topic Re: Gripe about DSS interface to PARDISO in Intel® oneAPI Math Kernel Library

Gripe about DSS interface to PARDISO

Dishaw__Jim — Sun, 03 Jun 2007 15:53:17 GMT

This is a problem that I have encountered with the DSS interface on MKL 8.0, 8.1, and the 9 beta.

I had set the values for the rowIndex and columns arrays, however, due to a bug one of the column positions for a few of the rows was wrong. The dss_define_structure and dss_reorder both execute without error. The dss_factor_real call results in an access violation--somewhere in _blklu_unsym_risc_pardiso() + 0x2bbb.

Catching the error before the access violation and generating a more descriptive error message would be handy--it might even be detectable in dss_reorder. Because bounds checking does have a performance impact, so it could optionally be enabled when calling dss_create

Re: Gripe about DSS interface to PARDISO

Intel_C_Intel — Tue, 05 Jun 2007 16:57:32 GMT

It would be valuable if you would submit an issue at http://primier.intel.com. This looks like it may be a bug and should be reported as such to assure it gets the attention it deserves.

Bruce

Re: Gripe about DSS interface to PARDISO

Dishaw__Jim — Wed, 13 Jun 2007 04:32:19 GMT

I tried to make a test case that has the problem and it does not generate the access violation.

What is is odd, when I reintroduce the bug back into my code, I get an access violation at a different location. The error occurs at the dss_reorder() instruction at the "di.dll!_inv_perm_mod_mod_pardiso() + 0xfc6" line.

So, I decided to investigate further. If I set the number of threads via omp_set_num_threads() to more than 1 on a single processor machine I get an access violation.

I am sending the example to Premier support--hopefully the bug will manifest itself.

Re: Gripe about DSS interface to PARDISO

Intel_C_Intel — Mon, 18 Jun 2007 15:22:09 GMT

Thanks for taking the time to submit the issue through your Premier account.

Bruce