GlobIter: Make GlobMemType a public typedef #635
Conversation
Codecov Report
@@ Coverage Diff @@
## development #635 +/- ##
===============================================
+ Coverage 84.94% 84.95% +<.01%
===============================================
Files 335 335
Lines 24821 24821
Branches 11236 11236
===============================================
+ Hits 21085 21087 +2
Misses 3733 3733
+ Partials 3 1 -2
|
|
@pascalj this is needed to get the executor in mephisto work again: mephisto-hpc/testing@60f6084#diff-0d729d073571dd4e46129f851327de2eR39 |
There was a problem hiding this comment.
Iterators and memory as independent concepts is a crucial principle in the STL.
The conceptual discrimination here is not formalese, it is most definitely existential.
Or to avoid category theory, in software engineering terms:
Let's remember that every implementation of the Global Iterator concept had to provide a public type memory_type which again is required to be a model of the Global Memory concept.
I had a look at the Mephisto code that motivated this request.
We actually need execution- and memory spaces as proposed (but not realized?) in Kokkos:
As a workaround for now, we could introduce dash::global_iterator_traits<GlobIterT>::memory_type, default-defined as void and specialized for ("befriended") global iterators which are actually memory-bound.
Ok, but than why did you removed @pascalj and @rkowalewski? @rkowalewski implemented the memory spaces in DASH and @pascalj implemented the mephisto part. I was only trying to get it to work again. |
|
The changes in this PR are not related to memory spaces or Mephisto but on iterator concepts. Global memory concepts, maybe (both developed and implemented by myself, but why should that be relevant?) |
No description provided.