[Mesa-dev] [RFC] panfrost: Add DRM backend
Tomeu Vizoso
tomeu.vizoso at collabora.com
Wed Mar 6 08:10:52 UTC 2019
On 3/5/19 4:24 PM, Alyssa Rosenzweig wrote:
>> One is that with kbase I don't see any noticeable pauses during the very
>> first scene in glmark2.
>
> ...That sounds like a good thing? :)
I was assuming that the horse should spin smoothly, and not pausing every
second or so, as it happens with the DRM driver.
>> Another is that with the DRM driver I don't see the wallpaper in GNOME Shell.
>
> GNOME Shell is very broken under Panfrost; that's not a kernel-space
> regression (but a userspace one).
Yeah, but with kbase instead of the DRM driver, at least the wallpaper
gets rendered :)
> Does glmark's -brefract and -bshadow demo work?
It's complicated :)
With glmark2-drm, the rendering seems to be the same with both kernel
drivers.
With glmark2-es2-wayland on Weston:
- With the DRM driver, -brefract doesn't render at all, I get lots of JS
fault interrupts.
- With the DRM driver, -bshadow renders a first frame only, but
interestingly enough I can see the horse's shadow, which I haven't seen
otherwise in any other combination (not even with kbase and glmark2-drm).
After the first frame I start getting MMU faults.
- With kbase, -brefract renders even better than on KMS. I suspect
something wrong with synchronization that causes unfinished content to be
presented on KMS.
- With kbase, -bshadow renders even better than on KMS, but the shadow
doesn't seem to be whole. I suspect something wrong with synchronization
that causes unfinished content to be presented on KMS, and maybe also on
Wayland (with GPU composition).
Any ideas about what are the underlying problems to be solved?
> I'm also curious about
> performance, though honestly I wouldn't be surprised if we end up with a
> slight advantage there when all is said and done..
Yeah, TBH, I'm running glmark2 with a hack to sync to refresh rate
because otherwise it tries to render at 1500FPS and stumbles upon all the
races in my crappy kernel code :)
Cheers,
Tomeu
More information about the mesa-dev
mailing list