Task 0 Spike Notes

June 17, 2026 · View on GitHub

Date: 2026-06-17
Device: OPPO P0110, arm64-v8a, API 36 (Android 16 preview)
Build type: user (NOT debuggable/userdebug/eng) — important for input below.
trace_processor: perfetto py 0.56.0 + trace_processor_shell v56.1 mac-arm64 (downloaded directly from commondatastorage.googleapis.com/perfetto-luci-artifacts/v56.1/mac-arm64/ — the Python package's bootstrap at get.perfetto.dev/trace_processor hit an SSL cert-verify failure on macOS Python 3.12; the direct-artifact curl bypasses it).
Test trace: /tmp/spike2.pftrace — 8s with manual adb input swipe activity, 17.7 MB. Captured via official/record_android_trace -c <min.pbtx> --no-open.

Confirmed facts

The minimal config with linux.ftrace + atrace_categories/atrace_apps (NOT a standalone android.atrace data source) parses and captures cleanly. ✅
android.surfaceflinger.frametimeline is accepted as a data source. ✅
android.input.inputevent is accepted as a data source. ✅

792 rows in the 8s swiping trace. Columns (confirmed): id, ts, dur, track_id, category, name, depth, parent_id, arg_set_id, display_frame_token, surface_frame_token, upid, layer_name, present_type, on_time_finish, gpu_composition, jank_type, jank_severity_type, prediction_type, jank_tag, jank_tag_experimental, jank_score, latched_unsignaled_count, addressable_unsignaled_latch_count, latched_fence_state
present_type values: Dropped Frame, Early Present, Late Present, On-time Present, Unspecified Present.
jank_type values: None, App Deadline Missed, Buffer Stuffing, Display HAL, Dropped Frame, Prediction Error, SurfaceFlinger Scheduling, Unknown Jank. Can be comma-separated (e.g. App Deadline Missed, Buffer Stuffing).
Multi-source is BUILT IN via layer_name: each producing surface is its own layer (TX - com.obric.quicksearch/...SearchActivity#292, TX - ...Launcher#264, TX - StatusBar#92, TX - InputMethod#281, TX - Task=905#272, ...). SurfaceView / TextureView surfaces would appear here too.
layer_name IS NULL = display (SF composited) frames — 327 rows, distinct from the 465 per-surface frames. These represent screen refreshes, not per-source production.

Useless on this device/version. The plan's multi-source BufferQueue path via frame_slice does NOT work here.

android_input_events / android_motion_events tables exist but require INCLUDE PERFETTO MODULE android.input; first.
On this user build: android_input_events has 533 rows but event_action is NULL for ALL of them. android_motion_events has 0 rows. → Structured input action is only populated on debuggable/userdebug/eng builds.
BUT: atrace slices DO carry the action with timestamps (from the input atrace category):
- dispatchInputEvent MotionEvent ACTION_DOWN deviceId=-1 source=0x1002 historySize=0 — 15 occurrences, with ts.
- dispatchInputEvent MotionEvent ACTION_UP deviceId=-1 source=0x1002 historySize=0 — 13 occurrences, with ts.
- publishMotionEvent(inputChannel=..., action=DOWN|UP|MOVE) also present. → ACTION_DOWN/UP can be extracted from atrace slices even on a user build, giving precise fling windows without the coarse device-clock fallback.

queueBuffer (1232), dequeueBuffer (869), acquireBuffer (713) all present in the gfx atrace category.
BUT their track.name is NULL — there is no layer association on these slices. → TextureView single-buffer overwrite detection via raw BufferQueue slices is NOT feasible here (can't attribute a queue/acquire to a layer). Dropped frames are still caught via present_type='Dropped Frame' in FrameTimeline.

Available. TO_REALTIME(0) = 1781656624538267536 (a unix-realtime nanosecond value).
Used to align frame ts (trace/boottime) with device-clock swipe markers.

Structured android_input_events.event_action (non-NULL) — debuggable builds.
atrace slice dispatchInputEvent ... ACTION_UP → next ACTION_DOWN — works on user builds (confirmed here). Preferred over device-clock when structured is empty.
Device-clock swipe markers from run_fps_test.sh (coarse, last resort).

Drop the frame_slice / per-layer BufferQueue path (empty / no layer attribution).
FPS is computed per layer_name from actual_frame_timeline_slice.
layer_name IS NULL rows are bucketed as source "display" (SF output refreshes); non-NULL rows keep their layer name as the source. SurfaceView/TextureView/video surfaces appear as their own layer_name automatically.

Drop the detect_overwrite_drops / _query_raw_buffer_events integration path (no layer attribution available). Dropped frames are caught by FrameTimeline's present_type. jank_type='Dropped Frame' mirrors it.
The pure-math detect_overwrite_drops/buffer_events_to_frames functions and their unit tests are KEPT (they are correct sequence logic and cheap), but analyze_trace no longer calls into the raw-BufferQueue path. They remain as a tested library for future use when a trace exposes layer-attributed buffer events.

dropped = present_type='Dropped Frame' → never on screen → excluded from FPS.
janky = jank_type IS NOT NULL AND jank_type != 'None' → presented but late → counts toward FPS, reported separately. (Note: when present_type is Dropped, jank_type is also 'Dropped Frame' — already excluded as dropped, don't double-count it as janky. So janky requires NOT dropped.)

02_jank_frame.pbtx keeps android.surfaceflinger.frametimeline (primary frame source) and android.input.inputevent (for structured input when available). The input atrace category is ALSO required — it's what populates the ACTION_UP/DOWN slices used on user builds. Both must be present.