procedures:boost_python

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision		 Previous revision
procedures:boost_python [2016/11/11 02:20] – xifprocedures:boost_python [2016/11/12 03:21] (current) – [NumPy arrays] xif
Line 1: Line 1:
 +====== Boost.Python on OS X with NumPy ======
  
 +Ur python program too slow at calculating the boiling point of tungsten under a 600bar atmosphere of francium fluoride ? Too baaad...\\
 +Well, let's implement some part of it in C++ using [[http://boostorg.github.io/python/|Boost.Python]] with its NumPy brand new module. That's easy, I promise.\\
 +This tuto uses Python 3.4, installed from [[https://www.python.org/downloads/mac-osx/]]. Should be easily adapted to other versions.
 +
 +===== Build Boost.Python =====
 +
 +  - Download Boost (dev version) : ''git clone --recursive https://github.com/boostorg/boost.git modular-boost''
 +  - ''./bootstrap.sh --with-libraries=python --with-python-version=3.4 --with-python-root=/Library/Frameworks/Python.framework/Versions/3.4/bin/python3.4''
 +  - ''sudo ./b2 toolset=clang cxxflags="-stdlib=libc++ -std=c++0x" linkflags="-stdlib=libc++" -j2 install''
 +  - If it works, you should see ''/usr/local/lib/libboost_python3.a'' and ''/usr/local/lib/libboost_numpy3.a''.
 +
 +===== Example Python module =====
 +
 +Exemple Makefile :
 +<code make>
 +NAME = gauss_seidel
 +CXX = clang++
 +CXXFLAGS += -Wall -std=c++1y
 +PYTHON_VER = 3.4
 +
 +NUMPY_ROOT = $(shell pip$(PYTHON_VER) show numpy | grep "Location:" | cut -d" " -f2-)
 +CXXFLAGS += -fPIC $(shell python$(PYTHON_VER)-config --cflags) -I$(NUMPY_ROOT)/numpy/core/include
 +LDFLAGS += -fPIC $(shell python$(PYTHON_VER)-config --ldflags)
 +#LDFLAGS += -lboost_python3 -lboost_numpy3  # Fails at module import
 +LDFLAGS += /usr/local/lib/libboost_python3.a /usr/local/lib/libboost_numpy3.a
 +
 +all: $(NAME).o
 + $(CXX) -shared $^ $(LDFLAGS) -o $(NAME).so
 +
 +$(NAME).o: $(NAME).cpp
 +
 +%.o: %.cpp
 + $(CXX) -o $@ -c $< $(CXXFLAGS) 
 +</code>
 +
 +Exemple code (''my_little_poney.cpp'') :
 +<code cpp>
 +#include <boost/python.hpp>
 +#include <boost/python/numpy.hpp>
 +
 +namespace py = boost::python;
 +namespace np = boost::python::numpy;
 +
 +np::ndarray shift_col (int k, np::ndarray M) {
 + size_t n = M.shape(0), p = M.shape(1);
 + np::ndarray R = np::empty(py::make_tuple(n,p), M.get_dtype());
 + for (size_t i = 0; i < n; ++i) {
 + for (size_t j = 0; j < p; ++j) {
 + R[i][(j+k)%p] = M[i][j];
 + }
 + }
 + return R;
 +}
 +
 +BOOST_PYTHON_MODULE(my_little_poney) {
 + np::initialize();
 + py::def("shift_col", shift_col);
 +}
 +</code>
 +
 +Then build with ''make''. It will create a ''my_little_poney.so'' dynamic lib which can be loaded as a python module with ''import my_little_poney''.
 +
 +==== NumPy arrays ====
 +
 +Here is the issue : all operation made with ''array[i][j]'' are executed through the python interpreter and are very inefficient.
 +You can use the ''xif::multiarr'' warper from [[http://dev.xif.fr:7979/xifutils/|xifutils]] to access directly and easily ''ndarray''s : <code cpp>
 +template <typename T, size_t dim>
 +struct nparray_accessor : public xif::multiarr<T,dim> {
 + nparray_accessor (np::ndarray& ndarr) : xif::multiarr<T,dim>( 
 + (T*)ndarr.get_data(), 
 + [&ndarr] (size_t d) -> size_t { return ndarr.shape(d); }
 + ) {}
 +};
 +</code>
 +''nparray_accessor<T,dim>'' takes ''T'' as the data type (should match the ''ndarray'''s ''dtype'') and ''dim'' the number of dimensions.
 +
 +Documentation about Boost.Python.NumPy : [[http://boostorg.github.io/python/develop/doc/html/numpy/]].