May be you can try removing nomkl in AmberTools/src/configure_python and
try again.
(I dont have intel compiler to test).
Hai
On Tue, Apr 4, 2017 at 2:28 AM Scott Brozell <sbrozell.rci.rutgers.edu>
wrote:
> Hi,
>
> On Fri, Mar 31, 2017 at 04:35:14PM -0400, Hai Nguyen wrote:
> > On Fri, Mar 31, 2017 at 4:18 PM, Scott Brozell <sbrozell.rci.rutgers.edu
> >
> > > 1. This is a relatively recent change since i frequently configure
> with
> > > -mkl and various forms of python without this problem. So probably the
> > > best to investigate are those that have made recent python building
> > > changes.
> >
> > Can you "ldd lib/libsander.so" to see if it is linked to
> lib_mkl_core.so?
> > Likely the libsander.so issue rather python binding.
>
> It is there:
> libmkl_core.so =>
> /opt/intel/compilers_and_libraries_2016.3.210/linux/mkl/lib/intel64_lin/libmkl_core.so
> (0x00007fdf9b6f8000)
>
> That file exists and has mkl_serv_getenv:
> /opt/intel/compilers_and_libraries_2016.3.210/linux/mkl/lib/intel64_lin/libmkl_core.so:000000000010ba10
> T mkl_serv_getenv
>
> All this makes sense since the non python tests using libsander worked.
>
> On the same machine but with the system python instead of amber's
> there are no test errors; so it still seems that something is
> wrong or different with Amber's python:
> ===
> ./configure -mkl --with-python /usr/local/anaconda2/bin/python intel
> ...
> Testing the Python LES API
> /usr/local/anaconda2/bin/python test.py
> Testing GB sander interface (diffcoords w/ RDT)
> PASSED
> ...
> 0 tests experienced errors
> ===
>
> Compare with
> ===
> ./configure -mkl intel
> ...
> Testing the Python LES API
> /tmp/amber/bin/amber.python test.py
> /tmp/amber/bin/amber.python: symbol lookup error:
> /opt/intel/compilers_and_libraries_2016.3.210/linux/mkl/lib/intel64_lin/libmkl_vml_avx2.so:
> undefined symbol: mkl_serv_getenv
> Program error
> ...
> 2 tests experienced errors
> ===
>
> > > 2. For this build ./configure -mkl intel
> > > I get a lot of hits grepping for getenv and then only 1 for grepping
> that
> > > for mkl:
> > > Binary file
> ./miniconda/pkgs/scipy-0.19.0-np112py27_nomkl_0/lib/python2.
> > > 7/site-packages/scipy/sparse/_
> > > sparsetools.so matches
> > >
> > Just FYI: If install python via Miniconda, numpy and scipy will be
> > installed without mkl (I added "nomkl" option
> > since MKL package added about 100 MB for miniconda build (got complain
> > about size before).
>
> Maybe i don't understand what you wrote, but it reads like that is the
> problem, ie, adding "nomkl" is what broke it.
>
> > > 3.
> > > IMO we should be encouraging use the -mkl compiler flag.
> > > The -mkl approach is far superior to MKL_HOME because the former
> follows
> > > the principle of information hiding and the latter typically forces at
> > > least
> > > someone to decrypt the mess of details of mkl .a's and .so's.
> >
> > I agree.
>
> Wow, we agree on a principle. I feel satisfied.
>
> :)
> scott
>
>
> > >
> > > Worth more than $0.02
> > > :)
> > > scott
> > >
> > >
> > >
> > >
> > > On Fri, Mar 31, 2017 at 11:24:43AM -0400, David Case wrote:
> > > > Hoping someone might recognize this symptom, and know what to do:
> > > >
> > > > Using Intel compilers (any recent version) there is a difference in
> > > > AmberTools between passing the "-mkl" flag to configure, and setting
> > > > the MKL_HOME environment variable. The former option is relatively
> new,
> > > > whereas the latter is what we have used for some time.
> > > >
> > > > As Scott and others have found, using "-mkl" breaks the sanderapi:
> > > > python cannot find mkl_serv_getenv, which is in lib_mkl_core.so
> > > > libraries, which are apparently not getting properly linked in to
> > > > libsander.so.
> > > >
> > > > If one instead sets MKL_HOME, things work fine, presumably because
> > > > configure2 creates a config.h file that explicitly links to
> > > libmkl_core.a.
> > > > I've tried adding -lmkl_core to the config.h file created by
> configure2,
> > > > but without success.
> > > >
> > > > Minor point: using -mkl rather than setting MKL_HOME leads to
> roundoff
> > > > differences in many of the sqm tests. But users of the intel
> compilers
> > > > can probably recognize these for what they are.
> > > >
> > > > If anyone can figure out how to get the pysander API to find
> > > mkl_serv_getenv
> > > > when the -mkl flag is set, that would be a big help.
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Apr 04 2017 - 06:30:02 PDT
