PC Drivers & Software

[pszilard]

We are seeing GROMACS kernels likely miscompiled by AMDGPU-PRO 18.10 for V; several of our unit and regression tests fail that otherwise pass on Hawaii, Fiji, and Baffin.

Where do I report such issues. Reproducing is easy: just follow the build instructions and run "make check".

[pokester]

fsadough​ Can probably help you.

[pszilard]

OK, thanks for the contact. How do things work, do I reach out in a PM?

I still can't find a bug tracker where I can file bug reports, do you have any?

[pokester]

Bug / Issue reports link is on the driver download page, but they don't respond to those it is just a notification.

Support email tickets that they do respond to are the bottom link in the contact page.

There are also links in Radeon Settings.

I included his contact in the thread, that sent him a notification he was mentioned. Usually he would respond to that and may still do that.

[pszilard]

Which "contact page", the forum's, the driver stack's, or AMD general?

I'm not sure what "Radeon Settings" is, presumably a gui tool I suspect. I live on the command line for using GPUs.

[pokester]

Global Technical Support

[fsadough]

Please provide detailed info on your OS, your product and how to reproduce the issue

[pszilard]

Here are the steps to reproduce the issue:

Installation guide — GROMACS 2018 documentation

i.e. get the code, build it and run "make check".

OS:

$ uname -r; cat /etc/lsb-release

4.13.0-41-generic

DISTRIB_ID=Ubuntu

DISTRIB_RELEASE=16.04

DISTRIB_CODENAME=xenial

DISTRIB_DESCRIPTION="Ubuntu 16.04.4 LTS"

Hardware:

$ lspci | grep -i Vega

03:00.0 VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] Vega 10 XTX [Radeon Vega Frontier Edition] (rev ff)

Not sure what you mean the "product"?

[fsadough]

Is this your product?
Radeon™ Vega Frontier Edition Graphics | AMD

[fsadough]

Which driver or ROCm version did you install?

[pszilard]

Yes, it is. Apologies, looks like there was a unannouced software swap and it was actually the ROCm 1.8 stack that I was testing with -- not the AMDGPU-PRO that used to be installed on the machine before.

Yes, the card is a Vega Frontier Edition; ROCm version is the latest from the Deb repos.

I guess I should go back to the github bugtracker and file the issues there?

Related: the Vega card does not seem to work with AMDGPU-PRO installed in legacy mode, is that normal?

[fsadough]

If you install ROCm, then there is no need to install the AMDGPU-Pro driver. ROCm from Github contains the gpu driver 

GitHub - RadeonOpenCompute/ROCm: ROCm - Open Source Platform for HPC and Ultrascale GPU Computing

[pszilard]

I know; we are evaluating ROCm, but at the moment the stability and correctness issues necessitate the use of AMDGPU-PRO. Can you advise on legacy VS PAL mode?

[fsadough]

GitHub - GPUOpen-Drivers/pal: Platform Abstraction Library