topic Tobiaz, in Intel® oneAPI Math Kernel Library

HPL benchmark compilation problem

Tobiasz_J_ — Wed, 12 Aug 2015 08:00:13 GMT

Hello,

I am trying to build from sources the Intel version of HPL benchmark which is included in MKL libraries. I am able to make the compilation using the 'non-offload' version. When I try to build using parameter version=offload (as suggested in the guide) I recieve an error when the libhpl_offload.a library is missing. If the missing file is included there are no errors and compilation finishes successfully.

1. Is it possible to make the compilation without using precompiled libraries like libhpl_offload.a?

2. Is the offload version apropriate for Xeon Haswell CPUs like E5-2680 v3 or E5-2650 v3? I am getting almost 20% better results with the offload version than 'non-offload', is that normal behavior?

Best regards,

Tobiasz Jarosiewicz

Tobiaz,

VipinKumar_E_Intel — Thu, 13 Aug 2015 03:54:06 GMT

Tobiaz,

You may also try our pre-built binaries of linpack which are available here:

https://software.intel.com/en-us/articles/intel-math-kernel-library-linpack-download/

Yes, we recommend using the offload version now for better efficiency, and we will be updating the documentation soon to avoid such confusions.

Vipin

Thank you, Vipin,

Tobiasz_J_ — Wed, 19 Aug 2015 08:58:52 GMT

Thank you, Vipin,

I am aware that the offload version should offer the best efficiency howewer, I would like to know if the compilation is possible without any precompiled libraries. If it is possible then probably it should be possible also to improve efficiency.

Regards,

Tobiasz