Writing retargetable parallel programs for low and high level vision using a global address space

A model for programming parallel and distributed computers in which the user is presented with a single linear address space, unlimited computational parallelism, and a global clock for synchronization is presented. This allows the user to express the abstract parallelism, without being tied to the target architecture and makes the resulting software transportable. An implementation of this model using C as the source language supplemented with a distributed data storage class and concurrent assignments is presented. An example of generating code for low-level image processing algorithms, including histogram equalization, is given. This example shows that parallel algorithms can be developed without architectural knowledge and the resulting programs can be compiled for a wide range of target processors