> Thanks everybody for your contribution. I'm right now working with
> Bob's approach because:
> - It handles 16 and 24 bits
> - The license is ok
> - I don't relly need top performance on those types, and the code is small.
> I tracked the origin of the previous code I was considering, and
> if anybody interested, is from a paper from Jeroen van der Zijp:
> ftp://www.fox-toolkit.org/pub/fasthalffloatconversion.pdf >
> Clean and fast code, used by MESA. But I am not sure about
> the license. So, in this case of doubt, I prefer the safe side.
> The code suggested by James Clos (thanks Jim!) is
> also ok, but is under BSD license and that would mix licenses,
> which would be a pain for some people using the package.
> So, thanks again Bob, I hope integrating both formats (16 and 24
> floating point) would take few days. Hopefully it will be ready
> for next release in July.