[x264-devel] Motion Estimation

[David Smith]

OK,

I am in the process of packing up dependencies so that I can pass the  
compute of x264_me_search_ref over to a PCIe x4 multi core DSP board  
with 2Gb global memory.  Assume a messaging system is already in place  
between the host PC and the dsp board.

A function replicating x264_me_search_ref is running on DSP core 0  
with job management that sends a x264_me_search_ref call to each core.

I have been trying to figure out the data dependancy with Understand  
v2 / Netbeans and how other threads change, if at all, the current  
thread structures *h & *m.  Would it simply a case of copying the *h  
*m structures for each call of x264_me_search_ref over to the dsp  
board memory and returning them back at the end for each thread?  I  
assume that would be an unnecessarily large replication of data for  
each call per thread.

How about if the structures are copied once on the first call per  
frame?  Would that be enough to satisfy the dependancy for the rest of  
the frame up until
h->mb.i_mb_xy = h->m.i_mb_count

Any thoughts or comments welcomed.

Thanks,
David.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x264-devel/attachments/20090713/78a41d67/attachment.htm>