基于GPU平台的联合迭代重构算法加速

Design and Optimization of Simultaneous Iterative Reconstruction Technique Based on GPU Platform

电子断层三维重构技术(Electron Tomography,ET)是在纳米尺度下研究不具有全同性的细胞或大分子三维结构的重要方法。迭代重构法是ET中重构效果最好的方法,但是其性能较差,重构大尺寸图像时需要数天的时间甚至更长,使其应用受到限制。迭代重构法中经典的方法是代数重构法(Algebraic Reconstruction Technique,ART)和联合迭代重构法(Simultaneous Iterative Reconstruction Technique,SIRT),SIRT算法总是收敛的并且比ART重构的效果更好。利用CUDA语言设计和实现了基于Tesla C1060GPU平台上的并行SIRT重构算法,并利用存储器合并访问、常量存储器、共享存储器等优化技术对并行算法进行优化,优化后的SIRT并行算法在Tesla C1060GPU平台上的最大加速比是Intel i7 920CPU上的串行算法的47倍,并且重构的质量没有任何下降。

Electron tomography（ET） is widely used in reconstructing non-uniform cells or macromolecules in nano scale.One of the best methods of ET is iterative reconstruction due to its outstanding quality of reconstruction,but it is limited by its huge computational requirements.A parallel simultaneous iterative reconstruction technique（SIRT） was designed and implemented based on GPU platform with Tesla C1060 using CUDA programming languages.Experimental results demonstrate the performance of optimized parallel SIRT algorithm.The maximum speedup of the parall el SIRT is 47 times of sequential SIRT approach,and it is not any loss of accuracy.