United States Patent
`
`[:9]
`
`Che
`
`[54] LOSSLESS DATA COMPRESSION CIRCUIT
`AND METHOD
`
`[75]
`
`Inventor: Ke-Chiang Chit. Saratoga, Calif.
`
`[73] Assignee: The Board of Trustees of the Leland
`Stanford Junior University, Stanford,
`Calif.
`
`[21] App]. No.1 670,281
`
`[22] Filed:
`
`Mar. 15. 1991
`
`Int. Cl.5 ............................................ .. HOSM 7/36
`[51]
`[52] U.S. Cl. ...................................... .. 382/56; 341/51;
`341/67
`Field of Search .................... .. 382/56; 341/51. 61?,
`341/95. l06; 353/261.]. 261.4, 261.2
`
`{S8}
`
`[56]
`
`References Cited
`U.S. PATENT DOCUMENTS
`
`4.558.302 ‘.2./I985 Welch ................................. .. 3-1-U95
`4.941.193 T/1990 Barnsley et al
`38256
`5.C0l.47B 311991 Nagy ............... ..
`341.351
`5.003.307
`3/1991 Whiting et al.
`34l;'5l
`5.051.745 9/1991 Kaiz
`341.351
`
`
`
`Primary Examiner-—Leo H. Bondreau
`Assistant Exam:‘ner—Steven P. Klocinski
`Attorney. Agent. or .Fl‘rm——Flehr. Hohbach, Test,
`Albritton & Herbert
`
`[57]
`
`ABSTRACI
`
`ll|||||||||l|I|||||||||l|lll|||||||||||||ll||||llllll||||||l||||l||||ll||||
`usoosi sn43oA
`
`[11] Patent Number:
`
`5,150,430
`
`[45] Date of Patent:
`
`Sep. 22, 1992
`
`data string with a set of comparison data. and produces
`a sequence of codewords representing a sequence of
`successive. non-overlapping substrings of the new data
`string. A shift register stores and shifts the comparison
`data until all of the characters in the comparison data
`have been compared with the new data string. A com-
`posite reproduction length circuit finds the maximum
`length string within the set of comparison characters
`matching substrings of characters in the new data string
`beginning at each position in the new data string. The
`composite reproduction length circuit produces a multi~
`plicity of data pairs. one for each position of the new
`data string. Each data pair comprises a maximum length
`value. corresponding to the maximum length matching
`comparison string found for the new data substring
`starting at the corresponding position. and a pointer
`value denoting where the maximurri length matching
`comparison string is located in the comparison data. A
`codeword generator then generates a sequence of code-
`words representing the new data string, each codeword
`including one of these data pairs and representing a
`substring of the new data string. By using N such data
`compression units in parallel, with each storing an idcn~
`tical new data string. and each unit's shift register stor-
`ing and shifting a different subset of a specified compari-
`son string, processing time for generating codewords is
`reduced by a factor of approximately (N— l)/N.
`
`A lossless data compression circuit compares a new
`
`9 Claims. 6 Drawing Sheets
`
`200
`2o2/"/
`
`NEW DATA
`
`0: MATCI-i0...55
`
`Oracle 1008
`
`Oracle 1008
`
`
`
`BUFFER CODEWORD
`
`
`
`

`
`U.S. Patent
`
`Sep. 22, 1992
`
`Sheet 1 of 6
`
`5,150,430
`
`EIlIflEllllIIBElIIl!— '“'°”T5T“"“G
`
`-:ptr,length,tai|_character>=codeword=<125.5."r">
`
`If
`I
`I
`
`-lI|flBIfl!flIllfl;';1BIlII—
`
`
`
`90
`
`HISTORY BUFFER
`
`100
`
`/V/120
`
`CLK
`
`MAXIMUM REPRODUCTION _
`LENGTH REGISTER
`
`POINTER REGISTER
`
`E0
`
`%
`3|;
`w|.L
`asO
`O
`
`

`
`U.S. Patent
`
`Sep. 22, 1992
`
`Sheet 2 of 5
`
`5,150,430
`
`1 54
`
`152
`
`CODEWORD REGISTER
`
`coogwoan BUFFER
`
`CLK2
`
`L
`
`‘I60
`
`T
`
`
`
`1 64
`
`
`
`ADDRESS GENERATOR
`
`
`
`
`
`
`170
`
`MESSAGE
`
`
`BUFFER
`
`neconen
`SYMBOLS
`
`HIST
`
`I N
`
`OFIYBUFFER
`
`MUX
`
`T
`
`
`
`
`
`CODEWOFID
`
`BUFFER
`
`

`
`U.S. Patent
`
`Sep. 22, 1992
`
`Sheet 3 of 5
`
`5,150,430 I
`
`LD_CW1 .
`LD_Cw2.
`cw Fm
`230
`
`200
`
`\/\ SHlFT_NEW_DATA
`
` COUNTER CODEWOHD
`GENEFIATO‘
`
`
`
`ll
`
`
`LEN
`
`SHII-'|'_NEW__DATA.
`SHIP-'l'_OLD_DATA
`
`u
`
`P E
`
`FHOFNTY
`NODDER
`
`
`I
`
`EXTERNAL
`cooewono
`INTERFACE
`
`3°34’
`NEW BAND}
`
`
`
`
`3°‘
`
`LD_CW2
`
`CW_HDY-—§
`
`FIGURE 10
`
`

`
`U.S. Patent
`
`Sep. 22, 1992
`
`Sheet 4 of 6
`
`5,150,430
`
`NEW DATA i-1
`
`NEW DATA i
`
`206
`
`
`
`
`
`ow DATA
`SHIFT
`REGISTER
`
`cmcum
`
`SH! FT_N EW_DATA
`
`OLD DATA i-1
`
`SHI FT_OL D_DATA
`CLK
`
`
`
`224
`
`W
`
`sun-r_uew
`__DA'l'A
`
`
`
` PTR(i—1)
`
`COUNTER
`(11 :0] FIGURE 9
`
`
`

`
`U.S. Patent
`
`Sep. 22, 1992
`
`Sheet 5 of 5
`
`5,150,430
`
`m?
`
`MATCH55
`
`MATCH54
`
`MATCH53
`
`MATCH52
`
`MATCH51
`
`MATCH50
`
`VCC
`
`
`
`In.I..III.IIInIn‘mvID._.¢2
`
`eififitl
`
`InI‘._IIIIIIImv_._Oh<E
`
`..f.ur.u.vm“
`
`vIO._.<2
`
`....rWm..mm§
`
`ufwfiwwfi
`
`235i.'..I..‘.lII.....I.I...IIo_IO.—.<s_
`
`
`
`....mm£m...w....v
`
`FIGURE 8B
`
`TALLYL]
`
`..uMWLAT
`
`FIGURE 8A
`
`
`

`
`U.S. Patent
`
`Sep. 22, 1992
`
`Sheet 6 of6
`
`5,150,430
`
`
`
`320-N
`
`
`
`II
`
`OLD DATA
`
`PT Fl
`
`
`
`
`
`
`
`NEW DATA
`
`OLD DATA
`
`NEW DATA
`
`
` CODEWORDT
`
`
`
`I -
`
`I —
`
`-
`
`
`
`MRL I
`
`
`
`BEST
`BEST
`conewono
`cooewono
`
`SELECTOR
`SELECTOR
`3557
`
`(wnm LARGEST
`(wnm LARGEST
`°°°E“'°F‘D
`
`MRL)
`MRL)
`
`
`
`FIGURE 11
`
`

`
`1
`
`5,150,430
`
`LOSSLESS DATA COMPRESSION CIRCUIT AND
`METHOD
`
`The present invention relates generally to data com-
`pression systems and circuits and particularly to a cir-
`cuit and method for performing a number of data com-
`pression operations in parallel so as to substantially
`reduce the computation time required for performing
`lossless data compression.
`BACKGROUND OF THE INVENTION
`
`The term “lossless data compression“ means that all
`of the information in an input string or message is re-
`tained when that information is compressed so as to be
`represented by a smaller amount of data. Thus, when
`the compressed data is decoded, the reconstructed mes-
`sage will be identical to the original input message rep-
`resented by the compressed data.
`The present invention is. in essence, a highly efficient
`system and method for implementing a well known
`lossless data compression technique, sometimes known
`as the "Lempel and Ziv noisless universal data compres-
`sion algorithm". This data compression algorithm is
`also sometimes called the “history buffer“ algorithm.
`THE LEMPEI. AND ZIV LOSSLBSS DATA
`COMPRESSION METHOD.
`
`Referring to FIG. 1. this data compression algorithm
`works as follows. The data compression unit (or pro-
`gram) stores the last X (e.g., I024) characters (some-
`times called symbols) that have already been encoded in
`a "history buffer“ 90, typically implemented as a circu-
`lar buffer. When new input data is encoded. the data
`compression unit looks for the longest string in the
`history buffer 9|] that matches the new input data, be-
`ginning with the first character of that new input data.
`For instance, if the new input data is the string
`
`"the front axle of the vehicle snapped when .
`
`.
`
`. "
`
`the data compression unit would look for the longest
`matching string beginning with the letters "th .
`.
`. ". If
`the best match found in the history buffer was, say. “the
`f" (e.g.. as pan‘. of the phrase “the lirsl: time .
`.
`. "), then
`the "reproduction length" of the longest matching
`string would be five characters; “t“, “h“, “e“. "", and
`"f". The form of the codeword used to represent the
`compressed data is as follows:
`
`(pointer. reproduction length. tail chal'acIe1').
`
`Thus the codeword comprises (1) a pointer to a position
`in the history buffer where a matching string can be
`found, (2) the length of the matching string. and (3) a
`tail character which is the next character in the input
`string. One reason for including the tail character is to
`handle cases in which a character is not found in the
`history buffer. In such cases, the pointer and length
`fields of the codeword are set equal to "0.0“ and the tail
`character field is equal to the first character of the input
`string.
`Each time that a portion of the input string is con-
`verted into a codeword, the input string is shifted so
`that the beginning of the input string is the first charac-
`ter that has yet to be convened into codeword form.
`Furthermore,
`the contents oi’ the history buffer are
`periodically updated by adding the portions ofthe input
`string that have been processed to the end of the history
`
`10
`
`15
`
`20
`
`25
`
`3|]
`
`35
`
`40
`
`45
`
`SD
`
`55
`
`60
`
`65
`
`2
`buffer. Since the length of the history buffer is typically
`constant. adding new data at the end of the buffer usu-
`ally means that data at the beginning of the buffer is
`erased or overwritten. unless the history buffer is not
`yet full; Depending on the particular implementation of
`the data compression algorithm,
`the contents of the
`history buffer may be updated each time that a new
`codeword is generated. or it may be updated each time
`that a certain number of input characters have been
`processed. For instance.
`the contents of the history
`buffer may be updated every time that another thirty-
`two input string characters have been processed by the
`data compression algorithm.
`The net result of the data compression algorithm is
`that an input string is convened into a sequence of
`codewords, as described above. To see how this pro-
`duces data compression, consider the following exam-
`ple. Assume that the history buffer contains H124 char-
`acters, that the maximum allowed reproduction length
`will be seven, and that it takes eight bits to represent an
`uncompressed character. The length of each codeword
`will be:
`
`it] hits
`3 bits
`3 bits
`2] bits
`
`Pointer to history buffer position
`Reproduction length value
`Tail character
`length of each codeword
`
`If. on average. each codeword has a reproduction
`length of four characters. each codeword will represent
`a total of five characters, including the tail character.
`Thus, on average, it will take 21 bits to represent five
`characters instead of 40 bits. representing a compression
`ratio of about two to one (2:1).
`DECODING COMPRESSED DATA
`
`Data decoding. which is the process of converting
`the codewords back into regular character strings,
`is
`very easy. The data decoding unit or program maintains
`a history buffer that is identical to the one maintained by
`the data compression unit, updating the contents of the
`history buffer with newly decoded characters. Each
`codeword is decoded simply by using the pointer in the
`codeword to locate a position in its history buffer. Then
`the specified number of characters in the history buffer
`are copied, following by the specified tail character. to
`regenerate the string represented by the codeword.
`Data decoding is very fast and uncomplicated.
`Note that we have not specified the contents of the
`history buffer when one first begins to compress or
`decompress data. This is a peripheral issue not relevant
`to the present invention, but is addressed here to en-
`hance the reader's understanding of the compression
`method. There are at least two standard methods of
`dealing with what to put in the history buffer at the
`beginning of the process: (I) initially clear the history
`buffer (e.g., fill it with null characters), or (2) always
`initially fill the history buffer in both the data compres-
`sion unit and in the data decoding unit with a predefined
`set of data that is known to produce reasonably good
`data compression with a wide variety of input data. The
`first solution results in a lower data compression ratio
`when small sets of data are compressed. while the sec-
`ond requires a certain amount of coordination between
`the data compression and data decoding units. Either
`
`

`
`5,150,430
`
`3
`solution. or any other such solution. can be used in
`conjunction with the present invention.
`As will be understood by those skilled in the art. there
`is a set of tradeoffs between the length of history buffer,
`the maximum allowed reproduction length, the com-
`pression ratio and the amount of processing time re-
`quired to perform the data compression. In particular, if
`the history buffer is made longer (say to 4096 charac-
`ters). the chances of finding good matching strings in-
`crease, and the average reproduction length will in-
`crease. However, this will also increase the length of
`each codeword, because more bits will be required to
`represent positions in the history buffer. and it will take
`more processing time to perform the compression be-
`cause it will take longer to find the best matching string
`in a larger history buffer.
`PROCESSING TIME FOR DATA COMPRESSION
`
`The present invention is primarily concerned with
`the problem that compressing data using the Lempel
`and Ziv is very time consuming. As will be explained
`next using some rlurnerical examples. the Lempel and
`Ziv data compression method has been used in a mini-
`ber of contexts in which the speed at which data com-
`pression is accomplished is not critical. such as transmit-
`ting data over telephone lines using standard medium
`speed modems. The Lernpel and Ziv data compression
`method has generally not been used in high speed appli-
`cations. such as data storage on hard disks. because the
`amount of time required for data compression prior to
`storage on disk has been assumed to be excessive.
`Consider the following example of the processing
`time required to compress data for transmission over a
`24-01] baud {i.e.. 2400 bits per second) transmission line.
`which is typical of modem transmission used by desktop
`computers. What we are computing in this example is
`the number of instructions per second that would have
`to be executed by a software data compression program
`in order to keep up with a 2400 baud modem. We will
`assume a data compression ratio of two to one, meaning
`that 4800 bits of data, or about 600 characters must be
`encoded for transmission each second. Also assume a
`history buffer of 1024 characters. and that the average
`codeword represents five characters. Since the average
`codeword represents five characters. the compression
`unit must generate about 120 codewords per second. To
`generate a codeword. the compression unit must try to
`match an input string with strings of characters in the
`1024 character history buffer. A ballpark figure for
`performing this would be about 10.801) instructions:
`about four instructions for performing each character to
`character comparison, with an average of about 1.5 to
`2.0 such comparisons being required for each character
`in the history buffer, plus a small number of instructions
`required to keep track of the best match and to con-
`struct the codeword.
`
`Thus the number of executed instructions required to
`generate the 120 codewords needed on average each
`second would be about: 120): 10,000 or about 1.2 mil-
`lion instructions per second, requiring a computer capa-
`ble of at least 1.2 MIPS. Using an average of three or
`four computer clock cycles per instruction, 1.2 MIPS is
`equivalent to a CPU cycle rate of about -1- or 5 mega-
`hertz (i.e., 4 or 5 million computer clock cycles per
`second}. This is well within the capabilities of many
`desktop computers. with the fastest desktop computers
`in 1990 having CPU cycle rates of approximately 40
`megahertz. In comparison, data decoding requires only
`
`4
`about a hundred instructions per codeword, and thus is
`extremely fast.
`Next, we consider the processing requirements for
`using the Lernpel and Ziv data compression method for
`storing and retrieving data to and from a hard disk or a
`comparable medium in a computer system. The idea
`here is that if data compression can be used with such
`hard disks, then one could approximately double the
`amount of data stored on any hard disk.
`The hard disks used in desktop computers, as of the
`end l990, can store up to about I million characters per
`second. High perforrnance hard disks used with more
`expensive computer systems can store up to about 3.5
`million characters per second. but we will use a data
`storage rate of 1 million characters per second for the
`purposes of this example. Using a data compression
`scheme with a two to one compression ratio, the rate of
`data storage would be about 2 million characters per
`second. or about 400,000 codewords per second (using
`an average of five characters per codeword}. In order
`to avoid slowing down the rate of data storage to disk,
`data compression would need to be accomplished about
`3300 times faster than in the above 2400 baud modem
`example. Thus. performing “searnless" data compres-
`sion for hard disk storage without impacting on a com-
`puter's data storage perfonnance is much, much more
`challenging. Furthermore. it is clear that such data
`compression cannot be accomplished witlt software,
`since it would require that the computer's processor
`have a CPU cycle rate on the order of 10 gigaherrz.
`The present invention provides a data compression
`circuit using parallel processing techniques which is
`sufficient fast that it can keep up with the data storage
`rates of current hard disks. That is, using 2140 megahertz
`clock rate and the current invention, one can compress
`two million bytes or characters per second into the
`codewords associated with the Lempel and Ziv lossless
`data compression method.
`_
`It is noted that there is always a trade-off between
`implementation feasibility. or processing complexity, of
`a data compression system and compression ratio. The
`present invention provides data compression apparatus
`that is feasible for high speed. using a simple. hardward
`implementation.
`SUMMARY OF THE INVENTION
`
`In summary, the present invention is a lossless data
`compression circuit which produces a sequence of
`codewords representing a sequence of successive, non-
`overlapping substrings of a specified set of data. A first
`shift register stores a new input string. while a second
`shift register stores and shifts a prefix string until all of
`the characters in the prefix string data have been com-
`pared with the new data string.
`A composite reproduction length circuit finds the
`maximum lgth strings within the prefix string which
`match substrings of characters in the new data string
`beginning at each position in the new data string. The
`composite reproduction length circuit produces amulti-
`plieity of data pairs. one for each position of the new
`data string. Each data paircomprises a msrtimurn length
`value. corresponding to the maximum length matching
`prefix string found for the new data substring starting at
`the corresponding position. and a pointer value denot-
`ing where the maximum length matching string is lo-
`cated in the prefix string. A codeword generator then
`generates a sequence of codewords representing the
`new data string, each codeword including one of these
`
`at
`
`IO
`
`15
`
`20
`
`25
`
`30
`
`35
`
`4-5
`
`55
`
`65
`
`

`
`5
`data pairs and representing a substring of the new data
`string.
`In the preferred embodiment, the composite repro-
`duction length circuit has the following components. A
`reproduction length register stores a separate length
`value corresponding to each position of the new data
`string, and a pointer register stores a separate pointer
`value corresponding to each position of the new data
`string. A counter generates a pointer value denoting
`which subset of the prefix characters are currently
`stored in the shift register. A modified tallying circuit
`generates a tally value for each position in the new data
`string, each tally value representing how many consec-
`utive prefix characters in the shift register, up to a pre-
`defined maximum value. match characters in the new
`data string beginning at a corresponding position in the
`new data string. Finally. a decoder compares each tally
`value with a corresponding one of the length values
`stored in the reproduction length register. and when the
`telly value exceeds the corresponding stored length
`value, it replaces the stored corresponding length value
`with the tally value and replaces the corresponding
`pointer value stored in the pointer register with the
`pointer value generated by the counter.
`By using N such data compression units in parallel,
`with each storing an identical new data string, and each
`unit’s shift register storing and shifting a different subset
`of a. specified prefix string, processing time for generat-
`ing codewords is reduced by a factor of (N — I)/N (e.g.,
`using three parallel data compression units would re
`duce processing time by a factor of approximately §).
`BRIEF DESCRIPTION OF THE DRAWINGS
`
`Additional objects and features of the invention will
`be more readily apparent from the following detailed
`description and appended claims when taken in con-
`junction with the drawings. in which:
`FIG, 1 conceptually depicts how a codeword is gen-
`erated from an input string and a prefix string.
`FIG. 2 is a conceptual diagram of a computer system
`incorporating the lossless data compression and data
`decoding circuits of the present invention in the data
`path between a hard disk storage system and the com-
`puter system's internal‘ bus.
`FIG. 3 is a block diagram of a simple hardware cir-
`cuit for lossless data compression.
`FIG. 4 is a block diagram of a circuit for decoding
`compressed data generated using the circuit of FIG. 3.
`FIG. 5 is a conceptual diagram of a high speed loss-
`less data compression circuit which speeds data com-
`pression using one type of parallel data processing.
`FIG. 6 is a block diagram of a preferred embodiment
`of a high speed lossless data compression circuit using
`the firs! type of parallel data processing.
`FIG. 7 is a block diagram of a data comparison cir-
`cuit.
`FIGS. 8A and 83 are circuit schematics for a modi-
`fied tally circuit.
`FIG. 9 is a block diagram of a decoder for selecting
`the longest string in a set of comparison data which
`matches a specified new data string.
`FIG. 10 is a block diagram of codeword generator
`circuit.
`FIG. 11 is a block diagram of a high speed lossless
`data compression circuit which speeds data compres-
`sion nsing a second type of parallel data processing.
`
`5,150.4-30
`
`6
`
`‘ 5
`
`10
`
`15
`
`20
`
`25
`
`30
`
`35
`
`45
`
`S0
`
`55
`
`DESCRIPTION OF THE PREFERRED
`EMBODIMENT
`
`Referring to FIG. 2, we show a typical computer
`system 100 in which the present invention can be used.
`The computer system 10!} includes a CPU 1il2, memory
`module 104, an internal bus Ill}, and a hard disk storage
`system 112 which stores data on a hard disk 114. In the
`preferred embodiment, data being written to the hard
`disk 114 are processed by a data encoding circuit 120
`and data being retrieved from the hard disk are pro-
`cessed by a data decoding circuit 150. In the preferred
`embodiment, both the encoder 129 and decoder 150 are
`implemented on a single, integrated circuit 116. The
`combined encoder/decoder circuit 116 is used in the
`data path between a computer's internal bus 110 and its
`hard disk storage system 112. causing data stored on
`hard disk 114 to be automatically encoded and com-
`pressed as it is being written to the hard disk 114, and to
`be automatically decoded as it is retrieved. If the en-
`coder circuit 120 is sufficiently fast at performing the
`necessary encoding operations, the use of the data com-
`pression and decoding circuits will be “seamless“. in
`that their presence would not aifect the performance of
`the computer system 100, except that the amount of
`data that could be stored on the hard disk 114 would be
`approximately doubled (assuming a compression ratio
`of 2:1).
`Referring to FIG. 3. we will first describe a very
`simple hardware implementation of a data compression
`circuit 120 for performing the Letnpel and Ziv lossless
`data compression algorithm. The first data encoding
`circuit 120 is approximately thirty times faster than a
`comparable software data compression program. Later,
`we will show a data compression circuit which is faster
`than this first circuit, on average, by a factor often. This
`means that, on average, the improved circuit reduces
`the number of clock cycles required to produce the
`same codewords by about a factor of ten. As will also be
`described below, the speed of data compression can be
`further increased by an additional factor of four to ten,
`if one is willing to use a commensurate amount of addi-
`tional circuitry.
`The term "character" as used in this document is
`defined to mean any symbol or other form of data
`which occupies a predefined amount of storage space
`(i.e., a predefined number of hits), such as a byte or
`“word".
`In the following description. the string of characters
`that is being compressed is represented by the letter S.
`The string S contains characters at positions beginning
`at position 0 and continuing to position N— 1, where -N
`is the length of the string. A substring ofS which starts
`at position i and ends at position j is denoted as S(i,j).
`For convenience,
`the string currently stored in the
`history buffer is labelled H, and substrings in the history
`buffer are denoted H(i,_i). where i and j are the begin-
`ning and ending positions of the substring in the history
`buffer. The contents of the history bufier are sometimes
`called a proper prefix of the string which is currently
`being processed for data compression.
`Given a string S, the reproduction length of the string
`8 at position i is the largest integer I..,- such that
`
`65
`
`sa.i+L.-— n=Ha=.k+L.--1}
`
`where H represents the contents of the history buffer
`and It is any position within the history buffer. The
`
`

`
`5,150,430
`
`7
`position P,-in the history buffer corresponds to the maxi-
`mum value of the reproduction length L,-at position i.
`Let S=s1s;. .
`. be the string of source symbols that is
`being compressed. The sequential encoding of S parses
`S into successive source words, S=S1S2 .
`.
`.
`. For each
`source word S,-, the encoder looks for the longest repro-
`duction length L,-, and assigns a codeword C,-.
`
`5
`
`IO
`
`20
`
`25
`
`30
`
`35
`
`45
`
`SD
`
`55
`
`65
`
`The encoding procedure can be formally described as
`follows.
`1. Initially. SR2 is reset to be 0. SR1 stores the string to
`be parsed.
`2. The history buffer string is shifted into SR2 sequen-
`tially, at the rate of one symbol per clock cycle. RLG
`122 compares SR1 against SR2 at each clock cycle. A
`reproduction length L is generated. At the end of
`every clock cycle, the contents of Mill. and pointer
`registers 124 and 126 are updated. if necessary.
`3. After N clock cycles. where N is the length of the
`history buffer string. a codeword C; is generated.
`Ct= < P.'.Lr.T:'>-
`4. To update the contents of SR], the symbols occupy-
`ing the first L,-+l positions are shifted out, while
`L,-+1 new source symbols are shifted in. Go back to
`step 1. and repeat the procedure until the end of the
`string to be parsed is reached.
`In the preferred embodiment, during the first CLK2
`cycle during which old symbols are shifted out of SR]
`and new symbols are shifted in, the contents of the
`MRI. and pointer registers 124 and 126 are latched into
`the codeword bulfer 134. and simultaneously these reg-
`isters are cleared. At the same time, the contents of SR2
`are also cleared. Thus. no additional clock cycles are
`required for codeword generation. Codewords are gen-
`erated. and registers are cleared for the new codeword
`generation cycle. simultaneously with the shifting in of
`new symbols into the SR] register. Using a history
`buffer of 1024 symbols, and data registers SR1 and SR2
`oflength Lrnax. it will take 1024+Lmax— l+LC clock
`cycles to generate a codeword which represents LC
`symbols, as well as to prepare for generating the next
`codeword.
`Assuming that codewords. on the average. each rep-
`resent five symbols,
`the data compression circuit of
`FIG. 3 requires about 206 clock cycles per character to
`compress a set of data. As a result, a computer system
`with a 4-D megahertz clock rate could compress data
`using this circuit at a rate of about 200,000 characters or
`symbols per second. This is about a factor often slower
`than the rate of processing required for seamless pro-
`cessing of data being stored to a hard disk at a rate of
`one million bytes per second. which is equivalent to
`about two million symbols per second for it system with
`a two to one compression ratio.
`The data compression circuits described below with
`reference to FIGS. 5 through 11 overcome the failure
`of the circuit of FIG. 3 to encode data with sufficient
`speed. First, however, we describe a compressed data
`decoder circuit. which while simple. is more than suiti-
`ciently fast to keep up with even the fastest hard disk
`systems.
`
`COMPRESSED DATA DECODER
`
`Referring to FIG. 4, the compressed data decoder
`circuit 150 temporarily stores received codewords in a
`buffer 152, which are then loaded into codeword regis-
`ter 154 using clock CLK2, which is generated by finite
`state machine (FSM) 160 each time the decoder circuit
`150 is ready to process the next codeword. The current
`contents of the history buffer or prefix string are stored
`in buffer 162. which is typically a memory circuit used
`as circular buffer. The FSM 160 keeps track of the
`current position of the beginning of the prefix string
`within the message buffer 162, and generates a corre-
`sponding OFFSET value that is added by address gen-
`erator circuit l6Il to the pointer value PTR from the
`
`C',--—- <8. Li. 71>
`
`where T; is the last symbol of 5;. To ensure that all
`codewords have the same number of bits, a maximum
`value Lrnax is imposed on the reproduction length,
`typically 1 or 15.
`The core of the Letnpcl-Ziv encoder shown in FIG.
`3, is implemented with two shift registers, SR1 and SR2,
`a reproduction length generator (RLG) 122, maximum
`reproduction length and pointer latches 114 and 126,
`and a counter 128. In addition, a memory circuit or
`message buffer 130 is used to store the data that is being
`compressed, a finite state machine (FSM) 132 controls
`the sequence of steps performed by the entire circuit
`120. and a codeword buffer 134 is used to temporarily
`hold portions of the codewords generated by the circuit
`120. The FSM 132 generates the two clock signals CLK
`and CLK2 which control the shift registers SR1 and
`SR2, and also generates a signal CW'_RDY whenever a
`new codeword has been generated.
`Both shift registers SR1 and SR2 are of the same
`length. In particular. they both store a number of char-
`acters which is equal to the maximum allowed repro-
`duction length. For specificity. we will assume that
`both SR1 and SR2 store seven characters at any one
`time.
`Shift register SR1 stores the substring to be parsed.
`After SR] is initially loaded with data, the contents of
`SR! are shifted only after the generation of each code-
`word. As SR] is shifted, new data is loaded into the shift
`register from the message buffer 130. In particular, if a
`codeword with a reproduction length of L is generated,
`SR1 is shifted L+ 1 times using clock signal CLK2. The
`last character shifted out of SI-‘.1 is held in the codeword
`buffer 134 as the tail character of the codeword.
`At the beginning of each codeword generation cycle.
`the contents of registers 124 and 126 are cleared. i.e., set
`to zero. In addition. the counter l23_ is also reset to a
`predefined beginning value. such as zero. Then the
`history buffer contents, also called the proper prefix
`substring,
`is shifted through SR2 using clock signal
`CLK. Lmax—l clock cycles are required to preload
`SR2. where Linux is the maximum allowed reproduc-
`tion length. and is also the number of characters posi-
`tions in SR1 and SR2. Then, each time that SR2 shifts,
`the reproduction length generator 122 compares the
`contents of SR] and SR2. and generates a reproduction
`length value L. Comparator 136 compares the current
`reproduction length value I. with the current value
`MR1. stored in the maximum reproduction length regis-
`ter 124. If the current reproduction length value I. is
`greater than MRL,
`the comparator 136 generates a
`LOAD signal, which causes the current reproduction
`length value I. to be stored in register 12!. and also
`causes the current pointer value generated by counter
`128 to be stored in pointer register 125.
`After the proper prefix substring has completely
`shifted through SR1, the codeword can be generated
`from the contents of MRL register 124 and pointer
`register 126. followed by the appropriate tail symbol.
`
`

`
`5,150,430
`
`9
`codeword being decoded. The resulting address is the
`address of the first character of the substring repre-
`sented by the codeword. The OFFSET value is then
`incremented by the FSM 160 L times, where L is the
`reproduction value from the codeword being decoded.
`during each new CLK clock cycle, until the decoding
`of a new codeword begins.
`After the initial CLK2 signal used to load a code-
`word into register 154, the FSM generates L+l clock
`signals CLK. is the reproduction length value obtained
`from the codeword being decoded. At each CLIC sig-
`nal, a new decoded symbol is generated. The new de-
`coded symbol is read from the output of a multiplexer
`166. which outputs either (I) the tail character T from
`the codeword being decoded, or (2) the symbol read
`from the message buffer 162 at the address generated by
`address generator 164. These decoded symbols are then
`stored at the appropriate position in the history buffer
`162. as well as in an output message buffer 170. Depend-
`ing on how the history buffer 162 is designed. the read-
`ing of a decoded symbol from one address and the stor-
`age of that symbol back into the history buffer at a
`different address can be accomplished in either one or
`two clock cycles. By using a history buffer that stores,
`say, 1024 bytes. but limiting the length of prefix string to
`say. 1020 or 1023 bytes, one of ordinary skill in the art
`can de