Hello, i am struggling for compiling a code that requires both CUDA and MPI (MPICH implementation). In particular, when cmake is linking CUDA device code i get the following errors:
gcc: error: unrecognized command line option ‘-Wl’; did you mean ‘-W’?
gcc: error: unrecognized command line option ‘-rpath’
gcc: error: unrecognized command line option ‘-Wl’; did you mean ‘-W’?
gcc: error: unrecognized command line option ‘-Wl’; did you mean ‘-W’?
These errors are caused by the following flag passed on the nvcc command: -Xcompiler=-Wl,-rpath,-Wl,/opt/modules/binary/mpich/3.3.2/lib,-Wl,--enable-new-dtags . To work it must be either wrapped in quotes or used with Xlinker instead of Xcompiler. However, i believe that the actual culprit is MPICH, because when i ask mpicxx for compile and linking flags i get the following output:
$ mpicc -link_info
gcc -I/opt/modules/binary/mpich/3.3.2/include -L/opt/modules/binary/mpich/3.3.2/lib -Wl,-rpath -Wl,/opt/modules/binary/mpich/3.3.2/lib -Wl,–enable-new-dtags -lmpi
$ mpicc -compile_info
gcc -I/opt/modules/binary/mpich/3.3.2/include -L/opt/modules/binary/mpich/3.3.2/lib -Wl,-rpath -Wl,/opt/modules/binary/mpich/3.3.2/lib -Wl,–enable-new-dtags -lmpi
if i guess correctly what happens is that cmake correctly forward all flags that mpich provides as compiler flags to the compiler, however this cause the error. I solved the issue by setting the following CMake policy (which ignores all the link flags as far as i understood) :
cmake_policy(SET CMP0105 OLD)
However, i don’t really like this solution because eventually it will become obsolete. Does someone have a better solution? I also tried something similar to the one proposed here: https://gitlab.kitware.com/cmake/cmake/-/issues/18897 ; but it seems that i can’t use generators with INTERFACE_COMPILE_OPTIONS .
I used CMake 3.19.6 and 3.20.0 (compiled from sources), nvcc 11.2.152, and MPICH 3.3.2.