北京航空航天大学学报 ›› 2008, Vol. 34 ›› Issue (10): 1177-1181.

• 论文 • 上一篇    下一篇

一种面向矢量视频处理器的并行存储结构

朱 笛, 申功勋   

  1. 北京航空航天大学 宇航学院, 北京 100191
  • 收稿日期:2007-11-12 出版日期:2008-10-30 发布日期:2010-09-16
  • 作者简介:朱 笛(1977-),男,山东嘉祥人,博士生,andy_zhudi@yahoo.com.cn.

2-D parallel memory architecture for video processor

Zhu Di, Shen Gongxun   

  1. School of Astronautics, Beijing University of Aeronautics and Astronautics, Beijing 100191, China
  • Received:2007-11-12 Online:2008-10-30 Published:2010-09-16

摘要: 视频编解码算法不仅复杂度非常高,对数据存储需要密集的矢量化访问.针对这一问题提出了一种线性斜移结构的二维存储方案,支持任意行或任意列的单指令存取,与SIMD(Simple Instruction Multiple Data)结构的矢量处理器相结合,能够成倍的加速视频计算的执行效率.给出了该方案的地址计算逻辑和存储结构,精简的地址计算逻辑和数据重排逻辑构成了视频处理器的矢量地址产生单元,解决了数据重排的高延迟和高复杂度问题.结合H.264视频编码标准对这一结构的性能进行了分析和对比,对该编码标准中的运动估计,分像素插值和去块效应滤波这3个计算复杂度最高的模块在该结构下的执行效率进行了分析和对比.

Abstract: Video codec has a very high computation complexity and features intensive vector accesses to memory. A 2-D parallel memory scheme based on linear skewing scheme was proposed. The memory scheme can be combined with simple instruction multiple data (SIMD) vector processor to address the computation challenge of video. The address generation logic and scheme-s micro architecture were analyzed. The scheme uses a data rotation unit to permute data element to form a vector. This method simplified data permutation network which is the bottleneck of most parallel memory schemes. The performance difference between parallel memory and traditional memory were also compared. The kernel module of H.264/AVC such as motion estimation, de-blocking filter and interpolation were investigated based on the proposed memory scheme.

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发