Performance Evaluation of Parallel Computing For Real-Time Image Processing

Abstract

Real-time applications based on image processing are both computationally data intensive. This fact presents real-time image processing as a strong candidate for parallelism. In this thesis, real-time image processing is exploited for parallel computing from two different views. The first is loop parallelization since loops are dominant control structure existing in image processing algorithms. The second is volume visualization which is taken a5 a real-time image processing workload. Concerning the loop parallelization aspect, a loop fusion algorithm is proposed. objective is to fuse loops of the same type (either DOALL or DO loops) considering identical and overlapping bounds. Data locality is enhanced by bringing references to the same data in different loops into a single loop and therefore the need for communication and data replication is reduced. In this work, the iteration spaces of loops before and after using the proposed loop fusion algorithm are. executed sequentially and on a multiprocessor computer based on message-passing . The proposed loo. p fusion algorithm proved to be effective for both sequential and parallel execution, fully and partially overlapped iteration Spaces and different types of dependencies. Increasing either the shift size (negative dependence) or the peel size (positive dependence) increases the execution time for the same number of processors due to the resulting reduction in the fused iteration space.