Christian Trott comments

Results 519 comments of


                                            Christian Trott

Hip backend compilation issues with debug plus boundscheck options enabled (rocm/4.2.0 compiler)

Ok closing since 5.2 is the minimum we support anyway

MDRangePolicy has no API to obtain the default tile size

Yeah, we actually probably need a max tile size as well. Possibly this needs to follow the TeamPolicy pattern of recommended and max team_size. I am putting this on our...

Add support for CUDA unified memory architectures i.e. Grace Hopper

Seeing the same thing. Will disable NEON SIMD for now if __NVCC__ is used. See: https://github.com/abseil/abseil-cpp/issues/1665

Add support for CUDA unified memory architectures i.e. Grace Hopper

@jbadwaik I pushed the NEON thing (disabling use of NEON SIMD if NVCC is the compiler) can you try (if you didn't) with: ``` -DKokkos_ENABLE_IMPL_CUDA_UNIFIED_MEMORY=ON ``` I am getting two...

Add support for CUDA unified memory architectures i.e. Grace Hopper

My full configure line: ``` cmake -DKokkos_ARCH_ARMV9_GRACE=ON -DKokkos_ARCH_HOPPER90=ON -DKokkos_ENABLE_TESTS=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ENABLE_IMPL_CUDA_UNIFIED_MEMORY=ON ../ ```

Add support for CUDA unified memory architectures i.e. Grace Hopper

@jbadwaik thanks! so its most likely something real. Now we have to figure out what it is ...

Add support for CUDA unified memory architectures i.e. Grace Hopper

@jbadwaik found the issue feel free to test again. But this should be good now.

cuda: reduction with `RangePolicy`: fix grid dimensions to work for large values and avoid overflow

Ok lets restrict to GPU backends and Release mode.

Kokkos + Legion interop

This is in support of customers at LANL, i.e. we will try and figure this out in the next couple months.

Kokkos + Legion interop

We are moving: https://github.com/kokkos/kokkos/pull/6091