kaspter has quit [Remote host closed the connection]
kaspter has joined #panfrost
megi has quit [Quit: WeeChat 3.1]
megi has joined #panfrost
danboid has joined #panfrost
<danboid>
I recently got an Amlogic S905X3 board with a G3
<danboid>
Mali G31 MP2
<danboid>
I have installed Manjaro on it and enabled glamour but the performance is v.bad - I can't even move windows smoothly (under MATE)
<danboid>
Is it still early days for G31 support in Mesa or is this another issue?
<danboid>
I've got a Allwinnher H6 board recently too, that has a Mali-T720 and I can move windows about smoothly under MATE but that doesn't work with HDMI hot swapping
<danboid>
That mesa release isn't 2 weeeks old yet but maybe upgrading to mesa-git could fix it?
<icecream95>
danboid: 21.0 is quite an old release in terms of Bifrost support, but note that Wayland compositors usually work better than Xorg on Panfrost
<danboid>
Is there any diffeence between GNOME wayland and KDE wayland for 'frost performance?
<danboid>
2 weeks is old? :)
<alyssa>
21.0 branched off in January
<alyssa>
so icecream95 is right
<HdkR>
and 21.1 is already released :)
<cphealy>
danboid: I bumped into poor performance with G31 on a different platform a few days ago when running glmark2. In my case it turned out that the GPU was seriously underclocked. Have you validated that the GPU is clocked at the correct frequency?
<danboid>
cphealy, No, how to check the GPU clock?
<danboid>
cphealy, Should that be in the xorg log? /proc?
<cphealy>
cat /sys/kernel/debug/clk/clk_summary
<danboid>
/sys even
<danboid>
Booting it up now...
<cphealy>
I'm not familiar with S905X3 but perhaps it's using devfreq driver for GPU frequency scaling. If so, you should be able to find a stack of devfreq files/directories in sysfs too. devfreq will show what the frequency is for the device the driver is for.
<alyssa>
cphealy: Wiring in some crude GPU memory accounting to panfrost userspace
<alyssa>
dunno if I'll be happy enough with it to land it but maybe will be of interest
<cphealy>
crude GPU memory accounting would be better than what we have today, correct? ;-)
<alyssa>
crude GPU memory accounting is better than no GPU memory accounting, yes ;)
<cphealy>
sold!
<cphealy>
alyssa: can you share any details of what this memory accounting does?
<alyssa>
right now? I just want to know total size allocated ;P
<danboid>
cphealy, Thats dumped loads of stuff. What t grep for?
<alyssa>
cphealy: and would love to know why opening webgl aquarium in chromium consumes 300MB of GPU mem >_>
<alyssa>
Oh, awful. Memory usage isn't even deterministic in the lower digits because of BO cache behaviour.
<alyssa>
bbrezillon: The real joy of elimiating struct set* is getting back some semblance of determinism :-p
<robclark>
alyssa: something like drm/msm's $debugfs/gem + gem bo debug labels is a useful thing for tracking down who is using how much gpu memory and for what
<alyssa>
robclark: Yeah I'm labeling everything now
<alyssa>
labeled
* alyssa
apologises to Vulkan
<HdkR>
poor vulkan :)
<danboid>
<Vulkan does it best Spok miffed face in return>