Running Mesa 9.2-devel + LLVM 3.3 SVN With The R600 Back-End

[phoronix]

Phoronix: Running Mesa 9.2-devel + LLVM 3.3 SVN With The R600 Back-End

The last time I extensively tested the AMD Radeon Gallium3D LLVM shader compiler back-end was last April. Since then the R600 LLVM back-end has matured quite a lot with new features and was merged into upstream LLVM. In the past few days I carried out some new tests on several different graphics cards using Mesa Git master of the R600 Gallium3D open-source graphics driver.

http://www.phoronix.com/vr.php?view=18510

[duby229]

Still though it would be testable by those who do actually own it.. Not everyone would be able to make comparisons, but those that have the game could. I would love to know how well some of the commercial games that are finally being made available for linux performs - regardless of whether or not I could duplicate the tests.

[zigfreed]

Or more demanding Trine 2 and Serious Sam 3.

You wouldn't be able to put that into PTS and compare it to your own hardware. S2Games and Epic have have games out for Linux, however if you can't compile them then you can't see how the entire system has evolved.

A DX9 game would be nice, however.

[kwahoo]

I don't know why Michael refuses to use better tests than he does.

Add in a WINE test, a Valve Source engine test, and Unigine, and you'd get 3 distinctly different engines apart from the Id engine he relies almost exclusively on. They'd all probably be bound on very different things performance-wise.

Or more demanding Trine 2 and Serious Sam 3.

[smitty3268]

Compiler influence on nexuiz and Doom3 are negligible (they are not shader bound), and I suspect this extend to any ioquake3 / doom3 engine powered benchmark.
I personnaly use Unigine Heaven 3.0 and Lightmark 2008 2.0 to measure performance of compiler.

I don't know why Michael refuses to use better tests than he does.

Add in a WINE test, a Valve Source engine test, and Unigine, and you'd get 3 distinctly different engines apart from the Id engine he relies almost exclusively on. They'd all probably be bound on very different things performance-wise.

[brosis]

I personnaly use Unigine Heaven 3.0 and Lightmark 2008 2.0 to measure performance of compiler.

Same here. Lightmark is pretty reliable.

[vljn]

Compiler influence on nexuiz and Doom3 are negligible (they are not shader bound), and I suspect this extend to any ioquake3 / doom3 engine powered benchmark.
I personnaly use Unigine Heaven 3.0 and Lightmark 2008 2.0 to measure performance of compiler.

[tmikov]

comparison against the regular backend

Would be useful to see how LLVM compares against the conventional backend.

[Drago]

In the next tests, include Vadim's back-end for r600g.