Please Speed UP This Algoritm!!!

hi i'd like to know if there's is a way to speed up this algoritm where:, lphi is a vector of 120 elements, IN4 is a matrix of 4x8192, aaa is a 3D matrix of 4x8192x120

for k = 1:lphi;

B4(k,:) = sum(IN4.*aaa(:,:,k));

end

B = sum(abs(B4).^2,2);

thanks, i don't know if in c++ it work faster any ideas?

[the cyclist]

It's not clear to me what your loop does. You say that lphi is a vector, but then you loop over k=1:lphi. That seems odd to me, and not really what you want.

[Paolo Rossi]

Sorry, lphi is not a vector is the length of a vector lphi = 120

Answer 1

[Jonathan Sullivan]

B4 = reshape(permute(bsxfun(@tims,IN4,aaa),[3 1 2]),size(aaa,3),[]);
B = sum(abs(B4).^2,2);

[Paolo Rossi]

In your way it takes about 0.3 sec to calculate,

in my way it takes about 0.1 sec

thanks anyway friend

i'd like to have 0.05 sec ... for real time processing, i don't know if in c++ it is more faster

[Sean de Wolski]

I believe, Jonathan meant @times

for the bsxfun call.

[Sean de Wolski]

And it still doesn't give the correct results.

[Sean de Wolski]

Like, I said, this does not produce correct results, nor does it work as written. PERMUTEing a big array will likely never be faster than a well written for-loop.

Answer 2

[Sean de Wolski]

bsxfun is lurking (update: and staying hidden)!

The fastest I could get it was to dynamically preallocate your for-loop.

    for k = 120:-1:1
        B5(k,:) = sum(IN4.*aaa(:,:,k));
    end
    B2 = sum(abs(B5).^2,2);

[Paolo Rossi]

i will study this function

Answer 3

[James Tursa]

Another option you can try using some FEX submissions (slightly better at memory management than your original code):

uninit; % One time only, to built the mex routine
mtimesx; % One time only, to build the mex routine

B4 = uninit(8192,1,lphi); % Fast create variable
for k = 1:lphi;
    B4(:,:,k) = sum(IN4.*aaa(:,:,k)).'; % Contiguous result
end
B = mtimesx(B4,'t',B4,'speedomp'); % Fast nD dot product
B = B(:); % Reshape result to column

You can find uninit and mtimesx here:

http://www.mathworks.com/matlabcentral/fileexchange/31362-uninit-create-an-uninitialized-variable-like-zeros-but-faster

http://www.mathworks.com/matlabcentral/fileexchange/25977-mtimesx-fast-matrix-multiply-with-multi-dimensional-support

Note: This won't give exactly the same answer as the MATLAB code because the last dot product operation is done internally in mtimesx differently from MATLAB.

Q: Is everything real? What is the point of the abs( )?

[Paolo Rossi]

No it's a complex Data, the point in the abs is becouse every elements must be elevated at 2

now i try your code thanks and then i will write the results

[James Tursa]

In that case use 'c' instead of 't' in the mtimesx call. E.g.,

B = mtimesx(B4,'c',B4,'speedomp');

[Paolo Rossi]

Thanks but the bottle neck is in the loop, infact the loop takes 0.1 sec...may be this is the best performance, i will try anyway your functions