OpenMP problems in .onLoad

[aitap]

Hello!

The current omp_thread_count() tries to start as many OpenMP threads as feasible and then count them one by one:

rsparse/src/utils.cpp

Lines 86 to 89 in 695d4eb

	#ifdef _OPENMP
	#pragma omp parallel reduction(+ : n)
	#endif
	n += 1;

This results in a NOTE on CRAN and has caused problems for at least two of the reverse dependencies of rsparse (tidymodels/textrecipes#251, recent R-pkg-devel thread) when the threads started by rsparse:::.onLoad caused more than 250% average CPU usage.

I think you can obtain the desired "default" number of threads without starting a parallel group, by choosing the minimal value between omp_get_max_threads() (which is documented to take OMP_NUM_THREADS into account if set) and omp_get_thread_limit() (which takes into account the additional OMP_THREAD_LIMIT).

Additionally, it would be beneficial to only set options(rsparse_omp_threads=...) when it is not already set:

rsparse/R/zzz.R

Line 20 in 695d4eb

options("rsparse_omp_threads" = n_omp_threads)

This will let your users set the option in their .Rprofile without having to load your package first.

[dselivanov]

Thanks for report, all looks valid. PR would be great. I honestly already don't remember why implemented own omp_thread_count instead of relying on R facilities. My guess is that R doesn't always count it correctly.