Jan Eilers | 2ffddda | 2021-02-03 09:14:30 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2021 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include "TestUtils.hpp" |
| 9 | |
| 10 | #include <armnn_delegate.hpp> |
| 11 | #include <armnn/DescriptorsFwd.hpp> |
| 12 | |
| 13 | #include <flatbuffers/flatbuffers.h> |
| 14 | #include <tensorflow/lite/interpreter.h> |
| 15 | #include <tensorflow/lite/kernels/register.h> |
| 16 | #include <tensorflow/lite/model.h> |
| 17 | #include <tensorflow/lite/schema/schema_generated.h> |
| 18 | #include <tensorflow/lite/version.h> |
| 19 | |
| 20 | #include <doctest/doctest.h> |
| 21 | |
| 22 | #include <string> |
| 23 | |
| 24 | namespace |
| 25 | { |
| 26 | |
| 27 | struct StridedSliceParams |
| 28 | { |
| 29 | StridedSliceParams(std::vector<int32_t>& inputTensorShape, |
| 30 | std::vector<int32_t>& beginTensorData, |
| 31 | std::vector<int32_t>& endTensorData, |
| 32 | std::vector<int32_t>& strideTensorData, |
| 33 | std::vector<int32_t>& outputTensorShape, |
| 34 | armnn::StridedSliceDescriptor& descriptor) |
| 35 | : m_InputTensorShape(inputTensorShape), |
| 36 | m_BeginTensorData(beginTensorData), |
| 37 | m_EndTensorData(endTensorData), |
| 38 | m_StrideTensorData(strideTensorData), |
| 39 | m_OutputTensorShape(outputTensorShape), |
| 40 | m_Descriptor (descriptor) {} |
| 41 | |
| 42 | std::vector<int32_t> m_InputTensorShape; |
| 43 | std::vector<int32_t> m_BeginTensorData; |
| 44 | std::vector<int32_t> m_EndTensorData; |
| 45 | std::vector<int32_t> m_StrideTensorData; |
| 46 | std::vector<int32_t> m_OutputTensorShape; |
| 47 | armnn::StridedSliceDescriptor m_Descriptor; |
| 48 | }; |
| 49 | |
| 50 | std::vector<char> CreateSliceTfLiteModel(tflite::TensorType tensorType, |
| 51 | const std::vector<int32_t>& inputTensorShape, |
| 52 | const std::vector<int32_t>& beginTensorData, |
| 53 | const std::vector<int32_t>& endTensorData, |
| 54 | const std::vector<int32_t>& strideTensorData, |
| 55 | const std::vector<int32_t>& beginTensorShape, |
| 56 | const std::vector<int32_t>& endTensorShape, |
| 57 | const std::vector<int32_t>& strideTensorShape, |
| 58 | const std::vector<int32_t>& outputTensorShape, |
| 59 | const int32_t beginMask, |
| 60 | const int32_t endMask, |
| 61 | const int32_t ellipsisMask, |
| 62 | const int32_t newAxisMask, |
| 63 | const int32_t ShrinkAxisMask, |
| 64 | const armnn::DataLayout& dataLayout) |
| 65 | { |
| 66 | using namespace tflite; |
| 67 | flatbuffers::FlatBufferBuilder flatBufferBuilder; |
| 68 | |
| 69 | std::array<flatbuffers::Offset<tflite::Buffer>, 4> buffers; |
| 70 | buffers[0] = CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})); |
| 71 | buffers[1] = CreateBuffer(flatBufferBuilder, |
| 72 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(beginTensorData.data()), |
| 73 | sizeof(int32_t) * beginTensorData.size())); |
| 74 | buffers[2] = CreateBuffer(flatBufferBuilder, |
| 75 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(endTensorData.data()), |
| 76 | sizeof(int32_t) * endTensorData.size())); |
| 77 | buffers[3] = CreateBuffer(flatBufferBuilder, |
| 78 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(strideTensorData.data()), |
| 79 | sizeof(int32_t) * strideTensorData.size())); |
| 80 | |
| 81 | std::array<flatbuffers::Offset<Tensor>, 5> tensors; |
| 82 | tensors[0] = CreateTensor(flatBufferBuilder, |
| 83 | flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(), |
| 84 | inputTensorShape.size()), |
| 85 | tensorType, |
| 86 | 0, |
| 87 | flatBufferBuilder.CreateString("input")); |
| 88 | tensors[1] = CreateTensor(flatBufferBuilder, |
| 89 | flatBufferBuilder.CreateVector<int32_t>(beginTensorShape.data(), |
| 90 | beginTensorShape.size()), |
| 91 | ::tflite::TensorType_INT32, |
| 92 | 1, |
| 93 | flatBufferBuilder.CreateString("begin_tensor")); |
| 94 | tensors[2] = CreateTensor(flatBufferBuilder, |
| 95 | flatBufferBuilder.CreateVector<int32_t>(endTensorShape.data(), |
| 96 | endTensorShape.size()), |
| 97 | ::tflite::TensorType_INT32, |
| 98 | 2, |
| 99 | flatBufferBuilder.CreateString("end_tensor")); |
| 100 | tensors[3] = CreateTensor(flatBufferBuilder, |
| 101 | flatBufferBuilder.CreateVector<int32_t>(strideTensorShape.data(), |
| 102 | strideTensorShape.size()), |
| 103 | ::tflite::TensorType_INT32, |
| 104 | 3, |
| 105 | flatBufferBuilder.CreateString("stride_tensor")); |
| 106 | tensors[4] = CreateTensor(flatBufferBuilder, |
| 107 | flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(), |
| 108 | outputTensorShape.size()), |
| 109 | tensorType, |
| 110 | 0, |
| 111 | flatBufferBuilder.CreateString("output")); |
| 112 | |
| 113 | |
| 114 | // create operator |
| 115 | tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_StridedSliceOptions; |
| 116 | flatbuffers::Offset<void> operatorBuiltinOptions = CreateStridedSliceOptions(flatBufferBuilder, |
| 117 | beginMask, |
| 118 | endMask, |
| 119 | ellipsisMask, |
| 120 | newAxisMask, |
| 121 | ShrinkAxisMask).Union(); |
| 122 | |
| 123 | const std::vector<int> operatorInputs{ 0, 1, 2, 3 }; |
| 124 | const std::vector<int> operatorOutputs{ 4 }; |
| 125 | flatbuffers::Offset <Operator> sliceOperator = |
| 126 | CreateOperator(flatBufferBuilder, |
| 127 | 0, |
| 128 | flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()), |
| 129 | flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()), |
| 130 | operatorBuiltinOptionsType, |
| 131 | operatorBuiltinOptions); |
| 132 | |
| 133 | const std::vector<int> subgraphInputs{ 0, 1, 2, 3 }; |
| 134 | const std::vector<int> subgraphOutputs{ 4 }; |
| 135 | flatbuffers::Offset <SubGraph> subgraph = |
| 136 | CreateSubGraph(flatBufferBuilder, |
| 137 | flatBufferBuilder.CreateVector(tensors.data(), tensors.size()), |
| 138 | flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()), |
| 139 | flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(), subgraphOutputs.size()), |
| 140 | flatBufferBuilder.CreateVector(&sliceOperator, 1)); |
| 141 | |
| 142 | flatbuffers::Offset <flatbuffers::String> modelDescription = |
| 143 | flatBufferBuilder.CreateString("ArmnnDelegate: StridedSlice Operator Model"); |
| 144 | flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, |
| 145 | BuiltinOperator_STRIDED_SLICE); |
| 146 | |
| 147 | flatbuffers::Offset <Model> flatbufferModel = |
| 148 | CreateModel(flatBufferBuilder, |
| 149 | TFLITE_SCHEMA_VERSION, |
| 150 | flatBufferBuilder.CreateVector(&operatorCode, 1), |
| 151 | flatBufferBuilder.CreateVector(&subgraph, 1), |
| 152 | modelDescription, |
| 153 | flatBufferBuilder.CreateVector(buffers.data(), buffers.size())); |
| 154 | |
| 155 | flatBufferBuilder.Finish(flatbufferModel); |
| 156 | |
| 157 | return std::vector<char>(flatBufferBuilder.GetBufferPointer(), |
| 158 | flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize()); |
| 159 | } |
| 160 | |
| 161 | template <typename T> |
| 162 | void StridedSliceTestImpl(std::vector<armnn::BackendId>& backends, |
| 163 | std::vector<T>& inputValues, |
| 164 | std::vector<T>& expectedOutputValues, |
| 165 | std::vector<int32_t>& beginTensorData, |
| 166 | std::vector<int32_t>& endTensorData, |
| 167 | std::vector<int32_t>& strideTensorData, |
| 168 | std::vector<int32_t>& inputTensorShape, |
| 169 | std::vector<int32_t>& beginTensorShape, |
| 170 | std::vector<int32_t>& endTensorShape, |
| 171 | std::vector<int32_t>& strideTensorShape, |
| 172 | std::vector<int32_t>& outputTensorShape, |
| 173 | const int32_t beginMask = 0, |
| 174 | const int32_t endMask = 0, |
| 175 | const int32_t ellipsisMask = 0, |
| 176 | const int32_t newAxisMask = 0, |
| 177 | const int32_t ShrinkAxisMask = 0, |
| 178 | const armnn::DataLayout& dataLayout = armnn::DataLayout::NHWC) |
| 179 | { |
| 180 | using namespace tflite; |
| 181 | std::vector<char> modelBuffer = CreateSliceTfLiteModel( |
| 182 | ::tflite::TensorType_FLOAT32, |
| 183 | inputTensorShape, |
| 184 | beginTensorData, |
| 185 | endTensorData, |
| 186 | strideTensorData, |
| 187 | beginTensorShape, |
| 188 | endTensorShape, |
| 189 | strideTensorShape, |
| 190 | outputTensorShape, |
| 191 | beginMask, |
| 192 | endMask, |
| 193 | ellipsisMask, |
| 194 | newAxisMask, |
| 195 | ShrinkAxisMask, |
| 196 | dataLayout); |
| 197 | |
| 198 | auto tfLiteModel = GetModel(modelBuffer.data()); |
| 199 | |
| 200 | // Create TfLite Interpreters |
| 201 | std::unique_ptr<Interpreter> armnnDelegate; |
| 202 | CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver()) |
| 203 | (&armnnDelegate) == kTfLiteOk); |
| 204 | CHECK(armnnDelegate != nullptr); |
| 205 | CHECK(armnnDelegate->AllocateTensors() == kTfLiteOk); |
| 206 | |
| 207 | std::unique_ptr<Interpreter> tfLiteDelegate; |
| 208 | CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver()) |
| 209 | (&tfLiteDelegate) == kTfLiteOk); |
| 210 | CHECK(tfLiteDelegate != nullptr); |
| 211 | CHECK(tfLiteDelegate->AllocateTensors() == kTfLiteOk); |
| 212 | |
| 213 | // Create the ArmNN Delegate |
| 214 | armnnDelegate::DelegateOptions delegateOptions(backends); |
| 215 | std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)> |
| 216 | theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions), |
| 217 | armnnDelegate::TfLiteArmnnDelegateDelete); |
| 218 | CHECK(theArmnnDelegate != nullptr); |
| 219 | |
| 220 | // Modify armnnDelegateInterpreter to use armnnDelegate |
| 221 | CHECK(armnnDelegate->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk); |
| 222 | |
| 223 | // Set input data |
| 224 | armnnDelegate::FillInput<T>(tfLiteDelegate, 0, inputValues); |
| 225 | armnnDelegate::FillInput<T>(armnnDelegate, 0, inputValues); |
| 226 | |
| 227 | // Run EnqueWorkload |
| 228 | CHECK(tfLiteDelegate->Invoke() == kTfLiteOk); |
| 229 | CHECK(armnnDelegate->Invoke() == kTfLiteOk); |
| 230 | |
| 231 | // Compare output data |
| 232 | armnnDelegate::CompareOutputData<T>(tfLiteDelegate, |
| 233 | armnnDelegate, |
| 234 | outputTensorShape, |
| 235 | expectedOutputValues); |
| 236 | |
| 237 | tfLiteDelegate.reset(nullptr); |
| 238 | armnnDelegate.reset(nullptr); |
| 239 | } // End of StridedSlice Test |
| 240 | |
| 241 | } // anonymous namespace |