Previewing The Radeon Gallium3D Shader Optimizations

[marek]

I think Michael did not set the environment variable R600_DEBUG=sb for the shader backend to be actually enabled.

[Bitiquinho]

Shouldn't mesa be exposing OpenGL 3.1 support on this test, rather then 3.0 ? Also, I think Michael have to set R600_LLVM env variable to 0, so the LLVM backend are not used.

[stalkerg]

Yes, something is suspicious here!
Vadim, do you have paypal? : )

Or Yandex.Money? And I can put a beer in the pub. (in one city because)

[vadimg]

Yes, something is suspicious here!

I agree with Vadim - even Phoronix readers have noticed ondemand governor throttling clocks, ie not detecting that there is demand, and causing bottleneck.
32bit tests on 64bit mesa are very probable too, I don't know how Ubuntu manages multilib, but ifs pack of prebuilt files instead of seperately recompiled i386, it is also possible them to interfere.

It would be very good if Michael retest the case, forcing performance governor and running only 64-bit tests on 64bit host.

Vadim, do you have paypal? : )

Yes, vadimgirlin at gmail dot com.

By the way, 32-bit tests are not a problem, you just need to build 32-bit mesa with r600-sb for them.

[brosis]

Yes, something is suspicious here!

I agree with Vadim - even Phoronix readers have noticed ondemand governor throttling clocks, ie not detecting that there is demand, and causing bottleneck.
32bit tests on 64bit mesa are very probable too, I don't know how Ubuntu manages multilib, but ifs pack of prebuilt files instead of seperately recompiled i386, it is also possible them to interfere.

It would be very good if Michael retest the case, forcing performance governor and running only 64-bit tests on 64bit host.

Vadim, do you have paypal? : )

[vadimg]

I suspect something is wrong with some of these results. Though many applications are not really expected to have noticeable improvements because they are not limited by shader performance (openarena, doom3, xonotic/nexuiz on low settings), but I'd expect different results at least with Unigine benchmarks.

Michael, did you take into account that Tropics and Sanctuary are 32-bit apps and use 32-bit drivers?

Also, probably not very important for these benchmarks, but I'd use performance governor for all benchmarks to avoid running them on different cpu frequencies due to different cpu loads, especially when you are comparing different drivers (versions, branches, etc). E.g. the driver with higher cpu load/overhead may end up running on higher frequency and produce better results. I think it's better to use fixed frequency for all benchmarks, otherwise it's like comparing the results of different drivers on different cpu's.

By the way, here are my results with some apps:

http://openbenchmarking.org/result/1304309-FO-57500792103#r-0

[stalkerg]

You joke? I have +30%, +40% in SC2, OilRush and ungine haven.On 6770.

[Drago]

My results are very different. I see consistently and noticeably improved performance, almost everywhere, with SB enabled. However, I have an APU which is bottlenecked heavily by shader performance. This might not be the case with those higher-end desktop GPUs.

I also have an APU. Can you share some of your findings?

[brent]

My results are very different. I see consistently and noticeably improved performance, almost everywhere, with SB enabled. However, I have an APU which is bottlenecked heavily by shader performance. This might not be the case with those higher-end desktop GPUs.

[vljn]

I'm a bit surprised with Unigine results, in the Heaven demo there is a significant performance improvement due to better shader, with llvm or sb backend.
Lightmark 2008.2 is also a benchmark where compiled shaders quality matters too.