基于多线程多GPU并行加速的最小二乘逆时偏移算法

最小二乘逆时偏移算法可对地下复杂构造精确成像,但由于计算量大,目前仍难以在实际资料处理中广泛推广应用,因此研究该方法的高效计算策略具有重要意义。结合Pthread标准,提出了多线程多图形处理器(Graphics Processing Unit,GPU)并行加速策略,在共炮点道集域分解计算任务,由多GPU并行计算并实时更新数据;并结合GPU存储器优化方法,调用GPU端共享存储和寄存器等高速存储器,提高波场模拟的计算效率;最终实现了二维空间的时域最小二乘逆时偏移算法大幅加速计算。分别对Marmousi2截断模型和Marmousi模型进行加速成像测试,结果表明:基于多线程多GPU并行加速的最小二乘逆时偏移算法具有普适性;随着数据规模的增加,该方法的加速效率可逐渐逼近线性加速,数据同步延迟小,加速效率显著。

 Least-squares reverse time migration (LSRTM) can accurately obtain images of subsurface structures.However,its application to actual data can be challenging owing to the large amount of calculation required.Therefore,a parallel acceleration strategy that combines Pthread standard with multithread-driven multi-graphics-processing-unit (multi-GPU) is proposed.The method decomposes computing tasks in common-shot gather domain with real-time update of data through multi-GPU parallel computing.The GPU memory optimization method is used to invoke high-speed memory,such as shared memory and registers,to increase the computational efficiency of wave field modeling.Finally,acceleration of 2D time-domain LSRTM is realized.The method was tested using the Marmousi2 truncation and Marmousi synthetic data.The results showed that the proposed method is applicable to different types of data.With the increase in data scale,the acceleration efficiency can gradually approximate linear acceleration,and is enhanced with small data synchronization delay.

国家自然科学基金项目(41574117,41474118,41804133),黑龙江省杰出青年科学基金项目(JC2016006)共同资助。