[FFmpeg-devel] [PATCH] SPARC VIS simple_idct try#8

Balatoni Denes dbalatoni
Fri Aug 31 20:28:05 CEST 2007


Hi!

Friday 31 August 2007 01:04-kor Michel Lespinasse ezt ?rta:
> Just for information, Belatoni, do you happen to have benchmark numbers
> against the mlib code ?

Sure. Simple_vis is slightly slower than mlib, but much more accurate - I 
think it's worth using it instead of mlib.

NASA abc HDTV 720p.mpg (mpeg2 HDTV)
1 run with mplayer

mlib
BENCHMARKs: VC: 264.861s VO:   0.058s A:   0.000s Sys:  11.586s =  276.505s
BENCHMARK%: VC: 95.7888% VO:  0.0210% A:  0.0000% Sys:  4.1902% = 100.0000%
simple_vis
BENCHMARKs: VC: 273.837s VO:   0.064s A:   0.000s Sys:  11.341s =  285.242s
BENCHMARK%: VC: 96.0015% VO:  0.0225% A:  0.0000% Sys:  3.9760% = 100.0000%
simple_c
BENCHMARKs: VC: 363.950s VO:   0.072s A:   0.000s Sys:  11.308s =  375.330s
BENCHMARK%: VC: 96.9680% VO:  0.0192% A:  0.0000% Sys:  3.0128% = 100.0000%
libmpeg2 
(it's own C idct-you have to add VC+VO numbers to compare with above numbers)
BENCHMARKs: VC: 287.611s VO:  21.027s A:   0.000s Sys:  11.127s =  319.766s
BENCHMARK%: VC: 89.9443% VO:  6.5758% A:  0.0000% Sys:  3.4798% = 100.0000%


delewis-broken-mlib.avi (mpeg4)
Best of 10 runs with mplayer

mlib
BENCHMARKs: VC:   8.908s VO:   0.005s A:   0.000s Sys:   1.275s =   10.188s
BENCHMARK%: VC: 87.4308% VO:  0.0515% A:  0.0000% Sys: 12.5177% = 100.0000%
simple_vis
BENCHMARKs: VC:   9.312s VO:   0.007s A:   0.000s Sys:   1.385s =   10.703s
BENCHMARK%: VC: 86.9998% VO:  0.0629% A:  0.0000% Sys: 12.9373% = 100.0000%
simple C
BENCHMARKs: VC:  11.669s VO:   0.006s A:   0.000s Sys:   1.496s =   13.171s
BENCHMARK%: VC: 88.5975% VO:  0.0433% A:  0.0000% Sys: 11.3592% = 100.0000%


logo.mpg  (mpeg2 HDTV)
Best of 10 runs with mplayer

mlib
BENCHMARKs: VC:  23.089s VO:   0.005s A:   0.000s Sys:   1.033s =   24.128s
BENCHMARK%: VC: 95.6949% VO:  0.0225% A:  0.0000% Sys:  4.2825% = 100.0000%
simple_vis
BENCHMARKs: VC:  23.954s VO:   0.005s A:   0.000s Sys:   1.125s =   25.084s
BENCHMARK%: VC: 95.4936% VO:  0.0218% A:  0.0000% Sys:  4.4846% = 100.0000%
simple_c
BENCHMARKs: VC:  29.675s VO:   0.007s A:   0.000s Sys:   1.124s =   30.806s
BENCHMARK%: VC: 96.3292% VO:  0.0233% A:  0.0000% Sys:  3.6475% = 100.0000%
libmpeg2
(it's own C idct-you have to add VC+VO numbers to compare with above numbers)
BENCHMARKs: VC:  23.110s VO:   2.438s A:   0.000s Sys:   1.092s =   26.641s
BENCHMARK%: VC: 86.7475% VO:  9.1527% A:  0.0000% Sys:  4.0997% = 100.0000%

Some accuracy numbers: (dct-test -i 0 output, these were posted before):

  144   101   394   286   421   176    11   126
   88   109   152    79    99    93   166    94
  233   140    94   134    83   131    57    74
  185   100   121    74    85   117    84    70
  207   134    74   129   129    89    66    88
  112    74    74    94    75    80   124   103
  126   144   130   111   120    82    87   104
   96   119    72    83    66   133   146    97
IDCT SIMPLE-VIS: err_inf=1 err2=0.02248828 syserr=0.02105000 maxout=260 
blockSumErr=8
IDCT SIMPLE-VIS: 1486.2 kdct/s

 2810  1725  3883  3476  1445  2977  1362  1943
 1263  1244  3402  2061  3266  2497  2936  2741
 3763  3555  1640  2765  1864  2615  2039  2391
 3648  1897  2893  2304  2447  2391  2482  2613
 1796  3132  2055  2615  2115  2426  2246  2373
 2690  2483  2701  2527  2544  2556  2554  2520
 1361  2738  2237  2556  2252  2520  2360  2349
 1882  2594  2395  2605  2279  2759  2274  2578
IDCT MLIBidct: err_inf=3 err2=0.38151094 syserr=0.19415000 maxout=260 
blockSumErr=40
IDCT MLIBidct: 2676.6 kdct/s

   92   156  -255   132  -189   129    -9   110
 -242    90   -97   100  -105    84  -104    80
  -89   157  -141    66  -165   107  -108    89
 -142   116   -92    81   -66   107  -102    67
 -163   128  -157    87   -92    59  -104    50
  -93   142   -78   123   -60   111  -107   109
 -149   148   -49    98   -43    64   -87    82
 -165   115  -115    66  -115    86  -112    96
IDCT WALKEN-VIS: err_inf=1 err2=0.02248906 syserr=0.01275000 maxout=260 
blockSumErr=8
IDCT WALKEN-VIS: 42.2 kdct/s


bye
Denes




More information about the ffmpeg-devel mailing list