Home | History | Annotate | Download | only in doc
      1 1. Compression algorithm (deflate)
      2 
      3 The deflation algorithm used by gzip (also zip and zlib) is a variation of
      4 LZ77 (Lempel-Ziv 1977, see reference below). It finds duplicated strings in
      5 the input data.  The second occurrence of a string is replaced by a
      6 pointer to the previous string, in the form of a pair (distance,
      7 length).  Distances are limited to 32K bytes, and lengths are limited
      8 to 258 bytes. When a string does not occur anywhere in the previous
      9 32K bytes, it is emitted as a sequence of literal bytes.  (In this
     10 description, `string' must be taken as an arbitrary sequence of bytes,
     11 and is not restricted to printable characters.)
     12 
     13 Literals or match lengths are compressed with one Huffman tree, and
     14 match distances are compressed with another tree. The trees are stored
     15 in a compact form at the start of each block. The blocks can have any
     16 size (except that the compressed data for one block must fit in
     17 available memory). A block is terminated when deflate() determines that
     18 it would be useful to start another block with fresh trees. (This is
     19 somewhat similar to the behavior of LZW-based _compress_.)
     20 
     21 Duplicated strings are found using a hash table. All input strings of
     22 length 3 are inserted in the hash table. A hash index is computed for
     23 the next 3 bytes. If the hash chain for this index is not empty, all
     24 strings in the chain are compared with the current input string, and
     25 the longest match is selected.
     26 
     27 The hash chains are searched starting with the most recent strings, to
     28 favor small distances and thus take advantage of the Huffman encoding.
     29 The hash chains are singly linked. There are no deletions from the
     30 hash chains, the algorithm simply discards matches that are too old.
     31 
     32 To avoid a worst-case situation, very long hash chains are arbitrarily
     33 truncated at a certain length, determined by a runtime option (level
     34 parameter of deflateInit). So deflate() does not always find the longest
     35 possible match but generally finds a match which is long enough.
     36 
     37 deflate() also defers the selection of matches with a lazy evaluation
     38 mechanism. After a match of length N has been found, deflate() searches for
     39 a longer match at the next input byte. If a longer match is found, the
     40 previous match is truncated to a length of one (thus producing a single
     41 literal byte) and the process of lazy evaluation begins again. Otherwise,
     42 the original match is kept, and the next match search is attempted only N
     43 steps later.
     44 
     45 The lazy match evaluation is also subject to a runtime parameter. If
     46 the current match is long enough, deflate() reduces the search for a longer
     47 match, thus speeding up the whole process. If compression ratio is more
     48 important than speed, deflate() attempts a complete second search even if
     49 the first match is already long enough.
     50 
     51 The lazy match evaluation is not performed for the fastest compression
     52 modes (level parameter 1 to 3). For these fast modes, new strings
     53 are inserted in the hash table only when no match was found, or
     54 when the match is not too long. This degrades the compression ratio
     55 but saves time since there are both fewer insertions and fewer searches.
     56 
     57 
     58 2. Decompression algorithm (inflate)
     59 
     60 2.1 Introduction
     61 
     62 The key question is how to represent a Huffman code (or any prefix code) so
     63 that you can decode fast.  The most important characteristic is that shorter
     64 codes are much more common than longer codes, so pay attention to decoding the
     65 short codes fast, and let the long codes take longer to decode.
     66 
     67 inflate() sets up a first level table that covers some number of bits of
     68 input less than the length of longest code.  It gets that many bits from the
     69 stream, and looks it up in the table.  The table will tell if the next
     70 code is that many bits or less and how many, and if it is, it will tell
     71 the value, else it will point to the next level table for which inflate()
     72 grabs more bits and tries to decode a longer code.
     73 
     74 How many bits to make the first lookup is a tradeoff between the time it
     75 takes to decode and the time it takes to build the table.  If building the
     76 table took no time (and if you had infinite memory), then there would only
     77 be a first level table to cover all the way to the longest code.  However,
     78 building the table ends up taking a lot longer for more bits since short
     79 codes are replicated many times in such a table.  What inflate() does is
     80 simply to make the number of bits in the first table a variable, and  then
     81 to set that variable for the maximum speed.
     82 
     83 For inflate, which has 286 possible codes for the literal/length tree, the size
     84 of the first table is nine bits.  Also the distance trees have 30 possible
     85 values, and the size of the first table is six bits.  Note that for each of
     86 those cases, the table ended up one bit longer than the ``average'' code
     87 length, i.e. the code length of an approximately flat code which would be a
     88 little more than eight bits for 286 symbols and a little less than five bits
     89 for 30 symbols.
     90 
     91 
     92 2.2 More details on the inflate table lookup
     93 
     94 Ok, you want to know what this cleverly obfuscated inflate tree actually
     95 looks like.  You are correct that it's not a Huffman tree.  It is simply a
     96 lookup table for the first, let's say, nine bits of a Huffman symbol.  The
     97 symbol could be as short as one bit or as long as 15 bits.  If a particular
     98 symbol is shorter than nine bits, then that symbol's translation is duplicated
     99 in all those entries that start with that symbol's bits.  For example, if the
    100 symbol is four bits, then it's duplicated 32 times in a nine-bit table.  If a
    101 symbol is nine bits long, it appears in the table once.
    102 
    103 If the symbol is longer than nine bits, then that entry in the table points
    104 to another similar table for the remaining bits.  Again, there are duplicated
    105 entries as needed.  The idea is that most of the time the symbol will be short
    106 and there will only be one table look up.  (That's whole idea behind data
    107 compression in the first place.)  For the less frequent long symbols, there
    108 will be two lookups.  If you had a compression method with really long
    109 symbols, you could have as many levels of lookups as is efficient.  For
    110 inflate, two is enough.
    111 
    112 So a table entry either points to another table (in which case nine bits in
    113 the above example are gobbled), or it contains the translation for the symbol
    114 and the number of bits to gobble.  Then you start again with the next
    115 ungobbled bit.
    116 
    117 You may wonder: why not just have one lookup table for how ever many bits the
    118 longest symbol is?  The reason is that if you do that, you end up spending
    119 more time filling in duplicate symbol entries than you do actually decoding.
    120 At least for deflate's output that generates new trees every several 10's of
    121 kbytes.  You can imagine that filling in a 2^15 entry table for a 15-bit code
    122 would take too long if you're only decoding several thousand symbols.  At the
    123 other extreme, you could make a new table for every bit in the code.  In fact,
    124 that's essentially a Huffman tree.  But then you spend too much time
    125 traversing the tree while decoding, even for short symbols.
    126 
    127 So the number of bits for the first lookup table is a trade of the time to
    128 fill out the table vs. the time spent looking at the second level and above of
    129 the table.
    130 
    131 Here is an example, scaled down:
    132 
    133 The code being decoded, with 10 symbols, from 1 to 6 bits long:
    134 
    135 A: 0
    136 B: 10
    137 C: 1100
    138 D: 11010
    139 E: 11011
    140 F: 11100
    141 G: 11101
    142 H: 11110
    143 I: 111110
    144 J: 111111
    145 
    146 Let's make the first table three bits long (eight entries):
    147 
    148 000: A,1
    149 001: A,1
    150 010: A,1
    151 011: A,1
    152 100: B,2
    153 101: B,2
    154 110: -> table X (gobble 3 bits)
    155 111: -> table Y (gobble 3 bits)
    156 
    157 Each entry is what the bits decode as and how many bits that is, i.e. how
    158 many bits to gobble.  Or the entry points to another table, with the number of
    159 bits to gobble implicit in the size of the table.
    160 
    161 Table X is two bits long since the longest code starting with 110 is five bits
    162 long:
    163 
    164 00: C,1
    165 01: C,1
    166 10: D,2
    167 11: E,2
    168 
    169 Table Y is three bits long since the longest code starting with 111 is six
    170 bits long:
    171 
    172 000: F,2
    173 001: F,2
    174 010: G,2
    175 011: G,2
    176 100: H,2
    177 101: H,2
    178 110: I,3
    179 111: J,3
    180 
    181 So what we have here are three tables with a total of 20 entries that had to
    182 be constructed.  That's compared to 64 entries for a single table.  Or
    183 compared to 16 entries for a Huffman tree (six two entry tables and one four
    184 entry table).  Assuming that the code ideally represents the probability of
    185 the symbols, it takes on the average 1.25 lookups per symbol.  That's compared
    186 to one lookup for the single table, or 1.66 lookups per symbol for the
    187 Huffman tree.
    188 
    189 There, I think that gives you a picture of what's going on.  For inflate, the
    190 meaning of a particular symbol is often more than just a letter.  It can be a
    191 byte (a "literal"), or it can be either a length or a distance which
    192 indicates a base value and a number of bits to fetch after the code that is
    193 added to the base value.  Or it might be the special end-of-block code.  The
    194 data structures created in inftrees.c try to encode all that information
    195 compactly in the tables.
    196 
    197 
    198 Jean-loup Gailly        Mark Adler
    199 jloup (a] gzip.org          madler (a] alumni.caltech.edu
    200 
    201 
    202 References:
    203 
    204 [LZ77] Ziv J., Lempel A., ``A Universal Algorithm for Sequential Data
    205 Compression,'' IEEE Transactions on Information Theory, Vol. 23, No. 3,
    206 pp. 337-343.
    207 
    208 ``DEFLATE Compressed Data Format Specification'' available in
    209 http://tools.ietf.org/html/rfc1951
    210