Abstract
This chapter focuses on power Fortran toolbox. It presents an analysis of individual loops for parallelization opportunities. Once the critical loops are identified, they are analysed for their parallelization potential. The situations in which loops can be rendered parallelizable generally center on the breaking of data dependence. A number of common situations exist that can be used as templates for making the appropriate changes in the loop. These changes can range from the declaration of the parallel region variable types with no changes in the actual FORTRAN code to significant restructuring of the statements in the loop. Recognizing these common situations can speed up the parallelization process. Sum reduction is both a very common procedure and highly data dependent. If the value of the loop index is needed beyond the parallel region, it is declared to be lastlocal. Lastlocal can be used with scalar variables to gain most of the speedup associated with the local variable type and avoid a data dependence for a scalar variable, provided that the variable needs only its value as determined in the last iteration.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call