optimizing byte-pair encoding

I've done work with optimizing a LZF compression implementation, and some of the same principles I used to improve performance are usable here.

To speed up performance on byte-pair encoding:

  1. Limit the block size to less than 65kB (probably 8-16 kB will be optimal). This guarantees not all bytes will be used, and allows you to hold intermediate processing info in RAM.
  2. Use a hashtable or simple lookup table by short integer (more RAM, but faster) to hold counts for a byte pairs. There are 65,656 2-byte pairs, and BlockSize instances possible (max blocksize 64k). This gives you a table of 128k possible outputs.
  3. Allocate and reuse data structures capable of holding a full compression block, replacement table, byte-pair counts, and output bytes in memory. This sounds wasteful of RAM, but when you consider that your block size is small, it's worth it. Your data should be able to sit entirely in CPU L2 or (worst case) L3 cache. This gives a BIG speed boost.
  4. Do one fast pass over the data to collect counts, THEN worry about creating your replacement table.
  5. Pack bytes into integers or short ints whenever possible (applicable mostly to C/C++). A single entry in the counting table can be represented by an integer (16-bit count, plus byte pair).

This is a summary of my progress so far:

Googling found this little report that links to the original code and cites the source:

Philip Gage, titled 'A New Algorithm for Data Compression', that appeared in 'The C Users Journal' - February 1994 edition.

The links to the code on Dr Dobbs site are broken, but that webpage mirrors them.

That code uses a hash table to track the the used digraphs and their counts each pass over the buffer, so as to avoid recomputing fresh each pass.

My test data is enwik8 from the Hutter Prize.

|----------------|-----------------|
| Implementation | Time (min.secs) |
|----------------|-----------------|
| bpev2          | 1.24            | //The current version in the large text benchmark
| bpe_c          | 1.07            | //The original version by Gage, using a hashtable
| bpev3          | 0.25            | //Uses a list, custom sort, less memcpy
|----------------|-----------------|

bpev3 creates a list of all digraphs; the blocks are 10KB in size, and there are typically 200 or so digraphs above the threshold (of 4, which is the smallest we can gain a byte by compressing); this list is sorted and the first subsitution is made.

As the substitutions are made, the statistics are updated; typically each pass there is only around 10 or 20 digraphs changed; these are 'painted' and sorted, and then merged with the digraph list; this is substantially faster than just always sorting the whole digraph list each pass, since the list is nearly sorted.

The original code moved between a 'tmp' and 'buf' byte buffers; bpev3 just swaps buffer pointers, which is worth about 10 seconds runtime alone.

Given the buffer swapping fix to bpev2 would bring the exhaustive search in line with the hashtable version; I think the hashtable is arguable value, and that a list is a better structure for this problem.

Its sill multi-pass though. And so its not a generally competitive algorithm.

If you look at the Large Text Compression Benchmark, the original bpe has been added. Because of it's larger blocksizes, it performs better than my bpe on on enwik9. Also, the performance gap between the hash-tables and my lists is much closer - I put that down to the march=PentiumPro that the LTCB uses.

There are of course occasions where it is suitable and used; Symbian use it for compressing pages in ROM images. I speculate that the 16-bit nature of Thumb binaries makes this a straightforward and rewarding approach; compression is done on a PC, and decompression is done on the device.