Instanced block rendering

[kpreid]

As of b4d42702f9fa7ed8d22d7ae3b1efc75388cca7e5, there is just-barely-working support for rendering Space contents using instanced block meshes instead of chunk meshes. This already provides great performance improvement for rendering — at least on desktop — but there's a lot of pieces left to do.

• Performance:
  • [ ] Investigate possible performance regression on web. (Not sure if it's real. If it is, it might be just the total number of WebGL state changes and draw calls increasing given the demo city's large number of distinct blocks — not sure what to do about that.)
  • [X] Optimize instance list data structure management to reduce the repeated work each frame.
  • [ ] Tune the threshold at which instances are used.
• Applicability:
  • [X] Use AnimationHint to decide whether to use an instance. (Done in 5dcfa46d2b0fdcb9efe88f398a2e5f4cf688d094.)
  • [X] When a change in the world only affects instances, only update instances. (Done in 7a16b293f89a8d577e12562abf3eb0b847b0b723, 434fc90b9dfee95afd042af97c71b01cc4124acb)
  • [ ] Use instances in glTF export.
  • [ ] When a change in the world occurs, consider just adding/replacing one instance instead of remeshing the containing chunk, even if we wouldn't use instances from scratch. (Blocked on teaching glTF about instances.)
    • [ ] If a chunk experiences lots of changes and then quiesces, consider turning its instances into a chunk mesh.
  • [ ] Decide what to do about rendering transparent blocks — perhaps they should just continue to be chunked and rendered using chunk depth-sorting (as is the status quo).
• Technical debt:
  • [X] In the wgpu renderer, we're not picking the size of the instance buffer to correctly account for all these block instances. It's working anyway. Why? There must be some overallocation. (Fixed in 51d9197916c0546bb050e6303e7b1c953b5ab2e5)
  • [x] Add tests in the mesh crate for instance behavior. Right now all the existing tests continue to pass because the test MeshTypes disables instancing.
  • [x] Tidy up sloppy API choices (types for block instance mesh and ChunkedSpaceMesh instance list).
  • [ ] Ensure all TODO(instancing) comments have been dealt with.

[kpreid]

In the wgpu renderer, we're not picking the size of the instance buffer to correctly account for all these block instances. It's working anyway. Why? There must be some overallocation.

This turned out to be overallocation because we allocate for all chunks and only push culled chunks, so it could fail given enough block instances. Fixed in 51d9197916c0546bb050e6303e7b1c953b5ab2e5.

[kpreid]

a0c547497ec9417ef161208a7c8c17c1041e9f53 fixes some of the mesh API and some of the inefficiency; it is now up to the user of ChunkedSpaceMesh to gather instances (so they can do it as part of the iteration they already must do), and none of the internal hash tables are exposed.

[kpreid]

Add tests in the mesh crate for instance behavior.

There are some tests now, added along with specific improvements. I'll call this done-enough-not-to-track even though there's more work ahead.

When a change in the world occurs, consider just adding/replacing one instance instead of remeshing the containing chunk.

This is partly supported via 7a16b293f89a8d577e12562abf3eb0b847b0b723 — instances are never used for blocks that wouldn't be instanced, but instance-only changes should skip the mesh rebuild. However, it doesn't work 100% of the time, for reasons I have not yet diagnosed.

Also of note for performance: 7627d9e630c3bd054ead324c7195999269528a3c added InstanceMap specifically so that removals were possible, but the strategy in 7a16b293f89a8d577e12562abf3eb0b847b0b723 is to sweep the entire chunk, so that we don't have to keep or update a complex todo data structure. So, one or the other should change.

[kpreid]

434fc90b9dfee95afd042af97c71b01cc4124acb makes instance-only updates work consistently.