[molpro-user] compiling problems/bug: multi-cpmcscf gradients with >50 atoms

Andy May MayAJ1 at cardiff.ac.uk
Mon Oct 21 23:07:18 BST 2013 

Sven,

Okay, I managed to reproduce the problem now and agree it only appears 
to happen with MKL, but without (-noblas -nolapack explicitly) the 
problem is not seen. I've made a bug report here:

https://www.molpro.net/bugzilla/show_bug.cgi?id=4195

The reason you aren't prompted for LAPACKLIB by configure when using MKL 
is that it knows the BLASLIB options for this version of MKL contain a 
full LAPACK library, so there is no need for further options. This 
wouldn't be the case if for instance BLASLIB were '-L/usr/lib64 -lblas'.

The fciqmc failures are known, actually I see more with ifort 14 
compared to 13, but they are reported here:

https://www.molpro.net/bugzilla/show_bug.cgi?id=4168
https://www.molpro.net/bugzilla/show_bug.cgi?id=4047

anyway.

Best wishes,

Andy

On 21/10/13 13:47, Sve N wrote:
> Thank you for your reply,
> I wasn't sure what things I did before, exactly, so I recompiled it as
> serial version without doing anything extra, last week - the problem is
> not the parallelism (the openmp part was, to be able to use the parallel
> mkl-libraries - I have some general questions there, but I'll ask them
> later, independent of this topic).
>
> This time I configured with:
> ./configure -prefix $insdirpath -blas -lapack -icc -ifort -blaspath
> /path/intel/mkl111/lib/intel64 -lapackpath /path/intel/mkl111/lib/intel64
> and used the default for every question asked, thereafter. (This, btw.
> results in not having a lapackpath set, there isn't even a question for
> this, and the entry in CONFIG is empty.)
>
> I used ifort 14.0 now, as it's listed as supported for linux x86_64, and
> again the mkl 11.1 (which is not, but from prior tests I know, is quite
> a bit faster). The compilation runs through normally and every test
> passes, except, again, the three fciqmc ones. And again I get the
> gradient error.
> I'll attach the CONFIG file, the errout of be2_odds_fciqmc and the
> out-file of my he-chain.
>
> I just compiled it without using mkl (so just ./configure -prefix
> $insdirpath -icc -ifort and hitting enter for all questions), which also
> gave a binary that ran the first testjob fine (I didn't do the whole
> suite, now), but failed to run be2_odds_fciqmc.
> Sven
>
>
>  > Date: Wed, 16 Oct 2013 10:02:36 +0100
>  > From: MayAJ1 at cardiff.ac.uk
>  > To: goretoffel at hotmail.de; molpro-user at molpro.net
>  > Subject: Re: [molpro-user] compiling problems/bug: multi-cpmcscf
> gradients with >50 atoms
>  >
>  > Sven,
>  >
>  > Is there any reason you have changed FFLAGS and CFLAGS? We only use the
>  > Fortran bindings of BLAS/LAPACK so I don't think the the -I options, or
>  > -DMKL_ILP64 will make any difference, and -i8 should be default on
>  > 64-bit anyway. The only flags I think are making a difference are the
>  > -openmp flags, although it is more convenient to just specify -openmp
>  > directly to configure and let it do the rest. I would suggest to try
>  > building without openmp since there is hardly any openmp code in Molpro,
>  > certainly the binaries are built without openmp.
>  >
>  > Best wishes,
>  >
>  > Andy
>  >
>  > On 15/10/13 14:10, Sve N wrote:
>  > > Dear molpro users and developers,
>  > > when compiling molpro myself, it finishes all testjobs of 'make test',
>  > > except the three fciqmc ones, properly (the fciqmc jobs finish running,
>  > > too, but end up giving too large errors in the energy, something like
>  > > 0.0006). Since I do not plan on using it, it didn't seem important to
>  > > me. My most recent configure-options (all of this mail holds true for
>  > > patch 0, 1 and 6 with various options I tried) are:
>  > >
>  > > ./configure -prefix $insdirpath -blas -lapack -icc -ifort -mpp
>  > > -auto-ga-mvapich2 -var 'BLASLIB=-Wl,--start-group
>  > > /path/intel/mkl111/lib/intel64/libmkl_intel_ilp64.a
>  > > /path/intel/mkl111/lib/intel64/libmkl_core.a
>  > > /path/intel/mkl111/lib/intel64/libmkl_intel_thread.a -Wl,--end-group
>  > > -lpthread -lm' -var 'LAPACKLIB=-Wl,--start-group
>  > > /path/intel/mkl111/lib/intel64/libmkl_intel_ilp64.a
>  > > /path/intel/mkl111/lib/intel64/libmkl_core.a
>  > > /path/intel/mkl111/lib/intel64/libmkl_intel_thread.a -Wl,--end-group
>  > > -lpthread -lm' -var 'CFLAGS=-DMKL_ILP64 -openmp
>  > > -I/path/intel/mkl111/include' -var 'FFLAGS=-i8 -openmp
>  > > -I/path/intel/mkl111/include'
>  > >
>  > > The compilers are ifort and icc 13.1, mkl's version is 11.1. The two
>  > > cpus I let it run on, are an Intel Xeon E5-2630L and an Intel Xeon
>  > > L5630, although it is compiled on an AMD Opteron 6128 HE (is that a
>  > > problem?).
>  > >
>  > > When doing a gradient calculation in a multi-block with more than 50
>  > > atoms, I get wrong gradients for the atoms after the 50th. This doesn't
>  > > happen with the prebuilt binaries, so it has to be a compiling problem.
>  > > I have no clue, how to fix this, so if you have any advice, I'd be
> thankful.
>  > >
>  > > As an example I calculated a He-chain, with the He-atoms seperated by 3
>  > > Angstrom, and little deviations in the y-direction.
>  > >
>  > > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>  > > ***
>  > > memory,50,m
>  > >
>  > > basis
>  > > default,6-31g
>  > > end
>  > >
>  > > angstrom
>  > > nosym
>  > > orient,noorient
>  > > geomtyp=xyz
>  > > geometry={
>  > > 52
>  > >
>  > > he 0. 0. 0.
>  > > he 3. 1. 0.
>  > > he 6. 0. 0.
>  > > [...]
>  > > he 147. 1. 0.
>  > > he 150. 0. 0.
>  > > he 153. 0. 0.
>  > > }
>  > >
>  > > hf
>  > >
>  > > force
>  > >
>  > > {multi
>  > > closed,51
>  > > occ,53
>  > > state,2
>  > > cpmcscf,grad,1.1
>  > > }
>  > >
>  > > force
>  > >
>  > > ---
>  > > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>  > >
>  > > This results in (regardless of the number of processors used):
>  > > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>  > > 1PROGRAM * RHF-SCF (CLOSED SHELL) Authors: W. Meyer, H.-J. Werner
>  > >
>  > >
>  > > NUMBER OF ELECTRONS: 52+ 52- SPACE SYMMETRY=1 SPIN
>  > > SYMMETRY: Singlet
>  > > [...]
>  > > 1PROGRAM * ALASKA (Gradient of the energy) Author: R. Lindh
>  > > [...]
>  > > SCF GRADIENT FOR STATE 1.1
>  > >
>  > > Atom dE/dx dE/dy dE/dz
>  > >
>  > > 1 0.000002476 0.000000825 0.000000000
>  > > 2 -0.000000001 -0.000001650 0.000000000
>  > > [...]
>  > > 49 -0.000016545 0.000000825 0.000000000
>  > > 50 -0.000000000 -0.000001650 0.000000000
>  > >
>  > > 51 0.000016546 0.000000825 0.000000000
>  > > 52 -0.000019022 0.000000000 0.000000000
>  > > [...]
>  > > 1PROGRAM * MULTI (Direct Multiconfiguration SCF) Authors: P.J.
>  > > Knowles, H.-J. Werner (1984) S.T. Elbert (1988)
>  > > [...]
>  > > 1PROGRAM * FORCE (Gradient of the energy)
>  > >
>  > > SA-MCSCF information from 5100.1
>  > > [...]
>  > > SA-MC GRADIENT FOR STATE 1.1
>  > >
>  > > Atom dE/dx dE/dy dE/dz
>  > >
>  > > 1 0.000002476 0.000000825 0.000000000
>  > > 2 -0.000000001 -0.000001650 0.000000000
>  > > [...]
>  > > 49 -0.000016545 0.000000825 0.000000000
>  > > 50 0.000000000 -0.000001650 0.000000000
>  > >
>  > > 51 0.000016546 0.000000825 289.128098107
>  > > 52 -0.000019022 0.000000000 289.128098107
>  > > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>  > > (The blank line between atom 50 and 51 appears in the out-file, like
>  > > it's shown here.)
>  > >
>  > > Thanks, Sven
>  > >
>  > >
>  > > _______________________________________________
>  > > Molpro-user mailing list
>  > > Molpro-user at molpro.net
>  > > http://www.molpro.net/mailman/listinfo/molpro-user
>  > >
>
>
> _______________________________________________
> Molpro-user mailing list
> Molpro-user at molpro.net
> http://www.molpro.net/mailman/listinfo/molpro-user
>

More information about the Molpro-user mailing list