Closed DSMishler closed 1 year ago
Can one of the admins verify this patch?
OK to test
New commit fixes a bug with vectorshift.cpp
in both single and double node.
Previously, vectorshift would copy over from vector b
into a_h
before checking, but this is no good! Vector b
contains the NUM_SHIFTS-1
th version of the vector since it is swapped with a
at the end of every loop. Instead, it is best to swap with a
. Alternatively, you could add -OFFSET
to the assert index check. -1
was a bit of a bandaid, hiding this issue because it happens to work when OFFSET
is its current default value of 1
.
Maybe some time someone could explain to me what Jenkins is doing, too.
@DSMishler it runs the suite. To see the results: click on Details
, then on Open Blue Ocean
on the left.
Your PR fails because of the format (see here)
Ah, I see now. Coding style checks. Thanks for that!
For the coding style, we use clang-format 8.0 with these rules. It is an old version of clang-format but that's what Kokkos uses.
Adding
Kokkos::PartitionedLayoutLeft
to the matvec and vectorshift examples, and the MPI rank determination to matvec. This allows the code to compile and run once again!On weaver with my current setup, I find I get the following message at runtime (though it seems to be benign):