A reconfigurable and hierarchical parallel processing architecture:performance results for stereo vision

摘要

A multiprocessor architecture called NETRA is discussed. It ishighly reconfigurable and does not involve the use of complexinterconnection schemes. The topology of this multiprocessor isrecursively defined and is therefore easily scalable from small to largesystems. It has a tree-type hierarchical architecture featuring leafnodes that consist of a cluster of small but powerful processorsconnected via a programmable crossbar with selective broadcastcapability. The architecture is simulated on a hypercube multiprocessorand the performance of one processor cluster is evaluated forstereo-vision tasks. The particular stereo algorithm selected forimplementation requires computation of the two-dimensional fast Fouriertransform (2-D FFT), template matching, histogram computation, andleast-squares surface fitting. Static partitioning of data is used forthe data-independent tasks such as 2-D FFT and dynamic scheduling, andload balancing is used for the data-dependent tasks of feature matchingand disambiguation