Subquad multi ff

A Generalized Method for Constructing Sub-quadratic Complexity
GF(2k
) Multipliers ∗
Draft Copy
B. Sunar
Atwater Kent Laboratory
100 Institute Road, Worcester, MA 01609
Worcester Polytechnic Institute
Worcester, MA 01609
January 2004
Abstract
We introduce a generalized method for constructing sub-quadratic complexity multipliers for even
characteristic field extensions. The construction is obtained by recursively extending short convolution
algorithms and nesting them. To obtain the short convolution algorithms the Winograd short convo-
lution algorithm is reintroduced and analyzed in the context of polynomial multiplication. We present
a recursive construction technique that extends any d point multiplier into an n = dk
point multiplier
with area that is sub-quadratic and delay that is logarithmic in the bit-length n. We present a thorough
analysis that establishes the exact space and time complexities of these multipliers. Using the recursive
construction method we obtain six new constructions, among which one turns out to be identical to the
Karatsuba multiplier. All six algorithms have sub-quadratic space complexities and two of the algorithms
have significantly better time complexities than the Karatsuba algorithm.
Keywords: Bit-parallel multipliers, finite fields, Winograd convolution
1 Introduction
Finite fields have numerous applications in digital signal processing [1, 2], coding theory [3, 4, 5] and
cryptography [6, 7, 8]. The efficient implementation of finite field operations, i.e. addition, multiplication,
and inversion, is therefore critical for many applications. The efficiency of field operations is intimately
related to the particular representation of the field elements (and hence the operands). It is customary
to view a finite field as a vector space, which facilitates the use of various bases for the representation of
the field elements. Among the many proposed bases, the canonical basis [4], the normal basis [8], and
the dual basis [9, 10, 11] are the most commonly used ones.
In this paper, we focus our attention on canonical basis bit-parallel multiplier architectures for even
characteristic field extensions, i.e. GF(2k
). In the canonical basis, field elements are expressed as
polynomials with coefficients from a base field (e.g. GF(2)) and field multiplication is carried out as
polynomial multiplication modulo an irreducible polynomial. Modular multiplication is usually carried
out in two steps: polynomial multiplication and reduction of the polynomial product with the irreducible
polynomial. Using the “paper and pencil” method the multiplication step may be performed in parallel
using n2
two input AND gates and (n − 1)2
two input XOR gates. This is followed by the reduction
step which requires only a number of additions depending on the number of non-zero coefficients of the
irreducible modulus polynomial. Using low-Hamming weight polynomials (e.g. trinomials, pentanomials
∗Appeared in the IEEE Transaction on Computers 53(9):1097-1105, September 2004.
1

etc.) the additions may be implemented with only 2n XOR gates. The total gate consumption becomes
2n2
− 1. On the other hand, when binary adder trees are utilized to accumulate the partial products,
the total delay of the bit-parallel multipler is found as T⊗ + (log2 n)T⊕ where T⊗ and T⊕ denote the
delays of a two input AND gate and a two input XOR gate, respectively. A variation to this theme
which combines the multiplication and reduction steps was proposed by Mastrovito in [12]. Although,
considerable research went into the design and optimization on this theme and its variations, the space
and time complexity figures have not improved much further than the quadratic space and logarithmic
time complexities cited above [13, 14, 15, 16, 17, 18].
An alternative approach for the initial multiplication step was developed by using the polynomial
version of the Karatsuba-Ofman algorithm [19]. The recursive splitting of polynomials and the special
reassembly of the partial products drastically reduces the number of AND gates required for the multi-
plication operation: nlog2 3
. Although, the reduction in the number of multiplications is accomplished by
sacrificing extra additions in each step, since the number of steps is reduced logarithmically the number
of XOR gates also enjoys a much improved asymptotic complexity of 6nlog2 3
− 8n + 2 [15]. While the
number of AND gates improves for all n ≥ 2, the number of XOR gates improves only after n ≥ 64 when
compared to the paper and pencil multiplication method. The main disadvantages of the Karatsuba mul-
tiplier are its time complexity which is about three times that of traditional multipliers: T⊗ +3(log2 n)T⊕,
and the restriction it places on the bit-length by requiring it to be a power of two.
In this paper, we pose and address the following question: Besides the Karatsuba algorithm are
there any other bit-parallel multipliers with subquadratic space complexity and perhaps with better time
complexity? To answer this question we first reintroduce the Winograd short convolution algorithm —
an old and well known digital signal processing technique — in the context of polynomial multiplication.
We then introduce a generalized recursive construction technique and present its complexity analysis.
We summarize the complexities for several constructions. Finally, we discuss the generalization of this
technique to support arbitrary polynomial lengths and present its complexity analysis.
2 Winograd Short Convolution
In this section we briefly describe the short convolution algorithm introduced by Winograd [20, 21]. Let
a(x) and b(x) represent d − 1 degree polynomials a(x) =
d−1
i=0
aixi
and b(x) =
d−1
i=0
bixi
defined over
an arbitrary field. The product of the polynomials is computed as
c(x) = a(x)b(x) =
2d−2
k=0
ckxk
where
ck =
i+j=k
aibj , for 0 ≤ k ≤ 2d − 2 . (1)
Alternatively, one can associate polynomials with sequences and view their coefficients as sequence ele-
ments:
a(x) → a = {a0, a1, a2, . . . , ad−1}
b(x) → b = {b0, b1, b2, . . . , bd−1} .
This representation enables us to define the elements of the linear convolution ¯c = a b of the two
sequences as follows
¯ck =
i+j=k
aibj , for 0 ≤ k ≤ 2d − 2 . (2)
It is easily seen that the polynomial multiplication in (1) and the linear convolution in (2) are equivalent
operations since ck = ¯ck for 0 ≤ k ≤ 2d − 2. The Winograd short convolution algorithm is based on
this association along with the following key observation: As introduced earlier a(x) and b(x) represent
polynomials of maximum degree d − 1. Their product c(x) = a(x)b(x) will have maximum degree 2d − 2.
Hence, when a made-up modulus m(x) with deg(m(x)) > 2d − 2 is introduced then their product is not
affected:
c (x) = a(x)b(x) mod m(x)
2

c(x) = a(x)b(x)
c(x) = c (x) , if deg(m(x)) > 2d − 2 .
The Winograd short convolution algorithm makes use of this fact. Choosing the factors of m(x) as
relatively prime polynomials enables efficient residue arithmetic with shorter operands. The final product
is obtained from the residues via the Chinese Remainder Theorem (CRT). The steps of the Winograd
convolution algorithm are summarized below:
Setup: Choose artificial modulus m(x) with pairwise relatively prime factors as
m(x) = m(1)
(x)m(2)
(x) . . . m(k)
(x).
Compute the polynomials M(i)
(x) given as follows.
M(i)
(x) =
m(x)
m(i)(x)
m(x)
m(i)(x)
−1
mod m(i)
(x) for i = 1, . . . , k . (3)
Step 1 Residue Computations: Find the residue representations of a(x) and b(x) with respect to
each factor of the modulus
a(i)
(x) = a(x) (mod m(i)
(x)) and (4)
b(i)
(x) = b(x) (mod m(i)
(x)) for i = 1, . . . , k . (5)
Step 2 Residue Products: Compute k parallel modular polynomial products
w(i)
(x) = a(i)
(x)b(i)
(x) (mod m(i)
(x)) for i = 1, . . . , k . (6)
Step 3 Inversion: Compute the CRT-inverse using
c (x) =
k
i=1
w(i)
(x)M(i)
(x) (mod m(x)) (7)
If deg(m(x)) > 2d−2 then c(x) = a(x)b(x) = c (x) and the algorithm will terminate with this step.
However, due to difficulties in selecting relatively prime factors the case deg(m(x)) = 2d − 2 is also
useful. The algorithm proceeds as described earlier. The computed product c (x) may not match
the actual product c(x) since in the product a(x)b(x) the term ad−1bd−1x2d−2
causes a reduction
unless ad−1bd−1 = 0. Hence, c(x) and c (x) are related to each other as follows.
c (x) = a(x)b(x) mod m(x) = a(x)b(x) − ad−1bd−1m(x)
= c(x) − ad−1bd−1m(x)
Fortunately, the desired product a(x)b(x) is easily recovered from c (x) by a simple correction step.
Step 4 Correction: If deg(m(x)) > 2d − 2 then c(x) = c (x) else if deg(m(x)) = 2d − 2 then compute
c(x) = c (x) + ad−1bd−1m(x) .
In the correction step due to the ad−1bd−1 term an additional multiplication is required. However the
degree of m(x) is reduced by one. It is customary to indicate the use of the correction technique by
adding a symbolic (x − ∞) factor to m(x).
In the following section we introduce a concrete example of a 4 × 4 Winograd short convolution
algorithm. In this construction and others described later in the paper the expressions become quite
cumbersome. To alleviate this problem we first introduce the following notation.
Notation: We define a product element as follows.
p = (
∀i∈
ai)(
∀j∈
bj) .
Here p denotes the product of two summations formed by the coefficients of a(x) and b(x), and denotes
the list of indexes of the summed coefficients. For example,
p013 = (a0 + a1 + a3)(b0 + b1 + b3) .
3

3 A Construction for d = 4
In this section we illustrate the construction of a 4×4 Winograd short convolution algorithm over a field
of characteristic 2. Our purpose is to construct an algorithm to compute the product
c(x) = a(x)b(x) = (a3x3
+ a2x2
+ a1x + a0)(b3x3
+ b2x2
+ b1x + b0) .
Setup: We apply the Winograd short convolution algorithm for d = 4 by selecting the modulus poly-
nomial as
m(x) = x2
(x2
+ 1)(x2
+ x + 1)(x − ∞) .
For CRT-inversion we compute the polynomials M(i)
(x) as given in (3) and find M(1)
(x) = x5
+
x3
+ x2
+ 1, M(2)
(x) = x5
+ x4
+ x3
, and M(3)
(x) = x4
+ x2
.
Residue Computations: We reduce a(x) and b(x) modulo m(i)
(x) for i = 1, 2, 3 and obtain the fol-
lowing residues.
a(1)
(x) = a1x + a0 b(1)
(x) = b1x + b0
a(2)
(x) = (a3 + a1)x + (a2 + a0), b(2)
(x) = (b3 + b1)x + (b2 + b0)
a(3)
(x) = (a2 + a1)x + (a3 + a2 + a0) b(3)
(x) = (b2 + b1)x + (b3 + b2 + b0)
Residue Products: The residue products are computed as
w(1)
(x) = (a1x + a0)(b1x + b0) = (a0b1 + a1b0)x + a0b0 (mod x2
)
= (p01 − p0 − p1)x + p0
w(2)
(x) = [(a3 + a1)x + (a2 + a0)][(b3 + b1)x + (b2 + b0)] (mod x2
+ 1)
= (a3 + a1)(b3 + b1)x2
+ [(a3 + a1)(b2 + b0) + (b3 + b1)(a2 + a0)]x
+(a2 + a0)(b2 + b0) (mod x2
+ 1)
= (p0123 + p03 + p13)x + (p02 + p13)
w(3)
(x) = [(a2 + a1)x + (a3 + a2 + a0)][(b2 + b1)x + (b3 + b2 + b0)] (mod x2
+ x + 1)
= (a2 + a1)(b2 + b1)x2
+ [(a2 + a1)(b3 + b2 + b0) + (a3 + a2 + a0)(b2 + b1)]x
+(a3 + a2 + a0)(b3 + b2 + b0) (mod x2
+ x + 1)
= (p013 + p023)x + (p023 + p12)
Inversion: Using (7) we derive the CRT-inverse c (x) =
5
i=0
cixi
as
c0 = p0
c1 = p01 + p0 + p1
c2 = p0123 + p023 + p12 + p01 + p02 + p13 + p1
c3 = p0123 + p023 + p013 + p0
c4 = p0123 + p023 + p12 + p01 + p0 + p1
c5 = p023 + p013 + p02 + p01 + p13 + p1
Correction: We implement the correction step c(x) = c (x) + p3m(x) and obtain the coefficients of the
desired product as
c0 = p0
c1 = p01 + p0 + p1
c2 = p0123 + p023 + p12 + p01 + p02 + p13 + p1 + p3
c3 = p0123 + p023 + p013 + p0 + p3
c4 = p0123 + p023 + p12 + p01 + p0 + p1
c5 = p023 + p013 + p02 + p01 + p13 + p1 + p3
c6 = p3
4

The final algorithm requires S⊗ = 10 multiplications:
p0123, p023, p013, p01, p12, p02, p13, p0, p1, p3
This is a significant improvement over the 16 multiplications required by the paper and pencil multipli-
cation method. However, this improvement comes at the price of an increased number of additions. First
we need to compute the summation elements that are multiplied together to obtain the product terms.
This is achieved by computing a0 +a1, a1 +a2, a0 +a2, a1 +a3 and then (a0 +a1)+a3, (a0 +a2)+a3, and
(a0 + a2) + (a1 + a3) in order. Note the we are reusing the previously computed summation terms. This
computation requires 7 additions. Likewise, another 7 additions are spent for preparing the coefficients
of b(x). Hence, the number of additions required to compute the ten product terms is Spre
⊕ = 14. We call
these additions pre-additions to differentiate from additions performed after the multiplications which
we call post-additions. The post-additions may be realized with the following algorithm.
c0 = p0
c1 = (p01 + p1) + p0
c2 = [(p0123 + p023) + (p3 + p12)] + [(p01 + p1) + (p02 + p13)]
c3 = (p0123 + p023) + (p013 + p3) + p0
c4 = [(p0123 + p023) + p12] + [(p01 + p1) + p0]
c5 = (p013 + p3) + (p01 + p1) + (p02 + p13) + p023
c6 = p3
The parentheses and brackets indicate shared subexpressions and hint the order of additions. Note that
the groupings of pre-additions and post-additions are achieved by inspection. The ad-hoc nature of this
step makes it much more difficult to realize for longer convolution algorithms. In this example, by reusing
common subexpressions the number of post-additions is reduced to Spost
⊕ = 16.
The delay associated with this algorithm may be analyzed in three parts: pre-additions, multiplica-
tions, and post-additions. Assuming two input XOR gates the pre-additions may be performed using
binary XOR trees. The longest delay path will be caused by the term p with the longest list . In both
multipliers the maximum delay is found for computing p0123 with length 4 and delay 2T⊕. Assuming all
multiplications are performed in parallel the delay is found as T⊗. For the post-additions the longest
delay path is found in coefficient c2 with 8 terms and 3 addition delays. The total delay is found as
T⊗ + 5T⊕.
4 Complexity of Winograd’s Algorithm
In this section we analyze the complexity of the Winograd short convolution algorithm and establish
conditions to achieve sub-quadratic complexity. We start by assuming that all residue products in Step
2 of Winograd’s algorithm are computed using the paper and pencil method and therefore require a
number of multiplications quadratic in the length. The entire convolution algorithm will have sub-
quadratic multiplicative complexity if S⊗ < d2
, and hence for our purposes the following condition needs
to be satisfied.
S⊗ =
k
i=1
d2
i < d2
Here di = deg (m(i)
(x)) and d = deg(m(x)). Also note that the Winograd algorithm establishes a lower
bound as follows.
k
i=1
di ≥ 2d − 1
Since
k
i=1
d2
i ≥
k
i=1
di we can combine the two conditions to bound S⊗ from both sides:
2d − 1 ≤
k
i=1
di ≤
k
i=1
d2
i < d2
2d − 1 ≤ S⊗ < d2
5

In setting up the Winograd algorithm we usually pick the equality case
k
i=1
di = 2d − 1 to eliminate
redundant multiplications. To minimize S⊗, one has to maximize the number of factors of m(x). Ideally,
if all factors are chosen as first degree polynomials then di = 1, k = 2d − 1 and
S⊗ = 2d − 1.
This represents the minimum number multiplications required to multiply two polynomials of length d.
As attractive as it may seem, for longer convolutions finding m(x) with linear factors becomes impossible
and the performance suffers degradation. Still there are sufficiently many relatively prime polynomials
that can be used for the construction of Winograd short convolution algorithms with subquadratic
multiplication complexity. The exact derivation of the addition complexity is much more complicated.
As our goal is to use shorter Winograd convolution algorithms to build longer ones we develop a
simple but useful parameterization. This will enable us in later sections to formulate the complexity
of longer convolution algorithms in terms of the parameters of the shorter convolution algorithms used
in the construction. As already used above, S⊗ denotes the number of multiplications incurred in Step
2 of the Winograd algorithm. There are two types of additions concentrated in Steps 1 and 3 of the
algorithm. These additions take place before and after the multiplications are computed. Hence, as
described earlier in Section 3 we call these additions pre-additions and post-additions and denote their
complexities by Spre
⊕ and Spost
⊕ , respectively. The space complexity of a short convolution algorithm is
identified by an ensemble of three parameters: (Spre
⊕ , S⊗, Spost
⊕ ).
The delay of the multiplier circuit is found in three parts. We use the parameter Dpre
to denote
the delay associated with computing the pre-additions. In a direct implementation the length of the
longest list of all product terms p appearing in the pre-additions will determine Dpre
. When binary
XOR trees are utilized to implement the pre-additions then the delay is found as Dpre
= log2 . In
the computation of the residue products all multiplications are performed in parallel and thus require
only one coefficient multiplication delay. The maximum delay in the post-additions is determined by the
coefficient that causes the longest delay. We associate the parameter Dpost
with this delay. In a direct
implementation this coefficient will be the one with the maximum number of product terms. As in the
pre-additions, if the post-additions are carried out using binary XOR trees the delay is determined as
Dpost
= log2 l . Here l denotes the maximum number of product terms in any coefficient of the product.
The complexity analysis presented in this section applies to the direct implementation of the Winograd
algorithm. However, there are many ad-hoc techniques which may be utilized to obtain improvements.
One of these is as follows.
Heuristic 1 (convolutional symmetry) The convolution operation is symmetric. More clearly when
two sequences are reversed their convolution is also reversed. Therefore, whenever a convolution sequence
is not symmetric, it may be possible to simplify the more complex sequence elements by deriving them
from their symmetric counterparts by applying the two sequences in reverse.
In a more mathematical form the heuristic is restated as follows. Consider the convolution of two
sequences {c} = {a} {b}, or the product computation of the polynomial product c(x) = a(x)b(x). We
treat the elements of the convolution sequence as functions of the two sequences:
cj = fj(a0, a1, . . . , ad−1; b0, b1, . . . , bd−1) , for j = 0, . . . , d − 1 .
Since the convolution operation is symmetric it follows that
c2d−2−j = fj(ad−1, ad−2, . . . , a0; bd−1, bd−2, . . . , b0)
In other words c2d−2−j may be obtained by replacing all ai with ad−1−i and all bi with bd−1−i in the
equation obtained for computing cj for any 0 ≤ j ≤ d − 1. For a sample application of this heuristic see
Case d = 4 in the Appendix.
5 The Karatsuba-Ofman Algorithm
In this section we show that a certain Winograd short convolution algorithm is essentially identical to
the Karatsuba-Ofman algorithm [19]. We state this formally as follows.
Theorem 1 The 2 × 2 Winograd short convolution algorithm constructed using the modulus m(x) =
x(x − 1)(x − ∞) is identical to the Karatsuba-Ofman algorithm.
6

Proof of Theorem 3 We aim to develop a multiplier to realize the product
c(x) = a(x)b(x) = (a0 + a1x)(b0 + b1x) .
Using the given modulus the residues are computed as
a(1)
(x) = a0 a(2)
(x) = a0 + a1
b(1)
(x) = b0 b(2)
(x) = b0 + b1
The residue products are formed as
w(1)
(x) = a0b0
w(2)
(x) = (a0 + a1)(b0 + b1)
The inverse polynomials are computed as M(1)
(x) = (x − 1)((x − 1)−1
mod x) = x − 1 and M(2)
(x) =
x(x−1
mod (x − 1)) = x. By using the CRT inversion formula shown in (7) we assemble the product as
follows
c (x) = a0b0(x − 1) + (a0 + a1)(b0 + b1)x .
Due to the (x − ∞) factor a correction step is needed. We add a1b1x(x − 1) to obtain the final product
as
c(x) = a1b1x2
+ [(a0 + a1)(b0 + b1) − a0b0 − a1b1]x + a0b0 .
2
6 Recursive Splitting
Using Winograd’s technique one may easily build a convolution algorithm for a desired length as shown
in the previous sections. However, as the size of the polynomials grows it becomes difficult to pick a
modulus with proper factors. The problem gets worse if the domain of the polynomial coefficients is
a small field such as GF(2). Another serious problem is the fast growth in the number of addition
operations required to realize the reductions and the CRT inverse computation. Both difficulties may be
overcome by recursively nesting such short convolution algorithms to achieve longer convolutions. The
same principle was successfully applied in the multilevel realization of the Karatsuba algorithm [15]. A
k-level recursive construction takes a d × d convolution (or multiplication) algorithm and extends it to a
n × n convolution algorithm where n = dk
.
Let the product to be computed be c(x) = a(x)b(x) where the two operands are represented as
a(x) = A
(0)
0 + A
(0)
1 x + A
(0)
2 x2
+ . . . + A
(0)
dk−1
xdk
−1
and
b(x) = B
(0)
0 + B
(0)
1 x + B
(0)
2 x2
+ . . . + B
(0)
dk−1
xdk
−1
with A
(0)
i , B
(0)
i ∈ GF(2). The algorithm proceeds by splitting the polynomials a(x) and b(x) into d
shorter polynomials. After the splitting a(x) and b(x) have d terms, each of length n/d = dk−1
, i.e.
a(x) = A
(1)
0 + A
(1)
1 xdk−1
+ A
(1)
2 x2dk−1
+ . . . + A
(1)
d−1x(d−1)dk−1
(8)
where each coefficient is a (dk−1
− 1)-st degree polynomial
A
(1)
i = A
(0)
idk−1 + A
(0)
idk−1+1
x + A
(0)
idk−1+2
x2
+ . . . + A
(0)
idk−1+dk−1−1
xdk−1
−1
for i = 0, 1, . . . , d−1. Since both operands have now d coefficients, to multiply the coefficients of a(x) and
b(x) a d × d convolution algorithm may be utilized. With the application of the convolution algorithm,
new products of the coefficients of a(x) and b(x) are formed. These are products of polynomials of length
dk−1
. To realize these products the splitting technique is used and another level of the d × d convolution
algorithm is applied on the coefficients. In fact, the splitting and convolution opertions may be applied
repeatedly, until the ground field GF(2) is reached. The coefficients in a level are related to the ones in
the previous level as follows.
A
(j)
i = A
(j−1)
idk−j + A
(j−1)
idk−j +1
x + A
(j−1)
idk−j +2
x2
+ . . . + A
(j−1)
idk−j +dk−j −1
xdk−j
−1
7

The d × d convolution algorithm is applied to all multiplications in each of the k levels until the
ground field is reached. Once the ground field is reached the products are computed and the polynomials
are appropriately reassembled according to the short convolution algorithm.
The recursive splitting technique not only allows a simple and uniform construction but effectively
reduces the asymptotic complexity of the multiplier. In the next section we present a detailed space and
time complexity analysis which derives the overall complexity in terms of the parameters of the short
convolution algorithm.
7 Complexity Analysis
The analysis assumes a short convolution algorithm characterized by the parameters (d, Spre
⊕ , S⊗, Spost
⊕ ,
Dpre
, Dpost
) as defined in Section 4. The analysis proceeds in three steps:
1. Pre-additions: In each recursion level the number of polynomials grows by a factor of S⊗, whereas
polynomial lengths shrink d times. Hence, the number of pre-additions performed in the overall
computation is found as:
Rpre
⊕ =
logd n
i=1
Si−1
⊗ Spre
⊕
n
di
=
Spre
⊕
S⊗ − d
(nlogd S⊗
− n) .
In each level of the recursion the pre-additions are performed with Dpre
delay. Since there are
logd n levels the delay in this stage is found as:
T pre
⊕ = T⊕Dpre
logd n .
2. Multiplications: The polynomials are split logd n times, the number of polynomials grows by a
factor of S⊗. Hence the total number of coefficient multiplications (after logd n levels of recursion)
is
R⊗ = S
logd n
⊗ = nlogd S⊗
All multiplications are computed in parallel and thus require only one multiplication delay.
T = T⊗
3. Post-additions: In calculating the total number of post-additions we need to consider the additions
required for the reassembly of the coefficients as well as the overlaps among the coefficients. Note
that product terms may be twice as long as coefficients, thus overlaps among coefficients will occur.
In the i-th iteration the number of polynomials grows to S
logd n−i
⊗ . Each polynomial has length
2di−1
−1. Since there are Spost
⊕ additions of such polynomials, multiplying these quantities together
gives the number additions as shown in the first term in the summation below. The product terms
will have 2d − 1 coefficients hence there are 2d − 2 intervals with overlaps each di−1
− 1 bits long.
Factoring in these overlaps results in the following overall figure for post-additions.
Rpost
⊕ =
logd n
i=1
S
logd n−i
⊗ [Spost
⊕ (2di−1
− 1) + (2d − 2)(di−1
− 1)]
The delay contributed by post-additions is found as
T post
⊕ = T⊕Dpost
logd n .
The overall complexities R⊗ and R⊕ = Rpre
⊕ + Rpost
⊕ are found as
R⊗ = nlogd S⊗
(9)
R⊕ = (u − v)nlogd S⊗
− un + v (10)
where
u =
2(d − 1) + 2Spost
⊕ + Spre
⊕
S⊗ − d
and v =
2(d − 1) + Spost
⊕
S⊗ − 1
.
8

d 2 5 10 100 1000 · · · ∞
logd(2d − 1) 1.584 1.365 1.278 1.149 1.100 · · · 1
Table 1: Change in the degree of the optimal asymptotic complexity
The total delay is found as follows
T = T⊗ + T⊕(Dpre
+ Dpost
) logd n . (11)
We observe that the polynomial degree of the space complexity is logd S⊗. When an optimally con-
structed Winograd short convolution algorithm is used, then S⊗ = 2d−1 and the asymptotic complexity
is optimized for that particular d. For instance the Karatsuba algorithm is optimally constructed for
d = 2 since S⊗ = 3. However, note that in Table 1 for higher values of d the exponent decreases and in
the limit case converges to 1. This means that as d increases the asymptotic complexity of a multiplier
built by extending an optimally constructed convolution algorithm will improve. Even if no optimal
construction can be found for higher d, it may still be possible to find sub-optimal (S⊗ < 2d − 1) short
convolution algorithms. When these algorithms are recursively extended, it may still be possible to ob-
tain asymptotic complexities that are better than the ones obtained for multipliers built using optimal
contructions for smaller d.
8 Several Constructions
Below Table 2 summarizes the complexities of several multiplier constructions obtained by recursively
extending the Winograd short convolution algorithms derived in this paper. The listed complexities are
obtained by a straightforward application of the formulae given in (9), (10), (11) on the convolution
algorithm parameters (S⊗, Spre
⊕ , Spost
⊕ , Dpre
, Dpost
). The complexities are given in terms of the length of
the multiplier n = dk
. The first algorithm is identical to the Karatsuba algorithm as derived in Section
5. The two algorithms for d = 4 are derived in Section 3 and in the Appendix. The algorithms for d = 3
and d = 5 are included in the Appendix. The stars in the two moduli indicate the improved versions of
the two algorithms as explained in the Appendix.
In Table 2 when compared to the paper and pencil method all six algorithms exhibit superior asymp-
totic complexities. Albeit very close, the degree of the asymptotic complexity of the Karatsuba algorithm,
i.e. log2 3 ≈ 1.58, is slightly better than that of the other six algorithms (log3 6 ≈ 1.63, log4 10 ≈ 1.66,
log5 14 ≈ 1.64). Nevertheless, the constant factor of the polynomial term in the additive complexities
R⊕ for the improved algorithms for d = 3 and d = 4 is smaller than that of the Karatsuba algorithm.
This translates into only slightly worse space complexity for certain practical values of n. For example,
the additions required for the improved algorithms for d = 3 and d = 4 are only about 10% and 20%
more that is required for the Karatsuba algorithm for n = 100. Also considering that the Karatsuba
algorithm requires n to be a power of two the algorithm for d = 3 and d = 5 clearly have practical value.
We see a different picture in the time complexities. With increasing d the base of the logarithms
increases. The increase in the constant factor is logarithmically related to the parameters Dpre
and
Dpost
. When all six complexities are converted to logarithms in base 2 and rounded appropriately the
logarithmic identities
5 log3 n = 3.15 log2 n , for d = 3
d m(x) R⊗ R⊕ T
2 x(x − 1)(x − ∞) nlog2 3 6nlog2 3 − 8n + 2 T⊗ + 3 log2 nT⊕
3 x(x + 1)(x2 + x + 1)(x − ∞) nlog3 6 97
15
nlog3 6 − 26
3
n + 11
5
T⊗ + 5 log3 nT⊕
3 x(x + 1)(x2 + x + 1)(x − ∞)∗ nlog3 6 16
3
nlog3 6 − 22
3
n + 2 T⊗ + 4 log3 nT⊕
4 x2(x2 + 1)(x2 + x + 1)(x − ∞) nlog4 10 56
9
nlog4 10 − 26
3
n + 22
9
T⊗ + 5 log4 nT⊕
4 x2(x2 + 1)(x2 + x + 1)(x − ∞)∗ nlog4 10 47
9
nlog4 10 − 22
3
n + 19
9
T⊗ + 5 log4 nT⊕
5 x(x2 + 1)(x2 + x + 1)(x3 + x + 1)(x − ∞) nlog5 14 982
117
nlog5 14 − 106
9
n + 44
13
T⊗ + 7 log5 nT⊕
Table 2: Recursive multiplier complexities of several constructions
9

4 log3 n = 2.52 log2 n , for d = 3
5 log4 n = 2.53 log2 n , for d = 4
7 log5 n = 3.01 log2 n , for d = 5
are obtained. Hence the time complexities of the four algorithms developed for d = 3 and d = 4 are
significantly better than that of the Karatsuba multiplier. Note that all six multipliers have worse time
complexities than the paper and pencil multiplication method which has only T⊗ + log2 nT⊕ delay.
We would like to emphasize that these particular constructions are only a very small fraction of the
multipliers that may be build by recursively extending short convolution algorithms obtained by Wino-
grad’s algorithm or by any other means. In the application of Winograd’s algorithm depending on the
choice of the artificial modulus m(x) the complexity parameters will change giving rise to a new recursive
multiplication algorithm. Furthermore, any short convolution algorithm obtained by any method besides
the Winograd algorithm may be recursively extended. The developed complexity formulae will work as
long as the same multiplier parameters can be identified.
9 Generalization to Arbitrary n
In many applications the multiplier length is determined by system constraints and hence n may not be
a prime power. For such applications it is crucial to build an efficient multiplier for arbitrary n. This
may be accomplished by recursively combining multiple level splitting algorithms of different lengths.
That is if n factorizes into prime powers as
n = de1
1 de2
2 . . . det
t ,
then as before the polynomial operands are recursively split in e1 levels and multiplied in each level
using a d1 × d1 convolution algorithm. However, the coefficients obtained in the final splitting are not
elements of GF(2) as before, but rather elements of GF(2d
e2
2
...d
et
t ). Thus another e2 levels of recursive
splitting and convolutions of length d2 × d2 are applied. This process is repeated for all prime factors in
the representation of n until the ground field is reached and the entire computation is expressed in terms
of GF(2) operations.
The complexity formulae developed in Section 7 still apply to each level of the recursive construction
of the multiplier. However, to determine the overall complexity, the complexity formulae should be nested
in accordance with the application order of the splitting:
R⊗,di = (dei
i )logdi
S⊗,di R⊗,di+1 = Sei
⊗,di
R⊗,di+1
R⊕,di = (ui − vi)(dei
i )logd S⊗,di − udei
i + vi R⊕,di+1
= (ui − vi)Sei
⊗,di
− uidei
i + vi R⊕,di+1
where
ui =
2(di − 1) + 2Spost
⊕,di
+ Spre
⊕,di
S⊗,di − di
and vi =
2(di − 1) + Spost
⊕,di
S⊗,di − 1
for i = 1, 2, . . . , t. The subscript di associates a particular parameter to a di × di convolution algorithm
and its application in the overall recursion. Note that R⊗,d1 and R⊕,d1 give the overal complexities
of the multiplier. Also note that R⊗,dt+1 and R⊕,dt+1 indicate the complexities of multiplication and
addition operations in GF(2), respectively. The closed form expressions are found as
R⊗ =
t
i=1
Sei
⊗,di
R⊕ =
t
i=1
(ui − vi)Sei
⊗,di
− uidei
i + vi .
By ignoring linear terms and constants the overal addition complexity may be approximated as follows
R⊕ ≈
t
i=1
(ui − vi)Sei
⊗,di
= R⊗
t
i=1
(ui − vi) .
10

The time complexity is derived by writing the latency of multiplication in a level in terms of the latency
of operations in a lower level as follows.
T⊗,di = T⊗,di+1 + T⊕,di+1 (Dpre
i + Dpost
i ) logdi
dei
i .
The total delay is found by substituting T⊗,di+1 until the ground field is reached. This gives a simple
summation which simplifies to the following closed form expression.
T = T⊗ + T⊕
d−1
i=0
(Dpre
i + Dpost
i )ei .
We also would like to note that, although not treated in the paper, the generalization of the recursive
extension technique and the presented complexity analysis to higher characteristic field extensions is
trivial. Furthermore, over higher characteristic fields selecting relatively prime polynomial factors in the
setup of the Winograd algorithm becomes easier and therefore the algorithm and the recursive extension
becomes more efficient. For instance, if p > 2n then any Winograd n × n convolution algorithm may be
realized using only 2n − 1 multiplications since all factors of m(x) can be selected as linear polynomials.
A final word on generalization is in order. It may appear that only a multiplier with a sufficiently
composite bit-length would benefit from the recursive construction technique. In certain applications
non-prime bit-lengths may be prohibited. In fact, an attack based on Weil descent was shown [22] to be
effective on elliptic curve discrete logarithm problems built over certain composite extension fields. Hence,
in practice composite extensions are avoided as much as possible in elliptic curve schemes. Fortunately,
by simply appending the polynomials to be multiplied with zero coefficients or by partitioning them
into shorter polynomials the recursive construction can be made to work. For instance, if one whishes
to multiply two polynomials each of length n = 251 then rather than finding a 251 × 251 convolution
algorithm one could partition a(x) and b(x) as
a(x) =
249
i=0
aixi
+ a250 and b(x) =
249
i=0
bixi
+ b250
and obtain the product in four parts as
a(x)b(x) = (
249
i=0
ai
249
i=0
bixi
) + a250
249
i=0
bixi
+ b250
249
i=0
aixi
+ a250b250 .
The implementation of the last three terms is trivial. The first term requires a 250 × 250 convolution
algorithm which could be build by using a 2 × 2 algorithm and a recursively extended 53
× 53
algorithm.
Alternatively one could append a(x) and b(x) with five zero coefficients and apply a recursively extended
28
× 28
convolution algorithm.
10 Conclusion
We introduced a generalized construction method for building low-complexity bit-parallel multipliers for
arbitrary bit-lengths. The construction is based on a divide and conquer strategy where through the
prime factorization of the bit-length a multiplier of arbitrary length is build using shorter multiplication
algorithms. The Winograd algorithm provides a straightforward construction method for such short
convolution algorithms. In Section 6 we have presented a recursive construction technique that extends
any d point multiplier with S⊗ multiplications into an n = dk
point multiplier with O(nlogd S⊗
) area and
O(logd n) delay. Note that the recursive combination technique and its complexity analysis presented
in Section 7 applies to any multiplication algorithm. The generalized construction method presented in
Section 9 nests such multipliers to build low-complexity bit-parallel multipliers for arbitrary bit-lengths.
The exact space and time complexities of the generalized multiplier are derived in terms of the parameters
of the short convolution algorithms. We provide a simple formula to compute the overall space and time
complexities of the multiplier from the short convolution parameters. In Section 9 we note that certain
applications restrict the bit-length to be a prime. We illustrate two simple techniques which may be used
to circumvent this restriction and make the constructions presented in this paper still relevant.
11

In Section 8 we presented six recursive multiplier constructions all of which exhibit sub-quadratic
asymptotic space complexities. Note that these multipliers are only a small fraction of the multipliers
that can be obtained by using the constructions introduced in this paper. Nevertheless, one of these
constructions obtained by extending a particular Winograd algorithm turns out to be identical to the
Karatsuba algorithm. We note that two of the other four constructions and the Karatsuba algorithm
require slightly more gates for practical values of the bit-length. On the hand, the time complexities of
two of the multipliers are signiﬁcantly better than that of the Karatsuba multiplier. The construction
for d = 3 and d = 5 complement the Karatsuba multiplier since the Karatsuba algorithm is limited may
only be used when the length of the multiplication is a power of two.
Finally we point out a possible disadvantage of the algorithms presented in this paper. When com-
pared to the paper and pencil method the algorithms with sub-quadratic complexity including the Karat-
suba algorithm appear to have less structure. This irregularity may cause additional wire delays in actual
VLSI implementation and may render the comparison in delay imprecise. We deﬁne a VLSI level perfor-
mance evaluation of these algorithms and comparison to the paper and pencil method as future work.
12

References
[1] R. E. Blahut. Fast Algorithms for Digital Signal Processing. Addison-Wesley, Reading, Mas-
sachusetts, USA, 1984.
[2] R. E. Blahut. Theory and Practice of Error Control Codes. Addison-Wesley, Reading, Massachusetts,
USA, 1983.
[3] S. B. Wicker and V. K. Bhargava, editors. Reed-Solomon Codes and Their Applications. IEEE
Press, 1994.
[4] E. R. Berlekamp. Algebraic Coding Theory. McGraw-Hill, New York, New York, USA, 1968.
[5] W. W. Peterson and E. J. Weldon Jr. Error-Correcting Codes. MIT Press, Cambridge, Mas-
sachusetts, USA, 1972.
[6] S.W. Golomb. Shift Register Sequences. Holden-Day, San Francisco, California, USA, 1967.
[7] R. J. McEliece. Finite Fields for Computer Scientists and Engineers. Kluwer Academic Publishers,
2nd edition, 1989.
[8] Blake, Gao, Mullin, Vanstone, and Yaghgoobin. Applications of Finite Fields. Kluwer Academic
Publisher, 1993.
[9] W. Geiselmann and D. Gollmann. Self-Dual Bases in Fqn . Designs, Codes and Cryptography,
3:333–345, 1993.
[10] S. T. J. Fenn, M. Benaissa, and D. Taylor. GF(2m
) Multiplication and Division Over the Dual
Basis. IEEE Transactions on Computers, 45(3):319–327, March 1996.
[11] S. T. J. Fenn, M. Benaissa, and D. Taylor. Finite Field Inversion Over the Dual Basis. IEEE
Transactions on Very Large Scale Integration (VLSI) Systems, 4(1):134–136, March 1996.
[12] E. D. Mastrovito. VLSI Architectures for Computation in Galois Fields. PhD thesis, Linköping
University, Department of Electrical Engineering, Linköping, Sweden, 1991.
[13] E. D. Mastrovito. VLSI Design for Multiplication over Finite Fields GF(2m
). In Lecture Notes
in Computer Science 357, Applied Algebra, Algebraic Algorithms and Error-Correcting Codes,
(AAECC-6), pages 297–309, Berlin, Germany, March 1989. Springer-Verlag.
[14] M. A. Hasan, M. Wang, and V. K. Bhargava. Modular Construction of Low Complexity Parallel
Multipliers for a Class of Finite Fields GF(2m
). IEEE Transactions on Computers, 41(8):962–971,
August 1992.
[15] C. Paar. Efficient VLSI Architectures for Bit-Parallel Computation in Galois Fields. PhD thesis,
(Engl. transl.), Institute for Experimental Mathematics, University of Essen, Essen, Germany, June
1994. ISBN 3–18–332810–0.
[16] Ç. K. Ko¸c and B. Sunar. Low-complexity bit-parallel canonical and normal basis multipliers for a
class of finite fields. IEEE Transactions on Computers, 47(3):353–356, March 1998.
[17] B. Sunar and Ç. K. Ko¸c. Mastrovito multiplier for all trinomials. IEEE Transactions on Computers,
48(5):522–527, May 1999.
[18] B. Sunar and Ç. K. Ko¸c. An efficient optimal normal basis type II multiplier. IEEE Transactions
on Computers, 50(1):83–87, January 2001.
[19] A. Karatsuba and Y. Ofman. Multiplication of Multidigit Numbers on Automata. Soviet Physics-
Doklady (English translation), 7(7):595–596, 1963.
[20] S. Winograd. Some bilinear forms whose multiplicative complexity depends on the field of constants.
Mathematical Systems Theory, 10:169–180, 1977.
[21] S. Winograd. Arithmetic complexity of computations. SIAM, 1980.
[22] M. Jacobson and A. J. Menezes, and A. Stein. Solving elliptic curve discrete logarithm problems
using Weil descent. CACR Technical Technical Report, CORR2001-31, University of Waterloo, May
2001.
13

A Appendix
In this section we summarize the results of the application of the Winograd short convolution algorithm
to the cases d = 3, 4 and d = 5 for selected moduli. The presented complexities are used in the paper for
deriving the complexities of the recursive application of these algorithms.
CASE d = 3
Modulus: m(x) = x(x + 1)(x2
+ x + 1)(x − ∞).
Residue Products:
w(1)
(x) = p0
w(2)
(x) = p012
w(3)
(x) = (p01 + p02)x + p02 + p12
Multiplication Algorithm:
c0 = p0
c1 = (p012 + p02) + (p12 + p2)
c2 = p012 + p01 + p12
c3 = (p012 + p02) + (p01 + p0)
c4 = p2
Product Terms:
p0, p2, p01, p12, p02, p012
Preadditions for one operand:
a0 + a1, a1 + a2, a0 + a2, (a0 + a1) + a2
Complexity parameters:
(S⊗, Spre
⊕ , Spost
⊕ , Dpre
, Dpost
) = (6, 8, 7, 2, 3)
Using the following identity
p012 = p01 + p02 + p12 + p0 + p1 + p2 ,
the coeﬃcients of x3
and x may further be simpliﬁed to yield a slightly better complexity:
c0 = p0
c1 = p01 + p0 + p1
c2 = p012 + p01 + p12
c3 = p12 + p2 + p1
c4 = p2
Product Terms:
p0, p1, p2, p01, p12, p012
Pre-additions for one operand:
a0 + a1, a1 + a2, (a0 + a1) + a2
(S⊗, Spre
⊕ , Spost
⊕ , Dpre
, Dpost
) = (6, 6, 6, 2, 2)
14

CASE d = 4
Improved version of the Algorithm derived in Section 3 using Heuristic 1:
Modulus: m(x) = x2
(x2
+ 1)(x2
+ x + 1)(x − ∞)
c0 = p0
c1 = (p01 + p0 + p1)
c2 = [(p0123 + p013) + p12] + (p23 + p2 + p3)
c3 = (p0123 + p023) + p013 + p0 + p3
c4 = [(p0123 + p023) + p12] + (p01 + p0 + p1)
c5 = (p23 + p2 + p3)
c6 = p3
Products:
p0, p1, p2, p3, p01, p12, p23, p023, p013, p0123
a0 + a1, a1 + a2, a2 + a3, a0 + (a2 + a3), (a0 + a1) + a3, (a0 + a1) + (a2 + a3)
(S⊗, Spre
⊕ , Spost
⊕ , Dpre
, Dpost
) = (10, 12, 13, 2, 3)
CASE d = 5
Modulus: m(x) = x(x2
+ 1)(x2
+ x + 1)(x3
+ x + 1)(x − ∞)
Residue Products:
w(1)
(x) = p0
w(2)
(x) = p024 + p13 + (p01234 + p024 + p13)x
w(3)
(x) = p023 + p124 + (p0134 + p023)x
w(4)
(x) = p123 + p134 + p24 + p03 + (p014 + p123 + p03)x + (p012 + p014 + p123 + p24)x2
Multiplication Algorithm:
c0 = p0
c1 = p014 + p023 + p124 + p024 + (p123 + p13) + p03 + p4
c2 = [(p0134 + p014) + (p124 + p24)] + p01234 + (p024 + p012) + (p123 + p13) + p0
c3 = [(p0134 + p014) + (p124 + p24)] + p134 + p4
c4 = p0134 + (p023 + p134) + p012 + p123 + p0
c5 = (p024 + p012) + p014 + (p123 + p13) + p24 + (p0 + p4)
c6 = [p01234 + (p023 + p134)] + p014 + (p124 + p24) + (p0 + p4)
c7 = [p01234 + (p023 + p134)] + (p024 + p012) + p03 + p13 + (p0134 + p014) + (p0 + p4)
c8 = p4
Product terms:
p01234, p0134, p134, p023, p014, p024, p012, p123, p124, p03, p13, p24, p0, p4
a0 + a3, a1 + a3, a2 + a4, a0 + a1, a1 + (a2 + a4), (a1 + a3) + a2,
(a0 + a1) + a2, a0 + (a2 + a4), (a0 + a1) + a4, a2 + (a0 + a3), (a1 + a3) + a4,
[(a1 + a3) + a4] + a0, [a0 + (a2 + a4)] + (a1 + a3)
Complexity Parameters:
(S⊗, Spre
⊕ , Spost
⊕ , Dpre
, Dpost
) = (14, 26, 36, 3, 4)
15

Subquad multi ff

More Related Content

What's hot (20)

Viewers also liked (18)

Similar to Subquad multi ff (20)

Recently uploaded (20)

Subquad multi ff