fixup! Fix load explosion: use subprocess for binary data plots to avoid thread conflict

* Seems we don't have to set so many variables, `OMP_NUM_THREADS` is enough.

Test: Annotate the code for setting other environment variables. It runs normally.

AMSS_NCKU_Program.py

@@ -66,8 +66,7 @@ if os.path.exists(File_directory):
     ## Prompt whether to overwrite the existing directory
     while True:
         try:
-            ## inputvalue = input()
+            inputvalue = input()
-            inputvalue = "continue"
             ## If the user agrees to overwrite, proceed and remove the existing directory
             if ( inputvalue == "continue" ):
                 print( " Continue the calculation !!! " )

AMSS_NCKU_source/MPatch.C

@@ -341,9 +341,8 @@ void Patch::Interp_Points(MyList<var> *VarList,
                           double *Shellf, int Symmetry)
 {
   // NOTE: we do not Synchnize variables here, make sure of that before calling this routine
-  int myrank, nprocs;
+  int myrank;
   MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
-  MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
 
   int ordn = 2 * ghost_width;
   MyList<var> *varl;
@@ -355,18 +354,24 @@ void Patch::Interp_Points(MyList<var> *VarList,
     varl = varl->next;
   }
 
-  memset(Shellf, 0, sizeof(double) * NN * num_var);
+  double *shellf;
+  shellf = new double[NN * num_var];
+  memset(shellf, 0, sizeof(double) * NN * num_var);
 
-  // owner_rank[j] records which MPI rank owns point j
+  // we use weight to monitor code, later some day we can move it for optimization
-  // All ranks traverse the same block list so they all agree on ownership
+  int *weight;
-  int *owner_rank;
+  weight = new int[NN];
-  owner_rank = new int[NN];
+  memset(weight, 0, sizeof(int) * NN);
-  for (int j = 0; j < NN; j++)
+
-    owner_rank[j] = -1;
+  double *DH, *llb, *uub;
+  DH = new double[dim];
 
-  double DH[dim], llb[dim], uub[dim];
   for (int i = 0; i < dim; i++)
+  {
     DH[i] = getdX(i);
+  }
+  llb = new double[dim];
+  uub = new double[dim];
 
   for (int j = 0; j < NN; j++) // run along points
   {
@@ -398,6 +403,12 @@ void Patch::Interp_Points(MyList<var> *VarList,
       bool flag = true;
       for (int i = 0; i < dim; i++)
       {
+// NOTE: our dividing structure is (exclude ghost)
+// -1 0
+//       1  2
+// so (0,1) does not belong to any part for vertex structure
+// here we put (0,0.5) to left part and (0.5,1) to right part
+// BUT for cell structure the bbox is (-1.5,0.5) and (0.5,2.5), there is no missing region at all
 #ifdef Vertex
 #ifdef Cell
 #error Both Cell and Vertex are defined
@@ -422,7 +433,6 @@ void Patch::Interp_Points(MyList<var> *VarList,
       if (flag)
       {
         notfind = false;
-        owner_rank[j] = BP->rank;
         if (myrank == BP->rank)
         {
           //---> interpolation
@@ -430,11 +440,14 @@ void Patch::Interp_Points(MyList<var> *VarList,
           int k = 0;
           while (varl) // run along variables
           {
-            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
+            //              shellf[j*num_var+k] = Parallel::global_interp(dim,BP->shape,BP->X,BP->fgfs[varl->data->sgfn],
+            //	  		                                    pox,ordn,varl->data->SoA,Symmetry);
+            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], shellf[j * num_var + k],
                             pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
             varl = varl->next;
             k++;
           }
+          weight[j] = 1;
         }
       }
       if (Bp == ble)
@@ -442,390 +455,104 @@ void Patch::Interp_Points(MyList<var> *VarList,
       Bp = Bp->next;
     }
   }
-  // Replace MPI_Allreduce with per-owner MPI_Bcast:
+
-  // Group consecutive points by owner rank and broadcast each group.
+  MPI_Allreduce(shellf, Shellf, NN * num_var, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-  // Since each point's data is non-zero only on the owner rank,
+  int *Weight;
-  // Bcast from owner is equivalent to Allreduce(MPI_SUM) but much cheaper.
+  Weight = new int[NN];
+  MPI_Allreduce(weight, Weight, NN, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+
+  //  misc::tillherecheck("print me");
+
+  for (int i = 0; i < NN; i++)
   {
-    int j = 0;
+    if (Weight[i] > 1)
-    while (j < NN)
     {
-      int cur_owner = owner_rank[j];
+      if (myrank == 0)
-      if (cur_owner < 0)
+        cout << "WARNING: Patch::Interp_Points meets multiple weight" << endl;
-      {
+      for (int j = 0; j < num_var; j++)
-        if (myrank == 0)
+        Shellf[j + i * num_var] = Shellf[j + i * num_var] / Weight[i];
-        {
-          cout << "ERROR: Patch::Interp_Points fails to find point (";
-          for (int d = 0; d < dim; d++)
-          {
-            cout << XX[d][j];
-            if (d < dim - 1)
-              cout << ",";
-            else
-              cout << ")";
-          }
-          cout << " on Patch (";
-          for (int d = 0; d < dim; d++)
-          {
-            cout << bbox[d] << "+" << lli[d] * DH[d];
-            if (d < dim - 1)
-              cout << ",";
-            else
-              cout << ")--";
-          }
-          cout << "(";
-          for (int d = 0; d < dim; d++)
-          {
-            cout << bbox[dim + d] << "-" << uui[d] * DH[d];
-            if (d < dim - 1)
-              cout << ",";
-            else
-              cout << ")" << endl;
-          }
-          MPI_Abort(MPI_COMM_WORLD, 1);
-        }
-        j++;
-        continue;
-      }
-      // Find contiguous run of points with the same owner
-      int jstart = j;
-      while (j < NN && owner_rank[j] == cur_owner)
-        j++;
-      int count = (j - jstart) * num_var;
-      MPI_Bcast(Shellf + jstart * num_var, count, MPI_DOUBLE, cur_owner, MPI_COMM_WORLD);
     }
-  }
+    else if (Weight[i] == 0 && myrank == 0)
-
-  delete[] owner_rank;
-}
-void Patch::Interp_Points(MyList<var> *VarList,
-                          int NN, double **XX,
-                          double *Shellf, int Symmetry,
-                          int Nmin_consumer, int Nmax_consumer)
-{
-  // Targeted point-to-point overload: each owner sends each point only to
-  // the one rank that needs it for integration (consumer), reducing
-  // communication volume by ~nprocs times compared to the Bcast version.
-  /*
-  double t_calc_end, t_calc_total = 0;
-  double t_calc_start = MPI_Wtime();*/
-  int myrank, nprocs;
-  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
-  MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
-
-  int ordn = 2 * ghost_width;
-  MyList<var> *varl;
-  int num_var = 0;
-  varl = VarList;
-  while (varl)
-  {
-    num_var++;
-    varl = varl->next;
-  }
-
-  memset(Shellf, 0, sizeof(double) * NN * num_var);
-
-  // owner_rank[j] records which MPI rank owns point j
-  int *owner_rank;
-  owner_rank = new int[NN];
-  for (int j = 0; j < NN; j++)
-    owner_rank[j] = -1;
-
-  double DH[dim], llb[dim], uub[dim];
-  for (int i = 0; i < dim; i++)
-    DH[i] = getdX(i);
-
-  // --- Interpolation phase (identical to original) ---
-  for (int j = 0; j < NN; j++)
-  {
-    double pox[dim];
-    for (int i = 0; i < dim; i++)
     {
-      pox[i] = XX[i][j];
-      if (myrank == 0 && (XX[i][j] < bbox[i] + lli[i] * DH[i] || XX[i][j] > bbox[dim + i] - uui[i] * DH[i]))
-      {
-        cout << "Patch::Interp_Points: point (";
-        for (int k = 0; k < dim; k++)
-        {
-          cout << XX[k][j];
-          if (k < dim - 1)
-            cout << ",";
-          else
-            cout << ") is out of current Patch." << endl;
-        }
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
-
-    MyList<Block> *Bp = blb;
-    bool notfind = true;
-    while (notfind && Bp)
-    {
-      Block *BP = Bp->data;
-
-      bool flag = true;
-      for (int i = 0; i < dim; i++)
-      {
-#ifdef Vertex
-#ifdef Cell
-#error Both Cell and Vertex are defined
-#endif
-        llb[i] = (feq(BP->bbox[i], bbox[i], DH[i] / 2)) ? BP->bbox[i] + lli[i] * DH[i] : BP->bbox[i] + (ghost_width - 0.5) * DH[i];
-        uub[i] = (feq(BP->bbox[dim + i], bbox[dim + i], DH[i] / 2)) ? BP->bbox[dim + i] - uui[i] * DH[i] : BP->bbox[dim + i] - (ghost_width - 0.5) * DH[i];
-#else
-#ifdef Cell
-        llb[i] = (feq(BP->bbox[i], bbox[i], DH[i] / 2)) ? BP->bbox[i] + lli[i] * DH[i] : BP->bbox[i] + ghost_width * DH[i];
-        uub[i] = (feq(BP->bbox[dim + i], bbox[dim + i], DH[i] / 2)) ? BP->bbox[dim + i] - uui[i] * DH[i] : BP->bbox[dim + i] - ghost_width * DH[i];
-#else
-#error Not define Vertex nor Cell
-#endif
-#endif
-        if (XX[i][j] - llb[i] < -DH[i] / 2 || XX[i][j] - uub[i] > DH[i] / 2)
-        {
-          flag = false;
-          break;
-        }
-      }
-
-      if (flag)
-      {
-        notfind = false;
-        owner_rank[j] = BP->rank;
-        if (myrank == BP->rank)
-        {
-          varl = VarList;
-          int k = 0;
-          while (varl)
-          {
-            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
-                            pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
-            varl = varl->next;
-            k++;
-          }
-        }
-      }
-      if (Bp == ble)
-        break;
-      Bp = Bp->next;
-    }
-  }
-  /*
-      t_calc_end = MPI_Wtime();
-      t_calc_total = t_calc_end - t_calc_start;*/
-  // --- Error check for unfound points ---
-  for (int j = 0; j < NN; j++)
-  {
-    if (owner_rank[j] < 0 && myrank == 0)
-    {
-      cout<<owner_rank[j-1]<<endl;
       cout << "ERROR: Patch::Interp_Points fails to find point (";
-      for (int d = 0; d < dim; d++)
+      for (int j = 0; j < dim; j++)
       {
-        cout << XX[d][j];
+        cout << XX[j][i];
-        if (d < dim - 1)
+        if (j < dim - 1)
           cout << ",";
         else
           cout << ")";
       }
       cout << " on Patch (";
-      for (int d = 0; d < dim; d++)
+      for (int j = 0; j < dim; j++)
       {
-        cout << bbox[d] << "+" << lli[d] * DH[d];
+        cout << bbox[j] << "+" << lli[j] * getdX(j);
-        if (d < dim - 1)
+        if (j < dim - 1)
           cout << ",";
         else
           cout << ")--";
       }
       cout << "(";
-      for (int d = 0; d < dim; d++)
+      for (int j = 0; j < dim; j++)
       {
-        cout << bbox[dim + d] << "-" << uui[d] * DH[d];
+        cout << bbox[dim + j] << "-" << uui[j] * getdX(j);
-        if (d < dim - 1)
+        if (j < dim - 1)
           cout << ",";
         else
           cout << ")" << endl;
       }
+#if 0
+       checkBlock();
+#else
+      cout << "splited domains:" << endl;
+      {
+        MyList<Block> *Bp = blb;
+        while (Bp)
+        {
+          Block *BP = Bp->data;
+
+          for (int i = 0; i < dim; i++)
+          {
+#ifdef Vertex
+#ifdef Cell
+#error Both Cell and Vertex are defined
+#endif
+            llb[i] = (feq(BP->bbox[i], bbox[i], DH[i] / 2)) ? BP->bbox[i] + lli[i] * DH[i] : BP->bbox[i] + (ghost_width - 0.5) * DH[i];
+            uub[i] = (feq(BP->bbox[dim + i], bbox[dim + i], DH[i] / 2)) ? BP->bbox[dim + i] - uui[i] * DH[i] : BP->bbox[dim + i] - (ghost_width - 0.5) * DH[i];
+#else
+#ifdef Cell
+            llb[i] = (feq(BP->bbox[i], bbox[i], DH[i] / 2)) ? BP->bbox[i] + lli[i] * DH[i] : BP->bbox[i] + ghost_width * DH[i];
+            uub[i] = (feq(BP->bbox[dim + i], bbox[dim + i], DH[i] / 2)) ? BP->bbox[dim + i] - uui[i] * DH[i] : BP->bbox[dim + i] - ghost_width * DH[i];
+#else
+#error Not define Vertex nor Cell
+#endif
+#endif
+          }
+          cout << "(";
+          for (int j = 0; j < dim; j++)
+          {
+            cout << llb[j] << ":" << uub[j];
+            if (j < dim - 1)
+              cout << ",";
+            else
+              cout << ")" << endl;
+          }
+          if (Bp == ble)
+            break;
+          Bp = Bp->next;
+        }
+      }
+#endif
       MPI_Abort(MPI_COMM_WORLD, 1);
     }
   }
 
-  // --- Targeted point-to-point communication phase ---
+  delete[] shellf;
-  // Compute consumer_rank[j] using the same deterministic formula as surface_integral
+  delete[] weight;
-  int *consumer_rank = new int[NN];
+  delete[] Weight;
-  {
+  delete[] DH;
-    int mp = NN / nprocs;
+  delete[] llb;
-    int Lp = NN - nprocs * mp;
+  delete[] uub;
-    for (int j = 0; j < NN; j++)
-    {
-      if (j < Lp * (mp + 1))
-        consumer_rank[j] = j / (mp + 1);
-      else
-        consumer_rank[j] = Lp + (j - Lp * (mp + 1)) / mp;
-    }
-  }
-
-  // Count sends and recvs per rank
-  int *send_count = new int[nprocs];
-  int *recv_count = new int[nprocs];
-  memset(send_count, 0, sizeof(int) * nprocs);
-  memset(recv_count, 0, sizeof(int) * nprocs);
-
-  for (int j = 0; j < NN; j++)
-  {
-    int own = owner_rank[j];
-    int con = consumer_rank[j];
-    if (own == con)
-      continue; // local — no communication needed
-    if (own == myrank)
-      send_count[con]++;
-    if (con == myrank)
-      recv_count[own]++;
-  }
-
-  // Build send buffers: for each destination rank, pack (index, data) pairs
-  // Each entry: 1 int (point index j) + num_var doubles
-  int total_send = 0, total_recv = 0;
-  int *send_offset = new int[nprocs];
-  int *recv_offset = new int[nprocs];
-  for (int r = 0; r < nprocs; r++)
-  {
-    send_offset[r] = total_send;
-    total_send += send_count[r];
-    recv_offset[r] = total_recv;
-    total_recv += recv_count[r];
-  }
-
-  // Pack send buffers: each message contains (j, data[0..num_var-1]) per point
-  int stride = 1 + num_var; // 1 double for index + num_var doubles for data
-  double *sendbuf = new double[total_send * stride];
-  double *recvbuf = new double[total_recv * stride];
-
-  // Temporary counters for packing
-  int *pack_pos = new int[nprocs];
-  memset(pack_pos, 0, sizeof(int) * nprocs);
-
-  for (int j = 0; j < NN; j++)
-  {
-    int own = owner_rank[j];
-    int con = consumer_rank[j];
-    if (own != myrank || con == myrank)
-      continue;
-    int pos = (send_offset[con] + pack_pos[con]) * stride;
-    sendbuf[pos] = (double)j; // point index
-    for (int v = 0; v < num_var; v++)
-      sendbuf[pos + 1 + v] = Shellf[j * num_var + v];
-    pack_pos[con]++;
-  }
-
-  // Post non-blocking recvs and sends
-  int n_req = 0;
-  for (int r = 0; r < nprocs; r++)
-  {
-    if (recv_count[r] > 0) n_req++;
-    if (send_count[r] > 0) n_req++;
-  }
-
-  MPI_Request *reqs = new MPI_Request[n_req];
-  int req_idx = 0;
-
-  for (int r = 0; r < nprocs; r++)
-  {
-    if (recv_count[r] > 0)
-    {
-      MPI_Irecv(recvbuf + recv_offset[r] * stride,
-                recv_count[r] * stride, MPI_DOUBLE,
-                r, 0, MPI_COMM_WORLD, &reqs[req_idx++]);
-    }
-  }
-  for (int r = 0; r < nprocs; r++)
-  {
-    if (send_count[r] > 0)
-    {
-      MPI_Isend(sendbuf + send_offset[r] * stride,
-                send_count[r] * stride, MPI_DOUBLE,
-                r, 0, MPI_COMM_WORLD, &reqs[req_idx++]);
-    }
-  }
-
-  if (n_req > 0)
-    MPI_Waitall(n_req, reqs, MPI_STATUSES_IGNORE);
-
-  // Unpack recv buffers into Shellf
-  for (int i = 0; i < total_recv; i++)
-  {
-    int pos = i * stride;
-    int j = (int)recvbuf[pos];
-    for (int v = 0; v < num_var; v++)
-      Shellf[j * num_var + v] = recvbuf[pos + 1 + v];
-  }
-
-  delete[] reqs;
-  delete[] sendbuf;
-  delete[] recvbuf;
-  delete[] pack_pos;
-  delete[] send_offset;
-  delete[] recv_offset;
-  delete[] send_count;
-  delete[] recv_count;
-  delete[] consumer_rank;
-  delete[] owner_rank;
-  /*
-  // 4. 汇总并输出真正干活最慢的 Top 4
-  struct RankStats {
-    int rank;
-    double calc_time; // 净计算时间
-  };
-
-  // 创建当前进程的统计数据
-  RankStats local_stat;
-  local_stat.rank = myrank;
-  local_stat.calc_time = t_calc_total;
-
-  // 为所有进程的统计数据分配内存
-  RankStats *all_stats = nullptr;
-  if (myrank == 0) {
-    all_stats = new RankStats[nprocs];
-  }
-
-  // 使用MPI_Gather收集所有进程的数据到rank 0
-  MPI_Gather(&local_stat, sizeof(RankStats), MPI_BYTE,
-             all_stats, sizeof(RankStats), MPI_BYTE,
-             0, MPI_COMM_WORLD);
-
-  // 准备输出前4个rank的信息（所有rank都参与，确保广播后一致）
-  int top10_ranks[10] = { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 };
-  double top10_times[10] = { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 };
-  int num_top10 = 0;
-
-  if (myrank == 0) {
-    // 按 calc_time（净计算时间）排序
-    std::sort(all_stats, all_stats + nprocs, [](const RankStats& a, const RankStats& b) {
-        return a.calc_time > b.calc_time;
-    });
-
-    // 取前4个
-    num_top10 = (nprocs < 10) ? nprocs : 10;
-    for (int i = 0; i < num_top10; i++) {
-      top10_ranks[i] = all_stats[i].rank;
-      top10_times[i] = all_stats[i].calc_time;
-    }
-
-    printf("\n--- Top %d Ranks by ACTIVE COMPUTATION (CPU Time) ---\n", num_top10);
-    for (int i = 0; i < num_top10; i++) {
-      printf("Rank [%4d]: Calc %.6f s\n", top10_ranks[i], top10_times[i]);
-    }
-
-    // 清理分配的内存
-    delete[] all_stats;
-  }
-
-  // 广播前4个rank的信息给所有进程
-  MPI_Bcast(&num_top10, 1, MPI_INT, 0, MPI_COMM_WORLD);
-  if (num_top10 > 0) {
-    MPI_Bcast(top10_ranks, 10, MPI_INT, 0, MPI_COMM_WORLD);
-    MPI_Bcast(top10_times, 10, MPI_DOUBLE, 0, MPI_COMM_WORLD);
-  }
-*/
 }
 void Patch::Interp_Points(MyList<var> *VarList,
                           int NN, double **XX,
@@ -846,22 +573,24 @@ void Patch::Interp_Points(MyList<var> *VarList,
     varl = varl->next;
   }
 
-  memset(Shellf, 0, sizeof(double) * NN * num_var);
+  double *shellf;
+  shellf = new double[NN * num_var];
+  memset(shellf, 0, sizeof(double) * NN * num_var);
 
-  // owner_rank[j] stores the global rank that owns point j
+  // we use weight to monitor code, later some day we can move it for optimization
-  int *owner_rank;
+  int *weight;
-  owner_rank = new int[NN];
+  weight = new int[NN];
-  for (int j = 0; j < NN; j++)
+  memset(weight, 0, sizeof(int) * NN);
-    owner_rank[j] = -1;
 
-  // Build global-to-local rank translation for Comm_here
+  double *DH, *llb, *uub;
-  MPI_Group world_group, local_group;
+  DH = new double[dim];
-  MPI_Comm_group(MPI_COMM_WORLD, &world_group);
-  MPI_Comm_group(Comm_here, &local_group);
 
-  double DH[dim], llb[dim], uub[dim];
   for (int i = 0; i < dim; i++)
+  {
     DH[i] = getdX(i);
+  }
+  llb = new double[dim];
+  uub = new double[dim];
 
   for (int j = 0; j < NN; j++) // run along points
   {
@@ -893,6 +622,12 @@ void Patch::Interp_Points(MyList<var> *VarList,
       bool flag = true;
       for (int i = 0; i < dim; i++)
       {
+// NOTE: our dividing structure is (exclude ghost)
+// -1 0
+//       1  2
+// so (0,1) does not belong to any part for vertex structure
+// here we put (0,0.5) to left part and (0.5,1) to right part
+// BUT for cell structure the bbox is (-1.5,0.5) and (0.5,2.5), there is no missing region at all
 #ifdef Vertex
 #ifdef Cell
 #error Both Cell and Vertex are defined
@@ -917,7 +652,6 @@ void Patch::Interp_Points(MyList<var> *VarList,
       if (flag)
       {
         notfind = false;
-        owner_rank[j] = BP->rank;
         if (myrank == BP->rank)
         {
           //---> interpolation
@@ -925,11 +659,14 @@ void Patch::Interp_Points(MyList<var> *VarList,
           int k = 0;
           while (varl) // run along variables
           {
-            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
+            //              shellf[j*num_var+k] = Parallel::global_interp(dim,BP->shape,BP->X,BP->fgfs[varl->data->sgfn],
+            //	  		                                    pox,ordn,varl->data->SoA,Symmetry);
+            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], shellf[j * num_var + k],
                             pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
             varl = varl->next;
             k++;
           }
+          weight[j] = 1;
         }
       }
       if (Bp == ble)
@@ -938,35 +675,97 @@ void Patch::Interp_Points(MyList<var> *VarList,
     }
   }
 
-  // Collect unique global owner ranks and translate to local ranks in Comm_here
+  MPI_Allreduce(shellf, Shellf, NN * num_var, MPI_DOUBLE, MPI_SUM, Comm_here);
-  // Then broadcast each owner's points via MPI_Bcast on Comm_here
+  int *Weight;
-  {
+  Weight = new int[NN];
-    int j = 0;
+  MPI_Allreduce(weight, Weight, NN, MPI_INT, MPI_SUM, Comm_here);
-    while (j < NN)
-    {
-      int cur_owner_global = owner_rank[j];
-      if (cur_owner_global < 0)
-      {
-        // Point not found — skip (error check disabled for sub-communicator levels)
-        j++;
-        continue;
-      }
-      // Translate global rank to local rank in Comm_here
-      int cur_owner_local;
-      MPI_Group_translate_ranks(world_group, 1, &cur_owner_global, local_group, &cur_owner_local);
 
-      // Find contiguous run of points with the same owner
+  //  misc::tillherecheck("print me");
-      int jstart = j;
+  //  if(lmyrank == 0) cout<<"myrank = "<<myrank<<"print me"<<endl;
-      while (j < NN && owner_rank[j] == cur_owner_global)
+
-        j++;
+  for (int i = 0; i < NN; i++)
-      int count = (j - jstart) * num_var;
+  {
-      MPI_Bcast(Shellf + jstart * num_var, count, MPI_DOUBLE, cur_owner_local, Comm_here);
+    if (Weight[i] > 1)
+    {
+      if (lmyrank == 0)
+        cout << "WARNING: Patch::Interp_Points meets multiple weight" << endl;
+      for (int j = 0; j < num_var; j++)
+        Shellf[j + i * num_var] = Shellf[j + i * num_var] / Weight[i];
     }
+#if 0 // for not involved levels, this may fail     
+     else if(Weight[i] == 0 && lmyrank == 0)
+     {
+       cout<<"ERROR: Patch::Interp_Points fails to find point (";
+       for(int j=0;j<dim;j++)
+       {
+	  cout<<XX[j][i];
+	  if(j<dim-1) cout<<",";
+	  else        cout<<")";
+       }
+       cout<<" on Patch (";
+       for(int j=0;j<dim;j++)
+       {
+	  cout<<bbox[j]<<"+"<<lli[j]*getdX(j);
+	  if(j<dim-1) cout<<",";
+	  else        cout<<")--";
+       }
+       cout<<"(";
+       for(int j=0;j<dim;j++)
+       {
+	  cout<<bbox[dim+j]<<"-"<<uui[j]*getdX(j);
+	  if(j<dim-1) cout<<",";
+	  else        cout<<")"<<endl;
+       }
+#if 0
+       checkBlock();
+#else
+  cout<<"splited domains:"<<endl;
+  {
+     MyList<Block> *Bp=blb;
+     while(Bp)
+     {
+	Block *BP=Bp->data;
+
+	for(int i=0;i<dim;i++)
+	{
+#ifdef Vertex
+#ifdef Cell
+#error Both Cell and Vertex are defined
+#endif    
+          llb[i] = (feq(BP->bbox[i]    ,bbox[i]    ,DH[i]/2)) ? BP->bbox[i]+lli[i]*DH[i]     : BP->bbox[i]    +(ghost_width-0.5)*DH[i];
+          uub[i] = (feq(BP->bbox[dim+i],bbox[dim+i],DH[i]/2)) ? BP->bbox[dim+i]-uui[i]*DH[i] : BP->bbox[dim+i]-(ghost_width-0.5)*DH[i];
+#else
+#ifdef Cell
+          llb[i] = (feq(BP->bbox[i]    ,bbox[i]    ,DH[i]/2)) ? BP->bbox[i]+lli[i]*DH[i]     : BP->bbox[i]    +ghost_width*DH[i];
+          uub[i] = (feq(BP->bbox[dim+i],bbox[dim+i],DH[i]/2)) ? BP->bbox[dim+i]-uui[i]*DH[i] : BP->bbox[dim+i]-ghost_width*DH[i];
+#else
+#error Not define Vertex nor Cell
+#endif
+#endif 
+	}       
+       cout<<"(";
+       for(int j=0;j<dim;j++)
+       {
+	  cout<<llb[j]<<":"<<uub[j];
+	  if(j<dim-1) cout<<",";
+	  else        cout<<")"<<endl;
+       }
+	if(Bp == ble) break;
+	Bp=Bp->next;
+     }
+  }
+#endif       
+       MPI_Abort(MPI_COMM_WORLD,1);
+     }
+#endif
   }
 
-  MPI_Group_free(&world_group);
+  delete[] shellf;
-  MPI_Group_free(&local_group);
+  delete[] weight;
-  delete[] owner_rank;
+  delete[] Weight;
+  delete[] DH;
+  delete[] llb;
+  delete[] uub;
 }
 void Patch::checkBlock()
 {

AMSS_NCKU_source/MPatch.h

@@ -39,10 +39,6 @@ public:
 
    bool Find_Point(double *XX);
 
-   void Interp_Points(MyList<var> *VarList,
-                      int NN, double **XX,
-                      double *Shellf, int Symmetry,
-                      int Nmin_consumer, int Nmax_consumer);
    void Interp_Points(MyList<var> *VarList,
                       int NN, double **XX,
                       double *Shellf, int Symmetry, MPI_Comm Comm_here);

AMSS_NCKU_source/NullShellPatch.h

@@ -24,7 +24,6 @@ using namespace std;
 #endif
 
 #include <mpi.h>
-#include <memory.h>
 #include "MyList.h"
 #include "Block.h"
 #include "Parallel.h"

AMSS_NCKU_source/Parallel.h

@@ -11,7 +11,7 @@
 #include <cmath>
 #include <new>
 using namespace std;
-#include <memory.h>
+
 #include "Parallel_bam.h"
 #include "var.h"
 #include "MPatch.h"
@@ -32,25 +32,6 @@ namespace Parallel
   int partition2(int *nxy, int split_size, int *min_width, int cpusize, int *shape); // special for 2 diemnsions
   int partition3(int *nxyz, int split_size, int *min_width, int cpusize, int *shape);
   MyList<Block> *distribute(MyList<Patch> *PatchLIST, int cpusize, int ingfsi, int fngfs, bool periodic, int nodes = 0); // produce corresponding Blocks
-
-  MyList<Block> *distribute_hard(MyList<Patch> *PatchLIST, int cpusize, int ingfsi, int fngfs, bool periodic, int nodes = 0); // produce corresponding Blocks
-  Block* splitHotspotBlock(MyList<Block>* &BlL, int _dim, 
-                                 int ib0_orig, int ib3_orig, 
-                                 int jb1_orig, int jb4_orig, 
-                                 int kb2_orig, int kb5_orig, 
-                                 Patch* PP, int r_1, int r_2, 
-                                 int ingfsi, int fngfsi, bool periodic,
-                                 Block* &split_first_block, Block* &split_last_block);
-
-Block* splitHotspotBlock(MyList<Block>* &BlL, int _dim, 
-                                 int ib0_orig, int ib3_orig, 
-                                 int jb1_orig, int jb4_orig, 
-                                 int kb2_orig, int kb5_orig, 
-                                 Patch* PP, int r_1, int r_2, int r_3, int r_4,
-                                 int ingfsi, int fngfsi, bool periodic,
-                                 Block* &split_first_block, Block* &split_last_block);
-  Block* createMappedBlock(MyList<Block>* &BlL, int _dim, int* shape, double* bbox,
-                        int block_id, int ingfsi, int fngfsi, int lev); 
   void KillBlocks(MyList<Patch> *PatchLIST);
 
   void setfunction(MyList<Block> *BlL, var *vn, double func(double x, double y, double z));
@@ -100,43 +81,53 @@ Block* splitHotspotBlock(MyList<Block>* &BlL, int _dim,
                    int Symmetry);
   void Sync(Patch *Pat, MyList<var> *VarList, int Symmetry);
   void Sync(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
-  void Sync_merged(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
 
-  struct SyncCache {
+  // Async Sync: overlap MPI communication with computation
-    bool valid;
+  struct TransferState
-    int cpusize;
+  {
-    MyList<gridseg> **combined_src;
-    MyList<gridseg> **combined_dst;
-    int *send_lengths;
-    int *recv_lengths;
-    double **send_bufs;
-    double **recv_bufs;
-    int *send_buf_caps;
-    int *recv_buf_caps;
     MPI_Request *reqs;
     MPI_Status *stats;
-    int max_reqs;
-    bool lengths_valid;
-    SyncCache();
-    void invalidate();
-    void destroy();
-  };
-
-  void Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, SyncCache &cache);
-  void transfer_cached(MyList<gridseg> **src, MyList<gridseg> **dst,
-                       MyList<var> *VarList1, MyList<var> *VarList2,
-                       int Symmetry, SyncCache &cache);
-
-  struct AsyncSyncState {
     int req_no;
-    bool active;
+    double **send_data;
-    AsyncSyncState() : req_no(0), active(false) {}
+    double **rec_data;
+    int cpusize;
+    MyList<gridseg> **transfer_src;
+    MyList<gridseg> **transfer_dst;
+    MyList<gridseg> **src;
+    MyList<gridseg> *dst;
+    MyList<var> *VarList1;
+    MyList<var> *VarList2;
+    int Symmetry;
+    bool owns_gsl; // true if this state owns and should free the GSLs
   };
+  struct SyncHandle
+  {
+    TransferState *states;
+    int num_states;
+  };
+  SyncHandle *SyncBegin(Patch *Pat, MyList<var> *VarList, int Symmetry);
+  SyncHandle *SyncBegin(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
+  void SyncEnd(SyncHandle *handle);
 
-  void Sync_start(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry,
+  // Cached GSL plan: pre-build grid segment lists once, reuse across multiple Sync calls
-                  SyncCache &cache, AsyncSyncState &state);
+  struct SyncPlanEntry
-  void Sync_finish(SyncCache &cache, AsyncSyncState &state,
+  {
-                   MyList<var> *VarList, int Symmetry);
+    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);
@@ -149,15 +140,6 @@ Block* splitHotspotBlock(MyList<Block>* &BlL, int _dim,
   void OutBdLow2Himix(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
                       MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
                       int Symmetry);
-  void Restrict_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
-                       MyList<var> *VarList1, MyList<var> *VarList2,
-                       int Symmetry, SyncCache &cache);
-  void OutBdLow2Hi_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
-                          MyList<var> *VarList1, MyList<var> *VarList2,
-                          int Symmetry, SyncCache &cache);
-  void OutBdLow2Himix_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
-                             MyList<var> *VarList1, MyList<var> *VarList2,
-                             int Symmetry, SyncCache &cache);
   void Prolong(Patch *Patc, Patch *Patf,
                MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
                int Symmetry);
@@ -227,18 +209,6 @@ Block* splitHotspotBlock(MyList<Block>* &BlL, int _dim,
 #if (PSTR == 1 || PSTR == 2 || PSTR == 3)
   MyList<Block> *distribute(MyList<Patch> *PatchLIST, int cpusize, int ingfsi, int fngfsi,
                             bool periodic, int start_rank, int end_rank, int nodes = 0);
-  
-  // Redistribute blocks with time statistics for load balancing
-  MyList<Block> *distribute(MyList<Patch> *PatchLIST, MyList<Block> *OldBlockL,
-                            int cpusize, int ingfsi, int fngfsi,
-                            bool periodic, int start_rank, int end_rank, int nodes = 0);
 #endif
-
+}
-  // Dynamic load balancing: split blocks for heavy ranks
+#endif /*PARALLEL_H */
-    void split_heavy_blocks(MyList<Patch> *PatL, int *heavy_ranks, int num_heavy,
-                            int split_factor, int cpusize, int ingfsi, int fngfsi);
-
-    // Check if load balancing is needed based on interpolation times
-    bool check_load_balance_need(double *rank_times, int nprocs, int &num_heavy, int *heavy_ranks);
-  }
-  #endif /*PARALLEL_H */

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

@@ -730,12 +730,6 @@ void bssn_class::Initialize()
     PhysTime = StartTime;
     Setup_Black_Hole_position();
   }
-
-  // Initialize sync caches (per-level, for predictor and corrector)
-  sync_cache_pre = new Parallel::SyncCache[GH->levels];
-  sync_cache_cor = new Parallel::SyncCache[GH->levels];
-  sync_cache_rp_coarse = new Parallel::SyncCache[GH->levels];
-  sync_cache_rp_fine = new Parallel::SyncCache[GH->levels];
 }
 
 //================================================================================================
@@ -987,32 +981,6 @@ bssn_class::~bssn_class()
   delete Azzz;
 #endif
 
-  // Destroy sync caches before GH
-  if (sync_cache_pre)
-  {
-    for (int i = 0; i < GH->levels; i++)
-      sync_cache_pre[i].destroy();
-    delete[] sync_cache_pre;
-  }
-  if (sync_cache_cor)
-  {
-    for (int i = 0; i < GH->levels; i++)
-      sync_cache_cor[i].destroy();
-    delete[] sync_cache_cor;
-  }
-  if (sync_cache_rp_coarse)
-  {
-    for (int i = 0; i < GH->levels; i++)
-      sync_cache_rp_coarse[i].destroy();
-    delete[] sync_cache_rp_coarse;
-  }
-  if (sync_cache_rp_fine)
-  {
-    for (int i = 0; i < GH->levels; i++)
-      sync_cache_rp_fine[i].destroy();
-    delete[] sync_cache_rp_fine;
-  }
-
   delete GH;
 #ifdef WithShell
   delete SH;
@@ -2213,7 +2181,6 @@ void bssn_class::Evolve(int Steps)
     GH->Regrid(Symmetry, BH_num, Porgbr, Porg0,
                SynchList_cor, OldStateList, StateList, SynchList_pre,
                fgt(PhysTime - dT_mon, StartTime, dT_mon / 2), ErrorMonitor);
-    for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 
 #if (REGLEV == 0 && (PSTR == 1 || PSTR == 2))
@@ -2429,7 +2396,6 @@ void bssn_class::RecursiveStep(int lev)
   GH->Regrid_Onelevel(lev, Symmetry, BH_num, Porgbr, Porg0,
                       SynchList_cor, OldStateList, StateList, SynchList_pre,
                       fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
-  for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 }
 
@@ -2608,7 +2574,6 @@ void bssn_class::ParallelStep()
   GH->Regrid_Onelevel(GH->mylev, Symmetry, BH_num, Porgbr, Porg0,
                       SynchList_cor, OldStateList, StateList, SynchList_pre,
                       fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
-  for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 }
 
@@ -2775,7 +2740,6 @@ void bssn_class::ParallelStep()
         GH->Regrid_Onelevel(lev + 1, Symmetry, BH_num, Porgbr, Porg0,
                             SynchList_cor, OldStateList, StateList, SynchList_pre,
                             fgt(PhysTime - dT_levp1, StartTime, dT_levp1 / 2), ErrorMonitor);
-        for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
         //               a_stream.clear();
         //               a_stream.str("");
@@ -2790,7 +2754,6 @@ void bssn_class::ParallelStep()
       GH->Regrid_Onelevel(lev, Symmetry, BH_num, Porgbr, Porg0,
                           SynchList_cor, OldStateList, StateList, SynchList_pre,
                           fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
-      for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
       //               a_stream.clear();
       //               a_stream.str("");
@@ -2809,7 +2772,6 @@ void bssn_class::ParallelStep()
           GH->Regrid_Onelevel(lev - 1, Symmetry, BH_num, Porgbr, Porg0,
                               SynchList_cor, OldStateList, StateList, SynchList_pre,
                               fgt(PhysTime - dT_lev, StartTime, dT_levm1 / 2), ErrorMonitor);
-          for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
           //               a_stream.clear();
           //               a_stream.str("");
@@ -2825,7 +2787,6 @@ void bssn_class::ParallelStep()
           GH->Regrid_Onelevel(lev - 1, Symmetry, BH_num, Porgbr, Porg0,
                               SynchList_cor, OldStateList, StateList, SynchList_pre,
                               fgt(PhysTime - dT_lev, StartTime, dT_levm1 / 2), ErrorMonitor);
-          for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
           //               a_stream.clear();
           //               a_stream.str("");
@@ -3074,6 +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)
@@ -3197,7 +3164,26 @@ void bssn_class::Step(int lev, int YN)
     }
     Pp = Pp->next;
   }
-  // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
+
+  // 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)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
 
 #ifdef WithShell
   // evolve Shell Patches
@@ -3341,16 +3327,27 @@ void bssn_class::Step(int lev, int YN)
 #endif
   }
 
-  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  // check error information
-  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
+    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)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
   }
 #endif
 
-  Parallel::AsyncSyncState async_pre;
+  // Complete async ghost zone exchange
-  Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
+  if (sync_pre) Parallel::SyncEnd(sync_pre);
 
 #ifdef WithShell
   if (lev == 0)
@@ -3369,23 +3366,6 @@ void bssn_class::Step(int lev, int YN)
     }
   }
 #endif
-  Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
-
-#ifdef WithShell
-  // Complete non-blocking error reduction and check
-  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
-  if (ERROR)
-  {
-    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
-#endif
 
 #if (MAPBH == 0)
   // for black hole position
@@ -3561,7 +3541,28 @@ void bssn_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
+    // 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)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
+                                << " variables at t = " << PhysTime 
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
 
 #ifdef WithShell
     // evolve Shell Patches
@@ -3701,16 +3702,28 @@ void bssn_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    // check error information
-    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
+      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)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" 
+                                << iter_count << " variables at t = " 
+                                << PhysTime << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
     }
 #endif
 
-    Parallel::AsyncSyncState async_cor;
+    // Complete async ghost zone exchange
-    Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
+    if (sync_cor) Parallel::SyncEnd(sync_cor);
 
 #ifdef WithShell
     if (lev == 0)
@@ -3729,25 +3742,6 @@ void bssn_class::Step(int lev, int YN)
       }
     }
 #endif
-    Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
-
-#ifdef WithShell
-    // Complete non-blocking error reduction and check
-    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
-    if (ERROR)
-    {
-      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
-                                << " variables at t = " << PhysTime
-                                << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
-#endif
 
 #if (MAPBH == 0)
     // for black hole position
@@ -3920,6 +3914,8 @@ void bssn_class::Step(int lev, int YN)
       Porg0[ithBH][2] = Porg1[ithBH][2];
     }
   }
+
+  Parallel::SyncFreePlan(sync_plan);
 }
 
 //================================================================================================
@@ -4059,7 +4055,22 @@ void bssn_class::Step(int lev, int YN)
     }
     Pp = Pp->next;
   }
-  // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
+  // check error information
+  {
+    int erh = ERROR;
+    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+  }
+  if (ERROR)
+  {
+    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime 
+                              << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
 
 #ifdef WithShell
   // evolve Shell Patches
@@ -4200,16 +4211,25 @@ void bssn_class::Step(int lev, int YN)
   }
 #endif
   }
-  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  // check error information
-  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
+    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+  }
+  if (ERROR)
+  {
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " 
+                              << PhysTime << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
   }
 #endif
 
-  Parallel::AsyncSyncState async_pre;
+  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
-  Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
 
 #ifdef WithShell
   if (lev == 0)
@@ -4223,26 +4243,8 @@ void bssn_class::Step(int lev, int YN)
       prev_clock = curr_clock;
       curr_clock = clock();
       cout << " Shell stuff synchronization used " 
-           << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC)
+      << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
-           << " seconds! " << endl;
+      << " seconds! " << endl;
-    }
-  }
-#endif
-  Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
-
-#ifdef WithShell
-  // Complete non-blocking error reduction and check
-  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
-  if (ERROR)
-  {
-    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime
-                              << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
     }
   }
 #endif
@@ -4405,7 +4407,23 @@ void bssn_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
+    // check error information
+    {
+      int erh = ERROR;
+      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    }
+    if (ERROR)
+    {
+      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
+                                << " variables at t = " << PhysTime 
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
 
 #ifdef WithShell
     // evolve Shell Patches
@@ -4545,16 +4563,25 @@ void bssn_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    // check error information
-    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
+      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    }
+    if (ERROR)
+    {
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count 
+                                << " variables at t = " << PhysTime << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
     }
 #endif
 
-    Parallel::AsyncSyncState async_cor;
+    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
-    Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
 
 #ifdef WithShell
     if (lev == 0)
@@ -4572,25 +4599,6 @@ void bssn_class::Step(int lev, int YN)
              << " seconds! " << endl;
       }
     }
-#endif
-    Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
-
-#ifdef WithShell
-    // Complete non-blocking error reduction and check
-    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
-    if (ERROR)
-    {
-      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
-                                << " variables at t = " << PhysTime
-                                << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
 #endif
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
@@ -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)
@@ -4956,21 +4970,17 @@ void bssn_class::Step(int lev, int YN)
 
   //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Predictor rhs calculation");
 
-  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  // Start async ghost zone exchange - overlaps with error check and BH position
-  MPI_Request err_req;
+  Parallel::SyncHandle *sync_pre = Parallel::SyncBeginWithPlan(sync_plan, SynchList_pre);
+
+  // check error information (overlaps with MPI transfer)
   {
     int erh = ERROR;
-    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev], &err_req);
+    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev]);
   }
-
-  //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor sync");
-
-  Parallel::Sync_cached(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev]);
-
-  // Complete non-blocking error reduction and check
-  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
   if (ERROR)
   {
+    Parallel::SyncEnd(sync_pre); sync_pre = 0;
     Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
     if (myrank == 0)
     {
@@ -4980,6 +4990,11 @@ void bssn_class::Step(int lev, int YN)
     }
   }
 
+  //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor sync");
+
+  // Complete async ghost zone exchange
+  if (sync_pre) Parallel::SyncEnd(sync_pre);
+
 #if (MAPBH == 0)
   // for black hole position
   if (BH_num > 0 && lev == GH->levels - 1)
@@ -5157,23 +5172,17 @@ void bssn_class::Step(int lev, int YN)
 
     //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector error check");
 
-    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    // Start async ghost zone exchange - overlaps with error check and BH position
-    MPI_Request err_req_cor;
+    Parallel::SyncHandle *sync_cor = Parallel::SyncBeginWithPlan(sync_plan, SynchList_cor);
+
+    // check error information (overlaps with MPI transfer)
     {
       int erh = ERROR;
-      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev], &err_req_cor);
+      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev]);
     }
-
-    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector sync");
-
-    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev]);
-
-    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Corrector sync");
-
-    // Complete non-blocking error reduction and check
-    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
     if (ERROR)
     {
+      Parallel::SyncEnd(sync_cor); sync_cor = 0;
       Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
       if (myrank == 0)
       {
@@ -5185,6 +5194,13 @@ void bssn_class::Step(int lev, int YN)
       }
     }
 
+    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector sync");
+
+    // Complete async ghost zone exchange
+    if (sync_cor) Parallel::SyncEnd(sync_cor);
+
+    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Corrector sync");
+
 #if (MAPBH == 0)
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
@@ -5297,6 +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);
 }
 
 //================================================================================================
@@ -5468,11 +5486,21 @@ void bssn_class::SHStep()
 #if (PSTR == 1 || PSTR == 2)
 //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor's error check");
 #endif
-  // Non-blocking error reduction overlapped with Synch to hide Allreduce latency
+  // check error information
-  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
+    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+  }
+
+  if (ERROR)
+  {
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
   }
 
   {
@@ -5490,19 +5518,6 @@ void bssn_class::SHStep()
     }
   }
 
-  // Complete non-blocking error reduction and check
-  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
-  if (ERROR)
-  {
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
-
   // corrector
   for (iter_count = 1; iter_count < 4; iter_count++)
   {
@@ -5645,11 +5660,21 @@ void bssn_class::SHStep()
         sPp = sPp->next;
       }
     }
-    // Non-blocking error reduction overlapped with Synch to hide Allreduce latency
+    // check error information
-    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
+      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    }
+    if (ERROR)
+    {
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count 
+                                << " variables at t = " << PhysTime << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
     }
 
     {
@@ -5667,20 +5692,6 @@ void bssn_class::SHStep()
       }
     }
 
-    // Complete non-blocking error reduction and check
-    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
-    if (ERROR)
-    {
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count
-                                << " variables at t = " << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
-
     sPp = SH->PatL;
     while (sPp)
     {
@@ -5809,7 +5820,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 //       misc::tillherecheck(GH->Commlev[GH->mylev],GH->start_rank[GH->mylev],a_stream.str());
 #endif
 
-      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
+      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
 
 #if (PSTR == 1 || PSTR == 2)
 //       a_stream.clear();
@@ -5819,11 +5830,21 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
 
 #if (RPB == 0)
+      Ppc = GH->PatL[lev - 1];
+      while (Ppc)
+      {
+        Pp = GH->PatL[lev];
+        while (Pp)
+        {
 #if (MIXOUTB == 0)
-      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
+          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
 #elif (MIXOUTB == 1)
-      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
+          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
 #endif
+          Pp = Pp->next;
+        }
+        Ppc = Ppc->next;
+      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, GH->bdsul[lev], Symmetry);
@@ -5860,7 +5881,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 //       misc::tillherecheck(GH->Commlev[GH->mylev],GH->start_rank[GH->mylev],a_stream.str());
 #endif
 
-      Parallel::Sync_cached(GH->PatL[lev - 1], SL, Symmetry, sync_cache_rp_coarse[lev]);
+      Parallel::Sync(GH->PatL[lev - 1], SL, Symmetry);
 
 #if (PSTR == 1 || PSTR == 2)
 //       a_stream.clear();
@@ -5870,11 +5891,21 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
 
 #if (RPB == 0)
+      Ppc = GH->PatL[lev - 1];
+      while (Ppc)
+      {
+        Pp = GH->PatL[lev];
+        while (Pp)
+        {
 #if (MIXOUTB == 0)
-      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
+          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SL, SL, Symmetry);
 #elif (MIXOUTB == 1)
-      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
+          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SL, SL, Symmetry);
 #endif
+          Pp = Pp->next;
+        }
+        Ppc = Ppc->next;
+      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SL,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->bdsul[lev], Symmetry);
@@ -5888,7 +5919,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
     }
 
-    Parallel::Sync_cached(GH->PatL[lev], SL, Symmetry, sync_cache_rp_fine[lev]);
+    Parallel::Sync(GH->PatL[lev], SL, Symmetry);
 
 #if (PSTR == 1 || PSTR == 2)
 //    a_stream.clear();
@@ -5946,14 +5977,24 @@ void bssn_class::RestrictProlong_aux(int lev, int YN, bool BB,
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SynchList_pre, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
+      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
 
 #if (RPB == 0)
+      Ppc = GH->PatL[lev - 1];
+      while (Ppc)
+      {
+        Pp = GH->PatL[lev];
+        while (Pp)
+        {
 #if (MIXOUTB == 0)
-      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
+          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
 #elif (MIXOUTB == 1)
-      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
+          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
 #endif
+          Pp = Pp->next;
+        }
+        Ppc = Ppc->next;
+      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, GH->bdsul[lev], Symmetry);
@@ -5968,21 +6009,31 @@ void bssn_class::RestrictProlong_aux(int lev, int YN, bool BB,
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync_cached(GH->PatL[lev - 1], SL, Symmetry, sync_cache_rp_coarse[lev]);
+      Parallel::Sync(GH->PatL[lev - 1], SL, Symmetry);
 
 #if (RPB == 0)
+      Ppc = GH->PatL[lev - 1];
+      while (Ppc)
+      {
+        Pp = GH->PatL[lev];
+        while (Pp)
+        {
 #if (MIXOUTB == 0)
-      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
+          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SL, SL, Symmetry);
 #elif (MIXOUTB == 1)
-      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
+          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SL, SL, Symmetry);
 #endif
+          Pp = Pp->next;
+        }
+        Ppc = Ppc->next;
+      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SL,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->bdsul[lev], Symmetry);
 #endif
     }
 
-    Parallel::Sync_cached(GH->PatL[lev], SL, Symmetry, sync_cache_rp_fine[lev]);
+    Parallel::Sync(GH->PatL[lev], SL, Symmetry);
   }
 }
 
@@ -6033,14 +6084,24 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB)
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, SynchList_pre, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
+      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
 
 #if (RPB == 0)
+      Ppc = GH->PatL[lev - 1];
+      while (Ppc)
+      {
+        Pp = GH->PatL[lev];
+        while (Pp)
+        {
 #if (MIXOUTB == 0)
-      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
+          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
+          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
 #endif
+          Pp = Pp->next;
+        }
+        Ppc = Ppc->next;
+      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6057,21 +6118,31 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB)
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, StateList, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync_cached(GH->PatL[lev - 1], StateList, Symmetry, sync_cache_rp_coarse[lev]);
+      Parallel::Sync(GH->PatL[lev - 1], StateList, Symmetry);
 
 #if (RPB == 0)
+      Ppc = GH->PatL[lev - 1];
+      while (Ppc)
+      {
+        Pp = GH->PatL[lev];
+        while (Pp)
+        {
 #if (MIXOUTB == 0)
-      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
+          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
+          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
 #endif
+          Pp = Pp->next;
+        }
+        Ppc = Ppc->next;
+      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],StateList,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, GH->bdsul[lev], Symmetry);
 #endif
     }
 
-    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_rp_fine[lev]);
+    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
   }
 }
 
@@ -6101,11 +6172,21 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
       }
 
 #if (RPB == 0)
+      Ppc = GH->PatL[lev - 1];
+      while (Ppc)
+      {
+        Pp = GH->PatL[lev];
+        while (Pp)
+        {
 #if (MIXOUTB == 0)
-      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
+          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
+          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
 #endif
+          Pp = Pp->next;
+        }
+        Ppc = Ppc->next;
+      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6114,11 +6195,21 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
     else // no time refinement levels and for all same time levels
     {
 #if (RPB == 0)
+      Ppc = GH->PatL[lev - 1];
+      while (Ppc)
+      {
+        Pp = GH->PatL[lev];
+        while (Pp)
+        {
 #if (MIXOUTB == 0)
-      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
+          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
+          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
 #endif
+          Pp = Pp->next;
+        }
+        Ppc = Ppc->next;
+      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],StateList,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6134,10 +6225,10 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
 #else
       Parallel::Restrict_after(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, StateList, Symmetry);
 #endif
-      Parallel::Sync_cached(GH->PatL[lev - 1], StateList, Symmetry, sync_cache_rp_coarse[lev]);
+      Parallel::Sync(GH->PatL[lev - 1], StateList, Symmetry);
     }
 
-    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_rp_fine[lev]);
+    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
   }
 }
 #undef MIXOUTB

AMSS_NCKU_source/bssn_class.h

@@ -126,11 +126,6 @@ public:
        MyList<var> *OldStateList, *DumpList;
        MyList<var> *ConstraintList;
 
-       Parallel::SyncCache *sync_cache_pre;  // per-level cache for predictor sync
-       Parallel::SyncCache *sync_cache_cor;  // per-level cache for corrector sync
-       Parallel::SyncCache *sync_cache_rp_coarse;  // RestrictProlong sync on PatL[lev-1]
-       Parallel::SyncCache *sync_cache_rp_fine;    // RestrictProlong sync on PatL[lev]
-
        monitor *ErrorMonitor, *Psi4Monitor, *BHMonitor, *MAPMonitor;
        monitor *ConVMonitor;
        surface_integral *Waveshell;

AMSS_NCKU_source/cgh.C

@@ -43,14 +43,6 @@ cgh::cgh(int ingfsi, int fngfsi, int Symmetry, char *filename, int checkrun,
   end_rank = 0;
 #endif
 
-  // Initialize load balancing variables
-  enable_load_balance = false;
-  load_balance_check_interval = 10;  // Check every 10 time steps
-  current_time_step = 0;
-  rank_interp_times = nullptr;
-  heavy_ranks = nullptr;
-  num_heavy_ranks = 0;
-
   if (!checkrun)
   {
     read_bbox(Symmetry, filename);
@@ -121,12 +113,6 @@ cgh::~cgh()
     delete[] Porgls[lev];
   }
   delete[] Porgls;
-
-  // Clean up load balancing memory
-  if (rank_interp_times)
-    delete[] rank_interp_times;
-  if (heavy_ranks)
-    delete[] heavy_ranks;
 }
 
 //================================================================================================
@@ -144,7 +130,7 @@ void cgh::compose_cgh(int nprocs)
   for (int lev = 0; lev < levels; lev++)
   {
     checkPatchList(PatL[lev], false);
-    Parallel::distribute_hard(PatL[lev], nprocs, ingfs, fngfs, false);
+    Parallel::distribute(PatL[lev], nprocs, ingfs, fngfs, false);
 #if (RPB == 1)
     // we need distributed box of PatL[lev] and PatL[lev-1]
     if (lev > 0)
@@ -1719,121 +1705,3 @@ void cgh::settrfls(const int lev)
 {
   trfls = lev;
 }
-
-//================================================================================================
-// Load Balancing Functions
-//================================================================================================
-
-// Initialize load balancing
-void cgh::init_load_balance(int nprocs)
-{
-  if (rank_interp_times)
-    delete[] rank_interp_times;
-  if (heavy_ranks)
-    delete[] heavy_ranks;
-
-  rank_interp_times = new double[nprocs];
-  heavy_ranks = new int[4];  // Maximum 4 heavy ranks
-  num_heavy_ranks = 0;
-
-  for (int i = 0; i < nprocs; i++)
-    rank_interp_times[i] = 0.0;
-}
-
-// Update interpolation time for a rank
-void cgh::update_interp_time(int rank, double time)
-{
-  if (rank_interp_times && rank >= 0)
-  {
-    rank_interp_times[rank] = time;
-  }
-}
-
-// Check and perform load balancing if needed
-bool cgh::check_and_rebalance(int nprocs, int lev,
-                               MyList<var> *OldList, MyList<var> *StateList,
-                               MyList<var> *FutureList, MyList<var> *tmList,
-                               int Symmetry, bool BB)
-{
-  int myrank;
-  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
-
-  // Only check at specified intervals
-  current_time_step++;
-  if (current_time_step % load_balance_check_interval != 0)
-    return false;
-
-  if (myrank == 0)
-  {
-    cout << "\n=== Checking load balance at time step " << current_time_step << " ===" << endl;
-  }
-
-  // Collect all rank times on rank 0
-  double *all_times = nullptr;
-  if (myrank == 0)
-  {
-    all_times = new double[nprocs];
-  }
-
-  MPI_Gather(rank_interp_times, 1, MPI_DOUBLE, all_times, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
-
-  bool need_rebalance = false;
-
-  if (myrank == 0)
-  {
-    // Check if load balancing is needed
-    need_rebalance = Parallel::check_load_balance_need(all_times, nprocs, num_heavy_ranks, heavy_ranks);
-
-    if (need_rebalance)
-    {
-      cout << "=== Load imbalance detected! Need to rebalance ===" << endl;
-      cout << "Top " << num_heavy_ranks << " heavy ranks: ";
-      for (int i = 0; i < num_heavy_ranks; i++)
-      {
-        cout << heavy_ranks[i] << " (" << all_times[heavy_ranks[i]] << " s) ";
-      }
-      cout << endl;
-
-      // Analyze blocks that need to be split
-      Parallel::split_heavy_blocks(PatL[lev], heavy_ranks, num_heavy_ranks, 2, nprocs, ingfs, fngfs);
-
-      // Set lev_flag to trigger recompose_cgh
-      cout << "=== Triggering recompose_cgh for level " << lev << " ===" << endl;
-    }
-    else
-    {
-      cout << "=== Load is balanced, no rebalancing needed ===" << endl;
-    }
-
-    delete[] all_times;
-  }
-
-  // Broadcast the decision to all ranks
-  MPI_Bcast(&need_rebalance, 1, MPI_C_BOOL, 0, MPI_COMM_WORLD);
-
-  if (need_rebalance)
-  {
-    // Broadcast heavy ranks information
-    MPI_Bcast(&num_heavy_ranks, 1, MPI_INT, 0, MPI_COMM_WORLD);
-    MPI_Bcast(heavy_ranks, num_heavy_ranks, MPI_INT, 0, MPI_COMM_WORLD);
-
-    // Perform recompose_cgh on the specified level
-    if (myrank == 0)
-    {
-      cout << "=== Performing recompose_cgh ===" << endl;
-    }
-
-    // Call recompose_cgh_Onelevel for the specified level
-    bool *lev_flag = new bool[1];
-    lev_flag[0] = true;
-    recompose_cgh_Onelevel(nprocs, lev, OldList, StateList, FutureList, tmList, Symmetry, BB);
-    delete[] lev_flag;
-
-    // Reset time counter after rebalancing
-    current_time_step = 0;
-
-    return true;
-  }
-
-  return false;
-}

AMSS_NCKU_source/cgh.h

@@ -87,21 +87,6 @@ public:
 #if (PSTR == 1 || PSTR == 2 || PSTR == 3)
    void construct_mylev(int nprocs);
 #endif
-
-   // Load balancing support
-   bool enable_load_balance;         // Enable load balancing
-   int load_balance_check_interval;  // Check interval (in time steps)
-   int current_time_step;            // Current time step counter
-   double *rank_interp_times;        // Store interpolation times for each rank
-   int *heavy_ranks;                 // Store heavy rank numbers
-   int num_heavy_ranks;              // Number of heavy ranks
-
-   void init_load_balance(int nprocs);
-   void update_interp_time(int rank, double time);
-   bool check_and_rebalance(int nprocs, int lev,
-                           MyList<var> *OldList, MyList<var> *StateList,
-                           MyList<var> *FutureList, MyList<var> *tmList,
-                           int Symmetry, bool BB);
 };
 
 #endif /* CGH_H */

AMSS_NCKU_source/diff_new.f90

@@ -69,8 +69,6 @@
   fy = ZEO
   fz = ZEO
 
-!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
-!DIR$ UNROLL PARTIAL(4)
   do k=1,ex(3)-1
   do j=1,ex(2)-1
   do i=1,ex(1)-1
@@ -373,8 +371,6 @@
   fxz = ZEO
   fyz = ZEO
 
-!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
-!DIR$ UNROLL PARTIAL(4)
   do k=1,ex(3)-1
   do j=1,ex(2)-1
   do i=1,ex(1)-1

AMSS_NCKU_source/fmisc.f90

@@ -883,17 +883,13 @@ subroutine symmetry_bd(ord,extc,func,funcc,SoA)
 
   integer::i
 
-!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
   funcc(1:extc(1),1:extc(2),1:extc(3)) = func
-!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
    do i=0,ord-1
       funcc(-i,1:extc(2),1:extc(3)) = funcc(i+1,1:extc(2),1:extc(3))*SoA(1)
    enddo
-!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
    do i=0,ord-1
       funcc(:,-i,1:extc(3)) = funcc(:,i+1,1:extc(3))*SoA(2)
    enddo
-!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
    do i=0,ord-1
       funcc(:,:,-i) = funcc(:,:,i+1)*SoA(3)
    enddo
@@ -1116,7 +1112,6 @@ end subroutine d2dump
 ! Lagrangian polynomial interpolation
 !------------------------------------------------------------------------------
 
-!DIR$ ATTRIBUTES FORCEINLINE :: polint
   subroutine polint(xa, ya, x, y, dy, ordn)
   implicit none
 

AMSS_NCKU_source/kodiss.f90

@@ -65,8 +65,6 @@ real*8,intent(in) :: eps
 !                       dx^4
 
 !  note the sign (-1)^r-1, now r=2
-!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
-!DIR$ UNROLL PARTIAL(4)
   do k=1,ex(3)
   do j=1,ex(2)
   do i=1,ex(1)

AMSS_NCKU_source/makefile.inc

@@ -10,15 +10,14 @@ filein  = -I/usr/include/ -I${MKLROOT}/include
 ## Added -lifcore for Intel Fortran runtime and -limf for Intel math library
 LDLIBS  = -L${MKLROOT}/lib -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lifcore -limf -lpthread -lm -ldl
 
-## Aggressive optimization flags + PGO Phase 2 (profile-guided optimization)
+## Aggressive optimization flags:
-## -fprofile-instr-use: use collected profile data to guide optimization decisions
+## -O3: Maximum optimization
-##   (branch prediction, basic block layout, inlining, loop unrolling)
+## -xHost: Optimize for the host CPU architecture (Intel/AMD compatible)
-PROFDATA     = ../../pgo_profile/default.profdata
+## -fp-model fast=2: Aggressive floating-point optimizations
+## -fma: Enable fused multiply-add instructions
 CXXAPPFLAGS  = -O3 -xHost -fp-model fast=2 -fma -ipo \
-               -fprofile-instr-use=$(PROFDATA) \
                -Dfortran3 -Dnewc -I${MKLROOT}/include
 f90appflags  = -O3 -xHost -fp-model fast=2 -fma -ipo \
-               -fprofile-instr-use=$(PROFDATA) \
                -align array64byte -fpp -I${MKLROOT}/include
 f90          = ifx
 f77          = ifx

AMSS_NCKU_source/surface_integral.C

@@ -11,8 +11,6 @@
 #include <strstream>
 #include <cmath>
 #include <map>
-#include <vector>
-#include <algorithm>
 using namespace std;
 #else
 #include <iostream.h>
@@ -222,9 +220,16 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
     pox[2][n] = rex * nz_g[n];
   }
 
+  double *shellf;
+  shellf = new double[n_tot * InList];
+
+  GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry);
+
   int mp, Lp, Nmin, Nmax;
+
   mp = n_tot / cpusize;
   Lp = n_tot - cpusize * mp;
+
   if (Lp > myrank)
   {
     Nmin = myrank * mp + myrank;
@@ -236,10 +241,8 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
     Nmax = Nmin + mp - 1;
   }
 
-  double *shellf;
+  //|~~~~~> Integrate the dot product of Dphi with the surface normal.
-  shellf = new double[n_tot * InList];
 
-  GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry, Nmin, Nmax);
   double *RP_out, *IP_out;
   RP_out = new double[NN];
   IP_out = new double[NN];
@@ -360,17 +363,8 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 
@@ -562,17 +556,8 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, Comm_here);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, Comm_here);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, Comm_here);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 
@@ -750,17 +735,8 @@ void surface_integral::surf_Wave(double rex, int lev, ShellPatch *GH, var *Rpsi4
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 
@@ -1008,17 +984,8 @@ void surface_integral::surf_Wave(double rex, int lev, ShellPatch *GH,
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 
@@ -1452,17 +1419,8 @@ void surface_integral::surf_Wave(double rex, int lev, ShellPatch *GH,
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 
@@ -1896,17 +1854,8 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH,
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 
@@ -2091,17 +2040,8 @@ void surface_integral::surf_Wave(double rex, int lev, NullShellPatch2 *GH, var *
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 
@@ -2286,17 +2226,8 @@ void surface_integral::surf_Wave(double rex, int lev, NullShellPatch *GH, var *R
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 
@@ -2383,9 +2314,25 @@ void surface_integral::surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var
     pox[2][n] = rex * nz_g[n];
   }
 
+  double *shellf;
+  shellf = new double[n_tot * InList];
+
+  // we have assumed there is only one box on this level,
+  // so we do not need loop boxes
+  GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry);
+
+  double Mass_out = 0;
+  double ang_outx, ang_outy, ang_outz;
+  double p_outx, p_outy, p_outz;
+  ang_outx = ang_outy = ang_outz = 0.0;
+  p_outx = p_outy = p_outz = 0.0;
+  const double f1o8 = 0.125;
+
   int mp, Lp, Nmin, Nmax;
+
   mp = n_tot / cpusize;
   Lp = n_tot - cpusize * mp;
+
   if (Lp > myrank)
   {
     Nmin = myrank * mp + myrank;
@@ -2397,20 +2344,6 @@ void surface_integral::surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var
     Nmax = Nmin + mp - 1;
   }
 
-  double *shellf;
-  shellf = new double[n_tot * InList];
-
-  // we have assumed there is only one box on this level,
-  // so we do not need loop boxes
-  GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry, Nmin, Nmax);
-
-  double Mass_out = 0;
-  double ang_outx, ang_outy, ang_outz;
-  double p_outx, p_outy, p_outz;
-  ang_outx = ang_outy = ang_outz = 0.0;
-  p_outx = p_outy = p_outz = 0.0;
-  const double f1o8 = 0.125;
-
   double Chi, Psi;
   double Gxx, Gxy, Gxz, Gyy, Gyz, Gzz;
   double gupxx, gupxy, gupxz, gupyy, gupyz, gupzz;
@@ -2531,13 +2464,15 @@ void surface_integral::surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var
     }
   }
 
-  {
+  MPI_Allreduce(&Mass_out, &mass, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double scalar_out[7] = {Mass_out, ang_outx, ang_outy, ang_outz, p_outx, p_outy, p_outz};
+
-    double scalar_in[7];
+  MPI_Allreduce(&ang_outx, &sx, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    MPI_Allreduce(scalar_out, scalar_in, 7, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  MPI_Allreduce(&ang_outy, &sy, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    mass = scalar_in[0]; sx = scalar_in[1]; sy = scalar_in[2]; sz = scalar_in[3];
+  MPI_Allreduce(&ang_outz, &sz, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    px = scalar_in[4]; py = scalar_in[5]; pz = scalar_in[6];
+
-  }
+  MPI_Allreduce(&p_outx, &px, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  MPI_Allreduce(&p_outy, &py, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  MPI_Allreduce(&p_outz, &pz, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
 
 #ifdef GaussInt
   mass = mass * rex * rex * dphi * factor;
@@ -2800,13 +2735,15 @@ void surface_integral::surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var
     }
   }
 
-  {
+  MPI_Allreduce(&Mass_out, &mass, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
-    double scalar_out[7] = {Mass_out, ang_outx, ang_outy, ang_outz, p_outx, p_outy, p_outz};
+
-    double scalar_in[7];
+  MPI_Allreduce(&ang_outx, &sx, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
-    MPI_Allreduce(scalar_out, scalar_in, 7, MPI_DOUBLE, MPI_SUM, Comm_here);
+  MPI_Allreduce(&ang_outy, &sy, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
-    mass = scalar_in[0]; sx = scalar_in[1]; sy = scalar_in[2]; sz = scalar_in[3];
+  MPI_Allreduce(&ang_outz, &sz, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
-    px = scalar_in[4]; py = scalar_in[5]; pz = scalar_in[6];
+
-  }
+  MPI_Allreduce(&p_outx, &px, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
+  MPI_Allreduce(&p_outy, &py, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
+  MPI_Allreduce(&p_outz, &pz, 1, MPI_DOUBLE, MPI_SUM, Comm_here);
 
 #ifdef GaussInt
   mass = mass * rex * rex * dphi * factor;
@@ -3083,13 +3020,15 @@ void surface_integral::surf_MassPAng(double rex, int lev, ShellPatch *GH, var *c
     }
   }
 
-  {
+  MPI_Allreduce(&Mass_out, &mass, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double scalar_out[7] = {Mass_out, ang_outx, ang_outy, ang_outz, p_outx, p_outy, p_outz};
+
-    double scalar_in[7];
+  MPI_Allreduce(&ang_outx, &sx, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    MPI_Allreduce(scalar_out, scalar_in, 7, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  MPI_Allreduce(&ang_outy, &sy, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    mass = scalar_in[0]; sx = scalar_in[1]; sy = scalar_in[2]; sz = scalar_in[3];
+  MPI_Allreduce(&ang_outz, &sz, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    px = scalar_in[4]; py = scalar_in[5]; pz = scalar_in[6];
+
-  }
+  MPI_Allreduce(&p_outx, &px, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  MPI_Allreduce(&p_outy, &py, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+  MPI_Allreduce(&p_outz, &pz, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
 
 #ifdef GaussInt
   mass = mass * rex * rex * dphi * factor;
@@ -3668,17 +3607,8 @@ void surface_integral::surf_Wave(double rex, cgh *GH, ShellPatch *SH,
   }
   //|------+  Communicate and sum the results from each processor.
 
-  {
+  MPI_Allreduce(RP_out, RP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP_out = new double[2 * NN];
+  MPI_Allreduce(IP_out, IP, NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    double *RPIP = new double[2 * NN];
-    memcpy(RPIP_out, RP_out, NN * sizeof(double));
-    memcpy(RPIP_out + NN, IP_out, NN * sizeof(double));
-    MPI_Allreduce(RPIP_out, RPIP, 2 * NN, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-    memcpy(RP, RPIP, NN * sizeof(double));
-    memcpy(IP, RPIP + NN, NN * sizeof(double));
-    delete[] RPIP_out;
-    delete[] RPIP;
-  }
 
   //|------= Free memory.
 

makefile_and_run.py

@@ -15,13 +15,12 @@ import time
 ## taskset ensures all child processes inherit the CPU affinity mask
 ## This forces make and all compiler processes to use only nohz_full cores (4-55, 60-111)
 ## Format: taskset -c 4-55,60-111 ensures processes only run on these cores
-#NUMACTL_CPU_BIND = "taskset -c 0-111"
+NUMACTL_CPU_BIND = "taskset -c 0-111"
-NUMACTL_CPU_BIND = "taskset -c 16-47,64-95"
 
 ## Build parallelism configuration
 ## Use nohz_full cores (4-55, 60-111) for compilation: 52 + 52 = 104 cores
 ## Set make -j to utilize available cores for faster builds
-BUILD_JOBS = 96
+BUILD_JOBS = 104
 
 
 ##################################################################
@@ -118,7 +117,6 @@ def run_ABE():
     
     if (input_data.GPU_Calculation == "no"):
         mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
-        #mpi_command         = " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
         mpi_command_outfile = "ABE_out.log"
     elif (input_data.GPU_Calculation == "yes"):
         mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABEGPU"
@@ -160,8 +158,7 @@ def run_TwoPunctureABE():
     print(                                                          )
     
     ## Define the command to run
-    #TwoPuncture_command         = NUMACTL_CPU_BIND + " ./TwoPunctureABE"
+    TwoPuncture_command         = NUMACTL_CPU_BIND + " ./TwoPunctureABE"
-    TwoPuncture_command         = " ./TwoPunctureABE"
     TwoPuncture_command_outfile = "TwoPunctureABE_out.log"
 
     ## Execute the command with subprocess.Popen and stream output

pgo_profile/PGO_Profile_Analysis.md

@@ -1,97 +0,0 @@
-# AMSS-NCKU PGO Profile Analysis Report
-
-## 1. Profiling Environment
-
-| Item | Value |
-|------|-------|
-| Compiler | Intel oneAPI DPC++/C++ 2025.3.0 (icpx/ifx) |
-| Instrumentation Flag | `-fprofile-instr-generate` |
-| Optimization Level (instrumented) | `-O2 -xHost -fma` |
-| MPI Processes | 1 (single process to avoid MPI+instrumentation deadlock) |
-| Profile File | `default_9725750769337483397_0.profraw` (327 KB) |
-| Merged Profile | `default.profdata` (394 KB) |
-| llvm-profdata | `/home/intel/oneapi/compiler/2025.3/bin/compiler/llvm-profdata` |
-
-## 2. Reduced Simulation Parameters (for profiling run)
-
-| Parameter | Production Value | Profiling Value |
-|-----------|-----------------|-----------------|
-| MPI_processes | 64 | 1 |
-| grid_level | 9 | 4 |
-| static_grid_level | 5 | 3 |
-| static_grid_number | 96 | 24 |
-| moving_grid_number | 48 | 16 |
-| largest_box_xyz_max | 320^3 | 160^3 |
-| Final_Evolution_Time | 1000.0 | 10.0 |
-| Evolution_Step_Number | 10,000,000 | 1,000 |
-| Detector_Number | 12 | 2 |
-
-## 3. Profile Summary
-
-| Metric | Value |
-|--------|-------|
-| Total instrumented functions | 1,392 |
-| Functions with non-zero counts | 117 (8.4%) |
-| Functions with zero counts | 1,275 (91.6%) |
-| Maximum function entry count | 386,459,248 |
-| Maximum internal block count | 370,477,680 |
-| Total block count | 4,198,023,118 |
-
-## 4. Top 20 Hotspot Functions
-
-| Rank | Total Count | Max Block Count | Function | Category |
-|------|------------|-----------------|----------|----------|
-| 1 | 1,241,601,732 | 370,477,680 | `polint_` | Interpolation |
-| 2 | 755,994,435 | 230,156,640 | `prolong3_` | Grid prolongation |
-| 3 | 667,964,095 | 3,697,792 | `compute_rhs_bssn_` | BSSN RHS evolution |
-| 4 | 539,736,051 | 386,459,248 | `symmetry_bd_` | Symmetry boundary |
-| 5 | 277,310,808 | 53,170,728 | `lopsided_` | Lopsided FD stencil |
-| 6 | 155,534,488 | 94,535,040 | `decide3d_` | 3D grid decision |
-| 7 | 119,267,712 | 19,266,048 | `rungekutta4_rout_` | RK4 time integrator |
-| 8 | 91,574,616 | 48,824,160 | `kodis_` | Kreiss-Oliger dissipation |
-| 9 | 67,555,389 | 43,243,680 | `fderivs_` | Finite differences |
-| 10 | 55,296,000 | 42,246,144 | `misc::fact(int)` | Factorial utility |
-| 11 | 43,191,071 | 27,663,328 | `fdderivs_` | 2nd-order FD derivatives |
-| 12 | 36,233,965 | 22,429,440 | `restrict3_` | Grid restriction |
-| 13 | 24,698,512 | 17,231,520 | `polin3_` | Polynomial interpolation |
-| 14 | 22,962,942 | 20,968,768 | `copy_` | Data copy |
-| 15 | 20,135,696 | 17,259,168 | `Ansorg::barycentric(...)` | Spectral interpolation |
-| 16 | 14,650,224 | 7,224,768 | `Ansorg::barycentric_omega(...)` | Spectral weights |
-| 17 | 13,242,296 | 2,871,920 | `global_interp_` | Global interpolation |
-| 18 | 12,672,000 | 7,734,528 | `sommerfeld_rout_` | Sommerfeld boundary |
-| 19 | 6,872,832 | 1,880,064 | `sommerfeld_routbam_` | Sommerfeld boundary (BAM) |
-| 20 | 5,709,900 | 2,809,632 | `l2normhelper_` | L2 norm computation |
-
-## 5. Hotspot Category Breakdown
-
-Top 20 functions account for ~98% of total execution counts:
-
-| Category | Functions | Combined Count | Share |
-|----------|-----------|---------------|-------|
-| Interpolation / Prolongation / Restriction | polint_, prolong3_, restrict3_, polin3_, global_interp_, Ansorg::* | ~2,093M | ~50% |
-| BSSN RHS + FD stencils | compute_rhs_bssn_, lopsided_, fderivs_, fdderivs_ | ~1,056M | ~25% |
-| Boundary conditions | symmetry_bd_, sommerfeld_rout_, sommerfeld_routbam_ | ~559M | ~13% |
-| Time integration | rungekutta4_rout_ | ~119M | ~3% |
-| Dissipation | kodis_ | ~92M | ~2% |
-| Utilities | misc::fact, decide3d_, copy_, l2normhelper_ | ~256M | ~6% |
-
-## 6. Conclusions
-
-1. **Profile data is valid**: 1,392 functions instrumented, 117 exercised with ~4.2 billion total counts.
-2. **Hotspot concentration is high**: Top 5 functions alone account for ~76% of all counts, which is ideal for PGO — the compiler has strong branch/layout optimization targets.
-3. **Fortran numerical kernels dominate**: `polint_`, `prolong3_`, `compute_rhs_bssn_`, `symmetry_bd_`, `lopsided_` are all Fortran routines in the inner evolution loop. PGO will optimize their branch prediction and basic block layout.
-4. **91.6% of functions have zero counts**: These are code paths for unused features (GPU, BSSN-EScalar, BSSN-EM, Z4C, etc.). PGO will deprioritize them, improving instruction cache utilization.
-5. **Profile is representative**: Despite the reduced grid size, the code path coverage matches production — the same kernels (RHS, prolongation, restriction, boundary) are exercised. PGO branch probabilities from this profile will transfer well to full-scale runs.
-
-## 7. PGO Phase 2 Usage
-
-To apply the profile, use the following flags in `makefile.inc`:
-
-```makefile
-CXXAPPFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
-              -fprofile-instr-use=/home/amss/AMSS-NCKU/pgo_profile/default.profdata \
-              -Dfortran3 -Dnewc -I${MKLROOT}/include
-f90appflags = -O3 -xHost -fp-model fast=2 -fma -ipo \
-              -fprofile-instr-use=/home/amss/AMSS-NCKU/pgo_profile/default.profdata \
-              -align array64byte -fpp -I${MKLROOT}/include
-```