Hommexx: Limit HIP team size in debug builds.

Also clean up the team_num_threads_vectors routine to consolidate
the rules.

Fixes SCREAM issue 2058.

[BFB]

components/homme/src/share/cxx/ExecSpaceDefs.cpp

@@ -173,39 +173,31 @@ team_num_threads_vectors (const int num_parallel_iterations,
   // fewer than 4 warps/thread block limits the thread occupancy to that
   // number/4. That seems to be in Cuda specs, but I don't know of a function
   // that provides this number. Use a configuration option that defaults to 4.
-  const int min_num_warps = HOMMEXX_CUDA_MIN_WARP_PER_TEAM;
+  int min_num_warps = HOMMEXX_CUDA_MIN_WARP_PER_TEAM;
+
+  int max_num_warps = HOMMEXX_CUDA_MAX_WARP_PER_TEAM;
+#ifdef KOKKOS_ENABLE_DEBUG
+  // In debug builds, team size must be smaller because of Kokkos-side data.
+  max_num_warps = std::min(max_num_warps, 8);
+#endif
+
 #ifdef KOKKOS_ENABLE_CUDA
   const int num_warps_device = Kokkos::Impl::cuda_internal_maximum_concurrent_block_count();
   const int num_threads_warp = Kokkos::Impl::CudaTraits::WarpSize;
-  // The number on P100 is 16, but for some reason
-  // Kokkos::Impl::cuda_internal_maximum_grid_count() returns 8. I may be
-  // misusing the function. I have an open issue with the Kokkos team to resolve
-  // this. For now:
-# ifdef KOKKOS_ENABLE_DEBUG
-  // Work around spurious "terminate called after throwing an instance
-  // of 'std::runtime_error'" message.
-  const int max_num_warps = 8;
-# else
-  const int max_num_warps = HOMMEXX_CUDA_MAX_WARP_PER_TEAM; //Kokkos::Impl::cuda_internal_maximum_grid_count();
-# endif
-
 #elif defined(KOKKOS_ENABLE_HIP)
-
-  //use 64 wavefronts per CU and 120 CUs
+  // Use 64 wavefronts per CU and 120 CUs.
   const int num_warps_device = 120*64; // no such thing Kokkos::Impl::hip_internal_maximum_warp_count();
-  const int max_num_warps = HOMMEXX_CUDA_MAX_WARP_PER_TEAM;
   const int num_threads_warp = Kokkos::Experimental::Impl::HIPTraits::WarpSize;
-
 #else
-
-  // I want thread-distribution rules to be unit-testable even when Cuda is
-  // off. Thus, make up a P100-like machine:
+  // I want thread-distribution rules to be unit-testable even when GPU spaces
+  // are off. Thus, make up a GPU-like machine:
   const int num_warps_device = 1792;
   const int num_threads_warp = 32;
-  const int max_num_warps = 16;
-
+  max_num_warps = 16;
 #endif
 
+  min_num_warps = std::min(min_num_warps, max_num_warps);
+
   return Parallel::team_num_threads_vectors_for_gpu(
     num_warps_device, num_threads_warp,
     min_num_warps, max_num_warps,