[FFmpeg-cvslog] r23095 - in trunk/libavcodec: Makefile allcodecs.c mpegaudio.h mpegaudio_tablegen.h mpegaudiodec.c mpegaudiodec_float.c

[Måns Rullgård]

Reimar D?ffinger <Reimar.Doeffinger at gmx.de> writes:

> On Thu, May 13, 2010 at 04:40:27PM +0100, M?ns Rullg?rd wrote:
>> Reimar D?ffinger <Reimar.Doeffinger at gmx.de> writes:
>> > On Wed, May 12, 2010 at 09:27:43PM +0100, M?ns Rullg?rd wrote:
>> >> Reimar D?ffinger <Reimar.Doeffinger at gmx.de> writes:
>> >> 
>> >> > Btw. going back to the original commit, what are your thoughts about an
>> >> > "automp3" codec that automatically selects whichever codec should be faster
>> >> > (could be compile-time for now)?
>> >> 
>> >> Compiling both doesn't really make much sense IMO.  I would suggest
>> >> making float-only the default and providing configure option for
>> >> switching to integer.  This could be done automatically for CPUs known
>> >> to lack an FPU (or have a slow one).
>> >
>> > Well, compiling both makes testing and comparison benchmarks easier
>> 
>> Easier testing and comparing is not reason enough to bloat the build
>> two copies of essentially the same code for everybody.
>
> I think I agree in the long term, however at the moment it _might_
> be beneficial if end-user can easily make that comparison, given how
> far your and Michael's benchmark results seem to be apart.

How about -flags +float or similar?

>> > and allows runtime-switching if we want to support devices where
>> > which one is faster might depend on which CPU extensions are
>> > available.
>> 
>> This is a valid point.  However, using different codec names is not a
>> good way to achieve this.  We should have one AVCodec selecting the
>> best implementation internally.
>
> Fair enough, I guess I like too much having everything fine-tuneable.
> However might that cause trouble for regression testing? I doubt the
> float and integer versions are bit-exact, and currently the whole
> testing infrastructure needs that...

The int and float versions give different outputs.  I don't know how
much the float version differs, if at all, between machines.

-- 
M?ns Rullg?rd
mans at mansr.com