On Sat, 28 Aug 1999, Carsten Pfeiffer wrote:
>
>
>Does it use MMX eventually?
>

BLAS libraries always make sure they perform the best possible on a computer
system.

Off course, only BLAS 3 achieves full speed on a computer system.
  
One of the things BLAS does is too make sure no cache trashing occurs. This
feature only will result in large speed gains.

>> MPICH is also a good candidate. Good implementations of MPI should scale on SMP
>> systems and clusters (if communication links > 100Mb are used). I am no expert
>> on MPI, .... so I guess we will need to contact someone else.
>
>I think we agreed not to use MPI because its use would be too limited in
>comparison to the effort and overhead involved.
>I mean, how many people have clusters for image editing? KIS' target
>(IMHO) is single computers with one or more CPUs, where MPI doesn't help
>much.
> 

Some MPI implementation do not need a cluster. They just model a SMP processor
using MPI concepts. This means that you write for a virtual MPI system. If you
are running kimageshop on an SMP system it will take full advantage of the SMP
system. However, the same codebase with a different MPI implementation library
(the interface, header files are the same) will port in 1 second (the time to
link the new MPI lib) to a cluster of 1000 PII .....

>> I work at the computer science departement of the university of louvain
>> (Belgium). I will ask our local MPI guru's for some information.
>
>Ok, if you get some real arguments pro MPI-usage, please tell me.
>

I am not pro MPI nor contra MPI. I am pro the-best-solution, if that is MPI
than it is MPI. If our MPI guru says MPI is not suitable for this task then we
can silently ignore MPI.


BDR