Re: [ft-devel] Fast bisecting BBox_Cubic_Check

[Alexei Podtelezhnikov]

On Wed, Feb 27, 2013 at 12:40 PM, Werner LEMBERG <address@hidden> wrote:
>
>
> > The original motivation was to kill the last use of FT_SqrtFixed.
>
> Ah, ok.
>
> > Then I read old emails that made perplexing assertion that bisection
> > is too slow to converge while never questioning iterations in the
> > sqrt algorithm.  So I saw a challenge and produced a fast
> > proof-of-concept code without sqrt.  It is a matter of improving
> > accuracy now.  The old algorithm uses upscaling to achieve its
> > precision and so can I.
>
> Great!  It's always good to have a challenge :-)
>

I think I am back online I think.
I want to wrap up the new implementation of BBox_Cubic_Check. As we
discussed, the integer Bezier split algorithms are not accurate, with
the lowest two bits being random. Freetype lives with this. So I
wonder what's the expectations for the BBox calculations? Do we really
need accuracy better than 2 bits?