Binary and Floating Point

Arithmetic
What is covered
Appendix A: Binary Numbers (P 631-643)
Appendix B: Floating-Point Numbers (P 643-653)

Lecture II: Binary Numbers 1

 Contents
• Finite Precision Numbers
• Radix Number System
• Conversion from One Radix to Another
• Negative Binary Numbers
• Binary Arithmetic
• Principles of Floating Point Numbers
• IEEE Floating Point Standard 754
Lecture II: Binary Numbers 2
 Finite Precision Numbers
• In arithmetic, no attention is paid to the amount of
memory taken to store numbers.
• Computers have finite memory, hence memory
matters – need a representation
• Algebra differs in finite precision arithmetic.
– Closure violated due to overflow and underflow
– Density lost in case of real and rational numbers

Lecture II: Binary Numbers 3

 Radix Number System
• Base 10 Example

Lecture II: Binary Numbers 4

• Requires k different symbols to represent
digits digits 0 through k-1.
• Decimal Numbers: (Base 10)
– Built from 0,1,2,3,4,5,6,7,8,9
• Binary Numbers: (Base 2)
– Built from 0,1
• Octal numbers: (Base 8)
– Built from 0,1,2,3,4,5,6,7
• Hexadecimal numbers: (Base 16)
– Built from 0,1,2,3,4,5,6,7,8,9,A,B,C,D,E,F
Lecture II: Binary Numbers 5
Examples

Lecture II: Binary Numbers 6

Lecture II: Binary Numbers 7
 Conversion from One Radix to Another
• Octal <-> Hexadecimal <-> Binary
– Easy
– Binary to Octal:
• Divide into groups of three bits –starting from immediate left (and right) of
the decimal point
• Convert each group into octal from 0 to 7.
• Example:
» Binary 1010111.011
» Octal 127.3
• May be necessary to add leading or trailing zeros to form groups of three
– Hexadecimal <-> Octal, Binary – Same procedure

Lecture II: Binary Numbers 8

More Examples

Lecture II: Binary Numbers 9

Decimal to Binary: First Method
• Take the largest power of 2 that divides the
number and subtract that amount
• Repeat process
• Example: Decimal 22
– Largest exponent of 2 is 4. I.e. 16
– 22 –16 = 6
– 6 = 4+2
– Hence Binary number is 10110
– I.e. 2**4 + 2**2 + 2**1

Lecture II: Binary Numbers 10

 Decimal to Binary
Second Method (Integers only)
• Divide the number by 2
– Quotient written directly under original number
– Remainder written next to quotient
• Repeat process until 0 is reached
• Directly use Euclidean Algorithm

Lecture II: Binary Numbers 11

Example

Lecture II: Binary Numbers 12

 Binary to Decimal
• Method 1- Sum up powers of 2
– Example: 10110 is 2**4 + 2**2 + 2**1 = 22
• Method 2: Successive doubling
– Write the binary number vertically
– Lines are numbered from bottom to top.
– Put 1 on line 1 (bottom)
– Entry on line n = 2*(entry on line n-1) + bit on
line n (0 or 1)

Lecture II: Binary Numbers 13

Example

Lecture II: Binary Numbers 14

 Negative Binary Numbers
• Signed Magnitude: Leftmost bit has sign
(0 is +, 1 is -)
• One’s Complement: To negate a number
switch all 0’s and 1’s (including sign bit)
Obsolete now.
– Example: 00000110 (+6)
– (-6 in ones complement) 11111001

Lecture II: Binary Numbers 15

 Negative Binary Numbers
• Two’s Complement: (Has sign bit 0 for + and 1
for -)
• Negating a number:
– Switch 0’s and 1’s (as in one’s complement)
– Add 1 to the result
• Example 00000110 (+6)
– (-6 in one;’s complement) 11111001
– (-6 in two’s complement) 11111010
– Any carry over from leftmost bit is thrown away

Lecture II: Binary Numbers 16

 Negative Binary Numbers
• Excess 2**(m-1): Represents a number
by (number) + 2**(m-1)
• Example: For 8-bit numbers –3 is
–3+127 = 125 in binary is 01111101
• Numbers from –128 to +127 mapped to 0 to
255, expressible as 8 bit positive integers
• Identical to two’s complement with sign bit
reversed
Lecture II: Binary Numbers 17
Lecture II: Binary Numbers 18
 Some Properties of Negative Numbers

• Signed magnitude and one’s complement has two

representatives for zero (+0 and –0)
• Two’s complement has only one zero
• Problem with two’s complement
negative of 100000 = 100000
• What is desired:
– Only one representation for zero
– Exactly the same number of + and – numbers
– Need to have an odd count if this is to be achieved.

Lecture II: Binary Numbers 19

 Binary Arithmetic
• Two binary (octal, hexadecimal) numbers
can be added just as decimal numbers
• The carry need to be taken to the next
position (left)
• In one’s complement, carry generated by
the leftmost bit is added to the rightmost bit.
• In two’s complement, carry generated by
the leftmost bit is thrown away
Lecture II: Binary Numbers 20
Binary Arithmetic

Lecture II: Binary Numbers 21

Examples of Binary Arithmetic

Lecture II: Binary Numbers 22

 Overflow
• If addend and augend are of opposite sides
no overflow occurs
• If both are of same sign, result is opposite
sign, overflow occurs
• In both one’s and two’s complement
overflow occurs iff carry into sign bit
differs from carry out of sign bit
• Most machines have overflow bit
Lecture II: Binary Numbers 23
 Floating Point Arithmetic
• Two Important issues in representing real (floating
point) numbers:
– Range: The length of the interval
– Precision: What small differences can be shown
• N = f* (10**e)
– Fraction = f: Number of digits here determines
precision
– Exponent = e : Determines the range
• Example: 3.14 = 0.314 * (10**1)
Lecture II: Binary Numbers 24
Two digit exponent and signed three
digit fraction
• Range +0.100*(10**-99) to +.999*(10+99)

Lecture II: Binary Numbers 25

 Overflow, Underflow
• Regions 1 and 7 represents overflows – answer
incorrect
• Regions 3 and 5 represents underflow errors –less
serious than overflow errors
• Problems with Floating Point Numbers:
– No density
– Do not form a continumm
– Spacing between two consecutive numbers not constant
throughout regions. I.e separation between
– +0.998* (10**98) and +0.999* (10**98) vastly
different from +0.998* (10**0) and +0.999* (10**0)
Lecture II: Binary Numbers 26
Lecture II: Binary Numbers 27
 Normalized Digits
• Shifting the number of digits between exponents
and fraction shifts boundaries of regions 2 and 6.
• Increasing number of digits in fraction increases
density – thus improves accuracy
• Increasing size of exponent increases regions 2 and
6, by shrinking others.
• In computers:
– Base 2,4, 8 or 16
– If leftmost digit is zero shift one place left and decrease
exponent by 1
– A fraction with nonzero leftmost digit is said to be
normalized.
Lecture II: Binary Numbers 28
 Normalized Digits
• Normalized Digits:
– There is only one normalized expression,
whereas there can be more than one non-
normalized floating point expressions.
• Example next slide

Lecture II: Binary Numbers 29

Lecture II: Binary Numbers 30
 IEEE Standard 754
• Designed by William Kahan of UCB
• Three representations:
– Single Precision: 32 bits
– Double Precision: 64 bits
– Extended Precision: 80 bits
• Both Single and Double precision uses
– Radix2 for fractions end excess notation for exponents
– Starts with sign bits (0 for +, 1 for -)

Lecture II: Binary Numbers 31

Lecture II: Binary Numbers 32
 Normal Fractions, Examples
• Normal Fraction:
– Decimal point,
– 1
– Other numbers
– Omit 1 to begin with (implied)
– Omit binary point (implied)
– Either 23 or 52 fraction bits: If all 0 then fraction 1.0
– If all 1, then fraction taken to be 2.0 (slightly less)
• Example
– 0.5 = 3F000000
– 1 = 3F800000
– 1.5 3FC00000
Lecture II: Binary Numbers 33
 Dealing with Underflow I
• If calculation results in a number smaller
than the smallest representable:Use
Denormalized Numbers
• Have exponent zero, and fraction given
by 23 or 52 bits
• These have bit left of decimal as 0
• The smallest denormalized number has 1 as
exponent and 0 as fraction = 1.0 *(2**-127)
• The largest denormalized number has 0 as
exponent and all 1’s = 1.0 *(2**-127)
Lecture II: Binary Numbers 34
 Dealing with Underflow II
• As numbers go further down, first few bits
become zero:
– exponent represents 2**(-127), and
– fraction representing 2**(-23)
– So the number represents 2**(-150)
• Gives graceful underfull without jumping to
zero
• Two zeros are present, with fraction 0 and
exponent 0
Lecture II: Binary Numbers 35
 Dealing with Overflow
• There are no bits to represent overflow
• Represent infinity with
– exponent with all 1’s
– Fraction 0
– Not a normalized number
• Behaves like mathematical infinity
• Infinity/Infinity cannot be determined,
represented by NaN
Lecture II: Binary Numbers 36
Lecture II: Binary Numbers 37
Lecture II: Binary Numbers 38
 Error Correcting Codes
• Due to occasional errors of voltage
fluctuations, can cause bit errors
• To prevent these errors, some memories and
networks use error correcting and error
detecting codes.
• Memory word of n bits has m data bits and r
check bits for error checking/correcting

Lecture II: Binary Numbers 39

 Hamming Distance
• How many bits differ: Take exclusive or:
• Example 10001001 and 10110001, EXOR
reveals three bit error.
• Hamming Distance: The distance in which
two code words differ.
• If two code words are a hamming distance
apart, it will take d single bit errors to
convert one to the other.
Lecture II: Binary Numbers 40
 Correcting vs. Detecting
• Error Detecting Codes: To detect d single bit errors
need distance d+1 apart codes
• To correct d single bit errors, need 2d+1 distance
apart code words.
• Parity Bit: Takes two errors to go from one correct
word to another.
• M data and r check bits allows 2**m legal memory
words, has n illegal word with distance 1, thus each
word has n+1 words dedicated to it. Since total bit
patters is 2**n So we get
m+r+1 < 2**r, Given m we get a lower bound for
r.
Lecture II: Binary Numbers 41
Lecture II: Binary Numbers 42