Improve DynRankView debug check #541
Assignees
Labels
Bug
Broken / incorrect code; it could be Kokkos' responsibility, or others’ (e.g., Trilinos)
Enhancement
Improve existing capability; will potentially require voting
Milestone
Comments
|
Would it be reasonable to check the extra indices are in fact zero? I think that would catch the errors Roger was running into. |
|
I added checks like @etphipp, tested on my laptop. Will run more rigorous tests and after passing will issue a pull request. |
ndellingwood
added a commit
to ndellingwood/kokkos
that referenced
this issue
Nov 15, 2016
This commit addresses Kokkos Github issue kokkos#541 If more arguments than the dynrankrank are passed to the parenthesis operator of a dynamic rank view, check that these are zero. This is allowed, but any non-zero value corresponds to an invalid address access, will abort with a message to user.
ndellingwood
added a commit
to ndellingwood/kokkos
that referenced
this issue
Nov 28, 2016
This commit addresses Kokkos Github issue kokkos#541 Enhancements made to debug bounds checking macros and routines to verify that arguments passed to an operator beyond the DynRankView's rank are 0 modified: src/Kokkos_DynRankView.hpp
ndellingwood
added a commit
to ndellingwood/kokkos
that referenced
this issue
Nov 29, 2016
This commit addresses Kokkos Github issue kokkos#541 Enhancements made to debug bounds checking macros and routines to verify that arguments passed to an operator beyond the DynRankView's rank are 0 modified: src/Kokkos_DynRankView.hpp
crtrott
added
Bug
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The
verify_dynrankview_rank( N , drv )routine used during debugging (specifically when bounds check is enabled) currently checks that the rank of the parenthesis operator ( the 'N' value above ) is greater than or equal to the dynamic rank, and aborts otherwise. This check, rather than equality, was chosen to allow for extra zero arguments to be passed to the DynRankView, which was needed for Intrepid2 refactoring and generic functors for various tensor/contraction operations if I remember correctly.With FAD types that are not created through the Sacado factory this can result in a situation where the results of this check pass but would be expected to fail.
@rppawlo @kyungjoo-kim @etphipp
The text was updated successfully, but these errors were encountered: