next up previous
Next: Precision Up: Results Previous: Results

Normalization

Figure 9 shows the effect of different vector normalization methods. So as to avoid precision problems the specular exponent is relatively small ($ n=8\,$).

Figure 9: Quality of Different Normalization Methods
\includegraphics[height=5.5cm]{D:/Studium/Per-Pixel-Lighting/text_cescg/image/normalization.eps}
Gouraud shading to the left is only shown for reference purposes. Without normalization it is clearly visible, that the shape of the highlight and intensity are not correct. With register combiner normalization the results are astonishingly good, although that only a linear square root approximation is used for normalization. The visible banding effects are primary caused by the limited precision for the exponentiation of the specular intensity. 8 bit cube map normalization works not as good as expected. One cause is that one more bit in comparison to register combiner normalization gets lost as the whole range $ [-1,1]$ must be range compressed to $ [0,255]$ (register combiners work internally with 8+1 bits so the range is $ [-255,255]$). The best render quality is reached by using a 16 bit normalization cube map in combination with a high precision dot product in the texture shaders.



Gerald Schröcker 2002-03-21