Intel MKL CSR -> BSR examples in C

[Filipe_Oliveira]

Hi there,

I am having some memory problems regarding the conversion of a CSR matrix to BSR. 

Can anyone tell me some links/books/examples regarding that MKL functionality?

Thank you in advance,

Filipe Oliveira

[mecej4]

In the CSR representation, only the non-zero elements are stored (exception: some routines may require that zero-value elements on the main diagonal should be stored).  In the BSR representation, only non-zero blocks are stored.

Thus, if the block-size = 1, the two representations are similar in storage requirements. When the block-size is not equal to 1, on the other hand, if there is at least one non-zero element in any block, all the elements in that block need to be stored (block-wise, it is "store all or none" rule that has to be followed).

As the block-size is increased, the number of within-block zeros that have to be kept will increase. Finally, when the block-size becomes equal to the matrix size, n, the entire n² elements of the n X n full matrix will have to be kept even in the BSR representation.

I hope that you see now why the memory footprint of a BSR matrix increases when the block-size is increased.

[Gennady_F_Intel]

pls have a look at the <mkldirectory>\examples\spblasc\source\dconverters.c  example.   

//* TASK 5    Obtain block sparse row matrix from compressed sparse row matrix

[Filipe_Oliveira]

hi there,

thank you for the quick reply. is there any limitation regarding the matrix dimension in CSR format? does it need to squared in order to be convertible to BSR? 

i am also getting an strange output of the mkl_scsrbsr(...) conversion. the float *absr array is returning empty even with the correct job config:

job[0] = 0;  //If job[0]=0, the matrix in the CSR format is converted to the BSR format;

168   job[1] = 0;  //If job[1]=0, zero-based indexing for the matrix in CSR format is used;

169   job[2] = 0;  //If job[2]=0, zero-based indexing for the matrix in the BSR format is used;

170   job[3] = 0;  //

171   job[4] = 0;  //

172   job[5] = 1;  //If job[5]>0, all output arrays absr, jab, and iab are filled in for the

regards,

Filipe Oliveira

[Alexander_K_Intel2 · Accepted solution]

Hi,

Have you try csr2bsr converter via new Inspector-executor SparseBlas API? The pseusocode for this converter could be represented by following scheme:

    sparse_status_t status;
    sparse_matrix_t  A_csr, A_bsr;

    status = mkl_sparse_d_create_csr ( &A_csr, SPARSE_INDEX_BASE_ZERO, m, n, ia, ia+1, ja, val );

    status = mkl_sparse_convert_bsr ( A_csr,   mblk, SPARSE_LAYOUT_ROW_MAJOR, SPARSE_OPERATION_NON_TRANSPOSE, &A_bsr );

    MKL_INT rows=0, cols=0, *ib, *rows_end, *jb, block_size;

    sparse_index_base_t    indexing; 
    sparse_layout_t        block_layout;
    double *b;

    status = mkl_sparse_d_export_bsr ( A_bsr, &indexing, &block_layout, &rows, &cols, &block_size, &ib, &rows_end, &jb, &b);

Thanks,

Alex

[Filipe_Oliveira]

Alexander Kalinkin (Intel) wrote:

Hi,

Have you try csr2bsr converter via new Inspector-executor SparseBlas API? The pseusocode for this converter could be represented by following scheme:

    sparse_status_t status;
    sparse_matrix_t  A_csr, A_bsr;

    status = mkl_sparse_d_create_csr ( &A_csr, SPARSE_INDEX_BASE_ZERO, m, n, ia, ia+1, ja, val );

    status = mkl_sparse_convert_bsr ( A_csr,   mblk, SPARSE_LAYOUT_ROW_MAJOR, SPARSE_OPERATION_NON_TRANSPOSE, &A_bsr );

    MKL_INT rows=0, cols=0, *ib, *rows_end, *jb, block_size;

    sparse_index_base_t    indexing; 
    sparse_layout_t        block_layout;
    double *b;

    status = mkl_sparse_d_export_bsr ( A_bsr, &indexing, &block_layout, &rows, &cols, &block_size, &ib, &rows_end, &jb, &b);

Thanks,

Alex

hi there, 

is there any relation between mblk and possible memory leaks or incorrect memory acesses? because depending on the block size i get sometimes segmentation fault errors in mkl_sparse_convert_bsr. 

Also, is there any mkl debugging tool  in OS X?

thank you in advance,

[Alexander_K_Intel2]

Hi,

mblk have to be divider of nrows and ncols. In this case could I ask you to send reproducer of this issue to me to understand reason of memory leaks/segfault?

Thanks,

Alex 

[Filipe_Oliveira]

Alexander Kalinkin (Intel) wrote:

mblk have to be divider of nrows and ncols. 

could you please rephrase this part?

If our matrix is  M rows by N columns, our mblk is n*m, N = k*n (k integer), and M = j*m (j integer), does k need to have the same value as j? 

Regards,

Filipe Oliveira and Sérgio Caldas