Interpolation-Decimation Circuit for Increasing or Decreasing Digital Sampling Frequency

FIG. 1 depicts, in block diagram form, a general purpose interpolator-decimator circuit which is helpful in understanding this invention;

FIG. 2 illustrates an embodiment of a single-stage interpolator-decimator circuit in accordance with this invention;

FIG. 3 depicts the timing diagram of the single-stage interpolator-decimator of FIG. 2; and

FIG. 4 depicts multistage embodiments of this invention employed as a narrow-band FIR filter.

FIG. 1 depicts a circuit for changing the sampling rate of an applied signal by the factor L/M. It will be recognized by those skilled in the art that this structure is effectively an interpolator circuit cascaded with a decimator circuit in which the interpolator low-pass filter and decimator low-pass filter have been combined in single low-pass filter 13. Since the circuit of FIG. 1 first increases the sampling frequency by an integer factor L and subsequently decreases the sampling frequency by an integer factor M, it can be observed that the sampling rate conversion of the circuit is L/M. Thus, the circuit of FIG. 1 is a general interpolation-decimation circuit in which the interpolation or decimation factor is determined by L and M. Accordingly, if L=1, the circuit becomes an integer decimator, and if M=1, the circuit becomes a typical integer interpolator.

In the operation of the circuit of FIG. 1, a digital input signal having a sampling frequency f.sub.r is applied to input terminal 11. Sampling rate increase unit 12 increases the sampling frequency to Lf.sub.r by inserting L-1 zero-valued samples between the samples of input signal x(n). This signal, denoted w(n) in FIG. 1, is low-pass filtered by filter 13 to produce the signal denoted as v(n). Low-pass filter 13 is preferably an FIR filter realized by a standard direct form implementation, since savings in computation can thereby be realized. When filter 13 is an FIR filter, the signal v(n) is related to the signal w(n) by ##EQU1## where h(m), m=0, 1 . . . , N-1 are the filter coefficients and N is the duration of the unit sample response of filter 13.

Sampling rate decrease circuit 14 decreases the sampling rate by effectively selecting every Mth sample of the interpolated signal v(n). Accordingly, the signal coupled to output terminal 16 exhibits the sampling frequency (L/M)f.sub.r, where L and M may be selected to provide decimation or interpolation by any rational number.

From Equation (1), it can be seen that the computation of an output point depends solely on past and present values of w(n) and not upon past values of any internal filter variables. Thus, the filter computations need only be performed for every Mth output signal. Furthermore, since w(n) is non-zero only for every Lth input point, only a single multiplication and addition need be performed for every Lth input point. Thus, in the interpolator-decimator of FIG. 1, the effective number of multiplications and additions perfomed by the circuit is N/(LM) per output sample instead of N, as would be predicted by a cursory examination of Equation (1).

Although the interpolation-decimation circuit of FIG. 1 could be realized by utilizing state-of-the art interpolator-decimator and filter circuits, such an implementation is generally very complicated, since it involves signals at several different sampling rates and also involves keeping track of the particular multiplication which must be performed.

Utilizing the concepts of the circuit of FIG. 1, we have discovered that a general interpolator-decimator circuit can be realized by relatively simple circuit structure which is similar to the direct form implementation of an FIR filter. If the unit sample response duration of low-pass filter 13 is N samples, L-1 out of every L samples of w(n) are zero-valued and thus the computation is proportional to N/L, since the unit sample response spans approximately N/L non-zero samples. If N is chosen to be N=QL where Q is a predetermined integer, then the unit sample response will span exactly Q non-zero samples of w(n) for each filter cycle. It can be seen that this condition can be easily satisfied since h(n) can be appended with a small number of zero-valued coefficients until N is equal to QL. Defining N in this manner, we have determined that each output point y(n) is of the form ##EQU2## where (nM) .sym. L denotes nM modulo L, [nM/L] denotes nM/L rounded up to the next largest interger, and the sequence h(n), n=0, 1, . . . , N-1 contains the coefficients of the direct form FIR filter, (or, equivalently, the samples of its impulse response).

It can be noted in Equation (2) that to compute each output point y(n), x(n) is sequentially addressed for Q of its values and that [(n+1)M/L] - [nM/L] new input samples are required for the calculation of the next output point, y(n+1). Examining the expression for h(n) in Equation (1), it can be further noted that during the calculation of any output sample y(n), h(n) must be addressed to arguments, (kL+(nM) .sym. L), where L), where K=0, 1, . . . Q-1 Thus, h(n) can be addressed sequentially only if the coefficients of h(n) are stored in an appropriate scrambled order. That is, to determine y(0) the sequence or set g.sub.0 (n) of Q coefficents g.sub.0 (n) = h(0), h(L), . . . , h((Q-1)L) is required. To determine y(1), the sequence or set g.sub.1 (n), of Q coefficients g.sub.1 (n) = h(M .sym. L), h(L+M .sym.L), . . . h((Q-1)L+M .sym. L) is required, whereas to compute y(L-1) the sequence g.sub.L-1 (n), of Q coefficients g.sub.L.sub.-1 (n) = h(((L-1)M) .sym. L), h(L+((L-1)M) .sym. L), . . . , h((Q-1)L + ((L-1)M) .sym. L) is needed. Finally, to compute y(L), the sequence g.sub.0 (n) is again required and the cycle repeats.

FIG. 2 depicts an embodiment of our invention employing the above principles to realize a general purpose interpolator-decimator. In FIG. 2, Q samples of the input signal are stored in shift register 27 and the h(n) coefficients are stored in memory 41 (e.g., a conventional read-only (ROM) memory) in the scrambled order described above.

A particular output sample y(n) is calculated by sequentially shifting an input sample from shift register 27 and the corresponding coefficient from memory 41 to the input terminals of multiplier 29. The product of the input sample and the coefficient is coupled to accumulator 31, which may be a conventional accumulator circuit which normally includes an adder and a register wherein the quantity contained in the register is added to the incoming signal and the resulting sum is stored in the register. Multiplier 29 can be any conventional multiplier circuit capable of handling the digital format employed in any particular embodiment.

The operation of the circuit of FIG. 2 can be best understood by assuming that a particular output sample y(n-1) has just been determined and the circuit is starting the sequence to determine the output sample y(n). At the beginning of such a sequence, signal samples which arrived during the calculation of y(n-1) are contained in buffer memory 22, which may be any conventional memory circuit of suitable capacity. As each of the signals stored in memory 22 arrived at input terminal 21, counter 23 was incremented one count and address circuit 24 directed each input sample into a separate addressable location within buffer memory 22. Thus, at the beginning of any given sequence, counter 23 contains the number of input samples which are contained in buffer memory 22 and must be loaded into shift register 27 for the new calculation sequence. In addition, at the end of the previous sequence, Q-counter 38 of sequencer 33 reached the count Q, activating selector circuit 26 to connect the input of shift register 27 to buffer memory 22.

The calculation sequence is initiated by countdown circuit 35 which effectively counts down from the system master clock frequency, Lf.sub.r, (applied to terminal 34) to supply a signal of frequency L/M f.sub.r, i.e., countdown circuit 35 counts down M samples from the master clock. It will be understood by those skilled in the art that a variety of such countdown circuits are available, e.g., logic gates may be connected directly to the system master clock to generate a single pulse each time the master clock reaches a predetermined count. The output pulse of countdown circuit 35 initiates the calculation sequence by transferring the data stored in buffer memory 22 into shift register 27. This transfer is effected by Clock 25 and counter 30 which operate in conjunction with counter 23 and address circuit 24. The output pulse of countdown circuit 35 starts clock 25 which operates at a frequency high enough to generate at least as many clock pulses as the maximum number of input samples which must be loaded into shift register 27 before Q-clock 37 begins to operate. Each clock pulse of clock 25 increments address circuit 24 and strobes buffer memory 22 to transfer the data samples in buffer memory 22 to shift register 27. These data samples are transferred to the shift register in the order in which they arrived at input terminal 21. The clock pulses generated by clock 25 are counted by counter 30. When the count reaches the count stored in counter 23, the last updating sample is transferred to shift register 27 and counter 30 stops clock 25 and resets counter 23 back to zero. Additionally, the output pulse of countdown circuit 35 switches the state of bistable multivibrator 36 which, in turn, starts Q-clock 37 running. Q-clock 37 is a conventional clock circuit which has a frequency high enough to generate Q pulses during the period M/(Lf.sub.r). Multivibrator 36 is a conventional multivibrator which operates with a sufficient delay to allow completion of the above-described loading sequence.

The clock pulses from Q-clock 37 are coupled to shift register 27 and address pointer 39. As will become apparent from the following discussion, address pointer 39 was automatically indexed to the storage location of the proper coefficient block at the conclusion of the previous calculation sequence. For example, if the previous calculation sequence determined output sample y(0), address pointer 39 would be initially indexed to the last coefficient stored in block 43 of memory unit 41 to begin the calculation of output sample y(1). In this case, shift register 27 would contain the input samples

it can be noted that regardless of which output sample was being calculated, shift register 27 would contain the proper input samples in an ordered sequence due to the previously discussed operation of memory 22, counters 23 and 30, address circuit 24, and clock 25.

During the calculation of each output sample y(n), clock pulses from Q-clock 37 sequentially shift input sample x([nM/L] - k) and coefficient h(kL + (nM) .sym. L) to multiplier 29. Thus, for example, in the calculation of output sample y(1), the first clock pulse from Q-clock 37 shifts x([M/L]-(Q-1)) to one input terminal of multiplier 29, while the coefficient h((Q-1)L + M .sym. L) is coupled to the second input terminal of multiplier 29. Since selector 26 was activated by multivibrator 36 at the same time Q-clock 37 was energized, each input sample is recirculated to the input terminal of shift register 27 as it is shifted to the input of multiplier 29. With each clock pulse from Q-clock 37, the next input sample is shifted to one input terminal of multiplier 29 and address pointer 39 is advanced to the next storage location within memory 41 to provide the proper coefficient to the second input terminal of multiplier 29.

Q-counter 38, which is a conventional counting circuit, counts the number of clock pulses generated by Q-clock 37. When Q-clock pulses have been generated, the calculation of the output sample will be complete and the value of the output sample will be stored in accumulator 31. Upon reaching the count of Q, Q-counter 38 causes bistable multivibrator 36 to be reset, shutting off Q-clock 37, while simultaneously activating selector 26 to connect the input terminal of shift register 27 to the output terminal of buffer memory 22. For example, in the calculation of the output sample y(1), after the (Q-1)th clock pulse address pointer 39 will be indexed to the coefficient h(M .sym. L) in coefficient block 43 of memory 41 and the input sample x([M/L]) will be in the final register of shift register 27. The Qth clock pulse will then couple this input signal and coefficient to multiplier 29 and the product will be added to the sum of the previous products by accumulator 31. At this point, calculation of a single output sample is complete and the value of the sample is stored in accumulator 31. It should be noted that the Qth clock pulse of clock 37 sets address pointer 39 to the last coefficient of the next output sequence g.sub.2 (n) and recirculates the last input sample of the interval calculated within shaft register 27. For example, in the previously referred to case of the determination of y(1), address pointer 39 would be indexed to the last coefficient in the g.sub.2 (n) block of coefficients within memory 41, i.e., h((Q-1)L + (2M) .sym. L) and shift register 27 would hold input samples x([M/L]), x([M/L] - 1), . . . ,x([M/L] - (Q-1)).

During this calculation sequence, counter 23 and address unit 24 have been in operation to count a number of arriving input samples and store each arriving sample in buffer memory 22. It should be noted that the frequency of Q-clock 37 is high enough to ensure that each calculation is complete prior to an output pulse from countdown circuit 35, that is, the Q calculations necessary to determine a particular output sample are completed in a period of time less than M/(Lf.sub.r). Thus, depending on the frequency of Q-clock 37, the input samples necessary to compute the next output sample may or may not have arrived at input terminal 21 when the calculation of a particular output sample is complete. In any case, the samples will arrive prior to an output pulse from countdown circuit 35. When countdown circuit 35 reaches the count M it stops counter 23 and address circuit 24 which, as previously described, in conjunction with clock 25 and counter 30, sequentially shift the input samples stored in buffer memory 22 into shift register 27. Simultaneously, countdown circuit 35 sets bistable multivibrator 36 which starts Q-clock 37, thus beginning the calculation of the next output sample.

It can be noted that the number of samples to be transferred from buffer memory 22 to shift register 27 varies not only with the particular interpolator-decimator ratio L/M, but may also vary during the calculation of particular output samples for a given ratio L/M. For example, it can be shown that where L/M is greater than one, i.e., the circuit is an interpolator, either zero or one update input sample is required for the calculation of any output point y(n). In the case of a decimator, however, several input samples may be required to update the system for a particular calculation. Accordingly, the storage capacity of buffer memory 22 is determined by the ratio L/M of a particular embodiment.

This relationship and the system operating sequence may be further understood by reference to FIG. 3, which illustrates system operation with L/M = 2/9. In FIG. 3, f.sub.r is the sampling frequency of the digital system, i.e., an input sample arrives at input terminal 21, with each pulse in the signal f.sub.r. As previously stated, the output of countdown logic 35, labeled f.sub.rLM in FIG. 3, is derived from the digital system master clock signal, Lf.sub.r, by counting down by the factor M. As shown in FIG. 3, the Q-clock signal f.sub.Q is initiated by each output pulse f.sub.rLM and Q-counter output f.sub.Qc turns Q-clock 37 off after Q-clock pulses have been generated.

It can be noted that FIG. 3 corresponds to the time interval in which two output samples are calculated. During the first subinterval of FIG. 3, denoted t.sub.0 to t.sub.1, four input samples arrive at input terminal 21 and are stored in buffer memory 22. Thus, at time t.sub.1 these four samples will be shifted into shift register 27 to update the system for the calculation of the next output sample. While the system is calculating this output, i.e., during the period of time from t.sub.1 until Q-counter 38 reaches the count Q at t.sub.1 ', three input samples arrive at input terminal 21 and are stored in buffer memory 22. During the period of time between t.sub.1 ', and t.sub.2 (when the next pulse is generated by countdown circuit 35), two more pulses arrive at input terminal 21 and are stored in buffer memory 22. Regardless of whether the input samples arrive during the calculation sequence or in the interval between the end of the calculation sequence and the next pulse from countdown circuit 35, counter 23 counts the arriving input samples. Thus, at time t.sub.2 counter 23 has reached the count of five, and the output pulse from countdown circuit 35 (which occurs just after time t.sub.2) signals clock 25 and counter 30 to load the five input pulses (in the order of arrival at input terminal 21) into shift register 27. Thus, it may be seen that, regardless of the ratio L/M and the corresponding variation in the number of updating input samples required, e.g., updating by 4 samples at t.sub.1 in FIG. 3 and updating by 5 samples at t.sub.2, buffer memory 22 always holds the proper number of updating input samples at the beginning of each new calculation sequence.

It will be recognized by those skilled in the art that a multistage interpolator or decimator is often advantageously employed rather than a single stage configuration. This is especially true in applications which require large changes in sampling rate, since multistage structures not only generally result in fewer circuit computations, but also result in less severe filtering constraints, i.e., the filter requirements for each stage of a multistage structure are less stringent than the requirements imposed on the single filter of a single stage structure.

FIG. 4 depicts an illustrative example of a multistage embodiment of our invention with n-stages in which L/M<1 (n decimators) are cascaded with n stages in which L/M>1 (n interpolators) to realize a narrow-band FIR filter. That is, FIG. 4 is a narrow-band FIR filter which is comprised of 2n circuits such as the circuit embodiment of FIG. 2 with stages i = 1,2, . . . ,n changing the sampling rate by a factor L.sub.i /M.sub.i <1 and stages i = n+1, n+2, . . . ,2n changing the sampling rate by a factor L.sub.i /M.sub.i >1. Generally, in such an FIR filter, the sampling rate of the output signal at terminal 32 will be identical to the sampling rate of the input signal at terminal 21. Accordingly, in most instances ##EQU3##

As depicted in FIG. 4, the master clock frequency applied to terminal 34-i of the ith decimation stage 13 L.sub.i f.sub.ri where ##EQU4##

The master clock frequency of each decimator stage may, of course, be determined from a single master clock frequency, L.sub.i f.sub.ri, by counting down from the clock frequency applied to stage (i - 1) by the factor M.sub.i. In a like manner, the master clock frequencies of each of the n interpolation stages may be determined from the clock frequency of the preceding interpolator stage by counting up by the appropriate factor M.sub.i.

As depicted in FIG. 4, each decimator-interpolator stage includes buffer memory 22-i, multiplier 29-i, accumulator 31-i, and coefficient memory 41-i. Since each of the n stages of decimation and each of the n stages of interpolation also include circuit sequencing means such as depicted in the embodiment of the single-stage interpolator-decimator of FIG. 2, the circuits effectively operate independently of one another. That is, the (i+1)th stage simultaneously processes the previous data output of the ith stage, while the ith stage processes signal samples of the next processing interval. Such a configuration is commonly denoted as a "pipeline" design.