Hello there! Hope to find a solution to my little issue here.
I'm really new to pure data and I'm not sure yet about its limits.
I'll go straight to the point. I'm trying to compute the autocorrelation function of a big array (4092 elements) and I'm trying to do it as quick as possible.
I made the patch (just a few math objects connected together) using a metro object as the main pulse of the index shifting needed to compute the autocorrelation, but I soon realized that the metro object has a limit in speed that is way under what I expected. I get the result I need but it takes a few minutes.
What I need to know is: how can I make quicker calculations?
Is this aim beyond the possibilities of pure data?
If the answer is yes, is there a way (a not so tricky one) to designate the calculation to some other softfware which is more appropriate? (MATLAB, for example?)

Thank you in advance

ps: this is somehow related to another issue I'm thinking about these day, which is how to write "for" cycles in a simple way in pure data...maybe I'm missing something big about how pure data works.

seems like you could use tabsend~ to an rfft~ subpatch and multiply it by its conjugate. http://en.wikipedia.org/wiki/Autocorrelation under Efficient computation

as fer for loops you could do something with until, or some kind of counter, some relational operations and select, I like the ones in ekext.

@sebfumaster said:

seems like you could use tabsend~ to an fft~ subpatch and multiply it by its conjugate. http://en.wikipedia.org/wiki/Autocorrelation under Efficient computation

Wow that seems to be a very elegant way to do it, I'll surely try, thank you.

It should solve the problem, ok, but still I don't understand how pure data works with fast index shifting in general. What if I find myself with a similar problem that has nothing to do with FFT, crosscorrelation or convolution?

i think that for things that are convolution-y like that which work with multiple indexes at once and need to store multiple add-multiply operations it's way just easier/faster in the frequency domain. You'd probably need a dedicated microprocessor (dsp?) if you want to do large array operations like that in the time domain quickly.
For table lookup though, the objects are tabread~ and tabread4~ or their non tilde'd equivalents. phasor~ will increment indices at audio rate

I'd use octave over matlab if you need it, it's free! I don't know how much faster it might be though

if you need to it might be possible to cycle over every computation window with a phasor~(or vline~) going into tabread~s and feed the results into a rpole~ (real pole) object with a coefficient of one in order to sum the results of whatever operation you're doing, clearing the rpole~ every time. There are some examples of this kind of use of rpole~ on this forum somewhere.

yop,

seems that this topic could be interresting for you : http://puredata.hurleur.com/sujet-6846-autocorrelation-wiener-khinchin-theorem-why-doesn-work

Katjav website is a goldmine : http://www.katjaas.nl/home/home.html#anchor2

Nau

Thank you all! Your answers are helping so much and making me understand more about my problem and about pure data in general.

And then there is this Katjav website...mindblowing! WOW!