On Tue, Apr 20, 2021 at 11:10:11AM -0400, Xu Zhao wrote:
Currently, I am using the official ROCM zypper repository provided by AMD: https://repo.radeon.com/rocm/zyp/zypper/
This way is similar to how Nvidia distributes nvidia UNIX drivers on openSUSE. Is there a reason we want to replicate this package in OSS?
Sorry for late reply. I just received this email so must be some issue with delivery. The packages you're linking to are AFAIK built for SLE15-SP2. I would like to have packages built against Factory and included in Tumbleweed. Thanks Patrik
Best, Xu
-- Xu Zhao nuk.zhao@utoronto.ca
On Tue, 20 Apr 2021, at 5:23 AM, Patrik Jakobsson wrote:
Hi,
I've been working on creating OpenSUSE packages for AMD's ROCm stack for some time. It consists of components for building and running compute workloads on AMD GPUs. One key component is a ROCm specific version of LLVM. This makes things tricky to distribute in normal fashion since we need to include an extra version of LLVM that do not collide with the rest of the system. The solution by AMD has been to put the entire stack in /opt/rocm.
AMD has also released a specific build of the stack called AOMP. The focus is on OpenMP but it also includes support for OpenCL, HIP, CUDA and Fortran. This is normally installed in /usr/lib/aomp-%{version}. AOMP also includes build scripts for building the entire stack which is very helpful.
What I've done is to create an AOMP package for Tumbleweed which is available at [1]. It tries to mimic what AMD has done for the package they provide for SLE15-SP1. I'm no expert on packaging so there's probably improvents to be made.
Is this something that can be included in Tumbleweed? I know the optimal solution would be to have all the ROCm specific parts upstreamed and included in the official LLVM package and then provide seperate packages with normal installation paths. Unfortunately we're not there yet and this would be a stop-gap solution until then.
Any comments are welcome
Thanks Patrik Jakobsson
[1] https://build.opensuse.org/package/show/science:GPU:ROCm/aomp