This tutorial will help you find some links and info on how you can improve your scene regarding rendering performance.
If you need node containers or transform nodes, do not use meshes but TransformNode instead. Use meshes only when associated with content to render.
The meshes need to go through an evaluation process where the camera checks if they are in the frustum. This is an expensive process so reducing the number of candidates by using TransformNode when possible is a good practice.
Babylon.js uses an advanced and automatic shaders engine. This system will keep shaders up to date regarding material options. If you are using a static material (ie. an immutable material) then you can let it know to Babylon.js by using the following code:
Once frozen, the shader will remain unchanged even if you change material's properties. You will have to unfreeze it to update the inner shader:
Every mesh has a world matrix to specify its position / rotation / scaling. This matrix is evaluated on every frame. You can improve performances by freezing this matrix. Any subsequent changes to position / rotation / scaling will then be ignore:
You can unfreeze a mesh with:
If you are CPU bound, you can decide to keep the list of active meshes unchanged and then free the time spent by the CPU to determine active meshes:
You can unfreeze the active meshes with:
Note that you can force a mesh to be in the active meshes before freezing the list with
mesh.alwaysSelectAsActiveMesh = true.
Freezing active meshes list may cause some things on the scene to stop updating. One example is RenderTargetTexture, if it's used by mesh materials. For that case RTT needs to be explicitly added to the list of active camera's custom render targets, which will guarantee that it ends up in the render list of the scene:
In conjonction with
mesh.alwaysSelectAsActiveMesh you can also decide to turn off bounding info synchronization. This way the world matrix computation will be faster as the bounding info will not be updated (this could be a problem if you want to use picking or collisions):
mesh.doNotSyncBoundingInfo = true;
As soon as you can please use instances as they are drawn with one single draw call.
If sharing the same material is a problem, you can then think about using clones which share the same geometry with
One remark regarding instances: If one of the instances has a world matrix with a different determinant (eg. one instance has a negative scale where others don't), babylon.js will be forced to remove the back face culling from their material.
By default, Babylon.js automatically clears the color, depth, and stencil buffers before rendering the scene. It also clears the depth and stencil buffers after switching to a new camera and before rendering a new RenderingGroup. On systems with poor fill rates, these can add up quickly and have a significant impact on performance.
If your scene is set up in such a way that the viewport is always 100% filled with opaque geometry (if you're always inside a skybox, for instance), you can disable the default scene clearing behavior with:
scene.autoClear = false; // Color buffer scene.autoClearDepthAndStencil = false; // Depth and stencil, obviously
If you know that the geometry in a particular RenderingGroup will always be positioned in front of geometry from other groups, you can disable buffer clearing for that group with the following:
scene.setRenderingAutoClearDepthStencil(renderingGroupIdx, autoClear, depth, stencil);
true to enable auto clearing. If
depth: Defaults to
true to enable clearing of the depth buffer
stencil: Defaults to
true to enable clearing of the stencil buffer
Go ahead and be aggressive with these settings. You'll know if it's not appropriate for your application if you see any smearing!
When dealing with complex scenes, it could be useful to use depth pre-pass. This technique will render designated meshes only in the depth buffer to leverage early depth test rejection. This could be used for instance when a scene contains meshes with advanced shaders.
To enable a depth pre-pass for a mesh, just call
mesh.material.needDepthPrePass = true.
By default Babylon.js uses indexed meshes where vertices can be reuse by faces. When vertex reuse is low and when vertex structure is fairly simple (like just a position and a normal) then you may want to unfold your vertices and stop using indices:
For example this works very well for a cube where it is more efficient to send 32 positions instead of 24 positions and 32 indices.
By default, Babylon.js does not adapt to device ratio anymore. It by default focuses on perf vs quality after receiving lots of community requests.
The drawback is that this could look low rea. You can turn it on with the fourth parameter of the Engine constructor:
var engine = new BABYLON.Engine(canvas, antialiasing, null, true);
By default the scene will keep all materials up to date when you change a property that could potentially impact them (alpha, texture update, etc...). To do so the scene needs to go through all materials and flag them as dirty. This could be a potential bottleneck if you have a lot of material.
To prevent this automatic update, you can execute:
scene.blockMaterialDirtyMechanism = true;
Do not forget to restore it to false when you are done with your batch changes.
Babylon.js processes speed depending on the current frame rate.
On low-end devices animations or camera movement may differ from high-end devices. To compensate this you can use:
The return value is higher on low frame rates.
Starting with version 3.1, Babylon.js can handle WebGL context lost event. This event is raised by the browser when the GPU needs to be taken away from your code. This can happen for instance when using WebVR in hybrid scenario (with multiple GPU). In this case, Babylon.js has to recreate ALL low level resources (including textures, shaders, program, buffers, etc.). The process is entirely transparent and done under the hood by Babylon.js.
As a developer you should not be concerned by this mechanism. However, to support this scenario, Babylon.js may need an additional amount of memory to keep track of resources creation. If you do not need to support WebGL context lost event, you can turn off the tracking by instantiating your engine with doNotHandleContextLost option set to true.
If you created resources that need to be rebuilt (like vertex buffers or index buffers), you can use the
engine.onContextRestoredObservable observables to keep track of the context lost and context restored events.
If you have a large number of meshes in a scene, and need to reduce the time spent when adding/removing those meshes to/from the scene, There are several options of the
Scene constructor that can help :
truewill speed-up the addition and removal of
Geometryin the scene.
truewill speed-up the disposing of
Materialby reducing the time spent to look for bound meshes.
truewill speed-up the disposing of
Meshby reducing the time spent to look for associated cloned meshes.
For each of this options turned on, Babylon.js will need an additional amount of memory.
Also, If you are disposing a large number of meshes in a row, you can save unnecessary computation by turning the scene property
blockfreeActiveMeshesAndRenderingGroups to true just before disposing the meshes, and set it back to
false just after, like this :
scene.blockfreeActiveMeshesAndRenderingGroups = true; /* * Dispose all the meshes in a row here */ scene.blockfreeActiveMeshesAndRenderingGroups = false;
The culling is the process to select whether a mesh must be passed to the GPU to be rendered or not. It's done CPU side.
If a mesh intersects the camera frustum in some way then it's passed to the GPU.
Depending on its accuracy (checking mesh bouding boxes or bouding spheres only, trying to include or to exclude fast the mesh from the frustum), this process can be time consuming.
In the other hand, reducing this process accuracy to make it faster can lead to some false positives : some meshes are passed to the GPU, are computed there and won't be finally visible in the viewport.
By default, BABYLON applies the most accurate test to check if a mesh is in the camera frustum.
You can change this behaviour for any mesh of your scene at any time (and change it back then, if needed) this the property
/** * Possible values : * - BABYLON.AbstractMesh.CULLINGSTRATEGY_STANDARD * - BABYLON.AbstractMesh.CULLINGSTRATEGY_BOUNDINGSPHERE_ONLY * - BABYLON.AbstractMesh.CULLINGSTRATEGY_OPTIMISTIC_INCLUSION * - BABYLON.AbstractMesh.CULLINGSTRATEGY_OPTIMISTIC_INCLUSION_THEN_BSPHERE_ONLY */ mesh.cullingStrategy = oneOfThePossibleValues;
Optimistic Inclusion modes give a little gain. They keep the same accuracy than the basic mode on what they are applied (standard or bSphereOnly).
BoundingSphereOnly modes, because they reduce a lot the accuracy, give a good perf gain. These should not be used with high poly meshes while sending false positives to the GPU has a real rendering cost. These can be very interesting for numerous low poly meshes instead. *Really useful if you are CPU bound**.
Instrumentation is a key tool when you want to optimize a scene. It will help you figure out where are the bottlenecks so you will be able to optimize what needs to be optimized.
The EngineInstrumentation class allows you to get the following counters:
instrumentation.captureGPUFrameTime = true.
instrumentation.captureShaderCompilationTime = true.
Here is an example of how to use engine instrumentation: https://www.babylonjs-playground.com/#HH8T00#1 -
Please note that each counter is PerfCounter object which can provide multiple properties like average, total, min, max, count, etc.
GPU timer require a special extension (EXT_DISJOINT_TIMER_QUERY) in order to work. This extension has been disabled due to Spectre and Meltdown on all major browsers. This is still possible to use by enabling the flag gfx.webrender.debug.gpu-time-queries on firefox at the moment. This should be re-enabled soon in the browsers.
The SceneInstrumentation class allows you to get the following counters (per scene):
instrumentation.captureActiveMeshesEvaluationTime = true.
instrumentation.captureRenderTargetsRenderTime = true.
instrumentation.captureFrameTime = true.
instrumentation.captureRenderTime = true.
instrumentation.captureInterFrameTime = true.
instrumentation.captureParticlesRenderTime = true.
instrumentation.captureSpritesRenderTime = true.
instrumentation.capturePhysicsTime = true.
instrumentation.captureCameraRenderTime = true.
Those counters are all reset to 0 at the beginning of each frame. Therefore it is easier to access them in the onAfterRender callback or observable.
When using BabylonJS with WebVR or WebXR, enabling Multiview is a quick way to almost double the rendering speed.