This application is a continuation of application Ser. No. 08/216,743 filed Mar. 23, 1994 now abandoned.

DESCRIPTION

Priority of application Ser. No. 93-03589 filed on Mar. 29, 1993 in France is claimed under 35 U.S.C. §119.

1. Technical Field

The present invention relates to a code point update device in the arithmetic coding method.

2. Prior Art

GENERAL EXPLANATION OF ARITHMETIC CODING

As described in the article by Glen G. Langdon, Junior entitled "An Introduction to Arithmetic Coding" (IBM Journal of Research and Development, 28(2), pp. 135-149, March 1984), arithmetic coding is a well known lossless data compression method.

Arithmetic coding establishes a correspondence between a sequence of symbols and an interval of real numbers. The starting point is the interval [0,1]. When each symbol is coded, the current interval is replaced by a subinterval thereof. After coding a sequence of symbols, the original sequence can be exactly reconstructed, no matter what the given point in the current interval.

In this type of coding it is conventional practice to store the leftmost point of the interval as the result, the value obtained being called the code point.

This coding algorithm is recursire. For each symbol to be coded, the algorithm divides the current interval as a function of occurrence probabilities and the order given by the list of symbols which can be coded. The algorithm then replaces the current interval by the subinterval corresponding to the symbol to be coded. This procedure is repeated for the number of times which is necessary for coding a given sequence.

MATHEMATICAL EXPLANATION OF ARITHMETIC CODING

Up to now a geometrical explanation has been given of arithmetic coding. Reference is made to intervals and their subdivision. For implementation in a computer, it is necessary to explain the method using numerical quantities and this explanation will now be given.

In the following description the following notations are adopted:

______________________________________S = {s0, s1 . . . , sN-1 }         set of codable N symbolsSi       symbol to be coded in stage iP(Si)    symbol occurrence probabilityP(Si)    cumulative symbol probabilityAi       current interval width         range of admissible values of AiXi       shift for representing p(Si).Ai inXmax     maximum possible value of XiRi       augend of P(Si).AiCi       value of current code point (location         of leftmost point of interval)ζi       as Ci + Ri unshiftedBi       1+Xi output bitsTk       fixed width bit block.______________________________________

As previously explained, in each stage the coder state is given by a subinterval of [0,1]. The process for coding a symbol consists of replacing the current interval by a subinterval of itself.

For representing the current interval in a computer, two arithmetic quantities are updated, mainly C_i, the left-most point of the interval, and A_i the interval width. Thus, for each stage the actual interval is [C_i,C_i +A_i ].

Thus, the process given hereinbefore for coding a symbol is expressed by the following equations:

Ai+1 =p(Si)•Ai                        ( 1)

Ci+1 =Ci +P(Si)•Ai               ( 2)

in which p(S_i) is the modellized probability of the symbol S_i and P(S_i) is the sum of the probabilities of all the symbols preceding S_i in the list of the source alphabet ##EQU1##

These equations arithmetically represent the same sub-division and selection procedure described above. The first relates to the recurrence of the interval width and the second to the recurrence of the code point.

EXPLANATION OF THE FINITE PRECISION ALGORITHM FOR ARITHMETIC CODING

These equations assume an infinite precision arithmetic. For carrying out arithmetic coding on finite precision arithmetic operations, use is made of the approach of Frank Rubin in an article entitled "Arithmetic Stream Coding Using Fixed Precision Registers" (IEEE Transactions an Information Theory, IT-25(6), pp. 672-675, November 1979). The number of bits used for representing each of the S_i,R_i,A_i,C_i,B_i,p(S_i) and P(S_i) is fixed, said quantities being designated by #s, #R, #A, #C, #B, #p and #P. It should be noted that #P=#p. The range R=[α,2α], in which α>0 of the admissible interval widths is also chosen. It is ensured that Ai ε R always applies and for this purpose the following operations are performed in each stage of coding:

1. Obtain p(S_i) and P(S_i), exactly compute p(S_i)•A_i.

2. Determine the shift X_i such that p(S_i)•A_i •2^X _i εR.

3. Find the lower approximation on #A bits closest to p(S_i)•A_i •2^X _i, which gives A_i+1.

4. Exactly compute R_i =P(S_i)•A_i.

5. Exactly compute ζ_i =C_i +R_i.

6. Emit the carry-out and X_i most significant bits of ζ_i, which forms the variable width block B_i.

It should be noted that the definitions of the quantities the algorithm determine certain length relations:

• by cumulative probability definition #P=#p,

• by stages 1 and 2, X_max =#p,

• by stage 4, #R=#P+#A=#p+#A,

• by stage 6, #B=1+X_max =1+#p (the width of B_i s is variable, but a large number of bits is required for representing the longest possible value of B_i),

• by stage 7, #C=#R=#p+#A.

Thus, the parameters #p and #A determine the length of all the other quantities in these equations.

Thus an explanation has been given as to how the sequence of variable width blocks B_i is produced by the algorithm. Each block contains 1+X_i bits, consisting of the carry-out and X_i most significant bits of the sum ζ=C_i +R_i. These variable length blocks can be assembled in a sequence of fixed length blocks T_k. The sequence of blocks T_k constitutes the output coding flow. The assembly process is performed by 8 standard method, which does not concern us in the remainder of the present document.

It is possible to summarize stages 1 to 7 of the finite precision algorithm given above by the following equations:

(Ai+1,Xi,Ri)=ƒ(Si,Ai)    (3)

(Ci+1,Bi)=g(Ci,Ri,Xi)             (4)

The function ƒ effects stages 1 to 4 and the function g stages 5 to 7 given above.

EXPLANATION OF THE CODE POINT UPDATE PRIOR ART

The updating of the code point C_i is described by the above equation (4). FIG. 5 shows the structure of a unit able to perform this equation. C_i is the present code point. R_i is the product P(s_i)•A_i and X_i is the shift value. Both are obtained from the interval width update unit (not shown in this drawing). The adder 19 computes the sum c_i =R_i +C_i. The shifter 20 removes the carry bit from this sum and the X_i most significant bits, which become the quantity B_i and shifts the remaining bits by X_i and which then become the new value of C_i.

Taking account of the loop existing in the unit of FIG. 5 and which passes from the register 14 across the adder, the shifter and then again to the register 14, the clock period τ of a circuit following this arrangement must satisfy the inequation

τ>τcary (#C)+τshift (Xmax)      (5)

if (5) is not respected, there is a risk of recording an incorrect value for a few bits of C_i. In this case τ_cary (n) is the time necessary for the propagation of the carry over across a n bit adder and τ_shift (m) is the time necessary for the passage of the resulting sum through a structure, which can carry out a shift able to reach m bits (as designated above #C is equal to #A+#p and X_max is equal to #p).

There are two well known methods for speeding up the synchronous digital logic, like that of FIG. 5, namely retiming and table lookup. It will be demonstrated that they cannot be used here.

Due to the existence of this loop, it is not possible to use the retiming technique as described in an article by Charles E. Leiserson and James B. Saxe entitled "Retiming ynchronous Circuitry", Algorithmica, 6(1), pp. 5-35, 1991, to said circuit.

Moreover, the implementation of this unit by a table lookup requires (1+X_max +#C)•2^#R+#C+#X memory bits. For usable values of these parameters, e.g. #C=#R=16, X_max =12 and #X=4, it would be necessary to have 232 GBytes, which is an excessive value for practical applications.

However, the value #p should be as high as possible for the following reason. Two conditions are necessary to enable our compression system to compress its input stream. Firstly the probability distribution of the symbols must be non-uniform. In addition, said non-uniformity must be reflected in the probability values used for the computation. The probabilities p(si) and P(si) are real numbers taken from the interval of [0,1], which can require an infinite precision in order to accurately represent everything. The computed quantities are binary finite representations of these values, each written with #p bits.

Obviously there is no control on the content of the input stream. However, given a favourable distribution, in order to exploit it it must be possible to represent the numbers just below 1, as well as those just above 0. It is therefore desirable to have #p as high as possible, so as to be able to represent the widest possible probability range. However, according to equation (5), the higher #p, the slower the circuit. It is therefore necessary to find a compromise between a fast circuit with mediocre compression and an effective compressor at reduced speed.

The invention offers a solution to this problem. It is a device for the construction of code point update units able to operate at high speed and independent of the size of the operands. The parameter #p can be taken as high as desired, without any negative consequence for the compression speed.

DESCRIPTION OF THE INVENTION

The present invention relates to a code point update device of an arithmetic coding system, characterized in that for maintaining a given number of precision bits, whilst maintaining a fixed number of logic levels between two registers, said device comprises a grid of n×n cells taken from among four basic cells:

n(n-1)/2 cells ad, above the principal diagonal of the grid, incorporating a multiplexer (33),

n-1 cells d, on said principal diagonal incorporating an adder (31) and a multiplexer (30) and at least one register (32),

a cell fd at the bottom end of the diagonal incorporating an adder (31), a multiplexer (30) and at least one register (32),

n(n-1)/2 output retiming cells or incorporating a register (35),

the lower limit of the clock period being such that

τ>max{τmux +τadd,Xmax •τprop }

τ_mux, τ_add being the setting up times of the multiplexers and adders, τ_prop the time lag necessary to enable a signal to propagate through the width of a cell and X_max the maximum shift.

In an advantageous embodiment, the adders are generalized adders each incorporating j+k+l input bits and which generate l+k output bits, in which j, k and l are non-negative integers and whose equations giving the outputs as a function of the inputs are completely arbitrary.

In a variant, the code point update device of an arithmetic coding system according to the invention comprises a grid of (2n-1)n cells taken from among six basic cells:

n(n-1)/2 input retiming cells ir incorporating two registers,

n-1 diagonal cells d incorporating an adder and a multiplexer,

a cell fd at the lower end of the diagonal incorporating an adder and a multiplexer,

n-1 cells fad located above the diagonal in the left-hand column of the grid and incorporating a multiplexer,

(n-2)(n-1)/2 cells ad located above the diagonal and incorporating a multiplexer,

n(n-1) output retiming cells or incorporating a register,

the lower limit of the period being such that

τ>max {τmux +τadd, τ'prop }

in which τ_mux, τ_add are the setting times of the multiplexers and adders, in which τ'_prop is the propagation time of a signal along any random wire.

In an advantageous embodiment the adders are generalized adders each incorporating j+k+l input bits and which generate l+k output bits, in which j, k and l are non-negative integers.

The invention makes it possible to improve the efficiency of coding non-adaptive arrangements for high speed arithmetic coding by proposing and advantageous design of the code point update circuit.

The present invention permits a higher data compression speed than the known arithmetic coding technique described in the article by Glen G. Langdon referred to hereinbefore and the article by Ian H. Witten, Radford M. Neal and John G. Cleary entitled "Arithmetic Coding for Data Compression" (Communications of the ACM, 30(6), pp. 520-540, June 1987). It can constitute the basis of a product used for data compression. This product could be used for the compression of data to be transmitted on a communication channel or for recording in a file system.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows the architecture of a prior art arithmetic code including a prior art code point update unit.

FIG. 2 shows a prior art adder.

FIG. 3 shows a prior art device for making a selection of a particular width for the circuit.

FIG. 4 shows a prior art circuit wherein for a chosen width of the circuit (4 bits) an output is introduced into the adder.

FIG. 5 shows a prior art version of a code point update unit.

FIG. 6a is a key of devices in FIGS. 6-17.

FIG. 6 shows an embodiment of the invention showing logic elements placed along the diagonal of a square.

FIG. 7 shows how retiming can be carried out.

FIG. 8 shows a retimed circuit by way of example.

FIG. 9 shows the circuit of FIG. 8 having a width of n bits.

FIG. 10 corresponds to the inventive version of FIG. 4.

FIG. 11 shows a code point update circuit for retiming.

FIGS. 12 and 13 show variants of a code point update circuit after retiming.

FIG. 14 shows a code point update circuit after retiming.

FIG. 15 shows how to construct a version of the arbitrary width circuit by forming a grid of size nx(2n-1).

FIG. 16 shows a general structure of a circuit using the logic units of FIG. 15.

FIG. 17 shows four logic units/adders used in a circuit.

FIG. 1 shows the architecture of a prior art arithmetic coder for coding the symbol s_i. It comprises an interval width update unit 10, a code point update unit 11 and optionally a buffer circuit 12.

The interval width update unit 10 supplies two signals X_i and R_i to the code point update unit 11. An output of the unit 10 connected to a register 13 makes it possible to supply to one input of said module the signal A_i. An output of the unit 11 connected to a register 14 makes it possible to supply C_i to another input of the module 11. The buffer circuit 12 receives on its two inputs the signals from the code point update unit 11, namely Xi_i and B_i, in order to supply a signal T_k.

The interval width update unit 10 makes it possible to update A_i in the register 13. For each stage of the algorithm it takes a new symbol s_i and the current value of the register 13. It generates the augend R_i, the shift X_i and the new value A_i for the register. It therefore keeps updated the current interval width A_i. The code point update unit 11 also updates C_i in the register 14. For each stage it takes into account the shift X_i, the augend R_i and the current value of the register 14 by producing a new value for R_i and a variable length block B_i of width 1+X_i. The value X_i is not modified.

The variable block buffer circuit 12 assembles the sequence of variable length blocks in the fixed length blocks T_k, which constitute the output of the coder.

The prior art interval width update unit 10 shown in FIG. 1 comprises a probability computation module 15, whose two outputs supplying the signals p(S_i) and P(S_i) are respectively connected to a first multiplier 16 followed by a standardization module 18 for supplying the signal X_i and to a second multiplier 17 supplying signal R_i. An output of the standardization module 18 is connected to an input of each multiplier 16 and 17 across the register 13.

The code point update unit 11 comprises an adder 19, which receives on a first input the signal R_i, connected to a shift circuit 20 receiving on another input the signal X_i and whereof a first output is relooped on a second input of the adder 19 across a register 14, a second output of the shift circuit 20 supplying a signal B_i.

The buffer circuit 12 comprises an adder receiving on its two inputs the output signals from the interval width update unit 10 and the code point update unit 11, namely X_i and B_i, for supplying the signal T_k.

The algorithm for the arithmetic coding must maintain a quantity designated by the code point C_i. Whenever a symbol is coded, the current code point must be incremented and optionally shifted. The shifted bits are emitted in the output stream and become the new value of the code point. FIG. 1 shows the elements of a circuit 11 maintaining the code point. This circuit performs the following algorithm:

1. Exact computation of ζ_i C_i +P(s_i)•A_i

2. Emission of the carry and the X_i most significant bits of ζ_i

3. Shifting by X_i positions the remaining bits of ζ_i, the result being C_i +1.

In this case C_i is the current code point, A_i the current interval width, s_i the following symbol to be coded, P(s_i) its cumulative probability and X_i the shift. R_i is written for the product P(s_i)•A_i. The quantities X_i and R_i are obtained from the interval width update.

A description will now be given of the invention and its operating principles. The invention relates to the code point update unit 11 of FIG. 1. FIG. 5 separately shows this part of FIG. 1.

The essence of the invention is a device permitting the application of transformation by retiming to the unit of FIG. 5. As explained hereinbefore, in the prior art it is not possible to apply said transformation to the unit. However, it is advantageously possible to apply this transformation twice, firefly for eliminating long combinational logic chains and secondly for eliminating long wires. The first stage is referred to as "logic retiming" and the second as "wire retiming".

The structure of digital circuits will now be given. The elements of these circuits are shown in FIG. 2 and are:

a full adder 21, whose outputs comply with the relations s=a⊕b⊕C_in and C_out =ab+bC_in +aC_in with a and b being input values, C_in the input carry and C_out the output carry,

a multiplexer or mux 22, whose output m complies with the relation m=it+-ie, with i, t, and e input values.

In this case the symbol ⊕ represents the "exclusive-OR" logic aperation, the symbol + represents the "inclusive-OR" logic operation, the symbol - represents the logic negation operation and the juxtaposition of variables represents the logic "AND" operation. In addition, a square symbol indicates a register, which delays its input by a clock stroke.

In the drawings an unrepresented input will be considered as transmitting a zero signal.

The first stage of the invention is to transform it to obtain the version of FIG. 3. The register 14 has been shifted in the loop, but this has no effect on the low limit of the clock cycle τ given in equation (5). In the cycle where R_i and X_i are introduced on the inputs, the register contains the quantity ζ_i =C_i +R_i.

On more particularly considering FIG. 3, a selection made of a particular width for the circuit (in this case 4 bits) and it is introduced into the adder structure. The carry over construction is used for the adder. The shifter is supplied with a unitary representation, i.e. a set of X_max wires (. . . x₂ x₁ x₀), which are all at zero, with the exception of X_u, in which u is the shift quantity. In the considered example X_max =4. The result is given in FIG. 4. Therefore the invention relates to a novel circuit integrating the desired addition, shift and recording functions and which can be synchronized at a frequency independent of #C and X_max.

It is now possible to explain the invention. This is dependent on a detailed device of the mixed adder--shifter unit given here. The invention consists of placing the logic elements along the diagonal of a square and to construct the device according to the invention for phase shifting thereon. This is shown in FIG. 6. A unitary representation of X_i is considered, in such a way that the shifter has a simple structure. However, this does not limit the generality of our invention, as is it is obvious to pass from a binary to a unitary notation. This is possible because the carry propagation direction and the direction of any shift are the same.

Two observations can be made: firsfly each loop in the drawing only has two logic stages. It is true that there are longer combinatory paths in the drawing such as the carry propagation chain. However, the second observation is that these long chains can be subdivided by retiming. This is what is referred to hereinbefore as logic retiming.

FIG. 7 explains how such a retiming can be carried out. The horizontal dotted lines representing the retiming limits indicate the locations where the registers must be introduced. This retiming is possible due to the uniform direction of the information flow through these lines.

FIG. 8 shows the retimed circuit (after displacing some of these registers for facilitating the drawing). In this case the elements 30 or 33 are multiplexers, 31 are full adders and 32, 34, 35, 36, 37, 38 and 39 are registers.

The essential characteristic resulting from this is that there are no longer two logic stages between two random registers. The lower limit of the clock period is

τ>max{τmux +τadd, Xmax •τprop }(6)

In this expression τ_mux and τ_add are setting times for the multiplexer and the full adder of FIG. 2, whilst τ_prop is the propagation time for a signal through a cell, said quantity being very low.

The structure of FIG. 8 can be extended to the general case of a circuit with a width of n bits, as shown in FIG. 9. This gives the example of the case n=5. There are several unconnected wires which can be ignored and which emanate from cells located on the left-hand edge of the grid.

Such a circuit of random size can be assembled by forming a grid n×n with four basic cells. Thus, a grid contains a total of n² cells distributed in the following way:

______________________________________Number of Cells         Type of Cell______________________________________n(n-1)/2      above the diagonal or (ad)n-1           within the diagonal, or (d)1             bottom end of the diagonal, or (fd)n(n-1)/2      output retiming, or (or).______________________________________

Up to now consideration has been given to X_max =n, which is the width of the circuit. This is the most general possible supposition, because a shift of more than n bits means that no bit of the register ζ_i will be restored to circulation and in this case the computation proceeds as if all the bits were 0. This effect can be obtained in the circuits by placing each bit of the unitary representation of X_i at 0. However, the case X_max <n is possible and allows a slight circuit simplification. The bits x_u are then eliminated from the unitary representation with u>X_max, together with the registers and multiplexers which they supply. FIG. 10 illustrates the code point update circuit for the case n=4 and X_max =2. This circuit can be retimed like that of FIG. 8 in order to obtain a rapid version.

ELIMINATION OF LONG TRANSMISSION PATHS

The length of the logic chains in all the tiles of the circuit shown in FIG. 9 is limited to two. However, the large tiles of this circuit with a high value for X_max contain long transmission paths. In particular the wire which passes out of the least significant adder must extend X_max columns to the left-hand end of the circuit. This gives the term X_max •τ_prop in the expression (6) for the lower limit of τ.

However, by adding supplementary registers and modifying the circuit geometry, it is possible to eliminate this disadvantage. This is what is referred to hereinbefore as retiming of the wires.

FIGS. 11, 12 and 13 illustrate the modification process and FIG. 14 is an example of the final version of the circuit.

FIG. 11 shows a code point update circuit for retiming. This circuit is topologically identical to that of FIG. 8. The inputs and outputs have merely been repositioned. The vertical dotted lines 40 indicate where it is necessary to retime the circuit. Such a modification makes it possible to subdivide the long horizontal paths into segments of limited length.

FIG. 12 shows a code point update circuit after retiming. FIG. 12 is identical to FIG. 11 with a register 41 inserted at the locations where a wire crosses a retiming limit. The long horizontal paths have been subdivided, but the wires top left and bottom right are of linear length n. By adding registers for cutting certain paths, other paths, which were previously short, were lengthened. A solution to this problem will now be described.

FIG. 13 is a code point update circuit after retiming. This variant of the previous drawing has input and output retiming registers in their respective columns. The wires supplying the most significant bit at the input and the most significant bit at the output are of linear length n.

FIG. 14 is a code point update circuit with constant wire lengths. This version is obtained by shifting each column of the circuit of FIG. 13 by a given quantity with respect to the adjacent column. All the wires now have a limited length. In addition, this limit is independent of the width of the bits n of the circuit.

As previously it is possible to construct versions of this arbitrary width circuit by forming a grid of size n×(2n-1) with the basic cells illustrated in FIG. 15. The latter contains an example for the case n=5. The slope of the diagonal in this case is more pronounced, because it is due to the shifting of each column of cells with respect to its neighbours. Such a grid consequently contains 2n² -n cells distributed in the following way:

______________________________________Number of Cells          Type of Cells______________________________________n(n-1)/2       input retiming or (ir)n-1            diagonal or (d)1              bottom end of diagonal or (fd)n-1            ends above the diagonal or (fad)(n-2)(n-1)/2   above diagonal or (ad)n(n-1)         output retiming or (or).______________________________________

The lower limit of the clock period of the circuit of this device is

τ>max{τmux +τadd, τ'prop }  (7)

in which τ'_prop is the propagation time of a signal along the longest wire among the cells ir, d, fd, fad, ad or or of FIG. 15. There is no longer any dependence of this limit on the bit width of the circuit.

The full adders of FIG. 2 can be generalized to the logic units of FIG. 15. Each unit to the left, referred to as a generalized adder, receives j+k+l input bits and produces k+l output bits. In this case j, k and l can be any random non-negative integer. In mathematical notation, if f represents the logic function performed by this unit, it is possible to write:

(d0, . . . ,dl-1,e0, . . . ,ek-1)=f(a0, . . . ,aj-1,b0, . . . ,bk-1,c0, . . . ,cl-1)

The outputs d₀, . . . ,d_l-1 are referred to as the generalized carries of this unit. FIG. 16 shows the general structure of a circuit using these units.

In the preceding paragraphs consideration has been given to a circuit having the structure of FIG. 16, where f has been the full adder function with j=k=l=1. However, nothing in our technique requires f to have any particular function (or the same function for each circuit column).

With a circuit having the structure of FIG. 17, in which the generalized carry propagation direction is the same as the shift direction of the shifter, it is possible to use the preceding devices for obtaining high speed operational circuits. τ_agen is the setting time of a generalized adder, τ_mgen the setting time of a generalized multiplexer and S_max the maximum shift. The lower limit of the clock period is obtained

τ>:max{τmgen +τagen,Smax •τprop }

by the device of FIG. 9 and

τ>:max{τmgen +τagen,τ'prop }

by the device of FIG. 15.