Skip to content

GitLab

Closed
Created by James Willenbring@jmwilleMaintainer

Tpetra::MultiVector constructor that takes a device View uses create_mirror instead of create_mirror_view

Created by: mhoemmen

@trilinos/tpetra @vbrunini

In most cases, Tpetra::MultiVector's host View is the result of Kokkos::create_mirror_view on its device View. However, in one case:

https://github.com/trilinos/Trilinos/blob/5c9a2ec8019b310b5b74a22e23fcbea11486f781/packages/tpetra/core/src/Tpetra_MultiVector_def.hpp#L427

Tpetra creates the host View from the input device View using Kokkos::create_mirror. This means that, unlike in all other cases, the host View is always a separate allocation (even if not building with CUDA!) from the device View. That's weird, and it may be causing issues relating to my fixes for #4626 (closed).

Note the comment above that line of code:

    // The difference between create_mirror and create_mirror_view, is
    // that the latter copies to host.  We don't necessarily want to
    // do that; we just want to allocate the memory.

git blame dates that to 2015. It could have been the case back then that create_mirror_view copied to host. It certainly does not now.