期刊名称：ELCVIA: electronic letters on computer vision and image analysis
印刷版ISSN：1577-5097
出版年度：2013
卷号：12
期号：1
页码：1-16
DOI：10.5565/rev/elcvia.508
语种：English
出版社：Centre de Visió per Computador
摘要：This paper provides an extensive analysis concerning runtime, accuracy and noise of High-Performance Computing (HPC) frameworks for Computed Tomography (CT) reconstruction tasks: "conventional" multi-core, multi threaded CPUs, the Compute Unified Device Architecture (CUDA) on GPUs, and the graphics pipeline of GPUs as facilitated by the DirectX or OpenGL programming interfaces, exploiting various built-in hardwired features like rasterization and texture filtering. We compare implementations of the Filtered Back-Projection (FBP) algorithm with fan-beam geometry on all these HPC frameworks. Specifically, an ACR-accredited phantom is reconstructed from the raw attenuation data acquired by a clinical CT scanner. Our analysis shows that a single GPU can run the FBP algorithm for reconstructing a 1024 x 1024 image considerably faster than a 64-core, multi-threaded CPU machine. Moreover, employing the graphics pipeline further increases performance as compared to CUDA, albeit with slightly lower accuracy due to "fast math" operations.
其他摘要：This paper provides an extensive analysis concerning runtime, accuracy and noise of High-Performance Computing (HPC) frameworks for Computed Tomography (CT) reconstruction tasks: "conventional" multi-core, multi threaded CPUs, the Compute Unified Device Architecture (CUDA) on GPUs, and the graphics pipeline of GPUs as facilitated by the DirectX or OpenGL programming interfaces, exploiting various built-in hardwired features like rasterization and texture filtering. We compare implementations of the Filtered Back-Projection (FBP) algorithm with fan-beam geometry on all these HPC frameworks. Specifically, an ACR-accredited phantom is reconstructed from the raw attenuation data acquired by a clinical CT scanner. Our analysis shows that a single GPU can run the FBP algorithm for reconstructing a 1024 x 1024 image considerably faster than a 64-core, multi-threaded CPU machine. Moreover, employing the graphics pipeline further increases performance as compared to CUDA, albeit with slightly lower accuracy due to "fast math" operations.
