The beginnings of a generic vectorized machine vision library implemented in AltiVec is discussed. The general vector loop for machine vision algorithms requiring only a single input pixel per input image to produce a single output pixel is discussed and analysed. Factors discussed that affect the speedup of the vector program are data alignment (when loading and storing), data streaming instructions (used to prefetch data), and function complexity. In addition, two simple operations are presented - an image doubling function and a threshold functor.
展开▼