[Carpet] cost of repeated calls on CCTK_SyncGroup()

[Thomas Radke]

Bela Szilagyi wrote:
> From what I understand
> 
> Carpet::SyncProlongateGroups (const cGH* cctkGH, const vector<int>& groups)
> 
> will not be reached from CCTK_Sync* calls.

No, all CCTK_Sync*() calls will eventually call 
Carpet::SyncProlongateGroups(). But in a flesh loop, with only one group 
at a time.
Currently there is no flesh-driver interface to pass a list of groups to 
be synchronised. There is only CCTK_SyncGroup() which must be overloaded 
by a driver. What you want is another overloadable function 
CCTK_SyncGroups().

> If so, then can we 
> 
> 1) modify the flesh CCTK_SyncGroupsI  function to call 
> Carpet::SyncProlongateGroups  once rather than calling Carpet::SyncGroup a 
> nubmer of times, in case multiple groups are lined up for syncronization
> 
> and/or
> 
> 2) provide a simple (aliased function) interface to 
> Carpet::SyncProlongateGroups.
> 
> I would expect that 2) is quicker, though 1) is preferred.  I guess "best" 
> would be both.

The solution should be driver-independent which excludes 2) (unless we 
add an equivalent aliased function in PUGH).
As mentioned above, another overloadable flesh function 
CCTK_SyncGroups() (rather than an aliased function) would be a general 
solution. I see two potential problems though:

  * CCTK_SyncGroup() takes the groupname as an argument, so
    CCTK_SyncGroups() would then take a array of groupnames.
    Is this possible to pass such an array from Fortran ?
  * We could instead add the overloadable function CCTK_SyncGroupsI()
    (which already exists in the flesh API as a wrapper for
    CCTK_SyncGroup()) and pass an array of group indices rather than
    names. This would then break the convention of passing group/variable
    names in overloadable functions.

I think that being able to efficiently synchronise a list of groups at 
the same time via the flesh API is more important than sticking with a 
convention of how to pass grid variables/groups. So let's add this new 
overloadable function CCTK_SyncGroupsI(). I'd write a patch for it if 
nobody objects.

-- 
Cheers, Thomas.