Re: [AD] div-by-zero... math3d.c

[Sven Sandberg <svens@xxxxxxxxxx>]

Ben Davis wrote:
> I honestly think the best solution would be to do the arithmetic using 
> double-precision values. A float can go down to 1e-38 before it becomes 
> subnormal (and slow) and starts losing precision. A double can go down to 
> 1e-some_three_digit_number. So the smallest float, cast to double, and 
> squared, would be will within the range of a double. As for efficiency, the 
> FPU on an x86 processor probably does everything using 10-bit (the 
> nonstandard 'long double') precision anyway, so we're not losing anything.

You are right, this is probably better.  sqrt() takes and returns 
doubles anyway, so assuming sqrt is the bottleneck we hardly lose much 
speed with your approach.

Sven