This paper descibes an infrastructure that enables transparentdevelopment of image processing software for parallel computers. Theinfrastructure's main component is an image processing librarycontaining operations capable of running on distributed memoryMIMD-style parallel hardware. Since the library has a programminginterface identical to that of a familiar sequential image processinglibrary the parallelism is completely hidden from the user. All libraryfunctions are based on an abstract parallel image processing machine(APIPM), introduced in this paper. To guide the process of automaticparallelization and optimization, a performance model is defined foroperations implemented using APIPM instructions. Experiments show thatfor realistic image operations performance predictions are highlyaccurate. These results suggest that the infrastructure's core forms apowerful basis for automatic parallelization and optimization ofcomplete applications
展开▼
