Hash Functions and GPU Algorithm of Infinite Grid Method for Contact Search

Abstract

The paper presents the memory-saving GPU algorithm of the infinite grid method with various hash functionsfor contact search in discrete element computations. The implemented hash table of fixed size has the addedbenefit of allowing the grid to be potentially infinite in size, which is particularly suitable for a large numberof discrete particles, moving in large computational domains with empty regions. Research on the applicationof various hash functions and hash table sizes was performed to preserve the computational performance ofthe developed memory-saving GPU algorithm and its OpenCL implementation. The performance of the developed software was evaluated by solving the hopper fill and discharge as well as the artificial avalanche problemson the NVIDIA® Tesla™ P100 GPU. The performance achieved by using the memory-saving implementationof contact search was compared with that attained by using the standard implementation of the uniform gridmethod. The performed analysis revealed that the developed GPU algorithm and its OpenCL implementationreduced the GPU memory consumed by the uniform grid method up to 69.7 times, which resulted in the contactsearch memory equal to 1.86% of the total memory required by DEM computations. Moreover, the applicationof the Morton hash function and the proper hash table size allowed for preserving high computational performance of the infinite grid method and the developed GPU software.