Extending VIPS

Extending — How to add operations to VIPS

Introduction

This section runs quickly through adding a simple operator to VIPS. For more information, see VipsOperation and VipsRegion. A good starting point for a new operation is a similar one in the VIPS library.

All VIPS operations are subclasses of VipsOperation, which in turn subclasses VipsObject and then GObject. You add an operation to VIPS by defining a new subclass of VipsOperation and arranging for its class_init() to be called, perhaps by calling its get_type() function.

The class and object structures

First you need to define a new object struct and a new class struct.

typedef struct _Negative {
  VipsOperation parent_instance;

  VipsImage *in;
  VipsImage *out;

  int image_max;

} Negative;

typedef struct _NegativeClass {
  VipsOperationClass parent_class;

  /* No new class members needed for this op.
   */

} NegativeClass;

This operation will find the photographic negative of an unsigned 8-bit image, optionally letting you specify the value which the pixels "pivot" about. It doesn't need any class members (ie. values common to all operations of this type), so the second struct is empty. See the source to vips_invert() for a more complete version of this operation that's actually in the library.

GObject has a handy macro to write some of the boilerplate for you.

G_DEFINE_TYPE( Negative, negative, VIPS_TYPE_OPERATION );

G_DEFINE_TYPE() defines a function called negative_get_type(), which registers this new class and returns its GType (a pointer-sized integer). negative_get_type() in turn needs two functions, negative_init(), to initialise a new instance, and negative_class_init(), to initialise a new class.

Class and object initialisation

negative_init() is very simple, it just sets the default value for our optional parameter.

static void
negative_init( Negative *negative )
{
  negative->image_max = 255;
}

negative_class_init() is more complicated: it has to set various fields in various superclasses and define the operation's parameters.

static void
negative_class_init( NegativeClass *class )
{
  GObjectClass *gobject_class = G_OBJECT_CLASS( class );
  VipsObjectClass *object_class = VIPS_OBJECT_CLASS( class );

  gobject_class->set_property = vips_object_set_property;
  gobject_class->get_property = vips_object_get_property;

  object_class->nickname = "negative";
  object_class->description = "photographic negative";
  object_class->build = negative_build;

  VIPS_ARG_IMAGE( class, "in", 1, 
    "Input", 
    "Input image",
    VIPS_ARGUMENT_REQUIRED_INPUT,
    G_STRUCT_OFFSET( Negative, in ) );

  VIPS_ARG_IMAGE( class, "out", 2, 
    "Output", 
    "Output image",
    VIPS_ARGUMENT_REQUIRED_OUTPUT, 
    G_STRUCT_OFFSET( Negative, out ) );

  VIPS_ARG_INT( class, "image_max", 4, 
    "Image maximum", 
    "Maximum value in image: pivot about this",
    VIPS_ARGUMENT_OPTIONAL_INPUT,
    G_STRUCT_OFFSET( Negative, image_max ),
    0, 255, 255 );
}

In GObject, it needs to set the getters and setters for this class. vips has a generic get/set system, so any subclass of VipsObject needs to use the vips ones.

In VipsObject, it needs to set the operation nickname and description, and set a build function (see below). nickname is used to refer to this operation in the API, description is used to explain this operation to users and will be translated into their language.

Finally, it needs to define the arguments the constructor for this class takes. There are a set of handy macros for doing this, see VIPS_ARG_INT() and friends.

The first few parameters are always the same and mean: class pointer for argument, argument name, argument priority (bindings expect required arguments in order of priority), long argument name (this one is internationalised and displayed to users), description (again, users can see this), some flags describing the argument, and finally the position of the member in the struct.

Integer arguments take three more values: the minimum, maximum and default value for the argument.

The build() function

The build function is the thing VipsObject calls during object construction, after all arguments have been supplied and before the object is used. It has two roles: to verify that arguments are correct, and then to construct the object. After build(), the object is expected to be ready for use.

static int
negative_build( VipsObject *object )
{
  VipsObjectClass *class = VIPS_OBJECT_GET_CLASS( object );
  Negative *negative = (Negative *) object;

  if( VIPS_OBJECT_CLASS( negative_parent_class )->build( object ) )
    return( -1 );

  if( vips_check_uncoded( class->nickname, negative->in ) ||
    vips_check_format( class->nickname, negative->in, VIPS_FORMAT_UCHAR ) )
    return( -1 );

  g_object_set( object, "out", vips_image_new(), NULL ); 

  if( vips_image_pipelinev( negative->out, 
    VIPS_DEMAND_STYLE_THINSTRIP, negative->in, NULL ) )
    return( -1 );

  if( vips_image_generate( negative->out, 
    vips_start_one, 
    negative_generate, 
    vips_stop_one, 
    negative->in, negative ) )
    return( -1 );

  return( 0 );
}

negative_build() first chains up to the superclass: this will check that all input arguments have been supplied and are sane.

Next, it adds its own checks. This is a demo operation, so we just work for uncoded, unsigned 8-bit images. There are a lot of convenience functions like vips_check_format(), see the docs.

Next, it creates the output image. This needs to be set with g_object_set() so that vips can see that it has been assigned. vips will also handle the reference counting for you.

vips_image_pipelinev() links our new image onto the input image and notes that this operation prefers to work in lines. You can request other input geometries, see VipsDemandStyle.

The geometry hint is just a hint, an operation needs to be able to supply any size VipsRegion on request. If you must have a certain size request, you can put a cache in the pipeline after your operation, see vips_linecache() and vips_tilecache(). You can also make requests to your operation ordered, see vips_sequential().

Finally, vips_image_generate() attaches a set of callbacks to the output image to generate chunks of it on request. vips_start_one() and vips_stop_one() are convenience functions that make the input region for you, see below.

The generate() function

The generate() function does the actual image processing. negative_generate() (of type VipsGenerateFn, supplied to vips_image_generate() above) is called whenever some pixels of our output image are required.

static int
negative_generate( VipsRegion *or, 
  void *vseq, void *a, void *b, gboolean *stop )
{
  /* The area of the output region we have been asked to make.
   */
  VipsRect *r = &or->valid;

  /* The sequence value ... the thing returned by vips_start_one().
   */
  VipsRegion *ir = (VipsRegion *) vseq;

  VipsImage *in = (VipsImage *) a;
  Negative *negative = (Negative *) b;
  int line_size = r->width * negative->in->Bands; 

  int x, y;

  /* Request matching part of input region.
   */
  if( vips_region_prepare( ir, r ) )
    return( -1 );

  for( y = 0; y < r->height; y++ ) {
    unsigned char *p = (unsigned char *)
      VIPS_REGION_ADDR( ir, r->left, r->top + y ); 
    unsigned char *q = (unsigned char *)
      VIPS_REGION_ADDR( or, r->left, r->top + y ); 

    for( x = 0; x < line_size; x++ ) 
      q[x] = negative->image_max - p[x];
  }

  return( 0 );
}

This has to calculate a section of the output image. The output VipsRegion, or, contains a VipsRect called valid which is the area needing calculation. This call to negative_generate() must somehow make this part of or contain pixel data.

vseq is the sequence value. This is the per-thread state for this generate, created (in this example) by vips_start_one(). In this simple case it's just a VipsRegion defined on the input image. If you need more per-thread state you can write your own start and stop functions and have a struct you create and pass as a sequence value. There are plenty of examples in the VIPS source code, see vips_rank().

a and b are the last two arguments to vips_image_generate() above. stop is a bool pointer you can set to stop computation early. vips_min() on an unsigned int image, for example, will set stop as soon as it sees a zero, and will not scan the entire image.

The first thing negative_generate() does is use vips_region_prepare() to ask for the corresponding pixels from the input image. Operations which do coordinate transforms or which need an area of input for each output point will need to calculate a new rect before calling vips_region_prepare().

Finally, it can calculate some pixels. negative_generate() loops over the valid area of the output and calls VIPS_REGION_ADDR() for each line. This macro is reasonaly quick, but it's best not to call it for each pixel. Once per line is fine though.

Adding to VIPS

To add the operation to vips, just call negative_get_type(). You can include the source in your program, or use GModule to make a binary plugin that will be loaded by libvips at startup. There are some example plugins available.

You can then use negative from any of the vips interfaces. For example, in Python you'd use it like this:

out = in.negative(image_max = 128)

From the command-line it'd look like this:

$ vips negative in.png out.tif --image-max 128

And from C like this:

VipsImage *in;
VipsImage *out;
if( vips_call( "negative", in, &out, "image_max", 128, NULL ) )
  ... error

Unfortunately that will do almost no compile-time type checking, so all vips operations have a tiny extra wrapper to add a bit of safety. For example:

static int 
negative( VipsImage *in, VipsImage **out, ... )
{
  va_list ap;
  int result;

  va_start( ap, out );
  result = vips_call_split( "negative", ap, in, out );
  va_end( ap );

  return( result );
}

And now you can write:

if( negative( in, &out, "image_max", 128, NULL ) )
  ... error

and it's at least a bit safer.

Other types of operation

Change the _build() function to make other types of operation.

Use vips_image_generate() with vips_start_many() to make operations which demand pixels from more than one image at once, such as image plus image.

Use vips_sink() instead of vips_image_generate() to loop over an image and calculate a value. vips uses this for the statistics operations, like vips_avg().

Use vips_image_wio_input() to get an entire image into memory so you can read it with a pointer. This will obviously not scale well to very large images, but some operations, like FFTs or flood-fill, need the whole image to be available at once.

Make area operations, like filters, by enlarging the VipsRect that _generate() is given before calling vips_region_prepare(). You can enlarge the input image, so that the output image is the same size as the original input, by using vips_embed() within the _build() function.

Make things like flips and rotates by making larger changes to the VipsRect in _generate().

Make zero-copy operations, like vips_insert(), with vips_region_region().