The present invention relates generally to the problem
of filtering, decimation or interpolation and frequency conversion in the digital
domain, and more particularly to the use of a modified fast convolution algorithm
in wideband multichannel receiver, channelization, and transmitter, de-channelization,
structures of a radio communication system.
In radio base station applications for cellular, Land Mobile
Radio (LMR), satellite, wireless local area networks (WLAN's) and other communication
systems, many receiving and transmitting channels are handled simultaneously. In
the future this will also become the situation for the terminals, i.e. mobile telephones.
There exist channelization and de-channelization structures in the receiver and
transmitter, respectively, in these radio systems. Channelization and de-channelization
can be defined as the filtering, decimation/interpolation and the frequency conversion
of the signals transmitted and received.
The traditional receiver architecture as seen in FIG. 1
can be explained in terms of the Radio Frequency (RF) signal being received by the
antenna 105 and then downconverted to an intermediate frequency (IF) by an RF front
end 110. The RF front end 110 consists of components such as Low Noise Amplifiers
(LNA's), filters and frequency conversion circuits. The desired channel is then
extracted by the receiver channelizer 120. The channelizer 120 also consists of
LNA's, frequency conversion circuits and filters.
The desired channel is then processed at baseband by the
RX baseband processing unit 130 to produce the received digital data stream. Today
baseband processing usually consists of analog-to-digital conversion, digital filtering,
decimation, equalization, demodulation, channel decoding, de-interleaving, data
decoding, timing extraction etc.
The traditional transmitter architecture in FIG. 1. is
the dual of the receiver architecture. The transmitted data is first processed by
the TX baseband processing unit 140 which consists of data coding, interleaving,
channel coding, modulation, interpolation filtering, digital-to-analog conversion
etc. The baseband channel is then converted to an IF frequency via the transmit
de-channelizer 150. The transmit de-channelizer 150 consists of filters, frequency
conversion circuits and low power amplifiers. The IF signal is then converted to
RF and amplified by the RF front end 160 which consists of frequency conversion
circuits, filters, and a high power amplifier. The signal is then transmitted by
the antenna 165.
Figure 1 illustrates the traditional architecture for a
single channel receiver and transmitter as used in a mobile terminal (i.e. mobile
phone) application. In the case of a basestation, multiple channels are processed
in a similar way. On the receiver end the path will split at some point to form
multiple paths for each channel being processed. On the transmitter end the channels
will be processed individual and then they will be combined at some point to form
a multichannel signal. The point of the split and combination varies, and therefore
a variety of basestation receiver and transmitter architectures can be created.
More importantly, though, the traditional analog and digital interface is currently
somewhere between the channelizer and baseband processing blocks.
The analog channelizer/dechannelizer is complex to design
and manufacture, and therefore costly. Therefore, in order to provide a cheaper
and more easily produced channelizer/de-channelizer, the future analog and digital
interface will lie, instead, somewhere between the RF front end and channelizer
blocks. Future radio receiver and transmitter structures of this type are called
a variety of names, including multistandard radio, wideband digital tuners, or wideband
radio and software defined radio, and they all require a digital channelizer/de-channelizer.
Efficient digital channelizer/de-channelizer structures,
consisting of filtering, decimation/interpolation and frequency conversion, are
very important in terms of power consumption and die area on a per channel basis.
With one of the main goals being to integrate as many channels into a single Integrated
Circuit (IC) as possible there are several known ways to achieve digital channelization/dechannelization.
In the following examples it is assumed that a wideband signal is sampled by an
ADC. The wideband signal is centered at an Intermediate Frequency (IF) and typically
consists of many Frequency Division Multiplexed (FDM) channels.
The most obvious way is shown in Figure 2. This receiver
architecture mimics the functions of a traditional analog channelizer with In-phase
and Quadrature(IQ) frequency conversion using e.g. sin/cos generators, decimating
and filtering on a per-channel basis. The bulk of the decimation filtering can be
done with computationally cheap CIC filters. Integrated circuits containing this
architecture are readily available from several manufacturers. The dual of this
architecture is also possible for the transmitter.
The IQ channelizer is flexible in that it can handle many
standards simultaneously and that the channels can be placed arbitrarily. Its main
drawback is the need for an IQ frequency conversion at a high input sampling frequency
and subsequent decimation filters for each channel. This means that the die area
and power consumption is relatively high per channel.
Another channelizer possibility is to build a decimated
filter bank in the receiver, as shown in Figure 3. This method shares a common polyphase
filter between many, or all, channels. The hardware cost for this structure is small
since it is split between many channels, and good filtering can be achieved. Filter
banks are also good for use in transmitter de-channelizers since they both interpolate
and add the channels together. An example of this is illustrated in
"Wideband FFT Channelizer".
Many satellite transponders are also built upon this principle.
Although these filter banks can be reconfigured to fit different standards, it is
still difficult to accommodate multiple channel spacings at the same time. The decimated
filter bank has a very low cost per channel, but only if all or the majority of
channels are used. This architecture is also very inflexible since the channels
have to lie on a fixed frequency grid and only one channel spacing is possible.
Multiple standards make the filter bank concept require multiple sampling rates,
which means multiple architectures, including the ADC and channelizer, are required
for simultaneous multiple standards.
A variation on the structure of the decimated filter bank,
called a subsampled filter bank, can lower the computational cost at the expense
of flexibility. For example, requirements for adaptive channel allocation, irregular
channel arrangements and frequency hopping precludes using subsampled filter banks,
since all channels must be available at the same time.
The third main channelization technique is based on the
fast convolution scheme of the overlap-add (OLA) or overlap-save (OLS) type. Fast
convolution is a means of using cyclic convolution to exactly perform linear convolution,
i.e. Finite Impulse Response (FIR) filtering. A state of the art fast convolution
algorithm is shown conceptually in Figure 4. The input data is divided into overlapping
blocks in the Block Generator. These blocks are discrete Fourier-transformed in
the DFT (Discrete Fourier Transform) and subsequently multiplied point-by-point
with a filter response in the frequency domain. This filter response can be obtained
by discrete Fourier-transforming the impulse response of a filter. The blocks are
then transformed back to the discrete time domain by the Inverse DFT (IDFT) and
added together in the Block Combiner. The advantage of this technique is the lower
computational requirement as compared to implementing the traditional form of linear
However, it is possible to modify the basic fast convolution
algorithm such that it is possible to simultaneously decimate/interpolate and frequency
convert, at the expense of then only approximately performing linear convolution.
If the standard fast convolution algorithm is modified so that it includes frequency
shifting and decimation/interpolation, it can be used for channelization and dechannelization.
Generally, one of the transforms is much smaller than the other when this type of
modified fast convolution algorithm ("MFC") is used. This reflects that the channels
are narrowband compared to the digitized spectrum. Figure 5 shows conceptually how
a modified fast convolution algorithm of the overlap-save type works in the function
of a channelizer. The modifications also reduce the computational complexity. The
stand-alone modified fast convolution algorithm, as illustrated in "
A Flexible On-board Demultiplexer/Demodulator", Proceedings of the 12th AIAA
International Communication Satellite Systems Conference, 1988, pp. 299-303
, is claimed to be a very computationally efficient technique for systems
containing a mixture of carrier bandwidths, although the technique discussed here
is limited to satellite systems.
The stand-alone modified fast convolution algorithm in
the prior art performs all the filtering alone, without any additional signal processing.
This method leads to various delays. However, delays are an inherent part of satellite
systems, due to the time to transmit to and from the satellite. Thus, delays due
to the filtering method effect the system proportionately less than if the stand-alone
modified fast convolution algorithm were to be used in a radio, e. g. cellular,
system. In most radio systems the delay becomes a much more crucial factor which
should be reduced as much as possible.
The stand-alone modified fast convolution algorithm, applied
to the receiver channelizer, chops the incoming data signal into blocks whose size
depends on the percentage of overlap (%overlap) and the length of the Discrete Fourier
Transform (DFT). The DFT is subsequently performed. The truncated filter response,
that is the number of filter coefficients (Ncoefficients) is less than
the length of the DFT (NDFT), is implemented directly in the frequency
domain. This is accomplished by multiplying the filter coefficients with the selected
output bins of the DFT. The result is then processed by an Inverse Discrete Fourier
transform (IDFT) of equal length to the truncated filter as a means to recover the
time domain samples of the desired channel. The blocks are then overlapped, depending
on the %overlap, and combined. The combination is either a process of adding the
overlapped section, overlap and add, or discarding the overlapped section, overlap
and save. Note that overlap/add and overlap/save can be considered two extremes,
and there are techniques known in the art that lie in-between these two.
A frequency domain implementation for replacing convolution
by multiplication of transforms is described by
E.R. Ferrara, "Frequency Domain Implementations of Periodically Time-Varying
Filters", IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 33,
no. 4, pages 883-892
. An overlap-save implementation having 50% overlap, employing M FFTs and
M inverse FFTs, is described. Half of the inverse FFT output samples are thrown
The truncation of the frequency response in the stand-alone
modified fast convolution algorithm distinguishes it from the standard fast convolution
approach. It causes the circular convolution algorithm to now only approximate linear
convolution, although with carefully chosen coefficients the error can be kept small.
Truncation of the frequency response also performs decimation by a factor of (Ncoefficients/NDFT),
and the frequency conversion is completed by centering the truncated filter coefficients
on the wanted channel.
The truncated frequency response also causes a dramatic
reduction in the computational complexity in the channel specific parts of the algorithm,
that is everything but the DFT. The number of multiplications needed to implement
the frequency filter and the size of the IDFT are reduced by approximately a factor
of (Ncoefficients/NDFT). The stand-alone modified fast convolution
algorithm can also be applied to the transmitter de-channelizer, containing all
the same attributes.
Other reductions in complexity that can be applied to standard
fast convolution, can also be applied here to the stand-alone modified fast convolution
algorithm. For example the DFT is a critical block in the operation. For efficiency
reasons it is usually implemented in the form of a Fast Fourier Transform (FFT).
Additionally, two real data blocks can be processed at the same time in one complex
DFT processor. Some extra adders and memory are then needed for post-processing.
This is more efficient than using two dedicated real DFTs.
Computational savings can also be made in the DFTs through
the use of pruning, since only a part of the DFT outputs need to be calculated.
Pruning refers to the process of cutting away branches in the DFT that do not affect
the output. The output points that are not needed are never computed.
A computational reduction can also be achieved if the complex
multiplication of the filter frequency response is replaced by real multiplication
and a subsequent circular shift of the IDFT output block of data before it is combined
to form the time domain samples of the desired channel. The amount of circular shift
depends only on the %overlap and the length of the IDFT.
There is still a problem with the above-identified systems,
especially in future systems involving the reception and transmission of many channels
simultaneously. As seen above, the choice of a digital channelizer, employed from
a few channels up to a large number of channels, is very dependant upon the target
radio communication system or systems. Invariably a trade-off between computational
cost and flexibility based on the radio systems requirements will make the ultimate
decision of which wideband channelizer algorithm to choose. There is still room
to improve these channelizer/dechannelizer structures in terms of computational
cost and flexibility so that they may be better suited for use in systems with many
A solution to the above-described problems was introduced
in the patent publication
"Digital Channeliser and De-Channeliser", R. Hellberg. In that patent
application an improved modified fast convolution algorithm is described which efficiently
handles the problems associated with conventional channelizers/de-channelizers (i.e.,
the problems with computational cost, flexibility and acceptable delay with respect
to designing those systems to handle multiple channels simultaneously).
The improved modified fast convolution algorithm as described
improves upon the stand-alone modified fast convolution when applied to
radio communication systems, see Figure 2. It splits the necessary filtering between
the MFC algorithm and additional channel filtering, thereby improving the power
consumption, die area and computational complexity when compared to the prior art
stand-alone MFC. It is also a very flexible algorithm in terms of designing it for
combination of different systems parameters, sampling frequency, channel bandwidth,
channel separation and bit-rate. A further advantage in
is that the MFC part of the algorithm processes smaller-sized blocks and
therefore produces smaller delays, delays which become acceptable for land-base
radio communication systems. The improved M FC algorithm is considered to be very
suitable for e. g. channelization/de-channelization in a wide variety of radio communication
systems. It is therefore a good choice for future hardware platforms that will support
multiple standards for more than a few channels activated at any one time.
The present invention in this application is applicable
to both the stand-alone MFC as found in"A Flexible On-Board Demultiplexer/Demodulator",
discussed above, and the improved MFC as found in
. When we use"modified fast convolution" ("MFC") in the remainder of the
application we use it to mean either a stand-alone MFC or the improved MFC disclosed
In the modified fast convolution algorithm as used in state
of the art channelizers, the number of points in the DFT and IDFT (computed by the
Fast Fourier Transform, FFT, and the Inverse Fast Fourier Transform, FFT, respectively)
are powers of two. Since the lengths of the input and output transforms are both
powers of two, overlaps of 50%, 75&, 25% (generally k*1/2n) are possible.
The decimation and interpolation ratios are limited to powers of two (NFFT/NIFFT).
The successive blocks are, in the state of the art algorithm, multiplied in the
frequency domain with identical filter responses, H(k).
Since the transform sizes are powers of two, many parameters
in the state-of-the-art modified fast convolution algorithm are fixed to the values
achievable with this set of transform sizes. Thus, a problem here is that one would
like to have more flexibility when it comes to the selection of the decimation ratio
and overlap. For example, if the input FFT is 1024 points and the output IFFT is
32 points, the decimation ratio is 1024/32, i.e. 32. The overlap can be 50%, 75%,
25%, down to k*1/32, where "k" is an integer. If the IFFT size is instead chosen
to be 64, the decimation ratio is 16 and the smallest granularity is considered
to one bin, therefore the granularity of the overlap is 1/64 of the block length
(possible overlaps k*1/64).
SUMMARY OF THE INVENTION
The present invention relates generally to the problem
of filtering, decimation or interpolation and frequency conversion in the digital
domain, to the use of a modified fast convolution algorithm in wideband multichannel
receiver, channelization, and transmitter, de-channelization, structures of a radio
communication system, and more particularly to the problems discussed above. The
means of solving these problems according to the present invention are summarized
in the following.
It can be seen above that the modified fast convolution
algorithm as used in the e. g. channelization/de-channelization structures of
is limited because the input and output transforms are both powers of
two. This limits the number of possibilities in the decimation and interpolation
ratios for channelization and dechannelization, respectively. Because these transform
sizes are limited to powers of two, other parameters in the algorithm are then fixed
to values achievable with this set of transform sizes.
Because transform sizes in the state of the art are powers
of 2, it is not obvious to see that it is possible to use transform sizes in either
end that also has other factors in them than two, like three or five. This works,
without substantial modifications to the algorithm, as long as both the input and
output transform have factors in common to provide for the overlap. For example,
a common factor of four makes possible 50%, 75% and 25% overlap, or a common factor
of three would make possible overlaps of 33% and 66%.
The expansion of the set to include transforms built with
other factors than two increases the applicability of the modified fast convolution
algorithm to beyond that of the state of the art and can be implemented with little
modification to the algorithm except the change from power of two lengths of one
of the transforms to lengths containing also other factors. This expansion of the
set of usable transform lengths is of course welcomed. However, since the large
size of the input and output transforms often has to be calculated in real time,
this transform in practice is limited to a length that is a power of two. This,
together with the requirement for having common factors in the input and output
transform lengths, makes the range of possible choices of transform size very limited.
One would like to have odd transform lengths on either end together with even length
or power of two lengths on the other end. One could have 1/3 or 3/11 or other unusual
overlaps with even length or power of two lengths on either end.
The basic problem when using input and output transforms
without common divisors in the modified fast convolution algorithm is that no overlap
fits both ends at the same time. For example, if the FFT size is 128 and the IFFT
size is 27, the input end would fit overlaps k*1/2n
and the output end would like to have overlaps of the kind k* 1/3n
. These are clearly incompatible.
The present invention makes it possible to use input and
output transforms in the modified fast convolution algorithm whose sizes contain
factors other than powers of two, e.g. three or five. This works, without substantial
modifications to the algorithm, as long as both the input and output transform have
factors in common to provide for the overlap. For example, a common factor of four
makes possible 50%, 75% and 25% overlap, or a common factor of three would make
possible overlaps of 33% and 66%. In the preferred embodiment, other modifications
will be added to the invention which will extend the use of the modified fast convolution
algorithm to input and output transforms which have no common factors.
Figure 6 shows the situations in which the present invention
are used. Across the top of the chart the MFC is divided into situations where the
input and output transforms either have a common factor (CF) or do not have a common
factor (No CF) in common with the overlap. On the left side of the chart are shown
three situations for the sizes of the transforms. In the first one both the input
and output transform sizes are a power of two. In the second one, one of the transforms
is a power of two while the other is a power of two further multiplied by some other
integer. In the third example both transform sizes can be any integer n1
and n2, where both of them may either have or not have a common factor
with the other. The prior art examples of the modified fast convolution algorithm
were limited to the case where the input and output transforms were both powers
of two, having factors in common with each other. The present invention also works
in the same cases as the prior art, but it also extends the MFC to cases where the
transform sizes may be any size whatsoever, with or without factors in common.
Accordingly it is an object of the present invention to
provide a method to increase the flexibility when it comes to the decimation ratio
and overlap in the modified fast convolution algorithm.
The object of the present invention is solved by the subject
matter of the independent claims. Advantageous embodiments are defined in the dependent
The invention can be described as consisting of essentially
2 steps. The first step is to make sure that we use different overlaps on consecutive
blocks that, on average, give the same overlap on both the input and output ends.
The second step is to align the signal in consecutive blocks of time. A third step
of an advantageous embodiment is then to compensate for phase shifts due to frequency
The essence of the invention is that it decouples the input
and output transform lengths in the modified fast convolution algorithm from each
other and from the overlap. it makes it possible to use any transform length on
the input together with any transform length on the output and at the same time
use any overlap. This provides an enormous amount of freedom compared with the limitations
of state of the art. Input and output sample rates can now be chosen with much finer
resolution, and decimation and interpolation ratios need no longer be powers of
Although the invention has been summarized above, the method
according to the present invention is defined according to appended claim 1. Various
embodiments are further defined in the dependent claims.
The present invention is not discussed in terms of any
particular system. It is particularly applicable to many radio base station applications
in e.g. cellular, Land Mobile Network (LMR), satellite, wireless local area networks
(WLAN's). However, it is not limited to these systems and may, in general, be used
in any system using the modified fast convolution algorithm. In addition, it's use
is not restricted to use in basestations, but may also be used in e.g. future mobile
terminals that are also capable of handling multiple channels simultaneously.
BRIEF DESCRIPTION OF THE DRAWINGS
The present invention will now be described in more detail
with reference to preferred embodiments of the present invention, given only by
way of example, and illustrated in the accompanying drawings, in which:
- FIG. 1 is a diagram of a traditional radio transmitter and receiver architecture.
- FIG. 2 is a diagram of a state of the art IQ-demodulating digital receiver.
- FIG. 3 is a diagram of a state of the art decimated filter bank.
- FIG. 4 is a diagram of a state of the art fast convolution algorithm.
- FIG. 5 is a diagram of a modified fast convolution algorithm of the overlap-save
- FIG. 6 illustrates the various combinations of input and output transform sizes
covered by the prior art and the present invention.
- FIG. 7 is an overview of the three basic steps of the present invention.
- FIG. 8 shows the phase of the frequency response and the corresponding delay
of the impulse response.
- FIG. 9 illustrates the accumulated difference in non-overlapping parts of the
input and output blocks.
- FIG. 10 illustrates the frequency shifts of the modified fast convolution algorithm.
- FIG. 11 illustrates a fully implemented system according to the present invention.
The method according to the present invention separates
into three varieties: (1) letting the input transform determine the overlap, (2)
letting the output transform determine the overlap, or (3) choosing an overlap that
is independent of both the input and output block lengths. These solutions contain
the same ingredients, as the inventive aspect is the same, but will look slightly
different in their specific implementations. Examples from the first two varieties
will be included in the following description of the solution steps.
There are essentially 2 steps in the method of the present
invention as shown in Figure 7. The first step 710 is making sure that we use different
overlaps on consecutive blocks that, on average, give the same overlap as on the
opposite end of the algorithm. The second step 720 is to align the signal in consecutive
blocks of time. A third step 730 of a advantageous embodiment is then to compensate
for phase shifts due to frequency shifting. These steps will be described in further
The first step 710 is making sure that we use different
overlaps on consecutive blocks that, on average, give the same overlap as on the
opposite end of the algorithm. This is due to the fact that in order to be able
to use transform lengths without common divisors (i. e. GCD [n1, n2]
= 1, "GCD" is "Greatest Common Divisor") we must first make sure that the input
and output rates are compatible with the bandwidth of the input and output transforms,
If the overlap is denoted l/m, then in general m different
overlaps will have to be used on one or both ends of the algorithm. If l and m have
a common factor, they can both first be reduced by this factor. We can then create
vectors of length m, representing the lengths of either the overlapping or the non-overlapping
parts of the blocks, that on average give the overlap l/m. The sequence of
overlaps (or non-overlaps) will repeat cyclically, although it is also possible
to implement the invention so that the overlaps come randomly, as long as the average
on both sides of the algorithm is the same. If m is a factor in the length of one
of the transforms, that end can have the same overlap for all blocks.
The first step 710 can be illustrated with an example where
we assume that the input transform has even length, the output transform has odd
length and the overlap is 50% on the input. The overlap on the output could then
be separated into two overlaps, one of odd length and one of even length, that are
used interleavingly on every other block .
As a second example we assume that the input transform
is 128 points and the output transform is 25 points. An overlap of 2/5 can then
be achieved through e.g. having the non-overlapping parts of the input 77, 77, 76,
77 and 77 samples long, respectively. The average of these numbers is 76.8, which
divided by 128 gives 3/5 (overlap = 2/5). The overlap on the output would be 2/5
of 25, i.e. 10, for all blocks.
In the second step 720 of the invention we align the signal
in the consecutive blocks in time. When using different overlaps the blocks' starting
points do not come regularly, as would occur when the overlaps are the same. For
example, when using a 27-point IDFT together with an even DFT and 50% overlap, the
time between the first sample of the blocks on the output end could be 13 and 14
samples, respectively. This yields an average of 13.5.
The time alignment is done by time-shifting the signal
within the block so that it compensates for the slight misalignment of the starting
time of the respective blocks. This can be done by multiplying the DFT samples coming
from different blocks by sinusoids with different incremental phase shifts (the
DFTs of the different delays). An equivalent, simpler and less computationally complex
approach would be to multiply the coefficients of the filter response in the frequency
domain, H(k), with the same incremental phase shift. This means that a set of m
filter responses, corresponding to the m different time shifts, is needed.
A time shift of the impulse response of x samples in a
block of length n is obtained by multiplying with a complex exponential, a sinusoid,
over the frequency response samples H(k), so that
We see that x/n is a measurement of the delay as
"part of the block length". The correspondence between the incremental phase shift
(-2pix/n in the equation) of the frequency response and the delay of the
impulse response is shown in Figure 8. Observe that a negative time shift of
n/5 (corresponding to x/n = -1/5) looks like a positive time shift
of 4n/5, since all shifts are cyclic within the block.
The time alignment is calculated via the difference between
the relative starting points for the blocks on the input and output ends of the
algorithm, which is the same as the accumulated difference between the input and
output non-overlapping parts. Figure 9 illustrates this concept.
The incremental phase shift to the bins of block p (p>1)
where n1 and n2 are the lengths of the input and output transforms
and nolp1(q) and nolp2(q), see 1190, 1195, respectively, Figure
10, are the lengths of the overlapping parts of the q-th input and output blocks,
respectively. The time compensation factor for block one, Tc(1) can often be set
to zero, but a certain time shift can also be added equally to all blocks (by setting
Tc(1) to a value other than zero) in order to minimize the maximum absolute time
The time alignment is implemented by multiplying the original
frequency response, H(k), with a sinusoid, ejTc(p)*k, , so that for the
filter response number p
An effort should be made to have a small accumulated difference
in (non-)overlap, the sum in the equation for Tc(p), through the design of the input
and output overlap vectors, while at the same time making the overlap as close as
possible to I/m. This increases the maximum length of the impulse response of the
filter that can be implemented.
In the example discussed above with a 27-point IDFT and
50% overlap we had alternating times between the first sample of each block of 13
and 14 bins, yielding an average of 13.5. We have to account for this extra half-sample
timeshift, and we would need a positive timeshift of half a sample (counted on the
output) on the blocks that come 13 samples after the previous block. This would
make the apparent starting points of the signal in the blocks separated 13.5 samples
(27/2), which is what we want. In practice this is achieved by having two frequency
responses with a difference in delay of half a sample. An incremental phase shift
of 2pi/27* 1/4 and 2pi/27*-1 /4 per bin, respectively, for the two frequency responses
is thus required.
For the example with a 128 point input and 25 point output
transform, one must compensate for the five different starting points of the input
blocks relative to the output blocks. These differences are 0, +1/5, +2/5, -1/5,
-2/5, sampled on the input end. This calls for having five filters with different
time shifts. These would be implemented as five different incremental phase shifts
in the frequency domain with the values 2pi/128*0,2pi/128*1/5,2pi/128*2/5, 2pi/128*-1/5,
2pi/128*-2/5 per bin, respectively.
Finally, a third step 730 of a method according to an advantageous
embodiment of the present invention is to compensate for phase shifts due to frequency
shifting. Figure 10 illustrates two shifts, Shift #1 and Shift #2, possible in the
modified fast convolution algorithm as implemented in a channelizer/de-channelizer.
In the modified fast convolution algorithm, when used for channelization/dechannelization,
a frequency shift, #1, is included which is performed by using a certain range of
the frequency domain samples coming from the DFT 1020 in a channelizer 1000 or by
inserting the filtered DFT samples at a certain place in the large IDFT of a dechannelizer.
In the channelizer 1000 this can be viewed as if the lowest selected bin of the
DFT 1020 is shifted down to zero frequency, and in the dechannelizer as if the zero
bin of the DFT is shifted up to the lowest frequency of the channel.
There is also the possibility of circularly shifting 1040
the bins within the selected range after the multiplication with the frequency response
1030. This is done in order to shift the center frequency of the signal within the
decimated frequency range. This technique, described further in the patent publication
US 6 167 102A
"NCO Size Reduction" to Hellberg, depends on the possibility to perform
this shift. Observe that Shift #1 in the case of a channelizer 1000 is a negative
frequency shift since the first bin of the range of bins going into the multiplication
with the frequency response is shifted down to zero frequency.
A shift of the bins in the frequency domain corresponds
to a multiplication of the time samples in a block of size n by a sinusoid, ej2pi*f/n*t,
where f is the frequency shift (an integer) and t is the number of a sample in the
block of size n. Over one whole transform block, all sinusoids corresponding to
different shifts in the frequency domain return to their initial phase. However,
if the blocks are overlapping, they are patched together with the next block at
an earlier point at which the sinusoid in general will not have returned to the
initial phase. This means that we will have a phase discontinuity between the blocks.
A compensation therefore has to be performed in order to align consecutive blocks
in phase to correct for phase errors due to the frequency shifting in the frequency
The phase compensation is done by calculating to which
phase the modulating sinusoid has moved during the non-overlapping part of the block
and shift the phase of the next block accordingly, by multiplying the whole block
with a constant phasor. After a certain number of blocks, in general the same as
the number of different timeshifts, m, the phase has returned to its initial value.
The phase compensation can also be incorporated into the
set of filter responses by multiplying the elements of each filter response with
a constant phasor, since the number of different phase shifts that are needed generally
is the same as the number of different time shifts. In addition to being dependent
on the length of the accumulated non-overlapping parts of the previous blocks the
phase compensation also depends on the frequency shift, which means that an individual
set of filter responses is generally required for each channel in the channelizer.
The-phase compensation for the p-th (p>l) block is
where n is the length of the transform, nolp(q) is the length of the
overlapping part of the q-th block and fshift is the frequency shift.
The phase compensation for the first block, Pc(1), can be set to zero. The values
of accumulated non-overlapping parts times the frequency shift that are above n
can be reduced modulo n, since this number represents one full circle of the phasor.
The values of accumulated non-overlapping parts themselves can of course also be
reduced modulo n.
The phase compensation is performed by multiplying all
elements in the time-aligned filter response of block p, Hp(k), with the constant
phasor ejPc(p), so that the phase compensated frequency response Hc,p(k)
The formula for the phase compensation is the same for
both Shift #1 and Shift #2, but the lengths of the transforms are in general different
and the accumulated non-overlapping parts are in general also different and have
to be calculated separately.
If both frequency shifts are included, two phase compensations
must be calculated. They can be summed and incorporated into the same phase compensation.
The formula for phase compensation then becomes:
where the phase compensation for the first block also here is assumed to be zero.
In a dechannelizer, the shift associated with inserting the filtered DFT samples
at a certain place in the large IDFT is a positive shift, which must be remembered
when using the formula for phase compensation.
For the example with the non-overlapping parts of the input
77,77,76,77 and 77 samples long, the accumulated value from the previous blocks
is 77,154,230 and 307 for blocks 2 through 5. Assuming only #1, the phase compensation
would then become 0, 2pi*77/128*fShift1, 2pi*154/128*fShift1,
2pi*230/128*fShift1 and 2pi*307/128*fShift1. Since 307+77,
384, is divisible by 128 the phase will return to zero after five blocks, and the
sequence of phase compensations can be repeated.
In the example system with 50% overlap previously described
the phase compensation due to the first shift is the same as for systems having
transforms with common factors and uniform overlaps. This would also be the case
for other systems where the overlap is determined by the DFT size in the case of
a receiver (channelizer) or determined by the IDFT size in the case of a transmitter
(dechannelizer). This phase compensation is quite simple to implement since it is
calculated modulo 2 in the case of 50% overlap and modulo 4 in the case of 75% and
25% overlaps. It is also computationally cheap since the multiplication of the blocks
by the two or four uniformly spaced phasors are just multiplications by plus and
minus one or multiplications by plus and minus one and swapping the real and imaginary
parts of the signal. This phase compensation has previously been thoroughly described
in U. S. Patent publication No.
US2001025291, titled "Flexibility Enhancement to the Modified Fast Convolution
Algorithm" filed on September 18, 1998 to Leyonhjelm et al.
On the other end, the phase compensation due to a cyclic
shift within the smaller range (#2 in Figure 10) would be dependent on the different
overlaps on this end. For the example system having a 27-point IDFT this compensation
would be 0 and 2pi*13/27*fShift2 on consecutive blocks.
Above, it was mentioned that the present invention can
be divided into three varieties: (1) letting the input transform determine the overlap,
(2) letting the output transform determine the overlap, or (3) choosing an overlap
that is independent of both the input and output block lengths. We will now discuss
an example of the third type: a fully implemented system where the overlap is independent
of both the input and output transform lengths. As an example of a modified fast
convolution system having transforms with no common factors and overlap independent
on either transform size we use n1=256 (=28), n2=23
(prime) and the overlap I/m=3/7.
Neither the input or output transform length can be divided
by 7, so both input and output (non-)overlap vectors of length 7 have to be created.
Since n1*(m-l)/m equals
we let the input vector of non-overlapping parts be [146 147 146 146 146 147 146]
which, as one of many possibilities, averages
In the same way n2*(m-l)/m equals
so the output non-overlap vector [13 13 13 14 13 13 13] is chosen.
The vector of incremental phase shifts, Tc(p), corresponding
to the time alignments of blocks 1 through 7 becomes (all numerators modulo 256
and 23, respectively)
and the phase compensation, Pc(p), for blocks 1-7 is accordingly
where the frequency shifts have been left variable. Remembering that fshift1
is a negative frequency shift in a channelizer, if the range to be filtered starts
at e.g. 97 the value of fshift1 would be -97.
The fully implemented system is shown conceptually in Figure
11. The figure illustrates that there are several frequency responses 1130 that
each has a time alignment dependent on the input and output overlap and transform
lengths. The phase compensations 1170, 1175, depend on the overlap, transform length
and shift on each end and are also performed on each of the frequency responses
The preferred implementation as described above uses m
different frequency responses 1130 into which all time alignments 1180 and phase
compensations 1170, 1175, are absorbed. This means that these frequency responses
1130 can be computed once and then used for a certain channel as long as desired,
which implies a low computational cost at the expense of increased memory needed
for storing these different responses 1130.
In alternate embodiments the time alignments 1180 and phase
compensations 1170, 1175, can be multiplied to the blocks in real time, minimizing
storage. In yet another embodiment it is possible to multiply the blocks only with
the phase compensations 1170, 1175, which differs between channels, in real time
and to use a set of filter responses 1130 with pre-multiplied time alignments 1180,
which do not differ between channels.
Although the examples in this document have been concentrated
towards channelizers 1100, the operations described work equally well for de-channelizers,
with slight modifications as indicated. They also work for both overlap-add and
overlap-save implementations, for arbitrary lengths of the input and output transforms
and for arbitrary overlaps I/m. Although it may appear from the description above
that the solution is only for the case GCD [n1, n2] = 1, in
fact all other cases when there are not enough common factors, or when the overlap
denominator, m, does not have a factor in common with n1 or n2,
are also covered. Even when there are factors in common between the input and output
transform lengths or between the transform lengths and the overlap denominator the
operations work. These are only special cases which may lead to fewer different
overlaps on either end or fewer alignments and compensations.
The embodiments described above serve merely as illustration
and not as limitation. It will be apparent to one of ordinary skill in the art that
departures may be made from the embodiments described above without departing from
the scope of the invention. The invention should not be regarded as being limited
to the examples described, but should be regarded instead as being equal in scope
to the following claims.