We're seeing DSYEVX in MKL give different results when the A matrix is aligned on a 16-byte boundary or not. A very simple test case is available at https://github.com/tomkcook/dsyevx_bug.
We're using the following product versions:
We've got some more info on this, but before I waste anyone's time with that, I'd appreciate if someone could confirm this is actually a bug an not just me doing something daft!
Thanks for the comment, but I'm fairly confident that's not what's happening here. Even if the lines were being silently truncated, that should cause a compile error as the line-continuation ampersands are at the ends of lines.
Check the discussions on this subject on the MKL forum, and consider whether you want to try the conditional numerical reproducibility options.
If you wish to minimize alignment dependencies in your own compiled code, set -fp:source or equivalent VS property.
Aligned results may be slightly more accurate, with differences indicative of the level of accuracy.
Tom, Could you check the problem with the latest version of MKL 11.3? This version fully supports unaligned input. But, please take into account - the #of threads should be the same from run to run.