Skip to content

Latest commit

 

History

History
50 lines (38 loc) · 2.45 KB

TODO.md

File metadata and controls

50 lines (38 loc) · 2.45 KB
Raw

TODO

A list of upcoming changes and features.

Fix 3DO

Each class is serializable, also expand that to vectors (e.g. save __class = "vec3")

Improve Doc

Light sources and a lot of other classes completely missing

Improve godrays

The current light disk approach performs barely acceptable, but can be further enhanced:

  • Tweak the disk, make rays texture less sharp
  • Instead of calculating pixels on the entire screen, convert the disks border into a radian stripe and only calculate that
  • Then use this 1D stripe for the rest of the screen, lookup using atan, no godray within the disk
  • This lookup table is generated for each light source (e.g.: 8px output), and then applied all at once, thus allowing easy multi source godrays if required

Performance

  • Finish LODs + example
  • One more attempt on transformation caching
    • If we had a scene tree, transformations could be cached
    • This only works if all objects are unique, which instantiate ensures
    • Cache transform, position and scale for objects and meshes by using a weak parent table .cache
    • All operations on objects now need to use the adder and remover functions to ensure integrity
      • Protect objects and meshes with metatable just to make sure
  • Add an experimental freezeShader function
    • The only way for the shader to change is...
      • Changes to the material, reflection, mesh, global shader and scene settings
      • A cache would add a subcache to the shader cache and is slightly faster at the expense of potential missed update errors
  • Threaded scenes
    • A scene would need a perfect reconstruction of all transforms and bounding boxes on a thread
      • That would work with the assumption of uniqueness from before!
    • The sync happens via a shared memory segment
    • The thread traverses the tree, places the transforms, does frustum culling, lods and sorting, and provides an argsort list.
    • The main thread syncs back with the thread and uses the array
    • Only issue: the main thread needs to do other work in the meantime.
      • The smoothest way are coroutines, render just adds its 1 or 2 (with alpha) scenes to the queue, together with the args
      • Once the thread reports back, finish that coroutine, in order to maintain reflections
      • In best case scenario, with 2 shadowed light sources, 4 threads can be used

Reflections

  • Non axis aligned box reflections
  • Easier to use reflection globes
  • Globe blending and proper multi globe