Computational fluid dynamics (CFD) codes using direct numerical simulation typically use fast fourier transforms (FFTs) as
part of their solvers. In multi-dimensional MPI-based codes this requires a transpose of the domain across all MPI ranks, which
can be a significant performance bottleneck.
cuDecomp is a library that provides routines for these transposes as well as halo cell communication on GPUs, however it only
supports NVIDIA GPUs. Several of the world’s largest supercomputers contain AMD GPUs and thus are excluded from being
used by cuDecomp-enabled codes.
In this project we will port cuDecomp to HIP, enabling it to run on AMD GPUs. The two libraries will be made available as single
package using e.g. Easybuild so it can be easily installed across many HPC systems with the European Environment for
Scientific Software Installations (EESSI), on either NVIDIA or AMD systems. Furthermore, we want to try and package it for the
most popular distros with tools like packpack and conda for local computer systems
