AMD Kaveri: Catalyst vs. RadeonSI Gallium3D On Linux

[phoronix]

Phoronix: AMD Kaveri: Catalyst vs. RadeonSI Gallium3D On Linux

With the recently released AMD A10-7800 Kaveri APU I carried out some new benchmarks comparing the open and closed-source Linux GPU driver performance for AMD with their Catalyst and RadeonSI Gallium3D solutions. When running the open-source Ubuntu driver tests, multiple versions of Mesa and the Linux kernel were used.

http://www.phoronix.com/vr.php?view=20756

[ChrisXY]

FYI: gimark works on mesa radeonsi when using

MESA_GL_VERSION_OVERRIDE=3.3 MESA_GLSL_VERSION_OVERRIDE=330

All necessary extensions seem to be supported and it's just using the compatibility context instead of core.

No unigine valley? In my experience this is the one that works best on open source drivers compared to catalyst and I think unlike heaven all features actually work (heaven does quite a lot with tesselation)

LLVM backend problems are missing VGPR register spilling I guess? AMD really needs to put some more people on the llvm backend to fix this mess. Tom Stellard is really good, but he is just doing way too much at once.

[przemoli]

FYI: gimark works on mesa radeonsi when using

MESA_GL_VERSION_OVERRIDE=3.3 MESA_GLSL_VERSION_OVERRIDE=330

All necessary extensions seem to be supported and it's just using the compatibility context instead of core.

No unigine valley? In my experience this is the one that works best on open source drivers compared to catalyst and I think unlike heaven all features actually work (heaven does quite a lot with tesselation)

LLVM backend problems are missing VGPR register spilling I guess? AMD really needs to put some more people on the llvm backend to fix this mess. Tom Stellard is really good, but he is just doing way too much at once.

There are updates there.

However Oibaf do not track LLVM git.

[smitty3268]

LLVM backend problems are missing VGPR register spilling I guess? AMD really needs to put some more people on the llvm backend to fix this mess. Tom Stellard is really good, but he is just doing way too much at once.

Well, it mostly is fixed now, as shown by the successful test runs of the test system when running LLVM3.5. But i agree Tom seems overwhelmed with all that he's working on right now, he needs some help.

Of some note, there are some potentially large performance improvements that should come with the 3.17 kernel. No idea exactly how large, but 10-20% possibly, at least on some tests.

[ChrisXY]

Well, it mostly is fixed now, as shown by the successful test runs of the test system when running LLVM3.5.

Not really, it's only a workaround to prevent crashing, "there may be some mis-rendering": https://bugs.freedesktop.org/show_bug.cgi?id=75276#c26

[dungeon]

Now this is interesting, i have better results with top Kabini than Michael here with near top Kaveri .

How so, mostly just enabled hyperz and governor performance .

edit: and i am not use mesa with out of tree patches. As i see gallium-nine patched mesa is used here, but to be honest not sure if that can hurt performance somehow .

[xeekei]

I would like to see a Radeon HD 6870 tested. It's the best card with open source drivers at the moment. Does it rival itself with Catalyst?

[molecule-eye]

Nice. I was curious about APU performance differences when running blob vs not. Lots of people are still running Trinity and Richland though so it would be interesting ti also have those numbers and cover the whole APU line worth covering.

How is perceived 2D performance with Catalyst? I cannot run it on my A10-5800K on KDE because composited 2D (smooth scrolling in Firefox and Chrome eg) is so choppy when it feels like butter with Gallium.

[drSeehas]

Radeon HD 6870

I would like to see a Radeon HD 6870 tested. ...

Me too.

[Michael]

I would like to see a Radeon HD 6870 tested. It's the best card with open source drivers at the moment. Does it rival itself with Catalyst?

From the linked article posted just last week.... http://www.phoronix.com/vr.php?view=20732