fresnel-fourier/phase0_evidence/2026-05-07/cross_validator_traces.md

# Cross-validator contract traces (per-codec) — fresnel 2026-05-07

Phase 0 deliverable #6. Per-codec V4L2 + media-request contract traces captured from an **independent V4L2 client** consuming each fixture, to validate (a) the kernel + driver path works for every locked codec and (b) anchor what V4L2 traffic per codec should look like for Phase 6 implementation work in the libva backend.

## Substrate change vs the locked deliverable

`phase0_findings.md` and `README.md` named **chromium-fourier 149** as the cross-validator. chromium-fourier is **not installed** on fresnel as of 2026-05-07 (`pacman -Ss chromium-fourier` returns nothing; only stock `chromium 147.0.7727.116-1` and `brave-bin 1.87.191-1`). The `marfrit-packages` repo is also not configured on fresnel (`grep -A3 "marfrit\|reauktion" /etc/pacman.conf` empty), so binary distribution to fresnel is currently manual.

**Substituted cross-validator: `ffmpeg -hwaccel v4l2request`** (ffmpeg `n8.1-13-gb57fbbe50c`, the `v4l2-request-n8.1` Kwiboo branch already in stock ALARM).

Rationale — this is actually a **better** cross-validator than chromium-fourier 149 for fresnel-fourier's purposes:

- **Different V4L2 client code path.** ffmpeg's `v4l2request` hwaccel lives in `libavcodec/v4l2_request*.c`. It uses no libva at all — it's a direct V4L2 + media-request consumer. So it cross-validates our libva backend's kernel-side traffic against an independent implementation, exactly the cross-validation function chromium-fourier 149's `media/gpu/v4l2/` was supposed to provide.
- **Already on the box.** No package install dance, no version locking exposure.
- **All five codecs implemented.** ffmpeg has `mpeg2_v4l2request`, `h264_v4l2request`, `hevc_v4l2request`, `vp8_v4l2request`, `vp9_v4l2request` decoder paths. chromium might or might not — irrelevant since we're not testing it.

The chromium-fourier 149 cross-validator is **not lost** as a Phase 1+ goal — once chromium-fourier is installed on fresnel via the standard marfrit-packages bootstrap, a follow-up trace pass per codec will cross-check ffmpeg-v4l2request and chromium contracts. For Phase 0 baseline, the ffmpeg-v4l2request anchor is sufficient.

## Headline finding: all 5 codecs work via the kernel direct path

```
$ ffmpeg -hwaccel v4l2request -i bbb_<codec>.<ext> -frames:v 2 -f null - ; echo "[exit $?]"

H.264  → exit 0
HEVC   → exit 0
VP9    → exit 0
MPEG-2 → exit 0
VP8    → exit 0
```

**Every locked codec on fresnel decodes end-to-end** when consumed via `ffmpeg -hwaccel v4l2request`. The Linux kernel + rkvdec + hantro-vpu drivers on RK3399 are solid for the campaign codec scope.

This dramatically shrinks the Phase 6 fork-side work scope: nothing needs fixing in the kernel. Every Phase 6 work item is in the libva backend (config.c profile validation, codec source files, control submission, image export).

## Per-codec status: libva (our backend, iter8) vs ffmpeg-v4l2request

Run via `mpv --hwdec=vaapi-copy ...` (libva path) and `ffmpeg -hwaccel v4l2request ...` (direct V4L2 path), 2 frames each:

| Codec | libva via iter8 | ffmpeg-v4l2request | Diagnosis |
|---|---|---|---|
| **H.264** | ✅ HW decode, pixel-perfect via DMA-BUF (T4) | ✅ exit 0 | Both paths work; campaign cell **passes**. |
| **HEVC** | ❌ `vaCreateConfig: 12 (UNSUPPORTED_PROFILE)` | ✅ exit 0 | `src/h265.c` excluded from build (commented out in `src/meson.build`); `src/config.c:151` enumerates HEVCMain anyway because the V4L2 probe of `V4L2_PIX_FMT_HEVC_SLICE` succeeds. Phase 6 fix: re-enable h265.c in build, audit it against rkvdec HEVC kernel contract. |
| **VP9** | ❌ `No support for codec vp9 profile 0` (not enumerated) | ✅ exit 0 | No `vp9.c` in fork at all. Phase 6 fix: write VP9 source file, model on FFmpeg's `v4l2_request_vp9.c` and kernel `rkvdec_vp9.c`. |
| **MPEG-2** | ❌ `vaCreateConfig: 12 (UNSUPPORTED_PROFILE)` | ✅ exit 0 | `mpeg2.c` IS compiled in fork. `src/config.c:64-65` has both VAProfileMPEG2Simple/Main cases. Yet vaCreateConfig fails. Suspects: device-discovery / V4L2 capability check in `RequestCreateConfig` runtime path; or a regression in mpeg2 path between Bootlin upstream and fork tip; or a routing-by-codec issue (the libva backend was bound to /dev/video5 via env var, but maybe the codec dispatch checks the wrong format list). Phase 2 source-read of config.c needed before Phase 6 implementation. |
| **VP8** | ❌ `No support for codec vp8 profile -99` (not enumerated) | ✅ exit 0 | No `vp8.c` in fork. Same shape as VP9: model on FFmpeg's `v4l2_request_vp8.c` and kernel `hantro_vp8.c`. |

So the boolean-correctness-per-codec scoreboard for Phase 0 is:

| Codec | Phase 0 boolean correctness | Phase 6 work item |
|---|---|---|
| H.264 | ✅ PASS | (already works; pixel-cache-coherency on vaDeriveImage path is a side issue — see T4 writeup) |
| HEVC | ❌ FAIL via libva, ✅ kernel-side OK | enable h265.c build + audit |
| VP9 | ❌ FAIL via libva, ✅ kernel-side OK | implement vp9.c |
| MPEG-2 | ❌ FAIL via libva, ✅ kernel-side OK | debug RequestCreateConfig MPEG-2 rejection |
| VP8 | ❌ FAIL via libva, ✅ kernel-side OK | implement vp8.c |

Phase 1 lock can take this directly: each codec is a separate iteration's binding cell, ordered by complexity (MPEG-2 likely cheapest if it's just a config-side bug, VP9 likely most expensive given the kernel control surface).

## Per-codec ioctl frequency (cross-validator anchor)

From `strace -ff -tt -y -e ioctl` on `ffmpeg -hwaccel v4l2request` decoding 2 frames each:

| ioctl | H.264 | HEVC | VP9 | MPEG-2 | VP8 |
|---|---|---|---|---|---|
| VIDIOC_DQBUF | 45 | 49 | 40 | 26 | 49 |
| VIDIOC_QBUF | 22 | 24 | 20 | 10 | 20 |
| VIDIOC_CREATE_BUFS | 17 | 17 | 17 | 12 | 17 |
| VIDIOC_QUERYBUF | 15 | 15 | 15 | 10 | 15 |
| VIDIOC_S_EXT_CTRLS | 13 | 14 | 11 | 5 | 10 |
| VIDIOC_EXPBUF | 11 | 11 | 11 | 6 | 11 |
| VIDIOC_QUERY_EXT_CTRL | 0 | 5 | 0 | 0 | 0 |
| MEDIA_IOC_REQUEST_ALLOC | 4 | 4 | 4 | 4 | 4 |
| MEDIA_REQUEST_IOC_QUEUE | varies | varies | varies | varies | varies |
| MEDIA_REQUEST_IOC_REINIT | 0 (not pooled) | 0 | 0 | 0 | 3 |
| DMA_BUF_IOCTL_SYNC | 0 | 0 | 0 | 4 | 0 |

Notable structural observations:

- **VIDIOC_EXPBUF** appears across all codecs (≥6 calls). ffmpeg-v4l2request exports CAPTURE buffers as DMA-BUF FDs to hand off to downstream consumers (or for cache-safe access). Our libva backend's iter8 trace did **not** show EXPBUF in the per-frame loop. This is an architectural divergence — ffmpeg uses DMA-BUF export by default for the readback-safe path; libva uses cached mmap via vaDeriveImage. Hypothesis (consistent with T4's findings): adding VIDIOC_EXPBUF + cache-aware DMA-BUF mapping to the libva backend's image-export path would fix the iter1 patch-0011 bug class on RK3399.
- **DMA_BUF_IOCTL_SYNC × 4 only on MPEG-2.** That's `DMA_BUF_SYNC_START | _END` calls — explicit cache management. Why only on MPEG-2? Probably because the MPEG-2 fixture is the only one that exercises the consumer's CPU readback path during these 2 frames; the others output direct via DMA-BUF FD passthrough. Not gating; useful detail.
- **VIDIOC_QUERY_EXT_CTRL × 5 only on HEVC.** ffmpeg's HEVC v4l2request decoder runtime-introspects the kernel's HEVC control struct (the `hevc_slice_params` dynamic-array, max 600 elements). H.264 control structs are version-stable so no QUERY_EXT_CTRL needed; HEVC isn't.
- **MEDIA_IOC_REQUEST_ALLOC = 4** for every codec — ffmpeg uses a pool of 4 request_fds, recycled. Our libva backend allocates 16 (iter6 per-OUTPUT-slot binding). 4× difference in pool depth.
- **MEDIA_REQUEST_IOC_REINIT** absent from H.264/HEVC/VP9/MPEG-2 ffmpeg traces, present 3× on VP8. ffmpeg's request-fd recycling pattern differs by codec (probably a function of how its decoder engages the request lifecycle). Our libva backend uses REINIT per-frame for all codecs (iter6 pattern). Both are valid.

## Per-codec ftrace v4l2 events (kernel perspective)

ftrace shows the kernel-side qbuf/dqbuf flow per codec:

| Codec | ftrace lines | minor (device) | Notes |
|---|---|---|---|
| H.264 | 56 | minor=3 (rkvdec) | Frame-threaded ffmpeg workers (av:h264:dfN). Standard contract. |
| HEVC | 60 | minor=3 (rkvdec) | Frame-threaded (av:hevc:dfN). Plus `dec0:0:hevc-NNNN` thread (ffmpeg's main decoder pump). Slightly more events than H.264 (HEVC needs slice_params per slice, hence more S_EXT_CTRLS). |
| VP9 | 52 | minor=3 (rkvdec) | Standard frame-cycle pattern, slightly fewer events than H.264 (VP9 has frame-level controls only, no slice-mode). |
| MPEG-2 | 32 | minor=5 (hantro-vpu-dec) | Single-threaded `dec0:0:mpeg2vid` (no frame-threading for MPEG-2 in ffmpeg). |
| VP8 | 52 | minor=5 (hantro-vpu-dec) | Frame-threaded (av:vp8:dfN). |

Sample HEVC per-frame pattern (from the ftrace head, decode loop):

```
av:hevc:df0  v4l2_qbuf  minor=3 OUTPUT_MPLANE  index=0  flags=...|0x800080
av:hevc:df0  v4l2_qbuf  minor=3 CAPTURE_MPLANE index=10 flags=MAPPED|QUEUED|...
av:hevc:df1  v4l2_dqbuf minor=3 OUTPUT_MPLANE  index=0  flags=...|0x800000
av:hevc:df1  v4l2_dqbuf minor=3 CAPTURE_MPLANE index=10 flags=MAPPED|...
```

Same OUTPUT/CAPTURE buffer-pair contract as H.264. The structural pattern transfers across codecs unchanged — what varies is the S_EXT_CTRLS payload (codec-specific control class + struct).

Sample MPEG-2 per-frame pattern:

```
dec0:0:mpeg2vid  v4l2_qbuf  minor=5 OUTPUT_MPLANE  index=0 flags=...|0x800080  timestamp=6000
dec0:0:mpeg2vid  v4l2_qbuf  minor=5 CAPTURE_MPLANE index=5 flags=MAPPED|QUEUED|...
dec0:0:mpeg2vid  v4l2_dqbuf minor=5 OUTPUT_MPLANE  index=0 ...
dec0:0:mpeg2vid  v4l2_dqbuf minor=5 CAPTURE_MPLANE index=5 ...
```

Hantro-vpu-dec (`minor=5`) follows the same QBUF/DQBUF pattern as rkvdec, just on a different V4L2 device. Confirms the contract is uniform across both decode blocks on RK3399.

## What this means for Phase 1 lock and beyond

1. **Phase 1 lock can proceed.** Boolean-correctness criterion is well-defined and measurable for each codec. The per-codec status table above tells us which iterations have what shape.
2. **Phase 6 work scope is purely libva-backend code.** Kernel + driver path is solid for all 5 codecs. No kernel patches needed; no upstream Linux engagement needed (good news per `feedback_no_upstream.md`).
3. **Iteration ordering** suggested for fresnel-fourier (subject to Phase 1 lock):
   - **iter1**: MPEG-2 — likely the cheapest fix (config.c-level path investigation; mpeg2.c already compiled). Fastest path to "second codec works."
   - **iter2**: HEVC — re-enable h265.c in build, audit against rkvdec kernel contract, address whatever caused it to be stripped originally.
   - **iter3**: VP8 — implement vp8.c, model on FFmpeg's `v4l2_request_vp8.c` and kernel `hantro_vp8.c` (similar shape to MPEG-2 path on hantro).
   - **iter4**: VP9 — implement vp9.c, the most code (kernel `rkvdec_vp9.c` has the largest control struct surface).
   - **iter5+**: per-codec follow-ups (HEVC 10-bit if silicon does it, VP9 profile 2/3 if relevant, perf metrics per codec).
4. **vaDeriveImage cache-stale bug** (T4 finding): orthogonal to per-codec work, applies to all codecs that succeed via libva. Phase 4 work item, can land independently of any single-codec iteration.

## Re-run incantation

```bash
ssh fresnel '
mkdir -p /tmp/cross_validator
for codec in h264 hevc vp9 mpeg2 vp8; do
  case "$codec" in
    h264)  FIXTURE=~/fourier-test/bbb_1080p30_h264.mp4;;
    hevc)  FIXTURE=~/fourier-test/bbb_720p10s_hevc.mp4;;
    vp9)   FIXTURE=~/fourier-test/bbb_720p10s_vp9.webm;;
    mpeg2) FIXTURE=~/fourier-test/bbb_720p10s_mpeg2.ts;;
    vp8)   FIXTURE=~/fourier-test/bbb_720p10s_vp8.webm;;
  esac
  mkdir -p /tmp/cross_validator/$codec
  
  sudo sh -c "echo 0 > /sys/kernel/tracing/tracing_on; \
              echo 0 > /sys/kernel/tracing/trace; \
              echo 1 > /sys/kernel/tracing/events/v4l2/enable; \
              echo 1 > /sys/kernel/tracing/tracing_on"
  
  strace -ff -tt -y -e trace=ioctl -o /tmp/cross_validator/$codec/ffmpeg.strace \
    ffmpeg -hide_banner -loglevel error -hwaccel v4l2request \
      -i "$FIXTURE" -frames:v 2 -f null - > /tmp/cross_validator/$codec/ffmpeg.stdout 2>&1
  
  sudo cp /sys/kernel/tracing/trace /tmp/cross_validator/$codec/ftrace_v4l2.txt
done
sudo sh -c "echo 0 > /sys/kernel/tracing/tracing_on; echo 0 > /sys/kernel/tracing/events/v4l2/enable"
'
```

## Evidence files

Under `phase0_evidence/2026-05-07/cross_validator/<codec>/`:

| File | Tracked? | Purpose |
|---|---|---|
| `ffmpeg.stdout` | yes | ffmpeg log (typically empty at -loglevel error; presence confirms a clean run) |
| `ffmpeg.strace.*` | gitignored | per-thread strace ioctl dump |
| `ftrace_v4l2.txt` | gitignored | kernel v4l2 tracepoint snapshot |

The .stdout files are tracked; the .strace.* and ftrace_v4l2.txt files are gitignored as raw data, regenerable from the re-run incantation above. Frequency tables in this document are derived from the strace dumps via the same `grep -oE "ioctl\([0-9]+<[^>]+>, [A-Z_]+" | sort | uniq -c` pipeline used in T4.