[cmake-developers] [CMake 0015314]: FindMPI.cmake can choose incompatible set of libraries for Intel MPI
Mantis Bug Tracker
mantis at public.kitware.com
Wed Dec 17 11:29:58 EST 2014
The following issue has been SUBMITTED.
======================================================================
http://public.kitware.com/Bug/view.php?id=15314
======================================================================
Reported By: Kelly Thompson
Assigned To:
======================================================================
Project: CMake
Issue ID: 15314
Category: CMake
Reproducibility: always
Severity: minor
Priority: normal
Status: new
======================================================================
Date Submitted: 2014-12-17 11:29 EST
Last Modified: 2014-12-17 11:29 EST
======================================================================
Summary: FindMPI.cmake can choose incompatible set of
libraries for Intel MPI
Description:
When using Intel MPI, CMake's FindMPI.cmake can choose the wrong MPI libraries
for linking because it does not differentiate between the regular and the thread
safe versions of the libraries (this might be related to
http://public.kitware.com/Bug/view.php?id=14991).
The problem can been observed by noting that FindMPI.cmake calls '@mpiicpc
-show@' to determine the MPI libraries that should be used during linking (these
will not be the thread safe versions). For example:
% mpiicpc -show
icpc -I/projects/opt/mpi/impi/5.0.1.035/intel64/include \
-L/projects/opt/mpi/impi/5.0.1.035/intel64/lib/release \
-L/projects/opt/mpi/impi/5.0.1.035/intel64/lib \
-Xlinker --enable-new-dtags \
-Xlinker -rpath \
-Xlinker /projects/opt/mpi/impi/5.0.1.035/intel64/lib/release \
-Xlinker -rpath -Xlinker /projects/opt/mpi/impi/5.0.1.035/intel64/lib \
-Xlinker -rpath -Xlinker /opt/intel/mpi-rt/5.0/intel64/lib/release \
-Xlinker -rpath -Xlinker /opt/intel/mpi-rt/5.0/intel64/lib \
-lmpicxx -lmpifort -lmpi -lmpigi -ldl -lrt -lpthread
However, if thread safe versions are desired, the command 'mpiicpc -mt_mpi
-show' should be used instead. Note that adding '-openmp' (my situation) will
also select the thread safe versions. These libraries are:
<pre>
% mpiicpc -mt_mpi -show
icpc -I/projects/opt/mpi/impi/5.0.1.035/intel64/include \
-L/projects/opt/mpi/impi/5.0.1.035/intel64/lib/release_mt \
-L/projects/opt/mpi/impi/5.0.1.035/intel64/lib \
-Xlinker --enable-new-dtags \
-Xlinker -rpath -Xlinker /projects/opt/mpi/impi/5.0.1.035/intel64/lib/release_mt
\
-Xlinker -rpath -Xlinker /projects/opt/mpi/impi/5.0.1.035/intel64/lib \
-Xlinker -rpath -Xlinker /opt/intel/mpi-rt/5.0/intel64/lib/release_mt \
-Xlinker -rpath -Xlinker /opt/intel/mpi-rt/5.0/intel64/lib \
-lmpicxx -lmpifort -lmpi_mt -lmpigi -ldl -lrt -lpthread
</pre>
The difference is the library folder location: 'release' vs. 'relase_mt'
Currently, when the CMAKE_<LANG>_COMPILER == MPI_<LANG>_COMPILER, FindMPI.cmake
will always add the non-thread safe library to the link line via
${MPI_<LANG>_LIBRARIES} so that if the compiler flags include '-mt_mpi' or
'-openmp', the link line will include both the thread safe libraries and the
normal libraries resulting in:
ld: MPIR_Thread: TLS definition in .../intel64/lib/release_mt/libmpi_mt.so
section .tbss mismatches non-TLS definition in .../intel64/lib/libmpi.so section
.bss
.../intel64/lib/release_mt/libmpi_mt.so: could not read symbols: Bad value
I think the correct resolution to this issue would be to disable the
interrogation of the wrapper when CMAKE_<LANG>_COMPILER == MPI_<LANG>_COMPILER
because all of the required include and link information will already be
available via the MPI compiler wrapper. In this model, MPI_<LANG>_LIBRARIES
would be an empty string if the mpi compiler wrapper is used as the project
primary compiler.
A potential fix would be a modification of some sample code that I recently
proposed via the cmake-developers mailing list:
--- a/Modules/FindMPI.cmake
+++ b/Modules/FindMPI.cmake
@@ -165,11 +165,28 @@ foreach (lang C CXX Fortran)
endif()
unset(_MPI_${id}_${lang}_COMPILER_NAMES) # clean up the namespace here
endforeach()
+
+ # If cmake_$lang_compiler matches a known mpi compiler wrapper name,
+ # prefer the provided name and prevent interrogation.
+ get_filename_component( compiler_wo_path "${CMAKE_${lang}_COMPILER}" NAME )
+ set( ${lang}_compiler_is_mpiwrapper false )
+ foreach( mpiwrapper ${_MPI_${lang}_COMPILER_NAMES} )
+ if( "${mpiwrapper}" STREQUAL "${compiler_wo_path}" )
+ set( ${lang}_compiler_is_mpiwrapper true )
+ endif()
+ endforeach()
+ if( ${lang}_compiler_is_mpiwrapper )
+ set( MPI_${lang}_COMPILER ${CMAKE_${lang}_COMPILER} )
+ set( MPI_${lang}_NO_INTERROGATE ${CMAKE_${lang}_COMPILER} )
+ endif()
endforeach()
Steps to Reproduce:
On Linux with Intel/15.0 and Intel MPI 5.0 (my testing used cmake/3.0.1 and the
git source version).
export CXX=`which mpiicpc`
export CC=`which mpiicc`
cmake <dir>
<dir> has 2 files: CMakeLists.txt and hw.cc
# CMakeLists.txt
cmake_minimum_required(VERSION 3.0)
project( hw_mpi CXX )
set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mt_mpi" )
find_package( MPI )
include_directories( ${MPI_INCLUDE_PATH} )
add_executable( hw hw.cc )
target_link_libraries( hw ${MPI_LIBRARIES} )
// hw.cc
#include <mpi.h>
#include <iostream>
int main(int argc, char* argv[])
{
MPI_Init(&argc, &argv);
int rank;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if (rank == 0)
{
int value = 17;
int result = MPI_Send(&value, 1, MPI_INT, 1, 0, MPI_COMM_WORLD);
if (result == MPI_SUCCESS)
std::cout << "Rank 0 OK!" << std::endl;
}
else if (rank == 1)
{
int value;
int result = MPI_Recv(&value, 1, MPI_INT, 0, 0, MPI_COMM_WORLD,
MPI_STATUS_IGNORE);
if (result == MPI_SUCCESS && value == 17)
std::cout << "Rank 1 OK!" << std::endl;
}
MPI_Finalize();
return 0;
}
Additionally, a short test like this should probably be added to the standard
suite of CMake regression tests.
Additional Information:
This issue is possibly related to
http://public.kitware.com/Bug/view.php?id=15182 and
http://public.kitware.com/Bug/view.php?id=14991.
======================================================================
Issue History
Date Modified Username Field Change
======================================================================
2014-12-17 11:29 Kelly Thompson New Issue
======================================================================
More information about the cmake-developers
mailing list