[QE-users] Bad performance of OpenMP builds linked to fftw3_omp

[Husak Michal]

Hi


I was getting horrible bad performance of QE 7.0 when build with

--enable-openmp option.

This did not happened with previous version (6.8) ...

There was an approx. 8x speed down ...


I had traced the problem to the link with fftw3_omp library witch was added in the 7.0 release

for the --enable-openmp option.

The code freeze on fftw routines almost.

The solution was to manually edit make.ins and set:

DFLAGS         =  -D__FFTW

The code was build for Windows by Cygwin64 and gfortran ...


Questions:


1) Is this an general issue with 7.0 and fftw3_omp  or only a Windows specific build problem ?


2) Is there anybody on this list with Windows QE build  optimization experiences ?

I need somebody for consultations ...


3) I was tiring to make a  pure "serial" build by setting --disable-openmp --disable-parallel option.

To my surprise the result run clearly on multiple cores an almost as fast as openmp enabled build.

Even with set OMP_NUM_THREADS=1. Any explanation ?


Michal Husak

UCT Prague