// Copyright (2019-2023) Paul Scherrer Institute

#include "hls_jfjoch.h"
#include "ip_header_checksum.h"

void generate_packet(STREAM_512 &data_out,
                     ap_uint<256> *d_hbm_p0,
                     ap_uint<256> *d_hbm_p1,
                     ap_uint<32> hbm_size_bytes,
                     ap_uint<32> frame,
                     ap_uint<8> module,
                     ap_uint<7> eth_packet,
                     ap_uint<48> src_mac_addr,
                     ap_uint<48> dest_mac_addr,
                     ap_uint<32> src_ipv4_addr,
                     ap_uint<32> dest_ipv4_addr,
                     ap_uint<64> bunchid,
                     ap_uint<32> exptime,
                     ap_uint<64> timestamp,
                     ap_uint<32> debug) {
#pragma HLS PIPELINE II=130
    ap_uint<720> header = 0;

    header(47 ,  0) = dest_mac_addr;
    header(95 , 48) = src_mac_addr;
    header(111, 96) = 0x0008; // ETHER_IP = IPv4
    header(eth_payload_pos+3  , eth_payload_pos    ) = 0x5; // header len of 5
    header(eth_payload_pos+7  , eth_payload_pos+4  ) = 0x4; // IPv4
    header(eth_payload_pos+31 , eth_payload_pos+16 ) = 0x4C20; // total length = 8268
    header(eth_payload_pos+79 , eth_payload_pos+72 ) = PROTOCOL_UDP; // UDP
    header(eth_payload_pos+127, eth_payload_pos+96 ) = src_ipv4_addr;
    header(eth_payload_pos+159, eth_payload_pos+128) = dest_ipv4_addr;
    header(eth_payload_pos+95 , eth_payload_pos+80 ) = computeCheckSum20B(header(eth_payload_pos + 159, eth_payload_pos));

    header(ipv4_payload_pos+47, ipv4_payload_pos+32) = 0x3820; // UDP length = 8248
    header(udp_payload_pos+63, udp_payload_pos)  = frame + 1;
    header(udp_payload_pos+95, udp_payload_pos+64)  = exptime;
    header(udp_payload_pos+127, udp_payload_pos+96) = eth_packet;
    header(udp_payload_pos+2*64+63, udp_payload_pos+2*64)  = bunchid;
    header(udp_payload_pos+3*64+63, udp_payload_pos+3*64)  = timestamp;
    header(udp_payload_pos+4*64+31, udp_payload_pos+4*64+16) = 2 * module;
    header(udp_payload_pos+5*64+31, udp_payload_pos+5*64) = debug;

    packet_512_t packet;
    packet.data = header(511 ,0);
    packet.last = 0;
    packet.dest = 0;
    packet.id = 0;
    packet.strb = UINT64_MAX;
    packet.keep = UINT64_MAX;
    packet.user = 0;

    data_out << packet;

    ap_uint<32> offset_hbm_0         = 20 * hbm_size_bytes / 32;
    ap_uint<32> offset_hbm_1         = 21 * hbm_size_bytes / 32;

    ap_uint<208> remainder = header(719, 512);
    for (int i = 0; i < 128; i++) {
        ap_uint<512> tmp;
        tmp(255,   0) = d_hbm_p0[offset_hbm_0 + (module % FRAME_GENERATOR_MODULES) * 128 * 128 + eth_packet * 128 + i];
        tmp(511, 256) = d_hbm_p1[offset_hbm_1 + (module % FRAME_GENERATOR_MODULES) * 128 * 128 + eth_packet * 128 + i];

        packet.data(207,   0) = remainder;
        packet.data(511, 208) = tmp(303, 0);
        data_out << packet;
        remainder = tmp(511, 304);
    }
    packet.data(207,   0) = remainder;
    packet.data(511, 208) = 0;
    packet.keep(63,   26) = 0;
    packet.last = 1;
    data_out << packet;
}

void frame_generator(STREAM_512 &data_out,
                     ap_uint<256> *d_hbm_p0,
                     ap_uint<256> *d_hbm_p1,
                     ap_uint<32> hbm_size_bytes,
                     ap_uint<32> frames,
                     ap_uint<5> modules,
                     ap_uint<48> src_mac_addr,
                     ap_uint<48> dest_mac_addr,
                     ap_uint<32> src_ipv4_addr,
                     ap_uint<32> dest_ipv4_addr,
                     ap_uint<64> bunchid,
                     ap_uint<32> exptime,
                     ap_uint<32> debug) {
#pragma HLS INTERFACE mode=s_axilite port=return
#pragma HLS INTERFACE mode=s_axilite port=frames
#pragma HLS INTERFACE mode=s_axilite port=modules
#pragma HLS INTERFACE mode=s_axilite port=bunchid
#pragma HLS INTERFACE mode=s_axilite port=exptime
#pragma HLS INTERFACE mode=s_axilite port=debug
#pragma HLS INTERFACE mode=s_axilite port=dest_mac_addr
#pragma HLS INTERFACE mode=s_axilite port=dest_ipv4_addr
#pragma HLS INTERFACE mode=ap_none port=src_mac_addr
#pragma HLS INTERFACE mode=ap_none port=src_ipv4_addr

#pragma HLS INTERFACE register ap_none port=hbm_size_bytes
#pragma HLS INTERFACE m_axi port=d_hbm_p0 bundle=d_hbm_p0 depth=512 offset=off \
        max_read_burst_length=16 max_write_burst_length=2 latency=120 num_write_outstanding=2 num_read_outstanding=9
#pragma HLS INTERFACE m_axi port=d_hbm_p1 bundle=d_hbm_p1 depth=512 offset=off \
        max_read_burst_length=16 max_write_burst_length=2 latency=120 num_write_outstanding=2 num_read_outstanding=9

#pragma HLS INTERFACE register both axis port=data_out
    for (uint32_t f = 0; f < frames; f++) {
        for (uint32_t p = 0; p < 128; p++) {
            for (uint32_t m = 0; m < modules; m++) {
                generate_packet(data_out,
                                d_hbm_p0,
                                d_hbm_p1,
                                hbm_size_bytes,
                                f, m, p,
                                src_mac_addr,
                                dest_mac_addr,
                                src_ipv4_addr,
                                dest_ipv4_addr,
                                bunchid + f,
                                exptime,
                                exptime * f,
                                debug);
            }
        }
    }
}