While trying to reduce the overhead of the mesh batch cache generation, I came to the conclusion that we tried too much to be efficient and became less efficient by doing so.
Basically we have 2 ways of drawing indexed geometry and if we use both, we double the cost of VBO filling.
So this task is about making all VBOs and IBOs (inside the Mesh Batch cache) use vert loops indexing.
The only issue is making sure Per vertex (not per loop) attributes are copied to all loops.
This could be the opportunity to unify all VBO generating functions and parallelize attributes filling.
Also a bit of time is spent on memcopy form the driver when uploading the data. We could try to use mapped memory instead (adding an exception in GPU).