Thread Subject:

dense to ell sparse format

Dear sirs
I used a simple intuitive script to generate an itpack (ell) sparse format
The simple script is:
[i,j,temp]=find(A);
for k=1:size(A,1)
ii=i(i==k);
jj=j(ii);
tt=temp(ii);
B(k,1:size(tt,1))=tt;
I(k,1:size(tt,1))=jj;
end

it works and is sufficiently fast for little sparse matrices but, of course, is very slow with huge matrices (nnz 1 million or more). Some useful suggestions?

Lpeter

"Piero Lanucara" <lanucara@caspur.it> wrote in message <imckec$nlc$1@fred.mathworks.com>...
> Dear sirs
> I used a simple intuitive script to generate an itpack (ell) sparse format
> The simple script is:
> [i,j,temp]=find(A);
> for k=1:size(A,1)
> ii=i(i==k);
> jj=j(ii);
> tt=temp(ii);
> B(k,1:size(tt,1))=tt;
> I(k,1:size(tt,1))=jj;
> end
>
> it works and is sufficiently fast for little sparse matrices but, of course, is very slow with huge matrices (nnz 1 million or more). Some useful suggestions?
>
> Lpeter

Your code doesn't look right. I guess ii setting should be

ii = i==k;

Bruno

"Bruno Luong" <b.luong@fogale.findmycountry> wrote in message
> Your code doesn't look right. I guess ii setting should be
>
> ii = i==k;
>

Assuming the code is with correction:

% Data
A=rand(5,6);
A(A<0.5)=0;

% Your engine
clear B I
[i,j,temp]=find(A);
for k=1:size(A,1)
    ii=i==k;
    jj=j(ii);
    tt=temp(ii);
    B(k,1:size(tt,1))=tt;
    I(k,1:size(tt,1))=jj;
end
B
I

% My engine
[j,i,temp]=find(A.'); % Better work with A.' so that "i" is sorted
ii = accumarray(i, 1, [], @(x) {1:size(x)});
ii = [ii{:}].';
B = accumarray([i,ii],temp)
I = accumarray([i,ii],j)

% Bruno

"Bruno Luong" <b.luong@fogale.findmycountry> wrote in message <imct21$kua$1@fred.mathworks.com>...
> "Bruno Luong" <b.luong@fogale.findmycountry> wrote in message
> > Your code doesn't look right. I guess ii setting should be
> >
> > ii = i==k;
> >
>
> Assuming the code is with correction:
>
> % Data
> A=rand(5,6);
> A(A<0.5)=0;
>
> % Your engine
> clear B I
> [i,j,temp]=find(A);
> for k=1:size(A,1)
> ii=i==k;
> jj=j(ii);
> tt=temp(ii);
> B(k,1:size(tt,1))=tt;
> I(k,1:size(tt,1))=jj;
> end
> B
> I
>
> % My engine
> [j,i,temp]=find(A.'); % Better work with A.' so that "i" is sorted
> ii = accumarray(i, 1, [], @(x) {1:size(x)});
> ii = [ii{:}].';
> B = accumarray([i,ii],temp)
> I = accumarray([i,ii],j)
>
> % Bruno



that's great!
it works fine wih huge symmetric matrix with something like 120Million of non zeros
the interesting thing is that the ell matrix in output is very compact (1 Million x 471 elements!) end this has nice property for solvers.....
Lpeter
p.s
I think my original release should work..the only difference is that

ii=i(i==k) %it generate a small vector with respect the original size
ii=i==k %it generate a vector of the same dimension of i

"Piero Lanucara" <lanucara@caspur.it> wrote in message <imd01k$hvf$1@fred.mathworks.com>...

> p.s
> I think my original release should work..the only difference is that
>
> ii=i(i==k) %it generate a small vector with respect the original size
> ii=i==k %it generate a vector of the same dimension of i

I disagree. ii=i(i==k) returns an array of repeated number of k, and later spoils jj and tt.

Bruno

"Bruno Luong" <b.luong@fogale.findmycountry> wrote in message <imd0o5$199$1@fred.mathworks.com>...
> "Piero Lanucara" <lanucara@caspur.it> wrote in message <imd01k$hvf$1@fred.mathworks.com>...
>
> > p.s
> > I think my original release should work..the only difference is that
> >
> > ii=i(i==k) %it generate a small vector with respect the original size
> > ii=i==k %it generate a vector of the same dimension of i
>
> I disagree. ii=i(i==k) returns an array of repeated number of k, and later spoils jj and tt.
>
> Bruno

I'm sorry I ckeck and you have right absolutely !

thanks
LPeter