Dynamic Piecewise Linear/Regular Algorithms

In this paper we present an extension of the class of piecewise linear algorithms (PLAs) in order to model one type of dynamic data dependencies. This extension significantly increases the range of applications which can be parallelized and mapped to massively parallel processor arrays. For instance, a lot of computational intensive applications for video and image processing consist of nested loop programs with only few and simple run-time dependent conditionals. Furthermore, we outline in which case these extensions can directly used - with slight changes - within traditional mapping methodologies based on loop parallelization in the polytope model. Additionally, we outline future research directions in the case existing methods will be inefficient.