xsyrk vs xxsyrk vs dsyrk (in netlib LAPACK)

[ajay-fuji]

Hi,

In the mkl function call, I have seen two types of call of same basic function, prefixed with x and xx .e.g, xsyrk and xxsyrk.
What is difference between these two type of functions and how is it different from function implementation in netlib LAPACK or OpenBLAS.

[Alexsandruss]

Reading MklLapack code you can find xx functions set number of threads to 1 locally:

template <CpuType cpu>
struct MklLapack<double, cpu>
{
    typedef DAAL_INT SizeType;

    static void xgetrf(DAAL_INT * m, DAAL_INT * n, double * a, DAAL_INT * lda, DAAL_INT * ipiv, DAAL_INT * info)
    {
        __DAAL_MKLFN_CALL(lapack_, dgetrf, (m, n, a, lda, ipiv, info));
    }

    static void xxgetrf(DAAL_INT * m, DAAL_INT * n, double * a, DAAL_INT * lda, DAAL_INT * ipiv, DAAL_INT * info)
    {
        int old_threads = fpk_serv_set_num_threads_local(1);
        __DAAL_MKLFN_CALL(lapack_, dgetrf, (m, n, a, lda, ipiv, info));
        fpk_serv_set_num_threads_local(old_threads);
    }
...
}

[ajay-fuji]

Thankyou @Alexsandruss for your response.
I still have two questions,

1. Why xx.. version of function set number of local thread to 1. What are the cases when this function will be used?
2. I can see xxgemm function call

   oneDAL/cpp/daal/src/externals/service_blas_mkl.h

   Line 140 in 23786e6

   static void xxgemm(const char * transa, const char * transb, const DAAL_INT * p, const DAAL_INT * ny, const DAAL_INT * n, const double * alpha,

   doesn't set thread number to 1 but instead call a different function xdgemm. Please comment on this.

Thanks!

[Braza]

@ajay-fuji is this still a pressing issue?

[ajay-fuji]

I still want to understand reason behind setting local threads to 1 in one case (xx) not in other case.

[icfaust]

@ajay-fuji Sorry for answering this almost a year later:

It looks like to me that in cases that routines are already multithreaded, they want to limit the number of threads taken by the blas/lapack calls to match the thread that called it. An example is: in svd there is
https://github.com/oneapi-src/oneDAL/blob/main/cpp/daal/src/algorithms/svd/svd_dense_default_batch_impl.i#L347
which in the multithread calls compute_QR_on_one_node_seq:
https://github.com/oneapi-src/oneDAL/blob/main/cpp/daal/src/algorithms/qr/qr_dense_default_impl.i#L110
which uses xx* calls.