[Openmp-commits] [PATCH] D95282: [OpenMP] Add source location information to the libomptarget profile

[Johannes Doerfert via Phabricator via Openmp-commits]

jdoerfert added inline comments.


================
Comment at: openmp/libomptarget/src/omptarget.cpp:217
                      TargetDataFuncPtrTy target_data_function) {
+  TIMESCOPE_WITH_IDENT(loc);
   DP("Calling the mapper function " DPxMOD "\n", DPxPTR(arg_mapper));
----------------
grokos wrote:
> I see some inconsistency here. `targetDataMapper` is an internal libomptarget function (not an API function) and yet it is getting profiled. Same for `targetDataContiguous` and `targetDataNonContiguous`. The other 3 internal data-related functions `targetDataBegin`, `targetDataEnd` and `targetDataUpdate` are not. Is this intentional?
Kind of.

There are a few options:
1) Simply profile the interface. We did this at first. The problem is the `omp target` entry point does 3 things and it is not too helpful to merge the results of h2d, kernel, and d2h, into a single number (IMHO).
2) Profile the interface and some internal functions. This is not a bad solution but it is unclear what we gain by also profiling the interface at that point.
3) What I tried to do here is to pick the functions that correspond to conceptual tasks, like h2d/d2h, kernel execution (which is not a function but scope with name in `target` below), ..., and then profile those.

Long story short, we can discuss this in the meeting at some point and I'm happy for improvement patches :)


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D95282/new/

https://reviews.llvm.org/D95282