[MPlayer-dev-eng] Nut startcodes

[Ville Saari]

On Thu, Apr 29, 2004 at 02:40:20AM +0200, Michael Niedermayer wrote:

> ok, heres the explanation on how this is possible, its not easy unless we 
> restrict ourselfs to seeking to frames with an immediatly preceeding 
> startcode, in which case it would be easy ...

That requires an assumption that there is only one stream or that all
the streams have their keyframe positions synchronized. In more general
case only one of the streams might have a keyframe immediately after the
startcode and all the other streams have to be decoded starting from
their last keyframes. That keyframe has to be located somehow.

If there can be an arbitrary number of startcodes between the seek
position and the last keyframe before the seek position, then finding
the keyframes is tricky.

You might want to optimize the amount of seeks on typical case by
starting form the second last startcode and then start decoding each
stream as the keyframes are met, but then in the worst case you have
to re-try multiple times from earlier startcodes and re-decode. That's
O(n^2) if you try all the startcodes (can be made O(n*log(n)) if the skips
backwards are progressively increased).

Another approach is to first find the keyframes and then seek back to
decode. Then the worst case search is linear time and no decoding, but
the typical case needs an extra seek back, which might be an issue if
the video comes from an optical medium.

-- 
 Ville