A study on a computational model for GPGPU algorithms and its application to medical image processing

本研究的目的是调查在何种程度上可以加快医疗图像处理在一个单一的PC与GPU，而不是PC集群。为此，我们开发了两个程序的锥束CT（计算机断层扫描）图像重建-一个使用OpenGL/Cg和一个使用CUDA。对于每个512 x 512像素的360个图像，我们重建了512 x 512 x 512体素体积。OpenGL/Cg版本的执行时间为8.3秒，CUDA版本为5.7秒。OpenGL/Cg版本比CPU版本快23.7倍（换句话说，它至少相当于24台PC集群的能力）。因此，我们确认，一个单一的PC与GPU可以提供足够的计算能力，实时操作的CT机。从2008年4月起，岛津公司开始销售使用该算法的CT机。CPU时间长的程序被称为“计算瓶颈”，目前的印象是CPU执行速度是计算的主要瓶颈。然而，即使在这些类型的程序中，也有一个不小的数量具有“内存瓶颈”。基于去年的成功和去年GPGPU性能模型测试的结果，我们相信GPGPU速度的改进也将适用于这种“内存瓶颈”程序。除了锥束CT图像重建，我们还使用CUDA实现了面向生物学的阵列对齐程序，以及图论中所有点之间的最短路径实现。与CPU相比，阵列对齐程序在GPU上的运行时间大大减少，1024长度的查询在GPU上花费约20秒，而在CPU上花费677秒。1024的查询长度适合于氨基酸的工作，我们认为这将有助于药物研究。

The purpose of this research is to investigate to what extent medical image processing can be sped up on a single PC with a GPU as opposed to a PC cluster. For this, we developed two programs for cone-beam CT (computed tomography) image reconstruction-one using OpenGL/Cg and one using CUDA. For 360 images of 512 x 512 pixels each, we reconstructed a 512x512x512 voxel volume. Execution time for the OpenGL/Cg version was 8.3 seconds, and for the CUDA version was 5.7 seconds. The OpenGL/Cg version was 23.7 times faster than the CPU version (in other words, it has equivalent power of at least a 24 PC cluster). Therefore we confirmed that a single PC with GPU can provide enough computing power for real-time operation of a CT machine. Starting in April 2008, a CT machine using this algorithm was being sold by the Shimadzu Corporation.Programs that requires much CPU time are said to have a "calculation bottleneck", and the current impression is that CPU execution speed is the main bottleneck in computation. However, even among these types of programs there is a not-insignificant number which have a "memory bottleneck". Based on last years success and results from the GPGPU performance model tested last year, we are confident that GPGPU speed improvements will also be applicable to such "memory bottleneck" programs.In addition to cone-beam CT image reconstruction, we also used CUDA to implement a biology oriented array alignment program, as well as a shortest-path-between-all-points implementation for graph theory. Run time of the array alignment program was substantially decreased on GPU compared to CPU, with a 1024 length query taking about 20 seconds on the GPU compared to 677 seconds on the CPU. A query length of 1024 is suitable for work with amino acids and we feel this will contribute to pharmaceutical research.