[arch-general] RFC: Catalyst and Hybrid Graphics

Armin K. krejzi at email.com
Tue Feb 5 14:45:34 EST 2013


Hello,

I have been playing with Catalyst on ArchLinux in order to get my Hybrid 
Graphics working. Catalyst supports PowerXpress 4.0+ in Linux, so 
muxless graphics can work.

It took me some time to get it to work, I had to workaround many stuff.

Here follows:

First thing that I was bothered with is that 
/usr/lib/fglrx/switchlib{GL,glx} were missing.

Their purpose is to switch /usr/lib/libGL.so.1 and 
/usr/lib/xorg/modules/extensions/libglx.so to either AMD (Catalyst) or 
Intel (Mesa and Xorg) libraries.

Each one is called with "intel" or "amd" as an argument, and both must 
exit sucessfully with "echo amd" or "echo intel" after getting the job done.

I've managed to "fool" Catalyst by creating dummy files.

#!/bin/bash

echo amd
exit 0

By default, aticonfig --initial will use AMD (Discrete) GPU, so the 
current installation seems to be okay. I'll continue with other issues 
that can arise later in this mail.

Next major problem is that fglrx will load intel driver, which, 
unfortunately can't be installed by default since it depends on 
intel-dri, which depends on libgl which is in conflict with 
catalyst-utils package (doh). I had to rebuild it without intel-dri 
dependency to get this working.

On why intel driver is necesary, I can answer the following:

Most newer hybrid graphics are muxless, they do not have hardware 
multiplexer and you can't select on which GPU you want to output. 
Instead, Discrete GPU is just wired to Integrated one and is meant only 
for Processing while it is on Integrated GPU to display anything passed 
from the Discrete GPU. For more information on this term, you may want 
to look at DRI PRIME and DMA BUF (An open source solution for this 
mess). So in fact, intel driver is needed since it is used for 2D and 
stuff, where fglrx only does the encoding/decoding stuff.

How to fix it?

Well, I couldn't think of any nice way to fix this. Debian has their 
alternatives system so they can have multiple libGL installed and set 
with alternatives. Gentoo does the same with their eselect stuff or 
whatever it's called.

Since aticonfig --px-dgpu tells xorg.conf to use Discrete (AMD) GPU, it 
also runs /usr/lib/fglrx/switchlib{GL,glx} with "amd" as parameter in 
order to "make" /usr/lib/libGL.so.1 and 
/usr/lib/xorg/modules/extensions/libglx.so point to AMD provided 
libraries. aticonfig --px-igpu would switch to Mesa libGL.so.1 and Xorg 
Server libglx.so which are, you know, not present when catalyst-utils 
are installed. (Xserver libglx.so is present, but it's useless without 
Mesa's libGL.so.1).

On how to solve this I can only think of making another package (part of 
Mesa source package) called something like intel-dri-hybrid (which 
depends on catalyst-utils and can only be used with it) that installs 
Intel DRI drivers and libGL.so.1 into some nonstandard place or just 
renamed to something else like libGL-intel.so.1. Also, it would be 
necesary to make /usr/lib/fglrx/switchlib{GL,glx} to correctly switch 
(symlinks should do it, but carefully) between AMD and Open Source libGL 
and libglx.

I haven't done this yet but some may want to use this - not everyone 
wants to use Discrete GPU always - especially when on battery. I can say 
that power consumption on my laptop has dropped even though I always 
disabled Discrete GPU via vgaswitcheroo (radeon driver).

Any comments and suggestions are welcome. I can try to implement 
something if necesary, it's just the question if my suggestions are good 
enough and will be accepted.

I appologize for any gramatical mistakes.


More information about the arch-general mailing list