Skip to content

media: apple: avd: T8142: 15 instruction FIFOs and the Hibiscus VP pool - #2219

Merged
DjDeveloperr merged 1 commit into
aurora-silicon:integration/aurorafrom
m0d16l14n1:j815/t8142-avd-hibiscus
Nov 19, 2026
Merged

DjDeveloperr merged 1 commit into
aurora-silicon:integration/aurorafrom
m0d16l14n1:j815/t8142-avd-hibiscus

Conversation

@m0d16l14n1

Copy link
Copy Markdown
Contributor

The M5's decoder is a Hibiscus part, not T8140's Tansy:

  • macOS's AppleAVD (27.2.1) starts T8142 as device type 0x1a8 and moves it to 0x1b0, the Hibiscus classes, when the low half of ctrl+0 (AVD+0x1100000) reads 0x20. The J815 reads 0x40020 there (the driver's "booting hw version" line).
  • CAvdApCommHibiscus asks for 15 instruction FIFOs (Tansy: 7).
  • The v5 tier 2 avd-fw build T8142 already runs reports 15 FIFOs and a 4/4/1/3 VP pool.

Change: T8142 uses 15 FIFOs and VP counts 4/4/1/3 (HEVC/H.264/VP9/AV1). Each codec's first VP stays where it was (0/4/8/9), now as the default contiguous layout, so vp_slot_base goes away.

Tested on J815 (MacBook Air 15" M5), two boots with 15 FIFOs against two earlier boots with 7:

  • Fluster: HEVC 143/147, AV1 235/242 and AVC unchanged. The same vectors fail either way: HEVC CONFWIN_A, RPS_E, VPSSPSPPS_A; TSUNEQBD is unequal bit depth.
  • VP9: 177/303, 4 more passing than with 7 FIFOs.
  • Real-world set: 59 decode checks (H.264, HEVC, VP9 and AV1 up to 4K; whole streams and seeks), each frame equal to libavcodec.
  • HEVC Main10: 10 of 11 JCT-VC vectors bit-exact in P010.

On one of the two boots, three AV1 vectors failed once (quantizer-44, size-202x196, svc-L1T2). They passed on the other boot and on the 7-FIFO boots, and they came right after a deliberately bad H.264 High 10 run that reset the decoder. They will be re-run on a fresh boot; I'll update here.

Part of the J815 bring-up; board results are summarised in discussion #135.

🤖 Generated with Claude Code

The M5's decoder is a Hibiscus part: macOS's AppleAVD moves T8142 to its
Hibiscus classes when the low half of ctrl+0 reads 0x20, which the
MacBook Air 15" M5 (J815) does (0x40020). CAvdApCommHibiscus asks for 15
instruction FIFOs, and the v5 tier 2 avd-fw build T8142 already runs
reports 15 FIFOs and a 4/4/1/3 VP pool, so use them instead of T8140's 7
FIFOs and the one-or-two VPs per codec. The first VP of each codec stays
where it was (0/4/8/9), now as the default contiguous layout.

Tested on the J815 on two boots: Fluster HEVC 143/147, AV1 235/242, VP9
177/303 and AVC unchanged from 7 FIFOs (the same vectors fail either
way; VP9 4 better), and 59 decode checks of a real-world set (H.264,
HEVC, VP9, AV1 up to 4K; whole streams and seeks), each frame equal to
software decoding.

Assisted-by: Claude Opus 5.5
Signed-off-by: Nick <nekreshen@yandex.ru>
@DjDeveloperr
DjDeveloperr merged commit df73b90 into aurora-silicon:integration/aurora Nov 19, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants