[gmx-users] Gromacs 4.6.3 installation Issue with Intel & CUDA

[Jim Strong]

Hello,

I have a problem compiling Gromacs 4.6.3 on RHEL 6.4 x64 with permutations
of the following programs:

Intel 13.1.0
CUDA 5.5
OpenMPI 1.7.2
Cmake 2.8.11.2
FFTW 3.3.3 (or Intel MKL)
GNU 4.8.1

Basically, once the Intel compiler is introduced into the picture, NVCC
seems to fail:

-----------
[  0%] Building NVCC (Device) object
src/gmxlib/cuda_tools/CMakeFiles/cuda_tools.dir//./cuda_tools_generated_copyrite_gpu.cu.o
gcc: error: unrecognized command line option ‘-ip’
CMake Error at cuda_tools_generated_copyrite_gpu.cu.o.cmake:206 (message):
  Error generating

/mnt/home/mod/SRC/gromacs/gromacs-4.6.3/build-cmake/src/gmxlib/cuda_tools/CMakeFiles/cuda_tools.dir//./cuda_tools_generated_copyrite_gpu.cu.o


make[2]: ***
[src/gmxlib/cuda_tools/CMakeFiles/cuda_tools.dir/./cuda_tools_generated_copyrite_gpu.cu.o]
Error 1
make[1]: *** [src/gmxlib/cuda_tools/CMakeFiles/cuda_tools.dir/all] Error 2
make: *** [all] Error 2
--------------

Because it seems CUDA doesn't support Intel compilers (or vice versa), I
set the Cmake CUDA_HOST_COMPILER to GCC, which in my understanding should
enable nvcc to compile the CUDA parts of the program with this GNU compiler
while leave the rest to Intel's ICC (through the OpenMPI wrappers). However
it doesn't work.

I have managed to Cmake compile with all kinds of permutations of settings
(CUDA On/Off, MPI On/Off, etc) with GNU, but when Intel is used, nothing
seems to work (well, Cmake builds the stuff, but make leads to the above).

So, my question is:
Has anybody succeeded compiling 4.6.3 with Intel compilers + CUDA + MPI on
a similar Linux-based system? If yes, how?

Thanks.