Before compiling in Windows - opengeostat/pygslib GitHub Wiki

Compiling pygslib in windows

Note: The explanation below is obsolete. The easy way to compile in Windows in August 2022 is installing Miniconda and running the following commands

conda create --name pygslib-build python=3.9
conda activate pygslib-build
conda install cython setuptools numpy VC=14 vs2015_runtime libpython pytest m2w64-toolchain numpy pandas matplotlib scipy vtk colour bokeh ezdxf jupyter jupyterlab ipython

Note that m2w64-toolchain contains gfortran and libgfortran. Other key libraries are VC=14, vs2015_runtime, and libpython.

This page explains how to prepare prerequisites to compile pygslib for Python3.8 in Windows. The support for older Python versions is now dropped. This is valid only for Python distribution compiled with MSC 64 bits, which are the one distributed in the official python website and other popular distributions such as Anaconda.

Compiling pygslib in Linux is an easy task. All GNU compilers are installed by default and are fully compatible with most python distributions.

Compiling pygslib in Mac is not that easy anymore. You will require gfortran_osx-64, available as a conda package, and mac XCODE.

Update setuptools

You may update setuptools before traying to compile pygslib. Try this conda update setuptools if you are a conda user.

Installing MSVC compiler

Any Python package with compiled code may match the Python interpreter used. Most python interpreters for windows are compiled with Microsoft visual studio C compiler. You will need the right Microsoft VC runtime libraries installed. However, the compilation of your modules can be with alternative compilers such as GNU GCC and gfortran.

Check requirements in the official python wiki. You may want to install the compact version Visual C++ 14.0 to compile pygslib for Python 3.8.

Installing MSVC compiler for Python 3.8

Install the community version of Microsoft Build Tools for Visual Studio 2017.

Install the Mingw-w64 compiler

You will require gfortran to compile the code in Fortran and optionally gcc to compile C code generated by Cython and f2py. This is the tricky part. You have to make sure your compiler uses the right libgfortran version if you are planning to distribute the code. For example, pygslib v 0.0.0.6.3 uses the conda package m2w64-gcc-libgfortran as a dependency; this contains libgfortran-3.dll. I used Mingw-64 compiler x86_64-6.4.0-posix-seh. Mingw-64 version 7 and above may contain libgfortran-4.dll, so you have to make sure you use the right version of Mingw-64 to compile pygslib. Mingw-64 is available at Mingw-w64 oficial page.

Broken dependencies of the Fortran compiled code produce a DLL load failed error. You can use Dependencies_x64 (for Windows) program to find the missing DLL.

Make sure the path to mingw64\bin\ is in the windows system environment variable PATH. There are two ways of testing this:

  • going to the path and making sure the line C:\Program Files\mingw-w64\x86_64-6.4.0-posix-seh-rt_v5-rev0\mingw64\bin\ is there, and
  • calling gfortran and gcc and making sure they are recognized by the system. For example, the block of code below shows that gfortran was recognized by the system and executed.
c:>gfortran
gfortran: fatal error: no input files
compilation terminated.

Other general prerequisites

For Python 3.8 you may need only vc=14 and vs2015_runtime. Note that many popular packages, such as pandas and numpy, depend on these two packages. Conda distributes vs2015_runtime, and vc=14.1, but you may have it already installed as part of your python distribution.

Pygslib dependencies

To build and run pygslib make sure you have all the dependencies required, for example:

  • numpy
  • pandas
  • matplotlib
  • scipy
  • vc =14.1
  • vs2015_runtime
  • colour
  • pyevtk
  • ipython
  • vtk
  • ipython
  • bokeh
  • jupyter
  • ezdxf

see https://github.com/opengeostat/pygslib/blob/master/pygslib_conda/pygslib_Win64_py3.8/meta.yaml to obtain a complete list of dependencies.

Useful notes

The compiler MinGW-w64, for some obscure reason, copy some dependency dll in the folder .libs. These files have to be in the gslib folder, along the compiled extension modules with code in Fortran (with extension *.pyd). To fix this, we added a script at the end of setup.py with the code:

# copy dll if so is windows
  if os.name=='nt':
      mayor = sys.version_info[0]
      minor = sys.version_info[1]
      os.system('copy build\\lib.win-amd64-{}.{}\\pygslib\\.libs\\*.dll build\\lib.win-amd64-{}.{}\\pygslib\\gslib'.format(mayor,minor,mayor,minor))
      os.system('del /s /q build\\lib.win-amd64-{}.{}\\pygslib\\.libs'.format(mayor,minor))

This may duplicate the dll files in the destination site-packages\pygslib-x.x.x.x.x folder after running python setup.py build and python setup.py install. The solution is not very elegant, but it works!

Important! in Windows you need to build the package before creating a distribution, otherwise the command to copy dependence *.dll will not include theses files in the distribution. For example, to create the wheels (required for) conda packages) you may need to run these two commands:

python setup.py build
python setup.py sdist bdist_wheel