Hi.
I am using Eigen v3.2 along with MKL 11.1 and I see have some (run-to-run) numerical reproducibility problems.
My program in general uses dynamic (dense and sparse) Eigen matrices for initialization and very basic matrix operations (addition, multiplication etc).
I use MKL routines for other complex things (e.g. diagonilization).

According to this https://software.intel.com/en-us/articl ... bility-cnr
MKL requires 16, 32 or 64bit memory alignment. Is Eigen compatible with that or do I have to use mkl_malloc as Intel recommends?

Anyone dealt with run-to-run reproducibility issues on Eigen?

Thank you

Dimitris

(Note that the only pre-processor directive I use is EIGEN_DEFAULT_DENSE_INDEX_TYPE = long long int)

Eigen 3.2 ensures 16 bytes alignement . With the devel branch, you can compile with -mavx to enable avx and 32 bytes alignment.

Thank you ggael,

It seems that my reproducibility issues come from other parts of my code.
Plus, I just found out that MKL 11.1 or later does not
require data alignment: http://sc13.supercomputing.org/sites/de ... p140s1.pdf