CULA DSYEV vs MKL DSYEV

[lumat]

Hello,

We've made some tests with the symmetric eigenvalue routines CULA DSYEV and MKL DSYEV. We've noticed the time to compute all eigenvalues and all eigenvectors greatly depends on the matrices when using CULA DSYEV. Whereas the time is pretty stable when using MKL DSYEV on the same matrices.

We guess that it's related to the underlying algorithm which needs more or less steps to achieve convergence. Could you tell us more about the algorithms in CULA DSYEV ? Are they of the same kind of those in MKL DSYEV (tridiagonalization then QR algorithm) ?

Best regards,

Philippe

PS : we've tested,
- CULA R16a on NVIDIA S2050 and MKL 11.0 on XEON 5650
- CULA R16a/R17 on NVIDIA K20 and MKL 11.0 on E5-2670

[john]

That is the method, yes. Give SYEVX a try, maybe.

[lumat]

Thanks a lot for your quick reply

We didn't use DSYEVX because we wanted the worst case i.e. all eigenvalues + all eigenvectors of a real double precision symmetric matrix. And while always faster than MKL, we've noticed the CULA computing time depends on the matrices.

So we wonder if CULA DSYEV is made up of the following routines :
- DSYRDB for tridiagonalization (finite number of steps) ?
- DSTEQR for the QR (QL) algorithm (undefined number of iterations) ?

In such a case, is the varying time in computing the eigenvalues + eigenvectors related to DSTEQR ?

Best regards,

Philippe

[john]

It's SYTRD, but then moves into STEQR as you noted.

[lumat]

Thanks a lot for your helpful insight

Best regards,

Philippe