[MPlayer-cvslog] r19836 - trunk/libmpcodecs/vf_yadif.c

[Loren Merritt]

On Thu, 14 Sep 2006, Jason Tackaberry wrote:
> On Thu, 2006-09-14 at 06:00 +0200, lorenm wrote:
>
>> mmx implementation of yadif. 3.5x faster.
>
> Awesome, thanks for this.
>
> Now if only mcdeint were an order faster. :) (But I suppose most of its
> work is done in libavcodec.)  It's by far the best deinterlacer I've
> seen anywhere.

Don't give up hope yet.
for -vf mcdeint=0:1:10 (i.e. fastest settings), the profile is

%        cum. %     symbol name
28.2325  28.2325    put_image
25.9386  54.1711    fast_memcpy
12.4428  66.6139    encode_frame
  8.1563  74.7702    mc_block
  4.8370  79.6072    encode_q_branch
  3.2598  82.8670    qpel_motion_search
  1.5675  84.4345    sad16_mmx2
  1.4897  85.9242    put_symbol
  1.2241  87.1483    ff_epzs_motion_search
  0.8493  87.9976    put_h264_qpel16_mc21_mmx2
  0.8267  88.8243    put_h264_qpel16_mc11_mmx2
  0.8102  89.6345    put_h264_qpel16_mc13_mmx2
  0.7951  90.4296    put_h264_qpel16_mc23_mmx2
  0.7271  91.1567    put_h264_qpel16_mc31_mmx2
...

so 54% of the time is spent outside of motion estimation.

--Loren Merritt