Daniel Arndt

> But similar things must have been done in the xSDK setup - how is the initialization handled there? `xSDK` just makes sure that `deal.II` can be built with other...

I agree that the new initialization function shouldn't initialize `Kokkos`.

I think it makes sense to always use the `Graph`s execution space as default for all its nodes.

> Do you have an idea what may be the issue? Essentially, in the test, we're creating three functors of increasing size. We launch them in a graph. We see...

@cwpearson Can you also run the benchmark with `SYCL` on `blake`?

I'm seeing ``` ======================================== Multidimensional View Index Benchmark ======================================== Concurrency: 1835008 Iterations: 100000 Backend: SYCL Team size: 64 League size: 28672 ======================================== warming up int32_t/int64_t... done Running int32_t test... done...

> Validity of the kernel launch parameters are checked by the Cuda, HIP, and SYCL APIs respectively (but the outcome of this is only checked by Kokkos on debug mode...

> @masterleinad Following [this blog post](https://leimao.github.io/blog/Proper-CUDA-Error-Checking/), I think it's safe to always check `cudaGetLastError` immediately after a kernel launch, which would capture synchronous errors (without fencing). Asynchronous errors, that require...

You shouldn't need to do anything special if you configure `Trilinos` with `OpenMP` support.

If you provide `BOOST_DIR`, it should set `BOOST_ROOT` internally and that should be picked up as hint for finding `Boost`, see https://cmake.org/cmake/help/latest/module/FindBoost.html. The warning suggests that you might want to...