Beruflich Dokumente
Kultur Dokumente
2
Computational Complexity (MOPS)
6000 where
5000 Time domain FIR
Frequency domain FIR 1−α 1+α
4000 m= and M =
3000
2T 2T
2000
1000 4 Methodology
0
0 20 40 60 80 100 120 140
FIR filter length In order to evaluate the performance of our proposed FIR
implementation method in the frequency domain, especially
Figure 1. Computational complexities of dif- in terms of immunity against intersymbol interference and
ferent FIR implementations noise rejection, two FIR filter implementations are consid-
ered:
paths before entering the filter into complex samples that • a time domain filter of which the coefficients were ob-
will be treated by the direct and inverse FFT then breaking tained by sampling the RRC impulse response with an
the output again into two paths. In fact, from the linearity oversampling factor of 4 and which processes the I and
property of the Fourier transform, we can state that: Q paths successively;
F IRf (XI + jXQ ) = F −1 [F (XI + jXQ ) H] • the frequency domain filter with coefficients extracted
= F −1 [F (XI ) H + jF (XQ ) H] as discussed in section 3.2 and which combines the I
= F IRf (XI ) + jF IRf (XQ ) (4) an Q paths into one complex signal and processes the
two paths simultaneously.
where F IRf denotes the FIR filtering in frequency domain
and XI and XQ are the inputs on the I and Q paths respec- Note that in designing the filter, the sampling rate also
tively. needs to be considered. The higher the sampling rate the
Comparing the computational complexities of the 2 dif- easier it is after a digital to analogue converter to remove
ferent implementations of the FIR filter for different filter high frequency components generated in the sampling pro-
lengths leads to Fig. 1. From this figure, we can find that cess. The higher the sampling rate, the faster the converter
it is more advantageous to implement FIR filters in the fre- must operate, which in general will lead to greater cost and
quency domain for lengths 44 and more which is generally greater power consumption. To keep our approach realistic,
the case of the UMTS standard. (Note that the computa- we chose an oversampling factor of 4 (thus a sampling rate
tional complexity of the frequency domain FIR could be fs = 4 × chip rate = 15.36M Hz) as a compromise be-
further reduced if we take into account trivial multiplica- tween high performance on a hand and cost and complexity
tions in the FFT, which makes this type of filtering feasible on the other hand.
for even shorter filters) We then created two real random input signals of length
5000 points with values between −1 and 1 to simulate in-
3.2 Filter coefficients puts on the I and Q paths, and we included 3 null samples
between each adjacent values to simulate an oversampled
Instead of performing the FFT of the filter’s coefficients signal by a factor of 4. These signals where passed through
extracted by sampling the RRC time representation (1) like a classical time domain filter to get the transmitted signal.
shown in section 2, we take the ideal square root of the The two filter implementations (the time domain and our
raised cosine filter transfer function shown in (5) and sam- proposed frequency domain) are then applied to the trans-
ple it to get the coefficients that should be used in the fre- mitted signal. Note that results shown in the next section
quency domain FIR filter. This operation leads to the mod- where obtained by repeating the same operations on 100
ule of the coefficients which we have to multiply by a linear different random input signals.
πf
phase ϕH (f ) = exp(− 2T ). The resultant coefficients are
then used in the frequency domain implementation of the
FIR filter. 5 Performance evaluation
T
T
πT 0 ≤ |f | ≤ m
H(f ) = 1 + cos (|f | − m) m ≤ |f | ≤ M Following in this section, is the comparison between the
2 α
0 |f | > M two filter implementation methods in terms of intersymbol
(5) interference and additive white Gaussian noise rejection.
3
90 40
Gain in immunity against ISI(%)
80 Time domain FIR
30
Frequency domain FIR
EVM(%)
70
20
60
Minimum and maximum gains
Average gain 10
50
40 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
FIR filter length FIR filter length
Figure 2. ISI immunity gain of frequency do- Figure 3. EVM of the two different filter imple-
main FIR implementation mentations
5.1 Intersymbol interference rejection mum requirement for the EVM: the EVM shall not exceed
17.5% in the user equipment radio trasmission and recep-
tion in the FDD mode [4]. This requirement is in fact one of
The main property of Nyquist filters is that they result
the key guidelines for choosing the appropriate length of the
in zero intersymbol interference (ISI) at the optimum sam-
FIR filter to use. The formula used to calculate the EVM is
pling point for filtered data [8]. This is why we present
presented in (6) where Perror is the root mean square power
here a comparison between a classical time domain FIR im-
of the error vector and Preference is the root mean square
plementation, and our frequency domain implementation.
power of the ideal transmitted signal.
The ISI is measured as the variance of the error between the
samples of the input signal and those of the received sig-
Perror
nals at the optimum sampling points. Fig. 2 shows the gain EV M (%) = × 100 (6)
in percentage of the frequency domain implementation over Preference
the time domain implementation for different filter lengths.
From this figure, we can see that no pertinent conclusion Simulations with different filter lengths and for the two
could be made about FIR filters of lengths 16 or less be- different FIR implementations discussed in this paper show
cause the gain in ISI rejection clearly varies with the input that the frequency domain implementation using the sam-
signal, the main reason is that such short filters do not re- pled ideal filter transfer function is always better than the
construct the input signal in a reliable way. For the other time domain implementation. From Fig. 3, we can see that
FIR filters (the kind of filters suitable for UMTS), the gain the frequency domain implementation ensures that the min-
is about 72%. This gain is augmented to 99% if we use imum requirement of EVM is met even with a FIR filter of
a frequency domain implemented filter in the transmission length 8. Of course, the EVM is not the only criterion to
chain as well. select the proper filter length, other decisive factors should
To have a better comparison of the immunity against ISI be considered like the filter immunity against noise and in-
of the two different filters, we compare two other parame- terference.
ters: the Error Vector Magnitude (EVM) and the peak dis-
tortion. 5.1.2 Peak distortion comparison
Another parameter commonly used to measure the ISI of
5.1.1 Error vector magnitude comparison a transmit-receive filter combination is the peak distortion
In literature discussing UMTS radio receivers, a parameter defined as follows [7]:
called EVM is commonly used to indicate the amount of
∞
ISI [7]. 3GPP technical specifications for the UMTS stan- 2 k=1 h N2 + kM
Dp(dB) = 20 log10 N (7)
dard provide the following definition of EVM: “The Error h
2
Vector Magnitude is a measure of the difference between
the reference waveform and the measured waveform. This In (7), h is the impulse response of the transmit-receive
difference is called the error vector. ... The EVM result is filter combination, N is the length of this response and M
defined as the square root of the ratio of the mean error vec- is the oversample factor. Furthermore it has been assumed
tor power to the mean reference power expressed as a per- that h is symmetric. In order to extract h, we passed a pulse
centage”. These technical specifications define also a mini- signal into two sets of transmit-receive filters. At first a
4
0 References
Peak distortion (dB)