You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
appimage: Investigate and ship Intel QSV, validated once on real hardware and pinned by a recorded fixture #1863
Problem. The AppImage leaves out obs-qsv11 (build-aux/appimage/README.md: "QSV probes the Intel GPU through the host's libvpl runtime at startup, which no CI runner can exercise"). build.py --qsv exists but defaults to false, and no QSV AppImage has been run on Intel hardware. Users importing a profile from upstream OBS or the Flatpak hit obs_qsv11_v2 as a missing encoder (#1720). We are a fork with a hybrid Rust libobs, so we can't assume upstream's QSV path works here unchanged.
Investigate.
Build the AppImage with --qsv true and list what obs-qsv11 loads: libvpl from the host or bundled, the oneVPL GPU runtime / Media SDK runtime, libva/libdrm, through hostlib like VA-API.
Check what the startup probe (obs-qsv-test or equivalent) does in an AppImage: helper path, APPDIR, LD_LIBRARY_PATH, the hostlib shim.
Check the encoder path against the Rust libobs (encoder registration, obs_encoder_* texture/frame paths QSV uses). List any differences from upstream behavior.
Validate once, then freeze it. We have no Intel machine in CI. The reporter in #1720 has an Intel HD Graphics 530 (Skylake, i915) with a working iHD driver and offered to help. Plan:
Ship them one test AppImage built with QSV and a capture mode. With an env var, the probe and encoder init record the host-facing calls they make (libvpl/libva queries and their results, device caps) to a JSON fixture, and OBS writes its log.
They run a fixed script: start, record 10 s with QSV H.264, stream-test, attach the fixture and the log.
Check the fixture into the repo. A replay shim answers those same calls from the fixture, so the probe, registration and the Missing Encoders logic run in the existing test binaries with no GPU and in milliseconds. If a change alters what we ask the runtime, the replay fails loudly and we know to revalidate on hardware. Otherwise the result stays sealed.
Gate it in test-appimage.sh: with the replay shim, the QSV encoder is registered from a saved profile that uses obs_qsv11_v2 (per the packaging rule in AGENTS.md).
Acceptance.
Written findings for items 1-3.
A QSV AppImage confirmed on the HD 530 (log and fixture attached here).
A recorded fixture plus replay tests in existing binaries, with no new CI lane and no measurable CI time added.
Claiming: #1863. Findings first, then one PR for the probe replay; enabling QSV in the default AppImage waits for the hardware run.
Findings (items 1-3 of the issue)
What obs-qsv11 loads in an AppImage.build.py --qsv true builds it against libvpl-dev, and libvpl2 (the VPL dispatcher) is bundled. libva and libdrm are excluded from the AppDir (assemble-appdir.sh EXCLUDE), so they come from the host, like VA-API. The dispatcher then dlopens the GPU runtime from the host:
libmfx-gen.so.1.2 (oneVPL GPU runtime, package libmfx-gen1.2 / intel-media-runtime): Tiger Lake (11th gen) and newer only.
libmfxhw64.so.1 (legacy Media SDK, package libmfx1): Broadwell to Ice Lake. The AnduinOS - Appimage build Missing Encoders linux ffmpeg_vaapi_tex, obs_qsv11_v2 #1720 reporter's HD 530 (Skylake) needs this one, plus the iHD driver they already installed.
Neither runtime is ever bundled, so the AppImage's QSV depends on the host's runtime, just as VA-API depends on the host's driver.
Startup probe. On Linux there is no obs-qsv-test helper (that's Windows only). obs_module_load() → check_adapters() → gs_enum_adapters() (implemented in this fork's EGL backends, gl-x11-egl.c.h / gl-wayland-egl.c.h) → for each render node vaInitialize, vaQueryVendorString, and the HEVC/AV1 VAEntrypointEncSlice[LP] rate-control attributes. A node with the legacy i965 driver is rejected. The encoders register only for what an Intel node reports. Nothing in the probe depends on APPDIR or a helper path.
Against the Rust libobs. The probe uses obs_enter_graphics, gs_get_adapter_count and gs_enum_adapters. Encoding uses the texture-encoder path (encode_texture2, NV12/P010 plane textures, qsv_host_texture_valid). Adapter enumeration and the texture calls go through libobs's C graphics layer (libobs/graphics/graphics.c.h → libobs-opengl), still upstream C here. What the Rust libobs does implement on this path is shared by every encoder: packet create/release and interleaving (obs_encoder_packet_*, obs_encoder_interleave_info), which x264 and VA-API already exercise. So there's no QSV-specific Rust difference to fix before a hardware run, and the hardware run checks the texture path end to end.
Plan
PR 1 (this claim): split the probe's decision out of the VA-API calls into plugins/obs-qsv11/qsv-probe.c.h (VA-API, VPL and libobs free). The plugin logs each node as a QSV probe: line, and test_qsv_probe replays log lines from real hardware through the same decision. The fixture starts as a provisional HD 530 line and is replaced verbatim by the reporter's log. It's one small cmocka executable that runs in milliseconds, with no lane and no GPU.
Then a qsv=true AppImage build (workflow_dispatch) for the reporter's one-time run.
Then, with their log: replace the fixture, flip qsv on for the AppImage, extend test-appimage.sh (QSV module loads and registers nothing on a non-Intel runner, a saved profile with obs_qsv11_v2 gives the dialog reason without crashing), and update the README "Left out" list.
Step 1 merged: #1879 (8fd9ae3). obs-qsv11 logs one QSV probe: line per render device, and test_qsv_probe replays them through the same decision. The fixture is still provisional. Step 2 is running: a QSV-enabled AppImage from master (appimage.yaml, qsv=true), run https://github.com/obs-rust/obs-studio/actions/runs/39899437010, for the one-time hardware run on #1720.
Step 2 done: the qsv build works again. #1888 (queued) installs libvpl 2.17.0 in the manylinux image (recipe obs-rust/obs-studio-libs#10, release libvpl-2.17.0, assets c74980e). Run 39901601164 built a QSV AppImage with obs-qsv11.so and a bundled libvpl.so.2. The test instructions are on #1720 (#1720 (comment)). The HD 530 needs the host libmfx1 (Media SDK) runtime, since libmfx-gen starts at Tiger Lake. Step 3 (replace the fixture with the real log, flip qsv on for the AppImage, extend test-appimage.sh) waits on that report; the artifact expires 2027-01-21.
A note for this claim, from #891's NixOS measurements: AppRun sets no ONEVPL_SEARCH_PATH. On NixOS the QSV runtime (libmfx-gen.so.1.2 / libmfxhw64.so.1 from hardware.graphics.extraPackages) sits in /run/opengl-driver/lib, which the bundled dispatcher won't search unless the app runs through appimage-run. Setting it when unset (like LIBVA_DRIVERS_PATH) may belong in this work; I'm leaving it to you rather than touching AppRun.
Triage: claimed and in progress (#1879 adapter-probe logging and replay, #1888 libvpl in the manylinux image; #1963 open). The issue's own acceptance: one validated run on real Intel hardware, pinned by a recorded fixture, plus an AppImage test that loads a saved profile using obs_qsv11_v2. Add the ONEVPL_SEARCH_PATH note from #891 for NixOS. P2: profiles imported from upstream or the Flatpak lose their encoder (#1720). size:M.
Activity
zackees commentedon Oct 23, 2026
Claiming: #1863. Findings first, then one PR for the probe replay; enabling QSV in the default AppImage waits for the hardware run.
Findings (items 1-3 of the issue)
build.py --qsv truebuilds it againstlibvpl-dev, andlibvpl2(the VPL dispatcher) is bundled. libva and libdrm are excluded from the AppDir (assemble-appdir.shEXCLUDE), so they come from the host, like VA-API. The dispatcher then dlopens the GPU runtime from the host:libmfx-gen.so.1.2(oneVPL GPU runtime, packagelibmfx-gen1.2/intel-media-runtime): Tiger Lake (11th gen) and newer only.libmfxhw64.so.1(legacy Media SDK, packagelibmfx1): Broadwell to Ice Lake. The AnduinOS - Appimage build Missing Encoders linux ffmpeg_vaapi_tex, obs_qsv11_v2 #1720 reporter's HD 530 (Skylake) needs this one, plus the iHD driver they already installed.Neither runtime is ever bundled, so the AppImage's QSV depends on the host's runtime, just as VA-API depends on the host's driver.
obs-qsv-testhelper (that's Windows only).obs_module_load()→check_adapters()→gs_enum_adapters()(implemented in this fork's EGL backends,gl-x11-egl.c.h/gl-wayland-egl.c.h) → for each render nodevaInitialize,vaQueryVendorString, and the HEVC/AV1VAEntrypointEncSlice[LP]rate-control attributes. A node with the legacy i965 driver is rejected. The encoders register only for what an Intel node reports. Nothing in the probe depends onAPPDIRor a helper path.obs_enter_graphics,gs_get_adapter_countandgs_enum_adapters. Encoding uses the texture-encoder path (encode_texture2, NV12/P010 plane textures,qsv_host_texture_valid). Adapter enumeration and the texture calls go through libobs's C graphics layer (libobs/graphics/graphics.c.h→ libobs-opengl), still upstream C here. What the Rust libobs does implement on this path is shared by every encoder: packet create/release and interleaving (obs_encoder_packet_*,obs_encoder_interleave_info), which x264 and VA-API already exercise. So there's no QSV-specific Rust difference to fix before a hardware run, and the hardware run checks the texture path end to end.Plan
plugins/obs-qsv11/qsv-probe.c.h(VA-API, VPL and libobs free). The plugin logs each node as aQSV probe:line, andtest_qsv_probereplays log lines from real hardware through the same decision. The fixture starts as a provisional HD 530 line and is replaced verbatim by the reporter's log. It's one small cmocka executable that runs in milliseconds, with no lane and no GPU.qsv=trueAppImage build (workflow_dispatch) for the reporter's one-time run.qsvon for the AppImage, extendtest-appimage.sh(QSV module loads and registers nothing on a non-Intel runner, a saved profile withobs_qsv11_v2gives the dialog reason without crashing), and update the README "Left out" list.obs-qsv11: GREEN: log the Linux adapter probe and decide from it
obs-qsv11: Log the Linux adapter probe and replay it in a test (obs-r…
zackees commentedon Oct 23, 2026
Step 1 merged: #1879 (8fd9ae3). obs-qsv11 logs one
QSV probe:line per render device, andtest_qsv_probereplays them through the same decision. The fixture is still provisional. Step 2 is running: a QSV-enabled AppImage from master (appimage.yaml,qsv=true), run https://github.com/obs-rust/obs-studio/actions/runs/39899437010, for the one-time hardware run on #1720.libvpl 2.17.0: linux-x86_64 (obs-qsv11, obs-rust/obs-studio#1863)
zackees commentedon Oct 23, 2026
Step 2 done: the
qsvbuild works again. #1888 (queued) installs libvpl 2.17.0 in the manylinux image (recipe obs-rust/obs-studio-libs#10, releaselibvpl-2.17.0, assets c74980e). Run 39901601164 built a QSV AppImage withobs-qsv11.soand a bundledlibvpl.so.2. The test instructions are on #1720 (#1720 (comment)). The HD 530 needs the hostlibmfx1(Media SDK) runtime, sincelibmfx-genstarts at Tiger Lake. Step 3 (replace the fixture with the real log, flipqsvon for the AppImage, extend test-appimage.sh) waits on that report; the artifact expires 2027-01-21.appimage: Install libvpl in the manylinux image so the qsv build conf…
libvpl 2.17.0 recipe (obs-qsv11) (#10)
zackees commentedon Oct 23, 2026
A note for this claim, from #891's NixOS measurements: AppRun sets no
ONEVPL_SEARCH_PATH. On NixOS the QSV runtime (libmfx-gen.so.1.2/libmfxhw64.so.1fromhardware.graphics.extraPackages) sits in/run/opengl-driver/lib, which the bundled dispatcher won't search unless the app runs through appimage-run. Setting it when unset (likeLIBVA_DRIVERS_PATH) may belong in this work; I'm leaving it to you rather than touching AppRun.zackees commentedon Oct 25, 2026
/ready
Triage: claimed and in progress (#1879 adapter-probe logging and replay, #1888 libvpl in the manylinux image; #1963 open). The issue's own acceptance: one validated run on real Intel hardware, pinned by a recorded fixture, plus an AppImage test that loads a saved profile using
obs_qsv11_v2. Add theONEVPL_SEARCH_PATHnote from #891 for NixOS. P2: profiles imported from upstream or the Flatpak lose their encoder (#1720). size:M.