基于模拟退火算法的浮点转定点自动位宽优化工具  

Automatic Word-Length Determination Tool Based on Simulated Annealing Algorithm

作　　者：黎渊[1] 蒋江[2] 张民选[1] 魏少军[3] 

机构地区：[1]国防科学技术大学计算机学院,长沙410073 [2]上海交通大学微电子学院,上海200240 [3]清华大学微电子研究所,北京100084

出　　处：《上海交通大学学报》2013年第1期76-80,85,共6页Journal of Shanghai Jiaotong University

基　　金：国家自然科学基金(60970036;61103011);教育部博士点基金(20114307110001);国家"核高基"重大专项(2009ZX01028-002-002)资助项目

摘　　要：开发了一套浮点转定点自动位宽优化软件系统(SATRANS),能够将用户输入的描述目标系统的浮点程序自动转换为位宽可配置的定点程序,并基于模拟退火算法进行自动位宽搜索,以得到满足精度要求的操作数定点位宽组合.同时,以IIR数字滤波器为例对SATRANS进行了实现与验证.结果表明,SATRANS的搜索结果优于传统贪心算法的搜索结果,并能够获得一系列满足精度要求的解,从而使得芯片设计人员能够在精度与复杂度等要素之间加以权衡,并选择一组最合适的位宽组合而用于芯片设计中.选择搜索结果中的面积最优解来配置IIR系统并在XilinxVirtex-6FPGA芯片中实现,相对于IEEE浮点单/双精度系统,其性能分别提高了12.4%和62.8%,面积的降幅分别为93.9%和97.9%.An automatic word-length determination tool （SATRANS） based on the simulated annealing al- gorithm was developed. SATRANS can automatically transform the system from floating-point model to fixed-point model and provide a series of word-length solutions that form a tradeoff curve for hardware complexity vs. signal quality. SATRANS was demonstrated to find word-length for an infinite impulse re- sponse filter （IIR）. The results show that SATRANS can provide better word-length solution in compari- son to the traditional search method based on greedy strategy. The word-length optimized IIR targeting Xilinx Virtex-6 FPGA device was implemented, which improves the performance by 12.4% and 62.8% while saves almost 93.9 %and 97.9 % of area in comparison to the IEEE single and double floating-point generators.

关 键 词：位宽优化 模拟退火算法 浮点转定点 数字滤波器 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]