Code runs slower when openmp is used

[JereKnuutinen]

For some reason, the code gets slower when OpenMP is used together with the release build compared to case when I just comment target_link_libraries(example PUBLIC OpenMP::OpenMP_CXX) in my cmakelist. Is there some reason for this?

[ddemidov]

There is probably too little work to hide openmp overhead. The only place in the library OMP is used is this:

mba/mba/mba.hpp

Line 366 in 9091b67

#pragma omp parallel

, so you could comment it out if that makes it faster in your case.

[JereKnuutinen]

Is there need for commenting since I think that compilers are suppose to ignore pragmas they do not understand?

[ddemidov]

I thought you want to keep using openmp in your outside project, but remove it from the mba code

[JereKnuutinen]

I also noticed that the MBA does not give always same result if OpenMP is used in compilation. How that can be possible? If I do not use openMP or If I use older commits where openMP is not utilized this does not happen.

[ddemidov]

The only reason I can think of is that when using openmp the order of summation is unpredictable here:

mba/mba/mba.hpp

Lines 414 to 419 in 3b296b2

	for(ptrdiff_t i = 0; i < m; ++i) {
	#pragma omp atomic
	delta[i] += t_delta[i];
	#pragma omp atomic
	omega[i] += t_omega[i];
	}

This could produce slightly different results with floating point arithmetic.