Using Quadruple Precision Arithmetic to Accelerate Krylov Subspace Methods on GPUs

Using Quadruple Precision Arithmetic to Accelerate Krylov Subspace Methods on GPUs
复制标题

使用四精度算术在 GPU 上加速 Krylov 子空间方法

DOI:
10.1007/978-3-642-55224-3_59
复制
发表时间:
2014
期刊:
Parallel Processing and Applied Mathematics-PPAM 2013, Part I, LNCS 8384
影响因子:
--
通讯作者:
Daisuke Takahashi
Daisuke Takahashi
中科院分区:
--
文献类型:
--
作者:
Daichi Mukunoki;Daisuke Takahashi

文献摘要

相似文献

Krylov子空间方法的收敛性受到舍入误差的影响。直到收敛的迭代次数可以通过使用四倍精度算术而不是双精度来减少舍入误差而减少。我们使用四倍精度算法实现了CG和BiCGStab方法,并在NVIDIA Tesla K20X GPU上与标准双精度实现进行了性能比较。我们的研究结果表明,在某些情况下,我们的实现使用四精度算法优于双精度版本。我们将证明,四倍精度算法对于GPU上的CG和BiCGStab方法来说并不昂贵,并且使用四倍精度算法可能是使用预处理的更有效的替代方案。
The convergence of the Krylov subspace methods is affected by round-off errors. The number of iterations until convergence may be decreased by reducing round-off errors through the use of quadruple precision arithmetic instead of double precision. We implemented the CG and BiCGStab methods using quadruple precision arithmetic and compared the performance with the standard double precision implementations on an NVIDIA Tesla K20X GPU. Our results show that in some cases our implementations using quadruple precision arithmetic outperform the double precision versions. We will show that quadruple precision arithmetic is not costly for the CG and BiCGStab methods on GPUs and the use of quadruple precision arithmetic may be a more effective alternative to the use of preconditioning.