//
// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "NetworkExecutionUtils/NetworkExecutionUtils.hpp"
#include "ExecuteNetworkProgramOptions.hpp"
#include <armnn/IAsyncExecutionCallback.hpp>
#include <AsyncExecutionCallback.hpp>

#include <armnn/Logging.hpp>
#include <Filesystem.hpp>
#include <InferenceTest.hpp>

#if defined(ARMNN_SERIALIZER)
#include "armnnDeserializer/IDeserializer.hpp"
#endif
#if defined(ARMNN_TF_LITE_PARSER)
#include "armnnTfLiteParser/ITfLiteParser.hpp"
#endif
#if defined(ARMNN_ONNX_PARSER)
#include "armnnOnnxParser/IOnnxParser.hpp"
#endif
#if defined(ARMNN_TFLITE_DELEGATE)
#include <armnn_delegate.hpp>
#include <DelegateOptions.hpp>

#include <tensorflow/lite/builtin_ops.h>
#include <tensorflow/lite/c/builtin_op_data.h>
#include <tensorflow/lite/c/common.h>
#include <tensorflow/lite/optional_debug_tools.h>
#include <tensorflow/lite/kernels/builtin_op_kernels.h>
#include <tensorflow/lite/interpreter.h>
#include <tensorflow/lite/kernels/register.h>
#endif

#include <future>
#if defined(ARMNN_TFLITE_DELEGATE)
int TfLiteDelegateMainImpl(const ExecuteNetworkParams& params,
                           const std::shared_ptr<armnn::IRuntime>& runtime = nullptr)
{
    using namespace tflite;

    std::unique_ptr<tflite::FlatBufferModel> model = tflite::FlatBufferModel::BuildFromFile(params.m_ModelPath.c_str());

    auto tfLiteInterpreter =  std::make_unique<Interpreter>();
    tflite::ops::builtin::BuiltinOpResolver resolver;

    tflite::InterpreterBuilder builder(*model, resolver);
    builder(&tfLiteInterpreter);
    tfLiteInterpreter->AllocateTensors();

    int status = 0;
    if (params.m_TfLiteExecutor == ExecuteNetworkParams::TfLiteExecutor::ArmNNTfLiteDelegate)
    {
        // Create the Armnn Delegate
        armnnDelegate::DelegateOptions delegateOptions(params.m_ComputeDevices);
        std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)>
                theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions),
                                 armnnDelegate::TfLiteArmnnDelegateDelete);
        // Register armnn_delegate to TfLiteInterpreter
        status = tfLiteInterpreter->ModifyGraphWithDelegate(std::move(theArmnnDelegate));
        if (status == kTfLiteError)
        {
            ARMNN_LOG(fatal) << "Could not register ArmNN TfLite Delegate to TfLiteInterpreter!";
            return EXIT_FAILURE;
        }
    }
    else
    {
        std::cout << "Running on TfLite without ArmNN delegate\n";
    }


    std::vector<std::string>  inputBindings;
    for (const std::string& inputName: params.m_InputNames)
    {
        inputBindings.push_back(inputName);
    }

    armnn::Optional<std::string> dataFile = params.m_GenerateTensorData
                                            ? armnn::EmptyOptional()
                                            : armnn::MakeOptional<std::string>(params.m_InputTensorDataFilePaths[0]);

    const size_t numInputs = inputBindings.size();

    for(unsigned int inputIndex = 0; inputIndex < numInputs; ++inputIndex)
    {
        int input = tfLiteInterpreter->inputs()[inputIndex];
        TfLiteIntArray* inputDims = tfLiteInterpreter->tensor(input)->dims;

        long inputSize = 1;
        for (unsigned int dim = 0; dim < static_cast<unsigned int>(inputDims->size); ++dim)
        {
            inputSize *=  inputDims->data[dim];
        }

        if (params.m_InputTypes[inputIndex].compare("float") == 0)
        {
            auto inputData = tfLiteInterpreter->typed_tensor<float>(input);

            if(inputData == NULL)
            {
                ARMNN_LOG(fatal) << "Input tensor is null, input type: "
                                    "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect.";
                return EXIT_FAILURE;
            }

            std::vector<float> tensorData;
            PopulateTensorWithDataGeneric<float>(tensorData,
                                                  params.m_InputTensorShapes[inputIndex]->GetNumElements(),
                                                  dataFile,
                                                  [](const std::string& s)
                                                  { return std::stof(s); });

            std::copy(tensorData.begin(), tensorData.end(), inputData);
        }
        else if (params.m_InputTypes[inputIndex].compare("qsymms8") == 0)
        {
            auto inputData = tfLiteInterpreter->typed_tensor<int8_t>(input);

            if(inputData == NULL)
            {
                ARMNN_LOG(fatal) << "Input tensor is null, input type: "
                                    "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect.";
                return EXIT_FAILURE;
            }

            std::vector<int8_t> tensorData;
            PopulateTensorWithDataGeneric<int8_t>(tensorData,
                                                  params.m_InputTensorShapes[inputIndex]->GetNumElements(),
                                                  dataFile,
                                                  [](const std::string& s)
                                                  { return armnn::numeric_cast<int8_t>(std::stoi(s)); });

            std::copy(tensorData.begin(), tensorData.end(), inputData);
        }
        else if (params.m_InputTypes[inputIndex].compare("int") == 0)
        {
            auto inputData = tfLiteInterpreter->typed_tensor<int32_t>(input);

            if(inputData == NULL)
            {
                ARMNN_LOG(fatal) << "Input tensor is null, input type: "
                                    "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect.";
                return EXIT_FAILURE;
            }

            std::vector<int32_t> tensorData;
            PopulateTensorWithDataGeneric<int32_t>(tensorData,
                                                   params.m_InputTensorShapes[inputIndex]->GetNumElements(),
                                                   dataFile,
                                                   [](const std::string& s)
                                                   { return std::stoi(s); });

            std::copy(tensorData.begin(), tensorData.end(), inputData);
        }
        else if (params.m_InputTypes[inputIndex].compare("qasymm8") == 0)
        {
            auto inputData = tfLiteInterpreter->typed_tensor<uint8_t>(input);

            if(inputData == NULL)
            {
                ARMNN_LOG(fatal) << "Input tensor is null, input type: "
                                    "\"" << params.m_InputTypes[inputIndex] << "\" may be incorrect.";
                return EXIT_FAILURE;
            }

            std::vector<uint8_t> tensorData;
            PopulateTensorWithDataGeneric<uint8_t>(tensorData,
                                                   params.m_InputTensorShapes[inputIndex]->GetNumElements(),
                                                   dataFile,
                                                   [](const std::string& s)
                                                   { return armnn::numeric_cast<uint8_t>(std::stoi(s)); });

            std::copy(tensorData.begin(), tensorData.end(), inputData);
        }
        else
        {
            ARMNN_LOG(fatal) << "Unsupported input tensor data type \"" << params.m_InputTypes[inputIndex] << "\". ";
            return EXIT_FAILURE;
        }
    }

    for (size_t x = 0; x < params.m_Iterations; x++)
    {
        // Run the inference
        status = tfLiteInterpreter->Invoke();

        // Print out the output
        for (unsigned int outputIndex = 0; outputIndex < params.m_OutputNames.size(); ++outputIndex)
        {
            auto tfLiteDelegateOutputId = tfLiteInterpreter->outputs()[outputIndex];
            TfLiteIntArray* outputDims = tfLiteInterpreter->tensor(tfLiteDelegateOutputId)->dims;

            long outputSize = 1;
            for (unsigned int dim = 0; dim < static_cast<unsigned int>(outputDims->size); ++dim)
            {
                outputSize *=  outputDims->data[dim];
            }

            std::cout << params.m_OutputNames[outputIndex] << ": ";
            if (params.m_OutputTypes[outputIndex].compare("float") == 0)
            {
                auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateOutputId);
                if(tfLiteDelageOutputData == NULL)
                {
                    ARMNN_LOG(fatal) << "Output tensor is null, output type: "
                                        "\"" << params.m_OutputTypes[outputIndex] << "\" may be incorrect.";
                    return EXIT_FAILURE;
                }

                for (int i = 0; i < outputSize; ++i)
                {
                    printf("%f ", tfLiteDelageOutputData[i]);
                }
            }
            else if (params.m_OutputTypes[outputIndex].compare("int") == 0)
            {
                auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<int32_t>(tfLiteDelegateOutputId);
                if(tfLiteDelageOutputData == NULL)
                {
                    ARMNN_LOG(fatal) << "Output tensor is null, output type: "
                                        "\"" << params.m_OutputTypes[outputIndex] << "\" may be incorrect.";
                    return EXIT_FAILURE;
                }

                for (int i = 0; i < outputSize; ++i)
                {
                    printf("%d ", tfLiteDelageOutputData[i]);
                }
            }
            else if (params.m_OutputTypes[outputIndex].compare("qsymms8") == 0)
            {
                auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<int8_t>(tfLiteDelegateOutputId);
                if(tfLiteDelageOutputData == NULL)
                {
                    ARMNN_LOG(fatal) << "Output tensor is null, output type: "
                                        "\"" << params.m_OutputTypes[outputIndex] << "\" may be incorrect.";
                    return EXIT_FAILURE;
                }

                for (int i = 0; i < outputSize; ++i)
                {
                    printf("%d ", tfLiteDelageOutputData[i]);
                }
            }
            else if (params.m_OutputTypes[outputIndex].compare("qasymm8") == 0)
            {
                auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<uint8_t>(tfLiteDelegateOutputId);
                if(tfLiteDelageOutputData == NULL)
                {
                    ARMNN_LOG(fatal) << "Output tensor is null, output type: "
                                        "\"" << params.m_OutputTypes[outputIndex] << "\" may be incorrect.";
                    return EXIT_FAILURE;
                }

                for (int i = 0; i < outputSize; ++i)
                {
                    printf("%u ", tfLiteDelageOutputData[i]);
                }
            }
            else
            {
                ARMNN_LOG(fatal) << "Output tensor is null, output type: "
                                    "\"" << params.m_OutputTypes[outputIndex] <<
                                 "\" may be incorrect. Output type can be specified with -z argument";
                return EXIT_FAILURE;
            }
            std::cout << std::endl;
        }
    }

    return status;
}
#endif
template<typename TParser, typename TDataType>
int MainImpl(const ExecuteNetworkParams& params,
             const std::shared_ptr<armnn::IRuntime>& runtime = nullptr)
{
    using namespace std::chrono;

    std::vector<std::vector<TContainer>> inputs;
    std::vector<std::vector<TContainer>> outputs;

    try
    {
        // Creates an InferenceModel, which will parse the model and load it into an IRuntime.
        typename InferenceModel<TParser, TDataType>::Params inferenceModelParams;
        inferenceModelParams.m_ModelPath                      = params.m_ModelPath;
        inferenceModelParams.m_IsModelBinary                  = params.m_IsModelBinary;
        inferenceModelParams.m_ComputeDevices                 = params.m_ComputeDevices;
        inferenceModelParams.m_DynamicBackendsPath            = params.m_DynamicBackendsPath;
        inferenceModelParams.m_PrintIntermediateLayers        = params.m_PrintIntermediate;
        inferenceModelParams.m_VisualizePostOptimizationModel = params.m_EnableLayerDetails;
        inferenceModelParams.m_ParseUnsupported               = params.m_ParseUnsupported;
        inferenceModelParams.m_InferOutputShape               = params.m_InferOutputShape;
        inferenceModelParams.m_EnableFastMath                 = params.m_EnableFastMath;
        inferenceModelParams.m_SaveCachedNetwork              = params.m_SaveCachedNetwork;
        inferenceModelParams.m_CachedNetworkFilePath          = params.m_CachedNetworkFilePath;
        inferenceModelParams.m_NumberOfThreads                = params.m_NumberOfThreads;
        inferenceModelParams.m_MLGOTuningFilePath             = params.m_MLGOTuningFilePath;
        inferenceModelParams.m_AsyncEnabled                   = params.m_Concurrent;
        inferenceModelParams.m_ThreadPoolSize                 = params.m_ThreadPoolSize;

        for(const std::string& inputName: params.m_InputNames)
        {
            inferenceModelParams.m_InputBindings.push_back(inputName);
        }

        for(unsigned int i = 0; i < params.m_InputTensorShapes.size(); ++i)
        {
            inferenceModelParams.m_InputShapes.push_back(*params.m_InputTensorShapes[i]);
        }

        for(const std::string& outputName: params.m_OutputNames)
        {
            inferenceModelParams.m_OutputBindings.push_back(outputName);
        }

        inferenceModelParams.m_SubgraphId          = params.m_SubgraphId;
        inferenceModelParams.m_EnableFp16TurboMode = params.m_EnableFp16TurboMode;
        inferenceModelParams.m_EnableBf16TurboMode = params.m_EnableBf16TurboMode;

        InferenceModel<TParser, TDataType> model(inferenceModelParams,
                                                 params.m_EnableProfiling,
                                                 params.m_DynamicBackendsPath,
                                                 runtime);

        const size_t numInputs = inferenceModelParams.m_InputBindings.size();

        armnn::Optional<QuantizationParams> qParams = params.m_QuantizeInput ?
                                                      armnn::MakeOptional<QuantizationParams>(
                                                          model.GetInputQuantizationParams()) :
                                                      armnn::EmptyOptional();

        for(unsigned int j = 0; j < params.m_SimultaneousIterations ; ++j)
        {
            std::vector<TContainer> inputDataContainers;
            for(unsigned int i = 0; i < numInputs; ++i)
            {
                armnn::Optional<std::string> dataFile = params.m_GenerateTensorData ?
                                                        armnn::EmptyOptional() :
                                                        armnn::MakeOptional<std::string>(
                                                            params.m_InputTensorDataFilePaths[(j * numInputs) + i]);

                unsigned int numElements = model.GetInputSize(i);
                if (params.m_InputTensorShapes.size() > i && params.m_InputTensorShapes[i])
                {
                    // If the user has provided a tensor shape for the current input,
                    // override numElements
                    numElements = params.m_InputTensorShapes[i]->GetNumElements();
                }

                TContainer tensorData;
                PopulateTensorWithData(tensorData,
                                       numElements,
                                       params.m_InputTypes[i],
                                       qParams,
                                       dataFile);

                inputDataContainers.push_back(tensorData);
            }
            inputs.push_back(inputDataContainers);
        }

        const size_t numOutputs = inferenceModelParams.m_OutputBindings.size();

        for (unsigned int j = 0; j < params.m_SimultaneousIterations; ++j)
        {
            std::vector <TContainer> outputDataContainers;
            for (unsigned int i = 0; i < numOutputs; ++i)
            {
                if (params.m_OutputTypes[i].compare("float") == 0)
                {
                    outputDataContainers.push_back(std::vector<float>(model.GetOutputSize(i)));
                } else if (params.m_OutputTypes[i].compare("int") == 0)
                {
                    outputDataContainers.push_back(std::vector<int>(model.GetOutputSize(i)));
                } else if (params.m_OutputTypes[i].compare("qasymm8") == 0)
                {
                    outputDataContainers.push_back(std::vector<uint8_t>(model.GetOutputSize(i)));
                } else if (params.m_OutputTypes[i].compare("qsymms8") == 0)
                {
                    outputDataContainers.push_back(std::vector<int8_t>(model.GetOutputSize(i)));
                } else
                {
                    ARMNN_LOG(fatal) << "Unsupported tensor data type \"" << params.m_OutputTypes[i] << "\". ";
                    return EXIT_FAILURE;
                }
            }
            outputs.push_back(outputDataContainers);
        }

        // Synchronous execution
        if (!params.m_Concurrent)
        {
            for (size_t x = 0; x < params.m_Iterations; x++)
            {
                // model.Run returns the inference time elapsed in EnqueueWorkload (in milliseconds)
                auto inference_duration = model.Run(inputs[0], outputs[0]);

                if (params.m_GenerateTensorData)
                {
                    ARMNN_LOG(warning) << "The input data was generated, note that the output will not be useful";
                }

                // Print output tensors
                const auto& infosOut = model.GetOutputBindingInfos();
                for (size_t i = 0; i < numOutputs; i++)
                {
                    const armnn::TensorInfo& infoOut = infosOut[i].second;
                    auto outputTensorFile = params.m_OutputTensorFiles.empty() ? "" : params.m_OutputTensorFiles[i];

                    TensorPrinter printer(inferenceModelParams.m_OutputBindings[i],
                                          infoOut,
                                          outputTensorFile,
                                          params.m_DequantizeOutput);
                    mapbox::util::apply_visitor(printer, outputs[0][i]);
                }

                ARMNN_LOG(info) << "\nInference time: " << std::setprecision(2)
                                << std::fixed << inference_duration.count() << " ms\n";

                // If thresholdTime == 0.0 (default), then it hasn't been supplied at command line
                if (params.m_ThresholdTime != 0.0)
                {
                    ARMNN_LOG(info) << "Threshold time: " << std::setprecision(2)
                                    << std::fixed << params.m_ThresholdTime << " ms";
                    auto thresholdMinusInference = params.m_ThresholdTime - inference_duration.count();
                    ARMNN_LOG(info) << "Threshold time - Inference time: " << std::setprecision(2)
                                    << std::fixed << thresholdMinusInference << " ms" << "\n";

                    if (thresholdMinusInference < 0)
                    {
                        std::string errorMessage = "Elapsed inference time is greater than provided threshold time.";
                        ARMNN_LOG(fatal) << errorMessage;
                    }
                }
            }
        }
        // Asynchronous execution using the Arm NN thread pool
        else if (params.m_ThreadPoolSize >= 1)
        {
            try
            {
                ARMNN_LOG(info) << "Asynchronous execution with Arm NN thread pool...  \n";
                std::vector<armnn::experimental::IAsyncExecutionCallbackPtr> callbacks;

                // Create callbacks that will be checked post scheduling
                for (size_t i = 0; i < params.m_SimultaneousIterations; ++i)
                {
                    // Point to ArmNN example implementation of AsyncExecutionCallback
                    callbacks.emplace_back(std::make_shared<armnn::experimental::AsyncExecutionCallback>());
                }

                // Declare the latest and earliest inference times here to be used when calculating overall time
                std::chrono::high_resolution_clock::time_point earliestStartTime;
                std::chrono::high_resolution_clock::time_point latestEndTime =
                    std::chrono::high_resolution_clock::now();

                // For the asynchronous execution, we are adding a pool of working memory handles (1 per thread) in the
                // LoadedNetwork with each scheduled inference having a specific priority
                for (size_t i = 0; i < callbacks.size(); ++i)
                {
                    model.RunAsync(inputs[i], outputs[i], callbacks[i]);
                }

                // Check the results
                unsigned int j = 0;
                for (armnn::experimental::IAsyncExecutionCallbackPtr cb : callbacks)
                {
                    // Get the results
                    auto endTime = time_point_cast<std::chrono::milliseconds>(cb->GetEndTime());
                    auto startTime = time_point_cast<std::chrono::milliseconds>(cb->GetStartTime());
                    auto inferenceDuration = endTime - startTime;

                    if (latestEndTime < cb->GetEndTime())
                    {
                        latestEndTime = cb->GetEndTime();
                    }

                    if (earliestStartTime.time_since_epoch().count() == 0)
                    {
                        earliestStartTime = cb->GetStartTime();
                    }
                    else if (earliestStartTime > cb->GetStartTime())
                    {
                        earliestStartTime = cb->GetStartTime();
                    }

                    if (params.m_GenerateTensorData)
                    {
                        ARMNN_LOG(warning) << "The input data was generated, note that the output will not be useful";
                    }

                    // Print output tensors
                    const auto& infosOut = model.GetOutputBindingInfos();
                    for (size_t i = 0; i < numOutputs; i++)
                    {
                        const armnn::TensorInfo& infoOut = infosOut[i].second;
                        auto outputTensorFile = params.m_OutputTensorFiles.empty()
                                                ? ""
                                                : params.m_OutputTensorFiles[(j * numOutputs) + i];

                        TensorPrinter printer(inferenceModelParams.m_OutputBindings[i],
                                              infoOut,
                                              outputTensorFile,
                                              params.m_DequantizeOutput);
                        mapbox::util::apply_visitor(printer, outputs[j][i]);
                    }

                    ARMNN_LOG(info) << "\nInference time: " << std::setprecision(2)
                                    << std::fixed << inferenceDuration.count() << " ms\n";

                     // If thresholdTime == 0.0 (default), then it hasn't been supplied at command line
                    if (params.m_ThresholdTime != 0.0)
                    {
                        ARMNN_LOG(info) << "Threshold time: " << std::setprecision(2)
                                        << std::fixed << params.m_ThresholdTime << " ms";
                        auto thresholdMinusInference =
                            params.m_ThresholdTime - duration<double, std::milli>(inferenceDuration).count();
                        ARMNN_LOG(info) << "Threshold time - Inference time: " << std::setprecision(2)
                                        << std::fixed << thresholdMinusInference << " ms" << "\n";

                        if (thresholdMinusInference < 0)
                        {
                            ARMNN_LOG(fatal) << "Elapsed inference time is greater than provided threshold time. \n";
                        }
                    }
                    ++j;
                }
                //print duration difference between overallStartTime and overallEndTime
                auto overallEndTime = time_point_cast<std::chrono::milliseconds>(latestEndTime);
                auto overallStartTime = time_point_cast<std::chrono::milliseconds>(earliestStartTime);
                auto totalInferenceDuration = overallEndTime - overallStartTime;
                ARMNN_LOG(info) << "\nOverall Inference time: " << std::setprecision(2)
                                << std::fixed << totalInferenceDuration.count() << " ms\n";
            }
            catch (const armnn::Exception& e)
            {
                ARMNN_LOG(fatal) << "Armnn Error: " << e.what();
                return EXIT_FAILURE;
            }
        }
        // Asynchronous execution using std::launch::async
        else
        {
            try
            {
                ARMNN_LOG(info) << "Asynchronous Execution with std::launch:async...  \n";
                std::vector<std::future<std::tuple<armnn::profiling::ProfilingGuid,
                    std::chrono::duration<double, std::milli>>>> inferenceResults;
                inferenceResults.reserve(params.m_SimultaneousIterations);

                // Create WorkingMemHandles for each inference
                std::vector<std::unique_ptr<armnn::experimental::IWorkingMemHandle>> workingMemHandles;
                workingMemHandles.reserve(params.m_SimultaneousIterations);
                for (unsigned int i = 0; i < params.m_SimultaneousIterations; ++i)
                {
                    workingMemHandles.push_back(model.CreateWorkingMemHandle());
                }

                // Run each inference in its own thread
                // start a timer
                const auto start_time = armnn::GetTimeNow();
                for (unsigned int i = 0; i < params.m_SimultaneousIterations; ++i)
                {
                    armnn::experimental::IWorkingMemHandle& workingMemHandleRef = *workingMemHandles[i].get();
                    inferenceResults.push_back(std::async(
                        std::launch::async, [&model, &workingMemHandleRef, &inputs, &outputs, i]() {
                            return model.RunAsync(workingMemHandleRef, inputs[i], outputs[i]);
                        }
                        ));
                }

                // Check the results
                for (unsigned int j = 0; j < inferenceResults.size(); ++j)
                {
                    // Get the results
                    auto inferenceResult = inferenceResults[j].get();
                    auto inferenceDuration = std::get<1>(inferenceResult);
                    auto inferenceID = std::get<0>(inferenceResult);

                    if (params.m_GenerateTensorData)
                    {
                        ARMNN_LOG(warning) << "The input data was generated, note that the output will not be useful";
                    }

                    // Print output tensors
                    const auto& infosOut = model.GetOutputBindingInfos();
                    for (size_t i = 0; i < numOutputs; i++)
                    {
                        const armnn::TensorInfo& infoOut = infosOut[i].second;
                        auto outputTensorFile = params.m_OutputTensorFiles.empty()
                                                ? ""
                                                : params.m_OutputTensorFiles[(j * numOutputs) + i];

                        TensorPrinter printer(inferenceModelParams.m_OutputBindings[i],
                                              infoOut,
                                              outputTensorFile,
                                              params.m_DequantizeOutput);
                        mapbox::util::apply_visitor(printer, outputs[j][i]);
                    }

                    ARMNN_LOG(info) << "\nInference time: " << std::setprecision(2)
                                    << std::fixed << inferenceDuration.count() << " ms\n";

                    // If thresholdTime == 0.0 (default), then it hasn't been supplied at command line
                    if (params.m_ThresholdTime != 0.0)
                    {
                        ARMNN_LOG(info) << "Threshold time: " << std::setprecision(2)
                                        << std::fixed << params.m_ThresholdTime << " ms";
                        auto thresholdMinusInference = params.m_ThresholdTime - inferenceDuration.count();
                        ARMNN_LOG(info) << "Threshold time - Inference time: " << std::setprecision(2)
                                        << std::fixed << thresholdMinusInference << " ms" << "\n";

                        if (thresholdMinusInference < 0)
                        {
                            ARMNN_LOG(fatal) << "Elapsed inference time is greater than provided threshold time. \n";
                        }
                    }
                    ARMNN_LOG(info) << "Asynchronous Execution is finished for Inference ID: " << inferenceID << " \n";

                }
                // finish timer
                const auto duration = armnn::GetTimeDuration(start_time);
                ARMNN_LOG(info) << "\nOverall Inference time: " << std::setprecision(2)
                                << std::fixed << duration.count() << " ms\n";
            }
            catch (const armnn::Exception& e)
            {
                ARMNN_LOG(fatal) << "Armnn Error: " << e.what();
                return EXIT_FAILURE;
            }
        }
    }
    catch (const armnn::Exception& e)
    {
        ARMNN_LOG(fatal) << "Armnn Error: " << e.what();
        return EXIT_FAILURE;
    }

    return EXIT_SUCCESS;
}


// MAIN
int main(int argc, const char* argv[])
{
    // Configures logging for both the ARMNN library and this test program.
    #ifdef NDEBUG
    armnn::LogSeverity level = armnn::LogSeverity::Info;
    #else
    armnn::LogSeverity level = armnn::LogSeverity::Debug;
    #endif
    armnn::ConfigureLogging(true, true, level);


    // Get ExecuteNetwork parameters and runtime options from command line
    ProgramOptions ProgramOptions(argc, argv);

    // Create runtime
    std::shared_ptr<armnn::IRuntime> runtime(armnn::IRuntime::Create(ProgramOptions.m_RuntimeOptions));

    std::string modelFormat = ProgramOptions.m_ExNetParams.m_ModelFormat;

    // Forward to implementation based on the parser type
    if (modelFormat.find("armnn") != std::string::npos)
    {
    #if defined(ARMNN_SERIALIZER)
        return MainImpl<armnnDeserializer::IDeserializer, float>(ProgramOptions.m_ExNetParams, runtime);
    #else
        ARMNN_LOG(fatal) << "Not built with serialization support.";
        return EXIT_FAILURE;
    #endif
    }
    else if (modelFormat.find("onnx") != std::string::npos)
    {
    #if defined(ARMNN_ONNX_PARSER)
        return MainImpl<armnnOnnxParser::IOnnxParser, float>(ProgramOptions.m_ExNetParams, runtime);
    #else
        ARMNN_LOG(fatal) << "Not built with Onnx parser support.";
        return EXIT_FAILURE;
    #endif
    }
    else if(modelFormat.find("tflite") != std::string::npos)
    {
        if (ProgramOptions.m_ExNetParams.m_TfLiteExecutor == ExecuteNetworkParams::TfLiteExecutor::ArmNNTfLiteParser)
        {
            #if defined(ARMNN_TF_LITE_PARSER)
                        return MainImpl<armnnTfLiteParser::ITfLiteParser, float>(ProgramOptions.m_ExNetParams, runtime);
            #else
                        ARMNN_LOG(fatal) << "Not built with Tensorflow-Lite parser support.";
                        return EXIT_FAILURE;
            #endif
        }
        else if (ProgramOptions.m_ExNetParams.m_TfLiteExecutor ==
                    ExecuteNetworkParams::TfLiteExecutor::ArmNNTfLiteDelegate ||
                ProgramOptions.m_ExNetParams.m_TfLiteExecutor ==
                    ExecuteNetworkParams::TfLiteExecutor::TfliteInterpreter)
        {
        #if defined(ARMNN_TF_LITE_DELEGATE)
            return TfLiteDelegateMainImpl(ProgramOptions.m_ExNetParams, runtime);
        #else
            ARMNN_LOG(fatal) << "Not built with Arm NN Tensorflow-Lite delegate support.";
            return EXIT_FAILURE;
        #endif
        }
    }
    else
    {
        ARMNN_LOG(fatal) << "Unknown model format: '" << modelFormat
                         << "'. Please include 'tflite' or 'onnx'";
        return EXIT_FAILURE;
    }
}