The ABINIT Users Mailing List ( CLOSED )

[Masayoshi Mikami <mmikami@rc.m-kagaku.co.jp>]

Salut !

Please let me have a similar report with MacOSX(Tiger, 10.4.9)
and gfortran (version 4.3.0 20061223 (experimental)).

v1-v3 : no fatal problems
v4: t04-t08, the same problem with the printout of
"the total DC energy", even after the remedy by Pierre-Matthieu.
v4:t86, t87 ... sudden death (NB: no problems with v.5.3.3 !)
v4:t88, the big numerical difference
v4: t92, sudden death, t93 many NaN, t94 sudden death
(NB: no problems of v4:t92-94 with v.5.3.3 on the same Mac,
 interestingly ...)
v5: t08, 09; the same problem with the DC total energy...

So something seemed to happen between v.5.3.3 and v.5.3.4 ...
at least for the MacOSX/gfortran.
(I did not try with MacOSX/XLF8.1. How was the results ?)

Cheers,
Masayoshi

On 2007/04/05, at 18:20, Anglade Pierre-Matthieu wrote:

> Hi guys,
>
> For me, under linux (fc6 on x86_64 nocona and core duo), with gfortran
> (4.2.0 20070110 and 4.2.0 20070317) the tests results for abinit 5.3.4
> is as follows:
> V1: OK
> V2: OK
> V3:
> - t19 gives a large discrepancy on an unconverged SCF cycle. This is
> a very old problem and maybe a bug.
> -  t83 has too many differents line but is in fact perfectly OK
> V4:
> T06, t07, t08, have a problem at printing the Total DC energy in eV.
> This is a bug. It is easily solved by  changing in 15comon/prtene.F90
> line 279 from:
> if (optdc==2) message=""
> to
> if (optdc==2) message="   "
> or by removing this line.
> The "" is misunterpeted by gfortran and causes a bug that appears only
> for this compiler (we may think the bug is in fact a gfortran bug). It
> hought I had commited a fix for that. Apparently this was not the
> case.
>
> t88: This is a real abinit/gfortran bug for GW self consistent. It is
> known since v5.3.2 (at least) and matteo giantomassi is investigating
> the case.
>
> V5:
> t08: same PAW problem as v4 t06-08
>
> So, unless you are conducting self-consistent GW calculations gfortran
> compiled (latest gfortran 4.2 I mean) is safe with abinit 5.3.4.
> What is the version o gfortran you use Nuno ? I
> reagards
>
> PMA
>
> On 4/4/07, Anglade Pierre-Matthieu <anglade@gmail.com> wrote:
> > HI,
> >
> > Those tests are (mostly v3, v4) PAW and GW tests. They undergoes
> > problem with gfortran before the final merges. Yet I was thinking  
> > most
> > of the problems with gfortran were fixed during 5.3.2-> 5.3.3. and
> > then fixed better again in the 5.3.3 -> 5.3.4 by Marc and Matteo. I
> > haven't already checked what's up with 5.3.4.+gfortran.
> > I'll tell you tomorrow.
> >
> > regards
> >
> > PMA
> >
> > On 4/4/07, Nuno A. G. Bandeira <nuno.bandeira@ist.utl.pt> wrote:
> > > gay@sfu.ca wrote:
> > > > I found recent versions of gfortran gave big errors on
> > > > tests v2, 61-66, which did not happen with earlier (Oct 06)
> > > > versions of gfortran. Do you find this?
> > > >
> > > Yes I do ! v2-21,58,61-66 have a failure report. Also  
> > v3-30,31,57 and 97
> > > fail and so too do v4-8,57,84-88 and v5-63.
> > > Does anyone on the development team see a pattern or are these  
> > tests
> > > totally unrelated ?
> > >
> > > I had noticed this with previous editions of abinit but since they
> > > weren't so many compared to the full range of tests I didn't   
> > pay too
> > > much attention.
> > >
> > > It has also been my experience that g95 makes pretty good  
> > binaries but
> > > they are painfully slow and the people at AMD have yet to make a  
> > g95
> > > compiled ACML for Windows. The one I use is a special hack of a g77
> > > binary libacml to become usable with gfortran. ACML makes it run  
> > much
> > > faster. The benchmark tests are run from an x86-64 platform and my
> > > humble pentium 4 can beat the timings for most of them.
> > >
> > > pgf90 is comparable in performance to gfortran+acml.
> > >
> > > --
> > > Nuno A. G. Bandeira, AMRSC
> > > Graduate researcher and molecular sculptor
> > > Inorganic and Theoretical Chemistry Group,
> > > Faculty of Science
> > > University of Lisbon - C8 building, Campo Grande,
> > > 1749-016 Lisbon,Portugal
> > > http://cqb.fc.ul.pt/intheochem/nuno.html
> > > Doctoral student @ IST,Lisbon
> > > --
> > >
> > >
> >
> >
> > --
> > Pierre-Matthieu Anglade
>
>
> --
> Pierre-Matthieu Anglade