Model-based programming for parallel image processing

We describe a programming environment which is being developed for the automatic generation of parallel image processing applications. Through the use of model-based software synthesis, we transparently create large grained data parallel applications which can be executed on arbitrary processor networks. The high-level abstractions provided by the modeling paradigm isolates the user from the complexity of the underlying implementation, allowing developers with little or no experience in parallel programming to rapidly create parallel applications. The data parallel modeling facilities perform the same tasks as the data alignment and distribution compiler directives of High Performance Fortran and the aggregate objects of pC++. However, we have found that by introducing the parallelism on the system level, instead of in the algorithm, we can use traditional compilers and leave the application specific code unchanged. This allows us to take advantage of existing well developed image processing code libraries. Here we describe a system which generates data parallel versions of applications created in Khoros, the popular image processing package developed by the University of New Mexico. This system retains the best qualities of Khoros: its interactive and experimental nature, and its visual interface, but adds the capability for automatically generating much higher performance parallel implementations when needed. This system demonstrate the suitability of the model-based approach for developing parallel imaging software