| 230 | newYieldedValues.push_back(it.value()); |
| 231 | } else { |
| 232 | // This used to be a tensor result of the masked op, but is now a memref |
| 233 | // that is defined outside of the vector.mask op. |
| 234 | newReturnValues[it.index()] = it.value(); |
| 235 | } |
| 236 | } |
| 237 | rewriter.modifyOpInPlace(yieldOp, [&]() { |
| 238 | yieldOp.getOperandsMutable().assign(newYieldedValues); |
| 239 | }); |
| 240 | |
| 241 | // Create a new vector.mask op. |
| 242 | ValueRange newYieldedValuesRange(newYieldedValues); |
| 243 | TypeRange newResultTypes(newYieldedValuesRange); |
| 244 | auto newOp = rewriter.create<vector::MaskOp>( |
| 245 | op->getLoc(), newResultTypes, maskOp.getMask(), maskOp.getPassthru(), |
| 246 | /*maskableOp=*/nullptr, |
| 247 | /*maskRegionBuilder=*/[](OpBuilder &b, Operation *) {}); |
| 248 | newOp.getRegion().takeBody(maskOp.getMaskRegion()); |
| 249 | |
| 250 | // Replace all uses of the old vector.mask op. |
| 251 | int idx = 0; |
| 252 | for (int i = 0; i < static_cast<int>(maskOp->getNumResults()); ++i) { |
| 253 | if (!newReturnValues[i]) |
| 254 | newReturnValues[i] = newOp->getResult(idx++); |
| 255 | } |
| 256 | replaceOpWithBufferizedValues(rewriter, maskOp, newReturnValues); |
| 257 | return success(); |
| 258 | } |
| 259 | }; |
| 260 | |
| 261 | /// Bufferization of vector.yield. Replaced with a new vector.yield that |
| 262 | /// operates on a memref. |
| 263 | struct YieldOpInterface |
| 264 | : public BufferizableOpInterface::ExternalModel<YieldOpInterface, |
| 265 | vector::YieldOp> { |
| 266 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 267 | const AnalysisState &state) const { |
| 268 | return true; |
| 269 | } |
| 270 | |
| 271 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 272 | const AnalysisState &state) const { |
| 273 | return false; |
| 274 | } |
| 275 | |
| 276 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
| 277 | const AnalysisState &state) const { |
| 278 | return {{op->getParentOp()->getResult(idx: opOperand.getOperandNumber()), |
| 279 | BufferRelation::Equivalent}}; |
| 280 | } |
| 281 | |
| 282 | bool mustBufferizeInPlace(Operation *op, OpOperand &opOperand, |
| 283 | const AnalysisState &state) const { |
| 284 | // Yield operands always bufferize inplace. Otherwise, an alloc + copy |
| 285 | // may be generated inside the block. We should not return/yield allocations |
| 286 | // when possible. |
| 287 | return true; |
| 288 | } |
| 289 | |
| 290 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 291 | const BufferizationOptions &options, |
| 292 | BufferizationState &state) const { |
| 293 | auto yieldOp = cast<vector::YieldOp>(op); |
| 294 | |
| 295 | // Only supported as a vector.mask terminator. |
| 296 | auto maskOp = dyn_cast<vector::MaskOp>(yieldOp->getParentOp()); |
| 297 | if (!maskOp) |
| 298 | return yieldOp->emitError("unsupported vector::YieldOp parent" ); |
| 299 | |
| 300 | // Do not bufferize if the masked op is not bufferizable. |
| 301 | Operation *maskedOp = &maskOp.getMaskRegion().front().front(); |
| 302 | if (!options.dynCastBufferizableOp(maskedOp)) |
| 303 | return success(); |
| 304 | |
| 305 | // Create a new terminator with the same number of operands. Some of these |
| 306 | // may get dropped during the bufferization of vector.mask. |
| 307 | SmallVector<Value> newResults; |
| 308 | for (Value value : yieldOp.getOperands()) { |
| 309 | if (isa<TensorType>(value.getType())) { |
| 310 | FailureOr<Value> maybeBuffer = |
| 311 | getBuffer(rewriter, value, options, state); |
| 312 | if (failed(maybeBuffer)) |
| 313 | return failure(); |
| 314 | newResults.push_back(*maybeBuffer); |
| 315 | } else { |
| 316 | newResults.push_back(value); |
| 317 | } |
| 318 | } |
| 319 | |
| 320 | replaceOpWithNewBufferizedOp<vector::YieldOp>(rewriter, op, newResults); |
| 321 | return success(); |
| 322 | } |
| 323 | }; |
| 324 | |
| 325 | } // namespace |
| 326 | } // namespace vector |
| 327 | } // namespace mlir |
| 328 | |
| 329 | void mlir::vector::registerBufferizableOpInterfaceExternalModels( |
| 330 | DialectRegistry ®istry) { |
| 331 | registry.addExtension(extensionFn: +[](MLIRContext *ctx, vector::VectorDialect *dialect) { |
| 332 | TransferReadOp::attachInterface<TransferReadOpInterface>(*ctx); |
| 333 | TransferWriteOp::attachInterface<TransferWriteOpInterface>(*ctx); |
| 334 | GatherOp::attachInterface<GatherOpInterface>(*ctx); |
| 335 | MaskOp::attachInterface<MaskOpInterface>(*ctx); |
| 336 | YieldOp::attachInterface<YieldOpInterface>(*ctx); |
| 337 | }); |
| 338 | } |
| 339 | |