[FFmpeg-devel] WMV2 decoding performance

[Loren Merritt]

On Sat, 10 Nov 2007, Diego Biurrun wrote:

> I notice that our WMV2 decoder is vastly faster for most of the samples
> I try it on, for example all the stuff from
>
> http://samples.mplayerhq.hu/V-codecs/WMV8/
>
> except for the following sample:
>
> http://samples.mplayerhq.hu/V-codecs/WMV8/vandread_ep13_op_wm8_ver.wmv
>
> Any clues what could be the reason?  Benjamin suspects a non-optimized
> IDCT.  The only difference I notice right off is that the above sample
> is much higher resolution that the others.

It's fast for me.
Core2 Duo 2.4GHz, gcc 4.2.2, Gentoo.

time mplayer-32 vandread_ep13_op_wm8_ver.wmv -nosound -benchmark -vo null -vfm vfw
5.41 +/- .03 sec

time mplayer-32 vandread_ep13_op_wm8_ver.wmv -nosound -benchmark -vo null -vfm dmo
3.49 +/- .03 sec

time mplayer-32 vandread_ep13_op_wm8_ver.wmv -nosound -benchmark -vo null -vfm ffmpeg
2.66 +/- .01 sec

time mplayer-64 vandread_ep13_op_wm8_ver.wmv -nosound -benchmark -vo null -vfm ffmpeg
2.38 +/- .02 sec

oprofile says:
samples  %        symbol name
178079   31.5680  msmpeg4_decode_block
48905     8.6694  ff_simple_idct_put_mmx
46937     8.3205  simple_idct84_add
43316     7.6786  wmv2_decode_mb
36894     6.5402  simple_idct48_add
25971     4.6039  dct_unquantize_h263_intra_mmx
22512     3.9907  ff_simple_idct_add_mmx
16078     2.8501  mpeg4_pred_ac
14519     2.5738  clear_blocks_mmx
12089     2.1430  wmv2_add_block
11930     2.1148  MPV_decode_mb
10293     1.8246  ff_emulated_edge_mc
10099     1.7902  MPV_motion
8504      1.5075  msmpeg4_decode_motion
6429      1.1397  decode_slice
6268      1.1111  put_pixels16_mmx
5881      1.0425  put_pixels_clamped_mmx
5071      0.8989  put_pixels16_xy2_mmx
5033      0.8922  ff_h263_update_motion_val
4694      0.8321  put_no_rnd_pixels16_xy2_mmx
4281      0.7589  add_pixels_clamped_mmx
4229      0.7497  ff_wmv2_add_mb
4094      0.7257  put_pixels8_xy2_mmx
3770      0.6683  put_no_rnd_pixels8_xy2_mmx
3452      0.6119  prefetch_mmx2
2932      0.5198  put_pixels8_mmx

--Loren Merritt