Open sjkelly opened 5 years ago
Might be worth checking out julia master with the nrw threading architecture!
I am building 1.2 now. Some rough changes are on this branch: https://github.com/JuliaGeometry/Meshing.jl/tree/sjk/threads
I think I am running into the closure inference bug on 1.1 because an Int
declaration seems to be getting boxed and is triggering the GC quite a bit. Hopefully 1.2 helps. I have seen ~12x performance improvements on the distance field sampling side with threads and SIMD, but meshing is still the performance bottleneck in my applications.
Alternatively I may port a GPU example and use CUDANative and/or OpenCL.
From what I can tell this should be possible in at least MarchingCubes and NaiveSurfaceNets. There are some other optimizations from the NaiveSurfaceNet implementation I would also like to apply to MarchingCubes and Marching Tetrahedra.