CPU/GPU协同计算在频率域二维全波形反演中的应用  

The Application of CPU/GPU Collaborative Computing in the Frequency Domain 2D Full Waveform Inversion

作　　者：王小青[1,2] 谭捍东[2] 许自龙[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,南京211103 [2]中国地质大学(北京)地球物理与信息技术学院,北京100083

出　　处：《CT理论与应用研究（中英文）》2016年第1期23-32,共10页Computerized Tomography Theory and Applications

基　　金：国家自然科学基金(41374078);国土资源部地质调查项目(12120113086100;12120113101300)

摘　　要：全波形反演利用波场的运动学和动力学信息重建地下物理参数,是建立高精度速度模型的有效手段,巨大的计算量是制约其实用化的瓶颈之一。本文针对全波形反演中频率域正演的复杂计算问题,采用粗细结合的并行策略,将MPI技术应用于多炮间并行计算,同时利用GPU技术加速正演过程中大型稀疏线性代数方程组的求解,以提高频率域全波形反演的计算效率。通过理论模型验证本文方法的正确性和有效性,给出不同数据量与GPU计算效率的相关分析结论,提出频率域全波形反演CPU/GPU协同并行计算的制约瓶颈和发展方向。The full waveform inversion（FWI） uses the kinematic and kinetic information of wave field to reconstruct the physical parameters underground, and it is an effective method to establish the high precision velocity model. A huge amount of calculation is one of the bottleneck of restricting its practicality. In this paper, we use the parallel strategy combined thick and thin to solve the complex computing problem of frequency domain FWI. To improve the computational efficiency of the FWI, the MPI technology was applied to compute multi shot forward parallel, while the GPU technology was used to speed up the solving of large sparse linear algebraic equations. The theoretical model verified the correctness and effectiveness of the method, and the correlation analysis results of different amount of data and GPU calculation efficiency are given. This paper proposed the bottleneck and the development direction of the CPU/GPU cooperative parallel computing in the frequency domain FWI.

关 键 词：频率域正演 稀疏矩阵 BI-CGSTAB CPU/GPU协同计算 

分 类 号：P631.4[天文地球—地质矿产勘探]