Japanese / English


論文の言語 英語
著者 Masakazu IWAMURA, Takayuki HONDO, Kazuto NOGUCHI, and Koichi KISE
論文名 An Attempt of CUDA Implementation of PCA-SIFT
論文誌名 IEICE Technical Report
Vol. 107
No. 281
発表番号 PRMU2007-117
ページ pp.149-154
年月 2007年10月
要約 GPGPU (General-Purpose computation on GPUs) is a paradigm to use GPUs (graphics processing units) for general computation. Due to recent remarkable improvement of GPU, GPU outperforms CPU in computation ability. However, most people could not use the ability for general computation because existing programming languages require knowledge about GPU hardware architectures and computer graphics for GPGPU computing. Recently, a new GPU language CUDA (Compute Unified Device Architecture) has been released from NVIDIA. The CUDA code is C language style and has less computational restriction. Thus, usual operations of C language can run on GPU without much special knowledge. In this report, we briefly introduce CUDA language programming and report a CUDA implemented of PCA-SIFT. Compared to a CPU implementation, our CUDA implementation reduced the processing time to around 1/4. In addition, we also report an interesting phenomenon results useful for practical use of CUDA.