Effective mapping of an SPH algorithm on massively parallel GPU architecture

In the present study, the performance of a Lagrangian, mesh-free, particle-based method called Smoothed Particle Hydrodynamics (SPH) is investigated on a General Purpose Graphics Processing Unit (GPGPU) architecture. A one-to-one mapping of host (CPU) function to device (GPU) kernel is particularly used. A new methodology of sorting the evolution of spatio-temporal data of particles based on cells is tested on GPU for efficiency measures such as speedup, Dynamic Random Access Memory (DRAM) utilization, warp execution, occupancy of each kernel with different grids, block sizes, etc. Thread-divergence caused by spline and Wendland families of weighting functions has been studied. In SPH algorithm, an overall speedup of 15× was achieved on GPU.