team_dot_complex_double fails on clang+cuda due to misalignment

[brian-kelley]

On kokkos-dev2, with Clang 8 and CUDA 10.0 (a spot check build), team_dot_complex_double fails because of cudaErrorMisalignedAddress. Since this passes in all other CUDA builds (even with 16 == alignof(Kokkos::complex<double>)), it seems likely this is just a compiler bug.

I can tell this error is caused by the KokkosBlas::Experimental::dot device function, since cudaGetLastError() reports the misaligned access right after the functor launches in the test Test_Blas1_team_dot.hpp, in impl_test_team_dot():

    Kokkos::fence();
    checkCuda();  // returns cudaSuccess
    Kokkos::parallel_for( "KokkosBlas::Test::TeamDot", policy, KOKKOS_LAMBDA ( const team_member &teamMember ) { 
       const int teamId = teamMember.league_rank();
       CHECKALIGN(&d_r(teamId));
       auto lhs = Kokkos::subview(a,Kokkos::make_pair(teamId*team_data_siz,(teamId < M-1)?(teamId+1)*team_data_siz:N));
       auto rhs = Kokkos::subview(b,Kokkos::make_pair(teamId*team_data_siz,(teamId < M-1)?(teamId+1)*team_data_siz:N));
       CHECKALIGN(lhs.data());
       CHECKALIGN(rhs.data());
       d_r(teamId) = KokkosBlas::Experimental::dot(teamMember, lhs, rhs);
    } );
    Kokkos::fence();
    checkCuda(); // returns cudaErrorMisalignedAddress

If it's not a compiler bug, I have no idea where the actual bug could be. This dot implementation is definitely calling TeamDot::team_dot in KokkosBlas1_team_dot_spec.hpp, and this takes actual views as input, not raw pointers. I checked that the X and Y passed to this functor are correctly aligned too.

Disabling KOKKOS_ENABLE_COMPLEX_ALIGN does make the test pass.