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/Arith/Transforms/BufferizableOpInterfaceImpl.h" |
10 | #include "mlir/Dialect/Arith/IR/Arith.h" |
11 | #include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h" |
12 | #include "mlir/Dialect/Bufferization/Transforms/BufferUtils.h" |
13 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
14 | #include "mlir/IR/Attributes.h" |
15 | #include "mlir/IR/Dialect.h" |
16 | #include "mlir/IR/Operation.h" |
17 | |
18 | using namespace mlir; |
19 | using namespace mlir::bufferization; |
20 | |
21 | namespace { |
22 | /// Bufferization of arith.constant. Replace with memref.get_global. |
23 | struct ConstantOpInterface |
24 | : public BufferizableOpInterface::ExternalModel<ConstantOpInterface, |
25 | arith::ConstantOp> { |
26 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
27 | const BufferizationOptions &options, |
28 | BufferizationState &state) const { |
29 | auto constantOp = cast<arith::ConstantOp>(op); |
30 | auto type = dyn_cast<RankedTensorType>(constantOp.getType()); |
31 | |
32 | // Only ranked tensors are supported. |
33 | if (!type) |
34 | return failure(); |
35 | |
36 | Attribute memorySpace; |
37 | if (auto memSpace = options.defaultMemorySpaceFn(type)) |
38 | memorySpace = *memSpace; |
39 | else |
40 | return constantOp->emitError("could not infer memory space" ); |
41 | |
42 | // Only constants inside a module are supported. |
43 | auto moduleOp = constantOp->getParentOfType<ModuleOp>(); |
44 | if (!moduleOp) |
45 | return failure(); |
46 | |
47 | // Create global memory segment and replace tensor with memref pointing to |
48 | // that memory segment. |
49 | FailureOr<memref::GlobalOp> globalOp = |
50 | getGlobalFor(constantOp, state.getSymbolTables(), |
51 | options.bufferAlignment, memorySpace); |
52 | if (failed(globalOp)) |
53 | return failure(); |
54 | memref::GlobalOp globalMemref = *globalOp; |
55 | replaceOpWithNewBufferizedOp<memref::GetGlobalOp>( |
56 | rewriter, op, globalMemref.getType(), globalMemref.getName()); |
57 | |
58 | return success(); |
59 | } |
60 | |
61 | bool isWritable(Operation *op, Value value, |
62 | const AnalysisState &state) const { |
63 | // Memory locations returned by memref::GetGlobalOp may not be written to. |
64 | assert(isa<OpResult>(value)); |
65 | return false; |
66 | } |
67 | }; |
68 | |
69 | struct IndexCastOpInterface |
70 | : public BufferizableOpInterface::ExternalModel<IndexCastOpInterface, |
71 | arith::IndexCastOp> { |
72 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
73 | const AnalysisState &state) const { |
74 | return false; |
75 | } |
76 | |
77 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
78 | const AnalysisState &state) const { |
79 | return false; |
80 | } |
81 | |
82 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
83 | const AnalysisState &state) const { |
84 | return {{op->getResult(idx: 0), BufferRelation::Equivalent}}; |
85 | } |
86 | |
87 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
88 | const BufferizationOptions &options, |
89 | BufferizationState &state) const { |
90 | auto castOp = cast<arith::IndexCastOp>(op); |
91 | auto resultTensorType = cast<TensorType>(castOp.getType()); |
92 | |
93 | FailureOr<Value> source = |
94 | getBuffer(rewriter, castOp.getIn(), options, state); |
95 | if (failed(Result: source)) |
96 | return failure(); |
97 | auto sourceType = cast<BaseMemRefType>(Val: source->getType()); |
98 | |
99 | // Result type should have same layout and address space as the source type. |
100 | BaseMemRefType resultType; |
101 | if (auto rankedMemRefType = dyn_cast<MemRefType>(sourceType)) { |
102 | resultType = MemRefType::get( |
103 | rankedMemRefType.getShape(), resultTensorType.getElementType(), |
104 | rankedMemRefType.getLayout(), rankedMemRefType.getMemorySpace()); |
105 | } else { |
106 | auto unrankedMemrefType = cast<UnrankedMemRefType>(sourceType); |
107 | resultType = UnrankedMemRefType::get(resultTensorType.getElementType(), |
108 | unrankedMemrefType.getMemorySpace()); |
109 | } |
110 | |
111 | replaceOpWithNewBufferizedOp<arith::IndexCastOp>(rewriter, op, resultType, |
112 | *source); |
113 | return success(); |
114 | } |
115 | }; |
116 | |
117 | /// Bufferization of arith.select. Just replace the operands. |
118 | struct SelectOpInterface |
119 | : public BufferizableOpInterface::ExternalModel<SelectOpInterface, |
120 | arith::SelectOp> { |
121 | bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
122 | const AnalysisState &state) const { |
123 | return false; |
124 | } |
125 | |
126 | bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
127 | const AnalysisState &state) const { |
128 | return false; |
129 | } |
130 | |
131 | AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand, |
132 | const AnalysisState &state) const { |
133 | return {{op->getOpResult(idx: 0) /*result*/, BufferRelation::Equivalent, |
134 | /*isDefinite=*/false}}; |
135 | } |
136 | |
137 | LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
138 | const BufferizationOptions &options, |
139 | BufferizationState &state) const { |
140 | auto selectOp = cast<arith::SelectOp>(op); |
141 | Location loc = selectOp.getLoc(); |
142 | |
143 | // Elementwise conditions are not supported yet. To bufferize such an op, |
144 | // it could be lowered to an elementwise "linalg.generic" with a new |
145 | // "tensor.empty" out tensor, followed by "empty tensor elimination". Such |
146 | // IR will bufferize. |
147 | if (!selectOp.getCondition().getType().isInteger(1)) |
148 | return op->emitOpError(message: "only i1 condition values are supported" ); |
149 | |
150 | // TODO: It would be more efficient to copy the result of the `select` op |
151 | // instead of its OpOperands. In the worst case, 2 copies are inserted at |
152 | // the moment (one for each tensor). When copying the op result, only one |
153 | // copy would be needed. |
154 | FailureOr<Value> maybeTrueBuffer = |
155 | getBuffer(rewriter, selectOp.getTrueValue(), options, state); |
156 | FailureOr<Value> maybeFalseBuffer = |
157 | getBuffer(rewriter, selectOp.getFalseValue(), options, state); |
158 | if (failed(Result: maybeTrueBuffer) || failed(Result: maybeFalseBuffer)) |
159 | return failure(); |
160 | Value trueBuffer = *maybeTrueBuffer; |
161 | Value falseBuffer = *maybeFalseBuffer; |
162 | |
163 | // The "true" and the "false" operands must have the same type. If the |
164 | // buffers have different types, they differ only in their layout map. Cast |
165 | // both of them to the most dynamic MemRef type. |
166 | if (trueBuffer.getType() != falseBuffer.getType()) { |
167 | auto targetType = |
168 | bufferization::getBufferType(value: selectOp.getResult(), options, state); |
169 | if (failed(targetType)) |
170 | return failure(); |
171 | if (trueBuffer.getType() != *targetType) |
172 | trueBuffer = |
173 | rewriter.create<memref::CastOp>(loc, *targetType, trueBuffer); |
174 | if (falseBuffer.getType() != *targetType) |
175 | falseBuffer = |
176 | rewriter.create<memref::CastOp>(loc, *targetType, falseBuffer); |
177 | } |
178 | |
179 | replaceOpWithNewBufferizedOp<arith::SelectOp>( |
180 | rewriter, op, selectOp.getCondition(), trueBuffer, falseBuffer); |
181 | return success(); |
182 | } |
183 | |
184 | FailureOr<BaseMemRefType> |
185 | getBufferType(Operation *op, Value value, const BufferizationOptions &options, |
186 | const BufferizationState &state, |
187 | SmallVector<Value> &invocationStack) const { |
188 | auto selectOp = cast<arith::SelectOp>(op); |
189 | assert(value == selectOp.getResult() && "invalid value" ); |
190 | auto trueType = bufferization::getBufferType( |
191 | value: selectOp.getTrueValue(), options, state, invocationStack); |
192 | auto falseType = bufferization::getBufferType( |
193 | value: selectOp.getFalseValue(), options, state, invocationStack); |
194 | if (failed(trueType) || failed(falseType)) |
195 | return failure(); |
196 | if (*trueType == *falseType) |
197 | return *trueType; |
198 | if (trueType->getMemorySpace() != falseType->getMemorySpace()) |
199 | return op->emitError(message: "inconsistent memory space on true/false operands" ); |
200 | |
201 | // If the buffers have different types, they differ only in their layout |
202 | // map. |
203 | auto memrefType = llvm::cast<MemRefType>(*trueType); |
204 | return getMemRefTypeWithFullyDynamicLayout( |
205 | RankedTensorType::get(memrefType.getShape(), |
206 | memrefType.getElementType()), |
207 | memrefType.getMemorySpace()); |
208 | } |
209 | }; |
210 | |
211 | } // namespace |
212 | |
213 | void mlir::arith::registerBufferizableOpInterfaceExternalModels( |
214 | DialectRegistry ®istry) { |
215 | registry.addExtension(extensionFn: +[](MLIRContext *ctx, ArithDialect *dialect) { |
216 | ConstantOp::attachInterface<ConstantOpInterface>(*ctx); |
217 | IndexCastOp::attachInterface<IndexCastOpInterface>(*ctx); |
218 | SelectOp::attachInterface<SelectOpInterface>(*ctx); |
219 | }); |
220 | } |
221 | |