[MPlayer-dev-eng] [PATCH] liba52 float support

[Zoltan Hidvegi]

D Richard Felker III wrote:
> On Mon, Feb 21, 2005 at 05:56:43PM +0100, Alex Beregszaszi wrote:
> > Hi,
> > 
> > > If any float processing gets done in libaf, it loads several
> > > conversion filters to convert back and forth between many different
> > > nonsense formats, generating 8bit samples in the end. And somewhere
> > > during all the processing, the audio becomes nothing but white noise.
> > > This has always happened ever since libaf was first written.
> > 
> > STOP trolling about libaf and try it out _again_. I have fixed the float
> > issue since some months for now.
> 
> OK, I will test again, but I've never seen it work...
> BTW can you tell me the patch that fixed it?

FYI, I've tested the float stuff on a 221 second 48kHz AC3 comparing
the output with and without -a52float (with my patch). 97% of the
sampes were identical, and in the remaining 3% one of the channels had
a difference of 1, which is expected, because liba52 and af_format use
different float to int conversion methods.  No samples were different
by more than 1.  And I did not see any MMX2 or SSE anywhere in libaf,
so it should work even on a K6.

But int should still be the default, the liba52 internal float to int
conversion is still a bit faster, as it saves some copying, and use
integer only (decodes with a 384 bias, and loads the float as an int).

Zoli