G.M. Megson
X. Chen
University of Newcastle upon Tyne. 1993
A methodology is proposed to partition and map an arbitrary computation graph onto a given regular array under the URE condition. It begins with a positive expressing basis to obtain canonical dependencies. Based on the canonical dependencies, two basis models of space projections, as well as, timing vectors are derived. The partitioning parallelepiped is scaled to map the original polytope within the given array. An LSGP method is used to improve efficiency. This methodology has significant advantages in mapping an arbitrary computation graph onto a given processor array while having high efficiency on both of communication and computation without extra restrictions.