High performance CPU/GPU multiresolution Poisson solver

We present a multipole-based N-body solver for 3D multiresolution, block-structured grids. The solver is designed for a single heterogeneous CPU/GPU compute node, and evaluates the multipole expansions on the CPU while offloading the compute-heavy particle-particle interactions to the GPU. The regular structure of the destination points is exploited for data parallelism on the CPU, to reduce data transfer to the GPU and to minimize memory accesses during evaluation of the direct and indirect interactions. The algorithmic improvements together with HPC techniques lead to 81% and 96% of the upper bound performance for the CPU and GPU parts, respectively.