@gracker/smartperfetto 1.0.13

package/skills/pipelines/android_view_multi_window.skill.yaml

@@ -0,0 +1,271 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+# Copyright (C) 2024-2026 Gracker (Chris)
+# This file is part of SmartPerfetto. See LICENSE for details.
+
+# =============================================================================
+# Android View Multi-Window Pipeline
+# =============================================================================
+# Same-process multi-window rendering (Dialog, PopupWindow).
+# RenderThread serialization may cause jank.
+# =============================================================================
+
+name: pipeline_android_view_multi_window
+version: "1.0"
+type: pipeline_definition
+category: rendering
+
+meta:
+  pipeline_id: ANDROID_VIEW_MULTI_WINDOW
+  display_name: "Android View 多窗口"
+  description: "同进程多窗口渲染 (Dialog/PopupWindow)，RenderThread 串行化"
+  icon: "window"
+  family: hwui
+  doc_path: "rendering_pipelines/android_view_multi_window.md"
+  # ----- Article-grounded signature (S06) -----
+  s_article_ref: "S06"
+  four_features:
+    producer_threads: ["main", "RenderThread"]   # 同进程多窗口共享单 RenderThread（进程级单例）
+    expected_layer_count: 2   # 主窗口 + 1+ 辅助（Dialog/PopupWindow/Activity Embedding 等）
+    bufferqueue_path: "BBQ_TRANSACTION_INPROC_MULTIPLE"   # 各窗口各自 BBQ
+    extra_rhythm_sources: []   # 同进程共享一条 vsync-app
+  deviation_anchors: "serialized_multi_viewroot_anchor_2_3_4_5"
+  # 锚点偏离：同一 doFrame 内多个 ViewRootImpl 的 ②③④⑤ 串行执行（callback 注册顺序决定）。
+  # 跨进程多窗口形态（split-screen/PiP/freeform）则是各进程独立 ①-⑥，⑧ SF 多 layer latch。
+  non_primary_note: |
+    此 pipeline ID 已在代码 NON_PRIMARY_PIPELINE_IDS（renderingPipelineDetectionSkillGenerator.ts:23）
+    中注册，不会被选为 primary，只会出现在 features list。
+  subvariants_note: |
+    文章 S06 严格区分（按窗口形态 + 进程拓扑两条轴）：
+    - MULTI_WINDOW_SAME_PROCESS（Dialog/PopupWindow/同 App 多 Activity/Activity Embedding）— 当前 ID 主要覆盖
+    - MULTI_WINDOW_SPLIT_SCREEN（分屏，通常跨进程）
+    - MULTI_WINDOW_PIP（Picture in Picture）
+    - MULTI_WINDOW_FREEFORM（自由窗口/桌面模式）
+    Phase E 拆分独立 ID。
+    决定性信号（同进程）：trace 同一 main thread 看到多次 performTraversals。
+    YAML 当前不支持"同 doFrame 内多次出现"语义，留 Phase B 主决策树用 SQL 时间窗口分组实现。
+
+detection:
+  required_signals:
+    - thread: RenderThread
+      min_count: 1
+    - thread: main
+      min_count: 1
+
+  scoring_signals:
+    - signal: has_dialog
+      slice_pattern: "*Dialog*"
+      weight: 20
+    - signal: has_popup_window
+      slice_pattern: "*PopupWindow*"
+      weight: 20
+    - signal: has_draw_frame
+      slice_pattern: "DrawFrame*"
+      weight: 20
+    - signal: has_perform_traversals_multi
+      slice_pattern: "*performTraversals*"
+      weight: 12
+      min_count: 2   # 至少 2 次以区分单窗口与多窗口（同 doFrame 多次的精确判断留 Phase B）
+
+teaching:
+  title: "Android View 多窗口渲染管线"
+  summary: |
+    同进程内存在多个窗口 (Dialog, PopupWindow, Toast)。由于共用一个 RenderThread，
+    多个窗口的渲染会串行化，可能导致帧率下降。需要关注窗口渲染顺序和时间分配。
+
+  mermaid: |
+    sequenceDiagram
+      participant VA as VSync-app
+      participant Main as App (main)
+      participant RT as RenderThread
+      participant BQ1 as BufferQueue (Main)
+      participant BQ2 as BufferQueue (Dialog)
+      participant VS as VSync-sf
+      participant SF as SurfaceFlinger
+
+      Note over VA,SF: 📍 Multi-Window (Dialog/PopupWindow)
+      VA->>Main: 🔔 VSync-app
+      activate Main
+      Main->>Main: 主窗口 Layout + Draw
+      Main->>RT: syncFrameState (主窗口)
+      Main->>Main: Dialog Layout + Draw
+      Main->>RT: syncFrameState (Dialog)
+      deactivate Main
+
+      activate RT
+      Note over RT: ⏱️ 串行处理多窗口
+      RT->>BQ1: queueBuffer (主窗口)
+      RT->>BQ2: queueBuffer (Dialog)
+      deactivate RT
+
+      VS->>SF: 🔔 VSync-sf
+      activate SF
+      SF->>SF: latchBuffer (多 Surface)
+      SF->>SF: HWC 合成
+      deactivate SF
+
+      Note over VA,SF: ⚠️ RenderThread 串行化可能导致卡顿
+  thread_roles:
+    - thread: main
+      role: "多窗口 UI 管理"
+      description: "管理多个窗口的 View 树"
+    - thread: RenderThread
+      role: "渲染线程 (串行)"
+      description: "串行渲染多个窗口，可能成为瓶颈"
+
+  key_slices:
+    - name: "Dialog"
+      thread: any
+      description: "对话框相关操作"
+    - name: "PopupWindow"
+      thread: any
+      description: "弹出窗口操作"
+    - name: "Choreographer#doFrame"
+      thread: main
+      description: "同一帧内可能连续驱动多个窗口的 layout/draw"
+    - name: "syncFrameState"
+      thread: RenderThread
+      description: "多个窗口共享 RenderThread，同步与提交容易串行化"
+    - name: "DrawFrame"
+      thread: RenderThread
+      description: "GPU 渲染阶段（多窗口会放大工作量）"
+
+auto_pin:
+  instructions:
+    # Time-ordered pinning (production -> transfer -> consume -> display)
+    - pattern: "^VSYNC-app$"
+      match_by: name
+      priority: 1
+      reason: "VSync (App 开始生产帧)"
+
+    - pattern: "^main(\\s+\\d+)?$"
+      match_by: name
+      priority: 2
+      reason: "App 主线程 (生产帧)"
+      expand: true  # 展开显示 Running 状态和 callstack
+      smart_filter:
+        enabled: true
+        description: "仅 Pin 活跃渲染进程的 main 线程"
+        detection_sql: |
+          SELECT DISTINCT p.name as process_name, COUNT(*) as frame_count
+          FROM slice s
+          JOIN thread_track tt ON s.track_id = tt.id
+          JOIN thread t ON tt.utid = t.utid
+          JOIN process p ON t.upid = p.upid
+          WHERE t.name = 'RenderThread'
+            AND s.name GLOB 'DrawFrame*'
+            AND p.name IS NOT NULL
+            AND p.name NOT LIKE 'com.android.systemui%'
+            AND p.name NOT LIKE 'system_server%'
+            AND p.name NOT LIKE '/system/%'
+          GROUP BY p.upid
+          HAVING frame_count > 5
+          ORDER BY frame_count DESC
+          LIMIT 10
+        fallback_sql: |
+          SELECT DISTINCT p.name as process_name, COUNT(*) as slice_count
+          FROM slice s
+          JOIN thread_track tt ON s.track_id = tt.id
+          JOIN thread t ON tt.utid = t.utid
+          JOIN process p ON t.upid = p.upid
+          WHERE t.name = 'main'
+          GROUP BY p.upid
+          HAVING slice_count > 10
+          ORDER BY slice_count DESC
+          LIMIT 10
+
+    - pattern: "^RenderThread(\\s+\\d+)?$"
+      match_by: name
+      priority: 3
+      reason: "App 渲染线程 (RenderThread)"
+      expand: true  # 展开显示 Running 状态和 callstack
+      smart_filter:
+        enabled: true
+        description: "仅 Pin 有活跃渲染的进程的 RenderThread"
+        detection_sql: |
+          SELECT DISTINCT p.name as process_name, COUNT(*) as frame_count
+          FROM slice s
+          JOIN thread_track tt ON s.track_id = tt.id
+          JOIN thread t ON tt.utid = t.utid
+          JOIN process p ON t.upid = p.upid
+          WHERE t.name = 'RenderThread'
+            AND s.name GLOB 'DrawFrame*'
+            AND p.name IS NOT NULL
+            AND p.name NOT LIKE 'com.android.systemui%'
+            AND p.name NOT LIKE 'system_server%'
+            AND p.name NOT LIKE '/system/%'
+          GROUP BY p.upid
+          HAVING frame_count > 5
+          ORDER BY frame_count DESC
+          LIMIT 10
+        fallback_sql: |
+          SELECT DISTINCT p.name as process_name, COUNT(*) as slice_count
+          FROM slice s
+          JOIN thread_track tt ON s.track_id = tt.id
+          JOIN thread t ON tt.utid = t.utid
+          JOIN process p ON t.upid = p.upid
+          WHERE t.name = 'RenderThread'
+            AND s.name GLOB 'DrawFrame*'
+          GROUP BY p.upid
+          HAVING slice_count > 10
+          ORDER BY slice_count DESC
+          LIMIT 10
+
+    - pattern: "^VSYNC-sf$"
+      match_by: name
+      priority: 5.5
+      reason: "VSync (SurfaceFlinger 消费/合成)"
+
+    - pattern: "^BufferTX"
+      match_by: name
+      priority: 6
+      reason: "BufferTX (SurfaceFlinger 事务)"
+
+    - pattern: "^[sS]urface[fF]linger"
+      match_by: name
+      priority: 7
+      reason: "SurfaceFlinger (最终合成/显示)"
+      main_thread_only: true
+
+analysis:
+  common_issues:
+    - id: rt_serialization
+      name: "RenderThread 串行化"
+      description: "多窗口导致 RenderThread 工作量翻倍"
+      detection_skill: render_thread_slices
+
+    # ----- Article-grounded issues (S06) -----
+    - id: traversal_callback_serialization
+      name: "performTraversals callback 串行竞争"
+      description: |
+        同一 doFrame 内多个 ViewRootImpl 的 performTraversals 在 main thread 串行执行
+        （注册顺序决定先后）。一个窗口 TRAVERSAL 变长会直接挤掉另一个窗口本帧的预算。
+        S06 §"同进程场景：多个 ViewRootImpl 在同一个 doFrame() 里串行执行"。
+      detection_skill: app_frame_production
+
+    - id: drawframe_serialization_on_renderthread
+      name: "RenderThread 进程级单例下 DrawFrame 串行排队"
+      description: |
+        HWUI RenderThread 是进程级单例（RenderThread::getInstance()），
+        同一进程所有窗口的 HardwareRenderer 都向同一个 RenderThread 投递任务。
+        多窗口 DrawFrame slice 在 RenderThread 上串行排队，主窗口滚动 + Dialog 动画同时发生时
+        哪一个先开始画对这一帧结果有影响。
+      detection_skill: render_thread_slices
+
+    - id: geometry_visibility_change_late
+      name: "窗口几何/可见性变化与 buffer 更新不同帧生效"
+      description: |
+        窗口大小/位置/可见性/裁剪通过 SurfaceControl.Transaction 进入系统。
+        SurfaceSyncGroup（API 34+）/ sync transaction 能让几何与 buffer 同帧生效，
+        但应用还没 draw / fence 未 ready / 跨进程 Producer 节奏不可控时，目标帧仍可能错过。
+      detection_skill: sf_composition_in_range
+
+    - id: same_process_callback_order_dependency
+      name: "callback 注册顺序决定窗口先后"
+      description: |
+        Choreographer.doFrame() 按 callback 注册顺序依次执行 pending callbacks——
+        先注册的窗口先执行，后注册的被挤后。Dialog 弹出时机不同会改变后续帧的 callback 顺序。
+      detection_skill: app_frame_production
+
+  recommended_skills:
+    - render_thread_slices
+    - jank_frame_detail

package/skills/pipelines/android_view_software.skill.yaml

@@ -0,0 +1,172 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+# Copyright (C) 2024-2026 Gracker (Chris)
+# This file is part of SmartPerfetto. See LICENSE for details.
+
+# =============================================================================
+# Android View Software Pipeline
+# =============================================================================
+# CPU-based Skia rendering without hardware acceleration.
+# No RenderThread - all drawing done on main thread.
+# =============================================================================
+
+name: pipeline_android_view_software
+version: "1.0"
+type: pipeline_definition
+category: rendering
+
+meta:
+  pipeline_id: ANDROID_VIEW_SOFTWARE
+  display_name: "Android View 软件渲染"
+  description: "CPU Skia 软件渲染，无 RenderThread，用于低端设备或特殊场景"
+  icon: "cpu"
+  family: hwui
+  doc_path: "rendering_pipelines/android_view_software.md"
+  # ----- Article-grounded signature (S07) -----
+  s_article_ref: "S07"
+  four_features:
+    producer_threads: ["main"]   # 无 RenderThread；主线程做 CPU Skia 栅格化
+    expected_layer_count: 1
+    bufferqueue_path: "ACQUIRE_FENCE_NONE_INPROC"  # acquire fence = NO_FENCE（CPU 写完即 ready）
+    extra_rhythm_sources: []
+  deviation_anchors: "skips_renderthread_anchor_5"
+  # 锚点 ⑤ 偏离：CPU 在 main thread 完成栅格化（不经 GPU）。其余按"最终可见对象"判断：
+  # 软件绘制到主窗口 Surface 时仍走 ⑥-⑫；纯离屏（截图/缩略图）则 ⑦-⑫ 完全缺席。
+  subvariants_note: |
+    文章 S07 把 software/离屏拆为 4 子变种：
+    1. SOFTWARE_RENDERING_CANVAS（整窗 lockCanvas，无 RenderThread）— 当前 ID 主要覆盖
+    2. SOFTWARE_RENDERING_LAYER（View.setLayerType(LAYER_TYPE_SOFTWARE)，混合路径，仍有 RenderThread + Bitmap upload）— 当前 exclude_if RenderThread 排除掉
+    3. OFFSCREEN_HARDWAREBUFFER（HardwareBufferRenderer 离屏 GPU，Android 14+/API 35）
+    4. OFFSCREEN_BITMAP（离屏 Bitmap + 纹理上传）
+    Phase E 会拆分独立 ID。
+    article-precise slice 名（用于 LLM 诊断参考）：
+    - nSurface_lockCanvas / nSurface_unlockCanvasAndPost（native 层 Surface 操作）
+    - drawSoftware（ViewRootImpl.drawSoftware()）
+    - copyBlt（dirty region 模式下从旧 buffer 抄 non-dirty 像素）
+
+detection:
+  required_signals:
+    - thread: main
+      min_count: 1
+
+  scoring_signals:
+    - signal: has_lock_canvas
+      slice_pattern: "*lockCanvas*"
+      weight: 55
+    - signal: has_unlock_canvas_post
+      slice_pattern: "*unlockCanvasAndPost*"
+      weight: 55
+
+  exclude_if:
+    - thread: RenderThread
+    - slice_pattern: "DrawFrame*"
+
+teaching:
+  title: "Android View 软件渲染管线"
+  summary: |
+    使用 CPU Skia 进行软件渲染，所有绘制都在主线程完成。
+    通常用于低端设备、特殊 View (如 SurfaceHolder.lockCanvas)，或硬件加速被禁用时。
+    性能较低，但兼容性最好。
+
+  mermaid: |
+    sequenceDiagram
+      participant VA as VSync-app
+      participant Main as App (main)
+      participant Canvas as Canvas (CPU)
+      participant BQ as BufferQueue
+      participant VS as VSync-sf
+      participant SF as SurfaceFlinger
+
+      Note over VA,SF: 📍 Software Rendering (无 RenderThread)
+      VA->>Main: 🔔 VSync-app 信号触发
+      activate Main
+      Main->>Main: Choreographer#doFrame
+      Main->>Canvas: lockCanvas
+      Main->>Canvas: Skia CPU 绘制 (drawXXX)
+      Main->>Canvas: unlockCanvasAndPost
+      Main->>BQ: queueBuffer
+      deactivate Main
+
+      VS->>SF: 🔔 VSync-sf 触发
+      activate SF
+      SF->>SF: latchBuffer
+      SF->>SF: HWC Composite
+      deactivate SF
+
+      Note over VA,SF: ⚠️ 无硬件加速，适用于简单 UI 或兼容模式
+  thread_roles:
+    - thread: main
+      role: "UI 构建 + 渲染"
+      description: "Measure/Layout/Draw 全在主线程，使用 CPU Skia 渲染"
+
+  key_slices:
+    - name: "lockCanvas"
+      thread: main
+      description: "获取 CPU 画布"
+    - name: "unlockCanvasAndPost"
+      thread: main
+      description: "提交绘制结果"
+
+auto_pin:
+  instructions:
+    # Time-ordered pinning (production -> consume -> display)
+    - pattern: "^VSYNC-app$"
+      match_by: name
+      priority: 1
+      reason: "VSync (App 开始生产帧)"
+
+    - pattern: "^main(\\s+\\d+)?$"
+      match_by: name
+      priority: 2
+      reason: "CPU 软件渲染 (Skia)"
+      expand: true  # 展开显示 Running 状态和 callstack
+      smart_filter:
+        enabled: true
+        description: "仅 Pin 活跃渲染进程的主线程（软件渲染通常由 lockCanvas/unlockCanvasAndPost 驱动）"
+
+    - pattern: "^VSYNC-sf$"
+      match_by: name
+      priority: 3
+      reason: "VSync (SurfaceFlinger 消费/合成)"
+
+    - pattern: "^[sS]urface[fF]linger"
+      match_by: name
+      priority: 4
+      reason: "SurfaceFlinger (最终合成/显示)"
+      main_thread_only: true
+
+analysis:
+  common_issues:
+    - id: cpu_bottleneck
+      name: "CPU 渲染瓶颈"
+      description: "软件渲染占用大量 CPU 时间"
+      detection_skill: cpu_analysis
+
+    # ----- Article-grounded issues (S07) -----
+    - id: dirty_region_copyblt_overhead
+      name: "Dirty region copyBlt 开销"
+      description: |
+        Surface::lock(dirty) 带 dirty region 调用时，frameworks/native/libs/gui/Surface.cpp
+        会通过 copyBlt(backBuffer, frontBuffer, copyback) 把上一帧 non-dirty 像素抄到新 buffer。
+        即使 dirty 区域很小，仍有 buffer 拷贝开销；新旧 buffer 尺寸/格式不一致时还会强制全重绘。
+        S07 §"CPU software rendering 这一段在做什么"。
+      detection_skill: cpu_analysis
+
+    - id: no_acquire_fence_release_fence_still_present
+      name: "无 acquire fence 但 release fence 仍存在"
+      description: |
+        软件路径下 unlockAndPost 提交时 acquire fence = NO_FENCE，SF latch 无需等待。
+        但 release fence 仍然存在：HWC presentDisplay 后通过 getReleaseFences() 返回，
+        SF 经 BufferQueue 回传给 Producer 通知可重新 dequeue。Producer 仍可能卡 dequeueBuffer 等 release。
+      detection_skill: render_thread_slices
+
+    - id: software_layer_bitmap_upload
+      name: "LAYER_TYPE_SOFTWARE 的 Bitmap 上传开销"
+      description: |
+        若某 View 设了 LAYER_TYPE_SOFTWARE，子树会被绘制到离屏 Bitmap 再由 HWUI 作为纹理上传参与 GPU 合成。
+        Trace 中表现为 RenderThread 仍存在（与纯软件不同），但额外伴随 Bitmap 分配 + uploadToTexture slice。
+        当前 ID 因 exclude_if RenderThread 排除此场景；Phase E 拆 SOFTWARE_RENDERING_LAYER 独立 ID。
+      detection_skill: render_thread_slices
+
+  recommended_skills:
+    - cpu_analysis
+    - scheduling_analysis