Dependency Injection in Games

Games and OOP

Only a few domains map to a programming paradigm so directly as computer games and object orientation. A Game class matches the concept of a virtual world where several different instances of a GameObject class reside. The simulation usually consists of a loop inside the Game class with three main goals: collect and interpret user or network input; update each GameObject instance based on input and/or a simulation step; draw visible game objects on the output device.

Inheritance vs. Composition in game objects (Deja-vu):

To represent different types of simulated objects, the programmer usually create subclasses of GameObject. There are problems with the inheritance approach since objects from different hierarquies sometimes have common functionality and characteristics. This is a common issue in object oriented software design and replacing inheritance by composition is strongly recomended [check previous posts]. The GameObject class then consists only of an unique identifier, some common attributes all game objects have such as position and orientation and, most important, a collection of components. Each class implementing the GameComponent interface represents a different aspect/behavior of a GameObject such as visuals, physics, AI or health, depending on game being implemented. These components are highly flexible and easier to maintain while also maximize code reuse with many of them being useful in several different game genres.

What is the problem with inter-dependent components?

Composition also has its drawbacks, mostly when there’s some coupling between components. For example, an AI component commonly depends on the existance of a health component to decide actions to take and also on a physics component to apply movements to. These dependencies are normally solved directly by the programmer, as shown in the following Java code, part of an AIComponent class, adapted from C++ example Cris Stoy put in his excelent article in Game Programmin Gems 6:

1 public void update() {

2 GameObject owner = getOwner();

3 HealthComponent health = owner.getComponent(HealthComponent.class);

4 if (health != null) {

5 // take appropriate actions

6 }

7 }

Aparently there’s nothing wrong with this code, but a closer inspection shows that:

a) lines 2-4 have nothing to do with the expected game logic of an AI update, being just boilerplate code;

b) if no instance of HealthComponent is initialized for this particular GameObject, no proper AI action (line 5) will ever be taken, making it hard to debug.

How to circunvent that?

What I think is a good idea is the use of dependency injection [check Martin Fowler's article] to automatically take care of the coupling between game components and safe initialization. The programmer will not need to manually check for dependencies or their nullability. It is easier to concentrate on the game logic to implement by replacing the above code with this:

1 @Inject(nullable=false)
2 private HealthComponent health;
3 public void update() {
4 // take appropriate actions
5 }

The above code is smaller, considerably cleaner and also safer since our framework will stop initialization and show an error log if there is no HealthComponent initialized for the GameObject that owns this AI component. With this approach we believe one can write better code and achieve faster prototyping which is always desired in game production pipelines.

In the next post more details about how GCore uses dependency injection to safely initialize game objects and its components…

Data-driven game development

I’ve been playing with JMonkey for the last 2 years at least and for the current project I’m using a data-driven aproach. For those who don’t know what this strange term is, data-driven means: reusable plug-and-play components you can attach to your game with declarative or script languages.

In my engine I’m using XML to define types, objects and gamestates, but this requires me explanation. GameStates enable you to encapsulate different scene concepts (or scene graphs) like: gameplay scene, inGame menus, main menus, each in its own state object. This aproach makes it easy to switch between them by enabling and disabling these GameState objects.

My idea is to define a Game as a collection of GameStates, each one composed by a collection of GameObject instances. Each of these GameObject instances is by itself a collection of pluggable Components. This GameObject design is based on an article from the book Game Programming Gems 6 – “Game Object Component System” – by Chris Stoy. One can find useful information in this post as well, by Robert Rose.

A GameObject can be anything like a player, with a 3D model, a static scene element or even a positional trigger of some type, it doesn’t matter, all of them are just this: GameObjects with a collection of Components defining their looks and behaviors.

The goal is to develop fine grained reusable Components that can be combined in ways to express the desired game design and behavior. I’ve already developed a lot of these reusable components, some independent and some dependent on other to work:

• VisualComponent – defines a 3D model to be imported, scaled, rotated and tranlated;
• PhysicsComponent – plugs a static physics behavior for a GameObject with a VisualComponent;
• DynamicPhysicsComponent – same as above, but with mass and dynamic physics;
• PositionalAudioComponent – allows the attachment of a positional audio clip in declarative form;
• TemporalTrigger – enables and disables child Components based on a timer;
• TerrainComponent – permits the declaration of heightmap based terrain;
• TerrainPhysics – for attaching physics to the terrain;
• TerrainTexture – for composing and glueing splatts, color and detail textures.

These are just some examples since I’ve provided ~15 different components already, the vast majority being reusable in different game contexts.

Last but not least: You can think the game architecture as a big mediator (design pattern) between gamestates switchs (gamestates defined as collection of game objects in XML). GameObjects and Components implement the Composite design pattern since both can have collection of Components. The XML definitions use the concepts of inheritance. One can define a generic type, with a collection of components and refine their attributes in inherited declarations.

Multiple Components of the same type are allowed inside the same collection, the order of declaration is guaranteed in initialization and I used a LOT of reflection in the parsing and gamestates building fases. Too much blah blah blah? So lets see some action:

Definition of a “game”:

<game name="Sertao 3D" init="caatinga">
 <!-- PATHS FOR RESOURCE LOCATOR -->
 <path dir="/sample/resource" type="texture">
 <path dir="/sample/resource" type="model">

 <!-- TYPE DEFINITIONS -->
 <include file="/sample/types-basic.xml">
 <include file="/sample/types-terrain.xml">
 <include file="/sample/types-player.xml">
 <include file="/sample/types-misc.xml">

 <!-- GAMESTATES -->
 <include file="/sample/caatinga.xml">
</game>

The abouve XML defines a game by including type-definitions and one gamestate. Lets see some samples of them:

Extract from types-player.xml:

<type name=”player”>

<component class=”gcore.component.VisualComponent”>

   <model value="./sample/resource/dwarf.jme">
   <scale x="0.03" y="0.03" z="0.03">
   <rotation y="-90">
</component>
<component class="gcore.component.DynamicPhysicsComponent">
  <mass value="2">
  <position y="2">
</component>
<component class="gcore.component.PlayerCameraComponent">
  <offset y="4" z="8">
</component>
<component class="component.ControllerComponent">
  <speed value="3.5">
</component>
</type>

This shows a player as being a 3D model, with physics behavior, followed by the camera and controlled by an especialized controller. To expose more lets see a n inheritance example with two extracts:

From types-basic.xml:

  <type name="visualPhysics">
 <component class="gcore.component.VisualComponent">
 <component class="gcore.component.PhysicsComponent">
 <triangleaccurate value="true">
 </component>
 </type>

And from types-misc.xml:

  <type name="oldfence" extends="visualPhysics">
 <component class="gcore.component.VisualComponent">
 <model value="./sample/resource/oldfence.jme">
 </component>
 <component class="gcore.component.PositionalAudioComponent">
 <file value="/sample/resource/audio/cricket2.wav">
 <playrange value="25">
 <stoprange value="25">
 <maxvolume value="0.15">
 </component>
 </type>

As you can see, the oldfence type inherits the components and attributes from the visualPhysics one and spectializes some things, including a positional audio clip and using a specific 3d model.

How can we use this in a GameState? An extract from the gamestate definition file caating.xml:

  <object name="fence1" extends="oldfence">
 <component class="gcore.component.VisualComponent">
 <position x="40" z="10">
 </component>
 </object>

As you can see, we just need to overhide the specific position information for the visual model and that’s it. You can download the entire source (zipped eclipse project) for the engine and the above example at this link.

Any comments are welcome.