基于MPI＋GPU的哼唱检索系统加速

MPI+GPU accelerated query by humming system

由于利用MIDI文件中提取的特征和耗时的匹配算法,当前的哼唱检索系统可以实时处理的规模很小。由于SPRING算法显著降低了子序列匹配的复杂度,通过将哼唱检索抽象为一个子序列匹配问题,然后利用SPRING算法作为核心的匹配算法对音高序列进行子序列匹配,大大缩短了匹配时间。此外,利用GPU对SPRING算法进行加速,算法与串行算法相比获得接近40倍的加速比,使单节点每秒可以匹配的序列数目达到几千个。最后利用集群对系统进行加速。结果表明,我们的系统具有很好的扩展能力,同时检索的准确率也指明了当前的问题和今后的方向。

The current query by humming system can hardly be extended to large massive database as most of them adopt the features extracted from MIDI files, which are not widely used, and the very time-consuming matching methods. Because the SPRING algorithm dramatically reduces the algorithm complexity of subsequence matching, we regard query by humming as a subsequenee similarity matching problem and exploit the SPRING algorithm as the core matching method to compare the melody feature extracted from polyphonic music, reducing the matching time greatly. Furthermore, accelerated by GPU, the SPRING algorithm achieves a near 40 times speedup over the serial version. The processing a- bility per node can reach thousands of sequences matching per second under down sampling. With the help of clusters, the processing scale can be extended heavily, which shows that our system has a good scalability. At the same time, the accuracy results of query by humming point out the encountered prob lems and the future direction.