Checkpoint Z4C CUDA throttling progress

AMSS_NCKU_source/Parallel.C

@@ -5319,6 +5319,10 @@ int Parallel::data_packer(double *data, MyList<Parallel::gridseg> *src, MyList<P
                     MPI_Abort(MPI_COMM_WORLD, 1);
                   }
                 }
+                else if (z4c_cuda_has_resident_state(dst->data->Bg) != 0)
+                {
+                  z4c_cuda_release_step_ctx(dst->data->Bg);
+                }
               }
               else
 #endif

AMSS_NCKU_source/Z4c_class.C

@@ -567,6 +567,34 @@ bool z4c_cuda_refresh_constraint_level(MyList<Patch> *PatL,
   return all_resident;
 }
 
+long long &z4c_constraint_output_counter()
+{
+  static long long counter = 0;
+  return counter;
+}
+
+int z4c_constraint_output_every()
+{
+  static int every = -1;
+  if (every < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_Z4C_CONSTRAINT_EVERY");
+    every = (env && atoi(env) > 0) ? atoi(env) : 1;
+  }
+  return every;
+}
+
+bool z4c_constraint_output_due_now()
+{
+  const int every = z4c_constraint_output_every();
+  return every <= 1 || (z4c_constraint_output_counter() % every) == 0;
+}
+
+void z4c_constraint_output_advance()
+{
+  z4c_constraint_output_counter()++;
+}
+
 } // namespace
 #endif
 
@@ -586,7 +614,8 @@ void Z4c_class::Step(int lev, int YN)
   int pre = 0, cor = 1;
   int ERROR = 0;
   const double dT_mon = dT * pow(0.5, Mymax(0, trfls));
-  const bool need_constraint_after_step = (LastConsOut + dT_mon >= AnasTime);
+  const bool need_constraint_after_step =
+      (LastConsOut + dT_mon >= AnasTime) && z4c_constraint_output_due_now();
 
   if (BH_num > 0 && lev == GH->levels - 1)
   {