关键词: 开源处理器, 硬件算法加速器, 平台验证, 异构SoC

Abstract: As the number of hidden layers of neural network increases, the amount of training calculation increases. Heterogeneous SoC including hardware algorithm accelerator has been proposed to improve the efficiency of the algorithm. The open source processor Rocket core project contains a core generator, which not only can customize the number of cores, but also contains the coprocessing extension interface, which is easy to research and design heterogeneous SoC. Based on the open source processor Rocket core and the open source project Si-Five Blocks, a simplified SoC is built with ReLU coprocessor and vector inner product accelerator as examples. The prototype verification of ReLU and vector dot product accelerator is completed with FPGA development board VC707 as the verification platform. The results show that the heterogeneous SoC is effective and practical for accelerating convolution neural network operation.

Key words: opensourceprocessor, hardwarealgorithmaccelerator, prototypeverification, heterogeneousSoC