RETRAN - A RECURRENT PARADIGM FOR DATA-PARALLEL COMPUTING

An applicative paradigm of data-parallel computing is proposed which i s based on nonscalar recurrence relations defining lazy sequences of o bjects (streams). It is argued that variable can be localised inside s imple recurrence relations to enable the compiler to infer the propert ies of their associated objects. A mechanism of overloading array oper ators with different-rank versions of an operation is proposed, called anti-currying, which effectively puts all arrays in a single index sp ace. Stream processing units can be connected to form a schema which c an then be transformed into a function and used at a higher level of s tream processing. The only significant constraint that the paradigm im poses on code is the lack of list recursion.