Merge pull request #386 from E3SM-Project/jgfouca/wsm_link_fix

WorkspaceManager: Make constant a CPP define to make it easier to link

Author: jgfouca

src/kokkos/ekat_team_policy_utils.hpp

@@ -258,6 +258,7 @@ class TeamUtils<ValueType,EkatGpuSpace> : public TeamUtilsCommonBase<ValueType,E
     _open_ws_slots("open_ws_slots", _need_ws_sharing ? _num_ws_slots : 0),
     _rand_pool()
   {
+    EKAT_REQUIRE_MSG(overprov_factor >= 1.0, "Makes no sense to have an overprov < 1");
     if (_need_ws_sharing) {
       _rand_pool = RandomGenerator(std::chrono::high_resolution_clock::now().time_since_epoch().count());
     }

src/kokkos/ekat_workspace.hpp

@@ -72,7 +72,9 @@ class WorkspaceManager
   //
 
   // Default overprov factor for large GPU problems, testing has shown 1.25 is optimal
-  static constexpr double GPU_DEFAULT_OVERPROVISION_FACTOR = 1.25;
+  // Some linkers were having trouble with the code below, so use a function
+  //static inline constexpr double GPU_DEFAULT_OVERPROVISION_FACTOR = 1.25;
+  static constexpr double GPU_DEFAULT_OVERPROVISION_FACTOR() { return 1.25; }
 
   //
   // ------- public API ---------
@@ -87,7 +89,7 @@ class WorkspaceManager
   //   policy: The team policy for Kokkos kernels using this WorkspaceManager
   //   overprov_factor: How many workspace slots to overprovision (only applies to GPU for large problems)
   WorkspaceManager(int size, int max_used, TeamPolicy policy,
-                   const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR);
+                   const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR());
 
   // Constructor, call from host
   //   Same as above, but here the user initializes the data.
@@ -97,14 +99,14 @@ class WorkspaceManager
   //   data is available, for which the get_total_slots_to_be_used()
   //   function can be helpful.
   WorkspaceManager(T* data, int size, int max_used, TeamPolicy policy,
-                   const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR);
+                   const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR());
 
   // Helper functions which return the number of bytes that will be reserved for a given
   // set of constructor inputs. Note, this does not actually create an instance of the WSM,
   // but is useful for when memory needs to be reserved in a different scope than the
   // WSM is created.
   static int get_total_bytes_needed(int size, int max_used, TeamPolicy policy,
-                                    const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR);
+                                    const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR());
 
   // call from host.
   //
@@ -116,14 +118,14 @@ class WorkspaceManager
   //
   // Setup routine for the WSM if the user used the empty constructor
   void setup(int size, int max_used, TeamPolicy policy,
-             const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR);
+             const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR());
 
   // call from host.
   //
   // Setup routine for the WSM if the user used the empty constructor.
   // Same as above, but here the user initializes the data.
   void setup(T* data, int size, int max_used, TeamPolicy policy,
-             const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR);
+             const double& overprov_factor=GPU_DEFAULT_OVERPROVISION_FACTOR());
 
   // call from host.
   //
@@ -377,9 +379,6 @@ class WorkspaceManager
   void operator() (const MemberType& team) const;
 }; // class WorkspaceManager
 
-template <typename T, typename D>
-constexpr double WorkspaceManager<T, D>::GPU_DEFAULT_OVERPROVISION_FACTOR;
-
 } // namespace ekat
 
 #include "ekat_workspace_impl.hpp"

tests/kokkos/workspace_mgr.cpp

@@ -39,7 +39,7 @@ static void unittest_workspace_overprovision()
   TeamUtils<double,ExeSpace> tu_temp(temp_policy);
   const int num_conc = tu_temp.get_max_concurrent_threads() / temp_policy.team_size();
 
-  constexpr double op_fact = WSM::GPU_DEFAULT_OVERPROVISION_FACTOR;
+  constexpr double op_fact = WSM::GPU_DEFAULT_OVERPROVISION_FACTOR();
   constexpr double explicit_op_fact = op_fact * 2.0;
 
   const int ni_under   = (num_conc / 2) + 1;