byte-size

Crates.iobyte-size
lib.rsbyte-size
version0.2.7
sourcesrc
created_at2023-02-01 16:12:16.36848
updated_at2023-02-06 01:21:18.005085
descriptionAn effective short string shrinker with total disregard for speed, memory usage and executable size
homepage
repositoryhttps://github.com/ray33ee/byte-size
max_upload_size
id773893
size5,168,029
Will (ray33ee)

documentation

README

Byte Size

Crates.io Docs.rs License

A short string compressor/decompressor that can store 20,000+ words in three bytes or less.

Similar to smaz, byte-size is able to compress small strings, something that other conventional compression algorithms struggle with.

However, byte-size is typically better than smaz, certainly for very commonly used words (out of 10000 most common words, less than 1% had better compression with smaz) byte-size can also represent numbers, repeated sequences and non-alphanumeric characters more efficiently than smaz. It can encode unicode characters, but not very efficiently. If your text includes a few unicode characters it should still compress better, but if your strings are mostly unicode characters, other schemes such as Unishox are better.

Cost

byte-size uses several tables with over 18000 total entries. Obviously this will incur a large runtime memory and binary file size cost, but if you have the memory available, it is worth it to compress more effectively.

Examples

Using examples directly from smaz we have:

[Insert examples]

We can see how every example is compressed more with byte-size than smaz.

How?

It's basically just two tables one of about a thousand most commonly used lemmas (expressible as 2 bytes) and another of 10s of thousands of lemmas (expressible as 3 bytes)

On top of that, we have a few commonly used 2 and 3 byte sequences expressible as just 1 byte, that can be used as lemma prefix/sufixes, or can be used to construct words not in either list.

There are 3 lists:

  • One byte wonders (OBW): Made up of the printable ascii characters (with a few control sequences), common prefix/suffixes and common bigrams.
  • Two byte common (TBC): Made up of 1793 of the most common lemmas (that aren't also in the OBW list)
  • Three byte uncommon (TBU): Made up of 16512 of the most common lemmas (that aren't in either previous lists)

These lists are stored in the package root directory. These lists can be modified and these modifications will work. Lists are represented as a file, where each line is a new lemma encoded via percent encoding (to allow non printable characters and unicode sequences)

Encoding

  • the one byte wonder sequences are taken from smaz
    • We filtered out certain sequences that don't work well
  • the two byte common and three byte uncommon words are chosen from here
    • Words are only chosen from the list if their representation is more compact (for example the word 'the' does not appear in either 2 byte or 3 byte tables as it can be represented in 1 byte)

The Snaz encoding is as follows:

  • The one byte wonders consist of 240 one byte sequences. These one byte wonders are the ascii values in their normal positions, with other common sequences filling the gaps to
    • This leaves 16 values remaining for multi-byte codes.
      • One of these values is used to indicate a Unicode scalar value will follow
      • This leaves 15*256=3840 combinations of two byte sequences, which are divided as such:
        • 3586 are used to encode the 2 byte common words, 1793 with a space prefix and 1793 without
        • 32 are used to encode custom words
        • 32 are used to encode sequence repetitions of anywhere between 4 and 35 repeating units
        • 32 are used to encode numbers. 32 values means 5 bits in total, 3 for the number of bytes used, and 2 for the number itself
        • 29 are used to encode the non-printable control characters
        • 129 are used for the 3 byte codes. 129*256 = 33024 combinations
          • All 33024 combinations are used for the 3 byte uncommon words, 16512 with a space prefix, and 16512 without
Commit count: 18

cargo fmt