Best Practices
Note: This documentation is for the old 0.8.0 version of A-Frame. Check out the documentation for the current 1.6.0 version
A-Frame
The core structure of A-Frame is entity-component (ECS). Place and
structure application code within purely A-Frame components for reusability,
modularity, composability, decoupling, encapsulation, declarativeness, and
testability. It’s okay to start out at first using content scripts
(<script>
), but look to move towards components.
WebVR
Performance
Performance is critical in VR. A high framerate must be maintained in order for people to feel comfortable. Here are some ways to help improve performance of an A-Frame scene:
- Use recommended hardware specifications.
- Use the stats component to keep an eye on various metrics (FPS, vertex and face count, geometry and material count, draw calls, number of entities). We want to maximize FPS and minimize everything else.
- Make use of the asset management system to benefit from browser caching and preloading. Trying to fetch assets while rendering is slower than fetching all assets before rendering.
- If using models, look to bake your lights into textures rather than relying on real-time lighting and shadows.
- Generally, the fewer number of entities and lights in the scene, the better.
- Make sure your textures’ resolutions are sized to powers of two (e.g., 256x256, 512x1024) in order to avoid the renderer having to resize the texture during runtime.
- Limit the number of faces and vertices on models.
- Some further techniques (not yet documented) include geometry instancing, geometry merging, level of detail (LOD).
- When using raycasters or colliders, select which entities are to be raycasted against rather than raycasting against every object in the scene.
- When adding continuously running behaviors, use A-Frame component
tick
handlers to hook into the global render loop. Use utilities such asAFRAME.utils.throttleTick
to limit the number of times thetick
handler is run if appropriate. - Use the background component instead of
a-sky
to define a solid color as the scene background. This prevents the creation of unnecessary geometry. - Update
position
,rotation
,scale
, andvisible
using at the three.js level (el.object3D.position
,el.object3D.position
,el.object3D.scale
,el.object3D.visible
) to avoid overhead on.setAttribute
.
GPU Texture Preloading
Until non-blocking texture uploads to the GPU are available, try to draw all materials and textures up front. When materials and textures are drawn for the first time, the browser will hang and block while uploading to the GPU. We can manually preload textures by calling:
document.querySelector('a-scene').renderer.setTexture2D(ourTexture, 0); |
We will try to come with a convenient API in A-Frame to do this automatically.
For example, this is apparent in the 360° image gallery. If we look at the browser performance tools, there will be frame drops when switching to a new image for the first time, but smooth transitions when switching back to images for the second time.
Reuse materials and textures as much as possible, aiming for a small number of unique materials. Texture atlases provide one efficient way to reuse materials while giving the impression of more variety. Simpler three.js materials such as MeshLambertMaterial or MeshBasicMaterial perform better and are often sufficient for low-poly scenes. In particular, pre-baked lighting on an unlit (Basic) material can significantly improve performance. A-Frame’s default PBR-based (Standard) material is more physically realistic, but also more expensive and often unnecessary in simple scenes.
Tools
There are browser add-ons available to look into what’s being sent to the GPU:
JavaScript
Avoid creating garbage and instantiating new JavaScript objects, arrays, strings, and functions as much as possible. In the 2D web, it is not as big of a deal to create a lot of JavaScript objects since there is a lot of idle time for the garbage collector to run. For VR, garbage collection may cause dropped frames. Learn more about detecting allocations and garbage collection in Firefox and Chrome performance tools.
Try to avoid patterns such as Object.keys(obj).forEach(function () { });
,
which create new arrays, functions, and callbacks versus using for (key in obj)
. Or for array iteration, avoid .forEach
and use a simple for
loop
instead. Avoid unnecessary copying of objects, using methods like
utils.extend(target, source)
instead of utils.clone
or .slice
.
If emitting an event, try to reuse the same object for event details:
AFRAME.registerComponent('foo', { |
All of the suggestions above are especially important in tick()
handlers,
where they will be running on every frame.
tick
Handlers
In component tick handlers, be frugal on creating new objects. Try to reuse objects. A pattern to create private reusable auxiliary variables is with a closure. Below we create a helper vector and quaternion and reuse them between frames, rather than creating new ones on each frame. Be careful that these variables do not hold state because they will be shared between all instances of the component. Doing this will reduce memory usage and garbage collection:
AFRAME.registerComponent('foo', { |
Also if we continuously modify a component in the tick, make sure we pass the same object for updating properties. A-Frame will keep track of the latest passed object and disable type checking on subsequent calls for an extra speed boost. Here is an example of a recommended tick function that modifies the rotation on every tick:
AFRAME.registerComponent('foo', { |
VR Design
Designing for VR is different than designing for flat experiences. As a new medium, there are new sets of best practices to follow, especially to maintain user comfort and presence. This has been thoroughly written about so we defer to these guides. Note that VR interaction design is fairly new, and nothing is absolute:
Some things to note:
- The common golden rule is to never unexpectedly take control of the camera away from users.
- Units (such as for position) should be considered meters. This is because the WebVR API returns pose in meters which is fed into most camera controls. By considering units as meters, we achieve expected scale.