FFmpeg/doc/README.tech

Technical notes:
---------------

Video:
-----

- The decision intra/predicted macroblock is the algorithm suggested
  by the mpeg 1 specification.

- only Huffman based H263 is supported, mainly because of patent
  issues.

- MPEG4 is supported, as an extension of the H263 encoder. MPEG4 DC
  prediction is used, but not AC prediction. Specific VLC are used for
  intra pictures. The output format is compatible with Open DIVX
  version 47.

- MJPEG is supported, but in the current version the huffman tables
  are not optimized. It could be interesting to add this feature for
  the flash format.

- To increase speed, only motion vectors (0,0) are tested for real
  time compression. NEW: now motion compensation is done with several
  methods : none, full, log, and phods. The code is mmx/sse optimized.

- In high quality mode, full search is used for motion
  vectors. Currently, only fcode = 1 is used for both H263/MPEG1. Half
  pel vectors are used.

Audio:
-----

- The mpeg audio layer 2 compatible encoder was rewritten from
  scratch. It is one of the simplest encoder you can imagine (800
  lines of C code !). It is also one of the fastest because of its
  simplicity. There are still some problems of overflow. A minimal
  psycho acoustic model could be added. Currently, stereo is
  supported, but not joint stereo.

- The AC3 audio encoder was rewritten from scratch. It is fairly
  naive, but the result are quiet interesting at 64 kbit/s. It
  includes extensions for low sampling rates used in some Internet
  formats. Differential and coupled stereo is not handled. Stereo
  channels are simply handled as two mono channels.

- The mpeg audio layer 3 decoder was rewritten from scratch. It uses
  only integers and can be 16 bit precision for the synthesis filter
  at the expense of a slight precision loss. A slower bit exact mode
  is available too for compliance testing.