Use ANGLE_instanced_arrays extension to reduce buffer sizes for layers with property functions

[mourner]

https://www.khronos.org/registry/webgl/extensions/ANGLE_instanced_arrays/

According to this presentation about WebGL2, this feature (available for WebGL 1.0 as extension) can lead to significant performance gains. Interesting bit:

Google Maps uses and requires instancing support — Reduces vertex buffer size by 6x (!)

The concept is:

To draw 1000 soldiers
1000 draw calls, with different position, posture, etc
Or, with instancing, only 1 draw call
Big perf improvement

cc @ansis @kkaefer @jfirebaugh @lucaswoj what do you think? Can we use it?

Also related: #1361 (mentioned in the slides as a notable perf feature too)

[lucaswoj]

If you are fill-rate bound or vertex-bound then instancing won't solve your problems. But if your seeing your app get hung up on making lots of draw calls while your GPU spins it's wheels waiting for Javascript to catch up, this is the API for you!

My impression is that it'd be worth prototyping and benchmarking a ANGLE_instanced_arrays-based rendering path for symbols, as they are nearly identical geometries, drawn multiple times, with multiple render calls.

I'm not clear if there's a way we could use this API to reduce the # of WebGL API calls for any other layer types.

[lucaswoj]

I wonder how exactly Google Maps uses this feature...

[ansis]

It could be very useful for reducing buffer sizes for data driven styling.

For example, if you had per-feature colours you could using instancing to store the colors per-triangle instead of per-vertex and use 1/3 the space. (I think you could potentially save even more space by making an instance 2 triangles and adding degenerate triangles to polygons with an odd number)

According to webglstats.com support is around 90% in browsers. I think all iOS devices support it. It's an extension on Android but it's hard to get stats on how well it's supported. If we use it we'd have to accept supporting less devices (what % less?) OR having two separate paths/a polyfill (how bad would this be?).

[mourner]

If GMaps requires it, I'd assume that support is good enough for us to require as well.