Data preferences

Ever wondered whether one should approach problem X with data structure Y or Z? This article covers a variety of topics related to these dilemmas.

INFO

This article makes references to "[something]-time" operations. This terminology comes from algorithm analysis' Big O Notation.

Long-story short, it describes the worst-case scenario of runtime length. In laymen's terms:

"As the size of a problem domain increases, the runtime length of the algorithm..."

• Constant-time, O(1): "...does not increase."

• Logarithmic-time, O(log n): "...increases at a slow rate."

• Linear-time, O(n): "...increases at the same rate."

• Etc.

Imagine if one had to process 3 million data points within a single frame. It would be impossible to craft the feature with a linear-time algorithm since the sheer size of the data would increase the runtime far beyond the time allotted. In comparison, using a constant-time algorithm could handle the operation without issue.

By and large, developers want to avoid engaging in linear-time operations as much as possible. But, if one keeps the scale of a linear-time operation small, and if one does not need to perform the operation often, then it may be acceptable. Balancing these requirements and choosing the right algorithm / data structure for the job is part of what makes programmers' skills valuable.

Array vs. Dictionary vs. Object

Godot stores all variables in the scripting API in the Variant class class. Variants can store Variant-compatible data structures such as Array and Dictionary as well as Object.

Godot implements Array as a Vector<Variant>. The engine stores the Array contents in a contiguous section of memory, i.e. they are in a row adjacent to each other.

INFO

For those unfamiliar with C++, a Vector is the name of the array object in traditional C++ libraries. It is a "templated" type, meaning that its records can only contain a particular type (denoted by angled brackets). So, for example, a PackedStringArray would be something like a Vector<String>.

Contiguous memory stores imply the following operation performance:

• Iterate: Fastest. Great for loops.

  • Op: All it does is increment a counter to get to the next record.

• Insert, Erase, Move: Position-dependent. Generally slow.

  • Op: Adding/removing/moving content involves moving the adjacent records over (to make room / fill space).

  • Fast add/remove from the end.

  • Slow add/remove from an arbitrary position.

  • Slowest add/remove from the front.

  • If doing many inserts/removals from the front, then...

  1. invert the array.

  2. do a loop which executes the Array changes at the end.

  3. re-invert the array.

  This makes only 2 copies of the array (still constant time, but slow) versus copying roughly 1/2 of the array, on average, N times (linear time).

• Get, Set: Fastest by position. E.g. can request 0th, 2nd, 10th record, etc. but cannot specify which record you want.

  • Op: 1 addition operation from array start position up to desired index.

• Find: Slowest. Identifies the index/position of a value.

  • Op: Must iterate through array and compare values until one finds a match.

  • Performance is also dependent on whether one needs an exhaustive search.

  • If kept ordered, custom search operations can bring it to logarithmic time (relatively fast). Laymen users won't be comfortable with this though. Done by re-sorting the Array after every edit and writing an ordered-aware search algorithm.

Godot implements Dictionary as an OrderedHashMap<Variant, Variant>. The engine stores a small array (initialized to 2^3 or 8 records) of key-value pairs. When one attempts to access a value, they provide it a key. It then hashes the key, i.e. converts it into a number. The "hash" is used to calculate the index into the array. As an array, the OHM then has a quick lookup within the "table" of keys mapped to values. When the HashMap becomes too full, it increases to the next power of 2 (so, 16 records, then 32, etc.) and rebuilds the structure.

Hashes are to reduce the chance of a key collision. If one occurs, the table must recalculate another index for the value that takes the previous position into account. In all, this results in constant-time access to all records at the expense of memory and some minor operational efficiency.

1. Hashing every key an arbitrary number of times.

   • Hash operations are constant-time, so even if an algorithm must do more than one, as long as the number of hash calculations doesn't become too dependent on the density of the table, things will stay fast. Which leads to...

2. Maintaining an ever-growing size for the table.

   • HashMaps maintain gaps of unused memory interspersed in the table on purpose to reduce hash collisions and maintain the speed of accesses. This is why it constantly increases in size exponentially by powers of 2.

As one might be able to tell, Dictionaries specialize in tasks that Arrays do not. An overview of their operational details is as follows:

• Iterate: Fast.

  • Op: Iterate over the map's internal vector of hashes. Return each key. Afterwards, users then use the key to jump to and return the desired value.

• Insert, Erase, Move: Fastest.

  • Op: Hash the given key. Do 1 addition operation to look up the appropriate value (array start + offset). Move is two of these (one insert, one erase). The map must do some maintenance to preserve its capabilities:

  • update ordered List of records.

  • determine if table density mandates a need to expand table capacity.

  • The Dictionary remembers in what order users inserted its keys. This enables it to execute reliable iterations.

• Get, Set: Fastest. Same as a lookup by key.

  • Op: Same as insert/erase/move.

• Find: Slowest. Identifies the key of a value.

  • Op: Must iterate through records and compare the value until a match is found.

  • Note that Godot does not provide this feature out-of-the-box (because they aren't meant for this task).

Godot implements Objects as stupid, but dynamic containers of data content. Objects query data sources when posed questions. For example, to answer the question, "do you have a property called, 'position'?", it might ask its Script or the ClassDB. One can find more information about what objects are and how they work in the Applying object-oriented principles in Godot article.

The important detail here is the complexity of the Object's task. Every time it performs one of these multi-source queries, it runs through several iteration loops and HashMap lookups. What's more, the queries are linear-time operations dependent on the Object's inheritance hierarchy size. If the class the Object queries (its current class) doesn't find anything, the request defers to the next base class, all the way up until the original Object class. While these are each fast operations in isolation, the fact that it must make so many checks is what makes them slower than both of the alternatives for looking up data.

INFO

When developers mention how slow the scripting API is, it is this chain of queries they refer to. Compared to compiled C++ code where the application knows exactly where to go to find anything, it is inevitable that scripting API operations will take much longer. They must locate the source of any relevant data before they can attempt to access it.

The reason GDScript is slow is because every operation it performs passes through this system.

C# can process some content at higher speeds via more optimized bytecode. But, if the C# script calls into an engine class' content or if the script tries to access something external to it, it will go through this pipeline.

NativeScript C++ goes even further and keeps everything internal by default. Calls into external structures will go through the scripting API. In NativeScript C++, registering methods to expose them to the scripting API is a manual task. It is at this point that external, non-C++ classes will use the API to locate them.

So, assuming one extends from Reference to create a data structure, like an Array or Dictionary, why choose an Object over the other two options?

1. Control: With objects comes the ability to create more sophisticated structures. One can layer abstractions over the data to ensure the external API doesn't change in response to internal data structure changes. What's more, Objects can have signals, allowing for reactive behavior.

2. Clarity: Objects are a reliable data source when it comes to the data that scripts and engine classes define for them. Properties may not hold the values one expects, but one doesn't need to worry about whether the property exists in the first place.

3. Convenience: If one already has a similar data structure in mind, then extending from an existing class makes the task of building the data structure much easier. In comparison, Arrays and Dictionaries don't fulfill all use cases one might have.

Objects also give users the opportunity to create even more specialized data structures. With it, one can design their own List, Binary Search Tree, Heap, Splay Tree, Graph, Disjoint Set, and any host of other options.

"Why not use Node for tree structures?" one might ask. Well, the Node class contains things that won't be relevant to one's custom data structure. As such, it can be helpful to construct one's own node type when building tree structures.

extends Object
class_name TreeNode

var _parent: TreeNode = null
var _children := []

func _notification(p_what):
    match p_what:
        NOTIFICATION_PREDELETE:
            # Destructor.
            for a_child in _children:
                a_child.free()

From here, one can then create their own structures with specific features, limited only by their imagination.

Enumerations: int vs. string

Most languages offer an enumeration type option. GDScript is no different, but unlike most other languages, it allows one to use either integers or strings for the enum values (the latter only when using the @export_enum annotation in GDScript). The question then arises, "which should one use?"

The short answer is, "whichever you are more comfortable with." This is a feature specific to GDScript and not Godot scripting in general; The languages prioritizes usability over performance.

On a technical level, integer comparisons (constant-time) will happen faster than string comparisons (linear-time). If one wants to keep up other languages' conventions though, then one should use integers.

The primary issue with using integers comes up when one wants to print an enum value. As integers, attempting to print MY_ENUM will print 5 or what-have-you, rather than something like "MyEnum". To print an integer enum, one would have to write a Dictionary that maps the corresponding string value for each enum.

If the primary purpose of using an enum is for printing values and one wishes to group them together as related concepts, then it makes sense to use them as strings. That way, a separate data structure to execute on the printing is unnecessary.

AnimatedTexture vs. AnimatedSprite2D vs. AnimationPlayer vs. AnimationTree

Under what circumstances should one use each of Godot's animation classes? The answer may not be immediately clear to new Godot users.

AnimatedTexture is a texture that the engine draws as an animated loop rather than a static image. Users can manipulate...

1. the rate at which it moves across each section of the texture (FPS).

2. the number of regions contained within the texture (frames).

Godot's RenderingServer then draws the regions in sequence at the prescribed rate. The good news is that this involves no extra logic on the part of the engine. The bad news is that users have very little control.

Also note that AnimatedTexture is a Resource unlike the other Node objects discussed here. One might create a Sprite2D node that uses AnimatedTexture as its texture. Or (something the others can't do) one could add AnimatedTextures as tiles in a TileSet and integrate it with a TileMapLayer for many auto-animating backgrounds that all render in a single batched draw call.

The AnimatedSprite2D node, in combination with the SpriteFrames resource, allows one to create a variety of animation sequences through spritesheets, flip between animations, and control their speed, regional offset, and orientation. This makes them well-suited to controlling 2D frame-based animations.

If one needs trigger other effects in relation to animation changes (for example, create particle effects, call functions, or manipulate other peripheral elements besides the frame-based animation), then will need to use an AnimationPlayer node in conjunction with the AnimatedSprite2D.

AnimationPlayers are also the tool one will need to use if they wish to design more complex 2D animation systems, such as...

1. Cut-out animations: editing sprites' transforms at runtime.

2. 2D Mesh animations: defining a region for the sprite's texture and rigging a skeleton to it. Then one animates the bones which stretch and bend the texture in proportion to the bones' relationships to each other.

3. A mix of the above.

While one needs an AnimationPlayer to design each of the individual animation sequences for a game, it can also be useful to combine animations for blending, i.e. enabling smooth transitions between these animations. There may also be a hierarchical structure between animations that one plans out for their object. These are the cases where the AnimationTree shines. One can find an in-depth guide on using the AnimationTree Using AnimationTree.