Prerequisites

The bindings (and demonstration applications) should work well both for python2.7 and python3.4 (or higher), and across platforms.

(Currently only tested on Ubuntu 14.04 and OS X 10.10 with python3.4 and a Halide source build).

Python requirements:

See requirements.txt (to be used with pip: pip install --user requirements.txt)

C++ requirements:

Halide compiled (assuming halide_git_clone_path/build). For a better python experience make sure Halide was compiled with WITH_EXCEPTIONS Boost.python (already available in most boost distributions) Recommended (used for I/O): Boost.Numpy https://github.com/ndarray/Boost.NumPy

Compilation instructions

Build using:

  mkdir build
  cd build
  ccmake ..
  make 

Documentation and Examples

The Python API reflects directly the C++ Halide API. Consult the module documentation.

Check out the code for the example applications in the apps/ and tutorial/ subdirectory.

You can run them as a batch via:

    run_apps.sh
    run_tutorial.sh

To run these examples, make sure the PYTHONPATH environment variable points to your build directory (e.g. export PYTHONPATH=halide_source/python_bindings/build:$PYTHONPATH).

License

The Python bindings use the same MIT license as Halide.

Python bindings provided by Connelly Barnes (2012-2013), Fred Rotbart (2014), Rodrigo Benenson (2015) and the Halide open-source community.

Why did we move from SWIG to Boost.Python ?

Problems with (old) SWIG bindings:

• P1) SWIG cannot parse C++15, and it will always lag
• P2) Python code is not compiled/tested together with C++ codebase
• P3) Albeit SWIG is ment to provide "automagic" conversion, current code base has explicit code for every class (and class member) of interest
• P4) Old swig bindings __init__.py file has a lot of "python magic" to get things running, code is spread around multiple pieces (py_util.cpp, cHalide.i, and __init__.py)
• P5) Old swig python bindings did not work anymore (swig issues, compilation issues)

The boost python approach helps because:

• a) P1 is solved out of the box, since no C++ parsing is involved
• b) P2 is mitigated since python bindings are also in C++. Any miss-match with generate a compilation error.
• c) P3, P4 and P5 are expected to be handled by having less spaghetti and aiming for tighter integration with rest of code-base (point b)