Parallel Image Processing on Heterogeneous SIMD-MIMD Machines

From a hardware standpoint a heterogeneous architecture such as an SIMD array coupled onto a MIMD system, will yield a powerful solution for real time embedded vision tasks. From a software standpoint there is a need to approach this heterogeneous system in a homogeneous way. We investigated the feasibility of a uniform parallel programming approach on a heterogeneous machine. Using the parallel language CC++ we can express functional parallelism with constructs like par and parfor and hide the data parallelism, like images distributed over SIMD PEs, in CC++ classes. We investigated a way of parallel programming using arrays of buckets. These bucket-arrays can be distributed over MIMD Processing Units, SIMD Processing Elements or both. A bucket array is hidden in a CC++ class and can be approached by independent producer and consumer threads of the parallel program. We analysed the feasibility and efficiency of this approach with the constrained Euclidean Distance Transform while varying the number of processors from 2 to 128.