Quick thoughts on Virtual Textures

1) I’d assume it wouldn’t be a bad idea to lay out the individual pages along a Hilbert or Peano curve, ensuring locality. Possibly one per MIP level?

2) Determining what to page in - shouldn’t we just do that one frame delayed? Operate on the current set of paged in tables, use the result of current render for next set of tables.

3) Ideally, that would mean no separate render pass to determine what to page in.

4) Having a full-screen render buffer to determine page misses seems… bad. GPU quad trees might be nice to look at. Or any other spatial query structure.