基于GPU并行计算的X射线动态显微CT的快速重构

Fast reconstruction of X-ray dynamic micro-CT based on GPU parallel computing

上海光源X射线成像线站的X射线动态显微计算机断层扫描(Computed Tomography,CT)的用户实验产生了海量投影数据,采用基于CPU串行计算的CT重构方法速度慢,将导致实验数据的大量积压。图形处理器(Graphics Processing Unit,GPU)因其多核架构、可大规模并行化的特点,在高性能计算领域得到广泛应用,将其用于CT重构,有望实现X射线动态显微CT的快速重构。采用CUDA编程标准,将滤波反投影(Filtered Back Projection,FBP)CT重构算法的反投影运算进行并行化设计,并在NVDIA RTX2080的GPU上成功实现了CT重构的并行计算。与基于CPU串行计算的CT重构时间的对比结果表明:基于GPU并行计算的CT重构获得了200倍左右的加速比,一套CT数据所需的重构时间由分钟级降低到了秒级。借助GPU强大的计算能力可以实现X射线动态显微CT的快速重构。

[Background]Massive projection data are produced by user experiments with the X-ray dynamic microcomputed tomography(CT)method on the X-ray imaging beamline at Shanghai synchrotron radiation facility(SSRF).CT reconstruction method based on central processing unit(CPU)serial computing is slow,which leads to a large backlog of experimental data.Graphics processing unit(GPU)is widely used in the field of high-performance computing because of its multi-core architecture and large-scale parallelization.[Purpose]This study aims to speed up CT reconstruction by making use of GPU parallel computation for X-ray dynamic micro-CT on the X-ray imaging beamline at SSRF.[Methods]According to computer unified device architecture(CUDA)programming standard,the back projection process of the filtered back projection(FBP)CT reconstruction algorithm was parallelized,and fast reconstruction of X-ray dynamic micro-CT was implemented on GPU of NVDIA RTX2080.CT reconstruction time was compared with that of CPU-based reconstruction method.[Results]The comparison of CT reconstruction time shows that the CT reconstruction based on GPU parallel computing achieves about 200 times speedup compared with the CT reconstruction based on CPU serial computing,and the reconstruction time of a set of CT projection data is reduced from minutes to seconds.[Conclusions]With the powerful computing power of GPU,the fast reconstruction of X-ray dynamic micro-CT can be realized.