HIP-RRTMG_SW: Accelerating a shortwave radiative transfer scheme under the heterogeneous-compute interface for portability (HIP) framework

Abstract

With the development of higher-resolution atmospheric circulation models, the amount of calculation increases polynomially with resolution, and the calculation accuracy of physical processes is increasing rapidly. The traditional parallel computing methods based on multi-core CPUs can no longer meet the requirements of high efficiency and real-time computing performance of climate models. In order to improve the computational efficiency and scalability of the Atmospheric General Circulation Model, it is urgent to study efficient parallel algorithms and performance optimization methods for radiation physical process with massive calculations. In this paper, a heterogeneous multidimensional acceleration algorithm is proposed for the shortwave radiation transfer model (RRTMG_SW) based on HIP. Then, the HIP version of RRTMG_SW is developed, namely HIP-RRTMG_SW. In addition, combined with the “MPI + HIP” hybrid programming model, a multi-GPU implementation of RRTMG_SW is also proposed, and it makes full use of the multi-node, multi-core CPU and multi-GPU computing capability of a heterogeneous high performance computing system. Experimental results show that HIP-RRTMG_SW achieves 7.05× of acceleration in the climate simulation with 0.25^∘ resolution using 16 AMD GPUs on the ORISE supercomputer compared with RRTMG_SW using 128 CPU cores. When using 1024 AMD GPUs, HIP-RRTMG_SW is 83.94× faster than RRTMG_SW with 128 CPU cores, indicating that the proposed multi-GPU acceleration algorithm has strong scalability.