基于交错网格Fourier伪谱微分矩阵算子的地震波场模拟GPU加速方案  

Accelerating seismic modeling with staggered-grid Fourier Pseudo-spectral differentiation matrix operator method on graphics processing unit

作　　者：龙桂华[1,2] 李小凡[2] 江东辉 

机构地区：[1]中国科学院深圳先进技术研究院,深圳518055 [2]中国科学院地质与地球物理研究所,北京100029 [3]中国石油化工股份有限公司上海海洋油气分公司研究院,上海200120

出　　处：《地球物理学报》2010年第12期2964-2971,共8页Chinese Journal of Geophysics

基　　金：国家高技术研究发展计划(863计划)(2007AA12Z164);国家自然科学基金项目(40874024);广东省自然科学基金项目(10478922035 X00491)联合资助

摘　　要：作为高精度波形反演或逆时偏移的重要组成部分,地震波数值模拟对计算速度和效率提出了更高要求.GPU通用计算技术的产生及其内在数据并行性,为高效地震波数值模拟应用和研究得以有效开展奠定了基础.本文借助交错网格的Fourier伪谱微分矩阵算子和GPU上高效矩阵乘法,实现了复杂介质中地震波模拟的高效算法.数值试验表明,优化后的GPU计算相比CPU单核计算在大规模二维地震波场计算中获得至少100x以上的加速比.这对我们快速分析目标反射层在地震剖面中同相轴位置,制定优化采集方案具有重要意义.As a crucial part of high-resolution waveform inversion and reverse time migration, seismic modeling demands a high efficiency of computation. The general-purpose GPU computing technology provides a fundamental basis as for GPU was naturally designed with data parallelism. By virtue of staggered-grid Fourier pseudo-spectral differentiation matrix operator, solving wave equation can be easily implemented on GPU by matrix multiplication. In large-scale twodimensional wavefield computation, optimized single GPU forward seismic modeling can achieve at least 100x faster than that of single CPU using only one core. The high speedup of GPU/CPU implementation enables us to quickly recognize seismic events of the reflectors of different layers in common shot profile, which is helpful for optimizing the acquisition configuration in field work.

关 键 词：伪谱微分矩阵算子 GPU通用计算 地震波高效正演模拟 

分 类 号：P631[天文地球—地质矿产勘探]