Hi,
On Wed, 1 Apr 2009, Joe Krahn wrote:
> lmod.f defines do_list_update as a parameter, but it is actually an
> intent(out) argument that gets passed down to nonbond_list(), where an
> attempt to modify the argument produces a segfault. Apparently, this part of
> the code is not getting run in the test scripts.
>
> There are also several instances of do_list_update as a parameter in debug.f.
> In looking at the code, I'm not sure that do_list_update is even useful. It
> seems to get set inside nonbond_list(), and is otherwise ignored. Maybe it
> was originally intent(in) to force an update?
Circa 2004, I had detailed control over list building in lmod.f:
select case ( list_control )
case ( NBL_UPDATE_ALWAYS )
do_list_update = .true.
ntnb = 1
case ( NBL_UPDATE_NEVER )
do_list_update = .false.
ntnb = 0
case ( NBL_UPDATE_NORMAL )
do_list_update = .true.
if ( mod(n_force_calls,nsnb) == 0 ) then
ntnb = 1
end if
case default
! invalid list_control
ASSERT( .false. )
end select
which I think was correct; at least lmod worked in 2004.
All of this was removed in 8.8. And then do_list_update was re-added as a
parameter in 8.9. After only a few minutes of looking at the do_list_update
control flow, it looks ok to just remove the parameter and see if it works.
A more thorough investigation is merited. Perhaps Ross and Mike can comment.
RCS file: /cvsroot/amber11/src/sander/lmod.f,v
revision 8.8
date: 2005/09/17 08:23:16; author: rcw; state: Exp; lines: +2 -6
RCW: Major performance updates to egb.f (>25%). gb_cox2 on 1cpu landau (ifort 8.1, mkl6.1). Amber 8 = 627.88s, Amber 9 17th Sept 2005 = 437.50s. This has been done by splitting the calculation of radii and optimising this routine seperately. Make much more use of vectorisation. Also re-ordered array for more cache hits. Also store and reuse inverse effective radii to avoid unecessary divisions. Improved initial part of main egb routine. Moved egb_les
out of egb.f and into its own file. Fixes to debug.f to make force calls consistently use the correct number of arguments. Removed do_list_update from force subroutine (it was never
used). Performance updates for qmgb=2. More egb performance updates to follow shortly.
revision 8.9
date: 2005/10/17 19:45:48; author: crowley; state: Exp; lines: +4 -2
MFC Psander code fixed, working for dhfr to 32 processors. test cases work.
As far as testing:
RCS file: /cvsroot/amber11/test/Makefile,v
revision 9.98
date: 2008/02/29 21:42:17; author: case; state: Exp; lines: +7 -7
don't run the lmod tests: they don't really "test" anything know (other than
that the program doesn't crash...)
comment out the tgtmd tests, until Carlos (or someone) can figure these out
I'll turn these back on after lmod.f is sorted out.
BTW, here is another plug for xmin: try it instead of the old Amber
minimization methods, you'll like it.
thanks for the report, Joe.
Scott
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Apr 03 2009 - 01:10:18 PDT
