The ABINIT Users Mailing List ( CLOSED )

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

Dear Marc and Matthieu,
        
Merci bien de vos commentaires !
Matthieu, sure, I did not change the compiler etc
between the v.5.3.x and v.5.2.x.
(I did not see big difference in the config.log (and no -O0 ;-)),
 which may be understandable the speed of "fourwf(pot)"
 did  not changed so much) , whereas I saw the speed of
 the nonlop is so slow with v.5.3.4.

Please let me have time to run the jobs suggested by Marc.

I should have test cases that I could disclose here...
(please let me think ...) Still, I could give some hints:
My model contains relatively large number of atoms
(over 20) and relatively large nbands (over 70)
due to pseudopotentials with "shallow core (s & p)".

Bien a vous, et bon weekend,
Masayoshi

On 2007/04/26, at 19:23, Marc Torrent wrote:

> Hi Masayoshi and Matthieu,
>
> You called for a PAW developper; here am I...
>
> I did a diff on abinit-5.2.4/src/03nonlocal and abinit-5.3.4/src/ 
> 13nonlocal directories and didn't see anything changed for nonlop  
> routine (and children), when it uses Legendre polynomials (which  
> the case for Masayoshi tests because he uses norm-conserving  
> pseudopotentials).
> Of course spherical harmonics version of nonlop has changed but it  
> is not called AT ALL in that case; and, it has been improved for  
> speed (at least on our machines).
> The only interaction between nonlop_pl and nonlop_ylm is the driver  
> "nonlop" which has, in v5.3.x, new arguments (like <p_i|cnk>  
> projected scalars) but the later are not transmitted to nonlop_pl.  
> Could it be possible that a memory problem occurs ? I don't think  
> so because these <p_i|cnk> are zero-dimensioned in norm-conserving  
> case.
>
> Masayoshi, could you perform following tests ?:
> - put useylm=1 in your input and see if Ylm version of nonlop has  
> the same behaviour as Legendre polynomials one.
> - try to follow the changes by testing v5.2.3, v5.2.4, v5.3.0,  
> v5.3.2 and v5.3.4...
>
> This could help us to track the problem...
>
>
> Cheers,
> Marc
>
>
> Matthieu Verstraete a ñÄrit :
> > This is horrendous! An order of magnitude slower!!!?
> > I trust mikami-san that your compilation options were the same,  
> > and linking to the same libs, same compiler version etc??? Has  
> > anyone else reproduced this on other machines?
> > The main changes to nonlop were probably paw related. Are you  
> > using this, and do the PAW developpers have any comment (maybe  
> > they haven't touched it at all)?
> > Matthieu