RE: [Algorithms] lights/shadows rendering architecture

[Tom F. <tom...@ee...>]

> From: Marc Fascia
>=20
> 3) 1 pass =3D n lights, shadow mapping done in scrreen space at the =
end
> This is a variation of 2). After the rendereing is done, i=20
> render a shadow
> map for a light and then apply it in screen space. I repeat=20
> that for all
> lights. This also requires n+1 render passes.
> 	+ limit the total ram needed to store shadow maps as they are
> applied one after another

That means you'll have a pipeline doing:

-Render shadowbuffer A
*stall*
-Use shadowbuffer A
-Render shadowbuffer B
*stall*
-Use shadowbuffer B
-Render shadowbuffer C
*stall*
-Use shadowbuffer C
etc.

As you can see, you're going to be stalling the pipeline a lot there. =
Also
be aware that some cards will lose hierarchical/easy-Z testing if you =
switch
away and back from a rendertarget (the framebuffer in this case) without
doing a clear in between, so passes B and C may be much slower. Even =
when
they do work, they tend to have to save/restore a bunch of state.

In general, the memory savings are not worth the effort on PCs. Now =
_maybe_
on a console with a specific arcitecture they would be, but I'd start =
with
assuming you can't, and then maybe you can optimise it later.

> 4) deffered rendering :
...
> 	- does it scales easily to translucent materials ?

No, terribly. I really wouldn't even consider deferred rendering unless
you're going for a fairly constrained world with a very particular look =
-
it's pretty hard. Ignacio mentioned the joys of only shading each pixel
once, and they're appealing, but you can get 90% of that benefit by =
doing
the ambient+unshadowed lighting pass first for the whole scene, and then
doing lighting passes. That way you get nice fast Z-rejects.

I would start with method #1, and maybe later you can hand-compile your =
ten
most popular shaders to use method #2 (if you have enough instruction
slots).


My own take on doing shadowbuffering is in my GDC 2004 "Practical =
Shadows"
notes here. Not directly related to this question, but might be useful:
http://www.eelpi.gotdns.org/papers/papers.html


TomF.