Eyevinn / hi265 / 37227603989

81%

main: 81%

Pull #12

github

fix(decoder): leave the edges between skip CUs unfiltered

internal/deblock gave every TU and CU edge a boundary strength of 2, the value for an intra edge. Between two zero-motion skip CUs with no residual, spec 8.7.2.4 gives 0, and the edge is not filtered. So every deblocked P picture was filtered next to its CU edges: a P-skip against x265 output decoded 132 of 12288 samples away from its IDR (max 2), where FFmpeg copies it exactly. Bs is now 2 where either side is intra and 0 otherwise. That is exact for skip CUs, the only inter CUs the decoder reconstructs.

x265's own P pictures were affected too. Two copies of one frame, coded at x265's defaults (deblocking, SAO and weighted prediction on) and decoded against FFmpeg, gave 11 decodable streams: testsrc2, smptebars, testsrc and rgbtestsrc at 128x64, 256x128 and 416x240, less one with motion that is refused. All 11 were up to 8 off and now match FFmpeg exactly.

testdata/pskip_deblock_128x64.265 is one of those streams, with an FFmpeg golden; TestDecodePSkipDeblock128x64 fails without the fix (754 samples off). TestEncodePSkipFromSPSPPS_WeightedPredDeblocked now takes its IDR from that vector instead of running x265, and checks pkg/decoder as well as FFmpeg, so it runs without either tool.

46 of 48 new or added lines in 3 files covered. (95.83%)

7174 of 8881 relevant lines covered (80.78%)

0.92 hits per line