Re: [gomp] Challenges in Implementing OpenMP

[Diego Novillo <dnovillo at redhat dot com>]

On Thu, 2004-10-21 at 08:09, Gabriel Dos Reis wrote:

> Allowing programmers to communicate more information for code
> generation purpose is good -- I don't think anyone is saying that is a
> bad idea.  However, I'm less convinced that the way to go is for GCC
> to pick that particular model.
> 
One thing to remember in this timeless debate of parallel paradigms is
that supporting a low-level abstraction such as OpenMP in the compiler
does not necessarily tie GCC to this particular model.  One can always
implement higher-level abstractions on top of it.

OpenMP is mostly just a threading library that happens to be built into
the compiler.  The big important advantage of having parallelism
expressed in such painstaking detail is that the optimizers know exactly
what they can and cannot do.  With a pure library approach, you pretty
much have to consider everything volatile.  If the compiler knows what's
going on, it can make better optimization decisions.

Note: I personally never found higher level abstractions to be all that
useful, particularly when trying to extract the last nanosecond of
performance out of the system.  But, hey, to each his own.  I've had
this religious discussion many times over the years.  It's fun, but it
ultimately boils down to personal preferences and the specific problem
you are trying to solve.


Diego.