[Openmp-commits] [PATCH] D19229: [STATS] Use partitioned timer scheme

[John Mellor-Crummey via Openmp-commits]

jmellorcrummey added a subscriber: jmellorcrummey.
jmellorcrummey added a comment.

I am somewhat surprised by this set of changes. The set of thread states being maintained is very close to what OMPT supports. Are the differences worthwhile? If so, why not have someone from Intel participate in the OMPT discussions and lobby for changes. If the differences are not worthwhile, then what is the point of having maintaining two independent implementations of state maintenance that largely duplicate one another? Why not just use the OMPT states? If states aren't being maintained properly for the OpenMP library (I know that not all of the OMPT state maintenance is implemented or correct), why doesn't someone from Intel with intimate knowledge of how the runtime works fix it? If I were to fix the OMPT implementation, I would consider simply putting OMPT state maintenance in the same places that Intel engineers who wrote the library put them (where they are in this patch).

After going to the effort of patching OMPT into this library, I've noticed the distinct lack of Intel investment in helping move the implementation forward. To me, this is a sign that Intel would prefer that the efforts to standardize the OMPT API would just wither and die. If I'm mistaken, please correct me and let's explore how to collaborate.


Repository:
  rL LLVM

http://reviews.llvm.org/D19229