Zero diagonals in diagonally dominant matrices generated

[seheracer]

Link:
Lines 179-183, kk_generate_diagonally_dominant_sparse_matrix function,
https://github.com/kokkos/kokkos-kernels/blob/1be37e8aee0ddfee5ad5076f3a049c4c473f5665/src/common/KokkosKernels_IOUtils.hpp

Related Code:

  for(int row=0;row<nrows;row++)
  {
    int varianz = (1.0*rand()/RAND_MAX-0.5)*row_size_variance;
    rowPtr[row+1] = rowPtr[row] + elements_per_row+varianz;
  }

Problem:
According to the last line of the above-given code, kk_generate_diagonally_dominant_sparse_matrix may create matrices with some diagonal entries being zero, becauseelements_per_row+varianz may be equal to 0, so, some rows may end up with all zero entries.

Is this behavior acceptable?

Edit:
Also, the loop in

kokkos-kernels/src/common/KokkosKernels_IOUtils.hpp

Line 191 in 1be37e8

for(SizeType k=rowPtr[row] ;k<rowPtr[row+1] - 1;k++)

does not check if the newly-computed column index already exists in the current row. We want to fix that, too, right?

[ndellingwood]

Is this behavior acceptable?

@seheracer For non-strictly diagonally dominant matrices a row of 0's would allow for a 0 at the diagonal, I don't see any comments in the files you reference to determine if this was intentionally allowed to occur (I doubt it, I'm guessing there would be added options to intentionally generate such a matrix since this would test corner cases in algorithms).

[ndellingwood]

Based on meeting discussion sounds like the path forward is to modify the code to ensure diagonal entries are present, rather than adding structural zeros as place holders for rows-of-zeros.

[brian-kelley]

@ndellingwood I agree with this, GS tests use this and expect the diagonal values to be nonzero.

[seheracer]

Sounds good. I will push the PR today.