[Ffmpeg-devel] [PATCH] Snow mc_block mmx optimization

[Oded Shimon]

On Sun, Mar 26, 2006 at 11:58:51PM +0200, Oded Shimon wrote:
> On Sun, Mar 26, 2006 at 11:21:08PM +0200, Oded Shimon wrote:
> > On Sun, Mar 26, 2006 at 11:13:37PM +0200, Michael Niedermayer wrote:
> > > first comes the spec/algo then comes the optimiation
> > > 
> > > whats the speed & PSNR/bitrate of the common test videos like forman if we
> > > would always use the mc_block() code and never the h.264 code?
> > 
> > No idea. How do I test this, remove all if's there and use mc_block ?
> 
> By a quick hack attempt of just removing everything and leaving only 
> mc_block in pred_block, I got about 10% speed loss in yblock (with my 
> optimized mc_block) :
> 
> (old is h264 and mc_block, new is pure mc_block)
> 
> old: 29762 dezicycles,   261510 runs,   634 skips (   778322746 total) in add_yblock
> new: 38523 dezicycles,   261584 runs,   560 skips (  1007706523 total) in add_yblock
> 
> other file:
> old: 47347 dezicycles,  1048264 runs,   312 skips (  4963238600 total) in add_yblock
> new: 55883 dezicycles,  1048303 runs,   273 skips (  5858310153 total) in add_yblock
> 
> during encoding:
> old: 55193 dezicycles,   523788 runs,   500 skips (  2890955984 total) in add_yblock
> new: 65341 dezicycles,   523849 runs,   439 skips (  3422894211 total) in add_yblock
> 
> as for psnr vs. bitrate:
> old: 2369047 bytes   PSNR: Y:44.64, Cb:46.19, Cr:46.30, All:45.11
> new: 2372830 bytes   PSNR: Y:44.62, Cb:46.15, Cr:46.28, All:45.09
> 
> 
> This is all about as quick and dirty a test as possible, just a single 
> file test, and maybe mc_block can be optimized further, I dunno.

Thoughts? What about patch?

- ods15