Re: ROCm for AMD GPUs at Ubuntu Summit 2024
Thanks for the vote of confidence! On 2025-02-11 02:47, Drew Parsons wrote: On 2025-02-11 09:57, PICCA Frederic-Emmanuel wrote: On 2025-02-11 09:32, Cordell Bloor wrote: On 2025-02-02 15:22, Cordell Bloor wrote: One of the most important things I'd like to share is the list of packages I found that have AMDGPU support upstream, and that could have support enabled within Debian. Same here, I have a bunch of packages which use OpenCL, and it would be great to add autopkgtests with the OpenCL AMD implementation. So all packages with pocl-opencl-icd in the autopkgtest Depends should be good candidate for this instrumentation. "me too". Debian's GPU support is in general somewhat underdone, partly due to cuda's historical nonfreeness. Improving our GPU support is a good theme for a GSOC project. I've submitted a GSoC project proposal [1]. If any Debian Science Team member would like to co-mentor, please feel free to add yourself. I can always refer a student to the appropriate resources for questions when I'm stumped myself, but it certainly doesn't hurt to have a dedicated co-mentor. Sincerely, Cory Bloor [1]: https://wiki.debian.org/SummerOfCode2025/Projects#SummerOfCode2025.2FApprovedProjects.2FEnhancingPackagesWithROCm.Enhancing_Debian_packages_with_ROCm_GPU_acceleration
Re: ROCm for AMD GPUs at Ubuntu Summit 2024
Hi Drew, On 2025-02-11 02:47, Drew Parsons wrote: If we get it running well, we'll start wanting to have debci GPU testing too to be sure it continues to work. Might be possible to build upon my GSOC project last year with Nikolaos Chatzikonstantinou for that step, though good also if debian's own CI infrastructure can cover GPU In the long term, I think the goal is to enhance Debian's own CI infrastructure. Right now, we're running a fairly large set of correctness tests on a GPU-enhanced fork of the DebCI software [1]. However, the validation suite is still not in a state that I'm happy with yet. At the moment, I'm just worrying about verifying correctness. In any case, I took a look at your GSoC project. It was interesting, though perhaps not immediately applicable. AMD upstream does extensive performance regression testing and it is quite a lot of work, so I appreciate how difficult that task is! Sincerely, Cory Bloor [1]: https://ci.rocm.debian.net/
Re: ROCm for AMD GPUs at Ubuntu Summit 2024
From: Drew Parsons Date: Tuesday, 11 February 2025 at 09:47 To: PICCA Frederic-Emmanuel Cc: Andrius Merkys , Cordell Bloor , debian-science Subject: Re: ROCm for AMD GPUs at Ubuntu Summit 2024 On 2025-02-11 09:57, PICCA Frederic-Emmanuel wrote: > Same here, I have a bunch of packages which use OpenCL, and it would be > great to add autopkgtests with the OpenCL AMD implementation. > > So all packages with pocl-opencl-icd in the autopkgtest Depends should > be good candidate for this instrumentation. > "me too". Debian's GPU support is in general somewhat underdone, partly due to cuda's historical nonfreeness. Improving our GPU support is a good theme for a GSOC project. Hypre has some GPU support, good to improve and test its activation. I think PETSc might have some too. If we get it running well, we'll start wanting to have debci GPU testing too to be sure it continues to work. Might be possible to build upon my GSOC project last year with Nikolaos Chatzikonstantinou for that step, though good also if debian's own CI infrastructure can cover GPU. Drew Just to add: I’ve added HIP/ROCm support to mpich 4.3.0 in experimental, (Its present in UCX so OpenMPI uses this); I’m working on enabling it on atlas-ecmwf (just working on one known. Bug), and am finalizing packaging of UCC – Unified Collectives Communication, which adds GPU (HIP) collectives support to UCX for MPI etc. Regards Alastair
Re: ROCm for AMD GPUs at Ubuntu Summit 2024
On 2025-02-11 09:57, PICCA Frederic-Emmanuel wrote: On 2025-02-11 09:32, Cordell Bloor wrote: On 2025-02-02 15:22, Cordell Bloor wrote: One of the most important things I'd like to share is the list of packages I found that have AMDGPU support upstream, and that could have support enabled within Debian. Same here, I have a bunch of packages which use OpenCL, and it would be great to add autopkgtests with the OpenCL AMD implementation. So all packages with pocl-opencl-icd in the autopkgtest Depends should be good candidate for this instrumentation. "me too". Debian's GPU support is in general somewhat underdone, partly due to cuda's historical nonfreeness. Improving our GPU support is a good theme for a GSOC project. Hypre has some GPU support, good to improve and test its activation. I think PETSc might have some too. If we get it running well, we'll start wanting to have debci GPU testing too to be sure it continues to work. Might be possible to build upon my GSOC project last year with Nikolaos Chatzikonstantinou for that step, though good also if debian's own CI infrastructure can cover GPU. Drew
Re: ROCm for AMD GPUs at Ubuntu Summit 2024
> On 2025-02-11 09:32, Cordell Bloor wrote: >> On 2025-02-02 15:22, Cordell Bloor wrote: >>> One of the most important things I'd like to share is the list of >>> packages I found that have AMDGPU support upstream, and that could >>> have support enabled within Debian. These mostly fall under the domain >>> of the Debian Science Team: adios2, blaspp, cp2k, cupy, dbcsr, >>> ectrans, elpa, gloo, hpx, hwloc, hypre, kokkos, lammps, lapackpp, >>> llama-cpp, magma, mfem, mpich, onnxruntime, papi, paraview, petsc, >>> pyfr, pytorch, slepc, spfft, sundials, superlu-dist, trilinos, and >>> whisper-cpp. Same here, I have a bunch of packages which use OpenCL, and it would be great to add autopkgtests with the OpenCL AMD implementation. So all packages with pocl-opencl-icd in the autopkgtest Depends should be good candidate for this instrumentation. An exemple of such instrumentation whcih should be simplify is here https://sources.debian.org/src/clpeak/1.1.3-1/debian/tests/ the upstream-binaries should be factorized for all packages :). This is something to see with the AMDGPU team. Cheers Fred
Re: ROCm for AMD GPUs at Ubuntu Summit 2024
Hi, On 2025-02-11 09:32, Cordell Bloor wrote: On 2025-02-02 15:22, Cordell Bloor wrote: One of the most important things I'd like to share is the list of packages I found that have AMDGPU support upstream, and that could have support enabled within Debian. These mostly fall under the domain of the Debian Science Team: adios2, blaspp, cp2k, cupy, dbcsr, ectrans, elpa, gloo, hpx, hwloc, hypre, kokkos, lammps, lapackpp, llama-cpp, magma, mfem, mpich, onnxruntime, papi, paraview, petsc, pyfr, pytorch, slepc, spfft, sundials, superlu-dist, trilinos, and whisper-cpp. I think that these package enhancements would make for a good Google Summer of Code project. I'd be happy to mentor a new contributor in preparing updates for these packages that enable AMD GPU support. I would imagine that's very likely something that the Debian Science Team would support, but I feel I should nevertheless ask explicitly before trying to engage a third party to do this work. So, is this something that the Debian Science Team would support? And, are there any other barriers to enablement that I should try to help with? I cannot speak for everyone, but I personally would be glad to see AMDGPU support enabled in packages I am interested in, in particular elpa, lammps, onnxruntime, pytorch and spfft. I maintain only the latter, though. Best, Andrius
Re: ROCm for AMD GPUs at Ubuntu Summit 2024
Howdy folks, On 2025-02-02 15:22, Cordell Bloor wrote: One of the most important things I'd like to share is the list of packages I found that have AMDGPU support upstream, and that could have support enabled within Debian. These mostly fall under the domain of the Debian Science Team: adios2, blaspp, cp2k, cupy, dbcsr, ectrans, elpa, gloo, hpx, hwloc, hypre, kokkos, lammps, lapackpp, llama-cpp, magma, mfem, mpich, onnxruntime, papi, paraview, petsc, pyfr, pytorch, slepc, spfft, sundials, superlu-dist, trilinos, and whisper-cpp. I think that these package enhancements would make for a good Google Summer of Code project. I'd be happy to mentor a new contributor in preparing updates for these packages that enable AMD GPU support. I would imagine that's very likely something that the Debian Science Team would support, but I feel I should nevertheless ask explicitly before trying to engage a third party to do this work. So, is this something that the Debian Science Team would support? And, are there any other barriers to enablement that I should try to help with? Sincerely, Cory Bloor
