Repository navigation
CI: the VideoOut display tests never run on Linux or Windows; under Xvfb they run in 10 s and caught the EOP flip regression #6899
Description
What
17 VideoOut tests skip on the Linux CI job, and Windows CI skips them as well: video_out_flip_{controls,eop,eop_aftergpu,eop_main,labels,labels_main,lifetime,overlay,reopen,unopened}, video_out_frame_dump_queue, video_out_hdr_present and guest_video_out_{hdr,high_frame_rate,output_mode,resolution_status,wqhd_detection}. They use SDL_VIDEODRIVER=offscreen (the ANYPS5_VIDEO_OUT_TEST_DRIVER default from #3810), and SDL2's offscreen driver has no Vulkan surface, so each one reports no display or Vulkan device. tools/check_skips.py accepts that reason, so the job stays green.
This hides real regressions. #1202 (pipelined draws on by default) broke presentation of EOP flips from #5539: video_out_flip_eop and _eop_main fail and _eop_aftergpu times out. CI could not see it; it was found by running the tests by hand and bisecting. The fix is #6616.
Running them under Xvfb
On Ubuntu 24.04 with lavapipe (Mesa 25.2.8):
Xvfb :99 -screen 0 1920x1080x24 -nolisten tcp &
cmake -S . -B build -DANYPS5_VIDEO_OUT_TEST_DRIVER=x11
DISPLAY=:99 ANYPS5_GPU=llvmpipe ctest --test-dir build -R video_out
On main plus #6616, 16 of the 17 run and pass, in 9.9 s for all 21 video_out tests. guest_video_out_high_frame_rate still skips, because Xvfb offers no 120 Hz mode. check_skips.py already accepts that reason. On main without #6616, the three EOP tests fail as described above.
A hardware Vulkan driver cannot present on Xvfb, which has no DRI3: Intel ANV fails with No DRI3 support detected. So this works with lavapipe, which is what the Linux job already uses.
Suggestion
On the self-hosted Linux runner: install xvfb, configure with -DANYPS5_VIDEO_OUT_TEST_DRIVER=x11, and run the ctest step under xvfb-run -a (or with a started Xvfb and DISPLAY set). Once that is in place, check_skips.py could stop accepting no display on the Linux job, so a lost display shows up as a failure instead of a skip. The runner is yours to configure, so I have not opened a workflow PR. I'm happy to send one if this direction is wanted.
AI-assisted: yes.
Activity
MarcZX10 commentedon Oct 16, 2026
I’m prototyping the Linux CI change from your suggestion: configure
ANYPS5_VIDEO_OUT_TEST_DRIVER=x11and run CTest underxvfb-run, withANYPS5_GPU=llvmpipe. The repo’s build instructions cover X11/XRandR development packages but not the self-hosted runner’s Xvfb package. Isxvfb-runalready installed on the[self-hosted, linux, x64, anyps5]runner, or should the workflow install/provision it? I’m keeping this separate from the EOP production fix in #6616.MarcZX10 commentedon Oct 16, 2026
I opened #6932 with the Linux X11/Xvfb CI change. The workflow checks for both
xvfb-runandxauth, and installsxvfb/xauthwith apt if either is missing. The focused VideoOut target builds locally, but runtime assertions on my host stop at the emulator CPU feature guard; I have not claimed a local CTest pass. The PR checks are still running, with the Windows relinker CTest step pending.MarcZX10 commentedon Oct 16, 2026
Update: #6932 has now passed its PR checks, including the Linux, macOS, and Windows relinker test jobs. The full Linux/Windows build suites are skipped on pull requests by the workflow conditions; the Xvfb VideoOut configuration will be exercised by the full Linux build on
mainafter merge. My local VideoOut test attempt remains blocked before assertions by this host’s CPU feature guard.MayankManas commentedon Oct 20, 2026
Windows results with a real display, for comparison with the Xvfb run above.
ANYPS5_VIDEO_OUT_TEST_DRIVER=windows,ctest -R video_out -VwithANYPS5_GPUper device,ANYPS5_REQUIRE_VULKAN=1,ANYPS5_NO_SHADER_CACHE=1,VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation. Windows 11 Pro 26200, WinLibs GCC 15.2.0 r7, Release,maine9cc4a8, display in HDR.With
offscreen, 13 of the 14video_out_flip_*tests skip (decodeneeds no window). Withwindows:video_out_hdr_probevideo_out_hdr_probe,video_out_flip_eop_aftergpuAll 14 flip tests,
guest_video_out_*andvideo_out_hdr_presentpass on the NVIDIA GPU. The skip isguest_video_out_wqhd_detection_no_vulkan. NoVUID-and nowas not found but was requested by env var VK_INSTANCE_LAYERSin either log.Both failures are in the tests, not in VideoOut, and CI cannot reach either of them:
video_out_hdr_probeexpects no Windows HDR monitor report and fails when the display is in HDR.video_out_flip_eop_aftergpu's 100M-iteration busy loop runs past the Windows GPU timeout on the UHD 770, and the test aborts withVK_ERROR_DEVICE_LOST.AI-assisted: yes
LeandroLP23 commentedon Oct 26, 2026
AMD and a second NVIDIA generation for the Windows half, following MayankManas's run above. Same setup:
-DANYPS5_VIDEO_OUT_TEST_DRIVER=windows,ctest -R video_out -VwithANYPS5_REQUIRE_VULKAN=1,ANYPS5_NO_SHADER_CACHE=1andVK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation(Vulkan SDK 1.4.363.0), WinLibs GCC 15.2.0 r7, Release,main035ac96edd, Windows 11 10.0.26300, displays in SDR. Both runs drew on the machine's real display in the logged-in session (on the desktop through an interactive-logon scheduled task, since an SSH session has no display).guest_video_out_wqhd_detection_no_vulkanVUID-lines and nowas not found but was requested by env var VK_INSTANCE_LAYERSin either log.video_out_flip_eop_aftergpupasses on the 860M (1.6 s), so the GPU timeout VideoOut tests: video_out_flip_eop_aftergpu trips the Windows GPU timeout on an Intel UHD 770 #8349 describes on the UHD 770 doesn't occur on this iGPU.video_out_hdr_probepasses on both, as expected with the displays in SDR (VideoOut tests: video_out_hdr_probe fails on Windows when the display is in HDR #8348 covers the HDR case).guest_video_out_high_frame_rateruns on both high-refresh panels:60 vblanks: default=0.998 s high=0.501 s restored=1.000 son the 860M, and0.997 / 0.500 / 1.001 son the 5060 Ti.With the Linux half done in #6932, what remains here is Windows CI, which has no display; on real Windows displays the suite now passes on AMD, NVIDIA (two generations) and Intel apart from the two test issues already filed.
AI-assisted: yes (Claude Code)
Abhi2april commentedon Nov 5, 2026
#6932 (merged) closed the Linux half: the Linux job runs ctest -R video_out under xvfb-run with -DANYPS5_VIDEO_OUT_TEST_DRIVER=x11 and ANYPS5_GPU=llvmpipe, and it passes 16 of 17 in 9.9 s. The Windows job still configures with offscreen (the default from #3810), so 13 of the 14 video_out_flip_* tests skip with no display or Vulkan device — only video_out_flip_decode runs (it needs no window) — and tools/check_skips.py accepts that reason, which keeps the job green. The guest_video_out_* tests skip the same way. That is the state the two manual runs by @MayankManas (Windows 11 26200, WinLibs GCC 15.2.0 r7, header main e9cc4a8, HDR display) and @LeandroLP23 (Windows 11 10.0.26300, Vulkan SDK 1.4.363.0, main 035ac96, SDR) already measured on AMD and two NVIDIA generations.
for the Windows job:
What I could not check from my side (I am a first-time contributor without runner access): whether the self-hosted Windows runner keeps a logged-in session with a live display, and whether the ctest step belongs in the Windows full-build job in build.yml or somewhere else. If the direction is fine, I open the PR and a collaborator confirms it with gh workflow run build.yml --repo boykopovar/AnyPS5 --ref main --field full=false-style run as in CONTRIBUTING.
On the PR checklist I would put Depends on: #6932 only if the shared check_skips.py changes conflict with their merged surface, Fixes: #6899
AI-assisted: yes.
LeandroLP23 commentedon Nov 5, 2026
One thing to check before the Windows step: the two manual runs you cite, mine and MayankManas's, used machines with a real GPU and Vulkan driver.
windows-latesthas neither.tools/check_skips.py --job windowsalready says so ("the runner has no Vulkan driver"), and the Windows ctest step inbuild.ymldoesn't setANYPS5_REQUIRE_VULKAN.On that runner,
ANYPS5_VIDEO_OUT_TEST_DRIVER=windowswould most likely move the flip tests from "no display" to "no usable Vulkan device", and they would still skip. AddingANYPS5_REQUIRE_VULKAN=1to that step would turn the device skips into failures: thevideo_outones, theagc*tests the job skips today, and any other test that reports no Vulkan device.So the Windows half also needs a Vulkan device on the runner. The counterpart of the Linux job's llvmpipe would be Mesa's lavapipe built for Windows, for example from mesa-dist-win, registered through
VK_DRIVER_FILES. I haven't tried that on a hosted runner, so I can't say whether thevideo_outtests pass on it or whether the runner's session gives them a display. One trial run of the job with that change would answer both.AI-assisted: yes (Claude Code)