基于ARM + DSP的异构平台优化加速算法

Optimization and Acceleration Algorithm for Heterogeneous Platforms Based on ARM + DSP

摘要: OpenCL编程模型应用于ARM + DSP异构多核平台存在核心利用率低、开发效率低等问题。本文基于AM5728异构开发平台，对OpenCL异构编程模型进行研究，提出了异构多核计算动态优化加速算法。分析了动态优化加速算法中的最优分配比例算法和数据划分原则，动态算法会根据运行情况动态调整相应参数。完成了测试系统的设计，对异构计算加速算法的相关参数进行测量，展示了Sobel算法、奇异值分解算法(SVD)分别采用计算加速驱动的结果和OpenCL异构编程模型的结果，分析两种不同方式下算法完成时间情况。测试结果表明优化加速算法使得Sobel算法执行时间降低至原执行时间的72.2%，SVD算法执行时间降低至原执行时间的80.2%。

Abstract: When OpenCL programming model is applied to ARM + DSP heterogeneous multi-core platform, there are still some problems such as low core utilization and low development efficiency. Based on AM5728 heterogeneous development platform, this paper studies the OpenCL heterogeneous pro-gramming model and proposes a dynamic optimization algorithm for heterogeneous multi- core computing. Analyze the calculation method of the optimal data segmentation ratio, determine the principle of data segmentation, and the dynamic algorithm will dynamically adjust the correspond-ing parameters according to the operation. The design of the test system is completed, and the re-sults of Sobel algorithm and Singular Value Decomposition (SVD) algorithm driven by computing acceleration and OpenCL heterogeneous programming model are shown. The completion time of the algorithm under two different modes is analyzed. The test results show that the optimization algo-rithm reduces the execution time of Sobel algorithm to 72.2% of the original execution time, and the SVD algorithm to 80.2% of the original execution time.