On Nov 10, 2011, at 4:10 PM, Jed Brown wrote:

> On Thu, Nov 10, 2011 at 16:05, Barry Smith <bsmith at mcs.anl.gov> wrote:
>   You could have ./configure check for MPI_IN_PLACE and generate an error if 
> it doesn't exist.  But do we want to say PETSc requires MPI 2.0 (or at least 
> parts of it)????  Since we already have support for not having MPI_IN_PLACE 
> I'd only like to remove that extra support if we knew that all (sane) users 
> had access to MPI_IN_PLACE. Do we know this?
> 
> If we found an implementation that doesn't have it, we could define the 
> MPI_IN_PLACE value and have our own wrapper that did the right thing 
> (allocated the work array).

   I don't see this being simple. We would need to implement many MPI functions 
that take MPI_IN_PLACE.

>  
> 
> 
> > In general, can we rely on any MPI-2.0 features that have been implemented 
> > in mpiuni and remove the macro checking?
> 
>    I don't have a clue what parts of MPI-2.0 that MPIUni does or does not 
> support :-(.  I'm afraid we'll need to find out one operation at a time. It 
> definitely doesn't have the one-sided stuff, can/should that be added? I 
> don't see a big upside in putting the one-sided stuff into MPI uni.
> 
> MPI one-sided is normally used to talk to remote processes, much like 
> MPI_Send, which we currently define to produce an error if called. I think it 
> would make sense to do the same thing with the one-sided stuff.


Reply via email to