Memory management
Mirroring load and unload
When working with Addressable Assets, the primary way to ensure proper memory management is to mirror your load and unload calls correctly. How you do so depends on your Asset types and load methods. In all cases, however, the release method can either take the loaded Asset, or an operation handle returned by the load. For example, during Scene creation (described below) the load returns a AsyncOperationHandle<SceneInstance>
, which you can release via this returned handle, or by keeping up with the handle.Result
(in this case, a SceneInstance
).
Asset loading
To load an Asset, use Addressables.LoadAssetAsync
(single asset) or Addressables.LoadAssetsAsync
(multiple assets).
Note: LoadAssetAsync
is intended for use with keys that map to single entries. If you provide a key that matches multiple entries (such as a widely used label) the method will load the first match it finds to the given key. This is not deterministic as it can be affected by build order.
This loads the Asset into memory without instantiating it. Every time the load call executes, it adds one to the ref-count for each Asset loaded. If you call LoadAssetAsync
three times with the same address, you will get three different instances of an AsyncOperationHandle
struct back, all referencing the same underlying operation. That operation has a ref-count of three for the corresponding Asset. If the load succeeds, the resulting AsyncOperationHandle
struct contains the Asset in the .Result
property. You can use the loaded Asset to instantiate using Unity's built-in instantiation methods, which does not increment the Addressables ref-count.
To unload the Asset, use the Addressables.Release
method, which decrements the ref-count. When a given Asset's ref-count is zero, that Asset is ready to be unloaded, and decrements the ref-count of any dependencies.
Note: The Asset may or may not be unloaded immediately, contingent on existing dependencies. For more information, read the section on when memory is cleared.
Scene loading
To load a Scene, use Addressables.LoadSceneAsync
. You can use this method to load a Scene in Single
mode, which closes all open Scenes, or in Additive
mode (for more information, see documentation on Scene mode loading.
To unload a Scene, use Addressables.UnloadSceneAsync
, or open a new Scene in Single
mode. You can open a new Scene by either using the Addressables interface, or using the SceneManager.LoadScene
or SceneManager.LoadSceneAsync
methods. Opening a new Scene closes the current one, properly decrementing the ref-count.
GameObject instantiation
To load and instantiate a GameObject Asset, use Addressables.InstantiateAsync
. This instantiates the Prefab located by the specified location
parameter. The Addressables system will load the Prefab and its dependencies, incrementing the ref-count of all associated Assets.
Calling InstantiateAsync
three times on the same address results in all dependent assets having a ref-count of three. Unlike calling LoadAssetAsync
three times, however, each InstantiateAsync
call returns an AsyncOperationHandle
pointing to a unique operation. This is because the result of each InstantiateAsync
is a unique instance. You would need to individually release each returned AsyncOperationHandle
or GameObject instance. Another distinction between InstantiateAsync
and other load calls is the optional trackHandle
parameter. When set to false
, you must keep the AsyncOperationHandle
to use while releasing your instance. This is more efficient, but requires more development effort.
To destroy an instantiated GameObject, use Addressables.ReleaseInstance
, or close the Scene that contains the instantiated object. This Scene can have been loaded (and thus closed) in Additive
or Single
mode. This Scene can also have been loaded using either the Addressables
or SceneManagement
API. As noted above, if you set trackHandle
to false
, you can only call Addressables.ReleaseInstance
with the handle, not with the actual GameObject.
Note: If you call Addressables.ReleaseInstance
on an instance that was not created using the Addressables
API, or was created with trackHandle==false
, the system detects that and returns false
to indicate that the method was unable to release the specified instance. The instance will not be destroyed in this case.
Addressables.InstantiateAsync
has some associated overhead, so if you need to instantiate the same objects hundreds of times per frame, consider loading via the Addressables
API, then instantiating through other methods. In this case, you would call Addressables.LoadAssetAsync
, then save the result and call GameObject.Instantiate()
for that result. This allows flexibility to call Instantiate
in a synchronous way. The downside is that the Addressables system has no knowledge of how many instances you created, which can lead to memory issues if not properly managed. For example, a Prefab referencing a texture would no longer have a valid loaded texture to reference, causing rendering issues (or worse). These sorts of problems can be hard to track down as you may not immediately trigger the memory unload (see section on clearing memory, below).
Data loading
Interfaces that do not need their AsyncOperationHandle.Result
released, will still need the operation itself to be released. Examples of these would be Addressables.LoadResourceLocationsAsync
and Addressables.GetDownloadSizeAsync
. They load data that you can access until the operation is released. This release should be done via Addressables.Release
.
Background interactions
Operations that do not return anything in the AsyncOperationHandle.Result
field have have an optional parameter to automatically release the operation handle on completion. If you have no further need for one of these operation handles after it has completed, set the autoReleaseHandle
parameter to true to make sure the operation handle is cleaned up. The scenario where you would want autoReleaseHandle
to be false would be if you needed to check the Status
of the operation handle after it has completed. Examples of these interfaces are Addressables.DownloadDependenciesAsync
and Addressables.UnloadScene
.
The Addressables Event Viewer
Use the Addressables Event Viewer window to monitor the ref-counts of all Addressables system operations. To access the window in the Editor, select Window > Asset Management > Addressables > Event Viewer.
Important: In order to view data in the Event Viewer, you must enable the Send Profiler Events setting in your AddressableAssetSettings
object's Inspector. Changes to Send Profiler Events will be reflected in the following build. This means that entering play mode when using the Use Existing Build play mode script will use the value set during the most recent build. Alternatively, entering play mode when using the Use Asset Database or Simulate Groups play mode scripts will pick up the current state, as those play mode scripts rebuild the settings data upon entering play mode.
The following information is available in the Event Viewer:
- A white vertical line indicates the frame in which a load request occurred.
- A blue background indicates that an Asset is currently loaded.
- The green part of the graph indicates an Asset's current ref-count.
Note that the Event Viewer is only concerned with ref-counts, not memory consumption (see section on clearing memory, below, for more information).
Listed under the Assets column, you will see a row for each of the following, per frame:
- FPS: The frames per second count.
- MonoHeap: The amount of RAM in use.
- Event Counts: The total number of events in a frame.
- Instantiation Counts: The total number of calls to Addressables.InstantiateAsync on a frame.
- Asset requests: Displays the reference count on an operation over time. If the Asset request has any dependencies, a triangle appears that you can click on to view the children's request operations.
You can click the left and right arrows in order to step through the frames one by one, or click Current to jump to the latest frame. Press the + button to expand a row for more details.
The information displayed in the Event Viewer is related to the build script you use to create Play mode data.
When using the Event Viewer, avoid the Use Asset Database built script because it does not account for any shared dependencies among the Assets. Use the Simulate Groups script or the Use Existing Build script instead, but the latter is better suited for the Event Viewer because it gives a more accurate monitoring of the ref-counts.
Connecting the Event Viewer to a standalone player
To connect the Event Viewer to a standalone player, go into the build menu, select the platform you wish to use, and ensure that Development Build and Autoconnect Profiler are both enabled. Next, open the Unity Profiler by selecting Window > Analysis > Profiler and select the platform you wish to build for on the top toolbar. Finally, select Build and Run in the Build Settings window and the Event Viewer will automatically connect to and display events for the standalone player selected.
When is memory cleared?
An Asset no longer being referenced (indicated by the end of a blue section in the profiler) does not necessarily mean that Asset was unloaded. A common applicable scenario involves multiple Assets in an AssetBundle. For example:
- You have three Assets (
tree
,tank
, andcow
) in an AssetBundle (stuff
). - When
tree
loads, the profiler displays a single ref-count fortree
, and one forstuff
. - Later, when
tank
loads, the profiler displays a single ref-count for bothtree
andtank
, and two ref-counts for thestuff
bundle. - If you release
tree
, it's ref-count becomes zero, and the blue bar goes away.
In this example, the tree
Asset is not actually unloaded at this point. You can load an AssetBundle, or its partial contents, but you cannot partially unload an AssetBundle. No Asset in stuff
will unload until the bundle itself is completely unloaded. The exception to this rule is the engine interface Resources.UnloadUnusedAssets
. Executing this method in the above scenario will cause tree
to unload. Because the Addressables system cannot be aware of these events, the profiler graph only reflects the Addressables ref-counts (not exactly what memory holds). Note that if you choose to use Resources.UnloadUnusedAssets
, it is a very slow operation, and should only be called on a screen that won't show any hitches (such as a loading screen).
AssetBundle dependencies
Loading an Addressable Asset loads all the AssetBundle dependencies and keeps them loaded until you call Addressables.Release
on the handle returned from the loading method.
AssetBundle dependencies are created when an asset in one AssetBundle references an asset in another AssetBundle. The dependencies of all these AssetBundles can be thought of as a dependency graph. During the catalog generation stage of the build process, Addressables walks this graph to calculate all the AssetBundles that must be loaded for each Addressable Asset. Because dependencies are calculated at the AssetBundle level, all Addressable Assets within a single AssetBundle have the same dependencies. Adding an Addressable Asset that has an external reference (references an object in another AssetBundle) to an AssetBundle adds that AssetBundle as a dependency for all the other Addressable Assets in the AssetBundle.
For Example:
BundleA
contains Addressable Assets Asset1
and Asset2
. Asset2
references Asset3
, which is contained in BundleB
. Even though Asset1
has no reference to BundleB
, BundleB
is still a dependency of Asset1
because Asset1
is in BundleA
, which has a reference on BundleB
.
NOTE
The dependency calculation for the required set of AssetBundles has changed as of 1.13.0. Prior to 1.13.0, only the AssetBundles that contained assets for the requested Addressable Asset were loaded. In the example above, Asset1
would not have had a dependency on BundleB
. This previous behavior was resulting in references breaking when an AssetBundle being referenced by another AssetBundle was unloaded and reloaded.