This patch series changes RD5/6 analysis (slower, veryslow, placebo) to analyze split CU cost before ME modes or intra, but after merge/skip. This basic change itself does not change much, but it allows us to use the results of the split CU analysis (aka the next lower depth) to limit unlikely work at the current depth. For instance it no longer measures intra at the current depth if none of the four split CUs did not pick intra. It also enables the --limit-refs CLI option to finally have an effect, but only for RD 5/6.
Example results: 1) before this patch series http://privatepaste.com/48bc42a800 2) after patch series (intra mode avoidance) http://privatepaste.com/9620a9c5be 3) after patch series --limit-refs 1 http://privatepaste.com/3245474bea 4) after patch series --limit-refs 2 http://privatepaste.com/33fa78fddf 5) after patch series --limit-refs 3 http://privatepaste.com/5d11be9fcd The biggest benefit of --limit-refs is that it will eventually allow us to increase refs for most presets for very limit performance cost, once we implement depth-first search for the remaining RD levels. See --limit-refs 3 --ref 6 http://privatepaste.com/0664d2816b Another benefit of evaluating splits before ME or intra is that pmode will be able to use nearly the same work-avoidance logic as normal analysis, so it will be nearly as work efficient as non-pmode encodes. _______________________________________________ x265-devel mailing list [email protected] https://mailman.videolan.org/listinfo/x265-devel
