R300 Gallium3D Performance Is Topping Out

[phoronix]

Phoronix: R300 Gallium3D Performance Is Topping Out

Recently I showed benchmarks of the Radeon Gallium3D driver for a mature Radeon HD 4870 graphics card over the past two years to look at the performance improvements made to this open-source Linux graphics driver. Up today are benchmarks of an old Radeon X1950PRO (R500 class) ATI graphics card when using the original "R300g" Gallium3D driver and testing every major Mesa release going back to Mesa 7.8 with the classic R300 driver.

http://www.phoronix.com/vr.php?view=17751

[wpoely86]

warsow

Does anyone know what the deal is with Warsow? Why is the classical driver almost 50% faster then gallium3d?

[ChemicalBrother]

Does anyone know what the deal is with Warsow? Why is the classical driver almost 50% faster then gallium3d?

Shaders not rendered correctly? Either way, that's what the tester should actually tell you.

Edit: I mean, the tester should tell you or write the reason, why it is so. At least a little bit more than "seems to work better with classic driver". Otherwise benchmarking warsow makes no sense.

[Veerappan]

Shaders not rendered correctly? Either way, that's what the tester should actually tell you.

Edit: I mean, the tester should tell you or write the reason, why it is so. At least a little bit more than "seems to work better with classic driver". Otherwise benchmarking warsow makes no sense.

But at least we know that something is weird.

Several possibilities:
1) There's a hardware feature that R300c implements, but r300g doesn't (and hits a cpu fallback path).
2) The gallium code is hitting a different (and possibly more complex) rendering path.
3) There's a bottleneck in the gallium code.
4) The classic driver could have had rendering glitches, which I'm assuming Michael would've mentioned... he usually does.

So, if someone wants to do some more diagnostics, it might be good to grab an apitrace of Warsow running on the r300 classic driver, and then on the gallium driver. Try to play back the classic trace on both the classic and the gallium driver, and see what the comparative performance is. If the performance difference disappears during playback, then it's likely that Warsow chose a less complex rendering path for the classic driver. If there's still a performance difference, then someone will need to dig into the driver code to investigate.

[bridgman]

Other possibility, I guess, is that 300g exposes a capability which 300c does not, Warsow uses that new capability, and the result is slower than not using it.

Similar things happened around the transition to GL 2.x IIRC, particularly with id games.

[xeros]

And I'm wondering what was so special with Mesa 8.0.4 that it brought such big performance boost in Xonotic, while on Mesa 8.1-devel FPS goes back to the state of 7.11.2.
Michael, do you have any lower end R300-R500 cards to test? And more tests with lower graphic resolution could say something more about performance.

[oibaf]

Lower FPS on a game may also be because of an optimization for another game. See for example this performance regression on 0 A.D. giving a performance improvement on Nexuiz.

[necro-lover]

Lower FPS on a game may also be because of an optimization for another game. See for example this performance regression on 0 A.D. giving a performance improvement on Nexuiz.

We really need a PTS test-profile for 0 A.D. that game is worth the attention!