基于CUDA的地震信号频率补偿并行算法

基于压缩感知的地震信号频率补偿算法可有效拓宽地震信号的频谱，提高地震资料的分辨率。虽然该算法具有良好的拓频效果，但对于高维度、大规模的地震数据时效较低。经过分析发现该算法的计算瓶颈在于计算反射系数部分的大量代数运算和重构信号部分的卷积运算，为此，提出一种基于CUDA的并行方案对该算法进行并行优化。首先，改变地震数据的组织形式，使其存取效率更高，且更适合并行处理；然后，重新设计计算反射系数串行代码，利用CUDA(Compute Unified Device Architecture)平台调用GPU大量轻量级线程对其中的代数运算进行并行化；最后，利用卷积定理改变了时域信号卷积计算方式，采用cufft库函数将两个时域信号的卷积转换到频域进行计算。结果表明，在保证计算精度的前提下，与串行算法相比，并行算法在PC端获得了4倍以上的总体加速比。

Parallel algorithm of seismic signal frequency compensation based on CUDA

The frequency compensation algorithm of seismic signals based on compressed sensing can effectively broaden the spectrum of seismic signals and improve the resolution of seismic data. Although the algorithm has a good frequency-broadening effect, it has low time efficiency in the face of high-dimensional and large-scale seismic data. Analysis shows that the bottleneck of the algorithm lies in massive algebraic operations for the reflection coefficient and convolution operations in signal reconstruction. Therefore, a parallel scheme based on CUDA is proposed for parallel optimization of the algorithm. Firstly, the organization form of seismic data is changed to make it more efficient and more suitable for parallel processing. Then, the serial code for computing the reflection coefficient is reconstructed, and a large number of lightweight threads of GPU are called by CUDA to parallelize the algebraic operations. Finally, the convolution calculation method of time-domain signals is changed by the convolution theorem, and the convolution operation of two time-domain signals is converted to the frequency domain by the cufft library function. The results reveal that the parallel algorithm achieves four times the overall speedup of the serial algorithm on the PC side on the pre-mise of ensuring computational accuracy.