// Copyright (2019-2023) Paul Scherrer Institute

#include "hls_jfjoch.h"

void eiger_reorder(STREAM_512 &data_in,
                   STREAM_512 &data_out,
                   hls::stream<axis_completion > &s_axis_completion,
                   hls::stream<axis_completion > &m_axis_completion) {
#pragma HLS INTERFACE ap_ctrl_none port=return
#pragma HLS INTERFACE register both axis port=data_in
#pragma HLS INTERFACE register both axis port=data_out
#pragma HLS INTERFACE register both axis port=m_axis_completion
#pragma HLS INTERFACE register both axis port=s_axis_completion

    packet_512_t packet;
    {
#pragma HLS PROTOCOL fixed
        data_in >> packet;
        ap_wait();
        data_out << packet;
        ap_wait();
    }

    ap_uint<512> memory_0[128];
#pragma HLS BIND_STORAGE variable=memory_0 type=ram_t2p impl=bram latency=3
    ap_uint<512> memory_1[128];
#pragma HLS BIND_STORAGE variable=memory_1 type=ram_t2p impl=bram latency=3

    packet_512_t packet_out;
    packet_out.user = 0;
    packet_out.last = 0;
    packet_out.id = 0;

    axis_completion cmpl;
    s_axis_completion >> cmpl;
    while (!cmpl.last) {
        m_axis_completion << cmpl;
        if (cmpl.detector_type == SLS_DETECTOR_TYPE_EIGER) {
            for (int i = 0; i < 16384; i++) {
#pragma HLS PIPELINE II=1
                data_in >> packet;

                size_t counter = i % 128;

                size_t col0 = (counter >= 64) ? 16 : 0;
                size_t line0 = (counter / 16) % 4;

                if (i < 8192)
                    line0 = 3 - line0; // second half is reversed

                size_t pos = line0 * 32 + col0 + counter % 16;
                if ((i / 128) % 2 == 0) {
                    memory_0[pos] = packet.data;
                    packet_out.data = memory_1[counter];
                } else {
                    memory_1[pos] = packet.data;
                    packet_out.data = memory_0[counter];
                }
                if (i >= 128)
                    data_out << packet_out;
            }

            for (int i = 0; i < 128; i++) {
#pragma HLS PIPELINE II=1
                packet_out.data = memory_1[i];
                if (i == 127)
                    packet_out.last = 1;
                data_out << packet_out;
            }
        } else {
            for (int i = 0; i < 16384; i++) {
#pragma HLS PIPELINE II=1
                data_in >> packet;
                data_out << packet;
            }
        }
        s_axis_completion >> cmpl;
    }
    m_axis_completion << cmpl;

    data_in >> packet;
    data_out << packet;
}