 Re: openmx runtime questions ( No.1 ) |
- Date: 2005/12/08 00:15
- Name: peter
- Here is the followup with the code-fixes:
---------------------------------
double VNAF(int Gensi, double R)
{
/*
..............
ORIGINAL CODE
..............
*/
/****************************************************
Spline like interpolation
****************************************************/
if (po==0){
// h1 = Spe_VPS_RV[Gensi][m-1] - Spe_VPS_RV[Gensi][m-2]; ERROR - ACCESS TO UNALLOCATED MEMORY
h2 = Spe_VPS_RV[Gensi][m] - Spe_VPS_RV[Gensi][m-1];
// h3 = Spe_VPS_RV[Gensi][m+1] - Spe_VPS_RV[Gensi][m]; ERROR - ACCESS TO UNALLOCATED MEMORY
// f1 = Spe_Vna[Gensi][m-2]; ERROR - ACCESS TO UNALLOCATED MEMORY
f2 = Spe_Vna[Gensi][m-1];
f3 = Spe_Vna[Gensi][m];
// f4 = Spe_Vna[Gensi][m+1]; ERROR - ACCESS TO UNALLOCATED MEMORY
/****************************************************
Treatment of edge points
****************************************************/
if (m==1)
{
h3 = Spe_VPS_RV[Gensi][m+1] - Spe_VPS_RV[Gensi][m];
f4 = Spe_Vna[Gensi][m+1];
h1 = -(h2+h3);
f1 = f4;
}
else if (m==(Spe_Num_Mesh_VPS[Gensi]-1))
{
h1 = Spe_VPS_RV[Gensi][m-1] - Spe_VPS_RV[Gensi][m-2];
f1 = Spe_Vna[Gensi][m-2];
h3 = -(h1+h2);
f4 = f1;
}
else
{
h1 = Spe_VPS_RV[Gensi][m-1] - Spe_VPS_RV[Gensi][m-2];
h3 = Spe_VPS_RV[Gensi][m+1] - Spe_VPS_RV[Gensi][m];
f1 = Spe_Vna[Gensi][m-2];
f4 = Spe_Vna[Gensi][m+1];
}
/*
..............
ORIGINAL CODE
..............
*/
}
double PhiF(double R, double *phi0, double *MRV, int Grid_Num)
{
/*
..............
ORIGINAL CODE
..............
*/
/****************************************************
Spline like interpolation
****************************************************/
if (po==0){
// h1 = MRV[m-1] - MRV[m-2]; ERROR - ACCESS TO UNALLOCATED MEMORY
h2 = MRV[m] - MRV[m-1];
// h3 = MRV[m+1] - MRV[m]; ERROR - ACCESS TO UNALLOCATED MEMORY
// f1 = phi0[m-2]; ERROR - ACCESS TO UNALLOCATED MEMORY
f2 = phi0[m-1];
f3 = phi0[m];
// f4 = phi0[m+1]; ERROR - ACCESS TO UNALLOCATED MEMORY
/****************************************************
Treatment of edge points
****************************************************/
if (m==1)
{
h3 = MRV[m+1] - MRV[m];
f4 = phi0[m+1];
h1 = -(h2+h3);
f1 = f4;
}
else if (m==(Grid_Num-1))
{
h1 = MRV[m-1] - MRV[m-2];
f1 = phi0[m-2];
h3 = -(h1+h2);
f4 = f1;
}
else
{
h1 = MRV[m-1] - MRV[m-2];
h3 = MRV[m+1] - MRV[m];
f1 = phi0[m-2];
f4 = phi0[m+1];
}
/*
..............
ORIGINAL CODE
..............
*/
}
 |
| Re: openmx runtime questions ( No.2 ) |
- Date: 2005/12/08 19:54
- Name: T.Ozaki
- Hi,
> A very substantial time openmx does "Fourier transform of PAO
> and projectors of VNL".
Its portion of the computational time depends on the system size.
Let me know the typical size of your 100 molecules.
> Do I understand right that these quantities
> can be pre-calculated for my molecules set provided that I know the
> atoms types in the set and all the molecules are calculated within
> the same size fft grid (all of my molecules are roughly of the same size)?
> Are there other "constant" quantities you could reccomend for pre-calculation
> and storing?
The Fourier-transformed quantities generated in "FT_*" can be pre-calculated
and re-used. if your systems consist of less than 30 atoms, then such a
scheme is maybe effective for reduction of the computational time.
But, I am not planning to implement the capability soon, while your contribution
is wellcome. If so, please e-mail me directry. Then, I will send you the latest
version of OpenMX.
> My MPI environment is sensitive to memory misallocations. There are
> few small "problems" in your code. Below I am sending you the two modified
> functions. There are comments in the code (the problem is related to treatment
> of the edges of arrays, both times there is an attempt to access unallocated
> memory). This isn't a great deal, though may cause trouble in a massive
> calculation, at least on our cluster. Hope it is helpful.
According to your suggestions, I have fixed the bug which lived in a dozen of
routines. It will be reflected to the next release. Thanks for your pointer.
> More than 30% of time openmx spends in Bessel functions evaluations.
> What actually is the reason for that specific realization of bessjp
> (from NR.COM I guess?). I would think that a realization of Bessel
> functions specifically optimizied for fast and accurate calculations
> for real arguments and integer indices might work faster. There other
> versions of besseljp (netlib, nist and so on). Do you think it is worth
> to go deep into?
I had also recognized that the evaluation of Bessel functions is time-consuming,
and have recently modified by implenting a much faster scheme.
Regards,
T.Ozaki
|
|
This thread is locked.Only browsing is available.
openmx runtime questions