Are there Non-DOE Fortran projects that are currently looking to run on AMD GPUs or transition to portable frameworks ? I’m familiar with projects that enable portable GPU programming for Fortran applications, but I’m interested in getting a feel for how many teams outside of the Department of Energy that are actively heading (or are interested in heading) that direction.
Not migrating to AMD GPUs, but I recently ported a project for NVIDIA GPUs (GitHub - grimme-lab/xtb: Semiempirical Extended Tight-Binding Program Package) with some mixed success. I’ve been looking for portable GPU programming frameworks for this projects to make it more sustainable in the long term, but didn’t really found a suitable one so far.
FWIW, Intel oneAPI is trying to be one programming model for everything: Intel® oneAPI: A Unified X-Architecture Programming Model
(Disclaimer: Intel is my employer, but my posts here are my own.)
1 Like
@awvwgk Focal is a project that exposes OpenCL to Fortran. I have limited experience with it on a demo “mini-app” but am impressed with this implementation. @LKedward on this forum leads that project.
I helped get hipfort off the ground (formerly hip-fortran) and have used this extensively in one of my projects ( SELF ).
If you’re interested in experimenting with either of these, I’m offering some help through free mentored sprints / hackathons this year. As a bonus, we’ve got support for compute cycles on GCP and on some AMD systems if you’re interested in testing. This could help you quickly get through some testing to help you make decisions for your project.
3 Likes
Using OpenCL based approaches turned out to be unsuitable due to missing library support so far.
The main issue I’m facing with GPU portation is finding a suitable provider for LAPACK like functionality. There are a few libraries available I’m aware of: cuSolver, Magma, hipMagma, clMagma (discontinued), clBLAS (dicontinued), clBlast, sycl-blas and now oneMKL.
hipfort might be an option together with hipMagma here, I will definitely look into it.
Definitely use the issue tracker on the hipfort github page to let us know what features would help you out and if you have any issue using what is there currently. We’re here to help!
Hello,
you may be interested in giving this year’s FOSDEM HPC devroom a good look: FOSDEM 2021 - HPC, Big Data and Data Science devroom
Specially: FOSDEM 2021 - Lessons in Programming Model Comparisons Using OpenMP and CUDA for Targetting GPUs
FOSDEM 2021 - Getting ready for the AMD GPUs
FOSDEM 2021 - GPU Computing Made Simple with the C++ Vulkan SDK & the C++ Kompute Framework (AMD, Qualcomm, NVIDIA & Friends)
Regards,
Fernando
1 Like