A preliminary study of OpenCL for accelerating CT reconstruction and image recognition

In recent years, different high performance techniques/platforms have been researched intensively in CT imaging, especially in CT image reconstruction. Muller and his group achieved lots of significant progress on using NVIDIA's GPU/GPGPU to accelerating CT reconstructions [13]. Kachelriess and his colleagues use Cell Broadband Engine (CBE) to accelerate the reconstruction and got a hyper fast result [4~5]. Recently, Xing et al proved AMD's GPGPU also can been used for accelerating CT reconstruction[6]. The introduction of these high-performance techniques/platforms also brings different concepts and coding environments, which leads to difficulties in software development for practical applications because these environments do not talk to each other and codes are not compatible. For example, codes for acceleration programmed with Nvidia CUDA will not fit on AMD GPU, AMD stream codes can't work on Nvidia GPU too, though they have plenty of common things. Open Computing Language (OpenCL), as an open standard and framework for writing programs that execute across those heterogeneous parallel platforms, makes it possible to solve these problems. In this study, our objective is to implement and test the NVIDIA's and AMD's early OpenCL compatibility implementation release on a back-projection step, the most time consuming part of an FBP reconstruction, and the Haralick's texture feature[7] extraction algorithm for CT images. The paper is arranged as follows. In section II, we will briefly introduce the model of OpenCL. Our implementation for BP and CT image recognition algorithm will be presented in section III. Finally, in section IV, the result and discussion will be given.