[BUG]: Segmentation fault with conda on MacOS

[enrico-lattuada]

Description When using the package built with C++ enabled inside a conda environment on MacOS, I get a segmentation fault error when I import the library from Python.

To Reproduce Steps to reproduce the behavior:

1. Create a conda environment configuration YAML file fastddm-test.yml.

   name: fddm-test
   channels:
     - defaults
   dependencies:
     - clang
     - clangxx
     - python>=3.8

2. Create the environment

   conda env create -f fastddm-env.yml

3. Activate the environment

   conda activate fddm-test

4. Use the environment compilers by exporting the relevant variables

   conda env config vars set CC=$CONDA_PREFIX/bin/clang
   conda env config vars set CXX=$CONDA_PREFIX/bin/clang++

5. Deactivate the environment

   conda deactivate

6. Reactivate the environment
7. Build and install the package from source with ENABLE_CPP=ON
8. Import the package in Python

Expected behavior Import without errors.

Desktop (please complete the following information):

• OS: MacOS (Intel CPU)

Additional context This might also be connected to the same problem (with or sometimes without conda environments) on other systems with multiple Python and C++ compiler versions installed.

[enrico-lattuada]

Possibly connected to this.

[enrico-lattuada]

The problem must be treated slightly differently for the different OSs.

The common part is using the MODULE option in the pybind11_add_module CMake function.

Linux

Linux automatically links the Python libraries when using the MODULE option. We don't need to make any changes.

MacOS

We must not explicitly link Python in the lower level libraries. We trust -undefined dynamic_lookup.

Add the following property for fddm (and for fddm_cuda as well, even though this should not be necessary since CUDA is not supported on macOS)

if(APPLE)
    set_target_properties(fddm PROPERTIES
        LINK_FLAGS "-undefined dynamic_lookup"
    )
endif(APPLE)

Windows

For Windows, we must explicitly link Python (at least, I found no source saying otherwise).

Add the following lines for both fddm and fddm_cuda

if(WIN32)
    target_link_libraries(fddm PUBLIC ${PYTHON_LIBRARIES})
endif(WIN32)

Also, since after pybind11_add_module the two options SHARED and MODULE are treated differently on Windows, we need to change the core and core_cuda target property from RUNTIME_OUTPUT_DIRECTORY_RELEASE to LIBRARY_OUTPUT_DIRECTORY_RELEASE.

[enrico-lattuada]

Closed by #207