Algorithms and Parallel Computing

3.5 SIMD PROCESSORS

SIMD could be classified as a special case of single program multiple data stream (SPMD) [29]. SIMD processors could belong to the class of shared memory multiprocessing system or distributed-memory multiprocessing system. SIMD machines built using shared memory are suited to applications that require frequent exchange of data where one processor acts as the producer of new data and many other processors act as the consumer of this data.

Each processor executes the same task in synchrony with the other processors. The task being executed could be a simple instruction, a thread, or a process. Distributing the memory among the processors reduces the memory bandwidth problem.

Many applications lend themselves to the SIMD processing model as long as the application is parallelizable. Applications include bioinformatics, biomedical diagnosis, fluid dynamics, image processing, and video processing. SIMD provides the ability to dramatically boost the performance of an application. Some computer manufacturers are adding SIMD extensions to their processors and can run existing applications without the need for recompilation. It is also easy-to-learn programming modifications that utilize SIMD architectures such as the Intel C++ parallel exploration compiler.

An example of a parallel algorithm that is suited to the shared memory model of SIMD is recursive filters described by the equation

(3.1)

where a(j) and b(j) are the filter coefficients and N is the filter order ...