[MPIWG Fortran] Profiling for MPI_SIZEOF?
Jeff Hammond
jeff.science at gmail.com
Thu May 15 16:00:05 CDT 2014
>>> That said, the optimal implementation would be to write all of the specific versions (I’m seeing 10) for a generic named MPI_SIZEOF. For each specific version, the executed code reduces to
>>>
>>> size = storage_size(x)/8
>>>
>>> which the compiler will evaluate at compile time based on the type of x.
>>
>> No, there's 10 * 7 versions, right? One for each type (10), and then 1 for each rank.
>>
>> The language in the MPI spec explicitly allows you to pass an array as the argument of x (although it does say that it will return the size of the base element -- not the array).
>>
>
> You are right. I just looked at one of the example specifics:
>
> subroutine MPI_Sizeof_xint8 (x, size, ierror)
> use,intrinsic :: iso_fortran_env, only: int8
> integer(int8),dimension(..) :: x
> INTEGER, intent(out) :: size
> INTEGER, optional, intent(out) :: ierror
>
> size = storage_size(x)/8
> if (present(ierror)) ierror = 0
> end subroutine MPI_Sizeof_xint8
>
> which is declared assumed-rank, but specific type (which is the interesting thing here). That’s how I got 10. To get them all without assumed-rank, you need 10*15 (the Fortran standard allows array ranks up to 15 now). All the more reason for no intercepting.
I vote for #1. Does anyone thing that there are linkers that can
handle 10^15 symbols?
Jeff
--
Jeff Hammond
jeff.science at gmail.com
http://jeffhammond.github.io/
More information about the mpiwg-fortran
mailing list