oneMKL gemm called within kernel

[TestDR]

Hi,

oneMKL gemm function can be called from the  "host" as below

gemm_done = oneapi::mkl::blas::gemm(main_queue, transA, transB, m, n, k, alpha, A, ldA, B, ldB, beta, C, ldC, gemm_dependencies);

can gemm function also be called within user's kernel code? For example,

sycl::queue queue;
queue.submit([&](sycl::handler& cgh) {
    cgh.parallel_for(range,[=](…) {
        oneapi::mkl::blas::gemm(...); // calling routine from user’s kernel code
    });

});

If so, do we need to pass "queue" again as the first argument to oneapi::mkl::blas::gemm(..)?

Thanks.

[PraneethA_Intel]

Hi Rajitha,

Thanks for posting in Intel communities.

By default, oneapi::mkl::blas::gemm can only be used on the host by passing the queue as an input parameter as per the example shown.

Thanks and Regards,

Praneeth Achanta

[Peter_C_Intel]

Hi @TestDR 

     As @PraneethA_Intel mentioned, oneMKL BLAS routines (including oneapi::mkl::blas::gemm) can only be called from host code currently, not from within a kernel.

The good news is that in the next oneAPI major release, we are planning to introduce experimental APIs that would let you call oneMKL GEMM within a kernel. If you're able to provide more details (how large the matrices are, how GEMM fits into the calculation your kernel performs), we could assess if these APIs would be a good fit for your use case.

Best, Peter

[PraneethA_Intel]

Hi Rajitha,

Has the information provided helped? Kindly let us know if you need any other help.

Thanks and Regards,

Praneeth Achanta

[PraneethA_Intel]

Hi Rajitha,

We assume that your issue is resolved. If you need any additional information, please post a new question as this thread will no longer be monitored by Intel.

Thanks and Regards,

Praneeth Achanta