AN1482 - APPLICATION NOTE

4/8

2 - PROFILING OF THE CODE

2.1 - Wang IDCT

This block has been taken directly from the Simple Profile implementation. Since inverse zig-zag/

alternate scan is already performed by the Huffman decoding/IQ block, there’s no need to unroll the first

loop of IDCT, and the code size is significantly reduced. The module wang_idct.st1* is made by two

functions, one for intra IDCT (that includes clipping and add 128) and one for non-intra IDCT. This module

uses 32 bytes of read-only data memory for constants (

are reported in Table 1.

This algorithm can be replaced by the high performance Huang IDCT, whose ST120 implementation is

now available.

*

For

more

information

on

the

MPEG2

code

implemented

on

ST120

please

contact

st100.marketing@st.com.

2.2 - Add_Prediction()

This function is located in the module add_pred.st1. Its purpose is to compute the sum of two 8x8 blocks

and clip the result in the range [0,255]. In fact, one of the two blocks is the predictor computed by

temporal prediction and the other is the difference block read from bitstream. The result overwrites the

predictor, and this highlights a global optimization that is very effective: Add_Prediction() is applied only

to the decoded blocks that are non-null, based on the

frame/field DCT coding is dealt with at this time: the Add_Prediction() routine can read the predictor both

in frame or field mode, and the result is written in the same fashion. Table 2 shows the number of clock

cycles needed for this function in both frame/field cases and its code size.

2.3 - Huffman Decoding/Inv Quant/Inverse Scan

This is the function that takes from the input bitstream the Huffman variable length codes of the data

coefficients for one 8x8 block and decodes them. Then, inverse quantization and reverse zig-zag/

alternate scan are applied. Another task performed by this function, is to clear the result block before

starting decoding. This is needed since the decoding process affects only the non-zero coefficients. Four

of these functions are needed to build a Main Profile decoder, that is MPEG-1 backward compatible. In

particular, we need different functions for intra and non-intra blocks, since they use different Huffman

tables. MPEG-1 backward compatibility requires to deal with a different kind of escape codes, whose

peculiarity is the fact that they do not have a fixed length (there is a sort of "escape in escape" increasing

the level of difficulty).

The computational cost of the four functions cannot be given as standalone, since it depends on the

number of symbols that are decoded. The only interesting figure is the number of cycles/symbol, that is

the same for all the four functions (Table 3).

Table 1 : Wang IDCT (code size and cycles)

Function

Code Size (Bytes)

Cycles/8x8 Block

Wang_idct()

1104

530

Wang_idct_inter()

996

469

Table 2 : Add prediction (code size and cycles)

Function

Code Size (Bytes)

Cycles/8x8 Block

Frame DCT Coding

Cycles/8x8 Block

Field DCT Coding

Add_Prediction()

612

117

115