基于GPU实现叠前时间偏移走时计算的并行算法

摘要

走时计算是叠前时间偏移计算中最耗时的部分,通过分析传统的串行走时算法,发现静态8点插值算法非常适合在GPU上运行.首先利用CUDA技术对静态8点插值算法进行并行化改造,设计静态8点并行插值算法,然后测试其正确性,统计其相对误差情况.实验表明此算法比工业生产上的动态插值算法更准确,最后我们利用体偏作性能测试.试验结果表明,运行在GPU上的静态8点并行插值算法内核性能是运行在CPU上的动态插值算法内核的22.76倍.这说明,静态8点并行插值算法适合进行走时计算,并且可以应用于工业生产上.%The computation of traveling time of Pre-Stack Time Migration (PSTM) is the most time-consuming part of whole procedure. In this paper, we present a novel parallel algorithm to low the time of traveling-time computation. By using static interpolation with step of 8 points instead of dynamic step length, the new designed algorithm performs more accuracy and performance. Experimental results show that the speed-up number reaches 22.76 times. The statistical relative error demonstrates the better accuracy comparing with the commercial codebases.