can support the fp16 or bf16 for cpu with 3.3.6 version?

[feixuedudiao]

can support the fp16 or bf16 for cpu with 3.3.6 version?

[shu1chen]

Hi, that's possibly because bf16/fp16 is not supported on the hardware. The hardware limitations session in oneDNN developer guide lists the minimal supported instruction set architecture (ISA) for the data types. You may find the supported ISA in the specifications of Intel(R) Core(TM) i7-10700 CPU.

[feixuedudiao]

@shu1chen, i have a question, for [Intel(R) Core(TM) i7-10700 CPU], rumtime is how select with omp, tbb and seq? i test the performance with seq. but i feel the performance is not good. If i run the convoluton example for runtime with omp, how can i set the omp_num_thread? thanks.

[shu1chen]

@feixuedudiao oneDNN CPU uses OpenMP threading by default. You may also set through -DONEDNN_CPU_RUNTIME=OMP when building oneDNN (See details).
On Windows, you may use set OMP_NUM_THREADS=4 to set the OMP thread number to 4, see more details in OpenMP Environment Variables.

[feixuedudiao]

@shu1chen i'm to try. thanks.

[feixuedudiao]

@shu1chen There have a question for omp_num_threads = 4, why set it to 4?

[shu1chen]

There have a question for omp_num_threads = 4, why set it to 4?

@feixuedudiao It's an example. You may set the desired number of cores to use

[feixuedudiao]

oh, thanks. But how i can set the omp_nump_thread with the api in code. I find a function with get_num_cores, but don't fin set_num_cores.

[shu1chen]

oh, thanks. But how i can set the omp_nump_thread with the api in code. I find a function with get_num_cores, but don't fin set_num_cores.

It might be omp_set_num_threads that you're looking for. Please see the OpenMP Functions in the doc.

[feixuedudiao]

@shu1chen
The function of omp_set_num_threads(4) is same to "set OMP_NUM_THREADS=4" for onednn?

[shu1chen]

The function of omp_set_num_threads(4) is same to "set OMP_NUM_THREADS=4" for onednn?

Yes, oneDNN CPU relies on OpenMP threading in this case. Both OMP_NUM_THREADS and omp_set_num_threads are from OMP and they are supposed to be the same, which is also shown in the OpenMP Environment Variables link.

[feixuedudiao]

@shu1chen
thanks. I look it as soon.