IP Core for Efficient Zero-Run Length Compression of CNN Feature Maps

摘要

Convolutional Neural Networks (CNNs) are becoming a fundamental tool for machine learning. High performance and energy efficiency are of great importance for deployments of CNNs in many embedded applications. Energy consumption during CNN processing is dominated by memory access and since large networks do not fit on on-chip storage, they require expensive DRAM access. This paper introduces an universal Output Stream Manager (OSM) which can be used to compress and format data coming from a CNN accelerator and reduce external memory access. The OSM exploits the sparsity of data and implements two Zero-Run Length encoding algorithms and can be easily reconfigured to optimize usage for different CNN layers.