摘要
The most popular hardware used for parallel depth migration is the PC-Cluster but its application is limited due to large space occupation and high power consumption. In this paper, we introduce a new hardware architecture, based on which the finite difference （FD） wavefield-continuation depth migration can be conducted using the Graphics Processing Unit （GPU） as a CPU coprocessor. We demonstrate the program module and three key optimization steps for implementing FD depth migration： memory, thread structure, and instruction optimizations and consider evaluation methods for the amount of optimization. 2D and 3D models are used to test depth migration on the GPU. The tested results show that the depth migration computational efficiency greatly increased using the general-purpose GPU, increasing by at least 25 times compared to the AMD 2.5 GHz CPU.
复杂介质情况下,地震波延多路径传播,此时基于波动方程延拓的深度成像方法,相对于Kirchhoff方法能够获得更为精确的成像效果,但是,该深度偏移方法由于高昂的计算消耗阻碍了它在生产中的应用。譬如,叠前深度偏移计算需要大规模的计算机集群,占地面积和电能消耗大。本文介绍了应用一种新的GPU计算架构来辅助CPU进行偏移计算。基于新架构的波动方程深度偏移提高了计算效率,而且机器占地面积和电能消耗也大幅度减少。本文以有限差分波动方程深度偏移为例,介绍了其编程模型和程序优化环节,提高了深度偏移计算效率。2D和3D测试表明,与相同单位个AMD2.5GHz CPU计算相比,该架构下的有限差分波动方程叠前深度偏移计算效率提高至少35倍。
基金
supported by the National Natural Science Foundation of China (Nos. 41104083 and 40804024)
Fundamental Research Funds for the Central Universities (No, 2011YYL022)
