Repository navigation
feat(ffmpeg-patches): take generic FFmpeg fixes from the shared fix series - #294
Merged
Merged
Conversation
…eries The Pelorus stack carried its own copies of fixes that are not Pelorus features: the Vulkan queue family fix (patch 0021) and the hevc_nvenc user data SEI budget inside patch 0022. vmafx needs the same fixes, and a fix kept in two stacks drifts. They now come from one place, the shared FFmpeg fix series VMAFx/ffmpeg-patches (ADR-0185). The stack applies on top of release v0.1.0-rc.1 of that series, pinned in build-config.env by tag, commit and tarball sha256. One script, scripts/fetch-ffmpeg-series.sh, downloads the release and fails closed unless the sha256 pin, the cosign signature of SHA256SUMS, the build attestation (gh attestation verify) and the FFmpeg base all check out. generate.sh, build-and-run.sh, qsv-roi-regression.sh, the tester image build and the CI stack jobs apply the series first, then series.txt, and check-build-config.py refuses a version of each that does not. Stack changes: - Patch 0021 is dropped: series patch 0001 is the same diff. Its number stays unused, and 0022 and 0023 keep theirs, because documents, ADRs, the tester and v0.4.0-rc.2 cite them. generate.sh formats the range above the series tip and maps its output to the names by position. - Patch 0022 becomes one call between the series' av_memdup() and nvenc_hevc_sei_fits(): it writes a Pelorus blob as its zero-free carrier and, on hevc_nvenc, strips the maps of a blob that does not fit. The budget and the drop are series patches 0003 and 0004. - Patches 0004, 0008 and 0011 change in context only. Behaviour that changes: h264_nvenc and hevc_nvenc no longer write a user data SEI from another producer that NVENC would truncate; they warn instead (series patch 0004). Regenerating or replaying the stack now needs network access, cosign and an authenticated gh. The NOTICE of the released patch archive names the series the stack needs first. Checked: the verifier refuses a flipped byte, a moved pin, a tampered SHA256SUMS and a wrong commit, and its offline checks refuse 13 planted defects in the fast suite; generate.sh is byte-stable; the full replay links FFmpeg and passes the NVENC smoke on an RTX 4090; the tester's sidedata_roundtrip passes on hevc_nvenc and h264_nvenc (research 0185). Refs #129 Signed-off-by: lusoris <lusoris@proton.me>
12 tasks done
lusoris
added a commit
that referenced
this pull request
Oct 11, 2026
Pelorus 0.4.0-rc.3 is the third release candidate of 0.4. The interop ABI stays at 1.5. Since rc.2 it adds `tiling=drm` output pools on the six filters that write new frames, so `hwmap` reaches VAAPI and QSV without a download (#292). The FFmpeg patch stack now applies on top of the shared FFmpeg fix series v0.1.0-rc.1 (ADR-0185, #294); with series patch 0004, NVENC no longer writes a foreign user data SEI that it would truncate. The documented recipes for hardware encoders no longer use `hwdownload` (#295), and the Praetor engine is pinned at 7cf9cbb2 and runs through its own launcher (ADR-0187, #297). The two known issues of rc.2 still hold and stay in the notes: the intermittent h264_qsv crash with SEI payloads inside Intel's QSV runtime, and VMAFx readers on ABI 1.4 ignoring the NVENC carrier until VMAFx re-pins (VMAFx/vmafx#2708). The six changelog fragments are rotated into the [0.4.0-rc.3] section, which the release build reads as the notes. meson.build and PELORUS_VERSION_STR carry 0.4.0-rc.3; the version numbers stay 0, 4, 0. Signed-off-by: lusoris <lusoris@proton.me>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What & why
The Pelorus FFmpeg stack now applies on top of the shared FFmpeg fix series instead of carrying its own copies of generic fixes. The series is VMAFx/ffmpeg-patches
v0.1.0-rc.1: four patches on the samen9.0.2base, which vmafx applies too. ADR-0185 records the decision and the alternatives.build-config.envpins the series by tag, commit and tarball sha256.scripts/fetch-ffmpeg-series.shfails closed unless the sha256 pin, the cosign signature ofSHA256SUMS, the build attestation and the FFmpeg base all check out. Every script, image build and CI job that patches FFmpeg applies the series first.v0.4.0-rc.2cite them.hevc_nvencbudget and the drop are series patches 0003 and 0004. Patches 0004, 0008 and 0011 change in context only.What changes for users:
h264_nvencandhevc_nvencwith-udu_sei 1no longer write a user data SEI from another producer that NVENC would truncate; they warn instead. Regenerating or replaying the stack needs network access,cosignand an authenticatedgh.Reviewer decisions:
Implementation: pendingline points at the epic [EPIC] Upstream and rebases (ongoing) #129, because no issue tracks this switch. Name another issue if you prefer one.Reproducer / smoke-test
SHA256SUMSand wrong commit each refused at their own check; with each check switched off, the self-test failsgenerate.sh, twicebuild-and-run.shsidedata_roundtriphevc_nvencandh264_nvencpass, 5 of 5 cases each; the tester's QSV legs werenot_runon this three-GPU hosthevc_qsv,h264_qsvsmoke (Arc A380)qsv-roi-regression.shbuildstageTester publishjobffmpeg-stack(verifier self-test with the default token, regeneration without diff, replay),lavapipeand the three tester kits passmake verify-all,reuse lintDetails are in research 0185. Raw logs are local, in
.workingdir/evidence/shared-series-switch/.Per-PR checklist (ADR-0100 / ADR-0108 — see CONTRIBUTING.md)
docs/adr/README.mddocs/development/build.md(new section),docs/development/tester-image.md,docs/usage/ffmpeg.md,docs/backends/vulkan.md,docs/licensing.md,README.md,ffmpeg-patches/README.mddocs/research/0185-shared-ffmpeg-fix-series.mdffmpeg-patches/AGENTS.mdinvariants 9 and 10, rootAGENTS.mdhard rule 5changelog.d/changed/0185-shared-ffmpeg-fix-series.md;--checkpassesdocs/rebase-notes.mdmake verify-allgreenmasterReviews
ffmpeg-patch-reviewer: no defect in the patch 0022 hook (ownership on every path, budget, interplay with the series' drop). Its findings are fixed in this commit: asei-fitcase for the hook path the GPU smoke cannot reach, a length guard that does not depend on the width ofsize_t, a patch-count assertion after the tester image's apply loop, and the notes in ADR-0181 and ADR-0183 moved to their headers.doc-reviewer: its findings are fixed: the script's variables and--unpackmode in the build guide, the series tarball in the-sourceimage list, the decoder wording indocs/usage/ffmpeg.md, and a note indocs/rebase-notes.mdthat older entries predate the 22-patch stack.pr-body-checker: deliverables, commit message and checklist pass; this summary was shortened on its finding.