Release Notes for LDPC Software

http://www.cs.toronto.edu/~radford/ldpc.software.html

Release of 2006-02-08.

The following new features have been added:

1. The make-ldpc program can new create LDPC codes with varying numbers of checks per bit. The examples now include a code of this sort.
2. New programs pchk-to-alist and alist-to-pchk have been added, which convert to and from alist format.
3. A run-examples script has been added, which runs all the examples and compares with the saved output that I obtained.

Release of 2001-11-18.

The following new feature has been added:

1. A new -t option has been added to print-pchk, which causes it to print the transpose of the parity check matrix.

1. The behaviour of mod2dense_invert_selected has been changed when the matrix has dependent rows. The result returned contains the inverse of a sub-matrix in its initial columns. The contents of the remaining columns up to the number of rows was previously undefined, but these columns are now set to contain zeros. This was done to fix bug (1) below.

1. A bug has been fixed in make-gen which sometimes produced invalid `dense' or `mixed' generator matrices when the parity check matrix had redundant rows. The `sparse' generator matrices were not affected. The output of the ex-dep example has changed slightly as a result.
2. The open mode for reading the file of random numbers has been changed from "r" to "rb", for compatibility with non-Unix systems.
3. mod2sparse_copy has been fixed so that it copies the lr and pr fields of entries.
4. The "set -v -e" in the command files for the examples has been replaced by two commands, "set -e" and "set -v", since some shells can't set two options at once.

Release of 2001-05-04.

1. Warnings have been added to the documentation about the dangers of using zero codewords for testing (due to the possibility of the decoder accidently being non-symmetric). The examples have been changed to not use zero codewords, except for ex-ham7a, which still demonstrates this facility.
2. Decoding by probability propagation has been changed so that if a probability ratio gets set to NaN ("not a number", the result of infinite but conflicting evidence from different sources), the probability ratio is changed to 1, in hopes that something sensible may result.
3. The decoding procedures have been changed to decode a bit to a 1 if there is a tie, so that any resulting asymmetry in decoding will make zero decoding less, rather than more, likely. See point (1) above.
4. All programs now allow a file name argument to be "-", which specifies standard input or standard output, as appropriate. The ex-ldpc36-5000a example has been changed to use this facility in order to avoid creating lots of files.
5. Channels with noise from a logistic distribution (AWLN) have been implemented. For details, see the channel documentation.
6. The make-gen program for "dense" or "mixed" options was modified so that when some parity checks are redundant, others are not declared redundant when they aren't. This makes the result compatible with what is obtained with the "sparse" option. The example in the discussion of linear dependence in parity check matrices has changed slightly as a result.
7. The messages displayed by the decode program have changed slightly: The block size is no longer displayed, and the percentage of bits changed is now shown. The data output when the -t option is specified now includes the number of bits changed in each block.
8. A new -T option to decode for displaying detailed information on decoding has been added.
9. Doing a make clean will now remove the files created by running the examples.
10. The files created by the examples now start with "ex-" and contain a ".", to make them easier to identify for cleanup, and the saved output file now ends in "-out" rather than ".out".
11. The scripts for the examples now use the -v option of set rather than -x, so that the command lines are echoed faithfully as they are read.
12. Two new examples have been added: ex-dep demonstrates what happens when parity check matrices have redundant row, and ex-wrong-model investigates what happens when the wrong model is used to decode.

1. Instructions have been added on how to modify the programs, and some code has been reorganized to facilitate modifications.
2. The makefile has been made trivial, simply recompiling everything from scratch every time. This is simpler and less error prone than putting in dependencies. Compiling all the programs takes only about five seconds on a modern PC.
3. The mod2dense_invert_selected procedure has been changed to take an additional argument in which a permutation of rows is returned. This extension is needed in order for the result when the matrix is not of full rank to include as many rows as possible, as needed for the above modification of make-gen.
4. New routines mod2dense_copyrows and mod2sparse_copyrows have been added, analogous to the previous routines for copying columns.
5. The decoding procedures have been changed to always return the bit probabilities (previously, one could pass a null pointer to suppress this). This makes things simpler, and facilitates the production of detailed trace output when the -T option is used.

1. A problem with print-gen prevented the -d option from working in most cases. The documentation was also fixed to note what the L and U matrices printed will look like.
2. The documentation for mod2sparse_copycols and mod2dense_copycols was ambiguous. It has been clarified. The dense and sparse implementations were inconsistent. The sparse version has been changed to match the dense version, which was and is the only version actually used in the LDPC software.
3. In decode.c, a variable declared to be float was printed using "%f", which strangely didn't work with one compiler. It's now declared as double.

Old release of 2000-03-19.