INTERFACES & ELEMENTS --------------------- 1) Introduction =============== Interfaces are descriptions on how to handle an object, without actually implementing the object. This allows for multiple objects to be instantiated based on this interface. Each of them can then be handled equally by an application. Glib, apparently (unchecked), has a way of creating interfaces, probably by means of a class struct without actually defining the object. The object, then, does not define a class and these two add up. Benjamin knows more about interfaces, I didn't study interfaces & glib too deeply, yet. I know them just from Java. Interfaces are cool! It allows for some sort of random element creation without needing to link to the implementation. This is similar to how GStreamer currently handles media plugins. GStreamer itself could be seen as an interface too, in that respect. 2) So why do we need interfaces? ================================ Because GStreamer doesn't handle it all. GStreamer in itself is a media framework for streams of data from one element to the next. There's lots of things that's media-related, but not handled in this description. Several examples will probably clarify this: think of the Xvideo output plugin. We can create an overlay here (Xv-based), and we currently control this X-connection using glib properties. However, what property name is associated with what control? And does it work the same as v4lsrc's overlay image control? The same goes for a mixer, for image control, audio control, and probably a lot more. The general idea is simple: *this needs to be documented*. But properties aren't all - they simply cannot do this all. Some things cannot be described in a simple one-argument property thing. Of course, we could give a pointer to a struct as argument, but that's merely a hack and requires both plugin and app to know the ABI of the struct. This kills the whole idea of making the plugin independent of the app. In short: we want interfaces for this. 3) How to integrate an interface in GStreamer ============================================= Let us start with some starting point: an interface is associated with an element. It is a feature exported by that specific element, not by a pipeline or anything more complex. Pipelines are already handled just fine by GStreamer (or you wouldn't be reading all this). Obviously, a pipeline can be a fallback for an interface. Imagine that we're looking for an audio sink that exposes a mixer, but our fakesink audio output doesn't ("I wonder why"). We could then create a pipeline with the volume element in it to "fake" a mixer. Ideally, the volume element would implement a mixer interface itself. How are we going to do that in programmatic way? We currently use properties. Their huge advantage is that we do not need to care about adding new functions or whatever. Their disadvantage is that they're limited to one argument. Anything more complex requires app/plugin knowledge about the shared data, and that defeats the point of them: to have no dependency on each other. This could be solved partially by using action signals, but that makes the whole picture quite complex (since you use multiple methods for doing one simple thing). Also, they are quite slow compared to functions because of the table lookups. In short: it'd work, but I'm not in facour of it... OK, so an element exposes interfaces. This allows us to think of the idea of embedding interfaces (dynamically, of course) in the GstElement object. Think of an object being able to register an indefinate number of interfaces per object instance, and a client application could then enumerate interfaces and instantiate one. Glib gives us GInterface for this purpose. The disadvantage of this is that it's on a per-class basis, not a per-instance basis. This is a problem in case of elements where it depends on several properties whether it supports an interface or not. This can be solved by simply making one generic virtual function "supported ()" in a generic derived object of GInterface (GstInterface?). GstInterface is then a generic thing that is inherited by specific interfaces (see examples). Obviously, the client will need to know about the ABI/API of this struct, but that'll happen either way. Surely, there needs to binary linkage, but I don't consider that a bad thing. It does improve performance compared to action signals! So an element contains interfaces. But where are these interfaces described? And who creates them? I suggest that we do that just as we handle gstvideo and gstaudio right now (these libs do *nothing* useful currently, so this'd make them a lot more interesting). These interfaces inherit from GstInterface. The functions that are needed, can be provided through a class object. The element is then responsible for storing variables and so on. gstvideo/audio provides wrapper functions for the class functions. That's also how glib suggest us to use GInterfaces. Plugin and application then handle and retrieve interfaces as documented in the glib documentation, which is available at: http://www.gnome.org/~mathieu/gobject/main.html So the most important part left is to document the interfaces and make sure all elements exporting them work equally. For this, I'll give two examples. 4) Examples =========== typedef struct _GstInterface { /* note: this struct indeed has no parent. That's normal. */ /* since we will mostly need a pointer to the element, * we put it here! */ GstElement *element; } GstInterface; /* This small extra virtual function is here to provide an * interface functionality on a per-instance basis rather * than a per-class basis, which is the case for glib. */ typedef struct _GstInterfaceClass { GTypeInterface parent; /* virtual functions */ gboolean (* supported) (GstInterface *iface); } GstInterfaceClass; There would probably be a convenience function that checks a specific interface's implementation (glib allows for this) and checks for ->supported () to be set and to return TRUE: gboolean gst_element_implements_interface (GstElement *element, GType iface_type) { if (G_TYPE_CHECK_INSTANCE_TYPE (G_OBJECT (element), type)) { GstInterface *iface; GstInterfaceClass *ifclass; iface = G_TYPE_CHECK_INSTANCE_CAST (G_OBJECT (element), type, GstInterface) ifclass = GST_INTERFACE_GET_CLASS (iface); if (ifclass->supported != NULL && ifclass->supported (iface) == TRUE) { return TRUE; } } return FALSE; } We could use this in the GST_IS_... () macros. For example, the macro GST_IS_MIXER () would then look like this: #define GST_IS_MIXER(obj) \ (G_TYPE_CHECK_INSTANCE_TYPE ((obj), GST_TYPE_MIXER) && \ gst_element_implements_interface (GST_ELEMENT (obj), \ GST_TYPE_MIXER)) So the application would just tread it with the known macro, and everything would look extremely simple to the end user. Also some convenience function to set the element: void gst_interface_set_element (GstInterface *iface, GstElement *element) { g_return_if_fail (GST_IS_INTERFACE (iface)); iface->element = element; } 4a) mixer --------- A mixer is a way of controlling volume and input/output channels. This doesn't mean that you control which channel is the subwoofer, all that is supposed to be done automatically. It is really meant as a way of representing system-level volumes and such. It could also be used to turn on/off certain outputs or inputs. As you've noticed, I'm not only talking about output, but also input. Indeed, I want both osssrc *and* osssink to export the same mixer interface! Or at least a very similar one. Volume control works the same for both. You could say that osssrc should enumerate the input channels (such as microphone, line-in). Of course, osssink should not. Or maybe it should, not sure... Maybe, we'd need a parent osselement which implements all mixer channels. And alsa* would surely implement the same interface. /* This is confusing naming... (i.e. FIXME) * A channel is referred to both as the number of simultaneous * sounds the input can handle as well as the in-/output itself */ #define GST_MIXER_CHANNEL_INPUT (1<<0) #define GST_MIXER_CHANNEL_OUTPUT (1<<1) #define GST_MIXER_CHANNEL_MUTE (1<<2) #define GST_MIXER_CHANNEL_RECORD (1<<3) typedef struct _GstMixerChannel { gchar *label; gint current_num_channels, max_num_channels, flags; } GstMixerChannel; typedef struct _GstMixer { GstInterface interface; } GstMixer; typedef struct _GstMixerClass { GstInterfaceClass klass; /* virtual functions */ GList * (* list_channels) (GstMixer *mixer); void (* set_volume) (GstMixer *mixer, GstMixerChannel *channel, gint *volumes); void (* get_volume) (GstMixer *mixer, GstMixerChannel *channel, gint *volumes); void (* set_mute) (GstMixer *mixer, GstMixerChannel *channel, gboolean mute); void (* set_record) (GstMixer *mixer, GstMixerChannel *channel, gboolean record); } GstMixerClass; Name for in the element list: "mixer". Gstaudio provides wrapper functions for each of the class' virtual functions. Possibly also some macros for GST_MIXER_CHANNEL_HAS_FLAG () or _get_channel (). How to register these? Let's say that we use osssrc as an example. typedef struct _GstOssMixer { GstMixer mixer; [.. more? ..] } GstOssMixer; typedef struct _GstOssMixerClass { GstMixerClass klass; } GstOssMixerClass; [..] static void gst_osssrc_init (GstOssSrc *osssrc) { GstOssMixer *mixer = GST_OSS_MIXER (osssrc); [..] gst_interface_set_element (GST_INTERFACE (mixer), GST_ELEMENT (osssrc)); [..] } The rest is done automatically, as described in the already- mentioned glib documentation for GInterface. This includes things like the class_init () function of the GstOssMixerClass, which fills all the virtual functions for the mixer, and the actual function implementations. The mixer, basically, operates as an element on its own. It gets the file descriptor from the interface->element (every oss... is a osscommon, etc.). 4b) overlay ----------- Overlay is used in both in- and output, too. Think of v4lsrc, v4l2src, v4lmjpegsrc, xvideosink - all overlays. But where do we position the overlay window? Control of this can be done at various levels: locational control (over the server, asynchronous) or XID control (but that makes you depend on X and limits the ability to broaden it over to non-X elements such as fbsink). However, simplicity *is* an issue here. Do we really care about overlay? In the end, users will have to link against either FB or X anyway, so we might want to create separate interfaces for both. On the other hand, we want to be general too... This is a decision that we need to make as early as possible in this process. For now, I propose making X- and FB-based interfaces. Let's assume that we take X as a basis. Then, overlay becomes as simple as one function. Possible extendible by providing inputs (like in the mixer) and norms, although that only applies to input-to-analog, not to-digital... Others simply return NULL. #define GST_OVERLAY_CHANNEL_INPUT (1<<0) #define GST_OVERLAY_CHANNEL_OUTPUT (1<<1) #define GST_OVERLAY_CHANNEL_TUNER (1<<2) typedef struct _GstOverlayChannel { gchar *label; gint flags; } GstOverlayChannel; typedef struct _GstOverlayNorm { gchar *label; gfloat framerate; } GstOverlayNorm; typedef struct _GstOverlay { GstInterface interface; } GstOverlay; typedef struct _GstOverlayClass { GstInterfaceClass klass; /* virtual functions */ GList * (* list_channels) (GstOverlay *overlay); void (* set_channel) (GstOverlay *overlay, gint index); gint (* get_channel) (GstOverlay *overlay); GList * (* list_norms) (GstOverlay *overlay); void (* set_norm) (GstOverlay *overlay, gint index); gint (* get_norm) (GstOverlay *overlay); void (* set_xwindowid) (GstOverlay *overlay, XID xid); } GstOverlayClass; That's all! 4c) user input -------------- And yes, user input could be an interface too. Even better, it should definately be. And wasn't this one of our key issues for 0.8.0? No code here. Go implement it, lazy ass! General ways of thinking: input can come from a plugin, or from the application (we don't have modules for joystick input et all). However, plugins handling input (such as dvdsrc) need to be able to handle each. So we get both input-to-application as well as input-from-application APIs. 5) Status of this document ========================== This is a proposal, nothing more. Nothing is implemented. Target release is 0.8.0 or any 0.7.x version. 6) Copyright and blabla ======================= (c) Ronald Bultje, 2003 under the terms of the GNU Free Documentation License. See http://www.gnu.org/ for details. And no, I'm not for hire. ;).