时域有限差分法中的GPU加速高效CPML方案

发布时间：2018-01-10 17:23

本文关键词：时域有限差分法中的GPU加速高效CPML方案　出处：《西安电子科技大学学报》2015年01期 　论文类型：期刊论文

【摘要】：针对并行CPML存在的计算冗余和访问冗余问题,提出了一种用于时域有限差分法的图形处理器加速无除法联合最小访存CPML更新方案.该方案通过重新安排CPML迭代公式,将除法操作吸收进公式的固定系数中,消去了图形处理器计算中负担繁重的除法操作.该方案进一步通过合并PML区域内时域有限差分法常规场值更新步骤和CPML更新步骤,剔除了这两个步骤中的重复访存,使算法的访存需求最小化.数值验证结果表明,在同等精度下,CPML更新过程和PML区域场值整体计算过程分别减少了70%和44%的计算时间.
[Abstract]:Aiming at the problem of computational redundancy and access redundancy in parallel CPML. In this paper, a new scheme of accelerated non-division and minimum memory access CPML updating for graphics processor for finite-difference time-domain (FDTD) method is proposed. The scheme reschedule the CPML iterative formula by rescheduling the CPML iteration formula. The division operation is absorbed into the fixed coefficient of the formula. It eliminates the burdensome division operation in the computation of graphics processor. The scheme further combines the conventional field value updating step and the CPML updating step of the finite-difference time-domain method in the PML region. The repeated memory access in these two steps is eliminated, which minimizes the memory access requirement of the algorithm. The numerical results show that the same precision is achieved. The CPML update process and the overall calculation process of the PML field value are reduced by 70% and 44%, respectively.
【作者单位】：西安电子科技大学电子工程学院;
【基金】：国家自然科学基金资助项目(30870577,61301288) 中央高校基本科研业务费资助项目(JB140218,K5051302057)
【分类号】：TP332
【正文快照】： 时域有限差分法(Finite Difference Time Domain,FDTD)作为主流电磁计算数值方法,广泛应用于波传播、电磁散射、天线辐射和微波电路等领域[1-3].时域有限差分法处理散射和辐射问题时,需要在计算区域的边界上设置吸收边界条件(Absorbing Boundary Condition,ABC).吸收边界条件

【参考文献】