| The time-domain finite integration method(TDFIT)has become an important calculation method for electromagnetic field simulation analysis.In recent years,computing large-scale electromagnetic problems or complex electromagnetic problems not only consumes a lot of computing resources,but also requires a lot of computing time.The time-domain finite integration method(TDFIT)has a natural parallel advantage,so adopting parallel computing technology is an effective way to improve the computational efficiency of the algorithm.Facing the international situation that is constrained by nondomestic GPU hardware,it is very necessary to study the parallel algorithm of finite integral in time domain based on domestic DCU,and provide a parallel computing solution for finite integral algorithm in time domain,making it a low-cost and efficient algorithm.The domestic accelerator DCU(Deep Computing Unit)integrates the concept of data stream parallel processing into scientific computing,which can process large amounts of data more efficiently.Due to the rapid development of computer technology,domestic DCUs are not much different from CPUs in price,but their computing power is dozens or even hundreds of times higher than that of CPUs.The HIP parallel programming model is a technology provided for realizing parallel programming based on DCU hardware.The time-domain finite integration method uses DCU to accelerate the calculation of a large amount of data in space,which can improve the performance of the program by dozens or even hundreds of times.This thesis mainly introduces the programming model and organizational structure of HIP through hardware mapping,software system,storage system,thread level and the relationship between host and device.Use the HIP programming model to perform singleDCU acceleration and multi-DCU acceleration for the two-dimensional finite-difference time-domain method,and perform single-DCU acceleration for the three-dimensional time-domain finite integration method,and simulate large-scale model block heterogeneous calculations based on multiple computing nodes.This article will specifically show its algorithm flow and some related codes to achieve acceleration.Through the numerical simulation results of the model,the running time and speedup ratio are obtained,which proves that,on a larger scale,using the DCU to accelerate the two-dimensional time-domain finite difference algorithm and the three-dimensional time-domain finite integration algorithm,compared with the serial algorithm,its calculation The speed generally reaches more than 20 times.The two-dimensional finitedifference time-domain method simulates the transmission line,and the threedimensional time-domain finite integration method simulates the dipole antenna and Mur absorption boundary.By comparing the simulation speed of single DCU and multi-DCU,a more cost-effective method can be selected for simulation. |
