[FFmpeg-devel] [PATCH] Implement the function cmdutils.c:parse_int_or_die

[Rich Felker]

On Wed, Feb 20, 2008 at 01:43:44PM +0100, Michael Niedermayer wrote:
> Also the normal (outside the academic dreamworld) way to handle float is
> not to assume floating point operations to be exact and keep in mind that
> they are represented by a finite number or mantisse and exponent bits.
> That worked quite well until the CS people found out that one can abstract
> that to the point of not having any relation to mathematics nor reality
> anymore and write 100 page papers about it.

It depends on your application. For multimedia, denormals, overflow,
underflow, etc. simply will not happen and the only issue of
conformance is rounding issues which will have small impact on output
anyway. But for intense numerical analysis, the subtle behaviors DO
matter, a lot! And precision can be very important. That's why C has a
function to compute e^x-1 in addition to just e^x; the former has
catastrophic loss of precision for small x.

I agree that for multimedia purposes we can just consider the floating
point machine without any rigor, but there are legitimate
applications where that approach does not work.

Rich