Tomasetti Romin comments

Results 152 comments of


                                            Tomasetti Romin

Cuda: fix long double suppression

I guess the only job failure (`nvcc` seg faulting during compilation) is not due to any change from this PR and can safely be ignored.

[DO NOT MERGE] Tpetra: Initial support for execution_space instances in MultiVector

> @romintomasetti Well the only test that failed was a randomly occurring one, so this looks like it worked. Next up: Reworking the rest of the relevant routines on MultiVector...

[WIP] Initial version of self similar interface for range loops

I'm wondering whether this new interface should be part of *core*. While it could be useful (it relieves you from writing specializations of `sum_views` for an execution space or a...

core(allocation): avoid copying shared allocation header on device if not needed

I cannot tell (and did not dig yet) if there are other device-side locations in `Kokkos` that rely on copying the shared allocation header. From what I can gather, the...

nvcc 12.9.41 / g++ 11.4.0 / c++20 compiler crash in unit tests, perf tests, and examples

I'd add that we found: - https://github.com/NVIDIA/cccl/issues/4629

Update bundled GoogleTest version to 1.14.0 and use FetchContent to build it

:partying_face: This makes me happy because you're bringing in `gmock` - that means a lot of test assertions can become much cleaner (thinking of strings for instance). For instance you...

core(graph): node properties and device handle

Let's first merge #8192.

core(graph): node properties and device handle

> I think it makes sense to always use the `Graph`s execution space as default for all its nodes. OK, so we agree :+1: My only concern is for the...

core(graph): node properties and device handle

> Assuming the following points are true right now: > > * we don't support mixing host/device nodes in the same graph We actually would (and need to) support it,...

core(graph): node properties and device handle

Next steps :foot: 1. Do not propagate the `Context`. Keep it at the user-facing API level. It will drastically reduce the changes. 2. Use `using device_handle_t = std::tuple;` instead of...