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

[Ross Towle]

On Fri, 12 Nov 2004 12:07:13 +0100, Lars Segerlund
<address@hidden> wrote:
> 
>  What you say might generally be true, but not if performance considerations 
> are taken into account, on a NUMA system for example.

Actually the direction I outlined does work for SMP and ccNUMA
systems. There are a number of commercial implimentations where the
user code is the same, the same calls are made to the OpenMP runtime
support, and the details are hidden in those routines.  And there is
performance.  If the target is specified then the actual schedule of
instructions in the outlined routine may be different and different
traditional optimizations may/may not be performed but the result of
the "lowering" of OpenMP constructs is independent of SMP or ccNUMA.
> 
>  Also if you have a look at the intel openMP compiler it certainly takes the 
> architecture of the target in mind, generate some code for a single proc 
> multi threading processor and a SMP multiproc, any differances ?

Since I don't have access to the Intel compilers (IA32 or IA64) right
now I will defer answering to next week.

> 
>  Anyhow I don't think it's very relevant for the implementation at this time, 
> since it's a long way to go before we reach the point at which we can start 
> to worry about details like that even if it's quite relevant.

But look at the requirements document from the point of somebody not
working on GOMP but working in the "middle end".  It might raise some
concerns that do not need to be raised.
> 
>  / regards, Lars Segerlund.
Cheers - Ross