OpenMP atomic memory order

Question

I started learning OpenMP and discovered the #pragma omp atomic directive.

I have basic understanding of C++11's atomics and know that you can pass a memory_order parameter to atomics's method. Correct me if I'm wrong, but I think this allows to use atomics as a synchronization point if using memory_order_seq_cst for example.

Some less restrictive memory order, like memory_order_relaxed just make sure that the operations on the atomic are synchronized and visible to others. It doesnt nothing about other memory updates.

I would like to know what memory order is used by the OpenMP's atomic directive. Will it only synchronize access to the atomic, or will it act as a memory synchronization's point?

My guess would be that it would be more like memory_order_relaxed, because critical's are here to provide a full synchronization.

I welcome any good explanation / informations. Thanks.

Answer 1

The OpenMP memory model has been evolving. Up to and including OpenMP 3.1, the model is based on "flush" operations. These are not really comparable to anything in the C++ memory model. Loosely speaking a flush roughly corresponds to atomic_thread_fence(x), where x is memory_order_seq_cst. But there's a gotcha where it's not that if the "intersection of the two flush-sets of the two flushes is empty".

OpenMP 4.0 adds the seq_cst clause, and answers the question directly:

Note – As with other implicit flush regions, Section 1.4.4 on page 20 reduces the ordering that must be enforced. The intent is that, when the analogous operation exists in C++11 or C11, a sequentially consistent atomic construct has the same semantics as a memory_order_seq_cst atomic operation in C++11/C11. Similarly, a non-sequentially consistent atomic construct has the same semantics as a memory_order_relaxed atomic operation in C++11/C11.

For more details, download the latest version of the spec (4.0 at this time) and read sections 1.4.4 "OpenMP Memory Consistency" and 2.12.6 "atomic Construct".