VOOZH about

URL: https://www.phoronix.com/review/ubuntu_1404regress_amd

⇱ Likely Radeon Gallium3D Regression On Linux 3.14 + Mesa 10.2 - Phoronix


👁 Phoronix

Likely Radeon Gallium3D Regression On Linux 3.14 + Mesa 10.2

Written by Michael Larabel in Display Drivers on 2 March 2014 at 12:46 AM EST. Page 1 of 4. 64 Comments.

For anyone currently experiencing a slowdown of their Radeon Gallium3D open-source driver stack or are interested in helping out track down a new performance issue, it appears the Linux 3.14 + Mesa 10.2 configuration is regressing on Ubuntu 14.04 LTS in place of Linux 3.13 + Mesa 10.1.

After running a 13-way R600/RadeonSI Gallium3D comparison a few days ago, I was continuing that comparison by comparing the tested Linux 3.13 + Mesa 10.1 configuration against the Linux 3.14 Git kernel with Mesa 10.2-devel Git. Fairly soon into the latest Radeon driver testing I noticed problems: OpenGL performance drops nearly across the board.

👁 Image

Compared to the Ubuntu 14.04 stock configuration that was used in the original 13-way graphics card comparison, the only system changes made since then were upgrading to the latest xorg-edgers packages that provided Mesa 10.2-devel and xf86-video-ati 7.3.99 and then installing the Ubuntu mainline kernel 3.14 package. Both of these upgrades were done on 27 February.

👁 Image

With the upgrade to the newer code, for many OpenGL benchmarks there's now a significant drop in performance for the five tested graphics cards (thus far) compared to the Linux 3.13 + Mesa 10.1 stable state. I'm mentioning this now in case anyone else has encountered similar issues with the newer code or looking to try a similar comparison on your own systems in hopes of spotting the issue. I'll likely get around to digging deeper in a few days time (and depending upon what other Phoronix readers report in as their experiences) to find out of it's a Mesa or kernel issue but as my time is already stretched very thin, that work unfortunately has to be punted to the back-burner until finding extra time or resources to dig deeper into this likely regression.

Here are the results thus far illustrating the problems with the newer code.