T0 for Computer Vision

International Computer Science Institute and the Computer Science Department of the University of California at Berkeley.

T0 is the first implementation of the Torrent architecture, which in turn is based on the industry standard MIPS-II RISC architecture. In Torrent, the integer RISC core is supplemented with a high performance fixed point vector coprocessor.

The initial application of T0 will be in the SPERT neural network and signal processing accelerator board for workstations.

One of T0's target application is the implemention of a full computer vision system. So far we coded key algorithms for T0 which are computational bottlenecks in an existing computer vision system. The algorithms are part of a traffic surveillance and a vehicle navigation system of the U.C. Berkeley PATH project. The most computational expensive modules are low-level feature extraction filters (graylevel gradient estimates), stereopsis (disparity calculations), and boundary tracking (snakes), which can't be performed in realtime on current workstations. Our code simulations showed that with t0 it is possible to run such a system in realtime. Compared to the existing C++ implementations on a SPARC/10 workstation, our t0 code implementations achieved a speed-up of 150 (Figure below). We note that algorithms like disparity calculations are general computations needed in stereopsis, optical flow, and other intermediate vision algorithms. The occurence of many conditional calculations usually makes implementations on DSP hardware difficult, but we experienced that the t0 instruction set was well suited and allowed short development time.