[CPMD-list] Some Numerical Mismatch with CPMD-TEST results in Apple-Mac-OS-X

[Ari P Seitsonen]

Dear Biswas,

  About the difference in the output: For example the FFT grid is
different (14x14x110 vs 16x16x120), that could cause such a tiny deviation
in the energy etc. In addition your job was not run in parallel like the
reference output has been. According to me there's nothing to worry about.

    Greetings from the-rainy-Zurich-again,

       apsi

-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-
 Ari Paavo Seitsonen / Ari.P.Seitsonen at iki.fi / http://www.iki.fi/~apsi/
 Tel +41 1 635 44 97 / Fax +41 1 635 68 38 / GSM +41 79 719 09 35      
 Anschrift: Physikalisch Chemisches Institut (PCI), Universität Zürich (UniZh)
 Indirizzo: Winterthurerstrasse 190, CH-8057 Zürich
 Address:   Schweiz / Svizzera / Suisse / Svizra / Switzerland

On Wed, 11 Aug 2004, Pradip Kumar Biswas wrote:

> Hi,
> 
> After making the CPMD code run on Apple MacOS-X (g4 machine) through some 
> alterations listed below, we find the end-results for Aluminium, for example, are quite 
> close (differ in the 5th place after decimal or above) except the electron density 
> gradient where the difference is most prominent. I attach our output file as well as the 
> CPMD-TEST output we are comparing with. DRHOMAX differ the most.  Below we 
> enlist the changes we made in the code to make it run in Apple MacOS.
> ##(Are they the source of the discrepancy? If so, how we can remove this discrepancy.
> Or nothing to be worried about those discrepancies ????)
> 
> 1. We figure out that the vecLib BLAS routine ZDOTC is leading to the segmentation 
> fault error. To overcome this,
> 
> a)  I changed the calls OVLAP_H in atomwf.f , calc_alm.f , friesner_c.f , friesner_c_p.f 
> by CALL OVLAP_C (which, perhaps, as I understood, does not enforce Hermitian 
> symmetry for the matrix)
> 
> b) The remaining use of ZDOTC in friesner-c.f is diverted by  using a renamed function 
> ZZDOTC defined as: ZZDOTC(NGWK,NKRY_BLOCK*NKRY_MAX,CSCR,IBN1,IBN0)
> 
> and we create a small routine for the function (given below) to make the dot product 
> between the columns given by IBN1 and IBN0.
> 
> 2) In an another attempt, we do not change the CALL OVLAP_H by CALL OVLAP_C 
> but within the SUBROUTINE OVLAP_H transfer the vectors to another routine to 
> perform the dot product. 
> 
> The two sets of results are identical indicating the replacement of the calls OVLAP_H 
> by OVLAP_C does not affect the solution (as perhaps the matrix is Hermitian).
> 
> For 1.b) we take the function as:
> 
>        Function zzdotc(n,m,zx,cj,ck)
> C
>        COMPLEX*16 zx,ztemp,zzdotc
>        DIMENSION ZX(N,M)
>        integer i,n,m,cj,ck
>        ztemp = (0.0d0,0.0d0)
>        zzdotc = (0.0d0,0.0d0)
>        if(n.le.0.or.m.le.0)return
> C
>         do 30 i=1,n
>         ztemp=ztemp+dconjg(zx(i,cj))*zx(i,ck)
>  30     continue
>         zzdotc=ztemp
>         return
>         end
> 
> Regards,
> 
> Biswas.
> 
> --
> Universidade do Vale do Paraíba - UNIVAP.
> http://www.univap.br/
> 
>