Files

T

claude-noether 695588aaaf Phase 1 v3 amendment: Janet PROCEED with 2 plan-text fixes

Janet re-review verdict on v2: PROCEED with no architecture changes.
Two minor plan-text amendments:

A. cap_pool init for vpu981 — at backend init (NOT lazy). Three
   explicit init calls in VA_DRIVER_INIT_FUNC. Lazy init creates race
   window where concurrent VP8 (legacy hantro) and AV1 (vpu981)
   sessions can alias cap_pool state across the wrong device.

B. Lock the F1 test vector: av1-1-b8-23-film_grain-50.ivf (AOM data
   set). Already used in Phase 0 kdirect; reuse via libva for byte-
   compare. Default 1080p encodes may be single-tile and never
   exercise F1; a named conformance vector locks the test surface.

Cross-compile hazard documented: _Static_assert validates against
COMPILE-TIME kernel headers. Sibling-campaign pattern is native arm64
build on ampere — no cross-compile in scope. Comment added near the
assert for future Makefile changes.

Phase 2 entry checklist locked. Plan stack: v1 → v2 (Q5 BLOCKER fix)
→ v3 amendment (Janet PROCEED + 2 fixes). Implementation starts with
request.c third-device scaffolding (smallest atomic change), then
small file edits, then NEW av1.h, then NEW av1.c (~700 LoC), then
build + Phase 3 hardware test.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

2026-05-17 07:51:56 +00:00

3.1 KiB

Raw Permalink Blame History

Phase 1 plan v3 amendment — Janet PROCEED with 2 plan-text fixes

Date: 2026-05-17 09:55. Janet re-review verdict on v2: PROCEED with 2 minor plan-text amendments. No architecture changes.

Amendment A: cap_pool init for vpu981 — at backend init, not lazy

v2 said "cap_pool extension needs the same pair-of-flags layout". Implicit: pool state for the vpu981 fd must be set up at backend init time alongside rkvdec + legacy hantro, NOT lazily on first AV1 RequestCreateConfig.

Why: lazy init creates a race window where a concurrent VP8 session (using legacy hantro fd) triggers cap_pool re-init on the legacy hantro device, while an AV1 session is mid-flight, can alias the pool state across the wrong device. Force init at backend open.

Concretely in VA_DRIVER_INIT_FUNC: after the three find_decoder_device_* calls succeed, immediately call cap_pool_init(driver_data, video_fd_rkvdec), cap_pool_init(driver_data, video_fd_hantro), cap_pool_init(driver_data, video_fd_vpu981) — three explicit init calls, one per device fd, before returning success to libva.

Amendment B: lock the F1 test vector

v2 said "Phase 3 verification must include multi-tile 1080p+ AV1". Janet flagged that a default-encoded 1080p might be single-tile and never exercise F1.

Specific test vector for F1: av1-1-b8-23-film_grain-50.ivf from the AOM test data set — exercises film_grain (F3) AND tiling (F1) simultaneously. URL: https://storage.googleapis.com/aom-test-data/av1-1-b8-23-film_grain-50.ivf.

Phase 0 already used this for kdirect verification (10/10 frames bit-perfect). Phase 3 reruns it via libva path and byte-compares.

Additional multi-tile coverage: AOM test vector av1-1-b8-22-svc-L2T1.ivf (if exists) for layered/tiled stress, and any 1080p AV1 source encoded with --tile-columns=2 --tile-rows=2.

Cross-compile build-time hazard (Janet flagged for documentation only)

The _Static_assert(sizeof(struct v4l2_ctrl_av1_tile_group_entry) == 16, ...) validates against kernel headers at COMPILE TIME. If the build moves cross-compile (e.g., builds on fresnel x86 against ampere arm64 kernel headers), the assert silently validates the wrong ABI.

Mitigation: builds happen on ampere (native arm64) per sibling-campaign pattern. No cross-compile in this campaign. Add a comment near the assert flagging the hazard for any future Makefile changes.

Carry-forward verdict

Janet PROCEED v2 → PROCEED with these two text amendments → Phase 2 code can open.

Phase 2 entry checklist

Plan v2 + v3 amendments committed and pushed (this commit)
Start with request.c scaffolding (vpu981 fd slots + capability probe) — smallest atomic change, easy to verify
Then codec.c + config.c + picture.c + surface.h (small edits, no new files)
Then NEW av1.h (~30 LoC, minimal interface)
Then NEW av1.c (~700 LoC — biggest single file, mirror Kwiboo + apply F1/F2/F3 corrections from v2)
Then Makefile.am + autoreconf + native build on ampere
Then Phase 3 hardware test starting with the existing 208×208 + 352×288 vectors (sanity), then av1-1-b8-23-film_grain-50.ivf (multi-tile F1 + film_grain F3 stress)

3.1 KiB Raw Permalink Blame History Unescape Escape