Re: [Gomp-discuss] GOMP Requirements v1.1

[Lars Segerlund]

 BIG +1 to Scott.

 It's not relevant what's right, we have to come up with the goods.

 As Seymore Cray once said, 'You don't have to be right, just to be fastest', ( 
refereing to floating point implementations, and justifying 0+ and 0- which 
happend to be inequal :-) ).

 / Lars Segerlund.

On Fri, 12 Nov 2004 09:41:49 -0500
Scott Robert Ladd <address@hidden> wrote:

> Ross Towle wrote:
> > Under the Effects section is:
> >       in some cases, how code is reorganized depends on the threading
> > model in use
> > 
> > I have a problem with this statement and its implications.
> > 
> > I very strongly believe that the outlined parallel region/work-sharing
> > region and the OpenMP runtime support routines called from the user
> > program should be EXACTLY the same no matter what threading model is
> > used.  If the runtime support routines are modeled after the OpenMP
> > spec model then  the threading model only affects how the support
> > routine is implimented.  To me, supporting multiple thread models and
> > multiple targets is much easier by modifying the guts of the runtime
> > support routines than the changes to be made in the compiler.
> 
> You may be correct.
> 
> However, we're going to take a significant performance hit as compared 
> to commercial compilers. Commercial vendors focus on a single threading 
> model and hardware architecture; they can fine-tune their compilers for 
> the vagaries of (for example) Opteron NUMA systems, accounting for 
> memory architectures and processor characteristics.
> 
> Perhaps I should reword the sentence to say:
> 
>      in some cases, how code is reorganized may depend on the target
>      architecture
> 
> Does that seem more sensivle?
> 
> -- 
> Scott Robert Ladd
> site: http://www.coyotegulch.com
> blog: http://chaoticcoyote.blogspot.com
> 
> 
> _______________________________________________
> Gomp-discuss mailing list
> address@hidden
> http://lists.nongnu.org/mailman/listinfo/gomp-discuss