1
0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-12 19:18:44 +02:00
Commit Graph

2 Commits

Author SHA1 Message Date
Michael Goulet
c02dd59cd3 lavc/dovi_rpu: Fix UB for possible left shift of negative values
It is undefined to left-shift a negative value.
2022-06-21 18:23:31 +02:00
Niklas Haas
fe04033739 lavc: Implement Dolby Vision RPU parsing
Based on a mixture of guesswork, partial documentation in patents, and
reverse engineering of real-world samples. Confirmed working for all the
samples I've thrown at it.

Contains some annoying machinery to persist these values in between
frames, which is needed in theory even though I've never actually seen a
sample that relies on it in practice. May or may not work.

Since the distinction matters greatly for parsing the color matrix
values, this includes a small helper function to guess the right profile
from the RPU itself in case the user has forgotten to forward the dovi
configuration record to the decoder. (Which in practice, only ffmpeg.c
and ffplay do..)

Notable omissions / deviations:
- CRC32 verification. This is based on the MPEG2 CRC32 type, which is
  similar to IEEE CRC32 but apparently different in subtle enough ways
  that I could not get it to pass verification no matter what parameters
  I fed to av_crc. It's possible the code needs some changes.
- Linear interpolation support. Nothing documents this (beyond its
  existence) and no samples use it, so impossible to implement.
- All of the extension metadata blocks, but these contain values that
  seem largely congruent with ST2094, HDR10, or other existing forms of
  side data, so I will defer parsing/attaching them to a future commit.
- The patent describes a mechanism for predicting coefficients from
  previous RPUs, but the bit for the flag whether to use the
  prediction deltas or signal entirely new coefficients does not seem to
  be present in actual RPUs, so we ignore this subsystem entirely.
- In the patent's spec, the NLQ subsystem also loops over
  num_nlq_pivots, but even in the patent the number is hard-coded to one
  iteration rather than signalled. So we only store one set of coefs.

Heavily influenced by https://github.com/quietvoid/dovi_tool
Documentation drawn from US Patent 10,701,399 B2 and ETSI GS CCM 001

Signed-off-by: Niklas Haas <git@haasn.dev>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
2022-01-04 11:59:41 +01:00