Support huge point clouds in the viewer

[emilk]

The Rerun Viewer is currently limited to only being able to display ~two million points at a time, and will have pretty terrible frame rate at 1 million points already.

The reason for this is that we upload the point cloud every frame. This is very simple, but obviously very wasteful.

A solution is to detect if the same points are being rendered this frame as the previous, and if so not re-upload them. This is similar to how we already handle tensors.

[Wumpf]

related to

• Remove limit on line and point count #3076
• Use (thick) postprocessing driven outlines for highlight selections #889
• Optimize point clouds #850

[EriKWDev]

I have been tackling rendering of point clouds for some time and here are some resources I've used to speed up rendering of point clouds:

• Only partially upload the point cloud and render it progressively using random or weighted subsets of the cloud: https://publik.tuwien.ac.at/files/publik_282669.pdf

• Using spacial hashing the cloud could be batched into smaller subsections on which various LOD techniques could be used: https://dl.acm.org/doi/pdf/10.1145/3543863 with related http://www.crs4.it/vic/data/papers/spbg04-lpc.pdf

While a bit specific to Unreal Engine 4, there are some insights in this master thesis project ( https://cgvr.cs.uni-bremen.de/theses/finishedtheses/point_cloud_rendering_in_unreal/thesis_FINAL_WEB.pdf ) about point order

[Wumpf]

We should be a bit more specific with the "huge" part. From what we learned so far most users are more than happy in the region of <8mio points which seems to be in the region where we don't need acceleration structures for rendering.

We need typically 24bytes per point right now (position, radius, color, index - could compress away the index!), so 8mio points would be merely 184Mb.
Given that upload speeds can be assumed to be at the very least 16GB/s (that's PCIe 3.0 16x; 5.0 is common and has 64GB/s and integrated GPUs are off the chart there) we get 273Mb/frame for 60fps.
So even if we upload ever frame we should still hit at least 30fps at that size if we do everything correctly.

Right now we have a hardcoded limited 4mio; point cloud renderer should be dynamic with this!

[abey79]

potree does large point clouds:

• demo: http://potree.org/potree/examples/viewer.html
• repo: https://github.com/potree/potree/
• paper: https://www.cg.tuwien.ac.at/research/publications/2016/SCHUETZ-2016-POT/SCHUETZ-2016-POT-thesis.pdf

[emilk]

The current state is around 1.5M points @30 fps on my M1 MacBook Pro.

We are bounded primarily by

• CPU (upload)
• overdraw when zoomed out a lot

A rough roadmap:

• Remove limit on line and point count #3076 - allowing bigger point cloud (without speeding them up)
• Optimizing our current code - we should be able to get to 2x-5x larger point clouds with a few days work
  • Arrow deser: Add zero-copy happy path #3109
  • Optimize the joining iterator
    • Related: Optimize out unnecessary joins across autogenerated instance keys #3107
• Cache uploads, so we only upload when the points change
  • Will move us completely to be GPU bound (memory and overdraw mostly). Maybe 100M points?
• Acceleration structures, streaming uploads etc: billions of points.

[teh-cmc]

potree does large point clouds:

• demo: http://potree.org/potree/examples/viewer.html
• repo: https://github.com/potree/potree/
• paper: https://www.cg.tuwien.ac.at/research/publications/2016/SCHUETZ-2016-POT/SCHUETZ-2016-POT-thesis.pdf

From the same authors, a drastically different approach by accumulating data across frames, removing the need for hierarchical datastructures, LOD, caching, etc: https://github.com/m-schuetz/Skye

[emilk]

examples/python/open_photogrammetry_format/main.py, showing just the 3D view, on --release, averaged over a few frames, MacBook Pro M1

• 0.8.2, ~47 ms/frame
  

• 0.9.0, 15.3 ms/frame
  

[zeeyang]

What about just rasterize it on the server side and serve as video stream.

[emilk]

Pixel rendering on the cloud is something we'd like to do in the future, but is orthogonal to the issue at hand.