[mpiwg-rma] {MAX,MIN}LOC in RMA

[Jeff Hammond]

> On Mar 13, 2014, at 8:42 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> 
> Jeff Hammond <jeff.science at gmail.com> writes:
> 
>> But this can't be done in hardware easily so how did we ever let it
>> into the RMA chapter? :-)
> 
> How 'bout MPI_PROD for complex?
> 

If NIC does FP for complex, this is fine. M**Loc isn't a commons HW atomic. 

> We use MINLOC to resolve sharing in unstructured meshes.  My RMA
> implementation is no longer default due to buggy MPI implementations of
> RMA datatypes, but we are using that semantic operation.