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

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

On Mon, Apr 9, 2018 at 4:39 PM, Peter <list@xxxxxxxxxxxxxxxxx> wrote:

Dear Gael,


Am 09.04.2018 um 16:13 schrieb Gael Guennebaud:

> It's more likely more accuracy in the lowest eigenvalues as we assume that eigenvalues lower than machine_epsilon * largest_eigenvalues are zero. 

actually my lowest eigen values are not zero, but of order one.
They get only shifted by the lowest eigen value such that the spectrum starts artificially at zero.
And since I already see this using  -DEIGEN_USE_BLAS only, it's at least not just the diagonalization.

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.

gael

 

Of course Eigen is tremendously superior to MKL as it runs flawlessly with gmp.
So let me thank you again for this great library.

Best regards,
Peter