[Pw_forum] Using fftw wrapper for intel mkl lib?
akohlmey at cmm.chem.upenn.edu
Mon Mar 5 22:22:40 CET 2007
On Mon, 5 Mar 2007, Detlef Schmicker wrote:
DS> Dear Paolo,
DS> thanks. I managed to link to mkl with v.2 and v.3 wrapper. Both are
DS> working in my serial execution setup.
DS> I could not reproduce spead up using v.3 wrapper and mkl. I have a
DS> Pentium 4 and espresso 3.2.
please keep in mind, that speedup frequently depends
also on the number of supported grids. fftw can handle
grids with roots up to 11, and does a very decent job
at that. many other ffts only support roots up to 5,
so it might frequently happen that with fftw you can
use smaller grids, which will compensate or even overcompensate
the speedup from a faster fft. so when you compare performance,
you should check wether you have the same grids.
in the case of - for example - fftw vs. essl, it can be that
you go from a 5% advantage of essl for an 128x128x128 grid
to a 20% deficit for an 132^3 grid (since essl has to use 144^3).
i stumbled across this with a different code, but it has
to apply to QE as well.
DS> Just as an information,
DS> thanks Detlef
DS> Am Montag, den 05.03.2007, 16:51 +0100 schrieb Paolo Giannozzi:
DS> > On Mar 3, 2007, at 18:19 , Detlef Schmicker wrote:
DS> > > did anybody manage to compile pwscf (pw.x) using the fftw-wrapper
DS> > > of mkl 9.0?
DS> > AFAIK there is no special trick to compile pw.x; just compile
DS> > for either v.2 or v.3 of FFTW and load the mkl wrappers,
DS> > provided by Intel as a separate package, to be compiled.
DS> > I heard that the FFTW v.2 wrappers work for both serial
DS> > and parallel execution, but you gain little; you gain much
DS> > more with v.3 wrappers, but they work only for the serial
DS> > case.
DS> > Paolo
DS> > ---
DS> > Paolo Giannozzi, Democritos and University of Udine, Italy
DS> > _______________________________________________
DS> > Pw_forum mailing list
DS> > Pw_forum at pwscf.org
DS> > http://www.democritos.it/mailman/listinfo/pw_forum
DS> Pw_forum mailing list
DS> Pw_forum at pwscf.org
Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu http://www.cmm.upenn.edu
Center for Molecular Modeling -- University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425
If you make something idiot-proof, the universe creates a better idiot.
More information about the users