Efficient neighbor search for particle methods on GPUs

Abstract

In this paper we present an efficient and general sorting-based approach for the neighbor search on GPUs. Finding neighbors of a particle is a common task in particle methods and has a significant impact on the overall computational effort– especially in dynamics simulations. We extend a space-filling curve algorithm presented in Connor and Kumar (IEEE Trans Vis Comput Graph, 2009) for its usage on GPUs with the parallel computing model Compute Unified Device Architecture (CUDA). To evaluate our implementation, we consider the respective execution time of our GPU search algorithm, for the most common assemblies of particles: a regular grid, uniformly distributed random points and cluster points in 2 and 3 dimensions. The measured computational time is compared with the theoretical time complexity of the extended algorithm and the computational time of its reference single-core implementation. The presented results show a speed up of factor of 4 comparing the GPU and CPU run times.