Abstract
New algorithms are proposed for implementing the high-performance discrete element method (DEM) on a graphics processing unit (GPU), and a GPU-DEM code is developed on the basis of the proposed algorithms. This newly developed GPU-DEM code considers the contact logic of Voigt model in normal and tangential directions, Coulomb type frictional slider, and rolling friction. The following two methods are designed to prevent con ict between GPU memory accesses: (1) generation of contact candidate pairs from neighboring particles and (2) force summation for each particle. In the first method, each particle is assigned a cell index in the domain where the particles exist. Then, a list of contact candidate pairs is prepared by pairing the particle labels that are sorted according to the cell index. In the summation process, a table is constructed to store the list of indexes of the particles in contact with contact candidate pairs. The contact forces acting on a particle for all the contacts are summed by referencing this table in order to apply the action-reaction law to the contact force. Using the two new algorithms, the global memory access con ict is avoided without additional redundant procedures. The calculation speed of our GPU-DEM code that uses the proposed algorithms is approximately 50 times faster than that of the conventional DEM implemented on a CPU.
