Multithreading lapack numeric differences

Multithreading lapack numeric differences

Hi,

I'm having different outputs with zgbsv LAPACK subroutine depending on the number of threads that are being used. My MKL dll version is 10.3.10.1. This numeric differences are in the order of the 10th decimal number, but they exist. Is this normal?

Thanks in advance,

Aurora

4 posts / 0 nouveau(x)
Dernière contribution
Reportez-vous à notre Notice d'optimisation pour plus d'informations sur les choix et l'optimisation des performances dans les produits logiciels Intel.

Hi,
yes, this is the expected behaviour with floating point operations. for your information, the latest version 11.0 introduces new functionality allowing users to have reproducibility in floating point calculations. Please look at this material regarding this topic:
http://software.intel.com/en-us/forums/topic/313020

--Gennady

Still, using mkl_cbwr_set(MKL_CBWR_COMPATIBLE) before LAPACK, it returns me different results with 3 or 4 threads for example. Am I using it fine?

MKL's CNR guarantees the bit to bit output results only if with the same #of threads.
Please participate this survey ( https://softwareproductsurvey.intel.com/survey/150072/1afd/ )
to influence how this feature will be extended in the future.

Laisser un commentaire

Veuillez ouvrir une session pour ajouter un commentaire. Pas encore membre ? Rejoignez-nous dès aujourd’hui