An adaptive dual-information FMBEM for 3D elasticity and its GPU implementation

Abstract

Combined the boundary element method (BEM) with the fast multipole method (FMM), the fast multipole BEM (FMBEM) is proposed to solve large scale problems. A key issue the FMBEM has to address is the element integrals, which usually consumes much time when the FMM for N -body problems is directly used. In order to accelerate element integrals, we present an adaptive FMBEM with a particular dual-information tree structure which contains both node and element information, and use it for 3D elasticity in this paper. In our adaptive FMBEM, the Multipole Expansions (ME), Moment-to-Local (M2L) translation, Local Expansions (LE), and the Near Field Direct Computation (NFDC) are level independent so that they are suitable for parallel computing. The examples show that the time of ME and NFDC in our FMBEM is almost 1/3 and 1/2 compared with that in a node-based FMBEM which deals with FMBEM in a particle interaction mode. We develop two GPU parallel strategies to accelerate the processes of ME, M2L and NFDC and implement them on a NVIDIA GTX 285 GPU, and the speedups to an Intel Core2 Q9550 CPU using 4 cores can reach 10.7 for ME, 16.2 for M2L, and 3.6 for NFDC.