AN1482 - APPLICATION NOTE

5/8

Table 3 : Huffman decoding/Inverse quantization/Inverse scan

All these functions are located in the module getblk.st1. The Huffman look-up tables needed for these

functions takes 1732 bytes (read-only data memory).

2.4 - Temporal Prediction

This function is basically made of three parts: the memory read from the external frame buffer, the

computation of horizontal and/or vertical interpolation and/or the average with the previously computed

predictor in case of bidirectional interpolation or dual prime prediction.

The first task is left to a DMA, that will be programmed with the address of the predictor just after the

macroblock header decoding, based on the prediction type information and on the motion vectors. For P

skipped macroblocks, the DMA will directly copy the predictor into the current frame buffer.

So, the profiling reported here only includes the computational cost of the interpolation and/or average.

An efficient implementation should envisage the use of a "powered" DMA, able to make itself this task, in

particular to cope with B skipped macroblocks. For these on, a simple memory copy is not enough, since

a B skipped macroblock inherit the prediction mode and the vectors from the previous one, whereas a P

skipped macroblock is simply patched with the corresponding one in the previous image.

Anyway, the software complexity of this task is computed here for the sake of completeness and for an

eventual implementation on a small video format. There are several possible combinations of

interpolations/average. The following Table 4 gives the number of clock cycles needed for a 16x16

luminance macroblock computation, in the different cases. The function is located in the module

motion_comp.st1 and its name is form_component_prediction().

The code size for this function is of 5524 bytes. A read-only table of 48 bytes is needed for the jump at the

beginning (switch structure). The algorithm for the chroma components is the same, but the cost is

reduced by a factor 4 for each component.

2.5 - Bitstream Parsing

This block represents a set of small functions that are used basically to decode the macroblock headers

(containing information about the contents of the macroblock and the motion vectors). These functions

are located in the modules mb_headers.st1 (

macroblock_modes(), Get_coded_block_pattern(),

The code size of these functions amounts to 2336 bytes. The data memory needed for the Huffman

tables is of 584 bytes. The computational cost of these functions cannot be provided as standalone, since

it depends on the bitstream to be decoded.

Function

Code Size (Bytes)

Cycles/Symbol

Decode_MPEG1_Intra_Block()

1272

24

Decode_MPEG1_Non_Intra_Block()

1040

24

Decode_MPEG2_Intra_Block()

1184

24

Decode_MPEG2_Non_Intra_Block()

896

24

Table 4 : Temporal prediction : the different cases

Operation

Clock Cycles For 16x16 Luminance

Simple copy

125

Copy with average

240

Vertical interpolation

274

Vertical interpolation + average

434

Horizontal interpolation

271

Horizontal interpolation + average

431

Vert+Hor interpolation

623

Vert+Hor interpolation + average

815