Repository navigation
v4.0.0 Number Nine "Pepper Pro" — H.264 and HEVC conformance, High 10, AV1 size changes
LatestSince v3.5.0:
- H.264: the PPS is rebuilt for every picture under the id the slices name, with the stream's scaling matrices (x264
--cqm, Blu-ray and broadcast encodes were slightly wrong at every bit depth, KI-17); field pictures and PAFF/MBAFF streams decode (second fields were dropped, KI-20; the output is weaved, not deinterlaced); CAVLC multi-slice pictures decode (KI-21); monochrome 4:0:0 streams no longer crash the player. JVT-AVC_V1 74 → 111 of 135 vectors bit-exact; JVT-FR-EXT's 48 4:2:0 vectors 6 → 48 of 48 (its 21 High 4:2:2 vectors are decoded by ffmpeg in software with any driver — an earlier version of this text counted them as "27 → 69 of 69"). - H.264 High 10: the initial QP that ffmpeg hands over with the bit-depth offset already added is corrected (
RKVA_H264_INITQP=coded|ffmpegoverrides the client recognition); both High 10 conformance vectors pass. - HEVC: slices that refer to SPS-level reference picture sets are rewritten with inline sets (the hardware stalled on them, KI-5), and the PPS is rebuilt per picture with the stream's scaling lists (KI-22). JCT-VC-HEVC_V1 17 → 141 of 147; the six left are hardware limits (PICSIZE_A/B, TSUNEQBD_A) and three vectors ffmpeg's own software decoder fails too (NUT_A, RPS_D, VPSSPSPPS_A).
- AV1 resolution change inside one decoder (KI-13): MPP decodes both libaom size-change vectors bit-exact; the driver no longer lets a client read past an image of the old size — players that export their surfaces (mpv
hwdec=vaapi, measured; Firefox uses the same export path but recreates its decoder on a size change) get a surface of the new size and the right picture,vaGetImageclients (ffmpeg, mpvvaapi-copy) a clean per-picture refusal instead of a crash. ffmpeg itself keeps its surface pool across an AV1 sequence change; that part is ffmpeg's. - Film grain is applied by the hardware with Plaid Pi's MPP
+pd35, published alongside this release for Ubuntu 24.04 and 26.04 (https://github.com/defcom5-rockchip/mpp/releases/tag/pd35) — the one-line fix from rockchip-linux/mpp issue #974 (pull request prepared); the film-grain conformance vectors (three streams) pass: libaom 215/242, Chromium 8-bit 11/13, 10-bit 23/23. - VP9: unchanged, 189/305 and 3/6 — identical vector by vector to 3.5.0; the failing classes (frames under 16 px, reference scaling, odd sizes, 4x4 tiles) are the next target.
- Review fixes across the four changes: PPS buffers sized for full scaling lists, a writer that refuses to ship a truncated header, the copy into a surface takes its destination under the surface lock, one-time log lines for every fallback. New switch
RKVA_H264_INITQP;RKVA_DUMP_H264andRKVA_DUMP_HEVCare now in the README switch table. - Evidence for every number is in
evidence/2026-10-25-*/; the kernel's decoder-reset counter did not move during the H.264 and HEVC runs.
Asset: rockchip-vaapi_4.0.0-1_arm64.deb, sha256 d0022adf792e7b59cf8f97fc3dafc8a117e6a2e060fa176991bd5edbd7f09e64 (SHA256SUMS).
Upgrade: sudo apt install ./rockchip-vaapi_4.0.0-1_arm64.deb (add --allow-change-held-packages if your image holds it).
Named after Number Nine's Pepper Pro series (1988–1990), the TIGA boards that followed the Pepper SGT.
An LLM assistant (Claude) co-engineered this release (credited per commit); four review agents checked the code, the claims against the evidence, the integration and the docs for each of the four changes and again for the release before it was tagged; every measurement was run on real hardware (Orange Pi 5B), the by-eye checks by the maintainer.