Thread Subject: Help me optimize a line of code?

Hello,

I really need to optimize this block of code:

for(i = 1:33)
E(frame,i)=sum(power(samp_borders(i):samp_borders(i+1)).^2);
end;

I wanted to use array indexing instead of the for loop, but everything I try does not work. power() is the array I'm working on. It is power spectrum data.

The squaring and the summing 33 times is just taking way too long for what I'm doing. (I'm computing the energy for 33 bands of a power spectrum, but I'm doing it thousands of times.)

Thanks.

bogfrog <jmcgraw@rcn.com> wrote in message
<1845647.1216798890667.JavaMail.jakarta@nitrogen.mathforum.org>...
> Hello,
>
> I really need to optimize this block of code:
>
> for(i = 1:33)
> E(frame,i)=sum(power(samp_borders(i):samp_borders(i+1)).^2);
> end;
>
> I wanted to use array indexing instead of the for loop,
but everything I try does not work. power() is the array I'm
working on. It is power spectrum data.
>
> The squaring and the summing 33 times is just taking way
too long for what I'm doing. (I'm computing the energy for
33 bands of a power spectrum, but I'm doing it thousands of
times.)
>
> Thanks.

okay. So you want to have the squared sum of each two
consecutive elements in your array samp_borders.

I'd suggest the following:

squared_samp_borders=samp_borders.^2;
squared_samp_borders_2=zeros(max(size(samp_borders))+1);
squared_samp_borders_2(2:end)=squared_samp_borders;

E(frame)= squared_samp_borders_2+squared_samp_borders;

would that work for you? I didn't try it, but it should be a
*lot* faster.

Note that the squaring is done only once. Alternatively, you
could make a shifted vector, and then square and sum both,
which would lead to a loss of time.

Cheers,

Gavrilo

bogfrog <jmcgraw@rcn.com> writes:

> I really need to optimize this block of code:
>
> for(i = 1:33)
> E(frame,i)=sum(power(samp_borders(i):samp_borders(i+1)).^2); end;
>
> I wanted to use array indexing instead of the for loop, but everything
> I try does not work. power() is the array I'm working on. It is power
> spectrum data.
>
> The squaring and the summing 33 times is just taking way too long for
> what I'm doing. (I'm computing the energy for 33 bands of a power
> spectrum, but I'm doing it thousands of times.)

Is there anything special about samp_borders? Are they evenly spaced?
Same number of samples per band? Just arbitrary? Are they static for
the whole program, or do they change?

If they're static, you could set up a matrix to multiply to do the
summation. Initialize that matrix like:

T = zeros(samp_borders(end), 33);

for i=1:33
  T(samp_borders(i):samp_borders(i+1), i) = 1;
end

And don't touch it again.

Then each of the 1000 loops, square "power" in one shot, and multiply by
T.

E(frame,:) = T * power.^2;

Have you preallocated E before the loop starts?

-Peter

> Is there anything special about samp_borders? Are
> they evenly spaced?
> Same number of samples per band? Just arbitrary?
> Are they static for
> the whole program, or do they change?

They are not evenly spaced, but they are logarithmically spaced. (so the number of samples per band is not the same) And they are static for the whole program.

 
> T = zeros(samp_borders(end), 33);
>
> for i=1:33
> T(samp_borders(i):samp_borders(i+1), i) = 1;
> end
>
> And don't touch it again.
>
> Then each of the 1000 loops, square "power" in one
> shot, and multiply by
> T.
>
> E(frame,:) = T * power.^2;

This seemed like a great idea. I was excited to try it, but it didn't speed things up at all! I'm pretty shocked. :(

 
> Have you preallocated E before the loop starts?

Yes. Otherwise it literally takes an eternity. lol