Graphics Processing Units Implementation for Fast Algorithms Solving Electromagnetic Integral Equations

摘要

Two important components for solving fast algorithm implementations for integral equation (IE) methods in electromagnetic problems are addressed, including dense and sparse matrix-vector multiplications. For dense multiplications we present a box adaptive integral method (B-AIM) with O(N log N) computational complexity, which is implemented on massively parallel systems with single and multiple Graphics Processing Units (GPUs). We also present single- and multi-GPU implementations of sparse matrix-vector multiplication (SpMVM). Our timing tests show a 100-300x speed-up comparing to CPU implementations for B-AIM and 42x for SpMVM on a single GPU. For multi-GPU implementations the parallelization efficiency between 65% and 85% is achieved on systems with 4 GPUs.