Critical sections (OpenMP)

In POSIX threads, mutual exclusion is explicit: you call pthread_mutex_lock before a shared update and pthread_mutex_unlock after.

Critical sections in OpenMP do the same thing with a pragma. The enclosed block is protected by an implicit mutex: only one thread at a time executes it, and threads that arrive while another is inside are blocked until it exits. Unlike `atomic`, which is limited to a single read-modify-write operation, critical can protect an arbitrary sequence of statements.

#pragma omp parallel for
for (int i = 0; i < N; i++) {
    int val = expensive_compute(i);
    #pragma omp critical
    {
        if (val > best) {
            best = val;
            best_idx = i;
        }
    }
}

Critical sections can be named (#pragma omp critical(name)) so that threads only contend with sections sharing the same name. Two differently-named critical sections can be entered simultaneously by different threads. Unnamed sections all share a single implicit lock, so mixing named and unnamed sections in the same program can cause unexpected serialisation.