Conversation
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
jserv
force-pushed
the
ci-refine
branch
3 times, most recently
from
d4bf16c to
2a230e8
Compare
.github/workflows/differential.yml: new workflow with five parallel
jobs (motif / violawww / mosaic / osiris / xfig) that each build the
downstream target twice on the same ubuntu-24.04 runner - once against
the apt-installed system libX11 stack and once against libx11-compat -
then screenshot-compare under Xvfb. Replaces the SSH-to-node11 path
the differential blocks in ci.yml used to gate on, so the differential
diff is reproducible on stock GH infrastructure without an external
host. Each job pulls its target-specific apt set on top of the shared
COMMON_BUILD_PKGS, SYSTEM_X11_DEV_PKGS, MOTIF_BUILD_PKGS,
OSIRIS_BUILD_PKGS, and DIFFERENTIAL_PKGS lists, shares the
upstream-src cache key with ci.yml so a warm CI run primes this
workflow, adds a 12 minute hard timeout, and the motif and osiris
jobs additionally cache the system-side build tree across runs keyed
on the configure inputs. SYSTEM_X11_DEV_PKGS includes libxpm-dev
(Motif configure requires xpm.pc via pkg-config) and xbitmaps (Motif
lib/Xm/I18List.c includes X11/bitmaps/gray directly).
DIFFERENTIAL_PKGS includes xfonts-base / 100dpi / 75dpi / scalable
because Motif demos and ViolaWWW request bitmap font families via
XLoadQueryFont and without these the system-side fall-through to 9x15
either crashes ViolaWWW's bundled xloadimage or blows up the diff
threshold against libx11-compat's SDL_ttf path.
.github/workflows/ci.yml: drops the steps.*-differential blocks and
their matching artifact uploads. The remaining ci.yml jobs (build,
sanitize, lint, debug-build, plus per-target compat-side smoke runs)
stay unchanged so the fast PR feedback loop is not delayed by the
heavier double-build pipeline.
scripts/run-{motif,mosaic,violawww,osiris,xfig}-differential-tests.py:
adds --local mode that short-circuits the rsync + ssh round trips and
runs the build / capture / compare pipeline on the current host via
"sh -s". The compat-side and system-side builds run in parallel
background subshells with explicit "status=0; wait \$pid || status=\$?"
propagation so a "set -eu" payload surfaces background failures
instead of swallowing diagnostics; the upstream source is pre-extracted
serially via a make stamp target so the two parallel sides do not
race on autoreconf / tarball extraction. Screenshot capture is split
the same way: two Xvfb instances on display N and N+1 let system-side
and compat-side captures run concurrently rather than serially, and
each capture_* shell helper derives display_num from the subshell's
\$DISPLAY env so the two sides target their own X server even though
the helper itself is shared. xfig's system-side configure does an
explicit cd into \$system_build/source first because autotools
generates Makefiles in the current working directory, not at the
configure script's path. ccache wrapping goes through
PATH=/usr/lib/ccache:\$PATH and CC=gcc so recursive Makefiles that
pass \$(CC) unquoted to sub-make (Mosaic's makefiles/Makefile.linux
is one) do not tokenize the value into "CC=ccache" + "gcc-target".
Osiris caps ninja parallelism via OSIRIS_NINJA="ninja -j\$JOBS" on
the compat side and "ninja -j\$JOBS -C" on the system side so the
default vCPU+2 heuristic does not launch 12 compiler processes across
both sides on a 4-vCPU runner.
The post-wait diagnostic blocks surface both build sides' failure
messages and tail both sides' logs when either side is non-zero, then
exit with the first non-zero status. Same pattern for the post-capture
diagnostic block. After the wait, Xvfb logs and any partial replay-*
trees are staged into \$remote_root/logs so the artifact upload picks
them up regardless of capture success. A CLI > env > --local-default
> SSH-default precedence chain resolves --compare-location and
--remote-root so XXX_DIFF_LOCAL=1 alone (no env override) switches
the comparison location to local without surprising the SSH-mode
developer flow.
The xfig differential captures the startup screen only. The
xfig-draw-line replay clicks (32, 96) targeting the line tool icon,
but xfig 3.2.9a's drawing toolbox now has a "Drawing" section label
that pushes the tool grid down so (32, 96) hits empty label space
and selects no tool. The compat-side internal input backend masks the
miss because libx11-compat's event-injection path handles the click
via widget translation, but xdotool on the system side sends a literal
X click at the wrong coord and no drawing happens. The smoke job
(mk/xfig.mk:check-smoke-xfig) keeps the compat-side draw-line
coverage; the differential compares the startup screen, which is the
meaningful library-parity check this workflow is supposed to gate.
mk/{motif,mosaic,violawww,osiris,xfig}.mk: adds XXX_DIFF_LOCAL knob
(default 0) that threads --local through to the script, and changes
XXX_DIFF_COMPARE_LOCATION default from unconditional "remote" to
"local" when XXX_DIFF_LOCAL=1 so the make wrapper does not silently
override the script's local-mode resolution. mk/motif.mk additionally
sets MOTIF_DIFF_MAE_THRESHOLD to 0.15 and MOTIF_DIFF_CHANGED_THRESHOLD
to 0.32 to admit the GH runner baseline; the worst observed deltas
(ColorSel/colordemo at MAE 0.133, changed 0.293) are real
widget-geometry parity gaps between system Motif Xm and
libx11-compat-built Motif Xm (ColorSel left-pane width, Ext18List row
metrics, TabStack header padding, workspace/wsm geometry) that need
follow-up work inside libx11-compat / our Motif Xm patches rather
than CI-side regressions.
tests/ui/replays/violawww-scroll-system.replay: wait-window budget
bumped from 3 seconds to 15 seconds with a 2 second post-wait-window
settle delay before the first interaction, and post-wheel delay from
1 second to 3 seconds. The smoke-job budgets were calibrated for a
single dedicated Xvfb; the differential workflow drives two Xvfb
instances and two captures concurrently in the same job, so vw
startup contends for CPU and disk against the system-side Motif and
vw builds, the window name can match before vw finishes parsing the
fixture HTML (so a wheel-down arriving immediately afterward gets
queued against a non-scrollable layout), and libx11-compat's
internal-backend event loop needs more wallclock to drain the
wheel-down sequence before the post-wheel screenshot is taken.
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.
.github/workflows/differential.yml: new workflow with five parallel jobs (motif / violawww / mosaic / osiris / xfig) that each build the downstream target twice on the same ubuntu-24.04 runner - once against the apt-installed system libX11 stack and once against libx11-compat - then screenshot-compare under Xvfb. Replaces the SSH-to-node11 path the differential blocks in ci.yml used to gate on, so the differential diff is reproducible on stock GH infrastructure without an external host. Each job pulls its target-specific apt set on top of the shared COMMON_BUILD_PKGS / SYSTEM_X11_DEV_PKGS / DIFFERENTIAL_PKGS lists, shares the upstream-src cache key with ci.yml so a warm CI run primes this workflow (and vice versa), and adds a 12-minute hard timeout so a stuck Xvfb cannot stall the queue. Motif and osiris additionally cache the system-side build tree across runs keyed on the configure inputs, which lets a cold PR restore the post-configure object set from the most recent main run instead of rerunning autoreconf and meson cold.
Summary by cubic
Run screenshot differential tests on GitHub-hosted runners in a new parallel workflow for Motif, ViolaWWW, Mosaic, Osiris, and Xfig. Drops the SSH host, runs double-builds and captures in parallel, and stabilizes deps, logs, thresholds, and timings for reproducible diffs.
New Features
.github/workflows/differential.ymlwith 5 parallel jobs onubuntu-24.04; each double-builds (system vslibx11-compat) and compares under Xvfb with a 12-minute timeout. Runs as separate checks;ci.ymldrops inline differential gates. Shares caches withci.yml; Motif/Osiris also cache the system-side build.--local/*_DIFF_LOCAL=1) with safe-path checks and local compare by default. Builds and captures run concurrently on two Xvfb displays; upstream sources are pre-extracted; failures surface with clear log tailing;ccacheis wrapped via PATH withCC=gcc. Osiris capsninjaparallelism for runner stability.XXX_DIFF_LOCALknobs that propagate to scripts and toggle the default compare location.Bug Fixes
libxpm-dev,xbitmaps, and legacy fonts (xfonts-base,xfonts-100dpi,xfonts-75dpi,xfonts-scalable) to avoid configure crashes and font-driven diffs.$DISPLAY(separate Xvfb), always stage both sides’replay-*andxvfb-*.log, and guard local runs against unsafe--remote-root.MAE=0.15,changed=0.32; ViolaWWW wait-window 15s + 2s settle + 3s post-wheel; Mosaic wait-window 15s.Written for commit 10d4301. Summary will update on new commits.