Thoughts on OpenMP and other extensions

[Jonathan Day]

Hi,

Apologies if this issue has been dragged to death,
already, but because I'm interested in OpenMP and
other extension to GCC, I thought I'l throw an idea in
the list's general idea.

OpenMP can be (crudely) described as parallelism
through extensions to the pre-processor, compiler,
optimizer and library. This suggests that OpenMP can
be loosely described as four distinct, otherwise
isolated, tasks and labelling the umbrella the
combination falls under.

On this basis, implementing "OpenMP" per-se may be the
wrong way. It may be possible to extend each of the
four levels to permit pluggable extensions. Each
section can then be written wholly independently.

Because this would be much more modular, it would
simplify a lot of the third-pary frontends. It would
also be easier to maintain.

Compiler-level parallel optimizers have been around
since the erly 1970s. They're not in mainstream use,
because although they're easy enough to write, they
are unbelievably hard to write well. There are simply
too many strategies, and not enough people with
experienc on which strategy to use for which
situation.

If OpenMP was implemented as a pluggable extension to
GCC, or a series of such extensions, then those
interested in doing their own reserch will have a
solid framework to design on, plus it would be much
easier to keep up-to-date as the specs change.

Jonathan


		
_______________________________
Do you Yahoo!?
Declare Yourself - Register online to vote today!
http://vote.yahoo.com