[FFmpeg-cvslog] r24213 - trunk/libavcodec/dca.c

[Vitor Sessak]

On 07/12/2010 11:17 PM, Christophe Gisquet wrote:
> Wasn't registered before, sorry for replying to another mail than the
> one intended.
>
> 2010/7/12 M?ns Rullg?rd<mans at mansr.com>:
>>>>> Won't this patch make it use double-precision floating point math when
>>>>> not needed?
>
> For the record, the only change on x86, where x87 FPU code is used
> anyway, is a flds instruction (load of float) changed to a fldl (load
> of double), which is pretty harmless speedwise.

I suppose you mean x86-32, because x86-64 uses as default SSE for 
floating-point operations. Also, even for x86-32, we should pass 
"-mfpmath=sse" to gcc to use SSE math in processors that support it, 
since I've measured up to 10% speedup in some codecs.

-Vitor