The Impact Of Switching To Linux 4.3 + Mesa-11.1/LLVM-3.8 On Ubuntu 15.10

[SystemCrasher]

Don't know the exact history but seems to be an incompatibility between fglrx install logic and gcc 5, and the first place the two fetched up together was U15.10. IIRC a fix is coming which uses gcc 4.9 for the install.

TBH, from my attempts to use catalyst I can admit it always followed very simple "logic", especially in Linux. This logic called SNAFU. Either it crashed kernel with GPF, and nobody bothered to fix it for years, or it failed to build DEB package all the time (I wonder if AMD tests Linux version of catalyst at all?), or it screwed up xorg config, incompletely removed itself, or something else. Ubuntu maintainers attempted to deal with rough edges at least in some places. But they honestly admit in package and repo description its proprietary software and they can't really fix it, whatever. And, hmm, polluting system with uselsess compiler version, potentially providing room for more troubles? SNAFU, as usually.

I think AMD should finally acknowledge obvious: this is just plain fucking wrong way of doing Linux drivers.

[drSeehas]

One number is not really enough as metric.

Agreed.

... min, max and average, ...

min and average are sufficient enough.

[MVinhas]

No problem, I will end up using open source drivers anyway.

It's funny, tbh. I used Windows for years and the only graphics drivers that messed up many times was the ones made by ATI/AMD. I still think that with a better driver, Fury X could leave Titan X in the dust. But guess what, better wait sitting.

[bridgman]

Don't know the exact history but seems to be an incompatibility between fglrx install logic and gcc 5, and the first place the two fetched up together was U15.10. IIRC a fix is coming which uses gcc 4.9 for the install.

[MVinhas]

Thanks

Lost a bit of performance with the newest Ubuntu and Mesa bundled with it (11.0.2), compared to Ubuntu 15.04+3.19+10.5.9. What a shame...

I will try now the Catalyst 15.9 and see who performs better.

edit: ...or not! I guess it's impossible to use fglrx in Ubuntu 15.10. Why in this world the driver is suggested on the Additional Drivers section?

[SystemCrasher]

My graphics card is the AMD HD5770, does it support OpenGL 4.1 with the latest Mesa delivered on Ubuntu 15.10?

From formal point of view, HD5770 does not supports OpenGL 4.1 so far with any MESA version, not even most recent git.

There're two major issues left, if I got it right:
1) Tesselation shaders. Without these it would be hard to call it GL 4.x I guess.
2) IIRC 5770 does not supports FP64 in hardware. So unless someone writes something like software emulation... well, games do not use it anyway, so you can try to fake GL version, even if it formally declared as 3.3. But it would only work adequately after tess shaders are here.

[Michael]

Even PTS has learned to use min, max and average, which are more informative metrics for 3D scenes - at least you can see worst case. Furthermore, PTS seems to be able to draw GRAPHS of frame times, at least for some loads. I think such data could be valuable to pinpoint weak spots and corner cases, etc?

The frame times is limited to engines that support outputting the frame-times in a meaningful way. Likewise, for the min/max support too. There is libframetime support for PTS, but it doesn't work well for all games/apps.

In the case of Marek's interest in GpuTest, that test doesn't even report FPS but rather reports its own "points" system without any min/max or other information being exposed. But like all PTS tests it runs it at least three times for accuracy and if the standard deviation is too great will run additional times automatically, so at least the average should be pretty solid.

[MVinhas]

My graphics card is the AMD HD5770, does it support OpenGL 4.1 with the latest Mesa delivered on Ubuntu 15.10?

[SystemCrasher]

- it prints one number: average fps

One number is not really enough as metric. Imagine you've got decent average fps, but it has regressed in some corner cases and some frames are taking way too long. So FPS drops somewhere to 10 FPS. While you'll get decent number reported as average, it would be unusable anyway. Users would swear when they face 10FPS in some scenes, and at this point its worth of nothing your average FPS has been okay. You've got major roadblock on the way and it could be really wrong idea to release with such regression. Somehow, Catalyst team seems to be bad at it. I.e. in catalyst one would have wild differences in min/max FPS, which makes catalyst unpleasant in games like 3D shooters.

Even PTS has learned to use min, max and average, which are more informative metrics for 3D scenes - at least you can see worst case. Furthermore, PTS seems to be able to draw GRAPHS of frame times, at least for some loads. I think such data could be valuable to pinpoint weak spots and corner cases, etc?

[Michael]

Also look at build-suite to select GpuTest Pixmark 1920x1080 (and any other tests) you want and then that suite's name is what you'd pass to phoronix-test-suite batch-run xxx. Feel free to let me know if any questions or issues. Tom might also be able to answer some of your questions as I think he had some sort of deployment going for his OpenCL tests.