Linux 4.0 Kernel Testing With AMD Radeon Graphics

[bridgman]

Since Michael isn't putting the old cards into the tests anymore, I ran a quick (read "not so accurate") test using the Tesseract game to compare the opensource radeon driver to catalyst driver on windows. Surprisingly/unfortunately catalyst 13.9 driver for radeonhd4890 on windows7 runs almost 3 times faster than mesa 10.4 on linux.

I was wondering what makes the game to run slower on linux? Is that because of shader optimizers not being still as versatile as catalyst? or do we have higher overhead (on CPU side) in mesa? And how can it (mesa, or use of opengl on game engines side) be improved to achieve almost same performance?

Thanks!

What were the FPS numbers ? Looking at Michael's results from this article, I'm guessing 60-70 FPS for the RV770 ?

[dungeon]

It is better to just run test Michael use in this article, so we can guess or conclude is it something on your side, is it some driver regression or is it what it is.

[dungeon]

Okay, I'll try to. This will take a few hours.

Why hours? Should be few minutes. Download http://www.phoronix-test-suite.com/?k=downloads , unpack and run:

phoronix-test-suite benchmark pts/tesseract

Not all tests from article, just tesseract i meant

[dungeon]

BTW, keep in mind that we don't have ability to conclude if it is renderend correctly or the same, neither for Windows or Linux

For example on radeonsi driver Tesseract with config which is in phoronix test, it only render correctly with llvm 3.7, etc... with earlier versions GI mid just not get rendered

https://bugs.freedesktop.org/show_bug.cgi?id=79155

That is what happened with Tesseract in this article, it is not regression but rendering is correct now

[przemoli]

Since Michael isn't putting the old cards into the tests anymore, I ran a quick (read "not so accurate") test using the Tesseract game to compare the opensource radeon driver to catalyst driver on windows. Surprisingly/unfortunately catalyst 13.9 driver for radeonhd4890 on windows7 runs almost 3 times faster than mesa 10.4 on linux.

I was wondering what makes the game to run slower on linux? Is that because of shader optimizers not being still as versatile as catalyst? or do we have higher overhead (on CPU side) in mesa? And how can it (mesa, or use of opengl on game engines side) be improved to achieve almost same performance?

Thanks!

What GPU? And why not Padoka PPA, but old 10.4?

Get used to using newest stuff for best performance.

I'm currently considering switching back to Ubuntu 14.10 from Fedora if I wont find Mesa-git+LLVM-git for my 5730M.

[dungeon]

Thanks, that saved a few hours! Here is the results with about 8? bots running around in the "complex" map: http://openbenchmarking.org/result/1...LI-SIAVASHSE61

If that is Kaveri it should be fine on eye, but it isn't . So that 39fps in fullhd looks like slower then it should be, if i compare with those 48.07 fps with 10.3 mesa and slower 4870, which AFAIK does not have hyperz enabled...

http://openbenchmarking.org/result/1410027-LI-GGRAPHICS93

I don't think that test is CPU bound, but anyway you might try how it goes with cpufreq on performance. Also, yeah that EXA sometimes trying to vsync fullscreen windows, even if you have vblank_mode=0... you might try xorg.conf options to turn off SwapbuffersWait and off EnablePageFlip

[SXX⁣]

I'm currently considering switching back to Ubuntu 14.10 from Fedora if I wont find Mesa-git+LLVM-git for my 5730M.

I'm sorry, but what for do you need latest LLVM on 5730M?

As far as I aware R600g don't use LLVM shader compiler backend by default because at least 6 month ago it's was in complete rubbish state for any non-GCN hardware. Is something changed since then?

Or you somehow using OpenCL via clover? If yes what's working on it?