HIP-CPU VS Taskflow

If I'm reading this right, this is Fortran's equivalent of HIP, i.e. a way to (semi-)automatically convert CUDA-based solution to a more backend-independent one so that the same source can be run both on CUDA and ROCm GPUs (and potentially more; e.g. they also have an experimental CPU backend).

So, I know that you asked about cuda, but this might actually be possible in hip, and you can convert your code to hip relatively easily. The path would be to use the CPU implementation (https://github.com/ROCm-Developer-Tools/HIP-CPU) and then run your code coverage on that.

For parallel programming nowadays, personally I reach for C++ Taskflow when I really care about performance, or a mix of core.async and running multiple load balanced instances when I’m doing more traditional web backend stuff in Clojure.

Also, you can take a look to good libraries. The problem is that very often libraries are heavily templated, so It could be hard. For example, I like the style of the Taskflow library, I think is very clear, is relatively small, while makes use of more advanced techniques: https://github.com/taskflow/taskflow

Cool. Thanks! How does it compare to taskflow?

I've not seen yet, but it's been a bit since I looked last, any evidence of being able to build a computation graph and "save" it to re-run on new inputs. Something like https://github.com/taskflow/taskflow

It seems you're describing something a generic parallel task framework. Check taskflow for a production ready example https://github.com/taskflow/taskflow/blob/master/

- https://github.com/taskflow/taskflow (I recommend to learn it first since its API and documentation are excellent)

If you want some sort of "job" system, where you submit items to a some sort of queue to be processed in parallel, try searching for a thread pool - there isn't one in the standard library, but there's about a million implementations online. There are more complicated versions of that idea, that describe computation as a directed acyclic graph, such as taskflow.

The tooling decisions affect engine design though. For example if you want to have visual representation of job graph as it happened in specific frame of interest you need to pass the information around about job relationships and output it to a tool of choice. For example see https://github.com/taskflow/taskflow

I am aware of DAGs of task based threading library like Taskflow and HPX however the benefit they have is not obvious to me, as the following sequential section depends on the parallel part being completed fully. If you want to suggest elaboration on the benefits of this approach would be welcome.