// SPDX-FileCopyrightText: 2024 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only

#include "ModuleSummation.h"
#include "JFJochException.h"

ModuleSummation::ModuleSummation(bool in_pixel_signed,uint32_t in_fpga_depth_bytes) {
    pixel_signed = in_pixel_signed;
    first = true;
    empty = false;
    output_depth_bytes = 4;
    fpga_depth_bytes = in_fpga_depth_bytes;

    if ((fpga_depth_bytes != 4)
        && (fpga_depth_bytes != 2)
        && (fpga_depth_bytes != 1))
        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid, "FPGA byte depth can be only 1, 2 or 4");

    output = std::make_unique<DeviceOutput>();
    memset(output.get(), 0, sizeof(DeviceOutput));
}

ModuleSummation::ModuleSummation(const DiffractionExperiment &experiment) :
ModuleSummation(experiment.IsPixelSigned(), experiment.GetByteDepthFPGA()) {
    if (!experiment.IsCPUSummation())
        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
            "CPU summation not turned on");
    if (experiment.GetByteDepthImage() != 4)
        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
            "CPU summation works only for 32-bit output");
}

const DeviceOutput& ModuleSummation::GetOutput() const {
    return *output;
}

DeviceOutput &ModuleSummation::GetOutput() {
    return *output;
}

template <class Tdst, class Tsrc>
void Add(DeviceOutput& output, const DeviceOutput& input) {
    auto in_pixel = reinterpret_cast<const Tsrc *>(input.pixels);
    auto out_pixel = reinterpret_cast<Tdst *>(output.pixels);

    Tdst err_val_dst = std::is_signed_v<Tdst> ? std::numeric_limits<Tdst>::min() : std::numeric_limits<Tdst>::max();
    Tsrc err_val_src = std::is_signed_v<Tsrc> ? std::numeric_limits<Tsrc>::min() : std::numeric_limits<Tsrc>::max();

    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        if ((in_pixel[i] == err_val_src) || (out_pixel[i] == err_val_dst))
            out_pixel[i] = err_val_dst;
        else if ((in_pixel[i] == std::numeric_limits<Tsrc>::max())
            || (out_pixel[i] == std::numeric_limits<Tdst>::max())
            || (static_cast<int64_t>(out_pixel[i]) + static_cast<int64_t>(in_pixel[i]) >= std::numeric_limits<Tdst>::max()))
            out_pixel[i] = std::numeric_limits<Tdst>::max();
        else
            out_pixel[i] = out_pixel[i] + in_pixel[i];
    }

}

void ModuleSummation::AddFPGAOutput(const DeviceOutput &input) {
    if (first) {
        output->module_statistics = input.module_statistics;
        first = false;
    } else {
        output->module_statistics.err_pixels += input.module_statistics.err_pixels;
        output->module_statistics.saturated_pixels += input.module_statistics.saturated_pixels;
        output->module_statistics.packet_count += input.module_statistics.packet_count;
    }

    // The only function not handled is spot finding

    if (fpga_depth_bytes == 2) {
        if (pixel_signed)
            Add<int32_t, int16_t>(*output, input);
        else
            Add<uint32_t, uint16_t>(*output, input);
    } else if (fpga_depth_bytes == 4) {
        if (pixel_signed)
            Add<int32_t, int32_t>(*output, input);
        else
            Add<uint32_t, uint32_t>(*output, input);
    } else if (fpga_depth_bytes == 1) {
        if (pixel_signed)
            Add<int32_t, int8_t>(*output, input);
        else
            Add<uint32_t, uint8_t>(*output, input);

    }
    for (int i = 0; i < FPGA_INTEGRATION_BIN_COUNT; i++) {
        output->integration_result[i].count += input.integration_result[i].count;
        output->integration_result[i].sum += input.integration_result[i].sum;
    }

    for (int i = 0; i < ADU_HISTO_BIN_COUNT; i++)
        output->adu_histogram[i] += input.adu_histogram[i];

    for (int i = 0; i < FPGA_ROI_COUNT; i++) {
        output->roi_counts[i].sum += input.roi_counts[i].sum;
        output->roi_counts[i].sum2 += input.roi_counts[i].sum2;
        output->roi_counts[i].good_pixels += input.roi_counts[i].good_pixels;
        output->roi_counts[i].sum_x_weighted += input.roi_counts[i].sum_x_weighted;
        output->roi_counts[i].sum_y_weighted += input.roi_counts[i].sum_y_weighted;
        if (output->roi_counts[i].max_value < input.roi_counts[i].max_value)
            output->roi_counts[i].max_value = input.roi_counts[i].max_value;
    }
}

void ModuleSummation::AddEmptyOutput() {
    empty = true;
}

bool ModuleSummation::IsEmpty() const {
    return empty;
}