Open-Source Radeon HD 6000 Series Still Borked

[darkbasic]

You sure you set Gallium up the same way Michael did?

Swapbufferwaits off, tiling and pageflipping on, vblank_mode 0, compositing off, I am not aware of any other optimization. Michael, did you tweak something else?

[crazycheese]

thanks for tests, darkbasic!
Yes looks extremely stupid...
You sure you set Gallium up the same way Michael did?
Core2K will of course spot some light...

[darkbasic]

Ok, here are the first results: http://openbenchmarking.org/result/1...GR-1107137LI76

I fear it's TERRIBLY cpu limited. It should be two times faster watching Michael's HD6870 results. Core i5 2500K results will follow and I will finally have an answer.

[darkbasic]

Why not use xonotic?

Because it isn't still mature enough and from snapshot to snapshot the timedemo does not work. Next week I will file some bugs regarding power management and the low performance I got. If want to see at least the 150 fps in nexuiz that Michael got with the 6870.

[crazycheese]

If the 6870 does 153 fps, the 5870 should be even better because it's a faster card.
I got something like ~ 70 fps @1024x768 normal quality.

5870, when using correctly programmed drivers and having an application thats not CPU bound and not serving one texture at time(I mean, the one capable to load 5pack GPU cores with tasks), should perform exactly as 6970, or fluctuate between 6970 and 6950.

20 fps in openarena with latest graphic stack and my HD5870.

Even before today's update it was too slow, Michael confirmed it:

Errata corrige: 20 fps the first time I run the benchmark, 90 fps the second time and 160 fps the third

Something is very wrong. I had at least 60 fps when I had 4770 with opensource drivers half-year ago(fullhd, amd athlon II x4) and that constantly, from first launch. I think it even was vsyncing... 4770 is same to 5670. 5870 should at minimum do 170+fps on that old opensource, theoretically. The problem was, when it came to heavy action, like having several players fragged in front of you, the fps went to 5-10 making game non-playable, but watchable.

When I had 4670 a year ago, with non-usable opensource drivers, catalyst performed at 160fps almost constantly. So, judging 5870 potential projecting from blob on 4670 - it should be around 400 fps.

And still a ridiculous 57 fps in nexuiz demo1, normal quality, 1920x1200

Dump the nexuiz altogether - its not being developed anymore due to main developer turning into a jerk (oss-style). Why not use xonotic?

[darkbasic]

20 fps in openarena with latest graphic stack and my HD5870.

Errata corrige: 20 fps the first time I run the benchmark, 90 fps the second time and 160 fps the third

And still a ridiculous 57 fps in nexuiz demo1, normal quality, 1920x1200

[darkbasic]

20 fps in openarena with latest graphic stack and my HD5870.

Even before today's update it was too slow, Michael confirmed it:



If the 6870 does 153 fps, the 5870 should be even better because it's a faster card.
I got something like ~ 70 fps @1024x768 normal quality.

[curaga]

Ok. 100 fps.
- Consider the rendering is better than OpenGL 2.1.
- Consider it is 1280x1024 instead of 1024x768
- Consider it has 16x Anisotropic filtering while mesa doesn't support it
- Consider it has 4xAA while mesa doesn't support anti aliasing.
- Consider the final FPS score is lower than a screenshot because sometimes there are big fps drop.

I was wrong, mesa is 30 times slower at least.

- different OS
- not OGL path

That does show a big difference, but results from the linux blob would be more comparable.

[darkbasic]

Hey, mesa is opensource! All the drawbacks you listed require huge crew of driver developers with access to hardware.

I know it very well and I knew it before buying every card, I bought them because proprietary drivers are no alternative and I'm already quite satisfied anyway.

[smitty3268]

Do you have a link to show that? In http://www.phoronix.com/scan.php?pag...e_heaven&num=2 even the best card, 4890 with the blob, did not break 80fps at 1024x768.

Would the 3.x vs 2.1 codepath really have that much impact?

I don't think there is a 2.1 vs 3.x codepath difference. It requires at least some GL 3 features, which Mesa has implemented, so I doubt they'd go to the trouble of creating alternative paths for other stuff. Maybe someone can point to a specific extension that Mesa is still lacking - the tesselation in 4.0 and AA are the main things I know of that aren't implemented yet, and that screenshot also shows tesselation off in the proprietary driver test.

Edit: I suppose they're probably using some geometry shaders as well. Forgot about that.