准对角矩阵与向量相乘在CPU+GPU异构集群上的实现与优化

Implementation and Optimization of Quasi Diagonal Matrix-Vector Multiplication on CPU +GPU Heterogeneous Cluster

稀疏矩阵与向量相乘(Sp MV)是科学计算和工程应用中一个重要问题,而且非常适宜进行并行计算,目前在GPU对Sp M V的实现和优化是一个研究热点.针对准对角矩阵存在的一些不规则性,采用CSR+DLA混合存储格式来进行Sp M V计算,能够提高压缩的效果.为了发挥CPU多核的并行计算能力,采用一种CPU+GPU混合计算模式,这样可以把混合存储格式不同格式的数据分割到CPU和GPU上,从而提高了资源的利用效能.本文另外还在分析CPU+GPU异构计算模式的特征基础上,提出一些优化策略,能够改进准对角矩阵与向量相乘在异构计算环境中的计算性能.

Sparse matrix vector multiplication （SpMV ） is an important issue in scientific computing and engineering applications, andis very suitable for parallel computing, the GPU implementation and optimization of SpMV is a research hotspot. In this paper we focuson a special SpMV, sparse quasi-diagonal matrix vector multiplication, which has irregular nonzero data distribution. We present a hy-brid diagonal storage format （hybrid DIA and CSR ） to get higher compression ratio than the DIA and CSR. It is possible to split thedata to the CPU and GPU on CPU ＋ GPU hybrid computing model to take full advantage of both CPU and GPU computing resourcesand be able to play the CPU and GPU computing features, which can enhance the resource utilization efficiency. Moreover based on a-nalysis of the characteristics of the CPU ＋ GPU heterogeneous computing model,proposed some optimization strategies can improveSQMVM performance computing in a heterogeneous computing environment.