Eyevinn / hi264 / 37922378987

88%

main: 88%

Build Type

Pull #34

github

Committed by tobbee

Commit Message

fix: report a CABAC slice that needs bits past the end of its data

cabac.Decoder.readBit returned 0 once the slice data ran out, so a truncated slice went on decoding bins from zeros up to the last macroblock of the frame. With #33's bound on mb_qp_delta this no longer hangs, but it could still decode up to 139264 phantom macroblocks and return a frame without an error: 22 of 59 truncations of a CABAC IDR with an 8192x4352 SPS did.

A conforming slice never needs bits past its end. Decoding end_of_slice_flag equal to 1 does no renormalization, and the last bit read into codIOffset is the rbsp_stop_one_bit (section 9.3.3.2.4), so no slack is needed. readBit still returns 0, keeping the arithmetic defined, but records that it went past the end, and the new Decoder.Err returns ErrEndOfData from then on. DecodeSliceData checks it after each macroblock and end_of_slice_flag, and the CABAC P_Skip loop after each end_of_slice_flag. The bound on mb_qp_delta stays, since a per-macroblock check cannot stop a loop inside one macroblock.

All golden streams still match FFmpeg, so none reads past its end. Across every truncation length of 31 golden CABAC IDRs, 2 generated ones and the 8192x4352 case, about 150000 cuts, every cut now returns an error, and the slowest 8192x4352 cut takes 0.12 s instead of 0.9 s. BenchmarkDecodeCABAC and BenchmarkDecodeHD are unchanged within noise.

Pull Request Pull Request #34: fix: report a CABAC slice that needs bits past the end of its data

Coverage Stats

13 of 15 new or added lines in 3 files covered. (86.67%)

5 existing lines in 1 file now uncovered.

8507 of 9646 relevant lines covered (88.19%)

1.02 hits per line