Questioning LOD

1
Questioning LOD

• Jonathan Blow
• Bolt Action Software
• jon_at_number-none.com

2
What is LOD?

• Complex models contain too many triangles to
  render quickly.
• Most of the triangles will not add much to the
  appearance of a scene.
• We can construct approximations to those models
  and render them instead.
• LOD stands for Level of Detail.

3
Types of LOD management

• Discrete LOD
• Rigid progressive meshes
• Malleable progressive meshes
• View-dependent mesh manipulation

4
Examples of LOD
5
Examples of LOD
6
Questioning the basic assumptions

• Complex models contain too many triangles to
  render quickly.
• Most of the triangles will not add to the
  appearance of the scene.
• We can construct approximations to those models
  and render them instead.
• Geometry management is an isolated problem.

7
Complex models contain too many triangles to
render quickly?

• Graphics card makers (e.g. nvidia) are giving us
  hardware with huge throughput potential (tens of
  millions of polygons / sec)
• You can only come near the marketing numbers if
  your data is rigid.
• Most types of LOD methods cripple the rendering
  system.

8
Questioning the basic assumptions

• Complex models contain too many triangles to
  render quickly.
• Most of the triangles will not add to the
  appearance of the scene.
• We can construct approximations to those models
  and render them instead.
• Geometry management is an isolated problem.

9
Most triangles will not add to the scene?

• ...but you spend CPU figuring out which ones.
• The set of desired triangles can change
  frequently
• requiring a lot of data structure updates
• disrupting rigidity that the graphics card
  likes.

10
Questioning the basic assumptions

• Complex models contain too many triangles to
  render quickly.
• Most of the triangles will not add to the
  appearance of the scene.
• We can construct approximations to those models
  and render them instead.
• Geometry management is an isolated problem.

11
We can construct approximations?

• Approximation algorithms are based on heuristics
  that often go awry.
• The approximations require more storage than the
  original model.
• This kind of code is complex and hard to write.
• It is hard to know when the code is correct.

12
Questioning the basic assumptions

• Complex models contain too many triangles to
  render quickly.
• Most of the triangles will not add to the
  appearance of the scene.
• We can construct approximations to those models
  and render them instead.
• Geometry management is an isolated problem.

13
Geometry management is not isolated from the rest
of your system.

• We must texture and light that geometry.
• We must push it through the rendering manager
  (vertex buffer manager, shader organizer)
• It must cooperate with unpredictable portions of
  your game engine.

14
Examples Terrain Lighting

• Vertex removals hate vertex-based lighting
  popping happens.
• Lightmaps solve this, but cost an extra texture
  pass.
• Oh wait! We want to bump-map now.
• Bump frames are pinned down by vertices...
• Were screwed unless we have really advanced
  dependent texture reads.

15
Examples Terrain Texturing

• Triangles cant cross texture boundaries.
• Unless we have an intricate texture LOD system,
  our geometric LOD is stifled.
• May involve dynamic texture generation or reading
  from a large disk store.
• Requires us to switch the textures on mesh
  triangles dynamically (this is not free!)

16
Examples Character Animation

• Skeletal animation is about weighted vertices
  attached to bones.
• If you create new vertices you need to synthesize
  new weights.
• If you dont create new vertices you will have
  weird stretching problems. (e.g. shoulder)
• It is difficult to know what the body will do in
  advance.

17
Massive scalabilityis a myth.

• Case study Shinys Messiah
• Game design imposes scaling limitations.
• Many developers consider 4-6x triangle count to
  be the maximum effective scalability
• 6x polygon count does not mean 6x better looking

18
But researchers say LOD is good!

• Many researchers have low expectations that are
  easily satisfied.
• Most printed research is targeted at old
  hardware.
• Most researchers dont have a clear idea of what
  is important to games.
• Most research is not used in production systems
  that have to work solidly.

19
Recommendations

• Terrain, far viewing distance Use
  Lindstrom-Koller or Isosurfaces.
• Terrain, short distance Use static-mesh TINs.
• Static Models Use VIPM.
• Animated characters Use no LOD if you can
  possibly get away with it.

20
Questioning LOD
Jonathan Blow Bolt Action Software jon_at_number-none
.com