Remove event system from KA and update to CUDA v4.1

[vchuravy]

This likely removes some of the single device parallelism,
but decreases the risk for user mistake (everything is stream-ordered)

This will cause severe merge conflicts with open PRs and depends on currently unreleased
versions of CUDA & KernelAbstractions. So I might regret doing it now, but I was curious
to see what the damage would be.

[simone-silvestri]

Nice, how do you go about launching overlapping asynchronous kernels this way?

[vchuravy]

Nice, how do you go about launching overlapping asynchronous kernels this way?

That's yet to be determined. I was talking to @maleadt about this. Maybe I will add a foreach variant to KernelAbstractions that implements this without requiring the use of tasks.

[maleadt]

Are the current options (@async blocks) not sufficient? I'd rather not have a parallel mechanism and have to maintain the APIs (explicitly passing streams/queues) to support it. Even right now it's relatively broken, only supporting kernel launches and memory copies (i.e. BLAS APIs do not take explicit stream arguments).

[vchuravy]

My goal right now is to do all of the changes without introducing new constructs and then evaluate performance from there.

[glwagner]

@vchuravy I recall long ago we discussed introducing a "stream" abstraction as an alternative to events for asynchronous computations? This would parallel CUDA's streams but I guess would have to be something entirely new for CPU?

[vchuravy]

In discussion with Tim I decided against it. I tried implementing it on the CPU, but that's basically why I required events everywhere.

See https://github.com/JuliaGPU/KernelAbstractions.jl/blob/main/examples/mpi.jl, but Tim and I are bullish on using Julia's native concurrency mechanism, to express what streams normally do.

[glwagner]

Epic work! It's not just an upgrade, the code is much simpler and I feel relieved. My comments are all cosmetic.

[glwagner]

We'll have to generate some new benchmarks after this especially for smaller / 1D / 2D problems

[navidcy]

Or even before this is merged?

[navidcy]

what's the difference of "0.4" and "^0.4"? I thought they are the same.