//
// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#pragma once

#include <armnn_delegate.hpp>

#include <flatbuffers/flatbuffers.h>
#include <tensorflow/lite/interpreter.h>
#include <tensorflow/lite/kernels/register.h>
#include <tensorflow/lite/model.h>
#include <tensorflow/lite/schema/schema_generated.h>
#include <tensorflow/lite/version.h>

#include <doctest/doctest.h>

namespace
{

std::vector<char> CreateActivationTfLiteModel(tflite::BuiltinOperator activationOperatorCode,
                                              tflite::TensorType tensorType,
                                              const std::vector <int32_t>& tensorShape)
{
    using namespace tflite;
    flatbuffers::FlatBufferBuilder flatBufferBuilder;

    std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers;
    buffers[0] = CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({}));

    std::array<flatbuffers::Offset<Tensor>, 2> tensors;
    tensors[0] = CreateTensor(flatBufferBuilder,
                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(), tensorShape.size()),
                              tensorType);
    tensors[1] = CreateTensor(flatBufferBuilder,
                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(), tensorShape.size()),
                              tensorType);

    // create operator
    const std::vector<int> operatorInputs{0};
    const std::vector<int> operatorOutputs{1};
    flatbuffers::Offset <Operator> unaryOperator =
        CreateOperator(flatBufferBuilder,
                       0,
                       flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
                       flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()));

    const std::vector<int> subgraphInputs{0};
    const std::vector<int> subgraphOutputs{1};
    flatbuffers::Offset <SubGraph> subgraph =
        CreateSubGraph(flatBufferBuilder,
                       flatBufferBuilder.CreateVector(tensors.data(), tensors.size()),
                       flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()),
                       flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(), subgraphOutputs.size()),
                       flatBufferBuilder.CreateVector(&unaryOperator, 1));

    flatbuffers::Offset <flatbuffers::String> modelDescription =
        flatBufferBuilder.CreateString("ArmnnDelegate: Activation Operator Model");
    flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, activationOperatorCode);

    flatbuffers::Offset <Model> flatbufferModel =
        CreateModel(flatBufferBuilder,
                    TFLITE_SCHEMA_VERSION,
                    flatBufferBuilder.CreateVector(&operatorCode, 1),
                    flatBufferBuilder.CreateVector(&subgraph, 1),
                    modelDescription,
                    flatBufferBuilder.CreateVector(buffers.data(), buffers.size()));

    flatBufferBuilder.Finish(flatbufferModel);

    return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
                             flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
}

void ActivationTest(tflite::BuiltinOperator activationOperatorCode,
                    std::vector<armnn::BackendId>& backends,
                    std::vector<float>& inputValues,
                    std::vector<float>& expectedOutputValues)
{
    using namespace tflite;
    const std::vector<int32_t> inputShape  { { 4, 1, 4} };
    std::vector<char> modelBuffer = CreateActivationTfLiteModel(activationOperatorCode,
                                                                      ::tflite::TensorType_FLOAT32,
                                                                      inputShape);

    const Model* tfLiteModel = GetModel(modelBuffer.data());
    // Create TfLite Interpreters
    std::unique_ptr<Interpreter> armnnDelegateInterpreter;
    CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
              (&armnnDelegateInterpreter) == kTfLiteOk);
    CHECK(armnnDelegateInterpreter != nullptr);
    CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk);

    std::unique_ptr<Interpreter> tfLiteInterpreter;
    CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
              (&tfLiteInterpreter) == kTfLiteOk);
    CHECK(tfLiteInterpreter != nullptr);
    CHECK(tfLiteInterpreter->AllocateTensors() == kTfLiteOk);

    // Create the ArmNN Delegate
    armnnDelegate::DelegateOptions delegateOptions(backends);
    std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)>
                        theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions),
                                         armnnDelegate::TfLiteArmnnDelegateDelete);
    CHECK(theArmnnDelegate != nullptr);
    // Modify armnnDelegateInterpreter to use armnnDelegate
    CHECK(armnnDelegateInterpreter->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk);

    // Set input data
    auto tfLiteDelegateInputId = tfLiteInterpreter->inputs()[0];
    auto tfLiteDelageInputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateInputId);
    for (unsigned int i = 0; i < inputValues.size(); ++i)
    {
        tfLiteDelageInputData[i] = inputValues[i];
    }

    auto armnnDelegateInputId = armnnDelegateInterpreter->inputs()[0];
    auto armnnDelegateInputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateInputId);
    for (unsigned int i = 0; i < inputValues.size(); ++i)
    {
        armnnDelegateInputData[i] = inputValues[i];
    }
    // Run EnqueWorkload
    CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
    CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);

    // Compare output data
    auto tfLiteDelegateOutputId = tfLiteInterpreter->outputs()[0];
    auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateOutputId);
    auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0];
    auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateOutputId);
    for (size_t i = 0; i < inputValues.size(); i++)
    {
        CHECK(expectedOutputValues[i] == doctest::Approx(armnnDelegateOutputData[i]));
        CHECK(tfLiteDelageOutputData[i] == doctest::Approx(armnnDelegateOutputData[i]));
    }
}

} // anonymous namespace