libgpuarray on Win10

This guide assumes:

1. Dev environment as described here
2. distutils.cfg set up to use mingw
3. cmake (installable with pacman)

To install libgpuarry and pygpu:

1. Run:

   LGA_DIR=$DEV/libgpuarray
   git clone https://github.com/Theano/libgpuarray.git
   cd libgpuarray
   mkdir build && cd build
   cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$LGA_DIR -G "Unix Makefiles"
   make && make install

2. libgpuarray should now be in $LGA_DIR. Add $LGA_DIR/bin to your $PATH

3. Go back to the libgpuarray source directory (in a new bash window, so that it has the updated $PATH)

4. Edit setup.py, and remove if sys.platform == 'win32' and not os.getenv('CONDA_BUILD') and the 9 lines after it

5. Run:

   python setup.py build_ext -L $LGA_DIR/lib -I $LGA_DIR/include
   python setup.py install

6. Test your install with DEVICE=cuda python -c "import pygpu;pygpu.test()". In my case, 40/6894 failed.