A classification of algorithms which are well suited for implementations on the DAP as a basis for further research on parallel programming

Abstract

Considering the programs which have been written for the DAP (Distributed Array Processor) by members of our research group or other institutes we are in contact with, one sees quickly that there are a few classes of algorithms which cover nearly all the work done on the DAP.

These classes are:

• Operations from Linear Algebra (matrix multiplication, matrix inversion, solution of large systems of linear equations)

• Bit-algorithms for the fast computation of function values

• Arithmetic on numbers with variable lengths of representation

• Simulations or similar computations on two-dimensional structures taking advantage of the north-east-south-west neighbourhood facilities of the DAP

• "Large scale"-pipelining (in contrast to: "small scale"-pipelining, i.e. pipeline architectures)

In addition to the original DAP facilities some very tricky programming techniques have been developed. With the examples of "linear routing" and "recursive doubling" we shall point out how these techniques can be made available to all users by supplying them with "algorithm patterns" at a somewhat higher language level than provided by DAP-FORTRAN.