Apparatus for Determining Booth Recoder Input Control Signals

The present invention relates to math processors for doing floating point divide and square root operations.

Typically, divide and square root operations are calculated in math coprocessors with a tedious method of subtracting and shifting, as with long division by hand. The disadvantage of this method is that it takes a long time to complete all of the necessary subtract and shift operations.

One method used to improve the speed of division and square root operations takes advantage of quadratic convergence. If one is trying to determine y/x, first, a determination of 1/x is made, but by methods other than the tedious long division. Then, a multiplication operation can be used (y multiplied by 1/x), which speeds up the calculation. One way of determining the value of 1/x would be to store all the possible values in a memory. However, for the number of bits of accuracy required, this would have to be a huge memory. Accordingly, the value of 1/x is calculated each time for different values of x. One method is to begin with a low accuracy initial approximation and improve it using a Newton-Raphson iteration. Such a method is described in S. Waser and M. J. Flynn, "Introduction to Arithmetic for Digital Systems Designers", CBS College Publishing, New York, 1982. The iteration uses the equation F.sub.K+1 =F.sub.K (2-x F.sub.K), where F.sub.K is an approximation to 1/x. By using an initial approximation of F.sub.K and doing a number of iterations, its accuracy can be improved to the point necessary through quadratic convergence. The x operand is typically normalized to within the range 1.ltoreq.x<2, and an initial approximation of 1/x can be chosen from values stored in a ROM. A technique similar to the one described above for 1/x is also used to compute square roots. This technique first computes 1/.sqroot.x and then multiplies it by x to produce .sqroot.x. If desired, the initial approximation to 1/.sqroot.x can be improved through a Newton-Raphson iteration similar to that above.

The present invention provides an improved apparatus and method for determining the value of 1/x and of 1/.sqroot.x for a given operand x, and for determining the value of y/x and .sqroot.x. This is accomplished using a polynomial equation. The coefficients of the polynomial equation are stored in a ROM. Different coefficients are used for different values of x. The most significant bits of the x operand are used as inputs to the ROM to select the proper coefficients.

In the preferred embodiment, the polynomial approximation is second order, and the coefficients have enough precision to give a final result for single precision accuracy. For double precision, the results of the polynomial are then used as the first approximation for the Newton-Raphson iteration. This provides a much higher degree of accuracy for the first approximation without the need for increasing the size of the ROM for the initial Newton-Raphson approximation. Accordingly, only one iteration of Newton-Raphson is typically required to produce the result with double precision accuracy.

The polynomial equation used is preferably in the form Ax.sup.2 +Bx+C. The present invention also uses a split multiplier array to speed up the calculation of this polynomial. After first calculating x.sup.2, the multiplication of A by x.sup.2 and B by x are done simultaneously in two halves of the multiplier array. This is possible because the coefficients stored are not needed to have full accuracy. Thus, they require only a portion of the multiplier array to multiply. The size of the coefficients correspond to what is needed to produce the proper precision result for each of the elements of Ax.sup.2 +Bx+C. In the preferred embodiment, A is 15 bits, B is 24 bits and C is 32 bits. Separate values for the coefficients A, B, and C must be stored for the 1/x approximation and for the 1/.sqroot.x approximation.

In a further aspect of the present invention, a special multiplier is provided which can accept an operand in carry/save format. This eliminates the need to convert partial products into a normal binary format by completely propagating the carry. Thus, the speed of a sequence of multiplications can be improved. This is achieved by providing Booth recoder logic which can accept operands in a normal binary or a carry/save format.

The present invention also provides an improved rounding technique which provides IEEE exact rounding by an operation that includes only one multiplication.

The invention thus speeds up the processor operation in several ways. First, the initial approximation is made very accurate by using the polynomial approximation. Second, the split multiplier array lets more than one operation be performed at once. Third, the Booth recoder logic is designed to accept intermediate operands in carry save format so they don't need to be put into pure binary format before the next multiply can be done. Lastly, the special rounding technique of this invention has very few computational steps.

The technique of this invention is used for divide, square root, inverse, and inverse square root operations, except the special rounding technique, which is not used for the inverse and inverse square root operations.

FIG. 1 is a block diagram of the overall chip electronics incorporating the present invention;

FIG. 2 is a graph illustrating the choosing of coefficients for the polynomial Ax.sup.2 +Bx+C;

FIGS. 3-3E are block diagrams illustrating the logic used for the polynomial calculation according to the present invention;

FIG. 4 is a timing chart illustrating the operations performed during different cycles for the polynomial calculation for single precision numbers;

FIG. 5 is a timing chart illustrating the polynomial calculation for double precision numbers

FIGS. 6a and 6b are a block diagram and a mathematical representation of the use of two sub-array multipliers to compute the partial products of a larger multiplication;

FIG. 7 illustrate a prior art N.times.M multiplier composed of M pairs muxes and adders;

FIG. 8a shows how an eight bit unsigned number is padded with zeros and grouped for radix 4 Booth recoding;

FIG. 8b shows how three bit groups are encoded through standard radix 4 Booth recoding;

FIG. 8c shows a block diagram representing a radix 4 Booth Recoder;

FIG. 8d shows five Booth Recoders of FIG. 8c arranged so as to compute Booth Recoding information for the eight bit unsigned number B.

FIG. 9 shows how a multiplier can be constructed from pairs of five input muxes and adders, controlled by radix 4 Booth recoding information;

FIG. 10 shows a block diagram for a carry save recoder;

FIG. 11 shows a block diagram in which carry save recoders and booth recoders are arranged so as to compute Booth Recoding information for the eight carry-save pair unsigned number B;

FIGS. 12A and 12B show block diagrams in which four-bit and two-bit adders, respectively, are used to apply carry save format bits to Booth recoders;

FIG. 13 shows a block diagram in which two-bit carry save recoders replace the two-bit adders of FIG. 12B;

FIG. 14 shows a block diagram of three-bit carry save recoders being used to generate inputs to radix 8 Booth recoders;

FIG. 15A shows a block diagram a one-bit carry save recoder being used in conjunction with adders to apply a carry save format number to Booth recoders; and

FIG. 15B shows one possible circuit schematic for a one-bit carry save recoder.

FIG. 1 is an overall block diagram of a semiconductor chip incorporating the present invention. A pair of operands are provided on input buses 12 and 14. These are provided to a register file 16. The register file is in turn connected to a separate ALU 18 and multiply unit 20. The outputs of these are provided through multiplexing circuitry 22 to an output 24. This structure allows the ALU and multiply unit to be used in parallel. The invention to which the present application is directed is contained within multiply unit 20.

FIG. 2 is a graph showing the curve for 1/x between x=1.0 and 2.0. As discussed in the background, for doing division operations with a given operand x, it is advantageous to calculate 1/x first, and then multiply by the numerator operand y. The determination of 1/x is started by an approximation for a given value of x between 1.0 and 2.0. All of the operations (inverse, division, inverse square root, and square root), except as noted, take as initial operands numbers according to IEEE standard 754, which are always normalized to be between 1.0 and 2.0.

In the present invention, the interval between 1.0 and 2.0 is divided into 256 smaller intervals, and the curve for 1/x in each interval i in FIG. 2 can be approximated by A.sub.i x.sup.2 +B.sub.i x+C.sub.i. Coefficients for each of these 256 intervals for A.sub.i, B.sub.i and C.sub.i are separately predetermined.

For each interval, the coefficients for the polynomial Ax.sup.2 +Bx+C are determined by choosing the mid-point of the interval and two other points which are .sqroot.3/2 times one-half the interval size in each direction from the mid-point. These three points are then used to produce three equations, which are solved simultaneously to produce A.sub.i, B.sub.i and C.sub.i.

The values of the coefficients for A.sub.1 through A.sub.256, B.sub.1 through B.sub.256 and C.sub.1 through C.sub.256 are to be stored in a ROM 30 as shown in FIG. 3A, and as explained further below. ROM 30 is addressed by the most significant 8 bits of the x operand on lines 32. One particular embodiment takes advantage of the fact that for each interval, the most significant 8 bits of x are constant. It does this by defining x=x.sub.0 +.DELTA.x, where x.sub.0 is the most significant 8 bits of x (which are constant over each interval) and where .DELTA.x is 8 zeros followed by the remaining bits of x. When this is expanded in the equation Ax.sup.2 +Bx+C, and the terms regrouped, the resulting equation is (A).DELTA.x.sup.2 +(2Ax.sub.0 +B).DELTA.x+(Ax.sub.0.sup.2 +Bs.sub.0 +C). The values in parentheses are constant over each interval, and can be stored in the ROM as modified A, B and C. When the initial approximation is done, the computation will be A.DELTA.x.sup.2 +B.DELTA.x+C, where A, B and C are these modified values. The advantage of this has to do with significant digits. .DELTA.x has 8 leading zeros, and .DELTA.x.sup.2 has 16 leading zeros, which get transferred to and lead products made from them. So, the most significant bit of A can effect only the 17th and lesser significant bits of the approximation, and the most significant bit of B can effect only the 9th and lesser significant bits of the approximation. Because this initial approximation needs to be accurate to 28 bits at most, the coefficients do not all need to be stored to full precision. In the preferred embodiment there are 15 bits of accuracy for A, 24 for B, and 32 for C. This provides 31 bits overall accuracy for Ax.sup.2 +Bx+C, because there are at least 31 bits of accuracy in each of the three elements being added. The coefficients A, B and C corresponding to this value x are then provided to registers 31, 33 and 34, respectively.

FIG. 3A shows a multiplier which is split into two halves 36 and 38. For the polynomial calculation, the halves are connected to act as a single multiplier for the x.sup.2 term, and are used as two separate multipliers to the Ax.sup.2 and Bx multiplication simultaneously as discussed below.

Referring to FIG. 4, the lookup of modified A, B and C is done during cycle 1. At the same time, the multiplication of .DELTA.x by itself to produce the square term is done with the two array halves acting as one. The upper order bits of operand .DELTA.x are provided through operand multiplexer 40 and multiplexer 52 to the "A" input of carry/save adder (CSA) multiplier array half 36, while the remaining lower order bits are provided to array half 38. The ax operand is also provided through multiplexer 41 to Booth recoder 50. Booth recoder 50 then supplies the Booth recoded version of the upper order bits of .DELTA.x to the "B" operand input of multiplier array 36, and the lower order bits to the "B" input of array 38. The product of the lower order multiplication in sum and carry form (CS) is provided to multiplexers 42 and 43, respectively. As can be seen, the control signal for this cycle selects these inputs, labeled "0". These are then shifted over, just as they would be in the multiplier array, and are applied to inputs 44, 45 of array half 36. Array 36 completes the calculation. The .DELTA.x.sup.2 result is provided from the product output of array half 36 through multiplexer 46 to registers 47 for the "first" pipeline (the "second " pipeline consists of carry-propagate (CP) adder 48 and second pipeline register 49. The .DELTA.x.sup.2 operand at this stage is stored in carry-save format.

In the second cycle shown in FIG. 4, and illustrated in FIG. 3B with the control signals for cycle 2, the .DELTA.x.sup.2 term from register 47 and the operand A are provided to multiplier array 36 to calculate the A.DELTA.x.sup.2 term. At the same time, the B operand and the .DELTA.x operand are provided to multiplier array 38, so that the B.DELTA.x multiplication can be done at the same time. It is possible to do this at the same time using two parts of a split multiplier array since a full 64 bits of the coefficients is not necessary. At the same that this is occurring in cycle 2, in the second pipeline consisting of carry propagate (CP) adder 48 and register 49, a CP rounding operation is done on .DELTA.x.sup.2 to put it into pure binary format. This purely automatic CP rounding is not needed until the final result, however, and the binary form for .DELTA.x.sup.2 and other intermediate results are ignored.

The .DELTA.x.sup.2 term from register 47 is provided in CS format to multiplexers 41 and 51, which connects to the CS inputs of Booth recoder 50, discussed in detail below. Two multiplexers are used because of the S,C format. The Booth coded output is then provided to the "B" operand input of array 36. This output is the most significant 36 bits of .DELTA.x.sup.2, with the rest of the Booth recoder providing bits to array 38. The A coefficient is provided through multiplexer 52 to the "A" operand input of array 36.

The B coefficient is provided from register 33 through multiplexer 41, through Booth recoder 50 to the "B" operand input of array 38. The .DELTA.x term from the MA input 32 is applied through multiplexer 40 to the "A" operand input of array 38.

Also during cycle 2, the B.DELTA.x product of array 38 is combined with the C coefficient from register 34 in carry save adder (CSA) 53 to provide the B.DELTA.x+C term. This output is provided to CSA 54, where it is added to the A.DELTA.x.sup.2 product of array 36, with the result being stored in register 47.

As can be seen, the x.sup.2 term is used in carry/save format, without being converted into a normal format through CP adder 48, since this would require additional time. This ability is provided by special Booth recoding logic (which supplies all the necessary information for one of the operands to the multiplier arrays 36 and 38) which accepts either a normal binary format or a CS format as will be described in detail below.

In the third cycle of a single precision calculation, shown in FIGS. 3C and 4, the operand y is multiplied times the polynomial Ax.sup.2 +Bx+C. This is done with the polynomial still in CS format, again saving translation time. For a single precision operation, no Newton-Raphson iteration is required because the polynomial provides the necessary level of precision to go straight to the multiplication of the 1/x term by the y term. The control signals used for the remaining cycles are shown in FIGS. 3D and 3E.

For double precision divide operations, as illustrated in the timing chart of FIG. 5, the operation proceeds similarly up to the third cycle. In the third cycle, a Newton-Raphson iteration is begun by multiplying a negative x value by the polynomial expression Ax.sup.2 +Bx+C (in CS form). The iteration is completed in cycle 4 to provide the precision necessary. Then, in cycle 5, the operand y is multiplied by the Newton-Raphson approximation of 1/x. Here, for double precision, only one iteration of Newton-Raphson is required due to the use of the polynomial equation of the present invention.

The Booth recoding referenced above is discussed in A. D. Booth, "Assigned Binary Multiplication Technique", Quarterly Journal Mechanics and Applied Mathematics, Vol. 4, Part 2, 1951. The logic of the present invention allows the Booth recoding logic to receive the operands in the carry/save format. Thus, for the x.sup.2 and other terms, it does not need to be put in a normal format, speeding up the operation, as explained further below.

The determination of a square root is done by also using a polynomial Ax.sup.2 +Bx+C, but to approximate the curve 1/.sqroot.x between X=0.5 and 2.0. 512 intervals are used, with a separate ROM, or a separate area of the same ROM 30, storing the coefficients for this inverse-square root approximation. The actual square root is determined from its inverse by multiplying the inverse by X, similar to the division procedure above.

The manner in which the Ax.sup.2 and the Bx term are shown to be computed above takes advantage of another aspect of the present invention. These terms are used in computing the single precision value for 1/x. This single precision result is computed to have an accuracy of only 28 bits; therefore, the values used in computing it need have no more than 28 bits of accuracy. Computing the Ax.sup.2 term and the Bx term each require one multiply operation, but again, because a result with only 28 bits of accuracy is needed, operands with only 28 bits of accuracy may be used. Thus, the computation of neither of these two terms requires the full 60.times.64 array, and it may be split so that the terms may be computed simultaneously. The particular embodiment shown in FIG. 3 illustrates this, with the multiplier array divided into two sub-arrays 36 and 38, which each have a 60 bit MA operand and which have a 36 bit and 28 bit MB operand, respectively. During the second step of the second order polynomial approximation, the product Ax.sup.2 is computed in the 60.times.36 sub-array at the same time that Bx is being computed in the 60.times.28 sub-array, saving one whole cycle over doing each calculation using the full array.

In general, multistep calculations often involve two intermediate calculations each of which has operands that are not dependent on the results of the other calculation. If so, the intermediate calculations should be examined to see how many bits of accuracy their results and operands need. If the sum of the number of bits needed for an operand of one intermediate calculation plus the number of bits needed for an operand of the other intermediate calculation does not exceed the maximum number of bits available in the array, the array may be split and both intermediate calculations performed at the same time. The same analysis indicates that for a more complex calculation the array could be split to allow three or more calculations to be performed at once.

The manner of splitting the array is fairly straightforward; each sub-array 36 and 38 is designed as a normal stand-alone multiplier array. For split-array operations each sub-array is therefore provided its operands, and delivers its result, in a standard manner. In order to combine the sub-arrays into a larger unified array, logic is provided to do the following: One of the two operands is provided in whole to each sub-array. The other operand is split into fragments, one for each sub-array. In this particular embodiment the fragmentation and providing of operands is done by the muxes 41, 51 and Booth recoder 50. The values that these sub-arrays 36 and 38 compute are used as partial products; they are shifted according to the location that the least significant digit of their fragment operand had within the pre-split whole operand. FIGS. 6a and 6b illustrate this. A is a 60 bit operand to be multiplied by B, a 64 bit operand, using a 60.times.28 multiplier array and a 60.times.36 multiplier array. A is multiplied by the least significant 28 bits of B in the 60.times.28 array to produce an 88 bit result C. At the same time, A is multiplied by the most significant 36 bits of B in the 60.times.36 array to produce a 96 bit result D. D is then logically left-shifted (the actual hardware does this with a right shift of C to make the layout easier) 28 bits to produce 124 bit F, the leftmost 96 bits of which are the 96 bits of D, and the rightmost 28 bits of which are zeros (E.sub.28 . . . E.sub.1). C and F are then added together to produce the result G. FIG. 6b illustrates the fact that C and D are partial products, and also makes apparent that D could be provided to the inputs of the adder in a left shifted manner, or C could be provided to the inputs of the adder in a right shifted manner, without needing an actual separate shifter.

In this particular embodiment, multiplexers 42 and 43 do the shifting of the product of array 38 by applying it to the inputs 44, 45 of array 36.

The process of multiplication can be thought of as having two main steps, the first being the generation of partial products, and the second being the reduction of these partial products to the final product. To add the first two partial products, one adder is needed, and generally, for each extra partial product an extra adder is needed. The simplest form of multiplication operates bit by bit of the multiplicator, and produces one partial product for each bit in the multiplicator.

An n by m bit multiplier along these lines generates m partial products to be added. A simple representation is as in FIG. 7. To build this multiplier a stack of m adders sums up the product. Each partial product is either A or 0, determined by the individual bits of B. This multiplier is simple but costly in size, and m additions are necessary to compute the product. It is possible to reduce the number of partial products to be added, however, by using Booth recoding. Booth recoding lets the multiplication process skip through contiguous strings of 1s and contiguous strings of 0s in the multiplicator; for an n bit multiplicator no more than n/2 partial products are created. Modified Booth recoding, more commonly used, processes overlapping groups of bits at a time, the number b of new bits per group being described by a radix 2.sup.b. Radix 2.sup.b Booth recoding generates (n/b)+1 partial products, and lends itself to parallel hardware implementation. The modified recoding method encodes these overlapping groups based on whether they start, continue, or end a string of 1s. To ensure proper termination of a string of 1s, a number to be recoded must be padded with one least significant 0 and at least one most significant 0. For radix 4 (2.sup.2) Booth recoding, each group has three bits, two bits of primary concern, plus an overlap bit from the right hand (less significant) side. FIG. 8a shows how an eight bit unsigned number, padded with one zero to the right and two zeros on the left, is grouped, and FIG. 8b shows how a three bit group is encoded through standard radix 4 Booth recoding. FIG. 8c shows a block diagram of Booth recoding logic which accepts three inputs, r.sub.0, r.sub.1, and r.sub.2. These correspond to bits Y.sub.i-1, Y.sub.i, and Y.sub.i+1, of FIG. 8b, respectively. So, for an eight bit number, five partial products are produced, and four adders are needed. FIG. 8d shows how five Booth recoders of FIG. 8c can be used to compute the Booth recoding information for the eight bit unsigned number of FIG. 8a. FIG. 9 illustrates the summing of partial products for A times B, where B has m bits. Note that only m/2 adders are needed.

A critical feature of standard Booth recoding is that it deals solely with numbers in binary format; this is a drawback in operations which require a series of multiplies, the result of one multiply being used as an operand in the next multiply operation. When the partial products from one multiply operation are first added, the result is initially in a redundant format, such as carry-save. Before this result may undergo standard Booth recoding it must be processed by a CP adder to put it in binary format--consuming valuable time. A CP adder needs to propagate a carry from the rightmost position to the leftmost, resulting in n delays for an n bit number to be put into binary format. The CP adder of the present invention has to CP round over-100-bit numbers, meaning over 100 CP add propagation delays. Performing such CP adds after each multiply of a series of multiplies consumes precious time.

One aspect of the present invention which significantly speeds up series-multiplies is a method of precoding redundant format numbers to provide Booth recoding inputs, but without requiring the long delay of a CP adder. The circuitry according to this aspect of the present invention accepts numbers in carry save format and processes them to produce inputs to a group of standard Booth recoders such as the one in FIG. 8c. The Booth recoders are organized in the basically same manner as in FIG. 8d, except that the digits, which are in carry save form rather than binary form, are processed by an intermediate layer of logic. This layer of logic consists of a number of identical units, the logic components for one such carry save recode unit being illustrated in FIG. 10, and indicated generally by reference numeral 200. The layer of Carry Save Recoding (CSR) logic does not simply determine the corresponding binary form of the carry save digits, because that would still take n delays for n pairs of carry save digits. Rather, the CSR units accept as inputs the two pairs of carry save bits of primary consideration (S.sub.1 S.sub.0 coupled to a.sub.1 a.sub.0 and C.sub.1 C.sub.0 coupled to b.sub.1 b.sub.0), and two inputs Ca and Cb from the CSR unit to the right (the next less significant group) Each CSR unit outputs r.sub.0, r.sub.1, and r.sub.2 (the information needed by the Booth recoding logic) and also the two bits Caout and Cbout needed by the next CSR unit. The logic is designed such that Cbout is determined by the sum of S.sub.1 S.sub.0 and C.sub.1 C.sub.0 and is completely independent of Ca and Cb. The sum of S.sub.1 S.sub.0 and C.sub.1 C.sub.0 can range from 0 to 6, and Cbout is set whenever this sum is greater than 1. Caout is set determined by the sum of S.sub.1 S.sub.0 and C.sub.1 C.sub.0 and also by Cb: Caout is set if the sum is 6, or if Cb is set and the sum is either 5 or 1. (The mathematical justification for this is explained below.) Caout is affected by Cb, but Cb was determined entirely by the preceding CSR unit, without reference to CSR units earlier (dealing with less significant bits) than that. Also note that Ca gets routed directly to r.sub.0, and that is the only use made of it. A CSR unit could direct the Caout output directly to the r.sub.0 input of the next more significant Booth recoder, and CSR units would then not need a Ca input.

The CSR unit does not have to wait for a valid Ca before computing its Cbout. Therefore, each CSR unit need wait only for inputs which originate in the unit next to it. No matter how many bits are involved, no matter how many CSR units are chained together, there is only one CSR unit propagation delay total before each CSR unit has all the necessary inputs to compute the information needed by the Booth recoders. FIG. 11 illustrates how carry save recoders and Booth recoders are combined to provide the Booth recoding information for a number consisting of 8 pairs of carry-save bits. The carry save recoders can also process normal binary numbers by setting the carry bits to zero.

The circuit of FIG. 10 is only one possible circuit for a CSR unit to provide inputs to radix 4 Booth recoding logic. As will become apparent below, the Ca and Cb outputs could be generated differently, and CSR units can also be constructed for Booth recoding logic having radix 8, 16, etc.

To understand the theory of operation of this aspect of the invention, examples for radix 4 Booth recoding will be used, and it should be noted that an alternative way of expressing the information of FIG. 8b is that the Y.sub.i+1 bit has a weight of -2, the Y.sub.i bit has a weight of 1, and the Y.sub.i-1 also has a weight of 1. (For radix 2.sup.b Booth recoding the bit Y.sub.i-1 has a weight of +1, the bit Y.sub.i+b-1 has a weight of -2.sup.b-1, and bits Y.sub.i+j (where 0.ltoreq.j.ltoreq.(b-2)) have a weight of 2.sup.j.

A straightforward way to Booth recode 4 carry save bit pairs is shown in FIG. 12A. The carry save bits are fed into a four bit adder 200 having a carry in C.sub.in and a carry out C.sub.out, and the sum total bits T.sub.0 through T.sub.3 are provided to Booth recoders 100 in a standard way. A very similar implementation is shown in FIG. 12B, in which two two-bit adders are used, having an intermediate carry C. This intermediate carry C is still dependent upon C.sub.in, so no propagation delays are saved by this configuration.

However, it is possible to make the intermediate carry C a predictive carry. That is, the predictive carry will be determined entirely by the two redundant format bits being added, without reference to the carry in. Of course, the predictive carry will not always correspond to the true carry, but it can be corrected for. This is because a carry, which gets added to the Y.sub.i position, carries the same weight and has the same net effect upon the final product as does the Y.sub.i-1 input r.sub.0. (Even though different Booth recoding output might get generated, the net effect of the Booth recoding output will be the same.) In order to use the overlap bit Y.sub.i-1 of the next more significant Booth recoder in a corrective manner, however, it must be decoupled from the T.sub.1 output, which goes to the Y.sub.i+1 input r.sub.2 of the current Booth recoder. This is shown in FIG. 13, in which Carry Save Recoder (CSR) units 200 have replaced the two bit adders 210 of FIG. 12B. These CSR units are modified adders that generate the sums T.sub.1 and T.sub.0, a predictive carry Cbout which is used as a carry in by the next more significant unit, and a corrective carry, or overlap bit, Caout, decoupled from the T.sub.1 output, that is used as the Y.sub.i-1 input to the next more significant Booth recoding unit.

Keeping in mind that in the configuration of FIG. 12B both the carry C and the T.sub.1 bit (as coupled to the Y.sub.i-1 input to the next more significant Booth recoder) have the same weight (+1), we can group their combined effects into three cases: (A) both the true carry and T.sub.1 are ones (combined weight +2); (B) exactly one of the true carry or T.sub.1 is a one (combined weight +1); and (C) neither the true carry nor T.sub.1 is a one (combined weight+0). Therefore, predictive carry Cbout and corrective carry Caout can be generated in any manner that creates the correct combined weights in these cases. Predictive carry Cbout must be generated in all configurations of a.sub.1 b.sub.1 a.sub.0 b.sub.0 which might result in case (A) (depending on the carry in), cannot be generated in any configurations that might result in case (C), and otherwise can be generated optionally. Corrective carry Caout depends upon the actual resultant case (A), (B), or (C), and is generated so that the corresponding combined weight (+2), (+1), (+0), respectively, will be produced by Caout and Cbout together.

Now the carry generation scheme of the circuit of FIG. 10 can be understood. As mentioned above, this circuit generates a Cbout signal whenever the sum of S.sub.1 S.sub.0 and C.sub.1 C.sub.0 is 2, 3, 4, 5 or 6. A sum of 0 definitely results in case (C) above, as would a sum of 1 if carry in Cb were not set, so Cbout must not be generated. A sum of 6 definitely results in case (A), as might a sum of 5, so Cbout must be generated for these. A sum of 2, 3 or 4 definitely results in case (B), and Cbout is set. Caout then takes into account the effect of carry in Cb, and is set either if the result is in case (A), or if the result is in case (B) and Cbout was not set. A sum of 6, or a sum of 5 with Cb set, results in case (A), so Caout is set for these. Also, a sum of 1 with Cb set results in case (B), and Cbout was not set for a sum of 1, so Caout is set. For all other situations Caout is not set. As mentioned above, a variety of other carry generation schemes could be used so long as they resulted in the correct weights for cases (A), (B) and (C). The particular carry generation scheme of FIG. 10 simplifies circuit layout.

This scheme can now be generalized for a b-bit CSR unit to replace a b-bit adder to create a predictive carry and inputs to Booth recoding logic. The generalized scheme has the same three cases with the same weights. Again, the cases are determined by the final total sum Tsum of the redundant format bits plus the carry in. Case (A) occurs when both Tsum.gtoreq.2.sup.b (there is a true carry out) and T.sub.b-1 (corresponding to the overlap bit) is set. Case (B) occurs when exactly one of these conditions is true, and Case (C) occurs when neither of them are true.

If we now apply these generalized rules to 3-bit CSR units, the circuit of FIG. 14 can be constructed, in which 3-bit CSR units are used with radix 8 Booth recoders to Booth recode a 16 bit number in carry save format, with required zero padding. Case (A) occurs when both Tsum.gtoreq.8 and T.sub.2 is set. Case (C) occurs when Tsum<4, and case (B) occurs all other times. A 3-bit CSR unit designed along the same lines as the 2-bit CSR unit of FIG. 10 would then generate Cbout whenever the sum of a.sub.2 a.sub.1 a.sub.0 and b.sub.2 b.sub.1 b.sub.0 (the carry save inputs) was greater than 3. Caout would be generated whenever the sum was 14, 13, 12, or 11 with Cb set, all corresponding to case (A), or if the sum was 3 with Cb set, corresponding to case (B) with Cbout not set.

CSR units thus replace adders to apply numbers in carry save format to Booth recoding logic, without requiring the full propagation delay that the adders would require. CSR units can even be mixed with CSR units of other sizes or with adders, although this might not be most efficient. The only significant requirement is that the Cbout carry correspond to a Y.sub.i input to a Booth recoder, and that the Caout corrector be coupled to the overlap (Y.sub.i-1) bit of that Booth recoder. Just as an example of the diverse configurations possible, in FIG. 15A a 1-bit CSR unit 300 is joined with 2-bit adders 310 and 3-bit adders 320 to provide inputs to radix 4 Booth recoders 100 for a 100 bit number in carry save format. If only adders were used and no CSR unit, 100 CP add propagations would be necessary. The use of CSR unit 300 generates a predictive carry in the middle of the number, reducing the propagation delay essentially by half. One possible schematic for CSR unit 300 is shown in FIG. 15B. This reduction by half of the propagation delay is an indication of the power of the predictive carry.

The preferred embodiment incorporates another aspect of the present invention, directed to providing IEEE exact rounding. By this we mean rounding which meets the requirements set forth in ANSI/IEEE Standard 754-1985. The IEEE standard sets forth two formats for floating point numbers--single precision and double precision. The floating point numbers have an exponent portion e and a fraction portion f, and their magnitude can be represented as f.times.2.sup.e, where f is 1.f.sub.1 f.sub.2 f.sub.3 . . . f.sub.n, where n is the number of bits in the fraction portion. Single precision numbers have 23 bits in the fraction portion and double precision numbers have 52 bits in the fraction portion.

The IEEE standard generally describes rounding as taking a number regarded as infinitely precise and, if necessary, modifying it to fit in the destination's format while signalling then that the result is inexact. The majority of mathematical operations are required to be performed as if an intermediate result is first produced with infinite precision and unbounded range, and which is then rounded according to one of a set of rounding modes.

The first rounding mode is called Round-to-Nearest. Under this mode the representable value nearest to the infinitely precise result shall be delivered. If the two nearest representable values are equally near, the one with its least significant bit zero shall be delivered. There are also three user-selectable rounding modes Round Toward +.infin., in which case the result shall be the format's value closest to and no less than the infinitely precise result; Round Toward -.infin., in which case the result shall be the format's value closest to and no greater than the infinitely precise result; and Round to Zero, in which case the result shall be the format's value closest to and no greater in magnitude than the infinitely precise result.

The IEEE standard applies in the same way to single precision and to double precision numbers; the reference point is always the least significant bit of the destination format. For single precision numbers this is the 23rd bit, for double precision, the 52nd. For ease of discussion, these two cases will not be differentiated, and the term "lsb" shall mean the least significant bit of the appropriate format. Also, with but minor exceptions the rounding procedure is the same for division as it is for computing square roots. So, all the discussion directed to the quotient will also apply to the square root, except as noted.

It should be noted that for round to nearest, in the worst case the invention needs to determine whether the real quotient falls just below a 1/2lsb point, just above it, or directly on it. For the worst case in the other rounding modes the invention needs to determine whether the real quotient falls just below an lsb point, just above it, or directly on it. This is a subset of the 1/2lsb points, where the value of the bit at 1/2lsb is zero. So, determining the relation of the real quotient to a 1/2lsb point will allow all rounding modes to be done.

For whatever calculation is being done, the present invention will compute an initial result which is accurate to several bits less significant than the lsb. At this point the quotient is truncated at the 1/4lsb to produce an original calculated quotient Q.sub.0. The error will be termed .epsilon., and is defined through the equation .sub.Qr =Q.sub.0-.epsilon., where Qr is the real quotient. The data path has been implemented such that .vertline..epsilon..vertline.<1/4lsb. It may also be that the magnitude of the error will be less than that, but that statement will definitely hold true, so -1/4lsb<.epsilon.<+1/4lsb.

Now, an extra error of 1/4lsb is deliberately introduced by adding 1/4lsb to Q.sub.0 to produce calculated quotient Q.sub.c. The range of .epsilon. is now 0<.epsilon.<1/2lsb. This now assures that the calculated quotient is strictly greater than the real quotient. We now have a better idea of where to "look" to find the real quotient.

For purposes of notation, let N be any number that has a least significant bit of lsb, whether that bit is a zero or a one. We can now make the general statement N.ltoreq.Q.sub.c <N+1 (where the units are understood to be one lsb and will be so from now on). This can be divided into two cases, (1) N.ltoreq.Q.sub.c <N+1/2, and (2) N+1/2.ltoreq.Q.sub.c <N+1. Let us consider these separately.

In all cases, 0.ltoreq..epsilon.<1/2, so for case (1) N-1/2<Q.sub.r <N+1/2. From above, all that needs be determined is Q.sub.r 's standpoint relative to 1/2lsb points. The only such point within this range is at N. Note the range of Q.sub.c for case (1): the only 1/2lsb point contained within it is also at N. So, Q.sub.c is truncated at the 1/2lsb point, producing a Q.sub.t equal to N.

For case (2), where N+1/2.ltoreq.Q.sub.c <N+1, the range of Q.sub.r is N<Q.sub.r <N+1. The only 1/2lsb point within this range is N+1/2lsb. Note that again, this is also the only 1/2lsb point within the range of Q.sub.c, so Q.sub.c is truncated at the 1/2lsb point to produce a Q.sub.t equal to N+1/2lsb. For both case (1) and case (2) it is seen that the critical value against which Q.sub.r needs to be compared can be obtained by truncating Q.sub.c at the 1/2lsb point to produce a Q.sub.t.

Once this critical value is identified, the relation of Q.sub.r to it can be determined by computing a Partial Remainder PR. For division, PR=Y-X.multidot.Q.sub.t ; for square roots, PR=X-Q.sub.t .multidot.Q.sub.t.

For the multiplication of X.multidot.Q.sub.t all of the significant bits of the result are retained for comparison to Y, and Y is padded with zeros in enough lesser significant digits to make it the same size as the X.multidot.Q.sub.t product. In all cases, therefore, it can be determined whether Q.sub.r lies just below, exactly on, or just above a 1/2lsb point, which provides sufficient information for IEEE exact rounding.

It is to be understood that the above description is intended to be illustrative and not restrictive. Many embodiments will be apparent to those of skill in the art upon reviewing the above description. The scope of the invention should, therefore, be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled.