摘要
对Smith-Waterman算法的计算公式进行了改进以适应GPU并行的特点,并提出新的基于BLOCK分块的并行前缀扫描法;通过UP-DOWN步骤、BLOCK间调整、Eij微调等步骤在O(logn)时间内计算出行中每一个元素的前缀最大值;最后将回溯过程置于GPU端,避免了CPU与GPU间内存的拷贝.与传统的Smith-Waterman算法相比,该算法在低端的GPU平台性能提升90倍;与同样基于GPU的SWAT算法相比,性能也有较大的提升.
The formulae of Smith-Waterman algorithm was improved to make it adapt to the parallel characteristics of GPU, and a novel strategy of parallel prefix scan was proposed. The prefix maximum of each element in the row within 0 （/ogn） time was eacu- fated through UP-DOWN steps and the adjustment between blocks and E/j fine tuning. At last, the backtrack procudure ran at GPU side to avoid the memory copies between GPU and CPU. Compared with traditional Smith-Waterman algorithm, this algorithm per- formance increased 90 times on the lower-end GPU platform, and also had larger ascension in comparison with SWAT algorithm.
出处
《福建农林大学学报（自然科学版）》
CSCD
北大核心
2015年第4期442-448,共7页
Journal of Fujian Agriculture and Forestry University:Natural Science Edition
基金
福建省自然科学基金资助项目(2013J01216
2014J01219)
