基于GPU并行计算的图像特征提取与匹配研究

Research on Image Feature Extraction and Matching Based on GPU Parallel Computing

GPU高性能计算以其强劲计算能力成为图像算法研究的首选计算平台,在实际的应用研究中往往将视频分析工作负载分布到CPU和GPU中执行,论文针对图像特征提取的应用提出一种从视频流数据中提取特征性能的性能预测模型。该模型实现CPU-GPU混合计算系统的总执行时间评价,并确定最佳工作负载比率以及如何针对给定工作负载使用CPU内核。通过实验分析,确认提出的方法可以提高三种典型工作负载分布的加速:CPU只有GPU,或CPU-GPU混合计算,工作负载比为50∶50。

GPU high performance computing with its strong computing ability become the preferred computing platform of im- age algorithm, in the applied research, video analysis workload is distributed to the implementation of CPU and GPU, this paper proposes a feature extraction from the video stream in the performance characteristics of the performance prediction model. The model implements the total execution time evaluation of CPU-GPU hybrid computing system, and determines the best working load ratio and how to use the CPU kernel for a given workload. Through experimental analysis, it is confirmed that the proposed method can improve the acceleration of three typical work load distributions： CPU, only GPU or CPU-GPU mixed calculation, and the work load ratio is 50： 50.