Optimising using normalisation Cubemaps

Is doing a cubemap lookup to normalise vectors much cheaper than doing a texcube sample on iOS? I haven't seen anyone really implement that type of shader.

I know doing a specular lookup doing a ramp texture is quite a bit cheaper.

 

I'd say you should always prefer ALU ops over texture lookups, but then I'm not so sure about mobiles. I guess the best way is to try it out.

 

Yeah just being a bit lazy I always thought normalise() was pretty expensive to use in the fragment shader.

 

I suspect it will be cheaper to use normalize(). Otherwise you're doing an indirect texture fetch on the cubemap and I don't think mobile hardware copes fantastically well with those.

Only way to find out is to try

 

normalize() seems to be implemented as rsqrt(dot(v,v)) * v. rsqrt() is quite fast since it's just an approximation. So I wouldn't worry about it.

 

I don't understand what the difference is between these two concepts.

 

Sorry I meant normalising a vector using a cubemap lookup or using the normalise cg function.

 

Would it still be an indirect texture fetch? I'm just re-normalising vectors that aren't normalised anymore because of the texture interpolators from vertex -> fragment?

It just basically

fixed4 frag(v2f i) : COLOR
{

fixed3 normalisedVector = texCUBE(_NormalisationCubeMap, i.normal); // i.normal is not normalised anymore

....do some other pizzaz

}

 

Ah, I don't think it's indirect if it's the value straight from v2f that you use as the UV coords.

I still don't think it'd be faster than
i.normal = normalize(i.normal);

But why not try it?