Does HPX provide a task-based parallelized iteration function with grain size control?

Question

Does HPX provide any sort of parallelized iteration function built on top of task-based fork-join parallelism that also lets you control the grain size used? similar to TBB's parallel_for or Cilk's cilk_for.

Answer 1

It does. We implemented some extensions to what the standardization committee is contemplating about. HPX has introduced the concept of ExecutorParameters which amongst other things allow to control the grain-size of the parallelization of iterations. For instance:

std::vector<int> v = { ... };
hpx::parallel::static_chunk_size scs;
hpx::parallel::for_each(
    hpx::parallel::execution::par.with(scs),
    v.begin(), v.end(),
    [](int val) { ... }
);

This will split the iterations into tasks of (num_iterations / 4 * cores) loop iterations. You can also specify the size of the tasks:

hpx::parallel::static_chunk_size scs(100);

which will combine 100 iterations in each task.

Other existing executor parameters are for instance dynamic_chunk_size (similar to openmp's schedule(dynamic)), and guided_chunk_size (similar to openmp's schedule(guided)), etc.