Make ICOPY_OPERATION threaded?

[ChipKerchner]

In some multi-threaded versions of OpenBLAS, the ICOPY_OPERATION (packing for A) seems to take 6-13X more time than the actual KERNEL_OPERATION(s). Is it possible to make this operation multi-threaded?

    /* Copy local region of A into workspace */
    START_RPCC();
    ICOPY_OPERATION(min_l, min_i, a, lda, ls, m_from, sa);
    STOP_RPCC(copy_A);

  48.09%  pt_main_thread  libomp.so                                               [.] bool __kmp_wait_template<kmp_flag_64<false, true>, true, false, true>(kmp_info*, kmp_flag_64<false, true>*, void*)
  18.23%  pt_main_thread  libgomp.so.1.0.0                                        [.] do_wait
   9.01%  pt_main_thread  libopenblasp-r0.3.27.dev.so                             [.] sbgemm_incopy_POWER10
   2.85%  pt_main_thread  libopenblasp-r0.3.27.dev.so                             [.] sbgemm_kernel_POWER10
   1.33%  pt_main_thread  libopenblasp-r0.3.27.dev.so                             [.] inner_thread

[martin-frbg]

I don't think anybody has tried yet, but there is probably no fundamental argument against doing it (apart from the fact that the copy operation itself is already called from a multithreaded region). The least invasive way of trying would probably be to add the blas_level1_thread() mechanism to an existing ICOPY kernel (similar to how some of the DOT
kernels are parallelized on arm64 and x86_64 targets). But I must admit I have not really thought this through, just a first impression from me. I think we already have an open issue about low performance of ICOPY, though that may have been more concerned with the use of outdated instructions on modern targets.