【摘要】 随着GPU在高性能计算领域更多地用于科学计算,采用GPU技术对大型稀疏线性方程组进行计算,从而满足人们对计算速度和计算精度要求的提高。NVIDIA Fermi架构的开发,大大提升了GPU的双精度浮点运算能力。拟极小残差法(QMR)作为高性能计算领域中的重要迭代算法,基于求解稀疏代数方程组对ELL算法进行GPU优化。通过对不同规模线性方程组计算分析表明,QMR-GPU的性能提升为原始QMR的3.5倍,与传统的BICG法相比,QMR并行算法具有速度和存储优势,可获得良好的并行加速比。更多还原

【Abstract】 As the GPU in high-performance computing is more used in scientific computing,GPU technology can be used for 3D electromagnetic problems in the large-scale sparse linear equations. The NVIDIA architectureFermi significantly improves floating point computation performance in double-precision. It is a necessity to optimize Minimum residual method( ELL) on GPU,which is an important iterative algorithm in high performance computing area. ELL solution of sparse linear equationson GPU is optimized,and ELL-GPU performance has 3. 5X improvement compared with implemented on CPU. Analysis of the calculations for linear equation sets of different sizes shows that,with better parallel speedup ratio,QMR parallel algorithm is superior to the traditional BICG method in speed and memory.更多还原