Re: [Algorithms] fast subdivision

[Paul F. <pf...@at...>]

Nb. Sorry if this post arrives twice, I seem to be having
trouble posting to this list :-(

Tom Forsyth wrote:

> That's pretty much the full info on N-Patches. With that info, you have a
> triangular bicubic Bezier patch, and there are gazillions of papers on
> subdividng them :-)

Ok...

> On things like the PS2, you have more flexibility, and it is perfectly
> possible to do the whole thing on the "GPU".

Yeah, I was mainly thinking of the PS2 given that you really need to give
the vector units as small as possible set of data and generate a bin-load
of tris from it.

I wanted to use Loop or Catmull-Clark as they produce really "nice"
surfaces but it looks like the calculation overhead is going to be too
great. It would be ideal if there were a parametric description for them,
but I fear that is almost impossible without fudges.

I was thinking of maybe using Catmull-Clark patches since most of a
model will consist of regular tensor-product b-spline patches and only
a small %age will be non-regular. If that small %age could be rendered
using a different technique, it might be a win - as long as the join was
seamless...

> For the PS2, similar schemes look perfectly feasable, though with the extra
> flexibility of the PS2 GPU, and its impressive triangle throughput, you may
> be able to move even more of it onto the GPU, and not bother with VIPM
> nearly as much - just draw everything. I have yet to experiment to find out
> which is better in different situations.

I should imagine that just drawing everything using a high tesselation
technique would be a win for PS2. Maybe using a VIPM-like
technique to somehow reduce the number of active bones in
sub-d SSC mesh based on some error metric could be cool too...

Cheers, Paul.