srzip 0.3 - Implementation notes


Stepan Roh

Version 0.3
Document revision 1.1 (2005/01/15)

Method bwt

Whole code for bwt method is in files m_bwt.[hc] (besides that code of method ahc from files m_ahc.[hc] is used). For generic introduction to bwt I recommend [1].

Individual compression stages

Compression is block-based with block size proportional to compression quality. Burrows-Wheeler transformation (see [2]) is implemented in function do_bwt(in_buf, out_buf, len, idx_buf), which does a transformation of input buffer in_buf (of length len) to output buffer out_buf. In comparison with basic transformation implementation there are these enhancements: ring buffer is not new buffer, bur array of pointers to input buffer (idx_buf), for quicker sorting (standard qsort() is used) eof character is appended to the input buffer which has higher value than any other character, this way it is not necessary to compare ring strings, but only input suffixes. It is necessary to retain this eof character during whole compression and decompression. This cause double speed gain. This enhancement is based on a thought from Mark Nelson's article (see [3]). Even with this enhancement the compression is very slow and that's because of quicksort. It has worst results on data with high ratio of repeated sections (it is necessary to compare long blocks). Sadly this is exactly the reason for this transformation's success and it is fairly normal for common data. For that reason different sorting method was tried and that's suffix tree. Sadly the time gain is not too high, but memory consumption is (around 40 times higher than quicksort). Complexity comparison:

  Complexity comparison    
     qsort     stree  
time    O(N*log(N))   O(N)  
space     O(N)     O(N)  

This does not look as bad, but in absolute numbers that's different: stree consumes around 20 bytes per node (in my implementation, normally it is 16 bytes, but with additional demands during tree walk) with max 2*N nodes. Suffix tree implementation is in directory bwt_as_stree, but beware that it sorts some data badly and thus generates wrong output.

Then follows Move-to-front coding implemented by function do_mtf(buffer, len), which is not differerent from common implementations.

Finally the output is coded by ahc (adaptive huffman coding) whereas at first zero strings are coded as pair (0, length-1), that's more or less what is suggested in [1]. Zero strings are very common in mtf's output so this improves compression ratio approximately by quarter.

Individual decompression stages

Compressed data are at first decompressed by ahc, coded zero strings are decoded on-the-fly.

Then goes reverse mtf (function rev_mtf(buffer, len)), nothing interesting.

Reverse Burrows-Wheeler transformation is implemented by function rev_bwt(in_buf, op, out_buf, buf_len, tbuf), which transforms in_buf to out_buf, op is original data position in sorted buffer so it's also a position of the only appeareance of eof character in the buffer. Transformation is done by constructing the transformation vector in tbuf.

Speed and achievements of the method

Compression is very slow: in comparison to gzip (which uses dictionary-based compression) or bzip2 (which uses bwt) it can be almost 50 times slower - this shows the slow sorting by qsort(), bzip2 uses qsort too, but modified (with higher memory consumption). Compression ratio is higher than gzip's and comparable with bzip2's (worse by tenth of percents). Decompression is very quick, comparable to other programs.



This document was generated using AFT v5.095