VLSI signal processing

Final report

FIR Filters with

ICS
R89921071
梁金權
l Introduction
Sigma-delta A/D converter is frequently used in many signal processing system.

This is largely because the single bit format simplifiers the analog circuit, and

decoding is by means of linear filtering. And conventionally, in the digital signal

processing system, the input is processed in pulse code modulated (PCM) format.

Therefore, the overall signal processing system is like fig. 1 describe. However, the

PCM digital filter suffers from several disadvantages: 1. The accuracy of a digital

filter depends on its word length, which is dictated by the resolution of the A/D

converter. 2. Conventional PCM digital filters require multibit multipliers and adders.

Therefore several sigma-delta modulated encoding FIR filters has been proposed [1]

[2] to solve the PCM problem and to replace the system in fig1 by the simplified

system in fig2.

Fig. 1

(i)oversampled bitsttram

(ii)nyquist multibit digital

(iii)nyquist multibit digital

(iv)oversampled multibit digital

(v)oversampled bitstream

(SDM: sigma-delta modulation)

1
Fig. 2

(i)oversampled bitsttram

(ii)oversampled multibit digital

(iii)oversampled bitstream

(SDSP: sigma-delta signal processing)

l Sigma-delta modulator
Here we discuss how sigma-delta converters use the technique of

noise shaping in addition to oversampling to allow high resolution conversion of

relatively low bandwidth.

The conventional PCM conversion system is show in Fig.3

Fig. 3

Assume the quantization noise e[n] is white uncorrelated noise,

, SNR can be obtained as follows.

2
  σ2   σ2 
SNR = 10 log  x2  = 10 log  x2  + 4.77 + 6.02 N ( dB)
 σe  V 

where N is the numbers of bits of the quantizatizer. For every increment of N, there is

about 6dB improvement in SNR.

The oversampled version of PCM conversion system is show in Fig. 4. And the

quantization noise power spectral density for nyquist rate PCM and oversampled

PCM conversion is show in Fig. 5.

Fig. 4 oversampled PCM conversion

Fig. 5 white noise power spectrum

From Fig. 5, we can see that after oversampling, the noise spectrum is suppressed

3
to be flat with a ratio of oversampling ratio. And after filtering out the out-band noise,

higher SNR is achieved. The SNR can be calculated as follows:

where fs/2fB=2r is the oversampling ratio. For every doubling the oversampling

ratio, i.e., for every increment in r, the SNR improves by about 3 dB, or the resolution

improves by one-half bit.

Then, we discuss the noise shaping concept. In a quantization system,if we can

predict and correct for next quantization error using previous quantization error, we

can reduced the quantization error, and it is the error feedback concept, the system is

show in Fig. 6

Fig 6

O(z)=STF(z)*I(z)+NTF(z)*Eq(z)

NTF(z)=1-H(z)

STF(z)=1

(NTF: noise transfer function)

And for H(z) =z-1, the Amplitude response of NTF=1-H(z) is show as Fig. 7

4
Fig. 7

We can see that the quantization errors at baseband frequencies is suppressed.

Sigma-delta modulator used the above two concepts (oversampled and noise

shaping) to achieve high SNR. A first order sigma-delta modulator and a overall A/D

system is show as Fig. 8.

Fig 8 first order sigma-delta modulator and a overall A/D system

The modulator output Y(z) is given by

and the corresponded time domain version of the modulator output is

5
For comparision the first order sigma-delta NTF and oversampled PCM NTF is

show in Fig. 9

Fig. 9

The in-band noise power of first order sigma-delta modulator is

 π2 
( ) ( )
SNR = 10 log σx2 − 10 log σe2 − 10 log   + 9.03r
 3 

For every increment in r, the SNR improves by 9 dB, or equivalently, the resolution

improves by 1.5 bits.

In addition, consider the second-order sigma-delta modulation

(double-loop) as Fig .10

6
Fig. 10

 π4 
( ) ( )
SNR = 10 log σ2x − 10 log σe2 − 10 log   + 15.05r
 5 

the integrator in the second loop can be view as another error feedback with a

more precisely error prediction,. Therefore, high order sigma-delta modulator can

achieve higher SNR under same oversampling ratio.

 π4 
( ) ( )
SNR = 10 log σ2x − 10 log σe2 − 10 log   + 15.05r
 5 

For every increment in oversampling r, the SNR improves by 15 dB, or

equivalently, the resolution improves by 2.5 bits.

l Sigma-delta modulation encoding FIR filters

An FIR filter is described by the convolution . This filter can be

7
Fig 11. (fN =1/TN is the nyquist frequency)

And three different type sigma-delta modulated encoding FIR filters are described as

follow.

(1) FIR filter implemented using PCM input signal and sigma-delta modulated filter

coefficients. ( as shown in Fig. 12)

Fig 12

dn is a sigma-delta modulated version of an. Since dn only takes on values if +1

and –1, each multiplier in Fig. 10 can be implemented using an inverter and a switch.

(in order to obtaining dn using sigma-delta modulation, it is necessary to upsample an

by R times.). In addition, a efficient two step method two interpolate the input signal

xn by R times is used in [1] as the block of “interpolate and append zeros”. And the

decoder is shown in Fig. 13.

8
Fig. 13 Decoder for signal modulated by single-loop sigma-delta modulator.

Remark: since the convolution of two signal in time domain is equivalent to the

product of the two signal in frequency domain. If the spectrum of dn is approximate to

an , the filtering process of Fig. 12 ( the sigma-delta encoding FIR ) and Fig. 11 ( the

original FIR ) is approximately equivalent. (Assume the upsampling of x doesn’t

affect the baseband spectrum itself).

(2) FIR filter implemented using sigma-delta modulated input signal and PCM filter

coefficients. ( as shown in Fig. 14)

Fig 14

Here xn is encoded using a ingle-loop sigma-delta modulator into a sequence vn which

only takes value of +1 and –1. Therefore each multiplier can be implemented by using

a switch to select either the stored coefficients an or –an , depending on the value vn .

The advantage of this configuration is that the sigma-delta modulator operates as both

a coder of signal and as an A/D converter. FIR filter implemented using sigma-delta

modulated input signal

9
and sigma-delta modulated filter coefficients. ( as shown in Fig. 15)

Fig. 15

l A filter design example [1] – simulation results

the impulse response is

by choosing α=0.2,β=0.25. the digital impulse response is

, where

If we implement the digital filter using infinite precision arithmetic,

10
The resulting transfer function would be

The magnitude response of the filter is shown in Fig. 16

Fig. 16

Frequency response if this filter where the impulse response is encoded by

single-stage sigma-delta modulator with oversampling ratio R=32,128,512 is shown n

Fig. 17

11
Fig. 17

and my simulation (using matlab) is as follows:

Fig. 18

12
compare with the results of Fig.17 in [1]. There are more passband ripple and

overshoot in my simulation result. I think it may be the initial value problem for the

sigma-delta modulator ( since a sigma-delta loop contains feedback, the output will

exhibit a starting and ending transient response) . In [4], a periodically extended FIR

impulse response as Fig 19 is used as the input to find the steady-state sigma-delta

modulation output.

Fig. 19

And Fig. 20 is the two stage version with R=8,16,32.

13
Fig. 20

And the two-stage sigma-delta modulator is as Fig.21

Fig. 21

(obviously its output is : –3 –1 1 3)

and my simulation (using matlab) is as follows:

14
Fig. 22

We can see that for higher oversampling ratio, the finite precision filter is more

approximately to the infinite precision filter. And for comparable performance, filter

realized using two-stage sigma-delta modulation require a substantially smaller

oversampling ratio (corresponding a smaller data rate) than those realized using

single-loop sigma-delta modulation.

l CSD sigma delta encoded FIR

sigma-delta loop, one bit resolution increment will cause 6dB improvement in SNR.

Consider the same filter example discussion in previously section but with 2-bits

sigma-delta encoding (as Fig. 23)

Fig. 23

. The simulation results is as follows:

15
Fig. 24

compare with one-bit sigma-delta encoding with the same oversampling ratio(32), the

magnitude response of the encoded FIR is much improved. But the multibit quantizer

suffer the disadvantage of the needing of multibit multiplier. However, by using

canonic signed-digit (CSD) as shows in Fig. 25. We can have 148 quantization levels

but requires only one addition as Fig. 24.

Fig. 25

16
and the magnitude response of the 2-bits CSD sigma-delta encoded FIR (the same

example as previously discussion) is as follows:

Fig 26

If a 50dB stopband attenuation is desired, the CSD sigma-delta encoded FIR filter

need only be sample by a factor of 8, but need only one addition for the

multiplication.

l Conclusion
The above discussion show a alternatively way to realize digital filtering and

overcome several disadvantage of PCM digital filtering.

And in the filter design example, we found that there is a trade-off on the

oversampling ratio and the order of the sigma-delta modulator. Considering both

effects leads us to conclude that there is a potential overall saving in hardware by

using two-stage sigma-delta modulation. Although it is in principle advantageous to

use higher order sigma-delta modulation systems for encoding, there might be

potential problems as the number of stages increase due to gain matching, component

tolerance, and other difficulties..

17
l References

[1] Ping Wah Wong, “FIR filters with sigma-delta modulation encoding”, IEEE

Transactions on Acoustics. Speech. And Signal Processing, VOL. 38, NO. 6, JUNE

1990

[2]Ping Wah Wong, “Fully sigma-delta modulation encoded FIR filters”, IEEE

Transactions on Signal Processing, VOL. 40, NO. 6, JUNE 1992

[3]Pervez M. Aziz,Henrik V. Sorensen,and Jan Van Der Spiegel,”An overview of

Sigma-Delta Converters” IEEE Signal Processing Magazine,pp.61-84,January 1996

[4]Scott R. Powell and, Paul M. Chau “Multiplierless FIR Filters With Discete-Value

Sigma-Delta Encoded Coefficients”, Signals, Systems and Computers, 1991. 1991

Conference Record of the Twenty-Fifth Asilomar Conference on , 1991

Page(s): 1010 -1014 vol.2

18