T

marfrit 2a5f5c296e iter1 phase0: substrate + prior-art survey (HEVC reclassified)

Phase 0 ran the operator-mandated upstream prior-art survey FIRST,
before any source-read or hypothesis. Headline finding: the HEVC
OOPS is fundamentally re-scoped from 'kernel bug' to 'userspace UAPI
gap' against the new 7.0 controls
V4L2_CID_STATELESS_HEVC_EXT_SPS_ST_RPS / _LT_RPS.

Survey + verification:
- Casanova/Collabora v8 series merged in Linux 7.0 added the two
  new V4L2 controls for VDPU381 HEVC; backend 7ac934e (June pre-iter38)
  predates the UAPI and grep returns zero hits for these CIDs.
- ampere linux-api-headers is still 6.19-1, doesn't define the
  constants — the backend literally cannot reference them without a
  headers bump.
- ampere kernel source rkvdec-hevc-common.c:500-509 looks up the new
  CIDs; if backend never set them, rkvdec_hevc_prepare_hw_st_rps
  reads invalid memory via memcmp — exactly the __pi_memcmp OOPS
  symptom.

VP9 still kernel-side per the v4 cover ('This patch only adds support
for H264 and H265 in both variants'). Multiple competing out-of-tree
starting trees: Sarma's android tree (working but Android-flavored),
dongioia/rock5bplus-rkvdec2 (mainline-style claims), Kwiboo (no VP9
on RK3588 yet). RKVDEC2 separate-driver path is dead — future VP9
extends the existing rkvdec driver's VDPU381 variant_ops.

Five open questions tabled for Phase 1 — most important being campaign
re-scope (HEVC moves to backend campaign; this stays VP9 kernel-only
OR becomes a meta-campaign coordinating both).

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

2026-05-16 08:10:40 +00:00

phase0_findings.md

iter1 phase0: substrate + prior-art survey (HEVC reclassified)

2026-05-16 08:10:40 +00:00

README.md

scaffold: kernel-side sibling campaign for RK3588 decoder enablement

2026-05-16 08:07:55 +00:00

README.md

ampere-kernel-decoders

Kernel-side sibling campaign to ampere-fourier (which validated the userspace libva backend on RK3588 at iter1 close, 3 codecs working) and fresnel-fourier (the RK3399 peer). Focuses on the kernel-side enablement required to unblock the three codec blockers iter1 of ampere-fourier surfaced:

HEVC kernel OOPS in rkvdec_hevc_prepare_hw_st_rps (__pi_memcmp fault, cascades to v4l2_mem2mem wedge). Tracked at marfrit/kernel-agent#11 [ka:experiment].
VP9 not exposed on RK3588 rkvdec — kernel doesn't register V4L2_PIX_FMT_VP9_FRAME on the VDPU381/383 variant_ops. Tracked at marfrit/kernel-agent#12 [ka:experiment].
(AV1 backend iter39 is userspace work, not in this campaign's scope; tracked at marfrit/libva-v4l2-request-fourier#2.)

Topology

Property	Value
Working tree (substrate kernel)	`boltzmann:~/src/linux-rockchip` branch `linux-rk3588-marfrit`
Board patches	`boltzmann:~/src/misc_patches/genbook/kernel/000*.patch` (6 board DTS/config patches; the ampere baseline package consumes these)
Experiment target host	ampere (CoolPi CM5 GenBook, RK3588)
Baseline kernel for regression-checks	`linux-ampere-fourier 7.0rc3.kafr1-1` (vanilla torvalds v7.0-rc3 + `genbook/kernel/000*` only — NO codec patches)
Patch destination	kernel-agent experiment branches per issue; promoted to `linux-ampere-fourier` only via explicit operator decision per `feedback_characterize_before_change` (campaign-experiment patches stay OUT of baseline)
Build host	boltzmann (primary aarch64 builder per kernel-agent README) + ampere as fallback secondary
Patch landing pad	`marfrit/kernel-agent/patches/{soc,driver}/...` scope-tagged per kernel-agent's tree convention

Scope (in flux until Phase 0 close)

In scope (iter1 candidate):

HEVC OOPS fix — minimal-blast-radius candidate patch for rkvdec_hevc_prepare_hw_st_rps, validated on ampere by re-running ampere-fourier Phase 3 with HEVC added.
Survey upstream prior art (linux-rockchip, linux-media, linux-mm, Kwiboo, Bootlin, Collabora) before writing any code — there may already be a fix in v7.0-rc4+ or in an out-of-tree branch.

Out of scope (likely iter2+):

VP9 enablement (broader work — VDPU381/383 variant_ops surface, DTS bindings, potentially backend codec table updates). Separate iteration.
Multi-core support for rkvdec/hantro (the "missing multi-core support, ignoring this instance" lines in dmesg). Upstream work; out of fleet-scope.

Explicitly NOT in scope:

Codec patches on the baseline linux-ampere-fourier package. Per operator policy 2026-05-16, ampere baseline stays clean mainline + board DTS. Codec enablement = kernel-agent experiment branches only.

Process

8(+1)-phase loop per feedback_dev_process.md, same as ampere-fourier / fresnel-fourier. Notable for this campaign:

Phase 0 includes an upstream prior-art survey (linux-rockchip + linux-media + linux-mm + Kwiboo + Bootlin). This is non-optional — writing a kernel patch without first checking whether the fix already exists upstream is the wrong workflow. Memory feedback_no_upstream says no PR/MR/RFC from us by default; that's about publishing, not consuming. Consume aggressively.
Phase 5 review uses the sonnet-architect subagent pattern (Plan with model: sonnet), same as ampere-fourier. Memory rule feedback_review_empirical_over_theoretical applies — test-compile reviewer-suggested struct/field mappings before adopting amendments.
Phase 6 implementation produces kernel patches in ~/src/linux-rockchip on boltzmann + accompanying kernel-agent experiment branch entries. No patches land in marfrit-packages/arch/linux-ampere-fourier/ directly.
Phase 7 verification re-runs ampere-fourier iter1's Phase 3 scripts with HEVC added to the codec list. Predicted outcome anchors against ampere-fourier iter1's baseline numbers.

Predecessor work this campaign builds on

../ampere-fourier/ iter1 — the baseline floor this campaign regresses against. Re-anchoring to ampere-fourier's N=3 FPS numbers + per-codec SSIM floors for any codec that flips from "blocked" to "validated."
marfrit/kernel-agent — the experiment-branch home + issue tracker.
marfrit/kernel-agent/issues/11 — HEVC bug ticket with the OOPS trace + reproducer.
marfrit/kernel-agent/issues/12 — VP9 enablement ticket.
Memory feedback_rkvdec_patch_reachability — VDPU381/383 vs RK3399 legacy path boundary.
Memory feedback_characterize_before_change — campaign-process rule from ampere-fourier iter1.

Operator-facing repo URL

git.reauktion.de/marfrit/ampere-kernel-decoders — to be created at iter1 close if there's something publish-worthy (matches the fresnel-fourier / ampere-fourier convention). Local-only at ~/src/ampere-kernel-decoders/ on noether for now.

Description

RK3588 decoder enablement on ampere (CoolPi CM5 GenBook). Sibling kernel-side campaign to marfrit/ampere-fourier (userspace consumer). Meta-campaign coordinating HEVC OOPS investigation (iter2: F1 negative-result), VP9 enablement (iter4), AV1 backend probe (sibling). Process: 8(+1)-phase loop.

Readme 194 KiB