A migrating data-driven architecture for multidimensional signal processing

摘要

Efficient architectures for multidimensional signal processing have to take into account algorithm requirements and technology bounds. Using a top-down approach, the algorithm is represented by modified Petri nets graphs so that both concurrency and execution control can be formally represented for its implementation on a data-driven machine. A modular architecture based on migrating-data-driven processors (MDDP) is then defined. Data packets are routed, in static ways, by the MDDPs, while computation operand packets can migrate between MDDPs depending on the computation load gradients. The architecture of the MDDP is discussed at the functional and logical levels, the most critical issues and blocks are presented.