OSX support

[pwolfram]

What are the requirements to provide OSX support (listed in README)? Great to see you put this on github Rusty!

[pwolfram]

I should note there is no rush on this and I'm happy to help as possible.

[rustychris]

Hi Phil - Good to hear from you.

I've been trying to get this to work in anaconda, but the main sticking point is compiling the python-cgal-bindings. The compilation process in CGAL/SWIG is a little too clever for anaconda, and some of the paths embedded in the libraries don't get updated, or at least that's my best interpretation.

Here is the error I get:

python ~/anaconda/envs/tom_test/lib/python2.7/site-packages/stomel/tom.py -b spinup-shoreline.shp -s density-spinup.shp 
/Users/rusty/anaconda/envs/tom_test/lib/python2.7/site-packages/matplotlib/cbook.py:133: MatplotlibDeprecationWarning: The matplotlib.delaunay module was deprecated in version 1.4. Use matplotlib.tri.Triangulation instead.
  warnings.warn(message, mplDeprecation, stacklevel=1)
CGAL unavailable.
dlopen(/Users/rusty/anaconda/envs/tom_test/lib/python2.7/site-packages/CGAL/_CGAL_Triangulation_2.so, 2): Library not loaded: /Users/rusty/anaconda/conda-bld/work/lib/libCGAL_Triangulation_2_cpp.dylib
  Referenced from: /Users/rusty/anaconda/envs/tom_test/lib/python2.7/site-packages/CGAL/_CGAL_Triangulation_2.so
  Reason: image not found

This is from a cleanly installed anaconda environment:
.condarc

channels:
 - ioos
 - rustychris
 - defaults 

Followed by conda create -n tom_test python=2.7, source activate tom_test, and conda install stomel.

If you have any thoughts, the recipe for python-cgal-bindings is here

[pwolfram]

@rustychris, I'm assuming that this works for you provided you do the local compile, right? I tried hacking this by updating all the library paths by hand from your binaries but it seems like something is off with the cmake script for CGAL so that it is platform dependent (due to boost?). Ultimately I ended up with a different error due to threading, like this:

In [2]: import CGAL._CGAL_Triangulation_2
Fatal Python error: PyThreadState_Get: no current thread
/Users/pwolfram/miniconda2/envs/tom_test/bin/python.app: line 3: 23297 Abort trap: 6           /Users/pwolfram/miniconda2/envs/tom_test/python.app/Contents/MacOS/python "$@"

I'll try building from source on my platform next which should hopefully solve the above problem (although obviously not in a general way).

[pwolfram]

It appears that https://code.google.com/r/rustychris-apollonius/ is down... (see rustychris/conda-recipes#3)

[pwolfram]

Part of the issue is likely due to using $ORIGIN on linux / @loader_path on OSX (http://conda.pydata.org/docs/building/shared-libraries.html)

[pwolfram]

Another potential problem is that it may be linking to the OSX version of python and not anaconda's: e.g., /System/Library/Frameworks/Python.framework/Versions/2.7/Python when the bindings are built from source. Essentially, it is linking to the clang python framework associated with OS X.

[pwolfram]

Here are some minor changes I needed to do to get this to compile: https://github.com/pwolfram/conda-recipes/tree/generalize, but I still have that Fatal Python error: PyThreadState_Get: no current thread error and don't know to resolve it. Google searches imply that it is a mismatch between the library and calling python versions but it is unclear to me what should be linked instead of what we have presently (if this is the problem):

└─▪ otool -L /Users/pwolfram/miniconda2/envs/tom_test/lib/python2.7/site-packages/CGAL/_CGAL_Triangulation_2.so
/Users/pwolfram/miniconda2/envs/tom_test/lib/python2.7/site-packages/CGAL/_CGAL_Triangulation_2.so:
    /usr/local/lib/libmpfr.4.dylib (compatibility version 6.0.0, current version 6.2.0)
    @rpath/./libgmp.10.dylib (compatibility version 12.0.0, current version 12.3.0)
    @rpath/libCGAL_ImageIO.10.dylib (compatibility version 10.0.0, current version 10.0.2)
    @rpath/libCGAL.10.dylib (compatibility version 10.0.0, current version 10.0.2)
    /usr/local/lib/libboost_thread-mt.dylib (compatibility version 0.0.0, current version 0.0.0)
    /usr/local/lib/libboost_system-mt.dylib (compatibility version 0.0.0, current version 0.0.0)
    /System/Library/Frameworks/AGL.framework/Versions/A/AGL (compatibility version 1.0.0, current version 1.0.0)
    /System/Library/Frameworks/OpenGL.framework/Versions/A/OpenGL (compatibility version 1.0.0, current version 1.0.0)
    /usr/lib/libz.1.dylib (compatibility version 1.0.0, current version 1.2.5)
    /Users/pwolfram/miniconda2/conda-bld/work/lib/libCGAL_Triangulation_2_cpp.dylib (compatibility version 0.0.0, current version 0.0.0)
    /System/Library/Frameworks/Python.framework/Versions/2.7/Python (compatibility version 2.7.0, current version 2.7.5)
    /Users/pwolfram/miniconda2/conda-bld/work/lib/libCGAL_Kernel_cpp.dylib (compatibility version 0.0.0, current version 0.0.0)
    /usr/lib/libstdc++.6.dylib (compatibility version 7.0.0, current version 60.0.0)

(e.g., perhaps something similar to Scenario 2 from https://github.com/Homebrew/homebrew-science/issues/3401)

It is probably due to not linking to the brew python framework, if you are using that for your version of boost (see also boostorg/build#78).

[pwolfram]

Summary

1. Paths are not installed correctly in OS X. Potentially solution is modifying the build.sh to use @loader_path or @rpath instead of $ORIGIN (http://conda.pydata.org/docs/building/shared-libraries.html, https://cmake.org/Wiki/CMake_RPATH_handling)
2. Shared object files, e.g., /Users/rusty/anaconda/envs/tom_test/lib/python2.7/site-packages/CGAL/_CGAL_Triangulation_2.so appear to link to system python framework /System/Library/Frameworks/Python.framework/Versions/2.7/Python that may be different than boost installation, e.g., /usr/local/lib/libboost_thread-mt.dylib, and result in Fatal Python error: PyThreadState_Get: no current thread errors. This appears to be due to linking to the system python framework, even though boost may have been built with a brew version of python (use -undefined dynamic_lookup on darwin in python_for_extensions boostorg/build#78). Essentially, the \*.so library cannot be linked to a python framework that is different than the other frameworks used to build the libraries linked to for CGAL and the python bindings. "The best solution is to provide -bundle -undefined dynamic_lookup, which has the same behavior as the Linux linker." (http://blog.tim-smith.us/2015/09/python-extension-modules-os-x/).

[pwolfram]

@rustychris, were you able to get a working copy of stomel on OSX, even with some hacking by hand? It seems like it should be possible.

[pwolfram]

Also, for future reference to demonstrate that this works well on linux, in general, with only minor modifications: see my fork at https://github.com/pwolfram/conda-recipes/tree/icc_mods which was used to build on institutional computing. Current anaconda channel packages for pwolfram are for this build (cc @vanroekel).

I tested it to make sure that tom_test.ipynb worked (no errors and good output), same was true for the new refine_existing_grid.ipynb.

[rustychris]

@pwolfram, I was able to eventually get a good compile on OSX, though have not yet been able to get the rpath stuff worked out. A few notes along the way:

• I do run into the PyThreadState issue on occasion, and while it's difficult to trace back to exact causes, I have been able to get around this by judicious recompiling. Sorry - not a lot of help, mostly just solidarity.
• Adding -DPYTHON_LIBRARY=$PREFIX/lib/libpython2.7.dylib -DPYTHON_INCLUDE_DIR=$PREFIX/include/python2.7 to the cmake command seems to help with grabbing the anaconda python
• I briefly tried compiling against CGAL 4.8.1, and encountered some impassable compilation errors. For now, I'm sticking with CGAL 4.3.0.
• I've been trying in meta.yaml adding build: \n detect_binary_files_with_prefix: True. Unclear whether this fixes anything, but it has a nice ring to it.
• there are some patches from @xantares which might also be relevant here - and they seem to know more of the inner workings. If I can't make headway with a few more shots in the dark, that's my next line of attack.

Thanks for the patience - let's keep the conversation going and get this working!

[xantares]

@rustychris cgal is available from https://anaconda.org/conda-forge/cgal

[pwolfram]

@rustychris, thanks for sharing your notes! I suspect we may have all the information needed now and all that is left is to give this a couple more tries.

Regarding @xantares's comment, is the challenge of using CGAL 4.8.1 building the bindings using SWIG? I'm wondering if a bigger challenge here that would benefit everyone, including the larger computational geometry community, is production of generalized python bindings that are "officially supported" by the CGAL community. I recognize the challenge here but think it may be worth considering.

If I recall correctly, when we build the bindings used in stomel not all all of CGAL is available, just some key parts needed by core stomel algorithms, right? What I'm trying to get at is would it be possible to switch to something like https://github.com/CGAL/cgal-swig-bindings in stomel? The currently supported packages appear to be https://github.com/CGAL/cgal-swig-bindings/wiki/Package_wrappers_available but I don't see Appolonius, for instance.

If we were easily able to switch to the general bindings we may be able to effectively "work out" this issue with the installation because the respective communities are providing conda-compatible installations. This seems like the best long-term solution but there may be technical challenges I am not fully appreciating. For instance, is it possible @xantares to install the cgal-swig-bindings via conda, e.g., is CGAL/cgal-swig-bindings#70 still unresolved or is there an unsupported work-around?

[xantares]

@pwolfram, yes, https://anaconda.org/conda-forge/cgal conda package does provide cgal and cgal-swig-bindings.