关键词: GIST特征, 统一计算设备架构（CUDA）, 线程池, 异构计算

Abstract: To extract the global feature of GIST, a heterogeneous CPU+GPU collaborative computing and optimization is firstly implemented: CPU is used to complete the tasks of small amount of calculations and irregular data operations, such as image quantization and linear extension, while using GPU to complete the tasks with compute-intensive and highly parallel data operations, such as filtering, Gabor feature extracting and dimension reducing. For processing image sequences, the thread pool technology is introduced on the CPU side. Through the use of each thread binding a CUDA stream for one image, the parallel stream computing for multiple images between CPU and GPU and the streaming data transmission delay hidden are achieved. Moreover thread pool technology also offers the methods of thread pre-creating, pre-allocating of resources and running thread number changing on resource, which can improve the computing efficiency of GPU scheduled by the CPU. Under the same computing accuracy, experiments show that GIST implementation on heterogeneous computing platforms for images reaches 8.35~9.31 times speedup of the running on traditional CPU platform, and has an upgrading rate of 10.0%~37.2% for image sequences data while using the thread pool.

Key words: GIST, Compute Unified Device Architecture（CUDA）, thread pool, heterogeneous computing