Re: [ft-devel] Significant optimization of ft_div64by32

[Alexei Podtelezhnikov]

>> I have just committed significant optimization of ft_div64by32..
>
> Thanks.  Do you have some numbers for you claim?

This is how long it takes to do 100 million ft_div64by32 calls on AMD
Opteron(TM) Processor 6238 depending how large is the high register:

HI REGISTER MASK, OLD SECONDS, NEW SECOBDS
0x000000FF, 10.4, 4.0
0x0000FFFF, 11.3, 7.0
0x00FFFFFF, 11.2, 9.2

> Additionally, I
> wonder whether your optimization also holds for older architectures
> where division and multiplication are much more expensive than
> operations like shifting and ORing.

It is hard to imagine that old architectures or new ARM implement
division other than through long division. So the explicit division
operator might still go through the same algorithm, i.e., we just
shift the work with some overhead. It might be interesting to
benchmark on ARM, but I do not have access to it.