The AppImage bundles libva from the build environment (Ubuntu 22.04,
VA-API 1.20). VA-API driver modules (e.g. Mesa's radeonsi_drv_video.so)
are dlopen()ed by libva and export an entrypoint named after the libva
minor version they were built against (e.g. __vaDriverInit_1_23). An
older libva cannot load newer drivers, so on any host more recent than
the build container va_openDriver() fails:
Trying .../radeonsi_drv_video.so
libva error: has no function __vaDriverInit_1_0
va_openDriver() returns -1
and Moonlight silently falls back to software decoding (reproduced on
Fedora 44 / Mesa 26.1 / RX 6500 XT; overriding the bundled libva with
LD_PRELOAD of the host libva immediately yields HW decode). Refreshing
the bundled libva at each release only postpones the breakage: the host
drivers always eventually outrun it (see #1398).
Exclude libva and its backends from linuxdeploy so Moonlight and the
bundled FFmpeg resolve them through the loader's normal search path and
pick up the host libva, which the distro's package manager keeps in step
with the host drivers. This mirrors how Mesa/GL/EGL are already
host-provided.
Hosts without a usable libva (e.g. bare NVIDIA systems) must still not
fail to launch, since libva.so.2 is a direct DT_NEEDED of Moonlight and
the bundled FFmpeg. Stage the build-environment libva in
opt/libva-fallback (outside every loader search path) together with
libva-probe, a tiny program whose only job is to carry the same
DT_NEEDED entries and VA_API_* versioned symbol references as the
binaries shipped in the AppImage. The generated AppRun runs the probe
first: if the dynamic loader can start it against the host, the host
libva is used (probe passing is, by construction, exactly Moonlight
being loadable); if not, it prepends the fallback directory and launches
with the bundled libva, matching pre-existing behavior. The build script
verifies that the probe references every version node required by the
shipped binaries, so the two cannot silently drift apart.
We should be using the EGL renderer's colorspace and color range
rather than the backend renderer's for fallbacks. The former is
what actually gets passed to the host PC during stream init.
Huge thanks to @oliverlj for identifying this longstanding bug.
This is likely responsible for many different issues with the DRM renderer randomly failing.
MoltenVK does a lazy fetch of the drawable only when it's being modified, so waitToRender() is completely ineffective as is. Rendering a small empty overlay in waitToRender() forces the next drawable wait to happen in waitToRender(), before we pick the current frame to render. This removes one display refresh period of latency when using V-sync.
Direct-to-display mode on some Macs (M4 Mac Mini) can retain our drawables, which cuts the frame rate in half if we're in double-buffered mode.
We don't want to use triple-buffered mode all the time because it increases latency when used if not required, so detect delayed presentation and enable it on the fly if needed.
Cameron Gutman
luanweslley77
Daniel Dunderfelt
dependabot[bot]
MiraandCameron Gutman