topic Using MKL with MKL_NUM_THREADS=1 in Intel® oneAPI Math Kernel Library

Using MKL with MKL_NUM_THREADS=1

DmitrySkachkov — Mon, 30 Dec 2024 01:13:40 GMT

Hello,

I am trying to use MKL to calculate on one core only, but within the cycle working in OpenMP parallel regime.

Program A !OMP PARALLEL DO do k=1,N2 ... call getrf(A,ipiv) ! MKL subroutine to work on one core call getrs(A,ipiv,B) ! MKL subroutine to work on one core ... enddo !OMP END PARALLEL DO

In job submission script I set

export OMP_NUM_THREADS=2

export MKL_NUM_THREADS=1

However this gives an error in MKL subroutines!

When I submit the job on one core only (OMP_NUM_THREADS=1) the MKL subroutines work correctly.

How to ask MKL library to calculate on one core correctly when the program runs in OpenMP parallel regime?

Re: Using MKL with MKL_NUM_THREADS=1

Ruqiu_C_Intel — Mon, 30 Dec 2024 01:44:41 GMT

For using oneMKL + OpenMP threading, we recommend you use intel OpenMP -liomp5 link option. For example,

${MKLROOT}/lib/libmkl_blas95_ilp64.a ${MKLROOT}/lib/libmkl_lapack95_ilp64.a -Wl,--start-group ${MKLROOT}/lib/libmkl_intel_ilp64.a ${MKLROOT}/lib/libmkl_intel_thread.a ${MKLROOT}/lib/libmkl_core.a -Wl,--end-group -liomp5 -lpthread -lm -ldl

More advisor, please check onemkl-link-line-advisor

Re: Using MKL with MKL_NUM_THREADS=1

DmitrySkachkov — Mon, 30 Dec 2024 03:23:45 GMT

Thank you!

With this options MKL works as necessary, in OpenMP regime MKL subroutine runs on one core without errors.