集成电路与嵌入式系统2025，Vol.25Issue(4)：30-38,9.DOI:10.20193/j.ices2097-4191.2024.0087
基于存内计算的卷积网络量化方法
Convolutional network quantization method based on in-memory computation
摘要
Abstract
This paper presents a quantization method for convolutional networks based on in-memory computation,to address the net-work performance degradation typically caused by statistical methods used for calculating analog-to-digital conversion coefficients when deployed on in-memory computing circuits.This method first quantifies the activation values and weight coefficients of the convolutional layer.Then,based on the characteristics of the single Tile data stream in the in-memory computing circuits,we design an analog-to-dig-ital conversion coefficient quantization network.Afterwards,a method based on KL divergence is developed to calculate the analog-to-digital conversion coefficients.Finally,the analog-to-digital conversion coefficients are mapped to conductance values and fused with the activation values and weight quantization coefficients in the convolutional layer.These values are then converted into shift and fixed-point multiplication forms to achieve the deployment of inference in the in-memory computing circuit of the convolutional network.The soft-ware simulation rusults show that compared with other methods for calculating analog-to-digital conversion coefficients,the designed quantization method results in less performance degradation and is suitable for multi-bit width mixed quantization in convolutional net-works.Due to the software simulation fully simulating the data flow process of in memory computing circuits,the proposed method can be applied in engineering implementations on in-memory computing circuits.关键词
存内计算/卷积网络/模/数转换系数/量化部署Key words
in-memory computation/convolutional network/analog to digital conversion coefficient/quantitative deployment分类
电子信息工程引用本文复制引用
桑贤侦,李敏,程虎,魏敬和,赵伟,王正行..基于存内计算的卷积网络量化方法[J].集成电路与嵌入式系统,2025,25(4):30-38,9.基金项目
国家自然科学基金项目(62174150,62204233) （62174150,62204233）
江苏省揭榜挂帅项目(BE2023005) （BE2023005）
江苏省自然科学基金项目(BK20211040,BK20211041) （BK20211040,BK20211041）
江苏省产业前瞻与关键核心技术重点项目(BE2021003-1). （BE2021003-1）
