A Fast Maximum-Likelihood

Decoder for Convolutional Codes

Jon Feldman Matteo Frigo Ibrahim Abou-Faycal
[email protected] [email protected] [email protected]

M.I.T. Vanu, Inc.

J. Feldman, VTC, 9/25/02 – p.1/13

 Convolutional Codes
Commonly used in TDMA/GSM cellular phones and
other wireless standards.
Simple linear time encoder.
Viterbi algorithm is an optimal decoder, but:
- Software radios cannot use parallelism,






- Running time
( = “constraint length”, = code length).



When SNR is high, decoding should be easier.
- Software radios can take advantage of running time
that (implicitly) depends on SNR...

J. Feldman, VTC, 9/25/02 – p.2/13

Convolutional Decoding Alternatives
Sequential Decoding:
- Suboptimal.
Decoding [ED ’96, HHC ’93, HCW ’02]:

- Time/block:









(high . . . low SNR).
 











- Worst-case running time worse than Viterbi.
- No experiments available to compare running time.
Our “Lazy Viterbi” algorithm:
- Time/block:
 







(high . . . low SNR).





- In software: up to 10 times faster than optimized

Viterbi.
- Works on “streaming” data (non-blocked data).
J. Feldman, VTC, 9/25/02 – p.3/13
 Outline

Convolutional codes, Viterbi algorithm.

decoding, advantages and limitations.

The Lazy Viterbi decoder.

J. Feldman, VTC, 9/25/02 – p.4/13

 Convolutional codes: the trellis
00 00 00 00 00 00
11
11 ... Info. bit 0
01 00 01 01 00 01 01
01 01 ... Info. bit 1
10 10 10 10 10 10
10 ...
11 01 11 11 11 11




Generator Polynomials: ,



Encoding: path through the trellis (labels = code bits).
Info bits:

  
 
 

States:












Code word:

 
 



J. Feldman, VTC, 9/25/02 – p.5/13

 Viterbi Algorithm

00 11 00 00 00 00
...
0
01 01 01 00 01 01
2 ...
01
10 10 10 1 10 10
...
11 11 11 11 11

Received: 11 10 01 ...
Branch Metric of edge: (code bits, received bits).
  

 


  


, , .







Viterbi Algorithm: Find path with lowest total metric.

J. Feldman, VTC, 9/25/02 – p.6/13

 Viterbi Algorithm is BFS
2 1 1
0 2 3 3
0 1 1
1 ...
2 1 2
2 0 ...
0 0 3 2 3
2 ...
0 0 0

Received: 11 10 01 ...
Computes accumulated error for every node.






Running time: (# states = ).




Breadth-First-Search (BFS) is only one possible
“shortest-path” algorithm.
J. Feldman, VTC, 9/25/02 – p.7/13
 Dijkstra’s Algorithm
2
0 2
0 ...
2 2
2 ...
0 0
2 ...
0 0 0

Received: 11 10 01 ...
Dijkstra’s Algorithm (Dijkstra, 1959):
- Repeatedly “expand” the node w/ lowest metric.
: Dijkstra’s algorithm with “heuristic” function.

Running time (# expansions) (time/expansion)

J. Feldman, VTC, 9/25/02 – p.8/13

Trellis Expansions as a Function of SNR

103
Avg. nodes expanded / output bit
Viterbi decoder

102
A* / Lazy decoder

1
10

0
10
-2 0 2 4 6 8
Signal-to-Noise Ratio Es/No (dB)

AWGN, QPSK, Rate = , (CDMA), gen. polys (753,541) (octal).






J. Feldman, VTC, 9/25/02 – p.9/13

Limitations of Previous Work
Cost of deciding which node to expand:
- Must maintain priority queue of trellis nodes.






- Cost = per operation.

 

- Not amenable to compile-time optimization.





High SNR: , time/block .

 






 

 







Low SNR: , time/block .

 




Only defined for block codes (rather than streaming
data).
No implementations given, no evidence useful in

practice.

J. Feldman, VTC, 9/25/02 – p.10/13

 The Lazy Viterbi Decoder
Constant time / expansion via new priority queue







- ( ).

 



“Lazier” version of Dijkstra’s algorithm
- (saves a few cycles per operation).
Represent trellis as sparse matrix;
- (saves memory, could improve cache performance).
Extension to data streams (finite traceback length).
Experimental study on different processors.

J. Feldman, VTC, 9/25/02 – p.11/13

Experimental Comparison (best case)
Decoder Pentium III PowerPC 7400 StrongARM
Lazy 6 201 200 226
Karn Generic 6 1143 626 892
Viterbi Optimized 6 316 239 310
Lazy 7 205 203 232
Karn Generic 7 2108 1094 1535
Karn Optimized 7 558 486 641
Karn SSE 7 108 N/A N/A
Lazy 9 235 225 343
Karn Generic 9 8026 3930 5561
Karn SSE 9 310 N/A N/A
(processor cycles per information bit)
J. Feldman, VTC, 9/25/02 – p.12/13
 Conclusion

A new fast convolutional decoder.

Decodes optimally (maximum-likelihood).
Generalizations:
- Soft-decision decoding.
- Data streams (non-block codes).
- Equalizers, or any other application of Viterbi
algorithm.
Takes advantage of the flexibility of software radio.

J. Feldman, VTC, 9/25/02 – p.13/13