Abstract:In this paper, we present the results of parallelizing the Data Encryption Standard (DES) algorithm. The data dependence analysis of the loop iterations was applied in order to parallelize this algorithm. The OpenMP standard is used for presenting a parallel algorithm. The classic DES algorithm can be divided into parallelizable and unparallelizable parts. As a result of our experiments it was stated that most of the "for" loops of the DES algorithm are well suitable for parallelization. The experiments with the parallel part of the algorithm against the sequential one using two-processors machine has shown that the speed-up is about 1.95. However, the rest code, containing I/O functions, is unparallelizable that reduces the speed-up of the parallel program running on multiprocessor computers.

Abstract. Possibilities of synthesizing parallel C/C++ codes into hardware are presented provided that the code parallelism is represented by means of the directives of OpenMP, a de-facto standard that specifies portable implementation of shared memory parallel programs. The limitations of the hardware realizations of OpenMP directives are described and implementation details are stressed.

No abstract

A technique, permitting us to find synchronization-free parallelism in non-uniform loops, is presented. It is based on finding affine space partition mappings. The main advantage of this technique is that it allows us to form constraints for finding mappings directly in a linear form while known techniques result in building non-linear constraints which should next be linearized. After finding affine space partition mappings, well-known code generation approaches can be applied to expose loop parallelism. The technique is illustrated with two examples.

A new method of parallel loop execution is presented. Firstly, all the loop iterations are executed in parallel. Then, the ends of pairs of dependent iterations are re-executed. The method requires no conversion of a source loop into an equivalent serial-parallel loop, involving iteration indices to be converted into new ones, that appears to be a typical requirement in the most known loop paralleling approaches such as unimodular and nonunimodular linear transformation methods. Possibilities of serial and parallel correction processes are discussed. Experimental results are considered. As follows from experiments, applying the method can be reasonable for the loops with a small fraction of dependent iterations.