Parallelization Realization Of Electron Tomography On GPU Platform

Electron Tomography (ET) provides the prospects of reconstructing non-uniform cells or macro molecules in nano scale. Iterative reconstruction method is the best way of ET reconstruction quality, but the performance of iterative reconstruction method is very poor and reconstruction of large data requires several days or even longer, so its application is limited. The classical methods of iterative reconstruction method are Algebraic Reconstruction Technique (ART) and Simultaneous Iterative Reconstruction Technique (SIRT), SIRT reconstruction is always convergent and better than ART reconstruction. This paper has designed and realized parallel SIRT reconstruction algorithm on Tesla C1060 GPU platform using CUDA programming languages and has developed ET software-ATOM using Qt.First, this paper has designed and realized the basic parallel SIRT reconstruction algorithm on Tesla C1060 GPU platform using CUDA. Second, we use optimization methods (coalesced global memory access, constant memory, shared memory, stream, etc) to optimize the basic parallel algorithm. Experiment results show optimized parallel SIRT algorithm can speed up 18-50 times over the sequential SIRT approach without and abvious loss of accuracy on Tesla C1060 GPU vs. Intel i7 920 CPU.Finally, we develope the first and free ET software——ATOM 1.0 using Qt and CUDA in China. Reconstructing 2048*2048 pixels data using the old SIRT need a few days, however, ATOM requires only 1-3 hours. ATOM improves greatly the performance of reconstruction, and the machine with a high-end GPU need only a few thousand to tens of thousands yuan, so ATOM reduces greatly the hardware cost. Therefore, ATOM software which is easy, high performance, low cost and portability strong is the first ET software in China, and it has a strong scientific and practical value.