Automatic Parallelization of Non-uniform Dependences

This report summarizes our current experiences with Automatic Program Parallelization tools for converting sequential Fortran code for use on a multiprocessor computer. A number of such tools were evaluated, including Parafrase, Adaptor, PAT, Petit and the SUIF compiler package. We evaluated the suitability of such tools for parallelizing Computational Fluid Dynamics code supplied by the Army Research Laboratory, Aberdeen Proving Grounds. SUIF was found to be most suitable by carrying out extensive tests on a suite of test programs and a matrix multiplication program. As a result of these experiments we suggest some modiications to the existing SUIF toolkit for eecient parallelization of the CFD code. Although SUIF does eecient loop partitioning for uniform dependences, it cannot handle nested loops with irregular dependencies eeciently. Unlike the case of nested loops with uniform dependencies these will have a complicated dependence pattern which forms a non-uniform dependence vector set. We propose to incorporate additional passes in SUIF based on results from our previous research, to generate code which will handle applications with non-uniform dependences.