A fully unsplit wave propagation algorithm for shallow water flows on GPUs

Donna Calhoun

Date : February 21, 2020

Dept. of Mathematics, Colorado School of Mines, Golden, CO (USA) (Invited)

Abstract

The focus of this talk is on a GPU implementation of the fully multi-dimensional patch solver based on the wave propagation algorithm (R. J. LeVeque, Clawpack). Our CUDA implementation is designed for use on small, fixed size patches (32x32) used in a larger the block-based adaptive code ForestClaw (D. Calhoun and C. Burstedde). To update patches on the GPU, we batch-process O(1000) patches per kernel call. Each patch is assigned a single CUDA thread block, eliminating the need for syncing between blocks. By redesigning the WPA, we are able to completely update the solution on each patch in a single batch kernel call. To avoid branch divergence, special attention is given to the implementation of wave limiters. Resulting time using the GPU is about 5-7x speedup over a single CPU. We will demonstrate our algorithm using examples from the shallow water wave equations implemented in ForestClaw.

Slides : mines-2020.pdf