Ian Graham

> We additionally need to consider GPU synchronization more carefully to ensure that user's threads are accessing the data they intend and not data from previous steps while asynchronous kernels...

Understood, though you would still image at most one stream per simulation, right? If that's all, I don't think the streams addition is that complicated code-wise. Just that if you...

Aaahh, I hadn't noticed that looking at the stream API, super simple! I think either model is fine, though naturally I would have thought the stream belongs with the `Device`...

Okay cool! I'll get separate PRs going for these!

Ahh, well that's cool to hear! But that's still quite a ways away, probably an October 2023 release? I don't mind spending some time with this since I need the...

Separately, I've made quite a bit of progress with refactoring a unique `stream` that is attached to the `ExecutionConfiguration`. One thing I'm intending to ignore here though is that there...

> Have you profiled with Nsight systems and verified that you get parallel execution on the separate streams? No not yet, but I certainly will. A couple more regex-replaces before...

> I'd be happy to merge your changes that release the GIL if it does so by default and you have tested and are reasonably sure that the GIL is...

I've also found some places where kernel drivers are not found in a `kernel` namespace, am I free to fix that?

Got it. Oh and it's no problem. That may be a better solution, since it might also apply to `thrust` operations. That was an issue I was anticipating having to...