#include <atomic>
#include <chrono>
#include <cstdint>
#include <cstring>
#include <exception>
#include <limits>

#include <mxl/flow.h>
#include <mxl/time.h>

#include "NodeBase.hpp"
#include "ST2110Receiver.hpp"
#include "V210.hpp"

namespace dmf {

namespace {

class RtpIndexMapper {
public:
    RtpIndexMapper(const mxlRational& rate, int latency_frames)
        : rate_(rate), latency_frames_(latency_frames) {}

    uint64_t index_for(uint32_t rtp_timestamp, uint64_t current_mxl_index) {
        const uint64_t rtp_ext = extend_rtp(rtp_timestamp);
        if (!anchored_) {
            anchored_ = true;
            base_rtp_ext_ = rtp_ext;
            base_mxl_index_ = current_mxl_index + static_cast<uint64_t>(latency_frames_);
            last_offset_ = 0;
            return base_mxl_index_;
        }

        const uint64_t rtp_delta = rtp_ext - base_rtp_ext_;
        const uint64_t denominator = 90'000ULL * static_cast<uint64_t>(rate_.denominator);
        const uint64_t numerator = rtp_delta * static_cast<uint64_t>(rate_.numerator);
        const uint64_t offset = (numerator + denominator / 2) / denominator;
        if (offset <= last_offset_) {
            duplicate_or_backwards_++;
            last_gap_frames_ = 0;
        } else if (offset > last_offset_ + 1) {
            last_gap_frames_ = offset - (last_offset_ + 1);
            rtp_gap_frames_ += last_gap_frames_;
        } else {
            last_gap_frames_ = 0;
        }
        last_offset_ = offset;
        return base_mxl_index_ + offset;
    }

    uint32_t last_rtp_delta() const { return last_rtp_delta_; }
    uint32_t expected_rtp_delta() const {
        return static_cast<uint32_t>(
            (90'000ULL * static_cast<uint64_t>(rate_.denominator)) /
            static_cast<uint64_t>(rate_.numerator));
    }
    uint64_t last_gap_frames() const { return last_gap_frames_; }
    uint64_t rtp_gap_frames() const { return rtp_gap_frames_; }
    uint64_t duplicate_or_backwards() const { return duplicate_or_backwards_; }
    uint64_t base_mxl_index() const { return base_mxl_index_; }

private:
    uint64_t extend_rtp(uint32_t rtp_timestamp) {
        if (!have_last_rtp_) {
            have_last_rtp_ = true;
            last_rtp_ = rtp_timestamp;
            return rtp_timestamp;
        }

        if (rtp_timestamp < last_rtp_ &&
            static_cast<uint32_t>(last_rtp_ - rtp_timestamp) > 0x80000000u) {
            rtp_cycles_ += 0x1'0000'0000ULL;
        }
        last_rtp_delta_ = rtp_timestamp - last_rtp_;
        last_rtp_ = rtp_timestamp;
        return rtp_cycles_ + rtp_timestamp;
    }

    mxlRational rate_{};
    int latency_frames_ = 2;
    bool anchored_ = false;
    bool have_last_rtp_ = false;
    uint32_t last_rtp_ = 0;
    uint64_t rtp_cycles_ = 0;
    uint64_t base_rtp_ext_ = 0;
    uint64_t base_mxl_index_ = 0;
    uint64_t last_offset_ = 0;
    uint64_t rtp_gap_frames_ = 0;
    uint64_t duplicate_or_backwards_ = 0;
    uint64_t last_gap_frames_ = 0;
    uint32_t last_rtp_delta_ = 0;
};

} // namespace

class ST2110In : public NodeBase {
    void run() override {
        ST2110ReceiverConfig cfg;
        try {
            cfg = parse_st2110_receiver_config(config());
        } catch (const std::exception& e) {
            log("config error: %s", e.what());
            return;
        }

        log("SMPTE 2110-20 RX %s:%u from %s on %s backend=%s local=%s %dx%d depth=%d @ %d/%d latency=%d index_mode=%s",
            cfg.mcast_ip.c_str(), cfg.udp_port, cfg.source_ip.c_str(), cfg.ifname.c_str(),
            cfg.backend.c_str(), cfg.local_ip.c_str(), cfg.width, cfg.height, cfg.depth,
            cfg.fps_num, cfg.fps_den, cfg.mxl_latency_frames, cfg.mxl_index_mode.c_str());

        try {
            MTLContext mtl(cfg);
            ST20RxSession rx(mtl.get(), mtl.port_name(), cfg);
            MXLVideoWriter writer(instance(), cfg, node_id());

            const mxlRational video_rate = {cfg.fps_num, cfg.fps_den};
            const uint32_t video_stride = writer.config().discrete.sliceSizes[0];
            const uint32_t uyvy_stride = static_cast<uint32_t>(cfg.width * 2);
            const auto expected_fmt = cfg.output_fmt();
            RtpIndexMapper rtp_mapper(video_rate, cfg.mxl_latency_frames);
            uint64_t last_published_index = std::numeric_limits<uint64_t>::max();
            uint64_t frames_written = 0;
            uint64_t incomplete_frames = 0;
            uint64_t unexpected_format_frames = 0;
            uint64_t mxl_open_failures = 0;
            uint64_t index_resyncs = 0;
            uint64_t skipped_indices = 0;
            uint64_t backwards_indices = 0;
            uint64_t last_report_frames = 0;
            auto last_report = std::chrono::steady_clock::now();

            log("MTL version=%s port=%s output_stride=%u input_stride=%u direct_v210=%d",
                mtl_version(), mtl.port_name().c_str(), video_stride,
                cfg.direct_v210() ? video_stride : uyvy_stride, cfg.direct_v210() ? 1 : 0);

            while (g_running.load(std::memory_order_relaxed)) {
                st_frame* frame = st20p_rx_get_frame(rx.get());
                if (!frame) {
                    continue;
                }

                if (!st_is_frame_complete(frame->status)) {
                    incomplete_frames++;
                    st20p_rx_put_frame(rx.get(), frame);
                    continue;
                }

                if (frame->fmt != expected_fmt) {
                    unexpected_format_frames++;
                    log("unexpected MTL frame fmt=%d; expected %d", frame->fmt, expected_fmt);
                    st20p_rx_put_frame(rx.get(), frame);
                    continue;
                }

                mxlGrainInfo grain{};
                uint8_t* video_buf = nullptr;
                const uint64_t current_index = mxlGetCurrentIndex(&video_rate);
                uint64_t video_index = 0;
                if (cfg.mxl_index_mode == "rtp") {
                    video_index = rtp_mapper.index_for(frame->rtp_timestamp, current_index);
                    if (rtp_mapper.last_gap_frames() > 0) {
                        log("RTP gap frame=%llu rtp=%u delta=%u expected=%u gap_frames=%llu",
                            static_cast<unsigned long long>(frames_written),
                            frame->rtp_timestamp,
                            rtp_mapper.last_rtp_delta(),
                            rtp_mapper.expected_rtp_delta(),
                            static_cast<unsigned long long>(rtp_mapper.last_gap_frames()));
                    }
                } else {
                    video_index = current_index + static_cast<uint64_t>(cfg.mxl_latency_frames);
                }

                if (last_published_index != std::numeric_limits<uint64_t>::max()) {
                    if (video_index <= last_published_index) {
                        backwards_indices++;
                        video_index = last_published_index + 1;
                    } else if (video_index > last_published_index + 1) {
                        skipped_indices += video_index - (last_published_index + 1);
                    }
                }

                mxlStatus st = mxlFlowWriterOpenGrain(writer.get(), video_index, &grain, &video_buf);
                if (st != MXL_STATUS_OK) {
                    mxl_open_failures++;
                    log("mxlFlowWriterOpenGrain failed (%s) index=%llu current=%llu", mxl_status_str(st),
                        static_cast<unsigned long long>(video_index),
                        static_cast<unsigned long long>(current_index));
                    st20p_rx_put_frame(rx.get(), frame);
                    continue;
                }

                if (last_published_index != std::numeric_limits<uint64_t>::max() &&
                    video_index > last_published_index + 1) {
                    index_resyncs++;
                    log("MXL index resync last=%llu next=%llu current=%llu",
                        static_cast<unsigned long long>(last_published_index),
                        static_cast<unsigned long long>(video_index),
                        static_cast<unsigned long long>(current_index));
                }

                if (cfg.direct_v210()) {
                    const auto* src = static_cast<const uint8_t*>(frame->addr[0]);
                    const size_t src_stride = frame->linesize[0] ? frame->linesize[0]
                                                                 : static_cast<size_t>(cfg.width / 6) * 16;
                    const size_t row_bytes = static_cast<size_t>(cfg.width / 6) * 16;
                    for (int y = 0; y < cfg.height; ++y) {
                        std::memcpy(video_buf + static_cast<size_t>(y) * video_stride,
                                    src + static_cast<size_t>(y) * src_stride, row_bytes);
                    }
                } else {
                    const uint32_t src_stride = frame->linesize[0]
                        ? static_cast<uint32_t>(frame->linesize[0])
                        : uyvy_stride;
                    v210::UYVYtoV210(static_cast<const uint8_t*>(frame->addr[0]), video_buf,
                                     cfg.width, cfg.height, src_stride, video_stride);
                }
                grain.flags = 0;
                grain.validSlices = grain.totalSlices;
                mxlFlowWriterCommitGrain(writer.get(), &grain);
                st20p_rx_put_frame(rx.get(), frame);

                frames_written++;
                last_published_index = video_index;
                const auto now = std::chrono::steady_clock::now();
                const auto elapsed = std::chrono::duration<double>(now - last_report).count();
                if (elapsed >= 5.0) {
                    const uint64_t delta = frames_written - last_report_frames;
                    const double measured_fps = static_cast<double>(delta) / elapsed;
                    log("stats frames=%llu fps=%.2f incomplete=%llu bad_fmt=%llu mxl_open_fail=%llu resync=%llu skipped=%llu backwards=%llu rtp_gap=%llu rtp_dup=%llu last_index=%llu current=%llu",
                        static_cast<unsigned long long>(frames_written), measured_fps,
                        static_cast<unsigned long long>(incomplete_frames),
                        static_cast<unsigned long long>(unexpected_format_frames),
                        static_cast<unsigned long long>(mxl_open_failures),
                        static_cast<unsigned long long>(index_resyncs),
                        static_cast<unsigned long long>(skipped_indices),
                        static_cast<unsigned long long>(backwards_indices),
                        static_cast<unsigned long long>(rtp_mapper.rtp_gap_frames()),
                        static_cast<unsigned long long>(rtp_mapper.duplicate_or_backwards()),
                        static_cast<unsigned long long>(last_published_index),
                        static_cast<unsigned long long>(mxlGetCurrentIndex(&video_rate)));
                    last_report = now;
                    last_report_frames = frames_written;
                }
            }

            log("stopped frames=%llu incomplete=%llu bad_fmt=%llu mxl_open_fail=%llu resync=%llu skipped=%llu backwards=%llu rtp_gap=%llu rtp_dup=%llu",
                static_cast<unsigned long long>(frames_written),
                static_cast<unsigned long long>(incomplete_frames),
                static_cast<unsigned long long>(unexpected_format_frames),
                static_cast<unsigned long long>(mxl_open_failures),
                static_cast<unsigned long long>(index_resyncs),
                static_cast<unsigned long long>(skipped_indices),
                static_cast<unsigned long long>(backwards_indices),
                static_cast<unsigned long long>(rtp_mapper.rtp_gap_frames()),
                static_cast<unsigned long long>(rtp_mapper.duplicate_or_backwards()));
        } catch (const std::exception& e) {
            log("error: %s", e.what());
        }
    }
};

} // namespace dmf

int main() {
    dmf::ST2110In node;
    return node.execute();
}