Carnegie Mellon

Binary Floating Point

COMP 222: Introduction to Computer Organization

Instructor:
Alan L. Cox

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition

 Carnegie Mello

Binary Floating Point

 Background: Fractional binary numbers
 IEEE floating point standard: Definition

 Example and properties

 Rounding, addition, multiplication

 Alternative formats

 Floating point in C

 Summary

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 2

 Carnegie Mello

Fractional Binary Numbers

 What is 1011.1012?

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 3

 Carnegie Mello

Fractional Binary Numbers

2i
2i-1

4
••• 2
1

bi-
bi ••• b2 b1 b0 b-1 b-2 b-3 ••• b-j
1 1/2
1/4 •••
1/8

 Representation 2-j
 Bits to right of “binary point” represent fractional powers of 2
 Represents rational number:

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 4

 Carnegie Mello

Fractional Binary Numbers:

Examples
 Value Representation
5 3/4 101.112
2 7/8 010.1112
1 7/16 001.01112

 Observations
 Divide by 2 by shifting right (unsigned)
 Multiply by 2 by shifting left
 Numbers of form 0.111111…2 are just below 1.0
 1/2 + 1/4 + 1/8 + … + 1/2i + … ➙ 1.0
 Use notation 1.0 – ε

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 5

 Carnegie Mello

Representable Numbers
 Limitation #1
 Can only exactly represent numbers of the form x/2k
 Other rational numbers have repeating bit representations
 Value Representation
 1/3 0.0101010101[01]…2
 1/5 0.001100110011[0011]…2
 1/10 0.0001100110011[0011]…2

 Limitation #2
 Just one setting of binary point within the w bits
 Limited range of numbers (very small values? very large?)

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 6

 Carnegie Mello

Binary Floating Point

 Background: Fractional binary numbers
 IEEE floating point standard: Definition

 Example and properties

 Rounding, addition, multiplication

 Alternative formats

 Floating point in C

 Summary

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 7

 Carnegie Mello

IEEE Floating Point

 IEEE Standard 754
 Established in 1985 as uniform standard for binary floating point arithmetic

Before that, …
Many idiosyncratic formats
Different results from the same program on different machines
 Major update in 2008, and minor revisions in 2019
Add decimal floating point for financial applications
 Supported by all major CPUs (and increasingly GPUs)
 Driven by numerical concerns
 Nice standards for rounding, overflow, underflow
 Hard to make fast in hardware
 Numerical analysts predominated over hardware designers in defining
standard

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 8

 Carnegie Mello

Binary Floating Point

Representation
 Numerical Form:
(–1)s M 2E
 Sign bit s determines whether number is negative or positive
 Significand M normally a fractional value in range [1.0,2.0).
 Exponent E weights value by power of two

 Encoding
 MSB s is sign bit s
 exp field encodes E (but is not equal to E)
 frac field encodes M (but is not equal to M)

s exp frac

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 9

 Carnegie Mello

Basic Formats (2008)

 binary32 (Single precision)
s exp frac
1 8-bits 23-bits
 binary64 (Double precision)
s exp frac
1 11-bits 52-bits
 binary128 (Quadruple precision)
s exp frac
1 15-bits 112-bits

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 10

 Carnegie Mello

“Normalized” Values v = (–1)s M 2E

 When: exp ≠ 000…0 and exp ≠ 111…1

 Exponent coded as a biased value: E = Exp – Bias

 Exp: unsigned value of exp field
 Bias = 2k-1 - 1, where k is number of exponent bits
 Single precision: 127 (Exp: 1…254, E: -126…127)
 Double precision: 1023 (Exp: 1…2046, E: -1022…1023)

 Significand coded with implied leading 1: M =

1.xxx…x2
 xxx…x: bits of frac field
 Minimum when frac=000…0 (M = 1.0)
 Maximum when frac=111…1 (M = 2.0 – ε)
 GetComputer
Bryant and O’Hallaron, extraSystems:
leading bit for
A Programmer’s “free”
Perspective, Third Edition 11
 Carnegie Mello

Normalized Encoding Example

v = (–1) s
M 2 E

E = Exp – Bias
 Value: float F = 15213.0;
 15213 = 11101101101101
10 2

= 1.1101101101101 x 2 2
13

 Significand
 M = 1.1101101101101 2

 frac = 11011011011010000000000 2

 Exponent
 E = 13
 Bias = 127
 Exp = 140 = 10001100 2

0 10001100 11011011011010000000000
s exp frac
Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 12
 Carnegie Mello

Denormalized (Subnormal)
v = (–1) M 2 s E

Values E = 1 – Bias
 Condition: exp = 000…0

 Exponent value: E = 1 – Bias (instead of E = 0 –

Bias)
 Significand coded with implied leading 0: M =

0.xxx…x2
 xxx…x: bits of frac
 Cases
 exp = 000…0, frac = 000…0
 Represents zero value
 Note distinct values: +0 and –0 (why?)
 exp = 000…0, frac ≠ 000…0
 Numbers closest to 0.0
 Equispaced
Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 13
 Carnegie Mello

Special Values
 Condition: exp = 111…1

 Case: exp = 111…1, frac = 000…0

 Represents value  (infinity)
 Operation that overflows
 Both positive and negative
 E.g., 1.0/0.0 = −1.0/−0.0 = +, 1.0/−0.0 = −

 Case: exp = 111…1, frac ≠ 000…0

 Not-a-Number (NaN)
 Represents case when no numeric value can be determined
 E.g., sqrt(–1),  − ,   0

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 14

 Carnegie Mello

Visualization: Floating Point Encodings

− +
−Normalized −Denorm +Denorm +Normalized

NaN NaN
0 +0

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 15

 Carnegie Mello

Binary Floating Point

 Background: Fractional binary numbers
 IEEE floating point standard: Definition

 Example and properties

 Rounding, addition, multiplication

 Alternative formats

 Floating point in C

 Summary

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 16

 Carnegie Mello

Tiny Floating Point Example

s exp frac
1 4-bits 3-bits

 8-bit Floating Point Representation

 the sign bit is in the most significant bit
 the next four bits are the exponent, with a bias of 7
 the last three bits are the frac

 Same general form as IEEE Format

 normalized, denormalized (subnormal)
 representation of 0, NaN, infinity

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 17

 Carnegie Mello

Dynamic Range (Positivev =Only)

(–1) M 2 s E

s exp frac E Value n: E = Exp –

0 0000 000 -6 0 Bias
0 0000 001 -6 1/8*1/64 = 1/512 d:closest
E=1 to –zero
Bias
Denormalized0 0000 010 -6 2/8*1/64 = 2/512
numbers …
0 0000 110 -6 6/8*1/64 = 6/512
0 0000 111 -6 7/8*1/64 = 7/512 largest denorm
0 0001 000 -6 8/8*1/64 = 8/512
smallest norm
0 0001 001 -6 9/8*1/64 = 9/512
…
0 0110 110 -1 14/8*1/2 = 14/16
0 0110 111 -1 15/8*1/2 = 15/16 closest to 1 below
Normalized 0 0111 000 0 8/8*1 = 1
numbers 0 0111 001 0 9/8*1 = 9/8 closest to 1 above
0 0111 010 0 10/8*1 = 10/8
…
0 1110 110 7 14/8*128 = 224
0 1110 111 7 15/8*128 = 240 largest norm
0 1111 000 n/a inf

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 18

 Carnegie Mello

Distribution of Values
 6-bit IEEE-like format
 e = 3 exponent bits
 f = 2 fraction bits s exp frac
 Bias is 23-1-1 = 3 1 3-bits 2-bits

 Notice how the distribution gets denser toward zero.

8 values

-15 -10 -5 0 5 10 15
Denormalized Normalized Infinity

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 19

 Carnegie Mello

Distribution of Values (close-up

view)
 6-bit IEEE-like format
 e = 3 exponent bits
 f = 2 fraction bits s exp frac
 Bias is 3 1 3-bits 2-bits

-1 -0.5 0 0.5 1
Denormalized Normalized Infinity

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 20

 Carnegie Mello

Special Properties of the IEEE

Encoding
 FP Zero Same as Integer Zero
 All bits = 0

 Can (Almost) Use Unsigned Integer Comparison

 Must first compare sign bits
 Must consider −0 = 0
 NaNs problematic

Will be greater than any other values
 What should comparison yield?
 Otherwise OK
 Denorm vs. normalized
 Normalized vs. infinity

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 21

 Carnegie Mello

Binary Floating Point

 Background: Fractional binary numbers
 IEEE floating point standard: Definition

 Example and properties

 Rounding, addition, multiplication

 Alternative formats

 Floating point in C

 Summary

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 22

 Carnegie Mello

Floating Point Operations:

Basic Idea
 x +f y = Round(x + y)

 x f y = Round(x  y)

 Basic idea
 First compute exact result
 Make it fit into desired precision
 Possibly overflow if exponent too large
 Possibly round to fit into frac

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 23

 Carnegie Mello

Rounding
 Rounding Modes (illustrate with $ rounding)

 $1.40 $1.60 $1.50 $2.50 –

$1.50
 Towards zero $1 $1 $1 $2 –$1
 Round down (−) $1 $1 $1 $2 –$2
 Round up (+) $2 $2 $2 $3 –$1
 Nearest Even (default) $1 $2 $2 $2 –$2

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 24

 Carnegie Mello

Closer Look at Round-To-Even

 Default Rounding Mode
 All others are statistically biased
 Sum of set of positive numbers will consistently be over- or under-
estimated

 Applying to Other Decimal Places / Bit Positions

 When exactly halfway between two possible values

Round so that least significant digit is even
 E.g., round to nearest hundredth
7.8949999 7.89 (Less than half way)
7.8950001 7.90 (Greater than half way)
7.8950000 7.90 (Half way—round up)
7.8850000 7.88 (Half way—round down)

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 25

 Carnegie Mello

Rounding Binary Numbers

 Binary Fractional Numbers
 “Even” when least significant bit is 0
 “Half way” when bits to right of rounding position = 100…2

 Examples
 Round to nearest 1/4 (2 bits right of binary point)
Value Binary Rounded Action Rounded Value
2 3/32 10.000112 10.002 (<1/2—down) 2
2 3/16 10.001102 10.012 (>1/2—up) 2 1/4
2 7/8 10.111002 11.002 ( 1/2—up) 3
2 5/8 10.101002 10.102 ( 1/2—down) 2 1/2

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 26

 Carnegie Mello

Floating Point Multiplication

 (–1)s1 M1 2E1 x (–1)s2 M2 2E2
 Exact Result: (–1)s M 2E

 Sign s: s1 ^ s2
 Significand M: M1 x M2
 Exponent E: E1 + E2

 Fixing
 If M ≥ 2, shift M right, increment E
 If E out of range, overflow
 Round M to fit frac precision

 Implementation
 Biggest chore is multiplying significands
Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 27
 Carnegie Mello

Floating Point Addition

 (–1)s1 M1 2E1 + (-1)s2 M2 2 E2
Get binary points lined up
Assume E1 > E2
E1–E2
(–1)s1 M1
 Exact Result: (–1) M 2 s E

Sign s, significand M:

+ (–1)s2 M2
Result of signed align & add
Exponent E: E1 (–1)s M

 Fixing
If M ≥ 2, shift M right, increment E
if M < 1, shift M left k positions, decrement E by k
Overflow if E out of range
Round M to fit frac precision
Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 28
 Carnegie Mello

Mathematical Properties of FP
Add
 Compare to those of Abelian Group
 Closed under addition? Yes
 But may generate infinity or NaN
 Commutative? Yes
 Associative? No
 Overflow and inexactness of rounding
 (3.14+1e10)-1e10 = 0, 3.14+(1e10-1e10) = 3.14
 0 is additive identity? Yes
 Every element has additive inverse? Almost
 Yes, except for infinities & NaNs

 Monotonicity
 a ≥ b ⇒ a+c ≥ b+c? Almost
 Except for infinities & NaNs

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 29

 Carnegie Mello

Mathematical Properties of FP
Mult
 Compare to Commutative Ring
 Closed under multiplication? Yes
 But may generate infinity or NaN
 Multiplication Commutative? Yes
 Multiplication is Associative? No
 Possibility of overflow, inexactness of rounding
 Ex: (1e20*1e20)*1e-20= inf, 1e20*(1e20*1e-20)= 1e20
 1 is multiplicative identity? Yes
 Multiplication distributes over addition? No
 Possibility of overflow, inexactness of rounding
 1e20*(1e20-1e20)= 0.0, 1e20*1e20 – 1e20*1e20 = NaN

 Monotonicity
 a ≥ b & c ≥ 0 ⇒ a * c ≥ b *c? Almost
 Except for infinities & NaNs
Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 30
 Carnegie Mello

Binary Floating Point

 Background: Fractional binary numbers
 IEEE floating-point standard: Definition

 Example and properties

 Rounding, addition, multiplication

 Alternative formats

 Floating point in C

 Constructing a floating-point number

 Summary

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 31

 Carnegie Mello

Application Driven
 Reduced range and precision sufficient for some
applications, for example,
 Image processing
 Neural networks

 IEEE binary16 (Half precision)

 Supported by recent CPUs from AMD, Arm, and Intel

s exp frac
1 5-bits 10-bits

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 32

 Carnegie Mello

Deep Learning Driven

 The Google Brain research group argued that the
range of binary16 was too limited for deep learning
 Need the same range as binary32
 Sacrifice precision for speed

 (de-facto standard) bfloat16

 Supported by Google's TPU and some recent CPUs
 Easy conversion from binary32: truncate frac

s exp frac
1 8-bits 7-bits

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 33

 Carnegie Mello

Statistics Driven
 Multiplying a large number of probabilities results in
underflow
 Hidden Markov Models

 Need greater range to avoid underflow

 Perform calculations in log space rather than linear space
 Replace x0x1x2...xn by log x0 + log x1 + log x2 + ... + log xn
Slow!
 Use Posits format
 Adds a new section to the format that allows for a varying partition
between the bits of the exp and frac

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 34

 Carnegie Mello

Binary Floating Point

 Background: Fractional binary numbers
 IEEE floating-point standard: Definition

 Example and properties

 Rounding, addition, multiplication

 Alternative formats

 Floating point in C

 Constructing a floating-point number

 Summary

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 35

 Carnegie Mello

Floating Point in C
 C provides three standard binary floating point
types
float binary32 (single precision)
double binary64 (double precision)
long double varies by machine, at least binary64
 binary128 on Arm64
 binary64-extended on x86-64 (80 bits)
s exp frac
1 15-bits 63 or 64-bits

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 36

 Carnegie Mello

Floating Point in C
 Conversions/Casting
 Casting between int, float, and double changes bit representation
 double/float → int

Truncates fractional part
 Like rounding toward zero
 Not defined when out of range or NaN: Generally sets to TMin
 int → double
 Exact conversion, as long as int has ≤ 53 bit word size
 int → float
 Will round according to rounding mode

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 37

 Carnegie Mello

Floating Point Puzzles

 For each of the following C expressions, either:
 Argue that it is true for all argument values
 Explain why not true
• x == (int)(float)x
• x == (int)(double)x
• f == (float)(double)f
int x = …;
• d == (double)(float)d
float f = …;
• f == -(-f);
double d = …;
• 2/3 == 2/3.0
Assume neither • d < 0.0 ⇒ ((d*2) < 0.0)
d nor f is NaN • d > f ⇒ -f > -d
• d * d >= 0.0
• (d+f)-d == f

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 38

 Carnegie Mello

Binary Floating Point

 Background: Fractional binary numbers
 IEEE floating-point standard: Definition

 Example and properties

 Rounding, addition, multiplication

 Alternative formats

 Floating point in C

 Constructing a floating-point number

 Summary

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 39

 Carnegie Mello

Creating Floating Point Number

 Steps s exp frac
 Normalize to have leading 1
1 4-bits 3-bits
 Round to fit within fraction
 Postnormalize to deal with effects of rounding

 Case Study
 Convert 8-bit unsigned numbers to tiny floating point format
Example Numbers
128 10000000
13 00001101
17 00010001
19 00010011
138 10001010
63 00111111
Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 40
 Carnegie Mello

Normalize s exp frac

1 4-bits 3-bits
 Requirement
 Set binary point so that numbers of form 1.xxxxx
 Adjust all to have leading one

Decrement exponent as shift left
Value Binary Fraction Exponent
128 10000000 1.0000000 7
13 00001101 1.1010000 3
17 00010001 1.0001000 4
19 00010011 1.0011000 4
138 10001010 1.0001010 7
63 00111111 1.1111100 5

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 41

 Carnegie Mello

Rounding 1.BBGRXXX
Guard bit: LSB of
result Sticky bit: OR of remaining bit
Round bit: 1st bit removed

 Round up conditions
 Round = 1, Sticky = 1 ➙ > 0.5
 Guard = 1, Round = 1, Sticky = 0 ➙ Round to even
Value Fraction GRS Incr? Rounded
128 1.0000000 000 N 1.000
13 1.1010000 100 N 1.101
17 1.0001000 010 N 1.000
19 1.0011000 110 Y 1.010
138 1.0001010 011 Y 1.001
63 1.1111100 111 Y 10.000

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 42

 Carnegie Mello

Postnormalize
 Issue
 Rounding may have caused overflow
 Handle by shifting right once & incrementing exponent
Value Rounded Exp Adjusted Result
128 1.000 7 128
13 1.101 3 13
17 1.000 4 16
19 1.010 4 20
138 1.001 7 144
63 10.000 5 1.000/6 64

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 43

 Carnegie Mello

Summary
 IEEE Floating Point has clear mathematical
properties
 Represents numbers of form M x 2E

 One can reason about operations independent of

implementation
 As if computed with perfect precision and then rounded
 Not the same as real arithmetic
 Violates associativity/distributivity
 Makes life difficult for compilers & serious numerical applications
programmers

Bryant and O’Hallaron, Computer Systems: A Programmer’s Perspective, Third Edition 44