Skip to content
fomics edited this page Jul 14, 2015 · 4 revisions

PI: Klaus Dolag

Institution: University of Munich, Leibniz Supercomputing Centre, others

Application Area: OpenMP/MPI parallelized TreePM-MHD-SPH code performing cosmological simulations within the astrophysics community

Experience: monotonic increase is OpenACC experience, but limited satisfaction with performance so far achieved.

Approach: Tree-walk has to be ported to GPUs, several different approaches, e.g., "core-guided" OpenACC calculation, attempted.

Issues: compiler problems with both PGI and Cray (reported). In particular ASYNC(tid) for ACC parallel region with OMP multithreaded region not properly supported on Cray.

Future: frustration comes from seeing the potential solution but not finding a way to work around compiler limitations, e.g., lack of reductions in parallel regions.

Final Presentation