Iteration 1 closes with all five Phase 1 boolean-correctness criteria
green. Second codec passes — campaign scoreboard 1/5 → 2/5 (H.264
in T4, MPEG-2 in iter1). Loop terminates per
feedback_dev_process.md Phase 8.
What landed:
Code (libva-v4l2-request-fourier master 65969da..229d6d1):
e7dad7a iter1 Phase 6 commit A: config.c break for MPEG-2 cases
5fe873c iter1 Phase 6 commit B: rewrite mpeg2.c against new V4L2 stateless API
3aab187 iter1 Phase 6 commit C: delete staging-era include/mpeg2-ctrls.h
229d6d1 iter1 Phase 6 commit D: drop missed mpeg2-ctrls.h include from context.c (fix-forward)
All four authored as Claude (noether) per feedback_gitea_as_claude_noether.md.
Campaign docs (fresnel-fourier):
f720c77 iter1 Phase 0 + Phase 1 lock
cc55a6e iter1 Phase 2 situation analysis (3 bugs)
b9625af iter1 Phase 3 baseline measurements (4 baselines)
3e996d0 iter1 Phase 4 plan (6 contract clauses)
0e2e1c2 iter1 Phase 5 sonnet review (6 findings, 4 amendments)
ec9133a iter1 Phase 7 verification (5/5 GREEN)
[this commit] iter1 Phase 8 close
Three lessons distilled to durable memory:
L1 — feedback_header_deletion_check.md
Phase 2 grep audit found 2 of 3 include sites for the
staging-era mpeg2-ctrls.h header; build broke on Commit C
delete because of context.c:42. Rule: when removing a header,
let the compiler enumerate includes authoritatively (clean
rebuild after include-removal patches, before git rm). Grep
is a hint; the compiler is the authority.
L2 — feedback_review_empirical_over_theoretical.md
Phase 5 reviewer S2 flagged a numerical mismatch in
vbv_buffer_size between Baseline C (1.31 MB) and predicted
post-fix (1 MB). I rejected with confident source-read
reasoning (slot->size = sizeimage = matches). Phase 7
byte-compare empirically showed the reviewer was right —
post-fix produces 1 MB, not 1.31 MB. Operational impact nil
(kernel ignores the field), but my Phase 5 rebuttal had a
source-read gap. Rule: when a reviewer cites a concrete
numerical discrepancy, defer to Phase 7 byte-compare; don't
reject on source-read alone.
L3 — feedback_rockchip_pixel_verify_path.md
Iter1 + T4 (H.264) both empirically confirm: libva backend's
vaDeriveImage / cached-mmap readback returns all-zero NV12
on RK3399 — same iter1 patch-0011 cache-coherency bug class
observed on RK3568. Pixel verification must use DMA-BUF GL
import (mpv --vo=image, ffmpeg-v4l2request DRM_PRIME +
hwdownload). NOT ffmpeg-vaapi+hwdownload (cache-stale on
Rockchip). Codec-agnostic; applies to all 5 codecs in
campaign scope.
Backlog items deferred (campaign-internal, not durable memory):
B1: V4L2 /dev/videoN numbering shuffles across reboots
(rkvdec moved video3+media1 → video1+media0 between
Phase 0/3 and Phase 7). Backend should probe /dev/media*
for driver match. Iter2+ Phase 4 cross-cutting candidate.
B2: mpv --hwdec=vaapi-copy silently filters MPEG-2 out before
libva is loaded. mpv --hwdec=vaapi (DMA-BUF) DOES engage.
Phase 1 criterion 3 ended up anchored on ffmpeg-direct.
Mpv-side investigation as separate follow-up.
B3: Latent surface-reuse bug — picture.c:287 h264.matrix_set
reset writes byte 240 of params union, lands inside
mpeg2.iqmatrix.chroma_intra_quantiser_matrix[20] (verified
via offsetof on fresnel via gcc + libva). Per-frame
RenderPicture overwrites this byte for ffmpeg-vaapi flows
that send VAIQMatrixBufferType every frame. Latent for
VAAPI clients that reuse a surface without re-sending
IQMatrix. Iter2+ candidate.
B4: src/context.c:142-155 H.264 device-init runs unconditionally
on every CreateContext, EINVALs on hantro. Intentional
best-effort but request_log fires "Unable to set
control(s)" cosmetically. Suppress-log candidate, low
priority.
B5: vbv_buffer_size = SOURCE_SIZE_MAX (1 MB) rather than
negotiated sizeimage. Kernel ignores. Polish candidate.
Phase 4 cross-cutting work items collected:
- Add VIDIOC_EXPBUF + DMA_BUF_IOCTL_SYNC to libva backend
image-export (fixes L3's vaDeriveImage cache-stale bug for
all codecs).
- V4L2 device-discovery probe (fixes B1).
- Picture.c BeginPicture profile-aware reset (fixes B3).
- Context.c H.264 device-init log suppression (fixes B4).
Campaign roadmap (codec iterations remaining):
iter2: HEVC on rkvdec — re-enable h265.c in build, audit against
rkvdec kernel HEVC contract.
iter3: VP8 on hantro — implement vp8.c.
iter4: VP9 on rkvdec — implement vp9.c (largest control surface).
Phase 5 review S2 historical-record correction: Phase 5 reviewer
was numerically right about vbv_buffer_size. My Phase 5 rebuttal
in phase5_iter1_review.md was empirically wrong. Acknowledged in
phase7_iter1_verification.md and phase8_iteration1_close.md;
Phase 5 doc preserved as-is for the historical record.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
12 KiB
Iteration 1 — Phase 8 (close)
Iteration 1 of the fresnel-fourier campaign closes 2026-05-08 with all five Phase 1 boolean-correctness criteria green. Second codec passes: campaign scoreboard advances 1/5 → 2/5 (H.264 from T4, MPEG-2 from iter1).
What iter1 fixed
MPEG-2 hardware-accelerated decode on Rockchip RK3399 / hantro-vpu-dec / libva-v4l2-request-fourier — bit-exact correct against software reference, when read via the cache-coherency-safe DMA-BUF GL import path.
Three independent bugs (per Phase 2 source-read), all in the libva backend (kernel + driver path was already solid per Phase 0 cross-validator sweep):
-
src/config.c:55-69fall-through todefault:— MPEG-2 case labels existed but lackedbreak;, sovaCreateConfig(VAProfileMPEG2Main)returnedVA_STATUS_ERROR_UNSUPPORTED_PROFILE(= 12). Fix: 5-linebreak;addition. -
src/mpeg2.cused staging-era UAPI —V4L2_CID_MPEG_VIDEO_MPEG2_SLICE_PARAMS(=V4L2_CID_MPEG_BASE+250= 0x9909fa) and the combined slice_params struct. Mainline kernel removed these in favor of the splitV4L2_CID_STATELESS_MPEG2_{SEQUENCE,PICTURE,QUANTISATION}(= 0xa409dc/dd/de) with three separate control structs. Fix: full rewrite (~120 lines) against the new API; one batchedVIDIOC_S_EXT_CTRLS count=3per frame matching FFmpeg'slibavcodec/v4l2_request_mpeg2.c:130-155reference. -
include/mpeg2-ctrls.hwas the staging-era local header — masking the kernel's modern<linux/v4l2-controls.h>definitions. Fix: delete the local header. (Required dropping#includefrom THREE files —src/config.c:37,src/mpeg2.c:38, andsrc/context.c:42. Phase 2 grep audit only found the first two; Commit D fix-forward caught the third.)
Contract that was followed
Per feedback_dev_process.md Phase 6 contract-before-code, citations from authoritative sources, mapped 1:1 to the patch hunks (full detail in phase4_iter1_plan.md Clauses 1–6):
- Linux mainline UAPI:
include/uapi/linux/v4l2-controls.h:1985-2105for the new MPEG-2 stateless control IDs, struct layouts, and flag macros. - FFmpeg Kwiboo branch:
libavcodec/v4l2_request_mpeg2.c:130-155for the batched submission shape (3v4l2_ext_controlentries, singleVIDIOC_S_EXT_CTRLS). - Linux mainline driver:
drivers/media/platform/verisilicon/hantro_mpeg2.c::hantro_mpeg2_dec_copy_qtable— confirms zigzag-to-raster permutation is kernel-side; libva backend's quantisation matrices stay in zigzag order (no permutation in userspace). - Empirical anchor: Phase 3 Baseline C (
phase0_evidence/2026-05-07/iter1_phase3/baseline_C_xvalidator/) — verbatim 256-byte QUANTISATION payload from ffmpeg-v4l2request decode of the BBB MPEG-2 fixture, used for byte-by-byte transcription ofmpeg2_default_intra_matrix[64](per Phase 5 S3 amendment that flagged spec-recall as unreliable).
What landed where
Code commits on git.reauktion.de/marfrit/libva-v4l2-request-fourier:
229d6d1 fresnel-fourier iter1 Phase 6 commit D: drop missed mpeg2-ctrls.h include from context.c (fix-forward)
3aab187 fresnel-fourier iter1 Phase 6 commit C: delete staging-era include/mpeg2-ctrls.h
5fe873c fresnel-fourier iter1 Phase 6 commit B: rewrite mpeg2.c against new V4L2 stateless API
e7dad7a fresnel-fourier iter1 Phase 6 commit A: config.c break for MPEG-2 cases
65969da iter8 Phase 4: tests/run_perf_binding_cell.sh — perf binding cell harness (libva-multiplanar tip)
All four commits authored as Claude (noether) per memory feedback_gitea_as_claude_noether.md.
Campaign documents on git.reauktion.de/marfrit/fresnel-fourier:
ec9133a iter1 Phase 7: verification — all 5 criteria GREEN, second codec PASS
0e2e1c2 iter1 Phase 5: sonnet-architect review — 6 findings, 4 amendments
3e996d0 iter1 Phase 4: plan — contract clauses, diff scope, Phase 7 harness
cc55a6e iter1 Phase 2: situation analysis — three bugs in MPEG-2 path
b9625af iter1 Phase 3: baseline measurements — Phase 2 confirmed empirically
f720c77 iter1 Phase 0 + Phase 1 lock: MPEG-2 boolean correctness on hantro
(Plus campaign Phase 0 setup commits + this Phase 8 close.)
Phase 1 binding-cell verification (Phase 7 verbatim)
| Criterion | Result |
|---|---|
| 1: vainfo MPEG-2 Simple+Main enumeration | ✅ both profiles with VAEntrypointVLD |
2: vaCreateConfig(VAProfileMPEG2Main) |
✅ VA_STATUS_SUCCESS (libva trace verbatim) |
| 3: ffmpeg-hwaccel-vaapi engages, exit 0 | ✅ 5 frames, exit 0, no Failed to create decode configuration |
| 4: DMA-BUF GL HW=SW byte-identical (+02s) | ✅ HW1=SW1=6e7873030dbf..., HW2=SW2=ccc7ce08810d..., frames differ |
| 5: T4 H.264 reference hashes match | ✅ HW + SW frames match f623d5f7... and 7d7bc6f2... |
Bonus: post-fix VIDIOC_S_EXT_CTRLS payload structurally matches Baseline C cross-validator (count=3, ctrl_class=0xf010000=V4L2_CTRL_CLASS_CODEC_STATELESS, sizes 12/32/256, QUANTISATION intra_matrix bytes byte-identical). One numerical divergence in vbv_buffer_size (1 MB post-fix vs 1.31 MB cross-validator) — kernel-ignored informational field, non-blocking.
Lessons (for Phase 8 memory update)
The 8-phase loop caught five lessons worth distilling. Three become durable memory entries (next section); two stay as campaign-internal backlog notes.
Worth a memory entry
L1 — Header deletion: completeness check is git rm + clean rebuild, not grep alone.
Phase 2 source-read used git grep "mpeg2-ctrls.h" to enumerate include sites. The grep returned src/config.c:37 and src/mpeg2.c:38 — Phase 4 plan and Commit B + C scope expanded only those two files. After Commit C (git rm include/mpeg2-ctrls.h), the build broke on src/context.c:42 — a third include the grep had missed (most likely because of grep tooling quirk + my path filtering). Cost: one fix-forward commit (D) and a moment of confused debugging.
Rule: when removing a header, schedule git rm during the patch series and verify with a clean rebuild. The compiler enumerates includes authoritatively; grep is best-effort. Memory entry: feedback_header_deletion_check.md.
L2 — When the Phase 5 reviewer flags a numerical discrepancy, lean empirical, not theoretical.
Phase 5 reviewer's S2 finding said: "post-fix vbv_buffer_size will be SOURCE_SIZE_MAX (1 MB), Baseline C shows 0x151800 (1.31 MB) — discrepancy unexplained." I responded with confident source-read reasoning that slot->size = sizeimage = 1382400 ergo equality holds, and rejected the finding. Phase 7 byte-compare empirically showed: post-fix produces 1 MB, NOT 1.31 MB. The reviewer was numerically right; my source-read was wrong (the actual slot->size resolution path doesn't equal the negotiated sizeimage in our backend's S_FMT setup).
Operational impact: nil (kernel ignores vbv_buffer_size per the doc). But the historical record of "I rejected this finding with the wrong reason" is itself the lesson.
Rule: when a reviewer cites a concrete numerical mismatch between two captured payloads, run the empirical byte-compare in Phase 7 BEFORE rejecting on source-read grounds. Source-read can have gaps; verbatim-vs-verbatim comparison can't lie. Memory entry: feedback_review_empirical_over_theoretical.md.
L3 — On RK3399 (and Rockchip in general?): vaDeriveImage cache-stale bug class is real and codec-agnostic.
Both H.264 (T4) and MPEG-2 (iter1) decode paths produce all-zero NV12 when read via libva's vaDeriveImage + cached mmap (the path that ffmpeg-vaapi + hwdownload uses). The kernel + driver decode is correct (verified by ffmpeg-v4l2request DRM_PRIME path producing matching pixels to SW reference). The bug is in the libva backend's image-export — same iter1 patch-0011 bug class observed on RK3568 (ohm) but pre-fixed there.
Rule: pixel verification on Rockchip platforms must use the DMA-BUF GL import path (mpv --vo=image or equivalent), NEVER the cached-mmap readback path. Treat ffmpeg-vaapi+hwdownload's correct-pixel guarantee as suspect on these SoCs until backend gains VIDIOC_EXPBUF + DMA_BUF_IOCTL_SYNC support. Memory entry: feedback_rockchip_pixel_verify_path.md.
Stays as campaign-internal backlog (not durable memory)
B1 — V4L2 /dev/videoN numbering shuffles across reboots on RK3399. Phase 0/3 had rkvdec=video3+media1, hantro=video5+media2; this boot has rkvdec=video1+media0, hantro=video3+media1. Hardcoded env-var paths in Phase 1 binding cells are fragile. Backend should probe /dev/media* for driver=hantro-vpu/rkvdec rather than rely on env var stability. Iter2+ Phase 4 cross-cutting candidate.
B2 — mpv --hwdec=vaapi-copy silently filters MPEG-2 out before libva is loaded. mpv's hwdec policy or ffmpeg-mpv glue layer rejects MPEG-2 from the vaapi-copy candidates, so the binding cell never engaged the libva backend. mpv --hwdec=vaapi (DMA-BUF, not -copy) DOES engage. Phase 1 criterion 3 anchored on ffmpeg-direct as a result; mpv-vaapi-copy-MPEG-2 is a separate follow-up task (mpv-side investigation, possibly an --hwdec-codecs filter override).
Latent bugs surfaced but deferred
B3 — src/picture.c:287 params.h264.matrix_set = false writes byte 240 of the surface's params union, which (per offsetof verification on fresnel via gcc + libva) lands inside mpeg2.iqmatrix.chroma_intra_quantiser_matrix[20]. For ffmpeg-vaapi-driven flows this byte is overwritten by the per-frame VAIQMatrixBufferType (which codec_store_buffer:140-146 copies wholesale). For VAAPI clients that reuse a surface without re-sending the IQMatrix on every cycle, the corruption persists. Latent bug; not exercised by iter1's binding cells. Iter2+ candidate — the fix is making the BeginPicture matrix_set reset profile-aware, or simpler memset(¶ms, 0, sizeof params).
B4 — src/context.c:142-155 H.264 device-init runs unconditionally on every CreateContext regardless of profile. Returns EINVAL on hantro-vpu-dec (which has no H.264 controls). Intentional best-effort behavior — return cast to (void) and discarded — but src/v4l2.c:484: request_log("Unable to set control(s): %s\n", strerror(errno)) still fires unconditionally, so every MPEG-2 decode session prints v4l2-request: Unable to set control(s): Invalid argument once at init. Cosmetic only; iter2+ cleanup candidate (suppress the log when caller intent is best-effort silent fail).
B5 — vbv_buffer_size = SOURCE_SIZE_MAX = 1 MB rather than the kernel-negotiated sizeimage. Informational field, kernel ignores. Low-priority post-iter1 polish.
What's next (campaign roadmap)
Per phase0_evidence/2026-05-07/cross_validator_traces.md suggested ordering, the campaign now has:
| Iter | Codec | Status |
|---|---|---|
| iter1 | MPEG-2 (hantro) | ✅ closed (this iteration) |
| iter2 | HEVC (rkvdec) | open — re-enable h265.c in build, audit against rkvdec kernel contract |
| iter3 | VP8 (hantro) | open — implement vp8.c |
| iter4 | VP9 (rkvdec) | open — implement vp9.c (largest kernel control surface) |
| iterN | Phase 4 cross-cutting | open — vaDeriveImage cache-stale fix; V4L2 device-discovery; B3-B5 polish |
Phase 8 doesn't pre-lock iter2; that decision is the next operator-driven moment. When iter2 opens, its Phase 0 substrate carries forward iter1's findings (especially B3-B5 if those iterations risk exercising them).
Memory updates (next section)
Three new memory entries land alongside this close:
feedback_header_deletion_check.md(L1)feedback_review_empirical_over_theoretical.md(L2)feedback_rockchip_pixel_verify_path.md(L3)
Plus an update to phase5_iter1_review.md to acknowledge S2 was numerically right (the historical-record correction).
Plus update to MEMORY.md index.
Iteration close acknowledgement
Per feedback_dev_process.md's Phase 8: "Loop terminates here." Iteration 1 of fresnel-fourier closes here. The 8-phase loop ran once end-to-end, no Phase 7 → Phase 4 loopbacks needed (one Phase 6 → Phase 4 mid-implementation surface for Commit D fix-forward). Phase 5 sonnet review caught two amendments worth landing (S3, Q4) plus one finding (S2) my response was wrong about. The campaign is on track.