1

DIVISION BY DIGIT RECURRENCE

SEVERAL DIVISION METHODS:

• DIGIT-RECURRENCE METHOD – studied in this chapter
• MULTIPLICATIVE METHOD (Chapter 7)
• VARIOUS APPROXIMATION METHODS (power series expansion),
• SPECIAL METHODS SUCH AS CORDIC (Chapter 11) AND CONTINUED
PRODUCT METHODS.
IMPLEMENTATIONS:
• SEQUENTIAL
• COMBINATIONAL
1. PIPELINED
2. NONPIPELINED
• COMBINATIONAL/SEQUENTIAL

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 2
DEFINITION AND NOTATION

x = q · d + rem

|rem| < |d| · ulp and sign(rem) = sign(x)

DIVIDEND x
DIVISOR d
QUOTIENT q
REMAINDER rem
• INTEGER QUOTIENT: ulp = 1,
• FRACTIONAL QUOTIENT: ulp = r−n
TWO TYPES OF DIVISION OPERATION:
1. INTEGER DIVISION, WITH INTEGER OPERANDS AND RESULT
USUALLY REQUIRES AN EXACT REMAINDER
2. FRACTIONAL DIVISION
TO AVOID QUOTIENT OVERFLOW: x < d
QUOTIENT ROUNDED, WHICH CAN RESULT IN A NEGATIVE REMAIN-
DER
Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
 3
FRACTIONAL DIVISION - ASSUMPTIONS

• OPERANDS/RESULT IN SIGN-AND-MAGNITUDE FORMAT =⇒CONSIDER

MAGNITUDES ONLY

1/2 ≤ d < 1; x < d; 0 < q < 1

• FOR SIMPLER SELECTION: qj ∈ Da = {−a, −a + 1, . . . , −1, 0, 1, . . . , a −

1, a}
• REDUNDANCY FACTOR
a 1
ρ = , <ρ≤1
r−1 2

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 4
RECURRENCE

• QUOTIENT AFTER j STEPS: q[j] = q[0] + Pji=1 qir−i

• FINAL QUOTIENT: q = q[n] = q[0] + Pni=1 qir−i
• FINAL QUOTIENT ERROR BOUND: 0 ≤ ǫq = xd − q < r−n
• ERROR AT STEP j:
x n
X −i a
ǫ[j] = − q[j] ≤ ǫ[n] + max(qi)r = ǫ[n] + (r−j − r−n)
d i=j+1 r−1
• FOR CONVERGENCE: ǫ[j] ≤ ρr−j
• FOR SIMPLER SELECTION, ALLOW NEGATIVE ERRORS
x
|ǫ[j]| = | − q[j]| ≤ ρr−j
d
• ELIMINATE DIVISION FROM ERROR BOUND: |x − dq[j]| ≤ ρdr−j

• DEFINE RESIDUAL (PARTIAL REMAINDER): w[j] = rj (x − dq[j])

• RESIDUAL RECURRENCE: w[j + 1] = rw[j] − dqj+1 w[0] = x

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 5
cont.

• BOUND ON w[j]: |w[j]| ≤ ρd

• SELECT QUOTIENT DIGIT TO KEEP w[j + 1] BOUNDED:

qj+1 = SEL(rw[j], d)

• REDUNDANCY IN QUOTIENT-DIGIT SET ALLOWS

c
qj+1 = SEL(y,
c d)

WHERE yc IS TRUNCATED rw[j], etc.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 6
IMPLEMENTATION OF DIVISION STEP

1. ONE DIGIT ARITHMETIC LEFT-SHIFT OF w[j] TO PRODUCE rw[j];

2. DETERMINATION OF THE QUOTIENT DIGIT qj+1

BY THE QUOTIENT-DIGIT SELECTION FUNCTION;

3. GENERATION OF THE DIVISOR MULTIPLE d × qj+1; and

4. SUBTRACTION OF dqj+1 from rw[j].

5. UPDATE OF THE QUOTIENT q[j] TO q[j + 1] BY THE ON-THE-FLY
CONVERSION

q[j + 1] = CON V (q[j], qj+1)

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 7

w[j] - present residual

ARITHMETIC LEFT SHIFT d - divisor

rw[j] - shifted residual

y d

SEL qj+1 dqj+1 subtraction

selection
qj+1
quotient update
DIVISOR MULTIPLE
(on-the-fly)
GENERATION
recurrence step time
rw[j] dqj+1
(b)
SUBTRACTION

w[j+1] - next residual

QUOTIENT CONVERSION qj+1

(a) q[j+1]

Figure 5.1: (a) COMPONENTS OF A DIVISION STEP. (b) TIMING.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 8
MAIN FACTORS AFFECTING COST AND EXECUTION TIME

• RADIX r
• QUOTIENT-DIGIT SET
1. CANONICAL: 0 ≤ qj ≤ r − 1
2. REDUNDANT: qj ∈ Da = {−a, −a + 1, . . . , −1, 0, 1, . . . , a − 1, a}
a 1
• REDUNDANCY FACTOR: ρ = r−1
, ρ> 2
• REPRESENTATION OF RESIDUAL:
1. NONREDUNDANT (e.g., 2’s complement)
2. REDUNDANT: carry-save, signed-digit

• QUOTIENT-DIGIT SELECTION FUNCTION

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 9
INITIALIZATION AND TERMINATION

Initialization:
• w[0] = x − dq[0] and |w[0]| ≤ ρd. Options:
– Make q[0] = 0 and
∗ For ρ = 1 we make w[0] = x/2.
∗ For 1/2 < ρ < 1 we make w[0] = x/4
Compensated in the termination step
– Make q[0] = 1 and w[0] = x − d. Applicable for ρ < 1 because q > 1 + ρ
not allowed.
Termination:
• QUOTIENT:



 q[N ] if w[N ] ≥ 0
q = 
q[N ] − r−N if w[N ] < 0
N – number of iterations
If dividend shifted in initialization - shift quotient (extra iteration)

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 10
ON-THE-FLY QUOTIENT CONVERSION

• j MS DIGITS OF CONVERTED QUOTIENT

j
X
Q[j] = qir−i
i=1

• UPDATE
Q[j + 1] = Q[j] + qj+1r−(j+1)

• SINCE qj+1 CAN BE NEGATIVE:


Q[j] + qj+1r−(j+1)


 if qj+1 ≥ 0
Q[j + 1] = 
Q[j] − r−j + (r − |qj+1|)r−(j+1) if qj+1 < 0

• DISADVANTAGE: SUBTRACTION Q[j] − r−j

REQUIRES THE PROPAGATION OF A BORROW – SLOW

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 11
ON-THE-FLY CONVERSION (cont.)

• DEFINE ANOTHER FORM QM [j]

QM [j] = Q[j] − r−j

• NEW CONVERSION ALGORITHM IS


Q[j] + qj+1r−(j+1)


 if qj+1 ≥ 0
Q[j + 1] = 
QM [j] + (r − |qj+1|)r−(j+1) if qj+1 < 0
• SUBTRACTION REPLACED BY LOADING THE FORM QM [j]
• UPDATE FORM QM [j] AS FOLLOWS:

QM [j + 1] = Q[j + 1] − r−(j+1)

 Q[j] + (qj+1 − 1)r
−(j+1)

 if qj+1 > 0
= 
QM [j] + ((r − 1) − |qj+1|)r−(j+1) if qj+1 ≤ 0
• ALL ADDITIONS ARE CONCATENATIONS

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 12
ON-THE-FLY CONVERSION ALGORITHM




 (Q[j], qj+1) if qj+1 ≥ 0
Q[j + 1] = 
(QM [j], (r − |qj+1|)) if qj+1 < 0

(Q[j], qj+1 − 1)


 if qj+1 > 0
QM [j + 1] = 
(QM [j], ((r − 1) − |qj+1|)) if qj+1 ≤ 0

INITIAL CONDITIONS Q[0] = QM [0] = 0 (for a positive quotient)

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 13
EXAMPLE OF RADIX-2 CONVERSION

j qj Q[j] QM [j]
0 0 0
1 1 0.1 0.0
2 1 0.11 0.10
3 0 0.110 0.101
4 1 0.1101 0.1100
5 -1 0.11001 0.11000
6 0 0.110010 0.110001
7 0 0.1100100 0.1100011
8 -1 0.11000111 0.11000110
9 1 0.110001111 0.110001110
10 0 0.1100011110 0.1100011101
11 1 0.11000111101 0.11000111100
12 0 0.110001111010 0.110001111001

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 14
IMPLEMENTATION OF THE CONVERSION

Q QM load
select
2-1 MUX

Qin

LOAD & SHIFT-IN

Q REG

CONTROL
q Q qj+1

QM load
select
2-1 MUX

QMin
QM REG

QM

Figure 5.2: Implementation of on-the-fly conversion.


shif t Q with insert (Qin)


 if Cshif tQ = 1
Q ← 
shif t QM with insert (Qin) if CloadQ = 1



 shif t QM with insert (QMin) if Cshif tQM = 1
QM ← 
shif t Q with insert (QMin) if CloadQM = 1

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 15
cont.




 qj+1 if qj+1 ≥ 0
Qin = 
r − |qj+1| if qj+1 < 0



 qj+1 − 1 if qj+1 > 0
QMin = 
(r − 1) − |qj+1| if qj+1 ≤ 0

′ ′
REGISTER CONTROL SIGNALS: CloadQ = Cshif tQ and CloadQM = Cshif tQM

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 16
EXAMPLE OF RADIX-4 CONVERSION

qj+1 Qin Cshif tQ Q[j + 1] QMin Cshif tQM QM [j + 1]

3 3 1 (Q[j], 3) 2 0 (Q[j], 2)
2 2 1 (Q[j], 2) 1 0 (Q[j], 1)
1 1 1 (Q[j], 1) 0 0 (Q[j], 0)
0 0 1 (Q[j], 0) 3 1 (QM [j], 3)
-1 3 0 (QM [j], 3) 2 1 (QM [j], 2)
-2 2 0 (QM [j], 2) 1 1 (QM [j], 1)
-3 1 0 (QM [j], 1) 0 1 (QM [j], 0)

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 17
x

INITIAL STEP

w[0]

RESIDUAL REG
x
w[j]

d INITIAL STEP
RECURRENCE
STEP w[0]
q[j+1] w[j+1]

w[N] RESIDUAL REG

TERMINAL STEP w[ik]

d qik+1
q REC. STEP
(a)
w[ik+1]

ON-THE-FLY CONVERTER
qik+2

& TERMINAL STEP

COMBINATIONAL
x REC. STEP
w[ik+2]
INITIAL STEP

w[0]
d q1
REC. STEP
w[1]
ON-THE-FLY CONVERTER

q(i+1)k
q2
& TERMINAL STEP

REC. STEP
COMBINATIONAL

REC. STEP
w[2] w[(i+1)k]

w[N] q
(c)

qN
REC. STEP
w[N]

(b) q

Figure 5.3: DIVISION IMPLEMENTATION: (a) TOTALLY SEQUENTIAL. (b) TOTALLY COMBINATIONAL. (c) COMBINED
IMPLEMENTATION.
Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
 18
EXAMPLES OF ALGORITHMS AND IMPLEMENTATIONS

• CELLS: delay function of the load; delay and area in terms of 2-NAND.
• DEGREE OF OPTIMIZATION: the same modules have been used in all de-
signs.
• INTERCONNECTIONS NOT INCLUDED: not considered the delay, area nor
load of interconnections.
• EXECUTION TIME AND THE AREA FOR 53-BIT OPERANDS AND RE-
SULT
• INCLUDED DELAY AND AREA OF REGISTERS

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 19
SCHEMES COMPARED

r2 Scheme Radix-2 with carry-save residual.

r4 Scheme Radix-4 with a = 2 and carry-save residual.
r8 Scheme Radix-8 with a = 7 and carry-save residual.
r16over Scheme Radix-16 with two overlapped radix-4 stages.
r512 Scheme Radix-512 with a = 511, carry-save residual, scaling, and quotient-
digit selection by rounding.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 20

2-input MUX 3-input MUX: [3:2] MODULE a b c a b c

Delay: 1.6 (FULL ADDER)
Area: 4.4
HA
4-input MUX:
g’ p
Delay: 1.8 p
Area: 5.6 g’ HA*

Delay: 1.4 vci+1 vsi

(a)
Area: 3.0
Delays:
a,b to vsi: 4.2
BUFFER REGISTER CELL
c to vsi: 2.2
D Q
a,b to vci+1 : 3.8
Delay: 1.8 Delay: 4.0 vci+1 vsi
Area:2.6 c to vci+1: 2.0
CK Area: 4.0
Area: 6.7
(b) SET (c)

a b b c c a a b c d
[4:2] MODULE

a b c d

ti+1 [4:2]
ti
Module
vci
vci+1 vsi ti+1 d ti
Delays:
0 1
to ti+1: 1.6 MUX

to vci+1: 5.6 vci

to vsi: 6.0
vci+1 (d) vsi
Area: 15.5

Figure 5.4: Basic modules: (a) Multiplexers. (b) Buffer and register cell. (c) Full-adder. (d) [4:2] module.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 21
RADIX-2 DIVISION WITH CS RESIDUAL

• REDUNDANT RESIDUAL w[j] = (W C[j], W S[j])

1. [Initialize]
W S[0] ← x/2; W C[0] ← 0; q0 ← 0;Q[-1]=0
2. [Recurrence]
for j = 0 . . . n + 1 (n + 2 iterations because of initialization and guard bit)
qj+1 ← SEL(y); c

(W C[j + 1], W S[j + 1]) ← CSADD(2W C[j], 2W S[j], −qj+1d);

Q[j] ← CON V ERT (Q[j − 1], qj )
end for
3. [T erminate]
If w[n + 2] < 0 then q = 2(CON V ERT (Q[n + 1], qn+2 − 1))
else q = 2(CON V ERT (Q[n + 1], qn+2))

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 22
RADIX-2 ALGORITHM (cont.)

• n is the precision in bits,

• SEL is the quotient-digit selection function:







1 if 0 ≤ yc ≤ 3/2

qj+1 = SEL(y)
c =



0 if y
c = −1/2


 −1 if −5/2 ≤ y
 c ≤ −1

The estimate yc has four bits (three integer bits and one fractional bit) of the
shifted residual in carry-save form,
• CSADD is carry-save addition
• −qj+1d is in 2’s complement form, and
• CON V ERT on-the-fly quotient conversion function

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 23

Dividend x = (0.10011111), divisor d = (0.11000101), scaled residual

2w[0] = 2(x/2) = x, qcomputed = q/2

2W S[0] = 000.10011111
∗
2W C[0] = 000.00000001 y[0]
c = 0.5 q1 = 1
−q1d = 11.00111010
2W S[1] = 111.01001000
2W C[1] = 000.01101100 y[0]
c = −1 q2 = −1
−q2d = 00.11000101
2W S[2] = 111.11000010
∗
2W C[2] = 001.00110001 y[1]
c = 0.5 q3 = 1
−q3d = 11.00111010
2W S[3] = 011.10010010
∗
2W C[3] = 100.11001001 y[2]
c =0 q4 = 1
−q4d = 11.00111010
2W S[4] = 000.11000010
2W C[4] = 110.01101000 y[4]
c = −1.5 q5 = −1
−q5d = 00.11000101
* for 2’s complement of qj+1d
q = 2(.11̄111̄) = .1101
Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
Figure 5.7: Example of radix-2 division with residual in carry-save form.
 24
Divisor d WC[0] = 0 WC[j+1] WS[0] = x /2 WS[j+1]

Reg. D Reg. WC Reg. WS

2WC[j] 2WS[j] 4
d d 4
buffers
qj+1 2-1 MUX
2 qSEL
c b a
{d, 0, d} q- 2 qj+1 q+ q-
[3:2]
WS[j+1]
WC[j+1]
1 1 0
SZ; Convert
0 0 0
qj+1 -1 0 1
q
[3:2] adder

2.2
qSEL buff MUX HA* REG
6.8 1.8 1.4 4

HA

critical path Modules defined in Fig. 5.4

Figure 5.8: IMPLEMENTATION OF RADIX-2 SCHEME.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 25
DELAY AND AREA OF RADIX-2 STAGE

element delay area

q-digit selection 6.8 50
buffers 1.8 5
MUX 1.4 160
CSA 2.2 360
registers (3) 4.0 650
Convert & Round (NC) 1360
Cycle time 16.2
Total area 2585
NC denotes a delay not in the critical path

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 26
RADIX-4 DIVISION WITH CS RESIDUAL

1. QUOTIENT DIGIT SET {-2,-1,0,1,2}

2. ρ < 1 initialize W S[0] ← x/4
3. THE NEXT RESIDUAL
(W C[j + 1], W S[j + 1]) ← CSADD(4W C[j], 4W S[j], −qj+1d)
4. QUOTIENT-DIGIT SELECTION DEPENDS ON ESTIMATES OF SHIFTED
RESIDUAL AND DIVISOR described in terms of SELECTION CONSTANTS
mk (i)

qj+1 = k if mk (i) ≤ ŷ < mk+1(i)

5. FINAL QUOTIENT = 4 x OBTAINED QUOTIENT

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 27
SELECTION CONSTANTS

• qj+1 = k if mk (i) ≤ ŷ < mk+1(i)

• i = 16dˆ and dˆ divisor truncated to the 4th fractional bit and
• ŷ is 4w[j] truncated to the 4th fractional bit.

i 8 9 10 11 12 13 14 15
m2(i)+ 12 14 15 16 18 20 20 24
m1(i)+ 4 4 4 4 6 6 8 8
m0(i)+ -4 -6 -6 -6 -8 -8 -8 -8
m−1(i)+ -13 -15 -16 -18 -20 -20 -22 -24
+: real value = shown value/16

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 28

Dividend x = (0.10101111), divisor d = (0.11000101) (i = 16(0.1100)2 = 12)

scaled residual 4w[0] = 4(x/4) = x, qcomputed = q/4

4W S[0]+ = 000.10101111
4W C[0]+ = 000.00000001 ∗
y[0]
c = 10/16 q1 = 1
−q1d+ = 11.00111010
W S[1] = 1.10010100
W C[1] = 0.01010110
4W S[1]+ = 110.01010000
4W C[1]+ = 001.01011000 y[1]
c = −6/16 q2 = 0
−q2d+ = 00.00000000
W S[2] = 1.00001000
W C[2] = 0.10100000
4W S[2]+ = 100.00100000
4W C[2]+ = 010.10000000 y[2]
c = −22/16 q3 = −2
−q3d+ = 01.10001010
w[3] = 0.00101010
* least-significant 1 for 2’s complement of qj+1d
+ only one integer bit used in the recurrence, because of the range of w[j + 1].
q[3] = .102̄4 = .0324
Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
Figure 5.9: Example of radix-4 division with residual in carry-save form.(On-the-fly conversion and termination not shown)
 29
DELAY AND AREA OF RADIX-4 STAGE

element delay area

q-digit selection 10.8 160
buffers 1.8 10
MUX 1.8 300
CSA 2.2 360
registers (3) 4.0 650
Convert & Round (NC) 1360
Cycle time 20.6
Total area 2840
NC denotes a delay not in the critical path

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 30
Divisor d WC[0]=0 WC[j+1] WS[0] = x /4 WS[j+1]

Reg. D Reg. WC Reg. WS

4WC[j] 4WS[j] 7
d d 7
buffers
qj+1 4-1 MUX 3
4 qSEL
{d}4
{2d, d, 0, d, 2d} c b a
(qj+1 < 0) 4
[3:2]
WS[j+1]
WC[j+1]
SZ; Convert
qj+1
q (q2+, q2-, q1+, q1-)
qj+1 = 0 coded as
(0, 0, 0, 0)
[3:2] adder

2.2
qSEL buff MUX HA* REG
10.8 1.8 1.8 4

HA

critical path Modules defined in Fig. 5.4

Figure 5.10: IMPLEMENTATION OF RADIX-4 SCHEME.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
 31
RADIX-8 DIVISION WITH CS RESIDUAL

• QUOTIENT DIGIT SET {-7, ..., 7} DECOMPOSED INTO

qH = {−8, −4, 0, 4, 8}
AND

qL = {−2, −1, 0, 1, 2}
element delay area
q-digit selection (qh) 12.2 610
buffers 1.8 20
MUXes 1.8 600
CSAh 2.2 360
CSAl 4.2 360
registers (3) 4.0 650
Convert & Round (NC) 1360
Cycle time 26.2
Total area 3960

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 32
Divisor d WC[0]=0 WC[j+1] WS[0]=x/2 WS[j+1]

Reg. D Reg. WC Reg. WS

8WC[j] 8WS[j] 8
d d 8
buffers
qH
j+1 4-1 MUX 3
4 qSEL
{d}4
{8d, 4d, 0, 4d, 8d}
(qH
j+1 < 0) 4 4
[3:2]
d d qj+1: qH
j+1 qLj+1
buffers
(decoded as
qLj+1 4-1 MUX in Fig. 5.10)
4

{2d, d, 0, d, 2d} (qLj+1 < 0)

[3:2]
WS[j+1]
WC[j+1]
SZ; Convert
(qH L
j+1,qj+1)

q
qH
j+1

2.2
qSEL buff MUX HA* HA REG
12.2 1.8 1.8 4
HA 4.2

qSEL buff MUX HA*

qLj+1

Modules defined in Fig. 5.4

critical path

Figure 5.11: Implementation of radix-8 scheme.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
 33
RADIX-16 DIVISION WITH TWO RADIX-4 OVERLAPPED STAGES

element delay area

CSA 4.2 220
q-digit selection 11.2 820
MUX 1.4
buffers 1.8 20
MUXes 1.8 600
CSA1 (NC) 360
CSA2 2.1 360
registers (3) 4.0 650
Convert & Round (NC) 1360
Cycle time 26.6
Total area 4390

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 34
Divisor d WC[0]=0 WC[j+2] WS[0]=x/4 WS[j+2]

Reg. D Reg. WC Reg. WS

d d 4WC[j] 4WS[j]
buffers
qj+1 4-1 MUX
4

{2d, d, 0, d, 2d}
[3:2] (qj+1 < 0)
d d
buffers WS[j+1] WC[j+1]

qj+2 4-1 MUX

4
4W[j+1]
{2d, d, 0, d, 2d}
[3:2] (qj+2 < 0)
WS[j+2] WC[j+2]

(a)
-2 -1 0. 1 2 3 4 5 6 7
4WS[j] x x x. x x x x x x x
{4W[j]}7
4WC[j] x x x. x x x x x x x

7 most 8 least
{4W[j]}7
{2d}7 {d}7 {d}7 {2d}7

8 least

7 most 8 8 8
3
{d}4 qSEL [3:2] [3:2] [3:2] [3:2]

7 7

qSEL qSEL qSEL qSEL qSEL

5-1 MUX

Conditional truncated
qj+1 qj+2 residuals
(b)

Figure 5.12: Implementation of radix-16 with radix-4 stages.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 35

CSA2

qj+1
qSEL buff MUX CSA1 HA REG
12.2 1.8 1.8 4
qj+2
2.2
[3:2]8 qSEL MUX buff MUX HA*
4.2 11.2 1.4 1.8 1.8

Modules defined in Fig. 5.4

critical path

Figure 5.13: Critical path in radix-16 scheme.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 36
RADIX-512 DIVISION WITH PRESCALING AND SELECTION BY
ROUNDING

Cycle 1 : COMPUTE M ≈ 1/d; compare d and x and set g

Cycle 2 : COMPUTE z = M d (in c-s form); v = 2−g x
Cycle 3 : INITIALIZE w[0] = M v (in c-s form); ASSIMILATE z;
Cycles 4 to 9 : ITERATE
qj+1 = round(ŷ); w[j + 1] = 512w[j] − qj+1z

Cycle 10 : CORRECTING AND ROUNDING.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 37
RADIX-512 DIVISON (cont.)

• RECTANGULAR MULTIPLIER-ACCUMULATOR
• DELAY-AREA
element delay area
M-module (NC) 1800
MUX 1.4
recoder 6.0 70
buffer 1.8
MUX 1.8 3000
2 levels of 4-2 CSA 12.0 3100
registers(3) 4.0 650
Convert & Round (NC) 1360
Cycle time 27
Total area 9980

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 38
x
15 d
Module
{d}15 (Initialization not shown)
M

M z
scale/iterate
MUX MUX

Round &
Recode
8+8 53+9+5 = 67
qj+1 M ; qj+1
512w[j]
MULTIPLIER
- ACCUMULATOR
Md; Mx; 512w -qz

WC[j+1] 67 67 WS[j+1]

WC[j] WS[j]

2(10+2)

CPA

WC[j+1] WS[j+1]

8+8
SZ; CONVERT qj+1

Figure 5.14: Implementation of radix-512 scheme.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
 39
OVERALL COMPARISONS

Scheme r2 r4 r8 r16 (overlapped) r512

Cycle-time factor 1.0 1.3 1.6 1.6 1.7
Number of cycles† 57 29 20 15 10
Speedup 1.0 1.5 1.8 2.4 3.4
Area factor 1.0 1.1 1.5 1.7 3.9
†Correction: two cycles for radix-2, one cycle for other cases.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 40
QUOTIENT-DIGIT SELECTION FUNCTION

Quotient-digit set:
qj+1 ∈ Da = {−a, −a + 1, . . . , −1, 0, 1, . . . , a − 1, a}
Redundancy factor:
a 1
ρ= , <ρ≤1
r−1 2
• TWO FUNDAMENTAL CONDITIONS FOR q-SELECTION
• CONTAINMENT – must guarantee bounded residual
• CONTINUITY – there must exist a valid choice of qj+1 in the range of shifted
residual

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 41
CONTAINMENT AND SELECTION INTERVALS

• RESIDUAL RECURRENCE
w[j + 1] = rw[j] − dqj+1 |w[j]| ≤ ρd

ρ = a/(r − 1) − a ≤ qj ≤ a
• SELECTION INTERVALS
• If rw[j] ∈ [Lk , Uk ] then qj+1 = k makes w[j + 1] bounded

Lk ≤ rw[j] ≤ Uk ⇒ ρd ≤ w[j + 1] = rw[j] − k · d ≤ ρd

• EXPRESSIONS FOR SELECTION INTERVALS

Uk = (k + ρ)d Lk = (k − ρ)d

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 42
ROBERTSON’S AND P-D DIAGRAMS

w[j+1]

ρd
-a -1 0 1 k a

-rρd rw[5] rρd

-ad -d d kd ad rw[j]
w[6]

-ρd
Lk Uk

(a)

Figure 5.16: ROBERTSON’S DIAGRAM

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 43
rw[j] Ua=rρd

k>0
Uk
selection region
for qj+1=k Lk

U0

d
0
L0

Uk

k<0 Lk

(b) L-a=-rρd

Figure 5.16: P-D DIAGRAM (for d > 0).

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
 44
CONTINUITY CONDITION, OVERLAP, AND Q-DIGIT SELECTION

qj+1 = SEL(w[j], d)
• SEL represented by the set {sk }, −a ≤ k ≤ a,
qj+1 = k if sk ≤ rw[j] ≤ sk+1 − ulp
• sk defined as the minimum value of rw[j] for which qj+1 = k
• sk ’s are functions of the divisor d
• CONTAINMENT: Lk ≤ sk ≤ Uk
• CONTINUITY: qj+1 = k − 1 for rw[j] = sk − ulp ≤ Uk−1
Uk ≥ Uk−1 + ulp → Lk ≤ sk ≤ Uk−1 + ulp or Lk ≤ sk ≤ Uk−1
• OVERLAP
Uk−1 − Lk = (k − 1 + ρ)d − (k − ρ)d = (2ρ − 1)d
RESULTING IN
ρ ≥ 2−1
• REDUNDANCY IN q-DIGIT SET → OVERLAP BETWEEN SELECTION
INTERVALS - SIMPLER SELECTION
Digital Arithmetic – Ercegovac/Lang 2003 5 – Division
 45

w[j+1]

qj+1=k-1 qj+1=k

rw[j]

Lk Uk
k
Lk-1 Uk-1
k-1
overlap
k or
k-1
(a)

Figure 5.17: Overlap between selection intervals.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 46

rw[j]
Uk-1

sk

qj+1=k
Lk

overlap
(k or k-1)

qj+1=k-1

d
1/2 1
normalized divisor range
(b)
Figure 5.17: Selection function.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 47
SELECTION USING CONSTANTS

• USE CONSTANTS mk , INDEPENDENT OF DIVISOR

max(Lk ) ≤ mk ≤ min(Uk−1) + ulp
max and min for the range 2−1 ≤ d < 1
• For k > 0
(k − ρ) ≤ mk ≤ (k − 1 + ρ)2−1 + ulp
which requires
k+1
ρ≥
3
• For k ≤ 0
(k − ρ)2−1 ≤ k − 1 + ρ
which requires
(−k) + 2
ρ≥
3

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 48

rw[j] Uk-1

mk
Lk

k>0

1/2 1 d
k<0

Uk-1
mk

Lk

Figure 5.18: BOUNDS ON mk .

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 49
RADIX-2 DIVISION WITH NON-REDUNDANT RESIDUAL

• EXTENSION OF NON-RESTORING DIVISION: {-1,0,1} SRT

• ALLOWS SKIPPING OVER ZEROS
U1 = 2d L1 = 0
U0 = d ≥ 1/2 L0 = −d ≤ −1/2
U−1 = 0 L−1 = −2d
• SELECTION CONSTANTS:
0 ≤ m1 ≤ 1/2, −1/2 ≤ m0 ≤ 0
Choose: m1 = 1/2 and m0 = −1/2
• THE QUOTIENT-DIGIT SELECTION FUNCTION







1 if 1/2 ≤ 2w[j]

qj+1 =  0 if −1/2 ≤ 2w[j] < 1/2


 −1 if 2w[j] <

− 1/2

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 50
w[j+1]
d
-1 0 1

-2d 1/2 2w[j]

-1/2
2d
-d d

-d

L-1 U-1 L1 U1

L0 U0

-1 0 1

(a)

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 51
2w[j]
2

U1

qj+1=1
1
U0

1/2 m1=1/2

U-1 d
1/2 L1 1
qj+1=0

-1/2 m0=-1/2
L0
-1

L-1 qj+1=-1

-2

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 52
STAIRCASE SELECTION FUNCTION

• FOR r > 2, mk DEPENDS ON DIVISOR

• DIVIDE RANGE OF DIVISOR INTO INTERVALS [di, di+1) with

1
d0 = , di+1 = di + 2−δ
2
δ MS FRACTIONAL BITS OF DIVISOR REPRESENT THE INTERVAL

• FOR EACH INTERVAL, THERE IS A SET of selection constants mk (i)

for d ∈ [di, di+1), qj+1 = k if mk (i) ≤ rw[j] ≤ mk+1(i) − ulp

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 53
DEFINITION OF mk (i)

rw[j]

qj+1=k+1
mk+1(i) ulp

qj+1=k

mk(i)

2-δ
di di+1

Figure 5.20: DEFINITION OF mk (i).

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 54
RANGE CONDITION FOR SELECTION CONSTANTS

max(Lk (di), Lk (di+1)) ≤ mk (i) ≤ min(Uk−1(di), Uk−1(di+1)) + ulp

min(Uk-1(di), Uk-1(di+1))
rw[j]

region for mk(i)

Uk-1

Lk
max(Lk(di), Lk(di+1))

2-δ
di di+1

Figure 5.21: SELECTION CONSTANT REGION.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 55
LOW-PRECISION SELECTION CONSTANTS

mk (i) = Ak (i)2−c
WHERE Ak (i) IS INTEGER

• CAN USE TRUNCATED RESIDUAL IN COMPARISONS

WITH SELECTION CONSTANTS
• RESIDUAL MUST BE IN 2’s COMPLEMENT
• SELECTION CONDITIONS

for k > 0 Lk (di + 2−δ ) ≤ Ak (i)2−c ≤ Uk−1(di)

(5.1)
for k ≤ 0 Lk (di) ≤ Ak (i)2−c ≤ Uk−1(di + 2−δ ) + ulp

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 56
USE OF SELECTION CONSTANTS

rw[j]
Uk-1

2-c
mk(4) Lk

mk(2,3)

mk(1)
mk(0)

d
2-δ
d0 d1 d2 d3 d4 d5

Figure 5.22: QUOTIENT-DIGIT SELECTION WITH SELECTION CONSTANTS.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 57
SELECTION WITH TRUNCATED RESIDUAL AND DIVISOR

{rw [j ]}c
log2 r + c + δ
1 + log2 r + c

SELECTION {d }δ
FUNCTION δ−1

q j+1
Figure 5.23: SELECTION WITH TRUNCATED RESIDUAL AND DIVISOR.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 58
LOWER BOUND ON δ

• CONSIDER CASE k > 0 (similar argument for k ≤ 0);

• FROM CONTINUITY CONDITION
Uk−1(di) − Lk (di + 2−δ ) ≥ 0

(ρ + k − 1)di − (−ρ + k)(di + 2−δ ) ≥ 0

=⇒
(2ρ − 1)di ≥ (k − ρ)2−δ
• Worst case: d ≥ 1/2 and k ≤ a:
2ρ − 1 2ρ − 1
2−δ ≤ =
2(a − ρ) 2ρ(r − 2)
• MINIMUM VALUE OF δ CAN RESULT IN A LARGE VALUE OF c
• OPTIMIZE THE VALUES OF δ AND c TOGETHER

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 59
RADIX-4 DIVISION WITH NONREDUNDANT RESIDUAL

• Known as Robertson’s division

• qj ∈ {−2, −1, 0, 1, 2}
• Uk = ( 32 + k)d Lk = (− 23 + k)d
• BOUND ON δ
−δ 2ρ − 1 1
2 ≤ =
2(a − ρ) 8

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 60

4w[j]

U2

2
2
U1

L2

w[j+1] 1
2d/3 1
U0
0 1 2
L1
d/3 4d/3 0 1
4w[j] d
2d/3 5d/3 8d/3
d=1/2
U-1

L0
-2d/3 -1
(a) -1

U-2

L-1
-2
-2

L-2

(b)

Figure 5.24: ROBERTSON’S AND PD DIAGRAMS FOR RADIX-4 AND a = 2.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 61
+
[di, di+1) [8, 9) [9, 10) [10, 11) [11, 12)
L2(di+1), U1(di)# 36, 40 40, 45 44, 50 48, 55
m2(i)∗ 6 7 8 8
L1(di+1), U0(di)# 9,16 10, 18 11, 20 12, 22
m1(i) 2 2 2 2
L0(di), U−1(di+1)# -16, -9 -18, -10 -20, -11 -22, -12
m0(i) -2 -2 -2 -2
L−1(di), U−2(di+1)# -40, -36 -45, -40 -50, -44 -55, -48
m−1(i) -6 -7 -8 -8
[di, di+1)+ [12, 13) [13, 14) [14, 15) [15, 16)
L2(di+1), U1(di)# 52, 60 56, 65 60, 70 64, 75
m2(i) 10 10 10 12
L1(di+1), U0(di)# 13, 24 14, 26 15, 28 16, 30
m1(i) 4 4 4 4
L0(di), U−1(di+1)# -24, -13 -26, -14 -28, -15 -30, -16
m0(i) -4 -4 -4 -4
L−1(di), U−2(di+1)# -60, -52 -65, -56 -70, -60 -75, -64
m−1(i) -10 -10 -10 -12

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 62
QUOTIENT-DIGIT SELECTION FOR RADIX-4 DIVISION;
NON-REDUNDANT RESIDUAL

{4w[j]}3 U1
{4w[j]}3 {d}4
m2(7) 12/8
xxx.xxx 0.1xxx
11/8 L2
m2(4,5,6) 10/8 6 3

9/8 qj+1=2

m2(2,3) 1 SELECTION
m2(1) 7/8
qj+1=1 3
m2(0) 6/8

5/8 qj+1
{d}4
1/2 10/16 12/16 14/16 1 (b)
9/16 11/16 13/16 15/16

(a)

Figure 5.25: QUOTIENT-DIGIT SELECTION: (a) FRAGMENT OF THE P-D DIAGRAM. (b) IMPLEMENTATION.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 63
USE OF REDUNDANT ADDER

• SO FAR: COMPUTE rw[j] IN FULL PRECISION, TRUNCATE, AND COM-

PARE WITH LOW-PRECISION CONSTANTS
• FULL-PRECISION ADDITION: significant portion of the cycle time
• OVERLAP BETWEEN SELECTION INTERVALS
=⇒COULD USE AN ESTIMATE OF rw[j]
• ERROR IN ESTIMATE:

ǫmin ≤ y − yc ≤ ǫmax

• BASIC CONSTRAINT: if we choose qj+1 = k for an estimate yc then

y ∈ [yc + ǫmin , yc + ǫmax]

L∗k = Lk − ǫmin
Uk∗ = Uk − ǫmax

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 64
CONSTRAINTS ON SELECTION CONSTANTS

max(L∗k (di), L∗k (di+1)) ≤ mk (i) ≤ min(Uk−1

∗ ∗
(di), Uk−1 (di+1))

Uk-1
U*
k-1

region for mk (i)

ε max

L*k

Lk
−εmin

di d i+1

Figure 5.26: CONSTRAINTS FOR SELECTION BASED ON ESTIMATES.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 65
MINIMUM OVERLAP AND RANGE OF ESTIMATE

• OVERLAP
∗ ∗
min(Uk−1 (di), Uk−1 (di+1)) − max(L∗k (di), L∗k (di+1)) ≥ 0
• RANGE
|rw[j]| ≤ rρd < rρ (for d < 1)

−rρ − ǫmax < yc < rρ − ǫmin

y=rw[j]

εmax -εmin

y (estimate)
-rρ rρ

Figure 5.27: RANGE OF ESTIMATE

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 66
REDUNDANT ADDER

-q d rw[j] (redundant)
j+1

(a) REDUNDANT ADDER

w[j+1] (redundant)

t t
ws x x x. x x x x x x x x x x x. x x x x x x x x
rw[j] rw[j]
wc x x x. x x x x x x x x + + +. + + + + + + + +
- + +. + + + + + + + + - - -. - - - - - - - -

y y y. y y y 0 < ε < 2-t+1-ulp y y y. y y y -(2-t-ulp) < ε < 2-t-ulp

- + +. + + + (truncation error) + + +. + + + (truncation error)
y^ (estimate) - - - - - -
y^ (estimate)

(b) (c)

Figure 5.28: USE OF REDUNDANT ADDER: (a) Redundant adder. (b) Carry-save case. (c) Signed-digit case.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 67
CARRY-SAVE ADDER

• ERRORS

ǫmin = 0 ǫmax = 2−t+1 − ulp

• RESTRICTED SELECTION INTERVAL

Uk∗ = Uk − 2−t+1 + ulp

L∗k = Lk

d ∗ −t −t
U k−1 = ⌊Uk−1 + 2 ⌋t = ⌊Uk−1 − 2 ⌋t

c ∗
L k = ⌈Lk ⌉t = ⌈Lk ⌉t

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 68
y Uk-1
^
Uk-1 (represented by )
-t
2 U*
k-1
Case B

-t+1 Case A
2

^
L k (represented by )
L k L*k

di d i+1 d i+2

Case A: U*k-1 is on the grid;

Uk-1= U*k-1+2-t on the grid

Case B: U*k-1 is off the grid;

Uk-1> U*k-1 on the grid

Digital Arithmetic – Ercegovac/Lang 2003

Figure 5.29: Ub and L
b for residual in carry-save form. 5 – Division
 69
LOWER BOUND FOR t and δ

(for positive k)
d c
U k−1(di) − Lk (di+1) ≥ 0

Uk−1(di) − 2−t − Lk (di+1) ≥ 0

2ρ − 1
− (a − ρ)2−δ ≥ 2−t
2
• RANGE
⌊−rρ − 2−t⌋t ≤ yc ≤ ⌊rρ − ulp⌋t
⌊z⌋t = 2−t⌊2tz⌋

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 70
RADIX-2 DIVISION WITH CARRY-SAVE ADDER

1
− 0 × 2−δ ≥ 2−t
2
c c d d
max(L (d
k i ), L (d
k i+1 )) ≤ mk (i) ≤ min( U (d
k−1 i ), U k−1(di+1 ))

c
L 1 (1) = 0
d
U 0(1/2) = 0
c
L 0 (1/2) = −1/2
d
U −1(1) = −1/2

c d
(L 1(1) = 0) ≤ m1 ≤ (U0(1/2) = 0)
c d
(L 0 (1/2) = −1/2) ≤ m0 ≤ ( U −1(1) = −1/2)
This results in the selection constants m1 = 0 and m0 = −1/2

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 71
SELECTION FUNCTION

⌊−2 − 2−1⌋1 ≤ yc ≤ ⌊2 − ulp⌋1

5 c
− ≤ y ≤ 3/2
2








1 if 0 ≤ yc ≤ 3/2

qj+1 =  0 if yc = −1/2


 −1 if −5/2 ≤ y
 c ≤ −1

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 72
RADIX-2 WITH CSA: PD PLOT

2w[j]

U1
2 ^ x x x. x
y[j] x x x. x
3/2
U0
1 4+4
^
^
1/2 L1 U*
U0 q j+1=1 m(1)=0 0
0
SELECTION
q j+1=0 d FUNCTION
^ ^
L 1 , U-1
y[j] -1/2 U-1
^
L0 m(0)=-1/2
-1 U* q j+1
q j+1=-1 -1
-3/2 L 0 , L*0
(b)
-2
L -1
-5/2
d=1/2 d=1-ulp
(a)

Figure 5.30: RADIX-2 DIVISION WITH CARRY-SAVE ADDER: (a) P-D PLOT. (b) SELECTION FUNCTION.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 73
IMPLEMENTATION

qj+1 = (qs, qm)

qm = (p−1p0p1)′ (5.2)
qs = p−2 ⊕ (g−1 + p−1g0 + p−1p0g1)
where
pi = ci ⊕ si gi = ci · si
and
(c−2, c−1, c0, c1)

(s−2, s−1, s0, s1)

qj+1 = 0 (qs, qm) = (1, 0)

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 74
RADIX-4 ALGORITHM WITH CARRY-SAVE ADDER

1 4 −δ
− 2 ≥ 2−t
6 3

−t 1 1 1
2 ≤ − =
6 12 12

⌊−8/3 − 1/16⌋4 ≤ ŷ ≤ ⌊8/3 − ulp⌋4

44 42
− ≤ ŷ ≤
16 16

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 75
+
[di, di+1) [8, 9) [9, 10) [10, 11) [11, 12)
Lb 2(di+1), U
c
1 (di )
+
12, 12 14, 14 15, 15 16, 17
m2 (i)+ 12 14 15 16
Lb 1(di+1), U
c
0 (di )
+
3, 4 4, 5 4, 5 4, 6
m1 (i) 4 4 4 4
Lb 0(di), U
c
−1 (di+1)
+
-5, -4 -6, -5 -6, -5 -7, -5
m0 (i) -4 -6 -6 -6
Lb −1(di), U
c
−2 (di+1)
+
-13, -13 -15, -15 -16, -1 6 -18, -17
m−1(i) -13 -15 -16 -18
[di, di+1)+ [12, 13) [13, 14) [14, 15) [15, 16)
Lb 2(di+1), U
c
1 (di )
+
18, 19 19, 20 20, 22 22, 24
m2 (i) 18 20 20 24
Lb 1(di+1), U
c
0 (di )
+
4, 7 5, 7 5, 8 6, 9
m1 (i) 6 6 8 8
Lb 0(di), U
c
−1 (di+1)
+
-8, -6 -8, -6 -9, -6 -10, -7
m0 (i) -8 -8 -8 -8
Lb −1(di), U
c
−2 (di+1)
+
-20, -19 -21, -20 -23, -2 1 -25, -23
m−1(i) -20 -20 -22 -24

+: real value = shown value/16; Lb k = ⌈Lk ⌉4 , U

c
k = ⌊Uk −
1
16 ⌋4 .

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division

 76

^
y[j] U1
U^1

12/8 L^2

L2
11/8

10/8
x x x.x x x x 0.1 x x x
y^ ^
d
7 3
9/8

SELECTION
1 FUNCTION

q
j+1
7/8
(b)

6/8

5/8

1/2 9/16 10/16 11/16 12/16 13/16 14/16 15/16 1 ^

d

m2 (i)
region of choice for (a)
m2 (i)

Figure 5.31: SELECTION FUNCTION FOR RADIX-4 SCHEME WITH CARRY-SAVE ADDER: (a) Fragment of P-D diagram. (b)
Implementation.

Digital Arithmetic – Ercegovac/Lang 2003 5 – Division