Thread Subject: Return empty sparse from a mex file

Hello!

I'm trying to assemble and return a sparse matrix in a mex-function. At the end of the function, I try to correct for the actual number of non-zero entries in the following way (where Btu = mxGetPr(plhs[0]);):

"nbytes = actual_number_of_nonzeros*sizeof(*Btu);
 newptr = mxRealloc(Btu,nbytes);
 mxSetPr(plhs[0], newptr);
        
 nbytes = actual_number_of_nonzeros*sizeof(*ir);
 newptr = mxRealloc(ir, nbytes);
 mxSetIr(plhs[0], newptr);"

To test my function I have tried , the, excellent, spok-function, see http://www.mathworks.com/matlabcentral/fileexchange/20186-spok-checks-if-a-matlab-sparse-matrix-is-ok, which gives an ok except when the output is empty, i.e. nnz = 0. spok then returns the error message "double, but with no values present". The resulting matrix still seem to work fine but it's a little disturbing to have an error which I don't understand.

My question is thus: What should my function return if actual_number_of_nonzeros = 0?

Best regards
Arne

"Arne Vagren" <arne.vagren@gmail.com> wrote in message <hc3s2q$fbd$1@fred.mathworks.com>...
> Hello!
>
> I'm trying to assemble and return a sparse matrix in a mex-function. At the end of the function, I try to correct for the actual number of non-zero entries in the following way (where Btu = mxGetPr(plhs[0]);):
>
> "nbytes = actual_number_of_nonzeros*sizeof(*Btu);
> newptr = mxRealloc(Btu,nbytes);
> mxSetPr(plhs[0], newptr);
>
> nbytes = actual_number_of_nonzeros*sizeof(*ir);
> newptr = mxRealloc(ir, nbytes);
> mxSetIr(plhs[0], newptr);"
>

There is a possibly problem in Jc array, which you didn't provide information.

Bruno

Thanks for your reply. Actually, in the case nnz = 0, I haven't touched the jc array, it's
simply an array of zeros of size [number of columns + 1]

"Arne Vagren" <arne.vagren@gmail.com> wrote in message <hc40j6$o3d$1@fred.mathworks.com>...
> Thanks for your reply. Actually, in the case nnz = 0, I haven't touched the jc array, it's
> simply an array of zeros of size [number of columns + 1]

Have you changed nzmax to zero accordingly (by using mxSetNzmax)?

Bruno

Actually I just check, it seems nzmax is never set to 0 by Matlab

>> s=sparse([],[],[],1,1,0)
>> nzmax(s)

ans =

     1

So I would Realloc Pr and Ir with *1* element, set nzmax to 1, and fill Jc with zeros (as you did).

Bruno

Ok, I added the line

    if (actual_number_of_nonzeros==0)
        actual_number_of_nonzeros = 1;

before reallocating Pr and Ir. It works fine now, although it seem a little odd to me. Thanks again for your comments!
/Arne

"Bruno Luong" <b.luong@fogale.findmycountry> wrote in message <hc42dq$ke8$1@fred.mathworks.com>...
> Actually I just check, it seems nzmax is never set to 0 by Matlab
>
> >> s=sparse([],[],[],1,1,0)
> >> nzmax(s)
>
> ans =
>
> 1
>
> So I would Realloc Pr and Ir with *1* element, set nzmax to 1, and fill Jc with zeros (as you did).
>
> Bruno

"Arne Vagren" <arne.vagren@gmail.com> wrote in message <hc3s2q$fbd$1@fred.mathworks.com>...
> Hello!
>
> I'm trying to assemble and return a sparse matrix in a mex-function. At the end of the function, I try to correct for the actual number of non-zero entries in the following way (where Btu = mxGetPr(plhs[0]);):
>
> "nbytes = actual_number_of_nonzeros*sizeof(*Btu);
> newptr = mxRealloc(Btu,nbytes);
> mxSetPr(plhs[0], newptr);
>
> nbytes = actual_number_of_nonzeros*sizeof(*ir);
> newptr = mxRealloc(ir, nbytes);
> mxSetIr(plhs[0], newptr);"
>
> To test my function I have tried , the, excellent, spok-function, see http://www.mathworks.com/matlabcentral/fileexchange/20186-spok-checks-if-a-matlab-sparse-matrix-is-ok, which gives an ok except when the output is empty, i.e. nnz = 0. spok then returns the error message "double, but with no values present". The resulting matrix still seem to work fine but it's a little disturbing to have an error which I don't understand.
>
> My question is thus: What should my function return if actual_number_of_nonzeros = 0?
>
> Best regards
> Arne

You are right to be troubled; spok is correctly telling you that you have constructed an invalid matrix. That matrix, with a NULL pointer for mxGetPr, will break MATLAB somewhere. Not everywhere, so it may seem fine, but it is not valid.

nnzmax(A) must never be zero, even if nnz(A) is zero. nnzmax(A)>=1 must always hold.

"Tim Davis" <davis@cise.ufl.edu> wrote in message <hcf36n$83n$1@fred.mathworks.com>...
>
> You are right to be troubled; spok is correctly telling you that you have constructed an invalid matrix. That matrix, with a NULL pointer for mxGetPr, will break MATLAB somewhere. Not everywhere, so it may seem fine, but it is not valid.
>
> nnzmax(A) must never be zero, even if nnz(A) is zero. nnzmax(A)>=1 must always hold.

FYI added info just to complete this thread. There is a difference between full matrices and sparse matrices on this point. If you examine the pr of a full empty matrix, it can be NULL (0), whereas for a sparse matrix it always points to a valid memory location.

James Tursa

Thanks! I guess that's a rather subtle error that could have caused me quite a bit of frustration...