/Zi makes MKL slow on Windows.

erling_andersen — Tue, 27 Mar 2018 08:42:03 GMT

I have a Windows application I build with Intel C 18.0.2 that calls MKL extensively. I want to profile it with Vtune.

If I build it with the /Zi optimization i.e. debug info. needed for profiling it seems to get 5 times slower when run as stand alone from the command line. Vtune tells me that much time is spend in dgemm. Could it add a fixed overhead per dgemm call?

How come? Any suggestions?

Erling

This is an unexpected case.

Gennady_F_Intel — Tue, 27 Mar 2018 10:34:23 GMT

This is an unexpected case. We expect to see the same mkl performance with and without debug optimization. You may try to measure mkl's ( dgemm in your case ) function performance

double t1 = dsecnd()

frot( i=0< m; ++i) dgemm();

double t2 = dsecnd()

and make some conclusion

topic This is an unexpected case. in Intel® oneAPI Math Kernel Library

/Zi makes MKL slow on Windows.

This is an unexpected case.