Laurent Carlier | 749294b | 2020-06-01 09:03:17 +0100 | [diff] [blame] | 1 | // |
Sadik Armagan | a9c2ce1 | 2020-07-14 10:02:22 +0100 | [diff] [blame] | 2 | // Copyright © 2017 Arm Ltd and Contributors. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 5 | |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 6 | #include "NetworkExecutionUtils/NetworkExecutionUtils.hpp" |
| 7 | #include "ExecuteNetworkProgramOptions.hpp" |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 8 | #include <armnn/IAsyncExecutionCallback.hpp> |
| 9 | #include <AsyncExecutionCallback.hpp> |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 10 | |
| 11 | #include <armnn/Logging.hpp> |
Rob Hughes | 9542f90 | 2021-07-14 09:48:54 +0100 | [diff] [blame] | 12 | #include <armnnUtils/Filesystem.hpp> |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 13 | #include <InferenceTest.hpp> |
| 14 | |
| 15 | #if defined(ARMNN_SERIALIZER) |
| 16 | #include "armnnDeserializer/IDeserializer.hpp" |
| 17 | #endif |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 18 | #if defined(ARMNN_TF_LITE_PARSER) |
| 19 | #include "armnnTfLiteParser/ITfLiteParser.hpp" |
| 20 | #endif |
| 21 | #if defined(ARMNN_ONNX_PARSER) |
| 22 | #include "armnnOnnxParser/IOnnxParser.hpp" |
| 23 | #endif |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 24 | #if defined(ARMNN_TFLITE_DELEGATE) |
| 25 | #include <armnn_delegate.hpp> |
| 26 | #include <DelegateOptions.hpp> |
| 27 | |
| 28 | #include <tensorflow/lite/builtin_ops.h> |
| 29 | #include <tensorflow/lite/c/builtin_op_data.h> |
| 30 | #include <tensorflow/lite/c/common.h> |
| 31 | #include <tensorflow/lite/optional_debug_tools.h> |
| 32 | #include <tensorflow/lite/kernels/builtin_op_kernels.h> |
| 33 | #include <tensorflow/lite/interpreter.h> |
| 34 | #include <tensorflow/lite/kernels/register.h> |
| 35 | #endif |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 36 | |
| 37 | #include <future> |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 38 | #if defined(ARMNN_TFLITE_DELEGATE) |
| 39 | int TfLiteDelegateMainImpl(const ExecuteNetworkParams& params, |
| 40 | const std::shared_ptr<armnn::IRuntime>& runtime = nullptr) |
| 41 | { |
| 42 | using namespace tflite; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 43 | |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 44 | std::unique_ptr<tflite::FlatBufferModel> model = tflite::FlatBufferModel::BuildFromFile(params.m_ModelPath.c_str()); |
| 45 | |
| 46 | auto tfLiteInterpreter = std::make_unique<Interpreter>(); |
| 47 | tflite::ops::builtin::BuiltinOpResolver resolver; |
| 48 | |
| 49 | tflite::InterpreterBuilder builder(*model, resolver); |
| 50 | builder(&tfLiteInterpreter); |
| 51 | tfLiteInterpreter->AllocateTensors(); |
| 52 | |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 53 | int status = 0; |
| 54 | if (params.m_TfLiteExecutor == ExecuteNetworkParams::TfLiteExecutor::ArmNNTfLiteDelegate) |
Sadik Armagan | 19a1c03 | 2021-01-20 12:17:00 +0000 | [diff] [blame] | 55 | { |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 56 | // Create the Armnn Delegate |
| 57 | armnnDelegate::DelegateOptions delegateOptions(params.m_ComputeDevices); |
| 58 | std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)> |
| 59 | theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions), |
| 60 | armnnDelegate::TfLiteArmnnDelegateDelete); |
| 61 | // Register armnn_delegate to TfLiteInterpreter |
| 62 | status = tfLiteInterpreter->ModifyGraphWithDelegate(std::move(theArmnnDelegate)); |
| 63 | if (status == kTfLiteError) |
| 64 | { |
| 65 | ARMNN_LOG(fatal) << "Could not register ArmNN TfLite Delegate to TfLiteInterpreter!"; |
| 66 | return EXIT_FAILURE; |
| 67 | } |
Sadik Armagan | 19a1c03 | 2021-01-20 12:17:00 +0000 | [diff] [blame] | 68 | } |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 69 | else |
| 70 | { |
| 71 | std::cout << "Running on TfLite without ArmNN delegate\n"; |
| 72 | } |
| 73 | |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 74 | |
| 75 | std::vector<std::string> inputBindings; |
| 76 | for (const std::string& inputName: params.m_InputNames) |
| 77 | { |
| 78 | inputBindings.push_back(inputName); |
| 79 | } |
| 80 | |
| 81 | armnn::Optional<std::string> dataFile = params.m_GenerateTensorData |
| 82 | ? armnn::EmptyOptional() |
| 83 | : armnn::MakeOptional<std::string>(params.m_InputTensorDataFilePaths[0]); |
| 84 | |
| 85 | const size_t numInputs = inputBindings.size(); |
| 86 | |
| 87 | for(unsigned int inputIndex = 0; inputIndex < numInputs; ++inputIndex) |
| 88 | { |
| 89 | int input = tfLiteInterpreter->inputs()[inputIndex]; |
Sadik Armagan | 15f7fae | 2020-11-18 09:37:03 +0000 | [diff] [blame] | 90 | TfLiteIntArray* inputDims = tfLiteInterpreter->tensor(input)->dims; |
| 91 | |
| 92 | long inputSize = 1; |
| 93 | for (unsigned int dim = 0; dim < static_cast<unsigned int>(inputDims->size); ++dim) |
| 94 | { |
| 95 | inputSize *= inputDims->data[dim]; |
| 96 | } |
| 97 | |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 98 | if (params.m_InputTypes[inputIndex].compare("float") == 0) |
| 99 | { |
| 100 | auto inputData = tfLiteInterpreter->typed_tensor<float>(input); |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 101 | |
Matthew Sloyan | f00f6c2 | 2020-12-07 13:33:24 +0000 | [diff] [blame] | 102 | if(inputData == NULL) |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 103 | { |
| 104 | ARMNN_LOG(fatal) << "Input tensor is null, input type: " |
| 105 | "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect."; |
| 106 | return EXIT_FAILURE; |
| 107 | } |
| 108 | |
Finn Williams | 5687018 | 2020-11-20 13:57:53 +0000 | [diff] [blame] | 109 | std::vector<float> tensorData; |
| 110 | PopulateTensorWithDataGeneric<float>(tensorData, |
| 111 | params.m_InputTensorShapes[inputIndex]->GetNumElements(), |
| 112 | dataFile, |
| 113 | [](const std::string& s) |
| 114 | { return std::stof(s); }); |
Sadik Armagan | 15f7fae | 2020-11-18 09:37:03 +0000 | [diff] [blame] | 115 | |
Finn Williams | 5687018 | 2020-11-20 13:57:53 +0000 | [diff] [blame] | 116 | std::copy(tensorData.begin(), tensorData.end(), inputData); |
| 117 | } |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 118 | else if (params.m_InputTypes[inputIndex].compare("qsymms8") == 0) |
Finn Williams | 5687018 | 2020-11-20 13:57:53 +0000 | [diff] [blame] | 119 | { |
| 120 | auto inputData = tfLiteInterpreter->typed_tensor<int8_t>(input); |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 121 | |
Matthew Sloyan | f00f6c2 | 2020-12-07 13:33:24 +0000 | [diff] [blame] | 122 | if(inputData == NULL) |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 123 | { |
| 124 | ARMNN_LOG(fatal) << "Input tensor is null, input type: " |
| 125 | "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect."; |
| 126 | return EXIT_FAILURE; |
| 127 | } |
| 128 | |
Finn Williams | 5687018 | 2020-11-20 13:57:53 +0000 | [diff] [blame] | 129 | std::vector<int8_t> tensorData; |
| 130 | PopulateTensorWithDataGeneric<int8_t>(tensorData, |
| 131 | params.m_InputTensorShapes[inputIndex]->GetNumElements(), |
| 132 | dataFile, |
| 133 | [](const std::string& s) |
| 134 | { return armnn::numeric_cast<int8_t>(std::stoi(s)); }); |
| 135 | |
| 136 | std::copy(tensorData.begin(), tensorData.end(), inputData); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 137 | } |
| 138 | else if (params.m_InputTypes[inputIndex].compare("int") == 0) |
| 139 | { |
| 140 | auto inputData = tfLiteInterpreter->typed_tensor<int32_t>(input); |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 141 | |
Matthew Sloyan | f00f6c2 | 2020-12-07 13:33:24 +0000 | [diff] [blame] | 142 | if(inputData == NULL) |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 143 | { |
| 144 | ARMNN_LOG(fatal) << "Input tensor is null, input type: " |
| 145 | "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect."; |
| 146 | return EXIT_FAILURE; |
| 147 | } |
| 148 | |
Finn Williams | 5687018 | 2020-11-20 13:57:53 +0000 | [diff] [blame] | 149 | std::vector<int32_t> tensorData; |
| 150 | PopulateTensorWithDataGeneric<int32_t>(tensorData, |
| 151 | params.m_InputTensorShapes[inputIndex]->GetNumElements(), |
| 152 | dataFile, |
| 153 | [](const std::string& s) |
| 154 | { return std::stoi(s); }); |
| 155 | |
| 156 | std::copy(tensorData.begin(), tensorData.end(), inputData); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 157 | } |
Mike Kelly | d7ed6d4 | 2021-07-21 09:42:43 +0100 | [diff] [blame] | 158 | else if (params.m_InputTypes[inputIndex].compare("qasymm8") == 0 || |
| 159 | params.m_InputTypes[inputIndex].compare("qasymmu8") == 0) |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 160 | { |
| 161 | auto inputData = tfLiteInterpreter->typed_tensor<uint8_t>(input); |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 162 | |
Matthew Sloyan | f00f6c2 | 2020-12-07 13:33:24 +0000 | [diff] [blame] | 163 | if(inputData == NULL) |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 164 | { |
| 165 | ARMNN_LOG(fatal) << "Input tensor is null, input type: " |
| 166 | "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect."; |
| 167 | return EXIT_FAILURE; |
| 168 | } |
| 169 | |
Finn Williams | 5687018 | 2020-11-20 13:57:53 +0000 | [diff] [blame] | 170 | std::vector<uint8_t> tensorData; |
| 171 | PopulateTensorWithDataGeneric<uint8_t>(tensorData, |
| 172 | params.m_InputTensorShapes[inputIndex]->GetNumElements(), |
| 173 | dataFile, |
| 174 | [](const std::string& s) |
| 175 | { return armnn::numeric_cast<uint8_t>(std::stoi(s)); }); |
| 176 | |
| 177 | std::copy(tensorData.begin(), tensorData.end(), inputData); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 178 | } |
Mike Kelly | d7ed6d4 | 2021-07-21 09:42:43 +0100 | [diff] [blame] | 179 | else if (params.m_InputTypes[inputIndex].compare("qasymms8") == 0) |
| 180 | { |
| 181 | auto inputData = tfLiteInterpreter->typed_tensor<int8_t>(input); |
| 182 | |
| 183 | if(inputData == NULL) |
| 184 | { |
| 185 | ARMNN_LOG(fatal) << "Input tensor is null, input type: " |
| 186 | "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect."; |
| 187 | return EXIT_FAILURE; |
| 188 | } |
| 189 | |
| 190 | std::vector<int8_t> tensorData; |
| 191 | PopulateTensorWithDataGeneric<int8_t>(tensorData, |
| 192 | params.m_InputTensorShapes[inputIndex]->GetNumElements(), |
| 193 | dataFile, |
| 194 | [](const std::string& s) |
| 195 | { return armnn::numeric_cast<int8_t>(std::stoi(s)); }); |
| 196 | |
| 197 | std::copy(tensorData.begin(), tensorData.end(), inputData); |
| 198 | } |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 199 | else |
| 200 | { |
| 201 | ARMNN_LOG(fatal) << "Unsupported input tensor data type \"" << params.m_InputTypes[inputIndex] << "\". "; |
| 202 | return EXIT_FAILURE; |
| 203 | } |
| 204 | } |
| 205 | |
| 206 | for (size_t x = 0; x < params.m_Iterations; x++) |
| 207 | { |
| 208 | // Run the inference |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 209 | status = tfLiteInterpreter->Invoke(); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 210 | |
| 211 | // Print out the output |
| 212 | for (unsigned int outputIndex = 0; outputIndex < params.m_OutputNames.size(); ++outputIndex) |
| 213 | { |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 214 | auto tfLiteDelegateOutputId = tfLiteInterpreter->outputs()[outputIndex]; |
Sadik Armagan | 15f7fae | 2020-11-18 09:37:03 +0000 | [diff] [blame] | 215 | TfLiteIntArray* outputDims = tfLiteInterpreter->tensor(tfLiteDelegateOutputId)->dims; |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 216 | |
Sadik Armagan | 15f7fae | 2020-11-18 09:37:03 +0000 | [diff] [blame] | 217 | long outputSize = 1; |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 218 | for (unsigned int dim = 0; dim < static_cast<unsigned int>(outputDims->size); ++dim) |
| 219 | { |
Sadik Armagan | 15f7fae | 2020-11-18 09:37:03 +0000 | [diff] [blame] | 220 | outputSize *= outputDims->data[dim]; |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 221 | } |
| 222 | |
| 223 | std::cout << params.m_OutputNames[outputIndex] << ": "; |
| 224 | if (params.m_OutputTypes[outputIndex].compare("float") == 0) |
| 225 | { |
| 226 | auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateOutputId); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 227 | if(tfLiteDelageOutputData == NULL) |
| 228 | { |
| 229 | ARMNN_LOG(fatal) << "Output tensor is null, output type: " |
| 230 | "\"" << params.m_OutputTypes[outputIndex] << "\" may be incorrect."; |
| 231 | return EXIT_FAILURE; |
| 232 | } |
| 233 | |
| 234 | for (int i = 0; i < outputSize; ++i) |
| 235 | { |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 236 | printf("%f ", tfLiteDelageOutputData[i]); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 237 | } |
| 238 | } |
| 239 | else if (params.m_OutputTypes[outputIndex].compare("int") == 0) |
| 240 | { |
| 241 | auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<int32_t>(tfLiteDelegateOutputId); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 242 | if(tfLiteDelageOutputData == NULL) |
| 243 | { |
| 244 | ARMNN_LOG(fatal) << "Output tensor is null, output type: " |
| 245 | "\"" << params.m_OutputTypes[outputIndex] << "\" may be incorrect."; |
| 246 | return EXIT_FAILURE; |
| 247 | } |
| 248 | |
| 249 | for (int i = 0; i < outputSize; ++i) |
| 250 | { |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 251 | printf("%d ", tfLiteDelageOutputData[i]); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 252 | } |
| 253 | } |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 254 | else if (params.m_OutputTypes[outputIndex].compare("qsymms8") == 0) |
Finn Williams | 5687018 | 2020-11-20 13:57:53 +0000 | [diff] [blame] | 255 | { |
| 256 | auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<int8_t>(tfLiteDelegateOutputId); |
| 257 | if(tfLiteDelageOutputData == NULL) |
| 258 | { |
| 259 | ARMNN_LOG(fatal) << "Output tensor is null, output type: " |
| 260 | "\"" << params.m_OutputTypes[outputIndex] << "\" may be incorrect."; |
| 261 | return EXIT_FAILURE; |
| 262 | } |
| 263 | |
| 264 | for (int i = 0; i < outputSize; ++i) |
| 265 | { |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 266 | printf("%d ", tfLiteDelageOutputData[i]); |
Finn Williams | 5687018 | 2020-11-20 13:57:53 +0000 | [diff] [blame] | 267 | } |
| 268 | } |
Mike Kelly | d7ed6d4 | 2021-07-21 09:42:43 +0100 | [diff] [blame] | 269 | else if (params.m_OutputTypes[outputIndex].compare("qasymm8") == 0 || |
| 270 | params.m_OutputTypes[outputIndex].compare("qasymmu8") == 0) |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 271 | { |
| 272 | auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<uint8_t>(tfLiteDelegateOutputId); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 273 | if(tfLiteDelageOutputData == NULL) |
| 274 | { |
| 275 | ARMNN_LOG(fatal) << "Output tensor is null, output type: " |
| 276 | "\"" << params.m_OutputTypes[outputIndex] << "\" may be incorrect."; |
| 277 | return EXIT_FAILURE; |
| 278 | } |
| 279 | |
| 280 | for (int i = 0; i < outputSize; ++i) |
| 281 | { |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 282 | printf("%u ", tfLiteDelageOutputData[i]); |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 283 | } |
| 284 | } |
| 285 | else |
| 286 | { |
| 287 | ARMNN_LOG(fatal) << "Output tensor is null, output type: " |
| 288 | "\"" << params.m_OutputTypes[outputIndex] << |
| 289 | "\" may be incorrect. Output type can be specified with -z argument"; |
| 290 | return EXIT_FAILURE; |
| 291 | } |
| 292 | std::cout << std::endl; |
| 293 | } |
| 294 | } |
| 295 | |
| 296 | return status; |
| 297 | } |
| 298 | #endif |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 299 | template<typename TParser, typename TDataType> |
| 300 | int MainImpl(const ExecuteNetworkParams& params, |
| 301 | const std::shared_ptr<armnn::IRuntime>& runtime = nullptr) |
| 302 | { |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 303 | using namespace std::chrono; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 304 | |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 305 | std::vector<std::vector<TContainer>> inputs; |
| 306 | std::vector<std::vector<TContainer>> outputs; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 307 | |
| 308 | try |
| 309 | { |
| 310 | // Creates an InferenceModel, which will parse the model and load it into an IRuntime. |
| 311 | typename InferenceModel<TParser, TDataType>::Params inferenceModelParams; |
| 312 | inferenceModelParams.m_ModelPath = params.m_ModelPath; |
| 313 | inferenceModelParams.m_IsModelBinary = params.m_IsModelBinary; |
| 314 | inferenceModelParams.m_ComputeDevices = params.m_ComputeDevices; |
| 315 | inferenceModelParams.m_DynamicBackendsPath = params.m_DynamicBackendsPath; |
| 316 | inferenceModelParams.m_PrintIntermediateLayers = params.m_PrintIntermediate; |
| 317 | inferenceModelParams.m_VisualizePostOptimizationModel = params.m_EnableLayerDetails; |
| 318 | inferenceModelParams.m_ParseUnsupported = params.m_ParseUnsupported; |
| 319 | inferenceModelParams.m_InferOutputShape = params.m_InferOutputShape; |
| 320 | inferenceModelParams.m_EnableFastMath = params.m_EnableFastMath; |
Matthew Sloyan | 4243211 | 2021-01-08 10:30:51 +0000 | [diff] [blame] | 321 | inferenceModelParams.m_SaveCachedNetwork = params.m_SaveCachedNetwork; |
| 322 | inferenceModelParams.m_CachedNetworkFilePath = params.m_CachedNetworkFilePath; |
Matthew Sloyan | 0a7dc6b | 2021-02-10 16:50:53 +0000 | [diff] [blame] | 323 | inferenceModelParams.m_NumberOfThreads = params.m_NumberOfThreads; |
Finn Williams | 4064632 | 2021-02-11 16:16:42 +0000 | [diff] [blame] | 324 | inferenceModelParams.m_MLGOTuningFilePath = params.m_MLGOTuningFilePath; |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 325 | inferenceModelParams.m_AsyncEnabled = params.m_Concurrent; |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 326 | inferenceModelParams.m_ThreadPoolSize = params.m_ThreadPoolSize; |
Keith Davis | f487486 | 2021-08-09 16:49:18 +0100 | [diff] [blame] | 327 | inferenceModelParams.m_OutputDetailsToStdOut = params.m_OutputDetailsToStdOut; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 328 | |
| 329 | for(const std::string& inputName: params.m_InputNames) |
| 330 | { |
| 331 | inferenceModelParams.m_InputBindings.push_back(inputName); |
| 332 | } |
| 333 | |
| 334 | for(unsigned int i = 0; i < params.m_InputTensorShapes.size(); ++i) |
| 335 | { |
| 336 | inferenceModelParams.m_InputShapes.push_back(*params.m_InputTensorShapes[i]); |
| 337 | } |
| 338 | |
| 339 | for(const std::string& outputName: params.m_OutputNames) |
| 340 | { |
| 341 | inferenceModelParams.m_OutputBindings.push_back(outputName); |
| 342 | } |
| 343 | |
| 344 | inferenceModelParams.m_SubgraphId = params.m_SubgraphId; |
| 345 | inferenceModelParams.m_EnableFp16TurboMode = params.m_EnableFp16TurboMode; |
| 346 | inferenceModelParams.m_EnableBf16TurboMode = params.m_EnableBf16TurboMode; |
| 347 | |
| 348 | InferenceModel<TParser, TDataType> model(inferenceModelParams, |
| 349 | params.m_EnableProfiling, |
| 350 | params.m_DynamicBackendsPath, |
| 351 | runtime); |
| 352 | |
| 353 | const size_t numInputs = inferenceModelParams.m_InputBindings.size(); |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 354 | |
| 355 | armnn::Optional<QuantizationParams> qParams = params.m_QuantizeInput ? |
| 356 | armnn::MakeOptional<QuantizationParams>( |
| 357 | model.GetInputQuantizationParams()) : |
| 358 | armnn::EmptyOptional(); |
| 359 | |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 360 | if (params.m_InputTensorDataFilePaths.size() > numInputs) |
| 361 | { |
| 362 | ARMNN_LOG(info) << "Given network has " << numInputs << " input/s. One input-tensor-data file is required " |
| 363 | << "for each input. The user provided " |
| 364 | << params.m_InputTensorDataFilePaths.size() |
| 365 | << " input-tensor-data file/s which will be used to fill the input/s.\n"; |
| 366 | } |
| 367 | |
| 368 | for(unsigned int j = 0; j < params.m_Iterations ; ++j) |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 369 | { |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 370 | std::vector<TContainer> inputDataContainers; |
| 371 | for(unsigned int i = 0; i < numInputs; ++i) |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 372 | { |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 373 | // If there are less input files given than required for the execution of |
| 374 | // params.m_Iterations we simply start with the first input file again |
| 375 | size_t inputFileIndex = j * numInputs + i; |
| 376 | if (!params.m_InputTensorDataFilePaths.empty()) |
| 377 | { |
| 378 | inputFileIndex = inputFileIndex % params.m_InputTensorDataFilePaths.size(); |
| 379 | } |
| 380 | |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 381 | armnn::Optional<std::string> dataFile = params.m_GenerateTensorData ? |
| 382 | armnn::EmptyOptional() : |
| 383 | armnn::MakeOptional<std::string>( |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 384 | params.m_InputTensorDataFilePaths.at(inputFileIndex)); |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 385 | |
| 386 | unsigned int numElements = model.GetInputSize(i); |
| 387 | if (params.m_InputTensorShapes.size() > i && params.m_InputTensorShapes[i]) |
| 388 | { |
| 389 | // If the user has provided a tensor shape for the current input, |
| 390 | // override numElements |
| 391 | numElements = params.m_InputTensorShapes[i]->GetNumElements(); |
| 392 | } |
| 393 | |
| 394 | TContainer tensorData; |
| 395 | PopulateTensorWithData(tensorData, |
| 396 | numElements, |
| 397 | params.m_InputTypes[i], |
| 398 | qParams, |
| 399 | dataFile); |
| 400 | |
| 401 | inputDataContainers.push_back(tensorData); |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 402 | } |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 403 | inputs.push_back(inputDataContainers); |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 404 | } |
| 405 | |
| 406 | const size_t numOutputs = inferenceModelParams.m_OutputBindings.size(); |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 407 | |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 408 | for (unsigned int j = 0; j < params.m_Iterations; ++j) |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 409 | { |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 410 | std::vector <TContainer> outputDataContainers; |
| 411 | for (unsigned int i = 0; i < numOutputs; ++i) |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 412 | { |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 413 | if (params.m_OutputTypes[i].compare("float") == 0) |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 414 | { |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 415 | outputDataContainers.push_back(std::vector<float>(model.GetOutputSize(i))); |
Mike Kelly | d7ed6d4 | 2021-07-21 09:42:43 +0100 | [diff] [blame] | 416 | } |
| 417 | else if (params.m_OutputTypes[i].compare("int") == 0) |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 418 | { |
| 419 | outputDataContainers.push_back(std::vector<int>(model.GetOutputSize(i))); |
Mike Kelly | d7ed6d4 | 2021-07-21 09:42:43 +0100 | [diff] [blame] | 420 | } |
| 421 | else if (params.m_OutputTypes[i].compare("qasymm8") == 0 || |
| 422 | params.m_OutputTypes[i].compare("qasymmu8") == 0) |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 423 | { |
| 424 | outputDataContainers.push_back(std::vector<uint8_t>(model.GetOutputSize(i))); |
Mike Kelly | d7ed6d4 | 2021-07-21 09:42:43 +0100 | [diff] [blame] | 425 | } |
| 426 | else if (params.m_OutputTypes[i].compare("qasymms8") == 0) |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 427 | { |
| 428 | outputDataContainers.push_back(std::vector<int8_t>(model.GetOutputSize(i))); |
| 429 | } else |
| 430 | { |
| 431 | ARMNN_LOG(fatal) << "Unsupported tensor data type \"" << params.m_OutputTypes[i] << "\". "; |
| 432 | return EXIT_FAILURE; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 433 | } |
| 434 | } |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 435 | outputs.push_back(outputDataContainers); |
| 436 | } |
| 437 | |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 438 | if (params.m_Iterations > 1) |
| 439 | { |
| 440 | std::stringstream msg; |
| 441 | msg << "Network will be executed " << params.m_Iterations; |
| 442 | if (params.m_Concurrent) |
| 443 | { |
| 444 | msg << " times in an asynchronous manner. "; |
| 445 | } |
| 446 | else |
| 447 | { |
| 448 | msg << " times successively. "; |
| 449 | } |
| 450 | msg << "The input-tensor-data files will be reused recursively if the user didn't provide enough to " |
| 451 | "cover each execution."; |
| 452 | ARMNN_LOG(info) << msg.str(); |
| 453 | } |
| 454 | |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 455 | // Synchronous execution |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 456 | if (!params.m_Concurrent) |
| 457 | { |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 458 | for (size_t x = 0; x < params.m_Iterations; x++) |
| 459 | { |
| 460 | // model.Run returns the inference time elapsed in EnqueueWorkload (in milliseconds) |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 461 | auto inference_duration = model.Run(inputs[x], outputs[x]); |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 462 | |
| 463 | if (params.m_GenerateTensorData) |
| 464 | { |
| 465 | ARMNN_LOG(warning) << "The input data was generated, note that the output will not be useful"; |
| 466 | } |
| 467 | |
| 468 | // Print output tensors |
| 469 | const auto& infosOut = model.GetOutputBindingInfos(); |
| 470 | for (size_t i = 0; i < numOutputs; i++) |
| 471 | { |
| 472 | const armnn::TensorInfo& infoOut = infosOut[i].second; |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 473 | |
| 474 | // We've made sure before that the number of output files either equals numOutputs, in which case |
| 475 | // we override those files when processing the results of each iteration (only the result of the |
| 476 | // last iteration will be stored), or there are enough |
| 477 | // output files for each output of each iteration. |
| 478 | size_t outputFileIndex = x * numOutputs + i; |
| 479 | if (!params.m_OutputTensorFiles.empty()) |
| 480 | { |
| 481 | outputFileIndex = outputFileIndex % params.m_OutputTensorFiles.size(); |
| 482 | ARMNN_LOG(info) << "Writing output " << i << " named: '" |
| 483 | << inferenceModelParams.m_OutputBindings[i] |
| 484 | << "' of iteration: " << x+1 << " to file: '" |
| 485 | << params.m_OutputTensorFiles[outputFileIndex] << "'"; |
| 486 | } |
| 487 | auto outputTensorFile = params.m_OutputTensorFiles.empty() |
| 488 | ? "" |
| 489 | : params.m_OutputTensorFiles[outputFileIndex]; |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 490 | |
| 491 | TensorPrinter printer(inferenceModelParams.m_OutputBindings[i], |
| 492 | infoOut, |
| 493 | outputTensorFile, |
| 494 | params.m_DequantizeOutput); |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 495 | mapbox::util::apply_visitor(printer, outputs[x][i]); |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 496 | } |
| 497 | |
| 498 | ARMNN_LOG(info) << "\nInference time: " << std::setprecision(2) |
| 499 | << std::fixed << inference_duration.count() << " ms\n"; |
| 500 | |
| 501 | // If thresholdTime == 0.0 (default), then it hasn't been supplied at command line |
| 502 | if (params.m_ThresholdTime != 0.0) |
| 503 | { |
| 504 | ARMNN_LOG(info) << "Threshold time: " << std::setprecision(2) |
| 505 | << std::fixed << params.m_ThresholdTime << " ms"; |
| 506 | auto thresholdMinusInference = params.m_ThresholdTime - inference_duration.count(); |
| 507 | ARMNN_LOG(info) << "Threshold time - Inference time: " << std::setprecision(2) |
| 508 | << std::fixed << thresholdMinusInference << " ms" << "\n"; |
| 509 | |
| 510 | if (thresholdMinusInference < 0) |
| 511 | { |
| 512 | std::string errorMessage = "Elapsed inference time is greater than provided threshold time."; |
| 513 | ARMNN_LOG(fatal) << errorMessage; |
| 514 | } |
| 515 | } |
| 516 | } |
| 517 | } |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 518 | // Asynchronous execution using the Arm NN thread pool |
Kevin May | 94dd4db | 2021-05-26 16:01:08 +0100 | [diff] [blame] | 519 | else if (params.m_ThreadPoolSize >= 1) |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 520 | { |
| 521 | try |
| 522 | { |
| 523 | ARMNN_LOG(info) << "Asynchronous execution with Arm NN thread pool... \n"; |
Finn Williams | f364d53 | 2021-06-09 17:07:33 +0100 | [diff] [blame] | 524 | armnn::AsyncCallbackManager callbackManager; |
| 525 | std::unordered_map<armnn::InferenceId, std::vector<TContainer>&> inferenceOutputMap; |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 526 | |
| 527 | // Declare the latest and earliest inference times here to be used when calculating overall time |
| 528 | std::chrono::high_resolution_clock::time_point earliestStartTime; |
| 529 | std::chrono::high_resolution_clock::time_point latestEndTime = |
| 530 | std::chrono::high_resolution_clock::now(); |
| 531 | |
| 532 | // For the asynchronous execution, we are adding a pool of working memory handles (1 per thread) in the |
| 533 | // LoadedNetwork with each scheduled inference having a specific priority |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 534 | for (size_t i = 0; i < params.m_Iterations; ++i) |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 535 | { |
Finn Williams | f364d53 | 2021-06-09 17:07:33 +0100 | [diff] [blame] | 536 | std::shared_ptr<armnn::AsyncExecutionCallback> cb = callbackManager.GetNewCallback(); |
| 537 | inferenceOutputMap.insert({cb->GetInferenceId(), outputs[i]}); |
| 538 | model.RunAsync(inputs[i], outputs[i], cb); |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 539 | } |
| 540 | |
| 541 | // Check the results |
| 542 | unsigned int j = 0; |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 543 | for (size_t iteration = 0; iteration < params.m_Iterations; ++iteration) |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 544 | { |
Finn Williams | f364d53 | 2021-06-09 17:07:33 +0100 | [diff] [blame] | 545 | auto cb = callbackManager.GetNotifiedCallback(); |
| 546 | |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 547 | // Get the results |
| 548 | auto endTime = time_point_cast<std::chrono::milliseconds>(cb->GetEndTime()); |
| 549 | auto startTime = time_point_cast<std::chrono::milliseconds>(cb->GetStartTime()); |
| 550 | auto inferenceDuration = endTime - startTime; |
| 551 | |
| 552 | if (latestEndTime < cb->GetEndTime()) |
| 553 | { |
| 554 | latestEndTime = cb->GetEndTime(); |
| 555 | } |
| 556 | |
| 557 | if (earliestStartTime.time_since_epoch().count() == 0) |
| 558 | { |
| 559 | earliestStartTime = cb->GetStartTime(); |
| 560 | } |
| 561 | else if (earliestStartTime > cb->GetStartTime()) |
| 562 | { |
| 563 | earliestStartTime = cb->GetStartTime(); |
| 564 | } |
| 565 | |
| 566 | if (params.m_GenerateTensorData) |
| 567 | { |
| 568 | ARMNN_LOG(warning) << "The input data was generated, note that the output will not be useful"; |
| 569 | } |
| 570 | |
| 571 | // Print output tensors |
| 572 | const auto& infosOut = model.GetOutputBindingInfos(); |
| 573 | for (size_t i = 0; i < numOutputs; i++) |
| 574 | { |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 575 | // We've made sure before that the number of output files either equals numOutputs, in which |
| 576 | // case we override those files when processing the results of each iteration (only the result |
| 577 | // of the last iteration will be stored), or there are enough |
| 578 | // output files for each output of each iteration. |
| 579 | size_t outputFileIndex = iteration * numOutputs + i; |
| 580 | if (!params.m_OutputTensorFiles.empty()) |
| 581 | { |
| 582 | outputFileIndex = outputFileIndex % params.m_OutputTensorFiles.size(); |
| 583 | ARMNN_LOG(info) << "Writing output " << i << " named: '" |
| 584 | << inferenceModelParams.m_OutputBindings[i] |
| 585 | << "' of iteration: " << iteration+1 << " to file: '" |
| 586 | << params.m_OutputTensorFiles[outputFileIndex] << "'"; |
| 587 | } |
| 588 | |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 589 | const armnn::TensorInfo& infoOut = infosOut[i].second; |
| 590 | auto outputTensorFile = params.m_OutputTensorFiles.empty() |
| 591 | ? "" |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 592 | : params.m_OutputTensorFiles[outputFileIndex]; |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 593 | |
| 594 | TensorPrinter printer(inferenceModelParams.m_OutputBindings[i], |
| 595 | infoOut, |
| 596 | outputTensorFile, |
| 597 | params.m_DequantizeOutput); |
Finn Williams | f364d53 | 2021-06-09 17:07:33 +0100 | [diff] [blame] | 598 | mapbox::util::apply_visitor(printer, inferenceOutputMap.at(cb->GetInferenceId())[i]); |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 599 | } |
| 600 | |
| 601 | ARMNN_LOG(info) << "\nInference time: " << std::setprecision(2) |
| 602 | << std::fixed << inferenceDuration.count() << " ms\n"; |
| 603 | |
| 604 | // If thresholdTime == 0.0 (default), then it hasn't been supplied at command line |
| 605 | if (params.m_ThresholdTime != 0.0) |
| 606 | { |
| 607 | ARMNN_LOG(info) << "Threshold time: " << std::setprecision(2) |
| 608 | << std::fixed << params.m_ThresholdTime << " ms"; |
| 609 | auto thresholdMinusInference = |
| 610 | params.m_ThresholdTime - duration<double, std::milli>(inferenceDuration).count(); |
| 611 | ARMNN_LOG(info) << "Threshold time - Inference time: " << std::setprecision(2) |
| 612 | << std::fixed << thresholdMinusInference << " ms" << "\n"; |
| 613 | |
| 614 | if (thresholdMinusInference < 0) |
| 615 | { |
| 616 | ARMNN_LOG(fatal) << "Elapsed inference time is greater than provided threshold time. \n"; |
| 617 | } |
| 618 | } |
| 619 | ++j; |
| 620 | } |
| 621 | //print duration difference between overallStartTime and overallEndTime |
| 622 | auto overallEndTime = time_point_cast<std::chrono::milliseconds>(latestEndTime); |
| 623 | auto overallStartTime = time_point_cast<std::chrono::milliseconds>(earliestStartTime); |
| 624 | auto totalInferenceDuration = overallEndTime - overallStartTime; |
| 625 | ARMNN_LOG(info) << "\nOverall Inference time: " << std::setprecision(2) |
| 626 | << std::fixed << totalInferenceDuration.count() << " ms\n"; |
| 627 | } |
| 628 | catch (const armnn::Exception& e) |
| 629 | { |
| 630 | ARMNN_LOG(fatal) << "Armnn Error: " << e.what(); |
| 631 | return EXIT_FAILURE; |
| 632 | } |
| 633 | } |
| 634 | // Asynchronous execution using std::launch::async |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 635 | else |
| 636 | { |
| 637 | try |
| 638 | { |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 639 | ARMNN_LOG(info) << "Asynchronous Execution with std::launch:async... \n"; |
Finn Williams | f364d53 | 2021-06-09 17:07:33 +0100 | [diff] [blame] | 640 | std::vector<std::future<std::tuple<unsigned int, |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 641 | std::chrono::duration<double, std::milli>>>> inferenceResults; |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 642 | inferenceResults.reserve(params.m_Iterations); |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 643 | |
| 644 | // Create WorkingMemHandles for each inference |
| 645 | std::vector<std::unique_ptr<armnn::experimental::IWorkingMemHandle>> workingMemHandles; |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 646 | workingMemHandles.reserve(params.m_Iterations); |
| 647 | for (unsigned int i = 0; i < params.m_Iterations; ++i) |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 648 | { |
| 649 | workingMemHandles.push_back(model.CreateWorkingMemHandle()); |
| 650 | } |
| 651 | |
| 652 | // Run each inference in its own thread |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 653 | // start a timer |
| 654 | const auto start_time = armnn::GetTimeNow(); |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 655 | for (unsigned int i = 0; i < params.m_Iterations; ++i) |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 656 | { |
| 657 | armnn::experimental::IWorkingMemHandle& workingMemHandleRef = *workingMemHandles[i].get(); |
Finn Williams | f364d53 | 2021-06-09 17:07:33 +0100 | [diff] [blame] | 658 | |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 659 | inferenceResults.push_back(std::async( |
| 660 | std::launch::async, [&model, &workingMemHandleRef, &inputs, &outputs, i]() { |
Finn Williams | f364d53 | 2021-06-09 17:07:33 +0100 | [diff] [blame] | 661 | return model.RunAsync(workingMemHandleRef, inputs[i], outputs[i], i); |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 662 | } |
| 663 | )); |
| 664 | } |
| 665 | |
| 666 | // Check the results |
| 667 | for (unsigned int j = 0; j < inferenceResults.size(); ++j) |
| 668 | { |
| 669 | // Get the results |
| 670 | auto inferenceResult = inferenceResults[j].get(); |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 671 | auto inferenceDuration = std::get<1>(inferenceResult); |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 672 | auto inferenceID = std::get<0>(inferenceResult); |
| 673 | |
| 674 | if (params.m_GenerateTensorData) |
| 675 | { |
| 676 | ARMNN_LOG(warning) << "The input data was generated, note that the output will not be useful"; |
| 677 | } |
| 678 | |
| 679 | // Print output tensors |
| 680 | const auto& infosOut = model.GetOutputBindingInfos(); |
| 681 | for (size_t i = 0; i < numOutputs; i++) |
| 682 | { |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 683 | // We've made sure before that the number of output files either equals numOutputs, in which |
| 684 | // case we override those files when processing the results of each iteration (only the result |
| 685 | // of the last iteration will be stored), or there are enough |
| 686 | // output files for each output of each iteration. |
| 687 | size_t outputFileIndex = j * numOutputs + i; |
| 688 | if (!params.m_OutputTensorFiles.empty()) |
| 689 | { |
| 690 | outputFileIndex = outputFileIndex % params.m_OutputTensorFiles.size(); |
| 691 | ARMNN_LOG(info) << "Writing output " << i << " named: '" |
| 692 | << inferenceModelParams.m_OutputBindings[i] |
| 693 | << "' of iteration: " << j+1 << " to file: '" |
| 694 | << params.m_OutputTensorFiles[outputFileIndex] << "'"; |
| 695 | } |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 696 | const armnn::TensorInfo& infoOut = infosOut[i].second; |
| 697 | auto outputTensorFile = params.m_OutputTensorFiles.empty() |
| 698 | ? "" |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 699 | : params.m_OutputTensorFiles[outputFileIndex]; |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 700 | |
| 701 | TensorPrinter printer(inferenceModelParams.m_OutputBindings[i], |
| 702 | infoOut, |
| 703 | outputTensorFile, |
| 704 | params.m_DequantizeOutput); |
| 705 | mapbox::util::apply_visitor(printer, outputs[j][i]); |
| 706 | } |
| 707 | |
| 708 | ARMNN_LOG(info) << "\nInference time: " << std::setprecision(2) |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 709 | << std::fixed << inferenceDuration.count() << " ms\n"; |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 710 | |
| 711 | // If thresholdTime == 0.0 (default), then it hasn't been supplied at command line |
| 712 | if (params.m_ThresholdTime != 0.0) |
| 713 | { |
| 714 | ARMNN_LOG(info) << "Threshold time: " << std::setprecision(2) |
| 715 | << std::fixed << params.m_ThresholdTime << " ms"; |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 716 | auto thresholdMinusInference = params.m_ThresholdTime - inferenceDuration.count(); |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 717 | ARMNN_LOG(info) << "Threshold time - Inference time: " << std::setprecision(2) |
| 718 | << std::fixed << thresholdMinusInference << " ms" << "\n"; |
| 719 | |
| 720 | if (thresholdMinusInference < 0) |
| 721 | { |
| 722 | ARMNN_LOG(fatal) << "Elapsed inference time is greater than provided threshold time. \n"; |
| 723 | } |
| 724 | } |
| 725 | ARMNN_LOG(info) << "Asynchronous Execution is finished for Inference ID: " << inferenceID << " \n"; |
| 726 | |
| 727 | } |
Kevin May | b4b3ac9 | 2021-05-21 16:42:21 +0100 | [diff] [blame] | 728 | // finish timer |
| 729 | const auto duration = armnn::GetTimeDuration(start_time); |
| 730 | ARMNN_LOG(info) << "\nOverall Inference time: " << std::setprecision(2) |
| 731 | << std::fixed << duration.count() << " ms\n"; |
Sadik Armagan | a04a9d7 | 2021-04-27 10:02:10 +0100 | [diff] [blame] | 732 | } |
| 733 | catch (const armnn::Exception& e) |
| 734 | { |
| 735 | ARMNN_LOG(fatal) << "Armnn Error: " << e.what(); |
| 736 | return EXIT_FAILURE; |
| 737 | } |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 738 | } |
| 739 | } |
| 740 | catch (const armnn::Exception& e) |
| 741 | { |
| 742 | ARMNN_LOG(fatal) << "Armnn Error: " << e.what(); |
| 743 | return EXIT_FAILURE; |
| 744 | } |
| 745 | |
| 746 | return EXIT_SUCCESS; |
| 747 | } |
| 748 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 749 | |
James Conroy | 7b4886f | 2019-04-11 10:23:58 +0100 | [diff] [blame] | 750 | // MAIN |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 751 | int main(int argc, const char* argv[]) |
| 752 | { |
| 753 | // Configures logging for both the ARMNN library and this test program. |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 754 | #ifdef NDEBUG |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 755 | armnn::LogSeverity level = armnn::LogSeverity::Info; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 756 | #else |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 757 | armnn::LogSeverity level = armnn::LogSeverity::Debug; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 758 | #endif |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 759 | armnn::ConfigureLogging(true, true, level); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 760 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 761 | |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 762 | // Get ExecuteNetwork parameters and runtime options from command line |
Jan Eilers | f17fcd5 | 2021-07-26 22:20:00 +0100 | [diff] [blame] | 763 | // This might throw an InvalidArgumentException if the user provided invalid inputs |
| 764 | ProgramOptions ProgramOptions; |
| 765 | try { |
| 766 | ProgramOptions.ParseOptions(argc, argv); |
| 767 | } catch (const std::exception &e){ |
| 768 | ARMNN_LOG(fatal) << e.what(); |
| 769 | return EXIT_FAILURE; |
| 770 | } |
Narumol Prangnawarat | d8cc811 | 2020-03-24 13:54:05 +0000 | [diff] [blame] | 771 | |
Keith Davis | f487486 | 2021-08-09 16:49:18 +0100 | [diff] [blame] | 772 | if (ProgramOptions.m_ExNetParams.m_OutputDetailsToStdOut && !ProgramOptions.m_ExNetParams.m_EnableProfiling) |
| 773 | { |
| 774 | ARMNN_LOG(fatal) << "You must enable profiling if you would like to output layer details"; |
| 775 | return EXIT_FAILURE; |
| 776 | } |
| 777 | |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 778 | // Create runtime |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 779 | std::shared_ptr<armnn::IRuntime> runtime(armnn::IRuntime::Create(ProgramOptions.m_RuntimeOptions)); |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 780 | |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 781 | std::string modelFormat = ProgramOptions.m_ExNetParams.m_ModelFormat; |
| 782 | |
| 783 | // Forward to implementation based on the parser type |
| 784 | if (modelFormat.find("armnn") != std::string::npos) |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 785 | { |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 786 | #if defined(ARMNN_SERIALIZER) |
| 787 | return MainImpl<armnnDeserializer::IDeserializer, float>(ProgramOptions.m_ExNetParams, runtime); |
| 788 | #else |
| 789 | ARMNN_LOG(fatal) << "Not built with serialization support."; |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 790 | return EXIT_FAILURE; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 791 | #endif |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 792 | } |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 793 | else if (modelFormat.find("onnx") != std::string::npos) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 794 | { |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 795 | #if defined(ARMNN_ONNX_PARSER) |
| 796 | return MainImpl<armnnOnnxParser::IOnnxParser, float>(ProgramOptions.m_ExNetParams, runtime); |
| 797 | #else |
| 798 | ARMNN_LOG(fatal) << "Not built with Onnx parser support."; |
| 799 | return EXIT_FAILURE; |
| 800 | #endif |
| 801 | } |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 802 | else if(modelFormat.find("tflite") != std::string::npos) |
| 803 | { |
Finn Williams | f806c4d | 2021-02-22 15:13:12 +0000 | [diff] [blame] | 804 | if (ProgramOptions.m_ExNetParams.m_TfLiteExecutor == ExecuteNetworkParams::TfLiteExecutor::ArmNNTfLiteParser) |
| 805 | { |
| 806 | #if defined(ARMNN_TF_LITE_PARSER) |
| 807 | return MainImpl<armnnTfLiteParser::ITfLiteParser, float>(ProgramOptions.m_ExNetParams, runtime); |
| 808 | #else |
| 809 | ARMNN_LOG(fatal) << "Not built with Tensorflow-Lite parser support."; |
| 810 | return EXIT_FAILURE; |
| 811 | #endif |
| 812 | } |
| 813 | else if (ProgramOptions.m_ExNetParams.m_TfLiteExecutor == |
| 814 | ExecuteNetworkParams::TfLiteExecutor::ArmNNTfLiteDelegate || |
| 815 | ProgramOptions.m_ExNetParams.m_TfLiteExecutor == |
| 816 | ExecuteNetworkParams::TfLiteExecutor::TfliteInterpreter) |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 817 | { |
| 818 | #if defined(ARMNN_TF_LITE_DELEGATE) |
| 819 | return TfLiteDelegateMainImpl(ProgramOptions.m_ExNetParams, runtime); |
| 820 | #else |
Finn Williams | bbbefec | 2020-11-25 14:32:42 +0000 | [diff] [blame] | 821 | ARMNN_LOG(fatal) << "Not built with Arm NN Tensorflow-Lite delegate support."; |
Sadik Armagan | 5d03e31 | 2020-11-17 16:43:56 +0000 | [diff] [blame] | 822 | return EXIT_FAILURE; |
| 823 | #endif |
| 824 | } |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 825 | } |
| 826 | else |
| 827 | { |
| 828 | ARMNN_LOG(fatal) << "Unknown model format: '" << modelFormat |
Nikhil Raj | 5d955cf | 2021-04-19 16:59:48 +0100 | [diff] [blame] | 829 | << "'. Please include 'tflite' or 'onnx'"; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 830 | return EXIT_FAILURE; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 831 | } |
| 832 | } |