The ABINIT Users Mailing List ( CLOSED )

[Jyh-Shyong <c00jsh00@nchc.org.tw>]

Hi Luigi,

After many attempts, I still failed to build parallel version of 
bigdft-1.3 with cuda support.
In order to have cuda support, I have to include compiler option -assume 
2underscores,
however, this option fails to pass the mpi environment checking in 
configure procedure.

Configure options:
export  FC="/opt/mvapich2/intel/bin/mpif90 -assume 2underscores"
export  FCFLAGS=""
export  LDFLAGS=" "
export  LIBS=""
export  CC=icc
export  CPPFLAGS=""
export  CXX=icpc

./configure -prefix=/package/chem/bigdft --enable-cuda-gpu  
--with-cuda-path=/opt/cuda --with-lib-cutil
s=/opt/cuda/lib --with-ext-linalg="-lmkl_intel_lp64 -lmkl_sequential 
-lmkl_core -lpthread -ldl -lsysfs"
--with-ext-linalg-path=-L/opt/intel/mkl/lib



Basics:
 Fortran90 compiler:        /opt/mvapich2/intel/bin/mpif90 -assume 
2underscores
 Fortran90 compiler name:   intel
 Fortran90 flags:           -O2
 Linker flags:                -L/opt/intel/mkl/lib 
-L$(top_builddir)/libXC/src
 Linked libraries:          -lxc -lmkl_intel_lp64 -lmkl_sequential 
-lmkl_core -lpthread -ldl -lsysfs

Build:
 Library ABINIT:            yes
 Library PSolver:           yes
 Library BigDFT:            yes
 Main binaries (cluster...):yes
 Minima hopping binary:     no

Options:
 With MPI:                  Warning
  | Include dir.:          
  | Linker flags:          
  | Linked libraries:      
  | MPI2 support:          
 With optimised conv.:      yes
 With Cuda GPU conv.:       yes
  | NVidia Cuda Compiler:   /opt/cuda/bin/nvcc
  | Cuda flags:             -O3 --compiler-options -fno-strict-aliasing
  | GCC41HACK (gcc path):  
  | ext. cutil library:     /opt/cuda/lib
  |      cutil path:        /opt/cuda/lib
 With libXC:                yes
  | internal built:         yes
  | include dir.:           -I$(top_builddir)/libXC/src

Installation paths:
 Source code location:      .
 Prefix:                    /package/chem/bigdft
 Exec prefix:               ${prefix}
 Binaries:                  ${exec_prefix}/bin
 Static libraries:          ${exec_prefix}/lib
 Fortran modules:           ${prefix}/include/intel

Warnings:
 MPI:                       disabled by tests


In file config.log:
....
configure:4965: checking for MPI capabilities
configure:4980: /opt/mvapich2/intel/bin/mpif90 -assume 2underscores -o 
conftest -O2      conftest.f90   >&5
/tmp/ifortSVHO4q.o: In function `MAIN__':
conftest.f90:(.text+0x37): undefined reference to `mpi_init__'
configure:4986: $? = 1
configure: failed program was:
| program main
|   include "mpif.h"
|
|   call MPI_INIT
| end
configure:5002: result: no
configure:5070: WARNING: MPI is not working and has been disabled.
configure:5072: WARNING:   If you don't want to see this warning,
configure:5074: WARNING:   set --disable-mpi in the configure line.

Eventually, the compilation built a serial version of bigdft with cuda 
support.


If -assume 2underscores is not included:

export  FC="/opt/mvapich2/intel/bin/mpif90"
export  FCFLAGS=""
export  LDFLAGS=" "
export  LIBS=""
export  CC=icc
export  CPPFLAGS=""
export  CXX=icpc

./configure -prefix=/package/chem/bigdft --enable-cuda-gpu  
--with-cuda-path=/opt/cuda --with-lib-cutil
s=/opt/cuda/lib --with-ext-linalg="-lmkl_intel_lp64 -lmkl_sequential 
-lmkl_core -lpthread -ldl -lsysfs"
--with-ext-linalg-path=-L/opt/intel/mkl/lib

....
Basics:
 Fortran90 compiler:        /opt/mvapich2/intel/bin/mpif90
 Fortran90 compiler name:   intel
 Fortran90 flags:           -O2
 Linker flags:                -L/opt/intel/mkl/lib 
-L$(top_builddir)/libXC/src
 Linked libraries:          -lxc -lmkl_intel_lp64 -lmkl_sequential 
-lmkl_core -lpthread -ldl -lsysfs

Build:
 Library ABINIT:            yes
 Library PSolver:           yes
 Library BigDFT:            yes
 Main binaries (cluster...):yes
 Minima hopping binary:     no

Options:
 With MPI:                  yes
  | Include dir.:          
  | Linker flags:          
  | Linked libraries:      
  | MPI2 support:           yes
 With optimised conv.:      yes
 With Cuda GPU conv.:       yes
  | NVidia Cuda Compiler:   /opt/cuda/bin/nvcc
  | Cuda flags:             -O3 --compiler-options -fno-strict-aliasing
  | GCC41HACK (gcc path):  
  | ext. cutil library:     /opt/cuda/lib
  |      cutil path:        /opt/cuda/lib
 With libXC:                yes
  | internal built:         yes
  | include dir.:           -I$(top_builddir)/libXC/src

Installation paths:
 Source code location:      .
 Prefix:                    /package/chem/bigdft
 Exec prefix:               ${prefix}
 Binaries:                  ${exec_prefix}/bin
 Static libraries:          ${exec_prefix}/lib
 Fortran modules:           ${prefix}/include/intel

The compilation failed since subroutines in libGPU.a can not be recognized.

What would be the right combination of configure parameters so the 
parallel version of bigdft which supports cuda?

Thans for your help

Jyh-Shyong Ho