An architecture for fast data movement in the FFP machine

Abstract

We propose an extension to the architecture of the FFP machine which permits storage management in O(log n) time and arbitrary data rearrangement in O(log² n) time, assuming messages are transmitted in unit time independent of wire length. This is achieved using a perfect shuffle sorter network with the FFP machine. The cost is only a constant amount of hardware per atomic symbol. We also propose mechanisms for early termination, that is, easy rearrangements may be performed faster than the worst case time. We propose a hierarchical structure to the FFP machine which is compatible with a modification of the perfect shuffle sorter network. This permits local operations to be faster than global operations, and permits the FFP machine to operate as a collection of independent processors when desired. A method for giving the programmer some control of the use of this hierarchy is proposed. The FFP machine selects certain subexpressions of the input string to be processed during each computation step. These subexpressions are effectively isolated from the rest of the input string during the computation. We show how this isolation may be accomplished in the perfect shuffle sorter network without giving unique names to each such subexpression. This architecture permits fast operations, such as the multiplication of two m by m matrices in O(log² m) time.