广义线性纵横波联合反演的GPU加速  

PP wave & PS wave joint inversion based on GPU high performance computation platform

作　　者：崔栋[1] 张玉洁[2] 宋建国[3] 邵丽[4] 

机构地区：[1]中国石油勘探开发研究院,北京100083 [2]中国科学院地质与地球物理研究所,北京100029 [3]中国石油大学(华东),山东青岛266555 [4]新疆油田公司采气一厂,新疆克拉玛依834000

出　　处：《石油地球物理勘探》2015年第4期650-657,3-4,共8页Oil Geophysical Prospecting

摘　　要：文中进行了广义线性反演(GLI)的GPU加速研究,通过理论分析和模型数据检验表明,通过应用L1模稀疏脉冲反褶积,数据质量得到了显著提高。该方法不仅对多波数据适用,对单一纵、横波数据都可以进行反演。反演结果表明:1应用L1模稀疏脉冲反褶积可以在反演前消除子波的效应,从而使GLI在反射系数层面上发挥更好的作用。2将GPU并行计算技术整合到反演过程中,能够极大地提高计算速度和效率,对单个角度域共成像点道集(ADCIG)的加速达数百倍以上;在处理海量的叠前反演数据时,这种优势将更加明显。3在反演中采用并行计算,必须摒弃细粒度的任务设计思路,采用较粗粒度的设计,才能尽量减少延迟,提高计算效率。This paper presents our work on GPU computation acceleration for generalized linear inversion(GLI).Both theoretical analysis and model tests indicate that the data quality has been significantly improved with sparse spike L1 norm deconvolution.The proposed method is not only applicable to multi-wave data,but also to PP data and PS data alone.The following observations are obtained based on inversion results:1L1sparse spike deconvolution can eliminate wavelet effects before inversion,so that the GLI plays a better role in thereflection coefficient levels.2GPU parallel computing technology can greatly improve the computation speed and efficiency. GPU speed may increase hundreds of times for a single ADCIG.Dealing with prestack data,this advantage will be more obvious.3Coarse-grained design would be adopted in parallel computing inversion rather than fine-grained design to minimize delays and improve efficiency.

关 键 词：纵横波联合反演 稀疏脉冲 广义线性反演 GPU 计算效率 

分 类 号：P631[天文地球—地质矿产勘探]