[HOWTO] Using recent muxless Intel/ATI hybrid graphics

Message

hirakendu · Post by **hirakendu** » Sat Jan 21, 2012 10:00 am

This howto attempts to provide some information about running OpenGL (and OpenCL) applications on the ATI discrete graphics card on many of the recent notebooks that come with a Intel/ATI hybrid graphics system. It is primarily aimed at the recent notebooks with the PowerXpress 4.0+ model, which are "muxless", i.e., don't have a hardware switch to toggle between integrated and dedicated graphics -- there is no way to turn off the integrated graphics completely (in the bios or using a physical switch) and use it as a simple single graphics card system with just the discrete graphics card. For a more or less current overview of the state of hybrid graphics in general, see [1].

Since I could not find much information on the web about using the discrete graphics on these systems, I decided to try out a few things on my machine and this howto is a summary of the results. One can use either the proprietary ATI drivers (fglrx) or the opensource radeon drivers to get graphics acceleration. However, I prefer using fglrx, since it offers more features, including nice OpenGL performance (which is expected out of these upper mid-range GPUs) and support for running OpenCL applications. Also, it is neatly done - we get to use the Intel KMS (Kernel Mode Setting) with the Radeon doing the rendering! A word of caution is that I still prefer using the Intel graphics usually since the machine runs hot with fglrx (or radeon) and I feel a slight lag while using fglrx (possibly due to the graphics from radeon being routed through the intel graphics).

To use the opensource radeon drivers, currently the only working way seems to be along the lines of the Bumblebee and Ironhide projects (see [2] and [3]) for Intel/Nvidia hybrid graphics (also muxless, known as Optimus and is anologous to powerexpress) that start a second X server on the discrete graphics card and then use VirtualGL (see [4]) for running (rendering?) OpenGL applications on the second X server while outputting images to the primary X server (running on integrated graphics). (PS: This is possibly an inaccurate lay man summary of things, as is most of this howto

.) This method also has its advantages - it is easy to switch off the discrete card using "vga_switcheroo" (described later) and turn it on when required. However, the graphics performance seems to be (only) on par with the Intel graphics - I think possibly due to the current state of mesa and drivers. Applications like glxgears run at half speed (250fps on radeon compared to 500 fps on intel, both on VirtualGL), although applications like Doom 3 run pretty much the same (about 20 fps at 1600x900), with the radeon being slightly faster and offering smoother frame rates. Another issue is that I think the image quality on discrete graphics is somewhat lower, possibly due to the image compression scheme(s) used by VirtualGL (which is geared towards efficient compression for transporting over networks).

I could not get to use fglrx using VirtualGL and also couldn't get to use open source radeon drivers to drive the primary X server. While I have tested things on my notebook, which is a Sony Vaio S 13.3 (VPCSA3AFX), these methods should work on similar notebooks with the only possible change of PCI Bus ID of the discrete graphics device (as described further along). Both driver methods are now described in some detail.

Method 1: ATI proprietary drivers (fglrx) method:

This method is fairly straight forward. In a nutshell, the system intially boots up with the Intel i915 driver running the display. But the X server is started with the fglrx driver that works in combination with the userspace component of the Intel drivers (xf86-video-intel) to do the actual graphics processing on the discrete GPU (and transferring the images to the Intel GPU, which is in turn the only device connected to the display outputs).

Hence, the first thing we need is a standard Intel graphics software stack - a recent kernel with i915 drivers, the xf86-video-intel package, with version <2.16 (preferably =2.15*) and a recent mesa and libdrm. The kernel configuration and other details are described later while discussing the second method that uses opensource drivers. The significant detail is that the i915 drivers should be preferably built into the kernel image, instead of being modular. (Even if built modular, it should be loaded before fglrx, perhaps using a initramfs. As explained later, the opensource radeon drivers in the kernel, if built, should preferably be modular and disabled at start using the option "radeon.disable=1" in the bootloader kernel commandline). Regarding the xf86-video-intel package, it is important to use a version <2.16, as was fortunately noted in the forum post [5] (by bojojo2020). This is because the current version of ati-drivers, 11.12 doesn't seem to work with newer versions of xf86-video-intel. (I was initially unsuccessful with 2.17, and in my opinion, this is perhaps the only detail due to which someone would fail in getting fglrx running.)

The rest is along the lines of what is described in the forum [5] and the Gentoo Wiki [6]. Here is a short summary. One emerges ati-drivers, preferably the latest 11.12 (as of now):

[HOWTO] Using recent muxless Intel/ATI hybrid graphics

[HOWTO] Using recent muxless Intel/ATI hybrid graphics

screen 1 deleted

Re: screen 1 deleted

Re: screen 1 deleted

CONFIG_DRM must be disabled