GALAEXI: An Architecture-Agnostic GPU-Acceleration Approach for Legacy Fortran Software

Join Spencer Starr at the Society for Industrial and Applied Mathematics (SIAM)Processing for Scientific Computing (PP26)confrence in Berlin for his talk on “GALAEXI: An Architecture-Agnostic GPU-Acceleration Approach for Legacy Fortran Software”

GALAEXI is a high-order discontinuous Galerkin spectral element method (DGSEM) computational fluid dynamics (CFD) code used for the study of compressible, turbulent flows. It is the device-accelerated version of an existing Fortran-based, CPU-only CFD framework. In this talk, the chosen strategy for porting the Fortran-based compute kernels in GALAEXI to CUDA/HIP C++ is covered. Included are discussions on managing device memory, retaining support for CPU computations, strategies for writing and testing kernels and how to handle GPU-to-GPU MPI communication. Specific emphasis is placed on how the strategies covered can be abstracted in an architecture-agnostic way to allow the greatest degree of portability. The information presented will then be distilled into advice for those seeking to perform a similar porting effort.