Pre-allocating a function since initialization is slow

Hi guys,

Consider this code:

clc; clear all;
t2=linspace(0,pi,10);
f2=cos(t2);

tic; cumtrapz(t2,f2); t1=toc
tic; cumtrapz(t2,f2); t2=toc
t1/t2

It turns out that the first call to cumtrapz takes around 7 times longer.

I want to run the cumtrapz function in real time within a function that is called by a nonlinear solver (IPOPT), many times +20000 in a single timestep. To be clear; the cumtrapz is only called ONCE every function call.

How can I make sure that everytime cumtrapz it is called, it does not take 7 times longer, due to initialisation stuff. In other words, can I e.g. preallocate the cumtrapz function in the workspace?

*Note that this has nothing to do with proallocating the output of cumtrapz since it is not stored to workspace in the code above.

Answer 1

[Daniel Shub]

What you are seeing is the reason that people suggest that you do not use

clear all

As long as you do not clear the cumtrapz function from memory (or its subfunctions) it won't take 7x as long.

[Martijn]

Ok, I understand, but I do not clear during the loop. Consider this:

----------------------

function testspeedpreallocation

t2=linspace(0,pi,10);

f2=cos(t2);

for i=1:5

t(i)=integrate(t2,f2)

end

function t=integrate(t,f)

tic; cumtrapz(t,f); t=toc;

end

end

--------------------

The first call is still about 2x slower.

[Jan]

@Daniel: Exactly. CLEAR ALL removes the already loaded functions from the memory. This is useful, if you have changed the contents of all M-files dynamically - a very rare condition.

I hope that MLINT will offer a warning for this frequently used command.

[Jan]

@Martijn: What happens if you use a subfuntion instead of a nested function?

[Martijn]

@Jan Simon: Same thing.

[Daniel Shub]

Timing things in MATLAB that take less than 100 microseconds tends not to be reliable. If I increase linspace from 10 points to 1e7 points the cumtrapz takes close to a second and I see no difference between the first and second call.