Reducing Atomic Contention

Replace global atomics with thread-local storage for better performance in concurrent loops.