| 1 | //===- RuntimeOpVerification.cpp - Op Verification ------------------------===// |
| 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/RuntimeOpVerification.h" |
| 10 | |
| 11 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 12 | #include "mlir/Dialect/Arith/Utils/Utils.h" |
| 13 | #include "mlir/Dialect/ControlFlow/IR/ControlFlow.h" |
| 14 | #include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h" |
| 15 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 16 | #include "mlir/Interfaces/RuntimeVerifiableOpInterface.h" |
| 17 | |
| 18 | using namespace mlir; |
| 19 | |
| 20 | namespace mlir { |
| 21 | namespace tensor { |
| 22 | namespace { |
| 23 | /// Generate a runtime check for lb <= value < ub. |
| 24 | Value generateInBoundsCheck(OpBuilder &builder, Location loc, Value value, |
| 25 | Value lb, Value ub) { |
| 26 | Value inBounds1 = builder.createOrFold<arith::CmpIOp>( |
| 27 | location: loc, args: arith::CmpIPredicate::sge, args&: value, args&: lb); |
| 28 | Value inBounds2 = builder.createOrFold<arith::CmpIOp>( |
| 29 | location: loc, args: arith::CmpIPredicate::slt, args&: value, args&: ub); |
| 30 | Value inBounds = |
| 31 | builder.createOrFold<arith::AndIOp>(location: loc, args&: inBounds1, args&: inBounds2); |
| 32 | return inBounds; |
| 33 | } |
| 34 | |
| 35 | struct CastOpInterface |
| 36 | : public RuntimeVerifiableOpInterface::ExternalModel<CastOpInterface, |
| 37 | CastOp> { |
| 38 | void generateRuntimeVerification(Operation *op, OpBuilder &builder, |
| 39 | Location loc) const { |
| 40 | auto castOp = cast<CastOp>(Val: op); |
| 41 | auto srcType = cast<TensorType>(Val: castOp.getSource().getType()); |
| 42 | |
| 43 | // Nothing to check if the result is an unranked tensor. |
| 44 | auto resultType = dyn_cast<RankedTensorType>(Val: castOp.getType()); |
| 45 | if (!resultType) |
| 46 | return; |
| 47 | |
| 48 | if (isa<UnrankedTensorType>(Val: srcType)) { |
| 49 | // Check rank. |
| 50 | Value srcRank = builder.create<RankOp>(location: loc, args: castOp.getSource()); |
| 51 | Value resultRank = |
| 52 | builder.create<arith::ConstantIndexOp>(location: loc, args: resultType.getRank()); |
| 53 | Value isSameRank = builder.create<arith::CmpIOp>( |
| 54 | location: loc, args: arith::CmpIPredicate::eq, args&: srcRank, args&: resultRank); |
| 55 | builder.create<cf::AssertOp>( |
| 56 | location: loc, args&: isSameRank, |
| 57 | args: RuntimeVerifiableOpInterface::generateErrorMessage(op, |
| 58 | msg: "rank mismatch" )); |
| 59 | } |
| 60 | |
| 61 | // Check dimension sizes. |
| 62 | for (const auto &it : llvm::enumerate(First: resultType.getShape())) { |
| 63 | // Static dim size -> static/dynamic dim size does not need verification. |
| 64 | if (auto rankedSrcType = dyn_cast<RankedTensorType>(Val&: srcType)) |
| 65 | if (!rankedSrcType.isDynamicDim(idx: it.index())) |
| 66 | continue; |
| 67 | |
| 68 | // Static/dynamic dim size -> dynamic dim size does not need verification. |
| 69 | if (resultType.isDynamicDim(idx: it.index())) |
| 70 | continue; |
| 71 | |
| 72 | Value srcDimSz = |
| 73 | builder.create<DimOp>(location: loc, args: castOp.getSource(), args: it.index()); |
| 74 | Value resultDimSz = |
| 75 | builder.create<arith::ConstantIndexOp>(location: loc, args: it.value()); |
| 76 | Value isSameSz = builder.create<arith::CmpIOp>( |
| 77 | location: loc, args: arith::CmpIPredicate::eq, args&: srcDimSz, args&: resultDimSz); |
| 78 | builder.create<cf::AssertOp>( |
| 79 | location: loc, args&: isSameSz, |
| 80 | args: RuntimeVerifiableOpInterface::generateErrorMessage( |
| 81 | op, msg: "size mismatch of dim " + std::to_string(val: it.index()))); |
| 82 | } |
| 83 | } |
| 84 | }; |
| 85 | |
| 86 | struct DimOpInterface |
| 87 | : public RuntimeVerifiableOpInterface::ExternalModel<DimOpInterface, |
| 88 | DimOp> { |
| 89 | void generateRuntimeVerification(Operation *op, OpBuilder &builder, |
| 90 | Location loc) const { |
| 91 | auto dimOp = cast<DimOp>(Val: op); |
| 92 | Value rank = builder.create<RankOp>(location: loc, args: dimOp.getSource()); |
| 93 | Value zero = builder.create<arith::ConstantIndexOp>(location: loc, args: 0); |
| 94 | builder.create<cf::AssertOp>( |
| 95 | location: loc, args: generateInBoundsCheck(builder, loc, value: dimOp.getIndex(), lb: zero, ub: rank), |
| 96 | args: RuntimeVerifiableOpInterface::generateErrorMessage( |
| 97 | op, msg: "index is out of bounds" )); |
| 98 | } |
| 99 | }; |
| 100 | |
| 101 | /// Verifies that the indices on extract/insert ops are in-bounds of the |
| 102 | /// tensor's index space: 0 <= index#i < dim#i |
| 103 | template <typename OpTy> |
| 104 | struct |
| 105 | : public RuntimeVerifiableOpInterface::ExternalModel< |
| 106 | ExtractInsertOpInterface<OpTy>, OpTy> { |
| 107 | void (Operation *op, OpBuilder &builder, |
| 108 | Location loc) const { |
| 109 | auto = cast<OpTy>(op); |
| 110 | |
| 111 | Value tensor; |
| 112 | if constexpr (std::is_same_v<OpTy, ExtractOp>) { |
| 113 | tensor = extractInsertOp.getTensor(); |
| 114 | } else if constexpr (std::is_same_v<OpTy, InsertOp>) { |
| 115 | tensor = extractInsertOp.getDest(); |
| 116 | } else { |
| 117 | llvm_unreachable("invalid op" ); |
| 118 | } |
| 119 | auto tensorType = cast<RankedTensorType>(Val: tensor.getType()); |
| 120 | auto rank = tensorType.getRank(); |
| 121 | if (rank == 0) { |
| 122 | // Nothing to check for 0-d tensors. |
| 123 | return; |
| 124 | } |
| 125 | |
| 126 | auto indices = extractInsertOp.getIndices(); |
| 127 | auto zero = builder.create<arith::ConstantIndexOp>(location: loc, args: 0); |
| 128 | Value assertCond; |
| 129 | for (auto i : llvm::seq<int64_t>(Begin: 0, End: rank)) { |
| 130 | Value dimOp = builder.createOrFold<tensor::DimOp>(location: loc, args&: tensor, args&: i); |
| 131 | Value inBounds = |
| 132 | generateInBoundsCheck(builder, loc, indices[i], zero, dimOp); |
| 133 | assertCond = |
| 134 | i > 0 ? builder.createOrFold<arith::AndIOp>(location: loc, args&: assertCond, args&: inBounds) |
| 135 | : inBounds; |
| 136 | } |
| 137 | builder.create<cf::AssertOp>( |
| 138 | location: loc, args&: assertCond, |
| 139 | args: RuntimeVerifiableOpInterface::generateErrorMessage( |
| 140 | op, msg: "out-of-bounds access" )); |
| 141 | } |
| 142 | }; |
| 143 | |
| 144 | struct |
| 145 | : public RuntimeVerifiableOpInterface::ExternalModel< |
| 146 | ExtractSliceOpInterface, ExtractSliceOp> { |
| 147 | void (Operation *op, OpBuilder &builder, |
| 148 | Location loc) const { |
| 149 | auto = cast<ExtractSliceOp>(Val: op); |
| 150 | RankedTensorType sourceType = extractSliceOp.getSource().getType(); |
| 151 | |
| 152 | // For each dimension, assert that: |
| 153 | // 0 <= offset < dim_size |
| 154 | // 0 <= offset + (size - 1) * stride < dim_size |
| 155 | Value zero = builder.create<arith::ConstantIndexOp>(location: loc, args: 0); |
| 156 | Value one = builder.create<arith::ConstantIndexOp>(location: loc, args: 1); |
| 157 | for (int64_t i = 0, e = sourceType.getRank(); i < e; ++i) { |
| 158 | Value offset = getValueOrCreateConstantIndexOp( |
| 159 | b&: builder, loc, ofr: extractSliceOp.getMixedOffsets()[i]); |
| 160 | Value size = getValueOrCreateConstantIndexOp( |
| 161 | b&: builder, loc, ofr: extractSliceOp.getMixedSizes()[i]); |
| 162 | Value stride = getValueOrCreateConstantIndexOp( |
| 163 | b&: builder, loc, ofr: extractSliceOp.getMixedStrides()[i]); |
| 164 | |
| 165 | // Verify that offset is in-bounds. |
| 166 | Value dimSize = builder.createOrFold<tensor::DimOp>( |
| 167 | location: loc, args: extractSliceOp.getSource(), args&: i); |
| 168 | Value offsetInBounds = |
| 169 | generateInBoundsCheck(builder, loc, value: offset, lb: zero, ub: dimSize); |
| 170 | builder.create<cf::AssertOp>( |
| 171 | location: loc, args&: offsetInBounds, |
| 172 | args: RuntimeVerifiableOpInterface::generateErrorMessage( |
| 173 | op, msg: "offset " + std::to_string(val: i) + " is out-of-bounds" )); |
| 174 | |
| 175 | // Verify that slice does not run out-of-bounds. |
| 176 | Value sizeMinusOne = builder.create<arith::SubIOp>(location: loc, args&: size, args&: one); |
| 177 | Value sizeMinusOneTimesStride = |
| 178 | builder.create<arith::MulIOp>(location: loc, args&: sizeMinusOne, args&: stride); |
| 179 | Value lastPos = |
| 180 | builder.create<arith::AddIOp>(location: loc, args&: offset, args&: sizeMinusOneTimesStride); |
| 181 | Value lastPosInBounds = |
| 182 | generateInBoundsCheck(builder, loc, value: lastPos, lb: zero, ub: dimSize); |
| 183 | builder.create<cf::AssertOp>( |
| 184 | location: loc, args&: lastPosInBounds, |
| 185 | args: RuntimeVerifiableOpInterface::generateErrorMessage( |
| 186 | op, msg: "extract_slice runs out-of-bounds along dimension " + |
| 187 | std::to_string(val: i))); |
| 188 | } |
| 189 | } |
| 190 | }; |
| 191 | } // namespace |
| 192 | } // namespace tensor |
| 193 | } // namespace mlir |
| 194 | |
| 195 | void mlir::tensor::registerRuntimeVerifiableOpInterfaceExternalModels( |
| 196 | DialectRegistry ®istry) { |
| 197 | registry.addExtension(extensionFn: +[](MLIRContext *ctx, tensor::TensorDialect *dialect) { |
| 198 | CastOp::attachInterface<CastOpInterface>(context&: *ctx); |
| 199 | DimOp::attachInterface<DimOpInterface>(context&: *ctx); |
| 200 | ExtractOp::attachInterface<ExtractInsertOpInterface<ExtractOp>>(context&: *ctx); |
| 201 | ExtractSliceOp::attachInterface<ExtractSliceOpInterface>(context&: *ctx); |
| 202 | InsertOp::attachInterface<ExtractInsertOpInterface<InsertOp>>(context&: *ctx); |
| 203 | |
| 204 | // Load additional dialects of which ops may get created. |
| 205 | ctx->loadDialect<arith::ArithDialect, cf::ControlFlowDialect>(); |
| 206 | }); |
| 207 | } |
| 208 | |