Abstract:
In the Mutual Information(MI) parallel computation method based on Compute Unified Device Architecture(CUDA),the execution efficiency is low because of the bank conflict.Aiming at this problem,an improved parallel computation method of MI is proposed in this paper.Firstly,CUDA threads hierarchy model and shared memory are used and the histogram is calculated by the parallel data access method with the same step.Secondly,combined with a shared memory,joint entropy is calculated by the twice merging method in the block,the spanning tree algorithm is adopted in the whole process of merging to resolve the bank conflict,and the instruction expand strategy is employed to optimize the execution efficiency.Finally,MI calculation is completed in accordance with the entropy and joint entropy.Experimental results show that this improved method can increase the computational efficiency of MI while avoiding bank conflicts effectively.
Key words:
Compute Unified Device Architecture(CUDA),
image histogram,
image entropy,
Normalized Mutual Information(NMI),
parallel computation
摘要： 基于计算统一设备架构(CUDA)的互信息并行计算方法存在因bank冲突而导致执行效率降低的问题。为此,提出一种改进的互信息并行计算方法。利用CUDA的线程层次模型和共享存储器,按等步长数据并行访问方式计算直方图,结合共享存储器,通过分块两次归并方法计算联合熵,采用生成树归并算法避免bank冲突,使用指令展开策略进一步优化执行效率,由熵和联合熵完成互信息计算。实验结果表明,该方法在避免bank冲突的同时,能有效提高互信息计算效率。
关键词:
计算统一设备架构,
图像直方图,
图像熵,
归一化互信息,
并行计算
CLC Number: