基于国产异构平台的光前哈密顿量对角化优化OA

To address the performance bottleneck encountered by the basis light-front quantization(BLFQ)method when solving the eigenvalue problems of large-scale sparse Hamiltonian matrices,this paper proposes an optimization scheme based on a domestic heterogeneous computing platform.This scheme employs a heterogeneous computing architecture that combines CPUs and GPU-like accelerator cards,utilizing message passing interface(MPI)for inter-process communication,offloading core com-putational tasks to the accelerator cards for execution.By porting the relevant interfaces of the AR-PACK library to make it compatible with the accelerator card environment,designs a parallel algorithm for sparse matrix-vector multiplication.The key improvements include replacing the original CPU-side linear algebra functions with accelerator-optimized functions,utilizing NVIDIA collective communication library(NCCL)for inter-accelerator communication,and adopting strategies such as thread coalescing memory access and load balancing to optimize parallel performance.Experimental results demonstrate that the optimization scheme achieves significant speedups across various test scales,with acceleration ratios ranging from 3.7 to 10.3.Additionally,communication overhead is significantly reduced,and load distribution becomes more balanced.The proposed scheme effectively enhances the performance of BLFQ in hadron physics computations and provides valuable insights for scientific computing applications involving large-scale sparse matrices in other domains.