WO1995017708A2

WO1995017708A2 - Calculation of a scalar product in a direct-type fir filter

Info

Publication number: WO1995017708A2
Application number: PCT/FI1994/000568
Authority: WO
Inventors: Eero Pajarre; Ville Eerola; Tapio SARAMÄKI; Tapani Ritoniemi; Timo Husu; Seppo Ingalsuo
Original assignee: Eero Pajarre; Ville Eerola; Saramaeki Tapio; Tapani Ritoniemi; Timo Husu; Seppo Ingalsuo
Priority date: 1993-12-17
Filing date: 1994-12-16
Publication date: 1995-06-29
Also published as: FI97002B; KR960701409A; EP0685127B1; JPH08506953A; DE69421073T2; FI935709A0; DE69421073D1; US6131105A; FI935709A; EP0685127A1; FI97002C; ATE185458T1; WO1995017708A3

Abstract

The invention relates to a direct-type FIR filter, a method for calculating a scalar product in a FIR filter, and a method for designing a direct-type FIR filter. Successive words of a digital input signal are delayed in a delay line having delays (50A-50D) of the duration of one word, and the scalar product between the variously delayed words derived from the delay line and the corresponding constant coefficients is calculated. In accordance with the invention, calculation of the scalar product comprises: (a) combining the bits of words at the input (X0) and outputs (X1-X4) of the delay line bit by bit in a network of bit-serial subtractor and/or adder elements (51-56) wherein at least one of the bit-serial elements is involved in the multiplication operation of at least two different coefficients; and (b) multiplying (49A-K) the multiplication results from the network by powers of two, and summing together (45-48) the results to yield the scalar product.

Description

Calculation of a scalar product in a direct-type FIR filter

Field of the Invention

The present invention relates to a digital direct- type FIR filter, and in particular to calculation of a scalar product in a direct-type FIR filter.

Background of the Invention

A digital filter is a software or a specially designed electronic circuit processing discrete digital signal samples to perform a desired transfer function operation on said signal. The Z transfer function of a digital, i.e. discrete time, FIR (Finite Impulse Response) filter has the generic form

where H(z) is the transfer function of the filter, Y(z) and X(z) represent the output and input of the filter respectively, a_± represent constant coefficients, i.e. tap coefficients, and z^" represents a delay of i samples. The properties of a FIR filter are solely dependent on the tap coefficients a_x, and thus determination of these coeffici¬ ents is required in order to obtain the desired character¬ istics for the filter. There are several methods for determining the coefficients. The non-recursive discrete time filter in accordance with equation (1) is normally represented as a block diagram as shown in Figures 1 and 2. Figure 1 illustrates a direct-type FIR (Finite Impulse Response) filter and Figure 2 a transposed FIR filter. The filtering function in accordance with equation (1) can be realized by both types of discrete time filter, but the present invention relates to a direct-type FIR filter according to Figure 1. As is apparent from Figures 1 and 2, the discrete time filter is illustrated as a block dia¬ gram wherein square blocks 1 perform delaying of the information by one sample z^"1, triangular blocks 2 repres¬ ent multiplication operations, and circles 3 represent adders. As stated previously, the characteristics of the filter are dependent on the values of the tap coefficients a_±. Prior art direct-type FIR filters exist in which a dis¬ crete multiplier unit for each tap coefficient is em¬ ployed. The drawback of this approach is the large number of multiplier units required, which occupy a considerable area on a semiconductor chip when embodied as an integ¬ rated circuit and are therefore costly. A prior art approach is known in which the tap coefficients are simple sums of powers of two, i.e. the coefficients are limited to the form 2^"a+2^"+2^"c. Such an approach is attended by the drawback of limitations in the possible coefficients to be realized. These limitations can substantially complicate the realization of the desired signal processing function H(z). Still another prior art solution entails the use of a fast multiplier and memory for realizing the filter. Such a solution is illustrated in Figure 3, wherein the necessary delays z^-1 are generated by buffering the values of the input signal X(z) into a RAM memory 41 prior to in- putting them to a multiplier 42, in which they are multi¬ plied by coefficients a_i derived from a ROM memory. There¬ after the multiplication results are supplied to an adder 44 wherein they are summed with the filter output Y(z). The drawback of such a solution is the chip area occupied by the fast multiplier unit 42. Further drawbacks include the high power consumption of the multiplier unit 42 and, in certain applications, the electromagnetic interference produced thereby in other circuitry. Furthermore, on account of the limited speed of the multiplier unit, only a limited number of coefficients a_A can be realized with one multiplier unit. Complex structures require several multiplier units and a complex control logic.

Disclosure of the Invention

The object of the present invention is a direct- type digital FIR filter that can be embodied as an integ¬ rated circuit with several coefficients so as to occupy substantially less chip area in integrated circuit con- figuration than the filters implemented by the prior art techniques.

Another object of the present invention is a digi¬ tal filter suitable for comparatively high clock fre¬ quencies. A further object of the present invention is a digital filter enabling realization of arbitrary coef¬ ficients automatically.

These and other objects and advantages of the invention are achieved with a method for calculating a scalar product in a direct-type digital FIR filter, the method comprising delaying successive words of a digital input signal in a delay line having delays of the duration of one word, and calculating the scalar product between the variously delayed words derived from the delay line and the corresponding coefficients, the method being characterized in that the calcula¬ tion step comprises combining the bits of words at the input and out- puts of the delay line bit by bit in a network of bit- serial subtractor and/or adder elements wherein at least one of said bit-serial subtractor and/or adder elements is involved in the multiplication operation of at least two different coefficients, multiplying the multiplication results from the network of bit-serial subtractor and/or adder elements by powers of two, and summing together the results to yield said scalar product.

Another aspect of the invention is a direct-type digital FIR filter, comprising a delay line having an input for receiving digital words in serial form, a plurality of one-word delays and an output after each delay; calculation means for calculating the scalar prod- uct between the words at the input and each output of the delay line and the corresponding coefficients; an output to which the calculated scalar product is applied, the filter being characterized in that said calculation means comprise a plurality of bit-serial subtractor and/or adder elements for combining bits of words at the input and each output of the delay line, said bit-serial subtractor and/ or adder elements forming a network wherein at least one of said bit-serial subtractor and/or adder elements is involved in the multiplication operation of at least two different coefficients, means for multiplying the multiplication results from the network of bit-serial subtractor and/or adder elements by powers of two and summing together the results to yield said scalar product.

One aspect of the invention is a method for design¬ ing a direct-type digital filter, comprising a step of determining the coefficients required in the filter. This method is characterized in accordance with the invention by further steps of designing a network of bit-serial subtractor and/or adder elements for the filter, wherein the number of said bit-serial elements is minimized taking into account per- formance criteria of the filter, so that a maximum number of said bit-serial elements are involved in the multi¬ plication operation of more than one different coeffici¬ ents, designing an output register performing multiplica- tion by a power of two and summing together the results from said network, said register comprising one-bit delay elements and bit-serial adder and subtractor elements.

In the present invention, the scalar product is calculated by combining values derived from the delay line in bit-serial adder and/or subtractor elements, so that at least some of the adder and/or subtractor elements are used to provide the mantissa of more than one coefficient. In other words, the "partial sum" or "partial difference" outputted by a specific adder or subtractor element can be used on the next level of the adder and/or subtractor ele¬ ment network to produce the mantissas of several coeffici¬ ents simultaneously. Furthermore, in the arrangement of the invention the combined use of adder and subtractor elements in producing the coefficients enables the number of calculation elements (+/- operators) to be minimized. The products given by the network of calculation elements are multiplied by the exponent of the respective coef¬ ficient and summed together to produce the final scalar product. The arrangement of the invention affords good round-off and truncation behaviour. The scalar product is rounded off or truncated only once, and thus the error in the scalar product is, on an average, only 1/2 least sig¬ nificant bits. Multiplication by a power of two and sum¬ mation are preferably performed on all coefficients in the same output register, which is comprised of one-bit delays and bit-serial adder and subtractor elements. Thus the number of necessary delays can be optimized as compared with a case where each coefficient has dedicated calcula¬ tion elements, delay elements, for multiplication by a power of two.

By means of the invention, the network of bit- serial adder and subtractor elements can be optimized by finding the sum and/or difference of powers of two for the coefficients required, so as to considerably diminish the requisite number of calculation elements in comparison with the prior art solutions. At the same time, the requi¬ site number of series-connected elements is characteris¬ tically diminished. With the construction of the inven¬ tion, arbitrary coefficients can be realized. Still an- other advantage of the invention is a low number of logic levels, and thus the maximum operating frequency is very high. When the invention is realized as an integrated cir¬ cuit, the silicon area occupation required is less than half the area required by the digital filter shown in Fig- ure 3 which includes a multiplier and RAM and ROM memories.

Brief Description of the Drawings

In the following the invention will be set forth by means of preferred embodiments with reference to the accompanying drawing, wherein

Figure 1 is a block diagram of a direct-type digi¬ tal FIR filter, Figure 2 is a block diagram of a transposed digital

FIR filter.

Figure 3 is a block diagram of a prior art digital filter implemented by a fast multiplier and memories.

Figure 4 is a block diagram of a digital filter of the invention with five coefficients. Figure 5 is a block diagram of an embodiment of the digital filter of Figure 4, and

Figure 6 is a block diagram of a bit-serial adder element.

Preferred embodiments of the Invention

Reference will now be made to Figure 4, showing a five-coefficient direct-type digital FIR filter imple- mented in accordance with the invention. After the input IN, the filter comprises a delay line having four delay blocks 50A, 50B, 50C and 50D, each having a length of one word (z^"1) . Each delay block is made up of N one-bit delay elements 49, as in delay block 50D in Figure 5. N is the word width of the filter. The number sequence X obtained at the input IN is supplied to delay line 50A-50D in bit- serial form, each value as N successive bits, the least significant bit LSB occurring first. The words are clocked through the delay line in such a way that each word is shifted one bit at each clock cycle. In other words, during N successive clock cycles each bit of word appears in turn at the one-bit output of the respective delay block. Thus the delay line 50A-50D (the input included) provides an output of five variously delayed values x₀, x₁, x₂, x₃ and x₄ in bit-serial form. The values thus obtained should be multiplied by the corresponding tap coefficients a₀, a_x, a₂, a₃ and a₄, whereafter the products obtained are summed together to give the desired scalar product, as in the schematic block diagram of a direct-type FIR filter of Figure 1.

A network of combining elements 51, 52, 53, 54, 55 and 56 is coupled to the outputs of the delay line for multiplication of the words x0-x4 by tap coefficients a₁-a₄ by combining the one-bit outputs of the delay line. The network of combining elements comprises bit-serial adder and subtractor elements 51, 52, 53, 54, 55 and 56 employ¬ ing bit-serial arithmetic on several levels. In accordance with the basic concept of the invention, it has been sought to minimize the number of calculation elements taking into account certain performance criteria for the filter, so that the same calculation elements are employed to produce more than one different tap coefficients. In the exemplary case of Figure 4, five coefficients are formed by employing only six bit-serial (one-bit) arith- etic elements. For instance, the inputs of the adder ele¬ ment 53 are provided by the outputs x₂ and x₃ of the delay line. On the other hand, the output value x₂ + x₃ of the adder element 53 provides an input for adder elements 52 and 54 on the next network level. The calculation elements 51, 52, 54, 55 and 56 respectively provide five outputs, which are then multiplied by the requisite power of two in multiplier units 57A, 57B, 57C, 57D and 57E. The outputs of multiplier units 57A-57E are summed together in an adder 58 to obtain the desired scalar product

^y = 1∑=0 "A ⁽²⁾

Delaying of a bit-serial value for the duration of one bit corresponds to multiplying the value by two. Therefore, the multiplier units 57A-57E have in one embodiment of the invention been configured by using one- bit delay elements, one delay element for each power of two. Thus for example multiplier unit 57A has nine one-bit delays and multiplier unit 57D three one-bit delays. Hence the embodiment of Figure 4 requires a total of 23 one-bit delay elements to implement the multiplier units 57A-57D.

Figure 5 shows the preferred embodiment of the invention, in which elements and functions similar to those in Figure 4 are denoted by the same reference numerals and symbols. In the embodiment of Figure 5, the delay line 50A-50D and the network of calculation elements 51-56 are identical with those of the embodiment of Figure 4. On the other hand, the multiplier units 57A-57D and adder 58 of Figure 4 are replaced by a common output register in Figure 5. The output register comprises a series connection of the following elements in the given order: three one-bit delay elements 49A, 49B and 49C, an adder 45, a one-bit delay element 49D, an adder 46, two one-bit delay elements 49E and 49F, a subtractor 47, three one-bit delay elements 49G, 49H and 49K, and an adder 48. In the network of calculation elements 51-56, the output of element 51 is coupled to the input of delay element 49A, the output of element 52 is coupled to adder 45 together with the output of delay element 49C, the output of element 54 is coupled to adder 46 together with the output of delay element 49D, the output of element 55 is coupled to subtractor 47 together with the output of delay element 49F, and the output of element 56 is coupled to adder 48 together with the output of delay element 49K. The output of adder 48 provides the output OUT for the filter. The calculation elements 45, 46, 47 and 48 are bit-serial calculation elements. The embodiment of Figure 5 employs common delays for different coefficients, thus enabling further reduction in the number of elements required for the filter. In the exemplary case, the output register comprises nine delay elements and four calcula- tion elements, while the corresponding operation in the embodiment of Figure 4 requires 23 delay elements and one calculation element.

The operation of the filter of the invention will be described hereinbelow with reference to the embodiment of Figure 5. All calculation elements 51-56, 45-48 and delay elements 49A-49K of the output register are reset to zero at the start of the calculation. The advantages of the filter in accordance with the invention include the fact that all these elements can be reset simultaneously, and the resetting can be performed with a single control signal. Thereafter a new value xO is clocked to the delay line 50A-50D, and thus the values x_x-x₄ clocked to the delay line during the previous calculation cycles are shifted one delay block forward. During the next N clock cycles, the values x₀, x_x, x₂, x₃ and x₄ with mutually different delays are thus derived bit by bit in bit-serial form as outputs from the delay line. The bit-serial values x₀-x₄ derived from the delay lines are applied to the bit- serial calculation network 51-56, the outputs of which yield bit-serial values x₀-x₂, x₁+x₂+x₃, x₀+x₂+x₃+x₄, x₀+x₄ and x₃-x₄. The values thus obtained from the calculation net¬ work 51-56 are applied in bit-serial form to the output register that combines them, simultaneously delaying them for various periods of time. As stated previously, delay- ing a bit-serial value for one bit corresponds to multi¬ plying the value by two. Hence in the embodiment of Figure 5 the output register delays the input values by 9, 6, 5, 3 and 0 bits, which corresponds to multiplying the values by 512, 64, 32, 8 and 1 respectively. Thus the value of the output OUT of the output register gives the value

512(x₀-x₂)+64(x₁+x₂+x₃)+32(x₀+x₂+x₃+x₄)-8(x₀+x₄)+(x₃-x₄)

= 536x₀+64x₁-416x₂+97x₃+23x₄=r (3)

After the entire word to be loaded in has been clocked bit by bit to the delay line 50A-50D and the cal¬ culation network 51-56 during N clock cycles, there are still numbers significant for the calculation in the out¬ put register and in the values stored in the calculation network 51-56. In order that the final product may be outputted from the filter, clocking of the delay line 50A- 50D and the calculation network 51-56 is continued for further ten clock cycles, whereupon the entire product is derived. When calculation is performed in accordance with the foregoing, the word width of the product in the filter increases considerably. If it is not desired to take advantage of this, n first bits of the product clocked from the adder 48 to the output OUT are omitted. This cor- responds to dividing the product by 2ⁿ. If for example in the embodiment of Figure 5 ten first output bits are omitted at the output OUT, the filter has the calculated function

512(x₀-x₂)+64(x₁+x₂+x₃)+32(x₀+x₂+x₃+x₄)-8(x₀-l-x₄)-l-(X₃-x₄)

1024 = | 0.5234375x₀+0.0625x₁-0.40625x₂+0.0947265625x₃+0.0224609375x₄ |

( 4 )

Figure 6 shows a block diagram of a bit-serial adder element. The adder element comprises a one-bit delay block 61, which in this exemplary case is implemented by a D flip-flop, and an adder 62 adding up two data bits and outputting a sum and a carry bit c_out. All signals shown in Figure 6 are one-bit signals, i.e. each of them can be implemented by a single signal line. The adder element shown in Figure 6 operates in the following way. The values to be summed together are applied to the adder 62 in serial form, the least significant bit (LSB) being the first. The adding up of two bits a and b and a carry bit c_ln gives as a result one sum bit sum and a carry bit c_out which is stored in the delay block 61 for summing together the next bits. The delay block 61 is reset between the addition of two successive N-bit values by way of the reset line.

The bit-serial subtractor element can be embodied similarly. The only difference is that instead of an adder

62, a subtractor is employed. Furthermore, the delay block 61 is set to the value 1 between the subtraction of two successive N-bit values.

The figures and the description relating to them are only intended to illustrate the present invention. In their details, the methods and filter of the invention can vary within the purview of the accompanying claims.

Claims

Claims :

1. A method for calculating a scalar product in a direct-type digital FIR filter, the method comprising delaying successive words of a digital input signal in a delay line having delays of the duration of one word, and calculating the scalar product between the variously delayed words derived from the delay line and the corresponding coefficients, c h a r a c t e r i z e d in that the calculation step comprises combining the bits of words at the input and out¬ puts of the delay line bit by bit in a network of bit- serial subtractor and/or adder elements wherein at least one of said bit-serial subtractor and/or adder elements is involved in the multiplication operation of at least two different coefficients, multiplying the multiplication results from the network of bit-serial subtractor and/or adder elements by powers of two, and summing together the results to yield said scalar product.

2. A direct-type digital FIR filter, comprising a delay line having an input (X0) for receiving digital words in serial form, a plurality of one-word delays (50A-50D) and an output (XI, X2, X3, X4) after each delay; calculation means for calculating the scalar product between the words at the input (X0) and each output (XI, X2, X3, X4) of the delay line and the cor- responding coefficients; an output (OUT) to which the calculated scalar product is applied, c h a r a c t e r i z e d in that said calculation means comprise a plurality of bit-serial subtractor and/or adder elements (51-56) for combining bits of words at the input (X0) and each output (XI, X2, X3, X4) of the delay line, said bit-serial subtractor and/or adder elements forming a network wherein at least one of said bit-serial subtractor and/or adder elements is involved in the multiplication operation of at least two different coefficients, means (57A-D, 58; 49A-49K, 45-47) for multiplying the multiplication results from the network of bit-serial subtractor and/or adder elements by powers of two and summing together the results to yield said scalar product.

3. A FIR filter as claimed in claim 2, c h a r ¬ a c t e r i z e d in that said means for multiplying by powers of two comprise an output register comprised of one-bit delay elements (49A-K) and bit-serial adder and subtractor elements (45-48), one one-bit delay element for each power of two and one bit-serial adder or subtractor element for each coefficient.

4. A FIR filter as claimed in claim 2, c h a r ¬ a c t e r i z e d in that each delay (50A-D) in the delay line comprises n one-bit delay elements (49), wherein n is the word width in bits.

5. A method for designing a direct-type digital filter, comprising a step of determining the coefficients required in the filter, c h a r a c t e r i z e d by further steps of designing a network of bit-serial subtractor and/or adder elements for the filter, wherein the number of said bit-serial elements is minimized taking into account per- formance criteria of the filter, so that a maximum number of said bit-serial elements are involved in the multi¬ plication operation of more than one different coefficients, designing an output register performing multiplica¬ tion by a power of two and summing together the results from said network, said register comprising one-bit delay elements and bit-serial adder and subtractor elements.