[PLASMA 5] KWin ⩾ 5.26 uses software renderer instead of OpenGL in X11

[Schwarzer Kater]

By coincidence I came across a KWin bug that affects all Plasma versions ⩾ 5.26.
That currently should include e.g. Kubuntu 22.10 with Kubuntu Backports and Plasma 5.27.x in Kubuntu 23.04, KDE neon, openSUSE Leap 15.5, openSUSE Tumbleweed, TUXEDO OS, Fedora 38, Debian 12 and Arch :
https://bugs.kde.org/show_bug.cgi?id=460404

The solution is to add

[QtQuickRendererSettings]
SceneGraphBackend=opengl​

to your $HOME/.config/kdeglobals

Afterwards for me the speed has increased by about 35-50% !
I could test this with Plasma 5.27.x in:
Kubuntu 23.04, KDE neon, TUXEDO OS, openSUSE Tumbleweed KDE, openSUSE Leap 15.5 KDE, Debian 12 KDE and Garuda KDE Dr460nized
so far.

I don't know if this also affects a Wayland session, perhaps somebody would like to test it?

[claydoh]

I think this may be either graphics card or system dependent, maybe? I remember this bug report (or maybe a similar one), but I am not affected by it on any of my systems, not counting the Arm device that needs a semi-custom kernel build.

[Schwarzer Kater]

Did you test it, e.g. like people did in the bug report?

All my tests were in X11 on x86-64.

[claydoh]

How specifically are you determining that you are using software rendering?

And I am going to assume this only effects plasma, Qt

[Schwarzer Kater]

I followed the method in the bug report repeatedly.

[claydoh]

I see the time commands to compare, so the only way to tell if it is using opengl or software is by how long it takes here?


Before bed:

$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.060s
user 0m0.040s
sys 0m0.015s

Config added

$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.054s
user 0m0.039s
sys 0m0.012s

x2

real 0m0.062s
user 0m0.048s
sys 0m0.012s

Edit removed:

real 0m0.060s
user 0m0.038s
sys 0m0.020s​

And it is slower in spots this morning

konsole: Missing value after '-e'.

real 0m0.070s
user 0m0.050s
sys 0m0.018s

And it is a bit variable:
One last time just now:

real 0m0.061s
user 0m0.045s
sys 0m0.014s

Back in wayland, the values are a tiny higher all around

This is using a PPA for a current mesa, cuz gamez but that shouldn't have an effect on this config edit.
Plasma 5.27.5 blah blah current blah
i5-13400F and RX6650

Differences, yes, but nowhere near the doubling that the bug reporter saw;

So I just tested with 5.27.0, the performance regression persists:
"time konsole -e" returns:

With "SceneGraphBackend=opengl" in .config/kdeglobals:
real 0m0,072s
user 0m0,046s
sys 0m0,023s

Without "SceneGraphBackend=opengl" in .config/kdeglobals:
real 0m0,125s
user 0m0,065s
sys 0m0,048s

I haven't tested on my much lower specc'ed laptops

I suspect I am 'blessed' by the Linux gods again, sprites, and mages. The must be keeping some of the demons and orks at bay, until I wander back into Chromebook and Arm lands.

[Schwarzer Kater]

I see the time commands to compare, so the only way to tell if it is using opengl or software is by how long it takes here?
[…]
And it is a bit variable:
[…]
I haven't tested on my much lower specc'ed laptops
[…]

I have only used time like in the bug report so far (and I am too stupid to understand the underlying mechanisms regarding KWin to do it another way… ).

Yes, there is variable - I have always tested it 5-10 times in the different distributions without the addition to $HOME/.config/kdeglobals and then 5-10 times with the addition to get some kind of average.

I suspect on older machines the difference could be a bit higher (perhaps I will be able to test it this weekend on an Intel i3-3225 with openSUSE Leap 15.5 and Plasma 5.27 - will post the results when I did).


In Plasma 5.24.7 it does not matter if you apply the addition - there is no difference at all between with and without the addition - the values are about the same as in Plasma 5.27 with the addition.
So I will always test this in Plasma ⩾ 5.26 in the future - I think the addition to $HOME/.config/kdeglobals is beneficial if it lowers the numbers.


PS:

[…]
Differences, yes, but nowhere near the doubling that the bug reporter saw;
[…]

One man's 50% is the other man's doubling…

On my desktop computer the differences are roughly about 35-45%.
And I just tested it in TUXEDO OS and it was exactly 50% for me there.

[GreyGeek]

Here are my Wayland results (using an Intel IRIS Xe GPU)
Before:

jerry@jerry-HP-Laptop-17-cn1xxx:~$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.042s
user 0m0.034s
sys 0m0.005s

After change:

jerry@jerry-HP-Laptop-17-cn1xxx:~$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.044s
user 0m0.036s
sys 0m0.004s​

However, doing the "time konsole=e" repeatedly produces results that vary as widely or more than those two results:

jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.044s
user 0m0.036s
sys 0m0.004s
jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.044s
user 0m0.033s
sys 0m0.007s
jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.042s
user 0m0.024s
sys 0m0.014s
jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.045s
user 0m0.025s
sys 0m0.018s
jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.043s
user 0m0.030s
sys 0m0.009s
jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.043s
user 0m0.029s
sys 0m0.010s
jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.045s
user 0m0.035s
sys 0m0.009s
jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.042s
user 0m0.035s
sys 0m0.005s
jerry@jerry-HP-Laptop-17-cn1xxx$ time konsole -e
konsole: Missing value after '-e'.

real 0m0.042s
user 0m0.026s
sys 0m0.013s

So on Wayland I'm given the "fix" a rating of zero impact.

[Schwarzer Kater]

Thank you very much!
Due to the bug report I already suspected that it could only concern X11 …

[Schwarzer Kater]

Results with an older CPU (Intel i3-3225 with iGPU, openSUSE Leap 15.5, KDE Plasma 5.27, X11):

average of 10 without additional
[QtQuickRendererSettings]
SceneGraphBackend=opengl
in $HOME/.config/kdeglobals

real    0m0,160s
user    0m0,134s
sys     0m0,022s​

average of 10 with additional
[QtQuickRendererSettings]
SceneGraphBackend=opengl
in $HOME/.config/kdeglobals

real    0m0,108s
user    0m0,092s
sys     0m0,013s​

-> performance gain is about 35%

I suppose the more modern a (dedicated) GPU is (in combination with newer kernel etc.) the more performance could be gained.

[GreyGeek]

I was tempted to switch to X11 and experiement, but at 82 my neurons take much more time to communicate than what your gain indicates, so I'll be staying with Wayland. In fact, I wouldn't switch for a 100% gain.

[Schwarzer Kater]

I don't think that Wayland has this error in KWin because the bug report speaks of GLX only - I was just curious…
Stock Kwin in Wayland should be roughly as fast as Kwin in X11 with the addition to $HOME/.config/kdeglobals I would suppose - perhaps I will test this next week if nobody beats me to it.

PS and a tiny bit OT: I would like to test this with Wayland in KDE neon - do I have to install additional packages to use Wayland there or is a standard KDE neon installation sufficient?

[claydoh]

PS and a tiny bit OT: I would like to test this with Wayland in KDE neon - do I have to install additional packages to use Wayland there or is a standard KDE neon installation sufficient?

Nope, just log in to a Wayland session

[GreyGeek]

Wayland update!
I have moved back to X11.
The main reason was because, while using Dolphin, the right mouse function was extremely temperamental, with the top 3 options being chosen randomly and automatically. Thus, while trying to add the execute permission to an AppImage, a copy was created instead. When I tried to delete the copy a copy of the copy was made. I ended up with 3 copies of the original file. Group selecting the 3 copies I managed to delete them. I again tried to add the execute permission to the original file and instead Dolphin (or Wayland) decided to delete it.

Wayland is much better than it was last year but I'll wait another year before I try it again.