Higher-order Fdtd To Solve The Scattering Problem Based On Gpu

Computational electromagnetic is based on the electromagnetic, computationalmathematics and computer science theories. Using the FDTD method to calculate thescattering problems is an important application in the computational electromagneticfield. For the classic FDTD, it has restrictions on the mesh size to suppress thenumerical dispersion. To improve the accuracy, the high-order FDTD method isintroduced. The high-order FDTD method used in this paper uses the symplecticintegrator propagator to complete the time differential, which can achieve4th orderaccuracy. In the space differential, it uses the DSC which has2M order accuracy. Theimproved FDTD is more accurate than the original FDTD by expand the samplingbandwidth when using the small mesh. However, it also takes much more time to do thecalculation.To reduce the computation time used by the higher order FDTD, this paper studieshow to shorten the time by using the parallel technology. As GPU is widely used in thePC and the parallel programming platform is becoming more and more perfect, it isused in the scientific computation a lot. This paper uses the GeForce9600GT as theGPGPU and CUDA as development tool to coding the parallel higher order FDTD.In this paper, first the metal square cylinder’s RCS in TM mode is calculated byusing the (2M,4) FDTD. After verify the correctness of the serial algorithm, the2Dscattering code is transplanted on the GPU. When the bandwidth M=1and M=4, thespeed up ratio are14.5and10.7respectively. If the code is optimized by the on-chipresources the ratio could be raised to15.45when M is4.After the2D calculation, the3D scattering problem is calculated in CPU. Twoclassic models’ RCS are gotten with Huygens principle, and the result is the same withother literature. According to the2D parallel algorithm, the3D FDTD data is divided inthe form of2D, then with the2D threads the3D scattering algorithm is run on the GPU,and finally the RCS is gotten by using the data generated by the GPU. The result showsthat the RCS obtained by the two methods are basically the same. But the speed up ratioin3D is not significant, it only runs5.44times faster than CPU when the number of the calculating points is50×50×50. When the number of the points is extended to84×84×84,the ratio is even smaller, it only runs3.55times faster than CPU.