// Copyright (2019-2023) Paul Scherrer Institute

#include "HLSSimulatedDevice.h"

#include <bitset>
#include <arpa/inet.h>
#include "../fpga/hls/datamover_model.h"
#include "../fpga/hls/hls_jfjoch.h"

uint16_t checksum(const uint16_t *addr, size_t count) {
/* Compute Internet Checksum for "count" bytes
 *         beginning at location "addr".
 */
    long sum = 0;

    for (int i = 0; i < count / 2; i++)
        sum += addr[i];

/*  Add left-over byte, if any */
    if (count % 2 == 1)
        sum += ((uint8_t *) addr)[count / 2];

/*  Fold 32-bit sum to 16 bits */
    while (sum>>16)
        sum = (sum & 0xffff) + (sum >> 16);

    return ~sum;
}

HLSSimulatedDevice::HLSSimulatedDevice(uint16_t data_stream, size_t in_frame_buffer_size_modules, int16_t numa_node)
        : FPGAAcquisitionDevice(data_stream),
          datamover_in(Direction::Input, nullptr),
          datamover_out(Direction::Output, nullptr),
          datamover_out_hbm_0(Direction::Output, (char *) hbm.data()),
          datamover_out_hbm_1(Direction::Output, (char *) hbm.data()),
          datamover_in_hbm_0(Direction::Input, (char *) hbm.data()),
          datamover_in_hbm_1(Direction::Input, (char *) hbm.data()),
          idle(true), hbm(hbm_if_size / 32 * hbm_if_count),
          dma_address_table(65536) {
    mac_addr = 0xCCAA11223344;
    ipv4_addr = 0x0132010A;

    max_modules = MAX_MODULES_FPGA;

    MapBuffersStandard(in_frame_buffer_size_modules, numa_node);

    auto in_mem_location32 = (uint32_t *) dma_address_table.data();

    for (int i = 0; i < buffer_device.size(); i++) {
        in_mem_location32[2 * i    ] = ((uint64_t) buffer_device[i]) & UINT32_MAX;
        in_mem_location32[2 * i + 1] = ((uint64_t) buffer_device[i]) >> 32;
    }
}

void HLSSimulatedDevice::CreateFinalPacket(const DiffractionExperiment& experiment) {
    CreateJFPacket(experiment, UINT64_MAX, 0, 0, nullptr, false);
}

void HLSSimulatedDevice::SendPacket(char *buffer, int len, uint8_t user) {
    auto obuff = (ap_uint<512> *)buffer;

    for (int i = 0; i < (len + 63) / 64; i++) {
        packet_512_t packet_in;
        if (i == (len + 63) / 64 - 1) packet_in.last = 1;
        else packet_in.last = 0;
        packet_in.keep = 0xFFFFFFFFFFFFFFFF;
        packet_in.user = user;
        packet_in.data = obuff[i];
        din_eth.write(packet_in);
    }
}

void HLSSimulatedDevice::CreateJFPacket(const DiffractionExperiment& experiment, uint64_t frame_number, uint32_t eth_packet,
                                        uint32_t module_number, const uint16_t *data, int8_t adjust_axis, uint8_t user) {
    char buff[256*64];
    memset(buff, 0, 256*64);

    auto packet = (jf_raw_packet *)buff;

    packet->ether_type = htons(0x0800);
    packet->sour_mac[0] = 0x00; // module 0

    uint64_t tmp_mac = mac_addr;
    for (int i = 0; i < 6; i++)
        packet->dest_mac[i] = (tmp_mac >> (8*i)) % 256;

    uint32_t half_module = 2 * module_number | ((eth_packet >= 64) ? 1 : 0);

    packet->ipv4_header_h = htons(0x4500); // Big endian in IP header!
    packet->ipv4_header_total_length = htons(8268); // Big endian in IP header!
    packet->ipv4_header_dest_ip = ipv4_addr;
    packet->ipv4_header_sour_ip = experiment.GetSrcIPv4Address(data_stream, half_module);

    packet->ipv4_header_ttl_protocol = htons(0x0011);
    packet->ipv4_header_checksum = checksum( (uint16_t *) &packet->ipv4_header_h, 20); // checksum is already in network order

    packet->udp_dest_port = htons(GetUDPPort()); // module number
    packet->udp_sour_port = htons(0xDFAC);
    packet->udp_length = htons(8248);

    // JF headers are little endian
    packet->jf.detectortype = SLS_DETECTOR_TYPE_JUNGFRAU;
    packet->jf.timestamp = 0xABCDEF0000FEDCBAL;
    packet->jf.bunchid   = 0x1234567898765431L;
    packet->jf.row = half_module;
    packet->jf.column = 0;
    packet->jf.framenum = frame_number;
    packet->jf.packetnum = eth_packet % 64;
    if (data != nullptr) {
        for (int i = 0; i < 4096; i++)
            packet->jf.data[i] = data[i];
    }
    packet->udp_checksum = htons(checksum( (uint16_t *) (buff+42), 8192+48));

    SendPacket(buff, (130+adjust_axis)*64, user);
}

void HLSSimulatedDevice::CreateJFPackets(const DiffractionExperiment& experiment, uint64_t frame_number_0, uint64_t frames,
                                         uint32_t module_number, const uint16_t *data) {
    for (uint64_t i = 0; i < frames; i++) {
        for (int j = 0; j < 128; j++)
            CreateJFPacket(experiment, frame_number_0 + i, j, module_number, data + (i * 128 + j) * 4096, 0, 0);
    }
}

void HLSSimulatedDevice::CreateEIGERPacket(const DiffractionExperiment &experiment, uint64_t frame_number,
                                           uint32_t eth_packet, uint32_t module_number, uint32_t col, uint32_t row,
                                           const uint16_t *data) {
    char buff[256*64];
    memset(buff, 0, 256*64);

    auto packet = (eiger_raw_packet *)buff;

    packet->ether_type = htons(0x0800);
    packet->sour_mac[0] = 0x00; // module 0

    uint64_t tmp_mac = mac_addr;
    for (int i = 0; i < 6; i++)
        packet->dest_mac[i] = (tmp_mac >> (8*i)) % 256;

    packet->ipv4_header_h = htons(0x4500); // Big endian in IP header!
    packet->ipv4_header_total_length = htons(4172); // Big endian in IP header!
    packet->ipv4_header_dest_ip = ipv4_addr;
    packet->ipv4_header_sour_ip = experiment.GetSrcIPv4Address(data_stream, 0);

    packet->ipv4_header_ttl_protocol = htons(0x0011);
    packet->ipv4_header_checksum = checksum( (uint16_t *) &packet->ipv4_header_h, 20); // checksum is already in network order

    packet->udp_dest_port = htons(GetUDPPort()); // module number
    packet->udp_sour_port = htons(0xDFAC);
    packet->udp_length = htons(4152);

    // JF headers are little endian
    packet->eiger.detectortype = SLS_DETECTOR_TYPE_EIGER;
    packet->eiger.timestamp = 0xABCDEF0000FEDCBAL;
    packet->eiger.bunchid   = 0x1234567898765431L;
    packet->eiger.row = (2 * module_number) | (row % 2);
    packet->eiger.column = col % 2;
    packet->eiger.framenum = frame_number;
    packet->eiger.packetnum = eth_packet % 64;
    if (data != nullptr) {
        for (int i = 0; i < 2048; i++)
            packet->eiger.data[i] = data[i];
    }
    packet->udp_checksum = htons(checksum( (uint16_t *) (buff+42), 4096+48));

    SendPacket(buff, 66*64, 0);
}

AXI_STREAM & HLSSimulatedDevice::OutputStream() {
    return dout_eth;
}

void HLSSimulatedDevice::HW_ReadActionRegister(DataCollectionConfig *job) {
    memcpy(job, &cfg, sizeof(DataCollectionConfig));
}

void HLSSimulatedDevice::HW_WriteActionRegister(const DataCollectionConfig *job) {
    memcpy(&cfg, job, sizeof(DataCollectionConfig));
}

void HLSSimulatedDevice::FPGA_StartAction(const DiffractionExperiment &experiment) {
    if (action_thread.joinable())
        action_thread.join();

    run_counter += 1;
    run_data_collection = 1;
    cancel_data_collection = 0;
    idle = false;

    while (!din_frame_generator.empty())
        din_frame_generator.read();

    datamover_out.ClearCompletedDescriptors();

    action_thread = std::thread(&HLSSimulatedDevice::HLSMainThread, this );
}

void HLSSimulatedDevice::FrameGeneratorFuture(FrameGeneratorConfig config) {
    frame_generator(din_frame_generator,
                    hbm.data(),
                    hbm.data(),
                    hbm_if_size,
                    mac_addr,
                    ipv4_addr,
                    cancel_data_collection,
                    config);
}

void HLSSimulatedDevice::HW_RunInternalGenerator(const FrameGeneratorConfig &config) {
    frame_generator_future = std::async(std::launch::async, &HLSSimulatedDevice::FrameGeneratorFuture, this, config);
}

void HLSSimulatedDevice::FPGA_EndAction() {
    if (action_thread.joinable())
        action_thread.join();
}

HLSSimulatedDevice::~HLSSimulatedDevice() {
    if (action_thread.joinable())
        action_thread.join();
}

bool HLSSimulatedDevice::HW_ReadMailbox(uint32_t *values) {
    std::unique_lock<std::mutex> ul(completion_mutex);

    ap_uint<32> tmp;
    bool ret = completion_stream.read_nb(tmp);
    values[0] = tmp;
    // equivalent to driver functionality
    if (ret) {
        uint32_t data_collection_id = (values[0] >> 16) & 0xFFFF;
        uint32_t handle = values[0] & 0xFFFF;
        if (handle == HANDLE_START)
            completion_count = 0;
        else if ((handle != HANDLE_END) && (data_collection_id != DATA_COLLECTION_ID_PURGE)) {
            completion_count++;
            while (completion_count * DMA_DESCRIPTORS_PER_MODULE > datamover_out.GetCompletedDescriptors())
                std::this_thread::sleep_for(std::chrono::milliseconds(1));
        }
    }
    return ret;
}

void HLSSimulatedDevice::Cancel() {
    cancel_data_collection = 1;
}

bool HLSSimulatedDevice::HW_IsIdle() const {
    return idle && datamover_out.IsIdle();
}


bool HLSSimulatedDevice::HW_SendWorkRequest(uint32_t handle) {
    work_request_stream.write(handle);
    return true;
}

inline uint32_t float2uint(float f) {
    float_uint32 fu;
    fu.f = f;
    return fu.u;
}

void HLSSimulatedDevice::HLSMainThread() {
    ap_uint<1> clear_counters = 0;

    uint64_t packets_processed;

    std::vector<std::thread> hls_cores;

    STREAM_512 ip1, udp1, udp2, icmp1, arp1;

    STREAM_512 network0;

    STREAM_768 stream_768_0;
    STREAM_768 stream_768_1;
    STREAM_768 stream_768_2;
    STREAM_768 stream_768_3;
    STREAM_768 stream_768_4;
    STREAM_768 stream_768_5;

    STREAM_512 data_0;
    hls::stream<ap_axiu<512, 1, 1, 1>, 2> data_1;
    hls::stream<ap_axiu<512, 1, 1, 1>, 2> data_2;
    STREAM_512 data_3;
    STREAM_512 data_4;
    STREAM_512 data_5;
    STREAM_512 data_6;
    STREAM_512 data_7;
    STREAM_512 data_8;
    STREAM_512 data_9;
    STREAM_512 data_10;
    STREAM_512 data_12;
    STREAM_512 data_13;

    hls::stream<axis_addr> addr0;
    hls::stream<axis_addr> addr1;
    hls::stream<axis_addr> addr2;
    hls::stream<axis_addr> addr3;

    hls::stream<axis_completion> axi_compl[12];

    hls::stream<ap_uint<16>> hbm_handles;
    hls::stream<ap_uint<512>> adu_histo_result;

    hls::stream<ap_axiu<64,1,1,1>> integration_result_0;
    hls::stream<ap_uint<512>> integration_result_1;

    hls::stream<ap_axiu<32,1,1,1>> spot_finder_result_0;
    hls::stream<ap_axiu<32,1,1,1>> spot_finder_result_1;
    hls::stream<ap_uint<32>> spot_finder_conn_0;
    hls::stream<ap_axiu<32,1,1,1>> spot_finder_result_2;
    hls::stream<ap_uint<512>> spot_finder_result_3;

    hls::stream<ap_uint<32>> spot_finder_mask_0;

    hls::stream<ap_uint<256>> roi_calc_result_0;

    hls::stream<ap_uint<UDP_METADATA_STREAM_WIDTH> > udp_metadata;
    volatile ap_uint<1> idle_data_collection = 1;
    ap_uint<1> load_to_hbm_idle;
    ap_uint<1> save_to_hbm_idle;
    ap_uint<1> integration_idle;
    ap_uint<1> stream_conv_idle;
    ap_uint<1> frame_summation_idle;

    volatile bool done = false;
    volatile bool udp_done = false; // done AND udp_idle
    volatile bool sls_done = false; // done AND sls_idle

    // Sent gratuitous ARP message
    arp(arp1, dout_eth, mac_addr, ipv4_addr, 1, 1);

    Logger logger_hls("HLS");

    volatile rcv_state_t state = RCV_INIT;
    // Start data collection
    data_collection_fsm(data_0, data_1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        cfg.mode,
                        float2uint(cfg.energy_kev),
                        cfg.nframes,
                        cfg.nmodules,
                        cfg.nstorage_cells,
                        cfg.nsummation , state);

    run_data_collection = 0;
    data_collection_fsm(data_0, data_1,
                        addr0, addr1,
                        run_data_collection,
                        cancel_data_collection,
                        idle_data_collection,
                        cfg.mode,
                        float2uint(cfg.energy_kev),
                        cfg.nframes,
                        cfg.nmodules,
                        cfg.nstorage_cells,
                        cfg.nsummation, state);

    hls_cores.emplace_back([&] {
        while (!udp_done) {
            if (din_eth.empty() && din_frame_generator.empty())
                std::this_thread::sleep_for(std::chrono::microseconds(10));
            else
                stream_merge(din_eth, din_frame_generator, network0, data_source);
        }
        logger_hls.Info("Stream_merge done");
    });

    hls_cores.emplace_back([&] {
        while (!udp_done) {
            if (network0.empty())
                std::this_thread::sleep_for(std::chrono::microseconds(10));
            else
                ethernet(network0, ip1, arp1, mac_addr, eth_packets, clear_counters);
        }
        logger_hls.Info("ethernet done");
    });

    hls_cores.emplace_back([&] {
        while (!udp_done) {
            if (ip1.empty())
                std::this_thread::sleep_for(std::chrono::microseconds(10));
            else
                ipv4(ip1, udp1, icmp1, ipv4_addr);
        }
        logger_hls.Info("ipv4 done");
    });

    hls_cores.emplace_back([&] {
        ap_uint<1> udp_idle = 1;
        while (!done || !udp_idle) {
            if (udp1.empty())
                std::this_thread::sleep_for(std::chrono::microseconds(10));
            else
                udp(udp1, udp2, udp_metadata, udp_packets, clear_counters, udp_idle);
        }
        udp_done = true;
        logger_hls.Info("udp done");
    });

    hls_cores.emplace_back([&] {
        ap_uint<1> sls_idle = 1;
        while (!done || !sls_idle) {
            if (udp2.empty())
                std::this_thread::sleep_for(std::chrono::microseconds (10));
            else
                sls_detector(udp2, udp_metadata, data_0, addr0, sls_packets, udp_eth_err, udp_len_err, current_pulse_id,
                             clear_counters, sls_idle);
        }
        sls_done = true;
        logger_hls.Info("sls_detector done");
    });

    // 1. Parse incoming UDP packets
    hls_cores.emplace_back([&] {
        while ((state != RCV_WAIT_FOR_START) || (idle_data_collection.read() == 0) || (!data_0.empty()))  {
            data_collection_fsm(data_0, data_1,
                                addr0, addr1,
                                run_data_collection,
                                cancel_data_collection,
                                idle_data_collection,
                                cfg.mode,
                                float2uint(cfg.energy_kev),
                                cfg.nframes,
                                cfg.nmodules,
                                cfg.nstorage_cells,
                                cfg.nsummation,
                                state);
        }
        done = true;

        logger_hls.Info("data collection done");
    });

    // 2. Cache images in HBM
    hls_cores.emplace_back([&] { save_to_hbm(addr1, axi_compl[0], hbm_handles,
                                             datamover_out_hbm_0.GetCtrlStream(),
                                             datamover_out_hbm_1.GetCtrlStream(),
                                             save_to_hbm_idle,
                                             hbm_if_size);});

    hls_cores.emplace_back([&] { save_to_hbm_data(data_1,
                                                  data_3,
                                                  datamover_out_hbm_0.GetDataStream(),
                                                  datamover_out_hbm_1.GetDataStream());});

    hls_cores.emplace_back([&] {frame_summation_reorder_compl(data_3, data_4, axi_compl[0], axi_compl[1]);});

    hls_cores.emplace_back([&] { load_from_hbm(data_4, data_5, axi_compl[1], axi_compl[2], hbm_handles,
                                               datamover_in_hbm_0.GetDataStream(), datamover_in_hbm_1.GetDataStream(),
                                               datamover_in_hbm_0.GetCtrlStream(), datamover_in_hbm_1.GetCtrlStream(),
                                               load_to_hbm_idle, hbm_if_size);});

    hls_cores.emplace_back([&] { pedestal(data_5, data_6, axi_compl[2], axi_compl[3],
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm.data(),
                                          hbm_if_size);});


    // 3. Calculate histogram of ADU values
    hls_cores.emplace_back([&] { adu_histo(data_6, data_7, adu_histo_result, axi_compl[3], axi_compl[4]);});

    // 4. Mask missing pixels
    hls_cores.emplace_back([&] { mask_missing(data_7, data_8, axi_compl[4], axi_compl[5]);});

    // 5. Handle EIGER packets properly
    hls_cores.emplace_back([&] { eiger_reorder(data_8, data_9, axi_compl[5], axi_compl[6]);});

    // 6. Apply pedestal & gain corrections
    hls_cores.emplace_back([&] { jf_conversion(data_9, stream_768_0,
                                               axi_compl[6], axi_compl[7],
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm.data(),
                                               hbm_if_size); });

    // 7. Frame summation
    hls_cores.emplace_back([&] { frame_summation(stream_768_0, stream_768_1, axi_compl[7], axi_compl[8], frame_summation_idle);});

    // 8. Integration of pixels
    hls_cores.emplace_back([&] { integration(stream_768_1, stream_768_2, integration_result_0, axi_compl[8], axi_compl[9],
                                             hbm.data(), hbm.data(), hbm.data(), hbm.data(), integration_idle, hbm_if_size);});
    hls_cores.emplace_back([&] { axis_64_to_512(integration_result_0, integration_result_1);});

    // 9. Spot finding
    ap_uint<32> tmp_snr_threshold = float2uint(spot_finder_parameters.snr_threshold);
    ap_uint<32> min_d = float2uint(spot_finder_parameters.min_d);
    ap_uint<32> max_d = float2uint(spot_finder_parameters.max_d);
    ap_int<32> tmp_count_threshold = spot_finder_parameters.count_threshold;
    ap_uint<32> min_pix_per_spot = spot_finder_parameters.min_pix_per_spot;

    hls_cores.emplace_back([&] {
        spot_finder_mask(stream_768_2,
                         stream_768_3,
                         spot_finder_mask_0,
                         axi_compl[9],
                         axi_compl[10],
                         hbm.data(),
                         hbm.data(),
                         min_d,
                         max_d,
                         hbm_if_size);
        logger_hls.Info("spot_finder_mask done");
    });

    hls_cores.emplace_back([&] {
        spot_finder(stream_768_3, spot_finder_mask_0, stream_768_4, spot_finder_result_0, tmp_count_threshold, tmp_snr_threshold);
        logger_hls.Info("spot_finder done");
    });

    hls_cores.emplace_back([&] {
        spot_finder_connectivity(spot_finder_result_0,
                                 spot_finder_result_1,
                                 spot_finder_conn_0);
        logger_hls.Info("spot_finder_connectivity done");
    });

    hls_cores.emplace_back([&] {
        spot_finder_merge(spot_finder_result_1,
                          spot_finder_conn_0,
                          spot_finder_result_2,
                          min_pix_per_spot);
        logger_hls.Info("spot_finder_merge done");
    });

    hls_cores.emplace_back([&] {
        axis_32_to_512(spot_finder_result_2, spot_finder_result_3);
        logger_hls.Info("axis_32_to_512 done");
    });

    hls_cores.emplace_back([&] {
        roi_calc(stream_768_4,
                 stream_768_5,
                 roi_calc_result_0,
                 axi_compl[10],
                 axi_compl[11],
                 hbm.data(),
                 hbm.data(),
                 hbm_if_size);
        logger_hls.Info("roi_calc done");
    });

    // 10. Reduce/extend 24-bit stream
    hls_cores.emplace_back([&] { stream_24bit_conv(stream_768_5, data_12, stream_conv_idle);});

    // 11. Prepare data to write to host memory
    hls_cores.emplace_back([&] {
        ap_uint<3> state;
        host_writer(data_12, adu_histo_result, integration_result_1, spot_finder_result_3, roi_calc_result_0,
                    axi_compl[11], datamover_out.GetDataStream(),
                    datamover_out.GetCtrlStream(), work_request_stream, completion_stream,
                    dma_address_table.data(), packets_processed, host_writer_idle, cancel_data_collection, state);
        logger_hls.Info("host_writer done");
    });

    for (auto &i : hls_cores)
        i.join();

    if (frame_generator_future.valid())
        frame_generator_future.get();

    // reset static counter
    arp(arp1, dout_eth, mac_addr, ipv4_addr, 0, 1);
    try {
        while (!din_eth.empty())
            din_eth.read();

        if (!addr1.empty())
            throw std::runtime_error("Addr1 queue not empty");

        if (!addr2.empty())
            throw std::runtime_error("Addr2 queue not empty");

        if (!addr3.empty())
            throw std::runtime_error("Addr3 queue not empty");

        if (!data_1.empty())
            throw std::runtime_error("data_1 queue not empty");

        if (!data_2.empty())
            throw std::runtime_error("data_2 queue not empty");

        if (!data_3.empty())
            throw std::runtime_error("data_3 queue not empty");

        if (!data_4.empty())
            throw std::runtime_error("data_4 queue not empty");

        if (!data_5.empty())
            throw std::runtime_error("data_5 queue not empty");

        if (!data_7.empty())
            throw std::runtime_error("data_7 queue not empty");

        if (!data_8.empty())
            throw std::runtime_error("data_8 queue not empty");

        if (!data_9.empty())
            throw std::runtime_error("data_9 queue not empty");

        if (!data_10.empty())
            throw std::runtime_error("data_10 queue not empty");



        if (!data_12.empty())
            throw std::runtime_error("data_12 queue not empty");

        if (!data_13.empty())
            throw std::runtime_error("data_13 queue not empty");

        for (auto &c: axi_compl) {
            if (!c.empty())
                throw std::runtime_error("Compl queue not empty");
        }

        if (!stream_768_0.empty())
            throw std::runtime_error("stream_768_0 queue not empty");

        if (!stream_768_1.empty())
            throw std::runtime_error("stream_768_1 queue not empty");

        if (!stream_768_2.empty())
            throw std::runtime_error("stream_768_2 queue not empty");

        if (!stream_768_3.empty())
            throw std::runtime_error("stream_768_3 queue not empty");

        if (!stream_768_4.empty())
            throw std::runtime_error("stream_768_4 queue not empty");

        if (!hbm_handles.empty())
            throw std::runtime_error("Handles queue not empty");

        if (!integration_result_0.empty())
            throw std::runtime_error("Integration result queue not empty");

        if (!integration_result_1.empty())
            throw std::runtime_error("Integration result queue not empty");

        if (!datamover_in.GetDataStream().empty())
            throw std::runtime_error("Datamover queue is not empty");

        while (!datamover_out.IsIdle())
            std::this_thread::sleep_for(std::chrono::milliseconds(100));
    } catch (const std::runtime_error &e) {
        if (logger)
            logger->ErrorException(e);
        throw e;
    }
    if (logger)
        logger->Info("Packets Eth {} UDP {} SLS {} Proc {}", eth_packets, udp_packets, sls_packets, packets_processed);
    idle = true;
}

DataCollectionStatus HLSSimulatedDevice::GetDataCollectionStatus() const {
    DataCollectionStatus status{};

    union {
        uint64_t u64;
        double d;
    } pulse_id_conv;
    pulse_id_conv.u64 = current_pulse_id;

    status.ctrl_reg = ap_uint<1>(host_writer_idle) ? (1 << 4) : 0;
    status.max_modules = max_modules;
    status.hbm_size_bytes = hbm_if_size;
    status.run_counter = run_counter;
    status.current_pulseid = pulse_id_conv.d;
    status.packets_udp = udp_packets;
    status.packets_eth = eth_packets;
    status.packets_sls = sls_packets;

    return status;
}

void HLSSimulatedDevice::HW_LoadCalibration(const LoadCalibrationConfig &config) {
    int ret = load_calibration(hbm.data(), hbm.data(),
                               config,
                               hbm_if_size,
                               datamover_in.GetCtrlStream(),
                               datamover_in.GetDataStream(),
                               dma_address_table.data());
    if (ret)
        throw JFJochException(JFJochExceptionCategory::AcquisitionDeviceError,
                              "Error in loading calibration " + std::to_string(ret));

    if (!datamover_in.GetDataStream().empty())
        throw std::runtime_error("Datamover queue is not empty");
}

void HLSSimulatedDevice::HW_SetSpotFinderParameters(const SpotFinderParameters &params) {
    spot_finder_parameters = params;
}

void HLSSimulatedDevice::HW_SetDataSource(uint32_t val) {
    data_source = val;
}

uint32_t HLSSimulatedDevice::HW_GetDataSource() {
    return data_source;
}

void HLSSimulatedDevice::CreateXfelBunchIDPacket(double pulse_id, uint32_t event_code) {

    union {
        uint64_t u64;
        double d;
    } pulse_id_conv;
    pulse_id_conv.d = pulse_id;

    bunchid_raw_packet packet{};

    packet.ether_type = htons(0x0800);
    for (int i = 0; i < 6; i++)
        packet.dest_mac[i] = 0xFF;

    packet.ipv4_header_h = htons(0x4500); // Big endian in IP header!
    packet.ipv4_header_total_length = htons(sizeof(bunchid_payload) + 8 + 20); // Big endian in IP header!
    packet.ipv4_header_dest_ip = UINT32_MAX;

    packet.ipv4_header_ttl_protocol = htons(0x0011);
    packet.ipv4_header_checksum = checksum( (uint16_t *) &packet.ipv4_header_h, 20); // checksum is already in network order

    packet.udp_length = htons(sizeof(bunchid_payload) + 8);
    packet.payload.magicn[0] = BUNCHID_MAGICN;
    packet.payload.magicn[1] = BUNCHID_MAGICN;
    packet.payload.magicn[2] = BUNCHID_MAGICN;
    packet.payload.bunchid_msb[0] = (pulse_id_conv.u64 >> 32) & UINT32_MAX;
    packet.payload.bunchid_msb[1] = (pulse_id_conv.u64 >> 32) & UINT32_MAX;
    packet.payload.bunchid_lsb[0] = pulse_id_conv.u64 & UINT32_MAX;
    packet.payload.bunchid_lsb[1] = pulse_id_conv.u64 & UINT32_MAX;
    packet.payload.montrig = event_code;

    SendPacket((char *) &packet, sizeof(bunchid_raw_packet));
}