Increasing Interconnection Network Connectivity for Reducing Operator Complexity in Asynchronous Vision Systems
Due to the restriction of SIMD mode to local operations in VLSI massively parallel vision chips, using programmable connections and asynchronous communications are key ingredients to support regional computations. Asynchronism implies using combinatorial multi-input operators having an important hardware cost. To reduce it, we propose to use a connection network having a connectivity level greater than the mesh being mapped. This solution allows to use only 2-inputs asynchronous operators having a reduced hardware cost in each pixel. Examples and results will be presented on the examples of the regional sum algorithm computed over a 4-connectivity squared mesh connected with a 6-connectivity interconnection network, and the regional sum computed over a 6-connectivity squared mesh connected with a 8-connectivity interconnection network.
KeywordsSpan Tree Interconnection Network Hardware Cost Single Instruction Multiple Data Full Adder
- 1.Moini, A.: Vision Chips. Kluwer Academic Publishers, Dordrecht (2000) (ISBN: 0-7923-8664-7)Google Scholar
- 2.Paillet, F., Mercier, D., Bernard, T.: Second generation programmable artificial retina. In: IEEE ASIC/SOC Conf, pp. 304–309 (1999)Google Scholar
- 3.Li, H., Stout, Q.: Reconfigurable Massively Parallel Computers. Prentice-Hall, Englewood Cliffs (1991)Google Scholar
- 6.Dulac, D., Mohammadi, S., Merigot, A.: Implementation and evaluation of a parallel architecture using asynchronous communications. In: CAMP, pp. 106–111 (1995)Google Scholar
- 7.Ducourthial, B., Merigot, A.: Graph embedding in the associative mesh model. Technical Report TR-96-02 (1996)Google Scholar