both the Fresnel and Distribution textures would have to be 3d to get a one-to-one mapping and an accurate specular spread.
I'm assuming that you are using the extinction=0 Fresnel approximation, which depends only on L.H. The Cook-Torrance distribution function is based only on N.H (acos(N.H) = alpha).
Therefore, the F*D texture only needs to be 2D.
To map the textures onto an object, you compute V, L, N, and H in the same 3D space (tangent, view, object, world, etc.), and then compute [ (dp+1)/2 ] to generate each texture coordinate.
I guess on GeForce 3 hardware, these values will probably need to be interpolated per-vertex to get everything running in 1 pass. On 8500, you could probably compute each dot product in the first phase, and then lookup the terms in the second phase. This is a non-issue on next-gen hardware.
Do you have to store the values from an Nth Taylor series in a texture? Or does the GPU integrate by summation? How many passes? Whats the error?
You store the values from an Nth taylor series (typically 4 terms, 8 if necessary) as a constant. Which is similar to storing as a texture. The error depends on the function you're approximating -- for atan, around pi/2, the series converges very slowly. There are better approximations for atan, though, then using an N-term taylor series.
