Porting Existing Radiation Code for GPU Acceleration

Abstract

Graphics processing units (GPUs) have proven very robust architectures for performing intensive scientific calculations, resulting in speedups as high as several hundred times. In this paper, the GPU acceleration of a radiation code for use in creating simulated satellite observations of predicted climate change scenarios is explored, particularly the prospect of porting an already existing and widely used radiation transport code to a GPU version that fully exploits the parallel nature of GPUs. The porting process is attempted with a simple radiation code, revealing that this process centers on creating many copies of variables and inlining function/subroutine calls. A resulting speedup of about 25x is reached. This is less than the speedup achieved from a radiation code built for CUDA from scratch, but it was achieved with an already existing radiation code using the PGI Accelerator to automatically generate CUDA kernels, and this demonstrates a possible strategy to speed up other existing models like MODTRAN and LBLRTM.