hi,
Am Freitag, den 14.08.2015, 16:48 +0800 schrieb Daniel van Vugt:
Hi all,
In testing performance optimisations on various phones, I keep running
into an annoying hurdle.
Although you can optimise your Mir server/clients in such a way that
they're smoother more often, there's an additional variable outside of
Mir and Unity that gets in the way. That seems to be frequency scaling
done by the kernel. Sometimes on desktops too, but I'm mostly concerned
about phones here.
I find it suspicious that on some devices you can turn stuttering into
smoothness just but touching the screen a lot. But the smoothness soon
goes away when you're not touching the screen.
usually android phones use the interactive or the userspace cpufreq
governor (the latter with a userspace daemon) with frequency scaling
directly tied to input/user events. we leave setting up this stuff to
the container under the assumption that the android setup already does
the best with the HW.
i bet you will see the frequency go up and down along when checking the
above again while watching it (which, along with the fact that someone
should probably look into optimizing it for our use-cases, probably also
points out that we are not doing enough of our processing in the GPU).
ciao
oli
