I have an algorithm that schedules subtasks between multiple devices to solve main_task. It enqueues kernel to all command queues and attaches an event with a callback.
In this callback I check whether my main_task is done and if it’s not, I enqueue one more subtask. Otherwise I set shared variable is_main_task_done to true to notify other command queues that they should stop enqueueing subtasks.
Meanwhile I have to somehow prevent host’s main thread from exit until is_main_task_done is set to true. Is there a cross platform solution? I’ve found only spinlock solution in AMD’s guide so far:
while (!is_main_task_done)
sleep(0)
But it appears to work in Windows only.
It’s a good idea to use semaphores here. Initialize the semaphore with the number of tasks you want to have running, and it will be ensured that only this number runs.