These are the states of the various algorithms I’ve been working at porting to Unreal. I’m sampling a 128x128x128 volume in this picture;
Marching cubes serial only works as of yet
Naive surface nets serial, parallel and vectorized for different archs*, and a GPU port*
Dual contouring serial only works as of yet
Still working on the other supposed faster polygonization methods such as FlyingEdges and Contour3DLinearGrid
Also still working on CMS, not sure why at this point, but might be interesting
*NSN gpu has a small bug that I’ll probably work out later today, however it’s the fastest of any polygonization method I’ve encountered. I run two passes through a custom compute shader. 95% of the 1-3ms time(for a 128x128x128 volume) is spent on memory passing, if someone kept the vertices on the GPU, this would be lightning fast. Definitely fast enough for realtime deformable meshes, I imagine even on older mobiles with ES 3.1+. I could port this to a vertex or fragment shader if someone needed to target older devices ES3.0, or WebGL.
NSN cpu threaded is slower than serial in this picture mostly because I have not optimized this yet, wait for this to get faster too.
If anyone is interested in any of these let me know and I can continue working on them to release.
If anyone knows of a faster method of polygonization without going overboard on triangles, in Unreal or not, please let me know.