#include <algorithm>
#include <atomic>
#include <chrono>
#include <condition_variable>
#include <cstdlib>
#include <cstring>
#include <iostream>
#include <memory>
#include <mutex>
#include <thread>
#include <vector>

#include "node.h"
#include "utils/dma_alloc.h"
#include "utils/logger.h"

#if defined(RK3588_ENABLE_RGA)
#include "im2d.hpp"
#include "im2d_buffer.h"
#include "im2d_type.h"
#endif

#if defined(RK3588_ENABLE_FFMPEG)
extern "C" {
#include <libavutil/imgutils.h>
#include <libswscale/swscale.h>
}
#endif

namespace rk3588 {

namespace {

inline int Align16(int v) { return (v + 15) & ~15; }

#if defined(RK3588_ENABLE_RGA)
int ToRgaFormat(PixelFormat fmt) {
    switch (fmt) {
        case PixelFormat::NV12: return RK_FORMAT_YCbCr_420_SP;
        case PixelFormat::YUV420: return RK_FORMAT_YCbCr_420_P;
        case PixelFormat::RGB: return RK_FORMAT_RGB_888;
        case PixelFormat::BGR: return RK_FORMAT_BGR_888;
        default: return RK_FORMAT_UNKNOWN;
    }
}

std::atomic<int>& GlobalRgaMaxInflightRef() {
    static std::atomic<int> v{2};
    return v;
}

int GlobalRgaMaxInflight() {
    static std::once_flag once;
    std::call_once(once, []() {
        const char* s = std::getenv("RK3588_RGA_MAX_INFLIGHT");
        if (!s || !*s) return;
        try {
            int v = std::stoi(s);
            if (v > 0 && v <= 32) GlobalRgaMaxInflightRef().store(v);
        } catch (...) {
        }
    });
    int v = GlobalRgaMaxInflightRef().load();
    return v > 0 ? v : 1;
}

void SetGlobalRgaMaxInflight(int v) {
    if (v <= 0) return;
    if (v > 32) v = 32;
    GlobalRgaMaxInflightRef().store(v);
}

class RgaGate {
public:
    void Acquire() {
        const int max_inflight = GlobalRgaMaxInflight();
        std::unique_lock<std::mutex> lock(mu_);
        cv_.wait(lock, [&]() { return in_flight_ < max_inflight; });
        ++in_flight_;
    }

    void Release() {
        std::lock_guard<std::mutex> lock(mu_);
        if (in_flight_ > 0) {
            --in_flight_;
        }
        cv_.notify_one();
    }

private:
    std::mutex mu_;
    std::condition_variable cv_;
    int in_flight_ = 0;
};

RgaGate& GlobalRgaGate() {
    static RgaGate* g = new RgaGate();
    return *g;
}

class ScopedRgaGate {
public:
    ScopedRgaGate() { GlobalRgaGate().Acquire(); }
    ~ScopedRgaGate() { GlobalRgaGate().Release(); }
    ScopedRgaGate(const ScopedRgaGate&) = delete;
    ScopedRgaGate& operator=(const ScopedRgaGate&) = delete;
};

void EnsureRgaInitializedOnce() {
    static std::once_flag once;
    std::call_once(once, []() {
        const IM_STATUS st = imcheckHeader();
        if (st != IM_STATUS_NOERROR && st != IM_STATUS_SUCCESS) {
            std::cerr << "[preprocess] imcheckHeader failed: " << imStrError(st) << "\n";
        }
    });
}
#endif

PixelFormat ParseFormat(const std::string& s) {
    if (s == "nv12" || s == "NV12") return PixelFormat::NV12;
    if (s == "yuv420" || s == "YUV420") return PixelFormat::YUV420;
    if (s == "rgb" || s == "RGB") return PixelFormat::RGB;
    if (s == "bgr" || s == "BGR") return PixelFormat::BGR;
    return PixelFormat::UNKNOWN;
}

size_t CalcImageSize(int w, int h, PixelFormat fmt) {
    switch (fmt) {
        case PixelFormat::NV12:
        case PixelFormat::YUV420:
            return static_cast<size_t>(w) * h * 3 / 2;
        case PixelFormat::RGB:
        case PixelFormat::BGR:
            return static_cast<size_t>(w) * h * 3;
        default:
            return 0;
    }
}

size_t CalcImageSizeStrided(int wstride, int hstride, PixelFormat fmt) {
    if (wstride <= 0 || hstride <= 0) return 0;
    const size_t ws = static_cast<size_t>(wstride);
    const size_t hs = static_cast<size_t>(hstride);
    switch (fmt) {
        case PixelFormat::NV12: {
            // Y: ws*hs, UV: ws*(hs/2)
            return ws * hs + ws * (hs / 2);
        }
        case PixelFormat::YUV420: {
            // Y: ws*hs, U/V: (ws/2)*(hs/2)
            const size_t y = ws * hs;
            const size_t uv = (ws / 2) * (hs / 2);
            return y + uv + uv;
        }
        case PixelFormat::RGB:
        case PixelFormat::BGR:
            return ws * hs * 3;
        default:
            return 0;
    }
}

bool CopyToStridedBuffer(const Frame& src, uint8_t* dst, size_t dst_size,
                         int dst_wstride, int dst_hstride) {
    if (!dst || dst_size == 0) return false;
    if (dst_wstride <= 0 || dst_hstride <= 0) return false;

    std::memset(dst, 0, dst_size);

    const int w = src.width;
    const int h = src.height;
    if (w <= 0 || h <= 0) return false;

    if (src.format == PixelFormat::NV12) {
        const size_t y_bytes = static_cast<size_t>(dst_wstride) * dst_hstride;
        const size_t uv_bytes = static_cast<size_t>(dst_wstride) * (dst_hstride / 2);
        if (y_bytes + uv_bytes > dst_size) return false;

        const uint8_t* src_y = src.planes[0].data ? src.planes[0].data : src.data;
        const uint8_t* src_uv = src.planes[1].data ? src.planes[1].data : nullptr;
        const int src_y_stride = src.planes[0].stride > 0 ? src.planes[0].stride : w;
        const int src_uv_stride = src.planes[1].stride > 0 ? src.planes[1].stride : w;
        if (!src_y) return false;
        if (!src_uv) {
            // Fallback: packed NV12 layout.
            if (!src.data) return false;
            src_uv = src.data + static_cast<size_t>(src_y_stride) * static_cast<size_t>(h);
        }

        for (int row = 0; row < h; ++row) {
            std::memcpy(dst + static_cast<size_t>(row) * dst_wstride,
                        src_y + static_cast<size_t>(row) * src_y_stride,
                        static_cast<size_t>(w));
        }

        uint8_t* dst_uv = dst + y_bytes;
        const int uv_rows = h / 2;
        for (int row = 0; row < uv_rows; ++row) {
            std::memcpy(dst_uv + static_cast<size_t>(row) * dst_wstride,
                        src_uv + static_cast<size_t>(row) * src_uv_stride,
                        static_cast<size_t>(w));
        }
        return true;
    }

    if (src.format == PixelFormat::YUV420) {
        const size_t y_bytes = static_cast<size_t>(dst_wstride) * dst_hstride;
        const size_t uv_stride = static_cast<size_t>(dst_wstride) / 2;
        const size_t uv_h = static_cast<size_t>(dst_hstride) / 2;
        const size_t u_bytes = uv_stride * uv_h;
        const size_t v_bytes = u_bytes;
        if (y_bytes + u_bytes + v_bytes > dst_size) return false;

        const uint8_t* src_y = src.planes[0].data ? src.planes[0].data : src.data;
        const uint8_t* src_u = src.planes[1].data ? src.planes[1].data : nullptr;
        const uint8_t* src_v = src.planes[2].data ? src.planes[2].data : nullptr;
        const int src_y_stride = src.planes[0].stride > 0 ? src.planes[0].stride : w;
        const int src_u_stride = src.planes[1].stride > 0 ? src.planes[1].stride : (w / 2);
        const int src_v_stride = src.planes[2].stride > 0 ? src.planes[2].stride : (w / 2);
        if (!src_y || !src_u || !src_v) return false;

        for (int row = 0; row < h; ++row) {
            std::memcpy(dst + static_cast<size_t>(row) * dst_wstride,
                        src_y + static_cast<size_t>(row) * src_y_stride,
                        static_cast<size_t>(w));
        }

        uint8_t* dst_u = dst + y_bytes;
        uint8_t* dst_v = dst + y_bytes + u_bytes;
        const int uv_rows = h / 2;
        const int uv_cols = w / 2;
        for (int row = 0; row < uv_rows; ++row) {
            std::memcpy(dst_u + static_cast<size_t>(row) * uv_stride,
                        src_u + static_cast<size_t>(row) * src_u_stride,
                        static_cast<size_t>(uv_cols));
            std::memcpy(dst_v + static_cast<size_t>(row) * uv_stride,
                        src_v + static_cast<size_t>(row) * src_v_stride,
                        static_cast<size_t>(uv_cols));
        }
        return true;
    }

    if (src.format == PixelFormat::RGB || src.format == PixelFormat::BGR) {
        const size_t need = static_cast<size_t>(dst_wstride) * dst_hstride * 3;
        if (need > dst_size) return false;

        const uint8_t* src_rgb = src.planes[0].data ? src.planes[0].data : src.data;
        const int src_stride = src.planes[0].stride > 0
                                   ? src.planes[0].stride
                                   : (src.stride > 0 ? src.stride : w * 3);
        if (!src_rgb) return false;

        const size_t dst_stride = static_cast<size_t>(dst_wstride) * 3;
        const size_t row_bytes = static_cast<size_t>(w) * 3;
        for (int row = 0; row < h; ++row) {
            std::memcpy(dst + static_cast<size_t>(row) * dst_stride,
                        src_rgb + static_cast<size_t>(row) * src_stride,
                        row_bytes);
        }
        return true;
    }

    return false;
}

}  // namespace

class PreprocessNode : public INode {
public:
    std::string Id() const override { return id_; }
    std::string Type() const override { return "preprocess"; }

    bool Init(const SimpleJson& config, const NodeContext& ctx) override {
        id_ = config.ValueOr<std::string>("id", "preprocess");
        dst_w_ = config.ValueOr<int>("dst_w", 640);
        dst_h_ = config.ValueOr<int>("dst_h", 640);
        keep_ratio_ = config.ValueOr<bool>("keep_ratio", false);
        dst_packed_ = config.ValueOr<bool>("dst_packed", false);
        std::string fmt_str = config.ValueOr<std::string>("dst_format", "");
        if (!fmt_str.empty()) {
            dst_fmt_ = ParseFormat(fmt_str);
        }

#if defined(RK3588_ENABLE_RGA)
        const int rga_max_inflight = config.ValueOr<int>("rga_max_inflight", 0);
        if (rga_max_inflight > 0) {
            SetGlobalRgaMaxInflight(rga_max_inflight);
        }
#endif
        const bool requested_use_rga = config.ValueOr<bool>("use_rga", true);
        use_rga_ = requested_use_rga;

        if (const SimpleJson* dbg = config.Find("debug"); dbg && dbg->IsObject()) {
            stats_log_ = dbg->ValueOr<bool>("stats", stats_log_);
            stats_interval_ = std::max<uint64_t>(1, static_cast<uint64_t>(dbg->ValueOr<int>("stats_interval", static_cast<int>(stats_interval_))));
        }

        input_queue_ = ctx.input_queue;
        if (!input_queue_) {
            std::cerr << "[preprocess] no input queue for node " << id_ << "\n";
            return false;
        }
        if (ctx.output_queues.empty()) {
            std::cerr << "[preprocess] no output queue for node " << id_ << "\n";
            return false;
        }
        output_queues_ = ctx.output_queues;

#if !defined(RK3588_ENABLE_RGA)
        if (requested_use_rga) {
            std::cerr << "[preprocess] use_rga=true but RGA not enabled at build time\n";
            return false;
        }
        use_rga_ = false;
#endif
#if !defined(RK3588_ENABLE_FFMPEG)
        if (!use_rga_) {
            std::cerr << "[preprocess] neither RGA nor FFmpeg enabled\n";
            return false;
        }
#endif
        return true;
    }

    bool Start() override {
        LogInfo("[preprocess] start id=" + id_ + " dst=" + std::to_string(dst_w_) + "x" + std::to_string(dst_h_) +
                (use_rga_ ? " (rga)" : " (swscale)"));
        return true;
    }

    void Stop() override {
#if defined(RK3588_ENABLE_FFMPEG)
        if (sws_ctx_) {
            sws_freeContext(sws_ctx_);
            sws_ctx_ = nullptr;
        }
#endif
    }

    NodeStatus Process(FramePtr frame) override {
        if (!frame) return NodeStatus::DROP;

#if defined(RK3588_ENABLE_RGA)
        if (use_rga_) {
            if (!ProcessRga(frame)) {
                return NodeStatus::ERROR;
            }
        } else {
            ProcessSwscale(frame);
        }
#elif defined(RK3588_ENABLE_FFMPEG)
        ProcessSwscale(frame);
#else
        ProcessPassthrough(frame);
#endif
        return NodeStatus::OK;
    }

private:
    void PushToDownstream(FramePtr frame) {
        for (auto& q : output_queues_) {
            q->Push(frame);
        }
    }

    void WarnMetaResizeOnce(const FramePtr& frame, int out_w, int out_h) {
        if (warned_meta_resize_) return;
        if (!frame) return;
        if (frame->width == out_w && frame->height == out_h) return;
        if (!frame->det && !frame->face_det && !frame->face_recog) return;
        warned_meta_resize_ = true;
        LogWarn("[preprocess] resized frame but forwarded det/face meta without coordinate scaling; ensure det/recog/osd use same resolution (id=" + id_ + ")");
    }

    void ProcessPassthrough(FramePtr frame) {
        PushToDownstream(frame);
        ++processed_;
        if (stats_log_ && stats_interval_ > 0 && (processed_ % stats_interval_) == 0) {
            LogInfo("[preprocess] passthrough frame=" + std::to_string(frame->frame_id) + " id=" + id_);
        }
    }

#if defined(RK3588_ENABLE_RGA)
    bool ProcessRga(FramePtr frame) {
        EnsureRgaInitializedOnce();

        PixelFormat out_fmt = (dst_fmt_ != PixelFormat::UNKNOWN) ? dst_fmt_ : frame->format;
        int out_w = dst_w_;
        int out_h = dst_h_;

        if (keep_ratio_ && frame->width > 0 && frame->height > 0) {
            float scale = std::min(static_cast<float>(dst_w_) / frame->width,
                                   static_cast<float>(dst_h_) / frame->height);
            out_w = static_cast<int>(frame->width * scale);
            out_h = static_cast<int>(frame->height * scale);
            out_w = (out_w + 1) & ~1;
            out_h = (out_h + 1) & ~1;
        }

        int src_fmt_rga = ToRgaFormat(frame->format);
        int dst_fmt_rga = ToRgaFormat(out_fmt);
        bool need_cvt = (src_fmt_rga != dst_fmt_rga);
        bool need_resize = (frame->width != out_w || frame->height != out_h);

        if (need_resize) {
            WarnMetaResizeOnce(frame, out_w, out_h);
        }

        // If no processing needed, passthrough directly
        if (!need_cvt && !need_resize) {
            PushToDownstream(frame);
            ++processed_;
            if (stats_log_ && stats_interval_ > 0 && (processed_ % stats_interval_) == 0) {
                LogInfo("[preprocess] passthrough frame=" + std::to_string(frame->frame_id) +
                        " " + std::to_string(frame->width) + "x" + std::to_string(frame->height) + " (no change)" +
                        " id=" + id_);
            }
            return true;
        }

        // Calculate proper strides.
        // IMPORTANT: For DMA-BUF frames (e.g. MPP decode output), the actual vertical stride
        // (ver_stride) may be larger than `height`. We must honor that, otherwise RGA will
        // read UV from the wrong offset (典型花屏/错色).
        int src_wstride = Align16(frame->width);
        int src_hstride = Align16(frame->height);
        if (frame->format == PixelFormat::NV12 || frame->format == PixelFormat::YUV420) {
            const int y_stride = frame->planes[0].stride > 0 ? frame->planes[0].stride
                                : (frame->stride > 0 ? frame->stride : frame->width);
            if (y_stride > 0) src_wstride = y_stride;
            if (frame->planes[0].size > 0 && y_stride > 0) {
                const int hs = frame->planes[0].size / y_stride;
                if (hs >= frame->height) src_hstride = hs;
            }
        } else if (frame->format == PixelFormat::RGB || frame->format == PixelFormat::BGR) {
            const int stride_bytes = frame->planes[0].stride > 0 ? frame->planes[0].stride
                                   : (frame->stride > 0 ? frame->stride : frame->width * 3);
            if (stride_bytes > 0 && (stride_bytes % 3) == 0) {
                src_wstride = stride_bytes / 3;
            }
            if (frame->planes[0].size > 0 && stride_bytes > 0) {
                const int hs = frame->planes[0].size / stride_bytes;
                if (hs >= frame->height) src_hstride = hs;
            }
        }
        int dst_wstride = Align16(out_w);
        // For AI input (RGB/BGR), allow a tightly packed output to avoid an extra per-frame memcpy
        // in downstream nodes (e.g. ai_yolo).
        if (dst_packed_ && (out_fmt == PixelFormat::RGB || out_fmt == PixelFormat::BGR)) {
            dst_wstride = out_w;
        }
        int dst_hstride = Align16(out_h);

        if (src_fmt_rga == RK_FORMAT_UNKNOWN || dst_fmt_rga == RK_FORMAT_UNKNOWN) {
            std::cerr << "[preprocess] unsupported format for RGA\n";
            return false;
        }

        size_t out_size = CalcImageSizeStrided(dst_wstride, dst_hstride, out_fmt);
        if (out_size == 0) {
            std::cerr << "[preprocess] invalid output size for RGA\n";
            return false;
        }

        // Use DMA-BUF allocation to avoid >4GB address issue with RGA
        auto dma_buf = DmaAlloc(out_size);
        if (!dma_buf || !dma_buf->valid()) {
            std::cerr << "[preprocess] DMA alloc failed\n";
            return false;
        }

        if (stats_log_ && processed_ < 3) {
            LogInfo("[preprocess] src: " + std::to_string(frame->width) + "x" + std::to_string(frame->height) +
                    " fmt=" + std::to_string(static_cast<int>(frame->format)) +
                    " rga_fmt=" + std::to_string(src_fmt_rga) +
                    " wstride=" + std::to_string(src_wstride) +
                    " hstride=" + std::to_string(src_hstride) +
                    " data_size=" + std::to_string(frame->data_size));
            LogInfo("[preprocess] dst: " + std::to_string(out_w) + "x" + std::to_string(out_h) +
                    " fmt=" + std::to_string(static_cast<int>(out_fmt)) +
                    " rga_fmt=" + std::to_string(dst_fmt_rga) +
                    " wstride=" + std::to_string(dst_wstride) +
                    " hstride=" + std::to_string(dst_hstride) +
                    " out_size=" + std::to_string(out_size));
        }

        rga_buffer_t src_buf{};
        rga_buffer_t dst_buf{};
        DmaBufferPtr src_dma_buf;  // keep alive if we allocate/copy

        // Lazily allocate tmp buffer only when we must fall back to a 2-step pipeline.
        // (resize + cvtcolor). Prefer a single improcess() call to reduce scheduling overhead.
        DmaBufferPtr tmp_dma;

        const bool can_cpu_read_src = (frame->data != nullptr && frame->data_size > 0);
        auto CopySrcToDmaIfPossible = [&]() -> bool {
            if (!can_cpu_read_src) return false;
            // Use tight/aligned strides for the copied source buffer.
            const int copy_wstride = Align16(frame->width);
            const int copy_hstride = Align16(frame->height);
            const size_t src_size = CalcImageSizeStrided(copy_wstride, copy_hstride, frame->format);
            src_dma_buf = DmaAlloc(src_size);
            if (!src_dma_buf || !src_dma_buf->valid()) {
                return false;
            }

            // If source is a DMA-BUF, sync it before CPU reads.
            if (frame->dma_fd >= 0) {
                DmaSyncStartFd(frame->dma_fd);
            }

            // CPU writes to a DMA-BUF must be flushed before RGA reads it.
            DmaSyncStartFd(src_dma_buf->fd);
            const bool ok = CopyToStridedBuffer(*frame, src_dma_buf->data(), src_dma_buf->size,
                                                copy_wstride, copy_hstride);
            DmaSyncEndFd(src_dma_buf->fd);

            if (frame->dma_fd >= 0) {
                DmaSyncEndFd(frame->dma_fd);
            }

            if (!ok) {
                src_dma_buf.reset();
                return false;
            }

            // Update strides used for RGA to match the copied buffer.
            src_wstride = copy_wstride;
            src_hstride = copy_hstride;
            return true;
        };

        // If there's no DMA fd, we must allocate/copy into a DMA-BUF for RGA.
        if (frame->dma_fd < 0) {
            if (!CopySrcToDmaIfPossible()) {
                std::cerr << "[preprocess] no dma_fd and src copy failed\n";
                return false;
            }
        }

        auto RunRgaOnce = [&](int src_fd, std::string& err) -> bool {
            // Serialize/limit librga/im2d usage; multiple pipelines call RGA concurrently.
            ScopedRgaGate guard;

            src_buf = wrapbuffer_fd_t(src_fd, frame->width, frame->height,
                                      src_wstride, src_hstride, src_fmt_rga);
            dst_buf = wrapbuffer_fd_t(dma_buf->fd, out_w, out_h,
                                      dst_wstride, dst_hstride, dst_fmt_rga);

            auto Check = [&](const rga_buffer_t& s, const rga_buffer_t& d) -> bool {
            const im_rect sr{0, 0, s.width, s.height};
            const im_rect dr{0, 0, d.width, d.height};
            // Do NOT call the imcheck(...) variadic macro with 0 extra args:
            // under -Wpedantic/-std=c++11 it expands to a zero-sized array.
            rga_buffer_t pat{};
            const im_rect pr{0, 0, 0, 0};
            const IM_STATUS chk = imcheck_t(s, d, pat, sr, dr, pr, 0);
            if (chk != IM_STATUS_NOERROR && chk != IM_STATUS_SUCCESS) {
                err = std::string("RGA imcheck failed: ") + imStrError(chk);
                return false;
            }
            return true;
        };

            IM_STATUS status = IM_STATUS_SUCCESS;

            if (need_resize && need_cvt) {
                // Try to fuse resize + CSC in a single call.
                if (Check(src_buf, dst_buf)) {
                    rga_buffer_t pat{};
                    const im_rect sr{0, 0, src_buf.width, src_buf.height};
                    const im_rect dr{0, 0, dst_buf.width, dst_buf.height};
                    const im_rect pr{0, 0, 0, 0};
                    status = improcess(src_buf, dst_buf, pat, sr, dr, pr,
                                       0, nullptr, nullptr, IM_SYNC);
                    if (status == IM_STATUS_SUCCESS) {
                        return true;
                    }
                }

                // Fallback: 2-step (resize + cvtcolor).
                if (!tmp_dma || !tmp_dma->valid()) {
                    tmp_dma = DmaAlloc(CalcImageSizeStrided(dst_wstride, dst_hstride, frame->format));
                    if (!tmp_dma || !tmp_dma->valid()) {
                        err = "DMA alloc for tmp failed";
                        return false;
                    }
                }

                rga_buffer_t tmp = wrapbuffer_fd_t(tmp_dma->fd, out_w, out_h,
                                                   dst_wstride, dst_hstride, src_fmt_rga);
                if (!Check(src_buf, tmp) || !Check(tmp, dst_buf)) {
                    return false;
                }
                status = imresize(src_buf, tmp, 0, 0, 0, 1, nullptr);
                if (status == IM_STATUS_SUCCESS) {
                    status = imcvtcolor(tmp, dst_buf, src_fmt_rga, dst_fmt_rga, IM_COLOR_SPACE_DEFAULT, 1, nullptr);
                }
            } else if (need_resize) {
                if (!Check(src_buf, dst_buf)) {
                    return false;
                }
                status = imresize(src_buf, dst_buf, 0, 0, 0, 1, nullptr);
            } else if (need_cvt) {
                if (!Check(src_buf, dst_buf)) {
                    return false;
                }
                status = imcvtcolor(src_buf, dst_buf, src_fmt_rga, dst_fmt_rga, IM_COLOR_SPACE_DEFAULT, 1, nullptr);
            }

            if (status != IM_STATUS_SUCCESS) {
                err = std::string("RGA failed: ") + imStrError(status);
                return false;
            }
            return true;
        };

        std::string rga_err;

        const int src_fd = (src_dma_buf && src_dma_buf->valid()) ? src_dma_buf->fd : frame->dma_fd;
        if (src_fd < 0 || !RunRgaOnce(src_fd, rga_err)) {
            std::cerr << "[preprocess] " << (rga_err.empty() ? "RGA failed" : rga_err) << "\n";
            return false;
        }

        auto out_frame = std::make_shared<Frame>();
        out_frame->width = out_w;
        out_frame->height = out_h;
        out_frame->format = out_fmt;
        out_frame->stride = (out_fmt == PixelFormat::RGB || out_fmt == PixelFormat::BGR)
                                ? (dst_wstride * 3)
                                : dst_wstride;
        out_frame->dma_fd = dma_buf->fd;
        out_frame->data = dma_buf->data();
        out_frame->data_size = dma_buf->size;
        out_frame->data_owner = dma_buf;  // DmaBuffer shared_ptr keeps fd alive
        out_frame->pts = frame->pts;
        out_frame->frame_id = frame->frame_id;
        out_frame->det = frame->det;
        out_frame->face_det = frame->face_det;
        out_frame->face_recog = frame->face_recog;
        out_frame->user_meta = frame->user_meta;

        SetupPlanes(*out_frame, out_fmt);
        PushToDownstream(out_frame);
        ++processed_;

        if (stats_log_ && stats_interval_ > 0 && (processed_ % stats_interval_) == 0) {
            LogInfo("[preprocess] rga frame=" + std::to_string(out_frame->frame_id) +
                    " " + std::to_string(frame->width) + "x" + std::to_string(frame->height) +
                    " -> " + std::to_string(out_w) + "x" + std::to_string(out_h));
        }

        return true;
    }
#endif

#if defined(RK3588_ENABLE_FFMPEG)
    void ProcessSwscale(FramePtr frame) {
        PixelFormat out_fmt = (dst_fmt_ != PixelFormat::UNKNOWN) ? dst_fmt_ : frame->format;
        int out_w = dst_w_;
        int out_h = dst_h_;

        if (keep_ratio_ && frame->width > 0 && frame->height > 0) {
            float scale = std::min(static_cast<float>(dst_w_) / frame->width,
                                   static_cast<float>(dst_h_) / frame->height);
            out_w = static_cast<int>(frame->width * scale);
            out_h = static_cast<int>(frame->height * scale);
            out_w = (out_w + 1) & ~1;
            out_h = (out_h + 1) & ~1;
        }

        if (frame->width != out_w || frame->height != out_h) {
            WarnMetaResizeOnce(frame, out_w, out_h);
        }

        AVPixelFormat src_av_fmt = ToAvFormat(frame->format);
        AVPixelFormat dst_av_fmt = ToAvFormat(out_fmt);

        if (src_av_fmt == AV_PIX_FMT_NONE || dst_av_fmt == AV_PIX_FMT_NONE) {
            PushToDownstream(frame);
            return;
        }

        if (!sws_ctx_ || frame->width != last_src_w_ || frame->height != last_src_h_ ||
            src_av_fmt != last_src_fmt_ || dst_av_fmt != last_dst_fmt_) {
            if (sws_ctx_) sws_freeContext(sws_ctx_);
            sws_ctx_ = sws_getContext(frame->width, frame->height, src_av_fmt,
                                     out_w, out_h, dst_av_fmt,
                                     SWS_BILINEAR, nullptr, nullptr, nullptr);
            last_src_w_ = frame->width;
            last_src_h_ = frame->height;
            last_src_fmt_ = src_av_fmt;
            last_dst_fmt_ = dst_av_fmt;
        }

        if (!sws_ctx_) {
            PushToDownstream(frame);
            return;
        }

        size_t out_size = CalcImageSize(out_w, out_h, out_fmt);
        auto buffer = std::make_shared<std::vector<uint8_t>>(out_size);

        uint8_t* src_data[4] = {nullptr};
        int src_linesize[4] = {0};
        uint8_t* dst_data[4] = {nullptr};
        int dst_linesize[4] = {0};

        SetupAvPlanes(frame.get(), src_data, src_linesize);
        av_image_fill_arrays(dst_data, dst_linesize, buffer->data(),
                             dst_av_fmt, out_w, out_h, 1);

        sws_scale(sws_ctx_, src_data, src_linesize, 0, frame->height,
                  dst_data, dst_linesize);

        auto out_frame = std::make_shared<Frame>();
        out_frame->width = out_w;
        out_frame->height = out_h;
        out_frame->format = out_fmt;
        out_frame->stride = (out_fmt == PixelFormat::RGB || out_fmt == PixelFormat::BGR)
                                ? (out_w * 3)
                                : out_w;
        out_frame->data = buffer->data();
        out_frame->data_size = buffer->size();
        out_frame->data_owner = buffer;
        out_frame->pts = frame->pts;
        out_frame->frame_id = frame->frame_id;
        out_frame->det = frame->det;
        out_frame->face_det = frame->face_det;
        out_frame->face_recog = frame->face_recog;
        out_frame->user_meta = frame->user_meta;

        SetupPlanes(*out_frame, out_fmt);
        PushToDownstream(out_frame);
        ++processed_;

        if (stats_log_ && stats_interval_ > 0 && (processed_ % stats_interval_) == 0) {
            LogInfo("[preprocess] swscale frame=" + std::to_string(out_frame->frame_id) +
                    " " + std::to_string(frame->width) + "x" + std::to_string(frame->height) +
                    " -> " + std::to_string(out_w) + "x" + std::to_string(out_h) +
                    " id=" + id_);
        }
    }

    static AVPixelFormat ToAvFormat(PixelFormat fmt) {
        switch (fmt) {
            case PixelFormat::NV12: return AV_PIX_FMT_NV12;
            case PixelFormat::YUV420: return AV_PIX_FMT_YUV420P;
            case PixelFormat::RGB: return AV_PIX_FMT_RGB24;
            case PixelFormat::BGR: return AV_PIX_FMT_BGR24;
            default: return AV_PIX_FMT_NONE;
        }
    }

    static void SetupAvPlanes(const Frame* f, uint8_t* data[4], int linesize[4]) {
        if (!f->data) return;
        if (f->format == PixelFormat::NV12) {
            data[0] = f->planes[0].data ? f->planes[0].data : f->data;
            data[1] = f->planes[1].data ? f->planes[1].data : (f->data + f->width * f->height);
            linesize[0] = f->planes[0].stride > 0 ? f->planes[0].stride : f->width;
            linesize[1] = f->planes[1].stride > 0 ? f->planes[1].stride : f->width;
        } else if (f->format == PixelFormat::YUV420) {
            data[0] = f->planes[0].data ? f->planes[0].data : f->data;
            int y_size = f->width * f->height;
            int uv_size = y_size / 4;
            data[1] = f->planes[1].data ? f->planes[1].data : (f->data + y_size);
            data[2] = f->planes[2].data ? f->planes[2].data : (f->data + y_size + uv_size);
            linesize[0] = f->planes[0].stride > 0 ? f->planes[0].stride : f->width;
            linesize[1] = f->planes[1].stride > 0 ? f->planes[1].stride : f->width / 2;
            linesize[2] = f->planes[2].stride > 0 ? f->planes[2].stride : f->width / 2;
        } else {
            data[0] = f->data;
            linesize[0] = f->stride > 0 ? f->stride : f->width * 3;
        }
    }
#endif

    void SetupPlanes(Frame& f, PixelFormat fmt) {
        if (!f.data) return;

        if (fmt == PixelFormat::NV12) {
            f.plane_count = 2;
            int y_stride = f.stride > 0 ? f.stride : f.width;
            if (y_stride <= 0) y_stride = f.width;

            size_t y_bytes = static_cast<size_t>(y_stride) * static_cast<size_t>(f.height);
            if (f.data_size > 0) {
                const size_t candidate = (f.data_size * 2) / 3;  // total = Y + UV = Y*3/2
                if (candidate >= y_bytes && candidate <= f.data_size &&
                    (candidate % static_cast<size_t>(y_stride)) == 0) {
                    y_bytes = candidate;
                }
            }

            size_t uv_bytes = y_bytes / 2;
            if (f.data_size > 0 && y_bytes + uv_bytes > f.data_size) {
                uv_bytes = f.data_size > y_bytes ? (f.data_size - y_bytes) : 0;
            }

            f.planes[0] = {f.data, y_stride, static_cast<int>(y_bytes), 0};
            f.planes[1] = {f.data + y_bytes, y_stride, static_cast<int>(uv_bytes), static_cast<int>(y_bytes)};
            return;
        }

        if (fmt == PixelFormat::YUV420) {
            f.plane_count = 3;
            int y_stride = f.stride > 0 ? f.stride : f.width;
            if (y_stride <= 0) y_stride = f.width;

            size_t y_bytes = static_cast<size_t>(y_stride) * static_cast<size_t>(f.height);
            size_t hstride = static_cast<size_t>(f.height);
            if (f.data_size > 0) {
                const size_t candidate = (f.data_size * 2) / 3;
                if (candidate >= y_bytes && candidate <= f.data_size &&
                    (candidate % static_cast<size_t>(y_stride)) == 0) {
                    y_bytes = candidate;
                    hstride = y_bytes / static_cast<size_t>(y_stride);
                }
            }

            size_t uv_stride = static_cast<size_t>(y_stride) / 2;
            size_t uv_h = hstride / 2;
            size_t u_bytes = uv_stride * uv_h;
            size_t v_bytes = u_bytes;
            size_t need = y_bytes + u_bytes + v_bytes;

            if (f.data_size > 0 && need > f.data_size) {
                // Fallback to tightly packed layout.
                y_stride = f.width;
                y_bytes = static_cast<size_t>(f.width) * static_cast<size_t>(f.height);
                uv_stride = static_cast<size_t>(f.width) / 2;
                uv_h = static_cast<size_t>(f.height) / 2;
                u_bytes = uv_stride * uv_h;
                v_bytes = u_bytes;
            }

            f.planes[0] = {f.data, y_stride, static_cast<int>(y_bytes), 0};
            f.planes[1] = {f.data + y_bytes, static_cast<int>(uv_stride), static_cast<int>(u_bytes), static_cast<int>(y_bytes)};
            f.planes[2] = {f.data + y_bytes + u_bytes, static_cast<int>(uv_stride), static_cast<int>(v_bytes), static_cast<int>(y_bytes + u_bytes)};
            return;
        }

        // RGB/BGR
        f.plane_count = 1;
        int stride_bytes = f.stride > 0 ? f.stride : (f.width * 3);
        f.planes[0] = {f.data, stride_bytes, static_cast<int>(f.data_size), 0};
    }

    std::string id_;
    int dst_w_ = 640;
    int dst_h_ = 640;
    bool keep_ratio_ = false;
    PixelFormat dst_fmt_ = PixelFormat::UNKNOWN;
    bool dst_packed_ = false;
    bool use_rga_ = true;

    bool stats_log_ = false;
    uint64_t stats_interval_ = 100;

    bool warned_meta_resize_ = false;

    std::shared_ptr<SpscQueue<FramePtr>> input_queue_;
    std::vector<std::shared_ptr<SpscQueue<FramePtr>>> output_queues_;
    uint64_t processed_ = 0;

#if defined(RK3588_ENABLE_FFMPEG)
    SwsContext* sws_ctx_ = nullptr;
    int last_src_w_ = 0;
    int last_src_h_ = 0;
    AVPixelFormat last_src_fmt_ = AV_PIX_FMT_NONE;
    AVPixelFormat last_dst_fmt_ = AV_PIX_FMT_NONE;
#endif
};

REGISTER_NODE(PreprocessNode, "preprocess");

}  // namespace rk3588