Speeding Up Convolution on Multi-cluster DSP in Deep Learning Scenarios

Recently, deep learning has achieved great success in artificial intelligent, whose superiority also brought new opportunity for the related research in embedded system. This paper focused on optimizing and speeding the convolution computing, the core operation within convolution neural network based on a multi-cluster digital signal processor, BWDSP. By taking advantage of the BWDSP’s architecture and characteristics of convolution computation, a suitable parallel algorithm was designed. Based on features of convolution neural network model structure, an automatic optimization tool for convolution computing with specific arguments was presented as well. The experimental result showed that the parallel algorithm given in this paper is 9.5x faster than GEMM-based algorithm commonly used in GPU and 5.7x faster than the traditional vectorization optimization algorithm. Meanwhile, a comparison was made between the parallel algorithm and tiled-base algorithm widely adopted in system with cache hierarchies, showing that the parallel one could achieve a better performance density of 1.55 times than that of later one, meaning that the work in this paper can make full use of computing resources to make them more efficient.