Exploring Improved Mali/Tensor Graphics Driver Paths for Android Emulation #1575
MissingxLink
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi,
I’m reaching out as a Pixel / Tensor G2 user experiencing consistent launch-level instability with Azahar on Android (Mali-G710). I understand that many of these failures likely originate below the emulator itself, at the Vulkan driver and shader compilation layer.
I’m aware that Turnip (Mesa Vulkan) is Adreno-specific and not portable to Mali, and that fully replacing vendor GPU drivers on stock Android is not feasible. That said, Android does allow limited user-space graphics driver path selection (system driver vs ANGLE), which suggests there may be room for improvement at the driver or translation layer rather than only emulator-side workarounds.
With that in mind, I wanted to ask whether any of the following are considered viable or interesting directions, either short- or long-term:
• Improving Mali compatibility via alternative Vulkan usage patterns known to be more robust on ARM drivers
• Exploring alignment with PanVK / open Mali Vulkan behavior (even if not directly deployable on stock Android) to inform safer code paths
• Identifying specific Mali driver failure points (extensions, SPIR-V patterns, pipeline creation assumptions) that could be mitigated by a compatibility layer
• Investigating whether an ANGLE-based or hybrid path could improve baseline stability for Mali devices
I fully understand that a true “open Mali driver on Android” is a long-term and upstream problem, and I’m not expecting a drop-in replacement or immediate parity with Adreno. My interest is in whether a driver-aware approach could reduce hard crashes and improve baseline stability for Tensor users, even if performance remains secondary.
I’m happy to assist by providing detailed logs, testing experimental builds, or supplying Tensor hardware for investigation, subject to maintainer approval and project norms.
If this isn’t the appropriate place to raise this or if another project or maintainer would be better suited to evaluate it, I’d really appreciate being pointed in the right direction.
Thanks for your time and for your work on this ecosystem.
All reactions