[petsc-dev] Custom reductions

[Barry Smith]

> On Apr 21, 2015, at 11:25 AM, Håkon Strandenes <haakon at hakostra.net> wrote:
> 
> The MPI_MIN, MPI_MAX and MPI_SUM only indicate that we want to have the smallest or largest value, or in case of MPI_SUM, the sum of all values participating in the reduction. It has nothing to do with floting point or integer types as such, they work with both.

   However as Jed notes MPI_MIN, MPI_MAX and MPI_SUM  do not work with __float128, hence to have PETSc code that works for --with-precision=__float128 one must use MPIU_MIN, MPIU_MAX, and MPIU_SUM for PetscReal and PetscScalar 


> 
> (In addition there is some boolean operations like MPI_LAND, MPI_LOR, MPI_LXOR, MPI_BAND, MPI_BOR that work with boolean types, and some operations like MPI_SUM and MPI_PROD can also be used with complex types, but I don't know if PETSc use this at all.)
> 
> Håkon
> 
> 
> On 21. april 2015 18:11, Matthew Knepley wrote:
>> I understood from Jed that MPIU_MIN and friends are for floating point,
>> but here
>> 
>> https://bitbucket.org/petsc/petsc/src/da407576a5bf29881cba0f2df6c198aa8caa98f0/src/sys/logging/plog.c?at=master#cl-1346
>> 
>> we use MPI_MIN. Is this wrong?
>> 
>>   Thanks,
>> 
>>      Matt
>> 
>> --
>> What most experimenters take for granted before they begin their
>> experiments is infinitely more interesting than any results to which
>> their experiments lead.
>> -- Norbert Wiener