Closed devreal closed 7 years ago
It's the local-to-global variants only, right? Global-to-local copying has been evaluated and benchmarked in lots of use cases, I would expect it's fine, Local-to-global has just been implemented for a single scenario and didn't get any love, yet. And, yes, that's annoying ... as you wrote yesterday, it's crucial to document "restricted" implementations like this one
The n-dim ranges and views can solve the underlying partitioning problem. So I prioritize documentation of views first and then power-wash the algorithms section.
It's the local-to-global variants only, right? Global-to-local copying has been evaluated and benchmarked in lots of use cases, I would expect it's fine, Local-to-global has just been implemented for a single scenario and didn't get any love, yet.
Yes, you're right. I seem to have missed the while
loop in the global-to-local overload of copy_impl
that iterates over the source units.
The n-dim ranges and views can solve the underlying partitioning problem. So I prioritize documentation of views first and then power-wash the algorithms section.
Sounds good to me :+1:
Closing this as a duplicate of #346
dash::copy
is broken for copy operations spanning more than the current node and its neighbors.Example:
with
dash::size()
> 3.Related: #386 for
dash::transform