C++ code generated by Matlab coder too slow

Hello,

I have an algorithm written in matlab, which detects some artefatcs in blood pulsewave. These algorithm is converted to C++ by maltab coder and I use these C++ files in my C# SW. It works fine.

But, I use several FIRs in my code, one of them is 10000 order FIR. The signal length is around 1M samples, so it take too much time to count it with filter(...) function or conv(..) function. I have tried to count the FIR with convolution theorem, I have used this:

fftLen = 2^nextpow2(length(values));
filteredValues = ifft(fft(values,fftLen).*fft(coefs,fftLen));

It works more then 30x faster in matlab. But if I convert the code to C++ by matlab Coder, it is not so fast. It is more then 10x slower than the matlab version. I use matlab 2017b. Why?

[Adam]

A lot of builtin code is well optimised in Matlab and utilises multi-threading. As far as I am aware C code created through code will not use multithreading and may lose many of the optimisations that are achieved by the Matlab code. I have no idea how Matlab coder works to turn Matlab code into C code, but it is not perfect. Sometimes it will create suboptimal C code.

[Josef Shrbeny]

I do not have parallel computing toolbox, hence, I quess, the matlab code also does not run multi threading

[Walter Roberson]

MATLAB detects some patterns of code, and if it finds that the arrays are "large enough" to make it worth while, automatically invokes BLAS or LAPACK libraries to do the work. Those libraries are multi-threaded even without the Parallel Computing toolbox.

[Preethi]

Hi Walter,

BLAS and LAPACK are applicable from which version of MATLAB

[Walter Roberson]

Multithreaded was added in R2007a according to Loren's blog of March 1, 2007.

Answer 1

[Ryan Livingston]

If the bottleneck in the generated code is in fft or other MATLAB builtin functions, I'd suggest taking a look at this answer:

https://www.mathworks.com/matlabcentral/answers/102725-why-is-my-matlab-coder-2-1-r2011b-generated-mex-file-slower-than-my-matlab-function

In short, those functions use highly optimized libraries in MATLAB as Walter mentioned and also for computing FFTs. The default generated code may be less optimized in favor of portability.

The answer describes that as of R2017b you can instruct MATLAB Coder and Simulink Coder to generate calls to FFTW to improve FFT performance:

https://www.mathworks.com/help/coder/ug/speed-up-fast-fourier-transforms-in-generated-standalone-code-by-using-fftw-library-calls.html

There are similar capabilities to generate calls to BLAS and LAPACK for matrix operations and linear algebra:

https://www.mathworks.com/help/coder/ug/speed-up-matrix-operations-in-generated-standalone-code-by-using-blas-calls.html

https://www.mathworks.com/help/coder/ug/generate-code-that-calls-lapack-functions.html

Note: When generating C++ code that calls FFTW there is a bug in R2017b that was fixed in R2017b_update5:

https://www.mathworks.com/matlabcentral/answers/429535-error-in-generating-code-with-custom-fft-library-in-matlab-coder#answer_346760

So if you would like to use FFTW with Coder and generate C++ in R2017b, I'd suggest installing that update.

[Josef Shrbeny]

Ryan,

I have tried FFTW library and it works very fast now

Tnank you

[Ryan Livingston]

Great to hear!