New SteamOS Beta Bundles Interesting AMDGPU-PRO Driver

[Kano]

I reuse SteamOS drivers for Kanotix too, but I can not test amdgpu personally, just if they install/uninstall correctly. Beware that glx-alternative-amdgpu-pro has a bug that prevents uninstall in case of no installed fglrx driver. The new Nvidia driver package has another bug, it required a xorg.conf file, because nvidia-drm-outputclass.conf is only triggered with nvidia-drm driver, but that was not loaded. You can hack nvidia-modprobe.conf to load nvidia-drm instead of nvidia, but that file looks really weird, would prefer a more complete rewrite of that file.

[tomtomme]

*Bridgman shakes head, wonders what part of "those drivers aren't tested for or intended for use outside of SteamOS" he is failing to communicate clearly

If you are saying "every driver we send to any partner for any purpose also needs to be made available for general use and broadly communicated as such even though we have no idea if it will even work" that's fair to ask but it seems like a really bad idea to me. Feel free to convince me I'm wrong.

poor bridgman. i am with you. For what its worth: I understood what you said from your first comment.

[eydee]

AMD did restrict them from being used by others by not releasing the information about the newer drivers through usual channels.

I wouldn't call that a restriction. They could have coded it to only boot with SteamOS kernel or something like that. They did not.

[bridgman]

Huh ? You're saying we can't have release branches where we test & bug fix for one specific partner ?

How could that possibly help ?

[FireBurn]

bridgman unless you haven't noticed by now, Guest has more or less been popping up everywhere stirring things when they do. Each time it happens it seems less like someone who wants to get involved and more like someone intent on trolling. Guest dev's only have limited bandwidth and it shouldn't be wasted on arguing with you over things only you seem to be getting upset over. I get there is a community, but you don't join that community by continuously throwing mud at it

[devius]

If they aren't intended for use outside of SteamOS then there needs to exist a separate "steamos" git/etc branch for that. If you merge the "steamos" branch into the master branch and publish the master branch for use outside of SteamOS, it will invalidate your claim "those drivers aren't intended for use outside of SteamOS". If there is no "steamos" branch and SteamOS-related changes are in the master branch then publishing the master branch for use outside of SteamOS will invalidate your claim "those drivers aren't intended for use outside of SteamOS".

So, you're basically telling AMD how they should organize their close-sourced and behind-the-doors repository where they work on AMDGPU-PRO, so that you feel better even though you will never have access to that repository.

That makes total sense...

[pszilard]

The NVIDIA 367.27 driver update is nice but not particularly noteworthy as it's been public for a week weeks

Actually, it is noteworthy. I had my suspicion about it, but had to confirm it first (did hint/note in the respective thread) and I did get my bug report accepted by NV, so I can safely say, there is a bizarre, possibly serious memory bandwidth issue on the 1080 and it's not clear whether it's a driver bug introduced by the 367.27 or it's a workaround for some other issue with a nasty side-effect.

What I've found is (and others confirmed) that the global memory bandwidth on the GTX 1080 drops by ~18% wrt the 367.08 release. I measure ~230 GB/s with the new driver, while ~278 GB/s with the old one. The memory does not clock up fully either, but that does not explain the full story. So, for my any bandwidth bound compute (and most probably graphics) kernel this means up to 20% slowdown and lower performance than a 980 Ti. I'm surprised gaming sites don't talk about it as apparently most (all?) Windows drivers suffer from the same (but then again, gaming benchmarks and reviews have become quite shallow these days, TBH).

[bridgman]

So, you're basically telling AMD how they should organize their close-sourced and behind-the-doors repository where they work on AMDGPU-PRO, so that you feel better even though you will never have access to that repository.

I'm OK with that in principle, just having trouble with the specific recommendation in this case.

Our core development always happens in a single branch (actually a pair of tightly coupled branches for open/hybrid kernel driver), but bug fixes and experimental performance tweaks for specific partner scenarios may end up being done initially in a release branch specific to that partner/customer and then merged back into the main tree after more testing. It's basically managing risk in an environment where you don't have infinite testing & merging bandwidth (eg the build we sent for SteamOS would probably not work properly on a Polaris card).

[Kano]

Testing vs Release AMD drivers... You can be sure that there is no diff in stability and as hardware support is mentioned: all drivers support a whole range of cards, the old whitelist for aticonfig and based on pci IDs was completely stupid. The differences between the same series can not be so huge that a card stops magically working that worked before. AMDGPU just seems to be less compatible in general with some output configurations from what I read on this site.

[beniwtv]

Hoping bridgman can shed some light here...

So I got my new RX 480 GB, to use it in my SteamOS machine.

I updated to brewmaster-beta and that gave me 16.40.x as driver (the one from the SteamOS repo).
That driver works, (although I had to hack the update-graphics script since that doesn't seem to support the RX 480 yet) I get OpenGL 4.5 and so on, but on some screens there are some flickering lines appearing across the screen.

This usually does not happen when in a game and the card is being used highly. Also does not relate to only the SteamOS Big Picture mode, as I get the same problem in the Gnome 3 session Steam OS provides.

I wanted to check if installing the driver for AMD.com would work... which is 16.30.x and get rid of the problem. Unfortunately, that does not install on Steam OS. First, it tries to install a package not shipping with the archive (amdgpu-pro-driver), fixing that gets you into a update-alternatives glx problem, and unconfigured packages.

So, am I to assume that the Steam OS 16.40.x driver is not yet fully supporting the RX 480?
Will you guys be providing them with new drivers?