Browsing through Twitter I found the ISC 2020 Digital event,
an online event that focuses on bringing the most critical developments and trends in high performance computing, machine learning and data analytics for the benefit of the global HPC community.
The event website is at https://www.isc-hpc.com/.
What caught my attention was the following presentantion (cross-posted at the OpenMP Twitter feed):
which examines how efficiently legacy Fortran codes can be ported to accelerators (GPUs) by leveraging OpenMP directives.
Note that registration is required to view the presentations.
5 Likes
And registration is free!
Interesting presentation and a good demonstration of the already existing capability to use GPUs from Fortran.
Directive-based approaches are often touted as a minimally-invasive option for implementing parallelism in existing codes - however, and as demonstrated in that presentation, significant code restructuring is very often required to properly exploit the GPU parallelism since the original code is usually optimised for cache-based architectures.
Moreover, this code restructuring requires a good knowledge of the underlying factors that affect parallel performance.
My preference is for a lower-level approach such as CUDA and OpenCL which are much more explicit by comparison and hence offer more control and opportunities for optimisation.
I’ve noticed two other presentations at Fortran-con on GPUs - I’m looking forward to seeing them!
2 Likes