X. Chen
G.M. Megson
University of Newcastle upon Tyne. 1994
This paper discusses methods of transforming sequential URE algorithms to parallel algorithms using an existing partitioning and mapping approach. The features of the so-called supernode domain are studied first, since the original computational body is clustered (to give large granularity) in order to be allocated within a given array. A method of code generation is developed for the situations where M-D processor arrays are given to implement a N-D nested loop structure. The resulting parallel algorithm is characterised by M DO ALL loops in the space domain and a FOR loop in the time domain. Let K = N - M. There are two cases: K > 1 and K = 1, which are treated separately. The generation of parallel code onto a given processor array can be derived by the integration of the partitioning and mapping techniques together with the code generation methods proposed here.