Research Of Several Technologies Under The Many-core System For Algorithm Optimization

Many-core architectures including GPU, GPU units on heterogeneous CPU andMIC acceleration components and so on are generally used to calculate the unit ofgeneral purpose computing. Recently CPU development has encountered a series ofbottlenecks such as frequency wall, power wall, and storage wall and so on. Therefore,the use of many-core computing architecture to achieve higher performance hasaroused more and more attention in this field. In fact, with the widespread usage ofmany-core architectures, some problems have been arising while during theexploration of high performance in applications. And the main problem focuses onhow to achieve high performance by taking full advantage of many-core architecture.That is to say, the lag of application software development wastes hardware, which isattributed by the blindness use of many-core architectures or developers’ unfamiliarityto these architectures. Moreover, many-core architectures reference documentation ofapplication development and open source are relatively few which certainly hinder thepromotion and application of many-core architectures.Research in this paper concentrates on the study and verification of several keytechnologies of algorithm optimization introduced here. These technologies includecommon optimization techniques targeting at smaller data association algorithm. Andabove of all, research in this paper focuses on how to achieve high performance of theiterative algorithms under the many-core architecture efficiently. In order to verify thevalidity of the technologies proposed in this paper, we apply these techniques tooptimize some algorithms of OpenCV library and use test cases of the OpenCVlibrary to testify the correctness and efficiency of the implement of after-optimizationalgorithms. Currently, the algorithm optimization cases introduced in this paper havebeen included in the main content of the OpenCV library. Experimental results show that, by using the algorithm optimization methods proposed in this paper, theimplement of after-optimization algorithms achieves higher speedup than that on CPUversion and similar performance to that on CUDA platform, which effectively verifythe validity and cross-platform ability of optimization methods introduced in thispaper.