Stabilize and wire BSSN-EScalar C path

2026-04-25 00:08:35 +08:00
parent b57d80ca61
commit 0f1d0de1e7
6 changed files with 731 additions and 369 deletions
--- a/AMSS_NCKU_source/Parallel.C
+++ b/AMSS_NCKU_source/Parallel.C
@@ -1,14 +1,50 @@

-#include "Parallel.h"
-#include "fmisc.h"
-#include "prolongrestrict.h"
-#include "misc.h"
-#include "parameters.h"
-
-int Parallel::partition1(int &nx, int split_size, int min_width, int cpusize, int shape) // special for 1 diemnsion
-{
-  nx = Mymax(1, shape / min_width);
-  nx = Mymin(cpusize, nx);
+#include "Parallel.h"
+#include "fmisc.h"
+#include "prolongrestrict.h"
+#include "misc.h"
+#include "parameters.h"
+
+namespace
+{
+enum { MAX_DATA_PACKER_VARS = 64 };
+
+int expand_var_list_pack_info(MyList<var> *src_list, MyList<var> *dst_list,
+                              int *src_sgfn, int *dst_sgfn, double **src_soa)
+{
+  int count = 0;
+  MyList<var> *src_it = src_list;
+  MyList<var> *dst_it = dst_list;
+
+  while (src_it && dst_it)
+  {
+    if (count >= MAX_DATA_PACKER_VARS)
+    {
+      cout << "Parallel::data_packer: too many variables in communication list." << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+    src_sgfn[count] = src_it->data->sgfn;
+    dst_sgfn[count] = dst_it->data->sgfn;
+    src_soa[count] = src_it->data->SoA;
+    count++;
+    src_it = src_it->next;
+    dst_it = dst_it->next;
+  }
+
+  if (src_it || dst_it)
+  {
+    cout << "error in short data packer, var lists does not match." << endl;
+    MPI_Abort(MPI_COMM_WORLD, 1);
+  }
+
+  return count;
+}
+}
+
+int Parallel::partition1(int &nx, int split_size, int min_width, int cpusize, int shape) // special for 1 diemnsion
+{
+  nx = Mymax(1, shape / min_width);
+  nx = Mymin(cpusize, nx);

  return nx;
 }
@@ -3711,11 +3747,11 @@ void Parallel::build_gstl(MyList<Parallel::gridseg> *srci, MyList<Parallel::grid
 }
 //   PACK: prepare target data in 'data'
 // UNPACK: copy target data from 'data' to corresponding numerical grids
-int Parallel::data_packer(double *data, MyList<Parallel::gridseg> *src, MyList<Parallel::gridseg> *dst, int rank_in, int dir,
-                          MyList<var> *VarLists /* source */, MyList<var> *VarListd /* target */, int Symmetry)
-{
-  int myrank;
-  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+int Parallel::data_packer(double *data, MyList<Parallel::gridseg> *src, MyList<Parallel::gridseg> *dst, int rank_in, int dir,
+                          MyList<var> *VarLists /* source */, MyList<var> *VarListd /* target */, int Symmetry)
+{
+  int myrank;
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);

  int DIM = dim;

@@ -3725,86 +3761,89 @@ int Parallel::data_packer(double *data, MyList<Parallel::gridseg> *src, MyList<P
    MPI_Abort(MPI_COMM_WORLD, 1);
  }

-  int size_out = 0;
-
-  if (!src || !dst)
-    return size_out;
-
-  MyList<var> *varls, *varld;
-
-  varls = VarLists;
-  varld = VarListd;
-  while (varls && varld)
-  {
-    varls = varls->next;
-    varld = varld->next;
-  }
-
-  if (varls || varld)
-  {
-    cout << "error in short data packer, var lists does not match." << endl;
-    MPI_Abort(MPI_COMM_WORLD, 1);
-  }
-
-  int type; /* 1 copy, 2 restrict, 3 prolong */
-  if (src->data->Bg->lev == dst->data->Bg->lev)
-    type = 1;
-  else if (src->data->Bg->lev > dst->data->Bg->lev)
-    type = 2;
-  else
-    type = 3;
-
-  while (src && dst)
-  {
-    if ((dir == PACK && dst->data->Bg->rank == rank_in && src->data->Bg->rank == myrank) ||
-        (dir == UNPACK && src->data->Bg->rank == rank_in && dst->data->Bg->rank == myrank))
-    {
-      varls = VarLists;
-      varld = VarListd;
-      while (varls && varld)
-      {
-        if (data)
-        {
-          if (dir == PACK)
-            switch (type)
-            {
-              // attention must be paied to the difference between src's llb,uub and dst's llb,uub
-            case 1:
-              f_copy(DIM, dst->data->llb, dst->data->uub, dst->data->shape, data + size_out,
-                     src->data->Bg->bbox, src->data->Bg->bbox + dim, src->data->Bg->shape, src->data->Bg->fgfs[varls->data->sgfn],
-                     dst->data->llb, dst->data->uub);
-              break;
-            case 2:
-              f_restrict3(DIM, dst->data->llb, dst->data->uub, dst->data->shape, data + size_out,
-                          src->data->Bg->bbox, src->data->Bg->bbox + dim, src->data->Bg->shape, src->data->Bg->fgfs[varls->data->sgfn],
-                          dst->data->llb, dst->data->uub, varls->data->SoA, Symmetry);
-              break;
-            case 3:
-              f_prolong3(DIM, src->data->Bg->bbox, src->data->Bg->bbox + dim, src->data->Bg->shape, src->data->Bg->fgfs[varls->data->sgfn],
-                         dst->data->llb, dst->data->uub, dst->data->shape, data + size_out,
-                         dst->data->llb, dst->data->uub, varls->data->SoA, Symmetry);
-            }
-          if (dir == UNPACK) // from target data to corresponding grid
-            f_copy(DIM, dst->data->Bg->bbox, dst->data->Bg->bbox + dim, dst->data->Bg->shape, dst->data->Bg->fgfs[varld->data->sgfn],
-                   dst->data->llb, dst->data->uub, dst->data->shape, data + size_out,
-                   dst->data->llb, dst->data->uub);
-        }
-        size_out += dst->data->shape[0] * dst->data->shape[1] * dst->data->shape[2];
-        varls = varls->next;
-        varld = varld->next;
-      }
-    }
-    dst = dst->next;
-    src = src->next;
-  }
-
+  int size_out = 0;
+
+  if (!src || !dst)
+    return size_out;
+
+  int src_sgfn[MAX_DATA_PACKER_VARS];
+  int dst_sgfn[MAX_DATA_PACKER_VARS];
+  double *src_soa[MAX_DATA_PACKER_VARS];
+  const int var_count = expand_var_list_pack_info(VarLists, VarListd, src_sgfn, dst_sgfn, src_soa);
+
+  int type; /* 1 copy, 2 restrict, 3 prolong */
+  if (src->data->Bg->lev == dst->data->Bg->lev)
+    type = 1;
+  else if (src->data->Bg->lev > dst->data->Bg->lev)
+    type = 2;
+  else
+    type = 3;
+
+  while (src && dst)
+  {
+    const bool rank_match =
+        (dir == PACK && dst->data->Bg->rank == rank_in && src->data->Bg->rank == myrank) ||
+        (dir == UNPACK && src->data->Bg->rank == rank_in && dst->data->Bg->rank == myrank);
+
+    if (rank_match)
+    {
+      const int segment_size = dst->data->shape[0] * dst->data->shape[1] * dst->data->shape[2];
+      int offset = size_out;
+
+      if (data)
+      {
+        if (dir == PACK)
+        {
+          switch (type)
+          {
+            case 1:
+              for (int iv = 0; iv < var_count; iv++, offset += segment_size)
+                f_copy(DIM, dst->data->llb, dst->data->uub, dst->data->shape, data + offset,
+                       src->data->Bg->bbox, src->data->Bg->bbox + dim, src->data->Bg->shape,
+                       src->data->Bg->fgfs[src_sgfn[iv]], dst->data->llb, dst->data->uub);
+              break;
+            case 2:
+              for (int iv = 0; iv < var_count; iv++, offset += segment_size)
+                f_restrict3(DIM, dst->data->llb, dst->data->uub, dst->data->shape, data + offset,
+                            src->data->Bg->bbox, src->data->Bg->bbox + dim, src->data->Bg->shape,
+                            src->data->Bg->fgfs[src_sgfn[iv]], dst->data->llb, dst->data->uub,
+                            src_soa[iv], Symmetry);
+              break;
+            case 3:
+              for (int iv = 0; iv < var_count; iv++, offset += segment_size)
+                f_prolong3(DIM, src->data->Bg->bbox, src->data->Bg->bbox + dim, src->data->Bg->shape,
+                           src->data->Bg->fgfs[src_sgfn[iv]], dst->data->llb, dst->data->uub,
+                           dst->data->shape, data + offset, dst->data->llb, dst->data->uub,
+                           src_soa[iv], Symmetry);
+              break;
+            default:
+              break;
+          }
+        }
+        else
+        {
+          for (int iv = 0; iv < var_count; iv++, offset += segment_size)
+            f_copy(DIM, dst->data->Bg->bbox, dst->data->Bg->bbox + dim, dst->data->Bg->shape,
+                   dst->data->Bg->fgfs[dst_sgfn[iv]], dst->data->llb, dst->data->uub,
+                   dst->data->shape, data + offset, dst->data->llb, dst->data->uub);
+        }
+      }
+
+      size_out = offset + ((!data) ? segment_size * var_count : 0);
+      if (data)
+        size_out = offset;
+    }
+    dst = dst->next;
+    src = src->next;
+  }
+
  return size_out;
 }
-int Parallel::data_packermix(double *data, MyList<Parallel::gridseg> *src, MyList<Parallel::gridseg> *dst, int rank_in, int dir,
-                             MyList<var> *VarLists /* source */, MyList<var> *VarListd /* target */, int Symmetry)
-{
-  int myrank;
-  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+int Parallel::data_packermix(double *data, MyList<Parallel::gridseg> *src, MyList<Parallel::gridseg> *dst, int rank_in, int dir,
+                             MyList<var> *VarLists /* source */, MyList<var> *VarListd /* target */, int Symmetry)
+{
+  int myrank;
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);

  int DIM = dim;

@@ -3814,33 +3853,22 @@ int Parallel::data_packermix(double *data, MyList<Parallel::gridseg> *src, MyLis
    MPI_Abort(MPI_COMM_WORLD, 1);
  }

-  int size_out = 0;
-
-  if (!src || !dst)
-    return size_out;
-
-  MyList<var> *varls, *varld;
-
-  varls = VarLists;
-  varld = VarListd;
-  while (varls && varld)
-  {
-    varls = varls->next;
-    varld = varld->next;
-  }
-
-  if (varls || varld)
-  {
-    cout << "error in short data packer, var lists does not match." << endl;
-    MPI_Abort(MPI_COMM_WORLD, 1);
-  }
-
-  int type; /* 1 copy, 2 restrict, 3 prolong */
-  if (src->data->Bg->lev == dst->data->Bg->lev)
-    type = 1;
-  else if (src->data->Bg->lev > dst->data->Bg->lev)
-    type = 2;
-  else
+  int size_out = 0;
+
+  if (!src || !dst)
+    return size_out;
+
+  int src_sgfn[MAX_DATA_PACKER_VARS];
+  int dst_sgfn[MAX_DATA_PACKER_VARS];
+  double *src_soa[MAX_DATA_PACKER_VARS];
+  const int var_count = expand_var_list_pack_info(VarLists, VarListd, src_sgfn, dst_sgfn, src_soa);
+
+  int type; /* 1 copy, 2 restrict, 3 prolong */
+  if (src->data->Bg->lev == dst->data->Bg->lev)
+    type = 1;
+  else if (src->data->Bg->lev > dst->data->Bg->lev)
+    type = 2;
+  else
    type = 3;

  if (type != 3)
@@ -3848,37 +3876,48 @@ int Parallel::data_packermix(double *data, MyList<Parallel::gridseg> *src, MyLis
    cout << "Parallel::data_packermix: error type " << type << " for data_packermix." << endl;
    MPI_Abort(MPI_COMM_WORLD, 1);
  }
-
-  while (src && dst)
-  {
-    if ((dir == PACK && dst->data->Bg->rank == rank_in && src->data->Bg->rank == myrank) ||
-        (dir == UNPACK && src->data->Bg->rank == rank_in && dst->data->Bg->rank == myrank))
-    {
-      varls = VarLists;
-      varld = VarListd;
-      while (varls && varld)
-      {
-        if (data)
-        {
-          if (dir == PACK)
-            f_prolongcopy3(DIM, src->data->Bg->bbox, src->data->Bg->bbox + dim, src->data->Bg->shape, src->data->Bg->fgfs[varls->data->sgfn],
-                           dst->data->llb, dst->data->uub, src->data->shape, data + size_out,
-                           src->data->llb, src->data->uub, varls->data->SoA, Symmetry);
-          if (dir == UNPACK) // from target data to corresponding grid
-            f_prolongmix3(DIM, dst->data->Bg->bbox, dst->data->Bg->bbox + dim, dst->data->Bg->shape, dst->data->Bg->fgfs[varld->data->sgfn],
-                          src->data->llb, src->data->uub, src->data->shape, data + size_out,
-                          dst->data->llb, dst->data->uub, varls->data->SoA, Symmetry, dst->data->illb, dst->data->iuub);
-        }
-        // the symmetry problem should be dealt in prolongcopy3,
-        // so we always have ghost_width for both sides
-        size_out += (src->data->shape[0] + 2 * ghost_width) * (src->data->shape[1] + 2 * ghost_width) * (src->data->shape[2] + 2 * ghost_width);
-        varls = varls->next;
-        varld = varld->next;
-      }
-    }
-    dst = dst->next;
-    src = src->next;
-  }
+
+  while (src && dst)
+  {
+    const bool rank_match =
+        (dir == PACK && dst->data->Bg->rank == rank_in && src->data->Bg->rank == myrank) ||
+        (dir == UNPACK && src->data->Bg->rank == rank_in && dst->data->Bg->rank == myrank);
+
+    if (rank_match)
+    {
+      const int segment_size =
+          (src->data->shape[0] + 2 * ghost_width) *
+          (src->data->shape[1] + 2 * ghost_width) *
+          (src->data->shape[2] + 2 * ghost_width);
+      int offset = size_out;
+
+      if (data)
+      {
+        if (dir == PACK)
+        {
+          for (int iv = 0; iv < var_count; iv++, offset += segment_size)
+            f_prolongcopy3(DIM, src->data->Bg->bbox, src->data->Bg->bbox + dim, src->data->Bg->shape,
+                           src->data->Bg->fgfs[src_sgfn[iv]], dst->data->llb, dst->data->uub,
+                           src->data->shape, data + offset, src->data->llb, src->data->uub,
+                           src_soa[iv], Symmetry);
+        }
+        else
+        {
+          for (int iv = 0; iv < var_count; iv++, offset += segment_size)
+            f_prolongmix3(DIM, dst->data->Bg->bbox, dst->data->Bg->bbox + dim, dst->data->Bg->shape,
+                          dst->data->Bg->fgfs[dst_sgfn[iv]], src->data->llb, src->data->uub,
+                          src->data->shape, data + offset, dst->data->llb, dst->data->uub,
+                          src_soa[iv], Symmetry, dst->data->illb, dst->data->iuub);
+        }
+      }
+
+      size_out = offset + ((!data) ? segment_size * var_count : 0);
+      if (data)
+        size_out = offset;
+    }
+    dst = dst->next;
+    src = src->next;
+  }

  return size_out;
 }