// SPDX-FileCopyrightText: 2024 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: CERN-OHL-S-2.0

#include "hls_jfjoch.h"

#ifdef JFJOCH_HLS_NOSYNTH
#include <thread>
#endif

void spot_finder_in_stream(STREAM_768 &data_in,
                           hls::stream<spot_finder_packet> &data_out,
                           volatile ap_int<32> &in_count_threshold,
                           volatile ap_uint<32> &in_snr_threshold) {
    packet_768_t packet_in;
    {
#pragma HLS PROTOCOL fixed
        data_in >> packet_in;
        ap_wait();
        data_out << spot_finder_packet{
            .data = packet_in.data,
            .strb = packet_in.strb,
            .mask = 0,
            .user = packet_in.user,
            .last = packet_in.last
        };
        ap_wait();
    }
    data_in >> packet_in;
    while (!packet_in.user) {
#pragma HLS PIPELINE II=1

        data_out << spot_finder_packet{
                .data = packet_in.data,
                .strb = packet_in.strb,
                .mask = packet_in.keep(31,0),
                .strong_pixel = packet_in.keep(63,32),
                .count_threshold = in_count_threshold,
                .snr_threshold = in_snr_threshold,
                .user = packet_in.user,
                .last = packet_in.last
        };
        data_in >> packet_in;
    }
    data_out << spot_finder_packet{
        .data = packet_in.data,
        .strb = packet_in.strb,
        .mask = 0,
        .user = packet_in.user,
        .last = packet_in.last
    };
}

void spot_finder_out_stream(hls::stream<spot_finder_packet> &data_in,
                            STREAM_768 &data_out,
                            hls::stream<ap_axiu<32,1,1,1>> &strong_pixel_out) {
    spot_finder_packet packet_in;
    {
#pragma HLS PROTOCOL fixed
        data_in >> packet_in;
        ap_wait();
        data_out << packet_768_t{.data = packet_in.data, .strb = packet_in.strb, .user = packet_in.user, .last = packet_in.last};
        ap_wait();
    }

    data_in >> packet_in;
    while (!packet_in.user) {
        ap_uint<32> count_threshold = packet_in.count_threshold;
        ap_uint<32> snr_threshold_u32 = packet_in.snr_threshold;
        for (int i = 0; i < RAW_MODULE_SIZE * sizeof(uint16_t) / 64; i++) {
#pragma HLS PIPELINE II=1

            ap_uint<64> keep;
            keep(31, 0) = packet_in.mask;
            keep(63,32) = packet_in.strong_pixel;

            data_out << packet_768_t{.data = packet_in.data, .keep = keep, .strb = packet_in.strb, .user = packet_in.user, .last = packet_in.last};
            strong_pixel_out << ap_axiu<32,1,1,1>{.data = packet_in.strong_pixel, .user = 0};
            data_in >> packet_in;
        }

        // Save module statistics
        strong_pixel_out << ap_axiu<32,1,1,1>{.data = count_threshold, .user = 0};
        strong_pixel_out << ap_axiu<32,1,1,1>{.data = snr_threshold_u32, .user = 0};
        for (int i = 0; i < 14;i++)
            strong_pixel_out << ap_axiu<32,1,1,1>{.data =  0, .user = 0};
    }
    strong_pixel_out << ap_axiu<32,1,1,1>{.data = 0, .user = 1};
    data_out << packet_768_t{.data = packet_in.data, .strb = packet_in.strb, .user = packet_in.user, .last = packet_in.last};
}

ap_int<SUM_BITWIDTH> calc_sum(ap_int<24> val[32], ap_uint<32> mask) {
#pragma HLS PIPELINE II=1
    ap_int<SUM_BITWIDTH> ret = 0;
    for (int i = 0; i < 32; i++) {
        if (mask[i])
            ret += val[i];
    }
    return ret;
}

ap_int<SUM2_BITWIDTH> calc_sum2(ap_int<24> val[32], ap_uint<32> mask) {
#pragma HLS PIPELINE II=1
    ap_int<SUM2_BITWIDTH> ret = 0;
    for (int i = 0; i < 32; i++) {
        if (mask[i])
            ret += val[i] * val[i];
    }
    return ret;
}

ap_int<VALID_BITWIDTH> calc_valid(ap_uint<32> mask) {
#pragma HLS PIPELINE II=1
    ap_int<VALID_BITWIDTH> ret = 0;
    for (int i = 0; i < 32; i++) {
        ret += mask[i];
    }
    return ret;
}

void spot_finder_prepare(hls::stream<spot_finder_packet> &data_in,
                         hls::stream<spot_finder_packet> &data_out,
                         hls::stream<ap_int<SUM_BITWIDTH>> &sum_out,
                         hls::stream<ap_int<SUM2_BITWIDTH>> &sum2_out,
                         hls::stream<ap_int<VALID_BITWIDTH>> &valid_out) {
    ap_uint<32> strong_pixel;

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

    ap_int<SUM_BITWIDTH> sum[32];
    ap_int<SUM2_BITWIDTH> sum2[32];
    ap_int<VALID_BITWIDTH> valid[32];

    ap_int<SUM_BITWIDTH> sum_old[32][LINES_PER_GO * 2 + 1];
    ap_int<SUM2_BITWIDTH> sum2_old[32][LINES_PER_GO * 2 + 1];
    ap_int<VALID_BITWIDTH> valid_old[32][LINES_PER_GO * 2 + 1];

    data_in >> packet;

    while (!packet.user) {
        int line_in_mem = 0;

        for (int j = 0; j < 32*32*16; j++) {
#pragma HLS PIPELINE II=1
            data_out << packet;

            int col = j % 32;
            int line = j / 32;

            ap_int<24> val[32];
            unpack32(packet.data, val);
            ap_uint<32> mask = packet.mask;

            auto tmp_sum = calc_sum(val, mask);
            auto tmp_sum2 = calc_sum2(val, mask);
            auto tmp_valid = calc_valid(mask);

            auto tmp2_sum = tmp_sum;
			auto tmp2_sum2 = tmp_sum2;
			auto tmp2_valid = tmp_valid;

            if ((line > 0) && (line < LINES_PER_GO * 2 + 1)) {
                tmp2_sum += sum[col];
                tmp2_sum2 += sum2[col];
                tmp2_valid += valid[col];
            } else if (line >= LINES_PER_GO * 2 + 1) {
            	tmp2_sum += sum[col] - sum_old[col][line_in_mem];
            	tmp2_sum2 += sum2[col] - sum2_old[col][line_in_mem];
                tmp2_valid += valid[col] - valid_old[col][line_in_mem];
            }

            if (line >= LINES_PER_GO) {
                sum_out << tmp2_sum;
                sum2_out << tmp2_sum2;
                valid_out << tmp2_valid;
            }

            sum[col] = tmp2_sum;
			sum2[col] = tmp2_sum2;
			valid[col] = tmp2_valid;

            sum_old[col][line_in_mem] = tmp_sum;
            sum2_old[col][line_in_mem] = tmp_sum2;
            valid_old[col][line_in_mem] = tmp_valid;
            
            if (col == 31) {
                if (line_in_mem >= LINES_PER_GO * 2 + 1 - 1)
                    line_in_mem = 0;
                else
                    ++line_in_mem;
            }

            data_in >> packet;
        }

        for (int i = 0; i < LINES_PER_GO * 32; i++) {
#pragma HLS PIPELINE II=1
            int col = i % 32;

            sum[col] -= sum_old[col][line_in_mem];
            sum2[col] -= sum2_old[col][line_in_mem];
            valid[col] -= valid_old[col][line_in_mem];
            sum_out << sum[col];
            sum2_out << sum2[col];
            valid_out << valid[col];

            if (col == 31) {
                if (line_in_mem >= LINES_PER_GO * 2 + 1 - 1)
                    line_in_mem = 0;
                else
                    ++line_in_mem;
            }
        }
    }

    data_out << packet;
}

ap_uint<32> spot_finder_snr_threshold(ap_int<24> val[32],
                                      ap_ufixed<20,16, AP_RND_CONV> snr_threshold_2,
                                      ap_int<SUM_BITWIDTH> sum,
                                      ap_int<SUM2_BITWIDTH> sum2,
                                      ap_int<VALID_BITWIDTH> valid_count) {
#pragma HLS PIPELINE II=1
    if (snr_threshold_2 == 0)
        return UINT32_MAX;

    ap_int<SUM2_BITWIDTH+12> variance = valid_count * sum2 - sum * sum; // This is variance * valid_count^2
    ap_fixed<SUM2_BITWIDTH+12+16+1, SUM2_BITWIDTH+12+16> threshold = variance * snr_threshold_2;
    // snr_threshold is in units of 0.25

    ap_uint<32> ret = 0;
    for (int j = 0; j < 32; j++) {
        ap_int<SUM_BITWIDTH+1> in_minus_mean = val[j] * valid_count - sum; // This is (pxl - mean) * valid_count

        // Aim is to compare pxl-mean with sqrt(variance) * threshold
        // however this would require sqrt and divisions, so
        // it is cheaper to compare ((pxl-mean) * valid_count)^2 with variance * valid_count^2 * threshold^2,
        // but need to make sure that (pxl - mean) is positive
        // Also assume that N ≈ (N-1)

        if ((in_minus_mean * in_minus_mean > threshold) &&
            (in_minus_mean > 0) &&
            (valid_count > 15 * 32 / 2)) // at least half of the pixels
            ret[j] = 1;
        else
            ret[j] = 0;
    }
    return ret;
}

ap_uint<32> spot_finder_count_threshold(ap_int<24> val[32], ap_int<32> &count_threshold) {
#pragma HLS PIPELINE II=1
    if (count_threshold <= 0)
        return UINT32_MAX;
    ap_uint<32> ret = 0;
    for (int j = 0; j < 32; j++) {
        if (val[j] > count_threshold)
            ret[j] = 1;
        else
            ret[j] = 0;
    }
    return ret;
}

void spot_finder_apply_threshold(hls::stream<spot_finder_packet> &data_in,
                                 hls::stream<spot_finder_packet> &data_out,
                                 hls::stream<ap_int<SUM_BITWIDTH>> &sum_in,
                                 hls::stream<ap_int<SUM2_BITWIDTH>> &sum2_in,
                                 hls::stream<ap_int<VALID_BITWIDTH>> &valid_in) {
    ap_uint<32> strong_pixel_prev;
    spot_finder_packet packet_in;
    {
#pragma HLS PROTOCOL fixed
        data_in >> packet_in;
        ap_wait();
        data_out << packet_in;
        ap_wait();
    }

    ap_int<SUM_BITWIDTH> sum;
    ap_int<SUM2_BITWIDTH> sum2;
    ap_int<VALID_BITWIDTH> valid;

    data_in >> packet_in;
    while (!packet_in.user) {
        ap_int<32> count_threshold = packet_in.count_threshold;
        ap_uint<32> snr_threshold_u32 = packet_in.snr_threshold;
        float_uint32 thr;
        thr.u = snr_threshold_u32;
        ap_ufixed<10,8, AP_RND_CONV> snr_threshold = thr.f;
        if (thr.f > 255)
            snr_threshold = 255;
        else if (thr.f <= 0)
            snr_threshold = 0;
        else
            snr_threshold = thr.f;
        ap_ufixed<20,16, AP_RND_CONV> snr_threshold_2 = snr_threshold * snr_threshold;
        ap_uint<32> strong_pixel_count = 0;
        for (int i = 0; i < RAW_MODULE_SIZE * sizeof(uint16_t) / 64; i++) {
#pragma HLS PIPELINE II=1

            sum_in >> sum;
            sum2_in >> sum2;
            valid_in >> valid;

            ap_int<24> data_unpacked[32];
            unpack32(packet_in.data, data_unpacked);

            ap_uint<32> strong_pixel = spot_finder_count_threshold(data_unpacked, count_threshold) &
                                       spot_finder_snr_threshold(data_unpacked, snr_threshold_2,
                                                                 sum, sum2, valid);

            if ((snr_threshold == 0) && (count_threshold <= 0))
                strong_pixel = 0;

            strong_pixel = strong_pixel & packet_in.mask;

            packet_in.mask = packet_in.mask & ~strong_pixel; // mask strong pixels
            packet_in.strong_pixel |= strong_pixel; // add strong pixels to the output

            data_out << packet_in;
            data_in >> packet_in;
        }
    }
    data_out << packet_in;
}

void spot_finder(STREAM_768 &data_in,
                 STREAM_768 &data_out,
                 hls::stream<ap_axiu<32,1,1,1>> &strong_pixel_out,
                 volatile ap_int<32> &in_count_threshold,
                 volatile ap_uint<32> &in_snr_threshold) {
#pragma HLS INTERFACE axis port=data_in
#pragma HLS INTERFACE axis port=data_out
#pragma HLS INTERFACE axis port=strong_pixel_out
#pragma HLS INTERFACE ap_none register port=in_count_threshold
#pragma HLS INTERFACE ap_none register port=in_snr_threshold
#pragma HLS DATAFLOW

    hls::stream<spot_finder_packet, 2> data_0;
    hls::stream<spot_finder_packet, 512> data_1; // 15*32 = 480 should be OK
    hls::stream<spot_finder_packet, 8> data_2;

#pragma HLS BIND_STORAGE variable=data_1 type=fifo impl=bram

    hls::stream<ap_int<SUM_BITWIDTH>, 32> sum_0;
    hls::stream<ap_int<SUM2_BITWIDTH>, 32> sum2_0;
    hls::stream<ap_int<VALID_BITWIDTH>, 32> valid_0;

#ifndef JFJOCH_HLS_NOSYNTH
    spot_finder_in_stream(data_in, data_0, in_count_threshold, in_snr_threshold);
    spot_finder_prepare(data_0, data_1, sum_0, sum2_0, valid_0);
    spot_finder_apply_threshold(data_1, data_2, sum_0, sum2_0, valid_0);
    spot_finder_out_stream(data_2, data_out, strong_pixel_out);
#else
    std::vector<std::thread> spot_finder_cores;
    spot_finder_cores.emplace_back([&] {spot_finder_in_stream(data_in,  data_0, in_count_threshold,
                                                              in_snr_threshold);});
    spot_finder_cores.emplace_back([&] {spot_finder_prepare(data_0, data_1, sum_0, sum2_0, valid_0);});
    spot_finder_cores.emplace_back([&] {spot_finder_apply_threshold(data_1, data_2, sum_0, sum2_0, valid_0);});
    spot_finder_cores.emplace_back([&] {spot_finder_out_stream(data_2, data_out, strong_pixel_out);});

    for (auto &i : spot_finder_cores)
        i.join();
#endif
}