architecture

MPEG2编解码的源代码.zip

This file contains a couple of (currently unstructured and incomplete) encoder
implementation notes.

Basic Assumptions

- data structures

Primary data structures are:

 - picture data arrays, containing either source or reconstructed pels

 - mbinfo array, containing all macroblock level side information

 - blocks array, containing DCT coefficients

mbinfo and blocks together are completely equivalent to the VLC encoded
picture_data() part of the MPEG stream, although in uncompressed and
therefore rather voluminous internal format.

a) picture data arrays

Frames are represented internally as in the following declaration:

unsigned char *frame[3];

i.e., an array of three pointers to the picture data proper. frame[0]
points to the luminance (Y) data, frame[1] to the first chrominance (Cb, U)
component, frame[2] to the second chrominance component (Cr, V).

Width and height of the luminance data array is 16*mb_width and 16*mb_height,
even if horizontal_size and vertical_size are not divisible by 16. The
actual pels are top left aligned and padded to the right and bottom (by
pixel replication) while being read into the encoder.

Width and height of the two chrominance arrays depends on the chroma_format.
For 4:2:0 data, they are half size in both directions, for 4:2:2 data only
width is half that of the luminance array, for 4:4:4 data, chrominance and
luminance are of identical size.

The dimensions are stored in the following global variables:

             | horizontal   | vertical
-------------+--------------+------------
luminance    | width        | height
chrominance  | chrom_width  | chrom_height

Picture data (array of pels) is always stored as frames, even when a
field picture sequence is to be generated. The two fields are stored
in interleaved from, that is alternating lines from both fields. In
most cases, however, it's easier to view the two fields as being stored
side to side of each other as an array of double width and half height.
The following picture demonstrates this:

 <----- width --------->
 T1 T1 T1 T1 T1 T1 T1 T1 <+
 B1 B1 B1 B1 B1 B1 B1 B1  |
 T2 T2 T2 T2 T2 T2 T2 T2  | height
 B2 B2 B2 B2 B2 B2 B2 B2  |
 T3 T3 T3 T3 T3 T3 T3 T3  |
 B3 B3 B3 B3 B3 B3 B3 B3 <+

 <------------- 2*width (width2) --------------->
 T1 T1 T1 T1 T1 T1 T1 T1  B1 B1 B1 B1 B1 B1 B1 B1 <+
 T2 T2 T2 T2 T2 T2 T2 T2  B2 B2 B2 B2 B2 B2 B2 B2  | height/2 (height2)
 T3 T3 T3 T3 T3 T3 T3 T3  B3 B3 B3 B3 B3 B3 B3 B3 <+

 Tn: top field pels
 Bn: bottom field pels

These are of cause only two different two-dimensional interpretations of
the same one-dimensional memory layout. Either the top or the bottom field
can be the earlier field in time.

The following table shows the relation between width, height, width2 and
height2 for frame and field pictures:

         | frame  | field
 --------+--------+---------
 width2  | width  | 2*width
 height2 | height | height/2

Using the convenience variables width2 and height2, many loops are
independent of frame / field picture coding.


b) mbinfo array

This array contains the complete encoded picture (field or frame),
except the DCT coefficients. This includes macroblock type, motion vectors,
motion vector type, dct type, quantization parameter etc. The number
of entries (size of the array) is identical to the number of macroblocks
in the picture.

c) blocks

This array contains all DCT coefficients of the picture. It is declared as

EXTERN short (*blocks)[64];

i.e. a pointer to an array whose elements are blocks of 64 short integers
each. The number of blocks is block_count times the number of macroblocks
in the picture. block_count depends on the chroma format (6, 8 or 12 blocks
per macroblock).

The actual content depends on the encoding stage. Either prediction error
(or intra block data), DCT transformed data, quantized DCT coefficients or
inverse transformed data is stored in blocks. This is done to keep
memory requirements reasonable.


Encoding procedure

The high-level structure of the bitstream is determined in putseq().
This includes writing the appropriate headers and extensions and the
decision which picture coding type to choose for each picture.

Encoding of a picture is divided into the following steps:

- motion estimation (motion_estimation())
- calculate prediction (predict())
- DCT type estimation (dct_type_estimation())
- subtract prediction from picture and perform DCT (transform())
- quantize DCT coefficients and generate VLC data (putpict())
- inverse quantize DCT coefficients (iquant())
- perform IDCT and add prediction (itransform())

Each of these steps is performed for the complete picture before
proceding to the next one. The intention is to keep these
steps as independent from each other as possible. They communicate
only via the above mentioned basic data structures. This should simplify
experimenting with different coding models.

Quantization and VLC generation could not be separated from each other,
as quantization parameters and output buffer content usually form a
closed loop on macroblock level.

- Motion Estimation

The procedures for motion estimation are in the file motion.c. The main
function motion_estimation() loops through all macroblocks of the current
picture calling frame_ME (for frame pictures) or field_ME (for field
pictures) to calculate motion vectors for each macroblock.

Motion estimation is currently done separately. More efficient schemes
like telescopic motion estimation, which span several pictures (e.g.
two I/P pictures and all intervening B pictures), are not yet implemented.

Motion estimation splits into three steps:

- calculation of optimum motion vectors for each of the possible motion
  compensation types (frame, field, 16x8, dual prime)

- selection of the best motion compensation type by calculating and comparing
  a prediction error based cost function

- selection of either motion compensation or any other possible encoding
  type (intra coding, No MC coding)

Motion vectors are estimated by integer pel full search in a search window
of user defined size. This full search is based on the original source
reference pictures. Subsequently the cost functions for the 9 motion vectors
with an offset of -0.5, 0, and 0.5 relative to the best integer pel vectors
are evaluated (using the reconstructed reference picture) and the vector with
smallest cost function is used as estimation.

To speed up full search, the cost function calculation is aborted if the
intermediate values exceed the cost function value of an earlier motion
vector in the same full search. This is most efficient if vectors of
potentially low cost are evaluated fist. Therefore full search is organized in
an outward spiral.