面向CPU+GPU异构计算的SIFT

依据图形处理器(GPU)计算特点和任务划分的特点,提出主从模型的CPU+GPU异构计算的处理模式.通过分析和定义问题中的并行化数据结构,描述计算任务到统一计算设备架构(CUDA)的映射机制,把问题或算法划分成多个子任务,并对划分的子任务给出合理的调度算法.结果表明,在GeForce GTX 285上实现的尺度不变特征变换(SIFT)并行算法相比CPU上的串行算法速度提升了近30倍.

Parallel Algorithm of CPU and GPU oriented Heterogeneous Computation in SIFT Feature Matching

As the development of space remote sensing technology witnessed a geometric growth in the data size of remote sensing images, the process of SIFT(Scale Invariant Feature Transform) feature matching is faced with such challenges as large data size, high computational complexity and large computational quantity, and so on. According to the basis of features about GPU(Graphic Processing Unit) computing and tasks dividing, the article tries to bring forward a method of Master/Slave CPU+GPU heterogeneous computing. This study analyzes and definites the parallel data structures, and describes the mapping mechanism for computing tasks on CUDA( Compute Unified Device Architecture ). A logical scheduling algorithm is proposed to divide an issue or algorithm into more little tasks to fulfill the features of the GPU and CPU and General-purpose computing ability of the GPU. The result shows that the speed of SIFT parallel algorithm in the Geforce GTX 285 is about 30-time of the serial algorithm running in the CPU. The comparison between GPU and CPU in SIFT feature matching algorithms shows the greater advance of the CUDA in high arithmetic intensity real-time processing and computing data processing and this provides new ideas to improve image matching performance and general-purpose computing on GPU.