clemisch

Bump :smiley_cat: > In this little test I only see x4 speedup vs. single-core numpy. Is this expected?

Thank you very much for clarifying!

Yes, I agree with you. It's still relevant to discuss them together since the user field being hidden but used can be very confusing.

I think the corresponing commit is 3ab334dd00401d087d9f6618c9bfce8570f54ea7. It seems to me that the prevalence of `user` over `login` is unintended.

Just now coming back to this. Does autobound have an interface for > non-sharp bounds based on the range of the kth derivative ? Or would I have to read...

Thanks, that looks good! Do I understand correctly that Kernex always uses array padding for border handling? Do you have experience with performance of padding vs. adapting the computed indices...

Ah, that's very interesting! I asked because I don't have good evidence myself. Naively, padding seems inefficient, but I don't know how it works under `jax.jit` and how the memory/flops...

PS: It could be interesting to compare performance of 1. Kernex' `padding="same"` with array padding, to 2. no padding/border handling at all and `.at(mode="clip")` mimicking `pad(mode="edge")`. If (2) is _significantly_...

Edit: see https://github.com/ASEM000/Kernex/issues/14

FYI, I can't re-open. I created a new issue: https://github.com/ASEM000/Kernex/issues/14