欢迎访问现代地质!

现代地质 ›› 2012, Vol. 26 ›› Issue (6): 1289-1293.

• 油气资源勘查 • 上一篇    下一篇

应用图形处理器快速计算逆时偏移

刘雅宁,刘国峰,张致付   

  1. 中国地质大学 地球物理与信息技术学院,北京100083
  • 收稿日期:2012-06-15 修回日期:2012-08-30 出版日期:2012-10-24 发布日期:2012-10-31
  • 通讯作者: 刘国峰,男,讲师,1981年出生,地球物理学专业,从事地震波成像、反演、高性能计算等研究工作。
  • 作者简介:刘雅宁,女,硕士研究生,1990年出生,地球物理学专业,从事复杂介质地震波传播成像研究。Email:sakuda728@sina.com。
  • 基金资助:

    国家自然科学基金项目(51074138,41104083,41074095);国家专项项目(SinoProbe-01,201011035);中央高校基本科研业务费专项资金资助项目。

Strategies of Accelerating Reverse Time Migration Using Graphic Processing Unit

LIU Ya ning, LIU Guo feng,ZHANG Zhi fu   

  1. School of Geophysics and Information Technology, China University of Geosciences, Beijing100083, China
  • Received:2012-06-15 Revised:2012-08-30 Online:2012-10-24 Published:2012-10-31

摘要:

逆时偏移是目前精度最高的地震数据叠前深度偏移方法,但高强度的计算需求限制了其在工业生产领域的大规模应用。可编程图形处理器的发展为逆时偏移的快速计算提供了一种新的计算选择。围绕如何在图形处理器上开展逆时偏移计算展开,总结了图形处理器计算的优化关键,并根据逆时偏移的特点着重介绍了两个优化环节:一个是应用随机边界条件,以计算换存储,减少数据在主机和图形处理器间的传输;二是应用共享存储器来存储正演计算的波场,相比全局存储器,提高了数据读取的带宽。应用Marmousi模型数据对经过上述优化后的程序进行了测试,结果表明,图形处理器逆时偏移程序得到了很好的优化,提高了计算效率。

关键词: 图形处理器, 逆时偏移, 随机边界, 共享存储器

Abstract:

Reverse-time migration is the most accurate seismic data prestack depth migration method, but the computing needs of its highintensity limit its largescale computing applications. The development of programmable Graphics Processing Unit(GPU) provides an alternative method for rapid calculation of the reverse-time migration. In this paper, we focus on the main steps using GPU to calculate the reverse time migration. We summarize the key points in GPU code optimization, and highlight two optimized components: one is the application of random boundary conditions. The data transmission between the host and GPU is reduced. The second is the application of shared memory to store the forward wave field. Compared to the global memory, data reading band widths increase greatly. Finally, we use the Marmousi data to test our code, and the results show that the reverse-time migration program running on GPU has been well optimized to improve the computational efficiency.

Key words: Graphic Processing Unit, reverse time migration, random boundary condition, share memory