How to indicate input size to dsp.BlockLMSFilter when using MATLAB Coder?

I would like to use dsp.BlockLMSFilter in code that will be converted to C code via MATLAB Coder. I understand that Length and BlockSize of dsp.BlockLMSFilter must be entered as literals in the constructor for the object. I have tried to feed the object signals whose lengths are equal to an integer multiple of BlockSize, but I get a warning about variable input size when I run codegen, even if I have "released" the object beforehand. I have the same problem if I feed the signal to the object in successive BlockSize-length blocks.

Is there some way to assure a dsp.BlockLMSFilter object (in code that will be converted to C and compiled) that the size of the incoming signal is acceptable?

Suppose that x and d (in the simplified example below) are guaranteed to have lengths that are multiples of the BlockSize. How do I let the code generator know this? Is there an assert that might help?

function wts = example(x,d,mu) %#codegen
    obj = dsp.BlockLMSFilter('Length',10,'BlockSize',10,'StepSizeSource','Input port');
    [y,errs,wts] = obj(x,d,mu);
end

[David Fink]

What specific warning or error are you seeing, and what commands are you using to generate code?

Note - if using the MATLAB Coder App, you can convert this to a code generation script via:

https://www.mathworks.com/help/coder/ug/generate-a-matlab-script-to-build-a-project.html

[Joe M]

@David Fink Thank you for responding. My question is part of a much larger project, so I will illustrate with an artificial m-file that captures just this particular issue's behavior. Below I include

• the m-file,
• main.c,
• my build-script,
• and the error message.

For my m-file, I borrow some content from the documentaion for dsp.BlockLMSFIlter.

% myexample.m
function myexample(N) %#codegen

filt = dsp.FIRFilter('Numerator',fir1(10,[0.5, 0.75]));
x = randn(N*100,1);
t = (0:length(x)-1)'/length(x);
d = filt(x) + sin(2*pi*8*t);

blms = dsp.BlockLMSFilter('Length',10,'BlockSize',5,'StepSize',0.01);

[y, err, w] = blms(x, d);
fprintf("norm(d - err): %f\n", norm(d - err));

end

/* main.c */
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include "myexample_initialize.h"
#include "myexample_terminate.h"

void example( const int N );

int main(int argc, char * const argv[])
{
    int N;
    N = atoi( argv[1] );
    myexample_initialize():
    myexample( N );
    myexample_terminate();
    
    return 0;
}

% buildscript.m
u = coder.typeof(int32(1));

cfg = coder.config('exe');
cfg.Toolchain = 'Intel IPP | make (64-bit Linux)';
cfg.CustomSource = 'main.c';
cfg.CustomInclude = pwd;

codegen -report -config cfg -args { u } myexample;

When I run buildscript.m, I receive the following error message.

??? Reported by dsp.BlockLMSFilter: This system object (or its current configuration) does not support variable input sizes.

Error in ==> myexample Line 11 Column 1
Code generation failed: View Error Report

Error using codegen

Error in buildscript (line 9)
codegen -report -config cfg -args { u } myexample;

Answer 1

[Pranav Venuprasad]

The dsp.BlockLMSFilter object does not support variable input sizes as mentioned in the error message. This begavior is not specific to codegen. For example consider the code snippet below:

blms = dsp.BlockLMSFilter(10,5);
filt = dsp.FIRFilter;
filt.Numerator = fir1(10,[.5, .75]);
x = randn(1000,1); % Noise
d = filt(x) + sin(0:.05:49.95)'; % Noise + Signal
[y, err] = blms(x, d);
[y2, err2] = blms([x;x], [d;d]); % double the size of the input

The above snippet produces the following error even though the size of the input in the second step is a multiple of the "BlockSize" property of the dsp.BlockLMSFilter object.

Error using dsp.BlockLMSFilter/parenReference
Changing the size on input 1 is not allowed without first calling the release() method.

In codegen to lock the input frame size, you can provide the size of the input as a constant when compiling. In the example mentioned above, if the desired value of N is say 10, you can generate the C code using the following snippet

N = int32(10);
codegen myexample -args {coder.Constant(N)};

[Joe M]

@Pranav Venuprasad: Thank you for the answer.

Unfortunately, a fixed signal-size is not an option for my work.

I wrote my own block LMS filter function, and I would like it to perform as much like the dsp.BlockLMSFilter system object as possible. This leads me to a question, for which I need to provide some background.

The MATLAB code that I inherited, uses a dsp.BlockLMSFilter object (named blms) and applies

for ii = 1:10

[y, err, w] = step(blms, x, d);

end

It is clear from my experiments that blms maintains an internal state, but even after reading the documentation I am unsure of the details. I suspect that the weights are part of the internal state, but I would like to know if there is more. Where can I find concrete answers?

[Pranav Venuprasad]

The dsp.BlockLMSFilter object maintains the internal state according to the update equations provided in the following documentation page: https://www.mathworks.com/help/dsp/ref/blocklmsfilter.html#f3-1110408 . If you would like to create a custom functionality for Block LMS filtering with variable size inputs, you can use make use of the equations outlined on the above page for the same.