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" |
| 8 | |
| 9 | #include <armnn/Logging.hpp> |
| 10 | #include <Filesystem.hpp> |
| 11 | #include <InferenceTest.hpp> |
| 12 | |
| 13 | #if defined(ARMNN_SERIALIZER) |
| 14 | #include "armnnDeserializer/IDeserializer.hpp" |
| 15 | #endif |
| 16 | #if defined(ARMNN_CAFFE_PARSER) |
| 17 | #include "armnnCaffeParser/ICaffeParser.hpp" |
| 18 | #endif |
| 19 | #if defined(ARMNN_TF_PARSER) |
| 20 | #include "armnnTfParser/ITfParser.hpp" |
| 21 | #endif |
| 22 | #if defined(ARMNN_TF_LITE_PARSER) |
| 23 | #include "armnnTfLiteParser/ITfLiteParser.hpp" |
| 24 | #endif |
| 25 | #if defined(ARMNN_ONNX_PARSER) |
| 26 | #include "armnnOnnxParser/IOnnxParser.hpp" |
| 27 | #endif |
| 28 | |
| 29 | #include <future> |
| 30 | |
| 31 | template<typename TParser, typename TDataType> |
| 32 | int MainImpl(const ExecuteNetworkParams& params, |
| 33 | const std::shared_ptr<armnn::IRuntime>& runtime = nullptr) |
| 34 | { |
| 35 | using TContainer = mapbox::util::variant<std::vector<float>, std::vector<int>, std::vector<unsigned char>>; |
| 36 | |
| 37 | std::vector<TContainer> inputDataContainers; |
| 38 | |
| 39 | try |
| 40 | { |
| 41 | // Creates an InferenceModel, which will parse the model and load it into an IRuntime. |
| 42 | typename InferenceModel<TParser, TDataType>::Params inferenceModelParams; |
| 43 | inferenceModelParams.m_ModelPath = params.m_ModelPath; |
| 44 | inferenceModelParams.m_IsModelBinary = params.m_IsModelBinary; |
| 45 | inferenceModelParams.m_ComputeDevices = params.m_ComputeDevices; |
| 46 | inferenceModelParams.m_DynamicBackendsPath = params.m_DynamicBackendsPath; |
| 47 | inferenceModelParams.m_PrintIntermediateLayers = params.m_PrintIntermediate; |
| 48 | inferenceModelParams.m_VisualizePostOptimizationModel = params.m_EnableLayerDetails; |
| 49 | inferenceModelParams.m_ParseUnsupported = params.m_ParseUnsupported; |
| 50 | inferenceModelParams.m_InferOutputShape = params.m_InferOutputShape; |
| 51 | inferenceModelParams.m_EnableFastMath = params.m_EnableFastMath; |
| 52 | |
| 53 | for(const std::string& inputName: params.m_InputNames) |
| 54 | { |
| 55 | inferenceModelParams.m_InputBindings.push_back(inputName); |
| 56 | } |
| 57 | |
| 58 | for(unsigned int i = 0; i < params.m_InputTensorShapes.size(); ++i) |
| 59 | { |
| 60 | inferenceModelParams.m_InputShapes.push_back(*params.m_InputTensorShapes[i]); |
| 61 | } |
| 62 | |
| 63 | for(const std::string& outputName: params.m_OutputNames) |
| 64 | { |
| 65 | inferenceModelParams.m_OutputBindings.push_back(outputName); |
| 66 | } |
| 67 | |
| 68 | inferenceModelParams.m_SubgraphId = params.m_SubgraphId; |
| 69 | inferenceModelParams.m_EnableFp16TurboMode = params.m_EnableFp16TurboMode; |
| 70 | inferenceModelParams.m_EnableBf16TurboMode = params.m_EnableBf16TurboMode; |
| 71 | |
| 72 | InferenceModel<TParser, TDataType> model(inferenceModelParams, |
| 73 | params.m_EnableProfiling, |
| 74 | params.m_DynamicBackendsPath, |
| 75 | runtime); |
| 76 | |
| 77 | const size_t numInputs = inferenceModelParams.m_InputBindings.size(); |
| 78 | for(unsigned int i = 0; i < numInputs; ++i) |
| 79 | { |
| 80 | armnn::Optional<QuantizationParams> qParams = params.m_QuantizeInput ? |
| 81 | armnn::MakeOptional<QuantizationParams>( |
| 82 | model.GetInputQuantizationParams()) : |
| 83 | armnn::EmptyOptional(); |
| 84 | |
| 85 | armnn::Optional<std::string> dataFile = params.m_GenerateTensorData ? |
| 86 | armnn::EmptyOptional() : |
| 87 | armnn::MakeOptional<std::string>( |
| 88 | params.m_InputTensorDataFilePaths[i]); |
| 89 | |
| 90 | unsigned int numElements = model.GetInputSize(i); |
| 91 | if (params.m_InputTensorShapes.size() > i && params.m_InputTensorShapes[i]) |
| 92 | { |
| 93 | // If the user has provided a tensor shape for the current input, |
| 94 | // override numElements |
| 95 | numElements = params.m_InputTensorShapes[i]->GetNumElements(); |
| 96 | } |
| 97 | |
| 98 | TContainer tensorData; |
| 99 | PopulateTensorWithData(tensorData, |
| 100 | numElements, |
| 101 | params.m_InputTypes[i], |
| 102 | qParams, |
| 103 | dataFile); |
| 104 | |
| 105 | inputDataContainers.push_back(tensorData); |
| 106 | } |
| 107 | |
| 108 | const size_t numOutputs = inferenceModelParams.m_OutputBindings.size(); |
| 109 | std::vector<TContainer> outputDataContainers; |
| 110 | |
| 111 | for (unsigned int i = 0; i < numOutputs; ++i) |
| 112 | { |
| 113 | if (params.m_OutputTypes[i].compare("float") == 0) |
| 114 | { |
| 115 | outputDataContainers.push_back(std::vector<float>(model.GetOutputSize(i))); |
| 116 | } |
| 117 | else if (params.m_OutputTypes[i].compare("int") == 0) |
| 118 | { |
| 119 | outputDataContainers.push_back(std::vector<int>(model.GetOutputSize(i))); |
| 120 | } |
| 121 | else if (params.m_OutputTypes[i].compare("qasymm8") == 0) |
| 122 | { |
| 123 | outputDataContainers.push_back(std::vector<uint8_t>(model.GetOutputSize(i))); |
| 124 | } |
| 125 | else |
| 126 | { |
| 127 | ARMNN_LOG(fatal) << "Unsupported tensor data type \"" << params.m_OutputTypes[i] << "\". "; |
| 128 | return EXIT_FAILURE; |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | for (size_t x = 0; x < params.m_Iterations; x++) |
| 133 | { |
| 134 | // model.Run returns the inference time elapsed in EnqueueWorkload (in milliseconds) |
| 135 | auto inference_duration = model.Run(inputDataContainers, outputDataContainers); |
| 136 | |
| 137 | if (params.m_GenerateTensorData) |
| 138 | { |
| 139 | ARMNN_LOG(warning) << "The input data was generated, note that the output will not be useful"; |
| 140 | } |
| 141 | |
| 142 | // Print output tensors |
| 143 | const auto& infosOut = model.GetOutputBindingInfos(); |
| 144 | for (size_t i = 0; i < numOutputs; i++) |
| 145 | { |
| 146 | const armnn::TensorInfo& infoOut = infosOut[i].second; |
| 147 | auto outputTensorFile = params.m_OutputTensorFiles.empty() ? "" : params.m_OutputTensorFiles[i]; |
| 148 | |
| 149 | TensorPrinter printer(inferenceModelParams.m_OutputBindings[i], |
| 150 | infoOut, |
| 151 | outputTensorFile, |
| 152 | params.m_DequantizeOutput); |
| 153 | mapbox::util::apply_visitor(printer, outputDataContainers[i]); |
| 154 | } |
| 155 | |
| 156 | ARMNN_LOG(info) << "\nInference time: " << std::setprecision(2) |
| 157 | << std::fixed << inference_duration.count() << " ms\n"; |
| 158 | |
| 159 | // If thresholdTime == 0.0 (default), then it hasn't been supplied at command line |
| 160 | if (params.m_ThresholdTime != 0.0) |
| 161 | { |
| 162 | ARMNN_LOG(info) << "Threshold time: " << std::setprecision(2) |
| 163 | << std::fixed << params.m_ThresholdTime << " ms"; |
| 164 | auto thresholdMinusInference = params.m_ThresholdTime - inference_duration.count(); |
| 165 | ARMNN_LOG(info) << "Threshold time - Inference time: " << std::setprecision(2) |
| 166 | << std::fixed << thresholdMinusInference << " ms" << "\n"; |
| 167 | |
| 168 | if (thresholdMinusInference < 0) |
| 169 | { |
| 170 | std::string errorMessage = "Elapsed inference time is greater than provided threshold time."; |
| 171 | ARMNN_LOG(fatal) << errorMessage; |
| 172 | } |
| 173 | } |
| 174 | } |
| 175 | } |
| 176 | catch (const armnn::Exception& e) |
| 177 | { |
| 178 | ARMNN_LOG(fatal) << "Armnn Error: " << e.what(); |
| 179 | return EXIT_FAILURE; |
| 180 | } |
| 181 | |
| 182 | return EXIT_SUCCESS; |
| 183 | } |
| 184 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 185 | |
James Conroy | 7b4886f | 2019-04-11 10:23:58 +0100 | [diff] [blame] | 186 | // MAIN |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 187 | int main(int argc, const char* argv[]) |
| 188 | { |
| 189 | // Configures logging for both the ARMNN library and this test program. |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 190 | #ifdef NDEBUG |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 191 | armnn::LogSeverity level = armnn::LogSeverity::Info; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 192 | #else |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 193 | armnn::LogSeverity level = armnn::LogSeverity::Debug; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 194 | #endif |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 195 | armnn::ConfigureLogging(true, true, level); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 196 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 197 | |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 198 | // Get ExecuteNetwork parameters and runtime options from command line |
| 199 | ProgramOptions ProgramOptions(argc, argv); |
Narumol Prangnawarat | d8cc811 | 2020-03-24 13:54:05 +0000 | [diff] [blame] | 200 | |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 201 | // Create runtime |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 202 | std::shared_ptr<armnn::IRuntime> runtime(armnn::IRuntime::Create(ProgramOptions.m_RuntimeOptions)); |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 203 | |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 204 | std::string modelFormat = ProgramOptions.m_ExNetParams.m_ModelFormat; |
| 205 | |
| 206 | // Forward to implementation based on the parser type |
| 207 | if (modelFormat.find("armnn") != std::string::npos) |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 208 | { |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 209 | #if defined(ARMNN_SERIALIZER) |
| 210 | return MainImpl<armnnDeserializer::IDeserializer, float>(ProgramOptions.m_ExNetParams, runtime); |
| 211 | #else |
| 212 | ARMNN_LOG(fatal) << "Not built with serialization support."; |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 213 | return EXIT_FAILURE; |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 214 | #endif |
Finn Williams | d7fcafa | 2020-04-23 17:55:18 +0100 | [diff] [blame] | 215 | } |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 216 | else if (modelFormat.find("caffe") != std::string::npos) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 217 | { |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 218 | #if defined(ARMNN_CAFFE_PARSER) |
| 219 | return MainImpl<armnnCaffeParser::ICaffeParser, float>(ProgramOptions.m_ExNetParams, runtime); |
| 220 | #else |
| 221 | ARMNN_LOG(fatal) << "Not built with Caffe parser support."; |
| 222 | return EXIT_FAILURE; |
| 223 | #endif |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 224 | } |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 225 | else if (modelFormat.find("onnx") != std::string::npos) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 226 | { |
Jan Eilers | 4527490 | 2020-10-15 18:34:43 +0100 | [diff] [blame] | 227 | #if defined(ARMNN_ONNX_PARSER) |
| 228 | return MainImpl<armnnOnnxParser::IOnnxParser, float>(ProgramOptions.m_ExNetParams, runtime); |
| 229 | #else |
| 230 | ARMNN_LOG(fatal) << "Not built with Onnx parser support."; |
| 231 | return EXIT_FAILURE; |
| 232 | #endif |
| 233 | } |
| 234 | else if (modelFormat.find("tensorflow") != std::string::npos) |
| 235 | { |
| 236 | #if defined(ARMNN_TF_PARSER) |
| 237 | return MainImpl<armnnTfParser::ITfParser, float>(ProgramOptions.m_ExNetParams, runtime); |
| 238 | #else |
| 239 | ARMNN_LOG(fatal) << "Not built with Tensorflow parser support."; |
| 240 | return EXIT_FAILURE; |
| 241 | #endif |
| 242 | } |
| 243 | else if(modelFormat.find("tflite") != std::string::npos) |
| 244 | { |
| 245 | #if defined(ARMNN_TF_LITE_PARSER) |
| 246 | return MainImpl<armnnTfLiteParser::ITfLiteParser, float>(ProgramOptions.m_ExNetParams, runtime); |
| 247 | #else |
| 248 | ARMNN_LOG(fatal) << "Not built with Tensorflow-Lite parser support."; |
| 249 | return EXIT_FAILURE; |
| 250 | #endif |
| 251 | } |
| 252 | else |
| 253 | { |
| 254 | ARMNN_LOG(fatal) << "Unknown model format: '" << modelFormat |
| 255 | << "'. Please include 'caffe', 'tensorflow', 'tflite' or 'onnx'"; |
| 256 | return EXIT_FAILURE; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 257 | } |
| 258 | } |