If configured with IntelMKL_FAIR_DISPATCH=ON will try to override Intel-centric processor detection in Intel MKL. Until MKL 2020.1 it was possible to use the undocumented environment variable MKL_DEBUG_CPU_TYPE (e.g. set it to 5 to force the use of AVX2 instruction set), but this variable no longer works. This adds the code from https://www.agner.org/optimize/intel_dispatch_patch.zip to try forcing the use of fair dispatch. Seems to work successfully on the Tinkercliffs cluster with AMD Rome processors.
If configured with
IntelMKL_FAIR_DISPATCH=ON
will try to override Intel-centric processor detection in Intel MKL. Until MKL 2020.1 it was possible to use the undocumented environment variableMKL_DEBUG_CPU_TYPE
(e.g. set it to5
to force the use of AVX2 instruction set), but this variable no longer works. This adds the code from https://www.agner.org/optimize/intel_dispatch_patch.zip to try forcing the use of fair dispatch. Seems to work successfully on the Tinkercliffs cluster with AMD Rome processors.P.S. see also http://public.clu2.fastmail.us.user.fm/icc_cpu_dispatch.html