fix: harden Quest VR compatibility

This commit is contained in:
XxJSONDeruloxX committed 2026-10-15 18:55:23 -04:00
1 parent 92f2696a0c
commit 5e9ebbc7df
17 files changed
+377 -55

No files matched your search

+11 -4
View File
@@ -96,8 +96,9 @@ then acknowledges that buffer. Missing acknowledgements quarantine the transfer
instead of allowing a producer to overwrite a possible reader. This still
includes a CPU read and GLES upload; it is not GPU image import into OpenXR.
The helper enables SteamVR's Linux Vulkan asynchronous queue and disables motion
smoothing. It excludes the pinned Steam Frame package's optional CVP backend,
The helper temporarily enables SteamVR's Linux Vulkan asynchronous queue and disables motion
smoothing, backing up those preferences and restoring them on stop or the next start
after an interrupted session. It excludes the pinned Steam Frame package's optional CVP backend,
whose library constructor exits on Quest's unsupported SoC. The original library
is retained with a `.droiddeck-disabled` suffix. SteamVR binaries are not patched.
Cardboard retains its existing direct-mode path; `questframes` remains an explicit
@@ -125,8 +126,14 @@ images per second despite a 90 Hz XR session.
That earlier measurement used the driver-direct presenter. With SteamVR's normal
compositor, a native Vulkan scene passed its first submission and continued
submitting stereo frames. Thumper and The Thrill of the Fight produced clean
immersive scenes. SUPERHOT's Windows OpenXR build still showed shared-image
corruption. A subsequent 60-second shared-buffer dashboard run delivered about
immersive scenes. SUPERHOT's Windows OpenXR build initially showed shared-image
corruption: DXVK supplied a restricted mutable-image format list while SteamVR
created its corresponding opaque shared images without one. Turnip can select
different memory layouts for those descriptions. On the Quest bridge, the layer
normalizes the known opaque-FD creation chains to SteamVR's unrestricted list,
without changing DMA-BUF/modifier images or other targets. A headset capture then
showed clean SUPERHOT scene geometry; interactive gameplay remains unverified.
A subsequent 60-second shared-buffer dashboard run delivered about
43 new stereo images per second. Reusing Android pixel buffers raised the next
60-second run to about 68 received and 65 uploaded images per second, with valid
focused tracking and no rejected frames. These dashboard results do not establish refresh-rate gameplay: game input,