9 files changed, 802 insertions, 0 deletions
diff --git a/src/boost/libs/python/example/numpy/Jamfile b/src/boost/libs/python/example/numpy/Jamfile
new file mode 100644
index 000000000..ac70a6ff0
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/Jamfile
@@ -0,0 +1,29 @@
+# Copyright Stefan Seefeld 2016.
+# Distributed under the Boost Software License, Version 1.0.
+# (See accompanying file LICENSE_1_0.txt or copy at
+# http://www.boost.org/LICENSE_1_0.txt)
+
+import python ;
+
+# Adjust the following if Boost.Python isn't installed in a default location
+lib boost_numpy
+  :
+  : <search>/usr/local/Boost/lib
+    <include>/usr/local/Boost/include
+    ;
+
+project numpy
+  : requirements
+    <include>/usr/local/Boost/include
+    <library>boost_numpy
+    <location>.
+    ;
+
+exe simple : simple.cpp boost_numpy /python//python ;
+exe dtype : dtype.cpp boost_numpy /python//python ;
+exe ndarray : ndarray.cpp /python//python ;
+exe fromdata : fromdata.cpp /python//python ;
+exe ufunc : ufunc.cpp /python//python ;
+exe wrap : wrap.cpp /python//python ;
+
+python-extension gaussian : gaussian.cpp ;
diff --git a/src/boost/libs/python/example/numpy/demo_gaussian.py b/src/boost/libs/python/example/numpy/demo_gaussian.py
new file mode 100644
index 000000000..0b1c78995
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/demo_gaussian.py
@@ -0,0 +1,37 @@
+# Copyright Jim Bosch 2010-2012.
+# Distributed under the Boost Software License, Version 1.0.
+#    (See accompanying file LICENSE_1_0.txt or copy at
+#          http://www.boost.org/LICENSE_1_0.txt)
+
+import numpy
+import gaussian
+
+mu = numpy.zeros(2, dtype=float)
+sigma = numpy.identity(2, dtype=float)
+sigma[0, 1] = 0.15
+sigma[1, 0] = 0.15
+
+g = gaussian.bivariate_gaussian(mu, sigma)
+
+r = numpy.linspace(-40, 40, 1001)
+x, y = numpy.meshgrid(r, r)
+
+z = g(x, y)
+
+s = z.sum() * (r[1] - r[0])**2
+print "sum (should be ~ 1):", s
+
+xc = (z * x).sum() / z.sum()
+print "x centroid (should be ~ %f): %f" % (mu[0], xc)
+
+yc = (z * y).sum() / z.sum()
+print "y centroid (should be ~ %f): %f" % (mu[1], yc)
+
+xx = (z * (x - xc)**2).sum() / z.sum()
+print "xx moment (should be ~ %f): %f" % (sigma[0,0], xx)
+
+yy = (z * (y - yc)**2).sum() / z.sum()
+print "yy moment (should be ~ %f): %f" % (sigma[1,1], yy)
+
+xy = 0.5 * (z * (x - xc) * (y - yc)).sum() / z.sum()
+print "xy moment (should be ~ %f): %f" % (sigma[0,1], xy)
diff --git a/src/boost/libs/python/example/numpy/dtype.cpp b/src/boost/libs/python/example/numpy/dtype.cpp
new file mode 100644
index 000000000..749a36b5e
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/dtype.cpp
@@ -0,0 +1,49 @@
+// Copyright Ankit Daftery 2011-2012.
+// Distributed under the Boost Software License, Version 1.0.
+//    (See accompanying file LICENSE_1_0.txt or copy at
+//          http://www.boost.org/LICENSE_1_0.txt)
+
+/**
+ *  @brief An example to show how to create ndarrays with built-in python data types, and extract
+ *         the types and values of member variables
+ *
+ *  @todo Add an example to show type conversion.
+ *        Add an example to show use of user-defined types
+ *        
+ */
+
+#include <boost/python/numpy.hpp>
+#include <iostream>
+
+namespace p = boost::python;
+namespace np = boost::python::numpy;
+
+int main(int argc, char **argv)
+{
+  // Initialize the Python runtime.
+  Py_Initialize();
+  // Initialize NumPy
+  np::initialize();
+  // Create a 3x3 shape...
+  p::tuple shape = p::make_tuple(3, 3);
+  // ...as well as a type for C++ double
+  np::dtype dtype = np::dtype::get_builtin<double>();
+  // Construct an array with the above shape and type
+  np::ndarray a = np::zeros(shape, dtype);
+  // Print the array
+  std::cout << "Original array:\n" << p::extract<char const *>(p::str(a)) << std::endl;
+  // Print the datatype of the elements
+  std::cout << "Datatype is:\n" << p::extract<char const *>(p::str(a.get_dtype())) << std::endl ;
+  // Using user defined dtypes to create dtype and an array of the custom dtype
+  // First create a tuple with a variable name and its dtype, double, to create a custom dtype
+  p::tuple for_custom_dtype = p::make_tuple("ha",dtype) ;
+  // The list needs to be created, because the constructor to create the custom dtype
+  // takes a list of (variable,variable_type) as an argument
+  p::list list_for_dtype ;
+  list_for_dtype.append(for_custom_dtype) ;
+  // Create the custom dtype
+  np::dtype custom_dtype = np::dtype(list_for_dtype) ;
+  // Create an ndarray with the custom dtype
+  np::ndarray new_array = np::zeros(shape,custom_dtype);
+
+}
diff --git a/src/boost/libs/python/example/numpy/fromdata.cpp b/src/boost/libs/python/example/numpy/fromdata.cpp
new file mode 100644
index 000000000..bd073cc69
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/fromdata.cpp
@@ -0,0 +1,48 @@
+// Copyright Ankit Daftery 2011-2012.
+// Distributed under the Boost Software License, Version 1.0.
+//    (See accompanying file LICENSE_1_0.txt or copy at
+//          http://www.boost.org/LICENSE_1_0.txt)
+
+/**
+ *  @brief An example to show how to access data using raw pointers.  This shows that you can use and
+ *         manipulate data in either Python or C++ and have the changes reflected in both.
+ */
+
+#include <boost/python/numpy.hpp>
+#include <iostream>
+
+namespace p = boost::python;
+namespace np = boost::python::numpy;
+
+
+int main(int argc, char **argv)
+{
+  // Initialize the Python runtime.
+  Py_Initialize();
+  // Initialize NumPy
+  np::initialize();
+  // Create an array in C++
+  int arr[] = {1,2,3,4} ; 
+  // Create the ndarray in Python
+  np::ndarray py_array = np::from_data(arr, np::dtype::get_builtin<int>() , p::make_tuple(4), p::make_tuple(4), p::object());
+  // Print the ndarray that we just created, and the source C++ array
+  std::cout << "C++ array :" << std::endl ;
+  for (int j=0;j<4;j++)
+  {
+    std::cout << arr[j] << ' ' ;
+  }
+  std::cout << std::endl << "Python ndarray :" << p::extract<char const *>(p::str(py_array)) << std::endl;
+  // Change an element in the python ndarray
+  py_array[1] = 5 ; 
+  // And see if the C++ container is changed or not
+  std::cout << "Is the change reflected in the C++ array used to create the ndarray ? " << std::endl ; 
+  for (int j = 0;j<4 ; j++)
+  {
+    std::cout << arr[j] << ' ' ;
+  }
+  // Conversely, change it in C++
+  arr[2] = 8 ;
+  // And see if the changes are reflected in the Python ndarray
+  std::cout << std::endl << "Is the change reflected in the Python ndarray ?" << std::endl << p::extract<char const *>(p::str(py_array)) << std::endl;
+
+}
diff --git a/src/boost/libs/python/example/numpy/gaussian.cpp b/src/boost/libs/python/example/numpy/gaussian.cpp
new file mode 100644
index 000000000..5f138b397
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/gaussian.cpp
@@ -0,0 +1,315 @@
+// Copyright Jim Bosch 2010-2012.
+// Distributed under the Boost Software License, Version 1.0.
+//    (See accompanying file LICENSE_1_0.txt or copy at
+//          http://www.boost.org/LICENSE_1_0.txt)
+
+#include <boost/python/numpy.hpp>
+
+#include <cmath>
+#include <memory>
+
+#ifndef M_PI
+#include <boost/math/constants/constants.hpp>
+const double M_PI = boost::math::constants::pi<double>();
+#endif
+
+namespace bp = boost::python;
+namespace bn = boost::python::numpy;
+
+/**
+ *  A 2x2 matrix class, purely for demonstration purposes.
+ *
+ *  Instead of wrapping this class with Boost.Python, we'll convert it to/from numpy.ndarray.
+ */
+class matrix2 {
+public:
+
+    double & operator()(int i, int j) {
+        return _data[i*2 + j];
+    }
+
+    double const & operator()(int i, int j) const {
+        return _data[i*2 + j];
+    }
+    
+    double const * data() const { return _data; }
+
+private:
+    double _data[4];
+};
+
+/**
+ *  A 2-element vector class, purely for demonstration purposes.
+ *
+ *  Instead of wrapping this class with Boost.Python, we'll convert it to/from numpy.ndarray.
+ */
+class vector2 {
+public:
+
+    double & operator[](int i) {
+        return _data[i];
+    }
+
+    double const & operator[](int i) const {
+        return _data[i];
+    }
+    
+    double const * data() const { return _data; }
+
+    vector2 operator+(vector2 const & other) const {
+        vector2 r;
+        r[0] = _data[0] + other[0];
+        r[1] = _data[1] + other[1];
+        return  r;
+    }
+
+    vector2 operator-(vector2 const & other) const {
+        vector2 r;
+        r[0] = _data[0] - other[0];
+        r[1] = _data[1] - other[1];
+        return  r;
+    }
+
+private:
+    double _data[2];
+};
+
+/**
+ *  Matrix-vector multiplication.
+ */
+vector2 operator*(matrix2 const & m, vector2 const & v) {
+    vector2 r;
+    r[0] = m(0, 0) * v[0] + m(0, 1) * v[1];
+    r[1] = m(1, 0) * v[0] + m(1, 1) * v[1];
+    return r;
+}
+
+/**
+ *  Vector inner product.
+ */
+double dot(vector2 const & v1, vector2 const & v2) {
+    return v1[0] * v2[0] + v1[1] * v2[1];
+}
+
+/**
+ *  This class represents a simple 2-d Gaussian (Normal) distribution, defined by a
+ *  mean vector 'mu' and a covariance matrix 'sigma'.
+ */
+class bivariate_gaussian {
+public:
+
+    vector2 const & get_mu() const { return _mu; }
+
+    matrix2 const & get_sigma() const { return _sigma; }
+
+    /**
+     *  Evaluate the density of the distribution at a point defined by a two-element vector.
+     */
+    double operator()(vector2 const & p) const {
+        vector2 u = _cholesky * (p - _mu);
+        return 0.5 * _cholesky(0, 0) * _cholesky(1, 1) * std::exp(-0.5 * dot(u, u)) / M_PI;
+    }
+
+    /**
+     *  Evaluate the density of the distribution at an (x, y) point.
+     */
+    double operator()(double x, double y) const {
+        vector2 p;
+        p[0] = x;
+        p[1] = y;
+        return operator()(p);
+    }
+
+    /**
+     *  Construct from a mean vector and covariance matrix.
+     */
+    bivariate_gaussian(vector2 const & mu, matrix2 const & sigma)
+        : _mu(mu), _sigma(sigma), _cholesky(compute_inverse_cholesky(sigma))
+    {}
+    
+private:
+
+    /**
+     *  This evaluates the inverse of the Cholesky factorization of a 2x2 matrix;
+     *  it's just a shortcut in evaluating the density.
+     */
+    static matrix2 compute_inverse_cholesky(matrix2 const & m) {
+        matrix2 l;
+        // First do cholesky factorization: l l^t = m
+        l(0, 0) = std::sqrt(m(0, 0));
+        l(0, 1) = m(0, 1) / l(0, 0);
+        l(1, 1) = std::sqrt(m(1, 1) - l(0,1) * l(0,1));
+        // Now do forward-substitution (in-place) to invert:
+        l(0, 0) = 1.0 / l(0, 0);
+        l(1, 0) = l(0, 1) = -l(0, 1) / l(1, 1);
+        l(1, 1) = 1.0 / l(1, 1);
+        return l;
+    }
+
+    vector2 _mu;
+    matrix2 _sigma;
+    matrix2 _cholesky;
+                        
+};
+
+/*
+ *  We have a two options for wrapping get_mu and get_sigma into NumPy-returning Python methods:
+ *   - we could deep-copy the data, making totally new NumPy arrays;
+ *   - we could make NumPy arrays that point into the existing memory.
+ *  The latter is often preferable, especially if the arrays are large, but it's dangerous unless
+ *  the reference counting is correct: the returned NumPy array needs to hold a reference that
+ *  keeps the memory it points to from being deallocated as long as it is alive.  This is what the
+ *  "owner" argument to from_data does - the NumPy array holds a reference to the owner, keeping it
+ *  from being destroyed.
+ *
+ *  Note that this mechanism isn't completely safe for data members that can have their internal
+ *  storage reallocated.  A std::vector, for instance, can be invalidated when it is resized,
+ *  so holding a Python reference to a C++ class that holds a std::vector may not be a guarantee
+ *  that the memory in the std::vector will remain valid.
+ */
+
+/**
+ *  These two functions are custom wrappers for get_mu and get_sigma, providing the shallow-copy
+ *  conversion with reference counting described above.
+ *
+ *  It's also worth noting that these return NumPy arrays that cannot be modified in Python;
+ *  the const overloads of vector::data() and matrix::data() return const references, 
+ *  and passing a const pointer to from_data causes NumPy's 'writeable' flag to be set to false.
+ */
+static bn::ndarray py_get_mu(bp::object const & self) {
+    vector2 const & mu = bp::extract<bivariate_gaussian const &>(self)().get_mu();
+    return bn::from_data(
+        mu.data(),
+        bn::dtype::get_builtin<double>(),
+        bp::make_tuple(2),
+        bp::make_tuple(sizeof(double)),
+        self
+    );  
+}
+static bn::ndarray py_get_sigma(bp::object const & self) {
+    matrix2 const & sigma = bp::extract<bivariate_gaussian const &>(self)().get_sigma();
+    return bn::from_data(
+        sigma.data(),
+        bn::dtype::get_builtin<double>(),
+        bp::make_tuple(2, 2),
+        bp::make_tuple(2 * sizeof(double), sizeof(double)),
+        self
+    );
+}
+
+/**
+ *  To allow the constructor to work, we need to define some from-Python converters from NumPy arrays
+ *  to the matrix/vector types.  The rvalue-from-python functionality is not well-documented in Boost.Python
+ *  itself; you can learn more from boost/python/converter/rvalue_from_python_data.hpp.
+ */
+
+/**
+ *  We start with two functions that just copy a NumPy array into matrix/vector objects.  These will be used
+ *  in the templated converted below.  The first just uses the operator[] overloads provided by
+ *  bp::object.
+ */
+static void copy_ndarray_to_mv2(bn::ndarray const & array, vector2 & vec) {
+    vec[0] = bp::extract<double>(array[0]);
+    vec[1] = bp::extract<double>(array[1]);
+}
+
+/**
+ *  Here, we'll take the alternate approach of using the strides to access the array's memory directly.
+ *  This can be much faster for large arrays.
+ */
+static void copy_ndarray_to_mv2(bn::ndarray const & array, matrix2 & mat) {
+    // Unfortunately, get_strides() can't be inlined, so it's best to call it once up-front.
+    Py_intptr_t const * strides = array.get_strides();
+    for (int i = 0; i < 2; ++i) {
+        for (int j = 0; j < 2; ++j) {
+            mat(i, j) = *reinterpret_cast<double const *>(array.get_data() + i * strides[0] + j * strides[1]);
+        }
+    }
+}
+
+/**
+ *  Here's the actual converter.  Because we've separated the differences into the above functions,
+ *  we can write a single template class that works for both matrix2 and vector2.
+ */
+template <typename T, int N>
+struct mv2_from_python {
+    
+    /**
+     *  Register the converter.
+     */
+    mv2_from_python() {
+        bp::converter::registry::push_back(
+            &convertible,
+            &construct,
+            bp::type_id< T >()
+        );
+    }
+
+    /**
+     *  Test to see if we can convert this to the desired type; if not return zero.
+     *  If we can convert, returned pointer can be used by construct().
+     */
+    static void * convertible(PyObject * p) {
+        try {
+            bp::object obj(bp::handle<>(bp::borrowed(p)));
+            std::auto_ptr<bn::ndarray> array(
+                new bn::ndarray(
+                    bn::from_object(obj, bn::dtype::get_builtin<double>(), N, N, bn::ndarray::V_CONTIGUOUS)
+                )
+            );
+            if (array->shape(0) != 2) return 0;
+            if (N == 2 && array->shape(1) != 2) return 0;
+            return array.release();
+        } catch (bp::error_already_set & err) {
+            bp::handle_exception();
+            return 0;
+        }
+    }
+
+    /**
+     *  Finish the conversion by initializing the C++ object into memory prepared by Boost.Python.
+     */
+    static void construct(PyObject * obj, bp::converter::rvalue_from_python_stage1_data * data) {
+        // Extract the array we passed out of the convertible() member function.
+        std::auto_ptr<bn::ndarray> array(reinterpret_cast<bn::ndarray*>(data->convertible));
+        // Find the memory block Boost.Python has prepared for the result.
+        typedef bp::converter::rvalue_from_python_storage<T> storage_t;
+        storage_t * storage = reinterpret_cast<storage_t*>(data);
+        // Use placement new to initialize the result.
+        T * m_or_v = new (storage->storage.bytes) T();
+        // Fill the result with the values from the NumPy array.
+        copy_ndarray_to_mv2(*array, *m_or_v);
+        // Finish up.
+        data->convertible = storage->storage.bytes;
+    }
+
+};
+
+
+BOOST_PYTHON_MODULE(gaussian) {
+    bn::initialize();
+
+    // Register the from-python converters
+    mv2_from_python< vector2, 1 >();
+    mv2_from_python< matrix2, 2 >();
+
+    typedef double (bivariate_gaussian::*call_vector)(vector2 const &) const;
+
+    bp::class_<bivariate_gaussian>("bivariate_gaussian", bp::init<bivariate_gaussian const &>())
+
+        // Declare the constructor (wouldn't work without the from-python converters).
+        .def(bp::init< vector2 const &, matrix2 const & >())
+
+        // Use our custom reference-counting getters
+        .add_property("mu", &py_get_mu)
+        .add_property("sigma", &py_get_sigma)
+
+        // First overload accepts a two-element array argument
+        .def("__call__", (call_vector)&bivariate_gaussian::operator())
+
+        // This overload works like a binary NumPy universal function: you can pass
+        // in scalars or arrays, and the C++ function will automatically be called
+        // on each element of an array argument.
+        .def("__call__", bn::binary_ufunc<bivariate_gaussian,double,double,double>::make())
+        ;
+}
diff --git a/src/boost/libs/python/example/numpy/ndarray.cpp b/src/boost/libs/python/example/numpy/ndarray.cpp
new file mode 100644
index 000000000..d7b57aa73
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/ndarray.cpp
@@ -0,0 +1,71 @@
+// Copyright Ankit Daftery 2011-2012.
+// Distributed under the Boost Software License, Version 1.0.
+//    (See accompanying file LICENSE_1_0.txt or copy at
+//          http://www.boost.org/LICENSE_1_0.txt)
+
+/**
+ *  @brief An example to show how to create ndarrays using arbitrary Python sequences.
+ *
+ *  The Python sequence could be any object whose __array__ method returns an array, or any
+ *  (nested) sequence.  This example also shows how to create arrays using both unit and
+ *  non-unit strides.      
+ */
+
+#include <boost/python/numpy.hpp>
+#include <iostream>
+
+namespace p = boost::python;
+namespace np = boost::python::numpy;
+
+#if _MSC_VER
+using boost::uint8_t;
+#endif
+
+int main(int argc, char **argv)
+{
+  // Initialize the Python runtime.
+  Py_Initialize();
+  // Initialize NumPy
+  np::initialize();
+  // Create an ndarray from a simple tuple
+  p::object tu = p::make_tuple('a','b','c') ;
+  np::ndarray example_tuple = np::array (tu) ; 
+  // and from a list
+  p::list l ;
+  np::ndarray example_list = np::array (l) ; 
+  // Optionally, you can also specify a dtype
+  np::dtype dt = np::dtype::get_builtin<int>();
+  np::ndarray example_list1 = np::array (l,dt);
+  // You can also create an array by supplying data.First,create an integer array
+  int data[] = {1,2,3,4} ;
+  // Create a shape, and strides, needed by the function
+  p::tuple shape = p::make_tuple(4) ;
+  p::tuple stride = p::make_tuple(4) ; 
+  // The function also needs an owner, to keep track of the data array passed. Passing none is dangerous
+  p::object own ;
+  // The from_data function takes the data array, datatype,shape,stride and owner as arguments
+  // and returns an ndarray
+  np::ndarray data_ex = np::from_data(data,dt,shape,stride,own);
+  // Print the ndarray we created
+  std::cout << "Single dimensional array ::" << std::endl << p::extract < char const * > (p::str(data_ex)) << std::endl ; 
+  // Now lets make an 3x2 ndarray from a multi-dimensional array using non-unit strides
+  // First lets create a 3x4 array of 8-bit integers
+  uint8_t mul_data[][4] = {{1,2,3,4},{5,6,7,8},{1,3,5,7}};
+  // Now let's create an array of 3x2 elements, picking the first and third elements from each row
+  // For that, the shape will be 3x2
+  shape = p::make_tuple(3,2) ;
+  // The strides will be 4x2 i.e. 4 bytes to go to the next desired row, and 2 bytes to go to the next desired column
+  stride = p::make_tuple(4,2) ; 
+  // Get the numpy dtype for the built-in 8-bit integer data type
+  np::dtype dt1 = np::dtype::get_builtin<uint8_t>();
+  // First lets create and print out the ndarray as is
+  np::ndarray mul_data_ex = np::from_data(mul_data,dt1, p::make_tuple(3,4),p::make_tuple(4,1),p::object());
+  std::cout << "Original multi dimensional array :: " << std::endl << p::extract < char const * > (p::str(mul_data_ex)) << std::endl ; 
+  // Now create the new ndarray using the shape and strides
+  mul_data_ex = np::from_data(mul_data,dt1, shape,stride,p::object());
+  // Print out the array we created using non-unit strides
+  std::cout << "Selective multidimensional array :: "<<std::endl << p::extract < char const * > (p::str(mul_data_ex)) << std::endl ; 
+
+}
+
+
diff --git a/src/boost/libs/python/example/numpy/simple.cpp b/src/boost/libs/python/example/numpy/simple.cpp
new file mode 100644
index 000000000..ad598bde4
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/simple.cpp
@@ -0,0 +1,32 @@
+// Copyright 2011 Stefan Seefeld.
+// Distributed under the Boost Software License, Version 1.0. (See
+// accompanying file LICENSE_1_0.txt or copy at
+// http://www.boost.org/LICENSE_1_0.txt)
+
+#include <boost/python/numpy.hpp>
+#include <iostream>
+
+namespace p = boost::python;
+namespace np = boost::python::numpy;
+
+int main(int argc, char **argv)
+{
+  // Initialize the Python runtime.
+  Py_Initialize();
+  // Initialize NumPy
+  np::initialize();
+  // Create a 3x3 shape...
+  p::tuple shape = p::make_tuple(3, 3);
+  // ...as well as a type for C++ float
+  np::dtype dtype = np::dtype::get_builtin<float>();
+  // Construct an array with the above shape and type
+  np::ndarray a = np::zeros(shape, dtype);
+  // Construct an empty array with the above shape and dtype as well
+  np::ndarray b = np::empty(shape,dtype);
+  // Print the array
+  std::cout << "Original array:\n" << p::extract<char const *>(p::str(a)) << std::endl;
+  // Reshape the array into a 1D array
+  a = a.reshape(p::make_tuple(9));
+  // Print it again.
+  std::cout << "Reshaped array:\n" << p::extract<char const *>(p::str(a)) << std::endl;
+}
diff --git a/src/boost/libs/python/example/numpy/ufunc.cpp b/src/boost/libs/python/example/numpy/ufunc.cpp
new file mode 100644
index 000000000..5fb692045
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/ufunc.cpp
@@ -0,0 +1,86 @@
+// Copyright Ankit Daftery 2011-2012.
+// Distributed under the Boost Software License, Version 1.0.
+//    (See accompanying file LICENSE_1_0.txt or copy at
+//          http://www.boost.org/LICENSE_1_0.txt)
+
+/**
+ *  @brief An example to demonstrate use of universal functions or ufuncs
+ *        
+ *
+ * @todo Calling the overloaded () operator is in a roundabout manner, find a simpler way
+ *       None of the methods like np::add, np::multiply etc are supported as yet
+ */
+
+#include <boost/python/numpy.hpp>
+#include <iostream>
+
+namespace p = boost::python;
+namespace np = boost::python::numpy;
+
+
+// Create the structs necessary to implement the ufuncs 
+// The typedefs *must* be made
+
+struct UnarySquare 
+{
+  typedef double argument_type;
+  typedef double result_type;
+
+  double operator()(double r) const { return r * r;}
+};
+
+struct BinarySquare
+{
+  typedef double first_argument_type;
+  typedef double second_argument_type;
+  typedef double result_type;
+
+  double operator()(double a,double b) const { return (a*a + b*b) ; }
+};
+
+int main(int argc, char **argv)
+{
+  // Initialize the Python runtime.
+  Py_Initialize();
+  // Initialize NumPy
+  np::initialize();
+  // Expose the struct UnarySquare to Python as a class, and let ud be the class object
+  p::object ud = p::class_<UnarySquare, boost::shared_ptr<UnarySquare> >("UnarySquare")
+    .def("__call__", np::unary_ufunc<UnarySquare>::make());
+  // Let inst be an instance of the class ud
+  p::object inst = ud();
+  // Use the "__call__" method to call the overloaded () operator and print the value
+  std::cout << "Square of unary scalar 1.0 is " << p::extract <char const * > (p::str(inst.attr("__call__")(1.0))) << std::endl ; 
+  // Create an array in C++
+  int arr[] = {1,2,3,4} ; 
+  // ..and use it to create the ndarray in Python
+  np::ndarray demo_array = np::from_data(arr, np::dtype::get_builtin<int>() , p::make_tuple(4), p::make_tuple(4), p::object());
+  // Print out the demo array
+  std::cout << "Demo array is " << p::extract <char const * > (p::str(demo_array)) << std::endl ; 
+  // Call the "__call__" method to perform the operation and assign the value to result_array
+  p::object result_array = inst.attr("__call__")(demo_array) ;
+  // Print the resultant array
+  std::cout << "Square of demo array is " << p::extract <char const * > (p::str(result_array)) << std::endl ; 
+  // Lets try the same with a list
+  p::list li ;
+  li.append(3);
+  li.append(7);
+  // Print out the demo list
+  std::cout << "Demo list is " << p::extract <char const * > (p::str(li)) << std::endl ; 
+  // Call the ufunc for the list
+  result_array = inst.attr("__call__")(li) ;
+  // And print the list out
+  std::cout << "Square of demo list is " << p::extract <char const * > (p::str(result_array)) << std::endl ; 
+  // Now lets try Binary ufuncs
+  // Expose the struct BinarySquare to Python as a class, and let ud be the class object
+  ud = p::class_<BinarySquare, boost::shared_ptr<BinarySquare> >("BinarySquare")
+    .def("__call__", np::binary_ufunc<BinarySquare>::make());
+  // Again initialise inst as an instance of the class ud
+  inst = ud();
+  // Print the two input listsPrint the two input lists
+  std::cout << "The two input list for binary ufunc are " << std::endl << p::extract <char const * > (p::str(demo_array)) << std::endl << p::extract <char const * > (p::str(demo_array)) << std::endl ; 
+  // Call the binary ufunc taking demo_array as both inputs
+  result_array = inst.attr("__call__")(demo_array,demo_array) ;
+  std::cout << "Square of list with binary ufunc is " << p::extract <char const * > (p::str(result_array)) << std::endl ; 
+}
+  
diff --git a/src/boost/libs/python/example/numpy/wrap.cpp b/src/boost/libs/python/example/numpy/wrap.cpp
new file mode 100644
index 000000000..33f71c85a
--- /dev/null
+++ b/src/boost/libs/python/example/numpy/wrap.cpp
@@ -0,0 +1,135 @@
+// Copyright Jim Bosch 2011-2012.
+// Distributed under the Boost Software License, Version 1.0.
+//    (See accompanying file LICENSE_1_0.txt or copy at
+//          http://www.boost.org/LICENSE_1_0.txt)
+
+/**
+ *  A simple example showing how to wrap a couple of C++ functions that
+ *  operate on 2-d arrays into Python functions that take NumPy arrays
+ *  as arguments.
+ *        
+ *  If you find have a lot of such functions to wrap, you may want to
+ *  create a C++ array type (or use one of the many existing C++ array
+ *  libraries) that maps well to NumPy arrays and create Boost.Python
+ *  converters.  There's more work up front than the approach here,
+ *  but much less boilerplate per function.  See the "Gaussian" example
+ *  included with Boost.NumPy for an example of custom converters, or
+ *  take a look at the "ndarray" project on GitHub for a more complete,
+ *  high-level solution.
+ *
+ *  Note that we're using embedded Python here only to make a convenient
+ *  self-contained example; you could just as easily put the wrappers
+ *  in a regular C++-compiled module and imported them in regular
+ *  Python.  Again, see the Gaussian demo for an example.
+ */
+
+#include <boost/python/numpy.hpp>
+#include <boost/scoped_array.hpp>
+#include <iostream>
+
+namespace p = boost::python;
+namespace np = boost::python::numpy;
+
+// This is roughly the most efficient way to write a C/C++ function that operates
+// on a 2-d NumPy array - operate directly on the array by incrementing a pointer
+// with the strides.
+void fill1(double * array, int rows, int cols, int row_stride, int col_stride) {
+    double * row_iter = array;
+    double n = 0.0; // just a counter we'll fill the array with.
+    for (int i = 0; i < rows; ++i, row_iter += row_stride) {
+        double * col_iter = row_iter;
+        for (int j = 0; j < cols; ++j, col_iter += col_stride) {
+            *col_iter = ++n;
+        }
+    }
+}
+
+// Here's a simple wrapper function for fill1.  It requires that the passed
+// NumPy array be exactly what we're looking for - no conversion from nested
+// sequences or arrays with other data types, because we want to modify it
+// in-place.
+void wrap_fill1(np::ndarray const & array) {
+    if (array.get_dtype() != np::dtype::get_builtin<double>()) {
+        PyErr_SetString(PyExc_TypeError, "Incorrect array data type");
+        p::throw_error_already_set();
+    }
+    if (array.get_nd() != 2) {
+        PyErr_SetString(PyExc_TypeError, "Incorrect number of dimensions");
+        p::throw_error_already_set();
+    }
+    fill1(reinterpret_cast<double*>(array.get_data()),
+          array.shape(0), array.shape(1),
+          array.strides(0) / sizeof(double), array.strides(1) / sizeof(double));
+}
+
+// Another fill function that takes a double**.  This is less efficient, because
+// it's not the native NumPy data layout, but it's common enough in C/C++ that
+// it's worth its own example.  This time we don't pass the strides, and instead
+// in wrap_fill2 we'll require the C_CONTIGUOUS bitflag, which guarantees that
+// the column stride is 1 and the row stride is the number of columns.  That
+// restricts the arrays that can be passed to fill2 (it won't work on most
+// subarray views or transposes, for instance).
+void fill2(double ** array, int rows, int cols) {
+    double n = 0.0; // just a counter we'll fill the array with.
+    for (int i = 0; i < rows; ++i) {
+        for (int j = 0; j < cols; ++j) {
+            array[i][j] = ++n;
+        }
+    }    
+}
+// Here's the wrapper for fill2; it's a little more complicated because we need
+// to check the flags and create the array of pointers.
+void wrap_fill2(np::ndarray const & array) {
+    if (array.get_dtype() != np::dtype::get_builtin<double>()) {
+        PyErr_SetString(PyExc_TypeError, "Incorrect array data type");
+        p::throw_error_already_set();
+    }
+    if (array.get_nd() != 2) {
+        PyErr_SetString(PyExc_TypeError, "Incorrect number of dimensions");
+        p::throw_error_already_set();
+    }
+    if (!(array.get_flags() & np::ndarray::C_CONTIGUOUS)) {
+        PyErr_SetString(PyExc_TypeError, "Array must be row-major contiguous");
+        p::throw_error_already_set();
+    }
+    double * iter = reinterpret_cast<double*>(array.get_data());
+    int rows = array.shape(0);
+    int cols = array.shape(1);
+    boost::scoped_array<double*> ptrs(new double*[rows]);
+    for (int i = 0; i < rows; ++i, iter += cols) {
+        ptrs[i] = iter;
+    }
+    fill2(ptrs.get(), array.shape(0), array.shape(1));
+}
+
+BOOST_PYTHON_MODULE(example) {
+    np::initialize();  // have to put this in any module that uses Boost.NumPy
+    p::def("fill1", wrap_fill1);
+    p::def("fill2", wrap_fill2);
+}
+
+int main(int argc, char **argv)
+{
+    // This line makes our module available to the embedded Python intepreter.
+# if PY_VERSION_HEX >= 0x03000000
+    PyImport_AppendInittab("example", &PyInit_example);
+# else
+    PyImport_AppendInittab("example", &initexample);
+# endif
+    // Initialize the Python runtime.
+    Py_Initialize();
+
+    PyRun_SimpleString(
+        "import example\n"
+        "import numpy\n"
+        "z1 = numpy.zeros((5,6), dtype=float)\n"
+        "z2 = numpy.zeros((4,3), dtype=float)\n"
+        "example.fill1(z1)\n"
+        "example.fill2(z2)\n"
+        "print z1\n"
+        "print z2\n"
+    );
+    Py_Finalize();
+}
+
+