International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248

Volume: 4 Issue: 3 338 – 342

_______________________________________________________________________________________________
Design and FPGA Implementation of Variable FIR Filters using the Spectral
Parameter Approximation and Time-Domain Approach

T. Pitchaiah P V Sridevi
Dept. of ECE Dept. of ECE
VFSTR AUCE, AU
Guntur, India Visakhapatnam, India
[email protected] [email protected]

Abstract—This brief present a design and FPGA implementation of variable FIR filters using time domain approach of the spectral parameter
approximation (SPA) technique. Farrow structure is used to implement the SPA-based filter. In the design of variable filters first design the
practical filters which satisfy the given transition bandwidth, passband ripple, and stopband attenuation specifications and then approximate the
coefficients of these filters by the impulse response of the Farrow structure. Least-squares technique is used to approximation problem. Various
design and implementation cases with FPGA synthesis results are presented.
Keywords- Farrow structure, FPGA, Reconfigurable FIR filter, Spectral parameter approximation technique, Variable cutoff frequency filter.
__________________________________________________*****_________________________________________________

I. INTRODUCTION α. The cutoff frequency (or equivalently the passband

frequency, as the transition bandwidth is fixed) of the SPA-
Variable digital filters can change their frequency response
based filter varies in the desired tuning range as α varies from
characteristics based on the application requirement. Such
0 to 1.
variable finite impulse response (FIR) filters are required in
many signal processing applications such as radar, sonar, and The Farrow structure implementation of the SPA-based
control systems, adaptive systems, vibration analysis, audio filter consists of L + 1 sub-filters, each of order N, as shown in
and biomedical signal processing, and wireless Fig. 1. For the FIR filter design case, the frequency response
communications. In a special case of the variable FIR filters, of the SPA-based filter, H (ω, α) is given by,
the cutoff frequency of the filter needs to be varied, and the 𝐻 𝜔, 𝛼 = 𝐿+1 𝑁
𝑘=1 𝑛=0 𝛼
𝑘−1
ℎ𝑠𝑘 (𝑛)𝑒 −𝑗𝑤𝑛 (1)
passband ripple, the stopband attenuation, and the transition where hsk(n) is the impulse response of the kth sub-filter.
bandwidth should be below predefined specified values.
For the variable coefficient filter, cutoff frequency is I/
controlled by changing all the coefficients of the filter. P
However, when the cutoff frequency needs to be changed P Sub- Sub- Sub- Sub-
frequently, this approach becomes unsuitable for higher order filter filter filter filter
filters due to the tedious updating routine (large number of L+1 L 2 1
memory access operations) and huge memory requirement (to
store all the filter coefficients corresponding to all the desired
responses). In practice, it is desirable to control the cutoff
frequency using fewer number of parameters to keep the o/p
updating routine simple. A number of such reconfigurable
filter design techniques have been developed to control the α
cutoff frequency by modifying the impulse response of the
fixed-coefficient prototype filter. Details and reviews of such Fig.1 Farrow Structure Implementation of SPA-based filter
reconfigurable filter design techniques can be found in [3,14]
and the references therein. The objective here is to find the optimal coefficients hsk (n)
In the spectral parameter approximation (SPA) technique such that the frequency response of the SPA-based filter will
[2,4–12], the variable filter is implemented using Farrow approximate the frequency response of the ideal filter for each
structure. The SPA-based filter has smaller group delay, fixed cutoff frequency (in the desired tuning range), i.e., for each
transition bandwidth, requires very small number of variable value of α (in 0–1). The ideal filter is defined as the filter with
multipliers (to be used for the variable weights), and its cutoff magnitude equal to 1 in its passband and 0 in its stopband. The
frequency can be controlled through only one variable approximation problem can be solved using the least-squares
parameter. The frequency response of the SPA-based filter is (LS) [4,7,8,12] or minimax techniques [6,9–11], by
modeled as a polynomial function of the cutoff frequency incorporating the desired peak to peak passband ripple (δp) and
parameter (α). The frequency responses of the fixed sub-filters stopband attenuation (δs) constraints in the problem
in the Farrow structure are equivalent to the coefficients of the formulation. The optimal solutions can be obtained from the
closed-form formulae [6–8,12], or by solving the system of
polynomial, and the frequency response (i.e., the cutoff
linear equations obtained by the discretization method
frequency) of the SPA-based filter is controlled by the variable
[2,4,5,9–11].
338
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 338 – 342
_______________________________________________________________________________________________
In this paper, we used a new approach for the design of proposed method, the first sub-filter in the Farrow structure is
SPA-based filter in the time domain, and model the designed to have cutoff frequency fcl and transition bandwidth
approximation problem such that the optimal coefficients for tbw, i.e.,
the sub-filters are found so as to approximate the impulse {hs0,1 . . . hs N ,1 }={ h0,1 . . . h N ,1 } (6)
responses of the practical filters, i.e., the impulse responses of Therefore, in the proposed method, the design problem further
the filters which satisfy the desired specifications on δp, δs, and reduces to finding the optimal coefficients for only the
transition bandwidth. In Section II, we present the time- remaining L sub-filters. Therefore, it is sufficient to model the
differences between the M − 1 practical filters and the first
domain approach in detail. Various design cases are presented
practical filter (i.e., the filter with cutoff frequency fcl) using
in Section III, along with the comparisons and some
the remaining L sub-filters of the Farrow structure. Therefore,
observations. The concluding remarks are presented in the
the matrices Gp, W , and Gs are modified as follows.
Section IV.
Matrix Gp:
Gp1(i,:)=Gp(1,:)-Gp(i,:) for i>1 (7)
II. TIME DOMAIN APPROACH FOR THE DESIGN OF Gp2(i,:)=Gp1(i+1 , :) for i=1 to M-1 (8)
SPA-BASED FILTER Matrix W:
W1(i,:)=W(i+1, :) for i=1 to M-1 (9)
Let the cutoff frequency tuning range of the SPA-based filter
be fcl to fcm, i.e., the cutoff frequency varies linearly as fcl ≤ fc ≤ W2(: , K)=W1(: , K+1) for k=1 to L (10)
fcm for 0 ≤ α ≤ 1. Let the corresponding passband and stopband Matrix Gs:
ranges be fpl to fpm and fsl to fsm, respectively. In this case, we Gs1(i , :) = Gs(i+1 , :) for i=1 to L (11)
will formulate the SPA-based filter design problem so as to The optimal coefficients hsi, j (for the remaining L sub-filters)
approximate the impulse response of the practical filter by the are then obtained by solving the system of linear equations,
impulse response of the SPA-based filter for the desired tuning given by Gp2 = W2 Gs1. Further, in the case of linear-phase
range. We use the discretization approach to formulate the filters, only half the matrices can be used due to the coefficient
approximation problem as follows. symmetry.
First, design M practical filters, each of order N , with cutoff The magnitude response of the SPA-based filter designed
frequencies fcl, fcl + Δf, fcl + 2Δ f, . . . , fcm, and the transition following above procedure is checked for the M values used
bandwidth tbw, where Δ f = ( fcm − fcl)/(M − 1). These filters for the filter design, as well as for arbitrary values of α, to
are designed using the standard filter design algorithm such as
the Remez exchange algorithm or the Parks–McClellan ensure that the desired specifications on δp and δs are satisfied.
algorithm [13]. If these specifications are not satisfied, the above mentioned
Let the matrix of the impulse responses of the practical filters design procedure can be repeated by varying N and/or L such
be denoted as that the magnitude response of the SPA-based filter satisfies
ℎ0,1 ℎ1,1 ℎ2,1 . ℎ𝑁,1 the desired specifications on δp and δs.
ℎ0,2 ℎ1,2 ℎ2,2 . ℎ𝑁,2
𝐺𝑝 = . . .. . . (2)
. . . . III. DESIGN CASES
ℎ0,𝑀 ℎ1,𝑀 ℎ2.𝑀 . ℎ𝑁,𝑀
We present five design cases of the SPA –based filter design
where hi,j denotes the ith sample of the impulse response of for different tuning ranges and different transition bandwidths.
the jth filter, i.e. ith coefficient of the jth filter. For every case, the SPA-based filters are designed using the
Discretize the cutoff frequency parameter, α, in M time-domain approach utilizing Least Squares technique. For
equidistant points in the range 0-1. The matrix of the every SPA-based filter, the weights are the powers of α and all
impulse responses of the SPA-based filter is given by, the subfilters of its Farrow structure are of equal order N. All
𝐺 = 𝑊𝐺𝑠 (3) the filter designs are based on the discretization method.
where Design Case.1 Let the desired cutoff frequency range be
1 𝛼1 𝛼12 . 𝛼1𝐿 fc1=0.3 to fcm=0.5 with passband attenuation p=0.1dB,
𝑊 = 1. 𝛼.2 𝛼2 . . 𝛼.2
2 𝐿
(4) stopband attenuation s= -45dB and transition bandwith of 0.2.
. (All the frequency values mentioned in this paper are
1 𝛼𝑀 𝛼𝑀 . 𝛼𝑀 2 𝐿
normalized with respect to π.)

where αji denotes the ith power of αj, with α1=0, and
ℎ𝑠0,1 ℎ𝑠1,1 ℎ𝑠2,1 . ℎ𝑠𝑁,1
ℎ𝑠0,2 ℎ𝑠1,2 ℎ𝑠2,2 . ℎ𝑠𝑁,2
𝐺𝑠 = .. .. .. . . (5)
. .
ℎ𝑠0,𝐿+1 ℎ𝑠1,𝐿+1 ℎ𝑠2.𝐿+1 . ℎ𝑠𝑁,𝐿+1

is the matrix of the coefficients of the sub-filters of the Farrow

structure and hsi,j denotes the ith coefficient of the jth
subfilter. The system of linear equations Gp=WGs , can be
solved using the LS technique to find the optimal coefficients
hsi,j.
Fig.2 Magnitude Response of the Practical filters for various
As, for α = 0 the impulse response of the SPA-based filter is
equal to the impulse response of the first sub-filter, in the α values with M=21
339
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 338 – 342
_______________________________________________________________________________________________

Fig.3 Magnitude Response of the SPA-based filter for

various values of α with M=21
Fig.7 Magnitude Response of the subfilter1 and total
response of all other subfilters with M=21

Design Case.3
Let the desired cutoff frequency range be fp1=0.3 to fcm=0.5
with passband attenuation p=0.1dB, stopband attenuation s=
-45dB and transition bandwith of 0.05.

Fig.4 Magnitude Response of the subfilter1 and total

response of all other subfilters with M=21
Design Case.2
Let the desired cutoff frequency range be fc1=0.3 to fcm=0.5
with passband attenuation p=0.1dB, stopband attenuation s=
-45dB and transition bandwith of 0.1. Fig.8 Magnitude Response of the Practical filters for various
α values with M=21

Fig.5 Magnitude Response of the Practical filters for various

α values with M=21
Fig.9 Magnitude Response of the SPA-based filter for
various values of α with M=21

Fig.6 Magnitude Response of the SPA-based filter for

various values of α with M=21
Fig.10 Magnitude Response of the subfilter1 and total
response of all other subfilters with M=21

340
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 338 – 342
_______________________________________________________________________________________________
Design Case.4
Let the desired passband frequency range be fc1=0.05 to
fcm=0.5 with passband attenuation p=0.1dB, stopband
attenuation s= -45dB and transition bandwith of 0.2.

Fig.14 Magnitude Response of the Practical filters for various

α values with M=46

Fig. 11 Magnitude Response of the Practical filters for various

α values with M=46

Fig.15 Magnitude Response of the SPA-based filter for

various values of α with M=46

Fig.12 Magnitude Response of the SPA-based filter for

various values of α with M=46

Fig.13 Magnitude Response of the subfilter1 and total

Fig.16 Magnitude Response of the subfilter1 and total
response of all other subfilters with M=46
response of all other subfilters with M=46
Design Case.5
Let the desired passband frequency range be fc1=0.05 to
fcm=0.5 with passband attenuation p=0.1dB, stopband IV. FPGA IMPLEMENTATION
attenuation s= -45dB and transition bandwith of 0.1.
The design of proposed SPA-based filters were developed
using Matlab, Simulink and HDL coder. The HDL code
generated was tested on FPGA. The device used is Virtex-6,
6vlx240tff1156-1. Synthesis results for all the five cases of
designed filters were given in the Table I and Table II. For the
comparison purpose, we consider the total number of
multipliers required for the SPA-based filter as the measure of
complexity. For a wider or moderately wide transition
bandwidths and small tuning range (Design cases 1,2, and 4)
requires less number of multipliers compared to narrower
341
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 338 – 342
_______________________________________________________________________________________________
transition bandwidth and wider tuning range filters (Design [2] H.H. Dam, A. Cantoni, S. Nordholm, K.L. Teo, Variable
cases 3 and 4). digital filter with group delay flatness speci-fication or phase
Compared to other technique, SPA-based filter design using constraints. IEEE Trans. Circuits Syst. II Express Br. 55(5),
frequency-domain [15] have less complex in Design cases 1, 2 442–446 (2008)
and 4, but in design cases 3 and 5 SPA-based filter design [3] S.J. Darak, A.P. Vinod, E.M.K. Lai, J. Palicot, H. Zhang,
using time domain have less complex. Linear-phase VDF design with unabridged bandwidth control
over the Nyquist band. IEEE Trans. Circuits Syst. II Express
Table I: FPGA Implementation of SPA filters- Comparison of Br. 61(6), 428–432 (2014)
Timing parameters [4] T.B. Deng, Weighted least-squares method for designing
Parameter Design Design Design Design Design arbitrarily variable 1-D FIR digital filters. Signal Process.
Case1 Case2 Case3 Case4 Case5
Timing Summary 80(4), 597–613 (2000)
Minimum 0.825 0.815 0.94 0.857 0.844 [5] T.B. Deng, Design of arbitrary-phase variable digital filters
Period(ns) using SVD-based vector-array decompo-sition. IEEE Trans.
Maximum 1212.1 1175.08 1063.40 1166.86 1184.83
Circuits Syst. I Regul. Pap. 52(1), 148–167 (2005)
Frequency 2 8 1 1 4
(MHz) [6] B. Dumitrescu, B.C. Sicleru, R. Stefan, Positive hybrid real-
Minimum 1.086 1.108 1.144 1.078 1.108 trigonometric polynomials and applications to adjustable
input arrival filter design and absolute stability analysis. Circuits Syst.
time before
clock (ns) Signal Process. 29(5), 881–899 (2010)
Maximum 66.95 106.135 209.376 93.456 184.757 [7] S.S. Kidambi, An efficient closed-form approach to the
output design of linear-phase FIR digital filters with variable-
required time bandwidth characteristics. Signal Process. 86(7), 1656–1669
after clock
(ns) (2006)
Maximum 47.622 68.416 136.118 78.421 147.038 [8] A. Kumar, S. Suman, G.K. Singh, A new closed form method
combinationa for design of variable bandwidth linear phase FIR filter using
l path delay
(ns)
different polynomials. Int. J. Electron. Commun. 68(4), 351–
360 (2014)
[9] P. Lowenborg, H. Johansson, Minimax design of adjustable-
Table II: FPGA Implementation of SPA filters- bandwidth linear-phase FIR filters. IEEE Trans. Circuits Syst.
Comparison of Device Utilization I Regul. Pap. 53(2), 431–439 (2006)
Parameter Design Design Design Design Design [10] C. Luo, J.H. McClellan, Adjustable bandwidth filter design
Case1 Case2 Case3 Case4 Case5 with generalized farrow structure. IEEE International
Conference on Acoustics, Speech and Signal Processing
Slice Logic Utilization: (ICASSP) (2011)
Number of Slice [11] C. Luo, L. Zhu, J.H. McClellan, A general structure for the
Registers 475 883 1707 441 883 design of adjustable FIR filters. IEEE International
(301440)
Conference on Acoustics, Speech and Signal Processing
Number of Slice
252 855 7682 301 589 (ICASSP) (2013)
LUTs (150720)
Number used as [12] C.K.S. Pun, S.C. Chan, K.S. Yeung, K.L. Ho, On the design
252 855 7682 301 589
Logic (150720) and implementation of FIR and IIR digital filters with
Specific Feature Utilization: variable frequency characteristics. IEEE Trans. Circuits Syst.
Number of II Analog Digit. Signal Process. 49(11), 689–703 (2002)
121 299 778 189 333
DSP48E1s (768) [13] L. Rabiner, J.H. McClellan, T.W. Parks, FIR digital filter
design techniques using weighted Chebyshev approximation.
V. CONCLUSIONS
Proc. IEEE. 63(4), 595–610 (1975)
This paper presented a FPGA implementation of variable [14] G. Stoyanov, M. Kawamata, Variable digital filters. J. Signal
filters by using spectral parameter approximation and time- Process. 1(4), 275–289 (1997)
domain approach. In the implemented method, the optimal [15] Dhabu, Sumedh, and A. Prasad Vinod. "A New Time-
coefficients for the sub-filters in the Farrow structure are Domain Approach for the Design of Variable FIR Filters
found so as to approximate the impulse responses of the Using the Spectral Parameter Approximation
practical filters by the impulse response of the Farrow Technique." Circuits, Systems, and Signal Processing 36.5
structure. Least Squares approximation problem is used in (2017): 2154-2165.
the optimization. The results show that the proposed time
domain method is effective compared to conventional
frequency domain approach when the desired
specifications are stringent.
REFERENCES
[1] M.G. Bellanger, On computational complexity in digital
transmultiplexer filters. IEEE Trans. Commun. 30(7), 1461–
1465 (1982)
342
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________