Re: [eigen] problem with MKL bindings for dense Eigen

[Peter <list@xxxxxxxxxxxxxxxxx>]

Dear Gael,


Am 10.04.2018 um 09:09 schrieb Gael Guennebaud:

> 
> interesting, so that's not related to the diagonalization algorithm. If you suspect FMA, then compile with -mfma to let Eigen use it for its BLAS-like routines.
> 

Interesting, it does make a difference.
I thought that it is sufficient to use -march=native .

The Eigen run without MKL  is  definitely more accurate when compiling '-mfma' compared to the run without.

Comparing the -mfma version with the MKL run, MKL is slightly more accurate in the
first half of the run, but in the long run the Eigen version without MKL but with
-mfma is more accurate, which  is actually the more important regime for me,
as I can improve on the first by increasing a parameter (leading to more CPU time),
but I can't avoid the second.

I'll look into tiḿings later.

Best regards,
Peter