topic I found that using ilp64 is in Intel® oneAPI Math Kernel Library

segmentation fault when mkl_sparse_d_mv for sparse matrix with many nonzeros

qiu__jiezhong — Sun, 01 Sep 2019 01:36:08 GMT

Hi,

I am computing y=Ax (mkl_sparse_d_mv) for a 81,000 x 81,000 sparse matrix A (csr format) with 2.3*10^9 nnzs, but MKL produce segmentation fault.

The sparse matrix is dumped from another program. I first read it from disk, and then call mkl_sparse_d_create_csr to create the matrix. mkl_sparse_d_create_csr returns SPARSE_STATUS_SUCCESS. But when I move to mkl_sparse_d_mv, it shows me segmentation fault.

My another observation is that when I decrease #nnz to be 1.5*10^9, MKL can finish the computation successfully.

My hypothesis is that MKL 2.3*10^9 > 2^31, and MKL may have constraints on #nnz in a sparse matrix.

Any help would be appreciated. Thanks!

Env: Linux, parallel_studio_xe_2019_update4.

I found that using ilp64 is

qiu__jiezhong — Sun, 01 Sep 2019 02:07:14 GMT

I found that using ilp64 is the solution.

https://software.intel.com/en-us/mkl-macos-developer-guide-using-the-ilp64-interface-vs-lp64-interface

Hi!Can you please try ILP64

MariaZh — Tue, 03 Sep 2019 23:52:00 GMT

Hi!
Can you please try ILP64 interface for MKL?
(https://software.intel.com/en-us/mkl-linux-developer-guide-using-the-ilp64-interface-vs-lp64-interface)

I believe it may help in your case, as number of nonzero elements is indeed big.

Best regards,
Maria

Quote:Zhukova, Maria (Intel)

qiu__jiezhong — Mon, 09 Sep 2019 20:03:31 GMT

Zhukova, Maria (Intel) wrote:
Hi!
Can you please try ILP64 interface for MKL?
(https://software.intel.com/en-us/mkl-linux-developer-guide-using-the-ilp...)
I believe it may help in your case, as number of nonzero elements is indeed big.
Best regards,
Maria

Yes, it helps. Thanks a lot!!!