A Resesrch Of EM-PIC's GPU Parallel Methods Based On CUDA

In recent years,with the rapid development of high-performance heterogeneous computing,the multi-core heterogeneous parallel computing technology represented by GPU has been widely used in scientific research,its unique organizational structure,powerful floating-point computing capability and huge data handling capability broken through the limitations of the traditional CPU on Moore's Law.In the field of electromagnetic particle-in-cellsimulation,the time domain finite difference method has highly parallelism and computational intensiveness,the calculation process of particle computing is huge time consuming.In order to effectively improve the calculation speed and shorten the development cycle of PIC research,this paper useing three-dimensional particles,which is based on the simulation software CHIPIC-3D,focused on transplanting dense electromagnetic field iteration and particle motion calculation to CUDA in particle simulation,and realizes computational acceleration by CPU-GPU heterogeneous paralleling structure.The code migration from serial program to parallel of CUDA process follows the principle of combining theoretical analysis with programming practice.Based on a deep understanding of GPU theory and hardware architecture,CUDA Fortran parallel programming software PGI Visual Fortran compiler is used as a tool to realize parallel computing of one-dimensional FDTD center difference algorithm,which is 10 times compared with single CPU serial program.With deeply understanding the theory and method of GPU parallel computing optimization,we utilize shared memory and constant memory to release more optimize in calculation;We theoretically analyzed the parallelism of FDTD central difference algorithm and time deviation sub-algorithm;We the mapping relationship between the partitioning of 3D mesh and CUDA thread.The parallel computing of these two algorithms is realized on CHIPIC-3D software.The simulation calculation of the portguide microwave model achieves a calculation acceleration of up to 10 times.The simulation calculation of the RBWO achieves a calculation acceleration of up to 11.2 times.The parallelism of the EM-PIC algorithm is theoretically analyzed,and the parallel calculation of the particles is partially realized in the CHIPIC-3D software,which is calculated by RBWO with particle emission,t he calculation acceleration of 4.5 times is obtained.The problems and bottlenecks encountered in the parallel migration process of the program are analyzed.For the problems that have not been solved yet are solved.We give solution the advice for future research.CUDA-based GPU parallel computing is the first application on the particle simulation software CHIPIC-3D with independent intellectual property rights.It has a certain speedup ratio for the simulation of electronic devices with particles.In the future,the particle computing part will be completely transplanted to the GPU for parallel computing to increase the calculation speedup more significantly.