marfrit/fresnel-fourier
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
104 Commits 1 Branch 0 Tags
70ddbd6c4b69a569d27500ca2bf66eb62093feaa
Commit Graph

104 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
claude-noether d8a9903ef4 phase 0 deliverable 4: H.264 baseline trace — PASS boolean correctness 
H.264 hardware decode on RK3399 / rkvdec / libva-v4l2-request-fourier
@ master tip 65969da (iter8 Phase 4) verified bit-exact correct against
software reference, when read via the cache-safe DMA-BUF GL import path.

Test method:

  - mpv --hwdec=vaapi --vo=image (DMA-BUF + EGL_EXT_image_dma_buf_import
    + glReadPixels + JPEG encode — cache-coherency-safe per the iter1
    patch-0011 lesson).
  - Decoded 2 frames at +30s seek (mid-content bunny motion, not BBB
    intro fade-in) so size + content variation is genuine.
  - Compared HW JPEGs vs SW reference JPEGs (same mpv invocation with
    --hwdec=no).

Result:

  HW frame 1 sha256 = f623d5f7...  (651,726 bytes)  byte-identical
  SW frame 1 sha256 = f623d5f7...  (651,726 bytes)  to SW reference
  HW frame 2 sha256 = 7d7bc6f2...  (630,433 bytes)  byte-identical
  SW frame 2 sha256 = 7d7bc6f2...  (630,433 bytes)  to SW reference

  Frames 1 vs 2 differ in size — real content change captured.

Phase 0 boolean-correctness criterion for H.264: PASS.

Contract trace:

The V4L2 + media-request ioctl sequence per H.264 frame is the
canonical iter6/iter7 pattern:

  S_EXT_CTRLS (CODEC_STATELESS class, request_fd=N)
  QBUF CAPTURE_MPLANE  index=K
  QBUF OUTPUT_MPLANE   index=K  (compressed slice)
  MEDIA_REQUEST_IOC_QUEUE   (request_fd=N)
  MEDIA_REQUEST_IOC_REINIT  (request_fd=N)  ← per-OUTPUT-slot reuse
  DQBUF OUTPUT_MPLANE  index=K
  DQBUF CAPTURE_MPLANE index=K

REINIT-before-DQBUF works because the kernel completes decode in
~0.6 ms (request → COMPLETE state), and mainline media_request_
ioctl_reinit accepts both IDLE and COMPLETE. iter7 cap_pool
instantiates 24 slots cleanly: "v4l2-request: cap_pool_init: 24
slots ready" in mpv stdout.

No EINVAL, no EBUSY, no errors observed across 5 frames. iter4's
frame-11 EINVAL bug from libva-multiplanar does not reproduce on
RK3399 in this short window (longer-run repro is Phase 1+ work).

Side finding — cache-stale readback bug present in libva-backend's
vaDeriveImage path on RK3399:

When pixels are read via the cached-mmap path (libva's vaDeriveImage
+ vaMapBuffer, used by ffmpeg -hwaccel vaapi -hwaccel_output_format
nv12), readback is corrupted in exactly the iter1 patch-0011 pattern:

  size=6,220,800 bytes (correct: 2 × 1920×1080×1.5 NV12)
  non-zero=544 (0.009%)
  pattern: 16 consecutive non-zero bytes at every 1920-byte row stride,
           rest of buffer reads as zero
  diff vs SW reference: 100% of bytes differ, MAE=53.3 per byte

This is the canonical stale-cached-mmap pattern. Kernel writes real
pixels (proven by DMA-BUF GL import readback succeeding), but the
libva backend's image-export path returns a cached pointer without
the correct cache-invalidation incantation. Userspace reads stale
all-zero memory punctuated by whichever cache lines happened to fetch
post-write.

Phase 4 work item: audit whether the iter1 patch-0011 cache-flush
fix is present, effective, or RK3399-routing-bypassed. Three
possibilities: (a) fix landed for RK3568 but cache topology differs
on RK3399, (b) fix is gated on something that's not true on RK3399,
or (c) RK3399 V4L2_MEMORY_MMAP page protection bypasses the flush.
Not gating Phase 0 — kernel-side decode is correct.

Phase 1+ binding cells must use the DMA-BUF GL import path for pixel
verification, not vaDeriveImage / cached-mmap. The iter1 lesson
restated: cached-mmap readback is unreliable on this hardware family.

Evidence files (under phase0_evidence/2026-05-07/h264_baseline_trace.md
and h264_baseline/):

  - mpv.stdout — libva log, vaapi-copy engaged, cap_pool_init
  - h264_baseline_trace.md — full writeup with re-run incantations
  - mpv.strace.* (gitignored) — 19 per-thread ioctl/openat traces
  - ftrace_v4l2.txt (gitignored) — kernel qbuf/dqbuf events
  - merged_ioctls.tsv (gitignored) — time-sorted V4L2/MEDIA/DRM
    ioctls across all threads
  - *.jpg (gitignored) — HW vs SW JPEG comparison artefacts
  - frames_hw_cached_readback.nv12 (gitignored) — broken nv12
    readback for forensic reference

gitignore: extended extension list (jpg, png, nv12, yuv, tsv, strace*).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-05-07 21:32:36 +00:00
claude-noether 60e62da666 phase 0 corrections: hantro RK3399 has no H.264; substrate is iter8 master 
Two empirical corrections to the morning-of-2026-05-07 phase 0 lock,
based on V4L2 inventory + iter8 fork build smoke captured this evening
on fresnel (kernel 6.19.9-99-eos-arm).

Correction 1 — hantro-vpu-dec on RK3399 does not advertise H.264.

phase0_findings.md morning lock claimed both rkvdec (/dev/video3)
and hantro-vpu-dec (/dev/video5) advertise H.264. Empirical
v4l2-ctl --list-formats-out shows hantro-vpu-dec exposes only
MG2S (MPEG-2) and VP8F (VP8) — no S264. Likely carryover from
RK3568 (ohm) hantro, which does support H.264; the RK3399 hantro
kernel variant in drivers/media/platform/verisilicon/ registers
a different codec list. Fix:

  - README.md hardware-target table: drop "+ H.264" from Decoder
    block 2.
  - README.md decode-side surface-area paragraph: note hantro is
    MPEG-2 + VP8 only and that there is exactly one bind for H.264
    (rkvdec).
  - phase0_findings.md mechanism table: drop H.264 from /dev/video5
    row; correct DT compatible to rockchip,rk3399-vpu (the actual
    parent device compatible — sysfs reports rockchip,rk3399-vpu,
    not rockchip,rk3399-vpu-dec which is just the v4l2 card type
    string).
  - phase0_findings.md "H.264 lands on both blocks" sentence:
    inverted to "H.264 lands only on rkvdec".
  - phase0_findings.md Open Question #2 (two-block H.264 routing):
    marked RESOLVED 2026-05-07 evening (null). Single bind, no
    routing decision, one test cell per codec.

Empirical evidence: phase0_evidence/2026-05-07/v4l2_inventory_findings.md
(distilled from v4l2_inventory.txt — the latter is gitignored as
raw data, regenerable via the v4l2-ctl invocation documented in
the findings file).

Correction 2 — substrate is iter8 master (65969da), not iter5-end.

phase0_findings.md morning lock framed the substrate as "iter5-end
fork." That was true on 2026-05-05 (iter5 close); between then and
the 2026-05-07 fresnel-fourier scaffold libva-multiplanar continued
through iter6 (per-OUTPUT-slot REINIT request_fd binding), iter7
(slot-leak fix, cap_pool harness, msync verify harness, OUTPUT-pool
teardown), and iter8 (perf binding cell harness, RK3566/3568 doc
fix). Building from master tip 65969da inherits all the iter6-iter8
hardening at zero cost. Fix:

  - phase0_findings.md substrate paragraph: strikethrough the
    "iter5-end" framing, add corrected paragraph naming master
    tip 65969da and listing what iter6/7/8 added.
  - phase0_findings.md top-of-doc: add an "Empirical corrections
    2026-05-07 evening" callout linking to the evidence files,
    so a reader spotting the locked-vs-corrected mismatch knows
    where the empirical update came from.

Empirical evidence: phase0_evidence/2026-05-07/iter8_build_smoke.md
(clean build, vainfo profile enumeration, HEVC anomaly write-up).

What's preserved on purpose:

The strikethrough rendering in phase0_findings.md keeps the original
locked text visible alongside the correction — campaign convention
treats locks as historical record, not editable state. A reader
landing on the file from a deep link sees both the morning's
intent and the evening's empirical update. Git history has the
clean diff if anyone wants the original without strikethrough.

What's not changed:

The codec scope in the locked research question stays correct in
count — five codecs (H.264 + HEVC + VP9 + MPEG-2 + VP8). The
routing table changes (H.264 → rkvdec only; MPEG-2 → hantro only;
no shared block) but the boolean-correctness pass/fail criterion
per codec is unaffected. Phase 1 lock can proceed on the corrected
map without re-opening scope.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-05-07 20:57:40 +00:00
claude-noether e83298f0da phase 0 deliverables 1-3: SDDM watchpoint, V4L2 inventory, iter8 fork smoke 
Three Phase 0 deliverables, three findings worth flagging.

Deliverable #1 — SDDM recovery (phase0_recovery_2026-05-07.md):

Closed as watchpoint, not as root-caused fix. Greeter is green on
this boot — mfritsche has been in Plasma Wayland on tty1 since 20:32.
The "Process crashed (exit code 1)" in journalctl -u sddm is the
post-login greeter teardown, not the pre-login crash described in
~/.claude/plans/dynamic-forging-piglet.md. No coredumps, no qFatal
strings. No package changed since the 2026-04-28 Syyuu; the only
difference between the failing boot and this one is a reboot —
likely a flaky panfrost/GBM cold-init that happened to succeed.
Plan procedures + /var/cache/pacman/pkg rollback candidates remain
ready if regression fires.

Deliverable #2 — V4L2 inventory (phase0_evidence/2026-05-07/
v4l2_inventory_findings.md, raw capture in v4l2_inventory.txt
which is gitignored as raw data):

Full v4l2-ctl --all + --list-ctrls-menus + --list-formats-out per
node, plus media-ctl topology, plus DT compatibles. Authoritative
codec map on running kernel 6.19.9-99-eos-arm:

  /dev/video3 (rkvdec, rockchip,rk3399-vdec):
    OUTPUT_MPLANE: S265 (HEVC), S264 (H.264), VP9F (VP9)
    CAPTURE_MPLANE: NV12

  /dev/video5 (hantro-vpu-dec, rockchip,rk3399-vpu):
    OUTPUT_MPLANE: MG2S (MPEG-2), VP8F (VP8)
    CAPTURE_MPLANE: NV12

This contradicts phase0_findings.md (and README.md), which both
claim hantro-vpu-dec on RK3399 also does H.264. It does not on
this kernel. Open Question #2 from phase0_findings.md ("two-block
H.264 routing") is null on RK3399. Correction commit follows.

Deliverable #3 — iter8 fork build + vainfo (phase0_evidence/
2026-05-07/iter8_build_smoke.md):

Built libva-v4l2-request-fourier master tip 65969da on fresnel
directly (no distcc per locked precedent), gcc 15.2.1, meson 1.11.1,
ninja 1.13.2, libva 1.23.0, libdrm 2.4.131. Clean build, 302 KB
.so, two harmless v4l2.h forward-decl warnings. Installed to
/usr/lib/dri/v4l2_request_drv_video.so. vainfo enumerates:

  rkvdec bind: H.264 {Main, High, ConstrainedBaseline, MultiviewHigh,
                      StereoHigh}, HEVCMain
  hantro-vpu-dec bind: MPEG-2 {Simple, Main}

Substrate is iter8 master, not iter5 as the campaign docs frame —
libva-multiplanar continued past iter5 into iter6-iter8 between the
2026-05-05 iter5 close and the 2026-05-07 fresnel-fourier scaffold.
Building from master inherits per-OUTPUT-slot REINIT, slot-leak
fix, cap_pool harness, msync verify harness, OUTPUT-pool teardown.
Correction commit follows.

HEVC anomaly worth flagging: src/config.c:146-151 probes
V4L2_PIX_FMT_HEVC_SLICE before adding VAProfileHEVCMain. rkvdec
advertises S265 so the probe succeeds; HEVCMain gets enumerated.
But src/meson.build excludes h265.c from the build. A consumer
that calls vaCreateConfig(VAProfileHEVCMain) will succeed (config.c
validation list includes HEVCMain) but actual decode will fault
at dispatch since no h265 symbols are linked. Phase 4 decision:
re-enable h265.c, gate enumeration on a compile-time #ifdef, or
strip enumeration honestly.

Per-codec routing confirmed one-env-var-per-process: request.c:149
reads LIBVA_V4L2_REQUEST_VIDEO_PATH once at init. A single backend
instance binds to either rkvdec or hantro-vpu-dec, not both. Phase 4
will need either a wrapper-script-per-consumer hack or a backend
probe-loop change to route by VAProfile across both decode nodes.

Build infrastructure (gitignore):

Switched from blanket phase*_evidence/ exclude to extension-based
allow-list — track narrative .md, ignore raw .txt/.log/.trace/.pcap/
.bin/.gz/.zst/.json/.dat/.ftrace/.strace. Keeps the V4L2 inventory
text untracked (reproducible from the v4l2-ctl invocation) while
preserving the findings narrative in-repo.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-05-07 20:54:38 +00:00
marfrit c9a96cbb38 fresnel-fourier campaign scaffold: peer to libva-multiplanar, RK3399 target 
Roll out libva-v4l2-request-fourier on fresnel (Pinebook Pro / RK3399)
with broader codec scope than ohm-side libva-multiplanar:

- H.264 + HEVC + VP9 via the rkvdec block (/dev/video3)
- MPEG-2 + VP8 via the hantro-vpu-dec block (/dev/video5)

Topology: peer campaign, independent 8(+1) loop. fresnel-fourier
results do not gate libva-multiplanar Phase 8 closes. Code-side work
in the shared fork (../libva-multiplanar/libva-v4l2-request-fourier/)
lands per Phase 2 source-read of each iteration.

Phase 0 task 1: recover fresnel from the SDDM greeter crash-loop
(per ~/.claude/plans/dynamic-forging-piglet.md). Recovery is bookkept
inside this campaign's Phase 0, not as an out-of-band prereq.

Verified 2026-05-07 via SSH:
- linux-eos-arm 6.19.9-99 has CONFIG_FTRACE=y + CONFIG_FUNCTION_TRACER=y
  + CONFIG_DYNAMIC_FTRACE=y + CONFIG_TRACING=y. /sys/kernel/tracing/
  populated. No kernel rebuild needed for trace work.
- /dev/video3 = rkvdec (NV12 capture, MPLANE).
- /dev/video5 = hantro-vpu, card type rockchip,rk3399-vpu-dec.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-05-07 20:11:04 +00:00