论文部分内容阅读
本文介绍地震波场高阶有限差分正演及其GPU计算问题,通过数值模拟技术实现地震波正演。对于声波方程,利用泰勒级数展开式得出波动方程的高阶有限差分格式及其离散表达式。运用C++语言和CUDA编写二维和三维GPU正演程序,使用共享存储器提升GPU线程间通信传输速度,并且改善了三维模型情况下共享存储器容量对有限差分阶数的限制问题。建立不同尺度模型针对二维和三维GPU正演程序和CPU正演程序进行计算测试,比较两个程序的计算效率。测试结果表明,无论是在二维和三维的模型下,GPU正演程序的计算耗时都远远小于CPU正演程序的计算耗时,且随着计算数据量的增大,加速效果越来越显著,测试结果可以很好地证明GPU程序相对于单CPU程序计算的高效性。
This paper introduces the high-order finite difference forward modeling of seismic wave field and its GPU computing problem, and realizes seismic wave forward by numerical simulation. For the acoustic wave equation, the higher-order finite difference scheme and its discrete expression of the wave equation are derived by Taylor series expansion. Using C ++ language and CUDA to write two-dimensional and three-dimensional GPU forward program, using shared memory to enhance the communication speed between GPU threads, and to improve the three-dimensional model of shared memory capacity limited finite difference order. Different scale models are established to calculate and test 2D forward and forward CPU forwards and compare the computational efficiency of the two programs. The test results show that, in both 2D and 3D models, the computation time of GPU forward programs is much less than that of CPU forward programs, and as the amount of computational data increases, the acceleration effect is more and more The more significant, the test results can be very good proof of GPU program relative to the single-CPU program to calculate the efficiency.