fix(hdr): make dynamic metadata scene-aware

docs/dolby_vision_profile81.md

@@ -180,8 +180,8 @@ L1 偏移也相同 —— 但这只是巧合，不应成为依赖。`configure()
 现有统计不能机械复制成 DV 的语义：
 
 - **`min_pq` 不能用绝对最小像素。** 游戏中一个黑色 UI 像素、黑边或透明合成区域就能
-  把整帧最小值钉死在零。用「原始最小值 + 低百分位亮度 + 黑像素比例 + 时间稳定器」
-  共同生成稳健最小值。
+  把整帧最小值钉死在零。分析器同时给出 PQ 第 1 百分位和首个 PQ 直方图 bin 的覆盖率；
+  覆盖率达到 1% 才报告零，否则采用第 1 百分位。旧分析结果没有扩展统计时回退到 P10。
 - **`avg_pq` 必须是 PQ 域平均，不是 `average_maxrgb`。** PQ 是凹函数，
   `PQ(mean(nits)) ≥ mean(PQ(nits))`，暗场带高光时差距是整个动态范围的大部分。
   现有 `stats.avg_maxrgb_pq` 正是逐像素累加的 PQ 域平均，直接用它。
@@ -425,8 +425,9 @@ OPPO 真机解码链路 + Sony 电视 Dolby Vision 点亮）；主机侧灰度
 统计缺失时的保守 RPU、RPU/frame_id 严格匹配、零分配优化。
 
 已落地：avg/max/min 推导与钳位（§3.2）、统计缺失复用上次有效值、首帧预热跳过、
-frame_index 严格绑定、队列溢出即停、稳态零分配。待实机调优：场景切换信号
-（当前 scene_refresh 恒 false，EMA 的切场检测尚未导出）。
+frame_index 严格绑定、队列溢出即停、稳态零分配。`scene_refresh` 由独立 GPU 样本的
+PQ 均值、P10/P90 与 HDR10+ 分位分布共同判定；重复使用同一分析样本不会重复刷新。
+待实机调优：按游戏类型校准切场阈值与近黑覆盖率阈值。
 
 ### Phase 3：正式协议协商与降级（协商层已落地）
 

src/platform/common.h

@@ -522,6 +522,12 @@ namespace platf {
     /// and the histogram is point-sampled per analysis cell. Zero alongside a nonzero
     /// avg_maxrgb means the analyzer did not produce one, which HDR Vivid rejects.
     float avg_maxrgb_pq = 0.0f;
+    /// First percentile and near-black coverage from the analyzer's PQ histogram.
+    /// `near_black_fraction` is the fraction in histogram bin zero (PQ < 1/256).
+    /// The validity bit distinguishes an older analyzer from a genuinely black frame.
+    float percentile_1_pq = 0.0f;
+    float near_black_fraction = 0.0f;
+    bool near_black_stats_valid = false;
     float percentile_10_pq = 0.0f;  ///< 10th percentile in normalized PQ signal space
     float percentile_90_pq = 0.0f;  ///< 90th percentile in normalized PQ signal space
     /// 99th percentile of maxRGB (nits). Reported as the HDR10+ maxSCL; see
@@ -531,6 +537,8 @@ namespace platf {
     float distribution_maxrgb[HDR10PLUS_PERCENTILES] = {};
     float analysis_max_nits = 0.0f; ///< Upper luminance bound used by the analyzer
     uint64_t sample_sequence = 0;   ///< Increments only when a new GPU readback completes
+    uint64_t analyzed_frame_sequence = 0; ///< Capture-side frame sequence represented by this sample
+    uint32_t sample_age_frames = 0; ///< Capture frames elapsed since analyzed_frame_sequence
     bool valid = false;         ///< Whether stats are available (false on first frame)
   };
 

src/platform/windows/display_vram.cpp

@@ -417,19 +417,25 @@ namespace platf::dxgi {
       if (!img.blank) {
         auto &img_ctx = img_ctx_map[img.id];
         const bool can_analyze_hdr_frame = hdr_analysis_enabled && img.linear_gamma && img.format == DXGI_FORMAT_R16G16B16A16_FLOAT;
+        const uint64_t hdr_current_frame_sequence = ++hdr_capture_frame_sequence;
 
         // Open the shared capture texture with our ID3D11Device
         if (initialize_image_context(img, img_ctx)) {
           return -1;
         }
 
         // Poll the previous analysis result before taking the capture mutex.
-        if (hdr_analysis_pending) {
-          read_hdr_analysis_results();
-          if (hdr_luminance_stats_out.valid && !runtime_status.scene_metadata_active) {
-            runtime_status.scene_metadata_active = true;
-            ::video::update_hdr_pipeline_status(runtime_status_id, runtime_status);
-          }
+        read_hdr_analysis_results(hdr_current_frame_sequence);
+        if (hdr_luminance_stats_out.valid) {
+          const uint64_t age = hdr_current_frame_sequence >= hdr_luminance_stats_out.analyzed_frame_sequence ?
+                                 hdr_current_frame_sequence - hdr_luminance_stats_out.analyzed_frame_sequence :
+                                 0;
+          hdr_luminance_stats_out.sample_age_frames = static_cast<uint32_t>(
+            std::min<uint64_t>(age, std::numeric_limits<uint32_t>::max()));
+        }
+        if (hdr_luminance_stats_out.valid && !runtime_status.scene_metadata_active) {
+          runtime_status.scene_metadata_active = true;
+          ::video::update_hdr_pipeline_status(runtime_status_id, runtime_status);
         }
 
         // Acquire encoder mutex to synchronize with capture code. Normal
@@ -600,7 +606,7 @@ namespace platf::dxgi {
         }
 
         if (hdr_analysis_source) {
-          dispatch_hdr_analysis(hdr_analysis_source);
+          dispatch_hdr_analysis(hdr_analysis_source, hdr_current_frame_sequence);
         }
       }
 
@@ -788,8 +794,10 @@ namespace platf::dxgi {
       ::video::unregister_hdr_pipeline_status(runtime_status_id);
       runtime_status_id = 0;
       hdr_luminance_stats_out = {};
-      hdr_analysis_pending = false;
+      hdr_staging_pending.fill(false);
+      hdr_staging_frame_sequences.fill(0);
       hdr_analysis_frame_index = 0;
+      hdr_capture_frame_sequence = 0;
       hdr_analysis_sample_sequence = 0;
 
       // init() builds the analyzer from the pixel format alone, because the
@@ -1712,16 +1720,20 @@ namespace platf::dxgi {
     uav_t hdr_final_result_uav;            // UAV view for pass 2 output
     buf_t hdr_global_histogram_buf;        // 256-bin PQ histogram accumulated by pass 1 atomics
     uav_t hdr_global_histogram_uav;        // Typed R32_UINT UAV (clearable + atomic-capable)
-    buf_t hdr_staging_buf;                 // Staging buffer for CPU readback (1 FinalResult)
+    // A small ring prevents a slow asynchronous Map() from making a later
+    // CopyResource overwrite the still-pending result it was meant to read.
+    std::array<buf_t, 4> hdr_staging_bufs;
+    std::array<uint64_t, 4> hdr_staging_frame_sequences {};
+    std::array<bool, 4> hdr_staging_pending {};
     buf_t hdr_analysis_cbuf;               // Constant buffer for pass 1 (analysis resolution)
     buf_t hdr_analysis_snapshot_cbuf;      // Shared converter/pass 1 params for the snapshot
     buf_t hdr_reduce_cbuf;                 // Constant buffer for pass 2 (numGroups)
     uint32_t hdr_analysis_width = 0;       // Analysis grid width (downsampled from source)
     uint32_t hdr_analysis_height = 0;      // Analysis grid height (downsampled from source)
     uint32_t hdr_num_groups = 0;           // Number of thread groups dispatched in pass 1
     uint64_t hdr_analysis_frame_index = 0; // Used to downsample analysis frequency
+    uint64_t hdr_capture_frame_sequence = 0; // Counts captured frames presented to conversion
     uint64_t hdr_analysis_sample_sequence = 0; // Counts completed, independent GPU samples
-    bool hdr_analysis_pending = false;     // Whether we have results ready to read
     bool hdr_analysis_ready = false;       // Whether the analyzer's GPU resources were created
     bool hdr_analysis_enabled = false;     // Whether analysis runs: resources exist and the stream can carry metadata
     ::video::hdr_metadata::formats_t hdr_metadata_formats;  // Dynamic metadata formats this stream may carry
@@ -1996,16 +2008,18 @@ namespace platf::dxgi {
         return -1;
       }
 
-      // --- Staging buffer for async CPU readback (1 FinalResult only) ---
+      // --- Staging ring for asynchronous CPU readback ---
       D3D11_BUFFER_DESC staging_desc = {};
       staging_desc.ByteWidth = sizeof(FinalResult);
       staging_desc.Usage = D3D11_USAGE_STAGING;
       staging_desc.CPUAccessFlags = D3D11_CPU_ACCESS_READ;
 
-      status = device->CreateBuffer(&staging_desc, nullptr, &hdr_staging_buf);
-      if (FAILED(status)) {
-        BOOST_LOG(warning) << "Failed to create HDR staging buffer: " << util::log_hex(status);
-        return -1;
+      for (auto &staging_buffer : hdr_staging_bufs) {
+        status = device->CreateBuffer(&staging_desc, nullptr, &staging_buffer);
+        if (FAILED(status)) {
+          BOOST_LOG(warning) << "Failed to create HDR staging buffer: " << util::log_hex(status);
+          return -1;
+        }
       }
 
       // Resources exist; whether they get used is init_output()'s call, once the
@@ -2015,7 +2029,8 @@ namespace platf::dxgi {
                       << ", analysis " << hdr_analysis_width << "x" << hdr_analysis_height
                       << ", " << hdr_num_groups << " groups (" << groups_x << "x" << groups_y << ")"
                       << ", interval 1/" << HDR_ANALYSIS_INTERVAL
-                      << ", staging: " << sizeof(FinalResult) << " bytes";
+                      << ", staging ring: " << hdr_staging_bufs.size() << " x "
+                      << sizeof(FinalResult) << " bytes";
       return 0;
     }
 
@@ -2052,9 +2067,16 @@ namespace platf::dxgi {
      * @param source Unified full-frame or snapshot analysis input.
      */
     void
-    dispatch_hdr_analysis(const HdrAnalysisSource &source) {
+    dispatch_hdr_analysis(const HdrAnalysisSource &source, uint64_t frame_sequence) {
       if (!hdr_analysis_enabled || !source) return;
 
+      const auto free_slot = std::find(hdr_staging_pending.begin(), hdr_staging_pending.end(), false);
+      if (free_slot == hdr_staging_pending.end()) {
+        BOOST_LOG(debug) << "HDR analysis readback ring full; skipping capture frame " << frame_sequence;
+        return;
+      }
+      const size_t staging_index = static_cast<size_t>(free_slot - hdr_staging_pending.begin());
+
       // Unbind render targets to avoid resource hazard (SRV vs RTV conflict)
       ID3D11RenderTargetView *null_rtv = nullptr;
       device_ctx->OMSetRenderTargets(1, &null_rtv, nullptr);
@@ -2108,10 +2130,11 @@ namespace platf::dxgi {
       device_ctx->CSSetConstantBuffers(0, 1, &null_cb);
       device_ctx->CSSetShader(nullptr, nullptr, 0);
 
-      // Copy final result to staging buffer for CPU readback next frame
-      device_ctx->CopyResource(hdr_staging_buf.get(), hdr_final_result_buf.get());
-
-      hdr_analysis_pending = true;
+      // Capture this dispatch in its own staging slot. GPU command ordering keeps
+      // the shared final buffer safe while the CPU may still be reading older slots.
+      device_ctx->CopyResource(hdr_staging_bufs[staging_index].get(), hdr_final_result_buf.get());
+      hdr_staging_frame_sequences[staging_index] = frame_sequence;
+      hdr_staging_pending[staging_index] = true;
     }
 
     /**
@@ -2120,23 +2143,48 @@ namespace platf::dxgi {
      * and computes PQ-domain percentiles from the histogram.
      */
     void
-    read_hdr_analysis_results() {
+    read_hdr_analysis_results(uint64_t current_frame_sequence) {
+      while (true) {
+        size_t staging_index = hdr_staging_pending.size();
+        for (size_t i = 0; i < hdr_staging_pending.size(); ++i) {
+          if (hdr_staging_pending[i] &&
+              (staging_index == hdr_staging_pending.size() ||
+               hdr_staging_frame_sequences[i] < hdr_staging_frame_sequences[staging_index])) {
+            staging_index = i;
+          }
+        }
+        if (staging_index == hdr_staging_pending.size()) {
+          return;
+        }
+
+        if (!read_hdr_analysis_result(staging_index, current_frame_sequence)) {
+          return;  // Oldest result is still on the GPU; later copies cannot be ready yet.
+        }
+      }
+    }
+
+    bool
+    read_hdr_analysis_result(size_t staging_index, uint64_t current_frame_sequence) {
       D3D11_MAPPED_SUBRESOURCE mapped = {};
-      HRESULT status = device_ctx->Map(hdr_staging_buf.get(), 0, D3D11_MAP_READ, D3D11_MAP_FLAG_DO_NOT_WAIT, &mapped);
+      HRESULT status = device_ctx->Map(
+        hdr_staging_bufs[staging_index].get(), 0,
+        D3D11_MAP_READ, D3D11_MAP_FLAG_DO_NOT_WAIT, &mapped);
 
       if (status == DXGI_ERROR_WAS_STILL_DRAWING) {
         // GPU hasn't finished yet — skip this readback, try next frame
-        return;
+        return false;
       }
 
       if (FAILED(status)) {
         BOOST_LOG(debug) << "HDR staging Map failed: " << util::log_hex(status);
-        return;
+        hdr_staging_pending[staging_index] = false;
+        return true;
       }
 
       auto *result = reinterpret_cast<const FinalResult *>(mapped.pData);
 
       if (result->pixelCount > 0) {
+        hdr_luminance_stats_out = {};
         hdr_luminance_stats_out.min_maxrgb = result->minMaxRGB;
         hdr_luminance_stats_out.max_maxrgb = result->maxMaxRGB;
         hdr_luminance_stats_out.avg_maxrgb = result->sumMaxRGB / static_cast<float>(result->pixelCount);
@@ -2164,6 +2212,9 @@ namespace platf::dxgi {
         // Retain P99 in nits for the independent HDR10+ path, and fill the nine
         // percentiles ST 2094-40 deployment profiles carry from the same walk.
         const uint32_t total = result->pixelCount;
+        hdr_luminance_stats_out.near_black_fraction =
+          static_cast<float>(result->histogram[0]) / static_cast<float>(total);
+        hdr_luminance_stats_out.near_black_stats_valid = true;
         const auto &percentages = ::video::hdr_metadata::hdr10plus_percentages;
         constexpr size_t kDistCount = percentages.size();
 
@@ -2187,6 +2238,9 @@ namespace platf::dxgi {
             if (!dist_found[p] && cumulative >= dist_targets[p]) {
               hdr_luminance_stats_out.distribution_maxrgb[p] =
                 ::video::hdr_metadata::pq_to_nits(pq_bin_center);
+              if (p == 0) {
+                hdr_luminance_stats_out.percentile_1_pq = pq_bin_center;
+              }
               dist_found[p] = true;
             }
           }
@@ -2212,11 +2266,18 @@ namespace platf::dxgi {
 
         hdr_luminance_stats_out.analysis_max_nits = hdr_analysis_max_nits;
         hdr_luminance_stats_out.sample_sequence = ++hdr_analysis_sample_sequence;
+        hdr_luminance_stats_out.analyzed_frame_sequence = hdr_staging_frame_sequences[staging_index];
+        const uint64_t age = current_frame_sequence >= hdr_luminance_stats_out.analyzed_frame_sequence ?
+                               current_frame_sequence - hdr_luminance_stats_out.analyzed_frame_sequence :
+                               0;
+        hdr_luminance_stats_out.sample_age_frames = static_cast<uint32_t>(
+          std::min<uint64_t>(age, std::numeric_limits<uint32_t>::max()));
         hdr_luminance_stats_out.valid = true;
       }
 
-      device_ctx->Unmap(hdr_staging_buf.get(), 0);
-      hdr_analysis_pending = false;
+      device_ctx->Unmap(hdr_staging_bufs[staging_index].get(), 0);
+      hdr_staging_pending[staging_index] = false;
+      return true;
     }
 
     // ===== Compute-shader RGB->P010 fast path (Phase 1) =====

src/video_dolby_vision.cpp

@@ -152,19 +152,27 @@ namespace video::dolby_vision {
   }
 
   std::optional<frame_metadata_t>
-  frame_metadata_from_stats(const platf::hdr_frame_luminance_stats_t &stats) {
+  frame_metadata_from_stats(
+    const platf::hdr_frame_luminance_stats_t &stats,
+    bool scene_refresh) {
     if (!stats.valid) {
       return std::nullopt;
     }
 
     const bool finite = std::isfinite(stats.avg_maxrgb_pq) &&
                         std::isfinite(stats.avg_maxrgb) &&
                         std::isfinite(stats.percentile_99) &&
-                        std::isfinite(stats.percentile_10_pq);
+                        std::isfinite(stats.percentile_10_pq) &&
+                        (!stats.near_black_stats_valid ||
+                          (std::isfinite(stats.percentile_1_pq) &&
+                           std::isfinite(stats.near_black_fraction)));
     if (!finite ||
         stats.avg_maxrgb_pq < 0.0f || stats.avg_maxrgb_pq > 1.0f ||
         stats.percentile_99 < 0.0f ||
-        stats.percentile_10_pq < 0.0f || stats.percentile_10_pq > 1.0f) {
+        stats.percentile_10_pq < 0.0f || stats.percentile_10_pq > 1.0f ||
+        (stats.near_black_stats_valid &&
+          (stats.percentile_1_pq < 0.0f || stats.percentile_1_pq > 1.0f ||
+           stats.near_black_fraction < 0.0f || stats.near_black_fraction > 1.0f))) {
       return std::nullopt;
     }
     // Zero PQ-domain mean beside a positive linear mean means the analyzer
@@ -174,11 +182,17 @@ namespace video::dolby_vision {
     }
 
     frame_metadata_t raw;
-    raw.min_pq = pq_signal_u12_rounded(stats.percentile_10_pq);
+    constexpr float meaningful_near_black_coverage = 0.01f;
+    const float robust_min_pq = stats.near_black_stats_valid ?
+                                  (stats.near_black_fraction >= meaningful_near_black_coverage ?
+                                     0.0f : stats.percentile_1_pq) :
+                                  stats.percentile_10_pq;
+    raw.min_pq = pq_signal_u12_rounded(robust_min_pq);
     raw.max_pq = pq_code_u12_rounded(stats.percentile_99);
     raw.avg_pq = pq_signal_u12_rounded(stats.avg_maxrgb_pq);
-    raw.scene_refresh = false;
-    return clamp_level1(raw.min_pq, raw.max_pq, raw.avg_pq);
+    auto result = clamp_level1(raw.min_pq, raw.max_pq, raw.avg_pq);
+    result.scene_refresh = scene_refresh;
+    return result;
   }
 
   bool
@@ -459,14 +473,23 @@ namespace video::dolby_vision {
     // Missing analysis reuses the last good values: once RPUs are flowing,
     // a frame without one would make the client's Dolby engine fall back to
     // static HDR10 mapping for that frame — a visible brightness step.
-    if (const auto metadata = frame_metadata_from_stats(stats)) {
+    const auto scene = scene_detector_.observe(stats);
+    if (scene.new_sample && scene.scene_change) {
+      luminance_filter_.reset();
+    }
+    luminance_filter_.update(stats);
+    if (const auto metadata = frame_metadata_from_stats(
+          luminance_filter_.smoothed(stats), scene.scene_change)) {
       last_metadata_ = metadata;
     }
     if (!last_metadata_) {
       return;  // cold analyzer: this frame ships without an RPU, like HDR10+ does
     }
 
-    if (!queue_.stage(frame_index, generator_, *last_metadata_)) {
+    const auto metadata_for_frame = *last_metadata_;
+    // Reused luminance is intentional, but a refresh belongs to one picture.
+    last_metadata_->scene_refresh = false;
+    if (!queue_.stage(frame_index, generator_, metadata_for_frame)) {
       // In-flight overflow means the encoder's output can no longer be
       // trusted to surface in order; stop rather than risk a stale RPU
       // landing on a newer picture (docs §3.5).
@@ -496,6 +519,8 @@ namespace video::dolby_vision {
   void
   rpu_injector_t::disable() {
     enabled_ = false;
+    scene_detector_.reset();
+    luminance_filter_.reset();
     last_metadata_.reset();
     queue_.clear();
     generator_.reset();