# 基准测试脚本片段（固定 1080p30 YUV420 10bit）
for preset in {4..8}; do
  ./SvtAv1EncApp -i input.yuv -w 1920 -h 1080 -b 10 -preset $preset 
    -rc 1 -tbr 5000 -fps 30 -tile_columns 2 -tile_rows 1 
    --stat-file preset_${preset}.json
done

{
  "rc_mode": "CBR",
  "target_bitrate": 5000,        // kbps
  "vbv_bufsize": 10000,          // 2× 目标码率，缓冲 2s
  "vbv_maxrate": 5500,           // 允许 10% 瞬时突发
  "qp_min": 18, "qp_max": 42,    // 防止极端 QP 导致画质崩塌
  "min_qp_allowed": 10,          // 关键帧最低 QP 保护
  "max_qp_allowed": 48
}

# 双路 2×32 核，NPS=4 模式，每 NUMA 节点跑 1 实例
numactl --cpunodebind=0 --membind=0 
  ./SvtAv1EncApp -tile_columns 2 -tile_rows 1 -pin 1 ...

numactl --cpunodebind=1 --membind=1 
  ./SvtAv1EncApp -tile_columns 2 -tile_rows 1 -pin 1 ...

┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│  输入采集    │────▶│  预处理/分发  │────▶│  编码调度器  │
│  (SDI/NDI)  │     │  (Scaler/    │     │  (决策引擎)  │
└─────────────┘     │   Colorspace)│     └──────┬──────┘
                    └──────────────┘            │
                           │                    ▼
                    ┌──────┴──────┐    ┌────────────────┐
                    ▼             ▼    ▼                ▼
             ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐
             │ CPU Worker│  │ CPU Worker│  │ GPU Enc  │  │ GPU Enc  │
             │ (Preset 6)│  │ (Preset 5)│  │ (NVENC)  │  │ (QSV)    │
             └──────────┘  └──────────┘  └──────────┘  └──────────┘
                    │             │            │            │
                    └──────┬──────┴────────────┴────────────┘
                           ▼
                    ┌──────────────┐
                    │  码流合规/封装 │
                    │  (MPEG-TS/    │
                    │   fMP4/HLS)  │
                    └──────────────┘

func (s *Scheduler) SelectEncoder(ctx context.Context, job *EncodeJob) (Encoder, error) {
    // 1. 特征提取
    feat := s.extractFeatures(job)
    
    // 2. 评分模型（轻量级 XGBoost/LightGBM）
    scores := s.scoringModel.Predict(feat)
    
    // 3. 硬约束过滤
    candidates := s.filterByHardConstraints(scores, job.SLA)
    
    // 4. 软约束加权：成本、延迟、质量
    best := s.softSelect(candidates, job.Priority)
    
    // 5. 熔断保护
    if s.isCircuitOpen(best.Type) {
        return s.fallbackEncoder(job)
    }
    return best, nil
}

# 多阶段构建：编译环境 → 运行环境
FROM intel/oneapi-basekit:2024.2 AS builder
ARG SVT_AV1_VERSION=v1.7.0
RUN git clone --depth 1 -b $SVT_AV1_VERSION https://github.com/AOMediaCodec/SVT-AV1.git 
 && cd SVT-AV1/Build/linux && cmake -DCMAKE_BUILD_TYPE=Release -DENABLE_AVX512=ON .. 
 && make -j$(nproc) && make install

FROM ubuntu:22.04 AS runtime
# 仅复制运行时依赖 + 编码器二进制
COPY --from=builder /usr/local/bin/SvtAv1EncApp /usr/local/bin/
COPY --from=builder /usr/local/lib/libSvtAv1Enc.so* /usr/local/lib/
# 硬件加速运行时依赖
RUN apt-get update && apt-get install -y --no-install-recommends 
    intel-media-va-driver-non-free vainfo clinfo ocl-icd-libopencl1 
 && rm -rf /var/lib/apt/lists/*
# 非 root 运行 + 资源限制
USER 1000:1000
ENTRYPOINT ["SvtAv1EncApp"]

resources:
  limits:
    cpu: "16"
    memory: "32Gi"
    nvidia.com/gpu: 1          # 或 intel.com/gpu: 1
  requests:
    cpu: "12"
    memory: "24Gi"
env:
- name: GOMAXPROCS
  value: "12"                  # 留 4 核给系统/网络中断
- name: SVT_AV1_PIN_THREADS
  value: "1"

┌─────────────────────────────────────────────────────────────┐
│ SVT-AV1 实时编码「一页纸」速查表                             │
├──────────────┬──────────────────────────────────────────────┤
│ 场景         │ 推荐配置                                       │
├──────────────┼──────────────────────────────────────────────┤
│ 720p30 直播  │ preset=7, tile_col=2, CBR 2.5M, vbv=5M       │
│ 1080p30 直播 │ preset=6, tile_col=2, CBR 5M, vbv=10M        │
│ 1080p60 电竞 │ preset=5, tile_col=4, CBR 8M, vbv=16M        │
│ 4K30 精品    │ preset=4, tile_col=4, CVBR 20M, lookahead=10 │
├──────────────┼──────────────────────────────────────────────┤
│ GPU 兜底     │ NVENC P6 / QSV veryfast / VCN balanced       │
│ NUMA 绑核    │ numactl --cpunodebind=N --membind=N          │
│ 监控红线     │ P99延迟>帧预算80% / VMAF<90 / GPU显存>90%    │
│ 熔断降级     │ Preset+1 → 切CPU → 降码率 → 丢帧保音频       │
└──────────────┴──────────────────────────────────────────────┘

# 伪代码：基于主观评分 (MOS) 的模型微调
def retrain_vmaf_model():
    # 1. 采样：从生产流量按内容标签分层抽样 2000 条/月
    samples = fetch_labeled_samples(strata=['sport', 'anime', 'movie'], n=2000)
    
    # 2. 编码：用当前主力 Preset 生成失真样本 (CRF 28-45)
    distorted = batch_encode(samples, presets=[5,6,7], rc_modes=['CBR','VBR'])
    
    # 3. 主观标注：众包/专家打分 (ITU-T P.910), 目标 15 人/样本
    mos_scores = crowdsourced_mos(distorted, annotators=15)
    
    # 4. 特征工程：VMAF 基础特征 + 时域池化统计 + 频域小波能量
    feats = extract_enhanced_features(distorted)
    
    # 5. 训练：LightGBM Ranker 优化 Spearman 相关系数
    model = lgb.LGBMRanker(objective='lambdarank', metric='ndcg')
    model.fit(feats, mos_scores, group=query_ids)
    
    # 6. 影子验证：新模型 vs 旧模型在线 A/B 测试 7 天
    if shadow_validate(model, threshold_srcc=0.94):
        deploy_to_feature_store(model, version=f"vmaf_custom_{datetime.now():%Y%m}")

// EbSvtAv1EncProcessFrame() 主循环重构
typedef struct FramePipelineStage {
    pthread_t       tid;
    TaskQueue      *in_q, *out_q;
    EbSvtAv1EncConfiguration *cfg;
} FramePipelineStage;

static void *stage_me_kernel(void *arg) {
    FramePipelineStage *s = arg;
    while (1) {
        PictureControlSet *pcs = task_queue_pop(s->in_q);
        if (pcs->frame_type == FRAME_END) break;
        
        // 1. 运动估计 (支持 WPP 行级回调)
        svt_av1_motion_estimation(pcs, s->cfg->wpp_enable);
        
        // 2. 提前推送下一阶段：不等待全帧 ME 完成，按 TileGroup 推流
        for (int tg = 0; tg < pcs->tile_group_count; tg++) {
            if (pcs->tg_me_done[tg]) task_queue_push(s->out_q, pcs->tg_task[tg]);
        }
    }
    return NULL;
}

graph LR
    A[采集设备 /dev/videoX] -->|V4L2 dmabuf export| B(DMA-BUF FD)
    B --> C[VAAPI vaCreateSurfacesFromDmaBufs]
    C --> D[SVT-AV1 VAAPI Input Surface]
    D --> E[编码内核]
    E --> F[VAAPI Coded Buffer]
    F -->|vaExportSurfaceHandle| G(DMA-BUF FD)
    G --> H[RTP 发送线程 / sendmsg MSG_ZEROCOPY]

// 导入 dmabuf 到 VAAPI Surface
VASurfaceAttrib attribs[] = {
    {.type = VASurfaceAttribMemoryType, .value = {.type = VAGenericValueTypeInteger, .value.i = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2}},
    {.type = VASurfaceAttribDrmPrimeFd, .value = {.type = VAGenericValueTypeInteger, .value.i = dmabuf_fd}},
    {.type = VASurfaceAttribWidth, .value = {.type = VAGenericValueTypeInteger, .value.i = width}},
    {.type = VASurfaceAttribHeight, .value = {.type = VAGenericValueTypeInteger, .value.i = height}},
    {.type = VASurfaceAttribFormat, .value = {.type = VAGenericValueTypeInteger, .value.i = VA_FOURCC_NV12}},
};
vaCreateSurfaces(va_dpy, VA_RT_FORMAT_YUV420, width, height, &surface_id, 1, attribs, 5);

输入帧 (1080p)
    │
    ▼
┌─────────────────────────────────────┐
│  共享分析层                          │
│  - 全分辨率 ME (1/4 亚像素)          │
│  - 场景切换检测 (SAD 峰值)           │
│  - 内容复杂度图 (CTU 级 SATD)        │
│  - ROI 检测 (人脸/文字/Logo)         │
└──────────────┬──────────────────────┘
               │ 共享元数据 (MV, Mode, SATD, ROI Mask)
               ▼
    ┌──────────┼──────────┐
    ▼          ▼          ▼
┌───────┐ ┌───────┐ ┌───────┐
│ 1080p │ │ 720p  │ │ 480p  │  ← 仅执行 RDO + 量化 + 熵编码
│ 6M    │ │ 3M    │ │ 1.2M  │
└───┬───┘ └───┬───┘ └───┬───┘
    │         │         │
    ▼         ▼         ▼
┌─────────────────────────────────────┐
│  码率控制协调器                       │
│  - 总带宽预算分配 (Water-filling)    │
│  - 跨层 QP 约束 (ΔQP ≤ 4)            │
│  - 关键帧强制对齐                    │
└─────────────────────────────────────┘

def generate_abr_ladder(content_features, total_bw_budget, device_caps):
    """
    content_features: {si, ti, duration, genre, target_devices}
    return: [{res, fps, target_bitrate, max_bitrate, preset}, ...]
    """
    # 1. 复杂度建模：预测各分辨率下的 R-D 曲线参数 (a, b, c) -> QP = a * log(R) + b * R + c
    rd_models = predict_rd_models(content_features)  # 轻量级 MLP, < 1ms
    
    # 2. 约束规划：整数规划求解
    # Max Σ Quality_i * Weight(device_i)
    # s.t. Σ Bitrate_i ≤ Budget
    #      Bitrate_i ∈ Discrete_Set (编码器支持档位)
    #      Quality_i = VMAF(rd_models[res_i], Bitrate_i)
    #      Resolution_Order: 1080p > 720p > 540p > 360p
    
    ladder = solve_knapsack_abr(rd_models, total_bw_budget, device_caps)
    
    # 3. 兜底策略：确保最低档位覆盖弱网设备
    if min(r['target_bitrate'] for r in ladder) > 400: # kbps
        ladder.append({'res': '320x180', 'fps': 15, 'target_bitrate': 200, 'preset': 8})
    
    return ladder

// encoder_control.proto
service EncoderControl {
  rpc StreamConfigUpdates(stream ConfigRequest) returns (stream ConfigResponse);
}

message ConfigRequest {
  string session_id = 1;
  repeated RenditionConfig renditions = 2;  // 当前生效的梯度配置
  GlobalRCConfig global_rc = 3;             // 总带宽预算、缓冲策略
  FeatureFlags features = 4;                // 开关：WPP, SAO, CDEF, Film_Grain
}

message RenditionConfig {
  string rendition_id = 1;
  int32 width = 2; int32 height = 3; int32 fps_num = 4; int32 fps_den = 5;
  int64 target_bps = 6; int64 max_bps = 7; int64 buf_size = 8;
  int32 preset = 9; int32 tile_cols = 10; int32 tile_rows = 11;
  bool force_idr = 12;  // 场景切换/求关键帧信令
}

type VirtualEncodingUnit interface {
    // 能力声明
    Capabilities() VEUCapability
    // 同步编码（低延迟路径）
    EncodeFrame(ctx context.Context, frame *VideoFrame, params EncodeParams) (*EncodedPacket, error)
    // 异步流式编码（高吞吐路径）
    StartStream(ctx context.Context, config StreamConfig) (StreamHandle, error)
    PushFrame(handle StreamHandle, frame *VideoFrame) error
    PullPacket(handle StreamHandle) (*EncodedPacket, error)
    // 生命周期
    HealthCheck() error
    Shutdown(graceful bool)
}

type VEUCapability struct {
    Codecs          []string  // ["av1", "hevc", "h264"]
    MaxResolution   Resolution
    MaxFramerate    int
    Presets         []int     // 支持的 preset 范围
    RCModes         []RCMode  // CBR, VBR, CVBR, CQP
    HardwareType    string    // "cpu-avx512", "nvidia-nvenc", "intel-qsv", "xilinx-vcu"
    PowerProfile    string    // "performance", "balanced", "power-saver"
    InstanceID      string    // 唯一标识
    NumaNode        int       // 亲和性提示
}

编码输出包 → [eBPF Hook] → 切片缓冲 (2s) → [异步推理: NSFW/暴恐/敏感人脸/Logo] → 
    ├─ 通过 → 释放缓冲 → CDN 推流
    └─ 拦截 → 注入替代内容 (色条/公益广告) → 触发告警 → 审计日志入库

#!/bin/bash
# 用法: ./diag_encoder.sh <pid> [duration_sec]
PID=$1
DUR=${2:-30}
OUT_DIR="diag_$(date +%Y%m%d_%H%M%S)"
mkdir -p $OUT_DIR

echo "[1/6] 采集 perf 火焰图..."
perf record -g -p $PID -- sleep $DUR -o $OUT_DIR/perf.data
perf script -i $OUT_DIR/perf.data | ./FlameGraph/stackcollapse-perf.pl | ./FlameGraph/flamegraph.pl > $OUT_DIR/flamegraph.svg

echo "[2/6] 采集延迟分布..."
perf stat -e cycles,instructions,cache-references,cache-misses,branch-misses -p $PID -- sleep $DUR 2>&1 | tee $OUT_DIR/perf_stat.txt

echo "[3/6] 采集内存/线程快照..."
gdb -batch -ex "set pagination off" -ex "info threads" -ex "thread apply all bt" -p $PID > $OUT_DIR/gdb_bt.txt 2>&1
cat /proc/$PID/status > $OUT_DIR/proc_status.txt
cat /proc/$PID/smaps_rollup > $OUT_DIR/smaps.txt

echo "[4/6] 采集 GPU 状态 (如适用)..."
nvidia-smi dmon -s pucvmet -d 1 -c $DUR > $OUT_DIR/nvidia_dmon.csv 2>/dev/null || true
intel_gpu_top -J -o $OUT_DIR/intel_gpu.json -d $DUR 2>/dev/null || true

echo "[5/6] 采集编码器内部统计 (需应用暴露 HTTP /metrics)..."
curl -s http://localhost:9090/metrics | grep -E 'encode_(latency|fps|error|bitrate|qp)' > $OUT_DIR/app_metrics.txt

echo "[6/6] 打包..."
tar czf ${OUT_DIR}.tar.gz $OUT_DIR
echo "诊断包: ${OUT_DIR}.tar.gz"