Migrating from old APIs

Migrating from old APIs

Migrating from the deprecated API that does not use viewports

First, some context. Until librsvg version 2.44, the only way to render an RsvgHandle into a Cairo context was with the functions rsvg_handle_render_cairo(handle, cairo_t) and rsvg_handle_render_cairo_sub(handle, cairo_t, id) — respectively, to render the whole document, and to render a single "layer" from it. Both functions assumed that the SVG document was to be rendered at its "natural size", or to the size overriden with rsvg_handle_set_size_callback(). Since the Cairo context can already have an affine transform applied to it, that transform can further change the size of the rendered image.

Librsvg 2.46 introduced the following functions, designed to replace the render_cairo ones:

All of those functions take a viewport argument. Let's see what this means. But first, some history.

Historical note: before librsvg supported viewports

When librsvg was first written, its API basically consisted of only functions to load an RsvgHandle, plus rsvg_handle_get_pixbuf() to render it directly to a GdkPixbuf image. Internally the library used libart (a pre-Cairo 2D rendering library), but did not expose it in the public API.

The only way to specify a size at which to render an RsvgHandle was with rsvg_handle_set_size_callback(), and the callback would run at an unspecified time during loading: when just enough of the SVG document had been loaded to read in the width/height attributes of the toplevel <svg> element, the callback would let the application override these values with its own desired size.

Some years later, Cairo was introduced, and it started to replace libart. Unlike libart, which could only render to in-memory RGBA buffers, Cairo had a notion of "backends": it could render to RGBA buffers, or it could translate its drawing model commands into PDF or PostScript. In Cairo's terms, one creates a cairo_surface_t of a particular kind (in-memory image surface, PDF surface, EPS surface, etc.), and then a cairo_t context for the surface. The context is what makes the drawing commands available.

Being able to render SVG documents directly to PDF or PostScript was clearly attractive, so librsvg's API of rsvg_handle_get_pixbuf() would clearly not be enough. It would be better to pass a cairo_t for an already-created surface, and have librsvg issue its drawing commands to it. Then the application would be in control of the surface type, or in the case of GTK widgets, they would already get a cairo_t passed to their drawing functions. Librsvg got modified to export a rsvg_handle_render_cairo(handle, cairo_t), and then it reimplemented the old rsvg_handle_get_pixbuf() in terms of Cairo.

At this point, librsvg still kept the notion of rendering SVG documents at their "natural size": the <svg> element's width and height attributes converted to pixels (e.g. converting from width="5cm" by using the dots-per-inch value from the RsvgHandle), or if those attributes don't exist, by using the viewBox as a pixel size. The assumption was that if you needed a different size, you could always start by setting the transformation matrix on your cairo_t and then rendering to that.

The problem with not having viewports

Most applications which use librsvg to render SVG assets for their user interface generally work in the same way. For example, to take an SVG icon and render it, they do something like this:

  1. Create an RsvgHandle by loading it from the SVG icon data.

  2. Ask the RsvgHandle for its dimensions.

  3. Divide the dimensions by the GUI's preferred size for icons.

  4. Translate a Cairo context so the icon will appear at the desired location. Scale the Cairo context by the result of the previous step to obtain the desired dimensions.

  5. Render the RsvgHandle in that Cairo context.

This is… too much work. The web world has moved on to using the CSS box model practically everywhere. To embed an image you specify where and at what size you want to place it, and it gets done automatically. You actually have to do extra work if you want to do non-standard things like scale an image non-proportionally.

The new rendering API that uses viewports

Starting with librsvg 2.46, the following functions are available:

typedef struct {
    double x;
    double y;
    double width;
    double height;
} RsvgRectangle;

gboolean rsvg_handle_render_document (RsvgHandle           *handle,
                                      cairo_t              *cr,
                                      const RsvgRectangle  *viewport,
                                      GError              **error);

gboolean rsvg_handle_render_layer    (RsvgHandle           *handle,
                                      cairo_t              *cr,
                                      const char           *id,
                                      const RsvgRectangle  *viewport,
                                      GError              **error);

gboolean rsvg_handle_render_element  (RsvgHandle           *handle,
                                      cairo_t              *cr,
                                      const char           *id,
                                      const RsvgRectangle  *element_viewport,
                                      GError              **error);

For brevity we will omit the rsvg_handle namespace prefix, and just talk about the actual function names. You can see that render_document is basically the same as render_cairo, but it has an extra viewport argument. The same occurs in render_layer versus render_cairo_sub.

In both of those cases — render_document and render_layer —, the viewport argument specifies a rectangle into which the SVG will be positioned and scaled to fit. Consider something like this:

RsvgRectangle viewport = {
    .x = 10.0,
    .y = 20.0,
    .width = 640.0,
    .height = 480.0,

rsvg_handle_render_document (handle, cr, &viewport, NULL);

This is equivalent to first figuring out the scaling factor to make the SVG fit proportionally in 640×480 pixels, then translating the cr by (10, 20) pixels, and then calling rsvg_handle_render_cairo(). If the SVG has different proportions than the width and height of the rectangle, it will be rendered and centered to fit the rectangle.

Note: rsvg_handle_render_element() is new in librsvg 2.46. It extracts a single element from the SVG and renders it scaled to the viewport you specify. It is different from render_layer (or the old-style render_cairo_sub) in that those ones act as if they rendered the whole document's area, but they only paint the layer you specify.

Even better: the old functions to get an SVG's natural dimensions, like rsvg_handle_get_dimensions(), returned integers instead of floating-point numbers, so you could not always get an exact fit. Please use the new get_geometry functions that take viewports; they will give you easier and better results:

New API for obtaining an SVG's dimensions

Per the previous section, you should seldom need to obtain the "natural size" of an SVG document now that you can render it directly into a viewport. But if you still need to know what the SVG document specifies for its own size, you can use the following functions, depending on the level of detail you require:

gboolean rsvg_handle_get_intrinsic_size_in_pixels (RsvgHandle *handle,
                                                   gdouble    *out_width,
                                                   gdouble    *out_height);

rsvg_handle_get_intrinsic_size_in_pixels() returns an exact width and height in floating-point pixels. You should round up to the next integer if you need to allocate a pixel buffer big enough, to avoid clipping the last column or row of pixels, which may be only partially covered. For example, if a document’s width is 41.3 CSS pixels, you should create a raster image 42 pixels wide so it fits without clipping the last pixel. You can do this with the ceil() function.

rsvg_handle_get_intrinsic_size_in_pixels() works by resolving the width/height attributes of the toplevel <svg> element against the handle's current DPI and the font-size that is defined for the <svg> element.

However, that is only possible if the width/height attributes actually exist and are in physical units. The function will return FALSE if the SVG has no resolvable units, for example if the width/height attributes are specified in percentages (e.g. width="50%"), since the function has no knowledge of the viewport where you will place the SVG, or if those attributes are not specified.

The other way of obtaining an SVG's dimensions is to actually query its "intrinsic dimensions", i.e. what is actually specified in the SVG document:

typedef enum {
} RsvgUnit;

typedef struct {
    double   length;
    RsvgUnit unit;
} RsvgLength;

void rsvg_handle_get_intrinsic_dimensions (RsvgHandle *handle,
                                           gboolean   *out_has_width,
                                           RsvgLength *out_width,
                                           gboolean   *out_has_height,
                                           RsvgLength *out_height,
                                           gboolean   *out_has_viewbox,
                                           RsvgRectangle *out_viewbox);

rsvg_handle_get_intrinsic_dimensions() will tell you precisely if the toplevel <svg> has width/height attributes and their values, and also whether it has a viewBox and its value.

Note: Remember that SVGs are scalable. They are not like raster images which have an exact size in pixels, and which you must always take into account to scale them to a convenient size. For SVGs, you can just render them to a viewport, and avoid working directly with their size — which is kind of arbitrary, and all that matters is the document’s aspect ratio.

SVGs with no intrinsic dimensions nor aspect ratio

SVG documents that have none of the width, height, or viewBox attributes are thankfully not very common, but they are hard to deal with: the software cannot immediately know their natural size or aspect ratio, so they cannot be easily scaled to fit within a viewport. If you need to measure the extents of all the objects in an SVG document, you can use rsvg_handle_get_geometry_for_element() by passing NULL for the target element’s id; this will measure all the elements in the document.

Migrating to the geometry APIs

Until librsvg 2.44, the available APIs to query the geometry of a layer or element were these:

struct _RsvgPositionData {
    int x;
    int y;

gboolean rsvg_handle_get_position_sub (RsvgHandle       *handle,
                                       RsvgPositionData *position_data,
                                       const char       *id);

struct _RsvgDimensionData {
    int width;
    int height;
    gdouble em;
    gdouble ex;

gboolean rsvg_handle_get_dimensions_sub (RsvgHandle        *handle,
                                         RsvgDimensionData *dimension_data,
                                         const char        *id);

These functions are inconvenient — separate calls to get the position and dimensions —, and also inexact, since they only return integer values, while SVG uses floating-point units.

Since librsvg 2.46, you can use these functions instead:

typedef struct {
    double x;
    double y;
    double width;
    double height;
} RsvgRectangle;

gboolean rsvg_handle_get_geometry_for_layer (RsvgHandle           *handle,
                                             const char           *id,
                                             const RsvgRectangle  *viewport,
                                             RsvgRectangle        *out_ink_rect,
                                             RsvgRectangle        *out_logical_rect,
                                             GError              **error);

gboolean rsvg_handle_get_geometry_for_element (RsvgHandle     *handle,
                                               const char     *id,
                                               RsvgRectangle  *out_ink_rect,
                                               RsvgRectangle  *out_logical_rect,
                                               GError        **error);

These functions return exact floating-point values. They also give you the ink rectangle, or area covered by paint, as well as the logical rectangle, which is the extents of unstroked paths (i.e. just the outlines).