Converting symbolic to MATLAB functions became 30x slower from R2018b to R2020a

I have code that converts a vector-valued 16x1 symbolic function of 50 variables into a MATLAB function stored in a file:

fxout=matlabFunction(fx,'Vars',order,'File','fxfun');

In R2018b, this ran in less than 2 seconds. In R2020a, to which I just upgraded, it takes 63 seconds. I profiled both and the difference is entirely in the execution time of mupadmex, a MEX file the contents of which obviously don't get profiled. The number of calls to mupadmex is higher in R2020a too, but by 25% or so, not by 30x.

I read the Symbolic Toolbox release notes for the last few versions and nothing jumps out at me as an explanation for this.

Anyone have any ideas? I don't even know where to start... I get that this should take a bit of time since it's a complex operation. I'm ok with 2 seconds. I'm not ok with 60+ seconds, especially since I need to do a similar operation for a much larger function, which used to take 2 minutes, and, if this scales linearly, would now take 1 hour.

Answer 1

[Paul Williams]

A colleague had a similar issue with symbolic toolbox between R2018 and R2019b. I investigated this and found that the degradation happens as of R2019a Update 4. This is due to a change in the lib.tar (toolbox\symbolic\mupad\lib). Inside the tar, a function was changed that optimizes the symbolic expressions. A special case was added for single input expressions in 2019a Update 4. Specifically, GENERATE\optimize.mu, commenting out the line:

equations[k+1..-1] := flattenOneArgFunctions(equations[k+1..-1]);

and re-taring the lib file restored the pre-update performance.

[Rik]

Have you reported this to Mathworks? It sounds like something they should know about, so this can be fixed in u5.

[Vadim Elenev]

Confirming that Paul's solution works. Rik, how does one go about submitting such a report?

[Rik]

I presume by filling out the contact form, or by filing a bug report (I believe you can choose to label it as an enhancement request instead of a bug).

Alternatively you could send an email to a Mathworks employee you may have had contact with and ask them to forward it to the correct department. But filling out the contact form is clearly more polite.