基于CUDA并行技术加速2维矩阵MOC方法的研究  

作　　者：郑勇 彭敏俊[2] 安萍 强胜龙[1] 芦韡 

机构地区：[1]中国核动力研究设计院核反应堆系统设计技术重点实验室,四川成都610213 [2]哈尔滨工程大学,黑龙江哈尔滨150001

出　　处：《科技视界》2021年第21期99-102,共4页Science & Technology Vision

摘　　要：在中子输运理论中,矩阵形式的特征线方法因减少了射线扫描次数,相较于它的标准形式在理论上具有更优良的性能。在以前的工作中,经预条件后的极小残余算法(PGMRES)用于求解得到的矩阵方程。为了进一步加速迭代过程,文章采用了CUDA并行机制实现了并行版的PGMRES算法。采用合并访存和共享内存的方式对稀疏矩阵-向量乘(SpMV)操作进行了并行优化,极大地提高了并行计算性能。基于对矩阵方程系数矩阵的分析,提出了3种并行优化策略来加速矩阵方程的求解,并且根据系数矩阵行向量尺寸的不同而采取不同的加速方案,以此提升SpMV运算的数据吞吐率。为验证加速策略的有效性,对2维C5G7基准题进行了计算和结果讨论。数值计算表明,在C5G7问题中,无论是否采用CMFD方案,并行策略都是有效的;在最优的并行策略中,问题得到的最高加速比为5.5。Due to the reduction of sweeping operations,the method of characteristics with matrix form has more favorable features contrasting to its standard form.In previous study,the preconditioned GMRES algorithm was used to solve the resulting matrix equations.In order to further accelerate the iteration process,the parallel version of the PGMRES algorithm is implemented using CUDA parallelism.The parallel optimization of sparse matrix-vector multiplication(SpMV)operations is implemented by combining memory access and shared memory,which greatly improves the parallel computing performance.After analysizing of coefficient matrix,three parallel optimization strategies are proposed to speed up the solution of the matrix equations,and different acceleration schemes are adopted according to the different vector sizes of the coefficient matrix rows to increase the data throughput rate of SpMV operations.In order to verify the effectiveness of the acceleration strategy,two-dimensional C5G7 benchmark questions was calculated and discussed.Numerical calculations show that in the C5G7 problem,the parallel strategy is effective whether or not the CMFD scheme is adopted;in the optimal parallel strategy,the highest speedup ratio of the two problems is 5.5.

关 键 词：特征线方法 CUDA并行技术 预条件GMRES算法 SpMV优化 

分 类 号：TL331[核科学技术—核技术及应用]