In the third part of his 12-part blog series about structured patterns as a basis for high-level parallel programming, Intel’s Michael McCool writes about the map pattern. If you disallow dependencies between loop iterations, he writes, you get a type of computation that can be parallelized easily: the map pattern. Here, a set of loop indices are generated and independent computations are performed for every unique index. According to McCool, these computations are not allowed to communicate with one another. At the end of the map, however, there is an implied barrier, and then other operations can depend on the output of the map operation as a whole.
To read the full blog post, click here.