reciprocal = 1.0f / value;

(x * reciprocal) == (x / value)

For texture mapping, you could precalculate the inverse Z1 and Z_delta, and linearly interpolate along there, and that way you wouldn't have to divide for each pixel, only multiply.

Also, for software perspective correct texture mapping, I recommend using the bilinear perspective correct method. This is where you recalculate the the correct coordinates every 8 or 16 pixels and then linearly interpolate between the coords. This is much faster and yields suprisingly decent results. This is what was done in the software rendering portions of Quake and Quake 2.

[This message has been edited by CodeDemon (edited November 22, 1999).]

Thanx all.

To do exact perspective texture mapping requires at one per-pixel divide by z. If every other component of the engine was optimized would one per-pixel divide be too much??

I look forward to your comments.

single-precision divide = 17 cycles
double-precision divide = 36 cycles
extended-precision divide = 56 cycles

So in answer to your question, Yes. But you probably won't have much time to do anything else!

I wrote a program recently that needed the effect of one floating-point divide per pixel. Rather than do the divide, I used a lookup table of the reciprocals described by CD.