基于GPU并行加速的探地雷达伪谱法三维正演模拟

随着探地雷达问题复杂度的加大以及对解释精度要求的提高,探地雷达三维正演模拟越来越重要,对高性能计算的需求也越来越大.常规三维正演模拟主要采用时域有限差分法开展,并且随着计算机技术的发展,GPU加速已经成为一种三维正演主流加速方法.但是GPU本身的显存相比CPU能利用的内存而言非常有限,要充分利用其加速性能,就需要尽量压...

A Three-dimensional Forward Modeling with Pseudo-spectral Method for Ground Penetrating Radar Accelerated by GPU

With the increasing complexity of ground penetrating radar(GPR)problems and the improvement of interpretation accuracy requirements,the three-dimensional forward modeling of GPR is becoming more and more important,and the demand for high-performance computing is also growing.Conventional 3D forward simulations are mainly carried out using finite-difference time-domain method(FDTD),and with the development of computer technology,GPU acceleration has become a mainstream computing acceleration method applied to it.However,the GPU′s own memory is very limited compared to the memory available to CPU.To make full use of GPU′s acceleration performance,it is necessary to compress the storage requirements as much as possible.Compared with FDTD,the pseudo-spectral time domain(PSTD)method is a wave equation forward modeling method with very high accuracy in space.Theoretically,only 2 spatial sampling points per wavelength are required to achieve a good accuracy.Due to this advantage,it can adopt large-size grids to compress memory requirements a lot,making using of the spatial characteristic of high accuracy.Therefore,this paper applies the PSTD method based on GPU acceleration to the three-dimensional forward modeling of GPR to compress its memory requirements,and makes full use of the high-performance floating-point computing capabilities of GPU for acceleration.The researches in this paper show that,compared with FDTD,the PSTD method can compress the memory demand by at least 8 times,so that the original problem that GPU acceleration cannot be used due to excessive memory demand can be solved and GPU acceleration performance can be fully used,which greatly improves computing efficiency and promotes the development of 3D reverse time migration imaging and full waveform inversion.