Re: [Algorithms] Multipass alpha-blended objects

[Wesley H. <we...@fu...>]

----- Original Message ----- 
From: "Miles Macklin" <gd...@co...>
>
>  > 2. Material color was very expensive to generate, so doing it once for
> alpha
> > and again for color was not acceptable. Some materials were themselves
> > multiple passes, so that was basically impracticle.
>
> Sorry I'm not sure I follow, why would you generate the material colour
for
> alpha and then generate it again for colour ?  But I understand if you
have
> a material colour that is very expensive and you need access to it in
> multiple passes you probably want a temporary copy.
>
Yeah, I meant your second point. Material color/alpha are generally the same
computation, so I'd have to do all the work to seed the framebuffer dest
alpha, then do it again when I apply the material color. Not impossible, and
at least you get the "correct" answer when doing it this way.


> I can't see why doing a bunch of additive passes followed by a green
overlay
> pass (for example) followed by a fog pass doesn't work.  In fact thats
> exactly what we do, it just breaks down to:
>
> result = (((light0 + light1 +
> light2)*green)*(1-fogIntensity))+(fogIntensity*fogColour)
>
Yeah, for opaque objects or offscreen buffers, it is easy and this works.
But think about the alpha-blended case we are talking about. For each pass,
you only have:
    * matalpha (in dest alpha)
    * srccolor (from shader)
    * accumulated color (in frame buffer)

Say you had only 2 additive passes, you'd want:

    (light0 + light1) * matalpha + fbcolor * (1-matalpha)

so split it into passes:

    pass 1: light0 * matalpha + fbcolor * (1-matalpha)
    pass 2: light1 * matalpha + 0

This works, but now add material modulation. You'd want:

    [(light0+light1) * matcolor] * matalpha + fbcolor * (1-matalpha):

try to split that into passes:

    pass 1: light0 * matalpha + fbcolor * (1-matalpha)
    pass 2: light1 * matalpha + 0
    pass 3: MatCol * matalpha * fbcolor * ???

There's no way to do this correctly. Yeah, add fog or other alpha blend
types and it's "even more impossible." It's impossible to construct an
expression tree using these inputs without requiring a temporary (ie, an
offscreen buffer).

-Wes