Daniel Arndt comments

Results 829 comments of


                                            Daniel Arndt

Add test for TeamPolicy array reduction and fix Cuda and HIP

Retest this please.

Add test for TeamPolicy array reduction and fix Cuda and HIP

Retest this please.

MDRangePolicy has no API to obtain the default tile size

Note that `m_tile` is public as of now (which we might change).

Optimized behavior of resize/realloc of views with aliasing mirrors is surprising

I think we mostly cared about the optimization in case the view count was 1 and restricting it to that case sounds sensible to me.

SYCL: Improve team_reduce implementation

For ```C++ #include constexpr int step_width = 16; template value_type team_reduce_v0(sycl::nd_item item, value_type value) { auto sg = item.get_sub_group(); const auto sub_group_range = sg.get_local_range()[0]; const auto vector_range = 1; const...

SYCL: Improve team_reduce implementation

The latest results for ```C++ #include constexpr int step_width = 16; #define DO_SHIFT_01(SHIFT) \ if (vector_range * SHIFT < sub_group_range)\ {\ const value_type tmp = sg.shuffle_down(value, vector_range * SHIFT); \...

SYCL: Improve team_reduce implementation

> @masterleinad can you add a changelog entry for this update to the 4.4 tracker #6914 ? done

unsupported option '-fopenmp-targets=' for language mode 'HIP'

> So, when I specify both Kokkos_ENABLE_HIP and Kokkos_ENABLE_OPENMP, the openmp flags are added after the HIP flags. Where is that order specified and is there a way to switch...

unsupported option '-fopenmp-targets=' for language mode 'HIP'

> Ok, but I could not find the place where -fopenmp is added to the kokkoscore target COMPILE_OPTIONS. `OpenMP` flags are not added to `COMPILE_OPTIONS` since we use `target_link_libraries` for...

`inline static` variables in headers break CUDA kernels of a library loaded with `RTLD_DEEPBIND`

https://github.com/kokkos/kokkos/pull/5473#issuecomment-1299102792 is related.