RE: [Algorithms] Game engine question

[Stefan B. <ste...@di...>]

  The old scene graph discussion! I think part of the problem is that people
have different ideas of what a "scene graph" really is. Many think it's the
uber-structure which is used to represent the entire world, some (including
myself) have more general definitions.
 
  A problem with many generic scene graph implementations is that in
practice, they simply perform worse than rendering systems with more
hard-wired behaviour. This is especially true for the PS2, where memory
access is something you really want to avoid. "Pure" OO designs often
perform like crud compared to more pragmatic approaches simply because with
OO you tend to group data in ways that make sense from a purely structural
perspective instead of a performance perspective.
 
  A prime example would be your reference to having a node per bone in a
skeleton. This does not perform well at all and offers little, if any,
benefit. You're much better off describing your skeleton hierarchy using an
array of transformation matrices and an array of indices describing the
relationships between these transformations.
 
  Another really bad idea is having hierarchical state propagation. For
example, having some kind of nodes for controlling whether backface culling
is on or off, which affects all nodes below it. This "feels" powerful but
just ends up complicating and slowing the system down by having to maintain
state stacks etc, and it's rarely useful in a real-life scenario.
 
  Well, I could go on forever actually but I have to go. This is more of a
SWEng-type discussion anyway really.
 
  In the end I think this is all about being pragmatic and not getting
carried away and over-generalizing systems if it has little benefit. The
problem is that often you won't know this until you tried both approaches ;)
 
/Stefan
 
Stefan Boberg
Chief Technical Officer
Digital Illusions CE AB