[petsc-dev] KNC build?

[Satish Balay]

On Tue, 4 Nov 2014, Satish Balay wrote:

> Ah - didn't check the configure comand closely enough..
> 
> > --with-mpicc=/global/babbage/nsg/opt/intel/impi/5.0.1.035/intel64/bin/mpicc
> > --with-mpicxx=/global/babbage/nsg/opt/intel/impi/5.0.1.035/intel64/bin/mpicxx
> > --with-mpif90=/global/babbage/nsg/opt/intel/impi/5.0.1.035/intel64/bin/mpif90
> 
> petsc configure doesn't provide any such options..
> 
> > --with-mpi-lib=/global/babbage/nsg/opt/intel/impi/5.0.1.035/intel64/lib
> 
> This would be the wrong way to provide MPI lib.
> 
> 
> The correct way to use Intel MPI is:
> 
> --with-cc=mpiipc --with-cxx=mpiicpc --with-fc=mpiifort [i.e no --with-mpi-include or --with-mpi-lib options]
> 
> But as Richard mentioned - the wrappers in intel impi-5 are
> broken. Alternative is to use impi-4 as suggested.
> 
> Or something like the following (for compilers/mpi):
> 
> ./configure --with-cc=icc --with-fc=ifort --with-cxx=0 --with-mpi-include=/global/babbage/nsg/opt/intel/impi/5.0.1.035/intel64/include --with-mpi-lib="-L/global/babbage/nsg/opt/intel/impi/5.0.1.035/intel64/lib/release -L/global/babbage/nsg/opt/intel/impi/5.0.1.035/intel64/lib -lmpifort -lmpi -lmpigi -ldl -lrt -lpthread"

BTW: I'm not sure what the correct list (for with-mpi-lib) would be
for KNC build. The above is for a regular CPU usage of impi.

So its better to just use impi-4 with mpiicc [the first suggestion with "-mmic" etc..]

Satish

> 
> Satish
> 
> On Tue, 4 Nov 2014, Richard Mills wrote:
> 
> > Hi Mark,
> > 
> > I noticed that you are using the wrong compiler wrappers.  Instead of
> > 'mpicc', 'mpif90', and 'mpicxx' (which will invoke GNU compilers), you
> > probably want 'mpiicc', 'mpiifort', and 'mpiicpc' (invoke the Intel
> > compilers).
> > 
> > You will probably also want to unload the IMPI 5.x and load the IMPI 4.1.3
> > module right now because of a stupid bug in the compiler wrappers (for
> > which I've filed a bug report--will be fixed in next IMPI release, sometime
> > before SC14).
> > 
> > --Richard
> > 
> > On Tue, Nov 4, 2014 at 12:59 PM, Mark Adams <mfadams at lbl.gov> wrote:
> > 
> > > Thanks Victor,
> > >
> > > I am getting cc error.  Any ideas?
> > > Mark
> > >
> > >
> > > On Tue, Nov 4, 2014 at 12:38 PM, Victor Eijkhout <eijkhout at tacc.utexas.edu
> > > > wrote:
> > >
> > >>
> > >> > On Nov 4, 2014, at 11:32 AM, Mark Adams <mfadams at lbl.gov> wrote:
> > >> >
> > >> > Has anyone built PETSc on KNC (Babbage at NERSC to be specific)?
> > >>
> > >> At TACC under intel 14:
> > >>
> > >> CEE_OPTIONS="-mmic -mkl -fp-model precise"
> > >> ./configure --PETSC_ARCH=mic --with-fc=0 --with-debug=0 \
> > >>     --with-batch=1 --CPPFLAGS=-mmic \
> > >>     --CFLAGS="${CEE_OPTIONS}" --CXXFLAGS="${CEE_OPTIONS}"
> > >> --FFLAGS="${CEE_OPTIONS}" \
> > >>     --with-mpi=1 --known-mpi-shared-libraries=1 \
> > >>         --with-mpi-include=${MPICH_HOME}/mic/include \
> > >>         --with-mpi-lib=${MPICH_HOME}/mic/lib \
> > >>         --with-mpicc=/opt/apps/intel13/impi/4.1.0.030/intel64/bin/mpicc \
> > >>         --with-mpicxx=/opt/apps/intel13/impi/4.1.0.030/intel64/bin/mpicxx
> > >> \
> > >>         --with-mpif90=/opt/apps/intel13/impi/4.1.0.030/intel64/bin/mpif90
> > >>
> > >> That first line is still giving me problems. Just nix the fp-model.
> > >>
> > >> Victor.
> > >>
> > >>
> > >
> > 
> 
>