Quotient Digit Selection Logic for Floating Point Division/Square Root

1. Field of the Invention

The present invention relates to the field of floating point dividers in microprocessors. Specifically, the present invention relates to quotient digit selection rules in Sweeny, Robertson, Tocher (SRT) division/square root implementations which prevent negative final partial remainders from occurring when results are exact, and which provide support for correct rounding in all rounding modes.

2. Discussion of the Related Art

The SRT algorithm provides one way of performing non-restoring division. See, J. E. Robertson, "A new class of digital division methods," IEEE Trans. Comput., vol. C-7, pp. 218-222, September 1958, and K. D. Tocher, "Techniques of multiplication and division for automatic binary computers," Quart. J. Mech. Appl. Math., vol. 11, pt. 3, pp. 364-384, 1958. Digital division takes a divisor and a dividend as operands and generates a quotient as output. The quotient digits are calculated iteratively, producing the most significant quotient digits first. In SRT division, unlike other division algorithms, each successive quotient digit is formulated based only on a few of the most significant partial remainder digits, rather than by looking at the entire partial remainder, which may have a very large number of digits. Since it is not possible to insure correct quotient digit selection without considering the entire partial remainder in any given iteration, the SRT algorithm occasionally produces incorrect quotient digit results. However, the SRT algorithm provides positive, zero, and negative quotient digit possibilities. If the quotient digit in one iteration is overestimated, then that error is corrected the next iteration by selecting a negative quotient digit. In SRT division, quotient digits must never be underestimated; quotient digits must always be overestimated or correctly estimated. By never underestimating any quotient digits, the partial remainder is kept within prescribed bounds so as to allow the correct final quotient to be computed. Because the SRT algorithm allows negative quotient digits, the computation of the final quotient output usually involves weighted adding and subtracting of the quotient digits, rather than merely concatenating all the quotient digits as in normal division.

The higher the radix, the more digits of quotient developed per iteration but at a cost of greater complexity. A radix-2 implementation produces one digit per iteration; whereas a radix-4 implementation produces two digits per iteration. FIG. 1 illustrates a simple SRT radix-2 floating point implementation. The simple SRT radix-2 floating point implementation shown in FIG. 1 requires that the divisor and dividend both be positive and normalized; therefore, 1/2.ltoreq.D, Dividend <1. The initial shifted partial remainder, 2PR›0!, is the dividend. Before beginning the first quotient digit calculation iteration, the dividend is loaded into the partial remainder register 100; thus, the initial partial remainder is the dividend. Subsequently, the partial remainders produced by iteration are developed according to the following equation.

In Equation 1, q.sub.i+1 is the quotient digit, and has possible values of -1, 0, or +1. This quotient digit q.sub.i+1 is solely determined by the value of the previous partial remainder and is independent of the divisor. The quotient selection logic 102 takes only the most significant four bits of the partial remainder as input, and produces the quotient digit. In division calculations, the divisor remains constant throughout all iterations. However, square root calculations typically involve adjustments to the divisor stored in the divisor register 101 after each iteration. Therefore, the independence of the quotient digit selection on the divisor is an attractive feature for square root calculations.

The partial remainder is typically kept in redundant carry save form so that calculations of the next partial remainder can be performed by a carry-save adder instead of slower and larger carry-propagate adders. The partial remainder is converted into non-redundant form after all iterations have been performed and the desired precision has been reached. Because the SRT algorithm allows overestimation of quotient digits resulting in a negative subsequent partial remainder, it is possible that the last quotient digit is overestimated, so that the final partial remainder is negative. In that case, since it is impossible to correct for the overestimation, it is necessary to maintain Q and Q-1, so that if the final partial remainder is negative, Q-1 is selected instead of Q. The quotient digits are normally also kept in redundant form and converted to non-redundant form at the end of all iterations. Alternatively, the quotient and quotient minus one (Q and Q-1) can be generated on the fly according to rules developed in M. D. Ercegovac and T. Lang, "On-the-fly rounding," IEEE Trans. Comput., vol. 41, no. 12, pp. 1497-1503, December 1992.

The SRT algorithm has been extended to square root calculations allowing the utilization of existing division hardware. The simplified square root equation looks surprisingly similar to that of division. See, M. D. Ercegovac and T. Lang, "Radix-4 square root without initial PLA," IEEE Trans. Comput., vol. 39, no. 8, pp. 1016-1024, August 1990. The iteration equation for square root calculations is as follows.

In Equation 2, the terms in parentheses are the effective divisor. For square root calculations, the so-called divisor is a function of Qi, which is a function of all the previous root digits q1 through qi. The root digits hereinafter will be referred to as "quotient digits" to maintain consistency in terminology. Therefore, in order to support square root calculation using the same hardware as used for division, on-the-fly quotient generation is required in order to update the divisor after each iteration.

Binary division algorithms are analogous to standard base 10 long division which is taught in grammar school. In R/D=Q, each quotient digit for Q is guessed. In order to determine the first quotient digit, a guess for the proper quotient digit is multiplied by the divisor, and that product is subtracted from the dividend to produce a remainder. If the remainder is greater than the divisor, the guess for the quotient digit was too small; if the remainder is negative, the guess for the quotient digit was too large. In either case, when the guess for the quotient digit is incorrect, the guess must be changed so that the correct quotient digit is derived before proceeding to the next digit. The quotient digit is correct when the following relation is true: 0.ltoreq.PR<D, in which PR stands for the partial remainder after subtraction of the quotient digit multiplied by the divisor.

The key to the SRT division algorithm is that negative quotient digits are permitted. For example, in base 10, in addition to the standard digits 0 through 9, quotient digits may take on values of -1 through -9. Consider the division operation 600.div.40. If the correct quotient digits are selected for each iteration, the correct result is 15. However, assume for the moment that during the first iteration, a quotient digit of 2 was incorrectly guessed instead of the correct digit of 1. The partial remainder after 2 has been selected as the first quotient digit is 600-(2*40*10.sup.1)=-200. According to SRT division, this error can be corrected in subsequent iterations, rather than having to back up and perform the first iteration again. According to SRT division, assume that the second quotient digit is correctly guessed to be -5. The partial remainder after that iteration will be -200-(-5*40*100)=0. When the partial remainder after an iteration is zero, the correct values for all the remaining digits are zeros. Thus, the computed result is 2*10.sup.1 +-5*10.sup.0 =15, which is the correct result. The SRT algorithm thus allows an overestimation of any given quotient digit to be corrected by the subsequent selection of one or more negative quotient digits. It is worth noting that the estimated quotient digit must not be more than one greater than the correct quotient digit in order to subsequently reduce the partial remainder to zero, thus computing the correct result. If errors greater than positive one were allowed in estimating quotient digits, then quotient digits less than -9 (for example -10, -11, etc.) would be required in base 10. Similarly, since the range of quotient digits is not expanded in the positive direction at all according to the SRT algorithm, underestimation of the correct quotient digit is fatal, because the resulting partial remainder will be greater than the divisor multiplied by the base, and a subsequent quotient digit higher than 9 (for example 10, 11, etc.) in base 10 would be required. Therefore, in order to keep the partial remainder within prescribed bounds, the quotient digit selection must never underestimate the correct quotient digit, and if it overestimates the quotient digit, it must do so by no more than one.

It is possible to guarantee that the above criteria for keeping the partial remainder within prescribed bounds will be satisfied without considering all the partial remainder digits. Only a few of the most significant digits of the partial remainder must be considered in order to choose a quotient digit which will allow the correct result to be computed.

SRT division requires a final addition after all quotient digits have been selected to reduce the redundant quotient representation into standard non-redundant form having only non-negative digits.

In binary (base 2) which is utilized in modern electrical computation circuits, SRT division provides quotient digits of +1, 0, or -1. The logic 102 which generates quotient selection digits is the central element of an SRT division implementation. Early research indicated that only the most significant three bits of redundant partial remainder are necessary inputs for a radix-2 quotient digit selection function. (See, S. Majerski, "Square root algorithms for high-speed digital circuits," Proc. Sixth IEEE Symp. Comput. Arithmetic., pp. 99-102, 1983; and D. Zuras and W. McAllister, "Balanced delay trees and combinatorial division in VLSI," IEEE J. Solid-State Circuits., vol. SC-21, no. 5, pp. 814-819, October 1986.) However more recent studies have shown that four bits are required to correctly generate quotient digit selection digits and keep the partial remainder within prescribed bounds. (See M. D. Ercegovac and T. Lang, Division and Square Root: Digit-recurrence Algorithms and Implementations, Kluwer Academic Publishers, 1994, ch. 3; S. Majerski, "Square-rooting algorithms for high-speed digital circuits," IEEE Trans. Comput., vol. C-34, no. 8, pp. 724-733, August 1985; P. Montuschi and L. Ciminiera, "Simple radix 2 division and square root with skipping of some addition Steps," Proc. Tenth IEEE Symp Comput. Arithmetic. pp. 202-209, 1991; and V. Peng, S. Samudrala, and M. Gavrielov, "On the implementation of shifters, multipliers, and dividers in floating point units," Proc. Eighth IEEE Symp. Comput. Arithmetic, pp. 95-101, 1987. The selection rules according to the prior art can be expressed as in the following equations in which PR represents the most significant four bits of the actual partial remainder, and in which the decimal point appears between the third and fourth most significant digits. The partial remainder is in two's complement form, so that the first bit is the sign bit.

q.sub.i+1 =1, if 0.ltoreq.2PR›i!.ltoreq.3/2,

q.sub.i+1 =0, if 2PR›i!=-1/2,

q.sub.i+1 =-1, if -5/2.ltoreq.2PR›i!.ltoreq.-1.

Because the partial remainder is stored in register 100 in carry-save form, the actual most significant four bits are not available without performing a full carry propagate addition of the carry and sum portions of the partial remainder. Because it is desirable to avoid having to perform a full carry propagate addition during each iteration in order to compute the most significant four bits of the partial remainder, quotient digit selection rules can be developed using an estimated partial remainder.

The estimated partial remainder (PR.sub.est) is computed using only a four-bit carry propagate adder that adds the most significant four bits of the carry and sum portions of the actual partial remainder. This simplification represents a significant savings of latency because the equivalent of a full 59 bit carry propagate addition would otherwise be required to compute the actual most significant four bits of the partial remainder. The estimated partial remainder PR.sub.est does not reflect the possibility that a carry might propagate into the bit position corresponding to the least significant bit position of the estimated partial remainder if a full 59-bit carry propagate addition had been performed. The truth table below describes the quotient selection rules according to the prior art where the most significant four bits of the estimated partial remainder are used to select the correct quotient digit. Thus, the truth table below takes into consideration the fact that the most significant four bits of the true partial remainder may differ from the most significant four bits of the estimated partial remainder.

In the above truth table, the four bits representing 2PR.sub.est are a non-redundant representation of the most significant four carry and sum bits of the partial remainder. The fourth bit is the fraction part, so that the resolution of the most significant four bits of the partial remainder is 1/2.

The quotient selection logic is designed to guess correctly or overestimate the true quotient result, e.g. predicting 1 instead of 0, or 0 instead of -1. The SRT algorithm corrects itself later if the wrong quotient digit has been chosen.

The prior art truth table for SRT radix-2 quotient selection logic has several don't care inputs because the partial remainder is constrained to -5/2.ltoreq.2PR›i!.ltoreq.3/2. The estimated partial remainder is always less than or equal to the true most significant bits of the partial remainder because the less significant bits are ignored. Therefore, there is a single case (marked with an asterisk in the above table) where the estimated partial remainder appears to be out of bounds. By construction, the real partial remainder is within the negative bound because the SRT algorithm as implemented will never produce an out of bounds partial remainder, so -1 is the appropriate quotient digit to select. There are two other cases (those corresponding to the entries for 111.0 and 111.1 in Table I) in which the quotient digit selected based on the estimated partial remainder differs from what would be chosen based on the real partial remainder. However, in both of these instances of "incorrect" quotient digit selection, the quotient digit is not underestimated and the partial remainder is kept within prescribed bounds, so that the final result will still be generated correctly.

The following table II illustrates the quotient selection logic described in Table I in a simplified form. In the table below, an "x" represents a "don't care" logic variable. Obviously, the third case, in which 1xx.x produces a -1 quotient digit does not apply when the estimated partial remainder is 111.1, such that the second entry applies, and the correct quotient digit is 0.

Floating point operations generate a sticky bit along with the result in order to indicate whether the result is inexact or exact. When the result is inexact, the sticky bit is asserted; conversely, when the result is exact, the sticky bit is deasserted. Essentially, the sticky bit indicates whether or not any of the bits of less significance are non-zero. The sticky bit is also used with the guard and round bits for rounding according to IEEE Standard 754. See, "IEEE standard for binary floating-point arithmetic," ANSI/IEEE Standard 754-1985, New York, The Institute of Electrical and Electronic Engineers, Inc., 1985.

For divide and square root operations, the sticky bit is determined by checking if the final partial remainder is non-zero. The final partial remainder is defined as the partial remainder after the desired number of quotients bits have been calculated. Since the partial remainder is in redundant form, a carry-propagate addition is performed prior to zero-detection. A circuit for computing the sticky bit is shown in FIG. 2. In FIG. 2, the carry 201 and sum 202 portions of the final partial remainder are added together by the carry propagate adder 200. The most significant bit output by the adder 200 is the sign bit 203 of the final partial remainder. As illustrated in FIG. 1, the division hardware accumulates the quotient Q and the quotient minus one Q-1. When the final partial remainder is negative, Q-1 is the proper quotient; when the final partial remainder is zero or positive, Q is the correct quotient. Thus, the sign bit 203 is used to select the correct quotient. Referring again to FIG. 2, the zero detector 204 determines if all bits of the non-redundant final partial remainder 205 are zeros and outputs a the sticky bit 206. The zero detector 204 is logically equivalent to a large 59-input OR gate.

At first glance, the above solution seems perfectly reasonable for all final partial remainder possibilities, positive or negative. However, in the rare case in which the result is exact, the final partial remainder will be equal to the negative divisor. For example, consider a number divided by itself, as illustrated in the table below, in which PR›i! represents the partial remainder after the ith quotient digit has been selected.

Since the dividend is always positive and normalized, the quotient digit from the first iteration is one. This is a consequence of the fact that a positive normalized number has a sign bit of zero and a most significant digit of one. When a positive normalized number is divided by two, presumably by right shifting by one bit position, the most significant bit necessarily becomes a zero. (If a negative number is divided by two, the most significant bit is one, because the most significant bits are sign extended so as to match the sign bit giving the correct two's complement representation.) When the most significant bit is a zero, Table II above dictates that a quotient digit of one should be selected.

For the second iteration shown in Table III, the partial remainder PR›1! is zero which causes the second quotient digit to be one. For all subsequent iterations, the partial remainder will equal the negative divisor and quotient digits of minus one will be selected. After the last iteration, performing a sign detect on the final partial remainder PR›n! determines that the final partial remainder is negative and indicates that Q-1 should be chosen. This is in fact the correct result. However, this same final partial remainder is non-zero which erroneously suggests an inexact result and erroneously suggests that the sticky bit should be asserted.

This problem extends to any division operation for which the result should be exact. Fundamentally, the problem is a consequence of the fact that the quotient selection logic is defined to guess positive for a zero partial remainder and correct for it later, as illustrated by Table II. The prior art dividers require one processor cycle for restoration of negative final partial remainders prior to sticky bit calculation. To insure correct rounding, it is necessary to correctly compute the sticky bit. It would be advantageous to develop a divider which did not restore negative final partial remainders, but that could nevertheless guarantee correct computation of the sticky bit.

The present invention provides the ability to correctly and efficiently compute the sticky bit during floating point division and square root computations when the final partial remainder is negative. The present invention further provides an optimized quotient selection circuit implementing quotient selection rules according to the present invention with minimized latency and size.

The present invention provides an enhanced quotient digit selection function which prevents the working partial remainder from becoming negative if the result is exact. This creates a one cycle savings since negative partial remainders no longer need to be restored before calculating the sticky bit.

According to the present invention, the quotient digit selection logic is modified so as to prevent a partial remainder equal to the negative divisor from occurring. To correctly and efficiently compute the sticky bit for exact division results, enhancing the quotient digit selection function so as to prevent the formation of a negative partial remainder equal to the negative divisor is an ideal solution because it saves hardware and improves latency. Extra hardware is eliminated because it is no longer necessary to provide any extra mechanism for restoring the preliminary final partial remainder. Latency is improved because additional cycle time is not required to restore negative preliminary partial remainders.

According to the present invention, the quotient digit selection logic is altered so as to choose a quotient digit of zero instead of a quotient digit of one when the actual partial remainder is zero. Using a five bit estimated partial remainder where the upper four bits are zero, a possible carry propagation into fourth most significant bit is detected. This can be accomplished by looking at the fifth most significant sum and carry bits of the redundant partial remainder. If they are both zero, then a carry propagation out of that bit position into the least significant position of the estimated partial remainder is not possible, and a quotient digit of zero is chosen. In the alternative case in which one or both of the fifth most significant carry or sum bits of the redundant partial remainder are ones, a quotient digit of one is chosen.

According to an alternative embodiment of the present invention, the quotient digit selection logic is additionally altered so as to choose a quotient digit of negative one when the actual partial remainder is negative one half or less. Using the five bit estimated partial remainder, a quotient digit of negative one is selected when the the upper four bits are ones while the fifth most significant bit is zero. This can be accomplished by looking at the fifth most significant sum and carry bits of the redundant partial remainder. If they are both zero, then a carry propagation out of that bit position into the least significant position of the estimated partial remainder is not possible, and a quotient digit of negative one is chosen. In the alternative case in which one or both of the fifth most significant carry or sum bits of the redundant partial remainder are ones, a quotient digit of zero is chosen.

These and other features and advantages of the present invention will be apparent from the FIGS. as explained in the detailed description of the invention.

FIG. 1 is a simplified logical illustration of a conventional hardware division module implementing the iterative SRT division algorithm.

FIG. 2 is a simplified logic diagram of a conventional sticky bit computation circuit.

FIG. 3 is a simplified logical illustration of one hardware division module which implements the iterative SRT division algorithm or alternatively restores negative final partial remainders.

FIG. 4 is a simplified logical illustration of another hardware division module which implements the iterative SRT division algorithm or alternatively restores negative final partial remainders.

FIG. 5 illustrates the output partial remainder and selected quotient digits as functions of the input partial remainder according to the prior art.

FIG. 6 illustrates the output partial remainder and selected quotient digits as functions of the input partial remainder divisor according to an embodiment of the present invention suitable for use in round to zero and round to nearest modes.

FIG. 7 illustrates the output partial remainder and selected quotient digits as functions of the input partial remainder according to an alternative embodiment of the present invention suitable for use in all rounding modes.

FIG. 8 illustrates the ranges of the output partial remainder corresponding to the three possible quotient digits as functions of the divisor as well as the ranges of the quotient digit selections according to the prior art, according to an embodiment of the present invention, and according to an alternative embodiment of the present invention.

FIG. 9 is a simplified logic diagram of a circuit implementing the quotient selection rules according to the present invention.

The detailed description of the invention explains the Figures in greater detail.

Several options must be considered in order to correctly compute the sticky bit in the case of negative final partial remainders in floating point division operations. The simplest solution is to restore negative final partial remainders by adding the divisor before performing zero-detection. Given the area expense of an additional carry-propagate adder, the solution should try to take advantage of existing hardware. Two ways to achieve this are shown in FIGS. 3 and 4.

The solution in FIG. 3 takes advantage of the carry save adder 300 that already exists to implement the basic iterative step. During normal operation, the final cycle signal 301 is low so that multiplexors 302 and 303 are configured so as to implement the standard iterative step described in conjunction with FIG. 1. However, during the final cycle, the final cycle signal 301 is high, and the multiplexor 303 directs the divisor D 304 to the adder 300 to be added to the partial remainder carry and sum portions stored in register 305. If the sign of the preliminary partial remainder was negative, multiplexor 306 outputs the result from adder 300 to the final partial remainder flip-flop 307.

The solution illustrated in FIG. 4 reuses the carry-propagate adder 400 that existed for the sticky bit computation shown in FIG. 2. After an iterative datapath such as that depicted in FIG. 1 has produced quotient digits to the desired precision, the redundant partial remainder's carry and sum portions (PRC and PRS, respectively) are input into multiplexors 403 and 402, respectively, while the final cycle signal is low. Carry propagate adder 400 then reduces the redundant partial remainder into a preliminary non-redundant final partial remainder. The most significant output bit 404 of the carry propagate adder 400 indicates the sign of the preliminary final partial remainder. The non-redundant preliminary final partial remainder is latched in flip-flop 406, then the final cycle signal 401 is asserted. If the sign bit 404 indicates that the preliminary final partial remainder is negative, multiplexor 405 is directed to supply the divisor to multiplexor 402 which supplies it to the adder 400. Meanwhile, multiplexor 403 is directing the preliminary non-redundant partial remainder to the other input of the adder 400. The zero detector 407 then produces the sticky bit using the restored partial remainder. If the sign bit indicated a non-negative partial remainder, then multiplexor 405 directs zero to be added to the partial remainder rather than the divisor D.

Both of the solutions illustrated in FIGS. 3 and 4 add extra multiplexor hardware and require an additional cycle to compute the sticky bit when the preliminary final partial remainder is negative. The solution in FIG. 3 especially impacts cycle time for basic iterations since the multiplexor 302 is on the partial remainder formation critical path. Moreover, variable latency instructions in a pipelined superscalar processor make instruction scheduling and bypass control logic much more complex and are generally undesirable. Therefore, even if only a fraction of the total distribution of division operations requires the extra cycle in order to correctly compute the sticky bit, it greatly complicates the processor design to attempt to save the extra cycle for those division instructions which do not result in a negative preliminary final partial remainder. Therefore, the net effect of restoring negative partial remainders is to add another cycle of latency for all divide and square root operations.

According to the present invention, the quotient digit selection logic is modified so as to prevent a partial remainder equal to the negative divisor from occurring. To correctly and efficiently compute the sticky bit for exact division results, enhancing the quotient digit selection function so as to prevent the formation of a negative partial remainder equal to the negative divisor is an ideal solution because it saves hardware and improves latency. Extra hardware is eliminated because it is no longer necessary to provide any extra mechanism for restoring the preliminary final partial remainder. Latency is improved because no additional cycle time is required to restore negative preliminary partial remainders.

According to the present invention, the quotient digit selection logic is altered so as to choose a quotient digit of zero instead of a quotient digit of one when the actual partial remainder is zero. This suggests choosing q›i+1!=0 for 2PR›i!.sub.est =000.0. However, since the quotient digit selection function works on an estimated partial remainder, rather than the actual partial remainder, careful design of the quotient selection logic is necessary. An estimated partial remainder could appear to be less than 1/2 when, in reality, adding the less significant redundant partial remainder bits causes a 1 to propagate into the least significant bit position of the most significant four bits of the actual partial remainder, so that the actual partial remainder is equal to or greater than 1/2.

If the actual partial remainder is 1/2 (or greater), q=1 should always be chosen since the divisor is constrained by 1/2.ltoreq.D<1. If the actual partial remainder is greater than 1/2 and q=0 is chosen, the next partial is derived by left shifting the actual partial remainder by one bit position without first adding or subtracting the divisor. This left shifting will produce a next actual partial remainder that exceeds the upper bound of 3/2 required to converge on the correct quotient.

If the divisor was greater than the actual partial remainder and positive one was chosen as the quotient digit when the correct quotient digit was zero, this error can be corrected in later iterations. However, if zero is chosen as the quotient digit when the estimated partial remainder is 000.0 and the correct quotient digit is one, there is no way to correct for this error. There is no way to correct for underestimation if q=0 is selected when q=1 was the correct quotient digit. The result will be irreversibly incorrect and the next partial remainder will be out of the prescribed bounds necessary to converge onto the correct result.

Performing binary addition on the full partial remainder eliminates the estimation problem, but defeats the timing and area benefits of SRT division. A quotient digit of zero could be chosen only when the full partial remainder is zero, but detection of zero for the full partial remainder would be counterproductive for the same reasons.

A simple alternative to the solutions proposed above is to detect a possible carry propagation into the least significant bit of the most significant four bits of the actual partial remainder. This can be accomplished by looking at the fifth most significant sum and carry bits of the redundant partial remainder, PR.sub.S,msb-.sub.4 and PR.sub.C,msb- 4, respectively. If they are both zero, then a carry propagation out of that bit position into the least significant position of the estimated partial remainder is not possible, and a quotient digit of zero should be chosen. In the alternative case in which one or both of the fifth most significant carry or sum bits of the redundant partial remainder are ones, a quotient digit of one should be chosen. Even though the lower bits of the partial remainder could be non-zero, the partial remainder is still within the prescribed bounds (-5/2.ltoreq.2PR<3/2), and the correct result will still be correctly computed.

The effect on latency of computation according to the present invention is minimized because the carry-propagate addition necessary to compute the estimated partial remainder is still performed only on the most significant four carry and sum bits of the redundant actual partial remainder. Table IV illustrates the truth table for the refined quotient selection logic according to one embodiment of the present invention.

Floating point units typically must support four rounding modes: round to zero, round to nearest, round toward plus infinity, and round toward minus infinity. Rounding to zero is essentially truncating the result so as to ignore less significant bits which underflow the result precision. Rounding to nearest involves rounding toward the representable result that differs from the true result by the smallest possible amount, and is the default rounding mode. Rounding toward plus infinity involves rounding the mantissa for a positive result up while rounding the mantissa for a negative result down. Rounding toward minus infinity involves rounding the mantissa for a positive result down while rounding the mantissa for a negative result up.

In certain floating-point divide operations where the expected result is exact, the divider using the quotient selection logic shown in Figure IV may err and generate an inexact flag. The inexact flag is asserted if the result is different from a result that has infinite precision. In practice, the inexact flag is set if the one or more of the guard, round, or sticky bits are set. If the rounding mode is set to direct rounding toward plus infinity or toward minus infinity, then this inexact flag will cause the division result to be rounded up for positive results or down for negative results, thereby causing an error of the smallest representable magnitude.

Table V shows the relationship between the inexact flag and the division result and describes the meaning of the inexact flag in presence of the bug. Since the inexact flag is set, the division result may be rounded incorrectly.

Table VI shows the relationship between the rounding mode and the division result.

The failing patterns causes the quotient select logic to err by failing to restore the partial remainder early enough, resulting in a negative partial remainder.

All subsequent quotient digits produced after the negative partial remainder is generated will be either 0 or -1, and the partial remainder cannot return to positive. If the final partial remainder is negative, then the inexact flag will be incorrectly set; if the final partial remainder is zero, then the inexact flag will be correct.

Exhaustive testing of single precision results failed to produce the above-described rounding error, thus the error cannot manifest itself in single precision results.

Since rounding errors occurs in the least significant bit and only in double precision operations, the relative magnitude of the error is approximately 2.sup.-52 which is approximately 2.2.times.10.sup.-6.

The potential error in the inexact flag and rounding in divide operations will not be noticeable to the majority of users for the following reasons. First, since floating-point operations are usually inexact, most applications do not examine the inexact flag or trap on inexact exceptions. Therefore, they will not be affected by erroneously asserting the inexact flag. Second, the default rounding mode is to round to nearest. Since the division error occurs when the rounding mode is set to rounding toward plus infinity or to rounding toward minus infinity, no errors will occur in the default rounding mode. Third, since the error has a magnitude equal to the least significant bit of a double precision division result, the relative error is some sixteen base ten orders of magnitude less than the result itself. This error is undetectable by the vast majority of applications, and has not been detected in any product tests of dividers using the quotient selection logic shown in Table IV.

The SRT division algorithm employed in the embodiments of the present invention are described below. ("SRT division" was named after D. Sweeney of IBM, J. E. Robertson of the University of Illinois, and T. D. Tocher of Imperial College of London, who discovered the algorithm independently.)

In radix-2 SRT Division, the following is true.

R.sub.j =partial remainder at the jth iteration

q.sub.j =quotient digit

P=dividend

D=divisor

Given two normalized operands, the following inequalities hold.

The recurrence relation for radix-2 SRT division algorithm is as follows.

This algorithm will converge to the correct solution if the remainder is restricted as follows.

The quotient digit q.sub.j+1 is a function of the partial remainder 2R.sub.j. In practice, only 4 bits from the partial remainder are needed for correct quotient digit selection. These four bits are referred to as the estimated (truncated) partial remainder, R'.sub.j.

In the above table, "NA" refers to logical outputs which are logical "don't cares" because the partial remainder is constrained so that the estimated partial remainder will never equal these values.

Table VII compares the prior art quotient digit selection truth table with that used in the embodiments according to the present invention; the difference being in the "000.00" entry, which is introduced to prevent the partial remainder from becoming negative. The truth table corresponding to the alternative embodiment is shown in the third column of Table VII. The alternative embodiment differs from the first embodiment in the quotient digit selection for the "111.10" case and is used to make the partial remainder more positive by selecting -1 instead of a 0.

Another useful way of viewing the quotient digit selection is through Robertson's Diagrams. Robertson's Diagrams show the input partial remainder 2R.sub.j along the x-axis and the resultant intermediate partial remainder R.sub.j+1 along the y-axis. The different quotient digit values for q.sub.j+1 are shown in relation to the input partial remainder 2R.sub.j such that the absolute value of the resultant partial remainder R.sub.j+1 is bounded positive and negative one.

FIG. 5 illustrates the Robertson's diagram for standard prior art quotient digit selection logic. Traces 501, 502, and 503 illustrate the partial remainder input/output relationship when the divisor D is 0.5. Because the initial partial remainder, 2R.sub.0 (the dividend), is constrained to be less than the divisor D, if the divisor is 0.5, then all subsequent resultant partial remainders are bounded by -0.5<R.sub.j+1 <+0.5, and all input partial remainders are bounded by -1.ltoreq.R.sub.j .ltoreq.+1. Thus the area of interest in FIG. 5 is bounded within the dotted rectangle 504 in FIG. 5 when the divisor is 0.5. When the divisor is 0.75, then line segments 505, 502, and 506 illustrates the input/output relationship between the partial remainders. Dotted line segments 507 and 508 illustrate the outer boundaries to the input/output partial remainder function. If the divisor were permitted to be +1, then dotted line segments 507 and 508 would be the transfer function. Since the divisor is constrained to be less than +1, the line segments 507 and 508 represent the boundaries. Thus, the dotted rectangle 509 illustrates the boundaries to the transfer function. Traces 501 and 507 represent lower and upper boundaries for the output partial remainder, and they correspond to the selection of a quotient digit of -1 for three different divisor values. Trace 502 illustrates the transfer function when a quotient digit of 0 is selected for all divisor values, because the divisor does not contribute to the output partial remainder when the input partial remainder is 0. Traces 503 and 508 illustrates the upper and lower boundaries for the transfer function when a quotient digit of +1 is selected. The diagonal dotted line 510 illustrates the upper and lower boundaries of both the output partial remainder and the input partial remainder for all values of the divisor. The dotted line 510 has a slope equal to inverse of the radix, thus the slope of dotted line 510 is 1/2. The breakpoints for the transfer functions correspond to the transitions between the different selections of quotient digits. The transition from q.sub.j+1 =-1 to q.sub.j+1 =0 occurs at 2R.sub.j =-0.5, which is noted by the dotted line segment 511. The bubble at the upper portion of the line segment 501 indicates that when the input partial remainder is exactly -0.5, then q.sub.j+1 =0 is chosen while, but for any value less than -0.5 by even a miniscule amount, the value q.sub.j+1 =-1 is chosen. The transition from q.sub.j+1 =0 to q.sub.j+1 =+1 occurs at the y-axis 512. The locations of these breakpoints 511 and 512 essentially specify the prior art quotient selection.

FIG. 6 illustrates the Robertson's diagram for a divider using the quotient selection logic according to an embodiment of the present invention. In FIG. 6, the breakpoints 601 occurs at -0.5, but the breakpoint 602 between q.sub.j+1 =0 to q.sub.j+1 =1 occurs at +0.25.

FIG. 7 illustrates the Robertson's diagram corresponding to the alternative embodiment of the present invention. The input partial remainder range mapping to -1 quotient digit selection is expanded in the alternative embodiment of the present invention so as to allow for the resulting partial remainder to tend towards positive values, thereby reducing the range where the partial remainder is not adjusted by adding the divisor. The breakpoint 701 corresponding to the transistion from a q.sub.j+1 =-1 to q.sub.j+1 =0 occurs at 2R.sub.j =-1/4, while the breakpoint corresponding to the transistion from q.sub.j+1 =0 to q.sub.j+1 =+1 occurs at 2R.sub.j =+1/4.

To find the overlap or redundancy among the partial remainders for a given divisor, the max and min ranges for the different partial remainders are found as follows. Since

is necessary to ensure convergence, we know that

max (R.sub.j+1)=D, and

min (R.sub.j+1)=-D.

Rewriting the recurrence relation,

The boundaries of the partial remainder are expressed in terms of the divisor and quotient digits.

FIG. 8 is a plot of the partial remainder vs. divisor ("P-D" plot) and is instructive in visualizing the regions of redundancy. The y-axis shows the actual input partial remainder 2R.sub.j, and the x-axis shows the divisor D. Because the divisor D is constrained by construction to normalized positive values, the only region of interest is the shaded regions between the D=1/2 boundary 801 and the D<1 boundary 802. Furthermore, because the SRT algorithm requires that the partial remainder 2R.sub.j be constrained to the range between -5/2 and 3/2, inclusive, in order to insure convergence, the quotient selection logic used need only maps quotient digits in the region between the upper boundary 2R.sub.j =3/2 illustrated by dotted line 803 and the lower boundary 2R.sub.j =-5/2 illustrated by the dotted line 804.

The quotient selection regions according to the prior art are illustrated in FIG. 8 by the line segment 805 corresponding to a quotient digit of +1, the line segment 806 corresponding to a quotient digit of 0, and the line segment 807 corresponding to a quotient digit of -1. According to an embodiment of the present invention, line segment 808 indicates the selection of +1 as the quotient digit, while line segment 809 indicates the range for selection of a zero segment, and line segment 810 indicates the range for selection of a -1 quotient digit. According to the alternative embodiment suitable for use in all rounding modes, line segment 811 indicates the selection of +1 as the quotient digit, while line segment 812 indicates the range for selection of a zero segment, and line segment 813 indicates the range for selection of a -1 quotient digit.

The regions used in one embodiment and the alternative embodiment of the present invention are very similar except in the case where the estimated partial remainder is 1/2. In one embodiment, a quotient digit of zero is used; in the alternative embodiment suitable for use in all rounding modes, a quotient digit of -1 is used.

In general SRT algorithms, the partial remainder from an exact divide operation is the negative of the divisor. In the prior art, the divisor is added to this partial remainder before a zero detect is performed to generate the sticky bit needed for performing rounding. A modification to the quotient selection logic was done in one embodiment of the present invention in the interest of saving an extra cycle and hardware. This modification to the quotient selection logic, as described above is incomplete in that it does not always produce correct double-precision division results in the modes for rounding to plus infinity and rounding to minus infinity.

According to an alternative embodiment of the present invention, the quotient selection logic should selected a -1 value for one particular estimated partial remainder instance in order for the final partial remainder to be reduced to zero. This modification to the quotient select is shown in Table VIII below.

In terms of implementation, optimized quotient selection logic cells corresponding to Table VII can be altered to generate correct results for all rounding modes by adding a single two-input NOR gate to implement the alternative embodiment.

FIG. 9 depicts a logical implementation of this enhanced quotient digit selection function. In FIG. 9, a four-bit carry propagate adder 900 takes the most significant four bits of the carry and sum portions of the redundant actual partial remainder (2PRC›msb:msb-3! and 2PRS›msb:msb-3!, respectively) as inputs and produces a non-redundant estimated partial remainder .times.›3:0!. An all-zero detector 901 determines if all four bits of the estimated partial remainder .times.›3:0! are zeros. If all bits of x›3:0! are zeros, an intermediate signal n0 is asserted. The all zero detector 901 is functionally equivalent to a four input NOR gate. An all-one detector 902 determines whether all bits of .times.›3:0! are ones. If all bits of .times.›3:0! are ones, an intermediate signal n1 is asserted. The all-one detector 902 is functionally equivalent to a four input AND gate. A NOR gate 903 takes the fifth most significant carry and sum bits of the redundant partial remainder (2PRC›msb-4! and 2PRS›msb-4!, respectively) as inputs and produces an intermediate signal i0, which is asserted only when 2PRC›msb-4! and 2PRS›msb-4! are both zeros.

A block of random logic 904 produces three signals (M, Z, and P) that each correspond to one of the possible quotient digit values according to SRT division. M corresponds to a quotient digit of minus one; Z corresponds to a quotient digit of zero; and P corresponds to a quotient digit of plus one. M is asserted if and only if either .times.›3! is asserted and n1 is not asserted or both i0 and n0 are asserted. Z is asserted if and only if either n1 is asserted and i0 is deasserted or both i0 and n0 are asserted. P is asserted if and only if x›3! is deasserted and either n0 or i0 is deasserted. Given their respective logical relationships, it is impossible for any two of the quotient digit signals M, Z, and P to be simultaneously asserted.

Although the implementation illustrated in FIG. 9 is logically correct, in practice, the four bit binary adder and subsequent logic can be merged into an integrated and optimized logic implementation. A five-stage logic implementation more efficiently utilizes area and produces the output with less delay. In such an implementation, not only M, Z, and P are computed, but also the logical inverses /M, /Z, and /P are computed as well. This is useful when the quotient signals are used to enable transmission gates such as in a CMOS multiplexor having decoded select inputs.

The number of additional gates needed to implement the new radix-2 quotient digit selection logic according to this embodiment of the present invention is relatively small. There is an implementation dependent timing trade-off between slower quotient selection logic and eliminating the partial remainder restoration cycle at the end. If the increased delay due to slightly more complicated quotient selection logic does not limit the processor cycle time, a performance gain will always be realized by saving the cycle of latency that would otherwise be required to restore negative final partial remainders. In a modern microprocessor which would include a hardware divider such as according to the present invention, it is likely that the processor cycle time will be determined by some other critical path or will be arbitrarily agreed upon prior to designing the processor. In the likely case that the increased delay of the quotient selection logic is not the longest path, a net savings of one cycle per division operation will be realized with no cost.

In an implementation of the altenative embodiment suitable for all rounding modes, the modified quotient selection logic algorithm slowed the internal floating point divider critical path by very little over the analogous optimized critical path using quotient selection rules according to the prior art. The limiting timing path on the processor was longer, so there was no negative impact on the processor cycle time from the improved quotient selection logic. The full benefit of eliminating the cycle for restoring the partial remainder prior to sticky detection was realized.

Using the quotient selection logic according to the present invention, it is still possible for negative final partial remainders to occur; however, when they occur, the quotient selection logic according to the present invention guarantees that the negative partial remainders are greater than the negative divisor. Therefore, whenever a negative final partial remainder occurs according to the present invention, Q-1 should be selected and the result is necessarily inexact, and the sticky bit can safely be asserted without fear that the negative final partial remainder equals the negative divisor.

It is possible to save hardware while also performing sign detection and sticky bit calculation in parallel. Instead of using a full 59-bit adder, a 59-bit sign detect adder can be used, slightly improving timing, but mainly saving area. Zero-detection can be done without an explicit addition to convert the redundant partial remainder into binary.

In the above equation, s.sub.i and c.sub.i are the sum and carry values of the final partial remainder. The sticky bit it then computed by:

This method generates inputs to the zero-detector with a 3-input XOR delay instead of the delay of a 59 bit carry-propagate adder, a significant net savings.

While the present invention has been disclosed with particular reference to its preferred embodiments, those embodiments are presented by way of example, not by way of limitation. Those of ordinary skill in the art would be enabled by this disclosure to add to or modify those embodiments of the present invention in various ways as needed and still be within the scope and spirit of the present invention as recited in the appended claims. For example, the optimized logical implementation of the quotient selection logic could be designed in a variety of equivalent manners. The logic described in FIG. 9 could compute both the positive and negative polarities of the quotient digits, thereby allowing the control of a CMOS transmission gate multiplexor rather than an n-channel pass transistor multiplexor. Different and yet optimized implementations would result from implementation in different technologies (for example, NMOS, ECL, etc.) or with different design rules and design tools within the same technology. The appended claims are intended to cover all circuit implementations of the truth Tables VII and VIII. Accordingly, various changes and modifications which are obvious to a person skilled in the art to which the invention pertains are deemed to lie within the spirit and scope of the invention as claimed.