基于CPU+GPU联合计算真地表叠前时间偏移实用化研究  

Kirchhoff pre-stack time migration of complex surface data on CPU+GPU

作　　者：马召贵[1,2] 赵改善[1] 武港山[2] 许自龙[1] 段文超[1] 王海波[1] 

机构地区：[1]中国石化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：《石油地球物理勘探》2014年第3期530-537,417,共8页Oil Geophysical Prospecting

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

摘　　要：为了提高起伏地表条件下的基尔霍夫叠前时间偏移算法的计算效率,针对CPU+GPU异构计算平台开展了算法移植与优化研究。首先分析了起伏地表条件下提高偏移成像精度的反假频、弯曲射线旅行时计算以及真地表旅行时校正的处理方法,然后在对算法的并行计算特征进行分析的基础上,针对CPU+GPU异构平台的算法移植进行了多级并行联合计算架构的设计,通过炮检距域的多进程数据域并行、地震数据I/O与偏移计算的异步并行、基于CUDA的超大规模线程成像域并行以及联合CPU计算的多线程成像域并行技术对算法进行了移植及优化。利用大规模测试数据集进行了计算效率测试,测试结果验证了多级并行联合计算架构以及分别针对CPU和GPU平台的算法优化技术能够极大地提高偏移处理效率。This paper discusses surface topography Kirchhoff pre-stack time migration of data in tight topography areas on CPU+GPU platform and its efficiency.We first analyze anti-aliasing,bending ray travel time calculation with high order,travel time corrections from the true surface,and the parallel framework of the original algorithms.Then we design a new parallel framework which is suitable to CPU+GPU platform.Offset domain calculations are conducted by multi-process parallelism,and seismic data I/O are carried out by asynchronous parallelism.Based on multi-thread parallelism with CPU and massively threads parallelism with CUDA in imaging domain,the original CPU algorithms are transplanted and optimized.Test results with large scale seismic data show that Kirchhoff pre-stack time migration calculations can be carried out much faster on the multi-level parallel framework based on CPU+GPU platform.

关 键 词：叠前时间偏移 起伏地表 高性能计算 并行效率 GPU移植 CUDA优化 

分 类 号：P631[天文地球—地质矿产勘探]