APR7/D

Rev. 2

Implementing
IIR/FIR Filters
with

Motorola's
DSP56000/SPS/DSP56001
Digital Signal
Processors

M o t o r o l a ’ s H i g h - P e r f o r m a n c e D S P T e c h n o l o g y
 Motorola
Digital Signal
Processors
Implementing IIR/FIR
Filters with Motorola’s
DSP56000/DSP56001

by
John Lane and Garth Hillman
Digital Signal Processing Division

MOTOROLA APR 7
© Motorola Inc. 1993

Motorola reserves the right to make changes without further notice to any products here-
in. Motorola makes no warranty, representation or guarantee regarding the suitability of
its products for any particular purpose, nor does Motorola assume any liability arising out
of the application or use of any product or circuit, and specifically disclaims any and all
liability, including without limitation consequential or incidental damages. “Typical” pa-
rameters can and do vary in different applications. All operating parameters, including
“Typicals” must be validated for each customer application by customer’s technical ex-
perts. Motorola does not convey any license under its patent rights nor the rights of oth-
ers. Motorola products are not designed, intended, or authorized for use as components
in systems intended for surgical implant into the body, or other applications intended to
support or sustain life, or for any other application in which the failure of the Motorola
product could create a situation where personal injury or death may occur. Should Buyer
purchase or use Motorola products for any such unintended or unauthorized application,
Buyer shall indemnify and hold Motorola and its officers, employees, subsidiaries, affili-
ates, and distributors harmless against all claims, costs, damages, and expenses, and
reasonable attorney fees arising out of, directly or indirectly, any claim of personal injury
or death associated with such unintended or unauthorized use, even if such claim alleg-
es that Motorola was negligent regarding the design or manufacture of the part. Motorola
and B are registered trademarks of Motorola, Inc. Motorola, Inc. is an Equal Opportunity/
Affirmative Action Employer.
 Table
of Contents
SECTION 1 1.1 Analog RCL Filter Types 1-4
Introduction 1.2 Analog Lowpass Filter 1-5
1.3 Analog Highpass Filter 1-9
1.4 Analog Bandstop Filter 1-9
1.5 Analog Bandpass Filter 1-12

SECTION 2 2.1 Digital Lowpass Filter 2-9

Second-Order 2.2 Digital Highpass Filter 2-11
Direct-Form IIR 2.3 Digital Bandstop Filter 2-18
Digital Filter 2.4 Digital Bandpass Filter 2-18
Sections 2.5 Summary of Digital Coefficients 2-22

SECTION 3 3.1 The Canonic-Form Difference

Single-Section Equation 3-1
Canonic Form 3.2 Analysis of Internal Node Gain 3-5
(Direct Form II) 3.3 Implementation on the DSP56001 3-10

SECTION 4 4.1 Gain Evaluation of Internal Nodes 4-1

Single-Section 4.2 Implementation on the DSP56001 4-2
Transpose Form

SECTION 5 5.1 Butterworth Lowpass Filter 5-1

Cascaded Direct 5.2 Cascaded Direct-Form Network 5-9
Form

MOTOROLA iii
 Table
of Contents
SECTION 6 6.1 Canonic Implementation 6-2
Filter Design 6.2 Transpose Implementation
and Analysis (Direct Form I) 6-14
System

SECTION 7 7.1 Linear-Phase FIR Filter Structure 7-2

FIR FILTERS 7.2 Linear-Phase FIR Filter Design
Using the Frequency Sampling
Method 7-8
7.3 FIR Filter Design Using FDAS 7-22
7.4 FIR Implementation on the
DSP56001 7-29

REFERENCES Reference-1

iv MOTOROLA
 Illustrations
Figure 1-1 s-Domain Analysis of Second-Order Lowpass
Analog Filter 1-6

Figure 1-2 Gain and Phase Response of Second-Order

Lowpass Analog Filter at Various Values
of Damping Factor, d 1-7

Figure 1-3 s-Domain Analysis of Second-Order Highpass

Analog Filter 1-10

Figure 1-4 Gain and Phase Response of Second-Order

Highpass Analog Filter at Various Values
of Damping Factor, d 1-11

Figure 1-5 s-Domain Analysis of Second-Order Bandstop

Analog Filter 1-13

Figure 1-6 Gain and Phase Response of Second-Order

Bandstop Analog Filter at Various Values of
Damping Factor, d 1-14

Figure 1-7 s-Domain Analysis of Second-Order Bandpass

Analog Filter 1-15

Figure 1-8 Gain and Phase Response of Second-Order

Bandpass Analog Filter at Various Values of
Damping Factor, d 1-16

Figure 2-1 Spectrum of Bandlimited Signal Repeated at

Multiples of the Sampling Frequency, fs 2-3

Figure 2-2 Direct-Form Implementation of Second-Order

Lowpass IIR Filter and Analytical Formulas
Relating Desired Response to Filter Coefficients 2-12

MOTOROLA v
 Illustrations
Figure 2-3 Gain and Phase Response of Second-Order
Lowpass IIR Filter 2-13
Figure 2-4 DSP56001 Code and Data Structures for
Second-Order Direct-Form Implementation
of a Lowpass IIR Filter 2-14
Figure 2-5 Direct-Form Implementation of Second Order
Highpass IIR Filter and Analytical Formulas
Relating Desired Response to Filter Coefficients 2-15
Figure 2-6 Gain and Phase Response of Second-Order
Highpass IIR Filter 2-16
Figure 2-7 DSP56001 Code and Data Structures for
Second-Order Direct-Form Implementation
of a Highpass IIR Filter 2-17
Figure 2-8 Direct-Form Implementation of Second-Order
Bandstop IIR Filter and Analytical Formulas
Relating Desired Response to Filter Coefficients 2-19
Figure 2-9 DSP56001 Code and Data Structures for
Second-Order Direct-Form Implementation
of a Bandstop IIR Filter 2-20
Figure 2-10 Gain and Phase Response of Second-Order
Bandstop IIR Filter 2-21
Figure 2-11 Direct-Form Implementation of Second-Order
Bandpass IIR Filter and Analytical Formulas
Relating Desired Response to Filter Coefficients 2-23
Figure 2-12 DSP56001 Code and Data Structures for
Second-Order Direct-Form Implementation
of a Bandpass IIR Filter 2-24

vi MOTOROLA
 Illustrations
Figure 2-13 Gain and Phase Response of Second-Order
Bandpass IIR Filter 2-25
Figure 2-14 Summary of Digital Coefficients for the Four
Basic Filter Types 2-26
Figure 2-15 Pole Location and Analysis of Second-Order
Section 2-27

Figure 3-1 The Second-Order Canonic (Direct Form II)

IIR Filter Network 3-2
Figure 3-2 Internal Node Transfer Function, Hw(z), of
Canonic (Direct Form II) Network 3-6
Figure 3-3 Internal Node Gain Analysis of Second-Order
Canonic Form 3-8
Figure 3-4 Internal Node Gain Analysis of Second-Order
Canonic Form 3-9

Figure 4-1 Network Diagram for Transpose-Form

(Direct Form I) Implementation of a
Single-Section Second-Order Filter 4-3
Figure 4-2 Gain Evaluation at First Internal Node, u(n), of
Transpose Network 4-4
Figure 4-3 Gain Evaluation at Second Internal Node,
v(n), of Transpose Network 4-5

MOTOROLA vii
 Illustrations
Figure 4-4 Total Gain and Gain at Internal Nodes of Lowpass
Transpose Filter Network Figure 4-1
for fs =1000Hz 4-6
Figure 4-5 DSP56001 Code and Data Structures for Single-
Section Second-Order Transpose Form 4-7

Figure 5-1 Composite Response of Cascaded Second-Order

Sections for Lowpass Butterworth Filter
(kth Section Damping Factor
According to Eqn. 5-2) 5-7
Figure 5-2 Network Diagram for Cascaded Direct-Form Filter
and Data Structures Used in Code
Implementation (Three-Section Example) 5-10
Figure 5-3 DSP56001 Code for Cascaded Direct-Form Filter 5-11

Figure 6-1 Log Magnitude Plot of Example Lowpass

Butterworth Filter 6-3
Figure 6-2 Phase Versus Frequency Plot for Example Filter 6-4
Figure 6-3 Zero/Pole Plot of Sixth-Order Lowpass Example
Filter 6-5
Figure 6-4 Group Delay Versus Frequency for FDAS IIR
Example 6-6

viii MOTOROLA
 Illustrations
Figure 6-5 FDAS.OUT File of Example Filter for Cascaded
Canonic Implementation 6-7
Figure 6-6 COEFF.OUT File of Example Filter Design – Scaled
for Cascaded Canonic Implementation 6-9
Figure 6-7 COEFF.ASM File Generated by MGEN for Example
Design and Cascaded Canonic Implementation 6-10
Figure 6-8 FDAS.OUT File of Example Filter for Cascaded
Transpose – Form Implementation 6-15
Figure 6-9 COEFF.FLT File for Example Filer Design – Scaled
for Cascaded Transpose Form 6-17
Figure 6-10 COEFF.ASM File Generated by MGEN for Example
Design and Cascaded Transpose – Form
Implementation 6-18

Figure 7-1 FIR Structure 7-3

Figure 7-2 Signal Data through a FIR Filter 7-5
Figure 7-3 Arbitrary Filter Example 7-14
Figure 7-4 Response Transformed from Polar Coordinates 7-16
Figure 7-5 FIR Coefficients from Eqn. 7-2 for Filter Example 7-17
Figure 7-6 Roots (Zeros) of Eqn. 7-21 for Filter Example 7-17
Figure 7-7 A 32-Point FIR Filter Example 7-18
Figure 7-8 Log Magnitude Response of Filter Example with
Larger N Values 7-20

MOTOROLA ix
 Illustrations
Figure 7-9 Window Function Effects on Filter Example 7- 21
Figure 7-10 FDAS Output for FIR Bandpass Filter Example
with a Kaiser Window 7-23
Figure 7-11 FDAS Output for FIR Bandpass Filter Example
with Equiripple Design 7-28
Figure 7-12 FIR Filter Example 7-31

x MOTOROLA
 SECTION 1

Introduction
“Two classes of SECTION 1
frequency-
selective digital
This application note considers the design of frequen-
cy-selective filters, which modify the frequency content
filters are
and phase of input signals according to some specifi-
considered:
cation. Two classes of frequency-selective digital filters
infinite impulse
are considered: infinite impulse response (IIR) and fi-
response (IIR)
nite impulse response (FIR) filters. The design process
and finite
consists of determining the coefficients of the IIR or FIR
impulse
filters, which results in the desired magnitude and
response (FIR)
phase response being closely approximated.
filters.”
Therefore, this application note has a two-fold purpose:

1. to provide some intuitive insight into digital filters,

particularly how the coefficients are calculated in
the digital domain so that a desired frequency
response is obtained, and

2. to show how to implement both classes of digital

filters (IIR and FIR) on the DSP56001.

It is assumed that most readers are analog designers

learning digital signal processing (DSP). The ap-
proach used reflects this assumption in that digital
filters are initially presented from an analog point of
view. Hopefully, this approach will simplify the transi-
tion from the analog s-domain transfer functions to the

MOTOROLA 1-1
 equivalent functions in the digital z-domain. In
keeping with this analog perspective, IIR filters will
be discussed first since the equivalent of FIR filters
are infrequently encountered in the analog world.

SECTION 1 is a brief review of lowpass, highpass,

bandpass, and bandstop analog filters. The s-domain
formulas governing the key characteristics, magni-
tude-frequency response, G(Ω), and phase-
frequency response, φ(Ω), are derived from first
principals. Damping factor, d, cutoff frequency, Ωc,
for lowpass and highpass filters, center frequency,
Ω0, for bandpass and bandstop filters, and quality
factor, Q, are defined for the various filter types.

SECTION 2 introduces the bilinear transformation so

that analog s-domain designs can be transformed
into the digital z-domain and the correct coefficients
thereby determined. The form of the formulas for the
z-domain filter coefficients thus determined are gen-
eralized in terms of the key filter characteristics in the
z-domain so that the engineer can design digital fil-
ters directly without the necessity of designing the
analog equivalent and transforming the design back
into the digital domain.

In the analog domain, the performance of the filter

depends on the tolerance of the components. Simi-
larly, in the digital domain, the filter performance is
limited by the precision of the arithmetic used to im-
plement the filters. In particular, the performance of
digital filters is extremely sensitive to overflow, which

1-2 MOTOROLA
occurs when the accumulator width is insufficient to
represent all the bits resulting from many consecu-
tive additions. This condition is similar to the
condition in the analog world in which the signal out-
put is larger than the amplifier power supply so that
saturation occurs. A short analysis of the gain at
critical nodes in the filters is given in SECTION 3
and SECTION 4 to provide some insight into the
scaling requirements for different forms of IIR filters.
For this reason, the signal flow graphs developed
are centralized about the accumulator nodes.

The analysis of IIR filters focuses on second-order

sections. Clearly, higher order filters are often re-
quired. Therefore, a brief discussion of how second
order sections can be cascaded to yield higher order
filters is given in SECTION 5. Because the analysis
becomes complex quickly, the discussion naturally
leads to using commercially available filter design
software such as Filter Design and Analysis System
(FDAS) from Momentum Data Systems, Inc. SEC-
TION 6 concludes by showing how the filter
coefficients just discussed can be used in DSP56000
code to implement practical digital filters. Examples
of complete filter designs are given, including the
code, coefficients, frequency response, and maxi-
mum sample frequency.

FIR filters are discussed in SECTION 7. Initially, FIR

filters are contrasted with IIR filters to show that in
many ways they are complementary, each satisfy-
ing weaknesses of the other. An intuitive approach

MOTOROLA 1-3
 is taken to calculating the filter coefficients by start-
ing from a desired arbitrary frequency response.
The importance of and constraint imposed by linear
phase is emphasized. Having developed an intuitive
appreciation of what FIR filter coefficients are, the use
of FDAS to accelerate the design process is de-
scribed. SECTION 7 concludes by showing how the
filter coefficients just determined can be used in
DSP56000 code to implement practical digital filters.
An example of a passband digital filter using a Kaiser-
window design approach is presented.

1.1 Analog RCL Filter Types

In the following paragraphs, the analog RCL filter
network will be analyzed for the four basic filter
types: lowpass, highpass, bandpass, and band-
stop. Analyzing analog filter types shows that
designing digital IIR filters is, in many cases, much
simpler than designing analog filters.

In this analysis, as in all of the following cases, the

input is assumed to be a steady-state signal con-
taining a linear combination of sinusoidal
components whose rms (or peak) amplitudes are
constant in time. This assumption allows simple an-
alytic techniques to be used in determining the
network response. Even though these results will
then be applied to real-world signals that may not
satisfy the original steady-state assumption, the de-

1-4 MOTOROLA
viation of the actual response from the predicted
response is small enough to neglect in most cases.
General analysis techniques consist of a linear
combination of steady-state and transient re-
sponse solutions to the differential equations
describing the network.

1.2 Analog Lowpass Filter

The passive RCL circuit forming a lowpass filter
network is shown in Figure 1-1 where the transfer
function, H(s), is derived from a voltage divider
analysis of the RCL network. This approach is valid
since the effect of C and L can be described as a
complex impedance (or reactance, Xc and XL) un-
der steady state conditions; s is a complex variable
of the complex transfer function, H(s). The filter fre-
quency response is found by evaluating H(s) with
s = jΩ, where Ω= 2πf and f is the frequency of a si-
nusoidal component of the input signal. The output
signal is calculated from the product of the input sig-
nal and H(jΩ). To facilitate analysis, the input and
output signal components are described by the
complex value, ejΩt = cos Ωt + j sin Ωt. The actual
physical input and output signal components are
found by taking the real part of this value. The input
is R{ejΩt} = cos Ωt; the output is R{H(jΩ)ejΩt} = G(Ω)
cos [Ωt + φ(Ω)]. The previous technique is based
upon the solution of the differential equations de-
scribing the network when the input is steady state.
Describing the circuit response by H(s) instead of

MOTOROLA 1-5
 solving the differential equation is a common simpli-
fication used in this type of analysis.

1-6 MOTOROLA
 V0 CX ||R Vi V0 Xc = 1/jΩC
----- = ------------
- L
Vi X L ||R C R

( R/jΩC ) / ( R + 1 /jΩC ) XL = jΩL

= --------------------------------------------------------------
jΩL + ( R/jΩC ) / ( R + 1/jΩC )

1
1/LC
= ------------------------------------------------- Ωc = -----------
LC
– Ω 2 + jΩ /RC + 1 / LC

Ω c2 L
= ------------------------------------------- d = ------------
– Ω 2 + jdΩ c Ω + Ω c2 R C
2

Let s = jΩ and define H(s) = Vo/Vi; then,

1
H(s)= ---------------------------------------------------
2
( s/Ω c ) + d ( s/Ω c ) + 1
which is the s-domain transfer function. The gain, G(Ω), of the filter is:

G(Ω) ≡ H ( s ) H∗ ( s )
s = jω

1
= ---------------------------------------------------------------
-
2 2 2 2
( 1 – Ω /Ω c ) + ( dΩ/Ω c )
where ∗ denotes complex conjugate.
The phase angle, φΩ, is the angle between the imaginary and
real components of H(s).
–1
φ ( Ω ) ≡ tan [ l { H ( s ) }/R { H ( s ) } ]

–1 d ( Ω/Ω c ) for Ω ≤ Ωc
= – tan ------------------------------
2
-
1 – ( Ω/Ω c )

–1 d ( Ω/Ω c ) for Ω > Ωc

= – π – tan ------------------------------
2
-
1 – ( Ω/Ω c )

Figure 1-1 s-Domain Analysis of Second-Order Lowpass Analog Filter

MOTOROLA 1-7
 1.5
d=0.707
d=1.0
1.0 d=1.41
Gain

d=2.0
0.5

0.1 1.0 10
Normalized Frequency (Ω/Ωc)

0 d=0.707
d=1.0
d=1.41
d=2.0
Phase

-π/2

-π

0.1 1.0 10
Normalized Frequency (Ω/Ωc)
Figure 1-2 Gain and Phase Response of Second-Order Lowpass Analog
Filter at Various Values of Damping Factor, d

The magnitude of H(s) is defined as the gain, G(Ω),

1-8 MOTOROLA
of the system; whereas, the ratio of the imaginary
part to real part of H(s), I{H(jΩ)}/R{H(jΩ)}, is the tan-
gent of the phase, φ(Ω), introduced by the filter. If
the input signal is Ak sin (Ωkt + φK), then the output
signal is AKG(ΩK) sin [ΩKt + φK+ φ(ΩK)] Figure 1-2
shows the gain, G(Ω), and phase, φ(Ω), plots for the
second-order lowpass network of Figure 1-1 for var-
ious values of damping factor, d; d also controls the
amplitude and position of the peak of the normal-
ized response curve.

The frequency corresponding to the peak amplitude

can be easily found by taking the derivative of G(Ω)
(from the equation for G(Ω) in Figure 1-1) with re-
spect to Ω and setting it equal to zero. Solving the
resultant equation for Ω then defines ΩM as the fre-
quency where the peak amplitude occurs. The peak
amplitude is then GM = G(ΩM):

2
Ω M = Ω c ( 1 – d /2 ) Eqn. 1-1

1
G M = ---------------------------------- Eqn. 1-2
2
d ( 1 – d /4 )

for d < 2 . For d > 2 , ΩM = 0 is the position of the

peak amplitude where GM = 1. When d = 2 , GM = 1,
which gives the maximally flat response curve used in
the Butterworth filter design (usually applies only to a
set of cascaded sections). Note that ΩC for a lowpass
filter is that frequency where the gain is G(Ωc) = 1/d
and the phase is φ(Ωc) = -π/2.

MOTOROLA 1-9
 1.3 Analog Highpass Filter
The passive RCL circuit forming a highpass filter
network is shown in Figure 1-3 where the transfer
function, H(s), is again derived from a voltage divid-
er analysis of the RCL network. The gain and phase
response are plotted in Figure 1-4 for different val-
ues of damping coefficient. As evidenced, the
highpass filter response is the mirror image of the
lowpass filter response.

1.4 Analog Bandstop Filter

The analog RCL network for a bandstop filter net-
work is simply the sum of the lowpass and highpass
transfer functions shown in Figure 1-5 where the
transfer function, H(s), is again derived from a volt-
age divider analysis of the RCL network. The gain
and phase response are plotted in Figure 1-6 for dif-
ferent values of quality factor, Q, (where Q = 1/d).
Neglecting the departure of real RCL components'
values from the ideal case, the attenuation at the
center frequency, f0, is infinite. Also, note that the
phase undergoes a 1 80-degree shift when passing
through the center frequency (zero in the s-plane).

Q for bandpass and bandstop filters is a measure

of the width, ∆Ω, of the stopband with respect to
the center frequency, ΩO, i.e., ∆Ω = Q-1Ωo. ∆Ω is
measured at the points where G(Ω) = 1/ 2 .

1-10 MOTOROLA
 V0 X ||R
L Vi V0 Xc = 1/jΩC
----- = ------------------------
-
Vi X C + X L ||R C
L R
j Ω LR / ( jΩ L + R ) XL = jΩL
= --------------------------------------------------------
1/j ΩC + j Ω LR / jΩ L + R

–Ω2 1
= ------------------------------------------------- Ωc = -----------
2 LC
– Ω + jΩ /RC + 1 / LC

–Ω2 L
d = ------------
= -------------------------------------------
2
– Ω + jdΩ c Ω + Ω c2 2
R C

Let s = jΩ and define H(s) = Vo/Vi; then,

2
( s/Ω c )
H(s)= ---------------------------------------------------
2
( s/Ω c ) + d ( s/Ω c ) + 1
which is the s-domain transfer function. The gain, G(Ω), of the filter is:

G(Ω) ≡ H ( s ) H∗ ( s )
s = jω 2
( Ω/Ω c )
= ---------------------------------------------------------------
-
2 2
( 1 – Ω 2 /Ω 2c ) + ( dΩ/Ω c )
where ∗ denotes complex conjugate.
The phase angle, φΩ, is the angle between the imaginary and
real components of H(s).
–1
φ ( Ω ) ≡ tan [ l { H ( s ) }/R { H ( s ) } ]

–1 d ( Ω/Ω c ) for Ω ≤ Ωc
= π – tan ------------------------------
2
-
1 – ( Ω/Ω c )

–1 d ( Ω/Ω c ) for Ω > Ωc

= – tan ------------------------------
2
-
1 – ( Ω/Ω c )

Figure 1-3 s-Domain Analysis of Second-Order Highpass Analog Filter

MOTOROLA 1-11
 1.5
d=0.707
d=1.0
d=1.414
1.0 d=2.0
Gain

0.5

0.1 1.0 10
Normalized Frequency (Ω/Ωc)

π d=0.707
d=1.0
d=1.414
d=2.0
Phase

π/2

0.1 1.0 10
Normalized Frequency (Ω/Ωc)

Figure 1-4 Gain and Phase Response of Second-Order Highpass Analog Filter
at Various Values of Damping Factor, d

1-12 MOTOROLA
1.5 Analog Bandpass Filter
The passive RCL circuit forming a bandpass filter
network is shown in Figure 1-7 where the transfer
function, H(s), is again derived from a voltage di-
vider analysis of the RCL network. The gain and
phase response are plotted in Figure 1-8 for differ-
ent values of Q. The lowpass gain approaches an
asymptotic function of G = (fc/f)2 for f/fO >> 1. The
highpass asymptotic gain is G = (fC/f)2 for f/fC << 1;
whereas, the bandstop case approaches unity at
zero and infinity with a true zero at the center fre-
quency. The bandpass gain, on the other hand,
approaches G = Q-1f/fC for f/fc << 1 and G = Q-1fc/f
for f/fc >> 1. The primary differences to note in the
bandpass response are:
1. the stopband attenuation is 6 dB/octave or
20 dB/decade (since it goes as 1/f); whereas,
the lowpass and highpass go as l/f2 (12 dB/
octave, 40 dB/decade)
2. the stopband attenuation asymptote is dependent
on the quality factor; whereas, for the lowpass and
highpass cases, the stopband attenuation
asymptote is independent of damping factor, d
3. the maximum value of gain is unity regardless of
the filter Q.

The specific features characterizing the bandpass,

lowpass, highpass, and bandstop analog networks
are found to be nearly identical in the digital IIR filter
equivalents when the sampling frequency is very
high as compared to the frequencies of interest. For
this reason, it is important to understand the basic
properties of the four filter types before proceeding
to the digital domain.

MOTOROLA 1-13
 L
V0 R Vi V0 Xc = 1/jΩC
----- = ------------------------
Vi x C ||X + R
L R
C
R XL = jΩL
= -------------------------------------------------------------------
R + ( jΩL/jΩC ) / ( jΩL + 1/jΩC )

– Ω 2 + 1/LC 1
Ωc = ----------
= -------------------------------------------------
2 LC
– Ω + jΩ /RC + 1 / LC

– Ω 2 + Ω 02 2
= ---------------------------------------------- Q = 1/d = R C/L
– Ω 2 + jΩ 0 Ω/Q + Ω 02

Let s = jΩ and define H(s) = Vo/Vi; then,

2
( s/Ω 0 ) + 1
H(s)= -----------------------------------------------
-
2
( s/Ω 0 ) + s/Ω 0 Q + 1
which is the s-domain transfer function. The gain, G(Ω), of the filter is:

G(Ω) ≡ H ( s ) H∗ ( s )
s = jω

2
– ( Ω/Ω 0 ) 1
= ---------------------------------------------------------------
2 2
( 1 – Ω 2 /Ω 20 ) + Ω/Ω 0 Q )
where ∗ denotes complex conjugate.
The phase angle, φΩ, is the angle between the imaginary and
real components of H(s).
–1
φ ( Ω ) ≡ tan [ l { H ( s ) }/R { H ( s ) } ]

–1
( Ω/Ω 0 Q )
= – tan --------------------------------
2
1 – ( Ω/Ω 0 )

Figure 1-5 s-Domain Analysis of Second-Order Bandstop Analog Filter

1-14 MOTOROLA
 1.0 d=0.707
d=1.0
d=1.414
d=2.0
Gain

0.5

0.1 1.0 10
Normalized Frequency (Ω/Ω0)

π/2

d=0.707
d=1.0
Phase

d=1.414
0
d=2.0

-π/2

0.1 1.0 10
Normalized Frequency (Ω/Ω0)
Figure 1-6 Gain and Phase Response of Second-Order Bandstop Analog Filter
at Various Values of Damping Factor, d

MOTOROLA 1-15
 V0 R Vi V0 Xc = 1/jΩC
----- = ----------------------------
Vi XL + XC + R L C R
XL = jΩL
R
= ---------------------------------
jΩL + 1/jΩ + R

1
jΩR/L Ωc = ----------
= ----------------------------------------------- LC
– Ω 2 + jΩ R/L + 1/LC

jΩ 0 Ω/Q
= ---------------------------------------------- 2
2 Q = 1/d = R C/L
– Ω + jΩ 0 Ω/Q + Ω 02

Let s = jΩ and define H(s) = Vo/Vi; then,

s/ Ω 0 Q
H ( s ) = -------------------------------------------------
2
( s/Ω 0 ) + s/Ω 0 Q + 1
which is the s-domain transfer function. The gain, G(Ω), of the filter is:

G(Ω) ≡ H ( s ) H∗ ( s )
s = jω

Ω/Ω Q 0
= ----------------------------------------------------------------
-
2 2 2 2
( 1 – Ω /Ω 0 ) + ( Ω/Ω 0 Q )

where ∗ denotes complex conjugate.

The phase angle, φΩ, is the angle between the imaginary and
real components of H(s).
–1
φ ( Ω ) ≡ tan [ l { H ( s ) }/R { H ( s ) } ]

–1 Ω/Ω 0 Q
= tan ------------------------------
-
2
1 – ( Ω/Ω 0 )

Figure 1-7 s-Domain Analysis of Second-Order Bandpass Analog Filter

1-16 MOTOROLA
 1.0
d=1.414
d=2.0
Gain

0.5
d=0.707
d=1.0

0.01 0.1 1.0

Normalized Frequency (Ω/Ωc)

π/2 d=0.707
d=1.0
d=1.41
d=2.0
Phase

-π/2

0.1 1.0 10
Normalized Frequency (Ω/Ωc)

Figure 1-8 Gain and Phase Response of Second-Order Bandpass Analog

Filter at Various Values of Damping Factor, d

1-17 MOTOROLA
1-18 MOTOROLA
 SECTION 2

Second-Order
Direct-Form IIR
Digital Filter Sections
“In the direct- The traditional approach to deriving the digital filter
form coefficients has been to start with the digital z-domain
implementation, description, transform to the analog s-domain to un-
the ai and bi are derstand how to design filters, then transform back to
used directly in the digital domain to implement the filter. This ap-
the difference proach is not used in this report. Instead, formulas are
equation, which developed relating the s-domain filter to the z-domain
can be easily filter so transformations to and from one domain to the
programmed on other are no longer necessary.
a high-speed
The Laplace or s-transform in the analog domain was
DSP such as the
developed to facilitate the analysis of continuous time
DSP56001.”
signals and systems. For example, using Laplace
transforms the concepts of poles and zeros, making
system analysis much faster and more systematic.
The Laplace transform of a continuous time signal is:

X(s) = L{x(t)}

∫ x ( t )e
– st Eqn. 2-3
dt
0

where: L = the Laplace transform operator and

implies the operation described in Eqn. 2-3

MOTOROLA 2-1
 In the digital domain, the continuous signal, x(t), is
first sampled, then quantized by an analog-to-digital
(A/D) converter before being processed. That is, the
signal is only known at discrete points in time, which
are multiples of the sampling interval, T = 1/fs,
where fs is the sampling frequency. Because of
the sampled characteristic of a digital signal, its
z-transform is given by a summation (as opposed
to an integral):

X(z) = Z{x(n)}

∞
–n
= ∑ x ( n )z Eqn. 2-4
n = –∞

where: Z is the z-transform operator as

described by the operation of Eqn. 2-4

x(n) is the quantized values from the A/D

converter of the continuous time
signal, x(t), at discrete times, t = nT

One property of the z-transform which will be used

later in this report is the time shifting property. The
time shifting property states:
––1  –k 
x(n – k) = Z z X(z)  Eqn. 2-5
 
The proof of this property follows directly from the
definition of the z-transform. An obvious question
arises: “If the s-domain of a signal, x(t), is known
and that same signal is digitized, what is the rela-
tionship between the s-domain transform and the
z-transform?” The relationship or mapping is not

2-2 MOTOROLA
unique and depends on the viewpoint used. It is ob-
vious that the trivial mapping, s = z, is inappropriate;
the signal has been sampled. When a bandlimited
signal is sampled (i.e., multiplied by a periodic im-
pulse function), the spectrum of the resulting signal is
repetitive as shown in Figure 2-9 (see Reference 10).

S(f)

usable
region

-fs/2 0 fs/2 fs 3fs/2

Frequency (f)

Figure 2-9 Spectrum of Bandlimited Signal Repeated at Multiples of the

Sampling Frequency, fs.

Clearly, the spectrum consists of the spectrum of

the bandlimited signal repeated at multiples of the
sampling frequency, ωs = 2πfs. That is, the resulting
spectrum is unique only between 0 and ωs/2 or mul-
tiples thereof; whereas, before the signal was
sampled, the energy at frequencies greater than
those in the signal was independent of the signal.
Therefore, acceptable mappings would either re-
flect the cyclic nature of the spectrum of the
sampled test or at least be linear over the frequen-
cies of interest.

MOTOROLA 2-3
 One mapping or transformation from the s-domain
to the z-domain discussed later in this report is the
bilinear transformation. To understand the origin of
this transformation, consider the simple first-order
linear analog filter with the system function:

Y(s) b
H ( s ) = ------------ = ------------ Eqn. 2-6
X(s) s+a

Recall the differentiation property of the s-transform

when x(t) = L-1{X(s)} (where L-1 is the inverse
Laplace transform operation); then the time deriva-
tive of x(t) is:

d –1
----- x ( t ) = L { sX ( s ) } Eqn. 2-7
dt

d
where: x ( t ) is the time derivative of x(t)
dt

Using the differentiation property of Eqn. 2-7, the

linear system described by Eqn. 2-6 can be ex-
pressed as follows:

d
----- y ( t ) + ay ( t ) = bx ( t ) Eqn. 2-8
dt

If this differential equation is solved by expressing

y(t) with the trapezoidal integration formula,
t
d
y(t) = ∫ -----
dτ
- y ( τ )dτ + y ( t 0 ) Eqn. 2-9
t0

2-4 MOTOROLA
where the approximate solution is given by:

1 d d
y ( t ) = --- ----- y ( t ) + ----- y ( t 0 ) ( t – t 0 ) + y ( t 0 ) Eqn. 2-10
2 dt dt

then using d y ( t ) from Eqn. 2-8 with t = nT and t0 =

dt
(n-1)T, Eqn. 2-10 can be expressed as follows:

( 2 + aT )y ( n ) – ( 2 – aT )y ( n – 1 ) = bT [ x ( n ) + x ( n – 1 ) ]

Eqn. 2-11

Taking the z-transform of this difference of this dif-

ference equation and using the time shifting
property of the z-transform, Eqn. 2-5 results in the
z-domain system function:

Y(z) b
H ( z ) = ------------ = ------------------------------------- Eqn. 2-12
X(z) –1
2 1 – z 
---  ------------------  + a
T
1+z 
–1

Clearly, the mapping between the s-plane and the

z-plane is:

2 1 – z 
–1
s = ---  -----------------
- Eqn. 2-13
T  1 + z –1 

This mapping is called bilinear transformation.

MOTOROLA 2-5
 Although this transformation was developed using a
first-order system, it holds, in general, for an Nth-order
system (see Reference 14). By letting s = σ + jΩ and
z = rejθ, it can be shown that the left-half plane in the
s-domain is mapped inside the unit r = 1 circle in the
z-domain under the bilinear transformation. More im-
portantly, when r = 1 and σ = 0, the frequencies in the
s-domain and the z-domain are related by:

2 θ
Ω = --- tan --- Eqn. 2-14
T 2

or equivalently:
– 1 ΩT
θ = 2tan -------- Eqn. 2-15
2

where: θ is the digital domain normalized

frequency equal to 2πf/fs

Ω is the analog domain frequency used

in the analysis of the previous section

On the jΩ axis or equivalently along the frequency

axis, the scale has been changed nonlinearly. The
gain and phase values depicted on the vertical axis
of Figure 1-2, Figure 1-4, Figure 1-6, and Figure 1-8
remain exactly the same in the digital domain (or z-
plane). The horizontal (frequency) axis is modified
so that an infinite frequency in the analog domain
maps to one-half of the sample frequency, fs/2, in
the digital domain; whereas, for frequencies much
less than fs/2, the mapping is approximately 1:1

2-6 MOTOROLA
with θ = Ω. In summary, the bilinear transformation
is a one-to-one nonlinear mapping from the s-do-
main into the z-domain in which high frequencies
(Ω > 2πfs/4) in the s-domain are compressed into a
small interval in the z-domain. Therefore, the gain
and phase expressions of the previous section can
be directly transformed into the digital domain by
simply substituting Eqn. 2-14 into the correspond-
ing expressions. This will be done for each filter
type in the following paragraphs.

First, it is appropriate to introduce the direct-form

implementation of a digital filter by noting that, in
general, if the bilinear transformation of Eqn. 2-13 is
substituted into the transfer function, H(s), of the
previous section, the resulting H(z) will have the fol-
lowing generalized form:

–1 –2
b0 + b1 z + b2 z
H ( z ) = ---------------------------------------------------- Eqn. 2-16
–1 –2
1 + a1 z + a2 z

where the digital domain coefficients, ai and bi, are

exactly related to the s-domain characteristics of
the system such as the center frequency, band-
width, etc. In the direct-form implementation, the ai
and bi are used directly in the difference equation,
which can be easily programmed on a high-speed
DSP such as the DSP56001. The time-domain dif-
ference equation is derived from the z-domain
transfer functions by applying the inverse z-trans-
form in general and the inverse time shifting

MOTOROLA 2-7
 property in particular as follows:
–1 –1
Z {H(z)} = Z { Y ( z )/X ( z ) }
–1  –1 –2 –1 
–2
= Z  b0 + b1 z + b2 z / 1 + a1 z + a2 z 
 
Eqn. 2-17

so that:

–1 –2  –1  –1
( z ) 1 + a1 z + a2 z  = Z  X ( z ) b0 + b1 z + b2
 

therefore, using the inverse time shifting property of

Eqn. 2-5:
–1  –k 
Z  X ( z )z  = {x(n – k)}
 
and
–1  –k 
Z  Y ( z )z  = {y(n – k)}
 

Eqn. 2-17 becomes:

y ( n ) = b0 x ( n ) + b1 x ( n – 1 ) + b2 x ( n – 2 ) + a1 y ( n – 1 ) – a2 y ( n – 2 )

Eqn. 2-18

2-8 MOTOROLA
Eqn. 2-18 can be directly implemented in software,
where x(n) is the sample input and y(n) is the corre-
sponding filtered digital output. When the filter
output is calculated using Eqn. 2-18, y(n) is calcu-
lated using the direct-form implementation of the
digital filter.

There are other implementations which can be

used for the same system (filter) transfer function,
H(z). The canonic-form implementation and the
transpose-form implementation are discussed in
subsequent sections. First, the directform imple-
mentation will be applied to the transfer function,
H(s), developed in SECTION 1 Introduction.

2.6 Digital Lowpass Filter

Using the analog transfer function, H(s), from Fig-
ure 1-1, Eqn. 2-13 and Eqn. 2-14, the digital transfer
function, H(z), becomes that shown in Figure 2-10,
where the coefficients α, β, and γ are expressed in
terms of the digital cutoff frequency, θc, and the
damping factor, d. The value of the transfer function
at θ = θc in the digital domain is identical to the value
of the s-domain transfer function at Ω= Ωc:

jθc
Hz ( e ) = H s ( jΩ c ) Eqn. 2-19

As shown in Figure 2-11, the digital gain and

phase response calculated from the equations of

MOTOROLA 2-9
 Figure 2-10 are similar to the analog plots shown
in Figure 1-2, except for the asymmetry introduced
by the zero at fs/2. That is, the frequency axis is
modified so that a gain of zero at f = ∞ in the s-do-
main corresponds to a gain of zero at f = fs/2 in the
z-domain. The fact that the magnitude of the trans-
fer functions, H(s) and H(z), is identical once the
proper frequency transformation is made is very
useful for understanding the digital filter and its re-
lationship to the analog equivalent. This fact is
also useful for purposes of scaling the gain since
the maximum magnitude of Gs(ΩM) = GZ(θM),
where ΩM and θM are related by Eqn. 2-14. In oth-
er words, scaling analysis of the digital transfer
function, H(z), can be done in the s-domain (the al-
gebra is often easier to manage). Scaling of the
gain is an essential part of digital filter implemen-
tation since the region of numeric calculations on
fixed-point DSPs such as the DSP56001 are usu-
ally restricted to a range of -1 to 1.

Using the formulas given in Figure 2-10 with Eqn. 1-

2 guarantees the behavior of the digital filter. Since
the gain is scaled to unity at f = 0 (DC), the input da-
ta, x(n), in Figure 2-10 must be scaled down by a
factor of 1/GM from Eqn. 1-2 if the entire dynamic
range of the digital network is to be utilized. The al-
ternative procedure is automatic gain control to
insure that x(n) is smaller than 1/GM before it arrives
at the filter input. For d ≥ 2 , the input does not re-
quire scaling since the gain of the filter will never
exceed unity.

2-10 MOTOROLA
The DSP56001 code to implement the second-or-
der lowpass filter section is shown in Figure 2-12.
The address register modifiers are initially set to
M0 = 4, M4 = 1, and M5 = 1 to allow use of the cir-
cular buffer or modulo addressing in this particular
implementation (see Reference 8). Typically, this
code would be an interrupt routine driven by the input
data (A/D converter, for example) sample rate clock.
The basic filter code and the interrupt overhead and
data l/O moves for this second-order filter could be
executed at a sample rate of nearly 1 MHz on the
DSP56001.

2.7 Digital Highpass Filter

The highpass filter is nearly identical to the lowpass
filter as shown by the formulas in Figure 2-13. As
with the analog case, the digital highpass filter is
just the mirror image of the lowpass filter (see Fig-
ure 2-14). The frequency transformation from high
to low in the analog case is Ω→l/Ω; whereas, in the
digital case, it is θ→π/θ.

The DSP56001 code is shown in Figure 2-15; as

seen by comparison to the code shown in Figure 2-
12, the same instruction sequence is used. The only
difference is the coefficient data, which is calculated
by the formulas given in Figure 2-13. The scaling
mode is turned on so that a move from the A or B
accumulator to the X or Y register or to memory re-
sults in an automatic multiply by two. The scaling
mode is used not only in the code for the lowpass

MOTOROLA 2-11
 case but also in the code for the highpass, band-
stop, and bandpass cases.

2-12 MOTOROLA
¥¥¥

Network Diagram
α
x(n) 2 y(n)

z-1 z-1

2α γ
Σ

z-1 z-1

α −β

Transfer Function –1 –2
α ( 1 + 2z + z )
H ( z ) = ---------------------------------------------
–1 –2
1/2 – γ z + βz

Gain ( 1 + cos θ ) ( 1 – cos θ c )

G ( θ ) = --------------------------------------------------------------------------------------------------
2 2 1/2
[ ( d sin θ sin θ c ) + 4 ( cos θ – cos θ c ) ]

Phase
 – 1 2 ( cos θ – cos θ c ) for θ≤θc
 tan -----------------------------------------
 d sin θ sin θ c
φ(θ) = 
 – 1 2 ( cos θ – cos θ c )
 – π + tan ----------------------------------------
d sin θ sin θ c
- for θ>θc


Coefficients
1 1 – d ⁄ 2 sin θ c
β =  ---  ----------------------------------
 2  1 + d ⁄ 2 sin θ c

γ = ( 1/2 + β ) cos θ c

α = ( 1/2 + β – γ )/4

Figure 2-10 Direct-Form Implementation of Second-Order Lowpass IIR Filter and

Analytical Formulas Relating Desired Response to Filter Coefficients

MOTOROLA 2-13
 fo/fny=0.1 fo/fny=0.65
1.5
Lowpass Gain

1.0
d=0.707
d=1.0
d=1.41
0.5 d=2.0

0.01 0.1 1.0

log (f/fny)

0
d=0.707
Lowpass Phase

d=1.0
d=1.41
d=2.0
-π/2 fo/fny=0.1 fo/fny=0.65

-π

0.01 0.1 1.0

log (f/fny)
NOTE: Nyquist frequency, fny, is equal to one-half the sample frequency fs.

Figure 2-11 Gain and Phase Response of Second-Order Lowpass IIR Filter

2-14 MOTOROLA
 Difference Equation

y ( n ) = 2 { α [ x ( n ) + 2x ( n – 1 ) + x ( n – 2 ) ] + γy ( n – 1 ) – βy ( n – 2 ) }

Data Structures

X:(R0) Y:(R4)
2α x(n-1)
α x(n-2)
γ
Y:(R5)
–β y(n-1)
α y(n-2)

DSP56001 Code
;Y1=x(n) (Input)
;X0=α
MPY X0,Y1,A X:(R0)+,X0 Y:(R4)+,Y0 ;A=αx(n)
MAC X0,Y0,A X:(R0)+,X0 Y:(R4),Y0 ;A=A+2αx(n-1)
MAC X0,Y0,A X:(R0)+,X0 Y:(R5)+,Y0 ;A=A+αx(n-2)
MAC X0,Y0,A X:(R0)+,X0 Y:(R5),Y0 ;A=A+γy(n-1)
MAC X0,Y0,A X:(R0)+,X0 Y1,Y:(R4) ;A=A-βy(n-2)
MOVE A,X1 A,Y:(R5) ;y(n-2)=2A (assumes scaling
;mode is set).
;X1 is Output.

Total Instruction Cycles

6 Icyc @ 20 MHz = 600ns

Figure 2-12
Figure 2-12 DSP56001 Code and Data Structures for Second-Order Direct-Form
Implementation of a Lowpass IIR Filter

MOTOROLA 2-15
 Network Diagram
α
x(n) 2 y(n)

z-1 z-1

−2α γ
Σ

z-1 z-1

α −β

Transfer Function –1 –2
α ( 1 –2 z + z )
H ( z ) = ---------------------------------------------
–1 –2
1/2 – γ z + βz

Gain ( 1 – cos θ ) ( 1 + cos θ c )

G ( θ ) = --------------------------------------------------------------------------------------------------
2 2 1/2
[ ( d sin θ sin θ c ) + 4 ( cos θ – cos θ c ) ]

Phase
 – 1 2 ( cos θ – cos θ c ) for θ≤θc
 π + tan -----------------------------------------
 d sin θ sin θ c
φ(θ) = 
 –1 2 ( cos θ – cos θ c )
 tan ----------------------------------------
d sin θ sin θ c
- for θ>θc


Coefficients
1 1 – d ⁄ 2 sin θ c
β =  ---  ----------------------------------
 2  1 + d ⁄ 2 sin θ c

γ = ( 1/2 + β ) cos θ c

α = ( 1/2 + β + γ ) ⁄ 4

Figure 2-13 Direct-Form Implementation of Second Order Highpass IIR Filter and
Analytical Formulas Relating Desired Response to Filter Coefficients

2-16 MOTOROLA
 fo/fny=0.1 fo/fny=0.65
1.5
HIGHPASS GAIN

1.0
d=0.707
d=1.0
0.5 d=1.41
d=2.0

0.01 0.1 1.0

log (f/fny)

π
HIGHPASS PHASE

d=0.707
d=1.0
d=1.41
π/2 d=2.0
fo/fny=0.1 fo/fny=0.65

0.01 0.1 1.0

log (f/fny)
NOTE: Nyquist frequency, fny, is equal to one-half the sample frequency fs.

Figure 2-14 Gain and Phase Response of Second-Order Highpass IIR Filter

MOTOROLA 2-17
 Difference Equation

y ( n ) = 2 { α [ x ( n ) – 2 x ( n – 1 ) + x ( n – 2 ) ] + γy ( n – 1 ) – βy ( n – 2 ) }

Data Structures

X:(R0) Y:(R4)
−2α x(n-1)
α x(n-2)
γ
Y:(R5)
–β y(n-1)
α y(n-2)

DSP56001 Code
;Y1=x(n) (Input)
;X0=α
MPY X0,Y1,A X:(R0)+,X0 Y:(R4)+,Y0 ;A=αx(n)
MAC X0,Y0,A X:(R0)+,X0 Y:(R4),Y0 ;A=A-2αx(n-1)
MAC X0,Y0,A X:(R0)+,X0 Y:(R5)+,Y0 ;A=A+αx(n-2)
MAC X0,Y0,A X:(R0)+,X0 Y:(R5),Y0 ;A=A+γy(n-1)
MAC X0,Y0,A X:(R0)+,X0 Y1,Y:(R4) ;A=A-βy(n-2)
MOVE A,X1 A,Y:(R5) ;y(n)=2A (assumes scaling
;mode is set).
;X1 is Output.

Total Instruction Cycles

6 Icyc @ 20 MHz = 600ns

Figure 2-15 DSP56001 Code and Data Structures for Second-Order Direct-Form
Implementation of a Highpass IIR Filter

2-18 MOTOROLA
2.8 Digital Bandstop Filter
The formulas and network diagram for the digital
bandstop filter are presented in Figure 2-16. The
DSP56001 code from Figure 2-17 is identical to that
for the lowpass and highpass cases except for the
coefficient data calculated from the equations of
Figure 2-16. Scaling of this filter is not a problem for
the singlesection case since the gain from the
equation in Figure 2-16 never exceeds unity (as is
true in the analog case as seen by the gain equation
from Figure 1-5). Figure 2-18 is the calculated gain
and phase of the digital filter, which should compare
to the response curves of the equivalent analog fil-
ter plotted in Figure 1-6.

2.9 Digital Bandpass Filter

Because there is one less coefficient in the bandpass
network (see Figure 2-19), one instruction can be
saved in the DSP56001 code implementation shown
in Figure 2-20. Otherwise, the instructions are identi-
cal to those in the other three filter routines. Like the
second-order bandstop network, the maximum re-
sponse at the center frequency, θ0, is unity for any
value of Q so that scaling need not be considered in
the implementation of a single-section bandpass fil-
ter. This is true when the formulas for α, β, and γ
(from Figure 2-19) are used in the direct-form imple-
mentation in Figure 2-20. Figure 2-21 is the
calculated gain and phase of the digital filter, which
should compare to the response curves of the equiv-
alent analog filter plotted in Figure 1-8.

MOTOROLA 2-19
 Network Diagram α
x(n) 2 y(n)

z-1 z-1
−2αcosθ0 γ
Σ

z-1 z-1

α −β

Transfer Function –1 –2
α ( 1 – 2 cosθ 0 z + z )
H ( z ) = --------------------------------------------------------
-
–1 –2
1/2 – γ z + βz

Gain cos θ – cos θ c

G ( θ ) = --------------------------------------------------------------------------------------------------
2 2 1/2
[ ( d sin θ sin θ 0 ) + 4 ( cos θ – cos θ 0 ) ]

Phase
–1 2 ( cos θ – cos θ 0 )
φ ( θ ) = tan ----------------------------------------
-
d sin θ sin θ 0

Coefficients 2tan ( θ 0 /2Q )

d = -------------------------------
sinθ 0

1 1 – tan ( θ 0 /2Q )
β =  ---  --------------------------------------
 2  1 + tan ( θ 0 /2Q )

γ = ( 1/2 + β ) cos θ 0

α = ( 1/2 + β )/2
Figure 2-16 Direct-Form Implementation of Second-Order Bandstop IIR Filter and
Analytical Formulas Relating Desired Response to Filter Coefficients

2-20 MOTOROLA
 Difference Equation

y ( n ) = 2 { α [ x ( n ) – 2 cos θ 0 x ( n – 1 ) + x ( n – 2 ) ] + γy ( n – 1 ) – βy ( n – 2 ) }

Data Structures

X:(R0) Y:(R4)
−2αcosθ0 x(n-1)
α x(n-2)
γ
Y:(R5)
–β y(n-1)
α y(n-2)

DSP56001 Code
;Y1=x(n) (Input)
;X0=α
MPY X0,Y1,A X:(R0)+,X0 Y:(R4)+,Y0 ;A=αx(n)
MAC X0,Y0,A X:(R0)+,X0 Y:(R4),Y0 ;A=A-2αcosq0x(n-1)
MAC X0,Y0,A X:(R0)+,X0 Y:(R5)+,Y0 ;A=A+αx(n-2)
MAC X0,Y0,A X:(R0)+,X0 Y:(R5),Y0 ;A=A+γy(n-1)
MAC X0,Y0,A X:(R0)+,X0 Y1,Y:(R4) ;A=A-βy(n-2)
MOVE A,X1 A,Y:(R5) ;y(n)=2A (assumes scaling
; mode is set).
;X1 is Output.

Total Instruction Cycles

6 Icyc @ 20 MHz = 600ns

Figure 2-17 DSP56001 Code and Data Structures for Second-Order Direct-Form
Implementation of a Bandstop IIR Filter

MOTOROLA 2-21
 1.5
BANDSTOP GAIN

fo/fny=0.1 fo/fny=0.65
1.0

0.5 d=0.707
d=1.0
d=1.41
0 d=2.0

0.01 0.1 1.0

log (f/fny)
BANDSTOP PHASE

fo/fny=0.1 fo/fny=0.65
π/2

d=0.707
d=1.0
π d=1.41
d=2.0

0.01 0.1 1.0

log (f/fny)
NOTE: Nyquist frequency, fny, is equal to one-half the sample frequency fs.

Figure 2-18 Gain and Phase Response of Second-Order Bandstop IIR Filter

2-22 MOTOROLA
2.10 Summary of Digital
Coefficients
Figure 2-22 gives a summary of the coefficient val-
ues for the four basic filter types. Note that the
coefficient, β has the same form for all four filter
types and that it can only assume values between
0 and 1/2 for practical filters. β is bounded by 1/2
because Q (or d) and θ0 are not independent. For
Q >> 1, β→1/2; whereas, for θ0 = fs/4 and Q = 1/2,
β→0. These properties are independent of the form
of implementation; they are only dependent on the
form of the transfer function. Alternate implementa-
tions (difference equations) will be described in the
following sections.

Note that the Q described in Figure 2-22 meets the

traditional requirements (i.e., Q is the ratio of the
bandwidth at the -3 dB points divided by the center
frequency). The formula for β can be modified in
the case of the bandpass or bandstop filter by re-
placing the damping coefficient, d, with the formula
for Q. When the coefficients are described in this
manner, a constant Q filter results. When the
bandwidth is any function of center frequency, this
relationship between d and Q makes it impossible
to implement a bandpass or bandstop filter by re-
placing Q with the desired function of bandwidth
and center frequency.

Figure 2-23 shows the relationship between the

pole of the second-order section and the center fre-
quency. Note that the pole is on the real axis for
d>2, where d is also constrained by d < 2/sin θ0.

MOTOROLA 2-23
 Network Diagram
α
x(n) 2 y(n)

z-1 z-1
γ
Σ

z-1 z-1

−α −β

Transfer Function –2
α(1 – z )
H ( z ) = ---------------------------------------------
–1 –2
1/2 – γ z + βz

Gain d sin θ o sin θ

G ( θ ) = --------------------------------------------------------------------------------------------------
-
2 2 1/2
[ ( d sin θ sin θ 0 ) + 4 ( cos θ – cos θ 0 ) ]

Phase
–1 2 ( cos θ – cos θ 0 )
φ ( θ ) = – tan -----------------------------------------
d sin θ sin θ 0

Coefficients 2 tan ( θ 0 /2Q )

d = -----------------------------------
-
sinθ 0

1 1 – tan ( θ 0 /2Q )
β =  ---  --------------------------------------
 2  1 + tan ( θ 0 /2Q )

γ = ( 1/2 + β ) cos θ 0

α = ( 1/2 – β )/2

Figure 2-19 Direct-Form Implementation of Second-Order Bandpass IIR Filter and

Analytical Formulas Relating Desired Response to Filter Coefficients

2-24 MOTOROLA
 Difference Equation

y ( n ) = 2 { α [ x ( n ) – x ( n – 2 ) ] + γy ( n – 1 ) – βy ( n – 2 ) }

Data Structures

X:(R0) Y:(R4)
−α x(n-1)
γ x(n-2)
–β
Y:(R5)
α y(n-1)
y(n-2)

DSP56001 Code
;Y1=x(n) (Input)
;X0=α
MPY X0,Y1,A X:(R0)+,X0 Y:(R4),Y0 ;A=αx(n)
MAC X0,Y0,A X:(R0)+,X0 Y:(R5)+,Y0 ;A=A-αx(n-2)
MAC X0,Y0,A X:(R0)+,X0 Y:(R5),Y0 ;A=A+γy(n-1)
MAC X0,Y0,A X:(R0)+,X0 Y1,Y:(R4)+ ;A=A-βy(n-2)
MOVE A,X1 A,Y:(R5) ;y(n)=2A (assumes scaling
; mode is set).
;X1 is Output.

Total Instruction Cycles

5 Icyc @ 20 MHz = 500ns

Figure 2-20 DSP56001 Code and Data Structures for Second-Order

Direct-Form Implementation of a Bandpass IIR Filter

MOTOROLA 2-25
 fo/fny=0.1 fo/fny=0.65
1.5
BANDPASS GAIN

d=0.707
d=1.0
d=1.41
1.0 d=2.0

0.5

0.01 0.1 1.0

log (f/fny)
BANDPASS PHASE

π/2
d=0.707
d=1.0
d=1.41
d=2.0
0
fo/fny=0.1 fo/fny=0.65

-π/2

0.01 0.1 1.0

log (f/fny)
NOTE: Nyquist frequency, fny, is equal to one-half the sample frequency fs.

Figure 2-21 Gain and Phase Response of Second-Order Bandpass IIR Filter

2-26 MOTOROLA
 Z-Domain Transfer Function

–1 –2
α ( 1 + µz + σz )
H ( z ) = -------------------------------------------------
–1 –2
1/2 – γ z + βz

Difference Equation (Direct Form)

y ( n ) = 2 { α [ x ( n ) + µx ( n – 1 ) + σx ( n – 2 ) ] + γy ( n – 1 ) – βy ( n – 2 ) }

Coefficients

1 1 – 1/2dsinθ 0 2tan ( θ 0 /2Q )

β = --- ----------------------------------- d = -------------------------------- γ = ( 1/2 + β ) cos θ 0
2 1 + 1/2dsinθ 0 sin θ 0

where 0 < β < 1/2 and θ 2πf 0 /f s f0

Q = -----0- = --------------------------------
- = --------------
-
∆0 2π ( f 2 – f 1 )/f s f2 – f1

where f0 is the center frequency of the bandpass or bandstop filter,

f1 and f2 are the half-power points (where gain is equal to 1/ 2), and
fs is the sample frequency. Note the f0 is replaced with fc in the lowpass
and highpass cases.

Numerator Coefficients
Type α µ σ Unity Gain at

Lowpass (1/2 + β−γ)/4 2 1 f=0

Highpass (1/2 + β+γ)/4 -2 1 f = fs/2
Bandpass (1/2 - β)/2 0 -1 f = f0
Bandstop (1/2 + β)/2 -2cosθ0 1 f = 0 and f = fs/2

NOTE: θ0 = 2πf0/fs

Figure 2-22 Summary of Digital Coefficients for the Four Basic Filter Types

MOTOROLA 2-27
 Pole Equation of H(z) Z p = r cos θ p + jr sin θ p

2
= γ ± j 2β – γ
For d<2
2
cos θ 0 ± j sin θ 0 1 – ( 1/2d )
= --------------------------------------------------------------------
1
1 + --- d sin θ 0
2
1 γ = ( 1/2 + β ) cos θ 0
where: β = --- ( 2 – d sin θ 0 )/ ( 2 + d sin θ 0 ) and
2
Distance from Origin to Pole is Z p = 2β

2
For d>2 Z p = γ – γ – 2β

2
cos θ 0 – sin θ 0 ( 1/2d ) – 1
= -----------------------------------------------------------------
-
1
1 + --- d sin θ 0
2
1
where: θ p = 0 To satisfy requirement 0<β<1/2 results in --- d sin θ 0 < 1
2

Im[z]

Center or Cutoff
Z0 Frequency

Zp θ0
Pole
θp

Re[z]
r = 2β

Figure 2-23 Pole Location and Analysis of Second-Order Section

2-28 MOTOROLA
APR7section3 Page 1 Monday, June 2, 1997 1:51 PM

SECTION 3

Single-Section
Canonic Form
(Direct Form II)

“The canonic The single-section canonic form network is dis-

(direct form II) cussed in the following paragraphs.
network has
trade-offs that
must be
carefully
understood and 3.11 The Canonic-Form
analyzed for the
particular
Difference Equation
application.” The direct-form difference equation, rewritten from
Eqn. 2-18, is:
2 2
y(n) = ∑ bi x ( n – 1 ) – ∑ ajy ( n – 1 ) Eqn. 3-20
i=0 j=1

Eqn. 3-20 can be represented by the diagram of Fig-

ure 3-24 (a). This diagram is the same as those
shown in Figure 2-10, Figure 2-13, Figure 2-16, and
Figure 2-19, except that the summations have been
separated to highlight the correspondence with Eqn.
3-20. From this diagram, it is clear that the direct-form
implementation requires four delay elements or,
equivalently, four internal memory locations.

MOTOROLA 3-1
 x(n) b0 y(n)
+ +

z-1 z-1
b1
x(n-1) + + y(n-1)
-a1
z-1 z-1
b2
x(n-2) y(n-2)
-a2
(a) Direct-Form Network Diagram of Equation 3-1

x(n) b0 y(n)
+ +

z-1 z-1
b1
+ +
-a1
z-1 z-1
b2
-a2
(b) Interchanged Left and Right Halves of Direct-form Network

x(n) w(n) b0 y(n)

+ +

z-1
b1
+ +
-a1
z-1
b2
-a2 w(n-2)
(c) Collapsing Delay Terms (part ‘a’ above) Resulting in Canonic-Form Diagram

Figure 3-24 The Second-Order Canonic (Direct Form II) IIR Filter Network

3-2 MOTOROLA
The diagram of Figure 3-24(b) represents the same
transfer function implemented by Eqn. 3-20, but
now the delay variable is w(n). Comparison of this
network with that of the direct-form network of Fig-
ure 3-24 (a) shows that interchanging the order of
the left and right halves does not change the overall
system response (see Reference 11).

The delay elements can then be collapsed to pro-

duce the final canonic-form network shown in Figure
3-24 (c). As a result, the memory requirements for
the system are reduced to the minimum (two loca-
tions); therefore, this realization of the IIR filter is
often referred to as the canonic form. The system
difference equations for the canonic realization are:
2
Eqn. 3-21a
y(n) = ∑ bi ( w ( n – i ) )
i=0
where:
2
w(n) = x(n)– ∑ aj ( w ( n – j ) ) Eqn. 3-2b
j=1

To prove that Eqn. 3-21a and Eqn. 3-2b are equiv-

alent to Eqn. 3-20, the following procedure can be
used. First, combine Eqn. 3-21a and Eqn. 3-2b:
2 2
y(n) = ∑ bi x ( n – i ) – ∑ ajw ( n – j – i )
i=0 j=1
2 2 2
= ∑ bi x ( n – i ) – ∑ aj ∑ bi w ( n – i – j )
i=0 j=1 i=0
2 2
= ∑ bi x ( n – i ) – ∑ ajy ( n – j ) Eqn. 3-3
i=0 j=1

MOTOROLA 3-3
 The last step uses the definition for y(n) from Eqn.
3-21a. The result is exactly equivalent to Eqn. 3-
20, the direct-form difference equation. To utilize
the scaling mode on the DSP56001, it is advanta-
geous to write Eqn. 3-21a and Eqn. 3-2b as
follows:

1 µ α 
y ( n ) = 2  --- w ( n ) + --- w ( n – 1 ) + ---w ( n – 2 ) 
 2 2 2 
Eqn. 3-4a
and
w ( n ) = 2 { αx ( n ) + γw ( n – 1 ) – βw ( n – 2 ) } Eqn. 3-4b

where: the coefficients have been substituted for

the aj and bj

From these equations, it can be seen that both y(n)

and w(n) depend on a sum of products and there-
fore are the output of an accumulator. The
accumulators on the DSP56001 have eight exten-
sion bits; thus, the sum can exceed unity by 255
without overflowing. However, when the contents of
a 56-bit accumulator are transferred to a 24-bit reg-
ister or memory location (i.e., delay element), an
overflow error may occur. The digital filter designer
must insure that only values less than unity are
stored. Of course, even if memory had the precision
of the accumulators, overflow can still occur if the
accumulator itself experiences overflow. However,
intermediate sums are allowed to exceed the ca-
pacity of the accumulator if the final result can be
represented in the accumulator (a result of the cir-

3-4 MOTOROLA
cular nature of twos-complement arithmetic).
To insure that overflow does not occur, it is necessary
to calculate the gain at the internal nodes (accumula-
tor output) of a filter network. Although canonic
realization is susceptible to overflow, it is advanta-
geous because of the minimum storage requirements
and implementation in fewer instruction cycles.

3.12 Analysis of Internal

Node Gain
Calculating the gain at any internal node is no differ-
ent than calculating the total network gain. In Figure
3-25, the transfer function of node w(n) is Hw(z).
Hw(z) represents the transfer function from the input
node, x(n), to the internal node, w(n). The gain,
Gw(θ), at w(n) is found in the standard manner by
evaluating the magnitude of Hw(z) as shown in Fig-
ure 3-26. Note that the flow diagram of Figure 3-25
uses the automatic scaling mode (multiply by a fac-
tor of two when transferring data from the
accumulator to memory) so that the coefficients are
by definition less than one. The peak gain, go, of
Gw(θ) is found by taking the derivative of Gw(θ) with
respect to θ and setting the result equal to zero. A
simple expression for go is derived as follows:
2
1 --- + β
α  2 
g o = ------------------------------ for --------------------- cos θ o ≤ 1 Eqn. 3-5
1  2β
--- – β sin θp
2 

MOTOROLA 3-5
 where: θp is the angle (see Figure 2-23) to the
pole of the filter

3-6 MOTOROLA
 Y(z)
H ( z ) = ------------
X(x)
X(z) Y(z)
Y(z)
H w ( z ) = ------------
X(z)
α W(z)
x(n) + + 2 y(n)
(Left Shift)
2 (Left Shift)
W(z)
w(n)

w ( z ) ) z-1
--------------
z w(n-1)
+ +
γ µ
---
w ( z ) ) z-1 2
--------------
2 w(n-2)
z
−β w(n-2) σ
---
2

α ( 1 + µz + σz )
–1 –2 α
H ( z ) = ------------------------------------------------- H w ( z ) = --------------------------------------
1 1 –1 –2
--- – γ z + βz
–1 –2 --- – γ z + βz
2 2

Figure 3-25 Internal Node Transfer Function, Hw(z), of Canonic

(Direct Form II) Network

MOTOROLA 3-7
 Since this result does not depend on the numerator
of the filter transfer function, H(z), the result is valid
for all four basic filter types. However, as shown by
the example given for a bandpass network in Figure
3-26, go may exceed unity by a large amount, espe-
cially for filters having poles at frequencies much
less than fs/2 where sin θp << 1. To compensate,
the input must be scaled down by an amount equal
to 1/go or guaranteed not to exceed 1/go before ar-
riving at the filter. This scaling aspect of the
canonic-form network is a disadvantage, but this
network has the advantage of being implemented in
one less instruction than the other filter realizations
for the lowpass, highpass, and bandstop filters; of
course, less memory is required since only two inter-
mediate variables are stored.

In general, the behavior of systems at internal nodes

can be unexpected. For instance, it is interesting to
note that the frequency at which the gain at the inter-
nal node of the canonic IIR filter section peaks is not
the same as the frequency at which the gain of the
filter peaks as given by the poles of the filter. From
Figure 2-13, this behavior is expressed by:

2β cos θ o
------------- = --------------- Eqn. 3-6
1 cos θ p
--- + β
2

The canonic (direct form II) network has trade-offs

that must be carefully understood and analyzed for
the particular application.

3-8 MOTOROLA
 Gain at w(n) is: W(z) α
H w ( z ) = ------------- = ---------------------------------------
X(z) 1 –1 –2
--- – γ z + βz
2

jθ – jθ
Gw ( θ ) = H w ( e )H w ( e )
α
= --------------------------------------------------------------------------------------------------------------
2 2
1 --- – β sin θ +  --- + β ( cosθ – cosθ )
2 1 2
2  2  0
1
where: γ = ( -2- + β) cos θ0 has been used.

d
Peak Gain is: ------G ( θ ) = 0
dθ w θ = θm

Frequency of peak gain is: cosθ m = ξcosθ 0 forξcosθ 0 ≤ 1

2 =1 otherwise
1 --- + β
2 
where: ξ ≡ ---------------------
2β
α forξcosθ 0 ≤ 1
g 0 ≡ G w ( θ m ) = ------------------------------------
2
1 γ
--- – β 1 – ------
2  2β

α otherwise
= ------------------------------------------------
1 --- + β ( 1 – cos θ 0 )
2 

EXAMPLE: Maximum Internal Node Gain for Bandpass Filter

1
g 0 = ------------------
2sinθp forξcosθ 0 ≤ 1

where γ2 = 2β2 cos2θp has been used and α= (1/2-β)/2 for a bandpass filter.
If sinθp > 1/2, then g0 <1; otherwise, an overflow (i.e., g0>1) may occur at the
internal node, w(n), unless the input is scaled down by 1/g0.

Figure 3-26 Internal Node Gain Analysis of Second-Order Canonic Form

MOTOROLA 3-9
 Difference Equation

1 µ σ 
y ( n ) = 2  --- w ( n ) + --- w ( n – 1 ) + ---w ( n – 2 ) 
 2 2 2 

w ( n ) = 2 { αx ( n ) + γw ( n – 1 ) – βw ( n – 2 ) }

Data Structures

X:(R0) Y:(R4)
w(n-1) γ
w(n-2) −β
µ/2
σ/2
α

DSP56001 Code
;Y1=x(n) (Input)
;X0=α
MPY X0,Y1,A X:(R0)+,X0 Y:(R4)+,Y0 ;A=αx(n)
MAC X0,Y0,A X:(R0),X1 Y:(R4)+,Y0 ;A=A+γw(n-1)
MACR X1,Y0,A X0,X:(R0)- Y:(R4)+,Y0 ;A=A-βw(n-2)
MAC X0,Y0,A A,X:(R0) Y:(R4)+,Y0 ;A=1/2w(n)+µ/2w(n-1)
MACR X1,Y0,A A,X0 Y:(R4)+,Y0 ;A=A+σ/2w(n-2). X0=2A
;(assumes scaling mode
;is set).X0 is Output.

Total Instruction Cycles

5 Icyc @ 20 MHz = 500ns

Figure 3-27 Internal Node Gain Analysis of Second-Order Canonic Form

3-10 MOTOROLA
 3.13 Implementation on the
DSP56001
Figure 3-27 shows the DSP56001 code and data
structures for implementation of the single-sec-
tion canonic-form network. Note that the modifier
register M4 is set equal to 4 to allow circular op-
eration for addressing coefficient data. M0 is set
to FFFF to turn off circular addressing for w(n-1)
and w(n-2). ■

3-11 MOTOROLA
3-12 MOTOROLA
 SECTION 4

Single-Section
Transpose Form
A third realization of IIR filters is the transpose form
“The modifier (direct form l) shown in Figure 4-1. This network imple-
register M4 is mentation can be derived directly from the direct-form
initially set to a difference equation (see Figure 4-1) or by taking the
value of 4 so that transpose of the canonic network (see References 10
a circular buffer and 11). The transpose realization is characterized by
can be three accumulator operations. One reason for the pop-
conveniently ularity of this realization is that, like the canonic
used to address realization, it only requires two memory locations;
the coefficient however, unlike the canonic realization, it is much less
data.” prone to overflow at internal nodes. The disadvantage
is that this realization requires more instructions to im-
plement.

4.14 Gain Evaluation of

Internal Nodes
Using the same techniques used to calculate Hw(z) for
the canonic realization, Hu(z) and Hv(z) are found as
shown in Figure 4-2 and Figure 4-3. The resulting ex-
pressions, unlike the canonic-form results, depend on
the numerator of the transfer function; thus, the internal
gains, Gu(θ) and Gv(θ), have different forms for the dif-
ferent filter types. In the case of the bandpass filter,
these results simplify significantly so that a closed-form

MOTOROLA 4-1
 expression for the maximum gain at the internal
nodes can be derived by calculating the maxima of
the gain functions. For the bandpass and bandstop
networks, the maximum of Gu(θ) and Gv(θ) is gm =
β + 1/2. Since, β < 1/2, then gm < 1 so that no over-
flow occurs at these nodes in the bandpass or
bandstop case.
Figure 4-4 contains example plots of Gu(θ) and Gv(θ)
for the second-order transpose-form lowpass filter
with various values of cutoff frequency, θc, and
damping factor, d. In most cases, Gu(θ) and Gv(θ)
never exceed the maximum value of G(θ), so that, if
the total gain does not exceed unity, the internal
nodes will not exceed unity (i.e., no overflow).

4.15 Implementation on the

DSP56001
The DSP56001 code and data structures for a single-
section second-order transpose-form network are
shown in Figure 4-5. Referring to the network diagram
of Figure 4-1, the diamond blocks enclosing 1/2 are
represented in the code by an accumulator shift right
(ASR) instruction. The 2 that is enclosed by a dia-
mond block can be implemented by the automatic
scaling mode (equivalent to a left shift of the accumu-
lator) feature of the DSP56001. The modifier register
M4 is initially set to a value of 4 so that a circular buffer
can be conveniently used to address the coefficient
data. Note that this network requires more instructions
than either of the previous two network forms.

4-2 MOTOROLA
 Network Diagram ασ −β
Σ

(Left Shift) 2
u(n)
z-1

1/2 (Right Shift)

αµ γ
x(n) Σ y(n)

(Left Shift) 2
v(n)
-1
z

(Right Shift) 1/2

α (Left Shift)
Σ 2

Direct-Form Difference Equation

y(n) = 2 { α [ x ( n ) + µx ( n – 1 ) + σ x ( n – 2 ) ] + γy ( n – 1 ) – βy ( n – 2 ) }

= 2 { αx ( n ) + αµx ( n – 1 ) + γy ( n – 1 ) + [ ασx ( n – 2 ) – βy ( n – 2 ) ] }
 1 
= 2  αx ( n ) + [ αµx ( n – 1 ) + γy ( n – 1 ) + -2 u ( n – 2 ) ] 
 
 1 
= 2  αx ( n ) + -2- v ( n – 1 ) 
 

Transpose-Form Difference Equation

 1 
y ( n ) = 2  αx ( n ) + - v ( n – 1 ) 
2
 1 
where v ( n ) = 2  αµx ( n ) + γy ( n ) + - u ( n – 1 ) 
2
 
and u ( n ) = 2 { ασx ( n ) – β y ( n ) }

Figure 4-1 Network Diagram for Transpose-Form (Direct Form I)

Implementation of a Single-Section Second-Order Filter

MOTOROLA 4-3
 GAIN AT INTERNAL NODE, u(n)
Transfer function of u(n) is U(z) = 2 [ ασX ( z ) – βY ( z ) ]
= Hu ( z ) X ( z )

where H u ( z ) = 2 [ ασ – βH ( z ) ]
–1
ασ ( 1/2 – γ z + βz –2 ) – αβ ( 1 + µz –1 + σ z –2 )
= 2 -------------------------------------------------------------------------------------------------------
–1 –2
1/2 – γ z + βz

–1
2α ( A – Bz )
= 2 --------------------------------------
–1 –2
1/2 – γ z + βz
where Hu(z) is the transfer function from input node, x(n), to internal mode, u(n), and
A=σ/2-β and B = σγ + µβ.

Gain of u(n) is Gu ( θ ) = Hu ( z )
jθ
z=e

2 2
2α A + B – 2AB cos θ
= -----------------------------------------------------------------------------------------------------------
2 2 2
( 1/2 – β ) sin 2 θ + ( 1/2 + β ) ( cos θ – cos θ 0 )

BANDPASS EXAMPLE
Bandpass Coefficients are σ = –1 µ = 0 α = ( 1/2 – β ) /2
so that A = – ( 1/2 + β )
B = –γ
= – ( 1/2 + β ) cos θ 0
2
( 1/2 – β ) ( 1/2 + β ) sin 2 θ + ( cos θ – cos θ 0 )
and Gu ( θ ) = -----------------------------------------------------------------------------------------------------------
-
2 2 2
( 1/2 – β ) sin 2 θ + ( 1/2 + β ) ( cos θ – cos θ 0 )

Peak Gain is gm = Gu ( θm )

d
which is found by evaluating ----- G u ( θ ) θ = θ = 0
dθ m

after evaluating derivative, θ m = θ 0

( 1/2 – β ) ( 1/2 + β ) sin θ 0

so that g m = ------------------------------------------------------------ = 1 ⁄ 2 + β < 1
( 1 ⁄ 2 – β sin θ 0 )

Figure 4-2 Gain Evaluation at First Internal Node, u(n), of Transpose Network

4-4 MOTOROLA
 GAIN AT INTERNAL NODE, v(n)
Transfer function of v(n) is V(z) = [ Y ( z ) – 2αX ( z ) ] /z –1
= Hv ( z ) X ( z )
1 H ( z ) – 2α
Hv ( z ) =
–1
where Y(z)= 2[ - V(z)z + αX ( z ) ] ----------------------
2 –1
z
–1 –2 –1 –2
α [ ( 1 + µ z + σz ) – 2 ( 1/2 + γ z + βz ) ]
and = -------------------------------------------------------------------------------------------------------
–1 –1 –2
z ( 1/2 – γ z + βz )
–1
2α ( C + Az )
= --------------------------------------
–1 –2
1/2 – γ z + βz
where Hv(z) is the transfer function from input node, x(n), to internal mode, v(n), and
A=σ/2-b and C = µ/2 + g.
Gain of v(n) is Gv ( θ ) = Hv ( z )
jθ
z=e

2 2
2α A + C + 2AC cos θ
= -----------------------------------------------------------------------------------------------------------
2 2 2
( 1/2 – β ) sin 2 θ + ( 1/2 + β ) ( cos θ – cos θ 0 )

BANDPASS EXAMPLE
Bandpass Coefficients are σ = –1 µ = 0 α = ( 1/2 – β ) /2
so that A = – ( 1/2 + β )
C = γ
= ( 1/2 + β ) cos θ 0
2
( 1/2 – β ) ( 1/2 + β ) sin 2 θ + ( cos θ – cos θ 0 )
and Gv ( θ ) = -----------------------------------------------------------------------------------------------------------
2 2 2
( 1/2 – β ) sin 2 θ + ( 1/2 + β ) ( cos θ – cos θ 0 )

Peak Gain is gm = Gv ( θm )

d
which is found by evaluating ----- G v ( θ ) θ = θ = 0
dθ m

after evaluating derivative, θ = θ

m 0

( 1/2 – β ) ( 1/2 + β ) sinθ

so that gm = ---------------------------------------------------0- = 1/2 +β<1
( 1/2 – β ) sinθ 0
(since β< 1/2

Figure 4-3 Gain Evaluation at Second Internal Node, v(n), of Transpose Network

MOTOROLA 4-5
 1.2
1.0
G
0.8
0.6
Gain

0.4
Gv
0.2
Gu
0.0
-0.2
0 50 100 150 200 250 300 350 400 450 500
Frequency

(a) fc = 450 Hz and d= 2

1.1
G
1.0
0.9
0.8
0.7 Gv
0.6
Gain

0.5
0.4
0.3
0.2 Gu
0.1

0 50 100 150 200 250 300 350 400 450 500

Frequency

(b) fc = 250 Hz and d= 2

2.5

2.0

1.5
G
Gain

1.0 Gv
0.5 Gu
0.0

0 50 100 150 200 250 300 350 400 450 500

Frequency

(c) fc = 250 Hz and d= 1/2

Figure 4-4 Total Gain and Gain at Internal Nodes of Lowpass Transpose
Filter Network Figure 4-1 for fs = 1000Hz

4-6 MOTOROLA
 Difference Equation
 1 
y ( n ) = 2  αx ( n ) + --- v ( n – 1 ) 
 2 
 1 
v ( n ) = 2  αµx ( n ) + γy ( n ) + --- u ( n – 1 ) 
 2 
u ( n ) = 2 { ασx ( n ) – βy ( n – 1 ) }

Data Structures

X:(R0) Y:(R4)
v(n-1) αµ
γ
ασ
X:(R1)
u(n-1) –β
α

DSP56001 Code
;Y1=x(n) (Input)
;Y0=α . A=v(n-1)/2
MACR Y0,Y1,A X:(R1),B Y:(R4)+,Y0 ;A=A+αx(n)
ASR B A,X0 ;y(n)=X0=2A (scale mode on)
MAC Y0,Y1,B Y:(R4)+,Y0 ;B=u(n-1)/2+αµx(n)
MACR X0,Y0,B Y:(R4)+,Y0 ;B=B+γy(n)
MPY Y0,Y1,B B,X:(R0) Y:(R4)+,Y0 ;y(n)=2B. B=ασx(n)
MACR X0,Y0,B X:(R0),A ;B=B- βy(n)
ASR A B,X:(R1) Y:(R4)+,Y0 ;A=v(n)/2. u(n)=2B
;Y0=α. y(n)=X0 (output)

Total Instruction Cycles

7 Icyc @ 20 MHz = 700ns

Figure 4-5 DSP56001 Code and Data Structures for Single-Section

Second-Order Transpose Form

MOTOROLA 4-7
4-8 MOTOROLA
 SECTION 5

Cascaded Direct Form

“. . . the By placing any of the direct-form second-order filter
cascaded direct- networks from Figure 2-10, Figure 2-13, Figure 2-16,
form network and Figure 2-19 in series (i.e., connect the y(n) of one
becomes to the x(n) of the next), a cascaded filter is created.
canonic as the The resulting order N of the network is two times the
filter order N number of second-order sections. An odd-order net-
increases.” work can be made simply by adding one first-order
section in the chain. In general, to achieve a particular
response, the filter parameters associated with each
second-order section are different, since generating a
predefined total response requires that each section
has a different response. This fact becomes more ob-
vious when discussing the special case of Butterworth
lowpass filters.

5.16 Butterworth Lowpass Filter

The Butterworth filter response is maximally flat in the
passband at the expense of phase linearity and steep-
ness of attenuation slope in the transition band. For
lowpass or highpass cascaded second-order sections,
all sections have the same center frequency (not the
case for bandpass filters). For this reason, it is easy to
design since all that remains to be determined are the

MOTOROLA 5-1
 damping factors, dk, of each individual kth section.
The damping coefficients, dk, are calculated from a
simple formula for any order N of response.

The s-domain transfer function for the Nth-order

lowpass Butterworth filter is:

  
H ( s ) =  ---------------------------------------------------------------  ---------------------------------------------------------------------- …
1 1
 2  2 
 ( s/Ω c ) + d 1 ( s/Ω c ) + 1   ( s/Ω c ) + d 2 ( 1 ) ( s/Ω c ) + 1 

N/2
∏
1
= -------------------------------------------------------------- Eqn. 5-7
2
k = 1 ( s/Ω c ) + d k ( s/Ω c ) + 1

where: dk is the kth damping coefficient,

s = jΩ
Ωc is the common cutoff frequency
(see Reference 14).

Only filter orders of even N will be considered in this

discussion to minimize the complexity of mathemat-
ical results. The analysis can be extended to
include odd values of N by inserting an additional
term of [(s/Ωc) + 1]-1 in Eqn. 5-7. Eqn. 5-7 can be
generalized if Ωc Of each second-order section is
an arbitrary value (corresponding to a different cut-
off frequency, Ωk, of each section). However, since
this discussion is limited to Butterworth polynomi-
als, Eqn. 5-7 will serve as the basis of all following
derivations. A filter of order N has N/2 second-order

5-2 MOTOROLA
sections. The second-order section of a Butterworth
filter can be derived from the simple RCL network of
Figure 1-1. However, the Butterworth damping factors
are predetermined values, which can be shown to
yield a maximally flat passband response (see Refer-
ence 14). The Butterworth damping coefficients are
given by the following equation:

( 2k – 1 )π
d k = 2 sin ------------------------ Eqn. 5-8
2N

Eqn. 5-8 is the characteristic equation that deter-

mines a Butterworth filter response. Note that for a
single-section (k = 1) second-order (N = 2) lowpass
filter, dk = 2 sin (π/4) = 2 as expected for a maxi-
mally flat response. For a fourth-order filter with two
second-order sections, d1 = 2 sin (π/8), where k = 1
and N = 4, and d2 = 2 sin (3π/8),
where k= 2 and N = 4.

Eqn. 5-7 represents an all-pole response (the only

zeros are at plus and minus infinity in the analog s-
domain). The poles of a second-order section are
the roots of the quadratic denominator as given by:

2 1/2
p k1 = – d k /2 – j 1 – d k /4 (a)
 
and
2 1/2
p k 2 = d k /2 + j 1 – d k /4 (b)
 
Eqn. 5-9

MOTOROLA 5-3
 Using , Eqn. 5-7 becomes:

N/2

∏ [--------------------------------------------------------------------
1
H(s) = - Eqn. 5-10
( s/Ω ) – p ] [ ( s/Ω ) – p ∗ ]
c k c k
k=1

where: pk = pk1 and

pk* = pk2 (complex conjugate of pk)

Eqn. 5-10 is useful in that the response of the sys-

tem can be analyzed entirely by studying the poles
of the polynomial. However, for purposes of trans-
forming to the z-domain, Eqn. 5-7 can be used as
previously shown in Figure 2-10.

To examine the gain and phase response (physically

measurable quantities) of the lowpass Butterworth
filter, the transfer function, H(s), of Eqn. 5-7 will be
converted into a polar representation. The magni-
tude of H(s) is the gain, G(Ω); the angle between the
real and imaginary components of H(s), φ(Ω), is the
arctangent of the phase shift introduced by the filter:

G(Ω) = H ( s )H∗ ( s )
s = jΩ

N/2
∏
1 Eqn. 5-11
= ---------------------------------------------------------------------------------
2
k = 1 ( Ω/Ω ) 2 – 1 + ( d Ω/Ω ) 2
c k c

5-4 MOTOROLA
 N/2 d k Ω/Ω c
–1
φ(Ω) = ∑ tan --------------------------------
2
Eqn. 5-12
k=1 ( Ω/Ω c ) – 1

Eqn. 5-11, and Eqn. 5-12 describe the response

characteristics of an Nth-order lowpass Butterworth
filter in the continuous frequency analog do-
main.Since the quantity of interest is usually 20 log
G(Ω), Eqn. 5-11 can be transformed into a sum
(over the second-order sections) of 20 log (Gk),
where Gk is the gain of the kth section. Similarly, the
total phase is just the sum of the phase contribution
by each section from Eqn. 5-12.

The bilinear transformation is used to convert the

continuous frequency domain transfer function into
the digital domain representation, where θ, the nor-
malized digital domain frequency equal to 2πf/fs,
can be thought of as the ratio of frequency to sam-
pling frequency scaled by 2π. Substituting s from
Eqn. 2-13 and Ωc from Eqn. 2-14 into the kth section
of Eqn. 5-7 yields the digital domain form of the But-
terworth lowpass filter (for the kth second-order
section):

–1 –2
α k ( 1 + 2z + z )
H k ( z ) = -------------------------------------------------- Eqn. 5-13
1 –1 –2
--- – γ z + β z
2 k k

MOTOROLA 5-5
 2
α k = tan ( θ c /2 ) /A k ( c ) (a)

2
β k = 1 – d k tan ( θ c /2 ) + tan ( θ c /2 ) /A k ( c ) (b)

2
γ k = 2 1 – tan ( θ c /2 ) /A k ( θc ) (c)

2
A k ( θ c ) = 2 1 + d k tan ( θ c /2 ) + tan ( θ c /2 ) (d)
Eqn. 5-14

(a)-(d) provides a complete description of the digital

lowpass Nth-order Butterworth filter. Given θc and
dk, these formulas allow precise calculation of the
digital coefficients, αk, βk, and γk, used to implement
each kth second-order section of the filter.
(a)-(d) can be further simplified into the following set
of formulas:

1 – ( d k /2 ) sin ( θ c )
β k = ------------------------------------------------------ (a)
2 [ 1 + ( d k /2 ) sin ( θ c ) ]

γ k = ( 1/2 + β k ) cos ( θ c ) (b)

α k = ( 1/2 + β k – γ k )/4 (c)

Eqn. 5-15
where: dk is given by Eqn. 5-8
θc is the digital domain cutoff frequency
(actual operating cutoff frequency of
the digital filter)

5-6 MOTOROLA
 20

15

10
k=1
20log(G) (db)

0
k=2
-5
k=3
Composite
-10
Response
-15

-20
0 100 200 300 400 500 600 700
Frequency (Hz)

(a) Analog Case

20

15

10
20log(G) (db)

0 k=1
k=2
-5 k=3
-10
Composite
-15 Response

-20
0 100 200 300 400 500 600 700
Frequency (Hz)

(b) Digital Filter with fs = 1200 Hz and fc = 425 Hz

Figure 5-28 Composite Response of Cascaded Second-Order Sections for

Lowpass Butterworth Filter (kth Section Damping Factor According to
Eqn. 5-8)
Figure 5-28 shows an example of a sixth-order low-

MOTOROLA 5-7
 pass Butterworth filter (three second-order sections)
in both the analog domain and digital domain. Note
that the gain of the first section (k = 1) is greater than
unity near the cutoff frequency but that the overall
composite response never exceeds unity. This fact
allows for easy implementation of the Butterworth fil-
ter in cascaded direct form (i.e., scaling of sections
is not needed as long as the sections are imple-
mented in the order of decreasing k). Overflow at the
output of any section is then guaranteed not to occur
(the gain of the filter never exceeds unity). Note that
the digital response (see Figure 5-28) is identical to
the analog response but warped from the right along
the frequency axis. Imagine the zero at plus infinity
in the analog response mapping into the zero at fs/2
in the digital case. Also note that, because of this
mapping, the digital response falls off faster than the
-12 dB/octave of the analog filter when the cutoff is
near fs/2.

The previous analysis is nearly identical to the case

of the highpass filter except the coefficients (see
Figure 3-24) have slightly different values. Since
the bandstop case is just the sum of a lowpass and
highpass case, it can be analyzed by these tech-
niques. The bandpass case, however, is more
difficult and requires considerably more work (see
Reference 14) because the center frequency of
each section is now different and the formula for
calculating these frequencies is not as simple as
the formulas for the previous filter types. In addi-
tion, to complicate matters further, scaling between
sections becomes more of a problem since the off-

5-8 MOTOROLA
set of the center of each section reduces the final
center response, which must be compensated at
some point in the filter network. For cases such as
higher order Butterworth bandpass filter designs,
commercially available filter design packages such
as FDAS are useful. The use of FDAS is discussed
in SECTION 6 Filter Design And Analysis Sys-
tem (FDAS) and SECTION 7 Fir Filters.

5.17 Cascaded Direct-

Form Network
Figure 5-29 shows the cascaded direct-form net-
work and data structures for the DSP56001 code
implementation of Figure 5-30. By cascading net-
work diagrams presented in SECTION 2 Second-
Order Direct-Form IIR Digital Filter Sections, the
set of delays at the output of one section can be
combined with the set of delays at the input of the
next section, thus reducing the total number of de-
lays by almost a factor of two. For this reason, the
cascaded direct-form network becomes canonic as
the filter order N increases. The DSP56001 code
(see Figure 5-30) shows an example of reading
data from a user-supplied memory-mapped analog-
to-digital converter (ADC) and writing it to a memo-
ry-mapped digital-to-analog converter (DAC). The
number of sections (nsec) in these examples is
three; thus, the filter order is six. The total instruction
time for this filter structure is 600 nsec + 800 ns, in-
cluding the data l/O moves (but excluding the
interrupt overhead).

MOTOROLA 5-9
 Network Diagram
x(n) α1 α2 α3 y(n)
2 2 2

z-1 z-1 z-1 z-1

α1µ1 γ1 α2µ2 γ2 α3µ3 γ3
Σ Σ Σ

z-1 z -1
z-1 z-1
α1σ1 −β1 α2σ2 −β2 α3σ3 −β3

Difference Equations
y1 ( n ) = 2 { αi [ xi ( n ) + µi xi ( n + 1 ) + σi xi ( n – 2 ) ] + γ i xi ( n – 1 ) – βi yi ( n – 2 ) }
xi + 1 ( n – k ) = yi ( n – k )

Data Structures coeff xbuf

X:(R0) Y:(R4)
−β1 x1(n-2) x(n-2)
α1 x1(n-1) x(n-1)
α1 σ1 x2(n-2) y1(n-2)
α1 µ1 x2(n-1) = y1(n-1)
γ1
x3(n-2) y2(n-2)
−β2 x3(n-1) y2(n-1)
α2
y(n-2) y3(n-2)
α2 σ2
y(n-1) y3(n-1)
α2 µ2
γ2
−β3
α3
α3 σ3
α3 µ3
γ3

Figure 5-29 Network Diagram for Cascaded Direct-Form Filter and Data
Structures Used in Code Implementation (Three-Section Example)

5-10 MOTOROLA
 DSP56001 CODE

MOVE #coeff, R0 ;Pointer to coefficients

MOVE #1,M4 ;Modulo of length 2 for xbuf
MOVE X:(R0)+, X1 ;β1
MOVE X:(R0)+, X0 ;α1
MOVE #2,N4
MOVE (R4)+ ;Point to next xbuf entry
;Input:Y1=x(n)
DO X: nsec, Sectn ;Loop on number of sections
MPY X0,Y1,A X:(R0)+,X0 Y:(R4)+,Y0 ;A=αixi(n)
MAC X0,Y0,A X:(R0)+,X0 Y:(R4)+N4,Y0 ;A=A+αiσixi(n-2)
MAC X0,Y0,A X:(R0)+,X0 Y:(R4)+,Y0 ;A=A+αiµixi(n-1)
MAC X0,Y0,A Y:(R4)-N4,Y0 ;A=A+γiyi(n-1)
MAC -X1,Y0,A X:(R0)+,X1 Y1,Y:(R4)+N4 ;A=A-βiyi(n-2). Save x(n)
MOVE A,Y1 X:(R0)+,X0 ;yi(n)=2A, assumes scale mode on

Sectn ;X1=βi+1. X0=αi+1

;Output:y(n)=Y1
MOVE X:Buflen, M4 ;Filter Order +1
NOP
MOVE Y1,Y:(R4)+N4 ;Save y(n)

Total Instruction Cycles

(6*nsec+10)Icyc @ 20 MHz = (0.6*nsec+1.0)ns

NOTE: nsec is number of sections.

Figure 5-30 DSP56001 Code for Cascaded Direct-Form Filter

MOTOROLA 5-11
5-12 MOTOROLA
 SECTION 6

Filter Design and

Analysis System

“The filter The following paragraphs discuss the design of a

example used is a cascaded filter in both the direct-form and canonic im-
sixth-order plementations, using a software package, QEDesign
Butterworth (formerly FDAS1), available from Momentum Data
lowpass filter Systems, Inc. The filter example used is a sixth-order
with a cutoff Butterworth lowpass filter with a cutoff frequency of
frequency of approximately 225 Hz and a sample frequency of
approximately 1000 Hz. Figure 6-6 is the log magnitude (gain) plot
225 Hz and a from the system output. Figure 6-7 is the phase as a
sample function of frequency in wrapped format (-π wraps to
frequency of +π). In addition, Figure 6-8 is a zero/pole plot, and Fig-
1000 Hz.” ure 6-9 is the group delay, which is the negative of the
derivative of the phase with respect to frequency.

FDAS will also generate an impulse response, step re-

sponse, and a linear magnitude plot. The results of the
design are written to a file, FDAS.OUT, which contains
much useful information. The coefficient data is written
to COEFF.FLT. The DSP56001 code generator
(MGEN) reads the COEFF.FLT file and generates a
DSP56001 assembly source file, COEFF.ASM, which

1. All references to FDAS in this application note refer to the

software package QEDesign.

MOTOROLA 6-1
 can be assembled by the DSP56001 assembler or
linker software. Examples of these files are shown
in Figure 6-10 to Figure 6-10.

6.18 Canonic Implementa-

tion
Figure 6-10 is the output file associated with the
FDAS design session, containing information on the
analog s-domain equivalent filter as well as the final
digital coefficients (listed again in Figure 6-6), which
have been properly scaled to prevent overflow at the
internal nodes and outputs of each cascaded sec-
tion. This procedure is done automatically by the
program in a matter of seconds. Executable code is
generated by MGEN (also from Momentum Data
Systems, Inc.), as shown in Figure 6-7. The code in-
ternal to each cascaded section is five instructions
long; thus, 500 ns (assuming a DSP56001 clock of
20 MHz) is added to the execution time for each ad-
ditional second-order section.

6-2 MOTOROLA
 0

-20
Log Magnitude (dB)

-40

-60

-80

-100
0.000 1.000 2.000 3.000 4.000 5.000
E+00 E+02 E+02 E+02 E+02 E+02
Frequency (Hz)

Passband Ripple in -dB -.5000 Lowpass Filter

Stopband Ripple in -dB -20.0000 IIR Design-Bilinear Transformation
Passband Cutoff Frequency 200.0000 Butterworth
Stopband Cutoff Frequency 300.0000 Filter Order: 6
Sampling Frequency 1000.0000
All Frequencies in Hertz
Quantization 24 Bits Fixed Point Fractional
Filter Design & Analysis System
Momentum Data Systems, Inc.

Figure 6-6 Log Magnitude Plot of Example Lowpass Butterworth Filter

MOTOROLA 6-3
 π

π/2
Phase (Radians)

-π/2

-π
0.000 1.000 2.000 3.000 4.000 5.000
E+00 E+02 E+02 E+02 E+02 E+02
Frequency (Hz)

Passband Ripple in -dB -.5000 Lowpass Filter

Stopband Ripple in -dB -20.0000 IIR Design-Bilinear Transformation
Passband Cutoff Frequency 200.0000 Butterworth
Stopband Cutoff Frequency 300.0000 Filter Order: 6
Sampling Frequency 1000.0000
All Frequencies in Hertz
Quantization 24 Bits Fixed Point Fractional
Filter Design & Analysis System
Momentum Data Systems, Inc.

Figure 6-7 Phase Versus Frequency Plot for Example Filter

6-4 MOTOROLA
 1.00
0

Passband Ripple in -dB -.5000 Lowpass Filter

Stopband Ripple in -dB -20.0000 IIR Design-Bilinear Transformation
Passband Cutoff Frequency 200.0000 Butterworth
Stopband Cutoff Frequency 300.0000 Filter Order: 6
Sampling Frequency 1000.0000
All Frequencies in Hertz
Quantization 24 Bits Fixed Point Fractional
Filter Design & Analysis System
Momentum Data Systems, Inc.

Figure 6-8 Zero/Pole Plot of Sixth-Order Lowpass Example Filter

MOTOROLA 6-5
 6.3
E-03

4.7
E-03
Seconds

3.1
E-03

1.5
E-03

0.0
0.000 1.000 2.000 3.000 4.000 5.000
E+00 E+02 E+02 E+02 E+02 E+02

Frequency (Hz)

Passband Ripple in -dB -.5000 Lowpass Filter

Stopband Ripple in -dB -20.0000 IIR Design-Bilinear Transformation
Passband Cutoff Frequency 200.0000 Butterworth
Stopband Cutoff Frequency 300.0000 Filter Order: 6
Sampling Frequency 1000.0000
All Frequencies in Hertz
Quantization 24 Bits Fixed Point Fractional
Filter Design & Analysis System
Momentum Data Systems, Inc.

Figure 6-9 Group Delay Versus Frequency for FDAS IIR Example

6-6 MOTOROLA
 Filter Type Low Pass
Analog Filter Type Butterworth
Passband Ripple In -dB -.5000
Stopband Ripple In -dB -20.0000
Passband Cutoff Frequency 200.000 Hertz
Stopband Cutoff Frequency 300.000 Hertz
Sampling Frequency 1000.00 Hertz
Filter Order: 6
Filter Design Method: Bilinear Transformation

Coefficients of Hd(z) are Quantized to 24 bits

Quantization Type: Fixed Point Fractional
Coefficients Scaled For Cascade Form II

NORMALIZED ANALOG TRANSFER FUNCTION T(s)

Numerator Coefficients Denominator Coefficients
***************************************** **********************************
S**2 TERM S TERM CONST TERM S**2 TERM S TERM CONST TERM
.000000E+00 .000000E+00 .100000E+01 .100000E+01.193185E+01 .100000E+01
.000000E+00 .000000E+00 .100000E+01 .100000E+01.141421E+01 .100000E+01
.000000E+00 .000000E+00 .100000E+01 .100000E+01.517638E+01 .100000E+01
INITIAL GAIN 1.00000000

UNNORMALIZED ANALOG TRANSFER FUNCTION T(s)

Numerator Coefficients Denominator Coefficients
***************************************** **********************************
S**2 TERM S TERM CONST TERM S**2 TERM S TERM CONST TERM
.000000E+00 .000000E+00 .299809E+07 .100000E+01.334500E+04 .299809E+07
.000000E+00 .000000E+00 .299809E+07 .100000E+01.244871E+04 .299809E+07
.000000E+00 .000000E+00 .299809E+07 .100000E+01.896290E+03 .299809E+07
INITIAL GAIN 1.00000000

DIGITAL TRANSFER FUNCTION Hd(z)

Numerator Coefficients Denominator Coefficients
***************************************** **********************************
Z**2 Z TERM CONST TERM Z**2 Z TERM CONST TERM
.2520353794 .5040707588 .252035379 1.00000 -.1463916302 .0225079060
.2364231348 .4728463888 .236423134 1.00000 -.1684520245 .1765935421
.3606387377 .7212774754 .360638737 1.00000 -.2279487848 .5921629667
INITIAL GAIN .876116210

Figure 6-10 FDAS.OUT File of Example Filter for Cascaded

Canonic Implementation (sheet 1 of 2)

MOTOROLA 6-7
 ZEROES OF TRANSFER FUNCTION Hd(z)
Real Part Imaginary Part Real Part Imaginary Part Radius
-.100000000E+01 .00000000E+00 -.100000000E+01 .00000000E+00 .100000000E+01
-.999290168E+00 .00000000E+00 -.100071034E+01 .00000000E+00 .999290168E+00
-.100000000E+01 .00000000E+00 -.100000000E+01 .00000000E+00 .100000000E+01

POLES OF TRANSFER FUNCTION Hd(z)

Real Part Imaginary Part Real Part Imaginary Part Radius
.731958151E-01 .130959072E+00 .731958151E-01 -.130959072E+00 .150026351E+00
.842260122E-01 .411703195E+00 .842260122E-01 -.411703195E+00 .420230344E+00
.113974392E+00 .761034036E+00 .113974392E+00 -.761034036E+00 .769521258E+00

NUMERATOR COEFFICIENTS - HIGHEST ORDER FIRST (in z)

.214893785E-01 .128936282E+00 .322340720E+00 .429787634E+00 .322340720E+00
.128936282E+00 .214893785E-01

DENOMINATOR COEFFICIENTS - HIGHEST ORDER FIRST (in z)

.100000000E+01 -.542792439E+00 .887692610E+00 .267088214E+00 .143235132E+0

-.184597152E-01 .235370025E-02

IMPULSE RESPONSE MIN = -.179722E+00, MAX = .452734E+00

Figure 6-5 FDAS.OUT File of Example Filter for Cascaded

Canonic Implementation (sheet 2 of 2)

6-8 MOTOROLA
 FILTER COEFFICIENT FILE
IIR DESIGN
FILTER TYPE LOW PASS
ANALOG FILTER TYPE BUTTERWORTH
PASSBAND RIPPLE IN -dB -.5000
STOPBAND RIPPLE IN -dB -20.0000
PASSBAND CUTOFF FREQUENCY .200000E+03 HERTZ
STOPAND CUTOFF FREOUENCY .300000E+03 HERTZ
SAMPLING FREQUENCY .100000E+04 HERTZ
FILTER DESIGN METHOD: BILINEAR TRANSFORMATION
FILTER ORDER 6 0006h
NUMBER OF SECTIONS 3 0003h
NO. OF QUANTIZED BITS 24 0018h
QUANTIZATION TYPE - FRACTIONAL FIXED POINT
COEFFICIENTS SCALED FOR CASCADE FORM 11
0 00000000 /* shift count for overall gain */
7349395 00702493 /* overall gain */
0 00000000 /* shift count for section 1 value */
2114226 002042B2 /* section1 coefficient B0 */
4228452 00408564 /* section1 coefficient B1 */
2114226 002042B2 /* section1 coefficient B2 */
1228022 0012BCF6 /* section1 coefficient A1 */
-188810 FFFD1E76 /* section1 coefficient A2 */
0 00000000 /* shift count for section 2 values */
1983261 001E431D /* section2 coefficient B0 */
3966523 003C863B /* section2 coefficient B1 */
1983261 001E431D /* section2 coefficient B2 */
1413078 00158FD6 /* section2 coefficient A1 */
-1481374 FFE96562 /* section2 coefficient A2 */
0 00000000 /* shift count for section 3 values */
3025257 002E2969 /* section3 coefficient B0 */
6050514 005C52D2 /* section3 coefficient B1 */
3025257 002E2969 /* section3 coefficient B2 */
1912173 001D2D6D /* section3 coefficient A1 */
-4967423 FFB43401 /* section3 coefficient A2 */
.2520353794097900D+00 3FD0215900000000 .25203538E+00 /* section 1 B0 */
.504D707588195801D+00 3FE0215900000000 .50407076E+00 /* section 1 B1 */
.2520353794097900D+00 3FDD215900000000 .25203538E+00 /* section 1 B2 */
.1463916301727295D+00 3FC2BCF600000000 .14639171E+00 /* section 1 A1 */
-.2250790596008301D-01 BF970C5000000000 -.22507921E-01 /* section 1 A2 */
.236423134803772OD+00 3FCE431D00000000 .23642325E+00 /* section 2 B0 */
.4728463888168335D+00 3FDE431D80000000 .47284651E+00 /* section 2 B1 */
.2364231348037720D+00 3FCE431D00000000 .23642325E+00 /* section 2 B2 */
.1684520244598389D+00 03FC58FD60000000 .168452D4E+D0 /* section 2 A1 */
.17659354Z0989990D+00 BFC69A9E000000D0 -.17659356E+00 /* section 2 A2 */
.3606387376785278D+00 3FD714B480000000 .36063876E+00 /* section 3 B0 */
.7212774753570557D+00 3FE714B480000000 .72127753E+00 /* section 3 Bl */
.3606387376785278D+00 3FD714B480000000 .36063876E+00 /* section 3 B2 */
.2279487848281860D+00 3FCD2D6DD0000000 .22794882E*00 /* section 3 A1 */
-.5921629667282104D+00 BFE2F2FFC0000000 -.59216306E+00 /* section 3 A2 */

Figure 6-6 COEFF.OUT File of Example Filter Design-Scaled for

Cascaded Canonic Implementation

MOTOROLA 6-9
 1 COEFF ident 1,0

6-10
2 include 'head2.asm'
3
4 page 132, 66, 0,10
5 pt cex, mex
6
7 ; This program implements an IIR filter in cascaded canonic sections
8 ; The coefficients of each section are scaled for cascaded canonic sections
9
10
11 00FFFF datin equ $ffff ;location in Y memory of input file
12 00FFFF datout equ $ffff ;location in Y memory of output file
13 00FFF0 m_scr equ $fff0 ;sci control register
14 00FFF1 m_ssr equ $fff1 ;sci status register
15 00FFF2 m_sccr equ $fff2 ;sci clock control register
16 00FFE1 m_pcc equ $ffe1 ;port c control register
17 00FFFF m_ipr equ $ffff ;interrupt priority register
18 000140 xx equ @cvi(20480/64*1.000)) ;timer interrupt value
19 FFD8F1 m_tim equ -9999 ;board timer interrupt value
20 000003 nsec equ 3 ;number of second order sectiona
21 include 'cascade2.asm'
22 ;
23 ; This code segment implements cascaded biquad sections in canonic form
24 ;
25
26 cascade2 macro nsec
27 m ;
28 m ;assumes each section's coefficients are divided by 2
29 m
30 m mpy y0,y1,a x:(r0)+,y0 y:(r4)+,y0 ;x0=1st section w(n-2),y0=a12/2
31 m do #nsec,_ends ;do each section
32 m mac x0,y0,a x:(r0)-,x1 y:(r4)+,y0 ;x1=w(n-1) ,y0=ai1/2
33 m macr x1,y0,a x1,x:(r0)+ y:(r4)+,y0 ;push w(n-1) to w(n-2),y0=bi2/2
34 m mac x0,y0,a a,x:(r0) y:(r4)+,y0 ;push w(n) to w(n-1),y0=bi1/2
35 m mac x1,y0,a x:(r0)+,x0 y:(r4)+,y0 ;get this iter w(n),y0=bi0/2
36 m mac x0,y0,a x:(r0)+,x0 y:(r4)+,y0 ;next iter:x0=w(n-2),y0=ai2/2
37 m _ends
38 m endm

Figure 6-7 COEFF.ASM File Generated by MGEN for Example Design andCascaded Canonic
Implementation (Sheet 1 of 4)

MOTOROLA
 39
40
41 ;
42 ;multiple shift left macro
43
44 mshl macro scount

MOTOROLA
45 m if scount
46 m rep #scount
47 m asl a
48 m endif
49 m endm
50 X:0000 org x:0
51 X:0000 states ds 2*nsec
52 Y:0000 org y:0
53 coef
54 d Y:0000 FE8F38 dc $FE8F3B ;a(*,2)/2 =-.01125395 section number 1
55 d Y:0001 095E7B dc $095E7B ;a(*,1)/2 = .07319582 section nunber 1
56 d Y:0002 102159 dc $102159 ;b(*,2)/2 = .12601769 section number 1
57 d Y:0003 2042B2 dc $2042B2 ;b(*,1)/2 = .25203538 section number 1
58 d Y:0004 D02159 dc $D02159 ;b(*,0)/2-0.5 =-.37398231 section number 1
59 d Y:0005 F4B2B1 dc $F4B2Bl ;a(*,2)/2 =-.08829677 section number 2
60 d Y:0006 0AC7EB dc $OAC7EB ;a(*,1)/2 = .08422601 section number 2
61 d Y:0007 0F218E dc $OF218E ;b(*,2)/2 = .11821157 section number 2
62 d Y:0008 1E4313 dc $1E431D ;b(*,1)/2 = .23642319 section number 2
63 d Y:0009 CF218E dc $CF218E ;b(*,0)/2-0.5 =-.38178843 section number 2
64 d Y:000A DAlA01 dc $DA1A01 ;a(*,2)/2 =-.29608148 section number 3
65 d Y:000B 0E96B6 dc $OE96B6 ;a(*,1)/2 = .11397439 section number 3
66 d Y:000c 1714B4 dc $1714B4 ;b(*,2)/2 = .18031937 section number 3
67 d Y:000D 2E2969 dc $2EZ969 ;b(*,1)/2 = .36063874 section number 3
68 d Y:000E D714B4 dc $0714B4 ;b(*,0)/2-0.5 =-.31968063 section number 3
69 Y:00c8 org y:200
70 d Y:00C8 381249 igain dc 3674697
71 000000 scount equ 0 ;final shift count
72 include 'bodyZ.asm'
73
74 000040 start equ $40 ;origin for user program
75
76 P:0000 org p:$0 ;origin for reset vector

Figure 6-7 COEFF.ASM File Generated by MGEN for Example Design and Cascaded Canonic
Implementation (Sheet 2 of 4)

6-11
 77 P:0000 0C0040 jmp start ;jump to 'start' on system reset

6-12
78
79 P:00lc org p:$lc ;origin for timer interrupt vector
80 P:00lc 0BF080 jsr filter ;jump to 'filter' on timer interrupt
000051
81
82 P:0040 org p:start ;origin for user program
83
84 P:0040 0003F8 ori #3,mr ;disable all interrupts
85 P:0041 08F4B2 movep #(xx-1),x:m_sccr ;cd=xx-1 for divide by xx
00013F
86 P:0043 08F4A1 movep #$7,x:m_pcc ;set cc(2;0) to turn on timer
000007
87 P:0045 08F4B0 movep #$2000,x:m_scr ;enable timer interrupts
002000
88 P:0047 08F4BF movep #$c000,x:m_ipr ;set interrupt priority for sci
00C000
89
90 P:0049 300000 move #states,r0 ;initialize internal state storage
91 P:004A 200013 clr a ;* set memory to zero
92 P:004B 0606A0 rep #nsec*2 ;*
93 P:004C 565800 move a,x:(rO)+ ;*
94
95 P:004D 4FF000 move y:igain,yl ;y1=initial gain/2
0000C8
96 P:004F 00FCB8 andi #$fc,mr ;allow interrupts
97 P:0050 0C0050 jmp * ;wait for interrupt
98
99 filter
100 P:0051 0008F8 ori #$08,mr ;set scaling node
101 P:0052 300000 move #states,r0 ;point to filter states
102 P:0053 340000 move #coef,r4 ;point to filter coefficients
103 P:0054 09463F movep y:datin,y0 ;get sample
104
105 cascade2 nsec ;do cascaded biquads
106 + ;
107 + ;assumes each section's coefficients are divided by 2
108 + ;

Figure 6-7 COEFF.ASM File Generated by MGEN for Example Design and Cascaded Canonic
Implementation (Sheet 3 of 4)

MOTOROLA
 109 + P:0055 F098B0 mpy y0,yl,a x:(r0)+ x0 y:(r4)+,y0 ;x0=1stsection w(n-2),y0=ai2/2
110 + P:0056 060380 do #nsec,_ends ;do each section
00005C
111 + P:0058 F490D2 mac x0,y0,a x:(r0)- ,xl y:(r4)+,y0 ;x1=w(n-1)y0=ai1/2
112 + P:0059 F418E3 macr xl,y0,a x1,x:(r0)+ y:(r4)+,y0 ;push w(n-1) to w(n 2),y0=bi2/2

MOTOROLA
113 + P:005A F800D2 mac x0,y0,a a,x:(r0) y:(r4)+,y0 ;push w(n) to w(n-1),y0=bi1/2
114 + P:005B F098E2 mac x1,y0,a x:(r0)+,x0 y:(r4)+,y0 :get this iter w(n),y0=bi0/2
115 + P:005C F098D2 mac x0,y0,a x:(r0)+,x0 y:(r4)+,y0 ;next iter:x0=w(n-2),yo=ai2/2
116 _ends
117 mshl scount
118 + if scount
121 + endif
122 P:005D 200011 rnd a ;round result
123 P:005E 09CE3F movep a,y:datout ;output sample
124 P:005F 000004 rti
125
126 end
0 Errors
0 Warnings

Figure 6-7 COEFF.ASM File Generated by MGEN for Example Design and Cascaded Canonic
Implementation (Sheet 4 of 4)

6-13
 6.2 Transpose
Implementation
(Direct Form I)
Figure 6-8 is the output file from FDAS for a trans-
pose-form implementation. As before, it contains
information on the analog s-domain equivalent filter
as well as the final digital coefficients (listed again
in Figure 6-9), which have been properly scaled to
prevent overflow at the internal nodes and outputs
of each cascaded section. Because of the stability
of the internal node gain of the transpose form and
because of the response of each second-order But-
terworth section, scaling was not done by the
program because it was not needed. Executable
code shown in Figure 6-10 is again generated by
MGEN. The code internal to each cascaded section
is seven instructions long; thus, 700 ns (assuming a
DSP56001 clock of 20 MHz) is added to the execu-
tion time for each additional second-order section.

6-14 MOTOROLA
 FILTER TYPE LOW PASS
ANALOG FILTER TYPE BUTTERWORTH
PASSBAND RIPPLE IN -dB -.5000
STOPBAND RlPPLE lN -dB -20.0000
PASSBAND CUTOFF FREQUENCY 200.000 HERTZ
STOPBAUD CUTOFF FREQUENCY 300.000 HERTZ
SAMPLING FREQUENCY 1000.00 HERTZ
FILTER ORDER: 6
FILTER DESIGN METHOD: BILINEAR TRANSFORMATION
COEFFICIENTS OF Hd(Z) ARE QUANTIZED TO 24 BITS
QUANTIZATION TYPE: FIXED POINT FRACTIONAL
COEFFICIENTS SCALED FOR CASCADE FORM I (TRANSPOSE FORM)

NORMALIZED ANALOG TRANSFER FUNCTION T(s)

NUMERATOR COEFFICIENTS DENOMINATOR COEFFICIENTS
********************************** ***********************************
S**2 TERM S TERM CONST TERM S**2 TERM S TERM CONST TERM
.000000E+00 .000000E+00 .100000E+01 .100000E+01 .193185E+01 .100000E+01
.000000E+00 .000000E+00 .100000E+01 .100000E+01 .141421E+01 .100000E+01
.000000E+00 .000000E+00 .100000E+01 .100000E+01 .517638E+00 .1000000E+01
INITIAL GAIN 1.00000000

UNNORMALIZED ANALOG TRANSFER FUNCTION T(s)

NUMERATOR COEFFICIENTS DENOMINATOR COEFFICIENTS
********************************** ***********************************
S**2 TERM S TERM CONST TERM S**2 TERM S TERM CONST TERM
.000000E+00.000000E+00 .299809E+07 .100000E+01 .334500E+04 .299809E+07
.000000E+00.000000E+00 .299809E+07 .100000E+01 .244871E+04 .299809E+07
.000000E+00.000000E+00 .299809E+07 .100000E+01 .896290E+03 .299809E+07
INITIAL GAIN 1.00000000

DIGITAL TRANSFER FUNCTION Hd(z)

NUMERATOR COEFFICIENTS DENOMINATOR COEFFICIENTS
********************************** ***********************************
Z**2 TERM Z TERM CONST TERM Z**2 TERM Z TERM CONST TERM
.2190289497.4380580187 .2190289497 1.000000 -.1463916302 .0225079060
.2520353794.5040707588 .2520353794 1.000000 -.1684520245 .1765935421
.3410534859.6821070910 .3410534859 1.000000 -.2279487848 .5921629667
INITIAL GAIN 1.00000000

ZEROES OF TRANSFER FUNCTION Hd(z)

REAL PART IMAGINARY PART REAL PART IMAGINARY PART RADIUS
-.999262530E+00 .000000000E+00 -.100073501E+01 .000000000E+00 .999262530E+00
-.100000000E+01 .000000000E+00 -.100000000E+01 .000000000E+00 .100000000E+01
-.999408962E+00 .000000000E+00 -.100059139E+01 .000000000E+00 .999408962E+00

Figure 6-8 FDAS.OUT File of Example Filter for Cascaded Transpose–

Form Implementation (sheet 1 of 2)

MOTOROLA 6-15
 POLES OF TRANSFER FUNCTION Hd(z)
Real Part Imaginary Part Real Part Imaginary Part Radius
-.731958151E-01 -.130959072E+00 -.731958151E-01 -.130959072E+00 -.150026351E+00
-.842260122E-01 -.411703195E+00 -.842260122E-01 -.411703195E+00 -.420230344E+00
-.113974392E+00 -.761034036E+00 -.113974392E-00 -.761034036E+00 -.769521258E+00

NUMERATOR COEFFICIENTS - HIGHEST ORDER FIRST (in z)

.188271907E-01 .112963161E+00.282407928E+00 .376543916E+00 .282407928E+OO
.112963161E+00 .188271907E-01

DENOMINATOR COEFFICIENTS - HIGHEST ORDER FIRST (in z)

.10000000E+01 -.542792439E+00.887692610E+00 -.267088214E+00 .143235132E+00
.184597152E-01 .235370025E-02

IMPULSE RESPONSE MIN = -.179722E+00, MAX = .452734E+00

Figure 6-8 FDAS.OUT File of Example Filter for Cascaded Transpose–

Form Implementation (sheet 2 of 2)

6-16 MOTOROLA
 FILTER COEFFICIENT FILE
IIR DESIGN
FILTER TYPE LOW PASS
ANALOG FILTER TYPE BUTTERWORTH
PASSBAND RIPPLE IN -dB -.5000
STOPBAND RIPPLE IN -dB -20.0000
PASSBAND CUTOFF FREQUENCY .200000E+03 HERTZ
STOPBAND CUTOFF FREQUENCY .300000E+03 HERTZ
SAMPLING FREQUENCY .100000E+04 HERTZ
FILTER DESIGN METHOD: BILINEAR TRANSFORMATION
FILTER ORDER 6 0006h
NUMBER OF SECTIONS 3 0003h
NO. OF QUANTIZED BITS 24 0018h
QUANTIZATION TYPE - FRACTIONAL FIXED POINT
COEFFICIENTS SCALED FOR CASCADE FORM I
1 00000001 /* shift count for overall gain */
4194304 00400000 /* overall gain */
0 00000000 /* shift count for section 1 values */
1837348 001C0924 /* section 1 coefficient B0 */
3674697 00381249 /* section 1 coefficient B1 */
1837348 001C0924 /* section 1 coefficient B2 */
1228022 00128CF6 /* section 1 coefficient A1 */
-188810 FFFDlE76 /* section 1 coefficient A2 */
0 00000000 /* shift count for section 2 values */
2114226 00204292 /* section 2 coefficient B0 */
4228452 00408564 /* section 2 coefficient B1 */
2114226 00204282 /* section 2 coefficient B2 */
1413078 00158FD6 /* section 2 coefficient A1 */
-1481374 FFE96562 /* section 2 coefficient A2 */
0 00000000 /* shift count for section 3 values */
2860964 002BA7A4 /* section 3 coefficient B0 */
5721929 00574F49 /* section 3 coefficient B1 */
2860964 002BA7A4 /* section 3 coefficient B2 */
1912173 001D2D6D /* section 3 coefficient Al */
-4967423 FFB43401 /* section 3 coefficient A2 */
.2190289497375488D+00 3FCC092400000000 .21902905E+00 /* section 1 B0 */
.4380580186843872D+00 3FDC092480000000 .43805810E+00 /* section 1 B1 */
.2190289497375488D+00 3FCC092400000000 .21902905E+00 /* section 1 B2 */
.146391630172 n95D+00 3FC2BCF600000000 .14639171E+00 /* section 1 A1 */
-.2250790596008301D-01 BF970C5000000000 -.22507921E-01 /* section 1 A2 */
.2520353794097900D+00 3FD0215900000000 .25203538E+00 /* section 2 B0 */
.5040707588195801D+00 3FE0215900000000 .50407076E+00 /* section 2 Bl */
.2520353794097900D+00 3FD0215900000000 .25ZD3538E+00 /* section 2 B2 */
.1684520244598389D+00 3FC58FD600000000 .16845204E+00 /* section 2 Al */
-.1765935420989990D+00 BFC69A9E00000000 -.17659356E+00 /* section 2 A2 */
.34105348587036130+00 3FDSD3D200000000 .34105356E+00 /* section 3 B0 */
.6821070909500122D+00 3FESD3DZ40000000 .68210712E+00 /* section 3 Bl */
.3410534858703613D+00 3FDSD3D200000000 .34105356E+00 /* section 3 B2 */
.2279487848281860D+00 3FCD2D6D00000000 .22794882E+00 /* section 3 Al */
-.5921629667282104D+00 BFE2F2FFC0000000 -.59216306E+00 /* section 3 A2 */

Figure 6-9 COEFF.FLT File for Example Filer Design–Scaled for Cascaded
Transpose Form

MOTOROLA 6-17
6-18
1 C0EFF ident 1,0
2 include 'head1.asm'
3
4 page132,66,0,10
5 optcex, mex
6 ;
7 ; This program implements an IIR filter in cascaded transpose sections
8 ; The coefficients of each section are scaled for cascaded transpose sections
9
10
11 00FFFF datin equ $ffff ;location in Y memory of input file
12 00FFFF datout equ $ffff ;location in Y memory of output file
13 00FFF0 m_scr equ $fff0 ;sci control register
14 00FFF1 m_ssr equ $fff1 ;sci status register
15 00FFF2 m_sccr equ $fff2 ;sci clock control register
16 00FFE1 m_pcc equ $ffe1 ;port c control register
17 00FFFF m_ipr equ $ffff ;interrupt priority register
18 000140 xx equ @cvi(20480/(64*1.000)) ;timer interrupt value
19 FFD8F1 m_tim equ -9999 ;board timer interrupt value
20 000003 nsec equ 3 ;number of second order sections
21 include 'cascade1.asm'
22 ;
23 ;This code segment implements cascaded biquad sections in transpose form
24 ;
25
26 cascade1 macro nsec
27 m
28 m ;assumes each section's coefficients are divided by 2
29 m
30 m do #nsec,_ends ;do each section
31 m macr y0,yl,a x:(r1),b y:(r4)+,y0 ;a=x(n)*bi0/2+wi1/2,b=wi2,y0=bi1/2
32 m asr b a,x0 ;b=wi2/2,x0=y(n)
33 m mac y0,yl,b y:(r4)+,y0 ;b=x(n)*bi1/2+wi2/2,y0=ai1/2
34 m macr x0,y0,b y:(r4)+,y0 ;b=b+y(n)*ai1/2,y0=bi2/2
35 m mpy y0,y1,b b,x:(r0)+ y:(r4)+,y0 ;b=x(n)*bi2/2,save wi1,y0=ai2
36 m macr x0,y0,b x:(r0),a a,y1 ;b=b+y(n)*ai2/2,a=next iter wi1,
37 m ;yl=output of section i

Figure 6-10 COEFF.ASM File Generated by MGEN for Example Design and Cascaded Transpose–Form
Implementation (sheet 1 of 4)

MOTOROLA
 38 m asr a b,x:(r1)+ y:(r4)+,y0 ;a=next iter wi1/2,save wi2,
39 m ;y0=next iter bi0
40 m _ends
41 m endm
42
43

MOTOROLA
44 X:0000 org x:0
45 X:0000 statel dsm nsec
46 X:0004 state2 dsm nsec
47 Y:0000 org y:0
48 coef
49 d Y:0000 0E0492 dc $0E0492 ;b(*,0)/2 = .10951447section number 1
50 d Y:0001 lC0924 dc $1C0924 ;b(*,1)/2 = .21902901 section number 1
51 d Y:0002 095E7B dc $095E7B ;a(*,1)/2 = .07319582 section number 1
52 d Y:0003 0E0492 dc $0E0492 ;b(*,2)/2 = .10951447 section number 1
53 d Y:0004 FE8F3B dc $FE8F3B ;a(*,2)/2 =-.01125395 section number 1
54 d Y:0005 102159 dc $102159 ;b(*,0)/2 = .12601769 section number 2
55 d Y:0006 2042B2 dc $2042B2 ;b(*,1)/2 = .25203538 section number 2
56 d Y:0007 OAC7EB dc $0AC7EB ;a(*,1)/2 = .08422601 section number 2
57 d Y:0008 102159 dc $102159 ;b(*,2)/2 = .12601769 section number 2
58 d Y:0009 F4B2B1 dc $F4B2B1 ;a(*,2)/2 =-.08829677 section number 2
59 d Y:000A lSD3D2 dc $l5D3D2 ;b(*,0)/2 = .17052674 section number 3
60 d Y:000B 2BA7A4 dc $2BA7A4 ;b(*,1)/2 = .34105355 section number 3
61 d Y:000c 0E96B6 dc $0E9686 ;a(*,1)/2 = .11397439 section number 3
62 d Y:0000 lSD3D2 dc $l5D3D2 ;b(*,2)/2 = .17052674 section number 3
63 d Y:000E DAlA01 dc $DAlA01 ;a(*,2)/2 =-.29608148 section number 3
64 include 'bodyl.asm'
65
66 000040 start equ $40 ;origin for user program
67
68 P:0000 org p:$0 ;origin for reset vector
69 P:0000 0C0040 jmp start ;jump to 'start' on system reset
70
71 P:001C org p:$lc ;origin for timer interrupt vector
72 P:00lC 0BF080 jsr filter ;jump to 'filter' on timer interrupt
000057
73
74 P:0040 org p:start ;origin for user program

Figure 6-10 COEFF.ASM File Generated by MGEN for Example Design and Cascaded Transpose–Form
Implementation (sheet 2 of 4)

6-19
6-20
75
76 P:0040 0003F8 ori #3,mr ;disable all interrupts
77 P:0041 08F4B2 movep #(xx-1),x:m_sccr ;cd=xx 1 for divide by xx
00013F
78 P:0043 08F4A1 movep #$7,x:m_pcc ;set cc(2;0) to turn on timer
000007
79 P:0045 08F4B0 movep #$2000,x:m_scr ;enable timer interrupts
002000
80 P:0047 08F4BF movep #$c000,x:m_ipr ;set interrupt priority for sci
00C000
81
82
83 P:0049 300000 move #state1,r0 ;point to filter state1
84 P:004A 310400 move #state2,rl ;point to filter state2
85 P:004B 340000 move #coef,r4 ;point to filter coefficients
86 P:004C 0502A0 move #nsec l,m0 ;addressing modulo nsec
87 P:004D 050EA4 move #5*nsec-l,m4 ;addressing modulo 5*nsec
88 P:004E 0502Al move #nsec-1,m1 ;addressing modulo nsec
89 P:004F 200013 clr a ;initialize internal state storage
90 P:0050 0603A0 rep #nsec ;* zero statel
91 P:0051 565800 move a,x:(r0)+ ;*
92 P:0052 0603A0 rep #nsec ;* zero state2
93 P:0053 565900 move a,x:(r1)+ ;*
94
95 P:0054 F88000 move x:(r0),a y:(r4)+,y0 ;a=wl (initially zero) ,y0=bl0/2
96
97 P:0055 00FCB8 andi #$fc,mr ;allow interrupts
98 P:0056 0C0056 jmp * ;wait for interrupt
99
100 filter
101 P:0057 0008F8 ori #$08,mr ;set scaling mode
102 P:0058 09473F movep y:datin,yl ;get sample
103
104 cascade1 nsec ;do cascaded biquads
105 + ;
106 + ; assumes each section's coefficients are divided by 2
107 + ;

Figure 6-10 COEFF.ASM File Generated by MGEN for Example Design and Cascaded Transpose–Form
Implementation (sheet 3 of 4)

MOTOROLA
 108 + P:0059 060380 do #nsec,_ends ;do each section
000061
109 + P:00SB FC81B3 macr y0,y1,a x:(r1),b y:(r4)+,y0 ;a=x(n)*bi0/2+wi1/2,b=wi2,y0=bi1/2
110 + P:OOSC 21C42A asr b a,x0 ;b=wi2/2,x0=y(n)
111 + P:005D 4EDCBA mac y0,yl,b y:(r4)+,y0 ;b=x(n)*bi1/2+wi2/2,y0=ail/2

MOTOROLA
112 + P:00SE 4EDCDB macr x0,y0,b y:(r4)+,y0 ;b=b+y(n)*ai1/2,y0=bi2/2
113 + P:00SF FC18B8 mpy y0,y1,b b,x:(r0)+ y:(r4)+,y0 ;b=x(n)*bi2/2,save wi1,y0=ai2
114 + P:0060 19A0DB macr x0,y0,b x:(r0),a a,y1 ;b=b+y(n)*ai2/2,a=next iter wi1,
115 + ;y1=output of section i
116 + P:0061 FC1922 asr a b,x:(r1)+ y:(r4)+,y0 ;a=next iter wi1/2,save wi2,
117 + ;y0=next iter bi0
118 + _ends
119
120 P:0D62 09C73F movep y1,y:datout ;output sample
121 _endp
122
123 P:0063 000004 rti
124 end
0 Errors
0 Warnings

6-21
6-22 MOTOROLA
 SECTION 7

FIR FILTERS
“The filter can If phase distortion is of secondary importance to
be efficiently magnitude response, the desired filter can generally
implemented on be implemented with less memory, less computation-
the DSP56001 by al complexity, and at the lowest cost using IIR
using modulo structures. On the other hand, in applications requir-
addressing to ing linear phase in the passband and a specific
implement the magnitude response, the specified filter is generally
shifting and best implemented using FIR structures. Examples of
parallel data applications requiring linear phase are as follows:
moves to load
4. Communication systems such as modems or
the multiplier- Integrated Services Digital Networks (ISDN) in
accumulator.” which the data pulse shape and relative timing in
the channel must be preserved

5. Ideal differentiators which provide a 90-degree

phase shift at all frequencies in addition to a
constant group delay

6. Hilbert transformers used to demodulate complex

signals such as those used in high-speed modems

7. Hi-fidelity audio systems in which phase distortion

of recorded music must be minimized to reproduce
the original sound with as much fidelity as possible

8. System synthesis in which the system impulse

response is known as a priori. FIR filters are also
important because they are all-zero filters (i.e., no
feedback) and are therefore guaranteed to be
stable.

MOTOROLA 7-1
 7.19 Linear-Phase FIR
Filter Structure
The basic structure of a FIR filter is simply a tapped
delay line in which the output from each tap is
summed to generate the filter output. This is shown
in Figure 7-31. This structure can be represented
mathematically as:

N–1
y(n) = ∑ h ( i )x ( n – i ) Eqn. 7-16
i=0

where: x(n) is the most recent (t = nT) input

signal sample
x(n-i) is signal samples delayed by i
sample periods (iT)
h(i) is the tap weights (or filter
coefficients)
y(n) is the filter output at time t = nT
From this structure, it is easy to see why the filter is
termed finite—the impulse response of the filter will
be identically zero after N sample periods because
an impulse input (i.e., x(n) = 1, x(n-i) = 0 for i ≠ 0) will
have traversed the entire delay line at time t = NT.
That is, the impulse response of FIR filters will only
last for a finite period of time. This response is in
contrast to IIR filters which will “ring” in response to
an impulse for an infinite period of time. The values
of the coefficients represent the impulse response
of the FIR filter as can be seen by evaluating Eqn.
7-16 over N sample periods for a single-unit input
pulse at time t = 0.

7-2 MOTOROLA
 Digital Delay Line

x(n) z-1 z-1 z-1

(Filter Input)

h0 x h1 x hN-1 x

Σ
y(n) (Filter Output)

Figure 7-31 FIR Structure

There are no feedback terms in the structure (i.e.,

Eqn. 7-16 has no denominator) but rather only N ze-
ros. By taking the z-transform of Eqn. 7-16, the
transfer function of the filter becomes:

N–1
Y(z) –1
H ( z ) = ------------ =
X(z) ∑ h ( i )z Eqn. 7-17
i=0

Eqn. 7-17 is a polynomial in z of order N. The roots

of this polynomial are the N zeros of the filter.

The same procedures used to calculate the magni-

tude and phase response of IIR filters can be
applied to FIR filters. Accordingly, the gain, G(θ),
can be obtained by substituting z = ejθ into Eqn. 7-
17 (where θ = 2π/fs is the normalized digital frequen-
cy), then taking the absolute value so that:

MOTOROLA 7-3
 N–1
– jiθ
G(θ) = ∑ h ( i )e
i=0 Eqn. 7-18

N–1
= ∑ [ h ( i ) cos iθ – jh ( i ) sin iθ ]
i=0

The phase response, φ(θ), is found by taking the in-

verse tangent of the ratio of the imaginary to real
components of G(θ) so that:

N–1
– ∑ h ( i ) sin iθ
–1 i=0
φ ( θ )=tan --------------------------------------
- Eqn. 7-19
N–1
∑ h ( i ) cos iθ
i=0

where: h(i) is implicitly assumed to be real

Intuitively, it can be reasoned that, for any pulse to
retain its general shape and relative timing (i.e.,
pulse width and time delay to its peak value) after
passing through a filter, the delay of each frequency
component making up the pulse must be the same
so that each component recombines in phase to re-
construct the original shape. In terms of phase, this
implies that the phase delay must be linearly related
to frequency or:
φ ( θ ) = – τθ Eqn. 7-20
This relationship is illustrated in Figure 7-2 in which
a pulse having two components, sin ωt and sin 2ωt,
passes through a linear-phase filter having a delay
of two cycles per Hz and a constant magnitude re-
sponse equal to unity. That is, the fundamental is

7-4 MOTOROLA
delayed by two cycles (4π) and the first harmonic is
delayed by four cycles (8π). The group delay of a
system, τg, is defined by taking the derivative of φ(θ):

dφ
τ g ≡ – ------ Eqn. 7-21
dθ

Filter
x1(t) = sin ωt
t
∆φ1 = 4π-8π = -4π

x2(t) = sin 2ωt

t

∆φ2 = 8π-16π = -8π

Composite input Stretched pulse within filter Composite output

pulse shape pulse shape

Legend:
Represents the filtered wave
Represents the unfiltered wave (i.e., no delay)

Figure 7-32 Signal Data through a FIR Filter

Since θ is the normalized frequency, τg in Eqn. 7-21

is a dimensionless quantity and can be related to
the group delay in seconds by dividing by the sam-
ple frequency, fs. For a linear-phase system, τg is
independent of θ and is equal to τ. This fact can be
seen by substituting Eqn. 7-20 into Eqn. 7-21.
In this example, τg is two cycles per Hz. Note that
the pulse shape within the filter has been retained

MOTOROLA 7-5
 but is twice the width of the original pulse; that is the
change in the pulse width is equal to the group de-
lay. The negative sign indicates the phase is
retarded or delayed (i.e., a causal system).

The impact of the requirement of linear phase on

the design of a FIR filter can be seen by substituting
Eqn. 7-19 into Eqn. 7-20 so that:

N–1
∑ h ( i ) sin iθ
i=0
tan θτ g = ------------------------------------ Eqn. 7-22
N–1
∑ h ( i ) cos iθ
i=0

which can be written as:

N–1
∑ h ( i ) ( cos iθ sin θτ g – sin iθ cos θ τ g ) = 0
i= 0

or

N–1
∑ h ( i ) sin [ θ ( τ g – i ) ] = 0 Eqn. 7-23
i=0

The solution to is the constraint on τg for a FIR filter

to be linear phase. The solution to can be found by
expanding the left-hand side (LHS) as follows:

7-6 MOTOROLA
 LHS = h 0 sin θτ g + h 1 sin θ ( τ g – 1 ) + h 2 sin θ ( τ g – 2 ) + …

+h N – 3 sin θ ( τ g – N + 3 ) + h N – 2 sin θ ( τ g – N + 2 ) + h N – 1 sin θ ( τ g – N + 1 )

so that if:
(N – 1)
τ g = ------------------ Eqn. 7-24
2

then for every positive argument there will be a cor-

responding negative argument. For example, the
argument for the h1 term becomes:

θ  ------------- – 1 = θ  ------------- 
N–1 N–3
 2   2 

which is the negative of the argument for the hN-2

term, i.e.,

θ  ------------- – N + 2 = – θ  ------------- 
N–1 N–3
 2   2 

so that if:

h ( i ) = h ( N – 1 – i ) for 0≤i≤N–1 Eqn. 7-25

then Eqn. 7-9 and Eqn. 7-25 represent the solution

to . By substituting Eqn. 7-9 into Eqn. 7-20, the
phase of a linear-phase FIR filter is given by:

φ ( θ ) = –  -------------  θ
N–1
Eqn. 7-26
 2 

MOTOROLA 7-7
 Therefore, a nonrecursive filter (FIR), unlike a recur-
sive filter (IIR), can have a constant time delay for all
frequencies over the entire range (from 0 to fs/2). It is
only necessary that the coefficients (and therefore
impulse response) be symmetrical about the mid-
point between samples (N-2)/2 and N/2 for even N or
about sample (N-1)/2 for odd N (see Eqn. 7-25).
When this symmetry exists, τg for the filter will be
N
-------------  T seconds.
–1
 2 

7.20 Linear-Phase FIR Filter

Design Using the
Frequency Sampling
Method
Implementing a FIR filter in DSP hardware such as
the DSP56001 is a relatively simple task. Determin-
ing a set of coefficients that describe a given
impulse response of a filter is also a straightforward
procedure. However, deriving the optimal coeffi-
cients necessary to obtain a particular response in
the frequency domain is not always as easy. The
following paragraphs introduce a simple method to
determine a set of coefficients based on a desired
arbitrary frequency response (often referred to as
the frequency sampling method). This method has
the distinct advantage of being done in real time.
However, the most efficient determination of coeffi-
cients is best done by utilizing a software filter

7-8 MOTOROLA
design system such as FDAS to perform numerical
curve-fitting and optimization, a procedure that ne-
cessitates using a computer. For example, the
inverse Fourier transform integral is used to deter-
mine the FIR coefficients from a response
specification in the frequency domain, which gener-
ally requires a numerical integration procedure. The
inverse discrete Fourier transform (IDFT), which
can be implemented using the inverse fast Fourier
transform (IFFT) algorithm, is discussed in the fol-
lowing paragraphs. One reason for choosing this
approach is to demonstrate a method that can be
used to determine coefficients in real time since the
fast Fourier transform (FFT) can be implemented in
the same DSP56001 hardware as the FIR filter.

The question is “How must the filter be specified in the

frequency domain so that Eqn. 7-25 and Eqn. 7-26 are
realized?”. That is, starting with the definition of the fil-
ter in the frequency domain, a method for calculating
the filter coefficients is required. Beginning with Eqn.
7-17, the z-transform of the FIR filter, and evaluating
this transfer function on the unit circle at N equally
spaced normalized frequencies (i.e., z = ej2πlk/N)
produces:
N–1
j2πk ⁄ N – j2πki ⁄ N
H(k) = H(e ) = ∑ h ( i )e
i=0
Eqn. 7-27
where: 0 ≤ k ≤ N-1

h(i) can be solved in terms of the frequency response

at the discrete frequencies, H(k), by multiplying both

MOTOROLA 7-9
 sides of by ej2πkm/N and summing over k as follows:

N–1 N – 1N – 1

∑ H ( k )e ∑ ∑ h ( i )e
j2πkm ⁄ N – j2πki ⁄ N j2πkm ⁄ N
= e
k=0 k=0i=0

N –1 N –1
– j2πki ⁄ N e j2πkm ⁄ N
= ∑ h(i) ∑ e

i = 0 k = 0

N –1
= ∑ h ( i ) Nδ
im
i = 0

= Nh ( m )

Eqn. 7-28
where: δim is the Kronecker delta, which is equal
to one when i = m but is zero otherwise

can be used to find h(i) by simply setting i = m.

N–1
1 j2πki ⁄ N
h ( i ) = ----
N ∑ H ( k )e Eqn. 7-29
k=0

Eqn. 7-29 is the IDFT of the filter response. In gen-

eral, the discrete Fourier coefficients are complex;
therefore, H(k) can be represented as:

jφ ( k )
H ( k ) = A ( k )e

where: A ( k ) ≡ H ( k )
so that Eqn. 7-29 becomes:

7-10 MOTOROLA
 N–1
1 jφ ( k ) j2πki ⁄ N
∑
Eqn. 7-30
h ( i ) = ---- A ( k )e e
N
k=0
At this point, the linear-phase constraints, Eqn. 7-25
and Eqn. 7-26, can be applied. The constraint that
the coefficients be real and symmetrical when h(i) is
complex (see Reference 1) implies that:
h ( i ) = h∗ ( N – 1 – i )

where: * signifies the complex conjugate

therefore:
N–1
1 – j φ ( k ) – j 2πk ( N – 1 – i ) ⁄ N
h∗ ( N – 1 – i ) = ----
N ∑ A ( k )e e
k=0

N–1
1 – j [ φ ( k ) + 2πk ( N – 1 ) ⁄ N ] j2πki ⁄ N
= ----
N ∑ A ( k )e e
k=0

Eqn. 7-31

will be identical to Eqn. 7-30 if:

2πk ( N – 1 )
φ ( k ) = – φ ( k ) – ----------------------------- + 2πr for r=0,1,2,...
N
or
(N – 1)
φ ( k ) = πr – πk ------------------ Eqn. 7-32
N

What are the constraints on r and A(k) which will

guarantee a purely real response for N even?
Substituting Eqn. 7-32 into Eqn. 7-30 yields:

MOTOROLA 7-11
 N–1
1 j [ πr – πk ( N – 1 ) ⁄ N + 2πki ⁄ N ]
h ( i ) = ----
N ∑ A ( k )e
Eqn. 7-33
k=0

Expanding Eqn. 7-33 for even N yields:

1 jπr j [ πr – π ( N – 1 ) ⁄ N + 2πi ⁄ N ]
( i ) = ---- { A ( 0 )e + A ( 1 )e +…
N

j [ πr – π ( N – 1 ) ⁄ 2 + πi ]
+A ( N ⁄ 2 )e +…

j [ πr – π ( N – 1 ) ( N – 1 ) ⁄ N + 2π ( N – 1 ) i ⁄ N ] 
+A ( N – 1 )e 


Consider the A(k) and A(N-k) terms; if r = 0 for the

A(k) term and r = 1 for the A(N-k) term, then the ar-
gument for the A(N-k) term is the negative of the
argument for the A(k) term, given that N is even.
That is:

j [ π – π ( N – k ) ( N – 1 ) ⁄ N + 2π ( N – k ) i ⁄ N ]
e
j [ πk ( N – 1 ) ⁄ N –2 πki ⁄ N ] j [ π – π ( N – 1 ) + 2πi ]
= e e
– j [ – π k ( N – 1 ) ⁄ N + 2πki ⁄ N ] – j [ π ( N – 2 ) + 2πi ]
= e e
– j [ – π k ( N – 1 ) ⁄ N – 2 π ki ⁄ N ]
= e

Therefore, if A(k) = A(N-k), then all imaginary com-

ponents in Eqn. 7-33 will cancel and h(i) will be

7-12 MOTOROLA
purely real, which is the desired result. In general,
for N even, the formulas for the frequency sampling
method can be reduced to:
N–1
1 j [ πr – πk ( N – 1 ) ⁄ N + 2πki ⁄ N ] 
h ( i ) = ---- { A ( 0 )+
N ∑ A ( k )e 

k=1

N⁄2–1
1 j [ πr – πk ( N – 1 ) ⁄ N + 2πki ⁄ N ] +
= ---- { A ( 0 )+
N ∑ Ak [ e
k=1 j [ π – π ( N – k ) ( N – 1 ) ⁄ N + 2π ( N – k ) i ⁄ N ] 
e ]


N⁄2–1
1 j –π kN – 1 ⁄ N + 2πki ⁄ N
= ---- { A ( 0 ) +
N ∑ Ak [ e +
k=1 – j [ – πk ( N – 1 ) ⁄ N + 2πki ⁄ N ] 
e ]


N⁄2–1
1 – j πk j [ πk ⁄ N + 2πki ⁄ N ]
= ---- { A ( 0 ) +
N ∑ A(k)[e e +
k=1 jπk – j [ πk ⁄ N + 2πki ⁄ N ] 
e e ]


 N⁄2–1 
1 k πk 
= ----  A ( 0 ) + 2 ∑ A ( k ) ( – 1 ) cos ------ ( 1 + 2i ) 
N N 
 k=1 

Eqn. 7-34
where: r= 0 for 0 ≤ k < N/2

MOTOROLA 7-13
 r= 1 for N/2 < k ≤ N-1
given
A(k) = A(N – k) N
for 1 ≤ k≤ ---- -1
2
with
A(N ⁄ 2) = 0
and
N–1 N
φ ( k ) = – πk ------------- for 0 ≤ k ≤ ---- -1
N 2

N–1 N
φ ( k ) = π – π k ------------- for ---- + ≤ k≤ N-1
N 2
In summary, if a filter with an even number of sym-
metrical real coefficients is desired, then the phase
must be linear, and the frequency response must
be symmetrical about N/2 and zero at N/2 (see Ref-
erence 1).

7-14 MOTOROLA
 1.0

N/2
0.5
H(k)

0
0 8 16 24 31

k(N=32)
(a) Arbitrary Input Specification for Frequency Response Magnitude

0
Phase φ

-14.5π

-28π
0 16 31

(b) Necessary Phase Specifications for Frequency Response Magnitude

Figure 7-33 Arbitrary Filter Example

As an example, consider the arbitrary filter speci-

fied in Figure 7-34. In this example, N = 32 and an
arbitrary lowpass and bandpass combination is
specified. Figure 7-34 shows the result of trans-
forming the polar coordinate filter specification into
rectangular coordinates, yielding real and imagi-
nary components of the transfer function. This
transformation is accomplished by treating the
magnitude of H(k) as the length of a vector in polar
coordinates and the phase as the angle. The x
component (real part) is the product of the length
(magnitude of H(k)) and the cosine of the phase.

MOTOROLA 7-15
 Likewise, the y component (imaginary part) is the
product of the length and the sine of the phase angle.
This transformation to rectangular coordinates is
necessary to perform the IDFT (or IFFT) calculation:

N–1
1 jkθn
h ( n ) = ----
N ∑ H ( k )e Eqn. 7-35
k=0

where: H(k) is a complex number and θn = 2πn/N

When the filter's transfer function, H(k), is described

as in Figure 7-34 and Figure 7-34 (i.e., the real part
symmetric about N/2 and the imaginary part asym-
metric about N/2), h(n) is strictly real. If h(n) were
complex, the FIR filter would be much more difficult
to implement since twice as many terms would be
present. Figure 7-35 shows the results of the IDFT
applied to the example arbitrary filter specification.

Note the symmetry of the coefficients in Figure 7-35

(symmetric about (N-1)/2). This symmetry is to be
expected for an even number of coefficients. Eqn.
7-36 has N-1 roots since it is a polynomial of order
N-1. These roots are plotted in Figure 7-36 and
must be found from a numerical algorithm such as
the Newton-Raphson root-finding recursion relation
or the Mueller method (see Reference 18). h(n) can
now be used to specify the filter response in the
continuous frequency domain by setting θ = 2πf/fs.

7-16 MOTOROLA
 0.5
Re[H(k)]

-0.5
0 16 31

k
(a) Real Part

0.5
Im[H(k)]

-0.5
0 16 31

k
(b) Imaginary Part
Figure 7-34 Response Transformed from Polar Coordinates

MOTOROLA 7-17
 0.2

0.1
h(n)

-0.1
0 16 31

Figure 7-35 FIR Coefficients from Eqn. 7-17 for Filter Example

Imaginary

Real

Figure 7-36 Roots (Zeros) of Eqn. 7-36 for Filter Example

7-18 MOTOROLA
The transfer function has exactly the same form as
the DFT of the h(n), but now the frequency is con-
tinuous up to fs/2:

N–1
– jnθ
H(θ) = ∑ h ( n )e Eqn. 7-36
n=0

where: θ = 2πf/fs

0
9 10 11
Filter Gain, G(θ), (dB)

0 1 2
-10
N/2
-20

-30

-40
3 4 5 6 7 8 12 13 14 15 16
0 π/4 π/2 3π/4 π

θ
(a) Gain (magnitude of Hz)) Plot

π/2
8 10
Phase φ(θ)

2 3 4 5 6 7
0
0
11 12 13 14 15
-π/2
9
1
-π

0 π/4 π/2 3π/4 π

θ
(b) Phase Response Corresponding to Gain Plot

Figure 7-37 A 32-Point FIR Filter Example

MOTOROLA 7-19
 The continuous frequency gain and phase re-
sponse of the 32-coefficient example filter, plotted
in Figure 7-37, are generated as follows:
1⁄2
G ( θ ) = H ( θ ) = [ H ( θ )H∗ ( θ ) ]

 N–1 2 N–1 2 1 ⁄ 2
 
=  ∑ h ( n ) cos nθ + ∑ h ( n ) sin nθ 
 n=0 n=0 
 

Eqn. 7-37
and
N–1 
 
 ∑ h ( n ) sin nθ 
–1  n = 0 
φ ( θ ) = ( – tan )  ----------------------------------------
- Eqn. 7-38
N – 1 
 ∑ h ( n ) cos nθ 
 
n = 0 

where: h(n) is the value obtained from

or equivalently from Eqn. 7-19
Note that the gain plot (see Figure 7-37) exactly in-
tersects the discrete frequency points originally
specified in Figure 7-34. Clearly, G(θ) and φ(θ) have
many discontinuities. Due to the symmetry of linear-
phase FIR filters, analytic expressions can be found
for H(ω) as shown in Eqn. 7-42 and Eqn. 7-43.

For comparison, Figure 7-38 shows the log magni-

tude response of the example filter with larger
values of N. The stopband attenuation is not im-
proved as much as expected when the number of
coefficients is increased; however, the sharpness of
the transition band is significantly enhanced be-
cause the approach discussed has implicitly

7-20 MOTOROLA
assumed a rectangular window function. A smooth-
ing window, w(n), can be used to improve this
situation (see Reference 10).

19 20 21 22 23 24
0 4 5
0 1 2 3
Filter Gain, G(θ), (dB)

-10

-20

-30

-40
6 7 8 9 10 1112 1314 15 16 17 18 25 26 27 282930 31 32
0 π/4 π/2 3π/4 π

θ
(a) N=64

0
Filter Gain, G(θ), (dB)

-10

-20
N=128
-30

-40

0 π/4 π/2 3π/4 π

θ
(b) N=128

Figure 7-38 Log Magnitude Response of Filter Example with Larger N Values

If h(n) is modified by a window function, w(n), hw(n)

= h(n)w(n) for 0 ≤ n ≤ N-1, the gain of Eqn. 7-37
(shown in Figure 7-34(a)) can be greatly improved.
The window function, w(n), goes to zero at both
ends (n = 0 and n = N) and is unity at the center; w(n)

MOTOROLA 7-21
 is symmetrical so as not to disrupt the linear-phase
characteristics of the filter. Figure 7-34(b) uses a
window function described by w(n) = sin2[πn/(N-1)]
(also known as a Hanning window) to demonstrate
the sensitivity of the gain to windowing. w(n) is basi-
cally an envelope function used to taper the ends of
h(n) smoothly. The rounding of the transition-band
edges in Figure 7-34(b) is the trade-off for window-
ing; however, in most applications, this trade-off is
well worthwhile.

0
-10
Filter Gain, G(θ), (dB)

-20 N=128
-30
-40
-50
-60
-70
-80
0 π/4 π/2 3π/4 π

θ
(a) Rectangular Window
Filter Gain, G(θ), (dB)

0
-10
-20
-30 SINE2 WINDOW
-40
-50
-60
-70
-80
0 π/4 π/2 3π/4 π

θ
(b) Sine2 (Hanning) Window
Figure 7-39 Window Function Effects on Filter Example

7-22 MOTOROLA
Window functions have a powerful effect on the stop-
band attenuation as well as the passband transition
slope.The best passband transition slope perfor-
mance is achieved with the rectangular window, but
this window results in very poor stopband perfor-
mance and often severe passband fluctuation (or
ripple). All window functions have the effect of in-
creasing the stopband attenuation and reducing the
passband ripple at the expense of increasing the
width of the transition region. However, for most
window functions, the passband ripple is relatively
insensitive to N. Windowing is described in virtually
any DSP textbook (see REFERENCES). Also, win-
dowing is discussed with practical examples in
Implementation of Fast Fourier Transforms on Mo-
torola's DSP56000/DSP56001 and DSP96002
Digital Signal Processors (see Reference 19).

7.21 FIR Filter Design

Using FDAS
Figure 7-40 shows an example (log magnitude and
impulse response) of a bandpass filter generated with
the FDAS software package using the Kaiser win-
dow. A totally different approach to FIR filter design,
the equiripple method, is based on finding an opti-
mum approximation to the ideal or desired response,
D(θ). An optimum approximation can be found be-
cause of the inherent symmetry of the coefficients of
linear-phase filters.

MOTOROLA 7-23
7-24
Figure 7-40 FDAS Output for FIR Bandpass Filter Example with a Kaiser Window

MOTOROLA
Recall that the continuous frequency response of
a FIR filter can be found by setting z = ejθ in the
z-transform so that:

N–1
– jθn
H(θ) = ∑ h ( n )e Eqn. 7-39
n=0

If the linear-phase factor is factored out, Eqn. 7-39

can be written as:

N–1
– jθ ( N – 1 ) ⁄ 2 jθ [ ( N – 1 ) ⁄ 2 – n ]
H(θ) = e ∑ h ( n )e
n=0

– jθ ( N – 1 ) ⁄ 2 jθ ( N – 1 ) ⁄ 2 jθ [ ( N – 1 ) ⁄ 2 ) – 1 ]
= e {h ( 0 )e + h ( 1 )e +…

– jθ ( N – 1 ) ⁄ 2 –j θ [(N – 1) ⁄ 2 ) – 1]
h ( N – 1 ) )e + h ( N – 2 )e }

Eqn. 7-40
Using Euler's identity, Eqn. 7-40 can be grouped
into sine and cosine terms as follows:

– jθ ( N – 1 ) ⁄ 2 (N – 1)
H(θ) = e { [ h ( 0 ) + h ( N – 1 ) ] cos θ ------------------
2

N–1 (N – 1)
+jh0 – hN – 1 sin θ ------------- + h ( 1 ) + h ( N – 2 ) ] cos θ ------------------ – 1
2 2

N–1
+j [ h ( 1 ) – h ( N – 2 ) ] sin θ ------------- – 1
2

Eqn. 7-41

MOTOROLA 7-25
 If it is assumed that linear phase is achieved by
even symmetry (i.e., h(n) = h(N-1-n) and that N is
even, reduces to:

N⁄2–1
– jθ ( N – 1 ) ⁄ 2
∑ 2 h ( n ) cos θ  ------------- – n
N–2
H(θ) = e
 2 
n=0

which, with a change of variable, k = N/2-n-1, can

be written as follows:

N⁄2–1
– jθ ( N – 1 ) ⁄ 2
2h ( n ) cos θ  ------------- – n
N–2
H(θ) = e ∑  2 
n=0
Eqn. 7-42

which is in the following form:

jPθ
H ( θ ) = A ( θ )e Eqn. 7-43

where: A(θ) is a real value amplitude function

P(θ) is a linear-phase function

The linear functions, A(θ) and P(θ), are to be con-

trasted with the inherently nonlinear absolute value
and arctangent functions in and Eqn. 7-38. For all
four types of linear-phase filters (N even or odd and
symmetry even or odd), A(θ) can be expressed as a
sum of cosines (see Reference 21).

Since an ideal filter cannot be realized, an approxima-

tion must be used. If the desired frequency response,
D(θ), can be specified in terms of a deviation, δ, from

7-26 MOTOROLA
 the ideal response, then the error function in Eqn.
7-44 can be minimized:

max
E ( θ ) = ----------- D ( θ ) – A ( θ ) Eqn. 7-44
θ

by finding the best A(θ). Because A(θ) can be ex-

pressed as a finite sum of cosines as shown in Eqn.
7-42, it can be shown (see Reference 18) that the
optimal A(θ) will be unique and will have at least N/2
+ 1 extremal frequencies, where extremal frequen-
cies are points such that for:
θ1 < θ2 < … < θN ⁄ 2 < θN ⁄ 2 + 1

E ( θe ) = –E ( θe + 1 ) N
fore = 1, 2, …, ---- + 1
2

and
max
E ( θ e ) = ----------- [ E ( θ ) ]
θ

Thus, the best approximation will exhibit an equirip-

ple error function. The problem reduces to finding
the extremal frequencies since, once they are
found, the coefficients, 2h(N/2-k-1), can be found
by solving the set of linear equations:

N⁄2–1
∑ 2h  ---- – k – 1 cos θ e k + --- 
N 1
D ( θ ) ±δ = A ( θe ) =
2   2
k=0
Eqn. 7-45

The Remez exchange algorithm is used to system-

atically find the extremal frequencies (see

MOTOROLA 7-27
 Reference 5). Basically, a guess is made for the ini-
tial N/2 + 1 extremal frequencies. (Usually, this
guess consists of N/2 + 1 equally spaced frequen-
cies in the Nyquist range.) Using this guess, is
solved for the coefficients and δ. Using these coef-
ficients, A(θ) is calculated for all frequencies and
the extrema, and frequencies at which the extrema
are attained are determined. If the extrema are all
equal and equal to or less than that specified in the
initial filter specification, the problem is solved.
However, if this is not the case, the frequencies at
which the extrema were attained are used as the
next guess. Note that the final extremal frequencies
do not have to be equally spaced. Clearly, the equir-
ipple design approach is calculation intensive.

What is the benefit of equiripple designs over win-

dow designs? In general, equiripple designs require
fewer taps for straightforward requirements. When
the specification requires a sharp cutoff and/or a
large stopband attenuation or a narrow bandpass,
the equiripple approach may fail to converge. In
general, when N is decreased, an equiripple design
tends to maintain its transition band while sacrific-
ing stopband attenuation; window designs tend to
do the opposite. Of the window alternatives, the
Kaiser window is preferred for designing filters be-
cause the passband ripple and stopband
attenuation can be varied relatively independent of
the transition width (see Reference 1). For spectral
analysis, the Blackman-Harris window is preferred
(see Reference 10).

7-28 MOTOROLA
MOTOROLA
Figure 7-41 FDAS Output for FIR Bandpass Filter Example with Equiripple Design

7-29
 Figure 7-41 shows an example of an equiripple de-
sign generated from the FDAS software package.
This example is the same as that used for the Kaiser
window example of Figure 7-40. The number of co-
efficients in the equiripple design is far less than that
generated by the Kaiser window method (179 ver-
sus 291). However, the passband ripple is larger.

7.22 FIR Implementation on

the DSP56001
The DSP56001 has several architectural features
that make it ideally suited for implementing FIR filters:
1. Dual Harvard architecture uses two data
memories with dedicated buses and address
generation units, allowing two addresses to be
generated in a single cycle. If one address is
pointing to data and another address is pointing
to coefficients, a word of data and a coefficient
can be fetched in a single cycle.

2. Modulo addressing makes the shifting of data

unnecessary. If an address pointer is
incremented (or decremented) with the modulo
modifier in effect, data shifting can be
accomplished by just “backing up” the address
register by one to overwrite the data that would
normally be shifted out. This procedure allows
very efficient addressing of operands without
wasting time shifting the data or reinitializing
pointers.

7-30 MOTOROLA
 3. Hardware DO loops execute without overhead
once the loop is started. After a three-cycle
initialization of the DO loop, the body of the loop
executes as if it were straight-line code. Since
the DO loop does not require any overhead
cycles for each pass, the need for straight-line
code is eliminated.

For a four-coefficient example of a linear-phase FIR

filter, the input-output difference equation can be
found by expanding Eqn. 7-16:

y(n) = h0x(n) + h1x(n-1) + h2x(n-2) + h3x(n-3)

Eqn. 7-46
This difference equation can be realized with the
discrete-time four-tap filter example shown in Figure
7-41. The filter can be efficiently implemented on the
DSP56001 by using modulo addressing to imple-
ment the shifting and parallel data moves to load the
multiplier-accumulator. The filter network is shown
in Figure 7-41(a); the memory map for the filter in-
puts and coefficients is shown in Figure 7-41(b). The
following DSP56001 code is used to implement the
direct form FIR filter:

CLR A X0,X:(R0)+ Y:(R4)+,Y0 ;Save input sample, fetch coef.

REP #NTAPS-1 ;Repeat next instruction.
MAC X0,Y0,A X:(R0)+,X 0 Y:(R4)+,Y0 ;FIR Filters.
MACR X0,Y0,A (R0)- ;Round result and adjust R0.

Register R0 points to the input variable buffer, M0

is set to three (modulo 4), R4 points to the coeffi-
cient buffer, and M4 is set to three. The input
sample is in X0.

MOTOROLA 7-31
 The CLR instruction clears accumulator A and per-
forms parallel data moves. The data move saves
the most recent input value to the filter (assumed to
be in XO) into the location occupied by the oldest
data in the shift register and moves the first coeffi-
cient in the filter (h0) into the data ALU.

The REP instruction repeats the next instruction

NTAPS-1 times. Since there are four taps in this fil-
ter, the next instruction is repeated three times.

x(n) z-1 z-1 z-1 + y(n)

b3

b0 b1 b2

+ +

(a) Direct-Form FIR Filter Network with Four Coefficients

X:(R0) Y:(R4)
M0 =ntaps-1
x(n) =3 b(0) M4 =ntaps-1
=3
R0 x(n-1) R0 b(1)
Input Data
x(n-2) b(2) Filter Coefficients

x(n-3) b(3)

x(n) b(4)

(b) Data Structures for Four-Tap FIR Example

Figure 7-42 FIR Filter Example

7-32 MOTOROLA
 The MAC instruction multiplies the data in X0 by the
coefficient in Y0 and adds the result to accumulator
A in a single cycle. The data move in this instruction
loads the next input data variable into X0 and the
next coefficient into Y0. Both address registers R0
and R4 are incremented. The MACR instruction cal-
culates the final tap of the filter, rounds the result
using convergent rounding, and address register
R0 is decremented.

Address register R4 is incremented once before the

REP instruction and three times due to the REP in-
struction for a total of four increments. Since the
modulus for R4 is four, the value of R4 wraps
around pointing back to the first coefficient.

The operation of RO is similar. The value input to

the filter is saved, and then RO is incremented
pointing to the first state. The REP instruction incre-
ments RO three times. Since the modulo on RO is
four and RO is incremented four times, the value in
RO wraps around pointing to the new input sample.
When the MACR instruction is executed, the value
of RO is decremented, pointing to the old value, x(n-
3). The next sample time overwrites the value of
x(n-3) with the new input sample, x(n). Thus, the
shifting of the input data is accomplished by simply
adjusting the address pointer, and the modular ad-
dressing wraps the pointer around at the ends of the
input data buffer. Instruction cycle counts for this fil-
ter are NTAPS+3. Thus, for a four-tap filter, seven
instructions are required. ■

7-33 MOTOROLA
7-34 MOTOROLA
 REFERENCES

4. Antoniou, Andreas. Digital Filters: Analysis and Design

New York, NY: McGraw-Hill, 1974.

5. Berlin, Howard M. Design of Active Filters with

Experiments. Indianapolis, IN: Howard W. Sams & Co.,
Inc., 1977.

6. Bogner and Constantinides, eds. Introduction to Digital

Filtering. New York, NY: John Wiley & Sons, 1975.

7. Brophy, J. J. Basic Electronics for Scientists. New York,

NY: McGraw-Hill, 1 966.

8. Cheney, E. W. Introduction to Approximation Theory.

New York, NY: McGraw-Hill, 1966.

9. Chrysafis, A. Fractional and Integer Arithmetic Using

the DSP56000 Family of General-Purpose Digital
Signal Processors (APR3/D). Motorola, Inc., 1988.

10. Digital Stereo 10-Band Graphic Equalizer Using the

DSP56001 (APR2/D). Motorola, Inc., 1988.

11. “Digital Filters on DSP56000/DSP56001.” Motorola

Technical Bulletin, 1988.

12. “Filter Design & Analysis System.” Version 1.3,

Momentum Data Systems, 1 985.

13. Harris, Fredric J. “On the Use of Windows for Harmonic

Analysis with the Discrete Fourier Transform.” Proc. of
IEEE, vol. 66, no. 1, 1978, pp.51-83.

14. Jackson, Leland B. Digital Filters and Signal

Processing. Boston, MA: Kluwer Academic Publishers,
1986.

15. Lancaster, D. Active Filter Cookbook. Indianapolis, IN:

Howard W. Sams & Co., Inc., 1975.

MOTOROLA Reference-1
 16. McClellan, J. H. and T. W. Parks. “A Unified Approach
to the Design of Optimum FIR Linear-Phase Digital
Filters.” IEEE Trans. Circuits Systems CT-20, 1973,
pp. 697-701.

17. Moschytzm, G. S. and P. Horn. Active Filter Design

Handbook. New York, NY: John Wiley & Sons, 1981.

18. Oppenheim, A. V. and R. W. Schafer. Digital Signal

Processing. Englewood Cliffs, NJ: Prentice-Hall,
1975.

19. Parks, T. W. and J. H. McClellan. “Chebyshev

Approximation for Nonrecursive Digital Filters with
Linear Phase.” IEEE Trans. Circuit Theory CT- 19,
1972, pp. 189-194.

20. Proakis, John G. et al. Introduction to Digital Signal

Processing. New York, NY: MacMillan, 1988.

21. Rabiner, L. R. and B. Gold. Theory and Application of

Digital Signal Processing. Englewood Cliffs, NJ:
Prentice-Hall, 1975.

22. Sohie, Guy R. L. Implementation of Fast Fourier

Transform on Motorola's DSP56000/DSP56001 and
DSP96002 Digital Signal Processors (APR4/D).
Motorola, Inc., 1989.

23. Strawn, J., et al. Digital Audio Signal Processing—An

Anthology. William Kaufmann, 1985.

24. Williams, Arthur B. Electronic Filter Design

Handbook. New York, NY: McGraw-Hill, 1981. ■

Reference-2 MOTOROLA