ld issue of "-libcm" on buidling gpu/user-cuda

LAMMPS LAMMPS Mailing List Mirror
1

Dear LAMMPS developers and experts,
I am trying to build either gpu or user-cude with the latest LAMMPS source from git (Aug 12). Our test machine is CentOS 5.8 with CUDA-5.0. The MPI library is based on the system-default gnu compiler (gcc,g++ and gfortran)and openmpi. We have two NVIDIA M2090 cards on the machine. We can build “libgpu.a” and “liblammpscuda.a” successfully. However, the building of executive binary for “lmp_jgpu” or “nvc_get_devices”. I always got an error message of missing library “libcm”. I am not sure about this library. Is it regarding the “gnome X library”? Does anyone face the same issue before? Thank you very much.

  1. Error log from STDOUT (user-cuda):
    -llammpscuda -lmpi -lcuda -lcudart -lm -o …/lmp_jgpu
    /usr/bin/ld: skipping incompatible /usr/local/cuda/lib/libcufft.so when searching for -lcufft
    /usr/bin/ld: skipping incompatible /usr/local/cuda/lib/libcudart.so when searching for -lcudart
    /usr/bin/ld: cannot find -libcm
    collect2: ld returned 1 exit status
    make[1]: *** […/lmp_jgpu] Error 1
    make[1]: Leaving directory `/home/jcchen/cuda/lammps/src/Obj_jgpu’
    make: *** [jgpu] Error 2

  2. Error log from STDOUT (gpu):
    mpic++ -DMPI_GERYON -DUCL_NO_EXIT -DMPICH_IGNORE_CXX_SEEK -O2 -D_SINGLE_SINGLE -I/usr/local/cuda/include -DUSE_CUDPP -Icudpp_mini -o nvc_get_devices ./geryon/ucl_get_devices.cpp -DUCL_CUDADR -L/usr/local/cuda/lib64 -lcuda
    /usr/bin/ld: cannot find -libcm
    collect2: ld returned 1 exit status
    make: *** [nvc_get_devices] Error 1

Sincerely,
James

2

Dear LAMMPS developers and experts,
I am trying to build either gpu or user-cude with the latest LAMMPS source
from git (Aug 12). Our test machine is CentOS 5.8 with CUDA-5.0. The MPI

the *current* git version is 16 Aug 2013.

library is based on the system-default gnu compiler (gcc,g++ and
gfortran)and openmpi. We have two NVIDIA M2090 cards on the machine. We can
build "libgpu.a" and "liblammpscuda.a" successfully. However, the building
of executive binary for "lmp_jgpu" or "nvc_get_devices". I always got an
error message of missing library "libcm". I am not sure about this library.
Is it regarding the "gnome X library"? Does anyone face the same issue
before? Thank you very much.

this looks a lot like you made a typo somewhere. can you compile and
link LAMMPS without any GPU support? try that first, then add the GPU
package, then the USER-CUDA package. this way you can find out, where
the typo is located.

i have very recently built LAMMPS on centos 5.x (and 6.x) with CUDA
5.0 and did not see this.

axel.