 claude-noether
|
6e8c970c1d
|
iter2 Phase 0 + Phase 1 lock: HEVC Main on rkvdec
Iteration 2 of the campaign 8(+1)-phase loop opens following iter1
close (dc69378). Per phase0_evidence/2026-05-07/cross_validator_
traces.md suggested ordering, iter2 attacks HEVC Main on rkvdec —
the kernel + driver path is verified working (cross-validator sweep
exit 0); broken link is the libva backend at five distinct sites:
src/config.c HEVCMain case fall-through (analogous to iter1 Bug 1)
src/picture.c HEVCMain explicit UNSUPPORTED_PROFILE reject (NEW)
src/h265.c uncompiled in build (presumably staging-era CIDs;
Phase 2 source-read decides scope of rewrite)
include/hevc-ctrls.h staging-era local header (deferred from
iter1 Phase 5 Nit 6; iter2 closes the loop)
src/meson.build h265.c commented out (re-enable)
Plus possible novel issues vs iter1's MPEG-2 work:
- HEVC has 10 stateless control IDs vs MPEG-2's 3 (much larger
rewrite if h265.c uses staging-era API)
- HEVC slice_params is dynamic-array (kernel rkvdec accepts up
to 600 entries) — different submission shape vs MPEG-2 single-
struct or H.264 fixed-shape
- HEVC SCALING_MATRIX is conditional (only when scaling_list_
enabled in SPS); mapping VAIQMatrixBufferHEVC to V4L2 control
- HEVC ENTRY_POINT_OFFSETS is in kernel surface (tile/slice
resync) but campaign fixture doesn't use tiles — defer
Locked research question:
Make HEVC Main the third codec to pass boolean-correctness on
fresnel via libva-v4l2-request-fourier — mpv --hwdec=vaapi
bbb_720p10s_hevc.mp4 engages backend cleanly and DMA-BUF GL
import yields HW pixels byte-identical to SW reference for the
same frames.
Phase 1 success criterion (5 boolean checks, all must pass):
1. vainfo enumerates VAProfileHEVCMain on rkvdec env binding
(regression check; already passes today).
2. vaCreateConfig(VAProfileHEVCMain, VLD) returns VA_STATUS_
SUCCESS. (Pre-iter2: VA_STATUS_ERROR_UNSUPPORTED_PROFILE.)
3. ffmpeg -hwaccel vaapi -i bbb_720p10s_hevc.mp4 -frames:v 5
-f null - exits 0 cleanly with no Failed-to-create-decode-
configuration lines and no S_EXT_CTRLS EINVAL on HEVC
controls. (Phase 1 criterion 3 anchored on ffmpeg-direct,
mirroring iter1 Phase 5 Q4 amendment for codecs mpv may
filter out.)
4. mpv --hwdec=vaapi --vo=image at +02s seek: 2 distinct frames
hash-equal to SW reference, hash-differ from each other (real
motion). DMA-BUF GL import path per memory feedback_rockchip_
pixel_verify_path.md (NOT ffmpeg-vaapi+hwdownload, which is
cache-stale on RK3399 for both H.264 and MPEG-2 per iter1
Phase 6/7 findings).
5. iter1 MPEG-2 + T4 H.264 reference hashes BOTH still match
(regression check on prior-iteration cells):
MPEG-2 +02s: HW1=6e7873030dbf... HW2=ccc7ce08810d...
H.264 +30s: HW1=f623d5f7a416... HW2=7d7bc6f2146d...
Substrate carry-over:
- libva-v4l2-request-fourier master tip post-iter1-close
(commits e7dad7a..229d6d1 stack on iter8 65969da).
- bbb_720p10s_hevc.mp4 fixture (620 KB, HEVC Main, 1280x720,
24fps, 10s, yuv420p; provenance phase0_evidence/2026-05-07/
test_fixtures.md).
- Cross-validator anchor: phase0_evidence/2026-05-07/
cross_validator/hevc/ — 14 S_EXT_CTRLS + 5 QUERY_EXT_CTRL
(HEVC slice_params dynamic-array introspection unique among
the 5 codecs) + 4 REQUEST_ALLOC.
- Memory carries forward: feedback_gitea_as_claude_noether,
feedback_no_session_termination_attempts, feedback_header_
deletion_check (iter1 lesson L1 — apply to hevc-ctrls.h
deletion), feedback_review_empirical_over_theoretical
(iter1 lesson L2 — apply to Phase 5 review responses),
feedback_rockchip_pixel_verify_path (iter1 lesson L3 —
DMA-BUF GL is the verifier, NOT cached-mmap).
Out-of-scope (LOCKED): VP9/VP8 (later iterations); HEVC Main 10
(silicon support unverified); HEVC Main Still Picture; performance
metrics; long-duration HEVC stress; tile / wavefront parallel
processing (ENTRY_POINT_OFFSETS); Phase 4 cross-cutting backlog
(B1 device-discovery, B3 BeginPicture profile-aware reset, B4
context.c log suppression, B5 vbv_buffer_size negotiation, L3
vaDeriveImage cache-stale fix); chromium-fourier 149 install;
src/context.c changes; upstream engagement.
Predecessor open questions:
- iter1 B3 latent surface-reuse bug (picture.c:287
h264.matrix_set=false hits union byte 240) — for HEVC, the
union member is params.h265.{picture,slice,iqmatrix,
iqmatrix_set}. params.h265 layout differs from params.mpeg2.
Phase 2 source-read action item: verify whether byte 240 lands
in a meaningful HEVC field. If so, iter2 may need to address
even though MPEG-2 didn't.
Phase 2 source-read targets (queued for next phase):
- src/h265.c (~267 lines) — current state, target API
- src/picture.c:204-206 (the explicit HEVC reject)
- src/config.c:55-69 (confirm HEVCMain fall-through)
- src/surface.h:103-108 (params.h265 struct)
- include/hevc-ctrls.h (staging-era; identify CID/struct refs)
- src/meson.build (commented-out h265.c)
- linux/v4l2-controls.h:2110+ (modern HEVC stateless UAPI)
- drivers/staging/media/rkvdec/rkvdec_hevc.c (rkvdec contract)
- libavcodec/v4l2_request_hevc.c (FFmpeg reference impl)
- va/va_dec_hevc.h (VAAPI HEVC buffer structs)
Predicted iter2 close shape: similar pattern to iter1 (config
break + h265.c new-API rewrite + header delete + meson re-enable
+ picture.c reject removal). Larger code change than iter1
(predicting 250-400 lines for h265.c rewrite vs iter1's ~120 lines
for mpeg2.c). One novel construct (slice_params dynamic-array)
worth Phase 4 contract-clause-level attention. Expect Phase 6
takes longer than iter1; Phase 7 harness re-uses iter1's pattern.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-05-08 09:44:30 +00:00 |
|