快速卷积定点运算中的量化噪声分析  

Quantization Noise Analysis in Fixed-point Operation of Fast Convolution

作　　者：邢姗姗 郑虹 戴乐 孟藏珍[2] XING Shanshan;ZHENG Hong;DAI Le;MENG Cangzhen(The Unit 93209 of PLA,Beijing 100085,China;Air Defense Early Warning Ecjuipment Department,Air Force Early Warning Academy,Wuhan 430019,China)

机构地区：[1]解放军93209部队,北京100085 [2]空军预警学院防空预警装备系,武汉430019

出　　处：《现代雷达》2020年第10期46-50,75,共6页Modern Radar

摘　　要：介绍了定点表示法的基本原理,分析了快速卷积定点运算中的量化误差,建立快速傅里叶变换(FFT)定点处理在分别采用不定标、预定标和逐级定标三种算法时,FFT点数以及系统字长对输出噪信比的影响关系模型,并进行了实验验证。结果表明,采用不定标算法时,输出噪信比随着FFT点数的增大先减小再增大,且系统字长越长,输出噪信比开始增大时所对应FFT点数越大;采用预定标和逐级定标算法时,输出噪信比随FFT点数的增大逐渐减小,最终趋近于某个值,且受系统字长影响不大。在实际FFT设计中,可根据FFT点数和系统字长选择不同的定标算法,进而满足输出噪信比的要求。The basic principle of fixed-point operation was introduced,quantitative error in fixed-point operation of fast convolution was analyzed,and the influence model of FFT points and system word length on output SNR was established when fixed-point processing of FFT adopted three algorithms of un-calibration,pre-calibration and stepwise calibration respectively,and the experimental verification was carried out.Experimental results showed that,when the un-calibrated algorithm was adopted,the output SNR first decreased and then increased with the increase of FFT points,and the longer the system word length,the larger the corresponding FFT points when the output SNR ratio began to increase.When the pre-calibration and stepwise calibration algorithms were adopted,the output SNR decreased gradually with the increase of FFT points,and finally approached to a certain value,which was not affected by the system word length.In the actual FFT design,different calibration algorithms can be selected according to FFT points and system word length,so as to meet the requirements of output SNR.

关 键 词：信号处理 快速傅立叶变换 定点运算 量化误差 

分 类 号：TN957.51[电子电信—信号与信息处理]