[CPMD-list] ovlap2_c

[Juerg Hutter]

Hi Thierry

usually xGEMM routines perform at the highest possible
rate on all computers. However, there might be the following
problems.
A bad implementation of ZGEMM on this computer. You could test
this by comparing to the DGEMM performance.
The other possibility is that in the FEMD code you call
ZGEMM very often with small dimensions. The VPP 5000 is
a vector machine that might not achieve peak performance
in this case.
A possible solution in this case would be to add special
code for N=1,2,... . There is already a special
code for the case of Matrix*Vector multiplication instead
of Matrix*Matrix (N=1).

greetings

Juerg

----------------------------------------------------------
Juerg Hutter                   Phone : ++41 1 635 4491
Physical Chemistry Institute   FAX   : ++41 1 635 6838
University of Zurich           E-mail: hutter at pci.unizh.ch
Winterthurerstrasse 190
CH-8057 Zurich, Switzerland
----------------------------------------------------------


On Tue, 27 May 2003, Thierry Deutsch wrote:

> Hello,
>
> 	For an aluminum interface (174 atoms, cutoff of 10Ry 310 states, free energy
> functional), I have the following behavior on a Fujitsu VPP 5000:
>
>  SUBROUTINE            CALLS         CPU TIME        ELAPSED TIME
>    OVLAP2_C             9060         30066.36            30067.66
>     EHPSI_C             8457          9696.90             9740.52
>     FFT-G/S          3038922          4757.54             4842.34
>  FRIESNER_C              240          4010.71             4021.63
>     S_FWFFT           725161          3545.42             3609.48
>    S_INVFFT           794300          3022.41             3091.97
>     OVLAP_H              881          2686.38             2686.41
>   CALC_BILN             8457          2379.40             2379.68
>       EVPSI           515741          1011.01             1033.20
>       RGS_C              639           638.15              638.18
>       VBETA              240           526.16              538.59
>
> You can see that ovalp2_C is used a lot.
> In this routine ZGEMM is used. I think it is not possible to optimize this
> routine.
> May I wrong?
>
> Au revoir,
>
> Thierry
> _______________________________________________
> CPMD-list mailing list
> CPMD-list at cpmd.org
> http://www.cpmd.org/mailman/listinfo/cpmd-list
>