前导0预测算法前缀模型的研究与实现  

Design and implementation of a prefix computation model of leading zero anticipator algorithm

作　　者：富坤[1,2] 魏思捷 耿跃华[3] 

机构地区：[1]河北工业大学计算机科学与软件学院,天津300401 [2]河北省大数据计算重点实验室,天津300401 [3]河北工业大学电气工程学院,天津300130

出　　处：《计算机工程与科学》2017年第10期1788-1793,共6页Computer Engineering & Science

基　　金：国家自然科学基金(51577131)

摘　　要：前导0预测算法对浮点加法运算的研究非常重要,通过对现有浮点加法分析发现,前导0预测模块处于运算的关键路径,提高前导0预测模块的性能会使整体性能提高。从前缀计算角度提出一系列的定义、定理,并证明了前导0检测算法本质上都可以归结为前缀计算问题,从而该问题可以采用二分递归方法求解。在前导0预测过程中,首先通过对加法运算量做"借位留存"的减法,把两个运算量的预测转换为{1,0,-1}数字字符集上的数字串的特征检测问题。接着采用消除连续"-1"串的重编码技术得到一种统一的位串形式。最后设计出基于位串构成形式前导0位数的二分递归判断方法。The leading zrro anticipator （LZA） algorithm is very important for studying floating-point addition operation,and the analysis of floating-point addition indicates that the LZA module plays a cri--ical role in the performance improvement of floating-point addition operations. We define a series of defi-nitions and theorems from the perspective of prefix computation,and prove that the LZA algorithm can be attributed to a prefix computation problem in nature. Thus,the problem can be omy recursive method. In LZA process,we firstly convert the prediction problem of two operations to a digital string feature detection problem on numeric character set of {1,0^ - 1} by performing the sub-traction of ＂Borrow Retirement＂ on addition operations. Secondly, a unified form of bit strings is ob-tained via the recoding technique which eliminates the continuous string of “-1”. Finally, the dichotomy recursive algorithm is designed based on the bit string structure of leading zero digits.

关 键 词：前导0预测 前缀模型 重编码 二分递归 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]