add GudhUI compilation flag

[DavidLapous]

This adds a flag WITH_GUDHI_GUDHUI to disable GudhUI compilation, without disabling everything.
This is very useful as Gudhi doesn't compile on a conda environment without qgl libraries, such as the following one :

channels:
- conda-forge
dependencies:
- auditwheel
- boost
- boost-cpp
- cgal-cpp
- cxx-compiler
- cython
- gudhi
- jupyterlab
- matplotlib
- numpy
- pybind11
- pytest
- python=3.11
- scikit-learn
- scipy
- sympy
- tbb
- tbb-devel
- tqdm

I suppose that cmake doesn't like some of the flags that comes with the cxx-compiler package, which gives the following error

[ 15%] Building CXX object src/GudhUI/CMakeFiles/GudhUI.dir/GudhUI_autogen/mocs_compilation.cpp.o
In file included from /user/dloiseau/home/git/gudhi-devel/build/src/GudhUI/GudhUI_autogen/include/ui_main_window.h:22,
                 from /user/dloiseau/home/git/gudhi-devel/build/src/GudhUI/GudhUI_autogen/DMHXEJ42XS/../../../../../src/GudhUI/gui/MainWindow.h:18,
                 from /user/dloiseau/home/git/gudhi-devel/build/src/GudhUI/GudhUI_autogen/DMHXEJ42XS/moc_MainWindow.cpp:10,
                 from /user/dloiseau/home/git/gudhi-devel/build/src/GudhUI/GudhUI_autogen/mocs_compilation.cpp:2:
/user/dloiseau/home/git/gudhi-devel/src/GudhUI/view/Viewer.h:14:10: fatal error: QGLViewer/qglviewer.h: No such file or directory
   14 | #include <QGLViewer/qglviewer.h>
      |          ^~~~~~~~~~~~~~~~~~~~~~~
compilation terminated.

[mglisse]

I am not opposed to a flag just for GudhUI, it makes some sense, but the priority should be understanding what is going wrong. CMakeLists.txt has

    find_package(QGLViewer QUIET)
    if ( QGLVIEWER_FOUND)

before the compilation of GudhUI, which seems to imply that QGLViewer was found by CMake. @VincentRouvreau could you try to reproduce?

[mglisse]

@DavidLapous what does CMakeCache.txt say about QGLViewer in that failing scenario?

[mglisse]

If I try on my laptop in a conda environment created from your .yml, CMake prints

++ GUDHI_MISSING_MODULES list is:"GudhUI;python"

(or python-documentation instead of python depending how I call cmake exactly)
and make does not try to build GudhUI.

[mglisse]

ok if Vincent is ok with it.
But I would really like to understand the failure, don't forget about it just because you have a way to skip it.

[DavidLapous]

Sorry for the delay. Here is my CMakeCache.txt.

[mglisse]

It does not reproduce on my computer because cmake finds libGL.so in /usr/lib64 on Fedora but not in /usr/lib/x86_64-linux-gnu on Debian (strange)... and thus it never even looks for QGLViewer on my computer.
The problem seems to come from a discrepancy between the paths that cmake considers and the paths that the compiler considers. 2 things that would go in the right direction, although they may not solve the problem:

• Use the version of CMake provided by conda (maybe they configure the default paths in a way better suited to the compilers they provide);
• in FindQGLViewer.cmake, we should remove default things like /usr/include or /usr/lib, normally CMake already knows about default directories to look at.

[DavidLapous]

I've tried with the make and cmake from conda (ie removing build + relaunch cmake), and it doesn't fix the problem.

[mglisse]

I managed to reproduce by telling CMake to look in /usr/lib/x86_64-linux-gnu.
My guess is that the following happens:

• cmake (even the one from conda) looks in /usr/include by default
• cmake considers /usr/include a default system directory, so it avoids passing -I/usr/include to compilers
• conda's compiler is configured not to look at /usr/include by default

Assuming my guess is close to the truth, that makes it a big bug in conda's cmake package (maybe related to conda-forge/cmake-feedstock#106 and conda-forge/cmake-feedstock#129 ?).

[mglisse]

I can confirm that the cmake in conda has a long CMAKE_SYSTEM_PREFIX_PATH that includes not just the conda paths but also /usr, /usr/local (twice!), /usr/X11R6, /opt, etc).
If I add include_directories("/usr/include") and include_directories("/tmp/y"), I only get -I/tmp/y in the compiler invocation, even though CMAKE_CXX_IMPLICIT_INCLUDE_DIRECTORIES only lists paths in the conda installation.

What conda+cmake does is bound to break...

In the mean time, when something that is installed on your system causes trouble for your conda environment, a natural workaround is to install this in conda as well and hopefully the build will use the right version.

[VincentRouvreau]

Thanks for the proposal @DavidLapous and for the investigation @mglisse !
Quite strange behaviour from cmake + conda to presume that some dependencies are installed in some specific folders...