Race condition in par_lr_interp.c if matrix is too small?

[craffael]

here the start and end for the current thread are computed:

hypre/src/parcsr_ls/par_lr_interp.c

Lines 1270 to 1274 in aa1f66d

	start = (n_fine / num_threads) * my_thread_num;
	if (my_thread_num == num_threads - 1)
	{ stop = n_fine; }
	else
	{ stop = (n_fine / num_threads) * (my_thread_num + 1); }

But this calculation is questionable when n_fine < num_threads, in which case all threads are running with start=stop=0. I have not studied the whole code, but it definitely leads to a write race condition on the following lines:

hypre/src/parcsr_ls/par_lr_interp.c

Lines 1401 to 1407 in aa1f66d

	#pragma omp barrier
	#endif
	P_diag_i[stop] = jj_counter;
	P_offd_i[stop] = jj_counter_offd;
	fine_to_coarse_offset[my_thread_num] = coarse_counter;
	diag_offset[my_thread_num] = jj_counter;
	offd_offset[my_thread_num] = jj_counter_offd;

(e.g. all threads may write to P_diag_i[0] at the same time)

[victorapm]

@craffael thanks for bringing this up. We are working on a PR to fix this issue