3 D GPGPU LBM Implementation on Non-Uniform Grids

General Purpose Graphics Processing Units (GPGPU) offer a high performance to price ratio compared to classical CPUs typically used in PC clusters. The standard lattice Boltzmann method (LBM) relies on a uniform Cartesian grid approach. Such uniform grid implementations have demonstrated very good performance on GPGPUs. To simulate test cases which require local grid refinement several solutions for LBM-related grid refinement already exist for CPU-based hardware. Yet, for GPGPUs it is necessary to adapt and reimplement the corresponding algorithmic extensions. Here we present a new approach for a GPGPU implementation of 3D hierarchic grid refinement. In addition, we present a new implementation which requires only one set of distribution functions and thus saves almost half of the required memory. This implementation is suitable for both CPU and GPGPU versions of typical LBM kernels.