IIR Filter

7.3. IIR Filter#

Consider an \(N\)-order IIR filter specified by (7.2) in the time domain or (7.4) in the \(z\)-domain. For the sake of drawing simpler SFGs, we assume below that the feedback order \(N\) is at least as large as the feedforward order \(M\), i.e., \(N \geq M\).
The transfer function \(H(z)\) of the IIR filter is given by

(7.12)#\[\begin{align} H(z) &= \frac{\sum_{k=0}^M b_k z^{-k}}{\sum_{k=0}^N a_k z^{-k}} & ~~~(a_0=1). \end{align} \]

Thus, the IIR filter can be interpreted as a cascade of a feedforward component, i.e., an \(M\)-order FIR filter with transfer function \(\sum_{k=0}^M b_k z^{-k}\), and a feedback component, i.e., an \(N\)-order IIR filter with transfer function \(\frac{1}{\sum_{k=0}^N a_k z^{-k}}\) and a trivial feedforward tap.

7.3.1. Direct-form Implementation#

The feedback component, i.e.,

\[\begin{equation*} Y(z) = - \sum_{k=1}^N a_k z^{-k} Y(z), \end{equation*}\]

can be implemented in the direct form similar to the development in Section 7.2.1.
When the feedforward FIR component is also implemented in the direct form, we obtain the following SFG of direct form I for the IIR filter:

\[\begin{align*} \!\bigcirc\kern-6.5pt\vcenter{\tiny x} \longrightarrow & \!\bigcirc\!\!\xrightarrow{\hspace{9pt}{\scriptsize b_0}\hspace{9pt}}\!\!\bigcirc\!\!\xrightarrow{\hspace{26pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{26pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc\kern-6.5pt\vcenter{\tiny y} \\[-0pt] {\scriptsize z^{-1}} & \Big\downarrow \hspace{25pt} \Big\uparrow \hspace{27pt} \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \\[-10pt] &\!\bigcirc\!\!\xrightarrow{\hspace{9pt}{\scriptsize b_1}\hspace{9pt}}\!\!\bigcirc \hspace{25.5pt} \!\bigcirc\!\!\xleftarrow{\hspace{7pt}{\scriptsize -a_1}\hspace{6pt}}\!\!\bigcirc \\[-0pt] {\scriptsize z^{-1}} & \Big\downarrow \hspace{25pt} \Big\uparrow \hspace{27pt} \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \\[-0pt] & \,\vdots \hspace{29pt} \vdots \hspace{29pt} \,\vdots \hspace{30pt} \vdots \\[-0pt] {\scriptsize z^{-1}} & \Big\downarrow \hspace{25pt} \Big\uparrow \hspace{27pt} \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \\[-10pt] &\!\bigcirc\!\!\xrightarrow{\hspace{8pt}{\scriptsize b_M}\hspace{8pt}}\!\!\bigcirc \hspace{25pt} \!\bigcirc\!\!\xleftarrow{\hspace{6pt}{\scriptsize -a_N}\hspace{5pt}}\!\!\bigcirc \end{align*}\]
Swapping the order of the feedforward and feedback components in the cascade above, we obtain following SFG of direct form II for the IIR filter:

\[\begin{align*} \!\bigcirc\kern-6.5pt\vcenter{\tiny x} \longrightarrow & \!\bigcirc\!\!\xrightarrow{\hspace{25pt}}\!\!\bigcirc \!\!\xrightarrow{\hspace{9pt}{\scriptsize b_0}\hspace{9pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc\kern-6.5pt\vcenter{\tiny y} \\[-0pt] & \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \\[-10pt] & \!\bigcirc\!\!\xleftarrow{\hspace{6.5pt}{\scriptsize -a_1}\hspace{7pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{9pt}{\scriptsize b_1}\hspace{9pt}}\!\!\bigcirc \\[-0pt] & \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \\[-0pt] & \,\vdots \hspace{30pt} \vdots \hspace{27.5pt} \,\vdots \\[-0pt] & \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \\[-10pt] &\!\bigcirc\!\!\xleftarrow{\hspace{5.5pt}{\scriptsize -a_M}\hspace{5pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{8pt}{\scriptsize b_M}\hspace{8pt}}\!\!\bigcirc \\[-0pt] & \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \\[-0pt] & \,\vdots \hspace{30pt} \vdots \\[-0pt] & \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \\[-10pt] &\!\bigcirc\!\!\xleftarrow{\hspace{5.5pt}{\scriptsize -a_N}\hspace{5.5pt}}\!\!\bigcirc \end{align*}\]

Below is an example HLS function that implements the direct-form II SFG above:

// This is a Chebyshev type-II lowpass filter with M=N=6 (L=7) 
const din_t b[ffL] = {
  0.168814352044553, 0.792090023955804, 1.723851142494651, 2.197013405545998,
  1.723851142494651, 0.792090023955804, 0.16881435204455
};

const din_t a[fbL] = {
  1.0, 1.776425956077127, 2.196211818205448, 1.583727976930950, 
  0.765387107901598, 0.216231507219810, 0.028540076201079
};

void iir(hls::stream<din_t> &in, hls::stream<dout_t> &out, int N) {

  static din_t w[fbL] = {};
#pragma HLS array_partition variable=w type=complete

  sample_loop: for (int n=0; n<N; n++) {
#pragma HLS loop_tripcount max=MAX_N
    delay_loop: for (int k=fbL-1; k>0; k--) {
      w[k] = w[k-1];
    }
    // Read in new sample from in
    dout_t y = in.read();
    dout_t by = 0.0;
#pragma HLS bind_op variable=y op=mul impl=fabric
#pragma HLS bind_op variable=by op=mul impl=fabric
    acc_loop: for (int k=1; k<fbL; k++) {
      y -= a[k]*w[k];
      if (k<ffL)
        by += b[k]*w[k];
    }
    w[0] = y;
    y *= b[0];

    // Write to out
    out.write(y+by);
  }
}

Vitis HLS gives a RTL implementation of sample_loop with II=2.
The bottleneck prevents achieving II=1 is the *carried dependence that updating w[0] in an iteration of sample_loop requires the accumulation of y in acc_loop to complete first. Although acc_loop is automatically unrolled by Vitis HLS, it still takes at least two clock cycles to complete accumulation on y for the example IIR filter of order \(6\). As a result, II=1 cannot be achieved for this filter.
This bottleneck problem is inherent to the direct-form structure of the IIR implementation, and would be more severe as the order of the IIR filter increases. Thus, the direct-from implementation may not be suitable for implementing higher-order IIR filters operating at a high sampling rate.

7.3.2. Transposed-form Implementation#

When both the feedforward and feedback components in the cascade are implemented in the transposed form similar to the development in Section 7.2.2, we obtain the following SFG of transposed from I:

\[\begin{align*} \!\bigcirc\kern-6.5pt\vcenter{\tiny x} \longrightarrow & \!\bigcirc\!\!\xrightarrow{\hspace{25pt}}\!\!\bigcirc \!\!\xrightarrow{\hspace{26pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{9pt}{\scriptsize b_0}\hspace{9pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc\kern-6.5pt\vcenter{\tiny y} \\[-0pt] {\scriptsize z^{-1}} & \Big\uparrow \hspace{25pt} \Big\downarrow \hspace{27pt} \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \\[-10pt] &\!\bigcirc\!\!\xleftarrow{\hspace{6pt}{\scriptsize -a_1}\hspace{6pt}}\!\!\bigcirc \hspace{25.5pt} \!\bigcirc\!\!\xrightarrow{\hspace{9pt}{\scriptsize b_1}\hspace{9pt}}\!\!\bigcirc \\[-0pt] {\scriptsize z^{-1}} & \Big\uparrow \hspace{25pt} \Big\downarrow \hspace{27pt} \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \\[-0pt] & \,\vdots \hspace{29pt} \vdots \hspace{29pt} \,\vdots \hspace{30pt} \vdots \\[-0pt] {\scriptsize z^{-1}} & \Big\downarrow \hspace{25pt} \Big\uparrow \hspace{27pt} \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \\[-10pt] &\!\bigcirc\!\!\xleftarrow{\hspace{5.5pt}{\scriptsize -a_N}\hspace{4.5pt}}\!\!\bigcirc \hspace{25pt} \!\bigcirc\!\!\xrightarrow{\hspace{8pt}{\scriptsize b_M}\hspace{9pt}}\!\!\bigcirc \end{align*}\]
Swapping the order of the feedforward and feedback components in the cascade above, we obtain following SFG of transposed form II for the IIR filter:

\[\begin{align*} \!\bigcirc\kern-6.5pt\vcenter{\tiny x} \longrightarrow & \!\bigcirc\!\!\xrightarrow{\hspace{10pt}{\scriptsize b_0}\hspace{9pt}}\!\!\bigcirc \!\!\xrightarrow{\hspace{24pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc\kern-6.5pt\vcenter{\tiny y} \\[-0pt] & \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \\[-10pt] & \!\bigcirc\!\!\xrightarrow{\hspace{10pt}{\scriptsize b_1}\hspace{9pt}} \!\!\bigcirc\!\!\xleftarrow{\hspace{6.5pt}{\scriptsize -a_1}\hspace{6pt}}\!\!\bigcirc \\[-0pt] & \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \\[-0pt] & \,\vdots \hspace{30pt} \vdots \hspace{27.5pt} \,\vdots \\[-0pt] & \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \\[-10pt] &\!\bigcirc\!\!\xrightarrow{\hspace{8.5pt}{\scriptsize b_M}\hspace{8pt}} \!\!\bigcirc\!\!\xleftarrow{\hspace{5pt}{\scriptsize -a_M}\hspace{5pt}}\!\!\bigcirc \\[-0pt] & \hspace{33pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \\[-0pt] & \hspace{33pt} \,\vdots \hspace{29.5pt} \vdots \\[-0pt] & \hspace{33pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \\[-10pt] & \hspace{33pt} \!\bigcirc\!\!\xleftarrow{\hspace{5.5pt}{\scriptsize -a_N}\hspace{5.5pt}}\!\!\bigcirc \end{align*}\]

Below is an example HLS function that implements the transposed-form II SFG above:

// This is a Chebyshev type-II lowpass filter with M=N=6 (L=7) 
const din_t b[ffL] = {
  0.168814352044553, 0.792090023955804, 1.723851142494651, 2.197013405545998,
  1.723851142494651, 0.792090023955804, 0.16881435204455
};

const din_t a[fbL] = {
  1.0, 1.776425956077127, 2.196211818205448, 1.583727976930950, 
  0.765387107901598, 0.216231507219810, 0.028540076201079
};

void iir(hls::stream<din_t> &in, hls::stream<dout_t> &out, int N) {

  static dout_t u[fbL-1] = {};
#pragma HLS array_partition variable=u type=complete

  sample_loop: for (int n=0; n<N; n++) {
#pragma HLS loop_tripcount max=MAX_N
    din_t x = in.read();
    dout_t bx = b[0]*x;
    dout_t y = bx+u[0];
    dout_t ay;
#pragma HLS bind_op variable=bx op=mul impl=dsp
#pragma HLS bind_op variable=ay op=mul impl=dsp
    // Write to out
    out.write(y);

    delay_add_loop: for (int k=1; k<fbL-1; k++) {
      ay = a[k]*y;
      if (k<ffL) {
        bx = b[k]*x;
        u[k-1] = u[k]+bx-ay;
      } else {
        u[k-1] = u[k]-ay;
      }
    }
    ay = a[fbL-1]*y;
    if (fbL==ffL) {
      bx = b[fbL-1]*x;
      u[fbL-2] = bx-ay;
    } else {
      u[fbL-2] = -ay;
    } 
  }
}

Vitis HLS gives a RTL implementation of sample_loop with II=1.
The structure of the transposed-form implementation “simplifies” the carried dependence in sample_loop to that the update y=bx+u[0] needs to complete before the calculation of a[1]*y for updating u[0] in the next iteration of the loop. Similar to the first-order IIR filter example in Section 4.2.1, the objective of II=1 can be achieved as long as the update y=bx+u[0] and the multiplication a[1]*y afterward can be achieved within a clock cycle. This condition is much easier to satisfied for example by using DSP slices to implement the multipliers.
In general, the transposed-form structure is more suitable for higher speed implementation of IIR filters since the carried dependence induced by the transposed-form structure involves only a single multiplication rather than the accumulation of many products.

7.3.3. Cascade-form Implementation#

As in Section 7.2.3, if the IIR filter taps in (7.12) are all real-valued, then the transfer function in (7.12) of the IIR filter can be factored into a product of the transfer functions of SoSs with real-valued taps in the form below:

(7.13)#\[\begin{align} H(z) &= b_0 \prod_{k=1}^K \frac{1+b_{k,1} z^{-1} + b_{k,2} z^{-2}}{1+ a_{k,1} z^{-1} + a_{k,2} z^{-2}} \end{align} \]

Hence, the IIR filter may be implemented as a cascade of all the SoSs.
Each SoS may be implemented in the direct form II or in the transposed form II. For example:
- Cascade-form SFG with direct-form II SoSs:
\[\begin{align*} \!\bigcirc\kern-6.5pt\vcenter{\tiny x} \xrightarrow{\hspace{8pt}{\scriptsize b_0}\hspace{7pt}} & \!\bigcirc\!\!\xrightarrow{\hspace{25pt}}\!\! \bigcirc\!\!\xrightarrow{\hspace{24.5pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc \!\!\xrightarrow{\hspace{25pt}}\!\!\bigcirc \!\!\xrightarrow{\hspace{24.5pt}}\!\!\bigcirc \!\!\longrightarrow \cdots \longrightarrow \!\bigcirc\!\!\xrightarrow{\hspace{25pt}}\!\! \bigcirc\!\!\xrightarrow{\hspace{25pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc\kern-6.5pt\vcenter{\tiny y} \\[-0pt] & \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \hspace{19pt} \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \hspace{54pt} \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \\[-10pt] & \!\bigcirc\!\!\xleftarrow{\hspace{5.5pt}{\scriptsize -a_{1,1}}\hspace{4pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{8pt}{\scriptsize b_{1,1}}\hspace{7pt}}\!\!\bigcirc \hspace{15pt} \bigcirc\!\!\xleftarrow{\hspace{5.5pt}{\scriptsize -a_{2,1}}\hspace{4pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{7pt}{\scriptsize b_{2,1}}\hspace{7pt}}\!\!\bigcirc \hspace{17pt} \cdots \hspace{19pt} \bigcirc\!\!\xleftarrow{\hspace{4.5pt}{\scriptsize -a_{K,1}}\hspace{3pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{6.5pt}{\scriptsize b_{K,1}}\hspace{6pt}}\!\!\bigcirc \\[-0pt] & \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \hspace{19pt} \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \hspace{54.5pt} \Big\uparrow \hspace{26pt} \Big\downarrow {\scriptsize z^{-1}} \hspace{15pt} \Big\uparrow \\[-10pt] & \!\bigcirc\!\!\xleftarrow{\hspace{5.5pt}{\scriptsize -a_{1,2}}\hspace{4pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{8pt}{\scriptsize b_{1,2}}\hspace{7pt}}\!\!\bigcirc \hspace{15.5pt} \bigcirc\!\!\xleftarrow{\hspace{5.5pt}{\scriptsize -a_{2,2}}\hspace{4pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{8pt}{\scriptsize b_{2,2}}\hspace{7pt}}\!\!\bigcirc \hspace{17pt} \cdots \hspace{18pt} \bigcirc\!\!\xleftarrow{\hspace{4.5pt}{\scriptsize -a_{K,2}}\hspace{3pt}} \!\!\bigcirc\!\!\xrightarrow{\hspace{6.5pt}{\scriptsize b_{K,2}}\hspace{6pt}}\!\!\bigcirc \end{align*}\]
- Cascade-form SFG with transposed-form II SoSs:
\[\begin{align*} \!\bigcirc\kern-6.5pt\vcenter{\tiny x} \xrightarrow{\hspace{8pt}{\scriptsize b_0}\hspace{7pt}} & \!\bigcirc\!\!\xrightarrow{\hspace{25pt}}\!\! \bigcirc\!\!\xrightarrow{\hspace{24.5pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc \!\!\xrightarrow{\hspace{25pt}}\!\!\bigcirc \!\!\xrightarrow{\hspace{24.5pt}}\!\!\bigcirc \!\!\longrightarrow \cdots \longrightarrow \!\bigcirc\!\!\xrightarrow{\hspace{25pt}}\!\! \bigcirc\!\!\xrightarrow{\hspace{25pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc\kern-6.5pt\vcenter{\tiny y} \\[-0pt] & \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \hspace{19pt} \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \hspace{54pt} \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \\[-10pt] & \!\bigcirc \!\!\xrightarrow{\hspace{8pt}{\scriptsize b_{1,1}}\hspace{7.5pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{4pt}{\scriptsize -a_{1,1}}\hspace{4pt}}\!\!\bigcirc \hspace{16pt} \bigcirc \!\!\xrightarrow{\hspace{8pt}{\scriptsize b_{2,1}}\hspace{7pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{4.5pt}{\scriptsize -a_{2,1}}\hspace{4pt}}\!\!\bigcirc \hspace{17pt} \cdots \hspace{19pt} \bigcirc \!\!\xrightarrow{\hspace{7pt}{\scriptsize b_{K,1}}\hspace{6pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{3.5pt}{\scriptsize -a_{K,1}}\hspace{3pt}}\!\!\bigcirc \\[-0pt] & \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \hspace{19pt} \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \hspace{54.5pt} \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{15pt} \Big\downarrow \\[-10pt] & \!\bigcirc \!\!\xrightarrow{\hspace{8pt}{\scriptsize b_{1,2}}\hspace{7.5pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{4.5pt}{\scriptsize -a_{1,2}}\hspace{4pt}}\!\!\bigcirc \hspace{15.5pt}\bigcirc \!\!\xrightarrow{\hspace{8pt}{\scriptsize b_{2,2}}\hspace{7pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{4.5pt}{\scriptsize -a_{2,2}}\hspace{4pt}}\!\!\bigcirc \hspace{17pt} \cdots \hspace{19.5pt} \bigcirc \!\!\xrightarrow{\hspace{7pt}{\scriptsize b_{K,2}}\hspace{6.5pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{3.5pt}{\scriptsize -a_{K,2}}\hspace{2.5pt}}\!\!\bigcirc \end{align*}\]

Below is an example HLS function that implements the cascade-form SFG with transposed-form II SoSs:

// This is a Chebyshev type-II lowpass filter with M=N=6 (L=7)
// expressed as SoSs. 
#define K 3
const din_t b[K][2]={
  {1.931625159186722, 0.999999999999988},
  {1.540424279915726, 1.000000000000037},
  {1.220028067120242, 0.999999999999978}
};
const din_t a[K][2]={
  {0.637344246881063, 0.128737933973358},
  {0.576772441445408, 0.316322235552652},
  {0.562309267750655, 0.700839985387954},
};
const din_t b0=0.168814352044553;

void iir2nd(dout_t &in, dout_t &out, dout_t &w1, dout_t &w2, 
            const din_t b[2], const din_t a[2]) {
#pragma HLS inline

  // Write to out
  out = in+w1;
  // Update register
  dout_t bx = b[0]*in;
  dout_t ay = a[0]*out;
#pragma HLS bind_op variable=bx op=mul impl=dsp
#pragma HLS bind_op variable=ay op=mul impl=dsp
  w1 = w2+bx-ay;
  bx = b[1]*in;
  ay = a[1]*out;
  w2 = bx-ay;
}

void iir(hls::stream<din_t> &in, hls::stream<dout_t> &out, int N) {

  static dout_t u[K+1];
  static dout_t w[K][2] = {};
#pragma HLS array_partition variable=u type=complete
#pragma HLS array_partition variable=w type=complete

  sample_loop: for (int n=0; n<N; n++) {
#pragma HLS loop_tripcount max=MAX_N
    u[0] = b0*in.read();
    cascade_loop: for (int k=0; k<K; k++) {
      iir2nd(u[k], u[k+1], w[k][0], w[k][1], &b[k][0], &a[k][0]);
    }
    out.write(u[K]);
  }
}

Vitis HLS gives a RTL implementation of sample_loop with II=1. The latency of sample_loop is higher than that achieved using the transposed-form implementation.

7.3.4. Parallel-form Implementation#

For an IIR filter (\(N \geq M\)) with real-valued taps and single-order poles, the transfer function \(H(z)\) in (7.12) can be expanded into a sum of partial fractions :

(7.14)#\[\begin{align} H(z) &= B_0 + \sum_{k=1}^K \frac{B_{k,0} + B_{k,1} z^{-1}}{1 + A_{k,1} z^{-1} + A_{k,2} z^{-2}} \end{align} \]

where all the \(A\) and \(B\) coefficients are real-valued.
Based on (7.14), we may implement the IIR filter as parallel second-order IIR components in the following SFG:

\[\begin{align*} & \!\bigcirc\!\!\xrightarrow{\hspace{32pt}{\scriptsize B_{0}}\hspace{31.5pt}}\!\!\bigcirc \\[-10pt] \nearrow \hspace{3pt} & \hspace{72pt} \searrow \\[-10pt] \!\bigcirc\kern-6.5pt\vcenter{\tiny x}\longrightarrow \!\!\bigcirc\!\!\longrightarrow & \!\bigcirc\!\!\xrightarrow{\hspace{7pt}{\scriptsize B_{1,0}}\hspace{6pt}}\!\! \bigcirc\!\!\xrightarrow{\hspace{25.5pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc\kern-6.5pt\vcenter{\tiny y} \\[-0pt] \Big| \hspace{19pt} & \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{16pt} \Big\downarrow \hspace{18.5pt} \Big\uparrow \\[-10pt] \big| \hspace{19.0pt} & \!\bigcirc \!\!\xrightarrow{\hspace{7pt}{\scriptsize B_{1,1}}\hspace{6.5pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{4pt}{\scriptsize -A_{1,1}}\hspace{4pt}}\!\!\bigcirc \hspace{18.2pt} \big| \\[-0pt] \Big| \hspace{19pt} & \hspace{33pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{16pt} \Big\downarrow \hspace{19.8pt} \Big| \\[-10pt] \Big\downarrow \hspace{17.5pt} & \hspace{31.5pt}\bigcirc \!\!\xleftarrow{\hspace{4.5pt}{\scriptsize -A_{1,2}}\hspace{4pt}}\!\!\bigcirc \hspace{17.8pt} \Big| \\[-0pt] \!\!\bigcirc\!\!\longrightarrow & \!\bigcirc\!\!\xrightarrow{\hspace{7pt}{\scriptsize B_{2,0}}\hspace{6pt}}\!\! \bigcirc\!\!\xrightarrow{\hspace{25.5pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc \\[-0pt] \Big| \hspace{19pt} & \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{16pt} \Big\downarrow \hspace{18.5pt} \Big\uparrow \\[-10pt] \big| \hspace{19.0pt} & \!\bigcirc \!\!\xrightarrow{\hspace{7pt}{\scriptsize B_{2,1}}\hspace{6.5pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{4pt}{\scriptsize -A_{2,1}}\hspace{4pt}}\!\!\bigcirc \hspace{18.2pt} \big| \\[-0pt] \Big| \hspace{19pt} & \hspace{33pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{16pt} \Big\downarrow \hspace{19.8pt} \Big| \\[-10pt] \Big\downarrow \hspace{17.5pt} & \hspace{31.5pt}\bigcirc \!\!\xleftarrow{\hspace{4.5pt}{\scriptsize -A_{2,2}}\hspace{4pt}}\!\!\bigcirc \hspace{17.8pt} \Big| \\[-0pt] \vdots \hspace{19.5pt} & \hspace{35pt} \vdots \hspace{54.8pt} \vdots \\[-0pt] \Big\downarrow \hspace{17.5pt} & \hspace{90.5pt} \Big\uparrow \\[-10pt] \!\!\bigcirc\!\!\longrightarrow & \!\bigcirc\!\!\xrightarrow{\hspace{5pt}{\scriptsize B_{K,0}}\hspace{6pt}}\!\! \bigcirc\!\!\xrightarrow{\hspace{25.5pt}}\!\!\bigcirc \!\!\longrightarrow\!\!\bigcirc \\[-0pt] & \Big\downarrow \hspace{26pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{16pt} \Big\downarrow \\[-10pt] & \!\bigcirc \!\!\xrightarrow{\hspace{5pt}{\scriptsize B_{K,1}}\hspace{6.5pt}}\!\!\bigcirc \!\!\xleftarrow{\hspace{4pt}{\scriptsize -A_{K,1}}\hspace{2pt}}\!\!\bigcirc \\[-0pt] & \hspace{33pt} \Big\uparrow {\scriptsize z^{-1}} \hspace{16pt} \Big\downarrow \\[-10pt] & \hspace{31.5pt}\bigcirc \!\!\xleftarrow{\hspace{4pt}{\scriptsize -A_{K,2}}\hspace{2.5pt}}\!\!\bigcirc \end{align*}\]

where each second-order component IIR filter is implemented in the transposed form II.

Below is an example HLS function that implements the parallel-form SFG with transposed-form II second-order components

// This is a Chebyshev type-II lowpass filter with M=N=6 (L=7)
// expressed as a parallel sum of second-order components. 
#define K 3
const din_t B[K][2]={
  { 0.182834372101273, -0.306910458916128},
  {-1.151247645072097, -1.125034655836861},
  {-4.777765413376408, -1.682130810244864}
};
const din_t A[K][2]={
  {0.562309267750655, 0.700839985387954},
  {0.576772441445408, 0.316322235552652},
  {0.637344246881063, 0.128737933973358},
};
const din_t B0=5.914993038391787;

void piir2nd(din_t &in, dout_t &out, dout_t &w1, dout_t &w2,
             const din_t B[2], const din_t A[2]) {
#pragma HLS inline

  dout_t bx = B[0]*in;
#pragma HLS bind_op variable=bx op=mul impl=dsp
  // Write to out
  out = bx+w1;

  // Update register
  dout_t ay = A[0]*out;
#pragma HLS bind_op variable=ay op=mul impl=dsp
  bx = B[1]*in;
  w1 = w2+bx-ay;
  ay = A[1]*out;
  w2 = -ay;
}

void iir(hls::stream<din_t> &in, hls::stream<dout_t> &out, int N) {
  dout_t w[K][2] = {};
#pragma HLS array_partition variable=w type=complete

  sample_loop: for (int n=0; n<N; n++) {
#pragma HLS loop_tripcount max=MAX_N
    din_t x = in.read();
    dout_t y = B0*x;
    parallel_loop: for (int k=0; k<K; k++) {
      dout_t yy;
      piir2nd(x, yy, w[k][0], w[k][1], &B[k][0], &A[k][0]);
      y += yy;
    }
    out.write(y);
  }
}

Vitis HLS gives a RTL implementation of sample_loop with II=1.