Re: [Gomp-discuss] Start of an OpenMP Implementation Spec

[Ross Towle]

On Thu, 26 Aug 2004 12:08:11 -0400, Scott Robert Ladd
<address@hidden> wrote:
> Biagio Lucini wrote:
> > which is where we languish: practical implementation. So far no-one has
> > provided a proof of concept, though many got close, including you and me.
> 
> I'd be willing to work on practical implementation if I had some
> response as to the best direction to take. Given that GCC is a community
> project, I hesitate to simply throw something out and say "here it is"
> without starting from some sort of consensus.
> 
> Also, I believe in design before implementation. To that end, I've made
> very specific proposals and offered sets of solutions; a lack of
> feedback tells me that I must either be a complete doofus, or that no
> one really cares.
> 

I am not sure what you are expecting in terms of  "proof of concept". 
If you mean does the
design hold together, I suggest that you look at my keynote
presentation at EWOMP99
(European Workshop on OpenMP) in Lund Sweden.  In it I discuss a successful 
implementation of OpenMP in a commercial suite of c, c++, and
Fortran90 compilers. The
general design is the same.  In my GOMP proposal, I took advantage of
two things. First,
GOMP does not have to support a pre-OpenMP parallelism model at the same time.
Second, OpenMP2.0 adds some interesting features that designing from
scratch gives a
cleaner design instead of expanding the OpenMP1.0 helper routines.  I
have used the
OpenMP Spec to be the abstract parallelism model which is directly
implimented by the
helper routines.

If you mean "show me the code", I have to agree with Scott that a
design is a REQUIREMENT
before going in and hacking code.  To me it is ESSENTIAL to have this
design.  I know
the amount of work it is going to take.  This is a project where you
must have help from
several other people.  The design is needed to communicate with the others.

I intended my proposal to be at a high enough level to communicate the
key ideas.  More
detail would obscure them.  Unless there is some sort of agreement
about the key ideas,
there is no reason to go into more detailed designs.

In the next few emails, I will deal with some specific questions.

-Ross A Towle