Asynchronous Loading

We’ve seen a basic way to load models in Model Files using loader.loadModel(). The major problem with this call is that it blocks the main thread while the model is being loaded, which means that all other tasks on the main thread (including Panda’s rendering task) are blocked until the model has finished loading. This is noticeable by the user as a jarring lag, especially when the application freezes for longer periods of time.

The following example demonstrates the naive way to load the scene models:

class Game(ShowBase):
    def __init__(self):
        ShowBase.__init__(self)

        self.loadScene()

    def loadScene(self):
        text = OnscreenText("Loading…")

        self.terrainModel = loader.loadModel("terrain")
        self.terrainModel.reparentTo(render)
        self.cityModel = loader.loadModel("city")
        self.cityModel.reparentTo(render)

        text.destroy()

You may notice that the “Loading” screen will never appear, because Panda3D never gets a chance to render it! We could force Panda3D to render a frame after creating the text object, but still, any operation that requires a re-render (such as resizing the window, or alt-tabbing to another application) would cause the window to become black and may even prompt the operating system to warn that the application is not responding.

Clearly, this does not provide a good user experience. Therefore, it is recommended that models are loaded in an asynchronous manner, in a separate thread of execution, so that the application can continue rendering while the load operation occurs in the background. Panda3D provides several ways of doing so.

Loading with callback

One of the ways to achieve asynchronous loading is with the callback argument to loader.loadModel(). If callback is not None, then the model load will be performed asynchronously. In this case, loadModel() will initiate a background load and return immediately. The return value will be an object that you may call .cancel() on to cancel the asynchronous request. At some later point, when the requested model(s) have finished loading, the callback function will be invoked with the n loaded models passed as its parameter list. It is possible that the callback will be invoked immediately, even before loadModel() returns.

If you use a callback, you may also specify a priority, which specifies the relative importance over this model over all of the other asynchronous load requests (higher numbers are loaded first).

The following example shows how to use this feature.

class Game(ShowBase):
    def __init__(self):
        ShowBase.__init__(self)

        self.accept('escape', self.quit)

        self.loadRequest = None
        self.startLoading()

    def startLoading(self):
        self.loadingText = OnscreenText("Loading…")

        self.loadRequest = loader.loadModel(["terrain", "city"], callback=self.finishLoading)

    def finishLoading(self, models):
        # Get rid of temporary objects
        self.loadRequest = None
        self.loadingText.destroy()
        del self.loadingText

        # Process the models that finished loading
        self.terrainModel, self.cityModel = models

        self.terrainModel.reparentTo(render)
        self.cityModel.reparentTo(render)

    def quit(self):
        if self.loadRequest:
            self.loadRequest.cancel()

        sys.exit()

Loading in a coroutine

As you can see, the previous approach made the code quite a bit more convoluted. We had to split up the load process into two methods, and also take special care to ensure that the load request was cancelled when necessary, and take care of where the intermediate variables were stored during the load operation. If we also wanted to handle exceptions in the load operation properly, it would get more complicated still!

A far more convenient way to do this is using Coroutines, introduced in Python 3.5 and supported as of Panda3D 1.10. These are special functions that can be suspended temporarily and resumed at a later point (pending the completion of an asynchronous operation). Instead, we can write our code as though it were synchronous, but we insert the await keyword where we want the task to be suspended while waiting for the following operation.

To make this possible, a few things are necessary:

1. We need to put async in front of our function.

2. We can no longer call the function directly, but rather need to schedule its execution using the task manager.

3. The asynchronous operation needs to return a future object. To get loader.loadModel() to do so, we need to pass the blocking=False parameter.

4. We need to use await on this future object to suspend the task while the operation is not yet done.

This may seem complicated at first, but it really allows us to write much more straightforward code:

class Game(ShowBase):
    def __init__(self):
        ShowBase.__init__(self)

        self.accept('escape', self.quit)

        self.taskMgr.add(self.loadScene())

    async def loadScene(self):
        text = OnscreenText("Loading…")

        # Load the models in the background, each time suspending this
        # method until they are done
        self.terrainModel = await loader.loadModel("terrain", blocking=False)
        self.cityModel = await loader.loadModel("city", blocking=False)

        self.terrainModel.reparentTo(render)
        self.cityModel.reparentTo(render)

        text.destroy()

    def quit(self):
        sys.exit()

Loading in a thread

Alternatively, it is possible to use a separate thread to initiate the model load. Panda3D’s scene graph is thread-safe and can safely handle model operations from any thread. See the Threading page for more details.

One thing to note is that you may want to make sure that you complete all model operations (positioning, material assignments, etc.) before attaching it into the scene graph. Otherwise, if Panda3D happens to render a frame in between those calls, there is a chance that the model may briefly appear in its original state.

On-demand texture loading

In addition, you can further ask textures to be loaded to the graphics card asynchronously. This means that the first time you look at a particular model, the texture might not be available; but instead of holding up the frame while we wait for it to be loaded, Panda can render the model immediately, with a very low-resolution version of the texture or even a flat color, and start loading of the full-resolution version in the background. When the texture is eventually loaded, it will be applied. This results in fewer frame-rate chugs, but it means that the model looks a little weird at first. It has the greatest advantage when you are using lazy-load textures as well as texture compression, because it means these things will happen in the background. Use these configuration options to enable this behavior:

preload-textures 0
preload-simple-textures 1
simple-image-size 16 16
compressed-textures 1
allow-incomplete-render 1

When converting models to .bam with preload-simple-textures active, simple textures will be baked into the model, so that Panda (starting with version 1.10.11) doesn’t need to load the textures from disk at all until they first come into view.

To test this process, you can set async-load-delay with a value in seconds, which artificially delays each individual texture load by the given amount. This is useful for simulating the user experience on older computers with slower hard drives. Set it to a value like 0.1 and you should see the textures pop in as you move around the scene.

You can use DisplayRegion.setTextureReloadPriority() if you want ensure that textures in some scenes are loaded with higher priority than other scenes.

Animation loading

A similar behavior can be enabled for Actors, so that when you have an Actor with a large number of animations (too many to preload them all at once), you can have the Actor load them on-demand, so that when you play an animation, the animation may not start playing immediately, but will instead be loaded in the background. Until it is ready, the actor will hold its last pose, and then when the animation is fully loaded, the actor will start playing where it would have been had the animation been loaded from the beginning. To make this work, you have to run all of the animations through egg-optchar with the -preload option, and you might also want to set:

allow-async-bind 1
restore-initial-pose 0

Configuration

All of the above asynchronous operations will take place on a separate task chain, automatically created by Loader. By default, one low-priority thread is created to serve these requests. To increase the number of available threads, or to increase their priority, these configuration variables can be changed:

# default is 1
loader-num-threads 2
# default is low
loader-thread-priority normal