高级检索

重力场反演中大型矩阵GPU加速运算的实现

Implementation of GPU-Accelerated Computation of Large Matrix in Gravity Field Inversion

  • 摘要: 针对重力场反演中大型矩阵解算任务量大、解算时间长的问题,提出一种基于多片GPU并行的矩阵运算方法,该方法将GPU和CUDA(Computer Unified Device Architecture)相结合,能实现密集运算的高度并行化,通过加速矩阵乘法和矩阵求逆的运算,极大缩短了行星重力场反演中大型矩阵求逆所需的时间,计算速率为使用CPU计算时的191倍,并且计算的精度较高,反演的误差数量级在10–17。将其应用到重力恢复与内部实验室(Gravity Recovery and Interior Laboratory,GRAIL)月球重力场反演的计算当中,在计算截断大小为50阶次和180阶次的矩阵时,比使用CPU的运算方法时间分别缩短了94.63%、99.51%,并在武汉大学高性能计算平台实现了900阶次矩阵的运算。采用的方法可有效缩短传统计算模型所需的时间,从而有助于建立高阶次、高精度的重力场模型。

     

    Abstract: To address the problem of heavy computational tasks and lengthy processing times in gravity field inversion, a parallel matrix computation method based on multi-GPU integration with CUDA was proposed. This method achieved highly parallel dense computation, significantly reducing the time required for inverting large matrices in planetary gravity field inversion by accelerating matrix multiplication and inversion operations. The computation rate was 191 times faster than that of using CPU. Moreover, it offered high computational accuracy, with inversion precision at the level of 10–17. Applied to the computation of GRAIL lunar gravity field inversion, the proposed method, when computing matrices of truncation orders 50 and 180 respectively reduced the processing time by 94.63% and 99.51% respectively compared to CPU-based methods. Furthermore, the method successfully computed a matrix of 900th order on the High-Performance Computing Platform at Wuhan University. The method employed in this paper can effectively reduce the time needed by traditional computing models, thereby conducive to aiding in the establishment of high-order, high-precision gravity field models.

     

/

返回文章
返回