1098 IEEE TRANSACTIONS ON COMPUTERS, VOL. 48, NO.

10, OCTOBER 1999

Theory and Applications of the Double-Base

Number System
Vassil S. Dimitrov, Graham A. Jullien, Senior Member, IEEE, and
William C. Miller, Senior Member, IEEE

AbstractÐIn this paper, we analyze some of the main properties of a double base number system, using bases 2 and 3; in particular,
we emphasize the sparseness of the representation. A simple geometric interpretation allows an efficient implementation of the basic
arithmetic operations and we introduce an index calculus for logarithmic-like arithmetic with considerable hardware reductions in look-
up table size. We discuss the application of this number system in the area of digital signal processing; we illustrate the discussion with
examples of finite impulse response filtering.

Index TermsÐDouble-base number system, index calculus, digital signal processing, FIR filters.

1 INTRODUCTION

I N many applications, the computational complexity of

algorithms crucially depends upon the number of zeros
of the input data in the corresponding number system [1],
from such a noncanonic form leads to interesting problems
in transcendental number theory. The canonic number
system, however, appears to provide very fast carry-free
[2], [3], [4], [5]. Number systems are often chosen to enable a addition and is also suitable for multiplication. We illustrate
reduction of the complexity of the arithmetic operations; the our ideas with applications to computing finite impulse
most popular are, perhaps, signed-digit number systems response filter inner products.
[5]. An analysis of the expected number of zeros in the
representation of arbitrary integers in the binary signed-
digit number system shows that on average, for long 2 PROPERTIES OF THE DBNS
wordlengths, 33 percent fewer adders are needed to Definition 1. The representation of a given integer x into the
perform multiplication than binary [6], [7]. In these number form:
systems, we need, on average, O…log N† nonzero digits to X
represent the integer N. xˆ di;j 2i 3j ; di;j 2 f0; 1g …2†
A number system using bases 2 and 3, allowing as digits i;j
only 0, 1, and requiring o…log N† nonzero digits is the
will be referred to as a double-base number system
double base number system (DBNS), i.e., a representation
(DBNS).
having the form of (1)
X
xˆ di;j 2i 3j : …1† The representation of a given integer as a sum of
i;j minimal number 2-integers (numbers of the form 2i 2j ,
smaller than or equal to x) will be referred to as the canonic
Clearly, the binary number system is a special case (and
double-base number representation (CDBNR). The main feature
valid member) of the above representation. In this paper,
of the CDBNR is the unusual sparsity of the representation.
we will deal with canonic (minimal number of nonzero
It is easy to check, for example, that 23 is the smallest
digits) and near canonic forms for the representation.
The DBNS has an unusually simple 2D geometric integer requiring three 2-integers. The smallest integer
interpretation, suitable for implementation via cellular requiring four 2-integers is 431, five 2-integers are needed to
automata [8], for example, and, in this paper, we introduce represent 18,431 and six 2-integers are needed to represent
an index calculus with which we can perform arithmetic 3,448,733. Up to this limit we can represent every integer as
using logarithmic-like computational units. a sum of at most five 2-integers. The procedure to find a
Arithmetic operations in this number system do not CDBNR of a given very large integer seems to be a very
guarantee that the results are obtained in the minimal, or complex task; however, in [9], we have proposed a greedy
canonic form, and the associated problem of conversion algorithm with the input as a positive integer x; and an
P
output of 2-integers, ai , such that i ai ˆ x. The algorithm
finds the largest 2-integer, w, smaller than or equal to x, and
. V.S. Dimitrov is with Helsinki University of Technology, PO Box
3000,FIN-02015 HUT. E-mail: vdimitro@wooster.hut.fi. recursively applies the same for x ÿ w until reaching zero.
. G.A. Jullien and W.C. Miller are with the VLSI Research Group, In [9], we proved the following theorem:
University of Windsor, Windsor, Ontario, Canada N9B 3P4.
E-mail: jullien@engn.uwindsor.ca. Theorem 1. The greedy algorithm terminates after k ˆ O…logloglogx x†
For information on obtaining reprints of this article, please send e-mail to: steps.
tc@computer.org, and reference IEEECS Log Number 109856.

0018-9340/99/$10.00 ß 1999 IEEE

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.
DIMITROV ET AL.: THEORY AND APPLICATIONS OF THE DOUBLE-BASE NUMBER SYSTEM 1099

TABLE 1
Column Reduction

Fig. 1. Occurrence of 2-integers for 215-bit numbers.

We also conjecture that the coefficient associated with the

complexity order is close to unity (this is based on
performing many numerical experiments).
Definition 2. We call the representation obtained by the greedy
algorithm a Near-Canonic DBNR (NCDBNR).

To confirm the utility of the greedy algorithm, we Table 2 demonstrates the application of the identity,
generated NCDBNRs for 1,000 randomly chosen 215-bit 2i 3j ‡ 2i 3j‡1 ˆ 2i‡2 3j , to remove consecutive active cells
lying in one row. This procedure is akin to the symbolic
integers. We predict (from Theorem 1) that the expected
substitution process used in optical computing.
number of nonzero digits is 27.75; the occurrence of the
number of 2-integers peaks at about 30, as shown in Fig. 1, 3.2 Generalized Reduction
which successfully demonstrates the efficacy of our algo- We can generalize the reduction problem using the purely
exponential Diophantine equation, (3), where l < k.
rithm. A variety of computational experiments shows that
the largest 2-integer, smaller than x, occurs in at least one of 2i1 3j1 ‡ 2i2 3j2 ‡ . . . ‡ 2ik 3jk ˆ 2m1 3n1 ‡ 2m2 3n2 ‡ . . . 2ml 3n1 :
the CDBNRs of x in about 80 percent of the cases. This …3†
observation, along with Theorem 1, allows an estimate that The problem of solving such Diophantine equations has
the greedy algorithm returns a CDBNR with probability been a subject of investigation over the last two decades
log x
0:8log log x ; fortunately, this tends to zero very slowly. [10], although some interesting results were obtained in the
We also find that the greedy algorithm produces a 30s and 40s [11].
representation requiring on average O…logloglogx x† 2-integers. We need only consider some special cases for k and l. For
This NCDBNR provides a sufficiently sparse representation example, using k ˆ 2, l ˆ 1, we can prove the following
of x to make it very useful. theorem:
Theorem 2. The Diophantine equation x ‡ y ˆ z, where
3 ARITHMETIC READY TRANSFORMATIONS GCD…x; y; z† ˆ 1 and x, y, and z are 6-integers (that is x,
The mechanism of finding the NCDBNR plays a crucial role y, z have the form 2x1 3x2 5x3 7x4 11x5 13x6 , with x  0,
in performing basic arithmetic operations. Along with i ˆ 1; 2; 3; 4; 5; 6) has exactly 545 solutions.
sparseness, we also require that nonzero digits be non- Proof. See [12]. u
t
consecutive in our mapping representation; this allows
addition to be mapped to a simple boolean operation. Based In our case, x3 ˆ x4 ˆ x5 ˆ x6 ˆ 0 and the only solutions
on this requirement, we provide the following definition: of x ‡ y ˆ z are …1; 2; 3†, …1; 3; 4†, and …1; 8; 9†. Therefore,
Definition 3. A DBNR that has no consecutive nonzero digits is these represent the only three cases where we can replace
defined as an addition ready DBNR (ARDBNR). two active cells with one. For k ˆ 3, l ˆ 1, an interesting
possibility for reducing the active cells follows from the
solution of the Pillai equation [13]:
3.1 ARDBNR Reduction Rules
We can use a geometrical interpretation, with orthogonal 2a  3b ˆ 3c  3d : …4†
dimensions for each of the bases, to represent numbers in
Pillai was able to solve all of the above four equations,
the DBNR. Nonzero DBNR digits are shown as black
excluding the equation 2a ÿ 2b ˆ 3c ÿ 3d on which he
squares (active cells). This interpretation allows us to
conjectured that the only solutions are …3; 1; 2; 1†,
demonstrate simple identities on special combinations of
active cells that provide a transformation of a DBNR to an …5; 3; 3; 1†, and …8; 4; 5; 1†. The conjecture was proven by
ARDBNR. Stroeker and Tijdeman [22]. For our purposes, only (5) and
For example, Table 1 shows the representation of the (6) are relevant:
identity 2i 3j ‡ 2i‡1 3j ˆ 2i 3j‡1 to remove consecutive active
2a ÿ 2b ˆ 3c ‡ 3d …5†
cells lying in one column.

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.
1100 IEEE TRANSACTIONS ON COMPUTERS, VOL. 48, NO. 10, OCTOBER 1999

TABLE 2 Definition 4. Ix …i; j† is the DBNS map of the integer x,

Row Reduction represented in the ARDBNR.

The image Iz …i; j† of the DBNS map of the number z ˆ

x ‡ y can be obtained using:

Iz …i ‡ 1; j† ˆ Ix …i; j† AND Iy …i; j† Rule …1†

Iz …i; j† ˆ Ix …i; j† XOR Iy …i; j† Rule …2†

Note, using the ARDBNR, if Ix …i; j† ˆ Iy …i; j† ˆ 1, then
Ix …i ‡ 1; j† ˆ Iy …i ‡ 1; j† ˆ 0 and, therefore, addition can be
accomplished using a symbolic substitution technique. To
reduce this result, it is sufficient to use the following rules
(see Tables 1 and 2.)

Iz …i; j ‡ 1† ˆ Iz …i; j† AND Iz …i ‡ 1; j† Rule …3†

Iz …i ‡ 2; j† ˆ Iz …i; j† AND Iz …i; j ‡ 1† Rule …4†

2a ‡ 2b ˆ 3c ÿ 3d : …6†
In the worst case, we have to perform a reduction (ªcarryº
Following [22], the solutions of (5) are …2; 1; 0; 0†, removal) for each row in the 2D representation. For an
…3; 1; 1; 1†, …5; 1; 1; 3†, …3; 2; 1; 0†, …5; 2; 0; 3†, …4; 2; 1; 2†, and equivalent n-bit binary number, there are n=…log2 3† rows in
…8; 2; 5; 3†, while (6) has only one solution …0; 0; 1; 0†.
the representation, which leads to the interesting result that
The total number of solutions of the equation
DBNS addition requires 37 percent fewer ªcarriesº than
x ‡ y ‡ z ˆ t, GCD…x; y; z; t† ˆ 1, in 2-integers is 27 and
binary addition.
they are tabulated in Table 3.
As an example of performing addition, let us consider
A powerful technique in transcendental number theory
the addition of the numbers 88 and 123 using the proposed
(studies of equations such as (3)), described by Baker [14],
technique. The representation of the numbers are
allows one to conclude that (3) has only a finite number of
NCDBNR, obtained via the greedy algorithm. For the
solutions. Existing methods for bounding the upper limits,
number, 88, the NCDBNR gives 88 = 81 + 6 + 1 (three
however, give very large upper bounds; therefore, in
active cells), while the CDBNR consists of only two active
searching for a CDBNR, we are forced to use methods that
cells (88 = 72 + 16 = 64 +24). The addition operation is
do not guarantee exact minimality but, rather, provide both
presented in Table 4.
near minimality and algorithmic realization.
The example is selected so that the application of the
A very intriguing special case of the number representa-
reduction rules, based on the solution of (3), gives a result
tion proposed in this paper has been extensively studied in
which is not optimal. In fact, 211 requires 3 ones in the
[23], [24], [25], [26].1 This is the so-called 3-smooth
representation of integers, where the expansion satisfies CDBNR; one of them could be found using more
the extra hypothesis that no 2i 3j divides a term 2k 3l . It is sophisticated substitution rules:
possible to demonstrate that the 3-smooth representation of Iz …i ‡ 3; j† AND Iz …i ‡ 1; j† ˆ Iz …i; j† AND Iz …i; j ‡ 1†
integers is not necessarily minimal, although our computa-
AND Iz …i ‡ 1; j ‡ 1†
tional experiments indicate that it provides a representation
which is even more sparse than the one obtained by the Rule …5†
greedy algorithm. The exact comparison between the two
algorithms is beyond the scope of this paper. Iz …i ‡ 5; j† ˆ Iz …i ‡ 1; j† AND Iz …i; j ‡ 1†
Rule …6†
AND Iz …i ‡ 1; j ‡ 3†
4 DBNS-MAP ADDITION AND MULTIPLICATION following from the solutions of the Pillai equation [13], (4).
4.1 Addition The final map is given in Table 5.
Let x and y be two integers in the CDBNR. We note that if x
4.2 Multiplication
and y contain the element 2i 3j , then the element 2i‡1 3j does
not exist, therefore, addition can be computed by simply Let x and y be integers, represented by DBNS maps in the
overlaying the corresponding DBNS maps; there will be no CDBNR. The CDBNR of their product, z, is an n-tuple of the
overlapping active cells. In order to prepare for another elements f2iz 3jz ˆ 2ix ‡iy 3jx ‡jy g, where the fix ; jx g and fiy ; jy g
addition, we ideally perform a reduction into minimal form. are the 2-integer index locations of the active cells in the
In practice, however, we only require an ARDBNR and the CDBNRs of x and y, respectively.
symbolic substitution methods described in the previous It is clear that the multiplication process simply
section can be effectively used. corresponds to 2D shifts and DBNS additions in an
equivalent way to that performed using binary arithmetic.
1. We are grateful to one of the referees for pointing out this work. The promise here, however, is that the number of

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.
DIMITROV ET AL.: THEORY AND APPLICATIONS OF THE DOUBLE-BASE NUMBER SYSTEM 1101

TABLE 3
The Solutions of the Equation x ‡ y ‡ z ˆ t in 2-Integers

operations is considerably reduced based on the sparseness Iz …i; j ‡ 3† ˆ Iz …i ‡ 1; j† AND Iz …i; j ‡ 2† AND Iz …i ‡ 4; j†
Rule …8†
of the representation.
Let us consider the multiplication of the numbers 79 and
The representation of the multiplication result is shown
107, represented via their DBNS maps as shown in Table 6. at the left side of Table 7 and the ARDBNR reduction, using
The final reduced forms of the product can be found by Rule (7) and Rule (8), at the right side of Table 7.
Efficient implementation of the NCDBNR (or ARDBNR)
using two specific solutions of the Pillai equation (4). reduction procedure is an ongoing project within our
Iz …i; j ‡ 4† ˆ Iz …i; j† AND Iz …i ‡ 4; j† AND Iz …i ‡ 6; j† group. An interesting implementation using analog Cellular
Neural Networks has recently been presented [15] using
Rule …7†
only the column and row reduction rules of Table 1 and
Table 2 to obtain ARDBNR forms.

TABLE 4
Example of the DBNS Addition Process 5 DBNS INDEX CALCULUS
The n-tuple, f2iz 3jz ˆ 2ix ‡iy 3jx ‡jy g, introduced in the pre-
vious section, immediately leads to an implementation of
multiplication using index addition, where the index
mapping of x, for example, is simply the n-tuple fix ; jx g.
For cases where we are approximating the reals by fixed
point numbers, it is possible to find a single index CDBNR
for any real number with arbitrary precision by using
diophantine approximation algorithms [12], [19]. This leads
us to a multiplication technique only involving a single 2D
shift, which corresponds to a pair of index additions. If only
one of the numbers to be multiplied is in the single index
form, multiplication will only require O…logloglogx x† addition
pairs. The following theorem proves the single index
mapping property.
Theorem 3. Let n and m be integers. The set An;m ˆ f2n 3m g is
dense over the nonnegative reals; that is, in every interval
‰1 ; 2 Š…1  0; 2 > 1 † at least one number of the form 2n 3m
appears.
Proof. The proof follows from the well-known fact that the
set of numbers: fn ‡ m
!g, n; m 2 Z, !-irrational is
dense over the reals. In our case, we have ! ˆ log2 3. t
u

Based on our geometrical interpretation, if we extend the

DBNS map in both directions, then every nonnegative real
number can be approximated with arbitrary small error
using only one active cell. We can show that a single-cell
coefficient can be represented by two …log2 n†=2-bit num-
bers, so we have not incurred any dynamic range
redundancy compared to the binary representation of the
number, even though our DBNS representation is inher-
ently very redundant (and sparse). The advantage, of

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.
1102 IEEE TRANSACTIONS ON COMPUTERS, VOL. 48, NO. 10, OCTOBER 1999

TABLE 5 TABLE 7
Pillai Solution Final Map The Reduction Rules Applied Here Are Rule (7) and Rule (8)

2a 3b ÿ 2c 3d ˆ 2a 3b …1 ÿ 2cÿa 3dÿb †
course, is the ability to use index calculus on the …10†
 2a 3b …c ÿ a; d ÿ b†:
representation.
We will, of course, precompute and store the functions
5.1 Implementing the Index Calculus containing the approximation of:
We represent a number, x, as a triple …sx ; bx ; tx †, where sx is
…x; y† ˆ 1 ‡ 2x 3y  2 3 …11†
the sign bit, and bx and tx are integers such that sx 2bx 2tx is an
acceptable approximation to x. More precisely, if " is the
…x; y† ˆ 1 ÿ 2x 3y  2 3 : …12†
error allowed, then jx ÿ sx 2bx 3tx j < ".
We have a low complexity implementation of multi- Addition (subtraction) of two numbers is now mapped
plication and division, namely, if: x ˆ …sx ; bx ; tx † and into the following two steps:
y ˆ …sy ; by ; ty †, then: 1. Find the corresponding element …; † in the table;
2. Add (subtract) …a; b† with …; †.
x
y ˆ ……sx ‡ sy † mod 2; bx ‡ by ; tx ‡ ty † …7†
5.2 Comparison with the Log Number System
x=y ˆ ……sx ‡ sy † mod 2; bx ÿ by ; tx ÿ ty †: …8† It is clear that this index calculus shares some similarities
with the Logarithmic Number System (LNS) [16]. Both
The implementation of addition and subtraction within
allow the mapping of multiplication and division to
this index calculus can be performed using the identities: addition and subtraction, uses an identity requiring the
2a 3b ‡ 2c 3d ˆ 2a 3b …1 ‡ 2cÿa 3dÿb † look-up of a unary function (e.g., (11) is similar to the LNS
…9† unary function log2 …1 ‡ 2…ÿ† † where the input to the table
 2a 3b …c ÿ a; d ÿ b†
is … ÿ †, the logarithms of the numbers being added).
There is a fundamental difference in the mapping,
however, since the DBNS involves indices over two
TABLE 6 orthogonal dimensions, where the logarithmic number
Example of the DBNS Multiplication Process
system involves only one ªindex.º In the DBNS, if the error
of the computations is fixed to be ", then, in approximating
some real number, x, we can expect the size of the integers
used to be smaller than the corresponding fixed-point
number in the LNS [17]. Therefore, the multiplications
(divisions) can be performed by two parallel additions
(subtractions) of two small numbers.
A disadvantage of the DBNS is that there is no direct
way to quickly compare two numbers of the form 2a 3b and
2c 3d . The simplest way seems to be to compare the numbers
a ‡ b
log2 3 and c ‡ d
log2 3, which requires one fixed-
number multiplication, one addition, and one final subtrac-
tion to obtain the result.
To make a fair comparison between both number
systems, we have to evaluate the size of the integers used

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.
DIMITROV ET AL.: THEORY AND APPLICATIONS OF THE DOUBLE-BASE NUMBER SYSTEM 1103

TABLE 8 tion of a real number, x. Let us assume, that x is represented

Comparison between the LNS and the DBNS with precision " ˆ 2ÿk , then the above theorem tells us that
there exist integers p and q requiring at most …k=2† ‡ 1 bits
and ensuring the approximation with desired accuracy. The
theorem also guarantees that the size of the look-up tables
will be the same as the size of the look-up tables in the LNS.
Table 8 summarizes the comparison analysis.

6 AN FIR FILTER EXAMPLE

An FIR filter implements the linear convolution:
X
yi ˆ xk
hiÿk : …13†
k

A well-established technique to reduce the complexity of

in representing a given real number, x, and the size of the (13) is to select filter coefficients with a small number of
look-up tables used. To do this, we use Theorem 4: nonzero binary digits. Canonic-signed-digit (CSD) repre-
Theorem 4. Given ;  reals and " > 0, then there exist integers sentations allow a greater choice of coefficients with no
p and q, such that: increase of the number of nonzero digits. Such an approach,
however, often leads to increases in the filter length to allow
jq ÿ p ÿ j < " a desired spectral envelope to be matched. The DBNS
representation typically allows single digit approximations
and
with much greater coefficient space support than the CSD
p  ÿ1
jp; qj  4
8" 2 : approach. There is also the advantage of reduced multi-
plication complexity, as discussed in the previous section.
Proof. See reference [18]. u
t We consider the following low-pass filter example to
demonstrate the efficacy of the DBNS index mapping:
There are many ways to easily find the corresponding p
1. passband and stopband edge frequencies of 0.021
and q using one of the large variety of approximation and 0.07 radians.
algorithms. Examples include the continued fraction algo- 2. passband ripple and stopband attenuation require-
rithm [19], LLL reduction [20], Ferguson-Forcade algorithm ments of 0.2 dB and 60 dB, respectively.
[21], Szekeres algorithm [22], etc. For our purposes, the The infinite precision 60-tap filter has a passband ripple
above theorem is relevant because it allows us to estimate of 0.2 dB and stopband attenuation 61.7 dB; the coefficients,
the size of the integers …bx ; tx † taking part in the approxima- along with a 10-bit DBNS mapping, are given in Table 9.

TABLE 9
DBNS10-Bit Pairs for Filter Example

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.
1104 IEEE TRANSACTIONS ON COMPUTERS, VOL. 48, NO. 10, OCTOBER 1999

Fig. 3. Screen dump of the DBNS 5-tap filter simulation.

This is the same restriction as for LNS implementations, and

still yields many practical DSP applications.
A very interesting feature of the IPSP in Fig. 2 is that the
mantissa is limited in the amount of right or left shift it will
undergo in the barrel shifter. This is a priori knowledge
since it depends on the precision of the original mapping.
Clearly, the inputs to the Exponent Sum adder, in general,
Fig. 2. Index calculus IPSP. will exceed the shift limit, but the sum output will always
be within this limit. We can therefore save on adder
The resulting stopband attenuation is 60.1 dB, still within hardware by only constructing modulo 2S adders for both
specifications. the Binary and Exponent Sum adders, where S is the
maximum left or right shift. The exponent part of the ROM
6.1 An Index Calculus IPSP output can also be limited to an S-bit (modulo 2S ) word.
A major building block for DSP processors is the inner
product step processor (IPSP). The basic structure of the 6.2 A DBNS FIR Filter Simulator
index calculus IPSP is shown in Fig. 2. As with LNS We have built a simulation of the IPSP in order to verify its
implementations, the sign of the data and coefficient inputs operation, particularly in regard to the word reduction of
and the zero data case are handled separately; we have not the ROM and Binary and Exponent Sum adders. The
included this logic block in order to keep Fig. 2 as simple as simulation tests the IPSP in a multitap systolic FIR filter
possible.
The function implemented is Si ÿ Siÿ1 ‡ …xi  yi †, where TABLE 10
xi ) i and yi ) i . Each of the mappings produces a First 10 Entries of the Data Conversion ROM
binary and ternary exponent. These exponents are added to
produce the sum exponents. Rather than map back to the
DBNS, we convert to a floating point-type representation
for the accumulation. Noting that the binary exponents
simply represent shifts, we only have to look up the
exponent and mantissa for the ternary components. There
are four binary adders, a …b=2 ‡ 1†  b0 -bit ROM, and a
barrel shifter. The word length of the ROM (b0 -bits) is the
concatenation of the mantissa and exponent for the
‰3ŠŠ ‰mŠ ‰eŠ
approximation 3i  yi  2yi . Note that the DBNS sub-
stitutes the LNS b-bit input ROM with the much smaller
address size ROM and a barrel shifter. This may represent a
considerable hardware reduction.
The input conversion can be performed with a single
table, providing the input data word-width is not too large.

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.
DIMITROV ET AL.: THEORY AND APPLICATIONS OF THE DOUBLE-BASE NUMBER SYSTEM 1105

the number of bits used in the Binary and Exponent Sum

adders of Fig. 2. The inputs on the left are the index and
sign data for the input data {DT, DB, DL} and the coefficient
{CT, CB, CL} together with the binary accumulating sum.
The leftmost IPSP receives a zero input for this sum. In this
example, all of the coefficients are set to the same value.
Latches in the IPSP allow the data to ªslideº over the
coefficient in this systolic implementation.
A screen dump of the essential elements of the IPSP is
shown in Fig. 4. For simplicity, the entire data and
coefficient words are input to the sign/zero logic detection
blocks and the corrected data is output to the final adder
from the Fix block.
The Binary and Exponent Sum adders are set to 5-bits
(mod 32 adders) which provides a maximum shift of 16
places in the barrel shifter; this exceeds the actual
requirement of 10 places. In this particular snap-shot,
the coefficient is 5, which is approximated as
2ÿ69 345 ˆ 5:0048. N o t e t h a t CB ˆ ÿ69…mod 32† ˆ ÿ5;
CT ˆ 45. The data is 44, with an optimal 10-bit mapping
of DB ˆ ÿ199…mod 32† ˆ ÿ7, DT ˆ 129. The result of the
multiplication operation is 5  44 ˆ 220, which is correctly
produced at the input to the Binary Accumulator. The ROM
looks up the prestored value 3129  3; 532  2264 and the
mantissa is loaded into the barrel shifter. The binary
exponent output of the ROM is only required to 5-bits.

264…mod 32† ˆ 8:
The shift input to the barrel shifter is ÿ4 and this value is
correctly obtained provided that it is contained within the
number of bits of the reduced binary adders.

7 CONCLUSIONS
In this paper, we have presented the theory of a double-
base number system as a technique for representing
numbers that allows potentially low complexity arithmetic
operations using a variety of implementation media.
We have explored an implementation procedure for
arithmetic operations using symbolic substitution. For
addition in the DBNS, we simply have to ensure that there
no consecutive active cells lying in any column on the
DBNS-map. For multiplication, using 2D shifts and addi-
tions, however, we seek representations with minimal
active cells located large 2D Hamming distances from each
other. Symbolic substitution provides the ability to seek
such representation, and our approach (presented as a set of
rules for the substitutions) produces near canonic forms for
Fig. 4. Screen dump of the IPSP simulator. the output representation.
An index calculus, which provides the same structure as
configuration. A screen dump of the essential elements of the Logarithmic Number System (LNS), is shown to
the simulator for a 5-tap filter test is shown in Fig. 3. considerably reduce the hardware required. An implemen-
The Data Conversion ROM tables were generated from a tation advantage over the LNS is that the index additions
custom integer programming package that selects binary and subtractions are reduced in complexity because the
and ternary exponents, a and b, based on minimizing the binary and ternary operations are completely independent.
error " ˆ jx ÿ 2a 3b j. The first 10 entries of this table for The use of this index calculus has been demonstrated using
an example of systolic FIR filtering.
integer values of x, with 10-bits of precision, is shown in
Table 10. These are the full values, not the modulo reduced
values. Note that only one data conversion ROM is required
for the entire IPSP filter chain. The top input to each IPSP is

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.
1106 IEEE TRANSACTIONS ON COMPUTERS, VOL. 48, NO. 10, OCTOBER 1999

ACKNOWLEDGMENTS [26] R. Blecksmith, M. McCallum, and J.L. Selfridge, ª3-Smooth

Representation of Integers,º Am. Math. Monthly, vol. 105, pp. 529-
The authors would like to acknowledge financial assistance 543, 1998.
from the Natural Sciences and Engineering Research
Council of Canada, the Micronet Network of Centres of
Excellence, and Gennum Corporation. Workstations and Vassil S. Dimitrov received the MSc degree in
software were provided by the Canadian Microelectronics computer science from the Technical University
of Sofia, Bulgaria, and the PhD in mathematics
Corporation. We also acknowledge the many valuable from the Mathematical Institute of the Bulgarian
comments, including additional references, from the anon- Academy of Sciences in 1995. From 1995 until
ymous reviewers. 1997, Dr. Dimitrov was a postdoctoral fellow in
the VLSI Research Group, University of Wind-
sor, Canada. From January 1998 until February
REFERENCES 1999, Dr. Dimitrov held the position of research
scientist at Reliable Software Technologies
[1] A. Borodin and P. Towari, ªOn the Decidability of Sparse Corporation in Sterling, Virginia, where he was involved with DARPA
Univariate Polynomial Interpolation,º Computational Complexity, sponsored projects on cryptoanalysis. Since March 1999, Dr. Dimitrov
vol. 1, pp. 67-90, 1991. has been with the Laboratory of Signal Processing and Computer
[2] P. Montgomery, ªA Survey of Modern Integer Factorization Technology at Helsinki University of Technology, Finland, where he
Algorithms,º CWI Quarterly, vol. 7, no. 4, pp. 337-366, 1994. teaches courses in cryptography, information security, and digital signal
[3] M.D. Ercegovac, T. Lang, J.G. Nash, and L.P. Chow, ªAn Area- processing. His research interests include number theoretic algorithms,
Time Efficient Binary Divider,º IEEE Int'l Conf. Computer Design, computer arithmetic, cryptography, fast algorithms for digital signal
pp. 645-648, Rye Brook, N.Y., Oct. 1987. processing, and related topics.
[4] P. Kornerup, ªComputer Arithmetic: Exploiting Redundancy in Dr. Dimitrov is a member of the New York Academy of Sciences.
Number Representations,º Proc. ASAP '95, Strasbourg, France,
1995. Graham A. Jullien was educated in the United
[5] A. Avizienis, ªSigned-Digit Number Representation for Fast Kingdom, receiving degrees in electrical engi-
Parallel Arithmetic,º IRE Trans. Electronic Computers, vol. 10, neering from the Universities of Loughborough,
pp. 389-400, 1961. Birmingham, and Aston (PhD, 1969). He was a
[6] H. Garner, ªNumber Systems and Arithmetic,º Advances in student engineer and data processing engineer
Computing, vol. 6, pp. 131-194, 1965. at English Electric Computers, United Kingdom,
[7] G.W. Reitwiesner, ªBinary Arithmetic,º Advances in Computing, from 1961 to 1966, and a visiting senior research
vol. 1, pp. 231-308, 1960. engineer at the Central Research Laboratories
[8] E. Swartzlander, ªDigital Optical Computing,º Applied Optics, of EMI Ltd., United Kingdom, from 1975 to 1976.
vol. 25, pp. 3,021-3,032, 1986. Since 1969, he has been with the Electrical
[9] V.S. Dimitrov, G.A. Jullien, and W.C. Miller, ªAn Algorithm for Engineering Department of the University of Windsor, Ontario, Canada,
Modular Exponentiation,º Information Processing Letters, vol. 66, and currently holds the rank of University Professor. He is also the
no. 3, pp. 155-159, 1998. director of the VLSI Research Group at the University of Windsor. He
[10] T.N. Shorey and R. Tijdeman, Exponential Diophantine Equations. was a member of the Board of Directors of the Canadian Microelec-
Cambridge Univ. Press, 1986. tronics Corporation from 1990 to 1993 and is a principle researcher and
[11] G. Hardy, Ramanujan. Cambridge Univ. Press, 1940. member of the Coordinating Committee and Board of Directors of the
[12] B.M.M. de-Weger, ªAlgorithms for Diophantine Equations,º CWI Micronet Network of Centres of Excellence.
Tracts-Amsterdam, vol. 65, 1989. He has published widely in the fields of computer arithmetic, digital
[13] S.S. Pillai, ªOn the Equation 2a ÿ 2b ˆ 3c ÿ 3d ,º Bulletin of the signal processing, and VLSI systems, and teaches courses in related
Calcutta Math. Soc., vol. 37, pp. 15-20, 1945. areas. He has served on the technical committees of many international
[14] A. Baker, ªThe Theory of Linear Forms in Logarithms,º conferences; he serves on the editorial board of the Journal of VLSI
Transcendental TheoryÐAdvances and Applications, A. Baker, ed., Signal Processing, and was an associate editor of the IEEE Transac-
pp. 1-27, Academic Press, 1987. tions on Computers from 1994 to 1996. He hosted and was program
[15] S. Sadeghi-Emamchaie, G.A. Jullien, V. Dimitrov, and W.C. Miller, cochair of the 11th IEEE Symposium on Computer Arithmetic; he was
ªDigital Arithmetic Using Analog Arrays,º Proc. Eighth Great Lakes also the program chair for the Eighth Great Lakes Symposium on VLSI,
Symp. VLSI, pp. 202-207, Lafayette, La., Feb. 1998. and is the program chair for the 1999 Asilomar Conference on Signals,
[16] E.E. Swartzlander, D.V. Chandra, H.T. Nagel, and S.A. Starks, Systems, and Computers. He is a senior member of the IEEE.
ªSign/Logarithmic for FFT Implementation,º IEEE Trans. Compu-
ters, vol. 32, pp. 526-534, 1983.
William C. Miller received the BSE degree in
[17] D. Lewis, ªAn Accurate LNS Arithmetic Unit Using Interleaved
electrical engineering from the University of
Memory Function Interpolator,º Proc. ARITH-11, pp. 2-9, Wind-
Michigan, Ann Arbor, in 1960, and the MASc
sor, Ontario, Canada, 1993.
and PhD degrees in electrical engineering from
[18] V.S. Dimitrov and T.V. Cooklev, ªTwo Algorithms for Modular
the University of Waterloo, Ontario in 1961 and
Exponentiation Using Nonstandard Arithmetic,º IEICE Trans.
1969, respectively. He joined the Department of
Fundamentals, pp. 82-87, 1995.
Electrical Engineering at the University of Wind-
[19] A.J. Brentjes, ªMulti-Dimensional Continued Fraction Algo-
sor in 1968, where he currently holds the rank of
rithms,º Mathematical Centre Tracts, Amsterdam, vol. 145, 1981.
professor.
[20] C.-Y. Chen, C.-C. Chang, and W.-P. Yang, ªHybrid Method for
His research interests are oriented toward
Modular Exponentiation with Precomputations,º IEE Electronics
digital signal processing and the design of massively parallel VLSI
Letters, vol. 32, no. 6, pp. 540-541, 1996.
processor architectures for application specific problems in the area of
[21] H.R.P. Ferguson and R.W. Forcade, ªGeneralization of the
image processing relating to machine vision. He also teaches courses in
Euclidean Algorithm for Real Numbers for All Dimensions
the area of circuit theory, signal processing and system theory. Dr. Miller
Higher than Two,º Bulletin Am. Math. Soc., vol. 1, pp. 912-914,
is also engaged extensively in industrial consulting work.
1979.
He was director of the CAD/CAM Centre at the University of Windsor
[22] R.J. Stroeker and R. Tijdeman, ªDiophantine Equations,º Compu-
for a two-year period ending in 1988, and currently serves on the Board
tational Methods in Number Theory, H. Lenstra and R. Tijdeman,
of Directors and Technical Advisory Committee of the Canadian
eds., Math. Centre Tracts-Amsterdam, vol. 155, pp. 321-369, 1987.
Microelectronics Corporation. Dr. Miller is a registered Professional
[23] P. Erdos, ªQuickies,º Math. Magazine, vol. 66, p. 67, 1994.
Engineer in the Province of Ontario and a senior member of the IEEE.
[24] P. Erdos and M. Lewin, ªd-Complete Sequences of Integers,º
Math. of Computation, vol. 65, pp. 837-840, 1996.
[25] G. Dombi and B. Valko, ªOn a Problem of Erdos,º Acta
Mathematica Hungarica, vol. 77, pp. 47-56, 1997.

Authorized licensed use limited to: the Leddy Library at the University of Windsor. Downloaded on March 22,2023 at 20:08:30 UTC from IEEE Xplore. Restrictions apply.