Hi Guys,
One of the reasons for keeping the locmem method around was for the ease of
broadcasting the whole data structure as a unit in the beginning.
There are ways of doing that now with mpi types which we should move to for
this kind of broadcasting of multiple things without doing multiple
broadcasts.
So I agree that it should go in a module now, but that the broadcast should
be figured out at the same time.
Cheers,
Mike
On 3/18/09 1:35 PM, "Ross Walker" <ross.rosswalker.co.uk> wrote:
> Hi Joe,
>
> I think this has been on the cards for a while, at least adding most things
> to modules and getting rid of all the common blocks etc. The thing that has
> slowed this down in some cases is there is a lot of subtle stuff implied by
> all of this. I.e. the forcing of things to be linear in memory etc. Since a
> lot of the MPI calls assume this to be the case. Hence I and others have
> been burnt by removing a common block and placing it in a module for
> example. The net result being that we have put it back the way it was and
> made a mental note to go back and do it when we have more time which often
> doesn't materialize.
>
> While adding the locmem stuff to a module is a good idea I think really what
> that needs is a complete overhaul. I.e. we should do away with the get stack
> approach etc. Mike Crowley was on his way to doing this before he got side
> tracked by DOE stuff and so it never got finished.
>
> So yes while I would applaud you adding this to a module I think really it
> would be better if we instead took the time to replace it rather than making
> it easy to use in which case the impetus to replace will be reduced and it
> will stay in therefore another decade.
>
> All the best
> Ross
>
>> > -----Original Message-----
>> > From: amber-developers-bounces.ambermd.org [mailto:amber-developers-
>> > bounces.ambermd.org] On Behalf Of Joe Krahn
>> > Sent: Wednesday, March 18, 2009 11:47 AM
>> > To: AMBER Developers Mailing List
>> > Subject: [AMBER-Developers] Fortran module for LOCMEM?
>> >
>> > It would be nice to be able to access LOCMEM globals without recursively
>> > passing the stack arrays from the main SANDER routine, by putting them
>> > into a module. PUPIL already does this, but it would be better to have a
>> > standard module, where SANDER, PUPIL and others can USE the stack arrays.
>> >
>> > The LOCMEM module can also include memory.h, so you can access a stack
>> > index without having to access the whole set of LOCMEM index variable
>> > names.
>> >
>> > It is pretty simple:
>> >
>> > #include "dprec.h"
>> > module memory
>> > #include "memory.h"
>> > _REAL_, dimension(:), allocatable :: x
>> > integer, dimension(:), allocatable :: ix
>> > character(len=4), dimension(:), allocatable :: ih
>> > end module memory
>> >
>> >
>> > then in sander.f
>> > use memory, only: x, ix, ih
>> >
>> > and pupildata.f:
>> > use memory, only: ixStack=>ix, realStack=>x, ihStack=>ih
>> >
>> > Or, skip the "only:" part, and don't #include "memory.h".
>> >
>> > It could also be used to hold general-purpose globals that are not part
>> > of the LOCMEM arrays. In fact, the module could be called "globals"
>> > instead of "memory".
>> >
>> > Joe
>> >
>> > _______________________________________________
>> > AMBER-Developers mailing list
>> > AMBER-Developers.ambermd.org
>> > http://lists.ambermd.org/mailman/listinfo/amber-developers
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Mar 20 2009 - 01:15:46 PDT
