摘要
针对高斯牛顿法地震全波形反演计算量大、计算速度慢的问题,采用图形处理器（GPU）对其加速.高斯牛顿法全波形反演耗时主要集中在波形正演模拟和矩阵乘法计算两个方面,而波形正演算法和矩阵乘法计算在算法特性上都满足并行性的要求.对于波形正演模拟的加速,研究并实现了基于CUDA平台的时域有限差分（FDTD）正演算法.对于矩阵乘法的加速,直接使用计算能力很强的CUB-LAS 库来完成计算.在台式PC 上对不同模型大小的反演区域做合成数据反演,所用显卡型号为GTX650ti,程序速度提升10-30倍,且随着模型增大,程序的加速比将进一步提高.二维Overthrust截取模型反演算例表明时间成本已经不再是影响高斯牛顿法全波形反演发展的主要问题.
The Gauss-Newton method for seismic full waveform inversion is extensive computational and time-consuming. The fast parallel platform of CUDA is applied to speed up the program on graphics processing unit (GPU). The time consuming parts of Gauss-Newton method full waveform inversion are waveform forward modeling and matrix multiplication, and they all meet the requirements of parallelism in the algorithm. For the acceleration of the wave forward modeling, we study and implement the finite-difference time-domain (FDTD) method based on CUDA platform; and for matrix multiplication, CUBLAS library with strong ability of calculation is directly used. Implementing the code of different model size on Personal Computer (PC) with GTX650ti GPU to test the speedups, the test shows that the GPU-based code is 10-30 times faster than the CPU-based code and it will perform faster when the model size is bigger. Numerical test of the Overthrust velocity model indicates the time cost is never a question to Guass-Newton method full waveform inversion.
出处
《科技通报》
北大核心
2016年第4期6-10,26,共6页
Bulletin of Science and Technology
基金
大型油气田及煤层气开发国家科技重大专项课题(2011ZX05003-003)
