Questions about "__c_mset16_sky" from the profiling summary

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

Hi,

I am making some profiling on a NVIDIA Tesla T4 GPU, using compiler and profiler from NVHPC 21.3.

As you can see from the part of the profiling summary below, I have a couple of __c_mset16_sky which I have no idea what they are. I would like to know what __c_mset16_sky means and how to trace it back in the source code.

  2.24%   298.4ms  | |     | rhs_m_add_flux_contribution_ (/home/x_ql/ess_openacc/src/rhs_m.F90:255 0xa31c)
  2.24%   298.4ms  | |     | | hyperviscosity_m_hyper_viscosity_ (/home/x_ql/ess_openacc/src/hyperViscosity_m.F90:0 0x156f)
  0.62%  82.316ms  | |     | |   hyperviscosity_m_traditionalsbp_dissipation_ (/home/x_ql/ess_openacc/src/hyperViscosity_m.F90:110 0xc83)
  0.39%  51.448ms  | |     | |   | __c_mset16_sky (0xcf02a157)
  0.15%  20.579ms  | |     | |   | sbp_operators_m_diff_xi_all_ (/home/x_ql/ess_openacc/src/sbp_operators_m.F90:892 0x68e)
  0.15%  20.579ms  | |     | |   | | sbp_operators_m_diff_xi_one_ (/home/x_ql/ess_openacc/src/sbp_operators_m.F90:736 0x41f)
  0.15%  20.579ms  | |     | |   | |   sbp_operators_m_dcsrmv_acc_ (/home/x_ql/ess_openacc/src/sbp_operators_m.F90:5047 0x4fb)
  0.15%  20.579ms  | |     | |   | |     __pgi_uacc_computedone (../../src/computeexitdone.c:90 0xd161002b)
  0.15%  20.579ms  | |     | |   | |       __pgi_uacc_computedone2 (../../src/computeexitdone.c:59 0xd1610196)
  0.15%  20.579ms  | |     | |   | |         __pgi_uacc_cuda_wait (../../src/cuda_wait.c:77 0xd119644c)
  0.15%  20.579ms  | |     | |   | |           cuStreamSynchronize (0xce4fc25f)
  0.08%   10.29ms  | |     | |   | __c_mset16_sky (0xcf02a168)
  0.54%  72.027ms  | |     | |   hyperviscosity_m_traditionalsbp_dissipation_ (/home/x_ql/ess_openacc/src/hyperViscosity_m.F90:111 0xca7)
  0.23%  30.869ms  | |     | |   | __c_mset16_sky (0xcf02a157)
  0.08%   10.29ms  | |     | |   | __c_mset16_sky (0xcf02a171)

Thanks in advance for your help.
Best
qiang

2

These look like host side memset calls. The compiler’s idiom recognition will often replace bulk copy operations (like array syntax “Array=value”) with calls to memset since they are more efficient than generating an implicit loop.

3

Thanks Mat.