2. Celeste Build Notes

Author:Benson Ma

Below is an assortment of build notes for handling different software dependencies and different platforms.

2.1. General

2.1.1. CUDA

  • Library code that is built on top of CUDA must be built as shared libraries; otherwise, linker errors will appear during building on Linux platforms. Hence, the entries in CMakeLists.txt specifying building CUDA-dependent libraries should be marked SHARED as such:

    CUDA_ADD_LIBRARY(CelesteFooCUDA SHARED foo.cu bar.cu)

  • The version of gcc installed may be a later version than the the latest officially-supported host compiler for nvcc. You will see an error like this:

    # in <PROJECT_ROOT>/target
localhost:target local$ cmake -D CMAKE_C_COMPILER=/opt/local/bin/gcc-mp-5 -D CMAKE_CXX_COMPILER=/opt/local/bin/g++-mp-5 -D CELESTE_GPU=1 ..
...
... generating Makefiles
...

localhost:target local$ make
...
... building code
...
/usr/local/cuda/include/host_config.h:115:2: error: #error -- unsupported GNU version! gcc versions later than 4.9 are not supported!

    While not recommended, this can be fixed by commenting out the appropriate #error macro in <CUDA_ROOT>/include/host_config.h:

    #if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ > 9)

// #error -- unsupported GNU version! gcc 4.10 and up are not supported!

#endif /* __GNUC__> 4 || (__GNUC__ == 4 && __GNUC_MINOR__ > 9) */

2.2. Linux

2.2.1. MPI

  • Installation of MPICH or OpenMPI may not include adding mpicc/mpic++ to the $PATH, resulting in the following error when cmake is invoked:

    Could NOT find MPI_C (missing: MPI_C_LIBRARIES MPI_C_INCLUDE_PATH)

    To resolve this, simply add the directory containing mpicc/mpic++ to the $PATH in the ENVIRONMENT or in the cmake` invocation:

    localhost:target local$ PATH=$PATH:/usr/lib64/mpich/bin cmake ..

2.3. Mac OS X

  • Unfortunately, with the newer versions of CUDA on the Mac, gcc is not a supported host compiler. Attempting to compile CUDA code using gcc as the host compiler will result in an error message that looks like this:

    # in <PROJECT_ROOT>/target
localhost:target local$ cmake -D CMAKE_C_COMPILER=/opt/local/bin/gcc-mp-5 -D CMAKE_CXX_COMPILER=/opt/local/bin/g++-mp-5 -D CELESTE_GPU=1 ..
...
... generating Makefiles
...

localhost:target local$ make
...
... building code
...
nvcc fatal   : GNU C/C++ compiler is no longer supported as a host compiler on Mac OS X.

    Intel’s ICC does not appear to be a supported host compiler for CUDA on Mac OS X either. Only Clang appears to be a supported host compiler, but this only applies to “AppleClang” (the version of Clang maintained by Apple). Attempting to use (newer versions of) mainline Clang will result in an error message that looks like this:

    # in <PROJECT_ROOT>/target
localhost:target local$ cmake -D CMAKE_C_COMPILER=/opt/local/bin/clang-mp-3.7 -D CMAKE_CXX_COMPILER=/opt/local/bin/clang++-mp-3.7 -D CELESTE_GPU=1 ..
...
... generating Makefiles
...

localhost:target local$ make
...
... building code
...
nvcc fatal   : The version ('30700') of the host compiler ('clang') is not supported

    While not recommended for ABI/linking reasons, issues such as this above can be resolved by specifying a _different_ compiler as the host compiler for nvcc:

    # where /usr/bin/clang symlinks to AppleClang
localhost:target local$ cmake -D CMAKE_C_COMPILER=/opt/local/bin/clang-mp-3.7 -D CMAKE_CXX_COMPILER=/opt/local/bin/clang++-mp-3.7 -D CELESTE_GPU=1 -D CUDA_HOST_COMPILER=/usr/bin/clang ..

2.4. Windows