[x264-devel] Re: Speed Issue

[Loren Merritt]

On Thu, 14 Oct 2004, Xiaojie Huang wrote:

> I think the speed is an important issue in any
> practical codec especially for H.264. The modification
> in my last mail can improve the encoding speed at
> least 1 fps.

My benchmark: 1.5% faster. ymmv.

$> time ./mencoder-r50 -ovc x264 -x264encopts qp_constant=20:subq=1:frameref=1:cabac in.avi
x264 [info]: analyse=64(44362ms) encode=16(11600ms) write=9(6812ms) filter=8(5900ms)
x264 [info]: slice I:25   Avg size:19203
x264 [info]: slice P:1405 Avg size:5874
x264 [info]: slice I      Avg I4x4:521   I16x16:444 
x264 [info]: slice P      Avg I4x4:72    I16x16:104   P:275   P8x8:31    PSKIP:481 
x264 [info]: kb/s:1171.5 fps:19.880
86.770u 1.780s 1:29.84 98.5%    0+0k 0+0io 2669pf+0w

$> time ./mencoder-xia -ovc x264 -x264encopts qp_constant=20:subq=1:frameref=1:cabac in.avi
x264 [info]: analyse=65(44144ms) encode=15(10708ms) write=9(6711ms) filter=8(5878ms)
x264 [info]: slice I:25   Avg size:19203
x264 [info]: slice P:1405 Avg size:5874
x264 [info]: slice I      Avg I4x4:521   I16x16:444 
x264 [info]: slice P      Avg I4x4:72    I16x16:104   P:275   P8x8:31    PSKIP:481 
x264 [info]: kb/s:1171.5 fps:20.204
85.670u 1.830s 1:28.19 99.2%    0+0k 0+0io 2662pf+0w

> If you think it influence the
> independency, we can define another variable to
> indicate whether we should encode the current MB again
> to solve this problem.
> If we allocate another small buffer 9*16*16, we can
> save the prediction data of 9 intra 4x4 modes into
> this buffer. When the best mode has been chosen,  the
> prediction data can be immediately used to encode the
> current MB instead of computing them again. Intra
> 16x16 also can employ the same method.

I think a better / more general way would be to combine the analyse and 
encode functions. This would also be required for a rate/distortion MB 
type decision (which I am trying to implement).

--Loren Merritt

-- 
This is the x264-devel mailing-list
To unsubscribe, go to: http://developers.videolan.org/lists.html