HDL Code Generation for LMS Filter

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% MATLAB Design: Adaptive Noise Canceler algorithm using Least Mean Square 
% (LMS) filter implemented in MATLAB
%
% Key Design pattern covered in this example: 
% (1) Use of function calls
% (2) Function inlining vs instantiation knobs available in the coder
% (3) Use of system objects in the testbench to stream test vectors into the design
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%#codegen
function [filtered_signal, y, fc] = mlhdlc_lms_fcn(input, ...
                                        desired, step_size, reset_weights)
% 'input'  : The signal from Exterior Mic which records the ambient noise.
% 'desired': The signal from Pilot's Mic which includes 
%            original music signal and the noise signal
% 'err_sig': The difference between the 'desired' and the filtered 'input'
%           It represents the estimated music signal (output of this block)
% 
% The LMS filter is trying to retrieve the original music signal('err_sig') 
% from Pilot's Mic by filtering the Exterior Mic's signal and using it to 
% cancel the noise in Pilot's Mic. The coefficients/weights of the filter 
% are updated(adapted) in real-time based on 'input' and 'err_sig'.

% register filter coefficients
persistent filter_coeff;
if isempty(filter_coeff)
    filter_coeff = zeros(1, 40);
end

% Variable Filter: Call 'mtapped_delay_fcn' function on path to create 
% 40-step tapped delay
delayed_signal = mtapped_delay_fcn(input);

% Apply filter coefficients 
weight_applied = delayed_signal .* filter_coeff;

% Call treesum function on matlab path to sum up the results
filtered_signal = mtreesum_fcn(weight_applied);

% Output estimated Original Signal
td = desired;
tf = filtered_signal;
esig = td - tf;
y = esig;

% Update Weights: Call 'update_weight_fcn' function on MATLAB path to 
% calculate the new weights
updated_weight = update_weight_fcn(step_size, esig, delayed_signal, ...
                                   filter_coeff, reset_weights);

% update filter coefficients register
filter_coeff = updated_weight;
fc = filter_coeff;

function y = mtreesum_fcn(u)
%Implement the 'sum' function without a for-loop
%  y = sum(u);

%  The loop based implementation of 'sum' function is not ideal for 
%  HDL generation and results in a longer critical path. 
%  A tree is more efficient as it results in
%  delay of log2(N) instead of a delay of N delay

%  This implementation shows how to explicitly implement the vector sum in 
%  a tree shape to enable hardware optimizations.

%  The ideal way to code this generically for any length of 'u' is to use 
%  recursion but it is not currently supported by MATLAB Coder


% NOTE: To instruct MATLAB Coder to compile an external function, 
% add the following compilation directive or pragma to the function code
%#codegen

% This implementation is hardwired for a 40tap filter.

level1 = vsum(u);
level2 = vsum(level1);
level3 = vsum(level2);
level4 = vsum(level3);
level5 = vsum(level4);
level6 = vsum(level5);
y = level6;

function output = vsum(input)

coder.inline('always');

vt = input(1:2:end);
    
for i = int32(1:numel(input)/2)
    k = int32(i*2);
    vt(i) = vt(i) + input(k);
end

output = vt;

function tap_delay = mtapped_delay_fcn(input)
% The Tapped Delay function delays its input by the specified number 
% of sample periods, and outputs all the delayed versions in a vector
% form. The output includes current input

% NOTE: To instruct MATLAB Coder to compile an external function, 
% add the following compilation directive or pragma to the function code
%#codegen

persistent u_d;
if isempty(u_d)
    u_d = zeros(1,40);
end


u_d = [u_d(2:40), input];

tap_delay = u_d;

function weights = update_weight_fcn(step_size, err_sig, ... 
            delayed_signal, filter_coeff, reset_weights)
% This function updates the adaptive filter weights based on LMS algorithm

%   Copyright 2007-2022 The MathWorks, Inc.

% NOTE: To instruct MATLAB Coder to compile an external function, 
% add the following compilation directive or pragma to the function code
%#codegen

step_sig = step_size .* err_sig;
correction_factor = delayed_signal .* step_sig;
updated_weight = correction_factor + filter_coeff;

if reset_weights
    weights = zeros(1,40);
else    
    weights = updated_weight;
end

% Returns an adaptive FIR filter System object,
% HLMS, that computes the filtered output, filter error and the filter
% weights for a given input and desired signal using the Least Mean
% Squares (LMS) algorithm.

%   Copyright 2011-2022 The MathWorks, Inc.
clear('mlhdlc_lms_fcn');

hfilt2 = dsp.FIRFilter(...
        'Numerator', fir1(10, [.5, .75]));
rng('default'); % always default to known state  
x = randn(1000,1);                              % Noise
d = step(hfilt2, x) + sin(0:.05:49.95)';         % Noise + Signal

stepSize = 0.01;
reset_weights =false;

hSrc = dsp.SignalSource(x);
hDesiredSrc = dsp.SignalSource(d);

hOut = dsp.SignalSink;
hErr = dsp.SignalSink;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%Call to the design
%%%%%%%%%%%%%%%%%%%%%%%%%%%%
while (~isDone(hSrc))
    [y, e] = mlhdlc_lms_fcn(step(hSrc), step(hDesiredSrc), ... 
                                        stepSize, reset_weights);
    step(hOut, y);
    step(hErr, e);
end

figure('Name', [mfilename, '_signal_plot']);
subplot(2,1,1), plot(hOut.Buffer), title('Noise + Signal');
subplot(2,1,2),plot(hErr.Buffer), title('Signal');

coder -hdlcoder -new mlhdlc_lms_nc