1

DIGITAL ARITHMETIC

Miloš D. Ercegovac
Computer Science Department
University of California Los Angeles
and
Tomás Lang
Department of Electrical and Computer Engineering
University of California at Irvine
Viewgraphs Copyright c 2003 by the authors
Copyright c 2004 by Elsevier Science (USA)

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 2
ABOUT THE BOOK

Objectives Theory and design of arithmetic algorithms for digital systems

Topics – not all covered in depth –
1. Overview [Chapter 1]
2. Two-Operand Addition [Chapter 2]
3. Multi-operand Addition [Chapter 3]
4. Multiplication [Chapter 4]
5. Division by Digit Recurrence [Chapter 5]
6. Square Root by Digit Recurrence [Chapter 6]
7. Reciprocal, Division, Reciprocal Square RToot, and Inverse Square Root
by Iterative Approximation [Chapter 7]
8. Floating-point Representation, Algorithms, and Implementations [Chapter
8]
9. Digit-Serial Arithmetic [Chapter 9]
10. Function Evaluation [Chapter 10]
Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction
 3
11. CORDIC Algorithm and Implementations [Chapter 11]

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 4
Chapter 1: Review of the Basic Number Representations and Arithmetic
Algorithms

• General-purpose processors

3 Main use: numerical computations

3 Address calculations
- basic operations
- fixed point and floating point
- IEEE standard
- vector processors

• Special-purpose (application-specific) processors

- for numerically intensive applications
- single computation or classes of computations

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 5
Application-specific processor

• Areas of application:
- signal processing
- embedded systems
- matrix computations
- graphics, vision, multi-media
- cryptography and security
- robotics, instrumentation; others?
• Features:
- better use of technology
- improvement in speed, area, power
- flexibility in
* implementation; decomposition into modules
* number systems and data formats; algorithms
• Need good design tools; difficult to change; FPGAs?

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 6
GENERAL-PURPOSE VS. APPLICATION-SPECIFIC

• Flexibility

• Matching specific applications

• Use of VLSI and special technology

• Use of hardware-level parallel processing

• Lower software overhead

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 7
ARITHMETIC PROCESSORS: USER’S VIEW

AP = (operands, operation, results, conditions, singularities)

• Numerical operands and results specified by
– Set of numerical values x ∈ N (finite and ordered set)
– Range Vmin ≤ x ≤ Vmax
– Precision
– Number representation system (NRS)
• Set of operations: addition, subtraction, multiplication, division, ...
• Conditions: values of the results – zero, negative, etc.
• Singularities: Illegal results – overflow, underflow, Nan, etc.

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 8
LEVELS OF DESCRIPTION AND IMPLEMENTATION

• Numerical computations (applications)

• Algorithms
• Arithmetic operations

Operations Operands-Result Algorithm

Numerical function Numbers Numerical Algorithm
↓ ↓
↓ N umber System Arithmetic design
↓ ↓
Digit-vector function Digit vector Digit-vector Algorithm
↓ ↓
↓ Digit Coding RT L design
↓ ↓
Bit-vector function Bit vector Bit-vector Algorithm
↓
Logic Design

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 9
NUMBER REPRESENTATION SYSTEMS

value: x ∈ N =⇒NRS =⇒X ∈ V : digit vector

• Digit Vector

X = (xa, . . . , xi, . . . , xb)

– indexing:

Leftward Zero Origin (LZ) (integers)

X = (xn−1, xn−2, . . . , x0)

Rightward One Origin (RO) (fractions)

X = (x1, x2, . . . , xn)

– Digit set Di - set of values for digit xi (usually consecutive integers)

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 10
NUMBER REPRESENTATION (cont.)

• Number of (unambiguously) representable numbers

|N | ≤ Π|Di|

• Number representation system

F :N →V

• Choose NRS to
– allow efficient computation(s)
– suitable interface with other systems
• Different implementation-performance constraints
=⇒ variety of NRS

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 11
SOME CHARACTERISTICS OF NRS

a) Range: finite set of digit-vector values

b) Unambiguity: two numbers should not have same representation

If x ∈ N, y ∈ N, x 6= y then F (x) 6= F (y)

c) Nonredundant/redundant
Redundant: F −1(X) = F −1(Y )

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 12
WEIGHTED NUMBER REPRESENTATION SYSTEMS

Integer x represented by digit vector X = (xn−1 . . . , x0),

n−1
x= xi · w i
X

i=0
where
W = (wn−1, . . . , w0) weight vector
Define
R = (rn−1, . . . , r0) radix vector
so that
w0 = 1 wi = wi−1ri−1

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 13
WEIGHTED NUMBER REPRESENTATION (cont.)

• Fixed-radix NRS
ri = r
Then wi = ri so that
n−1
x= xi r i
X

i=0

• Canonical digit set

Di = {0, 1, 2, . . . , |ri| − 1}
• Conventional number system

– Fixed radix positive

– Canonical digit set

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 14
NON-CONVENTIONAL FIXED-RADIX SYSTEM

• Negative radix
n−1
r = −2, x = Pi=0 xi(−2)i
1011 = (-8) + 0 + (-2) + 1 = -9
0111 = 0 + 4 + (-2) + 1 = 3

• Complex radix
√
r = 2j, j = −1, xi ∈ {0, 1, 2, 3} (Knuth’s quarter imaginary NRS)
W: -8j -4 +2j +1
1231 =⇒1 × (−8j) + 2 × (−4) + 3 × (2j) + 1 × 1 = −7 − 2j
• Non-canonical digit set, r = 2, {-1,0,1} or {0,1,2}
Example: radix 4 D = {−3, −2, −1, 0, 1, 2, 3}

x = 27 represented by (1, 2, 3) or (2, −2, 3)

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 15
REDUNDANT NRS

• Fixed radix r
• Non-canonical digit set

D = {−a, −a + 1, . . . , −1, 0, 1, . . . , b − 1, b}

• Symmetric if a = b
• Redundant a + b + 1 > r (a, b ≤ r − 1)
– ”standard” a, b ≤ r − 1
– over-redundant a, b > r − 1
– Redundancy factor
a 1
ρ = , ρ>
r−1 2

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 16
EXAMPLES OF REDUNDANT DIGIT SETS

r a Digit set ρ Comment

2 1 { -1, 0, 1} 1 minimally/maximally redundant
4 2 {-2, -1, 0, 1, 2} 2/3 minimally redundant
4 3 {-3, -2, . . . , 2, 3} 1 maximally redundant
4 4 {-4, . . . , 4} 4/3 over-redundant
9 4 {-4, . . . , 4} 1/2 non-redundant
10 5 {-5, . . . , 5} 5/9 minimally redundant
10 6 {-6, . . . , 6} 2/3 redundant
10 9 {-9, . . . , 9} 1 maximally redundant
10 13 {-13, . . . , 13} 13/9 over-redundant

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 17
MIXED-RADIX NUMBER SYSTEM

• ri 6= rj
• Example: Representation of time R = (31, 24, 60, 60)
• Example: Factorial number system

ri = i + 2, i = 0, . . . , n − 1
R = (n + 1, n, . . . , 3, 2)
wi = (i + 1)!

Canonical digit set

Integers in range 0 ≤ x ≤ (n + 1)! − 1

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 18
NON-WEIGHTED NUMBER SYSTEMS: RESIDUE (RNS)

• Base vector B of moduli mi

B = (mn−1, mn−2, . . . , m0)

mi positive integers and pairwise relatively prime

• Integer x is represented by vector

X = (xn−1, xn−2, . . . , x0)

where xi = x mod mi
• Represents uniquely integers in the range
n−1
0 ≤ x < Πi=0 mi

(more later)

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 19
REPRESENTATION OF SIGNED INTEGERS

A. Directly in the number representation

Examples: Signed-Digit Number System
B. With an extra symbol: Sign-and-magnitude
C. Additional mapping on positive integers
Signed integers x
↓
Positive integers xR
↓
Digit-vectors X

Examples:
• True-and-Complement (TC):
– 2’s complement
– 1s’ complement
• Biased representation

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 20
TRUE-AND-COMPLEMENT SYSTEM

• −k ≤ x ≤ k signed integer (implicit value)

• xR positive integer (representation value)
• C – complementation constant
• Mapping xe = x mod C
• Unambiguous if k < C/2
• Equivalent to


 x if x ≥ 0
xR = 


C − |x| if x < 0
• Converse mapping


 xR if xR < C/2
x = 


xR − C if xR > C/2

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 21
NUMBER REPRESENTATION (cont.)

• xR represented in any number system

• In fixed-radix system two common choices:
– 2’s complement: C = r n (Range-complement system)
– 1s’ complement: C = r n − 1 (Diminished-radix-complement system)

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 22
BIASED REPRESENTATION

• B – bias
• −k ≤ x ≤ k
• xR = x + B
•B≥k

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 23
TYPES OF ARITHMETIC ALGORITHMS

• Bottom-up development
Primitives
+ Addition/subtraction
+ Multioperand addition
+ Arithmetic shifts
+ Multiplication by digit
+ Result-digit selection (PLA)
+ Table look-up
+ Multiplication
• Algorithms
+ Composition of primitives

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 24
TYPES (cont.)

• (Digit) Recurrences (continued sums)

3 Residual recurrence: R[i + 1] = f (R[i], X, Y, Z[i], zi+1)
Uses: Add/sub, single-position shifts, multiplication by digit
3 Output digit selection: zi+1 = g(R[i], X, Y, Z[i])
(keep R[i + 1] bounded)
Uses: Comparisons, PLA
3 Result recurrence
Z[i + 1] = Z[i] + zi+1ri+1 (continued sum)

Uses: Concatenation

• Examples:
3 multiplication R[i + 1] = 1r (R[i] + X · r nyi)
3 division R[i + 1] = rR[i] − qi+1Y qi+1 = g(R[i + 1], Y )

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 25
Types of algorithms

• Continued product recurrences

R[i + 1] = f (R[i], X, Y, Z[i], zz+1)
Uses: Add/sub, variable shifts, mult. by digit

zi+1 = g(R[i], X, Y, Z[i])(keep R[i + 1] bounded)

Uses: Comparisons, PLA
Z[i + 1] = Z[i](1 + zi+1r−(i+1)) (continued product)
Uses: Variable shift, addition

• Example:
3 division
R[i + 1] = rR[i](1 + qi+1r−(i+1)) + qi+1
qi+1 = g(R[i], Y )
Q[i + 1] = Q[i](1 + qi+1r−(i+1))

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 26
TYPES (cont.)

• Iterative Approximations
Z[i + 1] = f (Z[i], X, Y ) until g(Z(i)) < ε

Example:
3 reciprocal
Z[i + 1] = Z[i](2 − Z[i]X)

• Polynomial Approximations
z = a 0 + a 1 x + a 2 x2 + · · ·

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 27
PERFORMANCE

• Measures
+ Execution time
+ Throughput
• Improving speed
a) Arithmetic level
+ Reducing number of steps
Example: higher radix
Example: combinational instead of sequential
+ Reducing time of step
Example: carry-save adder instead of carry-propagate
+ Overlap steps (concurrency/pipelining)
Example: multiple generation and addition (in mult.)
Example: simultaneous additions (in mult.)
b) Implementation level
+ Reduce number of logic levels

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction

 28
POWER AND COST

• Measures
+ Packaging
+ Interconnection complexity
+ Number of pins
+ Number of chips and types of chips
+ Number of gates and types of gates
+ Area
+ Design cost; verification and testing cost
+ Power dissipation
+ Power consumption
• Reduction of cost

Digital Arithmetic - Ercegovac/Lang 2003 1 – Introduction