| 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/Linalg/Transforms/BufferizableOpInterfaceImpl.h" |
| 10 | #include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h" |
| 11 | #include "mlir/Dialect/Bufferization/IR/DstBufferizableOpInterfaceImpl.h" |
| 12 | #include "mlir/Dialect/Linalg/IR/Linalg.h" |
| 13 | #include "mlir/Dialect/SparseTensor/IR/SparseTensor.h" |
| 14 | #include "mlir/IR/Dialect.h" |
| 15 | #include "mlir/IR/Operation.h" |
| 16 | #include "mlir/Interfaces/DestinationStyleOpInterface.h" |
| 17 | |
| 18 | using namespace mlir; |
| 19 | using namespace linalg; |
| 20 | using namespace mlir::bufferization; |
| 21 | |
| 22 | namespace { |
| 23 | |
| 24 | /// Generic conversion for any DestinationStyleOpInterface on tensors. |
| 25 | static LogicalResult bufferizeDestinationStyleOpInterface( |
| 26 | RewriterBase &rewriter, DestinationStyleOpInterface op, |
| 27 | const BufferizationOptions &options, const BufferizationState &state) { |
| 28 | // Take a guard before anything else. |
| 29 | OpBuilder::InsertionGuard g(rewriter); |
| 30 | rewriter.setInsertionPoint(op); |
| 31 | |
| 32 | // Nothing to do. This op is already bufferized. |
| 33 | if (op.hasPureBufferSemantics()) |
| 34 | return success(); |
| 35 | |
| 36 | // Ensure op has only tensors. Allow mixed tensor-buffer mode on a per-need |
| 37 | // basis. |
| 38 | if (!op.hasPureTensorSemantics()) |
| 39 | return op->emitError() << "op does not have pure tensor semantics" ; |
| 40 | |
| 41 | // New input operands for the cloned op. |
| 42 | SmallVector<Value> newInputBuffers; |
| 43 | newInputBuffers.reserve(N: op.getNumDpsInputs()); |
| 44 | for (OpOperand *opOperand : op.getDpsInputOperands()) { |
| 45 | if (op.isScalar(opOperand)) { |
| 46 | newInputBuffers.push_back(Elt: opOperand->get()); |
| 47 | continue; |
| 48 | } |
| 49 | FailureOr<Value> buffer = |
| 50 | getBuffer(rewriter, value: opOperand->get(), options, state); |
| 51 | if (failed(Result: buffer)) |
| 52 | return failure(); |
| 53 | newInputBuffers.push_back(Elt: *buffer); |
| 54 | } |
| 55 | |
| 56 | // New output operands for the cloned op. |
| 57 | SmallVector<Value> newOutputBuffers; |
| 58 | for (OpResult opResult : op->getOpResults()) { |
| 59 | OpOperand *opOperand = op.getDpsInitOperand(i: opResult.getResultNumber()); |
| 60 | FailureOr<Value> resultBuffer = |
| 61 | getBuffer(rewriter, value: opOperand->get(), options, state); |
| 62 | if (failed(Result: resultBuffer)) |
| 63 | return failure(); |
| 64 | newOutputBuffers.push_back(Elt: *resultBuffer); |
| 65 | } |
| 66 | |
| 67 | // Merge input/output operands. |
| 68 | SmallVector<Value> newOperands = newInputBuffers; |
| 69 | newOperands.append(in_start: newOutputBuffers.begin(), in_end: newOutputBuffers.end()); |
| 70 | |
| 71 | // Set insertion point now that potential alloc/dealloc are introduced. |
| 72 | rewriter.setInsertionPoint(op); |
| 73 | // Clone the op, but use the new operands. Move the existing block into the |
| 74 | // new op. Since the new op does not have any tensor results, it does not |
| 75 | // return anything. |
| 76 | assert(op->getNumRegions() == 1 && "expected that op has 1 region" ); |
| 77 | OperationState opState(op->getLoc(), op->getName(), newOperands, TypeRange{}, |
| 78 | op->getAttrs()); |
| 79 | opState.addRegion(); |
| 80 | Operation *newOp = Operation::create(state: opState); |
| 81 | newOp->getRegion(index: 0).getBlocks().splice(where: newOp->getRegion(index: 0).begin(), |
| 82 | L2&: op->getRegion(index: 0).getBlocks()); |
| 83 | |
| 84 | // We don't want the rewriter tracks an incomplete operation, so insert new |
| 85 | // operation after op was fully constructed. |
| 86 | rewriter.insert(op: newOp); |
| 87 | |
| 88 | // Replace the results of the old op with the new output buffers. |
| 89 | replaceOpWithBufferizedValues(rewriter, op, values: newOutputBuffers); |
| 90 | |
| 91 | return success(); |
| 92 | } |
| 93 | |
| 94 | /// Bufferization of linalg.generic. Replace with a new linalg.generic that |
| 95 | /// operates entirely on memrefs. |
| 96 | template <typename OpTy> |
| 97 | struct LinalgOpInterface |
| 98 | : public DstBufferizableOpInterfaceExternalModel<LinalgOpInterface<OpTy>, |
| 99 | OpTy> { |
| 100 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 101 | const AnalysisState &state) const { |
| 102 | // Operand is read if it is used in the computation. |
| 103 | auto linalgOp = cast<linalg::LinalgOp>(Val: op); |
| 104 | return linalgOp.payloadUsesValueFromOperand(opOperand: &opOperand); |
| 105 | } |
| 106 | |
| 107 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 108 | const AnalysisState &state) const { |
| 109 | // Operand is written to if it is not an input/init. |
| 110 | auto dpsOp = cast<DestinationStyleOpInterface>(Val: op); |
| 111 | return dpsOp.isDpsInit(opOperand: &opOperand); |
| 112 | } |
| 113 | |
| 114 | bool bufferizesToElementwiseAccess(Operation *op, const AnalysisState &state, |
| 115 | ArrayRef<OpOperand *> opOperands) const { |
| 116 | auto linalgOp = cast<linalg::LinalgOp>(Val: op); |
| 117 | |
| 118 | // Accesses into sparse data structures are not necessarily elementwise. |
| 119 | if (sparse_tensor::hasAnySparseOperand(op: linalgOp)) |
| 120 | return false; |
| 121 | |
| 122 | // All loops must be parallel. |
| 123 | if (linalgOp.getNumLoops() != linalgOp.getNumParallelLoops()) |
| 124 | return false; |
| 125 | |
| 126 | // All index maps of tensors must be identity maps. |
| 127 | SmallVector<AffineMap> indexingMaps = linalgOp.getIndexingMapsArray(); |
| 128 | assert(linalgOp->getNumOperands() == indexingMaps.size() && |
| 129 | "unexpected number of indexing maps" ); |
| 130 | for (auto [operand, map] : |
| 131 | llvm::zip(t: linalgOp->getOpOperands(), u&: indexingMaps)) { |
| 132 | // Non-tensors do not participate in bufferization, so they can be |
| 133 | // ignored. |
| 134 | if (!isa<RankedTensorType, MemRefType>(Val: operand.get().getType())) |
| 135 | continue; |
| 136 | // Only consider operands in `opOperands`. |
| 137 | if (!llvm::is_contained(Range&: opOperands, Element: &operand)) |
| 138 | continue; |
| 139 | // TODO: This could be generalized to other indexing maps. (All indexing |
| 140 | // must be the same.) |
| 141 | if (!map.isIdentity()) |
| 142 | return false; |
| 143 | } |
| 144 | |
| 145 | return true; |
| 146 | } |
| 147 | |
| 148 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 149 | const BufferizationOptions &options, |
| 150 | BufferizationState &state) const { |
| 151 | return bufferizeDestinationStyleOpInterface( |
| 152 | rewriter, op: cast<DestinationStyleOpInterface>(Val: op), options, state); |
| 153 | } |
| 154 | }; |
| 155 | |
| 156 | /// Helper structure that iterates over all LinalgOps in `OpTys` and registers |
| 157 | /// the `BufferizableOpInterface` with each of them. |
| 158 | template <typename... Ops> |
| 159 | struct LinalgOpInterfaceHelper { |
| 160 | static void registerOpInterface(MLIRContext *ctx) { |
| 161 | (Ops::template attachInterface<LinalgOpInterface<Ops>>(*ctx), ...); |
| 162 | } |
| 163 | }; |
| 164 | |
| 165 | struct SoftmaxOpInterface |
| 166 | : public DstBufferizableOpInterfaceExternalModel<SoftmaxOpInterface, |
| 167 | linalg::SoftmaxOp> { |
| 168 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 169 | const AnalysisState &state) const { |
| 170 | // Output operand is not read. |
| 171 | auto softmaxOp = cast<linalg::SoftmaxOp>(Val: op); |
| 172 | return &opOperand == &softmaxOp.getInputMutable(); |
| 173 | } |
| 174 | |
| 175 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 176 | const BufferizationOptions &options, |
| 177 | BufferizationState &state) const { |
| 178 | auto softmaxOp = cast<linalg::SoftmaxOp>(Val: op); |
| 179 | FailureOr<Value> inputBuffer = |
| 180 | getBuffer(rewriter, value: softmaxOp.getInput(), options, state); |
| 181 | if (failed(Result: inputBuffer)) |
| 182 | return failure(); |
| 183 | FailureOr<Value> outputBuffer = |
| 184 | getBuffer(rewriter, value: softmaxOp.getOutput(), options, state); |
| 185 | if (failed(Result: outputBuffer)) |
| 186 | return failure(); |
| 187 | rewriter.create<linalg::SoftmaxOp>(location: softmaxOp.getLoc(), |
| 188 | /*result=*/args: TypeRange(), args&: *inputBuffer, |
| 189 | args&: *outputBuffer, args: softmaxOp.getDimension()); |
| 190 | replaceOpWithBufferizedValues(rewriter, op, values: *outputBuffer); |
| 191 | return success(); |
| 192 | } |
| 193 | }; |
| 194 | } // namespace |
| 195 | |
| 196 | void mlir::linalg::registerBufferizableOpInterfaceExternalModels( |
| 197 | DialectRegistry ®istry) { |
| 198 | registry.addExtension(extensionFn: +[](MLIRContext *ctx, linalg::LinalgDialect *dialect) { |
| 199 | // Register all Linalg structured ops. `LinalgOp` is an interface and it is |
| 200 | // not possible to attach an external interface to an existing interface. |
| 201 | // Therefore, attach the `BufferizableOpInterface` to all ops one-by-one. |
| 202 | LinalgOpInterfaceHelper< |
| 203 | #define GET_OP_LIST |
| 204 | #include "mlir/Dialect/Linalg/IR/LinalgStructuredOps.cpp.inc" |
| 205 | |
| 206 | >::registerOpInterface(ctx); |
| 207 | |
| 208 | SoftmaxOp::attachInterface<SoftmaxOpInterface>(context&: *ctx); |
| 209 | }); |
| 210 | } |
| 211 | |