[x264-devel] x264 multithread predictors

[Jason Garrett-Glaser]

On Tue, Jan 8, 2013 at 9:09 AM, Nadir Raimondo <raimondo at ismb.it> wrote:
> Dear all,
>
> I'm trying to integrate into the x264 library a custom motion estimation
> algorithm for special sequences of images with low overlap.
>
> I calculate new predictors using ancillary data provided by some sensors to
> increase encoding accuracy. Given the low initial overlap, motion vectors
> (and predictor) are very long (about 260 pixels, 1040 in qpel).
>
> All works fine for P frames, with frames sequence IPPPPPPP the gain is about
> 1 dB. Now I'm trying to extend the algorithm for B frames (GOP IBPBPBPBP)
> but I have a problem in multithread scenario (with single thread all work
> correctly).
> It seems that inside x264_me_search_ref function in me.c the MV predictors
> between B and P frames are limited by a mv_y_max_qpel value. As a
> consequence the output MV is not correct and macroblock is encoded as intra.
>
> I thought x264 multithread encoding was performed on minigop so first thread
> 1 encode I, thread 2 encode P and than B, thread 3 encode next P and B and
> so on. Does this predictor limitation occur when the B frame encoding has
> already started while P frame encoding is still pending? or am I doing
> something wrong?

http://akuvian.org/src/x264/sliceless_threads.txt

Jason