#include <algorithm>
#include <atomic>
#include <chrono>
#include <cmath>
#include <cstddef>
#include <cstring>
#include <limits>
#include <memory>
#include <set>
#include <thread>
#include <vector>

#include "hw/i_image_processor.h"
#include "hw/i_infer_backend.h"
#include "node.h"
#include "utils/dma_alloc.h"
#include "utils/logger.h"

#if defined(RK3588_ENABLE_RKNN)
#include "rknn_api.h"
#endif

namespace rk3588 {

namespace {

constexpr int kObjClassNum = 80;
constexpr int kPropBoxSizeV5 = 5 + kObjClassNum;  // YOLOv5: x,y,w,h,conf + 80 classes
constexpr int kPropBoxSizeV8 = 4 + kObjClassNum;  // YOLOv8: x,y,w,h + 80 classes (no conf)
constexpr int kMaxDetections = 64;

// YOLOv5 anchors
const int kAnchor0[6] = {10, 13, 16, 30, 33, 23};
const int kAnchor1[6] = {30, 61, 62, 45, 59, 119};
const int kAnchor2[6] = {116, 90, 156, 198, 373, 326};

enum class YoloVersion { V5, V8 };
enum class V8BoxFormat { Auto, CxCyWh, XyXy, XyWh };
enum class V8ClsActivation { Auto, None, Sigmoid };

inline int Clamp(float val, int min_val, int max_val) {
    return val > min_val ? (val < max_val ? static_cast<int>(val) : max_val) : min_val;
}

struct DetCoordContext {
    bool has_transform = false;
    int out_w = 0;
    int out_h = 0;
    float scale_x = 1.0f;
    float scale_y = 1.0f;
    float pad_x = 0.0f;
    float pad_y = 0.0f;
    float fallback_scale_w = 1.0f;
    float fallback_scale_h = 1.0f;
};

DetCoordContext BuildDetCoordContext(const Frame& frame, int model_input_w, int model_input_h) {
    DetCoordContext ctx{};
    ctx.fallback_scale_w = frame.width > 0 ? static_cast<float>(model_input_w) / frame.width : 1.0f;
    ctx.fallback_scale_h = frame.height > 0 ? static_cast<float>(model_input_h) / frame.height : 1.0f;
    ctx.out_w = frame.width;
    ctx.out_h = frame.height;

    const bool infer_input_matches_frame = (frame.width == model_input_w && frame.height == model_input_h);

    if (infer_input_matches_frame &&
        frame.transform_meta && frame.transform_meta->valid &&
        frame.transform_meta->src_w > 0 && frame.transform_meta->src_h > 0 &&
        frame.transform_meta->scale_x > 1e-6f && frame.transform_meta->scale_y > 1e-6f) {
        ctx.has_transform = true;
        ctx.out_w = frame.transform_meta->src_w;
        ctx.out_h = frame.transform_meta->src_h;
        ctx.scale_x = frame.transform_meta->scale_x;
        ctx.scale_y = frame.transform_meta->scale_y;
        ctx.pad_x = frame.transform_meta->pad_x;
        ctx.pad_y = frame.transform_meta->pad_y;
    }
    return ctx;
}

struct BboxExpandConfig {
    bool enable = false;
    int class_id = 0;
    float left = 0.05f;
    float right = 0.05f;
    float top = 0.05f;
    float bottom = 0.12f;
};

Rect DecodeToOutputRect(float x, float y, float w, float h, const DetCoordContext& ctx) {
    float ox = x;
    float oy = y;
    float ow = w;
    float oh = h;

    if (ctx.has_transform) {
        ox = (x - ctx.pad_x) / ctx.scale_x;
        oy = (y - ctx.pad_y) / ctx.scale_y;
        ow = w / ctx.scale_x;
        oh = h / ctx.scale_y;
    } else {
        ox = x / ctx.fallback_scale_w;
        oy = y / ctx.fallback_scale_h;
        ow = w / ctx.fallback_scale_w;
        oh = h / ctx.fallback_scale_h;
    }

    Rect r{};
    const int out_w = std::max(1, ctx.out_w);
    const int out_h = std::max(1, ctx.out_h);
    r.x = static_cast<float>(Clamp(static_cast<int>(ox), 0, out_w));
    r.y = static_cast<float>(Clamp(static_cast<int>(oy), 0, out_h));
    r.w = static_cast<float>(Clamp(static_cast<int>(ow), 0, out_w - static_cast<int>(r.x)));
    r.h = static_cast<float>(Clamp(static_cast<int>(oh), 0, out_h - static_cast<int>(r.y)));
    return r;
}

Rect ExpandRect(const Rect& in, int img_w, int img_h, const BboxExpandConfig& cfg) {
    if (!cfg.enable || img_w <= 0 || img_h <= 0) return in;

    const float ex = in.w * cfg.left;
    const float ey = in.h * cfg.top;
    const float ew = in.w * (cfg.left + cfg.right);
    const float eh = in.h * (cfg.top + cfg.bottom);

    Rect out{};
    out.x = std::max(0.0f, in.x - ex);
    out.y = std::max(0.0f, in.y - ey);
    out.w = std::min(static_cast<float>(img_w) - out.x, in.w + ew);
    out.h = std::min(static_cast<float>(img_h) - out.y, in.h + eh);
    out.w = std::max(0.0f, out.w);
    out.h = std::max(0.0f, out.h);
    return out;
}

inline int32_t ClipFloat(float val, float min_val, float max_val) {
    return static_cast<int32_t>(val <= min_val ? min_val : (val >= max_val ? max_val : val));
}

inline int8_t QuantizeF32ToAffine(float f32, int32_t zp, float scale) {
    float dst_val = (f32 / scale) + zp;
    return static_cast<int8_t>(ClipFloat(dst_val, -128, 127));
}

inline float DequantizeAffineToF32(int8_t qnt, int32_t zp, float scale) {
    return (static_cast<float>(qnt) - static_cast<float>(zp)) * scale;
}

inline float Sigmoid(float x) {
    return 1.0f / (1.0f + std::exp(-x));
}

// FP16 (half) to FP32 conversion.
// Uses arithmetic reconstruction to avoid undefined behavior on subnormals.
inline float Fp16ToFp32(uint16_t h) {
    const int sign = (h & 0x8000) ? -1 : 1;
    const int exp = (h >> 10) & 0x1F;
    const int mant = h & 0x03FF;

    if (exp == 0) {
        if (mant == 0) return sign < 0 ? -0.0f : 0.0f;
        // subnormal: mant * 2^-24
        return static_cast<float>(sign) * std::ldexp(static_cast<float>(mant), -24);
    }
    if (exp == 0x1F) {
        if (mant == 0) return sign < 0 ? -INFINITY : INFINITY;
        return std::numeric_limits<float>::quiet_NaN();
    }
    // normal: (mant + 1024) * 2^(exp-25)
    return static_cast<float>(sign) *
           std::ldexp(static_cast<float>(mant + 1024), exp - 25);
}

float CalculateIoU(float x1_min, float y1_min, float x1_max, float y1_max,
                   float x2_min, float y2_min, float x2_max, float y2_max) {
    float w = std::fmax(0.f, std::fmin(x1_max, x2_max) - std::fmax(x1_min, x2_min) + 1.0f);
    float h = std::fmax(0.f, std::fmin(y1_max, y2_max) - std::fmax(y1_min, y2_min) + 1.0f);
    float inter = w * h;
    float area1 = (x1_max - x1_min + 1.0f) * (y1_max - y1_min + 1.0f);
    float area2 = (x2_max - x2_min + 1.0f) * (y2_max - y2_min + 1.0f);
    float uni = area1 + area2 - inter;
    return uni <= 0.f ? 0.f : (inter / uni);
}

void QuickSortDescending(std::vector<float>& values, int left, int right, std::vector<int>& indices) {
    if (left >= right) return;
    float pivot = values[left];
    int pivot_idx = indices[left];
    int low = left, high = right;
    while (low < high) {
        while (low < high && values[high] <= pivot) high--;
        values[low] = values[high];
        indices[low] = indices[high];
        while (low < high && values[low] >= pivot) low++;
        values[high] = values[low];
        indices[high] = indices[low];
    }
    values[low] = pivot;
    indices[low] = pivot_idx;
    QuickSortDescending(values, left, low - 1, indices);
    QuickSortDescending(values, low + 1, right, indices);
}

void NMS(int valid_count, std::vector<float>& boxes, std::vector<int>& class_ids,
         std::vector<int>& order, int filter_id, float threshold) {
    for (int i = 0; i < valid_count; ++i) {
        int n = order[i];
        if (n < 0 || n >= valid_count) continue;
        if (class_ids[n] != filter_id) continue;
        for (int j = i + 1; j < valid_count; ++j) {
            int m = order[j];
            if (m < 0 || m >= valid_count) continue;
            if (class_ids[m] != filter_id) continue;
            float x1_min = boxes[n * 4 + 0];
            float y1_min = boxes[n * 4 + 1];
            float x1_max = x1_min + boxes[n * 4 + 2];
            float y1_max = y1_min + boxes[n * 4 + 3];
            float x2_min = boxes[m * 4 + 0];
            float y2_min = boxes[m * 4 + 1];
            float x2_max = x2_min + boxes[m * 4 + 2];
            float y2_max = y2_min + boxes[m * 4 + 3];
            if (CalculateIoU(x1_min, y1_min, x1_max, y1_max, x2_min, y2_min, x2_max, y2_max) > threshold) {
                order[j] = -1;
            }
        }
    }
}

#if defined(RK3588_ENABLE_RKNN)
// YOLOv5 feature map processing (anchor-based)
int ProcessFeatureMapV5(int8_t* input, const int* anchor, int grid_h, int grid_w,
                        int /*model_h*/, int /*model_w*/, int stride,
                        std::vector<float>& boxes, std::vector<float>& obj_probs,
                        std::vector<int>& class_ids, float conf_thresh, int32_t zp, float scale) {
    int valid_count = 0;
    int grid_len = grid_h * grid_w;
    int8_t thresh_i8 = QuantizeF32ToAffine(conf_thresh, zp, scale);

    for (int a = 0; a < 3; ++a) {
        for (int i = 0; i < grid_h; ++i) {
            for (int j = 0; j < grid_w; ++j) {
                int8_t box_conf = input[(kPropBoxSizeV5 * a + 4) * grid_len + i * grid_w + j];
                if (box_conf >= thresh_i8) {
                    int offset = (kPropBoxSizeV5 * a) * grid_len + i * grid_w + j;
                    int8_t* ptr = input + offset;

                    float bx = DequantizeAffineToF32(*ptr, zp, scale) * 2.0f - 0.5f;
                    float by = DequantizeAffineToF32(ptr[grid_len], zp, scale) * 2.0f - 0.5f;
                    float bw = DequantizeAffineToF32(ptr[2 * grid_len], zp, scale) * 2.0f;
                    float bh = DequantizeAffineToF32(ptr[3 * grid_len], zp, scale) * 2.0f;

                    bx = (bx + j) * stride;
                    by = (by + i) * stride;
                    bw = bw * bw * anchor[a * 2];
                    bh = bh * bh * anchor[a * 2 + 1];
                    bx -= bw / 2.0f;
                    by -= bh / 2.0f;

                    int8_t max_cls_prob = ptr[5 * grid_len];
                    int max_cls_id = 0;
                    for (int k = 1; k < kObjClassNum; ++k) {
                        int8_t prob = ptr[(5 + k) * grid_len];
                        if (prob > max_cls_prob) {
                            max_cls_id = k;
                            max_cls_prob = prob;
                        }
                    }

                    if (max_cls_prob > thresh_i8) {
                        float score = DequantizeAffineToF32(max_cls_prob, zp, scale) *
                                      DequantizeAffineToF32(box_conf, zp, scale);
                        obj_probs.push_back(score);
                        class_ids.push_back(max_cls_id);
                        boxes.push_back(bx);
                        boxes.push_back(by);
                        boxes.push_back(bw);
                        boxes.push_back(bh);
                        ++valid_count;
                    }
                }
            }
        }
    }
    return valid_count;
}

uint32_t TensorTypeSizeBytes(rknn_tensor_type t) {
    switch (t) {
        case RKNN_TENSOR_INT8:
        case RKNN_TENSOR_UINT8:
            return 1;
        case RKNN_TENSOR_FLOAT16:
            return 2;
        case RKNN_TENSOR_FLOAT32:
            return 4;
        default:
            return 1;
    }
}

int DefaultV8NumBoxes(int model_h, int model_w) {
    if (model_h <= 0 || model_w <= 0) return 0;
    return (model_h / 8) * (model_w / 8) +
           (model_h / 16) * (model_w / 16) +
           (model_h / 32) * (model_w / 32);
}

struct V8LayoutInfo {
    int num_boxes = 0;
    bool channels_first = true;  // true: CxN, false: NxC
};

float ScoreBoxCandidate(float x, float y, float w, float h, int model_w, int model_h) {
    float s = 0.0f;
    if (w > 0.0f && h > 0.0f) s += 3.0f;
    if (w <= model_w * 1.2f) s += 1.0f;
    if (h <= model_h * 1.2f) s += 1.0f;
    if (x >= -model_w * 0.1f) s += 1.0f;
    if (y >= -model_h * 0.1f) s += 1.0f;
    if ((x + w) <= model_w * 1.2f) s += 1.0f;
    if ((y + h) <= model_h * 1.2f) s += 1.0f;
    return s;
}

bool SeemsNormalized(float a, float b, float c, float d) {
    auto in_range = [](float v) { return v >= -0.05f && v <= 2.5f; };
    return in_range(a) && in_range(b) && in_range(c) && in_range(d);
}

const char* V8BoxFormatName(V8BoxFormat fmt) {
    switch (fmt) {
        case V8BoxFormat::CxCyWh: return "cxcywh";
        case V8BoxFormat::XyXy: return "xyxy";
        case V8BoxFormat::XyWh: return "xywh";
        default: return "auto";
    }
}

void DecodeV8Box(float a, float b, float c, float d, int model_w, int model_h, V8BoxFormat fmt,
                 float& out_x, float& out_y, float& out_w, float& out_h, V8BoxFormat* used_fmt = nullptr) {
    if (SeemsNormalized(a, b, c, d)) {
        a *= static_cast<float>(model_w);
        b *= static_cast<float>(model_h);
        c *= static_cast<float>(model_w);
        d *= static_cast<float>(model_h);
    }

    auto decode_cxcywh = [&](float& x, float& y, float& w, float& h) {
        x = a - c / 2.0f;
        y = b - d / 2.0f;
        w = c;
        h = d;
    };
    auto decode_xyxy = [&](float& x, float& y, float& w, float& h) {
        x = a;
        y = b;
        w = c - a;
        h = d - b;
    };
    auto decode_xywh = [&](float& x, float& y, float& w, float& h) {
        x = a;
        y = b;
        w = c;
        h = d;
    };

    if (fmt == V8BoxFormat::CxCyWh) {
        decode_cxcywh(out_x, out_y, out_w, out_h);
        if (used_fmt) *used_fmt = V8BoxFormat::CxCyWh;
        return;
    }
    if (fmt == V8BoxFormat::XyXy) {
        decode_xyxy(out_x, out_y, out_w, out_h);
        if (used_fmt) *used_fmt = V8BoxFormat::XyXy;
        return;
    }
    if (fmt == V8BoxFormat::XyWh) {
        decode_xywh(out_x, out_y, out_w, out_h);
        if (used_fmt) *used_fmt = V8BoxFormat::XyWh;
        return;
    }

    float x1 = 0.0f, y1 = 0.0f, w1 = 0.0f, h1 = 0.0f;
    float x2 = 0.0f, y2 = 0.0f, w2 = 0.0f, h2 = 0.0f;
    float x3 = 0.0f, y3 = 0.0f, w3 = 0.0f, h3 = 0.0f;
    decode_cxcywh(x1, y1, w1, h1);
    decode_xyxy(x2, y2, w2, h2);
    decode_xywh(x3, y3, w3, h3);

    const float s1 = ScoreBoxCandidate(x1, y1, w1, h1, model_w, model_h);
    const float s2 = ScoreBoxCandidate(x2, y2, w2, h2, model_w, model_h);
    const float s3 = ScoreBoxCandidate(x3, y3, w3, h3, model_w, model_h);
    if (s2 >= s1 && s2 >= s3) {
        out_x = x2; out_y = y2; out_w = w2; out_h = h2;
        if (used_fmt) *used_fmt = V8BoxFormat::XyXy;
    } else if (s3 >= s1 && s3 >= s2) {
        out_x = x3; out_y = y3; out_w = w3; out_h = h3;
        if (used_fmt) *used_fmt = V8BoxFormat::XyWh;
    } else {
        out_x = x1; out_y = y1; out_w = w1; out_h = h1;
        if (used_fmt) *used_fmt = V8BoxFormat::CxCyWh;
    }
}

bool ResolveV8ApplySigmoid(const float* output, int num_boxes, int num_classes, bool channels_first,
                           V8ClsActivation act_mode) {
    if (act_mode == V8ClsActivation::None) return false;
    if (act_mode == V8ClsActivation::Sigmoid) return true;
    if (!output || num_boxes <= 0 || num_classes <= 0) return false;

    const int num_channels = 4 + num_classes;
    const int sample_boxes = std::min(num_boxes, 64);
    float min_v = 1e9f;
    float max_v = -1e9f;
    for (int i = 0; i < sample_boxes; ++i) {
        for (int c = 0; c < num_classes; ++c) {
            const float v = channels_first ? output[(4 + c) * num_boxes + i]
                                           : output[i * num_channels + (4 + c)];
            if (v < min_v) min_v = v;
            if (v > max_v) max_v = v;
        }
    }
    // If class outputs clearly look like logits, enable sigmoid.
    return (min_v < -0.1f || max_v > 1.5f);
}

V8LayoutInfo ResolveV8Layout(const std::vector<uint32_t>& dims, size_t byte_size,
                             rknn_tensor_type type, int num_classes,
                             int model_h, int model_w) {
    V8LayoutInfo info;
    const int num_channels = 4 + num_classes;
    if (num_channels <= 0) return info;

    const uint32_t elem_bytes = TensorTypeSizeBytes(type);
    const size_t total_elems = elem_bytes > 0 ? (byte_size / elem_bytes) : 0;
    const size_t max_boxes_from_data = static_cast<size_t>(num_channels) > 0
                                           ? (total_elems / static_cast<size_t>(num_channels))
                                           : 0;

    int ch_idx = -1;
    for (size_t i = 0; i < dims.size(); ++i) {
        if (dims[i] == static_cast<uint32_t>(num_channels)) {
            ch_idx = static_cast<int>(i);
            break;
        }
    }

    if (ch_idx >= 0 && total_elems >= static_cast<size_t>(num_channels)) {
        info.num_boxes = static_cast<int>(max_boxes_from_data);

        int prev_non1 = 1;
        for (int i = ch_idx - 1; i >= 0; --i) {
            if (dims[static_cast<size_t>(i)] > 1U) {
                prev_non1 = static_cast<int>(dims[static_cast<size_t>(i)]);
                break;
            }
        }
        int next_non1 = 1;
        for (size_t i = static_cast<size_t>(ch_idx + 1); i < dims.size(); ++i) {
            if (dims[i] > 1U) {
                next_non1 = static_cast<int>(dims[i]);
                break;
            }
        }

        if (next_non1 > 1 && prev_non1 <= 1) {
            info.channels_first = true;
        } else if (prev_non1 > 1 && next_non1 <= 1) {
            info.channels_first = false;
        } else if (next_non1 > 1 && prev_non1 > 1) {
            info.channels_first = next_non1 >= prev_non1;
        } else {
            info.channels_first = true;
        }
    } else if (dims.size() >= 3) {
        // Compatibility with old rank-3 assumptions.
        if (dims[1] == static_cast<uint32_t>(num_channels)) {
            info.num_boxes = static_cast<int>(dims[2]);
            info.channels_first = true;
        } else if (dims[2] == static_cast<uint32_t>(num_channels)) {
            info.num_boxes = static_cast<int>(dims[1]);
            info.channels_first = false;
        }
    }

    if (info.num_boxes <= 0 && max_boxes_from_data > 0) {
        info.num_boxes = static_cast<int>(max_boxes_from_data);
    }
    if (info.num_boxes <= 0) {
        info.num_boxes = DefaultV8NumBoxes(model_h, model_w);
    }
    if (info.num_boxes <= 0) {
        info.num_boxes = 8400;
    }

    if (max_boxes_from_data > 0 && static_cast<size_t>(info.num_boxes) > max_boxes_from_data) {
        info.num_boxes = static_cast<int>(max_boxes_from_data);
    }
    if (info.num_boxes < 0) info.num_boxes = 0;
    return info;
}

// YOLOv8 output processing (anchor-free, single output tensor)
int ProcessOutputV8(float* output, int num_boxes, int num_classes,
                    int model_h, int model_w,
                    std::vector<float>& boxes, std::vector<float>& obj_probs,
                    std::vector<int>& class_ids, float conf_thresh,
                    bool channels_first, V8BoxFormat box_format, bool apply_sigmoid,
                    bool debug_decode, int* debug_left) {
    int valid_count = 0;
    const int num_channels = 4 + num_classes;

    for (int i = 0; i < num_boxes; ++i) {
        float max_score = 0.0f;
        int max_cls_id = 0;
        for (int c = 0; c < num_classes; ++c) {
            float score = channels_first ? output[(4 + c) * num_boxes + i]
                                         : output[i * num_channels + (4 + c)];
            if (apply_sigmoid) score = Sigmoid(score);
            if (score > max_score) {
                max_score = score;
                max_cls_id = c;
            }
        }

        if (max_score >= conf_thresh) {
            const float a = channels_first ? output[0 * num_boxes + i] : output[i * num_channels + 0];
            const float b = channels_first ? output[1 * num_boxes + i] : output[i * num_channels + 1];
            const float c = channels_first ? output[2 * num_boxes + i] : output[i * num_channels + 2];
            const float d = channels_first ? output[3 * num_boxes + i] : output[i * num_channels + 3];
            if (!std::isfinite(a) || !std::isfinite(b) || !std::isfinite(c) || !std::isfinite(d)) {
                continue;
            }
            float x1 = 0.0f, y1 = 0.0f, w = 0.0f, h = 0.0f;
            V8BoxFormat used_fmt = box_format;
            DecodeV8Box(a, b, c, d, model_w, model_h, box_format, x1, y1, w, h, &used_fmt);
            if (!std::isfinite(x1) || !std::isfinite(y1) || !std::isfinite(w) || !std::isfinite(h)) {
                continue;
            }
            if (w <= 1e-3f || h <= 1e-3f) continue;
            if (debug_decode && debug_left && *debug_left > 0) {
                --(*debug_left);
                LogInfo("[ai_yolo] v8 decode f32: raw4(" + std::to_string(a) + "," + std::to_string(b) + "," +
                        std::to_string(c) + "," + std::to_string(d) + ") fmt=" + V8BoxFormatName(used_fmt) +
                        " -> xywh(" + std::to_string(x1) + "," + std::to_string(y1) + "," +
                        std::to_string(w) + "," + std::to_string(h) + ") cls=" +
                        std::to_string(max_cls_id) + " score=" + std::to_string(max_score));
            }

            boxes.push_back(x1);
            boxes.push_back(y1);
            boxes.push_back(w);
            boxes.push_back(h);
            obj_probs.push_back(max_score);
            class_ids.push_back(max_cls_id);
            ++valid_count;
        }
    }
    return valid_count;
}

// YOLOv8 INT8 output processing
int ProcessOutputV8Int8(int8_t* output, int num_boxes, int num_classes,
                        int model_h, int model_w,
                        std::vector<float>& boxes, std::vector<float>& obj_probs,
                        std::vector<int>& class_ids, float conf_thresh,
                        int32_t zp, float scale, bool channels_first, V8BoxFormat box_format,
                        bool debug_decode, int* debug_left) {
    int valid_count = 0;
    int8_t thresh_i8 = QuantizeF32ToAffine(conf_thresh, zp, scale);
    const int num_channels = 4 + num_classes;

    for (int i = 0; i < num_boxes; ++i) {
        int8_t max_score_i8 = -128;
        int max_cls_id = 0;
        for (int c = 0; c < num_classes; ++c) {
            int8_t score = channels_first ? output[(4 + c) * num_boxes + i]
                                          : output[i * num_channels + (4 + c)];
            if (score > max_score_i8) {
                max_score_i8 = score;
                max_cls_id = c;
            }
        }

        if (max_score_i8 >= thresh_i8) {
            float a = DequantizeAffineToF32(
                channels_first ? output[0 * num_boxes + i] : output[i * num_channels + 0], zp, scale);
            float b = DequantizeAffineToF32(
                channels_first ? output[1 * num_boxes + i] : output[i * num_channels + 1], zp, scale);
            float c = DequantizeAffineToF32(
                channels_first ? output[2 * num_boxes + i] : output[i * num_channels + 2], zp, scale);
            float d = DequantizeAffineToF32(
                channels_first ? output[3 * num_boxes + i] : output[i * num_channels + 3], zp, scale);
            float max_score = DequantizeAffineToF32(max_score_i8, zp, scale);
            if (!std::isfinite(a) || !std::isfinite(b) || !std::isfinite(c) || !std::isfinite(d)) {
                continue;
            }
            float x1 = 0.0f, y1 = 0.0f, w = 0.0f, h = 0.0f;
            V8BoxFormat used_fmt = box_format;
            DecodeV8Box(a, b, c, d, model_w, model_h, box_format, x1, y1, w, h, &used_fmt);
            if (!std::isfinite(x1) || !std::isfinite(y1) || !std::isfinite(w) || !std::isfinite(h)) {
                continue;
            }
            if (w <= 1e-3f || h <= 1e-3f) continue;
            if (debug_decode && debug_left && *debug_left > 0) {
                --(*debug_left);
                LogInfo("[ai_yolo] v8 decode int8: raw4(" + std::to_string(a) + "," + std::to_string(b) + "," +
                        std::to_string(c) + "," + std::to_string(d) + ") fmt=" + V8BoxFormatName(used_fmt) +
                        " -> xywh(" + std::to_string(x1) + "," + std::to_string(y1) + "," +
                        std::to_string(w) + "," + std::to_string(h) + ") cls=" +
                        std::to_string(max_cls_id) + " score=" + std::to_string(max_score));
            }

            boxes.push_back(x1);
            boxes.push_back(y1);
            boxes.push_back(w);
            boxes.push_back(h);
            obj_probs.push_back(max_score);
            class_ids.push_back(max_cls_id);
            ++valid_count;
        }
    }
    return valid_count;
}
#endif

}  // namespace

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

    bool Init(const SimpleJson& config, const NodeContext& ctx) override {
        id_ = config.ValueOr<std::string>("id", "ai_yolo");
        model_path_ = config.ValueOr<std::string>("model_path", "");
        conf_thresh_ = config.ValueOr<float>("conf", 0.25f);
        nms_thresh_ = config.ValueOr<float>("nms", 0.45f);
        model_input_w_ = config.ValueOr<int>("model_w", 640);
        model_input_h_ = config.ValueOr<int>("model_h", 640);
        num_classes_ = config.ValueOr<int>("num_classes", 80);
        {
            const std::string bf = config.ValueOr<std::string>("v8_box_format", "cxcywh");
            if (bf == "xyxy") {
                v8_box_format_ = V8BoxFormat::XyXy;
            } else if (bf == "xywh") {
                v8_box_format_ = V8BoxFormat::XyWh;
            } else if (bf == "cxcywh") {
                v8_box_format_ = V8BoxFormat::CxCyWh;
            } else {
                v8_box_format_ = V8BoxFormat::Auto;
            }
        }
        {
            const std::string act = config.ValueOr<std::string>("v8_cls_activation", "auto");
            if (act == "sigmoid") {
                v8_cls_activation_ = V8ClsActivation::Sigmoid;
            } else if (act == "none") {
                v8_cls_activation_ = V8ClsActivation::None;
            } else {
                v8_cls_activation_ = V8ClsActivation::Auto;
            }
        }

        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_))));
            debug_det_ = dbg->ValueOr<bool>("detections", debug_det_);
        }

        // Optional inference throttle. 0 = run every frame.
        infer_interval_ms_ = std::max<int64_t>(0, static_cast<int64_t>(config.ValueOr<int>("infer_interval_ms", 0)));
        if (infer_interval_ms_ <= 0) {
            const double infer_fps = config.ValueOr<double>("infer_fps", 0.0);
            if (infer_fps > 0.0) {
                infer_interval_ms_ = static_cast<int64_t>(1000.0 / infer_fps);
                if (infer_interval_ms_ < 1) infer_interval_ms_ = 1;
            }
        }
        infer_phase_ms_ = std::max<int64_t>(0, static_cast<int64_t>(config.ValueOr<int>("infer_phase_ms", 0)));
        if (infer_interval_ms_ > 0 && infer_phase_ms_ >= infer_interval_ms_) {
            infer_phase_ms_ %= infer_interval_ms_;
        }
        use_dma_input_ = config.ValueOr<bool>("use_dma_input", use_dma_input_);

        std::string ver = config.ValueOr<std::string>("model_version", "auto");
        if (ver == "v5") {
            yolo_version_ = YoloVersion::V5;
        } else if (ver == "v8") {
            yolo_version_ = YoloVersion::V8;
        } else {
            yolo_version_ = YoloVersion::V8;
            auto_detect_version_ = true;
        }

        if (const SimpleJson* filter = config.Find("class_filter")) {
            for (const auto& item : filter->AsArray()) {
                class_filter_.insert(item.AsInt(-1));
            }
        }

        if (const SimpleJson* expand = config.Find("bbox_expand"); expand && expand->IsObject()) {
            bbox_expand_.enable = expand->ValueOr<bool>("enable", false);
            bbox_expand_.class_id = expand->ValueOr<int>("class_id", bbox_expand_.class_id);
            bbox_expand_.left = expand->ValueOr<float>("left", bbox_expand_.left);
            bbox_expand_.right = expand->ValueOr<float>("right", bbox_expand_.right);
            bbox_expand_.top = expand->ValueOr<float>("top", bbox_expand_.top);
            bbox_expand_.bottom = expand->ValueOr<float>("bottom", bbox_expand_.bottom);
        }

        input_queue_ = ctx.input_queue;
        if (!input_queue_) {
            LogError("[ai_yolo] no input queue for node " + id_);
            return false;
        }
        if (ctx.output_queues.empty()) {
            LogError("[ai_yolo] no output queue for node " + id_);
            return false;
        }
        output_queues_ = ctx.output_queues;

        infer_backend_ = ctx.infer_backend;
        if (!infer_backend_) {
            LogError("[ai_yolo] no infer backend for node " + id_);
            return false;
        }
        image_processor_ = ctx.image_processor;

#if defined(RK3588_ENABLE_RKNN)
        if (model_path_.empty()) {
            LogError("[ai_yolo] model_path is required");
            return false;
        }

        std::string err;
        model_handle_ = infer_backend_->LoadModel(model_path_, err);
        if (model_handle_ == kInvalidModelHandle) {
            LogError("[ai_yolo] failed to load model: " + err);
            return false;
        }

        ModelInfo info;
        if (infer_backend_->GetModelInfo(model_handle_, info)) {
            model_input_w_ = info.input_width;
            model_input_h_ = info.input_height;
            n_output_ = info.n_output;

            if (auto_detect_version_) {
                if (n_output_ == 1) {
                    yolo_version_ = YoloVersion::V8;
                } else if (n_output_ >= 3) {
                    yolo_version_ = YoloVersion::V5;
                }
            }
        }

        LogInfo("[ai_yolo] model loaded via InferBackend: " + model_path_ +
                " (handle=" + std::to_string(model_handle_) + ", version=" +
                (yolo_version_ == YoloVersion::V5 ? "v5" : "v8") + ")");
#else
        LogWarn("[ai_yolo] RKNN disabled, will passthrough frames");
#endif
        return true;
    }

    bool Start() override {
        LogInfo("[ai_yolo] start id=" + id_ + " conf=" + std::to_string(conf_thresh_) +
                " nms=" + std::to_string(nms_thresh_));
        return true;
    }

    void Stop() override {
#if defined(RK3588_ENABLE_RKNN)
        if (model_handle_ != kInvalidModelHandle) {
            infer_backend_->UnloadModel(model_handle_);
            model_handle_ = kInvalidModelHandle;
        }
#endif
        LogInfo("[ai_yolo] stop id=" + id_);
    }

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

        if (infer_interval_ms_ > 0 && frame->pts > 0) {
            const int64_t pts_ms = static_cast<int64_t>(frame->pts / 1000ULL);
            const int64_t effective_pts_ms = pts_ms + infer_phase_ms_;
            const int64_t delta_ms = effective_pts_ms - last_infer_pts_ms_;
            if (last_infer_pts_ms_ > 0 && delta_ms > 0 && delta_ms < infer_interval_ms_) {
                PushToDownstream(frame);
                ++processed_;
                return NodeStatus::OK;
            }
            last_infer_pts_ms_ = effective_pts_ms;
        }

#if defined(RK3588_ENABLE_RKNN)
        RunInference(frame);
#endif
        PushToDownstream(frame);
        ++processed_;

        // Stats logging disabled
        (void)stats_log_;
        (void)stats_interval_;
        return NodeStatus::OK;
    }

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

#if defined(RK3588_ENABLE_RKNN)
    void ResizeRgbBilinear(const uint8_t* src, int src_w, int src_h, int src_stride,
                           uint8_t* dst, int dst_w, int dst_h, int dst_stride) {
        const float scale_x = static_cast<float>(src_w) / static_cast<float>(dst_w);
        const float scale_y = static_cast<float>(src_h) / static_cast<float>(dst_h);

        for (int y = 0; y < dst_h; ++y) {
            const float fy = static_cast<float>(y) * scale_y;
            const int y0 = static_cast<int>(fy);
            const int y1 = std::min(y0 + 1, src_h - 1);
            const float dy = fy - static_cast<float>(y0);

            for (int x = 0; x < dst_w; ++x) {
                const float fx = static_cast<float>(x) * scale_x;
                const int x0 = static_cast<int>(fx);
                const int x1 = std::min(x0 + 1, src_w - 1);
                const float dx = fx - static_cast<float>(x0);

                for (int c = 0; c < 3; ++c) {
                    const float v00 = src[y0 * src_stride + x0 * 3 + c];
                    const float v01 = src[y0 * src_stride + x1 * 3 + c];
                    const float v10 = src[y1 * src_stride + x0 * 3 + c];
                    const float v11 = src[y1 * src_stride + x1 * 3 + c];
                    const float v = v00 * (1.0f - dx) * (1.0f - dy) +
                                    v01 * dx * (1.0f - dy) +
                                    v10 * (1.0f - dx) * dy +
                                    v11 * dx * dy;
                    dst[y * dst_stride + x * 3 + c] = static_cast<uint8_t>(v);
                }
            }
        }
    }

    void RunInference(FramePtr frame) {
        if (!frame->data || frame->data_size == 0) return;

        bool is_rgb = (frame->format == PixelFormat::RGB || frame->format == PixelFormat::BGR);
        if (!is_rgb) {
            LogWarn("[ai_yolo] input must be RGB/BGR, got other format");
            return;
        }

        const int w = frame->width;
        const int h = frame->height;
        const size_t packed_row = static_cast<size_t>(w) * 3;
        const size_t packed_size = packed_row * static_cast<size_t>(h);

        const uint8_t* src = frame->planes[0].data ? frame->planes[0].data : frame->data;
        int src_stride = frame->planes[0].stride > 0 ? frame->planes[0].stride
                         : (frame->stride > 0 ? frame->stride : static_cast<int>(packed_row));
        if (!src || src_stride <= 0) return;

        InferInput input;
        const bool exact_model_input = (w == model_input_w_ && h == model_input_h_);
        Frame resized_frame;
        if (exact_model_input && static_cast<size_t>(src_stride) == packed_row && frame->data_size >= packed_size) {
            input.data = src;
            input.size = packed_size;
            input.width = w;
            input.height = h;

            // Best-effort RKNN DMA-BUF zero-copy path.
            if (frame->DmaFd() >= 0 && frame->data) {
                const ptrdiff_t off = src - frame->data;
                if (off >= 0 && static_cast<size_t>(off) + packed_size <= frame->data_size) {
                    input.dma_fd = frame->DmaFd();
                    input.dma_offset = static_cast<int>(off);
                }
            }
        } else {
            if (frame->data_size < static_cast<size_t>(src_stride) * static_cast<size_t>(h)) {
                LogWarn("[ai_yolo] invalid RGB buffer size/stride (data_size=" + std::to_string(frame->data_size) +
                        ", stride=" + std::to_string(src_stride) +
                        ", h=" + std::to_string(h) + ")");
                return;
            }

            const size_t input_row = static_cast<size_t>(model_input_w_) * 3;
            const size_t input_size = input_row * static_cast<size_t>(model_input_h_);
            resized_input_.resize(input_size);

            if (exact_model_input) {
                for (int y = 0; y < h; ++y) {
                    memcpy(resized_input_.data() + static_cast<size_t>(y) * input_row,
                           src + static_cast<size_t>(y) * static_cast<size_t>(src_stride),
                           input_row);
                }
            } else if (image_processor_ && w > 0 && h > 0) {
                Frame src_frame;
                src_frame.width = w;
                src_frame.height = h;
                src_frame.format = PixelFormat::RGB;
                src_frame.stride = src_stride;
                src_frame.data = const_cast<uint8_t*>(src);
                src_frame.data_size = frame->data_size;
                src_frame.plane_count = 1;
                src_frame.planes[0] = {src_frame.data, src_stride, static_cast<int>(frame->data_size), 0};

                resized_frame.width = model_input_w_;
                resized_frame.height = model_input_h_;
                resized_frame.format = PixelFormat::RGB;

                Status st = image_processor_->Resize(src_frame, resized_frame);
                if (st.Ok()) {
                    const uint8_t* resized_data = resized_frame.planes[0].data ? resized_frame.planes[0].data : resized_frame.data;
                    const int resized_stride = resized_frame.planes[0].stride > 0
                                                   ? resized_frame.planes[0].stride
                                                   : (resized_frame.stride > 0 ? resized_frame.stride : static_cast<int>(input_row));
                    if (use_dma_input_ && resized_frame.DmaFd() >= 0 && resized_data) {
                        input.data = resized_data;
                        input.size = resized_frame.data_size > 0 ? resized_frame.data_size : input_size;
                        input.width = model_input_w_;
                        input.height = model_input_h_;
                        input.dma_fd = resized_frame.DmaFd();
                        input.dma_offset = 0;
                    } else if (resized_data && resized_stride > 0) {
                        for (int y = 0; y < model_input_h_; ++y) {
                            memcpy(resized_input_.data() + static_cast<size_t>(y) * input_row,
                                   resized_data + static_cast<size_t>(y) * static_cast<size_t>(resized_stride),
                                   input_row);
                        }
                        input.data = resized_input_.data();
                        input.size = input_size;
                        input.width = model_input_w_;
                        input.height = model_input_h_;
                        input.dma_fd = -1;
                        input.dma_offset = 0;
                    } else {
                        ResizeRgbBilinear(src, w, h, src_stride,
                                          resized_input_.data(), model_input_w_, model_input_h_,
                                          static_cast<int>(input_row));
                        input.data = resized_input_.data();
                        input.size = input_size;
                        input.width = model_input_w_;
                        input.height = model_input_h_;
                        input.dma_fd = -1;
                        input.dma_offset = 0;
                    }
                } else {
                    ResizeRgbBilinear(src, w, h, src_stride,
                                      resized_input_.data(), model_input_w_, model_input_h_,
                                      static_cast<int>(input_row));
                    input.data = resized_input_.data();
                    input.size = input_size;
                    input.width = model_input_w_;
                    input.height = model_input_h_;
                    input.dma_fd = -1;
                    input.dma_offset = 0;
                }
            } else {
                ResizeRgbBilinear(src, w, h, src_stride,
                                  resized_input_.data(), model_input_w_, model_input_h_,
                                  static_cast<int>(input_row));
                input.data = resized_input_.data();
                input.size = input_size;
                input.width = model_input_w_;
                input.height = model_input_h_;
                input.dma_fd = -1;
                input.dma_offset = 0;
            }
        }
        input.is_nhwc = true;

        auto result = infer_backend_->InferBorrowed(model_handle_, input);
        if (!result.success) {
            LogWarn("[ai_yolo] inference failed: " + result.error);
            return;
        }

        PostProcessBorrowed(result.outputs, frame);
    }

    void PostProcessBorrowed(const std::vector<AiScheduler::BorrowedOutput>& outputs, FramePtr frame) {
        std::vector<float> boxes;
        std::vector<float> obj_probs;
        std::vector<int> class_ids;
        int valid_count = 0;

        if (yolo_version_ == YoloVersion::V5) {
            if (outputs.size() < 3) return;
            if (!outputs[0].data || !outputs[1].data || !outputs[2].data) return;

            int stride0 = 8, stride1 = 16, stride2 = 32;
            int grid_h0 = model_input_h_ / stride0, grid_w0 = model_input_w_ / stride0;
            int grid_h1 = model_input_h_ / stride1, grid_w1 = model_input_w_ / stride1;
            int grid_h2 = model_input_h_ / stride2, grid_w2 = model_input_w_ / stride2;

            int cnt0 = ProcessFeatureMapV5(reinterpret_cast<int8_t*>(const_cast<uint8_t*>(outputs[0].data)), kAnchor0,
                                           grid_h0, grid_w0, model_input_h_, model_input_w_, stride0,
                                           boxes, obj_probs, class_ids, conf_thresh_,
                                           outputs[0].zp, outputs[0].scale);
            int cnt1 = ProcessFeatureMapV5(reinterpret_cast<int8_t*>(const_cast<uint8_t*>(outputs[1].data)), kAnchor1,
                                           grid_h1, grid_w1, model_input_h_, model_input_w_, stride1,
                                           boxes, obj_probs, class_ids, conf_thresh_,
                                           outputs[1].zp, outputs[1].scale);
            int cnt2 = ProcessFeatureMapV5(reinterpret_cast<int8_t*>(const_cast<uint8_t*>(outputs[2].data)), kAnchor2,
                                           grid_h2, grid_w2, model_input_h_, model_input_w_, stride2,
                                           boxes, obj_probs, class_ids, conf_thresh_,
                                           outputs[2].zp, outputs[2].scale);
            valid_count = cnt0 + cnt1 + cnt2;
        } else {
            if (outputs.empty()) return;
            if (!outputs[0].data || outputs[0].size == 0) return;

            const V8LayoutInfo layout = ResolveV8Layout(outputs[0].dims, outputs[0].size,
                                                        outputs[0].type, num_classes_,
                                                        model_input_h_, model_input_w_);
            const int num_boxes = layout.num_boxes;
            int debug_decode_left = (debug_det_ && processed_ < 20) ? 5 : 0;
            if (num_boxes <= 0) return;
            if (debug_det_ && processed_ < 5) {
                std::string dims_s;
                for (size_t di = 0; di < outputs[0].dims.size(); ++di) {
                    dims_s += (di == 0 ? "[" : ",");
                    dims_s += std::to_string(outputs[0].dims[di]);
                }
                dims_s += "]";
                LogInfo("[ai_yolo] v8 out type=" + std::to_string(static_cast<int>(outputs[0].type)) +
                        " size=" + std::to_string(outputs[0].size) +
                        " dims=" + dims_s +
                        " num_boxes=" + std::to_string(num_boxes) +
                        " layout=" + std::string(layout.channels_first ? "CxN" : "NxC"));
            }

            if (outputs[0].type == RKNN_TENSOR_FLOAT32) {
                const bool apply_sigmoid = ResolveV8ApplySigmoid(
                    reinterpret_cast<float*>(const_cast<uint8_t*>(outputs[0].data)),
                    num_boxes, num_classes_, layout.channels_first, v8_cls_activation_);
                if (debug_det_ && processed_ < 5) {
                    LogInfo("[ai_yolo] v8 cls activation=" + std::string(apply_sigmoid ? "sigmoid" : "none"));
                }
                valid_count = ProcessOutputV8(reinterpret_cast<float*>(const_cast<uint8_t*>(outputs[0].data)),
                                              num_boxes, num_classes_,
                                              model_input_h_, model_input_w_,
                                              boxes, obj_probs, class_ids, conf_thresh_,
                                              layout.channels_first, v8_box_format_, apply_sigmoid,
                                              debug_det_, &debug_decode_left);
            } else if (outputs[0].type == RKNN_TENSOR_FLOAT16) {
                // Convert FP16 to FP32
                size_t num_elements = outputs[0].size / sizeof(uint16_t);
                fp32_buffer_.resize(num_elements);
                const uint16_t* fp16_data = reinterpret_cast<const uint16_t*>(outputs[0].data);
                for (size_t i = 0; i < num_elements; ++i) {
                    fp32_buffer_[i] = Fp16ToFp32(fp16_data[i]);
                }
                const bool apply_sigmoid = ResolveV8ApplySigmoid(
                    fp32_buffer_.data(), num_boxes, num_classes_, layout.channels_first, v8_cls_activation_);
                if (debug_det_ && processed_ < 5) {
                    LogInfo("[ai_yolo] v8 cls activation=" + std::string(apply_sigmoid ? "sigmoid" : "none"));
                }
                valid_count = ProcessOutputV8(fp32_buffer_.data(),
                                              num_boxes, num_classes_,
                                              model_input_h_, model_input_w_,
                                              boxes, obj_probs, class_ids, conf_thresh_,
                                              layout.channels_first, v8_box_format_, apply_sigmoid,
                                              debug_det_, &debug_decode_left);
            } else {
                valid_count = ProcessOutputV8Int8(reinterpret_cast<int8_t*>(const_cast<uint8_t*>(outputs[0].data)),
                                                  num_boxes, num_classes_,
                                                  model_input_h_, model_input_w_,
                                                  boxes, obj_probs, class_ids, conf_thresh_,
                                                  outputs[0].zp, outputs[0].scale,
                                                  layout.channels_first, v8_box_format_,
                                                  debug_det_, &debug_decode_left);
            }
        }

        if (valid_count <= 0) return;

        std::vector<int> indices(valid_count);
        for (int i = 0; i < valid_count; ++i) indices[i] = i;

        QuickSortDescending(obj_probs, 0, valid_count - 1, indices);

        std::set<int> class_set(class_ids.begin(), class_ids.end());
        for (int c : class_set) {
            NMS(valid_count, boxes, class_ids, indices, c, nms_thresh_);
        }

        const DetCoordContext coord_ctx = BuildDetCoordContext(*frame, model_input_w_, model_input_h_);

        auto det_result = std::make_shared<DetectionResult>();
        det_result->img_w = coord_ctx.out_w;
        det_result->img_h = coord_ctx.out_h;
        det_result->model_name = (yolo_version_ == YoloVersion::V5) ? "yolov5" : "yolov8";

        for (int i = 0; i < valid_count && det_result->items.size() < kMaxDetections; ++i) {
            if (indices[i] == -1) continue;
            int n = indices[i];
            int cls_id = class_ids[n];

            if (!class_filter_.empty() && class_filter_.find(cls_id) == class_filter_.end()) {
                continue;
            }

            float x1 = boxes[n * 4 + 0];
            float y1 = boxes[n * 4 + 1];
            float w = boxes[n * 4 + 2];
            float h = boxes[n * 4 + 3];

            Detection det;
            det.cls_id = cls_id;
            det.score = obj_probs[i];
            det.bbox = DecodeToOutputRect(x1, y1, w, h, coord_ctx);
            if (bbox_expand_.enable && det.cls_id == bbox_expand_.class_id) {
                det.bbox = ExpandRect(det.bbox, coord_ctx.out_w, coord_ctx.out_h, bbox_expand_);
            }
            det.track_id = -1;

            if (debug_det_ && det_result->items.size() < 5 && processed_ < 20) {
                LogInfo("[ai_yolo] det: raw(" + std::to_string(x1) + "," + std::to_string(y1) + "," +
                        std::to_string(w) + "," + std::to_string(h) + ") -> bbox(" +
                        std::to_string(det.bbox.x) + "," + std::to_string(det.bbox.y) + "," +
                        std::to_string(det.bbox.w) + "," + std::to_string(det.bbox.h) + ") cls=" +
                        std::to_string(cls_id) + " score=" + std::to_string(det.score));
            }

            det_result->items.push_back(det);
        }
        if (debug_det_ && processed_ < 20) {
            LogInfo("[ai_yolo] det summary: valid_count=" + std::to_string(valid_count) +
                    " final=" + std::to_string(det_result->items.size()));
        }

        frame->det = det_result;
    }

    void PostProcess(std::vector<InferOutput>& outputs, FramePtr frame) {
        std::vector<float> boxes;
        std::vector<float> obj_probs;
        std::vector<int> class_ids;
        int valid_count = 0;

        if (yolo_version_ == YoloVersion::V5) {
            if (outputs.size() < 3) return;

            int stride0 = 8, stride1 = 16, stride2 = 32;
            int grid_h0 = model_input_h_ / stride0, grid_w0 = model_input_w_ / stride0;
            int grid_h1 = model_input_h_ / stride1, grid_w1 = model_input_w_ / stride1;
            int grid_h2 = model_input_h_ / stride2, grid_w2 = model_input_w_ / stride2;

            int cnt0 = ProcessFeatureMapV5(reinterpret_cast<int8_t*>(outputs[0].data.data()), kAnchor0,
                                           grid_h0, grid_w0, model_input_h_, model_input_w_, stride0,
                                           boxes, obj_probs, class_ids, conf_thresh_,
                                           outputs[0].zp, outputs[0].scale);
            int cnt1 = ProcessFeatureMapV5(reinterpret_cast<int8_t*>(outputs[1].data.data()), kAnchor1,
                                           grid_h1, grid_w1, model_input_h_, model_input_w_, stride1,
                                           boxes, obj_probs, class_ids, conf_thresh_,
                                           outputs[1].zp, outputs[1].scale);
            int cnt2 = ProcessFeatureMapV5(reinterpret_cast<int8_t*>(outputs[2].data.data()), kAnchor2,
                                           grid_h2, grid_w2, model_input_h_, model_input_w_, stride2,
                                           boxes, obj_probs, class_ids, conf_thresh_,
                                           outputs[2].zp, outputs[2].scale);
            valid_count = cnt0 + cnt1 + cnt2;
        } else {
            if (outputs.empty()) return;

            const V8LayoutInfo layout = ResolveV8Layout(outputs[0].dims, outputs[0].data.size(),
                                                        outputs[0].type, num_classes_,
                                                        model_input_h_, model_input_w_);
            const int num_boxes = layout.num_boxes;
            int debug_decode_left = (debug_det_ && processed_ < 20) ? 5 : 0;
            if (num_boxes <= 0) return;
            if (debug_det_ && processed_ < 5) {
                std::string dims_s;
                for (size_t di = 0; di < outputs[0].dims.size(); ++di) {
                    dims_s += (di == 0 ? "[" : ",");
                    dims_s += std::to_string(outputs[0].dims[di]);
                }
                dims_s += "]";
                LogInfo("[ai_yolo] v8 out(type copy) type=" + std::to_string(static_cast<int>(outputs[0].type)) +
                        " size=" + std::to_string(outputs[0].data.size()) +
                        " dims=" + dims_s +
                        " num_boxes=" + std::to_string(num_boxes) +
                        " layout=" + std::string(layout.channels_first ? "CxN" : "NxC"));
            }

            if (outputs[0].type == RKNN_TENSOR_FLOAT32) {
                const bool apply_sigmoid = ResolveV8ApplySigmoid(
                    reinterpret_cast<float*>(outputs[0].data.data()),
                    num_boxes, num_classes_, layout.channels_first, v8_cls_activation_);
                if (debug_det_ && processed_ < 5) {
                    LogInfo("[ai_yolo] v8 cls activation(copy)=" + std::string(apply_sigmoid ? "sigmoid" : "none"));
                }
                valid_count = ProcessOutputV8(reinterpret_cast<float*>(outputs[0].data.data()),
                                              num_boxes, num_classes_,
                                              model_input_h_, model_input_w_,
                                              boxes, obj_probs, class_ids, conf_thresh_,
                                              layout.channels_first, v8_box_format_, apply_sigmoid,
                                              debug_det_, &debug_decode_left);
            } else if (outputs[0].type == RKNN_TENSOR_FLOAT16) {
                // Convert FP16 to FP32
                size_t num_elements = outputs[0].data.size() / sizeof(uint16_t);
                fp32_buffer_.resize(num_elements);
                const uint16_t* fp16_data = reinterpret_cast<const uint16_t*>(outputs[0].data.data());
                for (size_t i = 0; i < num_elements; ++i) {
                    fp32_buffer_[i] = Fp16ToFp32(fp16_data[i]);
                }
                const bool apply_sigmoid = ResolveV8ApplySigmoid(
                    fp32_buffer_.data(), num_boxes, num_classes_, layout.channels_first, v8_cls_activation_);
                if (debug_det_ && processed_ < 5) {
                    LogInfo("[ai_yolo] v8 cls activation(copy)=" + std::string(apply_sigmoid ? "sigmoid" : "none"));
                }
                valid_count = ProcessOutputV8(fp32_buffer_.data(),
                                              num_boxes, num_classes_,
                                              model_input_h_, model_input_w_,
                                              boxes, obj_probs, class_ids, conf_thresh_,
                                              layout.channels_first, v8_box_format_, apply_sigmoid,
                                              debug_det_, &debug_decode_left);
            } else {
                valid_count = ProcessOutputV8Int8(reinterpret_cast<int8_t*>(outputs[0].data.data()),
                                                  num_boxes, num_classes_,
                                                  model_input_h_, model_input_w_,
                                                  boxes, obj_probs, class_ids, conf_thresh_,
                                                  outputs[0].zp, outputs[0].scale,
                                                  layout.channels_first, v8_box_format_,
                                                  debug_det_, &debug_decode_left);
            }
        }

        if (valid_count <= 0) return;

        std::vector<int> indices(valid_count);
        for (int i = 0; i < valid_count; ++i) indices[i] = i;

        QuickSortDescending(obj_probs, 0, valid_count - 1, indices);

        std::set<int> class_set(class_ids.begin(), class_ids.end());
        for (int c : class_set) {
            NMS(valid_count, boxes, class_ids, indices, c, nms_thresh_);
        }

        const DetCoordContext coord_ctx = BuildDetCoordContext(*frame, model_input_w_, model_input_h_);

        auto det_result = std::make_shared<DetectionResult>();
        det_result->img_w = coord_ctx.out_w;
        det_result->img_h = coord_ctx.out_h;
        det_result->model_name = (yolo_version_ == YoloVersion::V5) ? "yolov5" : "yolov8";

        for (int i = 0; i < valid_count && det_result->items.size() < kMaxDetections; ++i) {
            if (indices[i] == -1) continue;
            int n = indices[i];
            int cls_id = class_ids[n];

            if (!class_filter_.empty() && class_filter_.find(cls_id) == class_filter_.end()) {
                continue;
            }

            float x1 = boxes[n * 4 + 0];
            float y1 = boxes[n * 4 + 1];
            float w = boxes[n * 4 + 2];
            float h = boxes[n * 4 + 3];

            Detection det;
            det.cls_id = cls_id;
            det.score = obj_probs[i];
            det.bbox = DecodeToOutputRect(x1, y1, w, h, coord_ctx);
            if (bbox_expand_.enable && det.cls_id == bbox_expand_.class_id) {
                det.bbox = ExpandRect(det.bbox, coord_ctx.out_w, coord_ctx.out_h, bbox_expand_);
            }
            det.track_id = -1;

            if (debug_det_ && det_result->items.size() < 5 && processed_ < 20) {
                LogInfo("[ai_yolo] det: raw(" + std::to_string(x1) + "," + std::to_string(y1) + "," +
                        std::to_string(w) + "," + std::to_string(h) + ") -> bbox(" +
                        std::to_string(det.bbox.x) + "," + std::to_string(det.bbox.y) + "," +
                        std::to_string(det.bbox.w) + "," + std::to_string(det.bbox.h) + ") cls=" +
                        std::to_string(cls_id) + " score=" + std::to_string(det.score));
            }

            det_result->items.push_back(det);
        }
        if (debug_det_ && processed_ < 20) {
            LogInfo("[ai_yolo] det summary(copy): valid_count=" + std::to_string(valid_count) +
                    " final=" + std::to_string(det_result->items.size()));
        }

        frame->det = det_result;
    }
#endif

    std::string id_;
    std::string model_path_;
    float conf_thresh_ = 0.25f;
    float nms_thresh_ = 0.45f;
    int model_input_w_ = 640;
    int model_input_h_ = 640;
    int num_classes_ = 80;
    V8BoxFormat v8_box_format_ = V8BoxFormat::CxCyWh;
    V8ClsActivation v8_cls_activation_ = V8ClsActivation::Auto;
    YoloVersion yolo_version_ = YoloVersion::V8;
    bool auto_detect_version_ = false;
    std::set<int> class_filter_;

    std::shared_ptr<SpscQueue<FramePtr>> input_queue_;
    std::vector<std::shared_ptr<SpscQueue<FramePtr>>> output_queues_;
    std::shared_ptr<IImageProcessor> image_processor_;
    std::shared_ptr<IInferBackend> infer_backend_;
    uint64_t processed_ = 0;

    bool stats_log_ = false;
    uint64_t stats_interval_ = 100;
    bool debug_det_ = false;
    BboxExpandConfig bbox_expand_{};

    int64_t infer_interval_ms_ = 0;
    int64_t infer_phase_ms_ = 0;
    int64_t last_infer_pts_ms_ = 0;
    bool use_dma_input_ = true;

#if defined(RK3588_ENABLE_RKNN)
    ModelHandle model_handle_ = kInvalidModelHandle;
    uint32_t n_output_ = 0;
    std::vector<uint8_t> resized_input_;
    std::vector<float> fp32_buffer_;  // For FP16 to FP32 conversion
#endif
};

REGISTER_NODE(AiYoloNode, "ai_yolo");

}  // namespace rk3588