Acceleration of asymptotic computational electromagnetics physical optics — shooting and bouncing ray (PO-SBR) method using CUDA

Abstract

The objective of this research is to accelerate the Physical Optics - Shooting and Bouncing Ray (PO-SBR), an asymptotic computational electromagnetics (CEM) method, on the recently emerged general purpose graphics processing unit (GPGPU) using NVIDIA???s CUDA environment. In modern engineering, simulation programs are used to aid the development of advanced devices, and this is where CEM plays the important role of simulating the propagation of electromagnetic (EM) waves and fields using mod-ern computers. In this thesis, CUDA on NVIDIA???s GPU is used to accelerate the PO-SBR method, which greatly reduces the computational time required for various problems.
\n\tStarting with the theoretical background, we introduce the PO-SBR method, includ-ing the ray tracing and the electromagnetic aspects of the method. Next, we discuss its implementation using the standard CPU C++ language and point out the computationally parallel nature of the method. NVIDIA GPU???s hardware architecture is then described to show the portability of the method onto GPU devices. Then, NVIDIA???s GPU pro-gramming environment, CUDA, is introduced for the implementation of the part of the method to be parallelized. Finally, this thesis presents a novel and flexible method of implementation which fully exploits the hardware architecture of the GPU devices, while at the same time remaining flexible and intelligent enough to be able to optimize itself even on different NVIDIA GPU hardware platforms. The acceleration reaches more than 50 times speedup as compared to the traditional CPU version of the code, and it is believed that a higher speedup can still be achieved with problems of increasing com-plexity.