The present invention relates to minimized area implementation of a sum of products expression.
The implementation of a Sum of Products (SOP) expression can be used to realize signalprocessing applications on hardware. A SOP implementation is hereby discussed with reference to a hardware realization of a digital filter. A digital filter is an electronic circuit processing discrete signal samples to perform a desired transfer function operation on the discrete signal samples. The digital filter is a Finite Impulse Response (FIR) or Infinite Impulse Response (IIR) filter.
The following components are used in the digital filter implementation.
Unit sample delay (11): A unit sample delay (Z^{−1}) is shown in FIG. 1(a). A unit sample delay is a Parallel In Parallel Out (PIPO) register as it captures the sample at its input on the next positive clock edge.
Multiplier (12): Takes in an input ‘x’ and gives a multiplied output ‘y’=‘a*x’, where ‘a’ is the multiplicand. Its symbol is a triangle with a multiplicand by its side as shown in FIG. 1(b).
‘P1’ bit adders (13): Takes in two inputs each of width ‘p1’ bits and gives an added output of ‘p1’ bits. Hence this adder is called the ‘p1’ bit parallel adder and its symbol is ‘+’ in a circle as shown in FIG. 1(b).
‘P1’ bit subtractor (15): Takes in two inputs each of width ‘p1’ bits and subtracts one output from the other and gives an output of ‘p1’ bits. Hence this subtractor is called as the ‘p1’ bit parallel subtractor and its symbol is shown in a circle with ‘+’ in it and ‘−’ on one of the inputs as shown in FIG. 1(b).
Parallel shifter (14): Takes in an input ‘x’ and gives a ‘z’ bit left shifted output. Suppose ‘x=101’ and ‘z=2’ then ‘y=10100’, then its symbol is a line with ‘<<z’ on it (where ‘z’ can be any positive number) as shown in FIG. 1(b).
The working of unit sample delay is evident from the timing diagram shown in FIG. 1(c). The samples X0, X1, X2 . . . X7 at the input of unit sample delay in FIG. 1(c) appear as delayed by one Clock (Clk) period as they are captured by the next positive edge of the Clk. The transfer function H (Z) of a FIR filter has the generic form as given in equation (2a);
where b_{K }represents the coefficients, and z^{−k }represents delay of k clock cycles.
The equation of a FIR filter is given below:
where y(n) is the output and x(n−K) is the delayed input.
The FIR filter has different types of implementations in hardware. Two of the important implementations of the FIR filter are the direct form of coefficient bank implementation as shown in FIG. 2(a) and the transposed form of implementation as shown in FIG. 2(b). The direct form of the FIR filter has an input X connected to the first unit sample delay. Unit sample delays (11) are connected serially to each other. Taps S_{0 }to S_{N }obtained from unit sample delays (11) are connected to respective coefficient multipliers (12) b_{0 }to b_{N}. The output of the multipliers (12) is added through a plurality of adders (13,15) to form the output Y of the filter.
The transposed form of the FIR filter (2b) has the input X connected to the plurality of coefficient multipliers b_{0 }to b_{N }(12) resulting in taps S_{0 }to S_{N}. The taps S_{0 }to S_{N }are connected to the plurality of unit sample delays (11) and adders (13,15) to form the output Y of the filter.
Given below is the equation of the IIR filter:
where H (Z) is the transfer function, b_{m }and a_{1 }are the coefficients and z^{−k }represent a delay of k clock cycles.
The equation of the IIR filter can also be written as in equation (2d);
where y(n) is the output, y(n−1) is the delayed output and x(n−m) is the delayed input.
The implementation of a coefficient is the implementation of a number in hardware. Hence the term coefficient and number are used interchangeably in the specification.
The structure of a direct and transposed form IIR filter is shown in FIG. 2(c). The direct form structure has a plurality of unit sample delays (11), and two sets of direct form coefficient banks (1) formed by plurality of adders (13,15) and plurality of the multipliers (12) connected to the taps. The transposed form of the IIR filter has plurality of unit sample delays (11), and two sets of transposed form coefficient banks (1) formed by a plurality of adders (13,15) and plurality of the multipliers (12) connected to the taps as shown in FIG. 2(c). The parallel direct or transposed form of coefficient bank is the area consuming part in the parallel digital filter implementations. An existing way of implementing the parallel direct or transposed form of coefficient bank using the shift (12) and add (13,15) has been described in FIG. 2(b). A left shift is ‘0’ cost in parallel implementations as it needs only appending lines set at ‘0’ on the left of the parallel bus. The remaining part of shift (12) and add (13,15) implementation comprises adders. The implementation of parallel multipliers using shift and add is shown in FIG. 2(d).
An existing method for implementing the direct form of coefficient bank is described as follows. The existing method of implementation of direct form of coefficient bank has two parts, part 1 is an algorithm shown in FIG. 2(e), which takes coefficients taps as input and gives the equation of direct form of coefficient bank and horizontal subexpressions as the output, and part 2 is the generalized structure shown in FIG. 2(g), which is obtained by mapping the equation of the direct form of coefficient bank, and horizontal subexpressions onto hardware.
The method is explained by using coefficient bank implementation for a given example.
Assuming that the example coefficients of an FIR filter are given by, Set 1={−79, 1044, −5890, 27916, 49362, −8382, 1628, −154, 63, 126}. The equation form of the direct form of coefficient bank for the example is as given below:
Y=[−79]S0+[1044]S1+[−5890]S2+[27916]S3+[49362]S4+[−8382]S5+[1628]S6+[−154]S7+[S3]S8+[126]S9 (2e)
FIG. 2(f1) shows the direct form of implementation of the example FIR filter. Structure (1) is the direct form of coefficient bank, that comprises unit sample delays (11), multipliers and adders (12,13,15). The execution of the existing method (FIG. 2(e)) is explained with reference the coefficients of Set 1.
Step 1: This step illustrates the conversion of coefficients into Canonical Signed Digit (CSD) representation. Here the value of ‘i’ is −1.
−79=i0i0001; 1044=10000010100; −5890=i01001000000i0; 27916=100i0i0100010i00 49362=10i0000010i010010; −8382=i0000i01000010; 1628=10i010i00i00; −154=i0i010i0; 63=100000i; 126=100000i0
The direct form of coefficient bank with CSD representation of coefficients is as given below:
Y=(2ˆ0−2ˆ4−2ˆ6)S0+(2ˆ2+2ˆ4+2ˆ10)S1+(−2ˆ1+2ˆ8+2ˆ11−2ˆ13)S2+(−2ˆ2+2ˆ4+2ˆ8−2ˆ10−2ˆ12+2ˆ15)S3+(2ˆ1+2ˆ4−2ˆ6+2ˆ8−2ˆ14+2ˆ16)S4+(2ˆ1+2ˆ6−2ˆ8−2ˆ13)S5+(−2ˆ2−2ˆ5+2ˆ7−2ˆ9+2ˆ11)S6+(−2ˆ1+2ˆ3−2ˆ5−2ˆ7)S7+(−2ˆ0+2ˆ6)S8+(−261+2ˆ7)S9 (2f1)
or
Y=S0−S8+2(−S2+S4+S5−S7−S9+2(S1−S3−S6+2(S7+2(−S0+S1+S3+S4+2(−S6−S7+2(−S0−S4+S5+S8+2(S6−S7+S9+2(S2+S3+S4−S5+2(−S6+2(S1−S3+2(S2+S6+2(−S3+2(−S2−S5+2(−S4+2(S3+2(S4))))))))))))))) (2f2)
The number of adders required to implement equation (2f2) are 38. Hereafter, Adders/Subtractors are referred as adders for the ease of explanation.
Step 2: Horizontal subexpressions are formed from the CSD representation of coefficients. Table 1 shows the horizontal optimization for the example. From Table1, it is seen that subexpression in row 2ˆ0 is −(S0S8), and in row 2ˆ6 it is (S0S8). Since S0S8 is
TABLE 1 
Showing Horizontal Optimization 


present in two ‘rows’, it is called as horizontal subexpression. 
Similarly S1−S3, −S2+S5, S3+S4 are other horizontal subexpressions. The equation of the direct form of coefficient bank of the example after horizontal optimization is:
Y=X1+2(X3+S4−S7−S9+2(X2−S6+2(S7+2(−S0+S1+X4+2(−S6−S7+2(−X1−S4+S5+2(S6−S7+S9+2(−X3+X4+2(−S6+2(X2+2(S2+S6+2(−S3+2(−S2−S5+2(S4+2(S3+2(S4)))))))))))))))) (2g).
The equation can also be written as equation 2(i) for showing hardware mapping.
Y=X1+2(E1+2(E2+2(S7+2(E3+2(−E4+2(E5+2(E6+2(E7+2(−S6+2(X2+2(E8+2(−S3+2(−E9+2(−S4+2(S3+2(S4)))))))))))))))) (2i),
where the expressions E1=X3+S4−S7−S9, E2=X2−S6, E3=−S0+S1+X4, E4=S6+S7, E5=−X1−S4+S5, E6=S6−S7+S9, E7=−X3+X4, E8=S2+S6, E9=S2+S5, and horizontal subexpressions X1=S0−S8, X2=S1−S3, X3=−S2+S5, X4=S3+S4. The number of adders required to implement the equation (2i) is 34.
The structure of the direct form of coefficient bank from the existing method of implementations for the above example is shown in FIG. 2 (f2). In FIG. 2(f2) substructures SS [X1 to X4] show the hardware mapping of horizontal subexpressions X1 to X4, substructure SS [E1 to E9] are the hardware mapping of the expressions E1 to Ep and substructure SS [Y] is the hardware mapping of the equation Y. The generalized structure of the direct form of coefficient bank from the existing method of implementations is shown in FIG. 2 (g), where the substructures SS [X1 to Xm] show the hardware mapping of horizontal subexpressions X1 to Xm, substructures SS [E1 to Ep] are the hardware mapping of expressions and E1 to Ep and SS [Y] is the hardware mapping of the equation.
FIG. 2(i1) shows the transposed form of implementation of the example FIR filter. Structure (2) is the transposed form of coefficient bank implementation. Individual equations for the transposed form implementation of the example coefficient bank are:
S0=−79X, S1=1044X, S2=−5890X, S3=27916X, S4=49362X, S5=−8382X, S6=1628X, S7=−154X, S8=63X, S9=126X.
The execution of the steps in the existing method (FIG. 2(h)) are explained below.
Step 1: Conversion of coefficients into
Canonical Signed Digit (CSD) or any other representation. The conversion of individual coefficients to CSD with reference to Set 1 is as given below:
S0=−79X=(i0i0001)X
S1=1044X=(10000010100)X
S2=−5890X=(i01001000000i0)X
S3=27916X=(100i0i0100010i00)X
S4=49362X=(10i0000010i010010)X
S5=−8382X=(i0000i01000010)X
S6=1628X=(10i010i00i00)X
S7=−154X=(i0i010i0)X
S8=63X=(100000i)X
S9=126X=(100000i0)X
Step 2: From the vertical subexpression, form the CSD representation of coefficients. Table 2 shows the vertical subexpressions for the above stated example. In Table 2, the expression (2ˆ−2ˆ6) is present in column S0 and in column S8 as −(2ˆ0−2ˆ6). The expression is also present in columns S3 and S9, as (−2ˆ2)(2ˆ0−2ˆ6) and (−2ˆ1)(2ˆ0−2ˆ6) respectively. As the expression (2ˆ0−2ˆ6) is present in different columns, it is called vertical subexpression. Similarly (1−(2ˆ2)) and (1+(2ˆ3)) are other vertical subexpressions.
TABLE 2 
Showing Vertical Optimization 


Equations of the coefficients after the vertical optimization:
S0=−79X=(Y1−(2ˆ4))X
S1=1044X=((1+(2ˆ2)+(2ˆ8))*(2ˆ2))X
S2=−5890X=((−1+(2ˆ7)+(2ˆ10)*(Y2))*(2ˆ1))X
S3=27916X=((−Y1+(2ˆ2)−(2ˆ8)−(2ˆ10)+(2ˆ13))*(2ˆ2))X
S4=49362X=((Y3−(2ˆ5)*(Y2)+(2ˆ13)*(Y2))*(2ˆ2))X
S5=−8382X=((1+(2ˆ5)*(Y2)−(2ˆ12))*(2ˆ1))X
S6=1628X=((−Y3+(2ˆ5)*Y2+2ˆ9)*(2ˆ2))X
S7=−154X=((−1+(2ˆ2)*Y2−(2ˆ6))*(2ˆ1))X
S8=63X=(Y1)X
S9=126X=(Y1*(2ˆ1))X,
where,
Y1=1−(2ˆ6), Y2=1−(2ˆ2), Y3=1+(2ˆ3) are vertical subexpressions.
It is observed from the above analyses that the number of adders required to implement the transposed form coefficient bank is 20. For the example the vertical subexpressions are mapped into hardware as substructures SS (Y1, Y2, Y3) as shown in FIG. 2(i2). The individual equations S0 to S9 of the transposed form of coefficient bank are mapped into hardware as substructures SS (S0 to S9) as shown in FIG. 2(i2). The substructures are formed from the plurality of adders 13,15 and shifter 14.
The generalized structure of the transposed form of the coefficient bank (2) from the existing method of implementation is shown in FIG. 2 (j). It is formed with substructures SS (Y1 to Ym) formed from mapping the vertical subexpressions Y1 to Ym into hardware and substructures SS (S0 to Sn) formed from mapping the individual equations S0 to Sn into hardware. The substructures are formed from the plurality of adders (13,15) and shifter (14). It is observed that the existing structure of the coefficient bank is inefficient, when high magnitude coefficients are required to be implemented.
The constraints in the existing method for SOP expression implementation are further illustrated by implementing a number with CSD representation. The CSD representation of 45=10i0i01 and structure of the CSD implementation of the number (45) is shown in FIG. 2(k1). It is observed that the number of adders required to implement 45 in CSD based implementation as in FIG. 2(k1) is 3.
In CSD implementation the input is shifted by a value of (2ˆz) (where ‘z’ is a positive number) and then added or subtracted from the other shifts of the input. Another existing way of implementation of a 45 in hardware is 9*5 and is shown in FIG. 2(k2). The number of adders required for implementing 45 as in FIG. 2(k2) is 2. The input x is multiplied by number 9 and then the result 9x is multiplied with 5 to get an output of 45x.
It is evident from the implementation of the coefficient multiplier 45 that there can be an area efficient way of implementing coefficients, so as to reduce the number of adder/subtractors in the resultant hardware. Thus, a requirement is felt for an efficient method to achieve a reduced or minimal area implementation of a Sum of Products expression.
It is an object of the present invention to provide a reduced or minimal area integrated circuit implementation of a sumofproducts expression.
It is another object of the present invention to implement a direct form of coefficient bank using the sumofproducts expression.
It is yet another object of the present invention to implement the transposed form of coefficient bank using the sumofproducts expression.
To achieve the objectives the present invention provides a device for implementing a sumofproducts expression comprising:
Further, the present invention provides a method for creating an implementation of a sumofproducts expression, comprising:
The invention will now be described with reference to the accompanying drawings.
FIG. 1(a) illustrates the existing structure of unit sample delay as in the prior art.
FIG. 1(b) illustrates the existing structures of ‘p1’ bit adder, ‘p1’ bit saturator, multiplier and a ‘z’ bit left shifter as in the prior art.
FIG. 1(c) illustrates timing diagram to explain working of unit sample delay as in the prior art.
FIG. 2(a) illustrates the existing structure of direct form FIR filter as in the prior art.
FIG. 2(b) illustrates the existing structure of the transposed form FIR filter as in the prior art.
FIG. 2(c) illustrates the existing structure of direct and transposed form IIR filter as in the prior art.
FIG. 2(d) illustrates the generalized implementation of sum of products expression using shifters and adders as in the prior art.
FIG. 2(e) illustrates the existing method for implementation of direct form of coefficient bank as in the prior art.
FIG. 2(f1) illustrates the existing direct form structure of FIR filter for a given example as in the prior art.
FIG. 2(f2) illustrates the existing structure of direct form of coefficient bank with Canonical Signed Digit (CSD) representation of coefficients for a given example as in the prior art.
FIG. 2(g) illustrates the existing method for implementation of direct form of coefficient bank as in the prior art.
FIG. 2(h) illustrates the existing method for implementation of transposed form of coefficient bank as in the prior art.
FIG. 2(i1) illustrates the existing transposed form structure of FIR filter for a given example as in the prior art.
FIG. 2(i2) illustrates the existing structure oftransposed form of coefficient bank for a given example as in the prior art.
FIG. 2(j) illustrates existing generalized structure of transposed form of coefficient bank as in the prior art.
FIG. 2(k1) illustrates the existing structure of a CSD represented single coefficient ‘45’ as in the prior art.
FIG. 2(k2) illustrates another existing structure of a single coefficient ‘45’ as in the prior art.
FIG. 3(a1) illustrates the method for implementation of a sum of products expression in accordance with the instant invention.
FIG. 3(a2) illustrates the method for implementation of direct form of coefficient bank with generalized equations in accordance with the instant invention.
FIG. 3(a2)1 illustrates the convergent 2SAD (2 input Shift and Add) implementation of Subexpression R1 in accordance with the instant invention.
FIG. 3(a2)2 illustrates the 1SAD (1 input Shift and Add) implementation of Subexpression Y1 in accordance with the instant invention.
FIG. 3(a3) illustrates the generalized structure for implementation of minimal area Sum of Products expression to realize the direct form coefficient bank.
FIG. 3(b1) illustrates the structure obtained after optimization 1 for implementation of the direct form of coefficient bank for a given example in accordance with the instant invention.
FIG. 3(b2) illustrates the structure obtained after optimization 2 for implementation of the direct form of coefficient bank for a given example in accordance with the given invention.
FIG. 3(b3) illustrates the structure obtained after expanding the cost 1 coefficients in optimization 2 in accordance with the instant invention.
FIG. 3(b4) illustrates the structure obtained after optimization 3 for implementation of the direct form of coefficient bank for a given example in accordance with the instant invention.
FIG. 3(c1) illustrates the method for implementation of the transposed form of coefficient bank (Using modified sum of products expression) in accordance with the instant invention.
FIG. 3(c2) illustrates the method for realization of the transposed form of coefficient bank with generalized equations in accordance with the instant invention.
FIG. 3(c3) illustrates the generalized structure for implementation of the minimal area Sum of Products expression to realize the transposed form coefficient bank.
FIG. 3(d) illustrates the structure for implementation of the transposed form of coefficient bank for a given example according to the instant invention.
Before going into details of the method for implementation of the minimal area sum and products expression, the terms ‘cost’ and ‘weight’ are introduced as follows:
Cost: It is the hardware cost required in terms of adders to implement a number. Adder, subtractor cost is referred as ‘adder cost’ for ease of explanation. The cost to implement power of 2 (shifting) in parallel implementations is ‘0’. The cost for implementing (2ˆz+1) or (2ˆz−1) (where ‘z’ is a positive number) in parallel implementations is ‘1’.
Weight: Weight of the cost 0's, cost 1's and numbers greater than cost 1's in a number set is fixed as follows: Weight of cost 0 (power of 2 coefficients)=1, weight of cost 1's ((2 z)+1) or (2ˆz)−1))=2 and weight of numbers greater than cost1's=100 (Assumed for the ease of implementation). The weight of the powers of two is fixed as ‘1’, because it requires a single adder to connect to the elements in the number set. The weight of Cost 1 coefficients is fixed as ‘2’.
The present invention provides a method for generating a reduced or minimal area structure for a sum of products expression. The method has three optimization steps involved in reduction of the number of adders required to implement the coefficient set and is illustrated as shown in the flow chart in FIG. 3(a1).
Before explaining the step by step execution of the method (FIG. 3(a1)), the instant invention is explained with the help of the optimization steps applied to the input set of coefficients as follows.
The input set of coefficients is a user defined complex SumofProducts expression or an equation represented by a high magnitude coefficient or a number, wherein the input subexpression in the instant case is
Y=b_{0}*S0+b_{1*}S1+b_{2}*S2+ . . . +b_{n}*Sn,
where {b_{0}, b_{1}, b_{2 }. . . b_{j}, b_{k}, b_{n}} is the predefined (user input) coefficients, and S0=X, S1=Z^{−1}X, S2=Z^{−2}X . . . Sn=Z^{−n}X. (Here X is a variable and Z^{−n }represents unit sample delay)
Optimization 1
This optimization is a recursive optimization (FIG. 3 (a2)) that generates the modified quotient and remainder sets in the SOP expression after optimizing the adder cost.
Steps 1, 2, 3 of optimization 1 continue till all the numbers in quotient and remainder set are optimized to Cost1s or powers of 2. Cost 1 coefficients are considered as common factors. Optimization is recursively applied so as to generate a minimal area output expression and is hereafter also referred as Recursive Optimization. Given below is the required form of the optimized SumofProducts expression.
Y=cf_{11}{cf_{12}{ . . . {cf_{1m}{QO_{1m}}+RE_{1m}}+cf_{21}{cf_{22}{ . . . {cf_{2m}{QO_{2m}}+RE_{2m}}+cf_{31}{cf_{32 }. . . } . . . }+RE (Similar type of equation as QO),
where cf_{11}, cf_{12}, . . . cf_{1m},cf_{21},cf_{22}, . . . cf_{2m}. are common factors, and QO_{1m},QO_{2m }. . . are quotient expressions, and RE_{1m},RE_{2m }. . . are remainder expressions,
Generalized QO_{1m},QO_{2m }. . . and . . . ,RE_{1m},RE_{2m}. are given as follows:
QO_{1m}=(Power of 2)*D_{q1m0}*S0+(Power of 2)*D_{q1m1}*S1+ . . . +(Power of 2)*D_{q1mn}*Sn+(Power of 2)*R1+ . . . +(Power of 2)*Ro . . .
QO_{2m}=(Power of 2)*D_{q2m0}*S0+(Power of 2)*D_{q2m1}*S1+ . . . +(Power of 2)*D_{q2mn}*Sn+(Power of 2)*R1+ . . . +(Power of 2)*Ro . . .
RE_{1m}=(Power of 2)*D_{r1m0}*S0+(Power of 2)*D_{r1mn}*S1+ . . . +(Power of 2)*D_{r1mn}*Sn+(Power of 2)*R1+ . . . +(Power of 2)*Ro . . .
RE_{2m}=(Power of 2)*D_{r2m0}*S0+(Power of 2)*D_{r2m1}*S1+ . . . +(Power of 2)*D_{r2mn}*Sn+(Power of 2)*R1+ . . . +(Power of 2)*Ro,
where D_{q1m0 }to D_{r2mn }etc., are cost 1's,
where R1 to Ro are recursive subexpressions of form R1=(Power of 2)*S0+ . . . (Power of 2)*Sn . . . . Ro=(Power of 2)*S0+ . . . (Power of 2)*Sn,
where (Power of 2) is any power of 2.
Optimization 2
The expressions QO_{1m }to RE_{2m }obtained by applying the method of the optimization 1 are tabulated in the form of Table A as follows in order to factorize vertically. Further, the optimization 2 is hereafter also referred to as Vertical Optimization.
TABLE A 

Traversing the table column wise and assuming that D_{q1m1}=D_{q2m1}, then the first vertical subexpression is given by Y1=D_{q1m1}*S1. Similarly there can be ‘p’ vertical subexpressions, where ‘p’ is any positive number.
Generalized equation after optimization 2 is given by:
Y=cf_{11}{cf_{12}{ . . . {cf_{1m}{QO_{1m}}+RE_{1m}}+cf_{21}{cf_{22}{ . . . {cf_{2m}{QO_{2m}}+RE_{2m}}+cf_{31}{cf_{32}. . . } . . . }+RE;
where cf_{11}, cf_{12}, . . . cf_{1m},cf_{21},cf_{22}, . . . cf_{2m }. . . etc., are common factors, and QO_{1m},QO_{2m }. . . are quotient expressions, and RE_{1m},RE_{2m }. . . are remainder expressions,
where QO_{1m}=(Power of 2)*D_{q1m0}*S0+ . . . +(Power of 2)*D_{q1mn}*Sn+(Power of 2)*R1+ . . . +(Power of 2)*Ro+(Power of 2)*Y1 . . . +(Power of 2)Yp.
RE_{1m}=(Power of 2)*D_{r1m0}*S0+ . . . +(Power of 2)*D_{r1mn}*Sn+(Power of 2)*R1+ . . . +(Power of 2)*Ro+(Power of 2)*Y1 . . . . +(Power of 2)*Yp;
where R1 to Ro are recursive subexpressions and Y1 to Yp are vertical subexpressions, Y1=D_{qmm1}*S1, Yn=D_{qnmn}*S_{n}.
Optimization 3
The input coefficient sets are subjected to a horizontal factorization here and the optimization is hereafter also referred to as a horizontal optimization. Expanding Cost 1s in QO_{nm }to RE_{nm }(nth quotient and remainder expressions), and in the vertical subexpressions Y1 to Yp to (2ˆz+1) or (2ˆz−1), where ‘z’ is any positive number.
Assuming that cost 1's in the expressions QO_{1m }to RE_{1m }have been expanded to D_{q1m0}=(2ˆA0)+(2ˆB0), D_{q1mn}=(2ˆAn)+(2ˆBn), D _{rm10}=(2ˆA0)+(2ˆB1), D_{r1mn}=(2ˆAn)+(2ˆBk). Substituting the expanded cost 1's in QO_{nm }and RE_{nm }results in the following subexpression. QO_{1m}=(2ˆA0)*S0+(2ˆB0)*S0 . . . (2ˆAn)*Sn+(2ˆBn)*Sn+(Power of 2)*R1+ . . . +(Power of 2)*Ro+(Power of 2)*Y1 . . . +(Power of 2)*Yp . . . RE_{1m}=(2ˆA0)*S0+(2ˆB1)*S0 . . . (2ˆAn)*Sn+(2ˆBk)*Sn+(Power of 2)*R1+ . . . +(Power of 2)*Ro+(Power of 2)*Y1 . . . +(Power of 2)Yp; where (2ˆA0), (2ˆB0), (2ˆAn), (2ˆBn), (2ˆB1), (2ˆBk) are powers of 2. The expressions QO_{1m }to RE_{1m }are represented in Table B as follows.
TABLE B 

Assuming that the expression ((2ˆA0)*S0+(2ˆAn)*Sn)) row 1 of Table B is equal to the expression ((2ˆA0)*S0+(2ˆAn)*Sn)) in row 2. Thus X1=((2ˆA0)*S0+(2ˆAn)*Sn) is the derived horizontal subexpression. Thus, the resultant horizontal subexpressions can be derived as follows:
Y=cf_{11}{cf_{12}{ . . . {cf_{1m}{QO_{1m}}+RE_{1m}}+cf_{21}{cf_{22}{ . . . {cf_{2m}{QO_{2m}}+RE_{2m}}+cf_{31}{cf_{32 }. . . } . . . }+RE (Similar type of equation as QO),
where QO_{1m}=(Power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp+(Power of 2)*X1+ . . . +(Power of 2)*Xq,
RE_{1m}=(Power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp+(Power of 2)*X1+ . . . +(Power of 2)*Xq,
QO_{2m}=(Power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp+(Power of 2)*X1+ . . . +(Power of 2)*Xq,
RE_{2m}=(Power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp+(Power of 2)*X1+ . . . +(Power of 2)*Xq,
where horizontal subexpressions X1=(power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp,
Xq=(power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp,
where R1 to Ro are recursive subexpressions, Y1 to Yp are vertical subexpressions and X1 to Xq are horizontal subexpressions. Thus, the coefficient equation ‘Y’ is implemented with expansion of common factors as (2ˆz+1) or (2ˆz−1) where z is a positive number.
The method is hereby described with reference to the structural implementation of the optimized sumofproducts expression, the embodiment herein described pertaining to the direct and transposed form of digital filter realization.
Mapping of the generalized subexpressions and final expression into a generalized hardware structure: The generalized recursive subexpressions can be written as;
R_{1}=(Power of 2)*S_{0}+ . . . (Power of 2)*S_{n},
R_{o}=(Power of 2)*S_{0}+ . . . (Power of 2)*Sn,
where, Power of 2 is an integral power of 2.
Mapping generalized recursive subexpression into hardware can be illustrated with the help of an example as follows.
Assuming that the recursive subexpression is represented by the equation; R1=S0+S1+4*S5+4*S8 and the optimized subexpression be R1=S0+S1+4*(S5+S8).
The area efficient way, of implementing the subexpression R1 is, implementing S0+S1 and adding 4*(S5+S8) to it. (The area efficiency here is in having incremental growth of precision at different levels. Any conventional algorithm resulting in incremental growth of precision of adders at different levels can be used to implement the subexpressions. Further, since the method mainly deals with reducing the number of adders not the precision of the adder's, the method to reduce the precision of adders is not discussed herein). This resultant structure in FIG. 3(a2)1 is called a convergent set of 2SAD (2input Shifter and Adder).
The input set to the convergent structure in FIG. (3a2)1 is {S1, S1, S5, S8}. As illustrated in FIG. 3(a2)1, the complete input set S0 to Sn is not fed to the 2SAD block, as subexpression R1 does not have all the inputs. Further the shifter structure in FIG. (3a2)1 would depend on the powers of 2 applied to the coefficients in the subexpression R1. The convergent set of 2SAD structure can be extended to any subexpression with (power of 2) coefficients for S0 to Sn. The subexpressions R1 to Ro map on to the converging substructure SS [R_{1}] to SS [R_{o}] in hardware as shown in FIG. 3(a3).
In the Vertical subexpression Y1=D_{qmm1}*S1, Yp=D_{qpmn}*S_{n}. D_{qmm1}, D_{qpmn }are the numbers that can be expanded as 2ˆz+1 or 2ˆz−1, where z is a positive number. Mapping the generalized vertical subexpression into hardware can be illustrated with the help of an example as follows.
Assuming that the vertical subexpression is represented by the equation Y1=33*S1=32*S1+S1. This can be implemented as 1SAD (1 input shift and add) as shown in FIG. (3a22). Since the Vertical Subexpressions in would result in the substructures as shown in FIG. (3a22), it is therefore observed that the Y1 to Yp would form substructures; SS[Y1] . . . SS[Yp] as illustrated in FIG. 3(a3). The left shifts in 1SAD would depend on the powers of two assigned to the coefficients D_{qmm1}, . . . D_{qpmn}.
The horizontal subexpressions for the instant predefined coefficient set inputs are given as below:
X1=(power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp, . . .
Xq=(power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp
The above subexpressions result in the similar convergent 2SAD substructure. The substructures SS[X1] to SS[Xq] are illustrated in FIG. 3(a3). The quotient and remainder subexpressions QO_{1m }. . . RE_{2m }given below:
QO_{1m}=(Power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp+(Power of 2)*X1+ . . . +(Power of 2)*Xq,
RE_{1m}=(Power of 2)*S0+ . . . (Power of 2)*Sn+(Power of 2)*R1+ . . . (Power of 2)*Ro+(Power of 2)*Y1+ . . . +(Power of 2)*Yp+(Power of 2)*X1+ . . . +(Power of 2)*Xq,
The resultant structure for the above stated subexpressions is analogous to the convergent 2SAD substructures. The substructure is further illustrated as SS [QO1m] . . . SS[RE1m] . . . in FIG. 3(a3). The final equation of the coefficient bank combining all the quotient, and remainder expressions can be written as Y=cf_{11 }{cf_{12 }{ . . . {cf_{1m}{QO_{1m}}+RE_{1m}}+cf_{21}{cf_{22 }{ . . . {cf_{2m}{QO_{2m}}+RE_{2m}}+cf_{31 }{cf_{32 }. . . }+RE (Similar type of equation as QO),
Expanding the common factor's cf_{11 }. . . as 2ˆz+1 or 2ˆz−1 would result in another expression with coefficients as power of 2's which will result in convergent set of 2SAD structure with input set {QO_{1m}, RE_{1m }. . . }. This structure is shown as SS[Y] in FIG. 3(a3).
Generalized Structure depicting the minimal area integrated circuit implementation of the sum of products expression.
The generalized structure is shown in FIG. 3(a3) to map the results of the algorithm. The generalized structure contains taps S0 to Sn from the plurality of unit sample delays (11) connected to a group of substructures SS[R1] to SS[Ro] with inputs S0 to Sn each generating plurality of signals R1 to Ro. The substructures SS[Y1] to SS[Yp] with inputs S0 to Sn, generate the plurality of signals Y1 to Yp, where p is any positive number; substructures SS[X1] to SS[Xq] with inputs S0 to Sn, R1 to Ro and Y1 to Yp each generating signals X1 to Xq, where q is any positive number, substructures SS[QO1m] . . . SS[RE1m] . . . with inputs S0 to Sn, R1 to Ro, Y1 to Yp and X1 to Xq generating signals QO1m . . . RE1m etc., and substructure SS[Y] formed from inputs QO1m . . . to RE1m . . . generating output Y, these substructures contain a plurality of adders (13, 15) and shifter 14, these substructures are connected together to form the coefficient bank of the direct form of FIR filter or IIR filter.
The invented method for the direct form implementation of the coefficient bank can be used to implement a single number also. This is done by applying a single number at the input of the algorithm in FIG. 3(a1). A single number given to the algorithm would result in an optimized shift and add hardware structure of the number.
The method of the instant invention is hereby described with reference to the given (Optimization 1) coefficient set as the input.
Direct form of Coefficient Bank Implementation:
Decompose high cost terms into some multiples of cost1's and power of 2's. The algorithm has three optimizations. The three optimizations are explained with the help of an example as follows:
Optimization 1: This is a recursive optimization. By this optimization an individual coefficient ‘coeff1’ will be decomposed to:
coeff1=cf(qo)+re
or
coeff1=cf(qo)−re;
where cf=common factor, qo=quotient, re=remainder.
Further, the coefficient set called ‘coeffset’ will be decomposed to coeffset=cf(QO)+RE. Note: The negative coefficients are not treated separately, negative coefficient are implemented as positive coefficients with a subtractor at the end of substructure while adding to other substructures. The SOP realization for the example coefficient set is described with reference to the steps involved in the method as illustrated in FIG. 3(a1).
Steps Involved in Optimization 1:
Step 1: Decompose the coefficient set into odd fundamentals and multiplication factor set:
{coefficient set}={odd fundamental set}*{multiplication factor set},
(Each coefficient is divided with 2 till it is odd. If it is of odd magnitude, then it is moved to the odd fundamental set. The number of 2's with which each coefficient is divided is moved into multiplication factor set along with the sign of the coefficient.)
For the given set of coefficients; the coefficient set is decomposed to an odd fundamental set and multiplication factor set as follows:
{−79,1044,−5890,27916,49362,−8382,1628,−154,63,126}={79,261,2945,6979,24681,4191,407,77,63,63}*{−1,4,−2,4,2,−2,4,−2,1,2};
where the Coefficient set={−79,1044,−5890,27916,49362,−8382,1628,−154, 63,126},
Odd fundamental set={79,261,2945,6979, 24681, 4191, 407, 77,63,63}, individual elements in the odd fundamental set are called odd fundamentals. Multiplication factor set={−1,4,−2,4,2,−2,4,−2,1,2}
Step 2: Searching for repeated odd fundamentals other than 0 &1 and removing them from the odd fundamental set and form recursive subexpressions. In the example, the odd fundamental set “63” is the repeated odd fundamental, and the example odd fundamental set changes as:
{79,261,2945,6979,24681,4191,407,77,0,0,63}* {−1,4,−2,4,2,−2,4,−2,0,0,1}.
Two 0's have come in place of S8, S9 and one 63 has come in position of S10. The recursive subexpression thus generated is R_{1}=S8+2S9.
Step 3: Checking the odd fundamental set has only 0's, 1's and cost1's and exiting from the optimization 1 with recursive subexpressions and equation of direct form of coefficient bank, else proceed to step 4 for optimizing the numbers greater than the cost 1 numbers.
Step 4: From the odd fundamental set compute the “prec”, precision of maximum odd fundamental (max_num).
prec=ceil(log 2(max_num))
Form the common factor set, which contains only cost1's, cost1's are formed by (2ˆz+1) or (2ˆz−1), where z varies from ‘1‘to ’ prec’.
For the example ‘max_num’=24681, ‘prec’=ceil (log 2(abs(24681)))=15 and the common factor set is (3, 5, 7, 9 . . . 32769} formed by 2ˆz+1 or 2ˆz−1, z varying from ‘1’ to ‘15’.
Step 5: Calculate the total weight of the numbers in the odd fundamental set wherein, the Total weight of numbers in the odd fundamental set=weight of each number in the odd fundamental set. The weight of each odd fundamental in the example odd fundamental set is in Table 3 as given below:
TABLE 3  
Number  Weight  
79  100  
261  100  
2945  100  
6979  100  
24681  100  
4191  100  
407  100  
77  100  
63  2  
The total weight of the numbers in the odd fundamental set for the example=802, along similar lines the weight of any number set can be calculated.
Step 6: For each common factor in the common factor set decompose the odd fundamental set into a common factor, quotient and remainder.
{Odd fundamental set}=cf*{QO}+{RE};
where ‘cf’common factor, ‘QO’ is quotient set and ‘RE’ is remainder set, total weight of decomposed odd fundamental set after setting common factor equal to the weight of common factor+weight of quotient set+weight of remainder set.
For each common factor of the example starting from 3 to 32769 calculate the total weight. Choose the best common factor, which gives the lowest possible total weight. The best common factor for given example is “63” as it reduces the total weight of the odd fundamental set from “802” to “323”, thus the lowest possible weight is achieved.
The odd coefficient set is decomposed into common factor, quotient set and remainder set as follows:
{79,261,2945,6979,24681,4191,407,77,0,0,63}=63{1,4,47,111,391,66,7,1,0,0,1}+{16,9,−16,−14,48,33,−34,14,0,0,0}
where the common factor=63,
quotient set={1,4,47,111,391,66,7,1,0,0,1}
remainder set={16,9,−16,−14,48,33,−34,14,0,0,0}
Step 7: Multiply the QO and RE with multiplication factor set and treat the resultant new_QO and new_RE as new coefficient sets. Go to step 1 with new coefficient sets ‘new_QO’ &‘new_RE’. For the given example ‘new_QO’={1,4,47,111,391,66,7,1,0,0,1}*{−1,4,−2,4,2,−2,4,−2,0,0,1}={−1,16,−94,444,782,−132,28,−2,0,0,1}—this number set passes through all the steps in optimization 1 again and stops with ‘7’ as common factor. Another number set ‘new_RE’={16,9,−16,−14,48,33,−34,14,0,0,0}*{−1,4,−2,4,2,−2,4,−2,0,0,1}={−16,36,32,−56,96,−66,−136,−28,0,0,0}—this contains all cost1's and pow_{—}2's and stops at step 3 while passing through the steps of optimization 1.
After the optimization 1 the coefficient equation of the example number set for direct form implementation of coefficient bank is in equation (3a).
Y=[−16]*S0+36*S1+32*S2+96*S4+[−66]*S5+[−136]*S6+[−28]*R2+63{[−1]*S0+16*S1+4*S2+[−4]*S3+[−2]*S4+[−132]*S5+[−2]*S7+1*R1+7{64*S3+4*S6+[−14]*R3}} (3a),
where recursive subexpressions for the given equation are:
R1=1*S8+2*S9, R2=2*S3+1*S7, R3=1*S2+[−8]*S4.
The number of adders obtained after optimization 1 in the above equation is 29. The coefficient equation (3a) can also be written as equation (3b).
Y=RE1+63(RE2+7(QO2)) (3b);
where expressions,
RE1=[−16]*S0+36*S1+32*S2+96*S4+[−66]*S5+[−136]*S6+[−28]*R2 (3c1),
RE2=[−1]*S0+16*S1+4*S2+[−4]*S3+[−2]*S4+[−132]*S5+[−2]*S7+1*R1 (3c2),
QO2=64*S3+4*S6+[−14]*R3 (3c3), and recursive subexpression in expressions (3c1, 3c2, 3c3) are
R1=1*S8+2*S9, R2=2*S3+1*S7, R3=1*S2+[−8]*S4.
The resultant structure after the optimization 1 from equation (3b) is shown in FIG. 3(b1). The input X, of the filter is given to the plurality of unit sample delays (11) resulting in signals S0 to S9. These signals are given to a group of substructures that are defined for the following input/output parameters:
Note: After the end of optimization 1 the remainder and quotient expressions contain only 0's, power of 2's, cost1's.
Optimization 2 (Vertical Optimization):
The table formed from the quotient expression, and remainder expressions is scanned column wise for the possibility of taking cost1 coefficients as the common subexpression. This would further reduce the total weight of the coefficient equation. This optimization is explained with the quotient and remainder expressions of an example as follows.
The coefficient equation of the example in consideration after optimization 1 is Equation (3a) as given below:
Y=[−16]*S0+36*S1+32*S2+96*S4+[−66]*S5+[−136]*S6+[−28]*R2+63{[−1]*S0+16*S1+4*S2+[−4]*S3+[−2]*S4+[−132]*S5+[−2]*S7+1*R1+7{64*S3+4*S6+[−14]*R3}} (3a);
where recursive subexpressions are R1=1*S8+2*S9, R2=2*S3+1*S7, R3=1*S2+[−8]*S4,
Step 8: separate the remainder and quotient expressions in the coefficient equation. Equation (3a) can also be written as
Y=RE1+63(RE2+7(QO2)) (3b)
where quotient (QO2) and remainder expressions (RE1, RE2) are
RE1=[−16]*S0+36*S1+32*S2+96*S4+[−66]*S5+[−136]*S6+[−28]*R2 (3c1)
RE2=[−1]*S0+16*S1+4*S2+[−4]*S3+[−2]*S4+[−132]*S5+[−2]*S7+1*R1 (3c2)
QO2=64*S3+4*S6+[−14]*R3 (3c3)
Step 9: Find the vertical subexpressions in the table formed from the separated quotient and remainder expressions by traversing the table column wise, looking for common odd fundamental cost 1s. Forming a table from the quotient (3c3) and remainder (3c1, 3c2) expressions for the given example. By writing coefficients in each equation column wise under respective signals into a table as shown in Table 4.
TABLE 4 
Expressions RE1, RE2, QO2 in form of a table 

In Table 4, column 7 has the subexpression −2*33 and subexpression −4*33, the odd fundamental for both of these subexpressions is 33. Hence, instead of calculating 33 twice we can calculate it only once by taking 33*S5 as vertical subexpression. This is called vertical subexpression as we are taking the common odd fundamental column wise in the table. The example coefficient equation (3b) after substituting the vertical subexpression results in equation 3(f):
Y=[−16]*S0+36*S1+32*S2+96*S4+[−2]*Y1+[−136]*S6+[−28]*R2+63{[−1]*S0+16*S1+4*S2+[−4]*S3+[−2]*S4+[−4]*Y1+[−2]*S7+1*R1+7{64*S3+4*S6+[−14]*R3}} (3f);
where Y1=33*S5,
The equation (3f) can be rewritten as follows:
Y=RE1+63(RE2+7(QO2)) (3g),
where RE1=[−16]*S0+36*S1+32*S2+96*S4+[−2]*Y1+[−136]*S6+[−28]*R2
RE2=[−1]*S0+16*S1+4*S2+[−4]*S3+[−2]*S4+[−4]*Y1+[−2]*S7+1*R1
QO2=64*S3+4*S6+[−14]*R3,
It is observed that the number of adders in the coefficient equation (3f) after optimizations 2 are 28.
The structure obtained after the optimization 2 (equation (3g)) is illustrated in FIG. 3(b2). The input X of the filter is given to the plurality of unit sample delays (11) resulting in signals S0 to S9, these signals are given to a group of substructures that operate for the following inputoutput parameters:
(a) Substructure SS [R1] with inputs S8, S9 results in signal R1,
(b) Substructure SS[R2] with inputs S7, S3 results in signal R2,
(c) Substructure SS[R3] with inputs S2, S4 results in signal R3,
(d) Substructure SS[Y1] with inputs S5 results in signal Y1,
(e) substructure SS[RE1] with inputs S0, S2, S4, R2, S1, S6, Y1 results in signal RE1,
(f) ubstructure SS[RE2] with inputs S0, S4, S7, S2, S3, R1, S1, Y1 results in signal RE2, substructure SS[QO2] with inputs S3, S6, R3 results in signal QO2;
(g) substructure SS[Y] with inputs RE1, RE2, QO2 results in output Y of the direct form of FIR filter, the group of substructures formed from plurality of adders (13,15) and shifters (14) is called direct form of coefficient bank (1).
Optimization 3 (Horizontal Optimization):
Step 10: Expand cost 1s coefficients in to quotient, and remainder expressions and vertical subexpression as (2ˆz+1) or (2ˆz−1) where ‘z’ is a positive number.
The resultant coefficient equation for the example is as given below:
Y={32*S1+4*S1+32*S2+128*S4+[−128]*S6+[−8]*S6+[−32]*R2+4*R2+[−2]*Y1+[−16]*S0+[−32]*S4+63{16*S1+4*S2+[−4]*S3+[−2]*S7+[−4]*Y1+[−1]S0+[−2]*S4+R1+7{64*S3+4*S6+[−16]*R3+2*R3}}} (3h)
or
Y=RE1+63(RE2+7(QO2)) (3i)
where expression RE1=32*S1+4*S1+32*S2+128*S4+[−128]*S6+[−8]*S6+[−32]*R2+4*R2+[−2]*Y1+[−16]*S0+[−32]*S4, RE2=16*S1+4*S2+[−4]*S3+[−2]*S7+[−4]*Y1+[−1] S0+[−2]*S4+R1
QO2=64*S3+4*S6+[−16]*R3+2*R3.
The resultant structure after the optimization 2 from equation (3i) is shown in FIG. 3(b3). The input X of the filter is given to plurality of unit sample delays (11) resulting in signals S0 to S9, these signals are given to a group of substructures, substructure SS [R1] with inputs S8, S9 results in signal R1, substructure SS [R2] with inputs S7, S3 results in signal R2, substructure SS [R3] with inputs S2, S4 results in signal R3, substructure SS [Y1] with inputs S5 results in signal Y1, substructure SS[RE1] with inputs S0, S2, S4, R2, S1, S6, Y1 results in signal RE1, substructure SS[RE2] with inputs S0, S4, S7, S2, S3, R1, S1, Y1 results in signal RE2, substructure SS[QO2] with inputs S3, S6, R3 results in signal QO2; substructure SS[Y] with inputs RE1, RE2, QO2 results in output Y of the direct form of FIR filter, the group of substructures formed from plurality of adders (13,15) and shifters (14) is called direct form of coefficient bank (1).
Step 11: Find the optimal horizontal subexpressions in the table formed from separated quotient and remainder expressions by traversing the table row wise and looking for common horizontal subexpressions, there is possibility of overlap in the horizontal subexpressions so choose the best horizontal subexpressions which reduce the total weight of the coefficient equation to the lowest as possible. The quotient expressions (QO2) and remainder expression (RE1, RE2) as given in equation (3i) in table form are in Table 5.
TABLE 5 
Expression RE1, RE2 and QO2 after step 10 expressed in the tabular form as follows: 

In table 5 the subexpression −16*S0−32*S4 of row2 and subexpression −S0−2*S4 of row3 can be treated as single subexpression ‘S0+2*S4’. This is called horizontal subexpressions as we are selecting common subexpression across the rows of the table. There are no overlaps for the given table. The resultant coefficient equation obtained after optimization 3 for the example is:
Y={32*S1+4*S1+32*S2+128*S4+[−128]*S6+[−8]*S6+[−32]*R2+4*R2+[−2]*Y1+[−16]*X1+63{[−1]*X1+[−2]*S7+4*S2+[−4]*S3+16*S1+R1+[−4]*Y1+7{4*S6+2*R3+64*S3+[−16]*R3}}X
or
Y=RE1+63(RE2+8QO2−QO2) (3h)
where expressions RE1=32*S1+4*S1+32*S2+128*S4+[−128]*S6+[−8]*S6+[−32]*R2+4*R2+[−2]*Y1+[−16]*X1
RE2=[−1]*X1+[−2]*S7+4*S2+[−4]*S3+16*S1+R1+[−4]*Y1
QO2=4*S6+2*R3+64*S3+[−16]*R3
Equation (3h) can further be simplified and written as:
Y=RE1+64(RE2+8QO2−QO2)−(RE2+8QO2−QO2) (3i),
The number of adders after all the three optimizations is 27. The resultant structure after the three optimizations (from equation (3i)) is shown in FIG. 3(b4). The input X of the filter is given to the plurality of unit sample delays (11) resulting in signals S0 to S9, these signals are given to a group of substructures, substructure SS[R1] with inputs S8, S9 results in signal R1, substructure SS[R2] with inputs S7, S3 results in signal R2, substructure SS[R3] with inputs S2, S4 results in signal R3, substructure SS[Y1] with inputs S5 results in signal Y1, substructure SS[X1] with inputs S0, S4 results in signal X1, substructure SS[RE1] with inputs X1, S2, R2, S1, S6, Y1 results in signal RE1, substructure SS[RE2] with inputs X1, S7, S2, S3, R1, S1, Y1 results in signal RE2, substructure SS[QO2] with inputs S3, S6, R3 results in signal QO2; substructure SS[Y] with inputs RE1, RE2, QO2 results in output Y of the direct form of FIR filter, the group of substructures formed from plurality of adders (13,15) and shifters (14) is called the direct form of coefficient bank (1).
Generalized Structure Depicting the Minimal Area Integrated Circuit Implementation of the Sum of Products Expression:
The generalized structure is shown in FIG. 3(a3) to map the results of the algorithm, the generalized structure contains taps S0 to Sn from the plurality of unit sample delays (11) connected to a group of substructures SS[R1] to SS[Ro] with inputs S0 to Sn each generating plurality of signals R1 to Ro. The substructures SS[Y1] to SS[Yp] with inputs S0 to Sn, generate a plurality of signals Y1 to Yp, where p is any positive number; substructures SS[X1] to SS[Xq] with inputs S0 to Sn, R1 to Ro and Y1 to Yp each generating signals X1 to Xq, where q is any positive number, substructures SS[QO1m] . . . SS[RE1m . . . ] with inputs S0 to Sn, R1 to Ro, Y1 to Yp and X1 to Xq generating signals QO1m . . . RE1m etc., and substructure SS[Y] formed from inputs QO1m to RE1m generating output Y, these substructures contain a plurality of adders (13, 15) and shifter 14, these substructures are connected together to form a direct form of FIR filter or IIR filter.
The invented method for direct form implementation of the coefficient bank can be used to implement a single number also. This is done by giving a single number at the input of the algorithm in FIG. 3(a1). A single number given to the algorithm would result in a optimized shift and add hardware structure of the number.
Method for implementation of sum of products expression to thereby implement the transposed form of coefficient bank: (An embodiment of the instant invention).
The algorithm involved in the reduction of the number of adders required to implement the coefficient set is shown in FIG. 3(c1). This algorithm is explained with the help of generalized equations in FIG. 3(c2) (extended to 2 pages). Part 2 of the method is a generalized structure to map the results obtained from the algorithm as shown in FIG. 3(c3).
The equations of the individual coefficients in transposed form of coefficient bank for the example are:
S1=1044X, S2=−5890X, S3=27916X, S4=49362X, S5=−8382×, S6=1628X, S7=−154X, S8=63X, S9=126X
The following discussion is about the optimization of the coefficients for transposed form of coefficient bank for the example using the algorithm in FIG. 3(c1).
Step 1: Decompose the coefficient set into odd fundamentals and multiplication factor set:
{coefficient set}={odd fundamental set}* {multiplication factor set}
(Each coefficient is divided with 2 till it is odd, when it is odd, magnitude of odd number is moved to odd fundamental set. The number of 2's with which each coefficient is divided is moved into the multiplication factor set along with the sign of the coefficient).
Example coefficient set decomposes as follows:
{−79,1044,−5890,27916,49362,−8382,1628,−154,63,126}={79,261,2945,6979,24681,4191,407,77,63,63}*{−1,4,−2,4,2,−2,4,−2,1,2},
where {coefficient set}={−79,1044,−5890,27916,49362,−8382,1628,−154,63,126},
{odd fundamental set}={79,261,2945,6979,24681, 4191,407,77,63,63}
{multiplication factor set}={−1,4,−2,4,2,−2,4,−2,1,2}.
Step 2: Check if all the elements of the odd fundamental set has all cost 1s and 1s as odd fundamentals. If yes implement cost 1 odd fundamental as 2ˆz+1 or 2ˆz−1, where z=ceil (log 2(odd fundamental)), and use cost1 to implement coefficients and ‘exit’, else proceed to step 3. In the example odd fundamental set has only one cost 1 coefficient (63), and rest all are higher cost numbers, hence proceed to step 3.
Step 3: Check if odd fundamental set has cost 1 odd fundamentals. If yes proceed to step 4, else proceed to step 5. In the example odd fundamental set has ‘63’ which is the cost1 odd fundamental. Hence proceed to step 4.
Note: Cost1 odd fundamental is computed as follows:
prec_of_number=ceil(log 2(odd fundamental)).
If at any stage, odd fundamental=2ˆz+1 or 2ˆz−1,z varying from 1 to prec_of_number, then that odd fundamental is cost 1.
Step 4: Decompose ‘odd fundamental set’ into ‘graph set’ and ‘incomplete set’, (‘graph set’ includes cost 1 odd fundamentals from the odd fundamental set, these cost1 odd fundamentals are implemented in a graph as 2ˆz+1 or 2ˆz−1, where z=ceil (log 2(odd fundamental))), (‘incomplete set’ includes unimplemented odd fundamentals).
For the given example the odd fundamental set {79,261,2945,6979, 24681, 4191, 407, 77,63,63} is decomposed to {63} and {79,261,2945,6979,24681,4191,407,77};
where graph set={63}, 63 is implemented in graph as (2ˆ6)−1. The signals S8 and S9 are obtained as S8=((2ˆ6)−1)X, S9=(2ˆ1)*((2ˆ6)−1)X, where coefficients 63=((2ˆ6)−1)&126=(2ˆ1)*((2ˆ6)−1), incomplete set={79,261,2945,6979, 24681, 4191, 407, 77}.
Step 5: Form the ‘incomplete set’ from the odd fundamentals greater than 1. When it is found in step 3 that there are no cost1 odd fundamentals, then there is no ‘graph set’. All numbers in the odd fundamental set which are greater than 1 are moved to the ‘incomplete set’. In the example under consideration step 5 is not applicable at this stage.
Step 6: Determine the cost and coefficient equation of the numbers in the incomplete set using the algorithm in FIG. 3(a1) and make a look up table with the cost and coefficient equation. Use the algorithm in FIG. 3(a1) with each number of the incomplete set separately to obtain the coefficient equation, ‘cost of implementation of a number in incomplete set’ (cost)=cost of implementation of each number in coefficient equation+number of adders or subtractors in the coefficient equation.
The incomplete set for the example is {79,261,2945,6979, 24681, 4191, 407, 77}, the cost & coefficient equation of each number in incomplete set found from the algorithm in FIG. 3(a1) is in look up table (Table 6).
TABLE 6  
Cost & coefficient equation of each number in  
incomplete set:  
Coefficient  
Number  Cost  equation  
79  2  63 + 16  
261  2  257 + 4  
2945  3  2049 + 896  
6979  4  7[−124 + 257(16)]  
24681  3  1023(24) + 127  
4191  2  127 * 33  
407  3  3(136) − 1  
77  3  −12 + 65(−1)  
The look up table is obtained using the algorithm of FIG. 3(a1) with each number processed by the algorithm separately, for example number 79 when processed by the algorithm of FIG. 3(a1) that gives common factor 63 with remainder as 16. The cost of implementation of 79 is 2. Since 63 is a cost1 number and there is a ‘+’ to add 63 and 16. Hence it is noted in the look up table 6 shows that 79 has cost of ‘2’ and coefficient equation of 79 is ‘63+16’. Similarly for all the other numbers in the incomplete set the coefficient equation and cost are noted in Table 6.
Step 7: Check if the adder distance* for the smallest cost, smallest magnitude number (found from look up table formed in step 6) in the incomplete set is 1. If yes proceed to step 8 else it means the adder distance is greater than 1 or it is not possible to implement the number from the graph set and proceed to step 9.
*Adder Distance:
Adder distance 1 of the selected number from the incomplete set=fundamental coefficients from the existing graph set+a power of 2.
For the example incomplete set={79,261,2945,6979, 24681, 4191, 407, 77}, the smallest cost, smallest magnitude number in the incomplete set from Table 6 is ‘79’, 79 has the adder distance ‘1’ as ‘79’=63 (existing fundamental from the graph set)+16.
Step 8: Use the graph set to implement the selected number (from step 7) in the ‘incomplete set’ and remove the number from the ‘incomplete set’ and move it to the ‘graph set’. For the given example the number 79 has an adder distance ‘1’ hence it is implemented in the form of the graph as ‘63+16’ (with ‘−1’ at the end of the structure as coefficient is −79), and 79 is removed from ‘incomplete set’ and is added to the ‘graph set’. Now the graph set={63,79}, the incomplete set={261,2945,6979, 24681, 4191, 407, 77}, now move to step 10.
Step 9: Use the coefficient equation in the look up table obtained in step 6 to implement the number in the incomplete set and add the odd fundamentals in the coefficient equation into the graph set. For the example in step 7 it is found that the number 79 has an adder distance 1. If the number to be implemented has an adder distance more than 1 then the implementation of the coefficient ‘79’ can be searched from the look up table.
Step 10: Check, whether the incomplete set is Null if yes exit, if not go to the step 7. For the example as the incomplete set {261,2945,6979, 24681, 4191, 407, 77} is not a null set then go to step 7 with the incomplete set. For the example step 7 select 261 as next number to be implemented, and the adder distance for the number ‘261’ is neither 1 nor 2. Hence step 9 is invoked and the implementation in look up table 6 for the number 261 i.e., (257+4)(with 4 at the end of the structure as coefficient is 4*261=1044) is taken, and the odd fundamental 257 and 261 are added to the graph.
The implementation of remaining numbers in incomplete set of the example using the step 7 to step 10 of algorithm in FIG. 3(c1) results in equations as follows,
S1=1044X=16+(257)*4
S2=−5890X=−(7)*256+(−2)*(2049)
S3=27916X=(7)*((−124)+(257)*16)
S4=49362X=(258)+(1023)*(48)
S5=−8382X=(127)*(−66)
S6=1628X=−4+(17)(96)
S7=−154X=(79−2)*(−2)
FIG. 3(c3) shows the generalized structure for the transposed form of the coefficient bank, where the structure has a plurality of substructures SS [S0], SS[S1] . . . SS[Sn]. SS[S0] has X as input and S0 as the output, SS[S1] has A and X as the input and S1 as output. The substructures are formed from adders (13,15) (adders or subtractors are referred as adders for the ease of explanation), shifter (14), to form the transposed form of coefficient bank (2). The taps obtained from the structure of the individual coefficient equations are connected to the unit sample delays (11) to form the transposed form of FIR/IIR filter.
The number of adders required to implement the graph for the example in FIG. 3(d) is 18. The resultant graph structure for all the coefficients of the example for the transposed form of the coefficient bank (2) is shown in FIG. 3(d). This structure is formed from the plurality of the adders (13,15) and the shifter (14) connected together depending on coefficients. Part 2 of the method is the generalized structure formed from the algorithm in part 1, the structure has plurality of substructures SS[S0], SS[S1] . . . SS[Sn], where SS[S0] has X as input and S0 as output, SS[S1] has A and X as the input and S1 as the output etc., the substructures are formed from adders (13,15), shifter 14, to form the transposed form of the coefficient bank (2), the taps obtained from structure of the individual coefficient equations are connected to unit sample delays (11) to form a transposed form of FIR filter or IIR filter.
Advantages of the Method:
The number of adders using the CSD based method and the system to implement the example coefficient bank are tabulated in Table 7.
TABLE 7  
Number of adders using CSD based method and  
invented method to implement the example coefficient bank:  
Existing (CSD)  Invented  
Direct  34  27  
Transposed  20  18  
It is evident from the above table that the invented method has less number of adders compared to existing method in implementation of direct and transposed form of coefficient bank. Thus, it is observed that the instant invention provides a minimal area implementation of a sum of products expression.