GPU concurrency: Weak behaviours and programming assumptions