Mail Archives: djgpp/1996/12/16/18:34:41
> It doesn't implement them completely. Nearly all the transcendental
> functions require you to "massage" the data to get it into the range
> the fpu supports. For example, sin(x) is only valid for small values
> of x; you must divide by 2PI and take the sin() of the remainder.
> That eats up a lot of cycles.
Ok... :) I was gettin over a lot of the hurdles with the limitations
by precomputing them: loading the inbuilt Pi, mul by 2, fld1,
fdiv ST, ST(1), and store as _1on2PI. Its messy, but it works... :)
Mind you, as I only do texture mapping and gourard shading etc, I
don't really need to use sin, cos, sincos, etc, so my experience with
these is limited...I just played with them at one stage for fun...
Oh well, I spose 34 cycles for a 4x4 matrix multiply good enough
nowadays anyway... :) And remember, it could be worse: we might
still be using external 386 FPU's... *GRIN*
Leathal.
- Raw text -