|
| 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/Shape/Transforms/BufferizableOpInterfaceImpl.h" |
| 10 | + |
| 11 | +#include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h" |
| 12 | +#include "mlir/Dialect/Bufferization/IR/Bufferization.h" |
| 13 | +#include "mlir/Dialect/Shape/IR/Shape.h" |
| 14 | +#include "mlir/IR/Dialect.h" |
| 15 | +#include "mlir/IR/Operation.h" |
| 16 | +#include "mlir/IR/PatternMatch.h" |
| 17 | + |
| 18 | +using namespace mlir; |
| 19 | +using namespace mlir::bufferization; |
| 20 | +using namespace mlir::shape; |
| 21 | + |
| 22 | +namespace mlir { |
| 23 | +namespace shape { |
| 24 | +namespace { |
| 25 | + |
| 26 | +/// Bufferization of shape.assuming. |
| 27 | +struct AssumingOpInterface |
| 28 | + : public BufferizableOpInterface::ExternalModel<AssumingOpInterface, |
| 29 | + shape::AssumingOp> { |
| 30 | + SmallVector<OpOperand *> |
| 31 | + getAliasingOpOperand(Operation *op, OpResult opResult, |
| 32 | + const BufferizationState &state) const { |
| 33 | + // AssumingOps do not have tensor OpOperands. The yielded value can be any |
| 34 | + // SSA value that is in scope. To allow for use-def chain traversal through |
| 35 | + // AssumingOps in the analysis, the corresponding yield value is considered |
| 36 | + // to be aliasing with the result. |
| 37 | + auto assumingOp = cast<shape::AssumingOp>(op); |
| 38 | + size_t resultNum = std::distance(op->getOpResults().begin(), |
| 39 | + llvm::find(op->getOpResults(), opResult)); |
| 40 | + // TODO: Support multiple blocks. |
| 41 | + assert(assumingOp.getDoRegion().getBlocks().size() == 1 && |
| 42 | + "expected exactly 1 block"); |
| 43 | + auto yieldOp = dyn_cast<shape::AssumingYieldOp>( |
| 44 | + assumingOp.getDoRegion().front().getTerminator()); |
| 45 | + assert(yieldOp && "expected shape.assuming_yield terminator"); |
| 46 | + return {&yieldOp->getOpOperand(resultNum)}; |
| 47 | + } |
| 48 | + |
| 49 | + // TODO: For better bufferization results, this could return `true` only if |
| 50 | + // there is a memory write in the region. |
| 51 | + bool isMemoryWrite(Operation *op, OpResult opResult, |
| 52 | + const BufferizationState &state) const { |
| 53 | + // Similar to scf.if, results of this op are always considered memory writes |
| 54 | + // in the analysis. This is a useful pattern for all ops that have tensor |
| 55 | + // OpResults but no tensor OpOperands. By default, `isMemoryWrite` is |
| 56 | + // implemented in terms of `bufferizesToMemoryWrite`, which does not work on |
| 57 | + // ops without OpOperands. |
| 58 | + return true; |
| 59 | + } |
| 60 | + |
| 61 | + LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 62 | + const BufferizationState &state) const { |
| 63 | + auto assumingOp = cast<shape::AssumingOp>(op); |
| 64 | + |
| 65 | + // Compute new result types. |
| 66 | + SmallVector<Type> newResultTypes; |
| 67 | + for (Type type : assumingOp->getResultTypes()) { |
| 68 | + if (auto tensorType = type.dyn_cast<TensorType>()) { |
| 69 | + newResultTypes.push_back(getMemRefType(tensorType, state.getOptions())); |
| 70 | + } else { |
| 71 | + newResultTypes.push_back(type); |
| 72 | + } |
| 73 | + } |
| 74 | + |
| 75 | + // Create new op and move over region. |
| 76 | + auto newOp = rewriter.create<shape::AssumingOp>( |
| 77 | + op->getLoc(), newResultTypes, assumingOp.getWitness()); |
| 78 | + newOp.getDoRegion().takeBody(assumingOp.getRegion()); |
| 79 | + |
| 80 | + // Update terminator. |
| 81 | + assert(newOp.getDoRegion().getBlocks().size() == 1 && |
| 82 | + "only 1 block supported"); |
| 83 | + Block *newBlock = &newOp.getDoRegion().front(); |
| 84 | + auto yieldOp = cast<shape::AssumingYieldOp>(newBlock->getTerminator()); |
| 85 | + rewriter.setInsertionPoint(yieldOp); |
| 86 | + SmallVector<Value> newYieldValues; |
| 87 | + for (const auto &it : llvm::enumerate(yieldOp.operands())) { |
| 88 | + Value val = it.value(); |
| 89 | + if (val.getType().isa<TensorType>()) { |
| 90 | + newYieldValues.push_back(rewriter.create<bufferization::ToMemrefOp>( |
| 91 | + yieldOp.getLoc(), newResultTypes[it.index()], val)); |
| 92 | + } else { |
| 93 | + newYieldValues.push_back(val); |
| 94 | + } |
| 95 | + } |
| 96 | + rewriter.replaceOpWithNewOp<shape::AssumingYieldOp>(yieldOp, |
| 97 | + newYieldValues); |
| 98 | + |
| 99 | + // Update all uses of the old op. |
| 100 | + rewriter.setInsertionPointAfter(newOp); |
| 101 | + SmallVector<Value> newResults; |
| 102 | + for (const auto &it : llvm::enumerate(assumingOp->getResultTypes())) { |
| 103 | + if (it.value().isa<TensorType>()) { |
| 104 | + newResults.push_back(rewriter.create<bufferization::ToTensorOp>( |
| 105 | + assumingOp.getLoc(), newOp->getResult(it.index()))); |
| 106 | + } else { |
| 107 | + newResults.push_back(newOp->getResult(it.index())); |
| 108 | + } |
| 109 | + } |
| 110 | + |
| 111 | + // Replace old op. |
| 112 | + rewriter.replaceOp(assumingOp, newResults); |
| 113 | + |
| 114 | + return success(); |
| 115 | + } |
| 116 | + |
| 117 | + BufferRelation bufferRelation(Operation *op, OpResult opResult, |
| 118 | + const BufferizationState &state) const { |
| 119 | + return BufferRelation::Equivalent; |
| 120 | + } |
| 121 | +}; |
| 122 | + |
| 123 | +/// Bufferization of shape.assuming_yield. Bufferized as part of their enclosing |
| 124 | +/// ops, so this is for analysis only. |
| 125 | +struct AssumingYieldOpInterface |
| 126 | + : public BufferizableOpInterface::ExternalModel<AssumingYieldOpInterface, |
| 127 | + shape::AssumingOp> { |
| 128 | + bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand, |
| 129 | + const BufferizationState &state) const { |
| 130 | + return true; |
| 131 | + } |
| 132 | + |
| 133 | + bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand, |
| 134 | + const BufferizationState &state) const { |
| 135 | + return false; |
| 136 | + } |
| 137 | + |
| 138 | + SmallVector<OpResult> |
| 139 | + getAliasingOpResult(Operation *op, OpOperand &opOperand, |
| 140 | + const BufferizationState &state) const { |
| 141 | + assert(isa<shape::AssumingOp>(op->getParentOp()) && |
| 142 | + "expected that parent is an AssumingOp"); |
| 143 | + return {op->getParentOp()->getResult(opOperand.getOperandNumber())}; |
| 144 | + } |
| 145 | + |
| 146 | + bool mustBufferizeInPlace(Operation *op, OpOperand &opOperand, |
| 147 | + const BufferizationState &state) const { |
| 148 | + // Yield operands always bufferize inplace. Otherwise, an alloc + copy |
| 149 | + // may be generated inside the block. We should not return/yield allocations |
| 150 | + // when possible. |
| 151 | + return true; |
| 152 | + } |
| 153 | + |
| 154 | + LogicalResult bufferize(Operation *op, RewriterBase &rewriter, |
| 155 | + const BufferizationState &state) const { |
| 156 | + // Op is bufferized as part of AssumingOp. |
| 157 | + return failure(); |
| 158 | + } |
| 159 | +}; |
| 160 | + |
| 161 | +} // namespace |
| 162 | +} // namespace shape |
| 163 | +} // namespace mlir |
| 164 | + |
| 165 | +void mlir::shape::registerBufferizableOpInterfaceExternalModels( |
| 166 | + DialectRegistry ®istry) { |
| 167 | + registry.addOpInterface<shape::AssumingOp, AssumingOpInterface>(); |
| 168 | + registry.addOpInterface<shape::AssumingYieldOp, AssumingYieldOpInterface>(); |
| 169 | +} |
0 commit comments