topic cblas_sgemv_batch_strided threading problem in Intel® oneAPI Math Kernel Library

cblas_sgemv_batch_strided threading problem

MicPac — Thu, 15 Feb 2024 17:45:20 GMT

Hi,

with intel oneApi 2023.0:

I try to force cblas_sgemv_batch_strided to run on ONE thread.

I set :

mkl_set_num_threads_local( 1 )

or
mkl_domain_set_num_threads(1, MKL_DOMAIN_BLAS)

mkl_domain_set_num_threads(1, MKL_DOMAIN_ALL);

or both

but I can see a lot of tbb thread in all cases

I can see one thread only if I set MKL_THREADING sequential before my cmake find_package

but it's not possible in my use case

Is it a bug of that gemv function ?

With my best regards,

Michel

SofeaAzrin_A_Intel — Thu, 07 Mar 2024 06:27:23 GMT

Based on your previous comments, you are using the openMP related routine.

mkl_set_num_threads_local( 1 )

or mkl_domain_set_num_threads(1, MKL_DOMAIN_BLAS)

Thus, you need to link with openMP but not with TBB. For further details and understanding, you can refer to our user guide or get started documents.

SofeaAzrin_A_Intel — Thu, 14 Mar 2024 06:14:56 GMT

Hi,

We recommend you to look in TBB runtime design. The function that you mentioned is not possible to do.

Thanks.