Conversation
lusoris
force-pushed
the
ci/retire-i686-lane
branch
from
September 19, 2026 08:47
4da21bd to
d2a20e9
Compare
26 tasks
added 14 commits
September 19, 2026 13:15
…lane (ADR-1258) ADR-0728 removed the i686 lane; the libvmaf/ -> core/ rename merge brought it back the same day and it ran compile-only since. ADR-1258 upholds ADR-0728: the lane and preflight's m32 stage go, and the x86-64-only intrinsics get 32-bit-safe forms as hygiene. ADR-0151 is marked superseded by ADR-0728.
…sources adm_avx2.c and adm_avx512.c called _mm_extract_epi64 on 12 lines each, next to an extract_epi64() fallback that covered only the 256-bit form, and psnr_avx2.c called _mm_cvtsi128_si64. GCC declares both only on x86-64, which is Netflix#1481: no 32-bit x86 build with asm. extract_epi64_128() maps to _mm_extract_epi64 on x86-64 and to a 32-bit fallback elsewhere; the PSNR sum is read with _mm_storel_epi64. x86-64 code is unchanged. The fork stays 64-bit only (ADR-1258); this keeps the sources portable.
…lines The 32-bit-safe extract_epi64_128() helper in adm_avx2.c and adm_avx512.c moves 19 baselined HISS-04 functions down by 10 lines. Their sizes are unchanged and the total stays at 1627 infractions, so only the recorded line numbers change. Recorded with the pinned engine (~/.cache/vmafx-praetor/praetorctl baseline --record, e4b35cb3c7fe) in a fresh clone of the rebased branch, per the ADR-1249 rebase note.
ADR-0691 retired 32-bit x86; the libvmaf/ -> core/ rename merge (384d97d) restored the lane the same day. It ran compile-only with -Denable_asm=false and no tests. Per ADR-1258 the fork stays 64-bit only: the matrix row, its dependency step and its matrix.i686 conditions go, and preflight drops the m32 stage that mirrored the lane. AGENTS.md (recompiled into CLAUDE.md and the other agent context files), the preflight page and the ADM page no longer describe 32-bit as covered. The shellcheck findings actionlint reports in this workflow are fixed on the way: quoted redirections and substitutions, and pkg-config flags read into arrays instead of relying on word splitting.
State rows T-CI-I686-LANE-RESURRECTED-2026-09-18 and T-X86-64-ONLY-INTRINSICS-2026-09-18 (closed), changelog fragments, a rebase note (no i686 row; keep extract_epi64_128() and the _mm_storel_epi64 read) and an x86 AGENTS invariant.
#1384 committed both .md.stub reservation files alongside the finished ADRs. A stub is a working-tree marker that the real ADR replaces (ADR-0535); these two only kept numbers 1223 and 1224 looking claimed twice.
9aa008e (#1564), the commit that removed the i686 lane, implemented ADR-0691. ADR-0728 (bfd4c43, #52) landed after the rename merge had already restored the lane and changed only its own ADR, a changelog fragment and deprecations.md. ADR-1258, its index row, the state row, the changelog fragment and the rebase note now say so, and the preflight page, preflight.sh and the ADM page cite ADR-1258 for the 64-bit-only rule. ADR-0151 is marked superseded by ADR-1258 instead of ADR-0728: ADR-0691's removal was undone the same day, so the lane ADR-0151 added ran until ADR-1258 removed it, and ADR-1258 is the record that is in force.
Four Accepted ADRs described a matrix CI does not run. ADR-0689 (0ab991a) and ADR-0691 (9aa008e) removed lanes that the libvmaf/ -> core/ rename merge 384d97d restored about an hour later. ADR-0710 (4e21173) added build.yml and sanitizers.yml but removed nothing and left the aggregator alone. ADR-0728 (bfd4c43) changed no workflow at all. ADR-1259 lists every lane in libvmaf-build-matrix.yml and build.yml with its required status and owning ADR, records the maintainer's three decisions (64-bit only via ADR-1258; the five lanes without a later ADR stay, not required; Cppcheck and the three Sanitizers jobs stay required) and supersedes all four ADRs. Their status lines and index rows are flipped, and the index rows for ADR-0691 and ADR-0710 now describe what those ADRs decided. It also records that two jobs report the required check name "Windows MSVC+CUDA", so the aggregator counts only whichever started last; that is opened as T-CI-MSVC-CUDA-SHARED-CHECK-NAME-2026-09-18.
ADR-1259 records the build matrix as it runs; these pages and fragments still described ADR-0689, ADR-0691, ADR-0710 and ADR-0728 as carried out. - deprecations.md: the ADR-0728 entry is marked withdrawn and says what became of each configuration it listed (MinGW64, Cppcheck and the three sanitizers are required checks); a 2026-09-18 entry records the i686 lane; tox does not run in the Linux all-backends lane. - changelog.d: the unreleased native-build-sunset, 0691 and 0689 fragments announced removals that never happened and are deleted; ci-slim-down-v2 now says build.yml and sanitizers.yml were added alongside the matrix. CHANGELOG.md is re-rendered. - release.md: the required-check inventory has 40 names, not 34 (including the Standards & Invariant Verification Gate from ADR-1249), and Semgrep is required. - ci-job-names.md: the "Previous Name" column now holds the names #1286 actually replaced (from the diff of f93a003), the eleven non-required matrix lanes and six later required checks are added, and the count is corrected to 40 names across eleven workflows. - ci.md: lists build.yml and sanitizers.yml and points at ADR-1259. - windows-d3d11-import.md: Windows CI is not "MinGW + no-SYCL"; the required Windows MSVC+SYCL lane compiles d3d11_import.cpp. - metal/index.md: the Metal lanes are not required checks. - .github/AGENTS.md: the MoltenVK lane section is replaced by the ADR-1259 invariant; macOS clang and macOS clang+DNN are the continue-on-error rows; current Tidy SYCL job name. - core/src/hip/AGENTS.md: the lane is "Ubuntu HIP" and installs ROCm 10. - rebase-notes.md: do not restore the deleted fragments or drop lanes in a merge resolution.
build.yml's header said it supersedes libvmaf-build-matrix.yml and that the ARM, static and DNN legs were retired; sanitizers.yml's said it replaced the per-PR sanitizer matrix. Neither happened (ADR-1259): both workflows run alongside the ones they were meant to replace, and none of their jobs is a required check. The headers now say so, and build.yml's notes that its Windows row shares the required name Windows MSVC+CUDA. actionlint findings in sanitizers.yml are fixed on the way: the workflow_dispatch trigger now declares the max_total_time input that fuzz-nightly already reads (default 60, as before, mirroring fuzz.yml), and the -max_total_time argument is quoted. No job, trigger or required check changes.
lanes that no longer exist: "Build — Ubuntu ARM clang (CPU)", "Build — Ubuntu HIP", "Build — macOS clang (CPU)", "Build — Windows MSVC + CUDA (build only)", "Build — Ubuntu SYCL". They now use the names the checks report under. Facts in the same sentences are corrected against the workflows: the ARM lane runs on ubuntu-24.04-arm, the Windows CUDA leg installs 13.3.1 (packages *_13.3, CUDA_PATH_V13_3), and oneAPI comes from Intel's apt repository at ONEAPI_VERSION (Linux) and a pinned offline installer (Windows), not from the intel/oneapi-runtime-toolkit action.
origin/master moved past 7cc0cc9 while ADR-1259 was written, so the evidence sentence in ADR-1259 and the T-CI-MSVC-CUDA-SHARED-CHECK-NAME state row name the master commit instead of calling it origin/master.
…ared names libvmaf-build-matrix.yml's required lane and build.yml's Windows row both reported as 'Windows MSVC+CUDA'. The required checks aggregator keeps one run per name, the newest, so either job could mask the other's failure. The build.yml job becomes 'Windows MSVC+CUDA (full)' (the maintainer's choice; the required name and the ruleset stay as they are). check-aggregator-names.sh compared sets of names and could not see the duplicate. It now also fails when more than one job reports a required name, ignoring step names and workflow titles. A fixture test covers the shared-name case and runs in rule-enforcement.yml. ADR-1259 and the CI pages record the rename; T-CI-MSVC-CUDA-SHARED-CHECK-NAME-2026-09-18 is closed.
The union resolution left three copies: the pre-correction wording that credited ADR-0728, and two copies of the corrected one that credits ADR-0691. Keeps the corrected row. Also carries the HISS baseline re-recorded from a clean clone with the pinned engine: the CAMBI SIMD dispatch merged in the meantime and moved the lines the line-keyed file tracks in cambi.c; no finding is new.
lusoris
force-pushed
the
ci/retire-i686-lane
branch
from
September 19, 2026 11:21
35f3820 to
4937dd3
Compare
The rebase left the rendered README carrying both the pre-correction ADR-1258 row and the corrected one, while _index_fragments/ held only the corrected fragment. Regenerated with concat-adr-index.sh and generate-adr-by-tag.sh, which the docs freshness gate compares against.
14 tasks
Contributor
Author
|
Folded into #1497 together with the other two, and closing here. The three kept invalidating each other: each one touches Nothing was dropped. Every source path of this branch is byte-identical in #1497, verified with |
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.
Summary
This PR brings the CI records in line with the matrix that actually runs, and carries out the maintainer's decision to keep the fork 64-bit only.
ADR-1258: the fork stays 64-bit only.
9aa008e70, #1564) removed theUbuntu i686 gcclane. Thelibvmaf/→core/rename squash (384d97d03) restored it an hour later by accident. It then ran compile-only with-Denable_asm=falseand no tests; ADR-1234's preflightm32stage was built on it.matrix.i686conditions, and preflight'sm32stage. This amends ADR-1234's stage list.extract_epi64_128()beside the existingextract_epi64(), and_mm_storel_epi64inpsnr_avx2.c. x86-64 code is unchanged; the fast suite passes 137/137.ADR-1259: the CI build matrix as it runs. It supersedes ADR-0689, ADR-0691, ADR-0710 and ADR-0728.
bfd4c436b, feat!: sunset legacy native build modes (BREAKING, ADR-0728) #52) changed only its own ADR, a changelog fragment anddeprecations.md. Its 14 lane removals and aggregator changes never happened.Records corrected to match.
deprecations.md;release.md(40 required checks);ci-job-names.md(65 wrong "Previous Name" rows);ci.md.build.ymlandsanitizers.ymlheaders;windows-d3d11-import.md;metal/index.md..github/AGENTS.md;core/src/hip/AGENTS.md; the backend pages.Also:
Found, recorded, not fixed here.
Windows MSVC+CUDAis reported by bothlibvmaf-build-matrix.yml(required) andbuild.yml, so the aggregator's newest-run rule can count the wrong job. The state row isT-CI-MSVC-CUDA-SHARED-CHECK-NAME-2026-09-18. Fixing it renames a required check, which is a maintainer decision.no docs needed: CI and decision records only; the docs changes above correct existing pages, and no user-facing option or output changes.
Type
feat— new featurefix— bug fixperf— performance improvementrefactor— no behavior changedocs— documentation onlytest— test-onlybuild/ci— tooling / infraport— cherry-pick from upstream Netflix/vmafsycl/cuda/simd— backend-specificChecklist
make format && make lintis green locally: pre-commit, both praetor audits,compile-context --verify,actionlinton the three touched workflows,check-aggregator-names.sh(40 match),mkdocs build --strict(0 warnings). clang-tidy and cppcheck report 0 findings on the three x86 files.meson test -C build: x86-64 fast suite 137/137.scripts/dev/preflight.shpasses all six stages./cross-backend-diffand the worst ULP is ≤ 2: 0. On x86-64 the ADM change maps back to_mm_extract_epi64through the macro, and the PSNR change stores the same low 64 bits..c/.cpp/.cu/.h/.hpp, it has the appropriate license header: no new files.!orBREAKING CHANGE:: not breaking; 32-bit x86 was already unsupported per ADR-0691.docs/adr/_index_fragments/<NNNN-slug>.mdand the slug is appended todocs/adr/_index_fragments/_order.txt: ADR-1258 and ADR-1259.Bug-status hygiene (ADR-0165)
docs/state.md:T-CI-I686-LANE-RESURRECTED-2026-09-18andT-X86-64-ONLY-INTRINSICS-2026-09-18closed;T-CI-MSVC-CUDA-SHARED-CHECK-NAME-2026-09-18opened.Netflix golden-data gate (ADR-0024)
assertAlmostEqual(...)score in the Netflix golden Python tests.Deep-dive deliverables (ADR-0108)
## Alternatives consideredindocs/adr/1258-keep-64-bit-only-retire-i686-lane.mdanddocs/adr/1259-ci-build-matrix-as-it-runs.md.core/src/feature/x86/AGENTS.md(no x86-64-only intrinsics),.github/AGENTS.md,core/src/hip/AGENTS.md.changelog.d/removed/ci-i686-lane-retired-again.md,changelog.d/fixed/x86-64-only-intrinsics.md.docs/rebase-notes.md: no i686 row (a merge that restores one is the bug that happened); keepextract_epi64_128()and the_mm_storel_epi64read.Reproducer
Known follow-ups
T-CI-MSVC-CUDA-SHARED-CHECK-NAME-2026-09-18: rename one of the twoWindows MSVC+CUDAjobs (a required check name).Ubuntu gcc,Ubuntu clang,macOS clang,Ubuntu ARM clang,Ubuntu CUDA) are recorded as running; the maintainer's keep-decision covered the advisory lanes it listed, and these five want an explicit confirmation.