Huffman Decoder

1
Huffman Decoder

• Team White Project Presentation
• Wei Jiang
• Suresh Polisetty
• Xiang Li

2
Background

• MPEG-1 layer-III, also known as MP3, is a highly
  compressed digital audio format.
• It was standardized in 1991 by the Moving
  Pictures Expert Group, popularly referred to as
  MPEG.
• Huffman Decoder is used in MP3 decoding process
• In the MPEG-1 layer-III, 32 Huffman tables based
  on statistics for audio information are defined.
  The Huffman table to use for the current frame in
  the decoding stage is defined in the side
  information.

3
MP3 Decoding Process
4
MP3 Compression

• Lossless coding like zip can only reduce file
  size by 15.
• Perceptual coding can remove 90 from the
  original file without a significant audible
  difference.
• MP3 (MPEG Audio Layer 3) uses both perceptual and
  lossless coding

5
Overview - Perceptual coding

• If a dominating frequency is present, it will
  suppress the perception of other frequencies in
  the vicinity. This is called frequency masking
• if a transient occurs in the audio signal, the
  human ear will for a short period become less
  sensitive to frequencies lower than the
  transient. This is called temporal masking.

6
MP3 Audio

• Audio compressed into frames (26 ms of sound)

• All frames have equal size
• Frame N can put some of its information in frame
  N-1 if there are free space, using the Bit
  Reservoir

7
Huffman Coding

• Huffman coding is one of the best known
  compression methods. The idea is to assign a
  variable-length encoding in bits to each symbol.
• Compression is achieved by assigning shorter
  codes to more frequent symbols.
• Decompression uniqueness is guaranteed because no
  code is a prefix of another.

8
Huffman Decoding

• A Huffman table can be viewed as a tree-like
  look-up table that must be processed from the
  root, referred to as a start node.

9
Huffman Decoding

• The stream of bits in the compressed file is
  traverse the Huffman compression tree, starting
  at the root. Whenever a leaf node is reached, the
  corresponding symbol is printed out and the tree
  traversal is restarted.

Example Huffman Table 1, 2, 2, 3, 0, 0, // x
0, y 0 2, 3, 0, 16, // x 1, y 0 0, 17, // x
1, y 1 0, 1 // x 0, y 1
Table1 of 32 Huffman tables
10
Huffman Decoding

• 576 bands are reproduced
• Only the most important values are huffman coded,
  the others are expected to be zero
• 31 tables are used for Bigvalues and only one
  table for count1.

11
Huffman Decoding
Actual output from Huffman Decoder
12
Huffman Implementation
13
32 Huffman Lookup Tables
LIBRARY IEEE use ieee.std_logic_1164.all entity
mylut is port( number IN
std_logic_vector(12 downto 0) dataout
OUT std_logic_vector(7 downto 0) ) end
mylut architecture behavior of mylut
is begin process(number) begin case number is
when "0000000000000" dataout when "0000000000001" dataout "00000000" when "0000000000010" dataout

dataout dataout "0000000000101" dataout when "0000000000110" dataout when "0000000000111" dataout "00000000" when "0000000001000" dataout

dataout dataout "0000000001011" dataout when "0000000001100" dataout when "0000000001101" dataout "00010001" when "0000000001110" dataout

dataout dataout "0000000010001" dataout when "0000000010010" dataout

Integrated as one table

5546 entries (13 bits)

8 bits output

Generated by MatLab from text file

Synthesized as a separate module

Use 20 logic cells in the chip

4 sub-tables are used for each frame

according to side info.

14
Decoding Procedure (BigValue Band)

• Wait for GO signal from the controller (Mem
  management)
• Moving inside the Huffman tree until it reaches
  the leaf
• Build 2/4 decoded values according to the side
  info
• Send the result (and the number of bits used)
  back to
• the controller and wait for the next bitstring.

15
Decoder Pre-layout Simulation
16
Decoder Post-layout Simulation
17
Decoder Layout
18
Communication between C VHDL

• 1. C code reads all the side info from frame
  header, and place them in 0 1 64-bits words.
• 2. C code reads the bit stream and place them
  from 2 65 words.
• 3. C code sends a start signal by writing to
  0XFF. C code will listen to OXFE for DONE signal.
• 4. VHDL start to work
• Read bit stream from DP ROM
• Write result to 100 244 words
• After done, send DONE signal by writing to 0XFE.
• 5. Go back to step 1 and restart for a new frame

19
Data Flow
20
FSM for Communication and Memory Management

• 1. 128 bits buffer to hold the bitstream
• 2. Each time send 64 bits to decoder module
• 3. Use an integer to trace the first bit in the
  buffer.

21
Test bench for pcore

• In our test bench, we simulate the communication
  process between C and VHDL
• We generate the side info and bitstream from a
  real frame and repeat this frame for N times
• The simulation result is correct.

22
Pre-layout simulation 1
23
Pre-layout simulation 2
24
Layout
25
Post-layout simulation
26
Final Virtex Layout
Virtex Layout, 45 of Chip used
27
Demo on pilchard

• 20 sec mp3 file(257k)
• Decoded into wav file (2.4 M,no header)
• Playable in MatLab
• Even frames correct, odd frames not

28
Whats next

• Implementation on ASIC tsmc1.8
• Fix the bug, demo on pilchard

29
References

• http//www.es.lth.se/home/tlt/dicp/2002/index.html
  
• http//www.dv.co.yu/mpgscript/mpeghdr.htm
• Modern Information Retrieval