A Variety Of OpenGL/OpenCL NVIDIA 367.27 vs. AMD Linux 4.7 + Mesa Git Benchmarks

[Ericg]

RHEL support is coming for sure - don't remember what current plans are for Fedora but my guess is that we may have to package for it as well - my guess is that the Fedora packagers are going to be less excited than most about packaging up closed source stuff



Ahh, interesting - your last comment there mentioned that the lockups were specific to a couple of apps (TF2 and Payday2) - is that still the case ?

I've had it happen a couple times on World Of Warcraft (Crossover), but I haven't tested every single game in my steam library. Team Fortress 2 is very consistent though, playing for a half hour or so will cause the lockups to begin.

[bridgman]

I've had it happen a couple times on World Of Warcraft (Crossover), but I haven't tested every single game in my steam library. Team Fortress 2 is very consistent though, playing for a half hour or so will cause the lockups to begin.

Thanks. Looks like a couple of separate bugs might have been getting conflated in the bug reports then - one related to DPM in general and another related to specific GL apps. The recently reported performance regression may be separate as well, or may be a side-effect of fixing the first (more serious) DPM issue.

[Ericg]

Thanks. Looks like a couple of separate bugs might have been getting conflated in the bug reports then - one related to DPM in general and another related to specific GL apps. The recently reported performance regression may be separate as well, or may be a side-effect of fixing the first (more serious) DPM issue.

Got any suggestions, Bridgman? I'm was originally hesitant to go upstream with this because of Fedora patching their kernel packages, but since it happens on 4.7rc4, then it's obviously upstream too... but they passed it off from kernel-space to userspace.

The only 'symptom' I have to go off of is: under certain workloads, the GPU gets into an inconsistent enough state that it hangs, and the kernel decides to say "Screw it all" and reset everything... which sometimes works and sometimes doesn't. Dmesg seems to blame DPM, but it happens even when I force the GPU into performance mode without DPM.

[finalzone]

RHEL support is coming for sure - don't remember what current plans are for Fedora but my guess is that we may have to package for it as well - my guess is that the Fedora packagers are going to be less excited than most about packaging up closed source stuff

RPM Fusion is very much a complement for the official Fedora repository which cannot include closed source stuff. For so long AMD closed source drive went missing due to complicated building at the point the packager gave up. Can't blame him as other distribution maintainers felt the same way and prefer working on closed source Nvidia driver.

If the AMD driver team developers provide the spec file to build rpm package and the instruction how to use AMD-PRO, that will be great as both Fedora packagers will help you improving it. It is possible to assign conditions to specific distribution on the same spec file in this example

%if 0%{?fedora} >= 23 || 0%{?rhel} > 7
Supplements:    %{name}-doc = %{version}-%{release}
%else
Requires:    %{name}-doc = %{version}-%{release}
%endif

You can also take a look at negativo17 repo source Nvidia which provides a more elegant approach for inspiration. Simply extract the src rpm to access to the spec files.

[Ericg]

RPM Fusion is very much a complement for the official Fedora repository which cannot include closed source stuff. For so long AMD closed source drive went missing due to complicated building at the point the packager gave up. Can't blame him as other distribution maintainers felt the same way and prefer working on closed source Nvidia driver.

If the AMD driver team developers provide the spec file to build rpm package and the instruction how to use AMD-PRO, that will be great as both Fedora packagers will help you improving it. It is possible to assign conditions to specific distribution on the same spec file in this example

%if 0%{?fedora} >= 23 || 0%{?rhel} > 7
Supplements: %{name}-doc = %{version}-%{release}
%else
Requires: %{name}-doc = %{version}-%{release}
%endif

You can also take a look at negativo17 repo source Nvidia which provides a more elegant approach for inspiration. Simply extract the src rpm to access to the spec files.

I pulled down the AMDGPU-PRO debs (The driver is like 20 different debs). They don't look TOO terrible-- no crazing scripting going on, just some ldconfigs and update-alternatives. I'll take a look at negativio's NVIDIA driver to see what the best way might be to hack-together an RPM for AMD.

[liam]

Hey, Bridgman, other than code contributions, is there anything that end-users can do to help get the performance of radeon/radeonsi/amdgpu in a better place? I hate to say it, but I found a 970 for <200 that I am about to pull the trigger on because my 290 is at a quarter of the performance I had on Windows, and a third of a 960... That's pretty sad. And that doesn't even mention the crashes & lockups that have been plaguing 290/390 users for over a year....

I believe that they've said they need traces of problematic games.

[dungeon]

I believe that they've said they need traces of problematic games.

What everybody wants is 100% reproducable case , if happens random maybe can you investigate what combo of software or user modes triggers it and maybe what not, if regression then bisection, etc... basicaly any straight path, cleaner so usable info that leads bug to be fixed.

[dungeon]

I pulled down the AMDGPU-PRO debs (The driver is like 20 different debs). They don't look TOO terrible-- no crazing scripting going on, just some ldconfigs and update-alternatives. I'll take a look at negativio's NVIDIA driver to see what the best way might be to hack-together an RPM for AMD.

You might try to made spec from steamos driver, as i see they just get newest and packaged amdgpu-pro 16.40.0-303090

Those newer versions does not appeared even on amd site, i guess it will be posted there soon too

[liam]

What everybody wants is 100% reproducable case , if happens random maybe can you investigate what combo of software or user modes triggers it and maybe what not, if regression then bisection, etc... basicaly any straight path, cleaner so usable info that leads bug to be fixed.

That's for bug fixing. I'm referring to performance issues given via apitrace.
Sorry for not being clearer

[bridgman]

Those newer versions does not appeared even on amd site, i guess it will be posted there soon too

Nope... I don't think those drivers will ever be published for general use. Next version released is likely to be the launch driver for Polaris.