COMPILING DISTRIBUTED LOOPS ONTO SPMD CODE
Vincent Van Dongen · Parallel Processing Letters · 1994
Given a loop written in a sequential language and a block-cyclic distribution, we present some techniques for generating SPMD code to run on distributed-memory machines. This is illustrated on three types of loops: parallel loops without internal dependence, parallel loops with internal dependences, and loops with hidden parallelism. Two models of distributions are considered: the data distribution and the computation distribution. We argue here that the data distribution model together with the owner-computes rule is only well adapted for the first class of loops. For the other loops, we present the idea of using a computation distribution instead, and we show how this can be compiled for distributed-memory machines.