1
$\begingroup$

I would like to use MKL to solve a sparse linear system. I chose the DSS (Direct Sparse Solver) interface, which implements the following steps:

//(1).define the non-zero structure of the matrix
dss_define_structure(handle, sym, rowIndex, nRows, nCols, columns, nNonZeros);
//(2).reorder the matrix
dss_reorder(handle, opt, 0);
//(3).factor the matrix
dss_factor_real(handle, type, values);
//(4).get the solution vector
dss_solve_real(handle, opt, rhs, nRhs, solValues);
//(5).deallocate solver storage
dss_delete(handle, opt);

According to my test, the DSS uses a column-major ordering. That means

//column = 3
{ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 }

is equivalent to

{{1,  6,  11},
 {2,  7,  12},
 {3,  8,  13},
 {4,  9,  14}
 {5, 10,  15}}

For instance, there is a sparase linear system $\mathbf A_{5\times 5}\mathbf X_{5\times 3}=\mathbf B_{5\times 3}$

where $\mathbf A$ is symmetric sparse array

enter image description here

//A stored with CSR3 format
rowIndex = { 0, 5, 6, 7, 8, 9 };
columns = { 0, 1, 2, 3, 4, 1, 2, 3, 4 };
values = { 9, 1.5, 6, .75, 3, 0.5, 12, .625, 16 };
//B
rhs[5*3] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}
//X
solValues[15]

By calling the above DSS interface, the solValues is

{-326.333, 983, 163.417, 398, 61.5, 
 -844.667, 2548, 423, 1028, 159, 
-1363, 4113, 682.583, 1658, 256.5}

Namely

$$ \mathbf X_{5\times 3}=\left( \begin{array}{ccc} -326.333 & -844.667 & -1363 \\ 983 & 2548 & 4113 \\ 163.417 & 423 & 682.583 \\ 398 & 1028 & 1658 \\ 61.5 & 159 & 256.5 \\ \end{array} \right) $$

In my application, the matrix is row-major. How to deal with this problem?

Improve this question
$\endgroup$
4
  • 2
    $\begingroup$ The example you show is a dense matrix. The DSS (obviously) deals with the sparse matrix format (shown here: software.intel.com/en-us/node/…). How are you storing the sparse matrix in your software? $\endgroup$
    – Bill Greene
    Mar 21, 2017 at 12:19
  • $\begingroup$ @BillGreene In my application, I also stores the sparse matrix with CSR3 format via zero-based indexing. $\endgroup$
    – xyz
    Mar 22, 2017 at 6:45
  • $\begingroup$ @BillGreene In addition, for a linear system $\mathbf A_{n \times n}\mathbf X_{n \times c}=\mathbf B_{n \times c}$ , the storing of $\mathbf X$ and $\mathbf B $ are in row-major ordering $\endgroup$
    – xyz
    Mar 22, 2017 at 6:51
  • $\begingroup$ According to the DSS documentation page I referenced, it requires CSR format. So you should be all set. $\endgroup$
    – Bill Greene
    Mar 22, 2017 at 11:36

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.