Daniel Arndt

We looked into that in #6276 and found https://github.com/kokkos/kokkos/pull/6276#discussion_r1358763491: > I explicitly need to link with svml in a SYCL+Cuda build I get slightly different results making the tests fail...

Thus, I think we would need to relax the test to allow for some small errors.

OK to test.

I am seeing ``` 53: [ RUN ] simd.host_math_ops 53: /home/darndt/kokkos/simd/unit_tests/include/SIMDTesting_Utilities.hpp:29: Failure 53: Expected equality of these values: 53: a 53: Which is: 0.736806 53: b 53: Which is: 0.736806...

We could do something like ``` diff --git a/simd/src/CMakeLists.txt b/simd/src/CMakeLists.txt index 8779112bc..6f667c909 100644 --- a/simd/src/CMakeLists.txt +++ b/simd/src/CMakeLists.txt @@ -22,6 +22,9 @@ KOKKOS_ADD_LIBRARY( SOURCES ${SIMD_SOURCES} HEADERS ${SIMD_HEADERS} ) + +target_link_libraries(kokkossimd PUBLIC...

> @vbrunini @ibaned I pushed this up for STK. Does this pull request improve performance for STK measurably?

Something like ```diff diff --git a/simd/src/CMakeLists.txt b/simd/src/CMakeLists.txt index 8779112bc..3460b388a 100644 --- a/simd/src/CMakeLists.txt +++ b/simd/src/CMakeLists.txt @@ -22,6 +22,16 @@ KOKKOS_ADD_LIBRARY( SOURCES ${SIMD_SOURCES} HEADERS ${SIMD_HEADERS} ) + +# At least for SYCL+Cuda,...

Only the HIP builds failed (which we now not to be working right now).

> , I suppose this issue depends on either the OpenMP version or vendor which is why it didn't show up in CI? We don't use a verbose `Makefile` there.

https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#table.half.math.functions is a list of supported math functions for half type.