A General-Purpose and Configurable Planar Data Processor for Energy-Efficient Pooling Computation

Abstract

Convolutional Neural Networks (CNN) have been widely used in artificial intelligence applications. A typical CNN contains both convolution and pooling layer, in which the convolution is to detect local conjunctions of features and the pooling is to merge similar patterns into one. It is necessary to make pooling operation, which plays a great role in CNN. Up to now, there have been numerous researches on CNN accelerators, however, most of the previous works are only focused on the acceleration of convolution layers, and the specific studies on pooling units are still lacking. Besides, the existing pooling designs are usually constrained by either the poor flexibility or the low energy/area efficiency. In this work, we propose a general purpose and energy-efficient planar data processor to support the pooling operation from different CNN structure. By using the configurable data path control method, the processor is able to support universal pooling operation with arbitrary input feature shape and arbitrary pooling kernel/stride/padding size. Besides, the processor exhibits high efficiency with hardware utilization ratio near 100% during operation, indicating good performance of the design. Most importantly, it is energy-efficient that exhibits 86%-off on power consumption and 62%-off on area utilization when compared with the separate pooling module of NVDLA (NVIDIA Deep Learning Accelerator), thus is particularly suitable for the resource-limited edge intelligent devices.