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

[Azpegath]

It's not. The HUD shows zero shaders created and zero compilations while it happens. But just now I noticed that the GTT usage drops a little bit, so I suspect something inefficient somewhere in the memory management. First thought is of course https://lists.freedesktop.org/archiv...st/001360.html but there can be no memory pressure on my 8 gigabyte rx 480...

You should post a bug on this issue, with all the relevant information you've gathered. I think there are quite a few performance issues that never get reported.

[pal666]

the only test which saddens me is bioshock. it greatly improved recently, but it needs more

[pal666]

The question may become relevant in the future.

it can't, because you missed one feature

[pal666]

what is the point of non free drivers?

paying customers

[eydee]

It's not. The HUD shows zero shaders created and zero compilations while it happens. But just now I noticed that the GTT usage drops a little bit, so I suspect something inefficient somewhere in the memory management. First thought is of course https://lists.freedesktop.org/archiv...st/001360.html but there can be no memory pressure on my 8 gigabyte rx 480...

This thing happens in pretty much all Source games. The issue may be in Valve's code.

[pal666]

I think that day has already passed.

that day will never come because mesa is not going to get compat profile support

[iznogood]

Michael, on Bioshock, why are you showing the results before the optimisation and not after? The results you showed on your previous article had all cards performing above the 80 fps, which is not the case in this test

[stqn]

Nice results for the FOSS driver; if non-overclocked RX 470 appear that need no more than 120W and the driver is improved enough so that games perform better than on a GTX 970, then I may buy one… At the moment I’m looking at getting a GTX 1060 instead because it performs much better, needs no more than 120W and I’m sure the drivers will work very well.

It makes no sense that the RX 470 performs just as well or better than the RX 480… It also costs the same… (Yes I know it’s an overclocked RX 470, but still.)

[dungeon]

He, he, steamos again get new pro driver... seems every month new blob checkout

[bwyan]

I just tried duplicating these results using phoronix-test-suite. For some reason, the AMDGPU-PRO driver always performs much worse for me on my new AMD R9 Fury on Ubuntu 16.04 when comparing my results to the ones Michael are getting (around ~25-30 percent lower). Is there something that I may be missing that I should be aware of or is my GPU simply being bottlenecked by my i5-2500k CPU?

For anyone interested, it turned out that my problem was that I had "non-stock" kernels 4.7 and 4.8rc2 installed (for experimentation with Mesa improvements).

Even though I had booted on stock 4.4 kernel when installing the amdgpu-pro driver, the amdgpu-pro-installer still attempted to build kernel modules for the newer kernels and threw errors for those. However, the kernel module for kernel 4.4 seemed to be built as expected and so, as glxinfo was reporting OpenGL 4.5 compatibility, I assumed that everything was dandy.

In the end, uninstalling kernel 4.7 and 4.8rc2 and then reinstalling the driver seems to have resolved my problems and the amdgpu-pro driver now performs more or less in line with Michael's results.