A fast bond-based peridynamic program based on GPU parallel computing

Peridynamic is an effective method for addressing fracture problems. However, the non-local theory makes it time-consuming. Although some techniques have been developed to improve computational efficiency, the acceleration effect remains relatively limited. This paper introduces a parallel algorithm for bond-based peridynamic using the GPU parallel CUDA programming technology. The calculation process is divided into functions with material points and bonds as the smallest calculation units. The loop of material points and bonds is changed to the index to achieve parallelism. A general horizon generation module is established to optimize storage. Additionally, a general register technique is proposed for high-speed access register memory to reduce global memory access. This technique not only eliminates the restriction on the number of horizon points, also suitable for nonuniform distribution of material points. Compared to serial and OpenMP parallel programs, the present algorithm can achieve up to 800-fold and 100-fold acceleration, respectively. In a typical simulation of one million particles, executing 4000 iterations can be completed in 5 minutes for single precision and 20 minutes for double precision on a low-end GPU PC.