Eyevinn / hi264 / 32481712975

88%

main: 88%

Pull #22

github

fix: write CAVLC trailing_ones_sign_flag in transmission order

levels is collected in reverse scan order, which is already the order the
trailing_ones_sign_flag syntax elements are transmitted in (spec 7.3.5.3.2),
but they were written back to front. Blocks with two or three trailing ones
of mixed sign therefore decoded with those coefficients' signs permuted.

For chroma DC a 1<->2 swap transposes the 2x2 DC array, so the TR and BL
sub-blocks pick up equal and opposite errors. Chroma is coded DC-only, so
the residual could not correct it and the error fed the next macroblock's
intra chroma prediction, propagating right and down: visible as colour
bleeding streaking diagonally down-right on PNG/JPEG input. Flat .gridimg
patterns were unaffected because a single DC coefficient never produces a
mixed-sign trailing-one pair, which is why the encoder verification tests
missed it.

With the fix the encoder's own reconstruction is bit-exact with both
hi264dec and FFmpeg for photo input (it was not before), and CAVLC output
is pixel-identical to CABAC, which never had this bug. On a 320x240
mandelbrot at QP 26 chroma PSNR vs the encoder target goes from
Cb 39.1 / Cr 40.4 dB to Cb 51.4 / Cr 50.9 dB.

Add round-trip tests through the FFmpeg-verified CAVLC reader: the targeted
sign cases, an exhaustive sweep of 4-coefficient chroma DC blocks, and a
fuzz pass over every (nC, maxNumCoeff) category the encoder emits.

4 of 4 new or added lines in 1 file covered. (100.0%)

8134 of 9300 relevant lines covered (87.46%)

1.01 hits per line