Closed mzuzek closed 2 years ago
One possible workaround is to disable the test in considered scenario, if we don't plan to support the problematic accessors in the Kokkos implementation any soon.
The scaled
and transposed
input matrix modifiers are fundamental to the proposal. Instead of disabling the failing test, would you consider a temporary fix that falls back to the sequential implementation if it encounters these accessors?
@mhoemmen
Instead of disabling the failing test, would you consider a temporary fix that falls back to the sequential implementation if it encounters these accessors?
Sure, please see https://github.com/kokkos/stdBLAS/pull/222. I've created new PR for the alternative solution, so this one can be just closed to mark this approach as rejected.
@MikolajZuzek Thanks for explaining! I'll close this PR then.
In build with default Kokkos execution space (both
LINALG_ENABLE_KOKKOS
andLINALG_ENABLE_KOKKOS_DEFAULT
enabled) the nativegemm
test runs Kokkos implementation and fails on missing overloads fortransposed
- andscaled
-wrapped inputs - see #218.One possible workaround is to disable the test in considered scenario, if we don't plan to support the problematic accessors in the Kokkos implementation any soon.
On this PR I'm proposing generic test exclusion list (
STDBLAS_TESTS_EXCLUDE
) to avoid "polluting" native tests CMake scripting with Kokkos code, which would happen with simpler solution like: