parallel_for with TeamPolicy::team_size_recommended with launch bounds not working -- reported by Daniel Holladay

[dholladay00]

On the CUDA backend, I need to make use of launch bounds to get rid of compile time errors regarding regcount. After this change, when the team_policy is constructed:

typedef Kokkos::TeamPolicy<schedule, Kokkos::LaunchBounds<64,4> > team_policy;
const team_policy policy(ns, team_policy::team_size_recommended(lambda));

The parallel_for with this team_policy does not execute. It does execute with this policy:

typedef Kokkos::TeamPolicy<schedule, Kokkos::LaunchBounds<64,4> > team_policy;
const team_policy policy(ns, Kokkos::AUTO);

[crtrott]

Clearly a bug where team_size_recommended does not apply LaunchBounds to its check.

[dholladay00]

I am showing similar for team_size_max, so could you check into that function as well?

[crtrott]

Yeah.

[crtrott]

Turns out this is slightly harder to fix than I thought. In particular the interface actually doesn't allow you to this for functors with more than one operator. Have to think about this a bit more and delay this to the next promotion.

[crtrott]

Ok here is a an idea interface wise:

team_policy.team_size_max(functor,parallel_for,TagType());
team_polciy.team_size_recommended(functor,parallel_reduce);

For this we need new overloads for parallel_for, parallel_reduce, parallel_scan something like

Impl::ParallelForTag parallel_for() {return ParallelForTag(); }

[crtrott]

Ok got that now in develop. We actually ended up directly using the Tags i.e.

team_policy.team_size_max(functor,Kokkos::ParallelForTag());
team_policy.team_size_recommended(functor,Kokkos::ParallelReduceTag());

Note that those are now member functions, and they take everything into account, including scratch memory etc.