Daniel Arndt

I'm having a hard time reproducing ```C++ TEST(TEST_CATEGORY, dualview_sequential_host_init) { DualViewType dv(Kokkos::view_alloc("myView", Kokkos::SequentialHostInit), 100); dv.resize(Kokkos::view_alloc(Kokkos::SequentialHostInit),200); ASSERT_EQ(dv.d_view.size(), 200); ASSERT_EQ(dv.h_view.size(), 200); dv.realloc(Kokkos::view_alloc(Kokkos::SequentialHostInit),300); ASSERT_EQ(dv.d_view.size(), 300); ASSERT_EQ(dv.h_view.size(), 300); } ``` seems to be fine

``` [ RUN ] cuda.dualview_sequential_host_init [ OK ] cuda.dualview_sequential_host_init (0 ms) ```

I can reproduce with ```C++ TEST(TEST_CATEGORY, dualview_sequential_host_init) { Kokkos::DualView dv(Kokkos::view_alloc("myView", Kokkos::SequentialHostInit), 100); dv.resize(Kokkos::view_alloc(Kokkos::SequentialHostInit),200); ASSERT_EQ(dv.d_view.size(), 200); ASSERT_EQ(dv.h_view.size(), 200); dv.realloc(Kokkos::view_alloc(Kokkos::SequentialHostInit),300); ASSERT_EQ(dv.d_view.size(), 300); ASSERT_EQ(dv.h_view.size(), 300); } ```

The unit test tried using `SharedSpace`.

@stanmoore1 Can you describe the use case some more? If we need/want `SequentialHostInit` on the host wouldn't we run into issues in the device view anyway?

I'd focus for now on improving behavior for 1 and make sure that in this case `modify` will actually set flags, a first `sync` fences and a second `sync` doesn't.

> Here is an option: we always make modify increment something, and sync without instance will fence if any of the modified flags are not 0. Impact: That affectively means...

> So with `Kokkos::Serial` as a backend the first case would now fence when it was not the case before. Yes, it would call `Kokkos::fence` on the first call no...

See https://github.com/dealii/dealii/pull/15731 and https://github.com/dealii/dealii/pull/15661.

> Since the build fails I think there is something (possible MueLu) missing in the dependencies docker. The error looks like the one fixed in https://github.com/trilinos/Trilinos/commit/553d6ebc00682cc84bfcb8724aeae37e7b77b563.