Multithreading

[djukic14]

Hi @krcools,
I’ve pushed a PR that introduces (a preliminary version of) basis function coloring for the multi-threaded assembly of system matrices.
If you're interested in incorporating this, I’d be happy to discuss the best approach for integration. Currently, the coloring is offered as a third option alongside the existing single-threaded and multi-threaded assemble! Multi-threading is enabled by passing a scheduler from OhMyThreads.jl to the threading argument:

T = assemble(op, X, X; threading=DynamicScheduler())

This also allows for more control, such as using only a subset of the available threads. When using SerialScheduler, no coloring is applied, and assembly proceeds in a single thread.
The coloring itself is performed using GraphsColoring.jl with the WorkstreamDSATUR algorithm, which aims to produce balanced colors.
Additionally, I’ve added a new progress bar and parallelized the computation of the potential!
I’d be glad to refine this further (and eliminate any potential bugs I introduced) based on your feedback.
Best regards,
Danijel

[krcools]

This looks really great. I am happy to merge this as is. Will add a test to make sure the matrices assembled with different multi-threading strategies always give the same result.

The main reason I held on to my poor man's progress bar is that is works well with both output to the terminal and logging into a file. But tbh never used the latter so happy to move towards this more advanced solution...

Maybe the threading option should represent that a coloring strategy is being used...

T = assemble(op, X, X; threading=:cellcoloring, scheduler=DynamicScheduler())

What do you think? The current algorithm could then be

T = assemble(op, X, X; threading=:dofsplitting, scheduler=DynamicScheduler())

[djukic14]

I am glad to hear you like the approach.
If the new progress bar proves inconvenient in practice, I can revert it to the old version.

I can also update the PR to reflect the API change you suggested, but I would then propose dropping the dedicated “single‑threaded” variant altogether. The pure single‑threaded path would simply be

assemble(op, X, X; threading=:cellcoloring, scheduler=SerialScheduler())

and

assemble(op, X, X; threading=:dofsplitting, scheduler=SerialScheduler())

If there’s anything else you’d like to tweak or any other detail, just let me know. I can incorporate this into the PR. Of course, if you’d prefer to make the adjustments yourself, that works for me as well.

[krcools]

I still would like to have the

assemble(T, X, X; threading=:single)

option, both for debugging, and clarity purposes. I'll take a look and push onto your PR branch.