Field of the Invention
This invention relates to arithmetic logic hardware of a digital computer. Specifically, it relates to bit-slice multiplication performed by arithmetic logic hardware of digital computers.
Background of the Invention
Multiplication, in binary arithmetic, is generally accomplished by successive shift and add operations. The number to be multiplied is known as the multiplicand (MD), while the number by which the MD is multiplied is known as the multiplier (MR). Since, in the binary number system, an individual binary digit (bit, in the vernacular) can assume the value of 0 or 1 only, multiplication of bits is simple and straight forward. As is the case with decimal multiplication, any number, x, multiplied by 1 is equal to itself x; any number, x, multiplied by 0 is 0.
To perform binary multiplication, the MD is multiplied by each bit in the MR. Thus, when the MR bit in question is a 0, the partial product is 0 (represented by a string of 0s as long as the number of bits in the MD). When the MR bit in question is a 1, the partial product is simply the MD. The mechanics of binary multiplication are similar to those of decimal multiplication. As a consequence, binary multiplication is accomplished in a series of multiplicative iterations, one iteration for each bit in the MR, producing a series of partial products for summation. Therefore, binary multiplication is, in its most elemental form, nothing more than the addition of partial products which are either 0 or the MD. Furthermore, each partial product is shifted to the left by the number of bits corresponding to the bit position of the MR bit in question.
A typical computer hardware implementation of a binary multiplier consists of three n-bit registers, an n-bit binary shifting mechanism, and an n-bit adder. The MD and MR are stored, for the duration of the multiplication operation, in two of the registers. The third register stores the result of each iteration of the multiplication procedure, known as the partial product. The partial product, for any given iteration of the multiplication, is equal to the sum of the partial product of the previous iteration and the product of the particular MR bit in question and MD. The sum is formed by the n-bit adder. The product of the MR bit and MD will equal zero if the MR bit equals zero and will equal the MD, shifted left by the number of places equal to the MR bit location, if the MR bit equals one.
Some computer hardware designers employ a "bit-sliced" design to implement a binary multiplier. In a typical "sliced-design," with the slice size equaling the radix of each iteration, several basic binary multipliers (slices) are cascaded. That is, the output from the adder of one multiplier is connected to the partial product register of its neighboring multiplier.
A basic problem with bit-sliced multipliers is that they often take many slice stages (cells) to implement. With emitter coupled logic (ECL) or Gallium Arsenide (GaAs) microchip technology, multipliers are generally implemented in a sliced-design because of the limited number of cells each chip can hold and the limited chip pin count.
In typical prior art bit-sliced multiplier designs, the interchip connection between slices introduce relatively long delays significantly lengthening the time needed to complete a multiplication operation. One particular cause of interchip delay stems from the fact that with each interchip connection, time is consumed in voltage magnitude translations between chips.
Summary of the Invention
It is an object of the present invention to accomplish bit-sliced multiplication in a way that avoids the need for time-consuming interchip connections at critical arithmetic paths thereby greatly improving operation speed.
The present invention eliminates undesirable interchip time delays associated with prior art bit-sliced multipliers by providing independent multiplier slices and by providing that each independent multiplier slice shall multiply each bit of multiplier data by each constituent multiplicand addend bit.
Brief Description of the Drawings
FIG. 1 depicts a prior art bit-sliced binary multiplier.
FIG. 2 depicts the present invention embodied in a four slice sample configuration.
FIG. 3 depicts a simplified version of FIG. 2, further including the apparatus for completing the multiplication.
FIG. 4 depicts a modified version of FIG. 2 showing a loop for passing bits of a multiplicand.
FIG. 5 depicts a modified version of FIG. 2 showing a centralized control of multiplicand bits for the slices.
FIG. 6 depicts a modified version of FIG. 2 showing a centralized control of multiplier bits for the slices.
Detailed Description
For clarity of explanation, the following description is directed to bit slices of one bit. The present invention can be extended to include multiple-bit slices. FIG. 1 depicts a typical prior art bit-sliced take advantage of the property that the product of two numbers is equal to the product of their constituent sums. That is, for a number X which equals A+B and a number Y which equals C+D, then X.times.Y=(A+B).times.(C+D). Further, the distributive law provides that: X.times.Y=(A.times.C)+(A.times.D)+(B.times.C)+(B.times.D). Such a sum of products may also be indicated as: AC+AD+BC+BD. Thus, in bit-sliced multiplication, the MD is parsed into constituent addends. For example, a 4-bit binary number ABCD can be represented as A000, OB00, 00C0 and 000D in the respective MD registers (9, 10, 11, 12) of the respective four slices. In operation, each slice multiplies each bit of the MR (for example, a 4-bit MR:WXYZ, contained in registers 13, 14, 15, 16) by its constituent of the MD in a binary multiplier (17, 18, 19, 20). Each slice multiplies its MD by bits Z, Y, X and W, in that order, passing on the sum of the product plus the partial product contained in registers 5, 6, 7 and 8 to the next slice in the design after each multiplicative iteration. The last slice passes its result to an accumulating circuit 21. The carry-out generates and propagates forwarded to a carry look ahead logic 25 which then generates a carry for each slice.
In order to eliminate the problems introduced by interchip signal transmission at the critical arithmetic path in the typical bit sliced multiplier, the present invention eliminates all such critical path interchip connections normally required during multiplier operation. For example, the interslice connection depicted in prior art FIG. 1 between the output of each slice's adder (1, 2 and 3) and its neighboring slice's partial product register (6, 7 and 8) is eliminated. In its place is an intraslice connection.
FIG. 2 depicts the reconfigured slices of the present invention. In the configuration of this invention, each adder 1'-4' is connected to its associated partial product register 5'-8'. Since critical arithmetic path interslice connections are eliminated, it is a further aspect of the present invention that each individual slice changes its "personality" with each iteration of the binary multiplication. The "personality" of each slice is determined by a personality counter 26, 27, 28, 29. The personality determines the contents of the MD.
In the prior art, each individual slice would multiply its constituent bit of the MD by each bit of the MR, with each multiplicative iteration varying the particular bit of the MR used in the multiplication. For example, in FIG. 1, slice I would perform four multiplicative iterations: the first being the product of its multiplicand constituent, A000, and 000Z; the second being the product of A000 and 00Y0; the third being the product of A000 and OX00; and the fourth being the product of A000 and W000. At the conclusion of each individual multiplicative iteration in the prior art, the results would be passed, by an interchip connection, to the partial product register of the succeeding slice which would also be performing multiplicative iterations concurrently on its data. The slices perform their multiplicative iterations in unison. Partial products from preceding slices would be included in each slice's iteration through each slice's adder (1-4). For example, in the first multiplicative iteration of slice II, the partial product register is cleared, i.e., contains all zeros, so the result of the iteration is simply the product of slice II's constituent of the MD, 0 B00, and 000Z. In the second multiplicative iteration, the results from slice I's first iteration appear in slice II's second iteration, added to the product of 00C0 and 00Y0.
In the prior art, each slice's constituent of the MD remained constant for the duration of the entire multiplication procedure i.e., for all required iterations. For example, in slice I, register 9 contains the datum A000, for all required iterations. Unlike the prior art, each slice of the present invention varies its constituent of the MD with each iteration. Each slice will begin the procedure with a different constituent bit of the MD. At the conclusion of each iteration, each slice will have loaded into its MD register the value of the next lower significant bit of the MD. If, during an iteration, a slice is addressing the least significant bit of the MD, its MD register shall be loaded with the most significant bit of the MD for the next iteration. Iterations continue until a slice has used each bit of the MD. Thus, in the present invention, for example, slices I', II', III', IV' will change their personality with each succeeding iteration. In the first iteration, slice I' will have, as its constituent of the MD the number A000. In its second iteration, slice I' will have, as its constituent of the MD, the number 0B00. In the third and fourth iterations, its constituent of the MD will be 00C0 and 000D, respectively. The other slices (II'-IV' ) also change their personalities on an iteration basis: slice II' cycles through the MD constituents 0B00, 00C0, 000D and A000; slice III' cycles through 00C0, 000D, A000 and 0B00; and slice IV' 219 cycles through 000D, A000, 0B00 and 00C0.
Although in the present invention the arithmetic critical path interslice connections are eliminated, there are still some interconnections between slices of a non-critical nature. Specifically, referring to FIG. 2, the carry-out bit from each slice's adder (1', 2', 3', 4') is fed through one one-bit register 21' to the carry-in port of the slice immediately to the left of the slice in question. The carry-out bit from slice I' is passed through one one-bit register 21' to the carry-in port of slice IV'.
After a slice I', II', III', IV' changes its personality from the least significant slice to the most significant slice, the present invention provides that the result of each such slice's adder 1', 2', 3', 4', is loaded into a low order partial result register 22', 23', 24', 25' and that each such slice initializes its DD/PR-Register (5', 6', 7', 8') and initializes the carry in signal. The loading and initialization of the low order partial result register 22', 23', 24', 25' and DD/PR Register 5', 6', 7', 8', respectively, is controlled by the personality counter 26, 27, 28, 29, which is counted by a control line 30, 31, 32, 33 to each of the low order partial result registers and the DD/PR register.
As a result of the individual slices cycling through all their iterations, each slice assuming the whole range of possible personalities, the answer to the multiplication can be determined by the concatenation of the DD/PR-Registers and the DD/PR data saved in the multiple word registers. Beginning with the DD/PR-Register of the most significant slice and moving right, each succeeding DD/PR-Register is concatenated in a summing register 26' until the registers from slices I', II', III' and IV' are placed side by side. The concatenation continues with the saved DD/PR data from the low order partial result registers (22', 23', 24', 25') with the most significant saved DD/PR data placed next to the least significant DD/PR-Register data already concatenated (in this case, that from slave slice IV'). Each lesser significant slice's saved DD/PR data is concatenated until all four of such DD/PR data words are in place. The result is a new large word comprised of, in order, eight pieces of data beginning with the DD/PR-Registers (5', 6', 7', 8') and ending with the saved DD/PR data from the low order partial result registers (22' , 23', 24', 25'). To the individual components of this large word are added the carry bits associated with such components: the one carry-in bit stored in registers 21' are added to the DD/PR-Register data. The result of the addition of these carry bits to the components of the concatenated large word is the answer to the multiplication problem.
FIG. 3 is a simplified version of FIG. 2 with further detail of the concatenation procedure. FIG. 3 includes the summing register 26' and depicts the interconnections between the summing register 26' and the multiple word registers (22', 23', 24', 25'), the DD/PR-Registers (5', 6', 7', 8') and the carry-out registers 21'.
FIG. 2 shows one way of varying the individual portions of the MD and MR on an iterative basis. In the circuit of FIG. 2, the control for doing this is associated with each MD and MR register in each slice. However, other techniques can also be employed.
For example, FIG. 4 shows a modified version of the circuit shown in FIG. 2, where the MD- registers are tied together in a loop (40, 41, 42, 43). Each MD-register 9', 10', 11', 12' holds the bit or bits of the MD needed by a slice for a given iteration. Once the slices I', II', III', IV' begin using the portion of the MD they need for their current iteration, each register can pass its contents to a neighboring slice's MD register for use in the next iteration, achieving the desired variations in slice personality. In this way, each individual MD- register 9', 10', 11', 12' need not maintain the entire MD; rather, the MD is shared among the slices I', II', III', IV'. Note that the MD register-to-MD register inter-chip connections 40, 41, 42, 43 are similar to the carry-out bit inter-chip connections discussed above, in terms of non-time-criticality, as a result of register-to-register data transfer.
Another example is shown in FIG. 5, where centralized control is shown. In this case, appropriate bits of the MD are supplied to each slice I', II', III', IV' for use in the next iteration by an MD Control Register 35 (which could comprise a circular shift register) by conductors 36, 37, 38, 39. The MD Control Register 35 varies the bits sent to each slice I', II', III', IV' to achieve the slice personality changes desired for each successive iteration. So, for example, slice I' receives bit A for use in its first iteration, bit B for its next iteration, bit C for the third iteration, and bit D for the last iteration. Slice II, receives, in order, bit B, bit C, bit D and bit A, and so on for slices III' and IV'. As is the case with the circuit shown in FIG. 4, this circuit can employ smaller MD registers, since they need not store the entire MD. Also, the connections 36, 37, 38, 39 between the MD Control Register 35 and each slice's MD- register 9', 10', 11', 12' are similar, in terms of non-time-criticality, to the inter-chip connections discussed in reference to FIG. 4.
Regarding the variation of the MR bits in the present invention, FIG. 6 shows how such variation may be accomplished by use of an MR Control Register 44. Over conductors 45, 46, 47 and 48, the MR Control Register 44 transfers the necessary bits to each slice's MR register 13', 14', 15', 16'. Each slice I', II', III', IV' receives the same bit or bits for use in each iteration. For example, slices I'-IV' each receive bit Z for use in the first iteration, bit Y for use in the second, bit X for use in the third, and finally bit W for use in the fourth. As is the case with the MD registers discussed in reference to FIGS. 4 and 5, the MR registers 13', 14', 15', 16' need only be large enough to hold the bit or bits used in each iteration, not the entire MR. Also, the connections 45, 46, 47, 48 between the MR Control Register 44 and each slice's MR- register 13', 14', 15', 16' have the same non-time-criticality as discussed previously, due to the use of register-to-register data transfer.