|
EnTT 3.14.0
|
EnTT offers a header-only, tiny and easy to use entity-component system module written in modern C++.
The entity-component-system (also known as ECS) is an architectural pattern used mostly in game development.
The library implements a sparse set based model that doesn't require users to specify the set of components neither at compile-time nor at runtime.
This is why users can instantiate the core class simply like:
In place of its more annoying and error-prone counterpart:
Furthermore, it isn't necessary to announce the existence of a component type. When the time comes, just use it and that's all.
The ECS module (as well as the rest of the library) is designed as a set of containers that are used as needed, just like a vector or any other container. It doesn't attempt in any way to take over on the user codebase, nor to control its main loop or process scheduling.
Unlike other more or less well known models, it also makes use of independent pools that are extended via static mixins. The built-in signal support is an example of this flexible design: defined as a mixin, it's easily disabled if not needed. Similarly, the storage class has a specialization that shows how everything is customizable down to the smallest detail.
Everything is designed around the principle that users only have to pay for what they want.
When it comes to using an entity-component system, the tradeoff is usually between performance and memory usage. The faster it is, the more memory it uses. Even worse, some approaches tend to heavily affect other functionalities like the construction and destruction of components to favor iterations, even when it isn't strictly required. In fact, slightly worse performance along non-critical paths are the right price to pay to reduce memory usage and have overall better performance.
EnTT follows a completely different approach. It gets the best out from the basic data structures and gives users the possibility to pay more for higher performance where needed.
As a rule of thumb, a T ** pointer (or whatever a custom pool returns) is always available to directly access all the instances of a given component type T.
This is one of the corner stones of the library. Many of the tools offered are designed around this need and give the possibility to get this information.
The entt::entity type implements the concept of entity identifier. An entity (the E of an ECS) is an opaque element to use as-is. Inspecting it isn't recommended since its format can change in future.
Components (the C of an ECS) are of any type, without any constraints, not even that of being movable. No need to register them nor their types.
Systems (the S of an ECS) are plain functions, functors, lambdas and so on. It's not required to announce them in any case and have no requirements.
The next sections go into detail on how to use the entity-component system part of the EnTT library.
This module is likely larger than what is described below. For more details, please refer to the inline documentation.
A registry stores and manages entities (or identifiers) and components.
The class template basic_registry lets users decide what the preferred type to represent an entity is. Because std::uint32_t is large enough for almost any case, there also exists the enum class entt::entity that wraps it and the alias entt::registry for entt::basic_registry<entt::entity>.
Entities are represented by entity identifiers. An entity identifier contains information about the entity itself and its version.
User defined identifiers are allowed as enum classes and class types that define an entity_type member of type std::uint32_t or std::uint64_t.
A registry is used both to construct and to destroy entities:
The create member function also accepts a hint. Moreover, it has an overload that gets two iterators to use to generate many entities at once efficiently. Similarly, the destroy member function also works with a range of entities:
In addition to offering an overload to force the version upon destruction.
This function removes all components from an entity before releasing it. There also exists a lighter alternative that doesn't query component pools, for use with orphaned entities:
As with the destroy function, also in this case entity ranges are supported and it's possible to force a version.
In both cases, when an identifier is released, the registry can freely reuse it internally. In particular, the version of an entity is increased (unless the overload that forces a version is used instead of the default one).
Users can then test identifiers by means of a registry:
Or inspect them using some functions meant for parsing an identifier as-is, such as:
Components are assigned to or removed from entities at any time.
The emplace member function template creates, initializes and assigns to an entity the given component. It accepts a variable number of arguments to use to construct the component itself:
The default storage detects aggregate types internally and exploits aggregate initialization when possible.
Therefore, it's not strictly necessary to define a constructor for each type.
The insert member function works with ranges and is used to:
Assign the same component to all entities at once when a type is specified as a template parameter or an instance is passed as an argument:
Assign a set of components to the entities when a range is provided (the length of the range of components must be the same of that of entities):
If an entity already has the given component, the replace and patch member function templates are used to update it:
When it's unknown whether an entity already owns an instance of a component, emplace_or_replace is the function to use instead:
This is a slightly faster alternative to the following snippet:
The all_of and any_of member functions may also be useful if in doubt about whether or not an entity has all the components in a set or any of them:
If the goal is to delete a component from an entity that owns it, the erase member function template is the way to go:
When in doubt whether the entity owns the component, use the remove member function instead. It behaves similarly to erase but it drops the component if and only if it exists, otherwise it returns safely to the caller:
The clear member function works similarly and is used to either:
Erases all instances of the given components from the entities that own them:
Or destroy all entities in a registry at once:
Finally, references to components are obtained simply as:
If the existence of the component isn't certain, try_get is the more suitable function instead.
By default, each storage comes with a mixin that adds signal support to it.
This allows for fancy things like dependencies and reactive systems.
The on_construct member function returns a sink (which is an object for connecting and disconnecting listeners) for those interested in notifications when a new instance of a given component type is created:
Similarly, on_destroy and on_update are used to receive notifications about the destruction and update of an instance, respectively.
Because of how C++ works, listeners attached to on_update are only invoked following a call to replace, emplace_or_replace or patch.
Runtime pools are also supported by providing an identifier to the functions above:
Refer to the following sections for more information about runtime pools.
In all cases, the function type of a listener is equivalent to the following:
All listeners are provided with the registry that triggered the notification and the involved entity. Note also that:
There are also some limitations on what a listener can and cannot do:
Please, refer to the documentation of the signal class to know about all the features it offers.
There are many useful but less known functionalities that aren't described here, such as the connection objects or the possibility to attach listeners with a list of parameters that is shorter than that of the signal itself.
Observing entities is also possible. In this case, the user must use the entity type instead of the component type:
Since entity storage is unique within a registry, if a name is provided it's ignored and therefore discarded.
As for the function signature, this is exactly the same as the components.
Entities support all types of signals: construct, destroy and update. The latter is perhaps ambiguous as an entity is not truly updated. Rather, its identifier is created and finally released.
Indeed, the update signal is meant to send general notifications regarding an entity. It can be triggered via the patch function, as is the case with components:
Destroying an entity and then updating the version of an identifier does not give rise to these types of signals under any circumstances instead.
Finally, note that listeners that observe the destruction of an entity are invoked after all components have been removed, not before. This is because the entity would be invalidated before deleting its elements otherwise, making it difficult for the user to write component listeners.
The destruction order of the storage classes and therefore the disconnection of the listeners is completely random.
There are no guarantees today and while a logic is easily discerned, it's not guaranteed that it will remain so in the future.
For example, a listener getting disconnected after a component is discarded as a result of pool destruction is most likely a recipe for problems.
Rather, it's advisable to invoke the clear function of the registry before destroying it. This forces the deletion of all components and entities without ever discarding the pools.
As a result, a listener that wants to access components, entities, or pools can safely do so against a still valid registry, while checking for the existence of the various elements as appropriate.
Signals are the basic tools to construct reactive systems, even if they aren't enough on their own. EnTT tries to take another step in that direction with its reactive mixin.
In order to explain what reactive systems are, this is a slightly revised quote from the documentation of the library that first introduced this tool, Entitas:
Imagine you have 100 fighting units on the battlefield but only 10 of them changed their positions. Instead of using a normal system and updating all 100 entities depending on the position, you can use a reactive system which will only update the 10 changed units. So efficient.
In EnTT, this means iterating over a reduced set of entities and components than what would otherwise be returned from a view or group.
On these words, however, the similarities with the proposal of Entitas also end. The rules of the language and the design of the library obviously impose and allow different things.
A reactive mixin can be used on a standalone storage with any value type (perhaps using an alias to simplify its use):
In this case, it must be provided with a reference registry for subsequent operations.
Alternatively, when using the value type provided by EnTT, it's also possible to create a reactive storage directly inside a registry:
In the latter case there is the advantage that, in the event of destruction of an entity, this storage is also automatically cleaned up.
Also note that, unlike all other storage, these classes don't support signals by default (although they can be enabled if necessary).
Once it has been created and associated with a registry, the reactive mixin needs to be informed about what it should observe.
Here the choice boils down to three main events affecting all elements (entities or components), namely creation, update or destruction:
It goes without saying that it's possible to observe multiple events of the same type or of different types with the same storage.
For example, to know which entities have been assigned or updated a component of a certain type:
Note that all configurations are in or and never in and. Therefore, to track entities that have been assigned two different components, there are a couple of options:
Create two reactive storage, then combine them in a view:
Use a reactive storage with a non-void value type and a custom tracking function for the purpose:
As highlighted in the last example, the reactive mixin tracks the entities that match the given conditions and saves them aside. However, this behavior can be changed.
For example, it's possible to capture all and only the entities for which a certain component has been updated but only if a specific value is within a given range:
This makes reactive storage extremely flexible and usable in a large number of cases.
Finally, once the entities of interest have been collected, it's possible to visit the storage like any other:
Wrapping it in a view and combining it with other views is another option:
In order to simplify this last use case, the reactive mixin also provides a specific function that returns a view of the storage already filtered according to the provided requirements:
The registry used in this case is the one associated with the storage and also available via the registry function.
It should be noted that a reactive storage never deletes its entities (and elements, if any). To process and then discard entities at regular intervals, refer to the clear function available by default for each storage type.
Similarly, the reactive mixin doesn't disconnect itself from observed storages upon destruction. Therefore, users have to do this themselves:
Destroying a reactive storage without disconnecting it from observed pools will result in undefined behavior.
Sorting entities and components is possible using an in-place algorithm that doesn't require memory allocations and is therefore quite convenient.
There are two functions that respond to slightly different needs:
Components are sorted either directly:
Or by accessing their entities:
There exists also the possibility to use a custom sort function object for when the usage pattern is known.
Components are sorted according to the order imposed by another component:
In this case, instances of movement are arranged in memory so that cache misses are minimized when the two components are iterated together.
As a side note, the use of groups limits the possibility of sorting pools of components. Refer to the specific documentation for more details.
The so called helpers are small classes and functions mainly designed to offer built-in support for the most basic functionalities.
The entt::null variable models the concept of a null entity.
The library guarantees that the following expression always returns false:
A registry rejects the null entity in all cases because it isn't considered valid. It also means that the null entity cannot own components.
The type of the null entity is internal and should not be used for any purpose other than defining the null entity itself. However, there exist implicit conversions from the null entity to identifiers of any allowed type:
Similarly, the null entity compares to any other identifier:
As for its integral form, the null entity only affects the entity part of an identifier and is instead completely transparent to its version.
Be aware that entt::null and entity 0 aren't the same thing. Likewise, a zero initialized entity isn't the same as entt::null. Therefore, although entt::entity{} is in some sense an alias for entity 0, none of them are used to create a null entity.
Similar to the null entity, the entt::tombstone variable models the concept of a tombstone.
Once created, the integral form of the two values is the same, although they affect different parts of an identifier. In fact, the tombstone only uses the version part of it and is completely transparent to the entity part.
Also in this case, the following expression always returns false:
Moreover, users cannot set the tombstone version when releasing an entity:
In this case, a different version number is implicitly generated.
The type of a tombstone is internal and can change at any time. However, there exist implicit conversions from a tombstone to identifiers of any allowed type:
Similarly, the tombstone compares to any other identifier:
Be aware that entt::tombstone and entity 0 aren't the same thing. Likewise, a zero initialized entity isn't the same as entt::tombstone. Therefore, although entt::entity{} is in some sense an alias for entity 0, none of them are used to create tombstones.
This function accepts a storage and an instance of a component of the storage type, then it returns the entity associated with the latter:
Where instance is a component of type position. A null entity is returned in case the instance doesn't belong to the registry.
The registry class is designed to create short circuits between its member functions. This greatly simplifies the definition of a dependency.
For example, the following adds (or replaces) the component a_type whenever my_type is assigned to an entity:
Similarly, the code below removes a_type from an entity whenever my_type is assigned to it:
A dependency is easily broken as follows:
There are many other types of dependencies. In general, most of the functions that accept an entity as their first argument are good candidates for this purpose.
The invoke helper allows to propagate a signal to a member function of a component without having to extend it:
All it does is pick up the right component for the received entity and invoke the requested method, passing on the arguments if necessary.
Connecting signals can quickly become cumbersome.
This utility aims to simplify the process by grouping the calls:
Runtime pools are also supported by providing an identifier when calling with, as shown in the previous snippet. Refer to the following sections for more information about runtime pools.
Obviously, this helper doesn't make the code disappear but it should at least reduce the boilerplate in the most complex cases.
A handle is a thin wrapper around an entity and a registry. It replicates the API of a registry by offering functions such as get or emplace. The difference being that the entity is implicitly passed to the registry.
It's default constructible as an invalid handle that contains a null registry and a null entity. When it contains a null registry, calling functions that delegate execution to the registry causes undefined behavior. It's recommended to test for validity with its implicit cast to bool if in doubt.
A handle is also non-owning, meaning that it's freely copied and moved around without affecting its entity (in fact, handles happen to be trivially copyable). An implication of this is that mutability becomes part of the type.
There are two aliases that use entt::entity as their default entity: entt::handle and entt::const_handle.
Users can also easily create their own aliases for custom identifiers as:
Non-const handles are also implicitly convertible to const handles out of the box but not the other way around.
This class is intended to simplify function signatures. In case of functions that take a registry and an entity and do most of their work on that entity, users might want to consider using handles, either const or non-const.
The organizer class template offers support for creating an execution graph from a set of functions and their requirements on resources.
The resulting tasks aren't executed in any case. This isn't the goal of this tool. Instead, they are returned to the user in the form of a graph that allows for safe execution.
All functions are added in order of execution to the organizer:
These are the parameters that a free function or a member function can accept:
entt::basic_view with any possible combination of storage classes.T (that is, a context variable).The function type for free functions and decayed lambdas passed as parameters to emplace is void(const void *, entt::registry &) instead. The first parameter is an optional pointer to user defined data to provide upon registration:
In all cases, it's also possible to associate a name with the task when creating it. For example:
When a function is registered with the organizer, everything it accesses is considered a resource (views are unpacked and their types are treated as resources). The constness of a type also dictates its access mode (RO/RW). In turn, this affects the resulting graph, since it influences the possibility of launching tasks in parallel.
As for the registry, if a function doesn't explicitly request it or requires a constant reference to it, it's considered a read-only access. Otherwise, it's considered as read-write access. All functions have the registry among their resources.
When registering a function, users can also require resources that aren't in the list of parameters of the function itself. These are declared as template parameters:
Similarly, users can override the access mode of a type again via template parameters:
In this case, even if renderable appears among the parameters of the function as not constant, it's treated as constant as regards the generation of the task graph.
To generate the task graph, the organizer offers the graph member function:
A graph is returned in the form of an adjacency list. Each vertex offers the following features:
ro_count and rw_count: the number of resources accessed in read-only or read-write mode.ro_dependency and rw_dependency: type info objects associated with the parameters of the underlying function.top_level: true if a node is a top level one (it has no entering edges), false otherwise.info: type info object associated with the underlying function.name: the name associated with the given vertex if any, a null pointer otherwise.callback: a pointer to the function to execute and whose function type is void(const void *, entt::registry &).data: optional data to provide to the callback.children: the vertices reachable from the given node, in the form of indices within the adjacency list.Since the creation of pools and resources within the registry isn't necessarily thread safe, each vertex also offers a prepare function which is used to setup a registry for execution with the created graph:
The actual scheduling of the tasks is the responsibility of the user, who can use the preferred tool.
Each registry has a context associated with it, which is an any object map accessible by both type and name for convenience. The name isn't really a name though. In fact, it's a numeric id of type id_type used as a key for the variable. Any value is accepted, even runtime ones.
The context is returned via the ctx functions and offers a minimal set of features including the following:
A context variable must be both default constructible and movable. If the supplied type doesn't match that of the variable when using a name, the operation fails.
For all users who want to use the context but don't want to create elements, the contains and find functions are also available:
Also in this case, both functions support constant types and accept a name for the variable to look up, as does at.
A context also supports creating aliases for existing variables that aren't directly managed by the registry. Const and therefore read-only variables are also accepted.
To do that, the type used upon construction must be a reference type and an lvalue is necessarily provided as an argument:
Read-only aliased properties are created using const types instead:
Note that insert_or_assign doesn't support aliased properties and users must necessarily use emplace or emplace_as for this purpose.
When insert_or_assign is used to update an aliased property, it converts the property itself into a non-aliased one.
From the point of view of the user, there are no differences between a variable that is managed by the registry and an aliased property. However, read-only variables aren't accessible as non-const references:
Aliased properties are erased as it happens with any other variable. Similarly, it's also possible to assign them a name.
This module comes with bare minimum support to serialization.
It doesn't convert components to bytes directly, there wasn't the need of another tool for serialization out there. Instead, it accepts an opaque object with a suitable interface (namely an archive) to serialize its internal data structures and restore them later. The way types and instances are converted to a bunch of bytes is completely in charge to the archive and thus to final users.
The goal of the serialization part is to allow users to make both a dump of the entire registry or a narrower snapshot, that is to select only the components in which they are interested.
Intuitively, the use cases are different. As an example, the first approach is suitable for local save/restore functionalities while the latter is suitable for creating client-server applications and for transferring somehow parts of the representation side to side.
To take a snapshot of a registry, use the snapshot class:
It isn't necessary to invoke all functions each and every time. What functions to use in which case mostly depends on the goal.
When getting an entity type, the snapshot class serializes all entities along with their versions.
In all other case, entities and components from a given storage are passed to the archive. Named pools are also supported:
There exists another version of the get member function that accepts a range of entities to serialize. It can be used to filter out those entities that shouldn't be serialized for some reasons:
Once a snapshot is created, there exist mainly two ways to load it: as a whole and in a kind of continuous mode.
The following sections describe both loaders and archives in details.
A snapshot loader requires that the destination registry be empty. It loads all the data at once while keeping intact the identifiers that the entities originally had:
It isn't necessary to invoke all functions each and every time. What functions to use in which case mostly depends on the goal.
For obvious reasons, what is important is that the data are restored in exactly the same order in which they were serialized.
When getting an entity type, a snapshot loader restores all entities with the versions that they originally had at the source.
In all other cases, entities and components are restored in a given storage. If the registry doesn't contain the entity, it's also created accordingly. As for the snapshot class, named pools are supported too:
Finally, the orphans member function releases the entities that have no components after a restore, if any.
A continuous loader is designed to load data from a source registry to a (possibly) non-empty destination. The loader accommodates in a registry more than one snapshot in a sort of continuous loading that updates the destination one step at a time.
Identifiers that entities originally had are not transferred to the target. Instead, the loader maps remote identifiers to local ones while restoring a snapshot. Wrapping the archive is a conveninent way of updating identifiers that are part of components automatically (see the example below).
Another difference with the snapshot loader is that the continuous loader has an internal state that must persist over time. Therefore, there is no reason to limit its lifetime to that of a temporary object:
It isn't necessary to invoke all functions each and every time. What functions to use in which case mostly depends on the goal.
For obvious reasons, what is important is that the data are restored in exactly the same order in which they were serialized.
When getting an entity type, a loader restores groups of entities and maps each entity to a local counterpart when required. For each remote identifier not yet registered by the loader, a local identifier is created so as to keep the local entity in sync with the remote one.
In all other cases, entities and components are restored in a given storage. If the registry doesn't contain the entity, it's also tracked accordingly. As for the snapshot class, named pools are supported too:
Finally, the orphans member function releases the entities that have no components after a restore, if any.
Archives must publicly expose a predefined set of member functions. The API is straightforward and consists only of a group of function call operators that are invoked by the snapshot class and the loaders.
In particular:
An output archive (the one used when creating a snapshot) exposes a function call operator with the following signature to store entities:
Where entt::entity is the type of the entities used by the registry.
Note that all member functions of the snapshot class also make an initial call to store aside the size of the set they are going to store. In this case, the expected function type for the function call operator is:
In addition, an archive accepts (const) references to the types of component to serialize. Therefore, given a type T, the archive offers a function call operator with the following signature:
The output archive can freely decide how to serialize the data. The registry isn't affected at all by the decision.
An input archive (the one used when restoring a snapshot) exposes a function call operator with the following signature to load entities:
Where entt::entity is the type of the entities used by the registry. Each time the function is invoked, the archive reads the next element from the underlying storage and copies it in the given variable.
All member functions of a loader class also make an initial call to read the size of the set they are going to load. In this case, the expected function type for the function call operator is:
In addition, an archive accepts references to the types of component to restore. Therefore, given a type T, the archive contains a function call operator with the following signature:
Every time this operator is invoked, the archive reads the next element from the underlying storage and copies it in the given variable.
EnTT comes with some examples (actually some tests) that show how to integrate a well known library for serialization as an archive. It uses Cereal C++ under the hood, mainly because I wanted to learn how it works at the time I was writing the code.
The code isn't production-ready and it isn't neither the only nor (probably) the best way to do it. However, feel free to use it at your own risk.
The basic idea is to store everything in a group of queues in memory, then bring everything back to the registry with different loaders.
Pools of components are specialized versions of the sparse set class. Each pool contains all the instances of a single component type and all the entities to which it's assigned.
Sparse arrays are paged to avoid wasting memory. Packed arrays of components are also paged to have pointer stability upon additions. Packed arrays of entities are not instead.
All pools rearranges their items in order to keep the internal arrays tightly packed and maximize performance, unless full pointer stability is enabled.
In EnTT, almost everything is customizable. Pools are no exception.
In this case, the standardized way to access all component properties is the component_traits class.
Various parts of the library access component properties through this class. It makes it possible to use any type as a component, as long as its specialization of component_traits implements all the required functionalities.
The non-specialized version of this class contains the following members:
in_place_delete: Type::in_place_delete if present, true for non-movable types and false otherwise.page_size: Type::page_size if present, ENTT_PACKED_PAGE for non-empty types and 0 otherwise.Where Type is any type of component. Properties are customized by specializing the above class and defining its members, or by adding only those of interest to a component definition:
The component_traits class template takes care of extracting the properties from the supplied type.
Plus, it's sfinae-friendly and also supports feature-based specializations.
An empty type T is such that std::is_empty_v<T> returns true. They also are the same types for which empty base optimization (EBO) is possible.
EnTT handles these types in a special way, optimizing both in terms of performance and memory usage. However, this also has consequences that are worth mentioning.
When an empty type is detected, it's not instantiated by default. Therefore, only the entities to which it's assigned are made available. There doesn't exist a way to get empty types from a storage or a registry. Views and groups never return their instances too (for example, during a call to each).
On the other hand, iterations are faster because only the entities to which the type is assigned are considered. Moreover, less memory is used, mainly because there doesn't exist any instance of the component, no matter how many entities it is assigned to.
More in general, none of the feature offered by the library is affected, but for the ones that require to return actual instances.
This optimization is disabled by defining the ENTT_NO_ETO macro. In this case, empty types are treated like all other types. Setting a page size at component level via the component_traits class template is another way to disable this optimization selectively rather than globally.
A void storage (or entt::storage<void> or entt::basic_storage<Type, void>), is a fully functional storage type used to create pools not associated with a particular component type.
From a technical point of view, it's in all respects similar to a storage for empty types when their optimization is enabled. Pagination is disabled as well as pointer stability (as not necessary).
However, this should be preferred to using a simple sparse set. In particular, a void storage offers all those feature normally offered by other storage types. Therefore, it's a perfectly valid pool for use with views and groups or within a registry.
This storage is such that the component type is the same as the entity type, for example entt::storage<entt::entity> or entt::basic_storage<Type, Type>.
For this type of pools, there is a specific specialization within EnTT. In fact, entities are subject to different rules with respect to components (although still customizable by the user if needed). In particular:
emplace as well as insert have a slightly different meaning than their counterparts for components. In the case of an entity storage, these functions generate or recycle identifiers rather than allowing them to be assigned to existing entities.each function iterates only the entities in use, that is, those not marked as ready for reuse. To iterate all the entities it's necessary to iterate the underlying sparse set instead.This kind of storage is designed to be used where any other storage is fine and can therefore be combined with views, groups and so on.
Within the registry, an entity storage is treated in all respects like any other storage.
Therefore, it's possible to add mixins to it as well as retrieve it via the storage function. It can also be used as storage in a view (for exclude-only views for example):
However, it's also subject to a couple of exceptions, partly out of necessity and partly for ease of use.
In particular, it's not possible to create multiple elements of this type.
This means that the name used to retrieve this kind of storage is ignored and the registry will only ever return the same element to the caller. For example:
In this case, the identifier is discarded as is. The call is in all respects equivalent to the following:
Because entity storage doesn't have a name, it can't be retrieved via the opaque storage function either.
It would make no sense to try anyway, given that the type of the registry and therefore its entity type are known regardless.
Finally, when the user asks the registry for an iterable object to visit all the storage elements inside it as follows:
Entity storage is never returned. This simplifies many tasks (such as copying an entity) and fits perfectly with the fact that this type of storage doesn't have an identifier inside the registry.
The ability to achieve pointer stability for one, several or all components is a direct consequence of the design of EnTT and of its default storage.
In fact, although it contains what is commonly referred to as a packed array, the default storage is paged and doesn't suffer from invalidation of references when it runs out of space and has to reallocate.
However, this isn't enough to ensure pointer stability in case of deletion. For this reason, a stable deletion method is also offered. This one is such that the position of the elements is preserved by creating tombstones upon deletion rather than trying to fill the holes that are created.
For performance reasons, EnTT favors storage compaction in all cases, although often accessing a component occurs mostly randomly or traversing pools in a non-linear order on the user side (as in the case of a hierarchy).
In other words, pointer stability is not automatic but is enabled on request.
The library offers out of the box support for in-place deletion, thus offering storage with completely stable pointers. This is achieved by specializing the component_traits class or by adding the required properties to the component definition when needed.
Views and groups adapt accordingly when they detect a storage with a different deletion policy than the default. In particular:
size_hint is available.In other words, the more generic version of a view is provided in case of stable storage, even for a single type view.
In no case a tombstone is returned from the view itself. Likewise, non-existent components aren't returned, which could otherwise result in an UB.
EnTT doesn't attempt in any way to offer built-in methods with hidden or unclear costs to facilitate the creation of hierarchies.
There are various solutions to the problem, such as using the following class:
However, it should be pointed out that the possibility of having stable pointers for one, many or all types solves the problem of hierarchies at the root in many cases.
In fact, if a certain type of component is visited mainly in random order or according to hierarchical relationships, using direct pointers has many advantages:
Furthermore, it's quite common for a group of elements to be created close in time and therefore fallback into adjacent positions, thus favoring locality even on random accesses. Locality that isn't sacrificed over time given the stability of storage positions, with undoubted performance advantages.
EnTT takes advantage of what the language offers at compile-time. However, this can have its downsides (well known to those familiar with type erasure techniques).
To fill the gap, the library also provides a bunch of utilities and feature that are very useful to handle types and pools at runtime.
Storage classes are fully self-contained types. They are extended via mixins to add more functionalities (generic or type specific). In addition, they offer a basic set of functions that already allow users to go very far.
The aim is to limit the need for customizations as much as possible, offering what is usually necessary for the vast majority of cases.
When a storage is used through its base class (for example, when its actual type isn't known), there is always the possibility of receiving a type_info object for the type of elements associated with the entities (if any):
Furthermore, all features rely on internal functions that forward the calls to the mixins. The latter can then make use of any information, which is set via bind:
The bind function accepts any element by reference or by value and forwards it to derived classes.
This is how a registry passes itself to all pools that support signals and also why a storage keeps sending events without requiring the registry to be passed to it every time.
Alongside these more specific things, there are also a couple of functions designed to address some common requirements such as copying an entity.
In particular, the base class behind a storage offers the possibility to take the value associated with an entity through an opaque pointer:
Similarly, the non-specialized push function accepts an optional opaque pointer and behaves differently depending on the case:
This means that, starting from a reference to the base, it's possible to bind components with entities without knowing their actual type and even initialize them by copy if needed:
This is particularly useful to clone entities in an opaque way. In addition, the decoupling of features allows for filtering or use of different copying policies depending on the type.
EnTT allows the user to assign a name (or rather, a numeric identifier) to a type and then create multiple pools of the same type:
If a name isn't provided, the default storage associated with the given type is always returned.
Since the storage are also self-contained, the registry doesn't duplicate its own API for them. However, there is still no limit to the possibilities of use:
Anything that can be done via the registry interface can also be done directly on the reference storage.
On the other hand, those calls involving all storage are guaranteed to also reach manually created ones:
Finally, a storage of this type works with any view (which also accepts multiple storages of the same type, if necessary):
The possibility of direct use of storage combined with the freedom of being able to create and use more than one of the same type opens the door to the use of EnTT at runtime, which was previously quite limited.
Views are a non-intrusive tool for working with entities and components without affecting other functionalities or increasing memory consumption.
Groups are an intrusive tool to use to improve performance along critical paths but which also has a price to pay for that.
There are mainly two kinds of views: compile-time (also known as view) and runtime (also known as runtime_view).
The former requires a compile-time list of component (or storage) types and can make several optimizations because of that. The latter is constructed at runtime using numerical type identifiers instead and is a bit slower to iterate.
In both cases, creating and destroying views isn't expensive at all since they don't have any type of initialization.
Groups come in three different flavors: full-owning groups, partial-owning groups and non-owning groups. The main difference between them is in terms of performance.
Groups can literally own one or more component types. They are allowed to rearrange pools so as to speed up iterations. Roughly speaking: the more components a group owns, the faster it is to iterate them.
Single type views and multi type views behave differently and also have slightly different APIs.
Single type views are specialized to give a performance boost in all cases. There is nothing as fast as a single type view. They just walk through packed (actually paged) arrays of elements and return them directly.
This kind of views also allow to get the exact number of elements they are going to return.
Refer to the inline documentation for all the details.
Multi type views iterate entities that have at least all the given components. During construction, they look at the number of elements available in each pool and use the smallest set in order to speed up iterations.
This kind of views only allow to get the estimated number of elements they are going to return.
Refer to the inline documentation for all the details.
Storing aside views isn't required as they are extremely cheap to construct. In fact, this is even discouraged when creating a view from a const registry. Since all storage are lazily initialized, they may not exist when the view is created. Thus, while perfectly usable, the view may contain pending references that are never reinitialized with the actual storage.
Views share the way they are created by means of a registry:
Filtering entities by components is also supported:
To iterate a view, either use it in a range-for loop:
Or rely on the each member functions to iterate both entities and components at once:
Note that entities can also be excluded from the parameter list when received through a callback and this can improve even further the performance during iterations.
Since they aren't explicitly instantiated, empty components aren't returned in any case.
As a side note, in the case of single type views, get accepts but doesn't strictly require a template parameter, since the type is implicitly defined. However, when the type isn't specified, the instance is returned using a tuple for consistency with multi type views:
Note: prefer the get member function of a view instead of that of a registry during iterations to get the types iterated by the view itself.
Views support lazy initialization as well as storage swapping.
An empty (or partially initialized) view is such that it returns false when converted to bool (to let the user know that it isn't fully initialized) but it also works as-is like any other view.
In order to initialize a view one piece at a time, it allows users to inject storage classes when available:
If there are multiple storages of the same type, it's possible to disambiguate using the index of the element to be replaced:
The ability to literally replace a storage in a view also opens up its reuse with different sets of entities.
For example, to filter a view based on two groups of entities with different characteristics, there will be no need to reinitialize anything:
Finally, it should be noted that the lack of a storage is treated to all intents and purposes as if it were an empty element.
Thus, a get storage (as in entt::get_t) makes the view empty automatically while an exclude storage (as in entt::exclude_t) is ignored as if that part of the filter didn't exist.
Exclude-only views aren't really a thing in EnTT.
However, the same result can be achieved by combining the right storage into a simple view.
If one gets to the root of the problem, the purpose of an exclude-only view is to return entities that don't meet certain requirements.
Since entity storage, unlike exclude-only views, is a thing in EnTT, users can leverage it for these kinds of queries. It's also guaranteed to be unique within a registry and is always accessible when creating a view:
The returned view is such that it will return only the entities that don't have the my_type component, regardless of what other components they have.
Views are combined with each other to create new and more specific queries.
The type returned when combining multiple views together is itself a view, more in general a multi component one.
Combining different views tries to mimic C++20 ranges:
The constness of the types is preserved and their order depends on the order in which the views are combined. For example, the pack above returns an instance of position first and then one of velocity.
Since combining views generates views, a chain can be of arbitrary length and the above type order rules apply sequentially.
By default, a view is iterated along the pool that contains the smallest number of elements.
For example, if the registry contains fewer velocitys than it contains positions, then the order of the elements returned by the following view depends on how the velocity components are arranged in their pool:
Moreover, the order of types when constructing a view doesn't matter. Neither does the order of views in a view pack.
However, it's possible to enforce iteration of a view by given component order by means of the use function:
On the other hand, if all a user wants is to iterate the elements in reverse order, this is possible for a single type view using its reverse iterators:
Unfortunately, multi type views don't offer reverse iterators. Therefore, in this case it's a must to implement this functionality manually or to use single type views to lead the iteration.
Multi type views iterate entities that have at least all the given components. During construction, they look at the number of elements available in each pool and use the smallest set in order to speed up iterations.
They offer more or less the same functionalities of a multi type view. However, they don't expose a get member function and users should refer to the registry that generated the view to access components.
Refer to the inline documentation for all the details.
Runtime views are pretty cheap to construct and should not be stored aside in any case. They should be used immediately after creation and then they should be thrown away.
To iterate a runtime view, either use it in a range-for loop:
Or rely on the each member function to iterate entities:
Performance are exactly the same in both cases.
Filtering entities by components is also supported for this kind of views:
Runtime views are meant for when users don't know at compile-time what types to use to iterate entities. The storage member function of a registry could be useful in this regard.
Groups are meant to iterate multiple components at once and to offer a faster alternative to multi type views.
Groups overcome the performance of the other tools available but require to get the ownership of components. This sets some constraints on their pools. On the other hand, groups aren't an automatism that increases memory consumption, affects functionalities and tries to optimize iterations for all the possible combinations of components. Users can decide when to pay for groups and to what extent.
The most interesting aspect of groups is that they fit usage patterns. Other solutions around usually try to optimize everything, because it is known that somewhere within the everything there are also our usage patterns. However this has a cost that isn't negligible, both in terms of performance and memory usage. Ironically, users pay the price also for things they don't want and this isn't something I like much. Even worse, one cannot easily disable such a behavior. Groups work differently instead and are designed to optimize only the real use cases when users find they need to.
Another nice-to-have feature of groups is that they have no impact on memory consumption, put aside full non-owning groups that are pretty rare and should be avoided as long as possible.
All groups affect to an extent the creation and destruction of their components. This is due to the fact that they must observe changes in the pools of interest and arrange data correctly when needed for the types they own.
In all cases, a group allows to get the exact number of elements it's going to return.
Refer to the inline documentation for all the details.
Storing aside groups isn't required as they are extremely cheap to create, even though valid groups can be copied without problems and reused freely.
A group performs an initialization step the very first time it's requested and this could be quite costly. To avoid it, consider creating the group when no components have been assigned yet. If the registry is empty, preparation is extremely fast.
To iterate a group, either use it in a range-for loop:
Or rely on the each member functions to iterate both entities and components at once:
Note that entities can also be excluded from the parameter list when received through a callback and this can improve even further the performance during iterations.
Since they aren't explicitly instantiated, empty components aren't returned in any case.
Note: prefer the get member function of a group instead of that of a registry during iterations to get the types iterated by the group itself.
A full-owning group is the fastest tool a user can expect to use to iterate multiple components at once. It iterates all the components directly, no indirection required.
This type of groups performs more or less as if users are accessing sequentially a bunch of packed arrays of components all sorted identically, with no jumps nor branches.
A full-owning group is created as:
Filtering entities by components is also supported:
Once created, the group gets the ownership of all the components specified in the template parameter list and arranges their pools as needed.
Sorting owned components is no longer allowed once the group has been created. However, full-owning groups are sorted using their sort member functions. Sorting a full-owning group affects all its instances.
A partial-owning group works similarly to a full-owning group for the components it owns, but relies on indirection to get components owned by other groups.
This isn't as fast as a full-owning group, but it's already much faster than a view when there are only one or two free components to retrieve (the most common cases likely). In the worst case, it's not slower than views anyway.
A partial-owning group is created as:
Filtering entities by components is also supported:
Once created, the group gets the ownership of all the components specified in the template parameter list and arranges their pools as needed. The ownership of the types provided via entt::get doesn't pass to the group instead.
Sorting owned components is no longer allowed once the group has been created. However, partial-owning groups are sorted using their sort member functions. Sorting a partial-owning group affects all its instances.
Non-owning groups are usually fast enough, for sure faster than views and well suited for most of the cases. However, they require custom data structures to work properly and they increase memory consumption.
As a rule of thumb, users should avoid using non-owning groups, if possible.
A non-owning group is created as:
Filtering entities by components is also supported:
The group doesn't receive the ownership of any type of component in this case. This type of groups is therefore the least performing in general, but also the only one that can be used in any situation to slightly improve performance.
Non-owning groups are sorted using their sort member functions. Sorting a non-owning group affects all its instances.
The registry class offers two overloads when it comes to constructing views and groups: a const version and a non-const one. The former accepts only const types as template parameters, the latter accepts both const and non-const types instead.
It means that views and groups generated by a const registry also propagate the constness to the types involved. As an example:
Consider the following definition for a non-const view instead:
In the example above, view is used to access either read-only or writable position components while velocity components are read-only in all cases.
Similarly, these statements are all valid:
It's not possible to get non-const references to velocity components from the same view instead. Therefore, these result in compilation errors:
The each member functions also propagates constness to its return values:
A caller can still refer to the position components through a const reference because of the rules of the language that fortunately already allow it.
The same concepts apply to groups as well.
Views and groups are narrow windows on the entire list of entities. They work by filtering entities according to their components.
In some cases there may be the need to iterate all the entities still in use regardless of their components. This is done by accessing entity storage:
As a rule of thumb, consider using a view or a group if the goal is to iterate entities that have a determinate set of components. These tools are usually much faster than filtering entities with a bunch of custom tests.
In all the other cases, this is the way to go. For example, it's possible to combine this view with the orphan member function to clean up orphan entities (that is, entities that are still in use and have no assigned components):
In general, iterating all entities can result in poor performance. It should not be done frequently to avoid the risk of a performance hit.
However, it's convenient when initializing an editor or to reclaim pending identifiers.
Most of the ECS available out there don't allow to create and destroy entities and components during iterations, nor to have pointer stability.
EnTT partially solves the problem with a few limitations:
In other terms, iterators are rarely invalidated. Also, component references aren't invalidated when a new element is added while they could be invalidated upon destruction due to the swap-and-pop policy, unless the type leading the iteration undergoes in-place deletion.
As an example, consider the following snippet:
The each member function won't break (because iterators remain valid) nor will any reference be invalidated. Instead, more attention should be paid to the destruction of entities or the removal of components.
Use a common range-for loop and get components directly from the view or move the deletion of entities and components at the end of the function to avoid dangling pointers.
For all types that don't offer stable pointers, iterators are also invalidated and the behavior is undefined if an entity is modified or destroyed and it's not the one currently returned by the iterator nor a newly created one.
To work around it, possible approaches are:
A notable side effect of this feature is that the number of required allocations is further reduced in most cases.
Groups are a faster alternative to views. However, the higher the performance, the greater the constraints on what is allowed and what is not.
In particular, groups add in some rare cases a limitation on the creation of components during iterations. It happens in quite particular cases. Given the nature and the scope of the groups, it isn't something in which it will happen to come across probably, but it's good to know it anyway.
First of all, it must be said that creating components while iterating a group isn't a problem at all and is done freely as it happens with the views. The same applies to the destruction of components and entities, for which the rules mentioned above apply.
The additional limitation arises instead when a given component that is owned by a group is iterated outside of it. In this case, adding components that are part of the group itself may invalidate the iterators. There are no further limitations to the destruction of components and entities.
Fortunately, this isn't always true. In fact, it almost never is and only happens under certain conditions. In particular:
In other words, the limitation doesn't exist as long as a type is treated as a free type (as an example with multi type views and partial- or non-owning groups) or iterated with its own group, but it can occur if the type is used as a main type to rule on an iteration.
This happens because groups own the pools of their components and organize the data internally to maximize performance. Because of that, full consistency for owned components is guaranteed only when they are iterated as part of their groups or as free types with multi type views and groups in general.
In general, the entire registry isn't thread safe as it is. Thread safety isn't something that users should want out of the box for several reasons. Just to mention one of them: performance.
Views, groups and consequently the approach adopted by EnTT are the great exception to the rule. It's true that views, groups and iterators in general aren't thread safe by themselves. Because of this users shouldn't try to iterate a set of components and modify the same set concurrently. However:
X or assign and removes that component from a set of entities, another thread can safely do the same with components Y and Z and everything work like just fine. As a trivial example, users can freely execute the rendering system and iterate the renderable entities while updating a physic component concurrently on a separate thread.This kind of entity-component systems can be used in single threaded applications as well as along with async stuff or multiple threads. Moreover, typical thread based models for ECS don't require a fully thread safe registry to work. Actually, users can reach the goal with the registry as it is while working with most of the common models.
Because of the few reasons mentioned above and many others not mentioned, users are completely responsible for synchronization whether required. On the other hand, they could get away with it without having to resort to particular expedients.
Finally, EnTT is configured via a few compile-time definitions to make some of its parts implicitly thread-safe, roughly speaking only the ones that really make sense and can't be turned around.
In particular, when multiple instances of objects referencing the type index generator (such as the registry class) are used in different threads, then it might be useful to define ENTT_USE_ATOMIC.
See the relevant documentation for more information.
A special mention is needed for the iterators returned by views and groups. Most of the time they meet the requirements of random access iterators, in all cases they meet at least the requirements of forward iterators.
In other terms, they are suitable for use with the parallel algorithms of the standard library. If it's not clear, this is a great thing.
As an example, this kind of iterators are used in combination with std::for_each and std::execution::par to parallelize the visit and therefore the update of the components returned by a view or a group, as long as the constraints previously discussed are respected:
This can increase the throughput considerably, even without resorting to who knows what artifacts that are difficult to maintain over time.
Unfortunately, because of the limitations of the current revision of the standard, the parallel std::for_each accepts only forward iterators. This means that the default iterators provided by the library cannot return proxy objects as references and must return actual reference types instead.
This may change in the future and the iterators will almost certainly return both the entities and a list of references to their components by default sooner or later. Multi-pass guarantee won't break in any case and the performance should even benefit from it further.
A const registry is also fully thread safe. This means that it's not able to lazily initialize a missing storage when a view is generated.
The reason for this is easy to explain. To avoid requiring types to be announced in advance, a registry lazily creates the storage objects for the different components. However, this isn't possible for a thread safe const registry.
Returned views are always valid and behave as expected in the context of the caller. However, they may contain dangling references to non-existing storage when created from a const registry.
As a result, such a view may misbehave over time if it's kept aside for a second use.
Therefore, if the general advice is to create views when necessary and discard them immediately afterwards, this becomes almost a rule when it comes to views generated from a const registry.
Fortunately, there is also a way to instantiate storage classes early when in doubt or when there are special requirements.
Calling the storage method is equivalent to announcing a particular storage, so as to avoid running into problems. For those interested, there are also alternative approaches, such as a single threaded tick for the registry warm-up, but these are not always applicable.
In this case, views never risk becoming invalid.
There are many other features and functions not listed in this document.
EnTT and in particular its ECS part is in continuous development and some things could be forgotten, others could have been omitted on purpose to reduce the size of this file. Unfortunately, some parts may even be outdated and still to be updated.
For further information, it's recommended to refer to the documentation included in the code itself or join the official channels to ask a question.
1.10.0