Hi, 2014-07-10 8:47 GMT+02:00 Christophe Gisquet <christophe.gisq...@gmail.com>: > The 4 first patches reduce memory allocations as, in a simplified view, > only the previous ctb line is needed for the various prediction mechanisms. > The actual required allocation is much smaller, but would require too many > changes. And the margin of this mail is too small to explain it. [...] > hevc: reduce allocation for split_cu_flag > hevc: reduce allocation for skip_flag > hevc: reduce allocation for tab_ipm > hevc: reduce cu depth allocation
The implementation of threading on slice/wavefront breaks with these patches: multiple threads are actually accessing these buffers. Having these buffers per thread may alleviate the issue, but that would quickly reduce the patches benefit. Anyway, they were more of a poking, so patches 1-4 dropped. PS: not handled in those patches, but tab_mvf_pool / rpl_tab_pool size might be divided by 16, see how x/y are rounded in temporal_luma_motion_vector. YMMV, seeing how threading was implemented for slices. -- Christophe _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel