Queue management for task graphs

A method, comprising:

• receiving, by a command processor, a fiber including one or more tasks and dependencies between the one or more tasks, the command processor including a first thread that manages a ready queue and a second thread that manages a sleep queue;
• dispatching, by the command processor, a task from the fiber for execution by a processing element array based on the fiber being enqueued in the ready queue and the dependencies of the task being resolved;
• enqueueing, by the command processor, the fiber in the sleep queue while the task is dispatched and unexecuted by the processing element array;
• enqueueing, by the command processor, the fiber in the ready queue based on receiving a completion signal indicating that the task has been executed by the processing element array;
• and
• enqueueing, by the command processor, a dependent fiber that depends from the fiber in the ready queue based on the fiber being enqueued in the ready queue and the one or more tasks of the fiber having been executed by the processing element array.

A system, comprising:

• an accelerator device that includes a command processor and a processing element array, the command processor including a first thread that manages a ready queue and a second thread that manages a check queue;
• and
• a host configured to compile operations for executing a fiber graph that includes fibers each having one or more tasks, the fiber graph indicating dependencies between the fibers and between the tasks within the fibers, the operations instructing the command processor to:
• dispatch a task from a fiber for execution by the processing element array based on the fiber being enqueued in the ready queue and the dependencies of the task being resolved;
• push the fiber to a tail of the ready queue based on the task being dispatched and unexecuted by the processing element array;
• enqueue the fiber in the check queue based on the fiber being enqueued in the ready queue and the one or more tasks of the fiber having been executed by the processing element array;
• and
• enqueue a dependent fiber that depends from the fiber in the ready queue based on the fiber being enqueued in the check queue.