面向高性能DSP的图像滤波库优化

Optimization of Image Filtering Library for High-Performance DSP

滤波函数在图像处理领域起着至关重要的作用.传统的滤波实现方法以窗口为处理单位,但窗口尺寸较小导致指令流水线频繁中断.针对该问题,本文提出了算法切片的优化方法.首先,对滤波算法进行切片,使得每个切片处理滤波窗口中的一个像素点;接着,根据切片处理的像素点在滤波窗口中的位置计算出有效处理范围,去除图像边缘处理中复杂的条件判断;最后,按列方向进行多簇流水,让流水线重复执行大量相同的指令序列.结合BWDSP 1042的特殊指令和硬件逻辑,对中值滤波等图像滤波函数进行了优化.实验结果表明:在四簇流水模式下,所有图像滤波函数性能均提升了51倍以上.

Filter functions play a significant role in image processing. The traditional implementation method takes the window as the processing unit, whose size is so small that the pipeline is interrupted frequently. This paper proposes an optimization method of algorithm slicing to settle this problem. First, the filtering algorithm is sliced so that each slice processes one pixel in the filter window. Then, the effective processing range is calculated from the position of the pixel in the filter window to remove the complex conditional judgment of the edge of the image. Finally, the software pipeline is carried out in the column direction, allowing the pipeline to repeat a large number of identical instruction sequences. Combined with BWDSP 1042 special instructions and hardware logic, the median filter and other image filtering functions are optimized. The experimental results show that the performance of the entire image filtering functions is improved by more than 51 times in the four-cluster pipeline mode.