几何操作的数据并行实现  

Data Parallel Implementation of Geometric Operations

作　　者：王光[1] 张发存[1] 沈绪榜[1] 

机构地区：[1]西安微电子技术研究所,西安710054

出　　处：《吉林大学学报（信息科学版）》2005年第6期650-656,共7页Journal of Jilin University（Information Science Edition）

基　　金："十五"国防863课题(2002AA714022)

摘　　要：为简化几何操作数据并行实现的复杂度,特别是简化几何操作中重采样数据并行实现的复杂度,提出了通过SIMD(S ingle Instruction Mu ltip le Data)处理元阵列内的数据并行传送,而不是通常所采用的处理元阵列与阵列存储器之间的数据流传送实现几何操作的处理方法。讨论了几何操作正向映射的零阶内插与一阶内插的数据并行实现方法;几何操作反向映射的零阶内插与一阶内插的数据并行实现方法;以及正向映射与反向映射的3种特殊情况。结果表明,该方法使几何操作的数据并行实现方法的复杂性为Ο(M+N),而不是数据流传送实现方法的Ο(MN),从而更好地满足了数字图像快速实时处理的需要。In order to simplify the complexity of data parallel implementation of geometric operations, especially to simplify the complexity of data parallel implementation of resampling in geometric operations, a new processing scheme is presented to implement geometric operations by moving data parallelly in PE （Processor Element） array, instead of moving data stream between PE （Processor Element） array and memory array which used by researchers at present. The data parallel implementation scheme of zero-order interpolation and first-order interpolation for forward mapping is discussed. Then, the data parallel implementation scheme of zero-order interpolation and first-order interpolation of backward mapping are discussed. Three kinds of special cases of forward and backward mapping are discussed at last. The result shows that the complexity of data parallel implementation scheme of geometric operations presented in this paper is O（M ＋ N） rather than O（MN） in the schemes implemented by data stream moving, thereby it can easily meet the requirements of the digital image processing in real

关 键 词：几何操作 重采样 数据并行实现 零阶内插计算 一阶内插计算 前向映射 后向映射 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]