// Copyright (2019-2023) Paul Scherrer Institute

#include <catch2/catch.hpp>
#include <random>

#include <bitshuffle/bitshuffle.h>
#include "../jungfrau/JFPedestalCalc.h"
#include "../receiver/HLSSimulatedDevice.h"
#include "FPGAUnitTest.h"

using namespace std::literals::chrono_literals;

TEST_CASE("HLS_C_Simulation_internal_packet_generator", "[FPGA][Full]") {
    const uint16_t nmodules = 4;

    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(4).PedestalG0Frames(0);

    HLSSimulatedDevice test(0, 64);
    test.SetInternalGeneratorFrame();
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    JFJochProtoBuf::AcquisitionDeviceStatistics device_statistics;
    REQUIRE_NOTHROW(test.SaveStatistics(x, device_statistics));
    REQUIRE(device_statistics.bytes_received() == 128 * nmodules * 4 * JUNGFRAU_PACKET_SIZE_BYTES);

    for (int image = 0; image < 4; image++) {
        for (int m = 0; m < nmodules; m++) {
            auto imageBuf = (uint16_t *) test.GetDeviceOutput(image, m)->pixels;
            for (int i = 0; i < RAW_MODULE_SIZE; i++)
                REQUIRE(imageBuf[i] == i % 65536);
        }
    }
    REQUIRE(test.GetCompletedDescriptors() == 2 * (4 + DELAY_FRAMES_STOP_AND_QUIT - 1) * nmodules);
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_custom_frame", "[FPGA][Full]") {
    const uint16_t nmodules = 32;
    const size_t nframes = 2;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    std::vector<uint16_t> test_frame(nmodules*RAW_MODULE_SIZE);

    std::mt19937 g1(1387);
    std::uniform_int_distribution<uint16_t> dist(0, 65535);

    for (auto &i: test_frame)
        i = dist(g1);

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(nframes).PedestalG0Frames(0);

    HLSSimulatedDevice test(0, 64);
    test.SetInternalGeneratorFrame(test_frame);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    REQUIRE(test.GetBytesReceived() == 128 * nmodules * nframes * JUNGFRAU_PACKET_SIZE_BYTES);

    for (int image = 0; image < nframes; image++) {
        for (int m = 0; m < nmodules; m++) {
            REQUIRE(test.Counters().GetCompletion(image, m).timestamp == INT_PKT_GEN_EXPTTIME * image);
            REQUIRE(test.Counters().GetCompletion(image, m).bunchid == INT_PKT_GEN_BUNCHID + image);
            REQUIRE(test.Counters().GetCompletion(image, m).exptime == INT_PKT_GEN_EXPTTIME);

            auto imageBuf = (uint16_t *) test.GetDeviceOutput(image, m)->pixels;
            for (int i = 0; i < RAW_MODULE_SIZE; i++) {
                REQUIRE(imageBuf[i] == test_frame[m * RAW_MODULE_SIZE + i]);
            }
        }
    }
}

TEST_CASE("HLS_C_Simulation_check_raw", "[FPGA][Full]") {
    std::vector<uint16_t> raw_frames(RAW_MODULE_SIZE*20);

    for (int i = 0; i < 20; i++) {
        LoadBinaryFile("../../tests/test_data/mod5_raw" + std::to_string(i)+".bin", raw_frames.data() + i * RAW_MODULE_SIZE, RAW_MODULE_SIZE);
    }

    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));
    uint16_t data[4096];

    x.Mode(DetectorMode::Raw);
    x.PedestalG0Frames(0).ImagesPerTrigger(5).NumTriggers(1);

    HLSSimulatedDevice test(0, 64);
    test.CreatePackets(x, 1, 5, 0, raw_frames.data(), true);
    test.CreateFinalPacket(x);
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.Counters().GetSlowestFrameNumber() == 0);

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == 5 * 128 * JUNGFRAU_PACKET_SIZE_BYTES);

    JFJochProtoBuf::AcquisitionDeviceStatistics device_statistics;
    REQUIRE_NOTHROW(test.SaveStatistics(x, device_statistics));
    REQUIRE(device_statistics.bytes_received() == 5 * 128 * JUNGFRAU_PACKET_SIZE_BYTES);
    REQUIRE(device_statistics.efficiency() == Approx(0.25));

    uint64_t diffs = 0;
    for (int image = 0; image < 5; image++) {
        for (int j = 0; j < RAW_MODULE_SIZE; j++) {
            if (raw_frames[image*RAW_MODULE_SIZE+j] != ((uint16_t *) test.GetDeviceOutput(image, 0)->pixels)[j]) diffs++;
        }
    }

    REQUIRE(diffs == 0);
}

TEST_CASE("HLS_C_Simulation_check_cancel", "[FPGA][Full]") {

    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[4096];

    x.Mode(DetectorMode::Raw);
    x.PedestalG0Frames(0).ImagesPerTrigger(5).NumTriggers(1);

    HLSSimulatedDevice test(0, 64);

    REQUIRE_NOTHROW(test.StartAction(x));
    test.Cancel();

    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.Counters().GetSlowestFrameNumber() == 0);

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == 0);
}


TEST_CASE("HLS_C_Simulation_check_cancel_conversion", "[FPGA][Full]") {
    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[4096];

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).ImagesPerTrigger(5).NumTriggers(1);

    HLSSimulatedDevice test(0, 64);

    REQUIRE_NOTHROW(test.StartAction(x));
    test.Cancel();

    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.Counters().GetSlowestFrameNumber() == 0);

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == 0);
}

TEST_CASE("HLS_C_Simulation_check_delay", "[FPGA][Full]") {
    std::vector<uint16_t> raw_frames(RAW_MODULE_SIZE*20);

    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[4096];

    x.Mode(DetectorMode::Raw);
    x.PedestalG0Frames(0).ImagesPerTrigger(3).NumTriggers(1);

    HLSSimulatedDevice test(0, 64);

    test.CreatePacketJF(x, 1, 0, 0, data, false);
    test.CreatePacketJF(x, 1, 0, 1, data, false);
    test.CreatePacketJF(x, 1, 0, 2, data, false);
    test.CreatePacketJF(x, 1, 0, 3, data, false);

    test.CreatePacketJF(x, 2, 0, 0, data, false);
    test.CreatePacketJF(x, 2, 0, 1, data, false);
    test.CreatePacketJF(x, 2, 0, 2, data, false);
    test.CreatePacketJF(x, 2, 0, 3, data, false);

    test.CreatePacketJF(x, 3, 0, 0, data, false);
    test.CreatePacketJF(x, 3, 0, 1, data, false);
    test.CreatePacketJF(x, 3, 0, 2, data, false);
    test.CreatePacketJF(x, 3, 0, 3, data, false);

    test.CreateFinalPacket(x);

    REQUIRE_NOTHROW(test.StartAction(x));

    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.Counters().CalculateDelay(0) == 2);
    REQUIRE(test.Counters().CalculateDelay(0, 0) == 2);
    REQUIRE(test.Counters().CalculateDelay(1) == 1);
    REQUIRE(test.Counters().CalculateDelay(1, 0) == 1);
    REQUIRE(test.Counters().CalculateDelay(2) == 0);
    REQUIRE(test.Counters().CalculateDelay(2, 0) == 0);
}

TEST_CASE("HLS_C_Simulation_check_lost_frame_raw", "[FPGA][Full]") {
    std::vector<uint16_t> raw_frames(RAW_MODULE_SIZE*20);

    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[4096];

    for (int i = 0; i < 4096; i++) data[i] = i;
    x.Mode(DetectorMode::Raw);
    x.PedestalG0Frames(0).ImagesPerTrigger(3).NumTriggers(1);

    HLSSimulatedDevice test(0, 64);

    test.CreatePacketJF(x, 1, 0, 0, data, false);
    test.CreateFinalPacket(x);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == JUNGFRAU_PACKET_SIZE_BYTES);

    REQUIRE(test.GetDeviceOutput(0,0)->pixels[0] == 0);
    REQUIRE(test.GetDeviceOutput(0,0)->pixels[1] == 1);
    REQUIRE(test.GetDeviceOutput(0,0)->pixels[4095] == 4095);
    REQUIRE(test.GetDeviceOutput(0,1)->pixels[0] == -1);
    REQUIRE(test.GetDeviceOutput(0,1)->pixels[RAW_MODULE_SIZE-1] == -1);
}

TEST_CASE("HLS_C_Simulation_check_lost_frame_conversion", "[FPGA][Full]") {
    std::vector<uint16_t> raw_frames(RAW_MODULE_SIZE*20);

    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[4096];

    for (int i = 0; i < 4096; i++) data[i] = i;
    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).ImagesPerTrigger(3).NumTriggers(1);

    HLSSimulatedDevice test(0, 64);

    test.CreatePacketJF(x, 1, 0, 0, data, false);
    test.CreateFinalPacket(x);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == JUNGFRAU_PACKET_SIZE_BYTES);

    REQUIRE(test.GetDeviceOutput(0,1)->pixels[0] == PIXEL_OUT_LOST);
    REQUIRE(test.GetDeviceOutput(0,1)->pixels[RAW_MODULE_SIZE-1] == PIXEL_OUT_LOST);
    REQUIRE(test.GetDeviceOutput(1,0)->pixels[0] == PIXEL_OUT_LOST);
    REQUIRE(test.GetDeviceOutput(2,1)->pixels[556] == PIXEL_OUT_LOST);
}

TEST_CASE("HLS_C_Simulation_check_single_packet", "[FPGA][Full]") {
    std::vector<uint16_t> raw_frames(RAW_MODULE_SIZE*20);

    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[4096];
    uint16_t wrong[4096];

    for (int i = 0; i < 4096; i++) {
        data[i] = (7 * i + 534 + 5 * i * i) % UINT16_MAX;
        wrong[i] = UINT16_MAX;
    }

    x.Mode(DetectorMode::Raw);
    x.PedestalG0Frames(0).ImagesPerTrigger(3).NumTriggers(1);
    HLSSimulatedDevice test(0, 64);

    test.CreatePacketJF(x, 1, 0, 0, data, false);
    test.CreatePacketJF(x, 1, 64, 0, data, false);
    test.CreatePacketJF(x, 1, 0, 2, data, false);

    test.CreatePacketJF(x, 1, 2, 3, data, false);
    test.CreatePacketJF(x, 1, 3, 3, data, false);

    test.CreatePacketJF(x, 1, 1, 3, data, false);
    test.CreatePacketJF(x, 1, 0, 3, data, false);
    test.CreatePacketJF(x, 1, 64, 3, data, false);

    test.CreatePacketJF(x, 1, 5, 0, data, false);
    test.CreatePacketJF(x, 1, 4, 0, data, false);

    test.CreatePacketJF(x, 1, 67, 1, data, false);
    test.CreatePacketJF(x, 1, 66, 1, data, false);
    test.CreatePacketJF(x, 1, 68, 1, data, false);

    test.CreatePacketJF(x, 3, 1, 0, data, false);
    test.CreatePacketJF(x, 2, 1, 0, data, false);

    test.CreatePacketJF(x, 4, 1, 0, data, false);

    test.CreateFinalPacket(x);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == 15 * JUNGFRAU_PACKET_SIZE_BYTES);

    REQUIRE(memcmp(test.GetDeviceOutput(0,0)->pixels, data, JUNGFRAU_PACKET_SIZE_BYTES) == 0);
    REQUIRE(memcmp(test.GetDeviceOutput(0,0)->pixels + 1 * JUNGFRAU_PACKET_SIZE_BYTES/2, wrong, JUNGFRAU_PACKET_SIZE_BYTES) == 0);
    REQUIRE(memcmp(test.GetDeviceOutput(0,0)->pixels + 2 * JUNGFRAU_PACKET_SIZE_BYTES/2, wrong,  JUNGFRAU_PACKET_SIZE_BYTES) == 0);
    REQUIRE(memcmp(test.GetDeviceOutput(0,0)->pixels + 3 * JUNGFRAU_PACKET_SIZE_BYTES/2, wrong, JUNGFRAU_PACKET_SIZE_BYTES) == 0);
    REQUIRE(memcmp(test.GetDeviceOutput(0,0)->pixels + 4 * JUNGFRAU_PACKET_SIZE_BYTES/2, data,  JUNGFRAU_PACKET_SIZE_BYTES) == 0);
    REQUIRE(memcmp(test.GetDeviceOutput(0,0)->pixels + 5 * JUNGFRAU_PACKET_SIZE_BYTES/2, data,  JUNGFRAU_PACKET_SIZE_BYTES) == 0);
    REQUIRE(memcmp(test.GetDeviceOutput(0,0)->pixels + 6 * JUNGFRAU_PACKET_SIZE_BYTES/2, wrong, JUNGFRAU_PACKET_SIZE_BYTES) == 0);
    REQUIRE(memcmp(test.GetDeviceOutput(0,0)->pixels + 64 * JUNGFRAU_PACKET_SIZE_BYTES/2, data,  JUNGFRAU_PACKET_SIZE_BYTES) == 0);

    REQUIRE(memcmp(test.GetDeviceOutput(0,1)->pixels + 67 * JUNGFRAU_PACKET_SIZE_BYTES/2, data,  JUNGFRAU_PACKET_SIZE_BYTES) == 0);
    REQUIRE(memcmp(test.GetDeviceOutput(0,1)->pixels + 65 * JUNGFRAU_PACKET_SIZE_BYTES/2, wrong,  JUNGFRAU_PACKET_SIZE_BYTES) == 0);

    JFJochProtoBuf::AcquisitionDeviceStatistics device_statistics;
    REQUIRE_NOTHROW(test.SaveStatistics(x, device_statistics));
    REQUIRE(device_statistics.good_packets() == 15);
    REQUIRE(device_statistics.bytes_received() == 15 * JUNGFRAU_PACKET_SIZE_BYTES);
}

TEST_CASE("HLS_C_Simulation_check_convert_full_range", "[FPGA][Full]") {
    std::vector<uint16_t> data(RAW_MODULE_SIZE);
    std::vector<double> gain(3 * RAW_MODULE_SIZE);

    JFModulePedestal pedestal_g0(0), pedestal_g1(14500), pedestal_g2(14500);

    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        data[i] = i % UINT16_MAX;
    }

    std::vector<double> energy_values = {6.0, 12.4, 17.7, 5, 4.5, 3.7};

    const uint16_t nmodules = 1;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Conversion);
    HLSSimulatedDevice test(0, 64);

    auto gain_from_file = GainCalibrationFromTestFile();


    for (const auto energy : energy_values) {
        x.PedestalG0Frames(0).NumTriggers(1).ImagesPerTrigger(1).PhotonEnergy_keV(energy);

        REQUIRE(x.GetPhotonEnergy_keV() == Approx(energy));

        JFCalibration c_in(x);
        c_in.Pedestal(0,0) = pedestal_g0;
        c_in.Pedestal(0,1) = pedestal_g1;
        c_in.Pedestal(0,2) = pedestal_g2;
        for (int i = 0; i < x.GetModulesNum(); i++)
            c_in.GainCalibration(i) = gain_from_file;

        test.InitializeCalibration(x, c_in);
        test.CreatePackets(x, 1, 1, 0, data.data(), false);
        test.CreateFinalPacket(x);

        REQUIRE_NOTHROW(test.StartAction(x));
        REQUIRE_NOTHROW(test.WaitForActionComplete());

        REQUIRE_NOTHROW(test.OutputStream().read());
        REQUIRE(test.OutputStream().size() == 0);
        REQUIRE(test.GetBytesReceived() == 128 * JUNGFRAU_PACKET_SIZE_BYTES);

        double mean_error = CheckConversion(x, c_in, data.data(), test.GetDeviceOutput(0,0)->pixels);

        REQUIRE(mean_error < 0.5);
    }
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_convert_full_range", "[FPGA][Full]") {
    double energy = 6.0;
    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules, 2, 8, 36, true)));

    std::vector<uint16_t> data(RAW_MODULE_SIZE);
    JFModulePedestal pedestal_g0, pedestal_g1, pedestal_g2;

    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        pedestal_g0.GetPedestal()[i] = 0 + (i / 65536) * 1000 + 100 * (i % 5);
        pedestal_g1.GetPedestal()[i] = 14500 - (i / 65536) * 1000 + 100 * (i % 3);
        pedestal_g2.GetPedestal()[i] = 14500 - + (i / 65536) * 1000;
    }

    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        data[i] = i % RAW_MODULE_SIZE;
    }

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(1).ImagesPerTrigger(1).UseInternalPacketGenerator(true).PhotonEnergy_keV(energy);
    REQUIRE(x.GetPhotonEnergy_keV() == Approx(energy));

    auto gain_from_file = GainCalibrationFromTestFile();

    JFCalibration c(x);
    for (int i = 0; i < x.GetModulesNum(); i++) {
        c.Pedestal(i, 0) = pedestal_g0;
        c.Pedestal(i, 1) = pedestal_g1;
        c.Pedestal(i, 2) = pedestal_g2;
        c.GainCalibration(i) = gain_from_file;
    }

    HLSSimulatedDevice test(0, 64);
    test.SetInternalGeneratorFrame();
    REQUIRE_NOTHROW(test.InitializeCalibration(x, c));
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == nmodules * 128 * JUNGFRAU_PACKET_SIZE_BYTES);

    double mean_error = CheckConversion(x, c, data.data(), test.GetDeviceOutput(0,0)->pixels);

    REQUIRE(mean_error < 0.5);
}


TEST_CASE("HLS_C_Simulation_internal_packet_generator_convert_full_range_adu_histo", "[FPGA][Full]") {
    double energy = 6.0;
    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules, 2, 8, 36, true)));

    std::vector<uint16_t> data(RAW_MODULE_SIZE);
    JFModulePedestal pedestal_g0, pedestal_g1, pedestal_g2;

    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        pedestal_g0.GetPedestal()[i] = 0 + (i / 65536) * 1000 + 100 * (i % 5);
        pedestal_g1.GetPedestal()[i] = 14500 - (i / 65536) * 1000 + 100 * (i % 3);
        pedestal_g2.GetPedestal()[i] = 14500 - + (i / 65536) * 1000;
    }

    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        data[i] = i % RAW_MODULE_SIZE;
    }

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(1).ImagesPerTrigger(1).UseInternalPacketGenerator(true).PhotonEnergy_keV(energy);
    REQUIRE(x.GetPhotonEnergy_keV() == Approx(energy));

    auto gain_from_file = GainCalibrationFromTestFile();

    JFCalibration c(x);
    for (int i = 0; i < x.GetModulesNum(); i++) {
        c.Pedestal(i, 0) = pedestal_g0;
        c.Pedestal(i, 1) = pedestal_g1;
        c.Pedestal(i, 2) = pedestal_g2;
        c.GainCalibration(i) = gain_from_file;
    }

    HLSSimulatedDevice test(0, 64);
    test.SetInternalGeneratorFrame();
    REQUIRE_NOTHROW(test.InitializeCalibration(x, c));
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == nmodules * 128 * JUNGFRAU_PACKET_SIZE_BYTES);

    uint32_t err = 0;
    const uint32_t *adu_histogram = test.GetDeviceOutput(0, 0)->adu_histogram;
    for (int i = 0; i < ADU_HISTO_BIN_COUNT; i++) {
        if(adu_histogram[i] != RAW_MODULE_SIZE / ADU_HISTO_BIN_COUNT)
            err++;
    }
    REQUIRE(err == 0);
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_apply_pixel_mask", "[FPGA][Full]") {
    double energy = 6.0;

    const uint16_t nmodules = 1;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    JFModulePedestal pedestal_g0, pedestal_g1, pedestal_g2;
    std::vector<double> gain(3 * RAW_MODULE_SIZE);

    std::vector<uint16_t> test_frame(RAW_MODULE_SIZE);

    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        test_frame[i] = 5000 | ((i > 128*1024) ? 16384:0) | ((i > 256*1024) ? 32768 : 0);
    }

    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        pedestal_g0.GetPedestal()[i] = 3000;
        pedestal_g1.GetPedestal()[i] = 14000;
        pedestal_g2.GetPedestal()[i] = 15000;
    }
    pedestal_g0.GetPedestalMask()[0] = 1;
    pedestal_g1.GetPedestalMask()[1] = 1;
    pedestal_g2.GetPedestalMask()[2] = 1;

    pedestal_g0.GetPedestalMask()[128*1024+1] = 1;
    pedestal_g1.GetPedestalMask()[128*1024+2] = 1;
    pedestal_g2.GetPedestalMask()[128*1024+3] = 1;

    pedestal_g0.GetPedestalMask()[256*1024+1] = 1;
    pedestal_g1.GetPedestalMask()[256*1024+2] = 1;
    pedestal_g2.GetPedestalMask()[256*1024+3] = 1;

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(1).ImagesPerTrigger(1).UseInternalPacketGenerator(true).PhotonEnergy_keV(energy)
            .ApplyPixelMaskInFPGA(true);
    REQUIRE(x.GetPhotonEnergy_keV() == Approx(energy));

    JFCalibration c(x);
    c.Pedestal(0,0) = pedestal_g0;
    c.Pedestal(0,1) = pedestal_g1;
    c.Pedestal(0,2) = pedestal_g2;
    c.Mask(3)          = UINT32_MAX;
    c.Mask(128*1024+3) = UINT32_MAX;
    c.Mask(256*1024+3) = UINT32_MAX;

    HLSSimulatedDevice test(0, 64);

    test.SetInternalGeneratorFrameForAllModules(test_frame);

    REQUIRE_NOTHROW(test.InitializeCalibration(x, c));
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == 128 * JUNGFRAU_PACKET_SIZE_BYTES);

    size_t err = 0;
    auto frame = test.GetDeviceOutput(0,0)->pixels;
    auto mask = c.CalculateMask(x);
    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        if (mask[i] && (frame[i] != INT16_MIN))
            err++;
        if (!mask[i] && (frame[i] == INT16_MIN))
            err++;
    }
    REQUIRE(err == 0);
}

TEST_CASE("HLS_C_Simulation_check_2_trigger_convert", "[FPGA][Full]") {
    std::vector<float> pedestal_g0(RAW_MODULE_SIZE), pedestal_g1(RAW_MODULE_SIZE), pedestal_g2(RAW_MODULE_SIZE);

    LoadBinaryFile("../../tests/test_data/mod5_pedeG0.bin", pedestal_g0.data(), RAW_MODULE_SIZE);
    LoadBinaryFile("../../tests/test_data/mod5_pedeG1.bin", pedestal_g1.data(), RAW_MODULE_SIZE);
    LoadBinaryFile("../../tests/test_data/mod5_pedeG2.bin", pedestal_g2.data(), RAW_MODULE_SIZE);

    std::vector<uint16_t> raw_frames(RAW_MODULE_SIZE*20);
    std::vector<float> conv_frames(RAW_MODULE_SIZE*20);

    for (int i = 0; i < 20; i++) {
        LoadBinaryFile("../../tests/test_data/mod5_raw" + std::to_string(i)+".bin", raw_frames.data() + i * RAW_MODULE_SIZE, RAW_MODULE_SIZE);
        LoadBinaryFile("../../tests/test_data/mod5_conv" + std::to_string(i)+".bin", conv_frames.data() + i * RAW_MODULE_SIZE, RAW_MODULE_SIZE);
    }

    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[4096];

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(2).ImagesPerTrigger(5);

    HLSSimulatedDevice test(0, 64);

    JFCalibration c(x);
    REQUIRE_NOTHROW(c.Pedestal(0, 0).LoadPedestal(pedestal_g0));
    REQUIRE_NOTHROW(c.Pedestal(0, 1).LoadPedestal(pedestal_g1));
    REQUIRE_NOTHROW(c.Pedestal(0, 2).LoadPedestal(pedestal_g2));
    REQUIRE_NOTHROW(c.GainCalibration(0) = GainCalibrationFromTestFile());

    REQUIRE_NOTHROW(test.InitializeCalibration(x, c));

    // send some frames without trigger (to be ignored)
    test.CreatePackets(x, 1, 10, 0, raw_frames.data(), true);
    test.CreateFinalPacket(x);

    REQUIRE_NOTHROW(test.StartAction(x));

    REQUIRE(!test.Counters().IsAcquisitionFinished());

    test.WaitForActionComplete();

    REQUIRE(test.Counters().IsAcquisitionFinished());

    // address properly aligned
    REQUIRE((uint64_t) test.GetDeviceOutput(0,0)->pixels % 128 == 0);

    REQUIRE(test.Counters().GetSlowestFrameNumber() == 0);
    REQUIRE(test.Counters().GetCurrFrameNumber(0) == 9);

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    JFJochProtoBuf::AcquisitionDeviceStatistics device_statistics;
    REQUIRE_NOTHROW(test.SaveStatistics(x, device_statistics));

    REQUIRE(device_statistics.efficiency() == Approx(0.25));
    REQUIRE(device_statistics.good_packets() == 128*10);
    REQUIRE(device_statistics.bytes_received() == 128 * 10 * JUNGFRAU_PACKET_SIZE_BYTES);

    double mean_error = 0.0;
    for (int image = 0; image < 10; image++) {
        auto img = test.GetDeviceOutput(image, 0)->pixels;
        for (int j = 0; j < RAW_MODULE_SIZE; j++) {
            if ((img[j] < 30000) && (img[j] > -30000)) {
                float diff = (conv_frames[image * RAW_MODULE_SIZE + j] - (float) img[j]);
                mean_error += diff * diff;
            }
        }
    }

    mean_error = sqrt(mean_error/ (10*RAW_MODULE_SIZE));
    std::cout << "Mean error " << mean_error << std::endl;
    REQUIRE(mean_error < 0.30);
}


TEST_CASE("HLS_C_Simulation_check_detect_last_frame", "[FPGA][Full]") {
    const uint16_t nmodules = 4;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[4096];

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(2).ImagesPerTrigger(5);

    HLSSimulatedDevice test(0, 64);

    test.CreatePacketJF(x, 15, 0, 0, data, true);
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE(!test.Counters().IsAcquisitionFinished());

    test.WaitForActionComplete();

    REQUIRE(test.Counters().IsAcquisitionFinished());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);
}

TEST_CASE("HLS_C_Simulation_check_wrong_packet_size", "[FPGA][Full]") {
    const uint16_t nmodules = 1;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    uint16_t data[8192];

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(1).ImagesPerTrigger(5);

    HLSSimulatedDevice test(0, 64);

    // send some frames with wrong size or tuser=1
    test.CreatePacketJF(x, 1, 0, 0, data, true, 0);
    test.CreatePacketJF(x, 1, 1, 0, data, true, -1);
    test.CreatePacketJF(x, 1, 2, 0, data, true, 2);
    test.CreatePacketJF(x, 1, 3, 0, data, true, 0);
    test.CreatePacketJF(x, 1, 4, 0, data, true, -5);
    test.CreatePacketJF(x, 1, 5, 0, data, true, 0, 1);
    test.CreatePacketJF(x, 1, 6, 0, data, true, 0);
    test.CreatePacketJF(x, 1, 7, 0, data, true, 0);
    test.CreatePacketJF(x, 1, 8, 0, data, true, 7);
    test.CreatePacketJF(x, 1, 9, 0, data, true, 100);
    test.CreatePacketJF(x, 1, 10, 0, data, true, 2);
    test.CreatePacketJF(x, 1, 11, 0, data, true, 0);
    test.CreatePacketJF(x, 1, 12, 0, data, true, -80);
    test.CreatePacketJF(x, 1, 13, 0, data, true, 100);
    test.CreatePacketJF(x, 1, 118, 0, data, true, 0);

    test.CreateFinalPacket(x);

    REQUIRE_NOTHROW(test.StartAction(x));

    REQUIRE(!test.Counters().IsAcquisitionFinished());

    test.WaitForActionComplete();

    REQUIRE(test.Counters().IsAcquisitionFinished());

    REQUIRE(test.GetBytesReceived() == 6 * JUNGFRAU_PACKET_SIZE_BYTES);
}

TEST_CASE("HLS_DataCollectionFSM","[OpenCAPI]") {
    DataCollectionConfig act_reg;

    STREAM_512 raw0;
    STREAM_512 raw1;

    hls::stream<axis_addr> addr0;
    hls::stream<axis_addr> addr1;

    ap_uint<1> run_data_collection = 0;
    ap_uint<1> cancel_data_collection = 0;
    ap_uint<1> idle_data_collection;
    uint32_t save_data_collection_counter;
    act_reg.mode = MODE_CONV;

    // state = WAIT_FOR_START

    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);
    REQUIRE(idle_data_collection == 1);
    REQUIRE(addr1.empty());
    REQUIRE(raw1.empty());

    run_data_collection = 1;

    // state = WAIT_FOR_START
    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);
    REQUIRE(idle_data_collection == 0);
    REQUIRE(addr1.empty());
    REQUIRE(raw1.empty());

    // state = WAIT_FOR_START_LOW
    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);
    REQUIRE(idle_data_collection == 0);
    REQUIRE(addr1.empty());
    REQUIRE(raw1.empty());

    // state = WAIT_FOR_START_LOW

    run_data_collection = 0;

    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);
    REQUIRE(idle_data_collection == 0);
    REQUIRE(addr1.empty());
    REQUIRE(raw1.empty());

    // state = START
    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);

    REQUIRE(idle_data_collection == 0);
    REQUIRE(addr1.size() == 1);
    REQUIRE(raw1.size() == 1);

    // state = INIT

    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);

    REQUIRE(idle_data_collection == 0);
    REQUIRE(addr1.size() == 1);
    REQUIRE(raw1.size() == 1);

    // state = INIT

    cancel_data_collection = 1;

    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);

    REQUIRE(idle_data_collection == 0);
    REQUIRE(addr1.size() == 1);
    REQUIRE(raw1.size() == 1);

    // state = LAST

    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);

    REQUIRE(idle_data_collection == 0);
    REQUIRE(addr1.size() == 2);
    REQUIRE(raw1.size() == 2);

    // state = WAIT_FOR_START

    data_collection_fsm(raw0, raw1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        act_reg.mode,
                        act_reg.one_over_energy,
                        act_reg.nframes,
                        act_reg.nmodules,
                        act_reg.nstorage_cells,
                        act_reg.nsummation);

    REQUIRE(idle_data_collection == 1);
    REQUIRE(addr1.size() == 2);
    REQUIRE(raw1.size() == 2);

    auto packet = raw1.read();
    REQUIRE(packet.last == 0);
    REQUIRE(packet.dest == 0);

    packet = raw1.read();
    REQUIRE(packet.last);
    REQUIRE(packet.dest == 0);
    auto addr = addr1.read();

    addr = addr1.read();
    REQUIRE(addr.last);
}


TEST_CASE("HLS_C_Simulation_internal_packet_generator_15_storage_cell_convert_G0", "[FPGA][Full]") {
    const uint16_t nmodules = 2;
    const uint16_t ntrigger = 2;
    const uint16_t nstoragecells = 15;

    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(ntrigger).ImagesPerTrigger(nstoragecells).UseInternalPacketGenerator(true)
            .PhotonEnergy_keV(10.0).StorageCells(nstoragecells);

    REQUIRE(x.GetImageNum() == ntrigger * nstoragecells);

    HLSSimulatedDevice test(0, ntrigger * nstoragecells);

    std::vector<double> tmp(3 * RAW_MODULE_SIZE, 50);
    JFModuleGainCalibration gain(tmp);

    std::vector<uint16_t> data(RAW_MODULE_SIZE);
    for (auto &i: data)
        i = 16000;
    REQUIRE_NOTHROW(test.SetInternalGeneratorFrameForAllModules(data));

    JFCalibration c(x);
    for (int i = 0; i < nstoragecells; i++) {
        for (int j = 0; j < RAW_MODULE_SIZE; j++) {
            c.Pedestal(0, 0, i).GetPedestal()[j] = (15 - i) * 500;
            c.Pedestal(1, 0, i).GetPedestal()[j] = i * 1000;
        }
    }
    c.GainCalibration(0) = gain;
    c.GainCalibration(1) = gain;

    REQUIRE_NOTHROW(test.InitializeCalibration(x, c));
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == ntrigger * nmodules * nstoragecells * 128 * JUNGFRAU_PACKET_SIZE_BYTES);

    for (int i = 0; i < ntrigger * nstoragecells; i++) {
        REQUIRE(test.GetDeviceOutput(i, 0)->pixels[511 * 764] == 32 - 15 + (i % nstoragecells));
        REQUIRE(test.GetDeviceOutput(i, 1)->pixels[200 * 145] == 32 - 2 * (i % nstoragecells));
    }
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_8_storage_cell_convert_G0", "[FPGA][Full]") {
    const uint16_t nmodules = 2;
    const uint16_t ntrigger = 2;
    const uint16_t nstoragecells = 8;

    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(ntrigger).ImagesPerTrigger(nstoragecells).UseInternalPacketGenerator(true)
            .PhotonEnergy_keV(10.0).StorageCells(nstoragecells);

    REQUIRE(x.GetImageNum() == ntrigger * nstoragecells);

    HLSSimulatedDevice test(0, ntrigger * nstoragecells);

    std::vector<double> tmp(3 * RAW_MODULE_SIZE, 50);
    JFModuleGainCalibration gain(tmp);

    std::vector<uint16_t> data(RAW_MODULE_SIZE);
    for (auto &i: data)
        i = 16000;
    REQUIRE_NOTHROW(test.SetInternalGeneratorFrameForAllModules(data));

    JFCalibration c(x);
    for (int i = 0; i < nstoragecells; i++) {
        for (int j = 0; j < RAW_MODULE_SIZE; j++) {
            c.Pedestal(0, 0, i).GetPedestal()[j] = (15 - i) * 500;
            c.Pedestal(1, 0, i).GetPedestal()[j] = i * 1000;
        }
    }
    c.GainCalibration(0) = gain;
    c.GainCalibration(1) = gain;

    REQUIRE_NOTHROW(test.InitializeCalibration(x, c));
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == ntrigger * nmodules * nstoragecells * 128 * JUNGFRAU_PACKET_SIZE_BYTES);

    for (int i = 0; i < ntrigger * nstoragecells; i++) {
        REQUIRE(test.GetDeviceOutput(i, 0)->pixels[511 * 764] == 32 - 15 + (i % nstoragecells));
        REQUIRE(test.GetDeviceOutput(i, 1)->pixels[200 * 145] == 32 - 2 * (i % nstoragecells));
    }
}


TEST_CASE("HLS_C_Simulation_internal_packet_generator_16_storage_cell_convert_G0", "[FPGA][Full]") {
    const uint16_t nmodules = 2;
    const uint16_t ntrigger = 4;
    const uint16_t nstoragecells = 16;

    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(ntrigger).ImagesPerTrigger(nstoragecells).UseInternalPacketGenerator(true)
            .PhotonEnergy_keV(10.0).StorageCells(nstoragecells);

    REQUIRE(x.GetImageNum() == ntrigger * nstoragecells);

    HLSSimulatedDevice test(0, ntrigger * nstoragecells);
    std::vector<double> tmp(3 * RAW_MODULE_SIZE, 50);
    JFModuleGainCalibration gain(tmp);

    std::vector<uint16_t> data(RAW_MODULE_SIZE);
    for (auto &i: data)
        i = 16000;
    REQUIRE_NOTHROW(test.SetInternalGeneratorFrameForAllModules(data));

    JFCalibration c(x);
    for (int i = 0; i < nstoragecells; i++) {
        for (int j = 0; j < RAW_MODULE_SIZE; j++) {
            c.Pedestal(0, 0, i).GetPedestal()[j] = (15 - i) * 500;
            c.Pedestal(1, 0, i).GetPedestal()[j] = i * 1000;
        }
    }
    c.GainCalibration(0) = gain;
    c.GainCalibration(1) = gain;

    REQUIRE_NOTHROW(test.InitializeCalibration(x, c));
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == ntrigger * nmodules * nstoragecells * 128 * JUNGFRAU_PACKET_SIZE_BYTES);

    for (int i = 0; i < ntrigger * nstoragecells; i++) {
        REQUIRE(test.GetDeviceOutput(i, 0)->pixels[511 * 764] == 32 - 15 + (i % nstoragecells));
        REQUIRE(test.GetDeviceOutput(i, 1)->pixels[200 * 145] == 32 - 2 * (i % nstoragecells));
    }
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_storage_cell_convert_G1", "[FPGA][Full]") {
    const uint16_t nmodules = 2;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Conversion);
    x.PedestalG0Frames(0).NumTriggers(1).ImagesPerTrigger(16).UseInternalPacketGenerator(true)
            .PhotonEnergy_keV(10.0).StorageCells(16);

    HLSSimulatedDevice test(0, 64);

    std::vector<uint16_t> data(RAW_MODULE_SIZE);
    for (auto &i: data)
        i = 16384 | 10;
    REQUIRE_NOTHROW(test.SetInternalGeneratorFrameForAllModules(data));

    JFCalibration c(x);
    for (int i = 0; i < 16; i++) {
        for (int j = 0; j < RAW_MODULE_SIZE; j++) {
            c.Pedestal(0, 1, i).GetPedestal()[j] = (17 - i) * 10;
            c.Pedestal(1, 1, i).GetPedestal()[j] = i * 20;
        }
    }

    std::vector<double> tmp(3 * RAW_MODULE_SIZE, -1);
    JFModuleGainCalibration gain(tmp);
    c.GainCalibration(0) = gain;
    c.GainCalibration(1) = gain;

    REQUIRE_NOTHROW(test.InitializeCalibration(x, c));
    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE_NOTHROW(test.OutputStream().read());
    REQUIRE(test.OutputStream().size() == 0);

    REQUIRE(test.GetBytesReceived() == 32*128*JUNGFRAU_PACKET_SIZE_BYTES);

    for (int i = 0; i < 16; i++) {
        REQUIRE(test.GetDeviceOutput(i, 0)->pixels[511 * 764] == 17 - (i % 16) - 1);
        REQUIRE(test.GetDeviceOutput(i, 1)->pixels[200 * 145] == 2 * (i % 16) - 1);
    }
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_integration", "[FPGA][Full]") {
    const uint16_t nmodules = 4;

    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(1).PedestalG0Frames(0);

    HLSSimulatedDevice test(0, 64);

    std::vector<uint16_t> frame(RAW_MODULE_SIZE);
    frame[0] = INT16_MAX;
    frame[1] = INT16_MIN;

    for (int i = 2; i < RAW_MODULE_SIZE; i++)
        frame[i] = 32754;

    test.SetInternalGeneratorFrameForAllModules(frame);

    std::vector<uint16_t> integration_map(nmodules * RAW_MODULE_SIZE, 54);
    for (int i = 0; i < RAW_MODULE_SIZE/2; i++) {
        integration_map[2 * i] = 0;
        integration_map[2 * i + 1] = FPGA_INTEGRATION_BIN_COUNT - 1;
    }

    integration_map[RAW_MODULE_SIZE - 1] = FPGA_INTEGRATION_BIN_COUNT;
    test.InitializeIntegrationMap(x, integration_map);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    JFJochProtoBuf::AcquisitionDeviceStatistics device_statistics;
    REQUIRE_NOTHROW(test.SaveStatistics(x, device_statistics));
    REQUIRE(device_statistics.bytes_received() == 128 * nmodules * JUNGFRAU_PACKET_SIZE_BYTES);

    auto imageBuf = test.GetDeviceOutput(0, 0)->pixels;
    REQUIRE(memcmp(imageBuf, frame.data(), RAW_MODULE_SIZE * sizeof(uint16_t)) == 0);

    auto integration_result = test.GetDeviceOutput(0, 0)->integration_result;
    CHECK(integration_result[0].sum == 32754LU * (RAW_MODULE_SIZE / 2 - 1) * (1LU<<24));
    CHECK(integration_result[0].count == RAW_MODULE_SIZE / 2 - 1);
    CHECK(integration_result[1].sum == 0);
    CHECK(integration_result[1].count == 0);
    CHECK(integration_result[FPGA_INTEGRATION_BIN_COUNT - 1].sum == 32754LU * (RAW_MODULE_SIZE / 2 - 2)* (1LU<<24));
    CHECK(integration_result[FPGA_INTEGRATION_BIN_COUNT - 1].count == RAW_MODULE_SIZE / 2 - 2);

    integration_result = test.GetDeviceOutput(0, 1)->integration_result;
    CHECK(integration_result[54].sum == 32754LU * (RAW_MODULE_SIZE - 2) * (1LU<<24));
    CHECK(integration_result[54].count == RAW_MODULE_SIZE - 2);

    integration_result = test.GetDeviceOutput(0, 2)->integration_result;
    CHECK(integration_result[54].sum == 32754 * (RAW_MODULE_SIZE - 2) * (1LU<<24));
    CHECK(integration_result[54].count == RAW_MODULE_SIZE - 2);

    integration_result = test.GetDeviceOutput(0, 3)->integration_result;
    CHECK(integration_result[54].sum == 32754 * (RAW_MODULE_SIZE - 2) * (1LU<<24));
    CHECK(integration_result[54].count == RAW_MODULE_SIZE - 2);
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_spot_finder_count_threshold", "[FPGA][Full]") {
    const uint16_t nmodules = 4;

    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(1).PedestalG0Frames(0);

    HLSSimulatedDevice test(0, 64);

    std::vector<uint16_t> frame(RAW_MODULE_SIZE, 0);
    frame [             0] = 11;
    frame [123*1024 + 578] = 10;
    frame [121*1024 + 800] = 12;
    frame [ 89*1024 + 300] =  8;
    frame [300*1024 +   0] =  9;

    test.SetInternalGeneratorFrameForAllModules(frame);

    test.SetSpotFinderParameters(10, 0.0);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    JFJochProtoBuf::AcquisitionDeviceStatistics device_statistics;
    REQUIRE_NOTHROW(test.SaveStatistics(x, device_statistics));
    REQUIRE(device_statistics.bytes_received() == 128 * nmodules * JUNGFRAU_PACKET_SIZE_BYTES);

    auto imageBuf = test.GetDeviceOutput(0, 0)->pixels;
    REQUIRE(memcmp(imageBuf, frame.data(), RAW_MODULE_SIZE * sizeof(uint16_t)) == 0);

    auto spot_finder_result = test.GetDeviceOutput(0, 0)->spot_finding_result;
    REQUIRE (spot_finder_result.strong_pixel_count == 3);
    REQUIRE (spot_finder_result.snr_threshold == 0);
    REQUIRE (spot_finder_result.count_threshold == 10);
    REQUIRE (spot_finder_result.strong_pixel[0] == (1<<0));
    REQUIRE (spot_finder_result.strong_pixel[(123*1024 + 578) / 8] == (1<<2)); // 578 % 8 == 2
    REQUIRE (spot_finder_result.strong_pixel[(121*1024 + 800) / 8] == (1<<0)); // 800 % 8 == 0
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_spot_finder_snr_threshold", "[FPGA][Full]") {
    const uint16_t nmodules = 4;

    DiffractionExperiment x((DetectorGeometry(nmodules)));

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(1).PedestalG0Frames(0);

    HLSSimulatedDevice test(0, 64);

    std::vector<uint16_t> frame(RAW_MODULE_SIZE);
    for (int i = 0; i < RAW_MODULE_SIZE; i++) {
        frame[i] = ((i / RAW_MODULE_COLS) + (i % RAW_MODULE_COLS)) % 2;
    }
    // Mean = 0.5
    // Std. dev. = 0.5
    // Threshold = 10 * std. dev. - 6 is minimum count

    frame [             0] = 8;
    frame [123*1024 + 578] = 5;
    frame [121*1024 + 800] = 4;
    frame [ 89*1024 + 300] = 7;
    frame [300*1024 +   0] = 3;

    test.SetInternalGeneratorFrameForAllModules(frame);

    test.SetSpotFinderParameters(0, 10);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    JFJochProtoBuf::AcquisitionDeviceStatistics device_statistics;
    REQUIRE_NOTHROW(test.SaveStatistics(x, device_statistics));
    REQUIRE(device_statistics.bytes_received() == 128 * nmodules * JUNGFRAU_PACKET_SIZE_BYTES);

    auto imageBuf = test.GetDeviceOutput(0, 0)->pixels;
    REQUIRE(memcmp(imageBuf, frame.data(), RAW_MODULE_SIZE * sizeof(uint16_t)) == 0);

    auto spot_finder_result = test.GetDeviceOutput(0, 0)->spot_finding_result;
    REQUIRE (spot_finder_result.strong_pixel_count == 2);
    REQUIRE (spot_finder_result.snr_threshold == 10 * 4);
    REQUIRE (spot_finder_result.count_threshold == 0);
    REQUIRE (spot_finder_result.strong_pixel[0] == (1<<0));
    REQUIRE (spot_finder_result.strong_pixel[(89*1024 + 300) / 8] == (1<<4));  // 300 % 8 == 4
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_32bit", "[FPGA][Full]") {
    const uint16_t nmodules = 1;
    const size_t nframes = 2;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    std::vector<uint16_t> test_frame(nmodules*RAW_MODULE_SIZE);

    std::mt19937 g1(1387);
    std::uniform_int_distribution<uint16_t> dist(0, UINT16_MAX);

    for (auto &i: test_frame)
        i = dist(g1);

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(nframes).PedestalG0Frames(0).FPGAOutputMode(JFJochProtoBuf::INT32);

    HLSSimulatedDevice test(0, 64);
    test.SetInternalGeneratorFrame(test_frame);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    REQUIRE(test.GetBytesReceived() == 128 * nmodules * nframes * JUNGFRAU_PACKET_SIZE_BYTES);

    auto test_frame_signed = (int16_t *) test_frame.data();

    for (int image = 0; image < nframes; image++) {
        for (int m = 0; m < nmodules; m++) {
            REQUIRE(test.Counters().GetCompletion(image, m).timestamp == INT_PKT_GEN_EXPTTIME * image);
            REQUIRE(test.Counters().GetCompletion(image, m).bunchid == INT_PKT_GEN_BUNCHID + image);
            REQUIRE(test.Counters().GetCompletion(image, m).exptime == INT_PKT_GEN_EXPTTIME);

            auto imageBuf = (int32_t *) test.GetDeviceOutput(image, m)->pixels;
            for (int i = 0; i < RAW_MODULE_SIZE; i++) {
                if (test_frame_signed[m * RAW_MODULE_SIZE + i] == INT16_MAX)
                    REQUIRE(imageBuf[i] == INT32_MAX);
                else if (test_frame_signed[m * RAW_MODULE_SIZE + i] == INT16_MIN)
                    REQUIRE(imageBuf[i] == INT32_MIN);
                else
                    REQUIRE(imageBuf[i] == test_frame_signed[m * RAW_MODULE_SIZE + i]);
            }
        }
    }
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_32bit_unsigned", "[FPGA][Full]") {
    const uint16_t nmodules = 1;
    const size_t nframes = 2;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    std::vector<uint16_t> test_frame(nmodules*RAW_MODULE_SIZE);

    std::mt19937 g1(1387);
    std::uniform_int_distribution<uint16_t> dist(0, UINT16_MAX);

    for (auto &i: test_frame)
        i = dist(g1);

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(nframes).PedestalG0Frames(0).FPGAOutputMode(JFJochProtoBuf::UINT32);

    HLSSimulatedDevice test(0, 64);
    test.SetInternalGeneratorFrame(test_frame);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    REQUIRE(test.GetBytesReceived() == 128 * nmodules * nframes * JUNGFRAU_PACKET_SIZE_BYTES);

    auto test_frame_signed = (int16_t *) test_frame.data();

    for (int image = 0; image < nframes; image++) {
        for (int m = 0; m < nmodules; m++) {
            REQUIRE(test.Counters().GetCompletion(image, m).timestamp == INT_PKT_GEN_EXPTTIME * image);
            REQUIRE(test.Counters().GetCompletion(image, m).bunchid == INT_PKT_GEN_BUNCHID + image);
            REQUIRE(test.Counters().GetCompletion(image, m).exptime == INT_PKT_GEN_EXPTTIME);

            auto imageBuf = (uint32_t *) test.GetDeviceOutput(image, m)->pixels;
            for (int i = 0; i < RAW_MODULE_SIZE; i++) {
                if (test_frame_signed[m * RAW_MODULE_SIZE + i] == INT16_MAX)
                    REQUIRE(imageBuf[i] == UINT32_MAX - 1);
                else if (test_frame_signed[m * RAW_MODULE_SIZE + i] == INT16_MIN)
                    REQUIRE(imageBuf[i] == UINT32_MAX);
                else if (test_frame_signed[m * RAW_MODULE_SIZE + i] < 0)
                    REQUIRE(imageBuf[i] == 0);
                else
                    REQUIRE(imageBuf[i] == test_frame_signed[m * RAW_MODULE_SIZE + i]);
            }
        }
    }
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_16bit_unsigned", "[FPGA][Full]") {
    const uint16_t nmodules = 1;
    const size_t nframes = 2;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    std::vector<uint16_t> test_frame(nmodules*RAW_MODULE_SIZE);

    std::mt19937 g1(1387);
    std::uniform_int_distribution<uint16_t> dist(0, UINT16_MAX);

    for (auto &i: test_frame)
        i = dist(g1);

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(nframes).PedestalG0Frames(0).FPGAOutputMode(JFJochProtoBuf::UINT16);

    HLSSimulatedDevice test(0, 64);
    test.SetInternalGeneratorFrame(test_frame);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    REQUIRE(test.GetBytesReceived() == 128 * nmodules * nframes * JUNGFRAU_PACKET_SIZE_BYTES);

    auto test_frame_signed = (int16_t *) test_frame.data();

    for (int image = 0; image < nframes; image++) {
        for (int m = 0; m < nmodules; m++) {
            REQUIRE(test.Counters().GetCompletion(image, m).timestamp == INT_PKT_GEN_EXPTTIME * image);
            REQUIRE(test.Counters().GetCompletion(image, m).bunchid == INT_PKT_GEN_BUNCHID + image);
            REQUIRE(test.Counters().GetCompletion(image, m).exptime == INT_PKT_GEN_EXPTTIME);

            auto imageBuf = (uint16_t *) test.GetDeviceOutput(image, m)->pixels;
            for (int i = 0; i < RAW_MODULE_SIZE; i++) {
                if (test_frame_signed[m * RAW_MODULE_SIZE + i] == INT16_MAX)
                    REQUIRE(imageBuf[i] == UINT16_MAX - 1);
                else if (test_frame_signed[m * RAW_MODULE_SIZE + i] == INT16_MIN)
                    REQUIRE(imageBuf[i] == UINT16_MAX);
                else if (test_frame_signed[m * RAW_MODULE_SIZE + i] < 0)
                    REQUIRE(imageBuf[i] == 0);
                else
                    REQUIRE(imageBuf[i] == test_frame_signed[m * RAW_MODULE_SIZE + i]);
            }
        }
    }
}

TEST_CASE("HLS_C_Simulation_internal_packet_generator_summation", "[FPGA][Full]") {
    const uint16_t nmodules = 1;
    const size_t nframes = 16;
    const size_t nsummation = 16;
    DiffractionExperiment x((DetectorGeometry(nmodules)));

    std::vector<uint16_t> test_frame(nmodules*RAW_MODULE_SIZE);

    std::mt19937 g1(1387);
    std::uniform_int_distribution<uint16_t> dist(0, 65535);

    for (auto &i: test_frame)
        i = dist(g1);

    x.Mode(DetectorMode::Raw);
    x.UseInternalPacketGenerator(true).ImagesPerTrigger(nframes).PedestalG0Frames(0).Summation(nsummation);

    HLSSimulatedDevice test(0, 64);
    test.SetInternalGeneratorFrame(test_frame);

    REQUIRE_NOTHROW(test.StartAction(x));
    REQUIRE_NOTHROW(test.WaitForActionComplete());

    REQUIRE(test.OutputStream().size() == 1);

    REQUIRE(test.GetBytesReceived() == 128 * nmodules * nframes * nsummation * JUNGFRAU_PACKET_SIZE_BYTES);

    auto test_frame_signed = (int16_t *) test_frame.data();
    for (int image = 0; image < nframes; image++) {
        for (int m = 0; m < nmodules; m++) {
            REQUIRE(test.Counters().GetCompletion(image, m).timestamp == INT_PKT_GEN_EXPTTIME * image * nsummation);
            REQUIRE(test.Counters().GetCompletion(image, m).bunchid == INT_PKT_GEN_BUNCHID + image * nsummation);
            REQUIRE(test.Counters().GetCompletion(image, m).exptime == INT_PKT_GEN_EXPTTIME);

            auto imageBuf = (int32_t *) test.GetDeviceOutput(image, m)->pixels;
            for (int i = 0; i < RAW_MODULE_SIZE; i++) {
                if (test_frame_signed[m * RAW_MODULE_SIZE + i] == INT16_MAX)
                    REQUIRE(imageBuf[i] == INT32_MAX);
                else if (test_frame_signed[m * RAW_MODULE_SIZE + i] == INT16_MIN)
                    REQUIRE(imageBuf[i] == INT32_MIN);
                else
                    REQUIRE(imageBuf[i] == test_frame_signed[m * RAW_MODULE_SIZE + i] * nsummation);
            }
        }
    }
}