Thread Subject: how to further vectorize ?

Hi,
how this can be further vectorized:
I need to remove from matrix M (example given below) the rows where K-L and I-J are swapped.

Example:
First row gives [K L] = [1 1] and [I J] = [2,2] ---> Row 6 should therefore be removed from M.
Second row gives [K L] = [1 1] and [I J] = [2,3] ---> Row 10 should therefore be removed from M.
Third row gives [K L] = [1 2] and [I J] = [2,3] ---> Row 11 should therefore be removed from M, etc.

I made a program (see below) that does the job, but it is not fully vectorized.

Any suggestion?

Cheers,
v.

Example:
--------
Matrix M:
        K L I J
        1 1 2 2 101.4 128.8 136.7
        1 1 2 3 23.5 29.9 68.6
        1 2 2 3 5.3 22.1 15.4
        1 1 3 3 63.5 80.7 87.6
        1 1 3 4 63.5 80.7 87.6
                                                         
        2 2 1 1 101.3 136.6 128.7
        2 2 1 2 22.8 30.7 95.1
        2 2 3 3 30.7 41.4 42.4
        2 2 3 4 6.6 9.0 22.1
                                                         
        2 3 1 1 23.5 68.5 29.9
        2 3 1 2 5.3 15.4 22.1
        2 3 3 3 7.1 20.8 9.8
                                                         
        3 3 1 1 63.4 87.5 80.6
        3 3 1 2 14.2 19.7 59.5
        3 3 1 3 0.2 0.3 3.0
        3 3 2 2 30.7 42.4 41.4
        3 4 1 1 0.7 2.4 1.4
                                                         
        1 3 6 18 0.0 0.1 0.5
        1 3 7 20 0.0 0.1 0.5
                                                         
        4 5 10 18 0.1 0.3 0.5
       24 3 4 20 0.2 0.4 0.6


Here is my code and I wonder how it can be further vectorized:

for K=1:max(M(:,1));
    IJ=M(M(:,1)==K,3:4);
    if ~isempty(IJ)
        tf = ismember(M(:,1:4),[IJ,K*ones(size(IJ,1),1),M(M(:,1)==K,2)],'rows');
        M(tf,:)=[];
    end
end


RESULT:
--------
   1 1 2 2 101.4 128.8 136.7
   1 1 2 3 23.5 29.9 68.6
   1 2 2 3 5.3 22.1 15.4
   1 1 3 3 63.5 80.7 87.6
   1 1 3 4 63.5 80.7 87.6
   2 2 1 2 22.8 30.7 95.1
   2 2 3 3 30.7 41.4 42.4
   2 2 3 4 6.6 9.0 22.1
   2 3 3 3 7.1 20.8 9.8
   3 3 1 2 14.2 19.7 59.5
   3 3 1 3 0.2 0.3 3.0
   1 3 6 18 0.0 0.1 0.5
   1 3 7 20 0.0 0.1 0.5
   4 5 10 18 0.1 0.3 0.5
  24 3 4 20 0.2 0.4 0.6

Dear Vladimir!

> I need to remove from matrix M (example given below) the rows where K-L and I-J are swapped.
> Here is my code and I wonder how it can be further vectorized:
>
> for K=1:max(M(:,1));
> IJ=M(M(:,1)==K,3:4);
> if ~isempty(IJ)
> tf = ismember(M(:,1:4),[IJ,K*ones(size(IJ,1),1),M(M(:,1)==K,2)],'rows');
> M(tf,:)=[];
> end
> end

Do you really need a vectorized method or should it be just faster?
This might be a little bit faster:

 for K = 1:max(M(:,1))
     IV = (M(:, 1) == K);
     if any(IV)
         tf = ismember(M(:,1:4), M(IV, [3, 4, 1, 2]), 'rows');
         M(tf, :) = [];
     end
 end

Is ISMEMBER the main problem? How large is the array?

Good night, Jan

Dear Jan,
thanks for your reply. Basically, I wanted to speed up the code, and in most cases a vectorization is the best way to do it. At the moment, it takes about 15s to finish the task, depending on the matrix size (which is about 80,000 rows).

Anyway, thanks for your concerns, your code is indeed a little bit faster than mine :)) !!

Cheers,
v.


"Jan Simon" <matlab.THIS_YEAR@nMINUSsimon.de> wrote in message <hblcdv$20m$1@fred.mathworks.com>...
> Dear Vladimir!
>
> > I need to remove from matrix M (example given below) the rows where K-L and I-J are swapped.
> > Here is my code and I wonder how it can be further vectorized:
> >
> > for K=1:max(M(:,1));
> > IJ=M(M(:,1)==K,3:4);
> > if ~isempty(IJ)
> > tf = ismember(M(:,1:4),[IJ,K*ones(size(IJ,1),1),M(M(:,1)==K,2)],'rows');
> > M(tf,:)=[];
> > end
> > end
>
> Do you really need a vectorized method or should it be just faster?
> This might be a little bit faster:
>
> for K = 1:max(M(:,1))
> IV = (M(:, 1) == K);
> if any(IV)
> tf = ismember(M(:,1:4), M(IV, [3, 4, 1, 2]), 'rows');
> M(tf, :) = [];
> end
> end
>
> Is ISMEMBER the main problem? How large is the array?
>
> Good night, Jan

Dear Vladimir!

> thanks for your reply. Basically, I wanted to speed up the code, and in most cases a vectorization is the best way to do it. At the moment, it takes about 15s to finish the task, depending on the matrix size (which is about 80,000 rows).

If each processing step depends on the previous steps, vectorization is not as straight as for independent steps.

ISMEMBER needs a lot of time for sorting the array again and again. So at first try to pre-sort the array (accelerates the sorting). In addition the inputs are processed with UNIQUE('rows') inside ISMEMBER. If you apply M = UNQIUE(M, 'rows') once, it should be a much faster to copy just the needed lines from ISMEMBER to create your own function avoiding further sorting and unificating.

Good luck, Jan