[CP2K:3948] pw_grid_assign error

[Noam Bernstein]

On Tue, Jul 31, 2012 at 7:37 AM,  <hut... at pci.uzh.ch> wrote:
> Hi
>
> I have also to guess for this one. Could be an additional
> constraint introduced by the QMMM solver that gets problematic
> for parallel jobs.
> Try
> NGRIDS 1
> or/and use FFTW. If the error still is there, we have
> to take a closer look.

Actually, I've tracked down the problem.  I'd be happy to solve it, but I
want to know how you want it solved.  Basically, my finest grid has
a size of 200, and I have 4 grid levels.  That means 100, 50, and 25
in addition.  The fftsg library, which is used to select grid sizes,
includes radixes of 200, 100, and 25, but not 50.  But when the
compatible grid size check happens (in pw_grid_find_n), it only checks
the top and bottom levels, i.e. 200 and 25, which check out OK.  Then,
when it gets around to making all the subgrids, pw_grid_init_setup fails
because 50 isn't a valid size.

So, I can think of two things that should perhaps be done.  One is
to add 50 to the fftsg list of valid radixes.  I don't understand enough
about the fft algorithm to tell if this is acceptable.  Also, the code in
pw_grid_find_n should check all the subgrids, not just the coarsest
one.  Perhaps it should print a warning if one of the middle subgrids
fails.  I'd be happy to code up both those changes, but I don't want to
do something (e.g. change the list of radixes) that'd break the code.

     Noam