Sparse matrix coordinate to CSC conversion

[Petros]

Hi,

I have a matrix in COO format and want to convert to CSC to be used with KLU solver from SuiteSparse. MKL comes with several converters, but I don't seem to find a mkl_?CSCCOO. Is the only way to do this is through :

mkl_dCSRCOO(...)
mkl_dCSRCSC(...)

Any faster way?

Thanks in advance, Petros.

[mecej4 · Accepted solution]

You may be able to skip the conversion altogether if the solver allows you to solve A^Tx = b in the solution phase, because the CSC representation of A is identical to the CSR representation of A^T.

You can use a COO ⇒ CSR converter to do a COO ⇒ CSC conversion by interchanging the COO row and column index vectors in the call. A couple of points to consider carefully: COO data may be in arbitrary order, whereas CSC and CSR have to be in prescribed order; COO data may contain multiple entries for any matrix position, and these entries need not appear consecutively. Make sure that the conversion routine is capable of sorting the COO data and combining multiple entries into a single entry.

[Petros]

mecej4 wrote:

You may be able to skip the conversion altogether if the solver allows you to solve A^Tx = b in the solution phase, because the CSC representation of A is identical to the CSR representation of A^T.

You can use a COO ⇒ CSR converter to do a COO ⇒ CSC conversion by interchanging the COO row and column index vectors in the call. A couple of points to consider carefully: COO data may be in arbitrary order, whereas CSC and CSR have to be in prescribed order; COO data may contain multiple entries for any matrix position, and these entries need not appear consecutively. Make sure that the conversion routine is capable of sorting the COO data and combining multiple entries into a single entry.

Thanks! I didn't realize that by swapping row and column of the COO I can get the CSC!