New installation guide for CUDA builds (#655)

docs/source/install/index.rst

@@ -21,6 +21,10 @@ DGL supports multiple tensor libraries (e.g. PyTorch, MXNet) as backends; refer
 `Working with different backends`_ for requirements on backends and how to select a
 backend.
 
+Starting from 0.3 DGL is separated into CPU and CUDA builds.  The builds share the
+same Python package name, so installing DGL with CUDA 9 build after installing the
+CPU build will overwrite the latter.
+
 Install from conda
 ----------------------
 One can either grab `miniconda <https://conda.io/miniconda.html>`_ or
@@ -31,16 +35,58 @@ Once the conda environment is activated, run
 
 .. code:: bash
 
-   conda install -c dglteam dgl
+   conda install -c dglteam dgl              # For CPU Build
+   conda install -c dglteam dgl-cuda9.0      # For CUDA 9.0 Build
+   conda install -c dglteam dgl-cuda10.0     # For CUDA 10.0 Build
 
 Install from pip
 ----------------
-One can simply run the following command to install via ``pip``:
+For CPU builds, one can simply run the following command to install via ``pip``:
 
 .. code:: bash
 
    pip install dgl
    
+For CUDA builds, one needs to specify the URL:
+
+.. code:: bash
+
+   pip install https://s3.us-east-2.amazonaws.com/dgl.ai/<BUILD-DIR>/dgl-0.3-<PYTHON-ABI>-<PLATFORM>.whl
+   
+where ``<BUILD-DIR>``, ``<PYTHON-ABI>`` and ``<PLATFORM>`` can take either of the following values:
+
++------------+---------------------+
+| Build Type | ``<BUILD>``         |
++============+=====================+
+| CUDA 9.0   | ``wheels/cuda9.0``  |
++------------+---------------------+
+| CUDA 10.0  | ``wheels/cuda10.0`` |
++------------+---------------------+
+
++------------------+------------------+
+| Python Version   | ``<PYTHON-ABI>`` |
++==================+==================+
+| Python 3.5       | ``cp35-cp35m``   |
++------------------+------------------+
+| Python 3.6       | ``cp36-cp36m``   |
++------------------+------------------+
+| Python 3.7       | ``cp37-cp37m``   |
++------------------+------------------+
+
++-------------+-----------------------+
+| OS/Arch     | ``<PLATFORM>``        |
++=============+=======================+
+| Linux x64   | ``manylinux1_x86_64`` |
++-------------+-----------------------+
+| Windows x64 | ``win_amd64``         |
++-------------+-----------------------+
+
+For instance, if one wants to install the CUDA 9.0 build on Linux, Python 3.5, then the command is
+
+.. code:: bash
+
+   pip install https://s3.us-east-2.amazonaws.com/dgl.ai/wheels/cuda9.0/dgl-0.3-cp35-cp35m-manylinux1_x86_64.whl
+
 Working with different backends
 -------------------------------