5bf6acb964
Fork tip c106d95 (was 70196f8). 165 LOC added / 57 removed in src/request.c. All 9 Phase 5 amendments (2 CRIT + 4 IMP + 3 MIN) incorporated. Fresnel offline at push time. Build + install + Phase 7 verify deferred until host returns. Phase 7 sweep ready to execute: vainfo + ffmpeg-vaapi + reboot stability + iter5b/iter6 regression check. Code review verified algorithm correctness against Phase 5 reviewer pseudocode + boltzmann's linux-rockchip source confirms MEDIA_ENT_F_PROC_VIDEO_DECODER is set on rkvdec.c:1382 + hantro_drv.c proc entities. Compile-time syntax untested (no va-api dev headers on noether). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3.9 KiB
3.9 KiB
Iteration 7 — Phase 6 (implementation)
Single fork commit landed 2026-05-13. Build + install + Phase 7 verify pending fresnel uptime (offline at commit-push time).
Commit
| SHA | Files | LOC | Summary |
|---|---|---|---|
c106d95 |
src/request.c |
+165 / -57 | Refactor find_video_node_via_topology → find_decoder_video_node_via_topology (entity-function discrimination + link-graph walk). Two-pass find_codec_device for rkvdec preference. |
Fork tip post-iter7-Phase-6: c106d95 (was 70196f8 at iter5b-β/iter6 close).
Phase 5 amendments incorporated
| ID | Amendment | Where landed |
|---|---|---|
| CRIT-1 | MEDIA_LNK_FL_INTERFACE_LINK link discrimination | find_decoder_video_node_via_topology data-link vs interface-link checks |
| CRIT-2 | Check both source_id AND sink_id of each link | Both inner loops check both endpoints |
| IMP-1 | Decoder-proc / encoder-proc distinction (algorithm safe by construction) | Confirmed correct; no additional code |
| IMP-2 | MEDIA_ENT_F_PROC_VIDEO_DECODER set by rkvdec.c + hantro_drv.c | Confirmed at code level on boltzmann |
| IMP-3 | 2-call MEDIA_IOC_G_TOPOLOGY pattern | Replaces 3-call pattern in old find_video_node_via_topology |
| IMP-4 | Two-pass fd open/close | for (pass = 0; pass < 2; pass++) outer loop |
| MIN-1 | media_out from path preserved |
Unchanged |
| MIN-2 | Single-bit flag test for interface links | flags & MEDIA_LNK_FL_INTERFACE_LINK |
| MIN-3 | Phase 7 MPEG-2/VP8 env-override mandatory | Phase 7 doc will enforce |
All amendments mechanical; no structural plan changes.
Status
- Fork commit
c106d95on noether + gitea. Pushed viaclaude-noetherSSH perfeedback_gitea_as_claude_noether.md. - Fresnel offline at commit time. Cannot syntax-check on noether (no va-api dev headers locally). Build error risk exists but is small — diff is mechanically derived from working code.
- Backend installed SHA on fresnel: unchanged from iter5b-β/iter6 close (
2c6ff82c…). Will update post-build.
Phase 7 plan (pending fresnel uptime)
When fresnel returns:
- Pull on fresnel:
cd ~/src/libva-v4l2-request-fourier && git pull --ff-only origin master. - Build:
ninja -C build. If syntax fails, report errors; fix via amend or new commit. - Install:
sudo install -m 644 build/src/v4l2_request_drv_video.so /usr/lib/dri/v4l2_request_drv_video.so. - Capture new SHA256.
- Phase 7 verification matrix:
- C5 item 1:
vainfo(no env override) lists H.264/HEVC/VP9 profiles (rkvdec). - C5 item 2: ffmpeg-vaapi HEVC fixture engages decoder (strace shows /dev/video opened).
- C5 item 3 (mandatory per Phase 5 MIN-3): MPEG-2/VP8 with explicit env override still decode correctly.
- C5 item 4: VP9 still PASS direct, MPEG-2 still PASS, no regression on Bugs 4/5/6 partial states.
- C5 item 5: reboot fresnel; re-run vainfo + ffmpeg. Verify auto-detect repeats.
- C5 item 1:
- If 5 of 5 PASS: write Phase 7 verification doc; proceed to Phase 8 close.
- If any fail: Phase 7 → Phase 4 loopback.
Risk register update
| Risk | Status |
|---|---|
| Syntax/compile error from Edit-based refactor | UNTESTED — fresnel offline |
| Algorithm bug in link traversal | Mechanically verified against Phase 5 reviewer pseudocode |
MEDIA_ENT_F_PROC_VIDEO_DECODER not set in kernel |
Verified at code level (rkvdec.c:1382, hantro_drv.c) |
| Two-pass changes behavior on rkvdec-only systems | Same as single-pass (pass 1 picks rkvdec, pass 2 never fires) |
| Two-pass on hantro-only systems | Pass 1 finds nothing; pass 2 picks hantro-decoder (works) |
| Encoder mis-detection | Eliminated by construction (entity-function check excludes encoder-only paths) |
Substrate state at Phase 6 close
- Fork tip
c106d95(iter7 Phase 6 commit). - Backend installed: still iter5b-β/iter6 SHA
2c6ff82c…. To be updated when fresnel returns. - Kernel
linux-fresnel-fourier 7.0-1. - Test fixtures unchanged.
- Phase 7 ready to execute once fresnel + build cycle complete.