And still no Shakespeare... |
Yeah, okay, me neither. (At least, not until I put together the demo for this post.)
But there's certainly occasions when it makes a lot of sense to redraw the same mesh many times with some minor tweak like position, orientation, color, etc. Simple examples of this are trees and other vegetation, streetlights, boxes in a warehouse, soldiers in an army, and so on.
Traditionally with WebGL the optimal way to draw those repeated meshes, commonly referred to as instances, would be something like the following pseudocode:
bindMeshArrays(gl);
for (i = 0; i < meshInstances.length; i++) {
var instance = meshInstances[i];
gl.bindUniform3fv(meshPosition, instance.position);
gl.bindUniform4fv(meshColor, instance.color);
gl.drawElements(gl.TRIANGLES, indexCount, gl.UNSIGNED_SHORT, 0);
}
This is good, because we only set up the vertex attributes (implied to be in bindMeshArrays) once, then make many draw calls in quick succession, only changing the properties that are different from instance to instance. In this case just position and color. This makes for a near minimal amount of overhead for each draw call, and in an environment like Javascript where each call is expensive that's important!
(This example is pretty simple, and in fact you can actually optimize it further by getting creative with how you provide the GPU with the data. Check out Gregg Tavares' wonderful Google I/O talk for more ideas about how to render lots of geometry very quickly.)
Even in this scenario, however, you have to make at least three WebGL calls for every instance of the mesh. That doesn't sound terrible but in the real world that would probably be quite a bit more, and the fact is that in Javascript every call into native code (Like the WebGL API) carries a certain expense. Thus while drawing repeated meshes in this way works, and generally works well, it could still be better.
And that's exactly why we have ANGLE_instanced_arrays.