NMG Raytracing Performance Improvement

From BRL-CAD

BRL-CAD uses two primary primitive types to represent mesh data - N-Manifold Geometry (nmg) and Bags of Triangles (BoTs). In principle, all planar nmg surfaces can gracefully downgrade to a BoT representation without introducing geometric changes to surfaces (indeed, this should be VERY straightforward, but a few bugs remain in BRL-CAD's implementation of that process.)

Recent work on raytracing of BoT models has yielded significant speed-ups for complex models, and BoT raytracing has always been faster than direct raytracing of nmg data. There are implementation issues with nmg raytracing, but an interesting alternative to directly raytracing nmg primitives would be to raytrace their BoT conversion. The actual raytracing code does not require the topological sophistication of the nmg representation (that's used elsewhere) so results should be identical to within floating point noise. This would allow one common raytracing code for nmg and bot primitives, simplifying the code and making improvements more general.

This task would involve extensive testing of nmg->bot conversion logic, setting up methods to allow the raytracer to work on a BoT representation of the nmg (need to consider approaches - can the raytracer call the nmg->bot routine on a per surface basis on the fly, does it need to pre-generate the whole bot, etc). It may very well be that the project would find fixing direct nmg raytracing is the best option for performance improvement, in which case the task would be to make those fixes.

References

  • src/librt/primitives/nmg
  • rt
    • our main raytracer, renders pictures
  • db
    • lots of sample geometry
  • mged 'facetize' command
    • useful for turning any geometry into NMG representation

Requirements

  • Familiarity with C
  • Good debugging/troubleshooting skills