Scribd
Upload
24 views

Finite Wordlength Effects: Professor A G Constantinides

Finite wordlength effects can cause errors in digital signal processing systems due to: (1) Input and output quantization from finite register lengths and A/D converters. (2) Coefficient (or multiplier) quantization from products of multiplication being truncated or rounded. (3) Limit cycles can occur when quantization causes the output of a system to assume a cyclic set of values within a "deadband" rather than converging to zero. (4) Finite precision arithmetic in systems like FFTs can cause the signals to grow exponentially with each pass, though on average it is 0.5 bits of growth per pass for FFTs.

Uploaded by

Antonio Leon
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
24 views

Finite Wordlength Effects: Professor A G Constantinides

Finite wordlength effects can cause errors in digital signal processing systems due to: (1) Input and output quantization from finite register lengths and A/D converters. (2) Coefficient (or multiplier) quantization from products of multiplication being truncated or rounded. (3) Limit cycles can occur when quantization causes the output of a system to assume a cyclic set of values within a "deadband" rather than converging to zero. (4) Finite precision arithmetic in systems like FFTs can cause the signals to grow exponentially with each pass, though on average it is 0.5 bits of growth per pass for FFTs.

Uploaded by

Antonio Leon
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 26

Finite Wordlength Effects

Finite register lengths and A/D converters


cause errors in:(i) Input quantisation.
(ii) Coefficient (or multiplier)
quantisation
(iii) Products of multiplication truncated
or rounded due to machine length

Professor A G Constantinides

Finite Wordlength Effects


Quantisation
Output

eo (k )
Q

ei (k )
Input

Q
Q
ei ,o (k )
2
2
2

Professor A G Constantinides

Finite Wordlength Effects


The pdf for e using rounding
1
Q

Noise power
or

2
Q
2

12

Q
2

Q
2
Q 2

e p(e).de E{e }
2

Q 2

Professor A G Constantinides

Finite Wordlength Effects


Let input signal be sinusoidal of unity
amplitude. Then total signal power P 1

b
Q

2
2
If b bits used for binary then

so that 2 22b 3
Hence
3 2b
2
P .2
2

or SNR 1.8 6b
4

dB
Professor A G Constantinides

Finite Wordlength Effects


Consider a simple example of finite
precision on the coefficients a,b of second
order system with poles e j
H ( z)

1
1

1 az bz

1
H ( z)
1
2 2
1 2 cos .z .z
5

where

a 2 cos

2
Professor A G Constantinides

Finite Wordlength Effects

bit pattern
000
001
010
011
100
101
110
111
1.0

2 cos , 2
0
0.125
0.25
0.375
0.5
0.625
0.75
0.875
1.0

0
0.354
0.5
0.611
0.707
0.791
0.866
0.935
1.0Professor A G Constantinides

Finite Wordlength Effects


Finite wordlength computations
+

OUTPU
T

INPUT
+

Professor A G Constantinides

Limit-cycles; "Effective Pole"


Model; Deadband
Observe that for H ( z ) 1

(1 b1 z 1 b2 z 2 )

instability occurs when b2 1


i.e. poles are
(i) either on unit circle when complex
(ii) or one real pole is outside unit
circle.
Instability under the "effective pole" model
is considered as follows
Professor A G Constantinides

Finite Wordlength Effects


In the time domain with H ( z ) Y ( z )
X ( z)

y (n) x(n) b1 y (n 1) b2 y (n 2)

With b2 1 for instability we have


Qb2 y (n 2) indistinguishable from y (n 2)
Where Q is quantisation

Professor A G Constantinides

Finite Wordlength Effects


With rounding, therefore we have
b2 y (n 2) 0.5

y ( n 2)

are indistinguishable (for integers)


or
b2 y (n 2) 0.5 y (n 2)
0.5
Hence
y (n 2)

1 b2

With both positive and negative numbers


10

0.5
y (n 2)
1 b2

Professor A G Constantinides

Finite Wordlength Effects


The range of integers

0.5
1 b2

constitutes a set of integers that cannot be


individually distinguished as separate or from the
asymptotic system behaviour.
The band of integers

0.5
0.5

,
1 b2
1 b2

is known as the "deadband".


In the second order system, under rounding, the
output assumes a cyclic set of values of the
deadband. This is a limit-cycle.
11

Professor A G Constantinides

Finite Wordlength Effects


Consider the transfer function
G( z) 1

(1 b1 z 1 b2 z 2 )

yk xk b1 yk 1 b2 yk 2

if poles are complex then impulse response


is given by hk
k
hk
.sin (k 1)
sin
12

Professor A G Constantinides

Finite Wordlength Effects


cos1 b1

2 b2

Where b2

If b2 1 then the response is sinusiodal


with frequency
1
1 b1
cos

Thus product quantisation causes instability


implying an "effective b2 1 .
13

Professor A G Constantinides

Finite Wordlength Effects


Consider infinite precision computations for
x0 10

yk xk yk 1 0.9 yk 2

xk 0 ; k 0

10
8
6
4
2
0
-2
-4
-6
-8
-10
-10

14

-5

10

Professor A G Constantinides

Finite Wordlength Effects


Now the same operation with integer
precision
10
8
6
4
2
0
-2
-4
-6
-8
-10
-10

15

-5

10

Professor A G Constantinides

Finite Wordlength Effects


Notice that with infinite precision the
response converges to the origin
With finite precision the reponse does not
converge to the origin but assumes
cyclically a set of values the Limit Cycle

16

Professor A G Constantinides

Finite Wordlength Effects


Assume e1 (k ) ,e2 (k ) .. are not
correlated, random processes etc.
2

0i e hi (k )
e
2

k 0

2
Q

12

Hence total output noise power


0

2
01

02

Where Q 2

22b 2 k sin 2 (k 1)
2.
.
12 k 0
sin 2

and

sin (k 1)
h1 (k ) h2 (k ) .
; k 0
sin
k

17

Professor A G Constantinides

Finite Wordlength Effects


ie
0

18

2b

2
1
1

2
4
2
6 1 1 2 cos 2

Professor A G Constantinides

Finite Wordlength Effects


For FFT

B(n+1)

A(n)
B(n)

B(n+1)

W(n)

A(n 1) A(n) W (n).B(n)


B(n 1) A(n) W (n).B(n)
A(n)

B(n)
19

B(n)W(n)

A(n+1)
B(n+1)
Professor A G Constantinides

Finite Wordlength Effects


FFT
A(n 1) B(n 1) 2
2

A(n 1) 2 A(n)
2

A(n) 2 A(n)

AVERAGE GROWTH: 1/2 BIT/PASS

20

Professor A G Constantinides

Finite Wordlength Effects


IMAG 1.0

FFT

1.0
REAL

-1.0

-1.0

Ax ( n 1) Ax ( n) Bx ( n)C ( n) B y ( n) S ( n)
Ax ( n 1) Ax ( n) Bx ( n) C ( n) B y ( n) S ( n)
Ax ( n 1)
1.0 C ( n) S ( n) 2.414....
Ax ( n)

PEAK GROWTH: 1.21.. BITS/PASS


21

Professor A G Constantinides

Finite Wordlength Effects


Linear modelling of product quantisation
x(n)

~
x ( n)

Modelled as
x(n)

~
x ( n) x ( n) q ( n)

q(n)

22

Professor A G Constantinides

Finite Wordlength Effects


For rounding operations q(n) is uniform
distributed between Q2 , Q
and where Q is
2
the quantisation step (i.e. in a wordlength of
bits with sign magnitude representation or
mod 2, Q 2 b ).
A discrete-time system with quantisation at
the output of each multiplier may be
considered as a multi-input linear system
23

Professor A G Constantinides

Finite Wordlength Effects


q1 (n)...q2 (n)...q p (n)

x(n)
Then

y (n)

h(n)

y ( n) x ( r ).h( n r ) q ( r ).h ( n r )

r 0
r 0
1
p

where h (n) is the impulse response of the


system from the output of the multiplier
to y(n).
24

Professor A G Constantinides

Finite Wordlength Effects


For zero input i.e. x(n) 0, n we can write
p

r 0

y (n) q . h (n r )

where q is the maximum of q (r ) , , r


which is not more than Q
2

ie
25

Q p
y (n) . h (n r )

2 1n0
Professor A G Constantinides

Finite Wordlength Effects


However

n 0

n 0

h (n) h(n)

And hence
pQ
y ( n)
. h( n)
2 n 0

26

ie we can estimate the maximum swing at


the output from the system parameters and
quantisation level
Professor A G Constantinides

You might also like