Re: [Algorithms] VIPM With T&L - what about roam

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

Martin Fuller wrote:

> VIPM is great for the reasons outlined below, you always transform used
> verts
> and no more. The problem with progressive meshing is that because it is view
> dependant it becomes very expensive.

In this case we can just chop the vipm'ed model into many smaller vipm segments
thereby approximating a VDPM but using a far smaller amount of cpu time.

> Here's a question though. If I want to VIPM a heightfield is there anyway to
> maintain a minimal indexed triangle strip. By minimal I mean as few strips
> and as few degenerate triangles as possible. A plain heightfield strips very
> well obviously with one strip per tile, (Assuming a quadtree split) with best
> case 1 degerenrate triangle per coulomb.

Hmmm... Triangle strips and VIPM tend not to mix very well, although having
said that there was talk a while ago about using degenrate tris to maintain
stip-ability through edge-collpases and all. Maybe someone else can go over
that idea again? :-)

> It's all swings and roundabouts, there are cruder LOD techniques for
> heightfields which strip very nicely so I suppose it depends on whether the
> bottleneck
> becomes the transforms and extra vertex data or the actual polygon drawing. My
>
> current thinking on PC and PS2 is that it's probably better to maintain longer
>
> strips and draw more polys. Any thoughts?

With gfx cards getting faster and faster there is less and less time to worry
about
what to draw, so we keep needing faster algos ;-)

I was thinking about a kind of ROAM on the PS2 the other day. If sending things
to the VU is the bottle-neck, then perhaps sending a bitmap and triangulating
that
in the VU is a good idea? As long as you spend hardly any time managing the
bitmap, I suppose....

Cheers, Paul.