An algorithm can be modeled as an index set and a set of dependence vectors. Each index vector in the index set indexes a computation of the algorithm. If the execution of a computation depends on the execution of another computation, then this dependency is represented as the difference between the index vectors of the computations. The dependence matrix corresponds to a matrix where each column is a dependence vector. An independent partition of the index set is such that there are no dependencies between computations that belong to different blocks of the partition. This report considers uniform dependence algorithms with any arbitrary kind of index set and proposes two very simple methods to find independent partitions of the index set. Each method has advantages over the other one for certain kind of application, and they both outperform previously proposed approaches in terms of computational complexity and/or optimality. Also, lower bounds and upper bounds of the cardinality of the maximal independent partitions are given. For some algorithms it is shown that the cardinality of the maximal partition is equal to the greatest common divisor of some subdeterminants of the dependence matrix. In an MIMD/multiple systolic array computation environment, if different blocks of ail independent partition are assigned to different processors/arrays, the communications between processors/arrays will be minimized to zero. This is significant because the communications usually dominate the overhead in MIMD machines. Some issues of mapping partitioned algorithms into MIMD/systolic systems are addressed. Based on the theory of partitioning, a new method is proposed to test if a system of linear Diophantine equations has integer solutions.
Date of this Version