interface ov::threading::ITaskExecutor¶

Overview¶

Interface for Task Executor. OpenVINO uses ov::ITaskExecutor interface to run all asynchronous internal tasks. Different implementations of task executors can be used for different purposes: More…

#include <itask_executor.hpp>

template ITaskExecutor
{
    // methods

    virtual void run(Task task) = 0;
    virtual void run_and_wait(const std::vector<Task>& tasks);
};

// direct descendants

template ITaskExecutor;
template IStreamsExecutor;
class ImmediateExecutor;

Detailed Documentation¶

Interface for Task Executor. OpenVINO uses ov::ITaskExecutor interface to run all asynchronous internal tasks. Different implementations of task executors can be used for different purposes:

To improve cache locality of memory bound CPU tasks some executors can limit task’s affinity and maximum concurrency.
The executor with one worker thread can be used to serialize access to acceleration device.
Immediate task executor can be used to satisfy ov::ITaskExecutor interface restrictions but run tasks in current thread. Implementation should guaranty thread safety of all methods

Synchronization¶

It is ov::ITaskExecutor user responsibility to wait for task execution completion. The c++11 standard way to wait task completion is to use std::packaged_task or std::promise with std::future. Here is an example of how to use std::promise to wait task completion and process task’s exceptions:

// std::promise is move only object so to satisfy copy callable constraint we use std::shared_ptr
auto promise = std::make_shared<std::promise<void>>();
// When the promise is created we can get std::future to wait the result
auto future = promise->get_future();
// Rather simple task
InferenceEngine::Task task = [] {std::cout << "Some Output" << std::endl; };
// Create an executor
InferenceEngine::ITaskExecutor::Ptr taskExecutor = std::make_shared<InferenceEngine::CPUStreamsExecutor>();
if (taskExecutor == nullptr) {
    // ProcessError(e);
    return;
}
// We capture the task and the promise. When the task is executed in the task executor context
// we munually call std::promise::set_value() method
taskExecutor->run([task, promise] {
    std::exception_ptr currentException;
    try {
        task();
    } catch(...) {
        // If there is some exceptions store the pointer to current exception
        currentException = std::current_exception();
    }

    if (nullptr == currentException) {
        promise->set_value();  //  <-- If there is no problems just call std::promise::set_value()
    } else {
        promise->set_exception(currentException);    //  <-- If there is an exception forward it to std::future object
    }
});
// To wait the task completion we call std::future::wait method
future.wait();  //  The current thread will be blocked here and wait when std::promise::set_value()
                //  or std::promise::set_exception() method will be called.

// If the future store the exception it will be rethrown in std::future::get method
try {
    future.get();
} catch(std::exception& /\*e\*/) {
    // ProcessError(e);
}

Methods¶

virtual void run(Task task) = 0

Execute ov::Task inside task executor context.

Parameters:

task	A task to start

virtual void run_and_wait(const std::vector<Task>& tasks)

Execute all of the tasks and waits for its completion. Default run_and_wait() method implementation uses run() pure virtual method and higher level synchronization primitives from STL. The task is wrapped into std::packaged_task which returns std::future. std::packaged_task will call the task and signal to std::future that the task is finished or the exception is thrown from task Than std::future is used to wait for task execution completion and task exception extraction.

run_and_wait() does not copy or capture tasks!

Parameters:

tasks

A vector of tasks to execute