No label defined (Q4935998)

This book presents a Ph.D. thesis that has been carried out in conjunction with a project on video signal processing at the Philips Research Laboratories. It presents the theory and implementation of methods to map signal flow graphs onto a network of video signal processors under a cyclostatic execution model. Video signal processing requires high performance in order to meet the real time constraints with high sampling rates. This performance can usually not be achieved with software on high speed DSP processors. The development of dedicated hardware for each video signal processing algorithm however is very expensive. Therefore an approach between these two sides by a programmable video processor network and a specification of the processing algorithm as a signal flow graph is an interesting alternative. Following this approach, methods are needed to translate a signal flow graph into program code of a given processor network that implements the functionality of the signal flow graph. This mapping problem is a very sophisticated task, because many constraints have to be met while performance has to be maximized. After a motivation of his approach in chapter 1 the author presents in chapter 2 a thorough formalization of all concepts involved, i.e. processor networks, signal flow graphs, the mapping problem and the constraints. Chapter 3 contains an analysis of the computational complexity of the mapping problem and several subproblems. This chapter motivates the problem decomposition and the heuristic strategies that are presented by chapter 4 to 7. All subproblems and solution strategies are accompanied by an experimental analysis of practical video signal processing problems on two different processor networks. The mathematical formalization of the mapping problem and the motivation and use of combinatorial optimization methods to find approximations of optimal solutions are the strength and the main contribution of this book.