dggev deterministic behaviour

[Konstantinos_K_]

Hello, 

I am using dggev (mkl versions 11.2.1 & 11.2.2) and i notice some non deterministic behaviour concerning eigenvalues near zero and infinite eigenvalues and eigenvectors. I use the parallel mkl version with gomp (not intel omp). I compile my code using gcc 4.4.7. It seems that there are (at least) 2 sets of output that differ in the regions mentioned. I also use mkl allocators to ensure my memory is aligned. It is a rare phainomenon, but still, could you have a look? 

Thanks! 

[TimP]

Did you follow suggestions posted on Intel site? such as

https://software.intel.com/en-us/articles/introduction-to-the-conditional-numerical-reproducibility-cnr

https://software.intel.com/sites/default/files/article/164389/fp-consistency-110212.pdf

If you didn't take precautions against varying data alignments, this might easily happen.

The differences you observe may be well within the uncertainties of your analysis.

[Konstantinos_K_]

I use mkl_malloc mkl_calloc and mkl_realloc for alignment

a call to mkl_cbwr_get(MKL_CBWR_ALL) returns 0

a call to mkl_cbwr_get(MKL_CBWR_BRANCH) returns 1 (= MKL_CBWR_BRANCH_OFF)

should i set it to AUTO?

[Konstantinos_K_]

I used mkl_cbwr_set(MKL_CBWR_AUTO) but did not work.

[Gennady_F_Intel]

what is the diff into these results? can you give us the example to check the problem?

[Konstantinos_K_]

In the attachment you will find the matrices in binary format, a description of the problem and 2 output sets that differ. 

Cheers!

[Gennady_F_Intel]

Konstantinos, we couldn't reproduce the problem on our side. Would you give the test to where you see the problem on your side?