Optimizing Algorithm of Sparse Linear Systems on GPU
Dongxu Yan, Haijun Cao, Xiaoshe Dong, Bao Zhang, Xingjun Zhang · 2011
Linear equations with large spare coefficient matrices arise in many practical scientific and engineering problems. Previous sparse matrix algorithms for solving linear equations based on single-core CPU are highly complex and time-consuming. To solve such problems, aiming at Jacobi iteration algorithm, in this paper we firstly implement a sparse matrix parallel iteration algorithm on a hybrid multi-core parallel system consisting of CPU and GPU, then an optimization scheme is proposed to carry out performance improvement in two ways, i.e., the multi-level storage structure and the memory access mode of CUDA. Experimental results show that the parallel algorithm on hybrid multi-core system can gain higher performance than the original linear Jacobi iteration algorithm on CPU. In addition, the optimization scheme is effective and feasible.