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

#include "armnn/Exceptions.hpp"
#include "armnn/LayerSupport.hpp"

#include "backends/CpuTensorHandle.hpp"
#include "backends/WorkloadFactory.hpp"

LayerTestResult<float,4> SimpleNormalizationTestImpl(armnn::IWorkloadFactory& workloadFactory,
                                                     armnn::NormalizationAlgorithmChannel normChannel,
                                                     armnn::NormalizationAlgorithmMethod normMethod)
{
    const unsigned int inputHeight = 2;
    const unsigned int inputWidth = 2;
    const unsigned int inputChannels = 1;
    const unsigned int inputNum = 2;

    unsigned int outputHeight = inputHeight;
    unsigned int outputWidth = inputWidth;
    unsigned int outputChannels = inputChannels;
    unsigned int outputNum = inputNum;

    unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth };
    unsigned int outputShape[] = { outputNum, outputChannels, outputHeight, outputWidth };

    auto inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::DataType::Float32);
    auto outputTensorInfo = armnn::TensorInfo(4, outputShape, armnn::DataType::Float32);

    LayerTestResult<float,4> ret(outputTensorInfo);

    auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({
        // Batch #0
        1.0f, 2.0f,
        3.0f, 4.0f,
        // Batch #1
        5.0f, 6.0f,
        7.0f, 8.0f
    }));

    float alpha = 1.f;
    float beta = 1.f;
    float kappa = 1.f;
    uint32_t normSize = 3;

    std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
    std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);

    armnn::NormalizationQueueDescriptor data;
    armnn::WorkloadInfo info;
    AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
    data.m_Parameters.m_NormChannelType = normChannel;
    data.m_Parameters.m_NormMethodType = normMethod;
    data.m_Parameters.m_NormSize = normSize;
    data.m_Parameters.m_Alpha = alpha;
    data.m_Parameters.m_Beta = beta;
    data.m_Parameters.m_K = kappa;

    armnn::PassthroughCpuTensorHandle refHandle(outputTensorInfo, &ret.outputExpected[0][0][0][0]);
    armnn::NormalizationQueueDescriptor refData = data;
    armnn::WorkloadInfo refInfo = info;
    SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, &refHandle);

    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateNormalization(data, info);

    inputHandle->Allocate();
    outputHandle->Allocate();

    CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]);

    workloadFactory.Finalize();
    workload->Execute();

    CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get());

    switch (normMethod)
    {
        case armnn::NormalizationAlgorithmMethod::LocalBrightness:
        {
            switch (normChannel)
            {
                case armnn::NormalizationAlgorithmChannel::Within:
                {
                    // When normalising within channels, the 3x3 kernel covers the entire 2x2 input at every index.
                    // Therefore, all output values should equal the inputs, but divided by:
                    // pow((kappa + (accumulatedScale * alpha)), beta)
                    // ...where accumulatedScale is the sum of every element squared.
                    float divisor[inputNum];
                    for(int i = 0; i < boost::numeric_cast<int>(inputNum); i++)
                    {
                        float accumulatedScale = input[i][0][0][0]*input[i][0][0][0] +
                                                 input[i][0][0][1]*input[i][0][0][1] +
                                                 input[i][0][1][0]*input[i][0][1][0] +
                                                 input[i][0][1][1]*input[i][0][1][1];
                        divisor[i] = powf((kappa + accumulatedScale * alpha), beta);
                    }
                    ret.outputExpected = MakeTensor<float, 4>(outputTensorInfo,
                                                              std::vector<float>({input[0][0][0][0]/divisor[0],
                                                                                  input[0][0][0][1]/divisor[0],
                                                                                  input[0][0][1][0]/divisor[0],
                                                                                  input[0][0][1][1]/divisor[0],
                                                                                  input[1][0][0][0]/divisor[1],
                                                                                  input[1][0][0][1]/divisor[1],
                                                                                  input[1][0][1][0]/divisor[1],
                                                                                  input[1][0][1][1]/divisor[1]}));
                    break;
                }
                case armnn::NormalizationAlgorithmChannel::Across:
                {
                    // When normalising across channels, all output values should equal the inputs, but multiplied by:
                    // pow((kappa + (accumulatedScale * alpha)), -beta)
                    // ...where accumulatedScale is the sum of the inputs for adjacent channels for this element squared
                    // ...where adjacent channels means within half the normSize for the channel
                    // The test data has only one channel, so this is simplified below.
                    std::vector<float> outputVector;
                    for (int n = 0; n < boost::numeric_cast<int>(inputNum); ++n)
                    {
                        for (int h = 0; h < boost::numeric_cast<int>(inputHeight); ++h)
                        {
                            for (int w = 0; w < boost::numeric_cast<int>(inputWidth); ++w)
                            {
                                float accumulatedScale = input[n][0][h][w]*input[n][0][h][w];
                                float scale = powf((kappa + accumulatedScale * alpha), -beta);
                                outputVector.push_back(input[n][0][h][w] * scale);
                            }
                        }
                    }
                    ret.outputExpected = MakeTensor<float, 4>(outputTensorInfo, outputVector);
                    break;
                }
                default:
                {
                    throw armnn::UnimplementedException("Unsupported normalisation channel type, "
                                                        "only Across and Within are supported");
                }
            }
            break;
        }
        case armnn::NormalizationAlgorithmMethod::LocalContrast: // NOTE: intentional fallthrough.
        default:
        {
            throw armnn::UnimplementedException("Unsupported normalisation method type, "
                                                "only LocalBrightness is supported");
        }
    }

    return ret;
}

LayerTestResult<float,4> CompareNormalizationTestImpl(armnn::IWorkloadFactory& workloadFactory,
                                                      armnn::IWorkloadFactory& refWorkloadFactory,
                                                      armnn::NormalizationAlgorithmChannel normChannel,
                                                      armnn::NormalizationAlgorithmMethod normMethod)
{
    constexpr unsigned int inputNum = 5;
    constexpr unsigned int inputChannels = 3;
    constexpr unsigned int inputHeight = 32;
    constexpr unsigned int inputWidth = 24;

    constexpr unsigned int outputNum = inputNum;
    constexpr unsigned int outputChannels = inputChannels;
    constexpr unsigned int outputHeight = inputHeight;
    constexpr unsigned int outputWidth = inputWidth;

    armnn::TensorInfo inputTensorInfo;
    armnn::TensorInfo outputTensorInfo;

    unsigned int inputShape[] = {inputNum, inputChannels, inputHeight, inputWidth};
    unsigned int outputShape[] = {outputNum, outputChannels, outputHeight, outputWidth};

    inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::DataType::Float32);
    outputTensorInfo = armnn::TensorInfo(4, outputShape, armnn::DataType::Float32);

    LayerTestResult<float,4> ret(outputTensorInfo);

    auto input = MakeRandomTensor<float, 4>(inputTensorInfo, 111234);

    constexpr float alpha = 1.f;
    constexpr float beta = 1.f;
    constexpr float kappa = 1.f;
    constexpr uint32_t normSize = 5;

    std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
    std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);

    armnn::NormalizationQueueDescriptor data;
    armnn::WorkloadInfo info;
    AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
    data.m_Parameters.m_NormChannelType = normChannel;
    data.m_Parameters.m_NormMethodType  = normMethod;
    data.m_Parameters.m_NormSize        = normSize;
    data.m_Parameters.m_Alpha           = alpha;
    data.m_Parameters.m_Beta            = beta;
    data.m_Parameters.m_K               = kappa;

    std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo);
    std::unique_ptr<armnn::ITensorHandle> inputHandleRef = refWorkloadFactory.CreateTensorHandle(inputTensorInfo);

    armnn::NormalizationQueueDescriptor refData = data;
    armnn::WorkloadInfo refInfo = info;
    SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get());
    SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get());

    // Don't execute if Normalization is not supported for the method and channel types, as an exception will be raised.
    armnn::Compute compute = workloadFactory.GetCompute();
    const size_t reasonIfUnsupportedMaxLen = 255;
    char reasonIfUnsupported[reasonIfUnsupportedMaxLen+1];
    ret.supported = armnn::IsNormalizationSupported(compute, inputTensorInfo, outputTensorInfo, data.m_Parameters,
                                                    reasonIfUnsupported, reasonIfUnsupportedMaxLen);
    if (!ret.supported)
    {
        return ret;
    }

    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateNormalization(data, info);
    std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateNormalization(refData, refInfo);

    outputHandleRef->Allocate();
    inputHandleRef->Allocate();

    inputHandle->Allocate();
    outputHandle->Allocate();

    CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]);
    CopyDataToITensorHandle(inputHandleRef.get(), &input[0][0][0][0]);

    workloadFactory.Finalize();
    workload->Execute();
    refWorkloadFactory.Finalize();
    workloadRef->Execute();

    CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get());
    CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get());

    return ret;
}