Daniel Arndt

icon indicating a website link https://www.ornl.gov/staff-profile/daniel-arndt icon indicating an email [email protected]

icon company Oak Ridge National Lab icon location Oak Ridge icon location Large Scale Simulations Computational Scientist. One of the developers of the deal.II finite element library and Kokkos.

Results 753 comments of Daniel Arndt

possible memory leak in Kokkos::initialize() cuda implementation

We are trying to record all memory allocations in `SharedAllocationRecord` so I wanted to check that first and came across leaking `InternalScratchBitset`, see https://github.com/kokkos/kokkos/pull/5300, and this weird entry. I have...

possible memory leak in Kokkos::initialize() cuda implementation

That looks as if you are trying to use the `master` branch instead of `develop`, see https://github.com/kokkos/kokkos/blob/d19aab9981a2c447e832a7b4eb7b16992328fb14/cmake/kokkos_test_cxx_std.cmake#L132 (master) vs. https://github.com/kokkos/kokkos/blob/d845b57307057c99b24628785a1a838b7848efb8/cmake/kokkos_test_cxx_std.cmake#L120 (develop).

Consider changing the behavior of `Kokkos::is_initialized` and providing `Kokkos::is_finalized`?

I would be in favor of at least adding `Kokkos::is_finalized`.

Consider changing the behavior of `Kokkos::is_initialized` and providing `Kokkos::is_finalized`?

I think we agree on introducing the functionality of `Kokkos::finalized`. We are mostly debating how `Kokkos::initialized` should behave.

Consider changing the behavior of `Kokkos::is_initialized` and providing `Kokkos::is_finalized`?

I am between 1 and 2. In principle, I like 2 better but we could do 1 without deprecating `is_initialized` which is argument in favor of it for me.

Trilinos nightly build errors: Stokhos package, static_assert: Layout is not constructible from extent arguments

To me, it looks like `tmp_src_type` must be constructed with a layout representing a contiguous View (like `LayoutRight` or `LayoutLeft` and not copying `src_type`'s layout (which is `LayoutStride` here).

Trilinos nightly build errors: Stokhos package, static_assert: Layout is not constructible from extent arguments

See https://github.com/trilinos/Trilinos/pull/10783.

Reorder header self contained tests

IIRC, we discussed splitting the test in the discussion for https://github.com/kokkos/kokkos/pull/2986#issue-608603930 and decided against it to avoid `CMake` code duplication.

Trilinos nightly test failure, Stokhos_SacadoMPVectorUnitTest_MPI_1 (various subtests)

It appears that the problem is that `Kokkos`' `max` function is a better match than `Sacado::MP::max` and the test intents to use the latter one.

Trilinos nightly test failure, Stokhos_SacadoMPVectorUnitTest_MPI_1 (various subtests)

Sounds similar to what @maartenarnst is observing in https://github.com/trilinos/Trilinos/issues/10740#issuecomment-1180885529.