- Preliminaries
- Running SWIG
- A tour of basic C/C++ wrapping
- Modules
- Functions
- Global variables
- Constants and enums
- Pointers
- Structures and C++ classes
- C++ inheritance
- C++ overloaded functions
- C++ operators
- Class extension with %extend
- C++ templates
- C++ Smart Pointers
- Directors (calling Octave from C++ code)
- Threads
- Memory management
- STL support
- Matrix typemaps

Octave is a high-level language intended for numerical programming that is mostly compatible with MATLAB. More information can be found at www.octave.org.

This chapter is intended to give an introduction to using the module. You should also read the SWIG documentation that is not specific to Octave. Also, there are a dozen or so examples in the Examples/octave directory, and hundreds in the test suite (Examples/test-suite and Examples/test-suite/octave).

The current SWIG implemention is based on Octave 2.9.12. Support for other versions (in particular the recent 3.0) has not been tested, nor has support for any OS other than Linux.

Let's start with a very simple SWIG interface file:

%module example %{ #include "example.h" %} int gcd(int x, int y); extern double Foo;

To build an Octave module, run SWIG using the `-octave` option. The `-c++` option is required (for now) as Octave itself is written in C++ and thus the wrapper code must also be.

$ swig -octave -c++ example.i

This creates a C/C++ source file `example_wrap.cxx`. The generated C++ source file contains the low-level wrappers that need to be compiled and linked with the rest of your C/C++ application (in this case, the gcd implementation) to create an extension module.

The swig command line has a number of options you can use, like to redirect it's output. Use `swig --help` to learn about these.

Octave modules are DLLs/shared objects having the ".oct" suffix. Building an oct file is usually done with the mkoctfile command (either within Octave itself, or from the shell). For example,

$ swig -octave -c++ example.i -o example_wrap.cxx $ mkoctfile example_wrap.cxx example.c

where example.c is the file containing the gcd() implementation.

mkoctfile can also be used to extract the build parameters required to invoke the compiler and linker yourself. See the Octave manual and mkoctfile man page.

mkoctfile will produce example.oct, which contains the compiled extension module. Loading it into Octave is then a matter of invoking

octave:1> example

Assuming all goes well, you will be able to do this:

$ octave -q octave:1> example octave:2> example.gcd(4,6) ans = 2 octave:3> example.cvar.Foo ans = 3 octave:4> example.cvar.Foo=4; octave:5> example.cvar.Foo ans = 4

The SWIG module directive specifies the name of the Octave module. If you specify `module example', then in Octave everything in the module will be accessible under "example", as in the above example. When choosing a module name, make sure you don't use the same name as a built-in Octave command or standard module name.

When Octave is asked to invoke `example`, it will try to find the .m or .oct file that defines the function "example". It will thusly find example.oct, that upon loading will register all of the module's symbols.

Giving this function a parameter "global" will cause it to load all symbols into the global namespace in addition to the `example` namespace. For example:

$ octave -q octave:1> example("global") octave:2> gcd(4,6) ans = 2 octave:3> cvar.Foo ans = 3 octave:4> cvar.Foo=4; octave:5> cvar.Foo ans = 4

It is also possible to rename the module namespace with an assignment, as in:

octave:1> example; octave:2> c=example; octave:3> c.gcd(10,4) ans = 2

All global variables are put into the cvar namespace object. This is accessible either as `my_module.cvar`, or just `cvar` (if the module is imported into the global namespace).

One can also rename it by simple assignment, e.g.,

octave:1> some_vars = cvar;

Global functions are wrapped as new Octave built-in functions. For example,

%module example int fact(int n);

creates a built-in function `example.fact(n)` that works exactly like you think it does:

octave:1> example.fact(4) 24

Global variables are a little special in Octave. Given a global variable:

%module example extern double Foo;

To expose variables, SWIG actually generates two functions, to get and set the value. In this case, Foo_set and Foo_set would be generated. SWIG then automatically calls these functions when you get and set the variable-- in the former case creating a local copy in the interpreter of the C variables, and in the latter case copying an interpreter variables onto the C variable.

octave:1> example; octave:2> c=example.cvar.Foo c = 3 octave:3> example.cvar.Foo=4; octave:4> c c = 3 octave:5> example.cvar.Foo ans = 4

If a variable is marked with the %immutable directive then any attempts to set this variable will cause an Octave error. Given a global variable:

%module example %immutable; extern double Foo; %mutable;

SWIG will allow the the reading of `Foo` but when a set attempt is made, an error function will be called.

octave:1> example octave:2> example.Foo=4 error: attempt to set immutable member variable error: assignment failed, or no method for `swig_type = scalar' error: evaluating assignment expression near line 2, column 12

It is possible to add new functions or variables to the module. This also allows the user to rename/remove existing functions and constants (but not linked variables, mutable or immutable). Therefore users are recommended to be careful when doing so.

octave:1> example; octave:2> example.PI=3.142; octave:3> example.PI ans = 3.1420

Because Octave doesn't really have the concept of constants, C/C++ constants are not really constant in Octave. They are actually just a copy of the value into the Octave interpreter. Therefore they can be changed just as any other value. For example given some constants:

%module example %constant int ICONST=42; #define SCONST "Hello World" enum Days{SUNDAY,MONDAY,TUESDAY,WEDNESDAY,THURSDAY,FRIDAY,SATURDAY};

This is 'effectively' converted into the following Octave code:

example.ICONST=42 example.SCONST="Hello World" example.SUNDAY=0 ....

C/C++ pointers are fully supported by SWIG. Furthermore, SWIG has no problem working with incomplete type information. Given a wrapping of the <file.h> interface: C/C++ pointers are fully supported by SWIG. Furthermore, SWIG has no problem working with incomplete type information. Given a wrapping of the <file.h> interface:

%module example FILE *fopen(const char *filename, const char *mode); int fputs(const char *, FILE *); int fclose(FILE *);

When wrapped, you will be able to use the functions in a natural way from Octave. For example:

octave:1> example; octave:2> f=example.fopen("w","junk"); octave:3> example.fputs("Hello world",f); octave:4> example.fclose(f);

Simply printing the value of a wrapped C++ type will print it's typename. E.g.,

octave:1> example; octave:2> f=example.fopen("junk","w"); octave:3> f f = { _p_FILE, ptr = 0x9b0cd00 }

As the user of the pointer, you are responsible for freeing it, or closing any resources associated with it (just as you would in a C program). This does not apply so strictly to classes and structs (see below).

octave:1> example; octave:2> f=example.fopen("not there","r"); error: value on right hand side of assignment is undefined error: evaluating assignment expression near line 2, column 2

SWIG wraps C structures and C++ classes by using a special Octave type called a `swig_ref`. A `swig_ref` contains a reference to one or more instances of C/C++ objects, or just the type information for an object.
For each wrapped structure and class, a `swig_ref` will be exposed that has the name of the type. When invoked as a function, it creates a new object of its type and returns a `swig_ref` that points to that instance. This provides a very natural interface. For example,

struct Point{ int x,y; };

is used as follows:

octave:1> example; octave:2> p=example.Point(); octave:3> p.x=3; octave:4> p.y=5; octave:5> p.x, p.y ans = 3 ans = 5

In C++, invoking the type object in this way calls the object's constructor.
`swig_ref` objects can also be acquired by having a wrapped function return a pointer, reference, or value of a non-primitive type.

The swig_ref type handles indexing operations such that usage maps closely to what you would have in C/C++. Structure members are accessed as in the above example, by calling set and get methods for C++ variables. Methods also work as expected. For example, code wrapped in the following way

class Point{ public: int x,y; Point(int _x,int _y) : x(_x),y(_y) {} double distance(const Point& rhs) { return sqrt(pow(x-rhs.x,2)+pow(y-rhs.y,2)); } void set(int _x,int _y) { x=_x; y=_y; } };

can be used from Octave like this

octave:1> example; octave:2> p1=example.Point(3,5); octave:3> p2=example.Point(1,2); octave:3> p1.distance(p2) ans = 3.6056

By using the `swig_this()` and `swig_type()` functions, one can discover the pointers to and types of the underlying C/C++ object.

octave:5> swig_this(p1) ans = 162504808 octave:6> swig_type(p1) ans = Point

Note that `swig_ref` is a reference-counted pointer to a C/C++ object/type, and as such has pass-by-reference semantics. For example if one has a allocated a single object but has two `swig_ref`'s pointing to it, modifying the object through either of them will change the single allocated object.
This differs from the usual pass-by-value (copy-on-write) semantics that Octave maintains for built-in types. For example, in the following snippet, modifying `b` does not modify `a`,

octave:7> a=struct('x',4) a = { x = 4 } octave:8> b=a b = { x = 4 } octave:9> b.y=4 b = { x = 4 y = 4 } octave:10> a a = { x = 4 }

However, when dealing with wrapped objects, one gets the behavior

octave:2> a=Point(3,5) a = { Point, ptr = 0x9afbbb0 } octave:3> b=a b = { Point, ptr = 0x9afbbb0 } octave:4> b.set(2,1); octave:5> b.x, b.y ans = 2 ans = 1 octave:6> a.x, a.y ans = 2 ans = 1

Depending on the ownership setting of a `swig_ref`, it may call C++ destructors when its reference count goes to zero. See the section on memory management below for details.

Single and multiple inheritance are fully supported. The `swig_ref` type carries type information along with any C++ object pointer it holds.
This information contains the full class hierarchy. When an indexing operation (such as a method invocation) occurs,
the tree is walked to find a match in the current class as well as any of its bases. The lookup is then cached in the `swig_ref`.

Overloaded functions are supported, and handled as in other modules. That is,
each overload is wrapped separately (under internal names), and a dispatch function is also emitted under the external/visible name.
The dispatch function selects which overload to call (if any) based on the passed arguments.
`typecheck` typemaps are used to analyze each argument, as well as assign precedence. See the chapter on typemaps for details.

C++ operator overloading is supported, in a way similar to other modules.
The `swig_ref` type supports all unary and binary operators between itself and all other types that exist in the system at module load time. When an operator is used (where one of the operands is a `swig_ref`), the runtime routes the call to either a member function of the given object, or to a global function whose named is derived from the types of the operands (either both or just the lhs or rhs).

For example, if `a` and `b` are SWIG variables in Octave, `a+b` becomes `a.__add(b)`. The wrapper is then free to implement __add to do whatever it wants. A wrapper may define the `__add` function manually, %rename some other function to it, or %rename a C++ operator to it.

By default the C++ operators are renamed to their corresponding Octave operators. So without doing any work, the following interface

%inline { struct A { int value; A(int _value) : value(_value) {} A operator+ (const A& x) { return A(value+x.value); } }; }

is usable from Octave like this:

a=A(2), b=A(3), c=a+b assert(c.value==5);

Octave operators are mapped in the following way:

__brace a{args} __brace_asgn a{args} = rhs __paren a(args) __paren_asgn a(args) = rhs __str generates string rep __not !a __uplus +a __uminus -a __transpose a.' __hermitian a' __incr a++ __decr a-- __add a + b __sub a - b __mul a * b __div a / b __pow a ^ b __ldiv a \ b __lshift a << b __rshift a >> b __lt a < b __le a <= b __eq a == b __ge a >= b __gt a > b __ne a != b __el_mul a .* b __el_div a ./ b __el_pow a .^ b __el_ldiv a .\ b __el_and a & b __el_or a | b

On the C++ side, the default mappings are as follows:

%rename(__add) *::operator+; %rename(__add) *::operator+(); %rename(__add) *::operator+() const; %rename(__sub) *::operator-; %rename(__uminus) *::operator-(); %rename(__uminus) *::operator-() const; %rename(__mul) *::operator*; %rename(__div) *::operator/; %rename(__mod) *::operator%; %rename(__lshift) *::operator<<; %rename(__rshift) *::operator>>; %rename(__el_and) *::operator&&; %rename(__el_or) *::operator||; %rename(__xor) *::operator^; %rename(__invert) *::operator~; %rename(__lt) *::operator<; %rename(__le) *::operator<=; %rename(__gt) *::operator>; %rename(__ge) *::operator>=; %rename(__eq) *::operator==; %rename(__ne) *::operator!=; %rename(__not) *::operator!; %rename(__incr) *::operator++; %rename(__decr) *::operator--; %rename(__paren) *::operator(); %rename(__brace) *::operator[];

The %extend directive works the same as in other modules.

You can use it to define special behavior, like for example defining Octave operators not mapped to C++ operators, or defining certain Octave mechanisms such as how an object prints. For example, the `octave_value::{is_string,string_value,print}` functions are routed to a special method `__str` that can be defined inside an %extend.

%extend A { string __str() { stringstream sout; sout<<$self->value; return sout.str(); } }

Then in Octave one gets,

octave:1> a=A(4); octave:2> a a = 4 octave:3> printf("%s\n",a); 4 octave:4> a.__str() 4

C++ class and function templates are fully supported as in other modules, in that the %template directive may used to create explicit instantiations of templated types. For example, function templates can be instantiated as follows:

%module example %inline { template<class __scalar> __scalar mul(__scalar a,__scalar b) { return a*b; } } %include <std_complex.i> %template(mul) mul<std::complex<double> > %template(mul) mul<double>

and then used from Octave

octave:3> mul(4,3) ans = 12 octave:4> mul(4.2,3.6) ans = 15.120 octave:2> mul(3+4i,10+2i) ans = 22 + 46i

Similarly, class templates can be instantiated as in the following example,

%module example %include <std_complex.i> %include <std_string.i> %inline { #include <sstream> template<class __scalar> class sum { __scalar s; public: sum(__scalar _s=0) : s(_s) {} sum& add(__scalar _s) { s+=_s; return *this; } std::string __str() const { std::stringstream sout; sout<<s; return sout.str(); } }; } %template(sum_complex) sum<std::complex<double> >; %template(sum_double) sum<double>;

and then used from Octave

octave:2> a=sum_complex(2+3i); octave:3> a.add(2) ans = (4,3) octave:4> a.add(3+i) ans = (7,4)

C++ smart pointers are fully supported as in other modules.

There is full support for SWIG Directors, which permits Octave code to subclass C++ classes, and implement their virtual methods.

Octave has no direct support for object oriented programming, however the `swig_ref` type provides some of this support. You can manufacture a `swig_ref` using the `subclass` function (provided by the SWIG/Octave runtime).

For example,

octave:1> a=subclass(); octave:2> a.my_var = 4; octave:3> a.my_method = @(self) printf("my_var = ",self.my_var); octave:4> a.my_method(); my_var = 4

`subclass()` can also be used to subclass one or more C++ types. Suppose you have an interface defined by

%inline { class A { public: virtual my_method() { printf("c-side routine called\n"); } }; void call_your_method(A& a) { a.my_method(); } }

Then from Octave you can say:

octave:1> B=@() subclass(A(),@my_method); octave:2> function my_method(self) octave:3> printf("octave-side routine called\n"); octave:4> end octave:5> call_your_method(B()); octave-side routine called

or more concisely,

octave:1> B=@() subclass(A(),'my_method',@(self) printf("octave-side routine called\n")); octave:2> call_your_method(B()); octave-side routine called

Note that you have to enable directors via the %feature directive (see other modules for this).

`subclass()` will accept any number of C++ bases or other `subclass()`'ed objects, `(string,octave_value)` pairs, and `function_handles`. In the first case, these are taken as base classes; in the second case, as named members (either variables or functions, depending on whether the given value is a function handle); in the third case, as member functions whose name is taken from the given function handle. E.g.,

octave:1> B=@(some_var=2) subclass(A(),'some_var',some_var,@some_func,'another_func',@(self) do_stuff())

You can also assign non-C++ member variables and functions after construct time. There is no support for non-C++ static members.

There is limited support for explicitly referencing C++ bases. So, in the example above, we could have

octave:1> B=@() subclass(A(),@my_method); octave:2> function my_method(self) octave:3> self.A.my_method(); octave:4> printf("octave-side routine called\n"); octave:5> end octave:6> call_your_method(B()); c-side routine called octave-side routine called

The use of threads in wrapped Director code is not supported; i.e., an Octave-side implementation of a C++ class must be called from the Octave interpreter's thread. Anything fancier (apartment/queue model, whatever) is left to the user. Without anything fancier, this amounts to the limitation that Octave must drive the module... like, for example, an optimization package that calls Octave to evaluate an objective function.

As noted above, `swig_ref` represents a reference counted pointer to a C/C++-side object. It also contains a flag indicating whether Octave or the C/C++ code owns the object. If Octave owns it, any destructors will be called when the reference count reaches zero. If the C/C++ side owns the object, then destructors will not be called when the reference count goes to zero.

For example,

%inline { class A { public: A() { printf("A constructing\n"); } ~A() { printf("A destructing\n"); } }; }

Would produce this behavior in Octave:

octave:1> a=A(); A constructing octave:2> b=a; octave:3> clear a; octave:4> b=4; A destructing

The %newobject directive may be used to control this behavior for pointers returned from functions.

In the case where one wishes for the C++ side to own an object that was created in Octave (especially a Director object), one can use the __disown() method to invert this logic. Then letting the Octave reference count go to zero will not destroy the object, but destroying the object will invalidate the Octave-side object if it still exists (and call destructors of other C++ bases in the case of multiple inheritance/`subclass()`'ing).

This is some skeleton support for various STL containers.

Octave provides a rich set of classes for dealing with matrices. Currently there are no built-in typemaps to deal with those. However, these are relatively straight forward for users to add themselves (see the docs on typemaps). Without much work (a single typemap decl-- say, 5 lines of code in the interface file), it would be possible to have a function

double my_det(const double* mat,int m,int n);

that is accessed from Octave as,

octave:1> my_det(rand(4)); ans = -0.18388