Problem Getting Embedded Codegen To Generate Code For ARM CMSIS-DSP

I've tested the "Verify FIR Filter on ARM Cortex-M Processor in MATLAB" example, which works fine and generates CMSIS-DSP function calls so I'm fairly certain all of the dependent packages are installed correctly: https://uk.mathworks.com/help/supportpkg/armcortexm/ref/Verify-FIR-Filter-on-ARM-Cortex-M-Processor-in-MATLAB-example.html

I've modified it to use a 256 point FFT, using DSP.fft() and abs(), and I'm hoping to generate code calling arm_rfft_fast_f32() and arm_abs_f32().

I'm observing two problems, with the compiled code, that I've been going round in circles trying to resolve, for days.

1/ arm_rfft_fast_f32() and arm_abs_f32() are not being called.

2/ The return array (X) is being declared as the maximum stack allocation size (65536), even if I pre-declare it using zeros() or coder.nullcopy(), so is eating up too much memory:

    void ec_fft(const double x[256], double X[65536])
    {
      static creal_T Y0[65536];

The test code is listed below. Its very short but the full source code is in the attachment.

Any help would be most appreciated.

Thanks very much,

John

FFT Function: (ec_fft.m)

function [X]=ec_fft(x) %#codegen

% X=zeros(256);

persistent ft;

if isempty(ft)

ft = dsp.FFT('FFTLengthSource','Property','FFTLength',256);

end

X = abs(ft(x));

end

Test App: (ec_fft_test.m)

Fs = 1000;

L = 256;

t = (0:L-1)'/Fs;

x = cos(2*pi*250*t);

X = coder.nullcopy(x);

X = ec_fft(x);

plot(Fs/2*linspace(0,1,128), X(1:128))

title('Single-sided amplitude spectrum of noisy signal y(t)')

xlabel('Frequency (Hz)'); ylabel('|X(f)|')

Script For Creating Library (run_codegen_arm.m):

ec_fft_test

cfgEx = coder.config('lib');

cfgEx.CodeReplacementLibrary = 'ARM Cortex-M';

cfgEx.HardwareImplementation.ProdHWDeviceType = 'ARM Compatible->ARM Cortex-M';

cfgEx.GenCodeOnly = true;

codegen -d ec_fft_lib ec_fft.m -report -rowmajor -config cfgEx -args {zeros(1,256)}

[John Edwards]

Attached, version that solves the output array size declaration problem.

Solution: Use:

types = coder.getArgTypes('ec_fft_test', 'ec_fft');

and:

-args types

Also uses single precision, which is the supported precision in codegen/CMSIS-DSP.

Just need to get DSP.fft() working with codegen now :-)

Best,

John

Answer 1

[John Edwards]

Thanks to the excellent, Mathworks support, I have a solution.

Essentially, DSP.fft() only supports complex input data, not real.

For reference, the following are the CRL conditions for dsp.FFT:

Input specifications:

1. Complex values
2. Data type: single
3. Single channel (input should be column vector)
4. Input length must equal FFT length (As per dsp.FFT property values requirements)

Property values requirement for dsp.FFT to replace:

1. FFTImplementation:  Radix-2 
2. Normalize :  False 
3. FFTLength:  16, 64, 256, 1024 
4. FFTLengthSource :  'Property' 
5. WrapInput :  True  or  False . Value does not matter because input length must equal FFT length

If anyone needs to do this in the future then the project files are attached with this mail with the filename project_files.zip.

Best regards,

John