[FFmpeg-devel] FATE & Regressions (and PPC is broken)

[Siarhei Siamashka]

On 16 March 2008, M?ns Rullg?rd wrote:

[...]

> > It still would be nice to keep C code as a reference implementation, so
> > that it could be always possible to check if an assembly optimized mdct
> > produces bitexact results.
>
> Floating-point isn't bit-exact, as has just been demonstrated.

Yes, it also demonstrates that md5 check in this particular regression test
does not make much sense :) It can break any time on any platform after, for
example, compiler upgrade which could result in changes in codegenerator.

Is it even worth fixing with your patch (also introducing performance
regression on some platforms)? Maybe a more intelligent regression test which
just verifies PSNR (taking some possible varuiation in last digits into
account) would be more suitable in this case?