Let's revisit these helpers in light of what happens in https://github.com/uliegecsm/kokkos-utils/issues/42.

Member Kokkos::Execution::ExecutionSpaceImpl::Scheduler< Exec >::OpState< Rcvr >::start () &noexcept

Check signature. And check whether we should move the receiver.

Member Kokkos::Execution::ExecutionSpaceImpl::ThenClosure< Exec, Label, Functor >::get_policy () const &noexcept

Implement policy update for the single policy in kokkos.

Struct Kokkos::Execution::GraphImpl::ExecOf< Kokkos::Impl::DeviceHandle< Exec > >

Enrich the Kokkos::Impl::DeviceHandle type with a execution_space alias.

Struct Kokkos::Execution::GraphImpl::GraphAddNodeEvent

Store the dispatch tag.

Struct Kokkos::Execution::GraphImpl::GraphOperationStateFor< OpState, Exec >

To be removed and done properly.

Member Kokkos::Execution::GraphImpl::Scheduler< Exec >::OpState< Rcvr >::start () &noexcept

Check signature. And check whether we should move the receiver.

Member Kokkos::Execution::GraphImpl::State< GraphComposition::Create, Exec >::m_root

The root node is stored to avoid reference counting incurred by https://github.com/kokkos/kokkos/blob/1945b637c3fab027fe90208753e8b2ec236302d4/core/src/Kokkos_Graph.hpp#L100.

Member Kokkos::Execution::GraphImpl::WhenAllOpState< Exec, Rcvr, Sndrs >::WhenAllOpState (stdexec::__tuple< Sndrs... > &&sndrs_, Rcvr &&rcvr_)

The graph will be created on the default device and submitted on the default execution space instance. The device selection for each node will still happen correctly according to each node properties. A possibility could be to ask the successor of stdexec::when_all for an execution space instance to submit the graph onto.

Member Kokkos::Execution::GraphImpl::WhenAllSender< Exec, Sndrs >::get_completion_signatures ()

Make it much more robust as per https://github.com/NVIDIA/stdexec/blob/da7640aa1002cfcd665254f03434d3c949844361/include/stdexec/__detail/__when_all.hpp#L426.

Member Kokkos::Execution::Impl::as_convertible_to_string (T &&value) noexcept(noexcept(std::forward< T >(value))) -> decltype(std::forward< T >(value))

While Kokkos::parallel_for allows for convertible-to std::string labels, Kokkos::parallel_reduce does not. See also:

Member Kokkos::Execution::Impl::RequiresSync< Exec, Rcvr >::operator() (const Exec &exec, const Rcvr &rcvr) const noexcept

Simplify by moving and using Tests::Utils::are_same_instances.

Member Kokkos::Execution::Impl::RequiresSync< Exec, Rcvr >::operator() (const Exec &, const Rcvr &) const noexcept

This is compile-time only and should go away.

Class Tests::ExecutionSpaceImpl::WhenAllTest_nested_with_inner_followed_by_other_Test

There is a missing fence. Our stdexec::continues_on puts the execution space instance in the environment because its preceding stdexec::when_all has at least one branch on Kokkos::Execution::ExecutionSpaceContext.

Class Tests::GraphImpl::WhenAllTest_schedule_sender_Test

Make sure that a bare continues_on(gctx) works as well.

Struct Tests::Utils::Functors::ReduceIndicesWithTeam< Exec, ValueType >

Because of the deduction logic in https://github.com/kokkos/kokkos/blob/dd9eb811539af4913ae8d4f29e17ab0c73676550/core/src/impl/Kokkos_FunctorAnalysis.hpp#L258-L303, we cannot template the team-handle call operator. Therefore, the member and value types must be known. See also https://github.com/kokkos/kokkos/issues/7794. Once solved, this functor can be fused with ReduceIndices.