Cache GSL in SyncPlan and apply async Sync to Z4c_class

Major optimization: Pre-build grid segment lists (GSLs) once per Step() call
via SyncPreparePlan(), then reuse them across all 4 RK4 substep SyncBegin calls
via SyncBeginWithPlan(). This eliminates the O(cpusize * blocks^2) GSL rebuild
cost that was incurred on every ghost zone exchange.

Applied async SyncBegin/SyncEnd overlap pattern to Z4c_class.C (ABEtype==2,
the default configuration), which was still using blocking Parallel::Sync.
Both the regular and CPBC variants of Z4c Step() are now optimized.

Co-authored-by: copilot-swe-agent[bot] <198982749+copilot@users.noreply.github.com>

AMSS_NCKU_source/Parallel.C

@@ -3869,6 +3869,7 @@ Parallel::SyncHandle *Parallel::SyncBegin(Patch *Pat, MyList<var> *VarList, int
   ts->VarList1 = VarList;
   ts->VarList2 = VarList;
   ts->Symmetry = Symmetry;
+  ts->owns_gsl = true;
 
   ts->dst = build_ghost_gsl(Pat);
   ts->src = new MyList<Parallel::gridseg> *[cpusize];
@@ -3909,6 +3910,7 @@ Parallel::SyncHandle *Parallel::SyncBegin(MyList<Patch> *PatL, MyList<var> *VarL
     ts->VarList1 = VarList;
     ts->VarList2 = VarList;
     ts->Symmetry = Symmetry;
+    ts->owns_gsl = true;
 
     ts->dst = build_ghost_gsl(Pp->data);
     ts->src = new MyList<Parallel::gridseg> *[cpusize];
@@ -3933,6 +3935,7 @@ Parallel::SyncHandle *Parallel::SyncBegin(MyList<Patch> *PatL, MyList<var> *VarL
     ts->VarList1 = VarList;
     ts->VarList2 = VarList;
     ts->Symmetry = Symmetry;
+    ts->owns_gsl = true;
 
     ts->dst = build_buffer_gsl(PatL);
     ts->src = new MyList<Parallel::gridseg> *[cpusize];
@@ -3961,7 +3964,9 @@ void Parallel::SyncEnd(SyncHandle *handle)
     TransferState *ts = &handle->states[i];
     transfer_end(ts);
 
-    // Cleanup grid segment lists
+    // Cleanup grid segment lists only if this state owns them
+    if (ts->owns_gsl)
+    {
       if (ts->dst)
         ts->dst->destroyList();
       for (int node = 0; node < ts->cpusize; node++)
@@ -3977,10 +3982,132 @@ void Parallel::SyncEnd(SyncHandle *handle)
       delete[] ts->transfer_src;
       delete[] ts->transfer_dst;
     }
+  }
 
   delete[] handle->states;
   delete handle;
 }
+//
+// SyncPreparePlan: Pre-build grid segment lists for a patch list.
+// The plan can be reused across multiple SyncBeginWithPlan calls
+// as long as the mesh topology does not change (no regridding).
+//
+Parallel::SyncPlan *Parallel::SyncPreparePlan(MyList<Patch> *PatL, int Symmetry)
+{
+  int cpusize;
+  MPI_Comm_size(MPI_COMM_WORLD, &cpusize);
+
+  // Count patches
+  int num_patches = 0;
+  MyList<Patch> *Pp = PatL;
+  while (Pp) { num_patches++; Pp = Pp->next; }
+
+  SyncPlan *plan = new SyncPlan;
+  plan->num_entries = num_patches + 1; // intra-patch + 1 inter-patch
+  plan->Symmetry = Symmetry;
+  plan->entries = new SyncPlanEntry[plan->num_entries];
+
+  // Intra-patch entries: ghost zone exchange within each patch
+  int idx = 0;
+  Pp = PatL;
+  while (Pp)
+  {
+    SyncPlanEntry *pe = &plan->entries[idx];
+    pe->cpusize = cpusize;
+    pe->dst = build_ghost_gsl(Pp->data);
+    pe->src = new MyList<Parallel::gridseg> *[cpusize];
+    pe->transfer_src = new MyList<Parallel::gridseg> *[cpusize];
+    pe->transfer_dst = new MyList<Parallel::gridseg> *[cpusize];
+    for (int node = 0; node < cpusize; node++)
+    {
+      pe->src[node] = build_owned_gsl0(Pp->data, node);
+      build_gstl(pe->src[node], pe->dst, &pe->transfer_src[node], &pe->transfer_dst[node]);
+    }
+    idx++;
+    Pp = Pp->next;
+  }
+
+  // Inter-patch entry: buffer zone exchange between patches
+  {
+    SyncPlanEntry *pe = &plan->entries[idx];
+    pe->cpusize = cpusize;
+    pe->dst = build_buffer_gsl(PatL);
+    pe->src = new MyList<Parallel::gridseg> *[cpusize];
+    pe->transfer_src = new MyList<Parallel::gridseg> *[cpusize];
+    pe->transfer_dst = new MyList<Parallel::gridseg> *[cpusize];
+    for (int node = 0; node < cpusize; node++)
+    {
+      pe->src[node] = build_owned_gsl(PatL, node, 5, Symmetry);
+      build_gstl(pe->src[node], pe->dst, &pe->transfer_src[node], &pe->transfer_dst[node]);
+    }
+  }
+
+  return plan;
+}
+//
+void Parallel::SyncFreePlan(SyncPlan *plan)
+{
+  if (!plan)
+    return;
+
+  for (int i = 0; i < plan->num_entries; i++)
+  {
+    SyncPlanEntry *pe = &plan->entries[i];
+    if (pe->dst)
+      pe->dst->destroyList();
+    for (int node = 0; node < pe->cpusize; node++)
+    {
+      if (pe->src[node])
+        pe->src[node]->destroyList();
+      if (pe->transfer_src[node])
+        pe->transfer_src[node]->destroyList();
+      if (pe->transfer_dst[node])
+        pe->transfer_dst[node]->destroyList();
+    }
+    delete[] pe->src;
+    delete[] pe->transfer_src;
+    delete[] pe->transfer_dst;
+  }
+  delete[] plan->entries;
+  delete plan;
+}
+//
+// SyncBeginWithPlan: Use pre-built GSLs from a SyncPlan to initiate async transfer.
+// This avoids the O(cpusize * blocks^2) cost of rebuilding GSLs on every call.
+//
+Parallel::SyncHandle *Parallel::SyncBeginWithPlan(SyncPlan *plan, MyList<var> *VarList)
+{
+  return SyncBeginWithPlan(plan, VarList, VarList);
+}
+//
+Parallel::SyncHandle *Parallel::SyncBeginWithPlan(SyncPlan *plan, MyList<var> *VarList1, MyList<var> *VarList2)
+{
+  SyncHandle *handle = new SyncHandle;
+  handle->num_states = plan->num_entries;
+  handle->states = new TransferState[handle->num_states];
+
+  for (int i = 0; i < plan->num_entries; i++)
+  {
+    SyncPlanEntry *pe = &plan->entries[i];
+    TransferState *ts = &handle->states[i];
+
+    ts->cpusize = pe->cpusize;
+    ts->VarList1 = VarList1;
+    ts->VarList2 = VarList2;
+    ts->Symmetry = plan->Symmetry;
+    ts->owns_gsl = false; // GSLs are owned by the plan, not this handle
+
+    // Borrow GSL pointers from the plan (do NOT free them in SyncEnd)
+    ts->transfer_src = pe->transfer_src;
+    ts->transfer_dst = pe->transfer_dst;
+    ts->src = pe->src;
+    ts->dst = pe->dst;
+
+    transfer_begin(ts);
+  }
+
+  return handle;
+}
 // collect buffer grid segments or blocks for the periodic boundary condition of given patch
 // ---------------------------------------------------
 // |con |                                       |con |

AMSS_NCKU_source/Parallel.h

@@ -98,6 +98,7 @@ namespace Parallel
     MyList<var> *VarList1;
     MyList<var> *VarList2;
     int Symmetry;
+    bool owns_gsl; // true if this state owns and should free the GSLs
   };
   struct SyncHandle
   {
@@ -107,6 +108,26 @@ namespace Parallel
   SyncHandle *SyncBegin(Patch *Pat, MyList<var> *VarList, int Symmetry);
   SyncHandle *SyncBegin(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
   void SyncEnd(SyncHandle *handle);
+
+  // Cached GSL plan: pre-build grid segment lists once, reuse across multiple Sync calls
+  struct SyncPlanEntry
+  {
+    int cpusize;
+    MyList<gridseg> **transfer_src;
+    MyList<gridseg> **transfer_dst;
+    MyList<gridseg> **src;
+    MyList<gridseg> *dst;
+  };
+  struct SyncPlan
+  {
+    SyncPlanEntry *entries;
+    int num_entries;
+    int Symmetry;
+  };
+  SyncPlan *SyncPreparePlan(MyList<Patch> *PatL, int Symmetry);
+  void SyncFreePlan(SyncPlan *plan);
+  SyncHandle *SyncBeginWithPlan(SyncPlan *plan, MyList<var> *VarList);
+  SyncHandle *SyncBeginWithPlan(SyncPlan *plan, MyList<var> *VarList1, MyList<var> *VarList2);
   void OutBdLow2Hi(Patch *Patc, Patch *Patf,
                    MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
                    int Symmetry);

AMSS_NCKU_source/Z4c_class.C

@@ -186,6 +186,12 @@ void Z4c_class::Step(int lev, int YN)
   int ERROR = 0;
 
   MyList<ss_patch> *sPp;
+
+  // Pre-build grid segment lists once for this level's patches.
+  // These are reused across predictor + 3 corrector SyncBegin calls,
+  // avoiding O(cpusize * blocks^2) rebuild each time.
+  Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
+
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
@@ -321,13 +327,17 @@ void Z4c_class::Step(int lev, int YN)
     }
     Pp = Pp->next;
   }
-  // check error information
+  // Start async ghost zone exchange - overlaps with error check and Shell computation
+  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
+
+  // check error information (overlaps with MPI transfer)
   {
     int erh = ERROR;
     MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
   }
   if (ERROR)
   {
+    Parallel::SyncEnd(sync_pre); sync_pre = 0;
     Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
     if (myrank == 0)
     {
@@ -475,6 +485,7 @@ void Z4c_class::Step(int lev, int YN)
   }
   if (ERROR)
   {
+    Parallel::SyncEnd(sync_pre); sync_pre = 0;
     SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
     if (myrank == 0)
     {
@@ -485,7 +496,8 @@ void Z4c_class::Step(int lev, int YN)
   }
 #endif
 
-  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
+  // Complete async ghost zone exchange
+  if (sync_pre) Parallel::SyncEnd(sync_pre);
 
 #ifdef WithShell
   if (lev == 0)
@@ -693,13 +705,17 @@ void Z4c_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // check error information
+    // Start async ghost zone exchange - overlaps with error check and Shell computation
+    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
+
+    // check error information (overlaps with MPI transfer)
     {
       int erh = ERROR;
       MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
     }
     if (ERROR)
     {
+      Parallel::SyncEnd(sync_cor); sync_cor = 0;
       Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
       if (myrank == 0)
       {
@@ -857,6 +873,7 @@ void Z4c_class::Step(int lev, int YN)
     }
     if (ERROR)
     {
+      Parallel::SyncEnd(sync_cor); sync_cor = 0;
       SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
       if (myrank == 0)
       {
@@ -868,7 +885,8 @@ void Z4c_class::Step(int lev, int YN)
     }
 #endif
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    // Complete async ghost zone exchange
+    if (sync_cor) Parallel::SyncEnd(sync_cor);
 
 #ifdef WithShell
     if (lev == 0)
@@ -1042,6 +1060,8 @@ void Z4c_class::Step(int lev, int YN)
       Porg0[ithBH][2] = Porg1[ithBH][2];
     }
   }
+
+  Parallel::SyncFreePlan(sync_plan);
 }
 #else
 // for constraint preserving boundary (CPBC)
@@ -1075,6 +1095,10 @@ void Z4c_class::Step(int lev, int YN)
   int ERROR = 0;
 
   MyList<ss_patch> *sPp;
+
+  // Pre-build grid segment lists once for this level's patches.
+  Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
+
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
@@ -1542,13 +1566,17 @@ void Z4c_class::Step(int lev, int YN)
   }
 #endif
   }
-  // check error information
+  // Start async ghost zone exchange - overlaps with error check
+  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
+
+  // check error information (overlaps with MPI transfer)
   {
     int erh = ERROR;
     MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
   }
   if (ERROR)
   {
+    Parallel::SyncEnd(sync_pre); sync_pre = 0;
     SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
     if (myrank == 0)
     {
@@ -1558,7 +1586,8 @@ void Z4c_class::Step(int lev, int YN)
     }
   }
 
-  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
+  // Complete async ghost zone exchange
+  if (sync_pre) Parallel::SyncEnd(sync_pre);
 
   if (lev == 0)
   {
@@ -2103,13 +2132,17 @@ void Z4c_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // check error information
+    // Start async ghost zone exchange - overlaps with error check
+    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
+
+    // check error information (overlaps with MPI transfer)
     {
       int erh = ERROR;
       MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
     }
     if (ERROR)
     {
+      Parallel::SyncEnd(sync_cor); sync_cor = 0;
       SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
       if (myrank == 0)
       {
@@ -2120,7 +2153,8 @@ void Z4c_class::Step(int lev, int YN)
       }
     }
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    // Complete async ghost zone exchange
+    if (sync_cor) Parallel::SyncEnd(sync_cor);
 
     if (lev == 0)
     {
@@ -2346,6 +2380,8 @@ void Z4c_class::Step(int lev, int YN)
 	  DG_List->clearList();
 	}
 #endif
+
+  Parallel::SyncFreePlan(sync_plan);
 }
 #endif
 #undef MRBD

AMSS_NCKU_source/bssn_class.C

@@ -3035,6 +3035,12 @@ void bssn_class::Step(int lev, int YN)
   int ERROR = 0;
 
   MyList<ss_patch> *sPp;
+
+  // Pre-build grid segment lists once for this level's patches.
+  // These are reused across predictor + 3 corrector SyncBegin calls,
+  // avoiding O(cpusize * blocks^2) rebuild each time.
+  Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
+
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
@@ -3160,7 +3166,7 @@ void bssn_class::Step(int lev, int YN)
   }
 
   // Start async ghost zone exchange - overlaps with error check and Shell computation
-  Parallel::SyncHandle *sync_pre = Parallel::SyncBegin(GH->PatL[lev], SynchList_pre, Symmetry);
+  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
 
   // check error information (overlaps with MPI transfer)
   {
@@ -3536,7 +3542,7 @@ void bssn_class::Step(int lev, int YN)
     }
 
     // Start async ghost zone exchange - overlaps with error check and Shell computation
-    Parallel::SyncHandle *sync_cor = Parallel::SyncBegin(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
 
     // check error information (overlaps with MPI transfer)
     {
@@ -3908,6 +3914,8 @@ void bssn_class::Step(int lev, int YN)
       Porg0[ithBH][2] = Porg1[ithBH][2];
     }
   }
+
+  Parallel::SyncFreePlan(sync_plan);
 }
 
 //================================================================================================
@@ -4830,6 +4838,12 @@ void bssn_class::Step(int lev, int YN)
   int ERROR = 0;
 
   MyList<ss_patch> *sPp;
+
+  // Pre-build grid segment lists once for this level's patches.
+  // These are reused across predictor + 3 corrector SyncBegin calls,
+  // avoiding O(cpusize * blocks^2) rebuild each time.
+  Parallel::SyncPlan *sync_plan = Parallel::SyncPreparePlan(GH->PatL[lev], Symmetry);
+
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
@@ -4957,7 +4971,7 @@ void bssn_class::Step(int lev, int YN)
   //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Predictor rhs calculation");
 
   // Start async ghost zone exchange - overlaps with error check and BH position
-  Parallel::SyncHandle *sync_pre = Parallel::SyncBegin(GH->PatL[lev], SynchList_pre, Symmetry);
+  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
 
   // check error information (overlaps with MPI transfer)
   {
@@ -5159,7 +5173,7 @@ void bssn_class::Step(int lev, int YN)
     //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector error check");
 
     // Start async ghost zone exchange - overlaps with error check and BH position
-    Parallel::SyncHandle *sync_cor = Parallel::SyncBegin(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
 
     // check error information (overlaps with MPI transfer)
     {
@@ -5299,6 +5313,8 @@ void bssn_class::Step(int lev, int YN)
 
   //     if(myrank==GH->start_rank[lev]) cout<<GH->mylev<<endl;
   //     misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"complet GH Step");
+
+  Parallel::SyncFreePlan(sync_plan);
 }
 
 //================================================================================================