Paolo Numb Repr

Computer Arithmetic
Number Representation
Why Representation? Need to map numbers to binary bits

Need a representation anyway
Representation is consequential
Paolo.Ienne@epfl.ch EPFL I&C LAP Vojin.Oklobdzija@epfl.ch EPFL CSDA and UC Davis ACSEL
Complexity of arithmetic operations depends heavily on the representation
Number representation is the heart of computer arithmetic

2 CompArith Number Representation Ienne, Oklobdzjia 2004
What Is a Number System?

A number is represented as an ordered n-tuple of symbols (digit vector)
Rule of Interpretation
Mapping from set of digit vectors to numbers (e.g., integers, reals)
an 1an 2 ...a2a1a0
Each symbol is a digit
an 1 , an 2 ,K a2 , a1 , a0
Digits usually represent integers from a given sete.g.,
12
twelve
ai {0, 1} or ai {0, 1,..., 9} or ai { 1, 0, 1}

3 CompArith Number Representation Ienne, Oklobdzjia 2004 4
digit vectors
CompArith Number Representation
N, Z, R,
Ienne, Oklobdzjia 2004
Redundancy
Positional Weighted Systems

The rule of interpretation is a scalar product
n 1 i =0
A = an 1an 2 ...a2a1a0 = ai wi
where
W = (wn 1 , wn 2 ,...w2 , w1 , w0 )
digit vectors N, Z, R, digit vectors N, Z, R,
is the weight vector
Nonredundant
5 CompArith Number Representation
Redundant
Ienne, Oklobdzjia 2004 6 CompArith Number Representation Ienne, Oklobdzjia 2004
Radix Systems
Weights are not arbitrary but related to a radix vector
Mixed-Radix Systems
Fixed-radix if all elements of R are identical
R = (rn 1 , rn 2 ,...r2 , r1 , r0 )
in the following way
wi = r A = ai r i
i i =0
n 1
A few mixed-radix systems are very commone.g., time
w0 = 1
wi = wi 1 ri 1 = rj
j =0
i 1
R = (24, 60, 60) W = (86400, 3600, 60, 1)
Common Decimal Notation

It is weighted
Common Decimal Notation

It is nonredundant because canonical
A = an 1an 2 ...a2a1a0 = ai wi
i =0
n 1
ai {0, 1, ..., r 1}
10n possible digit vectors to represent 10n values
It is positional
wi depends only on i
It is fixed-radix Weighted, positional, fixed-radix, nonredundant also called conventional systems
wi = r i
10
Digit Set
Canonical set
Why Redundant Representations?

Remember paper & pencil addition?
ai {0, 1, ..., r 1}
A canonical system is nonredundant
an-1 bn-1 FA Cout sn-1 In binary:
an-2 bn-2 FA sn-2
a1 b1 FA s1
a0 b0 FA s0
Cin
Any other choice is noncanonical
ai { 1, 0, 1, 2}
ai { 1, 0, 1}
ai { ( r / 2 1),..., 0,..., r / 2 1, r / 2}
Slowcritical path contains n full-adders: nave delay is O(n) Must be performed right to left (LSB to MSB)
CompArith Number Representation Ienne, Oklobdzjia 2004
Redundancy May Simplify Some Operations

Radix-2, noncanonical {0, 1, 2, 3} , 2 bits per digit ai ai-1 bi bi-1 FA
digit ai 0 1 2 3 encoding a i 0 0 1 1 ai 0 1 0 1
Redundancy and MSB-first?

Now that the carry chain is broken, output digit i depends exclusively on input digits i, i 1, i2 Hence I can receive the operands in serial form and, when I have seen the first three input digits, I can produce the first output digit; afterwards, every new input digit, I can produce a new output digit Online arithmetic or MSDF digit-serial arithmetic
A
2n
B
2n
ai ai-1 bi bi-1 FA
ai-1 ai-2 bi-1 FA
+
2n
FA si si
Critical path now contains only two full-adders: delay is now O(1)
FA si si
13
14
Very Large Choice of Weighted Positional Number Systems
Binary Representation and Negative Numbers

Simplest signed system: Sign-and-magnitude
Source: Parhami, Oxford 2000
One bit represents the sign (typ. 1 neg.) Absolute value represented as unsigned
A = san 2 ...a2 a1a0 = (1) s ai 2i

i =0
n2
Exactly as humans do in decimal
A = an 2 ...a2a1a0 = ai 10i
i =0
n 2
Of course, we could also write 0 for + and 9 for

15 CompArith Number Representation Ienne, Oklobdzjia 2004 16 CompArith Number Representation Ienne, Oklobdzjia 2004
Sign-and-Magnitude Representation Some advantages and disadvantages

Familiar for users Simple nave multiplication Adders are not the most efficient Redundant zero (+0 and 0) may cause some problems (e.g. testing for a variable =0)
Negative Numbers As Result of Paper & Pencil Subtraction

Consider a normal paper-and-pencil subtraction
1 1 1 1
0 0 0 0 1 1 1
0 0 1 1
0 1 1 1
1 0 1 1
0 0 0 0
1 02 0 12 0 12 0 12
+20
1010 1710
27 +26 +25 +24 +23
710
17
18
Intuitive Derivation of the Twos Complement System

Twos complement representation
n2 i =0
Comparison of Binary Representations on 3 Bits

Decimal Binary (Unsigned) 111 110 101 100 011 010 001 000 011 010 001 000 / 100 101 110 111 011 010 001 000 111 110 101 100
A = an 1an 2 ...a2 a1a0 = an 1 2 Amax = 2 n 1 1
n 1
+ ai 2
Sign and Magnitude
Twos Complement
7 6 5 4 3 2 1 0 -1 -2 -3 -4
Largest (positive): all ones except an-1 = 0
Smallest (negative): all zeros except an-1 = 1
Amin = 2
n 1
Asymmetric range, non-redundant Standard in computers for signed integers nowadays

Twos Complement Representation of Negative Numbers

It shifts the second half of the unsigned range into the negative by 2n E.g., on 8 bits:
N
True-and-Complement Systems for Signed Numbers

Transform signed numbers in unsigned and then use conventional systems
00000000 0 10000000 -128
01111111 10000000 127 128 01111111 127
11111111 255 minus two six
(unsigned)
110
11111111 00000000 -1 0
(signed)
AR
that is, A is represented as the unsigned integer 2n A

Z (signed)
N (unsigned)
digit vectors
Mapping from Signed to Unsigned

A number A is represented through the unsigned number AR defined through a complementation constant
Inverse Mapping
Inversely
AR if AR < C / 2 A= AR C if AR > C / 2
AR = A mod C
This is equivalent (for |A| < C) to
True forms
A if A 0 AR = C A = C + A if A < 0
For the representation to be unambiguous, it must be
Complement forms
If AR = C / 2 can be represented, it can be assigned either to A = C / 2 or to A = C / 2, but the representation is then asymmetric ( system not close under sign change operation) If AR = C is can be represented, there are two representations for zero
A < C/2
or at most equal
Range Complement and Digit Complement Systems

Two particular cases of true-and-complement systems
Comparison of More Binary Representations on 3 Bits

Decimal Binary (Unsigned) 111 110 101 100 011 010 001 000 011 010 001 000 / 100 101 110 111 011 010 001 000 111 110 101 100
Range complement: C = r n Digit complement: C = r n 1
Sign and Magnitude
Twos Complement
Ones Complement
7 6 5 4 3 2 1 0 -1
In radix-2
Twos complement: C = 2n Ones complement: C = 2n 1
011 010 001 000 / 111 110 101 100
Twos complement is asymmetric and nonredundant Ones complement is symmetric with a redundant zero
-2 -3 -4
Equivalence of the Two Twos Complement Definitions

Using the true-and-complement definition, if AR < C / 2
Similar Property for Ones Complement

Using the true-and-complement definition, if AR < C / 2
A = AR an 1 = 0
If AR C / 2
A = ai 2i
i =0
n 2
A = AR an 1 = 0
If AR C / 2
n 2 i =0
A = ai 2i
i =0
n 2
A = AR C an 1 = 1
A = 2n + ai 2i = 2n + 2n 1 + ai 2i =
i =0
n 1
A = AR C an 1 = 1
A = (2n 1) + ai 2i = 2n + 2n 1 + 1 + ai 2i =
i =0 i =0
n 1
n 2
= 2n 1 + ai 2i
i =0
n 2
= 2n 1 + 1 + ai 2i
i =0
n 2
Hence the property
Hence the property

n 1
A = an 1 2
27
+ ai 2
i =0
n 2
A = an 1 (2
Ienne, Oklobdzjia 2004 28
n 1
+ 1) + ai 2i
i =0
n 2
Sign
0 if A 0 sign( A) = 1 if A 0
For sign-and-magnitude is clearly trivial For true-and-complement too, it is
Implied Digits in the Unsigned Representation

Unsigned numbers (like naturals in decimal) have infinite leading zeros in front of them:
1 23410 00 001 23410 1 01012 0 0000 0001 01012

Changing the representation from n bits to m bits (m > n) is just a matter of padding the number with leading zeros
0 if AR < C / 2, that is if an 1 = 0 sign( A) = 1 if AR C / 2, that is if an 1 = 1

For all these representations
Sign bit
sign( A) = an 1
Changing the Number of Bits in Ones and Twos Complement

Signed numbers can be thought as having infinite replicas of the MSB/sign in front of them
Sign Extension and Shift Right

Twos complement representation on n bits
(n)
= an 1an 2 ...a2 a1a0 = an 1 2
n 1
12
+20
+ ai 2i
i =0
n2
4 bits
23 +22
310
8 bits
Sign extended twos complement representation on m bits (m > n)
A( m ) = an 1...an 1an 1an 2 ...a2 a1a0 = = an 1 2

negative #)
m 1
1
27
1
+26
1
+25
1
+24
1
+23
1
+22
12
+20
310
m2
j = n 1
n 1
2 + ai 2i
j i =0
n2
Minus three on 4 bits is not represented as minus three on 8 bits!

To extend from n to m bits, one adds (for positive #) zero or (for
( 2
m 1
+ 2 m 2 + ... + 2 n 1 2 n 1 0
) (
32
Addition in True-and-Complement
If there is no overflow (that is, we assume the result of the sum to be representable), S = A + B can be simply computed as
Twos Complement Addition

Particularly comfortable in twos complement, because C = 2n and performing mod 2n implies just to ignore the carry out bit
S R = ( AR + BR ) mod C
Easy to prove using (X mod Y) mod Y = X mod Y
Cout
Cin
( AR + BR ) mod C = ( A mod C + B mod C ) mod C = = ( A + B ) mod C = S mod C = S R

S
Remember the intuitive derivation of the twos complement representation
Ones Complement Addition

A little more complex in ones complement, because C = r n 1 and the modulo of WR = AR + BR is less easy
wn = 0 and WR mod (r n 1) = WR wn = 0 and WR mod (r n 1) = 0 if AR + BR < r n 1 if AR + BR = r n 1 if r n 1 < AR + BR 2(r n 1)
Ones Complement Addition

End-around carry A B
wn = 1 and WR mod (r n 1) = WR r n + 1
Cout
Hence
If wn = 0 S R = WR If wn = 1 S R = WR + 1, subtracting r n = 2n simply by ignoring wn
Cin
S
Negation
Bitwise negation
Complement and Negation in True-and-Complement

Fundamental property in n-digit r-radix systems
A = not( A) = an 1 an 2 K a2 a1 a0
where, as usual in binary,
because an 1 an 2 an 1 an 2 r 1 r 1 1
38
A + A + 1 = rn
a = ( r 1) a
Example in radix-2 on 4 bits:
a2 a2
a1 a1
a0 a0 1
+ = + =
A A 1 rn
r 1 r 1 r 1 0 0 0
610 = not(610 ) = 0 1 1 0 = 1001 = 10012 = 910

Change of Sign
In ones complement
Change of Sign in Twos Complement

Another way of verifying the property
A= A
A = r n 1 A = A if A 0 A = r n 1 + A = r n ( r n 1 + A) 1 = A if A < 0
A + A = 1
Use twos complement property/definition
n2 n2 n 1 i n 1 an 1 2 + ai 2 + an 1 2 + ai 2i = i =0 i =0
In twos complement
A = A +1
A = r n A = A + 1 if A 0 A + 1 = r n + A + 1 = r n ( r n + A) 1 + 1 = A if A < 0
= an 1 + an 1 2 n 1 + ai + ai 2i = 2 n 1 + 2i = 1
i =0 i =0
n2
n2
Twos Complement Subtracter

A B = A + ( B ) = A + B + 1
A B
Twos Complement Adder/Subtracter

A B
Sub
Cout
Cin
Overflow?!
Cout
Cin
S
S
Overflows for Unsigned Addition

Limited range, thus sum of two numbers can exceed the range 1 + 1 1 1 1 1 0 0 0 0 0 0 1 1 1 0 0 0 0 0 12 0 02 0 12 20110 20010 40110
(>25510)
Overflows for Signed Addition

The situation is somehow different and the same simple rule does not hold anymore 1 + 1 1 1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0 0 0 12 02 12 12 -5510 -5610 -11110 -11110
Overflows if carry out is 1that is, if the first bit which cannot be stored is not null
Carry out is 1, but now it is ok because the result is negative (think of sign extension)
Overflows for Signed Addition

Overflows makes sum of positives look like negative or sum of negatives look like positive
9010 = 0101 10102
Overflows Addition and Subtraction

No overflow possible if adding operands of different sign Overflow only if
Op A 0 <0 0 <0 B 0 <0 <0 0 Result <0 0 <0 0 A+B A+B AB AB
10510 = 0110 10012 19510 = 1100 00112 -6110 = 1100 00112

10000000 11111111 00000000 11111111 10000000 11111111
Hence
OVF+ = an 1 bn 1 sn 1 + an 1 bn 1 sn 1 OVF = an 1 bn 1 sn 1 + an 1 bn 1 sn 1
-128
-1 0
127 128
255
45
Overflows Addition
Considering all possible cases around the last FA
an 1 0 0 0 0 1 1 1 1 bn 1 0 0 1 1 0 0 1 1 cn 1 0 1 0 1 0 1 0 1 cn 0 0 0 1 0 1 1 1 sn 1 0 1 1 0 1 0 0 1 Overflow No Yes No No No No Yes No
Fixed Point
In real life, one does not only need integers If one adds a fractional point in a fixed position, hardware for integers works just as well 0 1 0 0 12 + 0 0 0 1 12 0 1 1 0 02
24 23 22 21 20
910 310 1210
0 1. 0 0 12 + 0 0. 0 1 12 0 1. 1 0 02
21 20 2-1 2-2 2-3
1.12510 0.37510 1.50010
OVF+ = cn cn 1
Its just a matter of representation!

Fixed Point Requires Attention from Programmers Multiplication often introduces the need of arithmetic shifts to the right
0. 1 0 1 02 0. 0 1 1 02 0. 0 0 1 1 1 1 0 02

References
M. D. Ercegovac and T. Lang, Digital Arithmetic, Morgan Kaufmann, 2004 I. Koren, Computer Arithmetic Algorithms, Peters, 2002 A. R. Omondi, Computer Arithmetic Systems Algorithms, Architecture and Implementation, Prentice Hall, 1991 B. Parhami, Computer ArithmeticAlgorithms and Hardware Designs, Oxford, 2000
0.62510 0.37510 0.23437510 0.187510
0. 0 0 1 12
Note the low accuracy!


Paolo Numb Repr

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Paolo Numb Repr

Загружено:

Авторское право:

Доступные форматы

Computer Arithmetic

Why Representation? Need to map numbers to binary bits

Complexity of arithmetic operations depends heavily on the representation

Number representation is the heart of computer arithmetic

What Is a Number System?

ai {0, 1} or ai {0, 1,..., 9} or ai { 1, 0, 1}

Positional Weighted Systems

is the weight vector

A few mixed-radix systems are very commone.g., time

R = (24, 60, 60) W = (86400, 3600, 60, 1)

CompArith Number Representation

Ienne, Oklobdzjia 2004

CompArith Number Representation

Ienne, Oklobdzjia 2004

Common Decimal Notation

Common Decimal Notation

CompArith Number Representation

Ienne, Oklobdzjia 2004

Why Redundant Representations?

an-1 bn-1 FA Cout sn-1 In binary:

an-2 bn-2 FA sn-2

Any other choice is noncanonical

Redundancy May Simplify Some Operations

Redundancy and MSB-first?

ai-1 ai-2 bi-1 FA

CompArith Number Representation

Ienne, Oklobdzjia 2004

CompArith Number Representation

Very Large Choice of Weighted Positional Number Systems

Binary Representation and Negative Numbers

A = san 2 ...a2 a1a0 = (1) s ai 2i

Exactly as humans do in decimal

Of course, we could also write 0 for + and 9 for

Sign-and-Magnitude Representation Some advantages and disadvantages

Negative Numbers As Result of Paper & Pencil Subtraction

27 +26 +25 +24 +23

CompArith Number Representation

Ienne, Oklobdzjia 2004

CompArith Number Representation

Ienne, Oklobdzjia 2004

Intuitive Derivation of the Twos Complement System

Comparison of Binary Representations on 3 Bits

A = an 1an 2 ...a2 a1a0 = an 1 2 Amax = 2 n 1 1

Sign and Magnitude

Largest (positive): all ones except an-1 = 0

Smallest (negative): all zeros except an-1 = 1

Asymmetric range, non-redundant Standard in computers for signed integers nowadays

Twos Complement Representation of Negative Numbers

True-and-Complement Systems for Signed Numbers

00000000 0 10000000 -128

01111111 10000000 127 128 01111111 127

11111111 255 minus two six

that is, A is represented as the unsigned integer 2n A

Mapping from Signed to Unsigned

Range Complement and Digit Complement Systems

Comparison of More Binary Representations on 3 Bits

Range complement: C = r n Digit complement: C = r n 1

Sign and Magnitude

011 010 001 000 / 111 110 101 100

Equivalence of the Two Twos Complement Definitions

Similar Property for Ones Complement

Hence the property

Hence the property