diff --git a/phase4_iter9_plan.md b/phase4_iter9_plan.md new file mode 100644 index 0000000..151045a --- /dev/null +++ b/phase4_iter9_plan.md @@ -0,0 +1,57 @@ +# Iteration 9 — Phase 4 (plan) + +The Phase 0 doc lays out α-7 in detail. Phase 4 just locks the implementation contract for Phase 6. + +## Implementation contract + +**File**: `src/picture.c::RequestEndPicture` (line 440 area). + +**Change**: replace +```c +gettimeofday(&surface_object->timestamp, NULL); +``` +with a monotonic per-context counter that produces small ns values. + +**State storage**: `driver_data->timestamp_counter` (u64). Init to 0 wherever `request_data` is first set up (likely `request.c::v4l2_request_init`). + +**Counter cadence**: increment by `1` (just enough to keep each frame's timestamp unique). For 3-frame test: ts will be 0, 1, 2 — followed by `v4l2_timeval_to_ns` returning `0`, `1000` (since 1 µs = 1000 ns), `2000` ... wait, `struct timeval` is sec + usec. Mapping a counter to timeval: + +```c +driver_data->timestamp_counter++; +surface_object->timestamp.tv_sec = 0; +surface_object->timestamp.tv_usec = driver_data->timestamp_counter; +``` + +This gives `tv_usec=1,2,3,...` → `v4l2_timeval_to_ns(tv) = tv_usec * 1000 = 1000, 2000, 3000, ...`. Matches kdirect's `0x2af8 = 10968` style. + +For long runs (>~71 minutes at 1000 fps), tv_usec wraps. Each wrap is a uniqueness collision risk. Mitigation: increment tv_sec on wrap. Code: + +```c +driver_data->timestamp_counter++; +surface_object->timestamp.tv_sec = driver_data->timestamp_counter / 1000000; +surface_object->timestamp.tv_usec = driver_data->timestamp_counter % 1000000; +``` + +This is collision-free for u64 / 1e6 ≈ 5.84 × 10^11 seconds = ~18,500 years of decode. Future-proof. + +LOC: ~10 (counter init in request.c + 4-line replacement in picture.c + new field in request.h). + +## Verification (Phase 7) + +1. Build + install on fresnel. +2. Run libva H.264 sweep. Hash against kdirect's `1e7a0bc9…`. +3. Re-strace and confirm DPB.reference_ts is now small (~1000-3000) matching kdirect's pattern. +4. Run 5-codec regression sweep. All 4 non-H.264 anchors must hold. + +## Risks recap (from Phase 0) + +- R-1 uniqueness: counter wraps in 18,500 years — non-issue. +- R-2 regression on VP9/MPEG-2/HEVC/VP8: change is uniform across codecs; same path. VP9/MPEG-2 currently PASS — switching to counter should be neutral or improve. +- R-3 consumer reads ts as wallclock: VAAPI surfaces don't expose timestamps to consumers. + +## Phase 5 review + +Quick review by the same reviewer agent. Key questions: +- Does any other backend code read `surface_object->timestamp` as a real wallclock? +- Does the V4L2 framework do anything weird with `timeval` vs `u64 ns` conversion that could cause libva's gettimeofday-based ts and DPB.reference_ts to actually differ at the kernel observation point? (Critical for M-C hypothesis.) +- Is there an even simpler reason giant ns fail — e.g., a 32-bit truncation in the V4L2 buffer struct itself?