基于GPU的并行APSP问题的研究

摘要

Floyd—Warshall算法是图论中APSP（All—Pair Shortest Paths）问题的经典算法，为了加快计算速度，提出使用GPU通用计算来实现。文章先从算法的原理入手，层层深入，提出了可以在GPU上运行的并行F—W算法。之后，又根据矩阵分块的原理和GPU共享存储器的使用，实现了改进的GPU并行F—W算法。通过大量测试实验，得到了该GPU并行程序相对于传统CPU并行程序产生超过百倍的加速比的结论。%How to use GPU-based Floyd-Warshall algorithm to deal with the APSP （All-Pair Shortest Paths） problem in graph theory is introduced. First, on the basis of the principle of F-W algorithm, its parallelized version is put forwards on GPU. Then, according to the matrix segmentation and the exploit of the shared memory in GPU, an improved parallel version of F-W algorithm on GPU is introdctced. At last, we make a comprehensive comparison and analysis of these algorithms, the speedup is over 100x.