Lumiera
The new emerging NLE for GNU/Linux
Note for the time being, this is a loose collection of Conventions and Policies deemed adequate for work on the Lumiera UI-Layer

Architecture

The UI is loaded as Plug-In, which instantiates a GtkLumiera object. All of the UI activity happens within the blocking function GtkLumiera::run(). This UI activity boils down to run the GTK event loop; we use a dedicated thread, and all of the GUI thus performs single-threaded and need not be thread-safe. All UI activities thus need to be short running and with deterministic outcome. It is strictly prohibited to spawn other threads from within the event loop.

We hold up a strict distinction between the UI mechanics and the core editing concerns. The latter must not be implemented within the UI-Layer. Rather, you need to send an act(CommandID) message over the UI-Bus, which causes the corresponding Steam-Layer command script to be dispatched within the Session thread. It is good practice to give an immediate visual clue regarding the fact of sending such a command (e.g. pressing a button). But be prepared that the actual feedback of invoking a command happens asynchronously.

UI-Model

In short: there is no dedicated UI-Model. Stateful working data goes directly into the widgets. However, some widgets or controllers are special, insofar they correspond to and reflect some entities within the Session model. These special entities must be implemented as subclasses of stage::model::Tangible. These are always connected to the UI-Bus, and they are populated and mutated by receiving diff messages pushed up from the Steam-Layer.

There is an inner circle of UI backbone services. These are all mutually dependent on each other and their start-up sequence is intricate. Implement global and cross-cutting concerns here. In case some widget or controller needs access to such a service, then you should prefer using lib::Depend<ServiceInterface>, and install the corresponding service via lib::DependInject<ServiceInterface>::ServiceInstance<ServiceImpl>.

Behaviour patterns

Parent Containers

A Parent Container is an object managing a collection of children, with respect to the structure of tangible UI elements. Irrespective if this container is also a Gtk::Container widget, or “just” a controller. The key point to turn something into a Parent Container is the fact that this entity creates the children in response to a (population) diff message.

Responsibilities
  • a Parent Container has to wire each child properly, so to enable the adequate use of the UI element protocol. This includes

    • to ensure the child is attached to the UI-Bus (usually enforced by ctor call)

    • to install a suitable Expander functor if the child supports expand/collapse

    • to install a suitable Revealer when it is relevant for the child to be brought into sight in response to some message (e.g. to indicate error state).

  • whenever a child is detached from the container, you need to ensure it is destroyed right away, before there is any chance of processing some other UI event.
    [ Invoking the destructor triggers a lot of magic here, especially it causes the child to be properly detached from signals. We want to ensure this happens as soon as the child is taken out of service. Do not use an “can collect garbage later” approach.]

Error handling

Be aware that GTK is written in C. And while GTKmm has some safeguards in place, better be sure no exception can emanate from event handling code.

Warning TODO: probably we’ll need a common wrapper to do so…

Code organisation

GTK is massive and compilation times tend to be problematic in the UI-Layer. Thus you should be diligent with the organisation of includes. Try to confine the more heavyweight includes within the implementation translation units. Use forward declarations and PImpl if possible. However, it is fine to write a mere implementation widget in header-only fashion, in case this improves the locality and readability of code.

Note using namespace Gtk and similar wildcard includes are prohibited.

We are forced to observe a tricky include sequence, due to NoBug’s ERROR macro, and also in order to get I18N right. Thus any actual translation unit should ensure that effectively the first include is that of stage/gtk-base.hpp.
[you need not include it literally. It is perfectly fine if you can be sure the first included header somehow drags in gtk-base.hpp before any other header.]