AMDGPU-PRO vs. Open-Source Gallium3D OpenGL Performance On Polaris Is A Very Tight Race

[duby229]

This was with the latest Mesa Git as of like two days ago. There may be some regression on Linux 4.8 / Mesa Git, ran into another odd situation this morning with a R7 260X and will likely have an article out shortly about that.

Those types of articles that identify weirdness/regressions and that try to point at a cause are really awesome. I don't know if it's true, but I bet those types of articles get a lot more page hits.

[Michael]

Those types of articles that identify weirdness/regressions and that try to point at a cause are really awesome. I don't know if it's true, but I bet those types of articles get a lot more page hits.

Nope, really needs to be some crazy regression for it to be worthwhile usually...

[iznogood]

Michael OK, but if your results differ so much with the same configuration just after a few days, then how can the test method considered reliable?

[Michael]

Michael OK, but if your results differ so much with the same configuration just after a few days, then how can the test method considered reliable?

Do note that the OS changed from 16.10 to 16.04, etc.

[Nille_kungen]

If there haven't been any code changes, do the latest set of Mesa results include the recent Bioshock performance patch? The latest results seem to have dropped from the previous article back to around the pre-patch values and in the case of the RX460 the results are significantly worse than either of the previous two sets of results?

Benchmark	Average Frame Rate
RX460 Old Mesa	62.82
RX460 New Mesa	82.80
RX460 Latest Results	32.47

Would it be possible to observe the AMDGPU-PRO Bioshock benchmarks running on the R9 Fury to determine if the frame rate really is rock solid at close to the maximum value or if there's a problem with the way the results are being collected? A difference of only 0.32% between the maximum (82.38 fps) and the average (82.11 fps) results sounds much to low unless the frame rate has been capped.

http://openbenchmarking.org/prospect/1608200-PTS-AMDGPUPR97/76c1d560ee4669f7cd81b1b56c6d92a067c5594f

http://openbenchmarking.org/prospect/1608138-LO-MESAGIT3992/76c1d560ee4669f7cd81b1b56c6d92a067c5594f

First look it looked close to pre-patch but the RX460 is way worse then that.
Looking at the system information it looks like the only difference is 11days of kernel development.
With the numbers from the patched benchmark mesa would have beaten the -pro stack.

[Nille_kungen]

Do note that the OS changed from 16.10 to 16.04, etc.

I didn't notice that.

[duby229]

I like it to and I think phoronix should be the backbone of testing and solve such bugs.
but sadly michael has no free time or/and no money for this.

Well, If the revenue/effort amount ratio is too low then I can understand. I do really like those types of articles though. It would be freaking awesome if oss driver devs released patches with a article style format including PTS benches that describe the situation. I understand that Michael decided it isn't financially worthwhile for him to do it. But it would be a very cool value add for someone to do it.

[pal666]

But why?

because it made such decision

[pal666]

Catalyst is more or less dead but i can see that amdgpu-pro can still be called catalyst since it's less confusing.

i am pretty sure he wanted to say 'blob' instead of catalyst and blob is not going away, though there is no need to use it for gaming already

[pal666]

Impressive! Makes me wonder whether the investment on the closed-source AMDGPU part wouldn't have been better spent on the open-source Gallium-3D driver instead.

closed-source part is developed by windows driver team, mesa is of no use for them