Supercomputing and Algorithms group, University of Almería, Almería, Spain.
PLoS One. 2012;7(11):e48261. doi: 10.1371/journal.pone.0048261. Epub 2012 Nov 6.
Tomography allows elucidation of the three-dimensional structure of an object from a set of projection images. In life sciences, electron microscope tomography is providing invaluable information about the cell structure at a resolution of a few nanometres. Here, large images are required to combine wide fields of view with high resolution requirements. The computational complexity of the algorithms along with the large image size then turns tomographic reconstruction into a computationally demanding problem. Traditionally, high-performance computing techniques have been applied to cope with such demands on supercomputers, distributed systems and computer clusters. In the last few years, the trend has turned towards graphics processing units (GPUs). Here we present a detailed description and a thorough evaluation of an alternative approach that relies on exploitation of the power available in modern multicore computers. The combination of single-core code optimization, vector processing, multithreading and efficient disk I/O operations succeeds in providing fast tomographic reconstructions on standard computers. The approach turns out to be competitive with the fastest GPU-based solutions thus far.
断层摄影术可以从一组投影图像中阐明物体的三维结构。在生命科学中，电子显微镜断层摄影术以几纳米的分辨率提供了关于细胞结构的宝贵信息。在这里，需要大的图像来将宽视场与高分辨率要求结合起来。然后，算法的计算复杂性以及大的图像尺寸将断层重建转化为计算上的难题。传统上，已经应用高性能计算技术来应对超级计算机、分布式系统和计算机群集上的此类需求。在过去几年中，趋势已经转向图形处理单元 (GPU)。在这里，我们详细描述并彻底评估了一种替代方法，该方法依赖于利用现代多核计算机提供的功能。通过对单核代码优化、向量处理、多线程和有效的磁盘 I/O 操作的结合，成功地在标准计算机上实现了快速断层重建。该方法的结果与迄今为止最快的基于 GPU 的解决方案具有竞争力。
