Porting an atmospheric simulator to GPUs using OpenACC

A global atmospheric model called SP-MIROC, written in Fortran 90, is ported to GPUs using the OpenACC directives. To hide kernel launch latencies, all kernels were executed asynchronously. Furthermore, loops containing long and complex codes were divided to simple collapsible ones for better parallelization. The ported code achieved a higher performance-per-watt on NVIDIA A100 GPUs compared to Fujitsu A64FX CPUs and various Intel Xeon CPUs on Wisteria/BDEC-01 and Oakbridge-CX, which are supercomputers operated by the University of Tokyo.