ATI R600 Classic Mesa 7.9 Performance

[kadambi]

The 3D textures performance seems to have gone down severely. By about a third.

I used code from here for testing. The FPS has gone down by about third

http://gpwiki.org/index.php/OpenGL:T...ls:3D_Textures

[kadambi]

The 3D textures performance seems to have gone down severely. By about a third.

I used code from here for testing. The FPS has gone down by about third

http://gpwiki.org/index.php/OpenGL:T...ls:3D_Textures

To be clear. I compared non KMS (on FreeBSD 7.2) with the latest on Linux.

[Perry3D]

last paragraph says that he plans on re-testing with 2.6.36, which is needed for tiling.

About warsow: it'd be interesting to see if warsow takes different render paths on newer gl versions or if those are actually speed improvements in mesa. Can you hack mesa-7.9 to advertise only gl-1.5 extensions and see if it's still faster than 7.7.1?

They use 2.6.32. That's too old for such a benchmark
Above all a lot of changes are done at the kernel side.

[xception]

not quite fair comparison

If I understood correctly what happened recently pretty much every new advancement in r600 requires a more recent kernel (not just the color tiling) because of massive drm changes so these comparisons are not quite as they should have been made.

For a better understanding of what has changed and how much the tests should be done with a more recent if not multiple kernel versions both with and without KMS enabled.
I noticed there was promise at the end of the article for more benchmarks with newer kernel, I hope they do multiple kernel versions if not for other reason than just to spot any possible drm regressions in these newer kernels.

[squirrl]

Can it handle Blender3D?

[chrisr]

Interesting, but it doesn't help me.

I'm going to need the S3TC support from Mesa's R300 driver(s) before the R600 drivers become a viable alternative.

[darkbasic]

I'm going to need the S3TC support from Mesa's R300 driver(s) before the R600 drivers become a viable alternative.

I agree.

[stupid character limit]

[drag]

Some extra benchmarks you can throw in are:

Speed-dreams, which is a fork from Torcs.
http://speed-dreams.sourceforge.net/

Torcs was about writing racing simulations and AI drivers, Speed-Dreams is aiming at making a more fun game for a wider audience.

Another one:
http://vdrift.net/

This one tries to make more use of graphics and is designed for more of the 'faster and the furiouser' crowd.

Maybe tile-racer.
http://tileracer.model-view.com/tl/index.php/news.html

And then also flightgear:
http://www.flightgear.org/


These ID-engine-based-only FPS benchmark artiles are killing me. This will at least get some more variation in the benchmarks with engines from entirely different code bases.

I am sure that if you work with the developers of those games a bit they can help you get a reproducible scripted game session. Tell them it will help find problems in OpenGL drivers before it hits their users and will help provide more exposure for their projects.

[The_Monkey_King]

More Testing!

I completely agree. testing needs to be done on 2D and 3D applications and video. Today's video card is expected to do so much more than even five years ago. If you are going to test a video interface, I expect testing for features and not just frame rates.

Benchmarks are nice, and we really appreciate the optimisation work, but in my opinion open source drivers' devs should target features and compatibility first, then speed. What i need is for the thing to work and be compatible with software, i don't care about fps so much. If i needed high performance 3d, i wouldn't use the opensource driver(or Linux at all...)