python/doc/tutorial.qbk

[article Boost.Python Tutorial
    [quickbook 1.6]
    [authors [de Guzman, Joel], [Abrahams, David]]
    [copyright 2002 2003 2004 2005 Joel de Guzman, David Abrahams]
    [category inter-language support]
    [id tutorial]
    [purpose
        Reflects C++ classes and functions into Python
    ]
    [license
        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]
    ]
]

[/ QuickBook Document version 0.9 ]

[def __note__       [$../images/note.png]]
[def __alert__      [$../images/alert.png]]
[def __tip__        [$../images/tip.png]]
[def :-)            [$../images/smiley.png]]
[def __jam__        [$../images/jam.png]]

[section QuickStart]

The Boost Python Library is a framework for interfacing Python and
C++. It allows you to quickly and seamlessly expose C++ classes
functions and objects to Python, and vice-versa, using no special
tools -- just your C++ compiler. It is designed to wrap C++ interfaces
non-intrusively, so that you should not have to change the C++ code at
all in order to wrap it, making Boost.Python ideal for exposing
3rd-party libraries to Python. The library's use of advanced
metaprogramming techniques simplifies its syntax for users, so that
wrapping code takes on the look of a kind of declarative interface
definition language (IDL).

[h2 Hello World]

Following C/C++ tradition, let's start with the "hello, world". A C++
Function:

    char const* greet()
    {
       return "hello, world";
    }

can be exposed to Python by writing a Boost.Python wrapper:

    #include <boost/python.hpp>

    BOOST_PYTHON_MODULE(hello_ext)
    {
        using namespace boost::python;
        def("greet", greet);
    }

That's it. We're done. We can now build this as a shared library. The
resulting DLL is now visible to Python. Here's a sample Python session:

[python]

    >>> import hello_ext
    >>> print hello_ext.greet()
    hello, world

[c++]

[:['[*Next stop... Building your Hello World module from start to finish...]]]

[endsect]
[section:hello Building Hello World]

[h2 From Start To Finish]

Now the first thing you'd want to do is to build the Hello World module and
try it for yourself in Python. In this section, we will outline the steps
necessary to achieve that. We will use the build tool that comes bundled
with every boost distribution: [*bjam].

[note [*Building without bjam]

Besides bjam, there are of course other ways to get your module built.
What's written here should not be taken as "the one and only way".
There are of course other build tools apart from [^bjam].

Take note however that the preferred build tool for Boost.Python is bjam.
There are so many ways to set up the build incorrectly. Experience shows
that 90% of the "I can't build Boost.Python" problems come from people
who had to use a different tool.
]

We will skip over the details. Our objective will be to simply create
the hello world module and run it in Python. For a complete reference to
building Boost.Python, check out: [@../building.html
building.html]. After this brief ['bjam] tutorial, we should have built
the DLLs and run a python program using the extension.

The tutorial example can be found in the directory:
[^libs/python/example/tutorial]. There, you can find:

* hello.cpp
* hello.py
* Jamroot

The [^hello.cpp] file is our C++ hello world example. The [^Jamroot] is
a minimalist ['bjam] script that builds the DLLs for us. Finally,
[^hello.py] is our Python program that uses the extension in
[^hello.cpp].

Before anything else, you should have the bjam executable in your boost
directory or somewhere in your path such that [^bjam] can be executed in
the command line. Pre-built Boost.Jam executables are available for most
platforms. The complete list of Bjam executables can be found
[@http://sourceforge.net/project/showfiles.php?group_id=7586 here].

[h2 Let's Jam!]
__jam__

[@../../../../example/tutorial/Jamroot Here] is our minimalist Jamroot
file. Simply copy the file and tweak [^use-project boost] to where your
boost root directory is and you're OK.

The comments contained in the Jamroot file above should be sufficient
to get you going.

[h2 Running bjam]

['bjam] is run using your operating system's command line interpreter.

[:Start it up.]

A file called user-config.jam in your home directory is used to
configure your tools. In Windows, your home directory can be found by
typing:

[pre
ECHO %HOMEDRIVE%%HOMEPATH%
]

into a command prompt window. Your file should at least have the rules
for your compiler and your python installation. A specific example of
this on Windows would be:

[pre
#  MSVC configuration
using msvc : 8.0 ;

#  Python configuration
using python : 2.4 : C:/dev/tools/Python/ ;
]

The first rule tells Bjam to use the MSVC 8.0 compiler and associated
tools. The second rule provides information on Python, its version and
where it is located. The above assumes that the Python installation is
in [^C:/dev/tools\/Python/]. If you have one fairly "standard" python
installation for your platform, you might not need to do this.

Now we are ready... Be sure to [^cd] to [^libs/python/example/tutorial]
where the tutorial [^"hello.cpp"] and the [^"Jamroot"] is situated.

Finally:

    bjam

It should be building now:

[pre
cd C:\dev\boost\libs\python\example\tutorial
bjam
...patience...
...found 1101 targets...
...updating 35 targets...
]

And so on... Finally:

[pre
   Creating library /path-to-boost_python.dll/
   Creating library /path-to-'''hello_ext'''.exp/
'''**passed**''' ... hello.test
...updated 35 targets...
]

Or something similar. If all is well, you should now have built the DLLs and
run the Python program.

[:[*There you go... Have fun!]]

[endsect]
[section:exposing Exposing Classes]

Now let's expose a C++ class to Python.

Consider a C++ class/struct that we want to expose to Python:

    struct World
    {
        void set(std::string msg) { this->msg = msg; }
        std::string greet() { return msg; }
        std::string msg;
    };

We can expose this to Python by writing a corresponding Boost.Python
C++ Wrapper:

    #include <boost/python.hpp>
    using namespace boost::python;

    BOOST_PYTHON_MODULE(hello)
    {
        class_<World>("World")
            .def("greet", &World::greet)
            .def("set", &World::set)
        ;
    }

Here, we wrote a C++ class wrapper that exposes the member functions
[^greet] and [^set]. Now, after building our module as a shared library, we
may use our class [^World] in Python. Here's a sample Python session:

[python]

    >>> import hello
    >>> planet = hello.World()
    >>> planet.set('howdy')
    >>> planet.greet()
    'howdy'

[section Constructors]

Our previous example didn't have any explicit constructors.
Since [^World] is declared as a plain struct, it has an implicit default
constructor. Boost.Python exposes the default constructor by default,
which is why we were able to write

    >>> planet = hello.World()

We may wish to wrap a class with a non-default constructor. Let us
build on our previous example:

[c++]

    struct World
    {
        World(std::string msg): msg(msg) {} // added constructor
        void set(std::string msg) { this->msg = msg; }
        std::string greet() { return msg; }
        std::string msg;
    };

This time [^World] has no default constructor; our previous
wrapping code would fail to compile when the library tried to expose
it. We have to tell [^class_<World>] about the constructor we want to
expose instead.

    #include <boost/python.hpp>
    using namespace boost::python;

    BOOST_PYTHON_MODULE(hello)
    {
        class_<World>("World", init<std::string>())
            .def("greet", &World::greet)
            .def("set", &World::set)
        ;
    }

[^init<std::string>()] exposes the constructor taking in a
[^std::string] (in Python, constructors are spelled
"[^"__init__"]").

We can expose additional constructors by passing more [^init<...>]s to
the [^def()] member function. Say for example we have another World
constructor taking in two doubles:

    class_<World>("World", init<std::string>())
        .def(init<double, double>())
        .def("greet", &World::greet)
        .def("set", &World::set)
    ;

On the other hand, if we do not wish to expose any constructors at
all, we may use [^no_init] instead:

    class_<Abstract>("Abstract", no_init)

This actually adds an [^__init__] method which always raises a
Python RuntimeError exception.

[endsect]
[section Class Data Members]

Data members may also be exposed to Python so that they can be
accessed as attributes of the corresponding Python class. Each data
member that we wish to be exposed may be regarded as [*read-only] or
[*read-write].  Consider this class [^Var]:

    struct Var
    {
        Var(std::string name) : name(name), value() {}
        std::string const name;
        float value;
    };

Our C++ [^Var] class and its data members can be exposed to Python:

    class_<Var>("Var", init<std::string>())
        .def_readonly("name", &Var::name)
        .def_readwrite("value", &Var::value);

Then, in Python, assuming we have placed our Var class inside the namespace
hello as we did before:

[python]

    >>> x = hello.Var('pi')
    >>> x.value = 3.14
    >>> print x.name, 'is around', x.value
    pi is around 3.14

Note that [^name] is exposed as [*read-only] while [^value] is exposed
as [*read-write].

    >>> x.name = 'e' # can't change name
    Traceback (most recent call last):
      File "<stdin>", line 1, in ?
    AttributeError: can't set attribute

[endsect]
[section Class Properties]

In C++, classes with public data members are usually frowned
upon. Well designed classes that take advantage of encapsulation hide
the class' data members. The only way to access the class' data is
through access (getter/setter) functions. Access functions expose class
properties. Here's an example:

[c++]

    struct Num
    {
        Num();
        float get() const;
        void set(float value);
        ...
    };

However, in Python attribute access is fine; it doesn't neccessarily break
encapsulation to let users handle attributes directly, because the
attributes can just be a different syntax for a method call. Wrapping our
[^Num] class using Boost.Python:

    class_<Num>("Num")
        .add_property("rovalue", &Num::get)
        .add_property("value", &Num::get, &Num::set);

And at last, in Python:

[python]

    >>> x = Num()
    >>> x.value = 3.14
    >>> x.value, x.rovalue
    (3.14, 3.14)
    >>> x.rovalue = 2.17 # error!

Take note that the class property [^rovalue] is exposed as [*read-only]
since the [^rovalue] setter member function is not passed in:

[c++]

    .add_property("rovalue", &Num::get)

[endsect]
[section Inheritance]

In the previous examples, we dealt with classes that are not polymorphic.
This is not often the case. Much of the time, we will be wrapping
polymorphic classes and class hierarchies related by inheritance. We will
often have to write Boost.Python wrappers for classes that are derived from
abstract base classes.

Consider this trivial inheritance structure:

    struct Base { virtual ~Base(); };
    struct Derived : Base {};

And a set of C++ functions operating on [^Base] and [^Derived] object
instances:

    void b(Base*);
    void d(Derived*);
    Base* factory() { return new Derived; }

We've seen how we can wrap the base class [^Base]:

    class_<Base>("Base")
        /*...*/
        ;

Now we can inform Boost.Python of the inheritance relationship between
[^Derived] and its base class [^Base].  Thus:

    class_<Derived, bases<Base> >("Derived")
        /*...*/
        ;

Doing so, we get some things for free:

# Derived automatically inherits all of Base's Python methods
  (wrapped C++ member functions)
# [*If] Base is polymorphic, [^Derived] objects which have been passed to
  Python via a pointer or reference to [^Base] can be passed where a pointer
  or reference to [^Derived] is expected.

Now, we will expose the C++ free functions [^b] and [^d] and [^factory]:

    def("b", b);
    def("d", d);
    def("factory", factory);

Note that free function [^factory] is being used to generate new
instances of class [^Derived]. In such cases, we use
[^return_value_policy<manage_new_object>] to instruct Python to adopt
the pointer to [^Base] and hold the instance in a new Python [^Base]
object until the the Python object is destroyed. We will see more of
Boost.Python [link tutorial.functions.call_policies call policies] later.

    // Tell Python to take ownership of factory's result
    def("factory", factory,
        return_value_policy<manage_new_object>());

[endsect]

[section Class Virtual Functions]

In this section, we will learn how to make functions behave polymorphically
through virtual functions. Continuing our example, let us add a virtual function
to our [^Base] class:

    struct Base
    {
        virtual ~Base() {}
        virtual int f() = 0;
    };

One of the goals of Boost.Python is to be minimally intrusive on an existing C++
design. In principle, it should be possible to expose the interface for a 3rd
party library without changing it. It is not ideal to add anything to our class
`Base`. Yet, when you have a virtual function that's going to be overridden in
Python and called polymorphically *from C++*, we'll need to add some
scaffoldings to make things work properly. What we'll do is write a class
wrapper that derives from `Base` that will unintrusively hook into the virtual
functions so that a Python override may be called:

    struct BaseWrap : Base, wrapper<Base>
    {
        int f()
        {
            return this->get_override("f")();
        }
    };

Notice too that in addition to inheriting from `Base`, we also multiply-
inherited `wrapper<Base>` (See [@../reference/high_level_components/boost_python_wrapper_hpp.html#high_level_components.boost_python_wrapper_hpp.class_template_wrapper Wrapper]). The
`wrapper` template makes the job of wrapping classes that are meant to
overridden in Python, easier.

[blurb __alert__ [*MSVC6/7 Workaround]

If you are using Microsoft Visual C++ 6 or 7, you have to write `f` as:

`return call<int>(this->get_override("f").ptr());`.]

BaseWrap's overridden virtual member function `f` in effect calls the
corresponding method of the Python object through `get_override`.

Finally, exposing `Base`:

    class_<BaseWrap, boost::noncopyable>("Base")
        .def("f", pure_virtual(&Base::f))
        ;

`pure_virtual` signals Boost.Python that the function `f` is a pure virtual
function.

[note [*member function and methods]

Python, like many object oriented languages uses the term [*methods].
Methods correspond roughly to C++'s [*member functions]]

[endsect]

[section Virtual Functions with Default Implementations]

We've seen in the previous section how classes with pure virtual functions are
wrapped using Boost.Python's [@../reference/high_level_components/boost_python_wrapper_hpp.html#high_level_components.boost_python_wrapper_hpp.class_template_wrapper class wrapper]
facilities. If we wish to wrap [*non]-pure-virtual functions instead, the
mechanism is a bit different.

Recall that in the [link tutorial.exposing.class_virtual_functions previous section], we
wrapped a class with a pure virtual function that we then implemented in C++, or
Python classes derived from it. Our base class:

    struct Base
    {
        virtual int f() = 0;
    };

had a pure virtual function [^f]. If, however, its member function [^f] was
not declared as pure virtual:

    struct Base
    {
        virtual ~Base() {}
        virtual int f() { return 0; }
    };

We wrap it this way:

    struct BaseWrap : Base, wrapper<Base>
    {
        int f()
        {
            if (override f = this->get_override("f"))
                return f(); // *note*
            return Base::f();
        }

        int default_f() { return this->Base::f(); }
    };

Notice how we implemented `BaseWrap::f`. Now, we have to check if there is an
override for `f`. If none, then we call `Base::f()`.

[blurb __alert__ [*MSVC6/7 Workaround]

If you are using Microsoft Visual C++ 6 or 7, you have to rewrite the line
with the `*note*` as:

`return call<char const*>(f.ptr());`.]

Finally, exposing:

    class_<BaseWrap, boost::noncopyable>("Base")
        .def("f", &Base::f, &BaseWrap::default_f)
        ;

Take note that we expose both `&Base::f` and `&BaseWrap::default_f`.
Boost.Python needs to keep track of 1) the dispatch function [^f] and 2) the
forwarding function to its default implementation [^default_f]. There's a
special [^def] function for this purpose.

In Python, the results would be as expected:

[python]

    >>> base = Base()
    >>> class Derived(Base):
    ...     def f(self):
    ...         return 42
    ...
    >>> derived = Derived()

Calling [^base.f()]:

    >>> base.f()
    0

Calling [^derived.f()]:

    >>> derived.f()
    42

[endsect]
[section Class Operators/Special Functions]

[h2 Python Operators]

C is well known for the abundance of operators. C++ extends this to the
extremes by allowing operator overloading. Boost.Python takes advantage of
this and makes it easy to wrap C++ operator-powered classes.

Consider a file position class [^FilePos] and a set of operators that take
on FilePos instances:

[c++]

    class FilePos { /*...*/ };

    FilePos     operator+(FilePos, int);
    FilePos     operator+(int, FilePos);
    int         operator-(FilePos, FilePos);
    FilePos     operator-(FilePos, int);
    FilePos&    operator+=(FilePos&, int);
    FilePos&    operator-=(FilePos&, int);
    bool        operator<(FilePos, FilePos);

The class and the various operators can be mapped to Python rather easily
and intuitively:

    class_<FilePos>("FilePos")
        .def(self + int())          // __add__
        .def(int() + self)          // __radd__
        .def(self - self)           // __sub__
        .def(self - int())          // __sub__
        .def(self += int())         // __iadd__
        .def(self -= other<int>())
        .def(self < self);          // __lt__

The code snippet above is very clear and needs almost no explanation at
all. It is virtually the same as the operators' signatures. Just take
note that [^self] refers to FilePos object. Also, not every class [^T] that
you might need to interact with in an operator expression is (cheaply)
default-constructible. You can use [^other<T>()] in place of an actual
[^T] instance when writing "self expressions".

[h2 Special Methods]

Python has a few more ['Special Methods]. Boost.Python supports all of the
standard special method names supported by real Python class instances. A
similar set of intuitive interfaces can also be used to wrap C++ functions
that correspond to these Python ['special functions]. Example:

    class Rational
    { public: operator double() const; };

    Rational pow(Rational, Rational);
    Rational abs(Rational);
    ostream& operator<<(ostream&,Rational);

    class_<Rational>("Rational")
        .def(float_(self))                  // __float__
        .def(pow(self, other<Rational>))    // __pow__
        .def(abs(self))                     // __abs__
        .def(str(self))                     // __str__
        ;

Need we say more?

[note What is the business of `operator<<`?
Well, the method `str` requires the `operator<<` to do its work (i.e.
`operator<<` is used by the method defined by `def(str(self))`.]

[endsect]
[endsect] [/ Exposing Classes ]

[section Functions]

In this chapter, we'll look at Boost.Python powered functions in closer
detail. We will see some facilities to make exposing C++ functions to
Python safe from potential pifalls such as dangling pointers and
references. We will also see facilities that will make it even easier for
us to expose C++ functions that take advantage of C++ features such as
overloading and default arguments.

[:['Read on...]]

But before you do, you might want to fire up Python 2.2 or later and type
[^>>> import this].

[pre
>>> import this
The Zen of Python, by Tim Peters
Beautiful is better than ugly.
Explicit is better than implicit.
Simple is better than complex.
Complex is better than complicated.
Flat is better than nested.
Sparse is better than dense.
Readability counts.
Special cases aren't special enough to break the rules.
Although practicality beats purity.
Errors should never pass silently.
Unless explicitly silenced.
In the face of ambiguity, refuse the temptation to guess.
There should be one-- and preferably only one --obvious way to do it
Although that way may not be obvious at first unless you're Dutch.
Now is better than never.
Although never is often better than *right* now.
If the implementation is hard to explain, it's a bad idea.
If the implementation is easy to explain, it may be a good idea.
Namespaces are one honking great idea -- let's do more of those!
]

[section Call Policies]

In C++, we often deal with arguments and return types such as pointers
and references. Such primitive types are rather, ummmm, low level and
they really don't tell us much. At the very least, we don't know the
owner of the pointer or the referenced object. No wonder languages
such as Java and Python never deal with such low level entities. In
C++, it's usually considered a good practice to use smart pointers
which exactly describe ownership semantics. Still, even good C++
interfaces use raw references and pointers sometimes, so Boost.Python
must deal with them. To do this, it may need your help. Consider the
following C++ function:

    X& f(Y& y, Z* z);

How should the library wrap this function? A naive approach builds a
Python X object around result reference. This strategy might or might
not work out. Here's an example where it didn't

    >>> x = f(y, z) # x refers to some C++ X
    >>> del y
    >>> x.some_method() # CRASH!

What's the problem?

Well, what if f() was implemented as shown below:

    X& f(Y& y, Z* z)
    {
        y.z = z;
        return y.x;
    }

The problem is that the lifetime of result X& is tied to the lifetime
of y, because the f() returns a reference to a member of the y
object. This idiom is is not uncommon and perfectly acceptable in the
context of C++. However, Python users should not be able to crash the
system just by using our C++ interface. In this case deleting y will
invalidate the reference to X. We have a dangling reference.

Here's what's happening:

# [^f] is called passing in a reference to [^y] and a pointer to [^z]
# A reference to [^y.x] is returned
# [^y] is deleted. [^x] is a dangling reference
# [^x.some_method()] is called
# [*BOOM!]

We could copy result into a new object:

[python]

    >>> f(y, z).set(42) # Result disappears
    >>> y.x.get()       # No crash, but still bad
    3.14

This is not really our intent of our C++ interface. We've broken our
promise that the Python interface should reflect the C++ interface as
closely as possible.

Our problems do not end there. Suppose Y is implemented as follows:

[c++]

    struct Y
    {
        X x; Z* z;
        int z_value() { return z->value(); }
    };

Notice that the data member [^z] is held by class Y using a raw
pointer. Now we have a potential dangling pointer problem inside Y:

    >>> x = f(y, z) # y refers to z
    >>> del z       # Kill the z object
    >>> y.z_value() # CRASH!

For reference, here's the implementation of [^f] again:

    X& f(Y& y, Z* z)
    {
        y.z = z;
        return y.x;
    }

Here's what's happening:

# [^f] is called passing in a reference to [^y] and a pointer to [^z]
# A pointer to [^z] is held by [^y]
# A reference to [^y.x] is returned
# [^z] is deleted. [^y.z] is a dangling pointer
# [^y.z_value()] is called
# [^z->value()] is called
# [*BOOM!]

[h2 Call Policies]

Call Policies may be used in situations such as the example detailed above.
In our example, [^return_internal_reference] and [^with_custodian_and_ward]
are our friends:

    def("f", f,
        return_internal_reference<1,
            with_custodian_and_ward<1, 2> >());

What are the [^1] and [^2] parameters, you ask?

    return_internal_reference<1

Informs Boost.Python that the first argument, in our case [^Y& y], is the
owner of the returned reference: [^X&]. The "[^1]" simply specifies the
first argument. In short: "return an internal reference [^X&] owned by the
1st argument [^Y& y]".

    with_custodian_and_ward<1, 2>

Informs Boost.Python that the lifetime of the argument indicated by ward
(i.e. the 2nd argument: [^Z* z]) is dependent on the lifetime of the
argument indicated by custodian (i.e. the 1st argument: [^Y& y]).

It is also important to note that we have defined two policies above. Two
or more policies can be composed by chaining. Here's the general syntax:

    policy1<args...,
        policy2<args...,
            policy3<args...> > >

Here is the list of predefined call policies. A complete reference detailing
these can be found [@../reference/function_invocation_and_creation/models_of_callpolicies.html here].

* [*with_custodian_and_ward]:  Ties lifetimes of the arguments
* [*with_custodian_and_ward_postcall]:  Ties lifetimes of the arguments and results
* [*return_internal_reference]:  Ties lifetime of one argument to that of result
* [*return_value_policy<T> with T one of:]
    * [*reference_existing_object]: naive (dangerous) approach
    * [*copy_const_reference]: Boost.Python v1 approach
    * [*copy_non_const_reference]:
    * [*manage_new_object]: Adopt a pointer and hold the instance

[blurb :-) [*Remember the Zen, Luke:]

"Explicit is better than implicit"

"In the face of ambiguity, refuse the temptation to guess"
]

[endsect]
[section Overloading]

The following illustrates a scheme for manually wrapping an overloaded
member functions. Of course, the same technique can be applied to wrapping
overloaded non-member functions.

We have here our C++ class:

    struct X
    {
        bool f(int a)
        {
            return true;
        }

        bool f(int a, double b)
        {
            return true;
        }

        bool f(int a, double b, char c)
        {
            return true;
        }

        int f(int a, int b, int c)
        {
            return a + b + c;
        };
    };

Class X has 4 overloaded functions. We will start by introducing some
member function pointer variables:

    bool    (X::*fx1)(int)              = &X::f;
    bool    (X::*fx2)(int, double)      = &X::f;
    bool    (X::*fx3)(int, double, char)= &X::f;
    int     (X::*fx4)(int, int, int)    = &X::f;

With these in hand, we can proceed to define and wrap this for Python:

    .def("f", fx1)
    .def("f", fx2)
    .def("f", fx3)
    .def("f", fx4)

[endsect]
[section Default Arguments]

Boost.Python wraps (member) function pointers. Unfortunately, C++ function
pointers carry no default argument info. Take a function [^f] with default
arguments:

    int f(int, double = 3.14, char const* = "hello");

But the type of a pointer to the function [^f] has no information
about its default arguments:

    int(*g)(int,double,char const*) = f;    // defaults lost!

When we pass this function pointer to the [^def] function, there is no way
to retrieve the default arguments:

    def("f", f);                            // defaults lost!

Because of this, when wrapping C++ code, we had to resort to manual
wrapping as outlined in the [link tutorial.functions.overloading previous section], or
writing thin wrappers:

    // write "thin wrappers"
    int f1(int x) { return f(x); }
    int f2(int x, double y) { return f(x,y); }

    /*...*/

        // in module init
        def("f", f);  // all arguments
        def("f", f2); // two arguments
        def("f", f1); // one argument

When you want to wrap functions (or member functions) that either:

* have default arguments, or
* are overloaded with a common sequence of initial arguments

[h2 BOOST_PYTHON_FUNCTION_OVERLOADS]

Boost.Python now has a way to make it easier. For instance, given a function:

    int foo(int a, char b = 1, unsigned c = 2, double d = 3)
    {
        /*...*/
    }

The macro invocation:

    BOOST_PYTHON_FUNCTION_OVERLOADS(foo_overloads, foo, 1, 4)

will automatically create the thin wrappers for us. This macro will create
a class [^foo_overloads] that can be passed on to [^def(...)]. The third
and fourth macro argument are the minimum arguments and maximum arguments,
respectively. In our [^foo] function the minimum number of arguments is 1
and the maximum number of arguments is 4. The [^def(...)] function will
automatically add all the foo variants for us:

    def("foo", foo, foo_overloads());

[h2 BOOST_PYTHON_MEMBER_FUNCTION_OVERLOADS]

Objects here, objects there, objects here there everywhere. More frequently
than anything else, we need to expose member functions of our classes to
Python. Then again, we have the same inconveniences as before when default
arguments or overloads with a common sequence of initial arguments come
into play. Another macro is provided to make this a breeze.

Like [^BOOST_PYTHON_FUNCTION_OVERLOADS],
[^BOOST_PYTHON_MEMBER_FUNCTION_OVERLOADS] may be used to automatically create
the thin wrappers for wrapping member functions. Let's have an example:

    struct george
    {
        void
        wack_em(int a, int b = 0, char c = 'x')
        {
            /*...*/
        }
    };

The macro invocation:

    BOOST_PYTHON_MEMBER_FUNCTION_OVERLOADS(george_overloads, wack_em, 1, 3)

will generate a set of thin wrappers for george's [^wack_em] member function
accepting a minimum of 1 and a maximum of 3 arguments (i.e. the third and
fourth macro argument). The thin wrappers are all enclosed in a class named
[^george_overloads] that can then be used as an argument to [^def(...)]:

    .def("wack_em", &george::wack_em, george_overloads());

See the [@../reference/function_invocation_and_creation/boost_python_overloads_hpp.html#function_invocation_and_creation.boost_python_overloads_hpp.macros overloads reference]
for details.

[h2 init and optional]

A similar facility is provided for class constructors, again, with
default arguments or a sequence of overloads. Remember [^init<...>]? For example,
given a class X with a constructor:

    struct X
    {
        X(int a, char b = 'D', std::string c = "constructor", double d = 0.0);
        /*...*/
    }

You can easily add this constructor to Boost.Python in one shot:

    .def(init<int, optional<char, std::string, double> >())

Notice the use of [^init<...>] and [^optional<...>] to signify the default
(optional arguments).

[endsect]
[section Auto-Overloading]

It was mentioned in passing in the previous section that
[^BOOST_PYTHON_FUNCTION_OVERLOADS] and [^BOOST_PYTHON_MEMBER_FUNCTION_OVERLOADS]
can also be used for overloaded functions and member functions with a
common sequence of initial arguments. Here is an example:

     void foo()
     {
        /*...*/
     }

     void foo(bool a)
     {
        /*...*/
     }

     void foo(bool a, int b)
     {
        /*...*/
     }

     void foo(bool a, int b, char c)
     {
        /*...*/
     }

Like in the previous section, we can generate thin wrappers for these
overloaded functions in one-shot:

    BOOST_PYTHON_FUNCTION_OVERLOADS(foo_overloads, foo, 0, 3)

Then...

    .def("foo", (void(*)(bool, int, char))0, foo_overloads());

Notice though that we have a situation now where we have a minimum of zero
(0) arguments and a maximum of 3 arguments.

[h2 Manual Wrapping]

It is important to emphasize however that [*the overloaded functions must
have a common sequence of initial arguments]. Otherwise, our scheme above
will not work. If this is not the case, we have to wrap our functions
[link tutorial.functions.overloading manually].

Actually, we can mix and match manual wrapping of overloaded functions and
automatic wrapping through [^BOOST_PYTHON_MEMBER_FUNCTION_OVERLOADS] and
its sister, [^BOOST_PYTHON_FUNCTION_OVERLOADS]. Following up on our example
presented in the section [link tutorial.functions.overloading on overloading], since the
first 4 overload functins have a common sequence of initial arguments, we
can use [^BOOST_PYTHON_MEMBER_FUNCTION_OVERLOADS] to automatically wrap the
first three of the [^def]s and manually wrap just the last. Here's
how we'll do this:

    BOOST_PYTHON_MEMBER_FUNCTION_OVERLOADS(xf_overloads, f, 1, 4)

Create a member function pointers as above for both X::f overloads:

    bool    (X::*fx1)(int, double, char)    = &X::f;
    int     (X::*fx2)(int, int, int)        = &X::f;

Then...

    .def("f", fx1, xf_overloads());
    .def("f", fx2)

[endsect]
[endsect] [/ Functions ]

[section:object Object Interface]

Python is dynamically typed, unlike C++ which is statically typed. Python
variables may hold an integer, a float, list, dict, tuple, str, long etc.,
among other things. In the viewpoint of Boost.Python and C++, these
Pythonic variables are just instances of class [^object]. We will see in
this chapter how to deal with Python objects.

As mentioned, one of the goals of Boost.Python is to provide a
bidirectional mapping between C++ and Python while maintaining the Python
feel. Boost.Python C++ [^object]s are as close as possible to Python. This
should minimize the learning curve significantly.

[$../images/python.png]

[section Basic Interface]

Class [^object] wraps [^PyObject*]. All the intricacies of dealing with
[^PyObject]s such as managing reference counting are handled by the
[^object] class. C++ object interoperability is seamless. Boost.Python C++
[^object]s can in fact be explicitly constructed from any C++ object.

To illustrate, this Python code snippet:

[python]

    def f(x, y):
         if (y == 'foo'):
             x[3:7] = 'bar'
         else:
             x.items += y(3, x)
         return x

    def getfunc():
       return f;

Can be rewritten in C++ using Boost.Python facilities this way:

[c++]

    object f(object x, object y) {
         if (y == "foo")
             x.slice(3,7) = "bar";
         else
             x.attr("items") += y(3, x);
         return x;
    }
    object getfunc() {
        return object(f);
    }

Apart from cosmetic differences due to the fact that we are writing the
code in C++, the look and feel should be immediately apparent to the Python
coder.

[endsect]
[section Derived Object types]

Boost.Python comes with a set of derived [^object] types corresponding to
that of Python's:

* list
* dict
* tuple
* str
* long_
* enum

These derived [^object] types act like real Python types. For instance:

    str(1) ==> "1"

Wherever appropriate, a particular derived [^object] has corresponding
Python type's methods. For instance, [^dict] has a [^keys()] method:

    d.keys()

[^make_tuple] is provided for declaring ['tuple literals]. Example:

    make_tuple(123, 'D', "Hello, World", 0.0);

In C++, when Boost.Python [^object]s are used as arguments to functions,
subtype matching is required. For example, when a function [^f], as
declared below, is wrapped, it will only accept instances of Python's
[^str] type and subtypes.

    void f(str name)
    {
        object n2 = name.attr("upper")();   // NAME = name.upper()
        str NAME = name.upper();            // better
        object msg = "%s is bigger than %s" % make_tuple(NAME,name);
    }

In finer detail:

    str NAME = name.upper();

Illustrates that we provide versions of the str type's methods as C++
member functions.

    object msg = "%s is bigger than %s" % make_tuple(NAME,name);

Demonstrates that you can write the C++ equivalent of [^"format" % x,y,z]
in Python, which is useful since there's no easy way to do that in std C++.

[blurb
    __alert__ [*Beware] the common pitfall of forgetting that the constructors
    of most of Python's mutable types make copies, just as in Python.
]

Python:
[python]

    >>> d = dict(x.__dict__)     # copies x.__dict__
    >>> d['whatever'] = 3        # modifies the copy

C++:
[c++]

    dict d(x.attr("__dict__"));  // copies x.__dict__
    d['whatever'] = 3;           // modifies the copy

[h2 class_<T> as objects]

Due to the dynamic nature of Boost.Python objects, any [^class_<T>] may
also be one of these types! The following code snippet wraps the class
(type) object.

We can use this to create wrapped instances. Example:

    object vec345 = (
        class_<Vec2>("Vec2", init<double, double>())
            .def_readonly("length", &Point::length)
            .def_readonly("angle", &Point::angle)
        )(3.0, 4.0);

    assert(vec345.attr("length") == 5.0);

[endsect]
[section Extracting C++ objects]

At some point, we will need to get C++ values out of object instances. This
can be achieved with the [^extract<T>] function. Consider the following:

    double x = o.attr("length"); // compile error

In the code above, we got a compiler error because Boost.Python
[^object] can't be implicitly converted to [^double]s. Instead, what
we wanted to do above can be achieved by writing:

    double l = extract<double>(o.attr("length"));
    Vec2& v = extract<Vec2&>(o);
    assert(l == v.length());

The first line attempts to extract the "length" attribute of the Boost.Python
[^object]. The second line attempts to ['extract] the [^Vec2] object from held
by the Boost.Python [^object].

Take note that we said "attempt to" above. What if the Boost.Python [^object]
does not really hold a [^Vec2] type? This is certainly a possibility considering
the dynamic nature of Python [^object]s. To be on the safe side, if the C++ type
can't be extracted, an appropriate exception is thrown. To avoid an exception,
we need to test for extractibility:

    extract<Vec2&> x(o);
    if (x.check()) {
        Vec2& v = x(); ...

__tip__ The astute reader might have noticed that the [^extract<T>]
facility in fact solves the mutable copying problem:

    dict d = extract<dict>(x.attr("__dict__"));
    d["whatever"] = 3;          // modifies x.__dict__ !


[endsect]
[section Enums]

Boost.Python has a nifty facility to capture and wrap C++ enums. While
Python has no [^enum] type, we'll often want to expose our C++ enums to
Python as an [^int]. Boost.Python's enum facility makes this easy while
taking care of the proper conversions from Python's dynamic typing to C++'s
strong static typing (in C++, ints cannot be implicitly converted to
enums). To illustrate, given a C++ enum:

    enum choice { red, blue };

the construct:

    enum_<choice>("choice")
        .value("red", red)
        .value("blue", blue)
        ;

can be used to expose to Python. The new enum type is created in the
current [^scope()], which is usually the current module. The snippet above
creates a Python class derived from Python's [^int] type which is
associated with the C++ type passed as its first parameter.

[note [*what is a scope?]

The scope is a class that has an associated global Python object which
controls the Python namespace in which new extension classes and wrapped
functions will be defined as attributes. Details can be found
[@../reference/high_level_components/boost_python_scope_hpp.html#high_level_components.boost_python_scope_hpp.class_scope here].]

You can access those values in Python as

[python]

    >>> my_module.choice.red
    my_module.choice.red

where my_module is the module where the enum is declared. You can also
create a new scope around a class:

[c++]

    scope in_X = class_<X>("X")
                    .def( ... )
                    .def( ... )
                ;

    // Expose X::nested as X.nested
    enum_<X::nested>("nested")
        .value("red", red)
        .value("blue", blue)
        ;

[def Py_Initialize          [@http://www.python.org/doc/current/api/initialization.html#l2h-652  Py_Initialize]]
[def Py_Finalize            [@http://www.python.org/doc/current/api/initialization.html#l2h-656  Py_Finalize]]
[def Py_XINCREF             [@http://www.python.org/doc/current/api/countingRefs.html#l2h-65     Py_XINCREF]]
[def Py_XDECREF             [@http://www.python.org/doc/current/api/countingRefs.html#l2h-67     Py_XDECREF]]
[def PyImport_AppendInittab [@http://www.python.org/doc/current/api/importing.html#l2h-137       PyImport_AppendInittab]]
[def PyImport_AddModule     [@http://www.python.org/doc/current/api/importing.html#l2h-125       PyImport_AddModule]]
[def PyModule_New           [@http://www.python.org/doc/current/api/moduleObjects.html#l2h-591   PyModule_New]]
[def PyModule_GetDict       [@http://www.python.org/doc/current/api/moduleObjects.html#l2h-594   PyModule_GetDict]]

[endsect]

[section:creating_python_object Creating `boost::python::object` from `PyObject*`]

When you want a `boost::python::object` to manage a pointer to `PyObject*` pyobj one does:

    boost::python::object o(boost::python::handle<>(pyobj));

In this case, the `o` object, manages the `pyobj`, it won’t increase the reference count on construction.

Otherwise, to use a borrowed reference:

    boost::python::object o(boost::python::handle<>(boost::python::borrowed(pyobj)));

In this case, `Py_INCREF` is called, so `pyobj` is not destructed when object o goes out of scope.

[endsect] [/ creating_python_object ]

[endsect] [/ Object Interface]

[section Embedding]

By now you should know how to use Boost.Python to call your C++ code from
Python. However, sometimes you may need to do the reverse: call Python code
from the C++-side. This requires you to ['embed] the Python interpreter
into your C++ program.

Currently, Boost.Python does not directly support everything you'll need
when embedding. Therefore you'll need to use the
[@http://www.python.org/doc/current/api/api.html Python/C API] to fill in
the gaps. However, Boost.Python already makes embedding a lot easier and,
in a future version, it may become unnecessary to touch the Python/C API at
all. So stay tuned... :-)

[h2 Building embedded programs]

To be able to embed python into your programs, you have to link to
both Boost.Python's as well as Python's own runtime library.

Boost.Python's library comes in two variants. Both are located
in Boost's [^/libs/python/build/bin-stage] subdirectory. On Windows, the
variants are called [^boost_python.lib] (for release builds) and
[^boost_python_debug.lib] (for debugging). If you can't find the libraries,
you probably haven't built Boost.Python yet. See
[@../../../building.html Building and Testing] on how to do this.

Python's library can be found in the [^/libs] subdirectory of
your Python directory. On Windows it is called pythonXY.lib where X.Y is
your major Python version number.

Additionally, Python's [^/include] subdirectory has to be added to your
include path.

In a Jamfile, all the above boils down to:

[pre
projectroot c:\projects\embedded_program ; # location of the program

# bring in the rules for python
SEARCH on python.jam = $(BOOST_BUILD_PATH) ;
include python.jam ;

exe embedded_program # name of the executable
  : #sources
     embedded_program.cpp
  : # requirements
     <find-library>boost_python <library-path>c:\boost\libs\python
  $(PYTHON_PROPERTIES)
    <library-path>$(PYTHON_LIB_PATH)
    <find-library>$(PYTHON_EMBEDDED_LIBRARY) ;
]

[h2 Getting started]

Being able to build is nice, but there is nothing to build yet. Embedding
the Python interpreter into one of your C++ programs requires these 4
steps:

# '''#include''' [^<boost/python.hpp>]

# Call Py_Initialize() to start the interpreter and create the [^__main__] module.

# Call other Python C API routines to use the interpreter.

[/ # Call Py_Finalize() to stop the interpreter and release its resources.]

[note [*Note that at this time you must not call Py_Finalize() to stop the
interpreter. This may be fixed in a future version of boost.python.]
]

(Of course, there can be other C++ code between all of these steps.)

[:['[*Now that we can embed the interpreter in our programs, lets see how to put it to use...]]]

[section Using the interpreter]

As you probably already know, objects in Python are reference-counted.
Naturally, the [^PyObject]s of the Python C API are also reference-counted.
There is a difference however. While the reference-counting is fully
automatic in Python, the Python C API requires you to do it
[@http://www.python.org/doc/current/c-api/refcounting.html by hand]. This is
messy and especially hard to get right in the presence of C++ exceptions.
Fortunately Boost.Python provides the [@../reference/utility_and_infrastructure/boost_python_handle_hpp.html#utility_and_infrastructure.boost_python_handle_hpp.class_template_handle handle] and
[@../reference/object_wrappers/boost_python_object_hpp.html#object_wrappers.boost_python_object_hpp.class_object object] class templates to automate the process.

[h2 Running Python code]

Boost.python provides three related functions to run Python code from C++.

    object eval(str expression, object globals = object(), object locals = object())
    object exec(str code, object globals = object(), object locals = object())
    object exec_file(str filename, object globals = object(), object locals = object())

eval evaluates the given expression and returns the resulting value.
exec executes the given code (typically a set of statements) returning the result,
and exec_file executes the code contained in the given file.

There are also overloads taking `char const*` instead of str as the first argument.

The [^globals] and [^locals] parameters are Python dictionaries
containing the globals and locals of the context in which to run the code.
For most intents and purposes you can use the namespace dictionary of the
[^__main__] module for both parameters.

Boost.python provides a function to import a module:

    object import(str name)

import imports a python module (potentially loading it into the running process
first), and returns it.

Let's import the [^__main__] module and run some Python code in its namespace:

    object main_module = import("__main__");
    object main_namespace = main_module.attr("__dict__");

    object ignored = exec("hello = file('hello.txt', 'w')\n"
                          "hello.write('Hello world!')\n"
                          "hello.close()",
                          main_namespace);

This should create a file called 'hello.txt' in the current directory
containing a phrase that is well-known in programming circles.

[h2 Manipulating Python objects]

Often we'd like to have a class to manipulate Python objects.
But we have already seen such a class above, and in the
[link tutorial.object previous section]: the aptly named [^object] class
and its derivatives. We've already seen that they can be constructed from
a [^handle]. The following examples should further illustrate this fact:

    object main_module = import("__main__");
    object main_namespace = main_module.attr("__dict__");
    object ignored = exec("result = 5 ** 2", main_namespace);
    int five_squared = extract<int>(main_namespace["result"]);

Here we create a dictionary object for the [^__main__] module's namespace.
Then we assign 5 squared to the result variable and read this variable from
the dictionary. Another way to achieve the same result is to use eval instead,
which returns the result directly:

    object result = eval("5 ** 2");
    int five_squared = extract<int>(result);

[h2 Exception handling]

If an exception occurs in the evaluation of the python expression,
[@../reference/high_level_components/boost_python_errors_hpp.html#high_level_components.boost_python_errors_hpp.class_error_already_set error_already_set] is thrown:

    try
    {
        object result = eval("5/0");
        // execution will never get here:
        int five_divided_by_zero = extract<int>(result);
    }
    catch(error_already_set const &)
    {
        // handle the exception in some way
    }

The [^error_already_set] exception class doesn't carry any information in itself.
To find out more about the Python exception that occurred, you need to use the
[@http://www.python.org/doc/api/exceptionHandling.html exception handling functions]
of the Python C API in your catch-statement. This can be as simple as calling
[@http://www.python.org/doc/api/exceptionHandling.html#l2h-70 PyErr_Print()] to
print the exception's traceback to the console, or comparing the type of the
exception with those of the [@http://www.python.org/doc/api/standardExceptions.html
standard exceptions]:

    catch(error_already_set const &)
    {
        if (PyErr_ExceptionMatches(PyExc_ZeroDivisionError))
        {
            // handle ZeroDivisionError specially
        }
        else
        {
            // print all other errors to stderr
            PyErr_Print();
        }
    }

(To retrieve even more information from the exception you can use some of the other
exception handling functions listed [@http://www.python.org/doc/api/exceptionHandling.html here].)

[endsect]
[endsect] [/ Embedding]

[section Iterators]

In C++, and STL in particular, we see iterators everywhere. Python also has
iterators, but these are two very different beasts.

[*C++ iterators:]

* C++ has 5 type categories (random-access, bidirectional, forward, input, output)
* There are 2 Operation categories: reposition, access
* A pair of iterators is needed to represent a (first/last) range.

[*Python Iterators:]

* 1 category (forward)
* 1 operation category (next())
* Raises StopIteration exception at end

The typical Python iteration protocol: [^[*for y in x...]] is as follows:

[python]

    iter = x.__iter__()         # get iterator
    try:
        while 1:
        y = iter.next()         # get each item
        ...                     # process y
    except StopIteration: pass  # iterator exhausted

Boost.Python provides some mechanisms to make C++ iterators play along
nicely as Python iterators. What we need to do is to produce
appropriate `__iter__` function from C++ iterators that is compatible
with the Python iteration protocol. For example:

[c++]

    object get_iterator = iterator<vector<int> >();
    object iter = get_iterator(v);
    object first = iter.next();

Or for use in class_<>:

    .def("__iter__", iterator<vector<int> >())

[*range]

We can create a Python savvy iterator using the range function:

* range(start, finish)
* range<Policies,Target>(start, finish)

Here, start/finish may be one of:

* member data pointers
* member function pointers
* adaptable function object (use Target parameter)

[*iterator]

* iterator<T, Policies>()

Given a container [^T], iterator is a shortcut that simply calls [^range]
with &T::begin, &T::end.

Let's put this into action... Here's an example from some hypothetical
bogon Particle accelerator code:

[python]

    f = Field()
    for x in f.pions:
        smash(x)
    for y in f.bogons:
        count(y)

Now, our C++ Wrapper:

[c++]

    class_<F>("Field")
        .property("pions", range(&F::p_begin, &F::p_end))
        .property("bogons", range(&F::b_begin, &F::b_end));

[*stl_input_iterator]

So far, we have seen how to expose C++ iterators and ranges to Python.
Sometimes we wish to go the other way, though: we'd like to pass a
Python sequence to an STL algorithm or use it to initialize an STL
container. We need to make a Python iterator look like an STL iterator.
For that, we use `stl_input_iterator<>`. Consider how we might
implement a function that exposes `std::list<int>::assign()` to
Python:

[c++]

    template<typename T>
    void list_assign(std::list<T>& l, object o) {
        // Turn a Python sequence into an STL input range
        stl_input_iterator<T> begin(o), end;
        l.assign(begin, end);
    }

    // Part of the wrapper for list<int>
    class_<std::list<int> >("list_int")
        .def("assign", &list_assign<int>)
        // ...
        ;

Now in Python, we can assign any integer sequence to `list_int` objects:

[python]

    x = list_int();
    x.assign([1,2,3,4,5])

[endsect]
[section:exception Exception Translation]

All C++ exceptions must be caught at the boundary with Python code. This
boundary is the point where C++ meets Python. Boost.Python provides a
default exception handler that translates selected standard exceptions,
then gives up:

    raise RuntimeError, 'unidentifiable C++ Exception'

Users may provide custom translation. Here's an example:

    struct PodBayDoorException;
    void translator(PodBayDoorException const& x) {
        PyErr_SetString(PyExc_UserWarning, "I'm sorry Dave...");
    }
    BOOST_PYTHON_MODULE(kubrick) {
         register_exception_translator<
              PodBayDoorException>(translator);
         ...

[endsect]
[section:techniques General Techniques]

Here are presented some useful techniques that you can use while wrapping code with Boost.Python.

[section Creating Packages]

A Python package is a collection of modules that provide to the user a certain
functionality. If you're not familiar on how to create packages, a good
introduction to them is provided in the
[@http://www.python.org/doc/current/tut/node8.html Python Tutorial].

But we are wrapping C++ code, using Boost.Python. How can we provide a nice
package interface to our users? To better explain some concepts, let's work
with an example.

We have a C++ library that works with sounds: reading and writing various
formats, applying filters to the sound data, etc. It is named (conveniently)
[^sounds].  Our library already has a neat C++ namespace hierarchy, like so:

    sounds::core
    sounds::io
    sounds::filters

We would like to present this same hierarchy to the Python user, allowing him
to write code like this:

    import sounds.filters
    sounds.filters.echo(...) # echo is a C++ function

The first step is to write the wrapping code. We have to export each module
separately with Boost.Python, like this:

    /* file core.cpp */
    BOOST_PYTHON_MODULE(core)
    {
        /* export everything in the sounds::core namespace */
        ...
    }

    /* file io.cpp */
    BOOST_PYTHON_MODULE(io)
    {
        /* export everything in the sounds::io namespace */
        ...
    }

    /* file filters.cpp */
    BOOST_PYTHON_MODULE(filters)
    {
        /* export everything in the sounds::filters namespace */
        ...
    }

Compiling these files will generate the following Python extensions:
[^core.pyd], [^io.pyd] and [^filters.pyd].

[note The extension [^.pyd] is used for python extension modules, which
are just shared libraries.  Using the default for your system, like [^.so] for
Unix and [^.dll] for Windows, works just as well.]

Now, we create this directory structure for our Python package:

[pre
sounds/
    \_\_init\_\_.py
    core.pyd
    filters.pyd
    io.pyd
]

The file [^\_\_init\_\_.py] is what tells Python that the directory [^sounds/] is
actually a Python package. It can be a empty file, but can also perform some
magic, that will be shown later.

Now our package is ready. All the user has to do is put [^sounds] into his
[@http://www.python.org/doc/current/tut/node8.html#SECTION008110000000000000000 PYTHONPATH]
and fire up the interpreter:

[python]

    >>> import sounds.io
    >>> import sounds.filters
    >>> sound = sounds.io.open('file.mp3')
    >>> new_sound = sounds.filters.echo(sound, 1.0)

Nice heh?

This is the simplest way to create hierarchies of packages, but it is not very
flexible. What if we want to add a ['pure] Python function to the filters
package, for instance, one that applies 3 filters in a sound object at once?
Sure, you can do this in C++ and export it, but why not do so in Python? You
don't have to recompile the extension modules, plus it will be easier to write
it.

If we want this flexibility, we will have to complicate our package hierarchy a
little. First, we will have to change the name of the extension modules:

[c++]

    /* file core.cpp */
    BOOST_PYTHON_MODULE(_core)
    {
        ...
        /* export everything in the sounds::core namespace */
    }

Note that we added an underscore to the module name. The filename will have to
be changed to [^_core.pyd] as well, and we do the same to the other extension modules.
Now, we change our package hierarchy like so:

[pre
sounds/
    \_\_init\_\_.py
    core/
        \_\_init\_\_.py
        \_core.pyd
    filters/
        \_\_init\_\_.py
        \_filters.pyd
    io/
        \_\_init\_\_.py
        \_io.pyd
]

Note that we created a directory for each extension module, and added a
\_\_init\_\_.py to each one. But if we leave it that way, the user will have to
access the functions in the core module with this syntax:

[python]

    >>> import sounds.core._core
    >>> sounds.core._core.foo(...)

which is not what we want. But here enters the [^\_\_init\_\_.py] magic: everything
that is brought to the [^\_\_init\_\_.py] namespace can be accessed directly by the
user.  So, all we have to do is bring the entire namespace from [^_core.pyd]
to [^core/\_\_init\_\_.py]. So add this line of code to [^sounds/core/\_\_init\_\_.py]:

    from _core import *

We do the same for the other packages. Now the user accesses the functions and
classes in the extension modules like before:

    >>> import sounds.filters
    >>> sounds.filters.echo(...)

with the additional benefit that we can easily add pure Python functions to
any module, in a way that the user can't tell the difference between a C++
function and a Python function. Let's add a ['pure] Python function,
[^echo_noise], to the [^filters] package. This function applies both the
[^echo] and [^noise] filters in sequence in the given [^sound] object. We
create a file named [^sounds/filters/echo_noise.py] and code our function:

    import _filters
    def echo_noise(sound):
        s = _filters.echo(sound)
        s = _filters.noise(sound)
        return s

Next, we add this line to [^sounds/filters/\_\_init\_\_.py]:

    from echo_noise import echo_noise

And that's it. The user now accesses this function like any other function
from the [^filters] package:

    >>> import sounds.filters
    >>> sounds.filters.echo_noise(...)

[endsect]
[section Extending Wrapped Objects in Python]

Thanks to Python's flexibility, you can easily add new methods to a class,
even after it was already created:

    >>> class C(object): pass
    >>>
    >>> # a regular function
    >>> def C_str(self): return 'A C instance!'
    >>>
    >>> # now we turn it in a member function
    >>> C.__str__ = C_str
    >>>
    >>> c = C()
    >>> print c
    A C instance!
    >>> C_str(c)
    A C instance!

Yes, Python rox. :-)

We can do the same with classes that were wrapped with Boost.Python. Suppose
we have a class [^point] in C++:

[c++]

    class point {...};

    BOOST_PYTHON_MODULE(_geom)
    {
        class_<point>("point")...;
    }

If we are using the technique from the previous session,
[link tutorial.techniques.creating_packages Creating Packages], we can code directly
into [^geom/\_\_init\_\_.py]:

[python]

    from _geom import *

    # a regular function
    def point_str(self):
        return str((self.x, self.y))

    # now we turn it into a member function
    point.__str__ = point_str

[*All] point instances created from C++ will also have this member function!
This technique has several advantages:

* Cut down compile times to zero for these additional functions
* Reduce the memory footprint to virtually zero
* Minimize the need to recompile
* Rapid prototyping (you can move the code to C++ if required without changing the interface)

Another useful idea is to replace constructors with factory functions:

    _point = point

    def point(x=0, y=0):
        return _point(x, y)

In this simple case there is not much gained, but for constructurs with
many overloads and/or arguments this is often a great simplification, again
with virtually zero memory footprint and zero compile-time overhead for
the keyword support.

[endsect]
[section Reducing Compiling Time]

If you have ever exported a lot of classes, you know that it takes quite a good
time to compile the Boost.Python wrappers. Plus the memory consumption can
easily become too high. If this is causing you problems, you can split the
class_ definitions in multiple files:

[c++]

    /* file point.cpp */
    #include <point.h>
    #include <boost/python.hpp>

    void export_point()
    {
        class_<point>("point")...;
    }

    /* file triangle.cpp */
    #include <triangle.h>
    #include <boost/python.hpp>

    void export_triangle()
    {
        class_<triangle>("triangle")...;
    }

Now you create a file [^main.cpp], which contains the [^BOOST_PYTHON_MODULE]
macro, and call the various export functions inside it.

    void export_point();
    void export_triangle();

    BOOST_PYTHON_MODULE(_geom)
    {
        export_point();
        export_triangle();
    }

Compiling and linking together all this files produces the same result as the
usual approach:

    #include <boost/python.hpp>
    #include <point.h>
    #include <triangle.h>

    BOOST_PYTHON_MODULE(_geom)
    {
        class_<point>("point")...;
        class_<triangle>("triangle")...;
    }

but the memory is kept under control.

This method is recommended too if you are developing the C++ library and
exporting it to Python at the same time: changes in a class will only demand
the compilation of a single cpp, instead of the entire wrapper code.

[note This method is useful too if you are getting the error message
['"fatal error C1204:Compiler limit:internal structure overflow"] when compiling
a large source file, as explained in the [@../faq/fatal_error_c1204_compiler_limit.html FAQ].]

[endsect]
[endsect] [/ General Techniques]