Suboptimal performance with EXT_mesh_gpu_instancing

[zeux]

I've been testing instancing using models provided in KhronosGroup/glTF#1699; attached is the ice stadium model encoded using EXT_mesh_gpu_instancing.

stadium.glb.zip

After conversion this scene has 583 nodes with a total of 610 mesh primitives, but many of these are instanced; the cumulative instance count is around 27K.

three.js has an open PR for support of this extension; using that, it takes ~140ms to fully deserialize the file including mesh parsing etc., and ~3ms each frame to submit draw commands to the GPU, with scene looking a bit like this

In comparison, Babylon.JS takes ~16 seconds (~114x longer) to deserialize the file, and ~55ms (~18x longer) to render each frame on the CPU. The GPU cost looks comparable, maybe better on Babylon side.

Here's a breakdown of loading process, with a separate section for createDefaultCamera (it seems to run into some sort of N^2 behavior?), and a few other expensive bits:

Here's a breakdown of a single render frame:

The most severe issues here are the problem with createDefaultCamera and the render frame performance. I'm wondering if Babylon.JS provides a more efficient way to represent instanced draws? Right now it looks like the loader just creates a node per mesh, and presumably then relies on auto-instancing functionality to dispatch draw calls more efficiently, but it looks like the scene graph checks are not very fast here.

[deltakosh]

Thanks a lot! We are currently working on improving (drastically) the situation. We will report here asap

[Popov72]

Fixed by #8208 and #8209.

Let us know how it comes for you.