Unfortunately even non-instances will still have to do this matrix multiplication. It could be templated to be skipped when the transform is a unit transform, but I would just be guessing, I haven't profiled to see if it's actually a bottleneck.

This function looks a little harder to parallelize now, I might have to think about restructuring this some.

Did you want to call this geo_groups or something like that?

typo (note)

How about you just remove all the groups from the vector above, that don't have a mesh. Then you don't need to check for it everywhere.

Looks like it would be a good idea to extract this loop to a separate function.

That's a good idea, I'll do that.

Should this be offset + i?

We shouldn't assert that anymore. We might get issues where point clouds have different builtin attributes than mesh.

same here

Is this as_span() necessary?

Don't change the parentheses here, because it will trigger an asan overflow warning sometimes. The addition should be done on unsigned integers (which allow overflow).

Can use a const Map?

Can use uninitialized_fill_n.

Oops, I didn't mean to change that. Good catch.

Oops, that's a typo of set_groups

Yes, I'm fine with removing this message, best to check if there are other places that have similar errors though.