"Bruno Cabasson" <bru### [at] alcatelaleniaspacefr> wrote:
> thus avoiding useless tree-traversal and intersection
> test?

I think it's the wrong approach. I'd instead suggest an investigation along
these lines:

1. Is intersection-finding a major bottleneck in rendering "standard"
scenes?
2. Is POVRay's spatial partitioning algorithm far less than optimal?
3. Is POVRay's implementation of the spatial partitioning algorithm
inefficient?
4. If 1, 2 and 3 are true, which of the many well-established,
well-researched techniques in the published literature (much of which is
freely available on line) could be applied to resolve the matter?

If something besides intersection-finding slows down most scenes, then
clearly we can stop worrying about it right now. However, for that to be
the case would be a great surprise to me unless there is something
seriously wrong with the rest of the code.

Since (as far as I can see) the partitioning method used in 3.6 and earlier
is BVH, the answer to 2 is almost certainly yes. I think I saw something
about 3.7 supporting another subdivision method though.

I've been finding POV's BVH implementation quite difficult to read, so I
don't want to say anything about 3 yet.

My point with 4 is that we really don't need to spend effort thinking of
entirely new algorithms as in the first part of this thread. There are
loads that we could try. For example, ompf have a good collection of links,
most of which describe algorithms (ignore the hardware-specific stuff, they
are focussed on realtime):

http://ompf.org/forum/viewtopic.php?t=9


Thanks,

Tom

Post a reply to this message