[FFmpeg-devel] [PATCH] Replace CLAMP_TO_SHORT with av_clip in adpcm.c

[Aurelien Jacobs]

On Sat, 11 Aug 2007 16:16:37 +0200
Aurelien Jacobs <aurel at gnuage.org> wrote:

> On Sat, 11 Aug 2007 12:02:38 +0200
> Michael Niedermayer <michaelni at gmx.at> wrote:
> 
> > Hi
> > 
> > On Sat, Aug 11, 2007 at 02:13:51AM -0300, Ramiro Ribeiro Polla wrote:
> > > Aurelien Jacobs wrote:
> > > > Hi,
> > > >
> > > > $subj
> > > > This fixes issue84.
> > > >
> > > >   
> > > 
> > > What about making av_clip_int16 in libavutil/common.h?
> > > There are many more places throughout FFmpeg where this is used.
> 
> Attached is an implementation of av_clip_int16().
> 
> > what is the speed/filesize effect of such a change?
> 
> I benchmarked the 3 versions (original CLAMP_TO_SHORT, initial patch
> using av_clip, new patch using av_clip_int16).
> 
> > speed being:
> > volatile int v=0;
> > START_TIMER
> > for(i=-40000; i<40000; i++)
> >     v+= av_clip_int16(i)
> > STOP_TIMER("foober")
> > printf("%d", v);
> 
> Here are the result of this speed benchmark:
> 
> 6184290 dezicycles in CLAMP_TO_SHORT, 1 runs, 0 skips
> 6184370 dezicycles in av_clip, 1 runs, 0 skips
> 5235530 dezicycles in av_clip_int16, 1 runs, 0 skips
> 
> av_clip_int16 is a clear winner here.
> 
> > filesize being object size with --enable-small
> 
> Object size of adpcm.o and ffmpeg with --enable-small:
>    text	   data	    bss	    dec	    hex	filename
>   14382	   2164	      0	  16546	   40a2	libavcodec/adpcm.o.orig
> 2493191	  45804	1017840	3556835	 3645e3	ffmpeg.orig
>   14377	   2164	      0	  16541	   409d	libavcodec/adpcm.o.av_clip
> 2493175	  45804	1017840	3556819	 3645d3	ffmpeg.av_clip
>   14209	   2164	      0	  16373	   3ff5	libavcodec/adpcm.o.av_clip_int16
> 2492967	  45804	1017840	3556611	 364503	ffmpeg.av_clip_int16
> 
> Once again, av_clip_int16 is a clear winner.
> 
> So I will commit soon this av_clip_int16 implementation.

Applied.

Aurel