//
// Copyright © 2021, 2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#pragma once

#include "TestUtils.hpp"

#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 <schema_generated.h>
#include <tensorflow/lite/version.h>

#include <doctest/doctest.h>

namespace
{
std::vector<char> CreateRoundTfLiteModel(tflite::BuiltinOperator roundOperatorCode,
                                         tflite::TensorType tensorType,
                                         const std::vector <int32_t>& tensorShape,
                                         float quantScale = 1.0f,
                                         int quantOffset = 0)
{
    using namespace tflite;
    flatbuffers::FlatBufferBuilder flatBufferBuilder;

    std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
    buffers.push_back(CreateBuffer(flatBufferBuilder));
    buffers.push_back(CreateBuffer(flatBufferBuilder));
    buffers.push_back(CreateBuffer(flatBufferBuilder));

    auto quantizationParameters =
        CreateQuantizationParameters(flatBufferBuilder,
                                     0,
                                     0,
                                     flatBufferBuilder.CreateVector<float>({quantScale}),
                                     flatBufferBuilder.CreateVector<int64_t>({quantOffset}));

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

    const std::vector<int32_t> operatorInputs({0});
    const std::vector<int32_t> operatorOutputs({1});

    flatbuffers::Offset<Operator> roundOperator;
    flatbuffers::Offset<flatbuffers::String> modelDescription;
    flatbuffers::Offset<OperatorCode> operatorCode;

    switch (roundOperatorCode)
    {
        case tflite::BuiltinOperator_FLOOR:
        default:
            roundOperator =
                CreateOperator(flatBufferBuilder,
                               0,
                               flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
                               flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()));
                modelDescription = flatBufferBuilder.CreateString("ArmnnDelegate: Floor Operator Model");
                operatorCode = CreateOperatorCode(flatBufferBuilder, tflite::BuiltinOperator_FLOOR);
            break;
    }
    const std::vector<int32_t> subgraphInputs({0});
    const std::vector<int32_t> 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(&roundOperator, 1));

    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());
}

template<typename T>
void RoundTest(tflite::BuiltinOperator roundOperatorCode,
               tflite::TensorType tensorType,
               std::vector<armnn::BackendId>& backends,
               std::vector<int32_t>& shape,
               std::vector<T>& inputValues,
               std::vector<T>& expectedOutputValues,
               float quantScale = 1.0f,
               int quantOffset = 0)
{
    using namespace tflite;
    std::vector<char> modelBuffer = CreateRoundTfLiteModel(roundOperatorCode,
                                                           tensorType,
                                                           shape,
                                                           quantScale,
                                                           quantOffset);

    const Model* tfLiteModel = GetModel(modelBuffer.data());

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

    std::unique_ptr<Interpreter> tfLiteDelegate;
    CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
              (&tfLiteDelegate) == kTfLiteOk);
    CHECK(tfLiteDelegate != nullptr);
    CHECK(tfLiteDelegate->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(armnnDelegate->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk);

    // Set input data
    armnnDelegate::FillInput<T>(tfLiteDelegate, 0, inputValues);
    armnnDelegate::FillInput<T>(armnnDelegate, 0, inputValues);

    // Run EnqueWorkload
    CHECK(tfLiteDelegate->Invoke() == kTfLiteOk);
    CHECK(armnnDelegate->Invoke() == kTfLiteOk);

    // Compare output data
    armnnDelegate::CompareOutputData<T>(tfLiteDelegate,
                                        armnnDelegate,
                                        shape,
                                        expectedOutputValues,
                                        0);

    tfLiteDelegate.reset(nullptr);
    armnnDelegate.reset(nullptr);
}

} // anonymous namespace