Update Iterative Tasking Infrastructure #917
Conversation
pdmullen
changed the title
| template <class T, class U, class... Args1, class... Args2> | ||
| TaskID AddTask(TaskID const &dep, TaskStatus (T::*func)(Args1...), U *obj, | ||
| Args2 &&...args) { | ||
| return this->AddTask_(TaskType::iterative, 1, dep, [=]() mutable -> TaskStatus { | ||
| return (obj->*func)(std::forward<Args>(args)...); | ||
| return (obj->*func)(std::forward<Args2>(args)...); |
There was a problem hiding this comment.
I don't immediately see why we need two separate variadic template parameters here. I am mostly just curious what issue the single list was causing. Maybe it was something to do with default arguments?
There was a problem hiding this comment.
Unfortunately I can't remember exactly what situation led to this change, but I can give you the basic idea. What happened was that I wrote a task with some set of reasonable arguments (I thought) but the compiler yelled at me about type mismatches when I actually tried to add it to a task list. Don't hold me to this, but I think it was a const vs non-const issue. I convinced myself at the time that there was no reason we should be preventing people from writing tasks the way I had written the problematic one, and this transformation generalized the interface to allow it.
|
Extended regression is failing upon enabling auto-merge. Is this definitely pointing to an issue with the changes in this PR? Or could it be unrelated? Asking before I jump down the rabbit hole. |
Given that automerge eventually let it go, sounds like the issue was just the test machine overloaded. |
PR Summary
I am submitting this PR on behalf of @jdolence who made these changes in
lroberts36/merge-sparse-with-jdolence-sparsesome time ago. In practice, this PR slightly modifies the task list infrastructure to update iterative tasking, in particular with regional dependencies. This PR fixes various failure modes exhibited in downstream codes whereby (1) iterative tasks failed to iterate, (2) hangs were experienced, or (3) unexpected results were obtained with MPI both with and without regional dependencies.PR Checklist