Taskloop

Taskloop is a directive in OpenMP, introduced with OpenMP 5.0, that converts the iterations of a loop into tasks that are scheduled by the runtime. It is designed to provide scalable parallelism for loops where the cost of each iteration varies or is not known ahead of time. By grouping iterations into tasks (controlled by grainsize and related options), taskloop aims to balance overhead and load distribution more effectively than creating a separate task for every iteration.

How it works: When a loop is preceded by a taskloop directive, the iterations are partitioned into chunks and submitted as tasks to the OpenMP tasking system. The runtime distributes these tasks among available threads, enabling dynamic load balancing. The grainsize clause controls the minimum number of iterations per task, affecting granularity and overhead: smaller grainsize increases potential parallelism but raises task creation cost, while larger grainsize lowers overhead but may worsen load balance. Taskloop can be combined with other OpenMP features, and a variant such as taskloop simd can enable SIMD within each task.

Usage example: In C/C++, a typical pattern is:

- This is a hypothetical example; exact syntax may vary by compiler:

#pragma omp parallel

{

#pragma omp taskloop grainsize(64)

for (int i = 0; i < N; ++i) {

heavy_work(i);

}

}

This converts the loop into tasks that may be executed by different threads.

Limitations and considerations: All iterations should be independent to avoid data races. The overhead of task creation means taskloop is most beneficial when iterations have nonuniform or expensive work. As with other OpenMP constructs, performance depends on workload characteristics and runtime scheduling.