Kirchhoff叠前时间偏移的GPU移植与性能优化技术  

GPU-based porting and optimization of Kirchhoff pre-stack time migration

作　　者：马召贵[1,2] 赵改善[1] 武港山[2] 岳承琪[1] 何恺[1] 王鹏[1] 

机构地区：[1]中国石油化工股份有限公司石油物探技术研究院,江苏南京211103 [2]南京大学软件新技术国家重点实验室,江苏南京210093

出　　处：《石油学报》2014年第4期700-705,共6页Acta Petrolei Sinica

基　　金：国家高技术研究发展计划(863)项目(2009AA01A140)资助

摘　　要：叠前时间偏移在工业生产中发挥着极其重要的作用,为了提高该算法的计算效率,开展了基于GPU异构计算平台的算法移植与优化。首先根据积分法偏移的算法特点制定了偏移距域的多进程数据域并行以及IO与计算异步并行总体并行策略;然后为了提高偏移核心计算部分在GPU上的计算效率,对偏移计算核在GPU上的并行方案进行了分析,选择了成像域超大规模线程并行方案对算法进行了移植和优化,并对不同优化手段在不同GPU硬件平台下获得的性能加速进行了对比测试;最后利用大规模计算节点及大规模地震数据体进行了移植后算法的应用测试,并对算法的计算效率、可扩展性以及精度误差进行了分析。大规模应用测试表明,积分法叠前时间偏移经过GPU移植后可获得较CPU平台近7倍的性能提升,具有很好的工业应用价值。Pre-stack time migration （PSTM） plays an important role in industrial production. To improve its computation efficiency, algorithm porting and optimization are performed based on the GPU heterogeneous computing platform. According to the character-istics of the Kirchhoff algorithm, a parallel strategy is initially designed with the multi-processes parallelism in common offset domain and the asynchronous parallelism of IO and computing. For improving the computation efficiency of migration kernel on GPU, this paper then analyzes the parallel strategy on GPU of migration kernel, performs algorithm porting and optimization with massive scale threads in imaging domain, and assesses the performance of different optimization methods for various GPU hardware platforms. Fi nally, the algorithm testing is carried out with large-scale computing nodes and seismic data followed by an analysis of computation efficiency, scalability, and precision error of the algorithm. Large-scale application test results show that after GPU porting, the performance of Kirchhoff pre-stack time migration is improved by nearly 7 fold compared with the CPU platform, thus having great value for industrial applications.

关 键 词：叠前时间偏移 高性能计算 并行效率 GPU移植 CUDA优化 

分 类 号：P631.443[天文地球—地质矿产勘探]