基于嵌入式GPU架构的并行形状特征提取算法优化

Optimization of parallel shape feature extraction algorithm based on embedded GPU architecture

为解决嵌入式平台上形状特征提取算法运算量大、耗时长,实时性不高的问题,提出一种基于嵌入式GPU架构的并行形状特征提取算法,并采用指令级优化、数据传输优化、访存优化等多种优化策略进行优化;设计串、并行算法的对比分析试验,并行算法相对于串行算法,最高加速比为596倍,试验结果表明,基于GPU并行优化形状特征提取算法在保证特征提取有效性的前提下,极大的提高算法的运行速度,进而能够提高整个模式识别过程的实时性。

In order to solve the problem that the shape feature extraction algorithm on the embedded platform has a large amount of computation,takes a long time,and has low real-time performance,we proposed a parallel shape feature extraction algorithm based on an embedded GPU structure.Using optimization strategies such as instruction-level optimization,data transmission optimization,and memory access optimization to optimize it.We designed comparative experiments to analysis the results of serial and parallel algorithms ,compared to the serial algorithm,the parallel algorithm's highest acceleration ratio can be 596 times.The results showed that on the premise of ensuring the validity of feature extraction,algorithm greatly improved the operating speed,and then the real-time performance of the entire pattern recognition process can be improved.