二维矩阵卷积在向量处理器中的设计与实现

摘要

为了加快卷积神经网络模型的计算速度,便于大规模神经网络模型在嵌入式微处理器中的实现,以FT-matrix2000向量处理器体系结构为研究背景,通过对多核向量处理器体系结构的分析和对卷积神经网络算法的深入研究,提出将规模较小的卷积核数据置于标量存储体,尺寸较大的卷积矩阵置于向量存储体的数据布局方案.针对矩阵卷积中数据难以复用的问题,提出根据卷积核移动步长的不同动态可配置的混洗模式,通过对所取卷积矩阵元素进行不同的移位操作,进而大幅提高卷积矩阵数据的复用率.针对二维矩阵卷积由于存在数据相关性进而难以多核并行的问题,提出将卷积矩阵多核共享,卷积核矩阵多核独享的多核并行方案.设计了卷积核尺寸不变、卷积矩阵规模变化和卷积矩阵尺寸不变、卷积核规模变化的两种计算方式,并在主流CPU、GPU、TI6678、FT-matrix2000平台进行了性能对比与分析.实验结果表明:FT-matrix2000相比CPU最高可加速238倍,相比TI6678可加速21倍,相比GPU可加速663 805倍.%In order to accelerate the computational speed of convolution neural network model and facilitate the implementation of large-scale neural network model in embedded microprocessor, the FT-matrix2000 vector processor architecture was taken as the research background.Through the analysis of the multi-core vector processor architecture and convolution neural network algorithm, a data layout scheme was proposed in which a smaller convolution kernel data was placed in a scalar memory bank and a larger convolution matrix was placed in a vector bank.Aimed at the problem that the data in the matrix convolution is hard to reuse, a dynamic shuffling pattern with different dynamic configurable parameters based on the moving steps of the convolution kernel was proposed, by carrying out different shift operations on the convolution matrix elements, the multiplexing rate of convolution matrix data was greatly improved.Aimed at the problem that two-dimensional matrix convolution is difficult to multi-core parallelism due to the existence of data correlation, a multi-core parallel scheme with convolution matrix sharing and convolution kernel matrix multi-core exclusive was proposed.Two computing methods of convolution kernel size unchanged, convolution matrix size changed and convolution matrix size unchanged and convolution kernel size changed were designed, a performance comparison and an analysis were carried out in mainstream CPU, GPU, TI6678 and FT-matrix2000.The final experimental results show that compared with the multi-core, the CPU can be accelerated up to 238 times, compared with TI6678, the speed can be accelerated 21 times, and compared with the high-performance GPU, the speed can accelerate 663 805 times.