Poor MEX performance when running filtfilt

Hi all,

I'm currently on the journey of converting my MATLAB data processing script into C. While testing some part of it (filtering data using filtfilt), I noticed that running the generated MEX file is MUCH slower than running the MATLAB function directly.

This is my MATLAB function:

function filter_test(data)
%$codegen
coder.inline('never');

% Initialise filtered data
f_data = data;

% Set the filter coeficients to remove the 50Hz and its harmonics
filter_coeficients = struct();
filter_coeficients.coef_50Hz = [...]; %1x4467 double
filter_coeficients.coef_100Hz = [...]; %1x4467 double
filter_coeficients.coef_150Hz = [...]; %1x4467 double
filter_coeficients.coef_200Hz = [...]; %1x4467 double
filter_coeficients.coef_250Hz = [...]; %1x4467 double
filter_coeficients.coef_300Hz = [...]; %1x4467 double
filter_coeficients.coef_350Hz = [...]; %1x4467 double
filter_coeficients.coef_400Hz = [...]; %1x4467 double
filter_coeficients.coef_450Hz = [...]; %1x4467 double

% Apply all the filters
filter_coeficients_id = fieldnames(filter_coeficients);
for c=1:numel(filter_coeficients_id)
    f_data = filtfilt(filter_coeficients.(filter_coeficients_id{c}), 1, f_data);
end

end

I generated the MEX file (.mexw64) with the following command:

codegen('filter_test', '-args', {coder.typeof(double(0), [Inf Inf])})

This was one of the test I made:

data = rand(30000,1);

tic
filter_test(data);  % Elapsed time is 1.914894 seconds.
toc

tic
filter_test_mex(data); % Elapsed time is 492.794533 seconds.
toc

The difference is abysmal! Forcing the code to coder.inline('always') made this even worse, as the run time reaches 997.7 seconds.

Am I doing something wrong? I guess the C performance will be similar to the MEX performance? Why is the MEX so slow when compared to MATLAB?

[Darshan Ramakant Bhat]

Two suggestions :

• Please make sure that you are running the MEX file (not the .m function). Typically codegen command on function 'foo' will produce a 'foo_mex' file. Similarly you will be having 'filter_test_mex' file.
• Using rand() function for testing performance may not be a good idea, since it my produce different numbers for two different cases. Try to save the values from rand() in a mat file and use the same values for different experiments.

[Diogo Tecelão]

Thanks for your reply. I have corrected both the points that you mentioned. I am indeed running the filter_test_mex function (from filter_test_mex.mexw64), and I'm now comparing the performance with the exact same data. I've also made some further testing.

Turns out that the MEX function is 257 times slower than MATLAB! Any idea?

[Diogo Tecelão]

I have converted the code into a DLL, and by running the DLL it takes 7 seconds to run. It is an improvement, but still much slower than MATLAB. I was expecting C to be faster. Is there any way I can optimize this?

[Wilson A N]

Hi Diogo,

Can you please share the complete reproduction steps along with the filter coefficients (seems to be large ~4467) ?

Additionally, which MATLAB release are you using?

I want to try to reproduce the issue and understand what is happening.

-Wilson

[Diogo Tecelão]

Hi Wilson! Thanks for your reply and attention. How can I share the filter coeficients? I cannot paste them here, as there's a character limit.

[Wilson A N]

Hi Diogo,

You can save the filter coefficients as a .mat file and attach it here.

[Wilson A N]

Hi Diogo,

I tried using some random filter coefficients to try and reproduce the issue but I was not able to observe the slowdown you had reported. Below are my results:

MATLAB Simulation time is

1.4649

MEX Simulation time is

3.4633

SIL Simulation time is

4.3738

(SIL simulation is actually checking lib/dll time)

Please update the MATLAB Release in which you are facing the issue along with the filter coefficients. This will help us to look into the issue further.

Answer 1

[Jan]

Just a remark: With FiltFiltM (https://www.mathworks.com/matlabcentral/fileexchange/32261-filterm ) my Matlab 2018b needs 6.0 instead of 8.0 seconds as for filtfilt.