segmentation fault in mkl_sparse_d_export_csr_i4_avx2

Pat__Nikolas — Mon, 01 Jun 2020 15:55:22 GMT

Hi *,

I am trying to use the pardiso direct solver in my code. My sparse matrix was stored in COO format so I am trying to convert it to CSR3 and run the pardiso solver. My problem is that I get the title's segmentation fault. I checked the input/output parameters and I think that I am not making any mistakes. Here is the code:

#include "mkl.h"
#include "mkl_spblas.h"
#include <iostream>

/* Internal solver memory pointer pt, */
/* 32-bit: int pt[64]; 64-bit: long int pt[64] */
/* or void *pt[64] should be OK on both architectures */
void *pt[64];
MKL_INT mtype = 11; // Real and nonsymmetric matrix
MKL_INT iparm[64];

sparse_matrix_t A_imkl_coo = 0;
sparse_matrix_t A_imkl_csr = 0;
MKL_INT *A_mkl_csr_ia = NULL;
MKL_INT *col_indx = NULL;
double *values = NULL;

// This one is in the initialization part:
for ( int i = 0; i < 64; ++i )
	pt = 0;
iparm[0] = 0;
pardisoinit( pt, &mtype, iparm );

=======================================

// These are in the computation loop

// Create a sparse MKL object in COO format
sparse_status_t ret_err;
if( A_imkl_coo != NULL )
        ret_err = mkl_sparse_destroy(A_imkl_coo); // I am running this in a loop so I do not want any mem leaks
ret_err = mkl_sparse_d_create_coo( &A_imkl_coo, SPARSE_INDEX_BASE_ONE, 2 * NN, 2 * NN, 4 * NNZ, IT, JT, T);
if( ret_err == SPARSE_STATUS_SUCCESS )
        printf( "Successful MKL sparse matrix operation: %s : %d\n", __FILE__, __LINE__);
else
        abort();

// Create a CSR handle from the COO
const sparse_operation_t operation = SPARSE_OPERATION_NON_TRANSPOSE;
ret_err = mkl_sparse_convert_csr( A_imkl_coo, operation, &A_imkl_csr);
if( A_imkl_csr != NULL )
        ret_err = mkl_sparse_destroy(A_imkl_csr);
if( ret_err == SPARSE_STATUS_SUCCESS )
        printf( "Successfully converted internal MKL sparse matrix representation from coo to csr: %s : %d\n", __FILE__, __LINE__);

// Now export the CSR data to be used in the pardiso solver
sparse_index_base_t indexing = SPARSE_INDEX_BASE_ZERO;
MKL_INT rows;
MKL_INT cols;
MKL_INT *rows_start = NULL;
MKL_INT *rows_end = NULL;
ret_err = mkl_sparse_d_export_csr(A_imkl_csr, &indexing, &rows, &cols, &rows_start, &rows_end, &col_indx, &values );
if( ret_err == SPARSE_STATUS_SUCCESS )
        printf( "Successfully exported internal MKL CSR representation: %s : %d\n", __FILE__, __LINE__);
else
        abort();
if( indexing == SPARSE_INDEX_BASE_ONE )
        printf("detected indexing base one\n");
else
        printf("detected indexing base zero\n");

// Pardiso solver expects CSR3 format so "convert" that one as well
A_mkl_csr_ia = (MKL_INT*)realloc(A_mkl_csr_ia, (rows+1)* sizeof(MKL_INT));
for ( MKL_INT i = 0; i < rows; ++i )
        A_mkl_csr_ia = rows_start;
A_mkl_csr_ia[rows] = rows_end[rows-1];


const MKL_INT maxfct = 1;
const MKL_INT mnum = 1;
const MKL_INT phase = 13;
const MKL_INT n = 2*NN;
const MKL_INT msglvl = 1;
      MKL_INT error;
      MKL_INT perm;
const MKL_INT nrhs = 1;
pardiso (pt, &maxfct, &mnum, &mtype, &phase, &n, values, A_mkl_csr_ia, col_indx, &perm, &nrhs, iparm, &msglvl, node->PHI/*void* b */, x, &error);
if( error != 0 ) {
	printf( "Error in intel mkl sparse linear solver: %d\n", error);
	exit(EXIT_FAILURE);
}

I am using gcc 9.3.0 on debian testing, mkl 2020.0.166 and these are the relevant link options:

-fopenmp -Wl,--start-group /opt/imkl-2020.0.166/compilers_and_libraries_2020.0.166/linux/mkl/lib/intel64/libmkl_gf_lp64.a /opt/imkl-2020.0.166/compilers_and_libraries_2020.0.166/linux/mkl/lib/intel64/libmkl_gnu_thread.a /opt/imkl-2020.0.166/compilers_and_libraries_2020.0.166/linux/mkl/lib/intel64/libmkl_core.a -Wl,--end-group -lgomp -lpthread -lm -ldl -pthread -lm  -lgfortran

I started converting the COO to CSR format on my own, but I would really like to find out what I am doing wrong :)

Thank you - Regards

Nikolas

edit: added compiler/imkl and linking information

Could you give us the

Gennady_F_Intel — Mon, 01 Jun 2020 16:45:04 GMT

Could you give us the standalone example which we could compile and launch for reproducing the problem?

Thank you for the reply.

Pat__Nikolas — Mon, 01 Jun 2020 22:15:00 GMT

Thank you for the reply. Unfortunately it is not that simple to come up with a minimum standalone reproducible code of the original code as it is a typical scientific computing spaghetti code. However, I get the same error when I try with the attached code. The compilation command I used is included in the beginning of the source. I tried running it step by step with gdb but it crashes as soon as it goes into the export_csr function.

Thanks again.

Hi Nikolas,I think the issue

MariaZh — Mon, 01 Jun 2020 23:24:28 GMT

Hi Nikolas,
I think the issue is happening because you destroy the handle for the matrix in CSR format (A_imkl_csr) before actually doing export from this handle.

ret_err = mkl_sparse_convert_csr( A_imkl_coo, operation, &A_imkl_csr);
if( A_imkl_csr != NULL )
ret_err = mkl_sparse_destroy(A_imkl_csr);
...
ret_err = mkl_sparse_d_export_csr(A_imkl_csr, &indexing, &rows, &cols, &rows_start, &rows_end, &col_indx, &values );

Best regards,
Maria

Hi Maria,

Pat__Nikolas — Mon, 01 Jun 2020 23:42:37 GMT

Hi Maria,

You are absolutely correct!

I am really embarrassed to have made such a mistake, but I really thank you for noticing.

Regards,

Nikolas

Hi Nikolas,No worries. Please

MariaZh — Tue, 02 Jun 2020 00:00:03 GMT

Hi Nikolas,
No worries. Please, let us know if you have any questions!

Best regards,
Maria

topic Hi Maria, in Intel® oneAPI Math Kernel Library

segmentation fault in mkl_sparse_d_export_csr_i4_avx2

Could you give us the

Thank you for the reply.

Hi Nikolas,I think the issue

Hi Maria,

Hi Nikolas,No worries. Please