Aura (obsolete)

This document is still good for a high level overview, with contact information, but many technical details are now obsolete; see the main Aura index for more details.

Project Goals

The goal is to produce a new desktop window manager and shell environment with modern capabilities. The UI must offer rich visuals, large-scale animated transitions and effects that can be produced only with the assistance of hardware acceleration.

Other constraints and goals:

• Cross platform code, should be able to build and run on Windows (and maybe eventually other platforms), even if we don't have an initial product need for them.
• Scalable performance characteristics depending on the target hardware capability.
• Provide the foundation of a flexible windowing system and shell for Chrome and ChromeOS on a variety of form factors.

Notable non-goals for the initial launch of this system include:

• Multiple monitor support. (This was added later, early 2013)
• Software rendering mode or any kind of remote desktop capability to the device.
• NPAPI plugin support. This will never be required. Pepper plugins only will be supported.

UI Design

Owner: Nicholas Jitkoff (alcor@) (UX) and Kan Liu (kanliu@) (PM)

Quick Chrome UI Implementation Backgrounder

Chrome UI for Chrome on Windows and Chrome OS is built using the Views UI framework that lives in src/views. The contents of a window is constructed from a hierarchy of views. View subclasses allow for implementation of various controls and components, like buttons and toolbars. Traditionally Chrome has used a mix of hand-rolled controls for aspects of its user interface where a custom look is desired, such as the browser toolbar and tabstrip, and native controls provided by the host platform where a more conventional look is desired, such as in dialog boxes and menus. When run on Windows, the Win32 API provides for native controls in the form of HWNDs, and on ChromeOS, the Gtk toolkit is used to provide native controls.

A view hierarchy is hosted within a Widget. A Widget is a cross-platform type, and relies on a NativeWidget implementation that is specific to each host environment to perform some duties. The NativeWidget implementation is actually the connection to the host environment. For example, on Windows a NativeWidgetWin wraps a HWND (via the WindowImpl class), receives Windows messages for event handling, painting, and other tasks. In the Gtk world, a NativeWidgetGtk wraps a GtkWidget and responds to signals. The NativeWidget is responsible for translating platform-specific notifications into cross platform views::Events and other types that the rest of Views code can respond to in a platform-independent fashion.

The Chrome UI was originally written for Windows, and so despite the relatively platform-neutral nature of the View hierarchy and much of the views code, Win32-isms did creep in. The philosophy on the Chrome team has always been "let not the perfect be the enemy of the good," so pathways to shorter-term success have been emphasized. The Mac and Desktop-Linux ports of Chrome pursued a different strategy for UI, more aggressively using the native toolkits offered on those platforms (Cocoa and Gtk), so at the start of the Chrome OS project there was still some considerable Win32 influence in Views code. Many of those Win32-isms have been augmented by ifdef'ed Gtkisms.

The reliance on platform widget systems has posed a problem though in that it prevents hardware acceleration of elements of the UI and arbitrary transformation of UI controls. The platform native frameworks are also peculiar in a number of ways, sharing constraints that are not relevant to desktop Chrome or Chrome OS. Before long a desire to eradicate our usage of them grew strong enough to begin work on doing so. An effort was spun up spanning several teams to start by removing Gtk usage in the Views frontend code. This has become one of the major sub-projects required for the Aura work described here.

Platform Native Control (aka Gtk/HWND) Elimination

Owner: Emmanuel Saint-Loubert-Bié (saintlou@)

Gtk/HWND use is pervasive. It is used everywhere from the NativeWidget implementations that host the View hierarchy down to individual dialog boxes. Here are examples of work that has been done to eliminate their use:

• Converted the Options UI to WebUI. The Options dialog boxes were massive platform-native constructs that used many Gtk widgets. Replacing the whole thing with a WebUI implementation has meant many fewer controls.
• Converted other dialogs to WebUI. In general, if something shows in a tab or in something resembling what could be a window.open() popup (top level window), it is a candidate for conversion to WebUI. While Options was the largest conglomeration of native controls, there is a long tail of other smaller dialogs that contribute to our reliance on Gtk.
• Written Views-based implementations of some native controls, like Textfield. Some places do not suit conversion to WebUI - e.g. the browser frame window itself. In these cases we have to write new Views-only (often referred to as "pure views" in code) versions of controls like the Textfield.
• Written Aura-based implementations of the RenderWidgetHostView. Traditionally the RenderWidgetHostView has been a HWND or GtkWidget, and is used as a parent for windowed NPAPI plugins. Since we are only supporting Pepper plugins going forward, we did not need a native window to parent NPAPI plugins and a synthetic implementation could be done.

While many of the major areas have been successfully tackled, this area remains a work in progress.

Hardware Accelerated Rendering/Compositor

Owner: Antoine Labour (piman@)

At the onset of this project Chrome was using two compositors - the compositor used by WebKit to hardware accelerate CSS transitions, and a "Browser Compositor" run in the UI thread of the browser process, used to implement Views transformations like whole-screen rotations.

For a number of reasons, it is desirable to unify our compositing efforts here and provide a single compositor. The primary reason is achieving acceptable performance on target hardware. It is necessary to have a single compositor and draw-pass instead of two as we have now. We would also like to unify the layer trees too at some point, although this was deemed less critical.

The Browser Compositor is implemented as implementations of a ui::Compositor interface, such as a GL one and a D3D one. Antoine has been proceeding by writing a new implementation that uses the WebKit-CC compositor. This way the UI can continue to use the ui::Layer API as its render target. As mentioned, we may eventually consolidate the API between UI and WebKit.

The compositor is a distinct component in Chrome code, consuming only gfx types, WebKit (obviously) and other low level components. In the fullness of time the WebKit compositor will be extracted from WebKit further so that we do not need to drag all of WebKit into Aura and Views.

Aura WM and Shell

Aura

Owner: Ben Goodger (beng@) and Scott Violet (sky@)

To allow us to perform large scale window transitions, we need to back Windows by compositor layers so that we can animate them without redrawing. This led to the development of a simple window type that supported an API compatible with (i.e. implementing the other side of the contract expected by) the Views NativeWidget type. We had initially tried to do this with a View-backed NativeWidget implementation (called NativeWidgetViews) called the views-desktop. However we still needed a platform-native widget (NativeWidgetWin/NativeWidgetGtk) to host the hierarchy. A big challenge was that pervasive in Chrome code is the concept of a gfx::NativeView/NativeWindow, which on Chrome OS and Windows was expected to resolve to a GtkWidget or an HWND. This assumption is also baked into NativeWidgetWin/NativeWidgetGtk and thus we were presented with many challenges parenting windows properly, since we could only ever offer the top level (desktop/screen-level window) as a parent to code that expected a NativeView, rather than a more localized (and probably more correct) window, because a views::View couldn't be a NativeView.

This, combined with some lingering issues with large View hierarchies led to the development of the simple aura::Window type. The aura::Window is what we consider a NativeView/NativeWindow (it typedefs thus). In the Views system, we have implemented a new NativeWidget targeting this type (NativeWidgetAura) that returns the bound aura::Window from its GetNativeView() method.

The aura::Window wraps a Compositor Layer. It also has a delegate which responds to events and paints to its layer.

aura::Windows are similar to Views, only simpler, they are a hierarchy that live within an aura::Desktop. The aura::Desktop is bound to an aura::DesktopHost, which is where the real platform-specific code lives. There is a DesktopHost that wraps an HWND and one that wraps an X window. There is no Gtk in this world. You can think of this as us having pushed the platform specific code one layer further away from Views, out to the screen edge (as far as ChromeOS is concerned). All windows within are synthetic. The DesktopHost window receives the low level platform events, cracks them to aura::Events, targets them to aura Windows, which pass them along to their delegates. On the Views side, NativeWidgetAura is a aura::Window delegate, receives the aura::Event (which it considers a platform native event), and constructs relevant views::Event types to propagate into the embedded View hierarchy.

Aura is responsible for the Window hierarchy, event cracking and propagation, and other basic window functionality (like focus, activation, etc).

Note that despite the fact that Aura is used by Views, it does not actually use Views itself. It is at a lower level of the onion. Think of it like a raw Win32 HWND or GtkWidget.

The Aura Shell and Chrome Integration

Owner: Zelidrag Hornung (zelidrag@) and David Moore (davemoore@)

A desktop environment is much more than just basic window types. We needed a playground to implement the higher level elements of the window manager, such as constraint-based moving and sizing, shell features such as the persistent launcher at the bottom of the screen, status areas, etc. Rather than build this directly into Chrome, which is huge and takes forever to link, we decided to build this as a separate component. Because it consists of UI components like the launcher and custom window frame Views, it would need to depend not just on Aura but also Views.

The product is a shell library (called aura_shell) that (eventually) we can use in Chrome when built with We also have a test runner, called aura_shell_exe. This instantiates the shell, and launches a few sample/example windows that allow us to build out and test functionality. Within the shell, models for components that would normally be populated with user data (such as apps in the launcher) come from mocked models. When instantiated in Chrome, the real data is provided.

The Chrome OS UI team has traditionally worked on many of these features and people from that team will contribute heavily to this effort.

Implementation Strategy

Since this is a complex project, there are several sub-efforts. The breakdown above covers the main areas: Compositor, Gtk-removal, Aura and the Aura Shell/Chrome Integration.

There is much work to be done, so we're pursuing a lot of it in parallel. While the two-compositor system in place at the start of the project isn't something we can put into production, it has let us start building out the Windowing system while the single compositor work proceeds. Likewise, getting a basic shell up and running with embedded Views widgets allows shell components like the launcher to be started while other elements of the window system are being designed and constructed. Similarly, Web-UI based components like the App List can be built in Chrome behind a flag independent of any of the rest of this work.

Since we're offering a new (native) widget system, our approach to implementing this new UI has been to consider it a new target platform for Chrome, and our work can be considered another "port".

You can build the code by setting use_aura=1 in your GYP_DEFINES. This should work from Linux or Windows. This switch should define everything else necessary to make the components above work.

Major Revision History

11/11/2011 - Ben Goodger and James Cook - revisions

10/5/2011 - Ben Goodger - initial revision