Fix BSSN C gauge RHS parity

Shell C functions must export Fortran-compatible symbols with trailing
underscore so bssn_rhs_ss.f90 and getnp4.f90 can link when WithShell is
active and USE_CXX_SHELL_KERNELS=1 replaces Fortran diff_new_sh.o.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

AMSS_NCKU_Input.py

@@ -16,9 +16,9 @@ import numpy
 File_directory   = "GW150914"                    ## output file directory
 Output_directory = "binary_output"               ## binary data file directory
                                                  ## The file directory name should not be too long
-MPI_processes    = 64                             ## number of mpi processes used in the simulation
+MPI_processes    = 2                             ## number of mpi processes used in the simulation
 
-GPU_Calculation  = "no"                          ## Use GPU or not 
+GPU_Calculation  = "yes"                          ## Use GPU or not 
                                                  ## (prefer "no" in the current version, because the GPU part may have bugs when integrated in this Python interface)
 CPU_Part         = 1.0
 GPU_Part         = 0.0
@@ -158,7 +158,7 @@ Detector_Rmax   = 160.0                  ## farest dector distance
 
 ## Setting the apprent horizon
 
-AHF_Find       = "no"                    ## whether to find the apparent horizon: choose "yes" or "no"
+AHF_Find       = "yes"                    ## whether to find the apparent horizon: choose "yes" or "no"
 
 AHF_Find_Every = 24
 AHF_Dump_Time  = 20.0

AMSS_NCKU_Program.py

@@ -58,31 +58,36 @@ File_directory = os.path.join(input_data.File_directory)
 
 ## If the specified output directory exists, ask the user whether to continue
 if os.path.exists(File_directory):
-    print( " Output dictionary has been existed !!!  "                                                              )
-    print( " If you want to overwrite the existing file directory, please input 'continue' in the terminal !! "     ) 
-    print( " If you want to retain the existing file directory, please input 'stop' in the terminal to stop the "   ) 
-    print( " simulation. Then you can reset the output dictionary in the input script file AMSS_NCKU_Input.py !!! " )
-    print(                                                                                                          )
-    ## Prompt whether to overwrite the existing directory
-    while True:
-        try:
-            inputvalue = input()
-            ## If the user agrees to overwrite, proceed and remove the existing directory
-            if ( inputvalue == "continue" ):
-                print( " Continue the calculation !!! " )
-                print(                                  )
-                break  
-            ## If the user chooses not to overwrite, exit and keep the existing directory
-            elif ( inputvalue == "stop" ):
-                print( " Stop the calculation !!! "    )
-                sys.exit() 
-            ## If the user input is invalid, prompt again
-            else:
+    auto_overwrite = str(getattr(input_data, "Auto_Overwrite_Output", "yes")).strip().lower()
+    if auto_overwrite in ("1", "yes", "y", "true", "on", "continue"):
+        print( " Output dictionary has been existed; Auto_Overwrite_Output=yes, continue the calculation. " )
+        print(                                                                                              )
+    else:
+        print( " Output dictionary has been existed !!!  "                                                              )
+        print( " If you want to overwrite the existing file directory, please input 'continue' in the terminal !! "     ) 
+        print( " If you want to retain the existing file directory, please input 'stop' in the terminal to stop the "   ) 
+        print( " simulation. Then you can reset the output dictionary in the input script file AMSS_NCKU_Input.py !!! " )
+        print(                                                                                                          )
+        ## Prompt whether to overwrite the existing directory
+        while True:
+            try:
+                inputvalue = input()
+                ## If the user agrees to overwrite, proceed and remove the existing directory
+                if ( inputvalue == "continue" ):
+                    print( " Continue the calculation !!! " )
+                    print(                                  )
+                    break  
+                ## If the user chooses not to overwrite, exit and keep the existing directory
+                elif ( inputvalue == "stop" ):
+                    print( " Stop the calculation !!! "    )
+                    sys.exit() 
+                ## If the user input is invalid, prompt again
+                else:
+                    print( " Please input your choice !!! "                   )
+                    print( " Input 'continue' or 'stop' in the terminal !!! " )
+            except ValueError:
                 print( " Please input your choice !!! "                   )
                 print( " Input 'continue' or 'stop' in the terminal !!! " )
-        except ValueError:
-            print( " Please input your choice !!! "                   )
-            print( " Input 'continue' or 'stop' in the terminal !!! " )
         
 ## Remove the existing output directory if present
 shutil.rmtree(File_directory, ignore_errors=True)
@@ -174,14 +179,11 @@ import generate_macrodef
 generate_macrodef.generate_macrodef_h()
 print( " AMSS-NCKU macro file macrodef.h has been generated. " )
      
-generate_macrodef.generate_macrodef_fh()
-print( " AMSS-NCKU macro file macrodef.fh has been generated. " )
-
-generate_macrodef.generate_build_config()
-print( " AMSS-NCKU build config AMSS_NCKU_build.mk has been generated. " )
-
-
-##################################################################
+generate_macrodef.generate_macrodef_fh()
+print( " AMSS-NCKU macro file macrodef.fh has been generated. " )
+
+
+##################################################################
 
 # Compile the AMSS-NCKU program according to user requirements
 
@@ -220,13 +222,11 @@ shutil.copytree(AMSS_NCKU_source_path, AMSS_NCKU_source_copy)
 
 # Copy the generated macro files into the AMSS_NCKU source folder
 
-macrodef_h_path  = os.path.join(File_directory, "macrodef.h")
-macrodef_fh_path = os.path.join(File_directory, "macrodef.fh")
-build_config_path = os.path.join(File_directory, "AMSS_NCKU_build.mk")
-
-shutil.copy2(macrodef_h_path,  AMSS_NCKU_source_copy)
-shutil.copy2(macrodef_fh_path, AMSS_NCKU_source_copy)
-shutil.copy2(build_config_path, AMSS_NCKU_source_copy)
+macrodef_h_path  = os.path.join(File_directory, "macrodef.h") 
+macrodef_fh_path = os.path.join(File_directory, "macrodef.fh") 
+
+shutil.copy2(macrodef_h_path,  AMSS_NCKU_source_copy)
+shutil.copy2(macrodef_fh_path, AMSS_NCKU_source_copy)
 
 # Notes on copying files:
 # shutil.copy2 preserves file metadata such as modification time.
@@ -263,7 +263,7 @@ print()
 if (input_data.GPU_Calculation == "no"):
     ABE_file = os.path.join(AMSS_NCKU_source_copy, "ABE")
 elif (input_data.GPU_Calculation == "yes"):
-    ABE_file = os.path.join(AMSS_NCKU_source_copy, "ABEGPU")
+    ABE_file = os.path.join(AMSS_NCKU_source_copy, "ABE_CUDA")
 
 if not os.path.exists( ABE_file ):
     print(                                                                                                  )

AMSS_NCKU_source/ABE.C

@@ -198,16 +198,16 @@ int main(int argc, char *argv[])
       if (myrank == 0)
       {
             string out_dir;
-            char filename[50];
-            map<string, string>::iterator iter;
-            iter = parameters::str_par.find("output dir");
-            if (iter != parameters::str_par.end())
-            {
-                  out_dir = iter->second;
-            }
-            sprintf(filename, "%s/setting.par", out_dir.c_str());
-            ofstream setfile;
-            setfile.open(filename, ios::trunc);
+            string filename;
+            map<string, string>::iterator iter;
+            iter = parameters::str_par.find("output dir");
+            if (iter != parameters::str_par.end())
+            {
+                  out_dir = iter->second;
+            }
+            filename = out_dir + "/setting.par";
+            ofstream setfile;
+            setfile.open(filename.c_str(), ios::trunc);
 
             if (!setfile.good())
             {
@@ -484,7 +484,11 @@ int main(int argc, char *argv[])
             cout << endl;
       }
 
-      delete ADM;
+      // Let the process teardown reclaim the simulation object.  Some derived
+      // equation classes keep MPI/CUDA-backed state whose destructor ordering
+      // is fragile at program shutdown.
+      if (getenv("AMSS_DELETE_ADM_ON_EXIT"))
+            delete ADM;
 
       //=======================caculation done=============================================================
 

AMSS_NCKU_source/Block.C

@@ -6,14 +6,68 @@
 #include <cstdio>
 #include <string>
 #include <cmath>
-#include <new>
-using namespace std;
-
-#include "Block.h"
-#include "misc.h"
-
-Block::Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fngfsi, int levi, const int cgpui) : rank(ranki), ingfs(ingfsi), fngfs(fngfsi), lev(levi), cgpu(cgpui)
-{
+#include <new>
+using namespace std;
+
+#include "Block.h"
+#include "misc.h"
+
+#if USE_CUDA_BSSN || USE_CUDA_Z4C
+#include <cuda_runtime_api.h>
+#endif
+
+namespace {
+
+bool cuda_pin_gridfuncs_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_PIN_GRIDFUNCS");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+double *alloc_gridfunc(size_t count, unsigned char &pinned)
+{
+  pinned = 0;
+#if USE_CUDA_BSSN || USE_CUDA_Z4C
+  if (cuda_pin_gridfuncs_enabled())
+  {
+    double *ptr = 0;
+    cudaError_t err = cudaMallocHost((void **)&ptr, count * sizeof(double));
+    if (err == cudaSuccess)
+    {
+      pinned = 1;
+      return ptr;
+    }
+    cudaGetLastError();
+  }
+#endif
+  return (double *)malloc(sizeof(double) * count);
+}
+
+void free_gridfunc(double *ptr, unsigned char pinned)
+{
+  if (!ptr)
+    return;
+#if USE_CUDA_BSSN || USE_CUDA_Z4C
+  if (pinned)
+  {
+    cudaFreeHost(ptr);
+    return;
+  }
+#else
+  (void)pinned;
+#endif
+  free(ptr);
+}
+
+}
+
+Block::Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fngfsi, int levi, const int cgpui) : rank(ranki), lev(levi), cgpu(cgpui), ingfs(ingfsi), fngfs(fngfsi), igfs(0), fgfs(0), fgfs_pinned(0)
+{
   for (int i = 0; i < dim; i++)
     X[i] = 0;
 
@@ -68,14 +122,15 @@ Block::Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fng
 #endif
     }
 
-    int nn = shape[0] * shape[1] * shape[2];
-    fgfs = new double *[fngfs];
-    for (int i = 0; i < fngfs; i++)
-    {
-      fgfs[i] = (double *)malloc(sizeof(double) * nn);
-      if (!(fgfs[i]))
-      {
-        cout << "on node#" << rank << ", out of memory when constructing Block." << endl;
+    int nn = shape[0] * shape[1] * shape[2];
+    fgfs = new double *[fngfs];
+    fgfs_pinned = new unsigned char[fngfs];
+    for (int i = 0; i < fngfs; i++)
+    {
+      fgfs[i] = alloc_gridfunc((size_t)nn, fgfs_pinned[i]);
+      if (!(fgfs[i]))
+      {
+        cout << "on node#" << rank << ", out of memory when constructing Block." << endl;
         MPI_Abort(MPI_COMM_WORLD, 1);
       }
       memset(fgfs[i], 0, sizeof(double) * nn);
@@ -103,17 +158,19 @@ Block::~Block()
   {
     for (int i = 0; i < dim; i++)
       delete[] X[i];
-    for (int i = 0; i < ingfs; i++)
-      free(igfs[i]);
-    delete[] igfs;
-    for (int i = 0; i < fngfs; i++)
-      free(fgfs[i]);
-    delete[] fgfs;
-    X[0] = X[1] = X[2] = 0;
-    igfs = 0;
-    fgfs = 0;
-  }
-}
+    for (int i = 0; i < ingfs; i++)
+      free(igfs[i]);
+    delete[] igfs;
+    for (int i = 0; i < fngfs; i++)
+      free_gridfunc(fgfs[i], fgfs_pinned ? fgfs_pinned[i] : 0);
+    delete[] fgfs;
+    delete[] fgfs_pinned;
+    X[0] = X[1] = X[2] = 0;
+    igfs = 0;
+    fgfs = 0;
+    fgfs_pinned = 0;
+  }
+}
 void Block::checkBlock()
 {
   int myrank;
@@ -184,12 +241,14 @@ void Block::swapList(MyList<var> *VarList1, MyList<var> *VarList2, int myrank)
   if (rank == myrank)
   {
     MyList<var> *varl1 = VarList1, *varl2 = VarList2;
-    while (varl1 && varl2)
-    {
-      misc::swap<double *>(fgfs[varl1->data->sgfn], fgfs[varl2->data->sgfn]);
-      varl1 = varl1->next;
-      varl2 = varl2->next;
-    }
+    while (varl1 && varl2)
+    {
+      misc::swap<double *>(fgfs[varl1->data->sgfn], fgfs[varl2->data->sgfn]);
+      if (fgfs_pinned)
+        misc::swap<unsigned char>(fgfs_pinned[varl1->data->sgfn], fgfs_pinned[varl2->data->sgfn]);
+      varl1 = varl1->next;
+      varl2 = varl2->next;
+    }
     if (varl1 || varl2)
     {
       cout << "error in Block::swaplist, var lists does not match." << endl;

AMSS_NCKU_source/Block.h

@@ -13,14 +13,15 @@ public:
    int shape[dim];
    double bbox[2 * dim];
    double *X[dim];
-   int rank; // where the real data locate in
-   int lev, cgpu;
-   int ingfs, fngfs;
-   int *(*igfs);
-   double *(*fgfs);
+   int rank; // where the real data locate in
+   int lev, cgpu;
+   int ingfs, fngfs;
+   int *(*igfs);
+   double *(*fgfs);
+   unsigned char *fgfs_pinned;
 
 public:
-   Block() {};
+   Block() : rank(0), lev(0), cgpu(0), ingfs(0), fngfs(0), igfs(0), fgfs(0), fgfs_pinned(0) {};
    Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fngfs, int levi, const int cgpui = 0);
 
    ~Block();

AMSS_NCKU_source/MPatch.C

@@ -11,12 +11,15 @@
 using namespace std;
 
 #include "misc.h"
-#include "MPatch.h"
-#include "Parallel.h"
-#include "fmisc.h"
-#ifdef INTERP_LB_PROFILE
-#include "interp_lb_profile.h"
-#endif
+#include "MPatch.h"
+#include "Parallel.h"
+#include "fmisc.h"
+#if USE_CUDA_BSSN
+#include "bssn_rhs_cuda.h"
+#endif
+#ifdef INTERP_LB_PROFILE
+#include "interp_lb_profile.h"
+#endif
 
 namespace
 {
@@ -154,8 +157,8 @@ void build_block_bin_index(Patch *patch, const double *DH, BlockBinIndex &index)
   index.valid = true;
 }
 
-int find_block_index_for_point(const BlockBinIndex &index, const double *pox, const double *DH)
-{
+int find_block_index_for_point(const BlockBinIndex &index, const double *pox, const double *DH)
+{
   if (!index.valid)
     return -1;
 
@@ -175,10 +178,448 @@ int find_block_index_for_point(const BlockBinIndex &index, const double *pox, co
   for (size_t bi = 0; bi < index.views.size(); bi++)
     if (point_in_block_view(index.views[bi], pox, DH))
       return int(bi);
-
-  return -1;
-}
-} // namespace
+
+  return -1;
+}
+
+inline int fortran_idint_local(double x)
+{
+  return int(x);
+}
+
+bool interp_fast_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_INTERP_FAST");
+    enabled = (!env || atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+bool interp_gpu_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_INTERP_GPU");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+bool interp_fast_compare_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_INTERP_FAST_COMPARE");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+double interp_fast_compare_tol()
+{
+  static double tol = -1.0;
+  if (tol < 0.0)
+  {
+    const char *env = getenv("AMSS_INTERP_FAST_COMPARE_TOL");
+    tol = (env && atof(env) > 0.0) ? atof(env) : 1.0e-11;
+  }
+  return tol;
+}
+
+long long interp_fast_compare_limit()
+{
+  static long long limit = -1;
+  if (limit < 0)
+  {
+    const char *env = getenv("AMSS_INTERP_FAST_COMPARE_LIMIT");
+    limit = (env && atoll(env) > 0) ? atoll(env) : 4096;
+  }
+  return limit;
+}
+
+struct FastInterpStencil
+{
+  int cxB[dim];
+  double cx[dim];
+  double wx[8];
+  double wy[8];
+  double wz[8];
+  int nsamples;
+  int loc[512];
+  unsigned char sign_mask[512];
+  double weight[512];
+};
+
+inline void lagrange_unit_weights(double x, int ordn, double *w)
+{
+  for (int i = 0; i < ordn; i++)
+  {
+    double num = 1.0;
+    double den = 1.0;
+    for (int j = 0; j < ordn; j++)
+    {
+      if (j == i)
+        continue;
+      num *= (x - double(j));
+      den *= double(i - j);
+    }
+    w[i] = num / den;
+  }
+}
+
+inline void z_unit_weights(double x, int ordn, double *w)
+{
+  if (ordn == 6)
+  {
+    static const double c_uniform[6] = {-1.0, 5.0, -10.0, 10.0, -5.0, 1.0};
+    for (int i = 0; i < 6; i++)
+    {
+      if (x == double(i))
+      {
+        for (int j = 0; j < 6; j++)
+          w[j] = (j == i) ? 1.0 : 0.0;
+        return;
+      }
+    }
+    double den = 0.0;
+    for (int i = 0; i < 6; i++)
+    {
+      w[i] = c_uniform[i] / (x - double(i));
+      den += w[i];
+    }
+    for (int i = 0; i < 6; i++)
+      w[i] /= den;
+    return;
+  }
+  lagrange_unit_weights(x, ordn, w);
+}
+
+inline bool fast_interp_map_index(int idx, int extent, int d,
+                                  int &mapped, unsigned char &mask)
+{
+  if (idx > 0)
+    mapped = idx;
+  else
+  {
+    mask |= (unsigned char)(1u << d);
+#ifdef Vertex
+#ifdef Cell
+#error Both Cell and Vertex are defined
+#endif
+    mapped = 2 - idx;
+#else
+#ifdef Cell
+    mapped = 1 - idx;
+#else
+#error Not define Vertex nor Cell
+#endif
+#endif
+  }
+  return mapped >= 1 && mapped <= extent;
+}
+
+bool prepare_fast_interp_stencil(Block *BP, const double *pox, int ordn,
+                                 int Symmetry, FastInterpStencil &st)
+{
+  if (!BP || ordn <= 0 || ordn > 8)
+    return false;
+
+  st.nsamples = 0;
+
+  const int NO_SYMM = 0;
+  const int OCTANT = 2;
+  int cmin[dim], cmax[dim], cxT[dim];
+  for (int d = 0; d < dim; d++)
+  {
+    const double *X = BP->X[d];
+    const double dX = X[1] - X[0];
+    const int cxI = fortran_idint_local((pox[d] - X[0]) / dX + 0.4) + 1;
+    st.cxB[d] = cxI - ordn / 2 + 1;
+    cxT[d] = st.cxB[d] + ordn - 1;
+    cmin[d] = 1;
+    cmax[d] = BP->shape[d];
+
+#ifdef Vertex
+#ifdef Cell
+#error Both Cell and Vertex are defined
+#endif
+    if (Symmetry == OCTANT && d < 2 && fabs(X[0]) < dX)
+      cmin[d] = -ordn / 2 + 2;
+    if (Symmetry != NO_SYMM && d == 2 && fabs(X[0]) < dX)
+      cmin[d] = -ordn / 2 + 2;
+#else
+#ifdef Cell
+    if (Symmetry == OCTANT && d < 2 && fabs(X[0]) < dX)
+      cmin[d] = -ordn / 2 + 1;
+    if (Symmetry != NO_SYMM && d == 2 && fabs(X[0]) < dX)
+      cmin[d] = -ordn / 2 + 1;
+#else
+#error Not define Vertex nor Cell
+#endif
+#endif
+
+    if (st.cxB[d] < cmin[d])
+    {
+      st.cxB[d] = cmin[d];
+      cxT[d] = st.cxB[d] + ordn - 1;
+    }
+    if (cxT[d] > cmax[d])
+    {
+      cxT[d] = cmax[d];
+      st.cxB[d] = cxT[d] + 1 - ordn;
+    }
+
+    if (st.cxB[d] > 0)
+      st.cx[d] = (pox[d] - X[st.cxB[d] - 1]) / dX;
+    else
+    {
+#ifdef Vertex
+#ifdef Cell
+#error Both Cell and Vertex are defined
+#endif
+      st.cx[d] = (pox[d] + X[1 - st.cxB[d]]) / dX;
+#else
+#ifdef Cell
+      st.cx[d] = (pox[d] + X[-st.cxB[d]]) / dX;
+#else
+#error Not define Vertex nor Cell
+#endif
+#endif
+    }
+  }
+
+  lagrange_unit_weights(st.cx[0], ordn, st.wx);
+  lagrange_unit_weights(st.cx[1], ordn, st.wy);
+  z_unit_weights(st.cx[2], ordn, st.wz);
+
+  for (int kk = 0; kk < ordn; kk++)
+  {
+    for (int jj = 0; jj < ordn; jj++)
+    {
+      for (int ii = 0; ii < ordn; ii++)
+      {
+        unsigned char mask = 0;
+        int ix, iy, iz;
+        if (!fast_interp_map_index(st.cxB[0] + ii, BP->shape[0], 0, ix, mask) ||
+            !fast_interp_map_index(st.cxB[1] + jj, BP->shape[1], 1, iy, mask) ||
+            !fast_interp_map_index(st.cxB[2] + kk, BP->shape[2], 2, iz, mask))
+          return false;
+        const int s = st.nsamples++;
+        st.loc[s] = (ix - 1) + (iy - 1) * BP->shape[0] +
+                    (iz - 1) * BP->shape[0] * BP->shape[1];
+        st.sign_mask[s] = mask;
+        st.weight[s] = st.wx[ii] * st.wy[jj] * st.wz[kk];
+      }
+    }
+  }
+  return true;
+}
+
+bool interpolate_var_list_with_stencil(Block *BP, MyList<var> *VarList,
+                                       int num_var, const double *pox,
+                                       int ordn, int Symmetry,
+                                       const FastInterpStencil &st,
+                                       double *out)
+{
+  if (num_var <= 0 || num_var > 128)
+    return false;
+
+  double *data_ptrs[128];
+  double *soa_ptrs[128];
+  var *vars[128];
+  MyList<var> *varl = VarList;
+  int k = 0;
+  while (varl)
+  {
+    if (k >= num_var)
+      return false;
+    vars[k] = varl->data;
+    data_ptrs[k] = BP->fgfs[vars[k]->sgfn];
+    soa_ptrs[k] = vars[k]->SoA;
+    out[k] = 0.0;
+    varl = varl->next;
+    k++;
+  }
+
+  if (k != num_var)
+    return false;
+
+  for (int s = 0; s < st.nsamples; s++)
+  {
+    const int loc = st.loc[s];
+    const double w = st.weight[s];
+    const unsigned char mask = st.sign_mask[s];
+    if (mask == 0)
+    {
+      for (int v = 0; v < num_var; v++)
+        out[v] += w * data_ptrs[v][loc];
+    }
+    else
+    {
+      for (int v = 0; v < num_var; v++)
+      {
+        const double *SoA = soa_ptrs[v];
+        double sgn = 1.0;
+        if (mask & 1u)
+          sgn *= SoA[0];
+        if (mask & 2u)
+          sgn *= SoA[1];
+        if (mask & 4u)
+          sgn *= SoA[2];
+        out[v] += w * sgn * data_ptrs[v][loc];
+      }
+    }
+  }
+
+  if (interp_fast_compare_enabled())
+  {
+    static int report_count = 0;
+    static long long compare_calls = 0;
+    if (compare_calls++ >= interp_fast_compare_limit())
+      return true;
+    const double tol = interp_fast_compare_tol();
+    varl = VarList;
+    k = 0;
+    while (varl)
+    {
+      var *vp = vars[k];
+      double ref = 0.0;
+      double x = pox[0], y = pox[1], z = pox[2];
+      f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2],
+                      BP->fgfs[vp->sgfn], ref,
+                      x, y, z, ordn, vp->SoA, Symmetry);
+      const double diff = fabs(ref - out[k]);
+      const double scale = 1.0 + fabs(ref);
+      if (diff > tol * scale && report_count < 32)
+      {
+        int rank = 0;
+        MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+        fprintf(stderr,
+                "[AMSS-INTERP-CMP][rank %d] var=%s diff=%.17e ref=%.17e fast=%.17e p=(%.17e,%.17e,%.17e)\n",
+                rank, vp->name, diff, ref, out[k], pox[0], pox[1], pox[2]);
+        report_count++;
+      }
+      varl = varl->next;
+      k++;
+    }
+  }
+
+  return true;
+}
+
+bool interpolate_var_list_fast(Block *BP, MyList<var> *VarList, int num_var,
+                               const double *pox, int ordn, int Symmetry,
+                               double *out)
+{
+  if (!interp_fast_enabled())
+    return false;
+
+  FastInterpStencil st;
+  if (!prepare_fast_interp_stencil(BP, pox, ordn, Symmetry, st))
+    return false;
+
+  return interpolate_var_list_with_stencil(BP, VarList, num_var, pox,
+                                           ordn, Symmetry, st, out);
+}
+
+struct CachedInterpPoint
+{
+  Block *bp;
+  int owner_rank;
+  FastInterpStencil stencil;
+};
+
+struct SurfaceInterpCache
+{
+  Patch *patch;
+  int NN;
+  int symmetry;
+  double key[9];
+  vector<CachedInterpPoint> points;
+
+  SurfaceInterpCache() : patch(0), NN(0), symmetry(-1) {}
+};
+
+bool surface_cache_key_matches(const SurfaceInterpCache &cache, Patch *patch,
+                               int NN, double **XX, int Symmetry)
+{
+  if (cache.patch != patch || cache.NN != NN || cache.symmetry != Symmetry ||
+      int(cache.points.size()) != NN || NN <= 0)
+    return false;
+  const int mid = NN / 2;
+  const int last = NN - 1;
+  const int ids[3] = {0, mid, last};
+  int p = 0;
+  for (int q = 0; q < 3; q++)
+    for (int d = 0; d < dim; d++)
+      if (cache.key[p++] != XX[d][ids[q]])
+        return false;
+  return true;
+}
+
+SurfaceInterpCache *find_surface_cache(Patch *patch, int NN, double **XX,
+                                       int Symmetry)
+{
+  static vector<SurfaceInterpCache> caches;
+  for (size_t i = 0; i < caches.size(); i++)
+    if (surface_cache_key_matches(caches[i], patch, NN, XX, Symmetry))
+      return &caches[i];
+  if (caches.size() >= 24)
+    caches.erase(caches.begin());
+  caches.push_back(SurfaceInterpCache());
+  return &caches.back();
+}
+
+bool build_surface_cache(SurfaceInterpCache &cache, Patch *patch, int NN,
+                         double **XX, int Symmetry, const double *DH,
+                         const BlockBinIndex &block_index, int ordn)
+{
+  int myrank = 0;
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+  cache.patch = patch;
+  cache.NN = NN;
+  cache.symmetry = Symmetry;
+  cache.points.clear();
+  cache.points.resize(NN);
+  const int mid = NN / 2;
+  const int last = NN - 1;
+  const int ids[3] = {0, mid, last};
+  int p = 0;
+  for (int q = 0; q < 3; q++)
+    for (int d = 0; d < dim; d++)
+      cache.key[p++] = XX[d][ids[q]];
+
+  for (int j = 0; j < NN; j++)
+  {
+    double pox[dim];
+    for (int d = 0; d < dim; d++)
+      pox[d] = XX[d][j];
+    const int block_i = find_block_index_for_point(block_index, pox, DH);
+    if (block_i < 0)
+    {
+      cache.points[j].bp = 0;
+      cache.points[j].owner_rank = -1;
+      continue;
+    }
+    Block *BP = block_index.views[block_i].bp;
+    cache.points[j].bp = BP;
+    cache.points[j].owner_rank = BP->rank;
+    cache.points[j].stencil.nsamples = 0;
+    if (BP->rank == myrank)
+    {
+      if (!prepare_fast_interp_stencil(BP, pox, ordn, Symmetry,
+                                       cache.points[j].stencil))
+        return false;
+    }
+  }
+  return true;
+}
+} // namespace
 
 Patch::Patch(int DIM, int *shapei, double *bboxi, int levi, bool buflog, int Symmetry) : lev(levi)
 {
@@ -561,22 +1002,26 @@ void Patch::Interp_Points(MyList<var> *VarList,
     if (block_i >= 0)
     {
       Block *BP = block_index.views[block_i].bp;
-      owner_rank[j] = BP->rank;
-      if (myrank == BP->rank)
-      {
-        //---> interpolation
-        varl = 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],
-                          pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
-          varl = varl->next;
-          k++;
-        }
-      }
-    }
-  }
+      owner_rank[j] = BP->rank;
+      if (myrank == BP->rank)
+      {
+        //---> interpolation
+        if (!interpolate_var_list_fast(BP, VarList, num_var, pox, ordn,
+                                       Symmetry, Shellf + j * num_var))
+        {
+          varl = 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],
+                            pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
+            varl = varl->next;
+            k++;
+          }
+        }
+      }
+    }
+  }
 
   // Replace MPI_Allreduce with per-owner MPI_Bcast:
   // Group consecutive points by owner rank and broadcast each group.
@@ -659,10 +1104,8 @@ void Patch::Interp_Points(MyList<var> *VarList,
     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[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;
@@ -670,12 +1113,117 @@ void Patch::Interp_Points(MyList<var> *VarList,
   double DH[dim];
   for (int i = 0; i < dim; i++)
     DH[i] = getdX(i);
-  BlockBinIndex block_index;
-  build_block_bin_index(this, DH, block_index);
-
-  // --- Interpolation phase (identical to original) ---
-  for (int j = 0; j < NN; j++)
-  {
+  BlockBinIndex block_index;
+  build_block_bin_index(this, DH, block_index);
+  SurfaceInterpCache *surface_cache = 0;
+  bool use_surface_cache = false;
+  if (interp_fast_enabled())
+  {
+    surface_cache = find_surface_cache(this, NN, XX, Symmetry);
+    use_surface_cache = surface_cache_key_matches(*surface_cache, this, NN, XX, Symmetry);
+    if (!use_surface_cache)
+      use_surface_cache = build_surface_cache(*surface_cache, this, NN, XX,
+                                              Symmetry, DH, block_index, ordn);
+  }
+
+  // --- Interpolation phase (identical to original) ---
+#if USE_CUDA_BSSN
+  const bool use_gpu_interp = interp_gpu_enabled() && use_surface_cache && num_var == 2 &&
+                              VarList && VarList->next && !VarList->next->next;
+#else
+  const bool use_gpu_interp = false;
+#endif
+  if (use_gpu_interp)
+  {
+#if USE_CUDA_BSSN
+    vector<vector<int> > local_points(block_index.views.size());
+    for (int j = 0; j < NN; j++)
+    {
+      for (int i = 0; i < dim; i++)
+      {
+        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);
+        }
+      }
+
+      CachedInterpPoint &cp = surface_cache->points[j];
+      Block *BP = cp.bp;
+      owner_rank[j] = cp.owner_rank;
+      if (BP && myrank == BP->rank)
+      {
+        for (size_t bi = 0; bi < block_index.views.size(); bi++)
+        {
+          if (block_index.views[bi].bp == BP)
+          {
+            local_points[bi].push_back(j);
+            break;
+          }
+        }
+      }
+    }
+
+    var *v0 = VarList->data;
+    var *v1 = VarList->next->data;
+    double soa6[6] = {
+        v0->SoA[0], v0->SoA[1], v0->SoA[2],
+        v1->SoA[0], v1->SoA[1], v1->SoA[2]};
+
+    for (size_t bi = 0; bi < local_points.size(); bi++)
+    {
+      const int count = int(local_points[bi].size());
+      if (count <= 0)
+        continue;
+
+      Block *BP = block_index.views[bi].bp;
+      vector<double> px(count), py(count), pz(count), out(2 * count);
+      for (int q = 0; q < count; q++)
+      {
+        const int j = local_points[bi][q];
+        px[q] = XX[0][j];
+        py[q] = XX[1][j];
+        pz[q] = XX[2][j];
+      }
+
+      const double dx = BP->X[0][1] - BP->X[0][0];
+      const double dy = BP->X[1][1] - BP->X[1][0];
+      const double dz = BP->X[2][1] - BP->X[2][0];
+      const int ok = bssn_cuda_interp_host_two_fields(
+          BP, BP->shape,
+          BP->fgfs[v0->sgfn], BP->fgfs[v1->sgfn],
+          BP->X[0][0], BP->X[1][0], BP->X[2][0],
+          dx, dy, dz,
+          &px[0], &py[0], &pz[0], count,
+          ordn, Symmetry, soa6, &out[0]);
+      if (ok != 0)
+      {
+        if (myrank == 0)
+          cout << "Patch::Interp_Points: CUDA two-field interpolation failed" << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+
+      for (int q = 0; q < count; q++)
+      {
+        const int j = local_points[bi][q];
+        Shellf[j * num_var] = out[2 * q];
+        Shellf[j * num_var + 1] = out[2 * q + 1];
+      }
+    }
+#endif
+  }
+  else
+  {
+  for (int j = 0; j < NN; j++)
+  {
     double pox[dim];
     for (int i = 0; i < dim; i++)
     {
@@ -692,28 +1240,59 @@ void Patch::Interp_Points(MyList<var> *VarList,
             cout << ") is out of current Patch." << endl;
         }
         MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
-
-    const int block_i = find_block_index_for_point(block_index, pox, DH);
-    if (block_i >= 0)
-    {
-      Block *BP = block_index.views[block_i].bp;
-      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 (use_surface_cache)
+    {
+      CachedInterpPoint &cp = surface_cache->points[j];
+      Block *BP = cp.bp;
+      owner_rank[j] = cp.owner_rank;
+      if (BP && myrank == BP->rank)
+      {
+        if (!interpolate_var_list_with_stencil(BP, VarList, num_var, pox,
+                                               ordn, Symmetry, cp.stencil,
+                                               Shellf + j * num_var))
+        {
+          MyList<var> *varl_fallback = VarList;
+          int k = 0;
+          while (varl_fallback)
+          {
+            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl_fallback->data->sgfn], Shellf[j * num_var + k],
+                            pox[0], pox[1], pox[2], ordn, varl_fallback->data->SoA, Symmetry);
+            varl_fallback = varl_fallback->next;
+            k++;
+          }
+        }
+      }
+    }
+    else
+    {
+      const int block_i = find_block_index_for_point(block_index, pox, DH);
+      if (block_i >= 0)
+      {
+        Block *BP = block_index.views[block_i].bp;
+        owner_rank[j] = BP->rank;
+      if (myrank == BP->rank)
+      {
+        if (!interpolate_var_list_fast(BP, VarList, num_var, pox, ordn,
+                                       Symmetry, Shellf + j * num_var))
+        {
+          MyList<var> *varl_fallback = VarList;
+          int k = 0;
+          while (varl_fallback)
+          {
+            f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl_fallback->data->sgfn], Shellf[j * num_var + k],
+                            pox[0], pox[1], pox[2], ordn, varl_fallback->data->SoA, Symmetry);
+            varl_fallback = varl_fallback->next;
+            k++;
+          }
+        }
+      }
+      }
+    }
+  }
+  }
 
 #ifdef INTERP_LB_PROFILE
   double t_interp_end = MPI_Wtime();
@@ -965,22 +1544,26 @@ void Patch::Interp_Points(MyList<var> *VarList,
     if (block_i >= 0)
     {
       Block *BP = block_index.views[block_i].bp;
-      owner_rank[j] = BP->rank;
-      if (myrank == BP->rank)
-      {
-        //---> interpolation
-        varl = 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],
-                          pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
-          varl = varl->next;
-          k++;
-        }
-      }
-    }
-  }
+      owner_rank[j] = BP->rank;
+      if (myrank == BP->rank)
+      {
+        //---> interpolation
+        if (!interpolate_var_list_fast(BP, VarList, num_var, pox, ordn,
+                                       Symmetry, Shellf + j * num_var))
+        {
+          varl = 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],
+                            pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
+            varl = varl->next;
+            k++;
+          }
+        }
+      }
+    }
+  }
 
   // Collect unique global owner ranks and translate to local ranks in Comm_here
   // Then broadcast each owner's points via MPI_Bcast on Comm_here

AMSS_NCKU_source/Parallel.h

@@ -104,6 +104,14 @@ namespace Parallel
     double **recv_bufs;
     int *send_buf_caps;
     int *recv_buf_caps;
+    unsigned char *send_buf_pinned;
+    unsigned char *recv_buf_pinned;
+    unsigned char *send_buf_is_dev;
+    unsigned char *recv_buf_is_dev;
+    int *send_buf_caps_dev;
+    int *recv_buf_caps_dev;
+    double **send_bufs_dev;
+    double **recv_bufs_dev;
     MPI_Request *reqs;
     MPI_Status *stats;
     int max_reqs;
@@ -111,12 +119,14 @@ namespace Parallel
     int *tc_req_node;
     int *tc_req_is_recv;
     int *tc_completed;
+    bool cuda_aware_mode;
     SyncCache();
     void invalidate();
     void destroy();
   };
 
   void Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, SyncCache &cache);
+  void Sync_ensure_cache(MyList<Patch> *PatL, int Symmetry, SyncCache &cache);
   void transfer_cached(MyList<gridseg> **src, MyList<gridseg> **dst,
                        MyList<var> *VarList1, MyList<var> *VarList2,
                        int Symmetry, SyncCache &cache);
@@ -179,13 +189,13 @@ namespace Parallel
   MyList<Parallel::gridseg> *clone_gsl(MyList<Parallel::gridseg> *p, bool first_only);
   MyList<Parallel::gridseg> *build_bulk_gsl(Patch *Pat); // similar to build_owned_gsl0 but does not care rank issue
   MyList<Parallel::gridseg> *build_bulk_gsl(Block *bp, Patch *Pat);
-  void build_PhysBD_gstl(Patch *Pat, MyList<Parallel::gridseg> *srci, MyList<Parallel::gridseg> *dsti,
-                         MyList<Parallel::gridseg> **out_src, MyList<Parallel::gridseg> **out_dst);
-  void PeriodicBD(Patch *Pat, MyList<var> *VarList, int Symmetry);
-  double L2Norm(Patch *Pat, var *vf);
-  void L2Norm7(Patch *Pat, var **vf, double *norms);
-  void checkgsl(MyList<Parallel::gridseg> *pp, bool first_only);
-  void checkvarl(MyList<var> *pp, bool first_only);
+  void build_PhysBD_gstl(Patch *Pat, MyList<Parallel::gridseg> *srci, MyList<Parallel::gridseg> *dsti,
+                         MyList<Parallel::gridseg> **out_src, MyList<Parallel::gridseg> **out_dst);
+  void PeriodicBD(Patch *Pat, MyList<var> *VarList, int Symmetry);
+  double L2Norm(Patch *Pat, var *vf);
+  void L2Norm7(Patch *Pat, var **vf, double *norms);
+  void checkgsl(MyList<Parallel::gridseg> *pp, bool first_only);
+  void checkvarl(MyList<var> *pp, bool first_only);
   MyList<Parallel::gridseg> *divide_gsl(MyList<Parallel::gridseg> *p, Patch *Pat);
   MyList<Parallel::gridseg> *divide_gs(MyList<Parallel::gridseg> *p, Patch *Pat);
   void prepare_inter_time_level(Patch *Pat,
@@ -217,12 +227,12 @@ namespace Parallel
   void aligncheck(double *bbox0, double *bboxl, int lev, double *DH0, int *shape);
   bool point_locat_gsl(double *pox, MyList<Parallel::gridseg> *gsl);
   void checkpatchlist(MyList<Patch> *PatL, bool buflog);
-
-  double L2Norm(Patch *Pat, var *vf, MPI_Comm Comm_here);
-  void L2Norm7(Patch *Pat, var **vf, double *norms, MPI_Comm Comm_here);
-  bool PatList_Interp_Points(MyList<Patch> *PatL, MyList<var> *VarList,
-                             int NN, double **XX,
-                             double *Shellf, int Symmetry, MPI_Comm Comm_here);
+
+  double L2Norm(Patch *Pat, var *vf, MPI_Comm Comm_here);
+  void L2Norm7(Patch *Pat, var **vf, double *norms, MPI_Comm Comm_here);
+  bool PatList_Interp_Points(MyList<Patch> *PatL, MyList<var> *VarList,
+                             int NN, double **XX,
+                             double *Shellf, int Symmetry, MPI_Comm Comm_here);
 #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);

AMSS_NCKU_source/ShellPatch.C

@@ -59,7 +59,7 @@ bool shell_fast_interp_enabled()
   if (enabled < 0)
   {
     const char *env = getenv("AMSS_SHELL_FAST_INTERP");
-    enabled = (!env || atoi(env) != 0) ? 1 : 0;
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
   }
   return enabled != 0;
 }
@@ -70,7 +70,7 @@ bool shell_parallel_interp_enabled()
   if (enabled < 0)
   {
     const char *env = getenv("AMSS_SHELL_PARALLEL_INTERP");
-    enabled = (!env || atoi(env) != 0) ? 1 : 0;
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
   }
   return enabled != 0;
 }

AMSS_NCKU_source/TwoPunctures.C

@@ -27,7 +27,7 @@ using namespace std;
 #endif
 
 #include "TwoPunctures.h"
-#include <mkl_cblas.h>
+#include <cblas.h>
 
 TwoPunctures::TwoPunctures(double mp, double mm, double b,
                            double P_plusx, double P_plusy, double P_plusz,

AMSS_NCKU_source/Z4c_class.C

@@ -3,6 +3,7 @@
 #include <sstream>
 #include <cstdio>
 #include <map>
+#include <string>
 using namespace std;
 #else
 #include <stdio.h>
@@ -28,6 +29,20 @@ using namespace std;
 #include "kodiss.h"
 #include "parameters.h"
 
+#ifndef USE_CUDA_Z4C
+#define USE_CUDA_Z4C 0
+#endif
+
+#if USE_CUDA_Z4C && (ABEtype == 2)
+#include "z4c_rhs_cuda.h"
+#endif
+#if USE_CUDA_BSSN
+#include "bssn_rhs_cuda.h"
+#ifdef WithShell
+#include "bssn_gpu.h"
+#endif
+#endif
+
 #ifdef With_AHF
 #include "derivatives.h"
 #include "myglobal.h"
@@ -37,6 +52,81 @@ using namespace std;
 
 // Define Z4c_class
 
+#if USE_CUDA_Z4C && (ABEtype == 2) && defined(WithShell)
+// GPU-accelerated Z4C shell RHS: same parameter signature as f_compute_rhs_Z4c_ss.
+// Internally calls gpu_rhs_z4c_ss which modifies trK→trKd before upload,
+// runs BSSN algebraic kernels, then applies Z4C post-processing (TZ_rhs, damping).
+extern "C" {
+static int cuda_compute_rhs_z4c_ss(
+    int *ex, double &T, double *crho, double *sigma, double *R,
+    double *X, double *Y, double *Z,
+    double *drhodx, double *drhody, double *drhodz,
+    double *dsigmadx, double *dsigmady, double *dsigmadz,
+    double *dRdx, double *dRdy, double *dRdz,
+    double *drhodxx, double *drhodxy, double *drhodxz, double *drhodyy, double *drhodyz, double *drhodzz,
+    double *dsigmadxx, double *dsigmadxy, double *dsigmadxz, double *dsigmadyy, double *dsigmadyz, double *dsigmadzz,
+    double *dRdxx, double *dRdxy, double *dRdxz, double *dRdyy, double *dRdyz, double *dRdzz,
+    double *chi, double *trK,
+    double *gxx, double *gxy, double *gxz, double *gyy, double *gyz, double *gzz,
+    double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
+    double *Gamx, double *Gamy, double *Gamz,
+    double *Lap, double *betax, double *betay, double *betaz,
+    double *dtSfx, double *dtSfy, double *dtSfz,
+    double *TZ,
+    double *chi_rhs, double *trK_rhs,
+    double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
+    double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
+    double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
+    double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
+    double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
+    double *TZ_rhs,
+    double *rho_mat, double *Sx, double *Sy, double *Sz,
+    double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
+    double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
+    double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
+    double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
+    double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
+    double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
+    double *Gmx_Res, double *Gmy_Res, double *Gmz_Res,
+    int &Symmetry, int &Lev, double &eps, int &sst, int &co)
+{
+    return gpu_rhs_z4c_ss(0, 0,  // calledby=ABE_main, mpi_rank=device_0
+                     ex, T, crho, sigma, R, X, Y, Z,
+                     drhodx, drhody, drhodz,
+                     dsigmadx, dsigmady, dsigmadz,
+                     dRdx, dRdy, dRdz,
+                     drhodxx, drhodxy, drhodxz, drhodyy, drhodyz, drhodzz,
+                     dsigmadxx, dsigmadxy, dsigmadxz, dsigmadyy, dsigmadyz, dsigmadzz,
+                     dRdxx, dRdxy, dRdxz, dRdyy, dRdyz, dRdzz,
+                     chi, trK,
+                     gxx, gxy, gxz, gyy, gyz, gzz,
+                     Axx, Axy, Axz, Ayy, Ayz, Azz,
+                     Gamx, Gamy, Gamz,
+                     Lap, betax, betay, betaz,
+                     dtSfx, dtSfy, dtSfz,
+                     TZ,
+                     chi_rhs, trK_rhs,
+                     gxx_rhs, gxy_rhs, gxz_rhs, gyy_rhs, gyz_rhs, gzz_rhs,
+                     Axx_rhs, Axy_rhs, Axz_rhs, Ayy_rhs, Ayz_rhs, Azz_rhs,
+                     Gamx_rhs, Gamy_rhs, Gamz_rhs,
+                     Lap_rhs, betax_rhs, betay_rhs, betaz_rhs,
+                     dtSfx_rhs, dtSfy_rhs, dtSfz_rhs,
+                     TZ_rhs,
+                     rho_mat, Sx, Sy, Sz,
+                     Sxx, Sxy, Sxz, Syy, Syz, Szz,
+                     Gamxxx, Gamxxy, Gamxxz, Gamxyy, Gamxyz, Gamxzz,
+                     Gamyxx, Gamyxy, Gamyxz, Gamyyy, Gamyyz, Gamyzz,
+                     Gamzxx, Gamzxy, Gamzxz, Gamzyy, Gamzyz, Gamzzz,
+                     Rxx, Rxy, Rxz, Ryy, Ryz, Rzz,
+                     ham_Res, movx_Res, movy_Res, movz_Res,
+                     Gmx_Res, Gmy_Res, Gmz_Res,
+                     Symmetry, Lev, eps, sst, co);
+}
+}
+// Redirect all Z4C shell RHS calls in Step/SHStep to GPU
+#define f_compute_rhs_Z4c_ss cuda_compute_rhs_z4c_ss
+#endif
+
 // This class inherits some members and methods from the parent `bssn_class` and modifies others.
 // The modified members and methods are defined below (and in the header Z4c_class.h).
 // The remaining members/methods are inherited from `bssn_class` (declared in bssn_class.h).
@@ -132,6 +222,13 @@ void Z4c_class::Initialize()
     PhysTime = StartTime;
     Setup_Black_Hole_position();
   }
+
+  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];
+  sync_cache_restrict = new Parallel::SyncCache[GH->levels];
+  sync_cache_outbd = new Parallel::SyncCache[GH->levels];
 }
 
 //================================================================================================
@@ -170,8 +267,573 @@ Z4c_class::~Z4c_class()
 #ifndef CPBC
 // for sommerfeld boundary
 
+#if USE_CUDA_Z4C && (ABEtype == 2)
+#if (MRBD == 2)
+#error "USE_CUDA_Z4C resident path does not support MRBD == 2"
+#endif
+
+namespace {
+
+static const int k_z4c_cuda_bh_state_indices[3] = {18, 19, 20};
+
+bool fill_z4c_cuda_views(Block *cg, MyList<var> *vars,
+                         double **host_views,
+                         double *propspeeds = 0,
+                         double *soa_flat = 0)
+{
+  int idx = 0;
+  while (vars && idx < Z4C_CUDA_STATE_COUNT)
+  {
+    host_views[idx] = cg->fgfs[vars->data->sgfn];
+    if (propspeeds)
+      propspeeds[idx] = vars->data->propspeed;
+    if (soa_flat)
+    {
+      soa_flat[3 * idx + 0] = vars->data->SoA[0];
+      soa_flat[3 * idx + 1] = vars->data->SoA[1];
+      soa_flat[3 * idx + 2] = vars->data->SoA[2];
+    }
+    vars = vars->next;
+    ++idx;
+  }
+  return idx == Z4C_CUDA_STATE_COUNT && vars == 0;
+}
+
+void z4c_cuda_download_level_state(MyList<Patch> *PatL, MyList<var> *vars, int myrank, bool release_ctx)
+{
+  MyList<Patch> *Pp = PatL;
+  while (Pp)
+  {
+    MyList<Block> *BP = Pp->data->blb;
+    while (BP)
+    {
+      Block *cg = BP->data;
+      if (myrank == cg->rank && z4c_cuda_has_resident_state(cg))
+      {
+        double *state_out[Z4C_CUDA_STATE_COUNT];
+        if (!fill_z4c_cuda_views(cg, vars, state_out))
+        {
+          cout << "CUDA Z4C state list mismatch on resident state download" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+        if (z4c_cuda_download_resident_state(cg, cg->shape, state_out))
+        {
+          cout << "CUDA Z4C resident state download failed" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+        if (release_ctx)
+          z4c_cuda_release_step_ctx(cg);
+      }
+      if (BP == Pp->data->ble)
+        break;
+      BP = BP->next;
+    }
+    Pp = Pp->next;
+  }
+}
+
+bool z4c_cuda_patch_contains_point(Patch *patch, const double *point)
+{
+  if (!patch)
+    return false;
+  for (int d = 0; d < dim; d++)
+  {
+    const double h = patch->getdX(d);
+    const double lo = patch->bbox[d] + patch->lli[d] * h;
+    const double hi = patch->bbox[dim + d] - patch->uui[d] * h;
+    if (point[d] < lo || point[d] > hi)
+      return false;
+  }
+  return true;
+}
+
+bool z4c_cuda_point_in_block(Patch *patch, Block *block,
+                             const double *point, const double *DH)
+{
+  if (!patch || !block)
+    return false;
+  for (int d = 0; d < dim; d++)
+  {
+    double llb;
+    double uub;
+#ifdef Vertex
+#ifdef Cell
+#error Both Cell and Vertex are defined
+#endif
+    llb = (feq(block->bbox[d], patch->bbox[d], DH[d] / 2))
+              ? block->bbox[d] + patch->lli[d] * DH[d]
+              : block->bbox[d] + (ghost_width - 0.5) * DH[d];
+    uub = (feq(block->bbox[dim + d], patch->bbox[dim + d], DH[d] / 2))
+              ? block->bbox[dim + d] - patch->uui[d] * DH[d]
+              : block->bbox[dim + d] - (ghost_width - 0.5) * DH[d];
+#else
+#ifdef Cell
+    llb = (feq(block->bbox[d], patch->bbox[d], DH[d] / 2))
+              ? block->bbox[d] + patch->lli[d] * DH[d]
+              : block->bbox[d] + ghost_width * DH[d];
+    uub = (feq(block->bbox[dim + d], patch->bbox[dim + d], DH[d] / 2))
+              ? block->bbox[dim + d] - patch->uui[d] * DH[d]
+              : block->bbox[dim + d] - ghost_width * DH[d];
+#else
+#error Not define Vertex nor Cell
+#endif
+#endif
+    if (point[d] - llb < -DH[d] / 2 || point[d] - uub > DH[d] / 2)
+      return false;
+  }
+  return true;
+}
+
+int z4c_cuda_interp_tile_start(const double *coords, int n, double x, double dx, int ordn)
+{
+  if (!coords || n <= ordn)
+    return 0;
+  int cxi = int((x - coords[0]) / dx + 0.4) + 1;
+  int start = cxi - ordn / 2;
+  if (start < 0)
+    start = 0;
+  const int max_start = n - ordn;
+  if (start > max_start)
+    start = max_start;
+  return start;
+}
+
+bool z4c_cuda_interp_bh_point_resident(MyList<Patch> *PatL,
+                                       int myrank,
+                                       const double *point,
+                                       var *forx, var *fory, var *forz,
+                                       int Symmetry,
+                                       double *shellf)
+{
+  const int ordn = 2 * ghost_width;
+  int owner_rank = -1;
+
+  shellf[0] = shellf[1] = shellf[2] = 0.0;
+
+  MyList<Patch> *PL = PatL;
+  while (PL)
+  {
+    Patch *patch = PL->data;
+    if (!z4c_cuda_patch_contains_point(patch, point))
+    {
+      PL = PL->next;
+      continue;
+    }
+
+    double DH[dim];
+    for (int d = 0; d < dim; d++)
+      DH[d] = patch->getdX(d);
+
+    MyList<Block> *BP = patch->blb;
+    while (BP)
+    {
+      Block *block = BP->data;
+      if (z4c_cuda_point_in_block(patch, block, point, DH))
+      {
+        owner_rank = block->rank;
+        if (myrank == owner_rank)
+        {
+          int interp_ordn = ordn;
+          int interp_sym = Symmetry;
+          double x = point[0];
+          double y = point[1];
+          double z = point[2];
+
+          if (z4c_cuda_has_resident_state(block) &&
+              block->shape[0] >= ordn && block->shape[1] >= ordn && block->shape[2] >= ordn)
+          {
+            const int sx = ordn;
+            const int sy = ordn;
+            const int sz = ordn;
+            const int region_all = sx * sy * sz;
+            const int i0 = z4c_cuda_interp_tile_start(block->X[0], block->shape[0], x, DH[0], ordn);
+            const int j0 = z4c_cuda_interp_tile_start(block->X[1], block->shape[1], y, DH[1], ordn);
+            const int k0 = z4c_cuda_interp_tile_start(block->X[2], block->shape[2], z, DH[2], ordn);
+            double *packed_fields = new double[3 * region_all];
+            var *vars[3] = {forx, fory, forz};
+            for (int f = 0; f < 3; f++)
+            {
+              if (z4c_cuda_pack_state_region_to_host_buffer(block,
+                                                            k_z4c_cuda_bh_state_indices[f],
+                                                            packed_fields + f * region_all,
+                                                            block->shape,
+                                                            i0, j0, k0,
+                                                            sx, sy, sz) != 0)
+              {
+                delete[] packed_fields;
+                cout << "CUDA Z4C BH tile download failed" << endl;
+                MPI_Abort(MPI_COMM_WORLD, 1);
+              }
+              int tile_shape[3] = {sx, sy, sz};
+              f_global_interp(tile_shape,
+                              block->X[0] + i0,
+                              block->X[1] + j0,
+                              block->X[2] + k0,
+                              packed_fields + f * region_all,
+                              shellf[f],
+                              x, y, z,
+                              interp_ordn,
+                              vars[f]->SoA,
+                              interp_sym);
+            }
+            delete[] packed_fields;
+          }
+          else
+          {
+            f_global_interp(block->shape, block->X[0], block->X[1], block->X[2],
+                            block->fgfs[forx->sgfn], shellf[0],
+                            x, y, z, interp_ordn, forx->SoA, interp_sym);
+            f_global_interp(block->shape, block->X[0], block->X[1], block->X[2],
+                            block->fgfs[fory->sgfn], shellf[1],
+                            x, y, z, interp_ordn, fory->SoA, interp_sym);
+            f_global_interp(block->shape, block->X[0], block->X[1], block->X[2],
+                            block->fgfs[forz->sgfn], shellf[2],
+                            x, y, z, interp_ordn, forz->SoA, interp_sym);
+          }
+        }
+        break;
+      }
+      if (BP == patch->ble)
+        break;
+      BP = BP->next;
+    }
+
+    if (owner_rank >= 0)
+      break;
+    PL = PL->next;
+  }
+
+  if (owner_rank < 0)
+    return false;
+
+  MPI_Bcast(shellf, 3, MPI_DOUBLE, owner_rank, MPI_COMM_WORLD);
+  return true;
+}
+
+bool z4c_cuda_compute_porg_rhs_resident(cgh *GH,
+                                        int ilev,
+                                        int myrank,
+                                        int BH_num,
+                                        double **BH_PS,
+                                        double **BH_RHS,
+                                        var *forx, var *fory, var *forz,
+                                        int Symmetry)
+{
+  for (int n = 0; n < BH_num; n++)
+  {
+    double shellf[3] = {0.0, 0.0, 0.0};
+    int lev = ilev;
+    while (lev >= 0 &&
+           !z4c_cuda_interp_bh_point_resident(GH->PatL[lev], myrank, BH_PS[n],
+                                              forx, fory, forz, Symmetry, shellf))
+    {
+      --lev;
+    }
+    if (lev < 0)
+      return false;
+    BH_RHS[n][0] = -shellf[0];
+    BH_RHS[n][1] = -shellf[1];
+    BH_RHS[n][2] = -shellf[2];
+  }
+  return true;
+}
+
+bool z4c_cuda_resident_step_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_Z4C_CUDA_RESIDENT");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+} // namespace
+#endif
+
 void Z4c_class::Step(int lev, int YN)
 {
+#if USE_CUDA_Z4C && (ABEtype == 2)
+  double dT_lev = dT * pow(0.5, Mymax(lev, trfls));
+#ifdef With_AHF
+  AH_Step_Find(lev, dT_lev);
+#endif
+  bool BB = fgt(PhysTime, StartTime, dT_lev / 2);
+  double ndeps = numepss;
+  if (lev < GH->movls)
+    ndeps = numepsb;
+  double TRK4 = PhysTime;
+  int iter_count = 0;
+  int pre = 0, cor = 1;
+  int ERROR = 0;
+
+#ifdef WithShell
+  if (bssn_cuda_use_resident_sync(lev))
+  {
+    for (int dl = 0; dl < GH->levels; dl++)
+      bssn_cuda_download_level_state_if_present(GH->PatL[dl], StateList, myrank);
+  }
+#endif
+
+  MyList<Patch> *Pp = GH->PatL[lev];
+  while (Pp)
+  {
+    MyList<Block> *BP = Pp->data->blb;
+    while (BP)
+    {
+      Block *cg = BP->data;
+      if (myrank == cg->rank)
+      {
+        double *state_in[Z4C_CUDA_STATE_COUNT];
+        double *state_out[Z4C_CUDA_STATE_COUNT];
+        double propspeed[Z4C_CUDA_STATE_COUNT];
+        double soa_flat[3 * Z4C_CUDA_STATE_COUNT];
+        if (!fill_z4c_cuda_views(cg, StateList, state_in, propspeed, soa_flat) ||
+            !fill_z4c_cuda_views(cg, SynchList_pre, state_out))
+        {
+          cout << "CUDA Z4C state list mismatch on predictor step" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+        int apply_bam_bc = 0;
+#if (MRBD == 0)
+#if (SommerType == 0)
+        apply_bam_bc = (lev == 0) ? 1 : 0;
+#endif
+#elif (MRBD == 1)
+        apply_bam_bc = 1;
+#endif
+        int keep_resident_state = z4c_cuda_resident_step_enabled() ? 1 : 0;
+        int apply_enforce_ga = 0;
+#if (AGM == 0)
+        apply_enforce_ga = 1;
+#endif
+        if (z4c_cuda_rk4_substep(cg,
+                                 cg->shape, cg->X[0], cg->X[1], cg->X[2],
+                                 state_in, state_out,
+                                 propspeed, soa_flat, Pp->data->bbox,
+                                 dT_lev, TRK4, iter_count, apply_bam_bc,
+                                 Symmetry, lev, ndeps, pre,
+                                 keep_resident_state, apply_enforce_ga, chitiny))
+        {
+          cout << "CUDA Z4C predictor substep failed in domain: ("
+               << cg->bbox[0] << ":" << cg->bbox[3] << ","
+               << cg->bbox[1] << ":" << cg->bbox[4] << ","
+               << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+          ERROR = 1;
+        }
+      }
+      if (BP == Pp->data->ble)
+        break;
+      BP = BP->next;
+    }
+    Pp = Pp->next;
+  }
+
+  {
+    int erh = ERROR;
+    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+  }
+  if (ERROR)
+  {
+    if (myrank == 0 && ErrorMonitor->outfile)
+      ErrorMonitor->outfile << "CUDA Z4C failed in predictor at t = " << PhysTime
+                            << ", lev = " << lev << endl;
+    MPI_Abort(MPI_COMM_WORLD, 1);
+  }
+
+  Parallel::Sync_cached(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev]);
+
+  if (BH_num > 0 && lev == GH->levels - 1)
+  {
+    compute_Porg_rhs(Porg0, Porg_rhs, Sfx0, Sfy0, Sfz0, lev);
+    for (int ithBH = 0; ithBH < BH_num; ithBH++)
+    {
+      f_rungekutta4_scalar(dT_lev, Porg0[ithBH][0], Porg[ithBH][0], Porg_rhs[ithBH][0], iter_count);
+      f_rungekutta4_scalar(dT_lev, Porg0[ithBH][1], Porg[ithBH][1], Porg_rhs[ithBH][1], iter_count);
+      f_rungekutta4_scalar(dT_lev, Porg0[ithBH][2], Porg[ithBH][2], Porg_rhs[ithBH][2], iter_count);
+      if (Symmetry > 0)
+        Porg[ithBH][2] = fabs(Porg[ithBH][2]);
+      if (Symmetry == 2)
+      {
+        Porg[ithBH][0] = fabs(Porg[ithBH][0]);
+        Porg[ithBH][1] = fabs(Porg[ithBH][1]);
+      }
+    }
+  }
+
+  if ((lev == a_lev) && (LastAnas + dT_lev >= AnasTime))
+    z4c_cuda_download_level_state(GH->PatL[lev], SynchList_pre, myrank, false);
+  if (lev == a_lev)
+    AnalysisStuff(lev, dT_lev);
+
+  for (iter_count = 1; iter_count < 4; iter_count++)
+  {
+    if (iter_count == 1 || iter_count == 3)
+      TRK4 += dT_lev / 2;
+    Pp = GH->PatL[lev];
+    while (Pp)
+    {
+      MyList<Block> *BP = Pp->data->blb;
+      while (BP)
+      {
+        Block *cg = BP->data;
+        if (myrank == cg->rank)
+        {
+          double *state_in[Z4C_CUDA_STATE_COUNT];
+          double *state_out[Z4C_CUDA_STATE_COUNT];
+          double propspeed[Z4C_CUDA_STATE_COUNT];
+          double soa_flat[3 * Z4C_CUDA_STATE_COUNT];
+          if (!fill_z4c_cuda_views(cg, SynchList_pre, state_in, propspeed, soa_flat) ||
+              !fill_z4c_cuda_views(cg, SynchList_cor, state_out))
+          {
+            cout << "CUDA Z4C state list mismatch on corrector step" << endl;
+            MPI_Abort(MPI_COMM_WORLD, 1);
+          }
+          int apply_bam_bc = 0;
+#if (MRBD == 0)
+#if (SommerType == 0)
+          apply_bam_bc = (lev == 0) ? 1 : 0;
+#endif
+#elif (MRBD == 1)
+          apply_bam_bc = 1;
+#endif
+          int keep_resident_state = z4c_cuda_resident_step_enabled() ? 1 : 0;
+          int apply_enforce_ga = 0;
+#if (AGM == 0)
+          apply_enforce_ga = 1;
+#elif (AGM == 1)
+          apply_enforce_ga = (iter_count == 3) ? 1 : 0;
+#endif
+          if (z4c_cuda_rk4_substep(cg,
+                                   cg->shape, cg->X[0], cg->X[1], cg->X[2],
+                                   state_in, state_out,
+                                   propspeed, soa_flat, Pp->data->bbox,
+                                   dT_lev, TRK4, iter_count, apply_bam_bc,
+                                   Symmetry, lev, ndeps, cor,
+                                   keep_resident_state, apply_enforce_ga, chitiny))
+          {
+            cout << "CUDA Z4C corrector substep failed in domain: ("
+                 << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                 << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                 << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+            ERROR = 1;
+          }
+        }
+        if (BP == Pp->data->ble)
+          break;
+        BP = BP->next;
+      }
+      Pp = Pp->next;
+    }
+
+    {
+      int erh = ERROR;
+      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+    }
+    if (ERROR)
+    {
+      if (myrank == 0 && ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "CUDA Z4C failed in RK4 substep#" << iter_count
+                              << " at t = " << PhysTime
+                              << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+
+    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev]);
+
+    if (BH_num > 0 && lev == GH->levels - 1)
+    {
+      if (z4c_cuda_resident_step_enabled())
+      {
+        if (!z4c_cuda_compute_porg_rhs_resident(GH, lev, myrank, BH_num,
+                                                Porg, Porg1,
+                                                Sfx, Sfy, Sfz, Symmetry))
+        {
+          if (myrank == 0 && ErrorMonitor->outfile)
+            ErrorMonitor->outfile << "CUDA Z4C failed to interpolate black-hole shift at t = "
+                                  << PhysTime << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+      }
+      else
+      {
+        compute_Porg_rhs(Porg, Porg1, Sfx, Sfy, Sfz, lev);
+      }
+      for (int ithBH = 0; ithBH < BH_num; ithBH++)
+      {
+        f_rungekutta4_scalar(dT_lev, Porg0[ithBH][0], Porg1[ithBH][0], Porg_rhs[ithBH][0], iter_count);
+        f_rungekutta4_scalar(dT_lev, Porg0[ithBH][1], Porg1[ithBH][1], Porg_rhs[ithBH][1], iter_count);
+        f_rungekutta4_scalar(dT_lev, Porg0[ithBH][2], Porg1[ithBH][2], Porg_rhs[ithBH][2], iter_count);
+        if (Symmetry > 0)
+          Porg1[ithBH][2] = fabs(Porg1[ithBH][2]);
+        if (Symmetry == 2)
+        {
+          Porg1[ithBH][0] = fabs(Porg1[ithBH][0]);
+          Porg1[ithBH][1] = fabs(Porg1[ithBH][1]);
+        }
+      }
+    }
+
+    if (iter_count < 3)
+    {
+      Pp = GH->PatL[lev];
+      while (Pp)
+      {
+        MyList<Block> *BP = Pp->data->blb;
+        while (BP)
+        {
+          Block *cg = BP->data;
+          cg->swapList(SynchList_pre, SynchList_cor, myrank);
+          if (BP == Pp->data->ble)
+            break;
+          BP = BP->next;
+        }
+        Pp = Pp->next;
+      }
+      if (BH_num > 0 && lev == GH->levels - 1)
+      {
+        for (int ithBH = 0; ithBH < BH_num; ithBH++)
+        {
+          Porg[ithBH][0] = Porg1[ithBH][0];
+          Porg[ithBH][1] = Porg1[ithBH][1];
+          Porg[ithBH][2] = Porg1[ithBH][2];
+        }
+      }
+    }
+  }
+
+  z4c_cuda_download_level_state(GH->PatL[lev], SynchList_cor, myrank, false);
+
+#if (RPS == 0)
+  RestrictProlong(lev, YN, BB);
+#endif
+
+  Pp = GH->PatL[lev];
+  while (Pp)
+  {
+    MyList<Block> *BP = Pp->data->blb;
+    while (BP)
+    {
+      Block *cg = BP->data;
+      cg->swapList(StateList, SynchList_cor, myrank);
+      cg->swapList(OldStateList, SynchList_cor, myrank);
+      if (BP == Pp->data->ble)
+        break;
+      BP = BP->next;
+    }
+    Pp = Pp->next;
+  }
+  if (BH_num > 0 && lev == GH->levels - 1)
+  {
+    for (int ithBH = 0; ithBH < BH_num; ithBH++)
+    {
+      Porg0[ithBH][0] = Porg1[ithBH][0];
+      Porg0[ithBH][1] = Porg1[ithBH][1];
+      Porg0[ithBH][2] = Porg1[ithBH][2];
+    }
+  }
+#else
   double dT_lev = dT * pow(0.5, Mymax(lev, trfls));
 #ifdef With_AHF
   AH_Step_Find(lev, dT_lev);
@@ -339,6 +1001,13 @@ void Z4c_class::Step(int lev, int YN)
   }
 
 #ifdef WithShell
+#if USE_CUDA_Z4C
+  if (bssn_cuda_use_resident_sync(lev))
+  {
+    for (int dl = 0; dl < GH->levels; dl++)
+      bssn_cuda_download_level_state_if_present(GH->PatL[dl], StateList, myrank);
+  }
+#endif
   // evolve Shell Patches
   if (lev == 0)
   {
@@ -1042,9 +1711,11 @@ void Z4c_class::Step(int lev, int YN)
       Porg0[ithBH][2] = Porg1[ithBH][2];
     }
   }
+#endif
 }
 #else
 // for constraint preserving boundary (CPBC)
+// Note: CPBC path uses CPU Fortran RHS; GPU resident sync is a no-op here.
 #ifndef WithShell
 #error "CPBC only supports Shell"
 #endif
@@ -1074,6 +1745,14 @@ void Z4c_class::Step(int lev, int YN)
   int pre = 0, cor = 1;
   int ERROR = 0;
 
+#if USE_CUDA_Z4C && defined(WithShell)
+  if (bssn_cuda_use_resident_sync(lev))
+  {
+    for (int dl = 0; dl < GH->levels; dl++)
+      bssn_cuda_download_level_state_if_present(GH->PatL[dl], StateList, myrank);
+  }
+#endif
+
   MyList<ss_patch> *sPp;
   // Predictor
   MyList<Patch> *Pp = GH->PatL[lev];
@@ -2404,6 +3083,11 @@ void Z4c_class::Check_extrop()
 
 //================================================================================================
 
+#if USE_CUDA_Z4C && (ABEtype == 2) && defined(WithShell)
+#undef f_compute_rhs_Z4c_ss
+#define f_compute_rhs_Z4c_ss compute_rhs_z4c_ss_
+#endif
+
 // this member function is used to compute and output constraint violation
 
 //================================================================================================
@@ -2679,11 +3363,12 @@ void Z4c_class::Interp_Constraint()
   }
 
   ofstream outfile;
-  char filename[50];
-  sprintf(filename, "%s/interp_constraint_%05d.dat", ErrorMonitor->out_dir.c_str(), int(PhysTime / dT + 0.5)); 
+  char suffix[64];
+  sprintf(suffix, "/interp_constraint_%05d.dat", int(PhysTime / dT + 0.5));
+  string filename = ErrorMonitor->out_dir + suffix;
   // 0.5 for round off
   
-  outfile.open(filename);
+  outfile.open(filename.c_str());
   outfile << "#  corrdinate, H_Res, Px_Res, Py_Res, Pz_Res, Gx_Res, Gy_Res, Gz_Res, ...." << endl;
   for (int i = 0; i < n; i++)
   {

AMSS_NCKU_source/bssnEScalar_class.h

@@ -51,7 +51,7 @@ public:
      void Compute_Psi4(int lev);
      void Step(int lev, int YN);
      void AnalysisStuff_EScalar(int lev, double dT_lev);
-     void Interp_Constraint(bool infg);
+     void Interp_Constraint();
      void Constraint_Out(); 
 
 protected:

AMSS_NCKU_source/bssn_class.h

@@ -31,19 +31,11 @@ using namespace std;
 #include "surface_integral.h"
 #include "checkpoint.h"
 
-extern void setpbh(int iBHN, double **iPBH, double *iMass, int rBHN);
-
-#ifndef BSSN_USE_TRANSFER_CACHE
-#define BSSN_USE_TRANSFER_CACHE 1
-#endif
-
-#ifndef BSSN_USE_ESCALAR_C_KERNEL
-#define BSSN_USE_ESCALAR_C_KERNEL 1
-#endif
-
-class bssn_class
-{
-public:
+extern void setpbh(int iBHN, double **iPBH, double *iMass, int rBHN);
+
+class bssn_class
+{
+public:
        int ngfs;
        int nprocs, myrank;
        cgh *GH;
@@ -56,6 +48,7 @@ public:
        double StartTime, TotalTime;
        double AnasTime, DumpTime, d2DumpTime, CheckTime;
        double LastAnas, LastConsOut;
+       bool cuda_level0_constraint_cache_valid;
        int *ConstraintRefreshLevels;
        double Courant;
        double numepss, numepsb, numepsh;
@@ -151,7 +144,7 @@ public:
        bssn_class(double Couranti, double StartTimei, double TotalTimei, double DumpTimei, double d2DumpTimei, double CheckTimei, double AnasTimei,
                   int Symmetryi, int checkruni, char *checkfilenamei, double numepssi, double numepsbi, double numepshi,
                   int a_levi, int maxli, int decni, double maxrexi, double drexi);
-       ~bssn_class();
+       virtual ~bssn_class();
 
        void Evolve(int Steps);
        void RecursiveStep(int lev);
@@ -175,25 +168,14 @@ public:
        void Setup_KerrSchild();
        void Enforce_algcon(int lev, int fg);
 
-       void testRestrict();
-       void testOutBd();
-       
-       bool check_Stdin_Abort(); 
-       bool use_transfer_cache() const;
-       void setup_transfer_caches();
-       void invalidate_transfer_caches();
-       void destroy_transfer_caches();
-       void sync_predictor_start(int lev, MyList<var> *VarList, Parallel::AsyncSyncState &async_state);
-       void sync_predictor_finish(int lev, Parallel::AsyncSyncState &async_state, MyList<var> *VarList);
-       void sync_corrector_start(int lev, MyList<var> *VarList, Parallel::AsyncSyncState &async_state);
-       void sync_corrector_finish(int lev, Parallel::AsyncSyncState &async_state, MyList<var> *VarList);
-       void sync_evolution(int lev, MyList<var> *VarList, Parallel::SyncCache *cache_array = 0);
-       void restrict_evolution(int lev, MyList<var> *src_var_list, MyList<var> *dst_var_list);
-       void outbdlow2hi_evolution(int lev, MyList<var> *src_var_list, MyList<var> *dst_var_list);
-
-       virtual void Setup_Initial_Data_Cao();
-       virtual void Setup_Initial_Data_Lousto();
-       virtual void Initialize();
+       void testRestrict();
+       void testOutBd();
+       
+       bool check_Stdin_Abort(); 
+
+       virtual void Setup_Initial_Data_Cao();
+       virtual void Setup_Initial_Data_Lousto();
+       virtual void Initialize();
        virtual void Read_Ansorg();
        virtual void Read_Pablo() {};
        virtual void Compute_Psi4(int lev);

AMSS_NCKU_source/bssn_escalar_rhs_c.C

@@ -1,169 +0,0 @@
-#include "macrodef.h"
-#include "bssn_rhs.h"
-#include "share_func.h"
-#include "tool.h"
-#include <vector>
-
-namespace
-{
-    // Reuse the temporary workspace across block calls to avoid repeated heap churn
-    // in the EScalar wrapper. MPI ranks execute this path sequentially, so a single
-    // process-local buffer is sufficient here.
-    std::vector<double> g_escalar_tmp_store;
-}
-
-#ifdef fortran1
-#define f_frpotential frpotential
-#endif
-#ifdef fortran2
-#define f_frpotential FRPOTENTIAL
-#endif
-#ifdef fortran3
-#define f_frpotential frpotential_
-#endif
-
-extern "C"
-{
-    void f_frpotential(int *, double *, double *, double *);
-}
-
-int f_compute_rhs_bssn_escalar_c(int *ex, double &T,
-                                 double *X, double *Y, double *Z,
-                                 double *chi, double *trK,
-                                 double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz,
-                                 double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
-                                 double *Gamx, double *Gamy, double *Gamz,
-                                 double *Lap, double *betax, double *betay, double *betaz,
-                                 double *dtSfx, double *dtSfy, double *dtSfz,
-                                 double *Sphi, double *Spi,
-                                 double *chi_rhs, double *trK_rhs,
-                                 double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
-                                 double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
-                                 double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
-                                 double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
-                                 double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
-                                 double *Sphi_rhs, double *Spi_rhs,
-                                 double *rho, double *Sx, double *Sy, double *Sz,
-                                 double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
-                                 double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
-                                 double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
-                                 double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
-                                 double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
-                                 double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
-                                 double *Gmx_Res, double *Gmy_Res, double *Gmz_Res,
-                                 int &Symmetry, int &Lev, double &eps, int &co)
-{
-    const int nx = ex[0], ny = ex[1], nz = ex[2];
-    const int all = nx * ny * nz;
-
-    const size_t workspace_size = size_t(all) * 17;
-    if (g_escalar_tmp_store.size() < workspace_size)
-        g_escalar_tmp_store.resize(workspace_size);
-
-    double *tmp_ptr = g_escalar_tmp_store.data();
-    auto alloc_tmp = [&](int n = 1) -> double *
-    {
-        double *ptr = tmp_ptr;
-        tmp_ptr += size_t(all) * n;
-        return ptr;
-    };
-
-    double *chix = alloc_tmp(), *chiy = alloc_tmp(), *chiz = alloc_tmp();
-    double *Kx = alloc_tmp(), *Ky = alloc_tmp(), *Kz = alloc_tmp();
-    double *fxx = alloc_tmp(), *fxy = alloc_tmp(), *fxz = alloc_tmp();
-    double *fyy = alloc_tmp(), *fyz = alloc_tmp(), *fzz = alloc_tmp();
-    double *Lapx = alloc_tmp(), *Lapy = alloc_tmp(), *Lapz = alloc_tmp();
-    double *V = alloc_tmp(), *dVdSphi = alloc_tmp();
-
-    const double ZEO = 0.0, ONE = 1.0, TWO = 2.0, HALF = 0.5;
-    const double SSS[3] = {1.0, 1.0, 1.0};
-
-    fderivs(ex, chi, chix, chiy, chiz, X, Y, Z, 1.0, 1.0, 1.0, Symmetry, Lev);
-    fderivs(ex, Lap, Lapx, Lapy, Lapz, X, Y, Z, 1.0, 1.0, 1.0, Symmetry, Lev);
-    fderivs(ex, Sphi, Kx, Ky, Kz, X, Y, Z, 1.0, 1.0, 1.0, Symmetry, Lev);
-    fdderivs(ex, Sphi, fxx, fxy, fxz, fyy, fyz, fzz, X, Y, Z, 1.0, 1.0, 1.0, Symmetry, Lev);
-
-    f_frpotential(ex, Sphi, V, dVdSphi);
-
-    for (int i = 0; i < all; ++i)
-    {
-        const double alpn1 = Lap[i] + ONE;
-        const double chin1 = chi[i] + ONE;
-        const double gxx = dxx[i] + ONE;
-        const double gyy = dyy[i] + ONE;
-        const double gzz = dzz[i] + ONE;
-        const double det = gxx * gyy * gzz + gxy[i] * gyz[i] * gxz[i] + gxz[i] * gxy[i] * gyz[i]
-                         - gxz[i] * gyy * gxz[i] - gxy[i] * gxy[i] * gzz - gxx * gyz[i] * gyz[i];
-        const double gupxx = (gyy * gzz - gyz[i] * gyz[i]) / det;
-        const double gupxy = -(gxy[i] * gzz - gyz[i] * gxz[i]) / det;
-        const double gupxz = (gxy[i] * gyz[i] - gyy * gxz[i]) / det;
-        const double gupyy = (gxx * gzz - gxz[i] * gxz[i]) / det;
-        const double gupyz = -(gxx * gyz[i] - gxy[i] * gxz[i]) / det;
-        const double gupzz = (gxx * gyy - gxy[i] * gxy[i]) / det;
-
-        Sphi_rhs[i] = alpn1 * Spi[i];
-
-        Spi_rhs[i] = gupxx * fxx[i] + gupyy * fyy[i] + gupzz * fzz[i]
-                   + TWO * (gupxy * fxy[i] + gupxz * fxz[i] + gupyz * fyz[i])
-                   - ((Gamx[i] + (gupxx * chix[i] + gupxy * chiy[i] + gupxz * chiz[i]) / TWO / chin1) * Kx[i]
-                   +  (Gamy[i] + (gupxy * chix[i] + gupyy * chiy[i] + gupyz * chiz[i]) / TWO / chin1) * Ky[i]
-                   +  (Gamz[i] + (gupxz * chix[i] + gupyz * chiy[i] + gupzz * chiz[i]) / TWO / chin1) * Kz[i]);
-
-        Spi_rhs[i] = Spi_rhs[i] * alpn1
-                   + gupxx * Lapx[i] * Kx[i] + gupxy * Lapx[i] * Ky[i] + gupxz * Lapx[i] * Kz[i]
-                   + gupxy * Lapy[i] * Kx[i] + gupyy * Lapy[i] * Ky[i] + gupyz * Lapy[i] * Kz[i]
-                   + gupxz * Lapz[i] * Kx[i] + gupyz * Lapz[i] * Ky[i] + gupzz * Lapz[i] * Kz[i];
-
-        Spi_rhs[i] = Spi_rhs[i] * chin1 + alpn1 * (trK[i] * Spi[i] - dVdSphi[i]);
-
-        rho[i] = chin1 * ((gupxx * Kx[i] * Kx[i] + gupyy * Ky[i] * Ky[i] + gupzz * Kz[i] * Kz[i]) * HALF
-               + gupxy * Kx[i] * Ky[i] + gupxz * Kx[i] * Kz[i] + gupyz * Ky[i] * Kz[i])
-               + Spi[i] * Spi[i] * HALF + V[i];
-        Sx[i] = -Spi[i] * Kx[i];
-        Sy[i] = -Spi[i] * Ky[i];
-        Sz[i] = -Spi[i] * Kz[i];
-
-        const double pressure = (rho[i] - Spi[i] * Spi[i]) / chin1;
-        Sxx[i] = Kx[i] * Kx[i] - pressure * gxx;
-        Sxy[i] = Kx[i] * Ky[i] - pressure * gxy[i];
-        Sxz[i] = Kx[i] * Kz[i] - pressure * gxz[i];
-        Syy[i] = Ky[i] * Ky[i] - pressure * gyy;
-        Syz[i] = Ky[i] * Kz[i] - pressure * gyz[i];
-        Szz[i] = Kz[i] * Kz[i] - pressure * gzz;
-    }
-
-    if (f_compute_rhs_bssn(ex, T, X, Y, Z,
-                           chi, trK,
-                           dxx, gxy, gxz, dyy, gyz, dzz,
-                           Axx, Axy, Axz, Ayy, Ayz, Azz,
-                           Gamx, Gamy, Gamz,
-                           Lap, betax, betay, betaz,
-                           dtSfx, dtSfy, dtSfz,
-                           chi_rhs, trK_rhs,
-                           gxx_rhs, gxy_rhs, gxz_rhs, gyy_rhs, gyz_rhs, gzz_rhs,
-                           Axx_rhs, Axy_rhs, Axz_rhs, Ayy_rhs, Ayz_rhs, Azz_rhs,
-                           Gamx_rhs, Gamy_rhs, Gamz_rhs,
-                           Lap_rhs, betax_rhs, betay_rhs, betaz_rhs,
-                           dtSfx_rhs, dtSfy_rhs, dtSfz_rhs,
-                           rho, Sx, Sy, Sz,
-                           Sxx, Sxy, Sxz, Syy, Syz, Szz,
-                           Gamxxx, Gamxxy, Gamxxz, Gamxyy, Gamxyz, Gamxzz,
-                           Gamyxx, Gamyxy, Gamyxz, Gamyyy, Gamyyz, Gamyzz,
-                           Gamzxx, Gamzxy, Gamzxz, Gamzyy, Gamzyz, Gamzzz,
-                           Rxx, Rxy, Rxz, Ryy, Ryz, Rzz,
-                           ham_Res, movx_Res, movy_Res, movz_Res,
-                           Gmx_Res, Gmy_Res, Gmz_Res,
-                           Symmetry, Lev, eps, co))
-        return 1;
-
-    lopsided_kodis(ex, X, Y, Z, Sphi, Sphi_rhs, betax, betay, betaz, Symmetry, SSS, eps);
-    lopsided_kodis(ex, X, Y, Z, Spi, Spi_rhs, betax, betay, betaz, Symmetry, SSS, eps);
-
-    for (int i = 0; i < all; ++i)
-    {
-        if (Sphi_rhs[i] != Sphi_rhs[i] || Spi_rhs[i] != Spi_rhs[i] || rho[i] != rho[i])
-            return 1;
-    }
-
-    return 0;
-}

AMSS_NCKU_source/bssn_gpu.h

@@ -1,73 +1,56 @@
-
-#ifndef BSSN_GPU_H_
-#define BSSN_GPU_H_
-#include "bssn_macro.h"
-#include "macrodef.fh"
-
-#define DEVICE_ID 0
-// #define DEVICE_ID_BY_MPI_RANK
-#define GRID_DIM 256
-#define BLOCK_DIM 128
-
-#define _FH2_(i, j, k) fh[(i) + (j) * _1D_SIZE[2] + (k) * _2D_SIZE[2]]
-#define _FH3_(i, j, k) fh[(i) + (j) * _1D_SIZE[3] + (k) * _2D_SIZE[3]]
-#define pow2(x) ((x) * (x))
-#define TimeBetween(a, b) ((b.tv_sec - a.tv_sec) + (b.tv_usec - a.tv_usec) / 1000000.0f)
-#define M_ metac.
-#define Mh_ meta->
-#define Ms_ metassc.
-#define Msh_ metass->
-
-// #define TIMING
-
-#define RHS_SS_PARA int calledby, int mpi_rank, int *ex, double &T, double *crho, double *sigma, double *R, double *X, double *Y, double *Z, double *drhodx, double *drhody, double *drhodz, double *dsigmadx, double *dsigmady, double *dsigmadz, double *dRdx, double *dRdy, double *dRdz, double *drhodxx, double *drhodxy, double *drhodxz, double *drhodyy, double *drhodyz, double *drhodzz, double *dsigmadxx, double *dsigmadxy, double *dsigmadxz, double *dsigmadyy, double *dsigmadyz, double *dsigmadzz, double *dRdxx, double *dRdxy, double *dRdxz, double *dRdyy, double *dRdyz, double *dRdzz, double *chi, double *trK, double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz, double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz, double *Gamx, double *Gamy, double *Gamz, double *Lap, double *betax, double *betay, double *betaz, double *dtSfx, double *dtSfy, double *dtSfz, double *chi_rhs, double *trK_rhs, double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs, double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs, double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs, double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs, double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs, double *rho, double *Sx, double *Sy, double *Sz, double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz, double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz, double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz, double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz, double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz, double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res, double *Gmx_Res, double *Gmy_Res, double *Gmz_Res, int &Symmetry, int &Lev, double &eps, int &sst, int &co
-
-/**  main function */
-int gpu_rhs(int calledby, int mpi_rank, int *ex, double &T,
-            double *X, double *Y, double *Z,
-
-            double *chi, double *trK,
-
-            double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz,
-
-            double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
-
-            double *Gamx, double *Gamy, double *Gamz,
-
-            double *Lap, double *betax, double *betay, double *betaz,
-
-            double *dtSfx, double *dtSfy, double *dtSfz,
-
-            double *chi_rhs, double *trK_rhs,
-
-            double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
-
-            double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
-
-            double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
-
-            double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
-
-            double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
-
-            double *rho, double *Sx, double *Sy, double *Sz, double *Sxx,
-            double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
-
-            double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
-
-            double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
-
-            double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
-
-            double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
-
-            double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
-            double *Gmx_Res, double *Gmy_Res, double *Gmz_Res,
-            int &Symmetry, int &Lev, double &eps, int &co);
-
-int gpu_rhs_ss(RHS_SS_PARA);
-
-/** Init GPU side data in GPUMeta. */
-// void init_fluid_meta_gpu(GPUMeta *gpu_meta);
-
-#endif
+
+#ifndef BSSN_GPU_H_
+#define BSSN_GPU_H_
+#include "bssn_macro.h"
+#include "macrodef.h"
+
+#define DEVICE_ID 0
+// #define DEVICE_ID_BY_MPI_RANK
+#define GRID_DIM 256
+#define BLOCK_DIM 128
+
+#define _FH2_(i, j, k) fh[(i) + (j) * _1D_SIZE[2] + (k) * _2D_SIZE[2]]
+#define _FH3_(i, j, k) fh[(i) + (j) * _1D_SIZE[3] + (k) * _2D_SIZE[3]]
+#define pow2(x) ((x) * (x))
+#define TimeBetween(a, b) ((b.tv_sec - a.tv_sec) + (b.tv_usec - a.tv_usec) / 1000000.0f)
+#define M_ metac.
+#define Mh_ meta->
+#define Ms_ metassc.
+#define Msh_ metass->
+
+// #define TIMING
+
+#define RHS_SS_PARA int calledby, int mpi_rank, int *ex, double &T, double *crho, double *sigma, double *R, double *X, double *Y, double *Z, double *drhodx, double *drhody, double *drhodz, double *dsigmadx, double *dsigmady, double *dsigmadz, double *dRdx, double *dRdy, double *dRdz, double *drhodxx, double *drhodxy, double *drhodxz, double *drhodyy, double *drhodyz, double *drhodzz, double *dsigmadxx, double *dsigmadxy, double *dsigmadxz, double *dsigmadyy, double *dsigmadyz, double *dsigmadzz, double *dRdxx, double *dRdxy, double *dRdxz, double *dRdyy, double *dRdyz, double *dRdzz, double *chi, double *trK, double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz, double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz, double *Gamx, double *Gamy, double *Gamz, double *Lap, double *betax, double *betay, double *betaz, double *dtSfx, double *dtSfy, double *dtSfz, double *chi_rhs, double *trK_rhs, double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs, double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs, double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs, double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs, double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs, double *rho, double *Sx, double *Sy, double *Sz, double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz, double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz, double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz, double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz, double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz, double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res, double *Gmx_Res, double *Gmy_Res, double *Gmz_Res, int &Symmetry, int &Lev, double &eps, int &sst, int &co
+
+/**  main function */
+int gpu_rhs(int calledby, int mpi_rank, int *ex, double &T,
+            double *X, double *Y, double *Z,
+            double *chi, double *trK,
+            double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz,
+            double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
+            double *Gamx, double *Gamy, double *Gamz,
+            double *Lap, double *betax, double *betay, double *betaz,
+            double *dtSfx, double *dtSfy, double *dtSfz,
+            double *chi_rhs, double *trK_rhs,
+            double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
+            double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
+            double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
+            double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
+            double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
+            double *rho, double *Sx, double *Sy, double *Sz, double *Sxx,
+            double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
+            double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
+            double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
+            double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
+            double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
+            double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
+            double *Gmx_Res, double *Gmy_Res, double *Gmz_Res,
+            int &Symmetry, int &Lev, double &eps, int &co);
+
+int gpu_rhs_ss(RHS_SS_PARA);
+
+#define Z4C_SS_PARA int calledby, int mpi_rank, int *ex, double &T, double *crho, double *sigma, double *R, double *X, double *Y, double *Z, double *drhodx, double *drhody, double *drhodz, double *dsigmadx, double *dsigmady, double *dsigmadz, double *dRdx, double *dRdy, double *dRdz, double *drhodxx, double *drhodxy, double *drhodxz, double *drhodyy, double *drhodyz, double *drhodzz, double *dsigmadxx, double *dsigmadxy, double *dsigmadxz, double *dsigmadyy, double *dsigmadyz, double *dsigmadzz, double *dRdxx, double *dRdxy, double *dRdxz, double *dRdyy, double *dRdyz, double *dRdzz, double *chi, double *trK, double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz, double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz, double *Gamx, double *Gamy, double *Gamz, double *Lap, double *betax, double *betay, double *betaz, double *dtSfx, double *dtSfy, double *dtSfz, double *TZ, double *chi_rhs, double *trK_rhs, double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs, double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs, double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs, double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs, double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs, double *TZ_rhs, double *rho, double *Sx, double *Sy, double *Sz, double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz, double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz, double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz, double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz, double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz, double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res, double *Gmx_Res, double *Gmy_Res, double *Gmz_Res, int &Symmetry, int &Lev, double &eps, int &sst, int &co
+
+int gpu_rhs_z4c_ss(Z4C_SS_PARA);
+
+#endif

AMSS_NCKU_source/bssn_gpu_class.h

@@ -1,210 +0,0 @@
-
-#ifndef BSSN_GPU_CLASS_H
-#define BSSN_GPU_CLASS_H
-
-#ifdef newc
-#include <iostream>
-#include <iomanip>
-#include <fstream>
-#include <cstdlib>
-#include <string>
-#include <cmath>
-using namespace std;
-#else
-#include <iostream.h>
-#include <iomanip.h>
-#include <fstream.h>
-#include <stdlib.h>
-#include <string.h>
-#include <math.h>
-#endif
-
-#include <mpi.h>
-
-#include "macrodef.h"
-#include "cgh.h"
-#include "ShellPatch.h"
-#include "misc.h"
-#include "var.h"
-#include "MyList.h"
-#include "monitor.h"
-#include "surface_integral.h"
-#include "checkpoint.h"
-
-// added by yangquan
-#include "bssn_macro.h"
-
-extern void setpbh(int iBHN, double **iPBH, double *iMass, int rBHN);
-
-class bssn_class
-{
-public:
-       // added by yangquan
-       //----------------------
-       int gpu_num_mynode;
-       int cpu_core_num_mynode;
-       int mpi_process_num_mynode;
-       int my_sequence_mynode;
-       int mynode_id;
-       int use_gpu;
-
-       virtual void Step_GPU(int lev, int YN);
-       virtual void Get_runtime_envirment();
-       // virtual void Step_OPENMP(int lev,int YN);
-       //----------------------
-
-       int ngfs;
-       int nprocs, myrank;
-       cgh *GH;
-       ShellPatch *SH;
-       double PhysTime;
-
-       int checkrun;
-       char checkfilename[50];
-       int Steps;
-       double StartTime, TotalTime;
-       double AnasTime, DumpTime, d2DumpTime, CheckTime;
-       double LastAnas, LastConsOut;
-       double Courant;
-       double numepss, numepsb, numepsh;
-       int Symmetry;
-       int maxl, decn;
-       double maxrex, drex;
-       int trfls, a_lev;
-
-       double dT;
-       double chitiny;
-
-       double **Porg0, **Porgbr, **Porg, **Porg1, **Porg_rhs;
-       int BH_num, BH_num_input;
-       double *Mass, *Pmom, *Spin;
-       double ADMMass;
-
-       var *phio, *trKo;
-       var *gxxo, *gxyo, *gxzo, *gyyo, *gyzo, *gzzo;
-       var *Axxo, *Axyo, *Axzo, *Ayyo, *Ayzo, *Azzo;
-       var *Gmxo, *Gmyo, *Gmzo;
-       var *Lapo, *Sfxo, *Sfyo, *Sfzo;
-       var *dtSfxo, *dtSfyo, *dtSfzo;
-
-       var *phi0, *trK0;
-       var *gxx0, *gxy0, *gxz0, *gyy0, *gyz0, *gzz0;
-       var *Axx0, *Axy0, *Axz0, *Ayy0, *Ayz0, *Azz0;
-       var *Gmx0, *Gmy0, *Gmz0;
-       var *Lap0, *Sfx0, *Sfy0, *Sfz0;
-       var *dtSfx0, *dtSfy0, *dtSfz0;
-
-       var *phi, *trK;
-       var *gxx, *gxy, *gxz, *gyy, *gyz, *gzz;
-       var *Axx, *Axy, *Axz, *Ayy, *Ayz, *Azz;
-       var *Gmx, *Gmy, *Gmz;
-       var *Lap, *Sfx, *Sfy, *Sfz;
-       var *dtSfx, *dtSfy, *dtSfz;
-
-       var *phi1, *trK1;
-       var *gxx1, *gxy1, *gxz1, *gyy1, *gyz1, *gzz1;
-       var *Axx1, *Axy1, *Axz1, *Ayy1, *Ayz1, *Azz1;
-       var *Gmx1, *Gmy1, *Gmz1;
-       var *Lap1, *Sfx1, *Sfy1, *Sfz1;
-       var *dtSfx1, *dtSfy1, *dtSfz1;
-
-       var *phi_rhs, *trK_rhs;
-       var *gxx_rhs, *gxy_rhs, *gxz_rhs, *gyy_rhs, *gyz_rhs, *gzz_rhs;
-       var *Axx_rhs, *Axy_rhs, *Axz_rhs, *Ayy_rhs, *Ayz_rhs, *Azz_rhs;
-       var *Gmx_rhs, *Gmy_rhs, *Gmz_rhs;
-       var *Lap_rhs, *Sfx_rhs, *Sfy_rhs, *Sfz_rhs;
-       var *dtSfx_rhs, *dtSfy_rhs, *dtSfz_rhs;
-
-       var *rho, *Sx, *Sy, *Sz, *Sxx, *Sxy, *Sxz, *Syy, *Syz, *Szz;
-
-       var *Gamxxx, *Gamxxy, *Gamxxz, *Gamxyy, *Gamxyz, *Gamxzz;
-       var *Gamyxx, *Gamyxy, *Gamyxz, *Gamyyy, *Gamyyz, *Gamyzz;
-       var *Gamzxx, *Gamzxy, *Gamzxz, *Gamzyy, *Gamzyz, *Gamzzz;
-
-       var *Rxx, *Rxy, *Rxz, *Ryy, *Ryz, *Rzz;
-
-       var *Rpsi4, *Ipsi4;
-       var *t1Rpsi4, *t1Ipsi4, *t2Rpsi4, *t2Ipsi4;
-
-       var *Cons_Ham, *Cons_Px, *Cons_Py, *Cons_Pz, *Cons_Gx, *Cons_Gy, *Cons_Gz;
-
-#ifdef Point_Psi4
-       var *phix, *phiy, *phiz;
-       var *trKx, *trKy, *trKz;
-       var *Axxx, *Axxy, *Axxz;
-       var *Axyx, *Axyy, *Axyz;
-       var *Axzx, *Axzy, *Axzz;
-       var *Ayyx, *Ayyy, *Ayyz;
-       var *Ayzx, *Ayzy, *Ayzz;
-       var *Azzx, *Azzy, *Azzz;
-#endif
-       // FIXME: uc = StateList, up = OldStateList, upp = SynchList_cor; so never touch these three data
-       MyList<var> *StateList, *SynchList_pre, *SynchList_cor, *RHSList;
-       MyList<var> *OldStateList, *DumpList;
-       MyList<var> *ConstraintList;
-
-       monitor *ErrorMonitor, *Psi4Monitor, *BHMonitor, *MAPMonitor;
-       monitor *ConVMonitor;
-       surface_integral *Waveshell;
-       checkpoint *CheckPoint;
-
-public:
-       bssn_class(double Couranti, double StartTimei, double TotalTimei, double DumpTimei, double d2DumpTimei, double CheckTimei, double AnasTimei,
-                  int Symmetryi, int checkruni, char *checkfilenamei, double numepssi, double numepsbi, double numepshi,
-                  int a_levi, int maxli, int decni, double maxrexi, double drexi);
-       ~bssn_class();
-
-       void Evolve(int Steps);
-       void RecursiveStep(int lev);
-#if (PSTR == 1)
-       void ParallelStep();
-       void SHStep();
-#endif
-       void RestrictProlong(int lev, int YN, bool BB, MyList<var> *SL, MyList<var> *OL, MyList<var> *corL);
-       void RestrictProlong_aux(int lev, int YN, bool BB, MyList<var> *SL, MyList<var> *OL, MyList<var> *corL);
-       void RestrictProlong(int lev, int YN, bool BB);
-       void ProlongRestrict(int lev, int YN, bool BB);
-       void Setup_Black_Hole_position();
-       void compute_Porg_rhs(double **BH_PS, double **BH_RHS, var *forx, var *fory, var *forz, int lev);
-       bool read_Pablo_file(int *ext, double *datain, char *filename);
-       void write_Pablo_file(int *ext, double xmin, double xmax, double ymin, double ymax, double zmin, double zmax,
-                             char *filename);
-       void AnalysisStuff(int lev, double dT_lev);
-       void Setup_KerrSchild();
-       void Enforce_algcon(int lev, int fg);
-
-       void testRestrict();
-       void testOutBd();
-
-       virtual void Setup_Initial_Data_Lousto();
-       virtual void Setup_Initial_Data_Cao();
-       virtual void Initialize();
-       virtual void Read_Ansorg();
-       virtual void Read_Pablo() {};
-       virtual void Compute_Psi4(int lev);
-       virtual void Step(int lev, int YN);
-       virtual void Interp_Constraint(bool infg);
-       virtual void Constraint_Out();
-       virtual void Compute_Constraint();
-
-#ifdef With_AHF
-protected:
-       MyList<var> *AHList, *AHDList, *GaugeList;
-       int AHfindevery;
-       double AHdumptime;
-       int *lastahdumpid, HN_num; // number of possible horizons
-       int *findeveryl;
-       double *xc, *yc, *zc, *xr, *yr, *zr;
-       bool *trigger;
-       double *dTT;
-       int *dumpid;
-
-public:
-       void AH_Prepare_derivatives();
-       bool AH_Interp_Points(MyList<var> *VarList,
-                             int NN, double **XX,
-                             double *Shellf, int Symmetryi);
-       void AH_Step_Find(int lev, double dT_lev);
-#endif
-};
-#endif /* BSSN_GPU_CLASS_H */

AMSS_NCKU_source/bssn_gpu_rhs_ss.cu

@@ -20,12 +20,14 @@ using namespace std;
 
 __device__ volatile unsigned int global_count = 0;
 
+#ifdef RESULT_CHECK
 void compare_result_gpu(int ftag1,double * datac,int data_num){
 	double * data = (double*)malloc(sizeof(double)*data_num);
 	cudaMemcpy(data, datac, data_num * sizeof(double), cudaMemcpyDeviceToHost);
 	compare_result(ftag1,data,data_num);
 	free(data);
 }
+#endif
 
 __global__ void sub_symmetry_bd_ss_partF(int ord, double * func, double *funcc)
 {
@@ -153,11 +155,11 @@ __global__ void sub_symmetry_bd_ss_partJ(int ord,double * func, double * funcc,d
 
 inline void sub_symmetry_bd_ss(int ord,double * func, double * funcc,double * SoA){
 	sub_symmetry_bd_ss_partF<<<GRID_DIM,BLOCK_DIM>>>(ord,func,funcc);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	sub_symmetry_bd_ss_partI<<<GRID_DIM,BLOCK_DIM>>>(ord,func,funcc,SoA[0]);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	sub_symmetry_bd_ss_partJ<<<GRID_DIM,BLOCK_DIM>>>(ord,func,funcc,SoA[1]);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 }
 
 __global__ void sub_fderivs_shc_part1(double *fx,double *fy,double *fz){
@@ -247,13 +249,13 @@ inline void sub_fderivs_shc(int& sst,double * f,double * fh,double *fx,double *f
 	//cudaMemset(Msh_ gy,0,h_3D_SIZE[0] * sizeof(double));
 	//cudaMemset(Msh_ gz,0,h_3D_SIZE[0] * sizeof(double));
 	sub_symmetry_bd_ss(2,f,fh,SoA1);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 			//compare_result_gpu(0,fh,h_3D_SIZE[2]);	
 	sub_fderivs_sh<<<GRID_DIM,BLOCK_DIM>>>(fh,Msh_ gx,Msh_ gy,Msh_ gz);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 	sub_fderivs_shc_part1<<<GRID_DIM,BLOCK_DIM>>>(fx,fy,fz);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 			//compare_result_gpu(1,fx,h_3D_SIZE[0]);
 			//compare_result_gpu(2,fy,h_3D_SIZE[0]);
 			//compare_result_gpu(3,fz,h_3D_SIZE[0]);
@@ -451,17 +453,17 @@ inline void sub_fdderivs_shc(int& sst,double * f,double * fh,
 	
 	//fderivs_sh
 	sub_symmetry_bd_ss(2,f,fh,SoA1);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 			//compare_result_gpu(1,fh,h_3D_SIZE[2]);	
 	sub_fderivs_sh<<<GRID_DIM,BLOCK_DIM>>>(fh,Msh_ gx,Msh_ gy,Msh_ gz);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 	//fdderivs_sh
 	sub_symmetry_bd_ss(2,f,fh,SoA1);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 			//compare_result_gpu(21,fh,h_3D_SIZE[2]);
 	sub_fdderivs_sh<<<GRID_DIM,BLOCK_DIM>>>(fh,Msh_ gxx,Msh_ gxy,Msh_ gxz,Msh_ gyy,Msh_ gyz,Msh_ gzz);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 			/*compare_result_gpu(11,Msh_ gx,h_3D_SIZE[0]);
 			compare_result_gpu(12,Msh_ gy,h_3D_SIZE[0]);
 			compare_result_gpu(13,Msh_ gz,h_3D_SIZE[0]);
@@ -472,7 +474,7 @@ inline void sub_fdderivs_shc(int& sst,double * f,double * fh,
 			compare_result_gpu(5,Msh_ gyz,h_3D_SIZE[0]);
 			compare_result_gpu(6,Msh_ gzz,h_3D_SIZE[0]);*/
 	sub_fdderivs_shc_part1<<<GRID_DIM,BLOCK_DIM>>>(fxx,fxy,fxz,fyy,fyz,fzz);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 			/*compare_result_gpu(1,fxx,h_3D_SIZE[0]);
 			compare_result_gpu(2,fxy,h_3D_SIZE[0]);
 			compare_result_gpu(3,fxz,h_3D_SIZE[0]);
@@ -496,9 +498,9 @@ __global__ void computeRicci_ss_part1(double * dst)
  inline void computeRicci_ss(int &sst,double * src,double* dst,double * SoA, Meta* meta)
 {
 	sub_fdderivs_shc(sst,src,Mh_ fh,Mh_ fxx,Mh_ fxy,Mh_ fxz,Mh_ fyy,Mh_ fyz,Mh_ fzz,SoA);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	computeRicci_ss_part1<<<GRID_DIM,BLOCK_DIM>>>(dst);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 }
 __global__ void sub_lopsided_ss_part1(double * dst)
@@ -516,9 +518,9 @@ __global__ void sub_lopsided_ss_part1(double * dst)
 inline void sub_lopsided_ss(int& sst,double *src,double* dst,double *SoA)
 {
 		sub_fderivs_shc(sst,src,Mh_ fh,Mh_ fxx,Mh_ fxy,Mh_ fxz,SoA);
-		cudaThreadSynchronize();
+		cudaDeviceSynchronize();
 		sub_lopsided_ss_part1<<<GRID_DIM,BLOCK_DIM>>>(dst);
-		cudaThreadSynchronize();
+		cudaDeviceSynchronize();
 }
 
 __global__ void sub_kodis_sh_part1(double *f,double *fh,double *f_rhs)
@@ -590,11 +592,11 @@ inline void sub_kodis_ss(int &sst,double *f,double *fh,double *f_rhs,double *SoA
 	}
 			//compare_result_gpu(10,f,h_3D_SIZE[0]);
 	sub_symmetry_bd_ss(3,f,fh,SoA1);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 			//compare_result_gpu(0,fh,h_3D_SIZE[3]);
 			
 	sub_kodis_sh_part1<<<GRID_DIM,BLOCK_DIM>>>(f,fh,f_rhs);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 			//compare_result_gpu(1,f_rhs,h_3D_SIZE[0]);
 }
 
@@ -1699,7 +1701,7 @@ void destroy_meta(Meta *meta,Metass *metass)
 	if(Msh_ gzz) cudaFree(Msh_ gzz);
 	
 #if (GAUGE == 2 || GAUGE == 3 || GAUGE == 4 || GAUGE == 5 || GAUGE == 6 || GAUGE == 7)
-	if(Mh_ reta) CUDA_SAFE_CALL(cudaFree(Mh_ reta));
+	if(Mh_ reta) cudaFree(Mh_ reta);
 
 #endif	
 	
@@ -1895,7 +1897,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
 
 	//1.2 local Data
 	cudaMalloc((void**)&(Mh_ gxx), matrix_size * sizeof(double));
-	CUDA_SAFE_CALL( cudaMalloc((void**)&(Mh_ gyy), matrix_size * sizeof(double)));
+	cudaMalloc((void**)&(Mh_ gyy), matrix_size * sizeof(double));
 	cudaMalloc((void**)&(Mh_ gzz), matrix_size * sizeof(double));
 	cudaMalloc((void**)&(Mh_ chix), matrix_size * sizeof(double));
 	cudaMalloc((void**)&(Mh_ chiy), matrix_size * sizeof(double));
@@ -2160,7 +2162,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
 	
 	double tmp_con2 = 1/Mass[0] - tmp_con;
 	cudaMemcpyToSymbol(C1, &tmp_con2, sizeof(double));
-	double tmp_con2 = 1/Mass[1] - tmp_con;
+	tmp_con2 = 1/Mass[1] - tmp_con;
 	cudaMemcpyToSymbol(C2, &tmp_con2, sizeof(double));
 		
 
@@ -2233,7 +2235,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
   		if((sst == 2 || sst == 4) && abs[1] < dYh) 
   		{
   			ijkmin_h[1] = -2; 
-  			ijkmin_h[1] = -3;
+  			ijkmin3_h[1] = -3;
   		}
   		if((sst == 3 || sst == 5) && abs_Y_ex2 < dYh)
   		{
@@ -2287,13 +2289,13 @@ int gpu_rhs_ss(RHS_SS_PARA)
 
 
 #ifdef TIMING1
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	gettimeofday(&tv2, NULL);
    	cout<<"TIME USED"<<TimeBetween(tv1, tv2)<<endl; 
 #endif	
 	//cout<<"GPU meta data ready.\n";
 	
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 
 
 //-------------get device info-------------------------------------
@@ -2306,7 +2308,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
 	//sub_enforce_ga(matrix_size);
 	//4.1-----compute rhs---------
 	compute_rhs_ss_part1<<<GRID_DIM,BLOCK_DIM>>>();
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 
 	sub_fderivs_shc(sst,Mh_ betax,Mh_ fh,Mh_ betaxx,Mh_ betaxy,Mh_ betaxz,ass);
 	sub_fderivs_shc(sst,Mh_ betay,Mh_ fh,Mh_ betayx,Mh_ betayy,Mh_ betayz,sas);
@@ -2322,7 +2324,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
 	sub_fderivs_shc(sst,Mh_ gyz,Mh_ fh,Mh_ gyzx,Mh_ gyzy,Mh_ gyzz, saa);
 	
 	compute_rhs_ss_part2<<<GRID_DIM,BLOCK_DIM>>>();
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 	sub_fdderivs_shc(sst,Mh_ betax,Mh_ fh,Mh_ gxxx,Mh_ gxyx,Mh_ gxzx,Mh_ gyyx,Mh_ gyzx,Mh_ gzzx,ass);
 	sub_fdderivs_shc(sst,Mh_ betay,Mh_ fh,Mh_ gxxy,Mh_ gxyy,Mh_ gxzy,Mh_ gyyy,Mh_ gyzy,Mh_ gzzy,sas);
@@ -2332,7 +2334,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
 	sub_fderivs_shc( sst,Mh_ Gamz, Mh_ fh,Mh_ Gamzx, Mh_ Gamzy, Mh_ Gamzz,ssa);
 	
 	compute_rhs_ss_part3<<<GRID_DIM,BLOCK_DIM>>>();
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 	computeRicci_ss(sst,Mh_ dxx,Mh_ Rxx,sss, meta);
 	computeRicci_ss(sst,Mh_ dyy,Mh_ Ryy,sss, meta);
@@ -2340,25 +2342,25 @@ int gpu_rhs_ss(RHS_SS_PARA)
 	computeRicci_ss(sst,Mh_ gxy,Mh_ Rxy,aas, meta);
 	computeRicci_ss(sst,Mh_ gxz,Mh_ Rxz,asa, meta);
 	computeRicci_ss(sst,Mh_ gyz,Mh_ Ryz,saa, meta);
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 	compute_rhs_ss_part4<<<GRID_DIM,BLOCK_DIM>>>();
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 	sub_fdderivs_shc(sst,Mh_ chi,Mh_ fh,Mh_ fxx,Mh_ fxy,Mh_ fxz,Mh_ fyy,Mh_ fyz,Mh_ fzz,sss);
 	
-	//cudaThreadSynchronize();
+	//cudaDeviceSynchronize();
 	//compare_result_gpu(0,Mh_ chi,h_3D_SIZE[0]);
 	//compare_result_gpu(1,Mh_ chi,h_3D_SIZE[0]);
 	//compare_result_gpu(2,Mh_ fyz,h_3D_SIZE[0]);
 	
 	compute_rhs_ss_part5<<<GRID_DIM,BLOCK_DIM>>>();
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 	sub_fdderivs_shc(sst,Mh_ Lap,Mh_ fh,Mh_ fxx,Mh_ fxy,Mh_ fxz,Mh_ fyy,Mh_ fyz,Mh_ fzz,sss);
 	
 	compute_rhs_ss_part6<<<GRID_DIM,BLOCK_DIM>>>();
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	
 #if (GAUGE == 2 || GAUGE == 3 || GAUGE == 4 || GAUGE == 5)
 	sub_fderivs_shc(sst,Mh_ chi,Mh_ fh, Mh_ dtSfx_rhs, Mh_ dtSfy_rhs, Mh_ dtSfz_rhs,sss);
@@ -2423,7 +2425,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
 	}
 	if(co == 0){
 		compute_rhs_ss_part7<<<GRID_DIM,BLOCK_DIM>>>();
-		cudaThreadSynchronize();
+		cudaDeviceSynchronize();
 
 		sub_fderivs_shc(sst,Mh_ Axx,Mh_ fh,Mh_ gxxx,Mh_ gxxy,Mh_ gxxz,sss);
 		sub_fderivs_shc(sst,Mh_ Axy,Mh_ fh,Mh_ gxyx,Mh_ gxyy,Mh_ gxyz,aas);
@@ -2432,7 +2434,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
 		sub_fderivs_shc(sst,Mh_ Ayz,Mh_ fh,Mh_ gyzx,Mh_ gyzy,Mh_ gyzz,saa);
 		sub_fderivs_shc(sst,Mh_ Azz,Mh_ fh,Mh_ gzzx,Mh_ gzzy,Mh_ gzzz,sss);
 		compute_rhs_ss_part8<<<GRID_DIM,BLOCK_DIM>>>();
-		cudaThreadSynchronize();
+		cudaDeviceSynchronize();
 	}
 	
 #if (ABV == 1)
@@ -2512,7 +2514,7 @@ int gpu_rhs_ss(RHS_SS_PARA)
 	//test kodis
 	//sub_kodis_sh(sst,Msh_ drhodx,Mh_ fh2,Msh_ drhody,sss);
 #ifdef TIMING
-	cudaThreadSynchronize();
+	cudaDeviceSynchronize();
 	gettimeofday(&tv2, NULL);
    	cout<<"MPI rank is: "<<mpi_rank<<" GPU TIME is"<<TimeBetween(tv1, tv2)<<" (s)."<<endl; 
 #endif
@@ -2522,4 +2524,55 @@ int gpu_rhs_ss(RHS_SS_PARA)
 	return 0;//TODO return
 }
 
+#if (ABEtype == 2)
+// Z4C Shell GPU: calls BSSN gpu_rhs_ss with trKd=trK+2*TZ, then applies
+// TZ_rhs = alpn1*Hcon/2 and constraint damping on CPU.
+int gpu_rhs_z4c_ss(Z4C_SS_PARA)
+{
+    int matrix_size = ex[0] * ex[1] * ex[2];
+    double k1 = 0.02, k2 = 0.0;
+
+    double *trKd_host = new double[matrix_size];
+    for (int _i = 0; _i < matrix_size; _i++)
+        trKd_host[_i] = trK[_i] + 2.0 * TZ[_i];
+
+    int result = gpu_rhs_ss(calledby, mpi_rank,
+        ex, T, crho, sigma, R, X, Y, Z,
+        drhodx, drhody, drhodz, dsigmadx, dsigmady, dsigmadz,
+        dRdx, dRdy, dRdz,
+        drhodxx, drhodxy, drhodxz, drhodyy, drhodyz, drhodzz,
+        dsigmadxx, dsigmadxy, dsigmadxz, dsigmadyy, dsigmadyz, dsigmadzz,
+        dRdxx, dRdxy, dRdxz, dRdyy, dRdyz, dRdzz,
+        chi, trKd_host, dxx, gxy, gxz, dyy, gyz, dzz,
+        Axx, Axy, Axz, Ayy, Ayz, Azz,
+        Gamx, Gamy, Gamz,
+        Lap, betax, betay, betaz,
+        dtSfx, dtSfy, dtSfz,
+        chi_rhs, trK_rhs,
+        gxx_rhs, gxy_rhs, gxz_rhs, gyy_rhs, gyz_rhs, gzz_rhs,
+        Axx_rhs, Axy_rhs, Axz_rhs, Ayy_rhs, Ayz_rhs, Azz_rhs,
+        Gamx_rhs, Gamy_rhs, Gamz_rhs,
+        Lap_rhs, betax_rhs, betay_rhs, betaz_rhs,
+        dtSfx_rhs, dtSfy_rhs, dtSfz_rhs,
+        rho, Sx, Sy, Sz, Sxx, Sxy, Sxz, Syy, Syz, Szz,
+        Gamxxx, Gamxxy, Gamxxz, Gamxyy, Gamxyz, Gamxzz,
+        Gamyxx, Gamyxy, Gamyxz, Gamyyy, Gamyyz, Gamyzz,
+        Gamzxx, Gamzxy, Gamzxz, Gamzyy, Gamzyz, Gamzzz,
+        Rxx, Rxy, Rxz, Ryy, Ryz, Rzz,
+        ham_Res, movx_Res, movy_Res, movz_Res,
+        Gmx_Res, Gmy_Res, Gmz_Res,
+        Symmetry, Lev, eps, sst, co);
+    delete[] trKd_host;
+    if (result != 0) return result;
+
+    for (int _i = 0; _i < matrix_size; _i++) {
+        double alp = Lap[_i] + 1.0;
+        TZ_rhs[_i] = alp * ham_Res[_i] * 0.5;
+        TZ_rhs[_i]  -= alp * (2.0 + k2) * k1 * TZ[_i];
+        trK_rhs[_i] += alp * k1 * (1.0 - k2) * TZ[_i];
+    }
+    return 0;
+}
+#endif // ABEtype == 2
+
 #endif //WithShell

AMSS_NCKU_source/bssn_rhs_c.C

@@ -1102,12 +1102,12 @@ int f_compute_rhs_bssn(int *ex, double &T,
             betaz_rhs[i] = FF * dtSfz[i];
 
             reta[i] =
-                gupxx[i] * dtSfx_rhs[i] * dtSfx_rhs[i]
-                + gupyy[i] * dtSfy_rhs[i] * dtSfy_rhs[i]
-                + gupzz[i] * dtSfz_rhs[i] * dtSfz_rhs[i]
-                + TWO * ( gupxy[i] * dtSfx_rhs[i] * dtSfy_rhs[i]
-                        + gupxz[i] * dtSfx_rhs[i] * dtSfz_rhs[i]
-                        + gupyz[i] * dtSfy_rhs[i] * dtSfz_rhs[i] );
+                gupxx[i] * chix[i] * chix[i]
+                + gupyy[i] * chiy[i] * chiy[i]
+                + gupzz[i] * chiz[i] * chiz[i]
+                + TWO * ( gupxy[i] * chix[i] * chiy[i]
+                        + gupxz[i] * chix[i] * chiz[i]
+                        + gupyz[i] * chiy[i] * chiz[i] );
 
             #if (GAUGE == 2)
             reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / pow( (ONE - sqrt(chin1[i])), 2.0 );
@@ -1120,12 +1120,12 @@ int f_compute_rhs_bssn(int *ex, double &T,
             dtSfz_rhs[i] = Gamz_rhs[i] - reta[i] * dtSfz[i];
             #elif (GAUGE == 4 || GAUGE == 5)
             reta[i] =
-                gupxx[i] * dtSfx_rhs[i] * dtSfx_rhs[i]
-                + gupyy[i] * dtSfy_rhs[i] * dtSfy_rhs[i]
-                + gupzz[i] * dtSfz_rhs[i] * dtSfz_rhs[i]
-                + TWO * ( gupxy[i] * dtSfx_rhs[i] * dtSfy_rhs[i]
-                        + gupxz[i] * dtSfx_rhs[i] * dtSfz_rhs[i]
-                        + gupyz[i] * dtSfy_rhs[i] * dtSfz_rhs[i] );
+                gupxx[i] * chix[i] * chix[i]
+                + gupyy[i] * chiy[i] * chiy[i]
+                + gupzz[i] * chiz[i] * chiz[i]
+                + TWO * ( gupxy[i] * chix[i] * chiy[i]
+                        + gupxz[i] * chix[i] * chiz[i]
+                        + gupyz[i] * chiy[i] * chiz[i] );
 
             #if (GAUGE == 4)
             reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / pow( (ONE - sqrt(chin1[i])), 2.0 );

AMSS_NCKU_source/bssn_rhs_cuda.h

@@ -0,0 +1,413 @@
+#ifndef BSSN_RHS_CUDA_H
+#define BSSN_RHS_CUDA_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+    BSSN_CUDA_STATE_COUNT = 24,
+    BSSN_ESCALAR_CUDA_STATE_COUNT = 26,
+    BSSN_EM_CUDA_STATE_COUNT = 32,
+    BSSN_EM_CUDA_SOURCE_COUNT = 4,
+    BSSN_CUDA_MATTER_COUNT = 10
+};
+
+int f_compute_rhs_bssn(int *ex, double &T,
+                       double *X, double *Y, double *Z,
+                       double *chi, double *trK,
+                       double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz,
+                       double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
+                       double *Gamx, double *Gamy, double *Gamz,
+                       double *Lap, double *betax, double *betay, double *betaz,
+                       double *dtSfx, double *dtSfy, double *dtSfz,
+                       double *chi_rhs, double *trK_rhs,
+                       double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
+                       double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
+                       double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
+                       double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
+                       double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
+                       double *rho, double *Sx, double *Sy, double *Sz,
+                       double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
+                       double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
+                       double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
+                       double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
+                       double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
+                       double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
+                       double *Gmx_Res, double *Gmy_Res, double *Gmz_Res,
+                       int &Symmetry, int &Lev, double &eps, int &co);
+
+int bssn_cuda_rk4_substep(void *block_tag,
+                          int *ex, double *X, double *Y, double *Z,
+                          double **state_host_in,
+                          double **state_host_out,
+                          double **matter_host,
+                          const double *propspeed,
+                          const double *soa_flat,
+                          const double *bbox,
+                          double &dT,
+                          double &T,
+                          int &RK4,
+                          int &apply_bam_bc,
+                          int &Symmetry,
+                          int &Lev,
+                          double &eps,
+                          int &co,
+                          int &use_zero_matter,
+                          int &keep_resident_state,
+                          int &apply_enforce_ga,
+                          double &chitiny);
+
+int bssn_escalar_cuda_rk4_substep(void *block_tag,
+                                  int *ex, double *X, double *Y, double *Z,
+                                  double **state_host_in,
+                                  double **state_host_out,
+                                  const double *propspeed,
+                                  const double *soa_flat,
+                                  const double *bbox,
+                                  double &dT,
+                                  double &T,
+                                  int &RK4,
+                                  int &apply_bam_bc,
+                                  int &Symmetry,
+	                                  int &Lev,
+	                                  double &eps,
+	                                  int &co,
+	                                  int &keep_resident_state,
+	                                  int &apply_enforce_ga,
+	                                  double &chitiny);
+
+int bssn_escalar_cuda_compute_constraints(int *ex, double *X, double *Y, double *Z,
+                                          double **state_host_in,
+                                          double **constraint_host_out,
+                                          int &Symmetry,
+                                          int &Lev,
+                                          double &eps);
+
+int bssn_em_cuda_rk4_substep(void *block_tag,
+                             int *ex, double *X, double *Y, double *Z,
+                             double **state_host_in,
+                             double **state_host_out,
+                             double **source_host,
+                             const double *propspeed,
+                             const double *soa_flat,
+                             const double *bbox,
+                             double &dT,
+                             double &T,
+                             int &RK4,
+                             int &apply_bam_bc,
+                             int &Symmetry,
+                             int &Lev,
+                             double &eps,
+                             int &co,
+                             int &keep_resident_state,
+                             int &apply_enforce_ga,
+                             double &chitiny);
+
+int bssn_em_cuda_resident_zero_fast_state(void *block_tag);
+
+int bssn_cuda_copy_state_region_to_host(void *block_tag,
+                                        int state_index,
+                                        double *host_state,
+                                        int *ex,
+                                        int i0, int j0, int k0,
+                                        int sx, int sy, int sz);
+
+int bssn_cuda_copy_state_region_from_host(void *block_tag,
+                                          int state_index,
+                                          double *host_state,
+                                          int *ex,
+                                          int i0, int j0, int k0,
+                                          int sx, int sy, int sz);
+
+int bssn_cuda_download_resident_state(void *block_tag,
+                                      int *ex,
+                                      double **state_host_out);
+
+int bssn_escalar_cuda_download_resident_state(void *block_tag,
+                                              int *ex,
+                                              double **state_host_out);
+
+int bssn_cuda_upload_resident_state_count(void *block_tag,
+                                          int *ex,
+                                          double **state_host_in,
+                                          int state_count);
+
+int bssn_escalar_cuda_upload_resident_state(void *block_tag,
+                                            int *ex,
+                                            double **state_host_in);
+
+int bssn_cuda_keep_only_resident_state_count(void *block_tag,
+                                             int *ex,
+                                             double **state_host_key,
+                                             int state_count);
+
+int bssn_escalar_cuda_keep_only_resident_state(void *block_tag,
+                                               int *ex,
+                                               double **state_host_key);
+
+int bssn_cuda_download_resident_state_count_if_present(void *block_tag,
+                                                       int *ex,
+                                                       double **state_host_out,
+                                                       int state_count);
+
+int bssn_cuda_download_resident_state_if_present(void *block_tag,
+                                                int *ex,
+                                                double **state_host_out);
+
+int bssn_cuda_download_constraint_outputs(int *ex,
+                                          double **constraint_host_out);
+
+int bssn_cuda_pack_state_region_to_host_buffer(void *block_tag,
+                                               int state_index,
+                                               double *host_buffer,
+                                               int *ex,
+                                               int i0, int j0, int k0,
+                                               int sx, int sy, int sz);
+
+int bssn_cuda_interp_state_point3(void *block_tag,
+                                  int *ex,
+                                  int state0,
+                                  int state1,
+                                  int state2,
+                                  double x0,
+                                  double y0,
+                                  double z0,
+                                  double dx,
+                                  double dy,
+                                  double dz,
+                                  double px,
+                                  double py,
+                                  double pz,
+                                  int ordn,
+                                  int symmetry,
+                                  double **state_host_key,
+                                  const double *soa3,
+                                  double *out3);
+
+int bssn_cuda_interp_host_two_fields(void *block_tag,
+                                     int *ex,
+                                     double *field0,
+                                     double *field1,
+                                     double x0,
+                                     double y0,
+                                     double z0,
+                                     double dx,
+                                     double dy,
+                                     double dz,
+                                     const double *px,
+                                     const double *py,
+                                     const double *pz,
+                                     int npoints,
+                                     int ordn,
+                                     int symmetry,
+                                     const double *soa6,
+                                     double *out_interleaved);
+
+int bssn_cuda_unpack_state_region_from_host_buffer(void *block_tag,
+                                                   int state_index,
+                                                   double *host_buffer,
+                                                   int *ex,
+                                                   int i0, int j0, int k0,
+                                                   int sx, int sy, int sz);
+
+int bssn_cuda_unpack_state_region_from_host_buffer_for_host_views(void *block_tag,
+                                                                  double **state_host_key,
+                                                                  int state_count,
+                                                                  int state_index,
+                                                                  double *host_buffer,
+                                                                  int *ex,
+                                                                  int i0, int j0, int k0,
+                                                                  int sx, int sy, int sz);
+
+int bssn_cuda_pack_state_batch_to_host_buffer(void *block_tag,
+                                              int state_count,
+                                              double *host_buffer,
+                                              int *ex,
+                                              int i0, int j0, int k0,
+                                              int sx, int sy, int sz);
+
+int bssn_cuda_pack_state_batch_to_host_buffer_for_host_views(void *block_tag,
+                                                             double **state_host_key,
+                                                             int state_count,
+                                                             double *host_buffer,
+                                                             int *ex,
+                                                             int i0, int j0, int k0,
+                                                             int sx, int sy, int sz);
+
+int bssn_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
+                                                  int state_count,
+                                                  double *host_buffer,
+                                                  int *ex,
+                                                  int i0, int j0, int k0,
+                                                  int sx, int sy, int sz);
+
+int bssn_cuda_unpack_state_batch_from_host_buffer_for_host_views(void *block_tag,
+                                                                 double **state_host_key,
+                                                                 int state_count,
+                                                                 double *host_buffer,
+                                                                 int *ex,
+                                                                 int i0, int j0, int k0,
+                                                                 int sx, int sy, int sz);
+
+int bssn_cuda_pack_state_batch_to_device_buffer(void *block_tag,
+                                                int state_count,
+                                                double *device_buffer,
+                                                int *ex,
+                                                int i0, int j0, int k0,
+                                                int sx, int sy, int sz);
+
+int bssn_cuda_pack_state_batch_to_device_buffer_for_host_views(void *block_tag,
+                                                               double **state_host_key,
+                                                               int state_count,
+                                                               double *device_buffer,
+                                                               int *ex,
+                                                               int i0, int j0, int k0,
+                                                               int sx, int sy, int sz);
+
+int bssn_cuda_unpack_state_batch_from_device_buffer(void *block_tag,
+                                                    int state_count,
+                                                    double *device_buffer,
+                                                    int *ex,
+                                                    int i0, int j0, int k0,
+                                                    int sx, int sy, int sz);
+
+int bssn_cuda_unpack_state_batch_from_device_buffer_for_host_views(void *block_tag,
+                                                                   double **state_host_key,
+                                                                   int state_count,
+                                                                   double *device_buffer,
+                                                                   int *ex,
+                                                                   int i0, int j0, int k0,
+                                                                   int sx, int sy, int sz);
+
+int bssn_cuda_pack_state_segments_to_device_buffer(void *block_tag,
+                                                   int state_count,
+                                                   double *device_buffer,
+                                                   int *ex,
+                                                   int segment_count,
+                                                   const int *segment_meta);
+
+int bssn_cuda_pack_state_segments_to_device_buffer_for_host_views(void *block_tag,
+                                                                  double **state_host_key,
+                                                                  int state_count,
+                                                                  double *device_buffer,
+                                                                  int *ex,
+                                                                  int segment_count,
+                                                                  const int *segment_meta);
+
+int bssn_cuda_unpack_state_segments_from_device_buffer(void *block_tag,
+                                                       int state_count,
+                                                       double *device_buffer,
+                                                       int *ex,
+                                                       int segment_count,
+                                                       const int *segment_meta);
+
+int bssn_cuda_unpack_state_segments_from_device_buffer_for_host_views(void *block_tag,
+                                                                      double **state_host_key,
+                                                                      int state_count,
+                                                                      double *device_buffer,
+                                                                      int *ex,
+                                                                      int segment_count,
+                                                                      const int *segment_meta);
+
+int bssn_cuda_restrict_state_segments_to_device_buffer(void *block_tag,
+                                                       int state_count,
+                                                       double *device_buffer,
+                                                       int *ex,
+                                                       int segment_count,
+                                                       const int *segment_meta);
+
+int bssn_cuda_restrict_state_segments_to_device_buffer_for_host_views(void *block_tag,
+                                                                      double **state_host_key,
+                                                                      int state_count,
+                                                                      double *device_buffer,
+                                                                      int *ex,
+                                                                      int segment_count,
+                                                                      const int *segment_meta,
+                                                                      const double *state_soa);
+
+int bssn_cuda_prolong_state_segments_to_device_buffer(void *block_tag,
+                                                      int state_count,
+                                                      double *device_buffer,
+                                                      int *ex,
+                                                      int segment_count,
+                                                      const int *segment_meta);
+
+int bssn_cuda_prolong_state_segments_to_device_buffer_for_host_views(void *block_tag,
+                                                                     double **state_host_key,
+                                                                     int state_count,
+                                                                     double *device_buffer,
+                                                                     int *ex,
+                                                                     int segment_count,
+                                                                     const int *segment_meta,
+                                                                     const double *state_soa);
+
+int bssn_cuda_restrict_state_batch_to_device_buffer(void *block_tag,
+                                                    int state_count,
+                                                    double *device_buffer,
+                                                    int *ex,
+                                                    int sx, int sy, int sz,
+                                                    int fi0, int fj0, int fk0);
+
+int bssn_cuda_restrict_state_batch_to_device_buffer_for_host_views(void *block_tag,
+                                                                   double **state_host_key,
+                                                                   int state_count,
+                                                                   double *device_buffer,
+                                                                   int *ex,
+                                                                   int sx, int sy, int sz,
+                                                                   int fi0, int fj0, int fk0,
+                                                                   const double *state_soa);
+
+int bssn_cuda_prolong_state_batch_to_device_buffer(void *block_tag,
+                                                   int state_count,
+                                                   double *device_buffer,
+                                                   int *ex,
+                                                   int sx, int sy, int sz,
+                                                   int ii0, int jj0, int kk0,
+                                                   int lbc_i, int lbc_j, int lbc_k);
+
+int bssn_cuda_prolong_state_batch_to_device_buffer_for_host_views(void *block_tag,
+                                                                  double **state_host_key,
+                                                                  int state_count,
+                                                                  double *device_buffer,
+                                                                  int *ex,
+                                                                  int sx, int sy, int sz,
+                                                                  int ii0, int jj0, int kk0,
+                                                                  int lbc_i, int lbc_j, int lbc_k,
+                                                                  const double *state_soa);
+
+int bssn_cuda_download_state_subset(void *block_tag,
+                                    int *ex,
+                                    int subset_count,
+                                    const int *state_indices,
+                                    double **state_host_out);
+
+int bssn_cuda_upload_state_subset(void *block_tag,
+                                  int *ex,
+                                  int subset_count,
+                                  const int *state_indices,
+                                  double **state_host_in);
+
+int bssn_cuda_prepare_inter_time_level(void *block_tag,
+                                       int *ex,
+                                       int state_count,
+                                       double **src1_host_key,
+                                       double **src2_host_key,
+                                       double **src3_host_key,
+                                       double **dst_host_key,
+                                       int source_count,
+                                       int tindex);
+
+int bssn_cuda_has_resident_state(void *block_tag);
+
+void bssn_cuda_release_step_ctx(void *block_tag);
+
+#ifdef __cplusplus
+}
+// C++-only helpers declared for derived equation classes (Z4C, etc.)
+// Defined in bssn_class.C. Requires MyList, Patch, var from including TU.
+bool bssn_cuda_use_resident_sync(int lev);
+void bssn_cuda_download_level_state_if_present(MyList<Patch> *PatL, MyList<var> *vars, int myrank);
+#endif
+
+#endif

AMSS_NCKU_source/checkpoint.C

@@ -76,8 +76,11 @@ checkpoint::checkpoint(bool checked, const char fname[], int myrank) : filename(
 
 	I_Print = (myrank == 0);
 
-	int i = strlen(fname);
-	filename = new char[i+30];
+	size_t filename_len = out_dir.size() + strlen(fname) + 32;
+#ifdef CHECKDETAIL
+	filename_len += 32;
+#endif
+	filename = new char[filename_len];
 	// cout << filename << endl;
 	// cout << i << endl;
 
@@ -100,12 +103,12 @@ checkpoint::checkpoint(bool checked, const char fname[], int myrank) : filename(
                 cout << " checkpoint class created " << endl;
         }
 }
-checkpoint::~checkpoint()
-{
-	CheckList->clearList();
-	if (I_Print)
-		delete[] filename;
-}
+checkpoint::~checkpoint()
+{
+	CheckList->clearList();
+	if (filename)
+		delete[] filename;
+}
 
 void checkpoint::addvariable(var *VV)
 {
@@ -136,7 +139,7 @@ void checkpoint::writecheck_cgh(double time, cgh *GH)
 	if (I_Print)
 	{
 		// char fname[50];
-		char fname[50+50]; 
+		char fname[4096]; 
 		sprintf(fname, "%s_cgh.CHK", filename);
 
 		outfile.open(fname, ios::out | ios::trunc);
@@ -195,7 +198,7 @@ void checkpoint::readcheck_cgh(double &time, cgh *GH, int myrank, int nprocs, in
 	int DIM = dim;
 	ifstream infile;
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_cgh.CHK", filename);
 
 	infile.open(fname);
@@ -297,7 +300,7 @@ void checkpoint::writecheck_sh(double time, ShellPatch *SH)
 
 	if (I_Print)
 	{
-		char fname[50];
+		char fname[4096];
 		sprintf(fname, "%s_sh.CHK", filename);
 
 		outfile.open(fname, ios::out | ios::trunc);
@@ -335,7 +338,7 @@ void checkpoint::readcheck_sh(ShellPatch *SH, int myrank)
 	int DIM = dim;
 	ifstream infile;
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_sh.CHK", filename);
 
 	infile.open(fname);
@@ -390,7 +393,7 @@ void checkpoint::write_Black_Hole_position(int BH_num_input, int BH_num, double
 
 	if (I_Print)
 	{
-		char fname[50];
+		char fname[4096];
 		sprintf(fname, "%s_BHp.CHK", filename);
 
 		outfile.open(fname, ios::out | ios::trunc);
@@ -417,7 +420,7 @@ void checkpoint::read_Black_Hole_position(int &BH_num_input, int &BH_num, double
 {
 	ifstream infile;
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_BHp.CHK", filename);
 
 	infile.open(fname);
@@ -461,7 +464,7 @@ void checkpoint::write_bssn(double LastDump, double Last2dDump, double LastAnas)
 
 	if (I_Print)
 	{
-		char fname[50];
+		char fname[4096];
 		sprintf(fname, "%s_bssn.CHK", filename);
 
 		outfile.open(fname, ios::out | ios::trunc);
@@ -481,7 +484,7 @@ void checkpoint::read_bssn(double &LastDump, double &Last2dDump, double &LastAna
 {
 	ifstream infile;
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_bssn.CHK", filename);
 
 	infile.open(fname);
@@ -506,7 +509,7 @@ void checkpoint::write_bssn(double LastDump, double Last2dDump, double LastAnas)
 	ofstream outfile;
 
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_bssn.CHK", filename);
 
 	outfile.open(fname, ios::out | ios::trunc);
@@ -527,7 +530,7 @@ void checkpoint::read_bssn(double &LastDump, double &Last2dDump, double &LastAna
 {
 	ifstream infile;
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_bssn.CHK", filename);
 
 	infile.open(fname);
@@ -551,7 +554,7 @@ void checkpoint::write_Black_Hole_position(int BH_num_input, int BH_num, double
 	ofstream outfile;
 
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_BHp.CHK", filename);
 
 	outfile.open(fname, ios::out | ios::trunc);
@@ -581,7 +584,7 @@ void checkpoint::read_Black_Hole_position(int &BH_num_input, int &BH_num, double
 {
 	ifstream infile;
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_BHp.CHK", filename);
 
 	infile.open(fname);
@@ -628,7 +631,7 @@ void checkpoint::writecheck_cgh(double time, cgh *GH)
 	ofstream outfile;
 
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_cgh.CHK", filename);
 
 	outfile.open(fname, ios::out | ios::trunc);
@@ -738,7 +741,7 @@ void checkpoint::readcheck_cgh(double &time, cgh *GH, int myrank, int nprocs, in
 	int DIM = dim;
 	ifstream infile;
 	// char fname[50];
-	char fname[50+50]; 
+	char fname[4096]; 
 	sprintf(fname, "%s_cgh.CHK", filename);
 
 	infile.open(fname);

AMSS_NCKU_source/fd_cuda_helpers.cuh

@@ -0,0 +1,412 @@
+#ifndef AMSS_NCKU_FD_CUDA_HELPERS_CUH
+#define AMSS_NCKU_FD_CUDA_HELPERS_CUH
+
+#ifndef ghost_width
+#error "ghost_width must be defined before including fd_cuda_helpers.cuh"
+#endif
+
+#if ghost_width < 2 || ghost_width > 5
+#error "CUDA finite-difference helpers support ghost_width 2..5"
+#endif
+
+#define AMSS_FD_CENTER_RADIUS (ghost_width - 1)
+#define AMSS_FD_LK_RADIUS     (ghost_width)
+
+__device__ __forceinline__ int fd_axis_radius(int qF, int qminF, int qmaxF)
+{
+#if AMSS_FD_CENTER_RADIUS >= 4
+    if (qF - 4 >= qminF && qF + 4 <= qmaxF) return 4;
+#endif
+#if AMSS_FD_CENTER_RADIUS >= 3
+    if (qF - 3 >= qminF && qF + 3 <= qmaxF) return 3;
+#endif
+#if AMSS_FD_CENTER_RADIUS >= 2
+    if (qF - 2 >= qminF && qF + 2 <= qmaxF) return 2;
+#endif
+    if (qF - 1 >= qminF && qF + 1 <= qmaxF) return 1;
+    return 0;
+}
+
+__device__ __forceinline__ int fd_common_radius(int iF, int jF, int kF,
+                                                int iminF, int jminF, int kminF,
+                                                int imaxF, int jmaxF, int kmaxF)
+{
+    int r = fd_axis_radius(iF, iminF, imaxF);
+    const int ry = fd_axis_radius(jF, jminF, jmaxF);
+    const int rz = fd_axis_radius(kF, kminF, kmaxF);
+    if (ry < r) r = ry;
+    if (rz < r) r = rz;
+    return r;
+}
+
+__device__ __forceinline__ double fd_first_coef(int r, int off)
+{
+    switch (r) {
+    case 1:
+        if (off == -1) return -1.0;
+        if (off ==  1) return  1.0;
+        return 0.0;
+    case 2:
+        if (off == -2) return  1.0;
+        if (off == -1) return -8.0;
+        if (off ==  1) return  8.0;
+        if (off ==  2) return -1.0;
+        return 0.0;
+    case 3:
+        if (off == -3) return  -1.0;
+        if (off == -2) return   9.0;
+        if (off == -1) return -45.0;
+        if (off ==  1) return  45.0;
+        if (off ==  2) return  -9.0;
+        if (off ==  3) return   1.0;
+        return 0.0;
+    case 4:
+        if (off == -4) return    3.0;
+        if (off == -3) return  -32.0;
+        if (off == -2) return  168.0;
+        if (off == -1) return -672.0;
+        if (off ==  1) return  672.0;
+        if (off ==  2) return -168.0;
+        if (off ==  3) return   32.0;
+        if (off ==  4) return   -3.0;
+        return 0.0;
+    default:
+        return 0.0;
+    }
+}
+
+__device__ __forceinline__ double fd_second_coef(int r, int off)
+{
+    switch (r) {
+    case 1:
+        if (off == -1) return  1.0;
+        if (off ==  0) return -2.0;
+        if (off ==  1) return  1.0;
+        return 0.0;
+    case 2:
+        if (off == -2) return  -1.0;
+        if (off == -1) return  16.0;
+        if (off ==  0) return -30.0;
+        if (off ==  1) return  16.0;
+        if (off ==  2) return  -1.0;
+        return 0.0;
+    case 3:
+        if (off == -3) return    2.0;
+        if (off == -2) return  -27.0;
+        if (off == -1) return  270.0;
+        if (off ==  0) return -490.0;
+        if (off ==  1) return  270.0;
+        if (off ==  2) return  -27.0;
+        if (off ==  3) return    2.0;
+        return 0.0;
+    case 4:
+        if (off == -4) return     -9.0;
+        if (off == -3) return    128.0;
+        if (off == -2) return  -1008.0;
+        if (off == -1) return   8064.0;
+        if (off ==  0) return -14350.0;
+        if (off ==  1) return   8064.0;
+        if (off ==  2) return  -1008.0;
+        if (off ==  3) return    128.0;
+        if (off ==  4) return     -9.0;
+        return 0.0;
+    default:
+        return 0.0;
+    }
+}
+
+__device__ __forceinline__ double fd_first_denom(int r)
+{
+    return (r == 4) ? 840.0 : ((r == 3) ? 60.0 : ((r == 2) ? 12.0 : 2.0));
+}
+
+__device__ __forceinline__ double fd_second_denom(int r)
+{
+    return (r == 4) ? 5040.0 : ((r == 3) ? 180.0 : ((r == 2) ? 12.0 : 1.0));
+}
+
+__device__ __forceinline__ double fd_fetch_axis(const double *src,
+                                                int iF, int jF, int kF,
+                                                int axis, int off,
+                                                int SoA0, int SoA1, int SoA2)
+{
+    if (axis == 0) iF += off;
+    else if (axis == 1) jF += off;
+    else kF += off;
+    return fetch_sym_ord2_direct(src, iF, jF, kF, SoA0, SoA1, SoA2);
+}
+
+__device__ __forceinline__ double fd_fetch_axis2(const double *src,
+                                                 int iF, int jF, int kF,
+                                                 int axis_a, int off_a,
+                                                 int axis_b, int off_b,
+                                                 int SoA0, int SoA1, int SoA2)
+{
+    if (axis_a == 0) iF += off_a;
+    else if (axis_a == 1) jF += off_a;
+    else kF += off_a;
+    if (axis_b == 0) iF += off_b;
+    else if (axis_b == 1) jF += off_b;
+    else kF += off_b;
+    return fetch_sym_ord2_direct(src, iF, jF, kF, SoA0, SoA1, SoA2);
+}
+
+__device__ __forceinline__ double fd_first_axis_radius(const double *src,
+                                                       int iF, int jF, int kF,
+                                                       int axis, int r, double h,
+                                                       int SoA0, int SoA1, int SoA2)
+{
+    if (r <= 0) return 0.0;
+    double s = 0.0;
+#pragma unroll
+    for (int off = -4; off <= 4; ++off) {
+        const double c = fd_first_coef(r, off);
+        if (c != 0.0) {
+            s += c * fd_fetch_axis(src, iF, jF, kF, axis, off, SoA0, SoA1, SoA2);
+        }
+    }
+    return s / (fd_first_denom(r) * h);
+}
+
+__device__ __forceinline__ double fd_second_axis_radius(const double *src,
+                                                        int iF, int jF, int kF,
+                                                        int axis, int r, double h,
+                                                        int SoA0, int SoA1, int SoA2)
+{
+    if (r <= 0) return 0.0;
+    double s = 0.0;
+#pragma unroll
+    for (int off = -4; off <= 4; ++off) {
+        const double c = fd_second_coef(r, off);
+        if (c != 0.0) {
+            s += c * fd_fetch_axis(src, iF, jF, kF, axis, off, SoA0, SoA1, SoA2);
+        }
+    }
+    return s / (fd_second_denom(r) * h * h);
+}
+
+__device__ __forceinline__ double fd_mixed_axis_radius(const double *src,
+                                                       int iF, int jF, int kF,
+                                                       int axis_a, int r_a, double h_a,
+                                                       int axis_b, int r_b, double h_b,
+                                                       int SoA0, int SoA1, int SoA2)
+{
+    if (r_a <= 0 || r_b <= 0) return 0.0;
+    double s = 0.0;
+#pragma unroll
+    for (int off_a = -4; off_a <= 4; ++off_a) {
+        const double ca = fd_first_coef(r_a, off_a);
+        if (ca == 0.0) continue;
+#pragma unroll
+        for (int off_b = -4; off_b <= 4; ++off_b) {
+            const double cb = fd_first_coef(r_b, off_b);
+            if (cb != 0.0) {
+                s += ca * cb * fd_fetch_axis2(src, iF, jF, kF, axis_a, off_a,
+                                              axis_b, off_b, SoA0, SoA1, SoA2);
+            }
+        }
+    }
+    return s / (fd_first_denom(r_a) * fd_first_denom(r_b) * h_a * h_b);
+}
+
+__device__ __forceinline__ void fd_compute_first3(const double *src,
+                                                  int iF, int jF, int kF,
+                                                  int iminF, int jminF, int kminF,
+                                                  int imaxF, int jmaxF, int kmaxF,
+                                                  int SoA0, int SoA1, int SoA2,
+                                                  double &fx, double &fy, double &fz)
+{
+#if ghost_width == 3
+    const int r = fd_common_radius(iF, jF, kF, iminF, jminF, kminF, imaxF, jmaxF, kmaxF);
+    fx = fd_first_axis_radius(src, iF, jF, kF, 0, r, d_gp.dX, SoA0, SoA1, SoA2);
+    fy = fd_first_axis_radius(src, iF, jF, kF, 1, r, d_gp.dY, SoA0, SoA1, SoA2);
+    fz = fd_first_axis_radius(src, iF, jF, kF, 2, r, d_gp.dZ, SoA0, SoA1, SoA2);
+#else
+    fx = fd_first_axis_radius(src, iF, jF, kF, 0, fd_axis_radius(iF, iminF, imaxF),
+                              d_gp.dX, SoA0, SoA1, SoA2);
+    fy = fd_first_axis_radius(src, iF, jF, kF, 1, fd_axis_radius(jF, jminF, jmaxF),
+                              d_gp.dY, SoA0, SoA1, SoA2);
+    fz = fd_first_axis_radius(src, iF, jF, kF, 2, fd_axis_radius(kF, kminF, kmaxF),
+                              d_gp.dZ, SoA0, SoA1, SoA2);
+#endif
+}
+
+__device__ __forceinline__ void fd_compute_second6(const double *src,
+                                                   int iF, int jF, int kF,
+                                                   int iminF, int jminF, int kminF,
+                                                   int imaxF, int jmaxF, int kmaxF,
+                                                   int SoA0, int SoA1, int SoA2,
+                                                   double &fxx, double &fxy, double &fxz,
+                                                   double &fyy, double &fyz, double &fzz)
+{
+#if ghost_width == 3
+    const int r = fd_common_radius(iF, jF, kF, iminF, jminF, kminF, imaxF, jmaxF, kmaxF);
+    const int rx = r, ry = r, rz = r;
+#else
+    const int rx = fd_axis_radius(iF, iminF, imaxF);
+    const int ry = fd_axis_radius(jF, jminF, jmaxF);
+    const int rz = fd_axis_radius(kF, kminF, kmaxF);
+#endif
+    fxx = fd_second_axis_radius(src, iF, jF, kF, 0, rx, d_gp.dX, SoA0, SoA1, SoA2);
+    fyy = fd_second_axis_radius(src, iF, jF, kF, 1, ry, d_gp.dY, SoA0, SoA1, SoA2);
+    fzz = fd_second_axis_radius(src, iF, jF, kF, 2, rz, d_gp.dZ, SoA0, SoA1, SoA2);
+    fxy = fd_mixed_axis_radius(src, iF, jF, kF, 0, rx, d_gp.dX, 1, ry, d_gp.dY, SoA0, SoA1, SoA2);
+    fxz = fd_mixed_axis_radius(src, iF, jF, kF, 0, rx, d_gp.dX, 2, rz, d_gp.dZ, SoA0, SoA1, SoA2);
+    fyz = fd_mixed_axis_radius(src, iF, jF, kF, 1, ry, d_gp.dY, 2, rz, d_gp.dZ, SoA0, SoA1, SoA2);
+}
+
+__device__ __forceinline__ bool fd_lop_fits(int qF, int qminF, int qmaxF,
+                                            int dir, int lo, int hi)
+{
+    for (int off = lo; off <= hi; ++off) {
+        const int q = qF + dir * off;
+        if (q < qminF || q > qmaxF) return false;
+    }
+    return true;
+}
+
+__device__ __forceinline__ double fd_lop_fetch_sum(const double *src,
+                                                   int iF, int jF, int kF,
+                                                   int axis, int dir,
+                                                   const double *coef,
+                                                   int lo, int hi,
+                                                   int SoA0, int SoA1, int SoA2)
+{
+    double s = 0.0;
+    for (int off = lo; off <= hi; ++off) {
+        const double c = coef[off - lo];
+        if (c != 0.0) {
+            s += c * fd_fetch_axis(src, iF, jF, kF, axis, dir * off, SoA0, SoA1, SoA2);
+        }
+    }
+    return s;
+}
+
+__device__ __forceinline__ double fd_lopsided_axis(const double *src,
+                                                   int iF, int jF, int kF,
+                                                   int axis, double speed,
+                                                   int qF, int qminF, int qmaxF,
+                                                   double h,
+                                                   int SoA0, int SoA1, int SoA2)
+{
+    if (speed == 0.0) return 0.0;
+    const int dir = (speed > 0.0) ? 1 : -1;
+    const double mag = (speed > 0.0) ? speed : -speed;
+
+#if ghost_width == 2
+    if (fd_lop_fits(qF, qminF, qmaxF, dir, 0, 2)) {
+        const double c[] = {-3.0, 4.0, -1.0};
+        return mag * fd_lop_fetch_sum(src, iF, jF, kF, axis, dir, c, 0, 2, SoA0, SoA1, SoA2) / (2.0 * h);
+    }
+    if (fd_lop_fits(qF, qminF, qmaxF, dir, 0, 1)) {
+        const double c[] = {-1.0, 1.0};
+        return mag * fd_lop_fetch_sum(src, iF, jF, kF, axis, dir, c, 0, 1, SoA0, SoA1, SoA2) / (2.0 * h);
+    }
+    return 0.0;
+#elif ghost_width == 3
+    if (fd_lop_fits(qF, qminF, qmaxF, dir, -1, 3)) {
+        const double c[] = {-3.0, -10.0, 18.0, -6.0, 1.0};
+        return mag * fd_lop_fetch_sum(src, iF, jF, kF, axis, dir, c, -1, 3, SoA0, SoA1, SoA2) / (12.0 * h);
+    }
+    const int r = fd_axis_radius(qF, qminF, qmaxF);
+    return speed * fd_first_axis_radius(src, iF, jF, kF, axis, r, h, SoA0, SoA1, SoA2);
+#elif ghost_width == 4
+    if (fd_lop_fits(qF, qminF, qmaxF, dir, -2, 4)) {
+        const double c[] = {2.0, -24.0, -35.0, 80.0, -30.0, 8.0, -1.0};
+        return mag * fd_lop_fetch_sum(src, iF, jF, kF, axis, dir, c, -2, 4, SoA0, SoA1, SoA2) / (60.0 * h);
+    }
+    if (fd_lop_fits(qF, qminF, qmaxF, dir, -1, 5)) {
+        const double c[] = {-10.0, -77.0, 150.0, -100.0, 50.0, -15.0, 2.0};
+        return mag * fd_lop_fetch_sum(src, iF, jF, kF, axis, dir, c, -1, 5, SoA0, SoA1, SoA2) / (60.0 * h);
+    }
+    const int r = fd_axis_radius(qF, qminF, qmaxF);
+    return speed * fd_first_axis_radius(src, iF, jF, kF, axis, r, h, SoA0, SoA1, SoA2);
+#else
+    if (fd_lop_fits(qF, qminF, qmaxF, dir, -3, 5)) {
+        const double c[] = {-5.0, 60.0, -420.0, -378.0, 1050.0, -420.0, 140.0, -30.0, 3.0};
+        return mag * fd_lop_fetch_sum(src, iF, jF, kF, axis, dir, c, -3, 5, SoA0, SoA1, SoA2) / (840.0 * h);
+    }
+    const int r = fd_axis_radius(qF, qminF, qmaxF);
+    return speed * fd_first_axis_radius(src, iF, jF, kF, axis, r, h, SoA0, SoA1, SoA2);
+#endif
+}
+
+__device__ __forceinline__ double fd_ko_coef(int r, int off)
+{
+    const int a = off < 0 ? -off : off;
+    if (r == 2) {
+        if (a == 0) return 6.0;
+        if (a == 1) return -4.0;
+        if (a == 2) return 1.0;
+    } else if (r == 3) {
+        if (a == 0) return -20.0;
+        if (a == 1) return 15.0;
+        if (a == 2) return -6.0;
+        if (a == 3) return 1.0;
+    } else if (r == 4) {
+        if (a == 0) return 70.0;
+        if (a == 1) return -56.0;
+        if (a == 2) return 28.0;
+        if (a == 3) return -8.0;
+        if (a == 4) return 1.0;
+    } else if (r == 5) {
+        if (a == 0) return -252.0;
+        if (a == 1) return 210.0;
+        if (a == 2) return -120.0;
+        if (a == 3) return 45.0;
+        if (a == 4) return -10.0;
+        if (a == 5) return 1.0;
+    }
+    return 0.0;
+}
+
+__device__ __forceinline__ double fd_ko_axis(const double *src,
+                                             int iF, int jF, int kF,
+                                             int axis, int r,
+                                             int SoA0, int SoA1, int SoA2)
+{
+    double s = 0.0;
+#pragma unroll
+    for (int off = -5; off <= 5; ++off) {
+        if (off < -r || off > r) continue;
+        const double c = fd_ko_coef(r, off);
+        if (c != 0.0) {
+            s += c * fd_fetch_axis(src, iF, jF, kF, axis, off, SoA0, SoA1, SoA2);
+        }
+    }
+    return s;
+}
+
+__device__ __forceinline__ double fd_ko_term(const double *src,
+                                             int iF, int jF, int kF,
+                                             int iminF, int jminF, int kminF,
+                                             int imaxF, int jmaxF, int kmaxF,
+                                             double eps_val,
+                                             int SoA0, int SoA1, int SoA2)
+{
+    const int r = AMSS_FD_LK_RADIUS;
+    if (eps_val <= 0.0) return 0.0;
+#if ghost_width >= 4
+    if (iF - r <= iminF || iF + r >= imaxF ||
+        jF - r <= jminF || jF + r >= jmaxF ||
+        kF - r <= kminF || kF + r >= kmaxF) {
+        return 0.0;
+    }
+#else
+    if (iF - r < iminF || iF + r > imaxF ||
+        jF - r < jminF || jF + r > jmaxF ||
+        kF - r < kminF || kF + r > kmaxF) {
+        return 0.0;
+    }
+#endif
+    double cof = 1.0;
+#pragma unroll
+    for (int n = 0; n < 2 * r; ++n) cof *= 2.0;
+    const double sign = (r & 1) ? 1.0 : -1.0;
+    const double dx = fd_ko_axis(src, iF, jF, kF, 0, r, SoA0, SoA1, SoA2);
+    const double dy = fd_ko_axis(src, iF, jF, kF, 1, r, SoA0, SoA1, SoA2);
+    const double dz = fd_ko_axis(src, iF, jF, kF, 2, r, SoA0, SoA1, SoA2);
+    return sign * eps_val * (dx / d_gp.dX + dy / d_gp.dY + dz / d_gp.dZ) / cof;
+}
+
+#endif

AMSS_NCKU_source/gaussj.C

@@ -17,8 +17,8 @@ using namespace std;
 #include <math.h>
 #endif
 
-// Intel oneMKL LAPACK interface
-#include <mkl_lapacke.h>
+// LAPACKE interface (AOCL for AOCC, oneMKL for Intel)
+#include <lapacke.h>
 /* Linear equation solution using Intel oneMKL LAPACK.
 a[0..n-1][0..n-1] is the input matrix. b[0..n-1] is input
 containing the right-hand side vectors. On output a is

AMSS_NCKU_source/gpu_rhsSS_mem.h

@@ -1,6 +1,6 @@
 #ifndef GPU_MEM_H_
 #define GPU_MEM_H_
-#include "macrodef.fh"
+#include "macrodef.h"
 
 #ifdef WithShell
 struct Metass
@@ -48,6 +48,8 @@ struct Meta
 	double * Gamx_rhs,*Gamy_rhs,*Gamz_rhs;//out
 	double * Lap_rhs, *betax_rhs, *betay_rhs, *betaz_rhs;//out
 	double * dtSfx_rhs,*dtSfy_rhs,*dtSfz_rhs;//out
+	double * TZ;                                  //in  (Z4C)
+	double * TZ_rhs;                              //out (Z4C)
 	double * rho,*Sx,*Sy,*Sz			;	//in
 	double * Sxx,*Sxy,*Sxz,*Syy,*Syz,*Szz;	//in
 	
@@ -132,6 +134,8 @@ __constant__ double SYM = 1.0;
 __constant__ double ANTI = -1.0; 
 __constant__ double FF = 0.75;
 __constant__ double eta = 2.0;
+__constant__ double kappa1_c = 0.02;
+__constant__ double kappa2_c = 0.0;
 __constant__ double F1o3;
 __constant__ double F2o3;
 __constant__ double F3o2 = 1.5;

AMSS_NCKU_source/monitor.C

@@ -1,7 +1,8 @@
 
-#ifdef newc
-#include <cstdio>
-using namespace std;
+#ifdef newc
+#include <cstdio>
+#include <sstream>
+using namespace std;
 #else
 #include <stdio.h>
 #endif
@@ -77,16 +78,17 @@ monitor::monitor(const char fname[], int myrank, string head)
       parameters::str_par.insert(map<string, string>::value_type("output dir", out_dir));
     }
     // considering checkpoint run
-    char filename[50];
-    sprintf(filename, "%s/%s", out_dir.c_str(), fname);
-    int i = 1;
-    while ((access(filename, F_OK)) != -1)
-    {
-      sprintf(filename, "%s/%d_%s", out_dir.c_str(), i, fname);
-      i++;
-    }
-
-    outfile.open(filename, ios::trunc);
+    string filename = out_dir + "/" + fname;
+    int i = 1;
+    while ((access(filename.c_str(), F_OK)) != -1)
+    {
+      stringstream ss;
+      ss << out_dir << "/" << i << "_" << fname;
+      filename = ss.str();
+      i++;
+    }
+
+    outfile.open(filename.c_str(), ios::trunc);
 
     time_t tnow;
     time(&tnow);
@@ -107,16 +109,17 @@ monitor::monitor(const char fname[], int myrank, const int out_rank, string head
   if (I_Print)
   {
     // considering checkpoint run
-    char filename[50];
-    sprintf(filename, "%s/%s", out_dir.c_str(), fname);
-    int i = 1;
-    while ((access(filename, F_OK)) != -1)
-    {
-      sprintf(filename, "%s/%d_%s", out_dir.c_str(), i, fname);
-      i++;
-    }
-
-    outfile.open(filename, ios::trunc);
+    string filename = out_dir + "/" + fname;
+    int i = 1;
+    while ((access(filename.c_str(), F_OK)) != -1)
+    {
+      stringstream ss;
+      ss << out_dir << "/" << i << "_" << fname;
+      filename = ss.str();
+      i++;
+    }
+
+    outfile.open(filename.c_str(), ios::trunc);
 
     time_t tnow;
     time(&tnow);

AMSS_NCKU_source/surface_integral.C

@@ -8,10 +8,11 @@
 #include <iostream>
 #include <iomanip>
 #include <fstream>
-#include <strstream>
-#include <cmath>
-#include <map>
-using namespace std;
+#include <strstream>
+#include <cmath>
+#include <map>
+#include <cstdlib>
+using namespace std;
 #else
 #include <iostream.h>
 #include <iomanip.h>
@@ -29,12 +30,26 @@ using namespace std;
 #include "fadmquantites_bssn.h"
 #include "getnpem2.h"
 #include "getnp4.h"
-#include "parameters.h"
-
-#define PI M_PI
-//|============================================================================
-//| Constructor
-//|============================================================================
+#include "parameters.h"
+
+#define PI M_PI
+
+namespace
+{
+bool amss_surface_timing_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_SURFACE_TIMING");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+}
+//|============================================================================
+//| Constructor
+//|============================================================================
 
 surface_integral::surface_integral(int iSymmetry) : Symmetry(iSymmetry),
                                                     wave_cache_spinw(-1),
@@ -484,9 +499,9 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
   delete[] IP_out;
   DG_List->clearList();
 }
-void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *Ipsi4,
-                                 int spinw, int maxl, int NN, double *RP, double *IP,
-                                 monitor *Monitor, MPI_Comm Comm_here) // NN is the length of RP and IP
+void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *Ipsi4,
+                                 int spinw, int maxl, int NN, double *RP, double *IP,
+                                 monitor *Monitor, MPI_Comm Comm_here) // NN is the length of RP and IP
 {
   //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"start surface_integral::surf_Wave");
 
@@ -720,10 +735,10 @@ void surface_integral::surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *
   delete[] IP_out;
   DG_List->clearList();
 }
-//|----------------------------------------------------------------
-//  for shell patch
-//|----------------------------------------------------------------
-void surface_integral::surf_Wave(double rex, int lev, ShellPatch *GH, var *Rpsi4, var *Ipsi4,
+//|----------------------------------------------------------------
+//  for shell patch
+//|----------------------------------------------------------------
+void surface_integral::surf_Wave(double rex, int lev, ShellPatch *GH, var *Rpsi4, var *Ipsi4,
                                  int spinw, int maxl, int NN, double *RP, double *IP,
                                  monitor *Monitor) // NN is the length of RP and IP
 {
@@ -3281,6 +3296,8 @@ void surface_integral::surf_WaveMassPAng(double rex, int lev, cgh *GH,
                                          var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs,
                                          double *Rout, monitor *Monitor, bool refresh_mass_fields)
 {
+  const bool timing = amss_surface_timing_enabled();
+  const double t_start = timing ? MPI_Wtime() : 0.0;
   if (Symmetry != 0 && Symmetry != 1)
   {
     surf_Wave(rex, lev, GH, Rpsi4, Ipsi4, spinw, maxl, NN, RP, IP, Monitor);
@@ -3325,6 +3342,7 @@ void surface_integral::surf_WaveMassPAng(double rex, int lev, cgh *GH,
       Pp = Pp->next;
     }
   }
+  const double t_refresh_done = timing ? MPI_Wtime() : 0.0;
 
   const int InList = 19;
   const int idx_rpsi4 = 0, idx_ipsi4 = 1;
@@ -3380,6 +3398,7 @@ void surface_integral::surf_WaveMassPAng(double rex, int lev, cgh *GH,
 
   double *shellf = new double[n_tot * InList];
   GH->PatL[lev]->data->Interp_Points(DG_List, n_tot, pox, shellf, Symmetry, Nmin, Nmax);
+  const double t_interp_done = timing ? MPI_Wtime() : 0.0;
 
   double *RP_out = new double[NN];
   double *IP_out = new double[NN];
@@ -3496,6 +3515,7 @@ void surface_integral::surf_WaveMassPAng(double rex, int lev, cgh *GH,
     if (Symmetry == 0)
       p_outz += f1o8 * Psi * (nx_g[n] * axz + ny_g[n] * ayz + nz_g[n] * azz) * theta_weight;
   }
+  const double t_integral_done = timing ? MPI_Wtime() : 0.0;
 
   for (int ii = 0; ii < NN; ii++)
   {
@@ -3534,6 +3554,7 @@ void surface_integral::surf_WaveMassPAng(double rex, int lev, cgh *GH,
     delete[] reduce_out;
     delete[] reduce_in;
   }
+  const double t_reduce_done = timing ? MPI_Wtime() : 0.0;
 
 #ifdef GaussInt
   mass = mass * rex * rex * dphi * factor;
@@ -3565,6 +3586,19 @@ void surface_integral::surf_WaveMassPAng(double rex, int lev, cgh *GH,
   Rout[5] = sy;
   Rout[6] = sz;
 
+  if (timing)
+  {
+    fprintf(stderr,
+            "[AMSS-SURFACE][rank %d] rex=%.6g lev=%d refresh=%.6f interp=%.6f integral=%.6f reduce=%.6f total=%.6f nlocal=%d ntotal=%d modes=%d\n",
+            myrank, rex, lev,
+            t_refresh_done - t_start,
+            t_interp_done - t_refresh_done,
+            t_integral_done - t_interp_done,
+            t_reduce_done - t_integral_done,
+            t_reduce_done - t_start,
+            Nmax - Nmin + 1, n_tot, NN);
+  }
+
   delete[] pox[0];
   delete[] pox[1];
   delete[] pox[2];

AMSS_NCKU_source/surface_integral.h

@@ -46,10 +46,10 @@ public:
 	surface_integral(int iSymmetry);
 	~surface_integral();
 
-	void surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *Ipsi4,
-				   int spinw, int maxl, int NN, double *RP, double *IP,
-				   monitor *Monitor); // NN is the length of RP and IP
-									  // this routine can only deal with the symmetry of Psi4
+	void surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *Ipsi4,
+				   int spinw, int maxl, int NN, double *RP, double *IP,
+				   monitor *Monitor); // NN is the length of RP and IP
+									  // this routine can only deal with the symmetry of Psi4
 	void surf_Wave(double rex, int lev, ShellPatch *GH, var *Rpsi4, var *Ipsi4,
 				   int spinw, int maxl, int NN, double *RP, double *IP,
 				   monitor *Monitor);

AMSS_NCKU_source/z4c_rhs_c.C

@@ -327,9 +327,6 @@ static int compute_rhs_z4c_cartesian(
     double Axxx[all], Axxy[all], Axxz[all], Axyx[all], Axyy[all], Axyz[all];
     double Axzx[all], Axzy[all], Axzz[all], Ayyx[all], Ayyy[all], Ayyz[all];
     double Ayzx[all], Ayzy[all], Ayzz[all], Azzx[all], Azzy[all], Azzz[all];
-#if (GAUGE == 2 || GAUGE == 3 || GAUGE == 4 || GAUGE == 5)
-    double reta[all];
-#endif
 
     const double SSS[3] = {1.0, 1.0, 1.0};
     const double AAS[3] = {-1.0, -1.0, 1.0};
@@ -479,181 +476,8 @@ static int compute_rhs_z4c_cartesian(
         dtSfy_rhs[idx] = ZEO;
         dtSfz_rhs[idx] = ZEO;
     }
-#elif (GAUGE == 2)
-    /* Variable-eta gamma-driver, chi-sqrt denominator */
-    for (int idx = 0; idx < all; ++idx)
-    {
-        const double chin1i = chin1[idx];
-        const double det = gxx[idx] * gyy[idx] * gzz[idx]
-                         + gxy[idx] * gyz[idx] * gxz[idx] * 2.0
-                         - gxz[idx] * gyy[idx] * gxz[idx]
-                         - gxy[idx] * gxy[idx] * gzz[idx]
-                         - gxx[idx] * gyz[idx] * gyz[idx];
-        const double idet = ONE / det;
-        const double upxx = (gyy[idx] * gzz[idx] - gyz[idx] * gyz[idx]) * idet;
-        const double upxy = -(gxy[idx] * gzz[idx] - gyz[idx] * gxz[idx]) * idet;
-        const double upxz = (gxy[idx] * gyz[idx] - gyy[idx] * gxz[idx]) * idet;
-        const double upyy = (gxx[idx] * gzz[idx] - gxz[idx] * gxz[idx]) * idet;
-        const double upyz = -(gxx[idx] * gyz[idx] - gxy[idx] * gxz[idx]) * idet;
-        const double upzz = (gxx[idx] * gyy[idx] - gxy[idx] * gxy[idx]) * idet;
-        const double grdchi2 =
-            upxx * chix[idx] * chix[idx] + upyy * chiy[idx] * chiy[idx] + upzz * chiz[idx] * chiz[idx]
-          + TWO * (upxy * chix[idx] * chiy[idx] + upxz * chix[idx] * chiz[idx] + upyz * chiy[idx] * chiz[idx]);
-        const double sqchi = sqrt(chin1i);
-        reta[idx] = 1.31 / TWO * sqrt(grdchi2 / chin1i) / ((ONE - sqchi) * (ONE - sqchi));
-        betax_rhs[idx] = FF * dtSfx[idx];
-        betay_rhs[idx] = FF * dtSfy[idx];
-        betaz_rhs[idx] = FF * dtSfz[idx];
-        dtSfx_rhs[idx] = Gamx_rhs[idx] - reta[idx] * dtSfx[idx];
-        dtSfy_rhs[idx] = Gamy_rhs[idx] - reta[idx] * dtSfy[idx];
-        dtSfz_rhs[idx] = Gamz_rhs[idx] - reta[idx] * dtSfz[idx];
-    }
-#elif (GAUGE == 3)
-    /* Variable-eta gamma-driver, chi-linear denominator */
-    for (int idx = 0; idx < all; ++idx)
-    {
-        const double chin1i = chin1[idx];
-        const double det = gxx[idx] * gyy[idx] * gzz[idx]
-                         + gxy[idx] * gyz[idx] * gxz[idx] * 2.0
-                         - gxz[idx] * gyy[idx] * gxz[idx]
-                         - gxy[idx] * gxy[idx] * gzz[idx]
-                         - gxx[idx] * gyz[idx] * gyz[idx];
-        const double idet = ONE / det;
-        const double upxx = (gyy[idx] * gzz[idx] - gyz[idx] * gyz[idx]) * idet;
-        const double upxy = -(gxy[idx] * gzz[idx] - gyz[idx] * gxz[idx]) * idet;
-        const double upxz = (gxy[idx] * gyz[idx] - gyy[idx] * gxz[idx]) * idet;
-        const double upyy = (gxx[idx] * gzz[idx] - gxz[idx] * gxz[idx]) * idet;
-        const double upyz = -(gxx[idx] * gyz[idx] - gxy[idx] * gxz[idx]) * idet;
-        const double upzz = (gxx[idx] * gyy[idx] - gxy[idx] * gxy[idx]) * idet;
-        const double grdchi2 =
-            upxx * chix[idx] * chix[idx] + upyy * chiy[idx] * chiy[idx] + upzz * chiz[idx] * chiz[idx]
-          + TWO * (upxy * chix[idx] * chiy[idx] + upxz * chix[idx] * chiz[idx] + upyz * chiy[idx] * chiz[idx]);
-        reta[idx] = 1.31 / TWO * sqrt(grdchi2 / chin1i) / ((ONE - chin1i) * (ONE - chin1i));
-        betax_rhs[idx] = FF * dtSfx[idx];
-        betay_rhs[idx] = FF * dtSfy[idx];
-        betaz_rhs[idx] = FF * dtSfz[idx];
-        dtSfx_rhs[idx] = Gamx_rhs[idx] - reta[idx] * dtSfx[idx];
-        dtSfy_rhs[idx] = Gamy_rhs[idx] - reta[idx] * dtSfy[idx];
-        dtSfz_rhs[idx] = Gamz_rhs[idx] - reta[idx] * dtSfz[idx];
-    }
-#elif (GAUGE == 4)
-    /* Variable-eta gamma-driver, first-order, chi-sqrt denominator */
-    for (int idx = 0; idx < all; ++idx)
-    {
-        const double chin1i = chin1[idx];
-        const double det = gxx[idx] * gyy[idx] * gzz[idx]
-                         + gxy[idx] * gyz[idx] * gxz[idx] * 2.0
-                         - gxz[idx] * gyy[idx] * gxz[idx]
-                         - gxy[idx] * gxy[idx] * gzz[idx]
-                         - gxx[idx] * gyz[idx] * gyz[idx];
-        const double idet = ONE / det;
-        const double upxx = (gyy[idx] * gzz[idx] - gyz[idx] * gyz[idx]) * idet;
-        const double upxy = -(gxy[idx] * gzz[idx] - gyz[idx] * gxz[idx]) * idet;
-        const double upxz = (gxy[idx] * gyz[idx] - gyy[idx] * gxz[idx]) * idet;
-        const double upyy = (gxx[idx] * gzz[idx] - gxz[idx] * gxz[idx]) * idet;
-        const double upyz = -(gxx[idx] * gyz[idx] - gxy[idx] * gxz[idx]) * idet;
-        const double upzz = (gxx[idx] * gyy[idx] - gxy[idx] * gxy[idx]) * idet;
-        const double grdchi2 =
-            upxx * chix[idx] * chix[idx] + upyy * chiy[idx] * chiy[idx] + upzz * chiz[idx] * chiz[idx]
-          + TWO * (upxy * chix[idx] * chiy[idx] + upxz * chix[idx] * chiz[idx] + upyz * chiy[idx] * chiz[idx]);
-        const double sqchi = sqrt(chin1i);
-        reta[idx] = 1.31 / TWO * sqrt(grdchi2 / chin1i) / ((ONE - sqchi) * (ONE - sqchi));
-        betax_rhs[idx] = Gamx_rhs[idx] - reta[idx] * betax[idx];
-        betay_rhs[idx] = Gamy_rhs[idx] - reta[idx] * betay[idx];
-        betaz_rhs[idx] = Gamz_rhs[idx] - reta[idx] * betaz[idx];
-        dtSfx_rhs[idx] = ZEO;
-        dtSfy_rhs[idx] = ZEO;
-        dtSfz_rhs[idx] = ZEO;
-    }
-#elif (GAUGE == 5)
-    /* Variable-eta gamma-driver, first-order, chi-linear denominator */
-    for (int idx = 0; idx < all; ++idx)
-    {
-        const double chin1i = chin1[idx];
-        const double det = gxx[idx] * gyy[idx] * gzz[idx]
-                         + gxy[idx] * gyz[idx] * gxz[idx] * 2.0
-                         - gxz[idx] * gyy[idx] * gxz[idx]
-                         - gxy[idx] * gxy[idx] * gzz[idx]
-                         - gxx[idx] * gyz[idx] * gyz[idx];
-        const double idet = ONE / det;
-        const double upxx = (gyy[idx] * gzz[idx] - gyz[idx] * gyz[idx]) * idet;
-        const double upxy = -(gxy[idx] * gzz[idx] - gyz[idx] * gxz[idx]) * idet;
-        const double upxz = (gxy[idx] * gyz[idx] - gyy[idx] * gxz[idx]) * idet;
-        const double upyy = (gxx[idx] * gzz[idx] - gxz[idx] * gxz[idx]) * idet;
-        const double upyz = -(gxx[idx] * gyz[idx] - gxy[idx] * gxz[idx]) * idet;
-        const double upzz = (gxx[idx] * gyy[idx] - gxy[idx] * gxy[idx]) * idet;
-        const double grdchi2 =
-            upxx * chix[idx] * chix[idx] + upyy * chiy[idx] * chiy[idx] + upzz * chiz[idx] * chiz[idx]
-          + TWO * (upxy * chix[idx] * chiy[idx] + upxz * chix[idx] * chiz[idx] + upyz * chiy[idx] * chiz[idx]);
-        reta[idx] = 1.31 / TWO * sqrt(grdchi2 / chin1i) / ((ONE - chin1i) * (ONE - chin1i));
-        betax_rhs[idx] = Gamx_rhs[idx] - reta[idx] * betax[idx];
-        betay_rhs[idx] = Gamy_rhs[idx] - reta[idx] * betay[idx];
-        betaz_rhs[idx] = Gamz_rhs[idx] - reta[idx] * betaz[idx];
-        dtSfx_rhs[idx] = ZEO;
-        dtSfy_rhs[idx] = ZEO;
-        dtSfz_rhs[idx] = ZEO;
-    }
-#elif (GAUGE == 6 || GAUGE == 7)
-    {
-        /* Jason's position-dependent damping: rational (6) or exponential (7) */
-        int BHN = 0;
-        double Porg[9] = {0.0};
-        double Mass[3] = {0.0};
-    #ifdef fortran1
-        extern "C" { void getpbh(int &, double *, double *); }
-    #elif defined(fortran2)
-        extern "C" { void GETPBH(int &, double *, double *); }
-    #else
-        extern "C" { void getpbh_(int &, double *, double *); }
-    #endif
-        {
-    #ifdef fortran1
-            getpbh(BHN, Porg, Mass);
-    #elif defined(fortran2)
-            GETPBH(BHN, Porg, Mass);
-    #else
-            getpbh_(BHN, Porg, Mass);
-    #endif
-        }
-        if (BHN == 2)
-        {
-            const double M = Mass[0] + Mass[1];
-            const double A = 2.0 / M;
-            const double w1 = 12.0, w2 = 12.0;
-            const double C1 = 1.0 / Mass[0] - A;
-            const double C2 = 1.0 / Mass[1] - A;
-            const double BH_sep2 = (Porg[3] - Porg[0]) * (Porg[3] - Porg[0])
-                                 + (Porg[4] - Porg[1]) * (Porg[4] - Porg[1])
-                                 + (Porg[5] - Porg[2]) * (Porg[5] - Porg[2]);
-            const double inv_BH_sep2 = 1.0 / BH_sep2;
-            for (int k0 = 0; k0 < nz; ++k0) {
-            for (int j0 = 0; j0 < ny; ++j0) {
-            for (int i0 = 0; i0 < nx; ++i0) {
-                const size_t idx = idx_ex(i0, j0, k0, ex);
-                const double xp = X[i0], yp = Y[j0], zp = Z[k0];
-                const double r1 = ((Porg[0]-xp)*(Porg[0]-xp) + (Porg[1]-yp)*(Porg[1]-yp) + (Porg[2]-zp)*(Porg[2]-zp)) * inv_BH_sep2;
-                const double r2 = ((Porg[3]-xp)*(Porg[3]-xp) + (Porg[4]-yp)*(Porg[4]-yp) + (Porg[5]-zp)*(Porg[5]-zp)) * inv_BH_sep2;
-            #if (GAUGE == 6)
-                const double reta_val = A + C1 / (1.0 + w1 * r1) + C2 / (1.0 + w2 * r2);
-            #else
-                const double reta_val = A + C1 * exp(-w1 * r1) + C2 * exp(-w2 * r2);
-            #endif
-                betax_rhs[idx] = FF * dtSfx[idx];
-                betay_rhs[idx] = FF * dtSfy[idx];
-                betaz_rhs[idx] = FF * dtSfz[idx];
-                dtSfx_rhs[idx] = Gamx_rhs[idx] - reta_val * dtSfx[idx];
-                dtSfy_rhs[idx] = Gamy_rhs[idx] - reta_val * dtSfy[idx];
-                dtSfz_rhs[idx] = Gamz_rhs[idx] - reta_val * dtSfz[idx];
-            }}}
-        }
-        else
-        {
-            fprintf(stderr, "z4c_rhs_c: GAUGE %d requires BHN=2, got BHN=%d\n", (int)GAUGE, BHN);
-            return 1;
-        }
-    }
 #else
-#error "z4c_rhs_c.C: unsupported GAUGE value"
+#error "z4c_rhs_c.C currently supports GAUGE == 0 or GAUGE == 1 for Z4C"
 #endif
 
     lopsided(ex, X, Y, Z, gxx, gxx_rhs, betax, betay, betaz, Symmetry, SSS);
@@ -681,7 +505,7 @@ static int compute_rhs_z4c_cartesian(
     lopsided(ex, X, Y, Z, betax, betax_rhs, betax, betay, betaz, Symmetry, ASS);
     lopsided(ex, X, Y, Z, betay, betay_rhs, betax, betay, betaz, Symmetry, SAS);
     lopsided(ex, X, Y, Z, betaz, betaz_rhs, betax, betay, betaz, Symmetry, SSA);
-#if (GAUGE == 0 || GAUGE == 2 || GAUGE == 3 || GAUGE == 6 || GAUGE == 7)
+#if (GAUGE == 0)
     lopsided(ex, X, Y, Z, dtSfx, dtSfx_rhs, betax, betay, betaz, Symmetry, ASS);
     lopsided(ex, X, Y, Z, dtSfy, dtSfy_rhs, betax, betay, betaz, Symmetry, SAS);
     lopsided(ex, X, Y, Z, dtSfz, dtSfz_rhs, betax, betay, betaz, Symmetry, SSA);
@@ -728,7 +552,7 @@ static int compute_rhs_z4c_cartesian(
         kodis(ex, X, Y, Z, betax, betax_rhs, ASS, Symmetry, eps);
         kodis(ex, X, Y, Z, betay, betay_rhs, SAS, Symmetry, eps);
         kodis(ex, X, Y, Z, betaz, betaz_rhs, SSA, Symmetry, eps);
-#if (GAUGE == 0 || GAUGE == 2 || GAUGE == 3 || GAUGE == 6 || GAUGE == 7)
+#if (GAUGE == 0)
         kodis(ex, X, Y, Z, dtSfx, dtSfx_rhs, ASS, Symmetry, eps);
         kodis(ex, X, Y, Z, dtSfy, dtSfy_rhs, SAS, Symmetry, eps);
         kodis(ex, X, Y, Z, dtSfz, dtSfz_rhs, SSA, Symmetry, eps);

AMSS_NCKU_source/z4c_rhs_cuda.h

@@ -0,0 +1,114 @@
+#ifndef Z4C_RHS_CUDA_H
+#define Z4C_RHS_CUDA_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+    Z4C_CUDA_STATE_COUNT = 25
+};
+
+int z4c_cuda_rk4_substep(void *block_tag,
+                         int *ex, double *X, double *Y, double *Z,
+                         double **state_host_in,
+                         double **state_host_out,
+                         const double *propspeed,
+                         const double *soa_flat,
+                         const double *bbox,
+                         double &dT,
+                         double &T,
+                         int &RK4,
+                         int &apply_bam_bc,
+                         int &Symmetry,
+                         int &Lev,
+                         double &eps,
+                         int &co,
+                         int &keep_resident_state,
+                         int &apply_enforce_ga,
+                         double &chitiny);
+
+int z4c_cuda_download_resident_state(void *block_tag,
+                                     int *ex,
+                                     double **state_host_out);
+
+int z4c_cuda_pack_state_region_to_host_buffer(void *block_tag,
+                                              int state_index,
+                                              double *host_buffer,
+                                              int *ex,
+                                              int i0, int j0, int k0,
+                                              int sx, int sy, int sz);
+
+int z4c_cuda_unpack_state_region_from_host_buffer(void *block_tag,
+                                                  int state_index,
+                                                  double *host_buffer,
+                                                  int *ex,
+                                                  int i0, int j0, int k0,
+                                                  int sx, int sy, int sz);
+
+int z4c_cuda_pack_state_batch_to_host_buffer(void *block_tag,
+                                             int state_count,
+                                             double *host_buffer,
+                                             int *ex,
+                                             int i0, int j0, int k0,
+                                             int sx, int sy, int sz);
+
+int z4c_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
+                                                 int state_count,
+                                                 double *host_buffer,
+                                                 int *ex,
+                                                 int i0, int j0, int k0,
+                                                 int sx, int sy, int sz);
+
+int z4c_cuda_pack_state_batch_to_device_buffer(void *block_tag,
+                                               int state_count,
+                                               double *device_buffer,
+                                               int *ex,
+                                               int i0, int j0, int k0,
+                                               int sx, int sy, int sz);
+
+int z4c_cuda_unpack_state_batch_from_device_buffer(void *block_tag,
+                                                   int state_count,
+                                                   double *device_buffer,
+                                                   int *ex,
+                                                   int i0, int j0, int k0,
+                                                   int sx, int sy, int sz);
+
+int z4c_cuda_restrict_state_batch_to_device_buffer(void *block_tag,
+                                                   int state_count,
+                                                   double *device_buffer,
+                                                   int *ex,
+                                                   int sx, int sy, int sz,
+                                                   int fi0, int fj0, int fk0,
+                                                   const double *state_soa);
+
+int z4c_cuda_prolong_state_batch_to_device_buffer(void *block_tag,
+                                                  int state_count,
+                                                  double *device_buffer,
+                                                  int *ex,
+                                                  int sx, int sy, int sz,
+                                                  int ii0, int jj0, int kk0,
+                                                  int lbc_i, int lbc_j, int lbc_k,
+                                                  const double *state_soa);
+
+int z4c_cuda_download_state_subset(void *block_tag,
+                                   int *ex,
+                                   int subset_count,
+                                   const int *state_indices,
+                                   double **state_host_out);
+
+int z4c_cuda_upload_state_subset(void *block_tag,
+                                 int *ex,
+                                 int subset_count,
+                                 const int *state_indices,
+                                 double **state_host_in);
+
+int z4c_cuda_has_resident_state(void *block_tag);
+
+void z4c_cuda_release_step_ctx(void *block_tag);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

BSSN_BUILD_CONFIG_MIGRATION.md

@@ -1,211 +0,0 @@
-# BSSN Build Config Migration
-
-This note records the build-configuration fix needed when replacing
-`AMSS_NCKU_Input.py` or `generate_macrodef.py` with a newer upstream version.
-
-## Problem
-
-`AMSS_NCKU_source/macrodef.h` is not the authoritative file used by normal
-runs. `AMSS_NCKU_Program.py` first generates macro files under
-`input_data.File_directory`, copies `AMSS_NCKU_source` to
-`<File_directory>/AMSS_NCKU_source_copy`, then copies the generated macro files
-into that copied source tree and compiles there.
-
-Therefore, makefile logic must not depend only on the stale
-`AMSS_NCKU_source/macrodef.h`. The actual equation path must be passed to the
-copied build tree from the same generation step that creates `macrodef.h`.
-
-The performance regression was caused by compiling/linking the
-`BSSN-EScalar` C wrapper into BSSN vacuum builds. For BSSN vacuum (`ABEtype=0`),
-the build must use:
-
-```make
-BSSN_USE_TRANSFER_CACHE=1
-BSSN_USE_ESCALAR_C_KERNEL=0
-```
-
-and must not link `bssn_escalar_rhs_c.o`.
-
-## Required Migration Steps
-
-### 1. Add an ABE type helper in `generate_macrodef.py`
-
-Add a helper that maps `input_data.Equation_Class` to the numeric `ABEtype`.
-Use the same mapping as `macrodef.h`:
-
-```python
-def get_abe_type():
-    if ( input_data.Equation_Class == "BSSN" ):
-        return 0
-    elif ( input_data.Equation_Class == "BSSN-EScalar" ):
-        return 1
-    elif ( input_data.Equation_Class == "BSSN-EM" ):
-        return 3
-    elif ( input_data.Equation_Class == "Z4C" ):
-        return 2
-    else:
-        raise ValueError("Equation_Class setting error!!!")
-```
-
-Update `generate_macrodef_h()` to print `#define ABEtype {get_abe_type()}`
-instead of duplicating the if/elif mapping.
-
-### 2. Generate a makefile fragment
-
-In `generate_macrodef.py`, add:
-
-```python
-def generate_build_config():
-    file1 = open(os.path.join(input_data.File_directory, "AMSS_NCKU_build.mk"), "w")
-    print("# Generated by generate_macrodef.py; do not edit manually.", file=file1)
-    print(f"ABE_TYPE := {get_abe_type()}", file=file1)
-    file1.close()
-```
-
-This file is the build-time authority for the equation path.
-
-### 3. Call and copy the generated build config
-
-In `AMSS_NCKU_Program.py`, after generating `macrodef.h` and `macrodef.fh`, call:
-
-```python
-generate_macrodef.generate_build_config()
-print(" AMSS-NCKU build config AMSS_NCKU_build.mk has been generated. ")
-```
-
-When copying generated files into `AMSS_NCKU_source_copy`, also copy:
-
-```python
-build_config_path = os.path.join(File_directory, "AMSS_NCKU_build.mk")
-shutil.copy2(build_config_path, AMSS_NCKU_source_copy)
-```
-
-### 4. Make the source makefile consume the generated config
-
-At the top of `AMSS_NCKU_source/makefile`, after `include makefile.inc`, add:
-
-```make
--include AMSS_NCKU_build.mk
-
-ABE_TYPE ?= $(shell awk '/^[[:space:]]*\#define[[:space:]]+ABEtype/ {print $$3; exit}' macrodef.h 2>/dev/null)
-```
-
-The generated `AMSS_NCKU_build.mk` is used during normal Python-driven builds.
-The fallback keeps manual source-tree builds usable.
-
-### 5. Gate path-specific build options by `ABE_TYPE`
-
-Use effective build switches:
-
-```make
-ifeq ($(USE_TRANSFER_CACHE),auto)
-ifeq ($(ABE_TYPE),0)
-EFFECTIVE_USE_TRANSFER_CACHE = 1
-else
-EFFECTIVE_USE_TRANSFER_CACHE = 0
-endif
-else
-EFFECTIVE_USE_TRANSFER_CACHE = $(USE_TRANSFER_CACHE)
-endif
-
-ifeq ($(USE_CXX_ESCALAR_KERNEL),1)
-ifeq ($(ABE_TYPE),1)
-EFFECTIVE_USE_CXX_ESCALAR_KERNEL = 1
-else
-EFFECTIVE_USE_CXX_ESCALAR_KERNEL = 0
-endif
-else
-EFFECTIVE_USE_CXX_ESCALAR_KERNEL = 0
-endif
-
-TRANSFER_CACHE_FLAG = -DBSSN_USE_TRANSFER_CACHE=$(EFFECTIVE_USE_TRANSFER_CACHE)
-ESCALAR_KERNEL_FLAG = -DBSSN_USE_ESCALAR_C_KERNEL=$(EFFECTIVE_USE_CXX_ESCALAR_KERNEL)
-```
-
-Only add `bssn_escalar_rhs_c.o` when the effective EScalar C kernel switch is
-enabled:
-
-```make
-ifeq ($(EFFECTIVE_USE_CXX_ESCALAR_KERNEL),1)
-CFILES += bssn_escalar_rhs_c.o
-endif
-```
-
-### 6. Use safe transfer-cache default
-
-In `AMSS_NCKU_source/makefile.inc`, keep:
-
-```make
-USE_TRANSFER_CACHE ?= auto
-```
-
-With the effective switch logic above, this enables cached transfer for BSSN
-vacuum while keeping non-BSSN paths on the uncached path by default.
-
-## Verification Checklist
-
-Run these checks after migrating:
-
-```bash
-python3 -c "import generate_macrodef; generate_macrodef.generate_build_config()"
-cat GW150914/AMSS_NCKU_build.mk
-```
-
-For BSSN, the generated file should contain:
-
-```make
-ABE_TYPE := 0
-```
-
-Dry-run the copied or source makefile:
-
-```bash
-make -n -B INTERP_LB_MODE=off ABE | grep -E 'BSSN_USE_TRANSFER_CACHE|BSSN_USE_ESCALAR_C_KERNEL|bssn_escalar_rhs_c'
-```
-
-Expected BSSN result:
-
-```text
--DBSSN_USE_TRANSFER_CACHE=1 -DBSSN_USE_ESCALAR_C_KERNEL=0
-```
-
-and no `bssn_escalar_rhs_c.o` in the final link command.
-
-Run the full workflow:
-
-```bash
-python3 AMSS_NCKU_Program.py
-```
-
-For the 10-step BSSN test, compare coordinate output:
-
-```bash
-python3 - <<'PY'
-from pathlib import Path
-old = Path('../GW150914-06457/AMSS_NCKU_output/bssn_BH.dat')
-new = Path('GW150914/AMSS_NCKU_output/bssn_BH.dat')
-
-def rows(path):
-    out = []
-    for line in path.read_text().splitlines():
-        if not line.strip() or line.lstrip().startswith('#'):
-            continue
-        out.append([float(x) for x in line.split()])
-    return out
-
-ro, rn = rows(old), rows(new)
-n = min(len(ro), len(rn))
-max_abs = 0.0
-for i in range(n):
-    for a, b in zip(ro[i], rn[i]):
-        max_abs = max(max_abs, abs(a - b))
-print(f"old_rows={len(ro)} new_rows={len(rn)} compared_rows={n}")
-print(f"max_abs_diff={max_abs:.17g}")
-PY
-```
-
-For the validated migration, the first 10 rows matched exactly:
-
-```text
-max_abs_diff=0
-```

generate_macrodef.py

@@ -12,37 +12,6 @@ import os
 import AMSS_NCKU_Input as input_data          ## import program input file
 
 
-##################################################################
-
-def get_abe_type():
-    if ( input_data.Equation_Class == "BSSN" ):
-        return 0
-    elif ( input_data.Equation_Class == "BSSN-EScalar" ):
-        return 1
-    elif ( input_data.Equation_Class == "BSSN-EM" ):
-        return 3
-    elif ( input_data.Equation_Class == "Z4C" ):
-        return 2
-    else:
-        raise ValueError("Equation_Class setting error!!!")
-
-
-##################################################################
-
-## Generate the makefile fragment used by the copied source tree.
-## The source-tree macrodef.h is not authoritative because macro files
-## are regenerated under File_directory for each run.
-
-def generate_build_config():
-
-    file1 = open( os.path.join(input_data.File_directory, "AMSS_NCKU_build.mk"), "w")
-
-    print( "# Generated by generate_macrodef.py; do not edit manually.", file=file1 )
-    print( f"ABE_TYPE := {get_abe_type()}",                             file=file1 )
-
-    file1.close()
-
-
 ##################################################################
 
 ## Generate the macro file macrodef.h according to user settings
@@ -89,10 +58,19 @@ def generate_macrodef_h():
     # 2: Z4c vacuum
     # 3: coupled to Maxwell field
 
-    try:
-        print( f"#define ABEtype {get_abe_type()}", file=file1 )
-        print(                                      file=file1 )
-    except ValueError:
+    if ( input_data.Equation_Class == "BSSN" ):
+        print( "#define ABEtype 0", file=file1 )
+        print(                      file=file1 )
+    elif ( input_data.Equation_Class == "BSSN-EScalar" ):
+        print( "#define ABEtype 1", file=file1 )
+        print(                      file=file1 )
+    elif ( input_data.Equation_Class == "BSSN-EM" ):
+        print( "#define ABEtype 3", file=file1 )
+        print(                      file=file1 )
+    elif ( input_data.Equation_Class == "Z4C" ):
+        print( "#define ABEtype 2", file=file1 )
+        print(                      file=file1 )
+    else:
         print( "Equation_Class setting error!!!"                )
         print()
         print( "# Equation type #define ABEtype setting error!!!", file=file1 )
@@ -226,7 +204,7 @@ def generate_macrodef_h():
     # use GPU or not
     
     if ( input_data.GPU_Calculation == "yes"):
-        print( "#define USE_GPU",   file=file1 )
+        print( "//#define USE_GPU",   file=file1 )
         print(                      file=file1 )
     elif ( input_data.GPU_Calculation == "no"):
         print( "//#define USE_GPU", file=file1 )

makefile_and_run.py

@@ -9,6 +9,8 @@
 
 
 import AMSS_NCKU_Input as input_data
+import os
+import shutil
 import subprocess
 import time
 
@@ -56,6 +58,157 @@ BUILD_JOBS = 64
 
 ##################################################################
 
+def _truthy(value, default=False):
+    if value is None:
+        return default
+    if isinstance(value, bool):
+        return value
+    text = str(value).strip().lower()
+    if text == "":
+        return default
+    return text in ("1", "yes", "y", "true", "on", "enable", "enabled")
+
+
+def _input_or_env(input_name, env_name, default=None):
+    if env_name in os.environ:
+        return os.environ[env_name]
+    return getattr(input_data, input_name, default)
+
+
+def _start_cuda_mps_if_requested(runtime_env):
+    if input_data.GPU_Calculation != "yes":
+        return False
+
+    default_auto_mps = int(getattr(input_data, "MPI_processes", 1)) > 1
+    auto_mps = _truthy(
+        _input_or_env("CUDA_Auto_MPS", "AMSS_CUDA_AUTO_MPS", default_auto_mps),
+        default=default_auto_mps,
+    )
+    if not auto_mps:
+        return False
+
+    mps_control = shutil.which("nvidia-cuda-mps-control")
+    if not mps_control:
+        print(" CUDA MPS control command was not found; running without MPS.")
+        return False
+
+    uid = os.getuid()
+    pipe_dir = str(_input_or_env("CUDA_MPS_PIPE_DIRECTORY", "CUDA_MPS_PIPE_DIRECTORY",
+                                 f"/tmp/amss-ncku-mps-{uid}"))
+    log_dir = str(_input_or_env("CUDA_MPS_LOG_DIRECTORY", "CUDA_MPS_LOG_DIRECTORY",
+                                f"/tmp/amss-ncku-mps-log-{uid}"))
+    os.makedirs(pipe_dir, exist_ok=True)
+    os.makedirs(log_dir, exist_ok=True)
+
+    mps_env = runtime_env.copy()
+    mps_env["CUDA_MPS_PIPE_DIRECTORY"] = pipe_dir
+    mps_env["CUDA_MPS_LOG_DIRECTORY"] = log_dir
+
+    if os.path.exists(os.path.join(pipe_dir, "control")):
+        runtime_env.update({
+            "CUDA_MPS_PIPE_DIRECTORY": pipe_dir,
+            "CUDA_MPS_LOG_DIRECTORY": log_dir,
+        })
+        print(f" Reusing CUDA MPS daemon: {pipe_dir}")
+        return False
+
+    print(f" Starting CUDA MPS daemon for this run: {pipe_dir}")
+    result = subprocess.run([mps_control, "-d"], env=mps_env, text=True,
+                            stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
+    if result.returncode != 0:
+        print(" CUDA MPS daemon did not start; running without MPS.")
+        if result.stdout:
+            print(result.stdout, end="")
+        return False
+
+    runtime_env.update({
+        "CUDA_MPS_PIPE_DIRECTORY": pipe_dir,
+        "CUDA_MPS_LOG_DIRECTORY": log_dir,
+    })
+    return True
+
+
+def _stop_cuda_mps(runtime_env):
+    mps_control = shutil.which("nvidia-cuda-mps-control")
+    if not mps_control:
+        return
+    subprocess.run([mps_control], input="quit\n", env=runtime_env, text=True,
+                   stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
+
+
+def _gpu_runtime_env():
+    runtime_env = os.environ.copy()
+    finite_difference = str(getattr(input_data, "Finite_Diffenence_Method", "4th-order")).strip()
+
+    defaults = {
+        "AMSS_EVOLVE_TIMING": "1",
+        "AMSS_ESCALAR_STEP_TIMING": "0",
+        "AMSS_INTERP_FAST": "1",
+        "AMSS_INTERP_GPU": "1",
+        "AMSS_ANALYSIS_MAP_EVERY": "1000000",
+        "AMSS_CUDA_AWARE_MPI": "1",
+        "AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP": "1",
+        "AMSS_CUDA_KEEP_ALL_LEVELS": "1",
+        "AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP": "1",
+        "AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS": "1",
+        "AMSS_CUDA_EM_CACHE_SOURCES": "1",
+        "AMSS_CUDA_EM_ZERO_FASTPATH": "1",
+        "AMSS_EM_ZERO_ANALYSIS_FASTPATH": "1",
+        "AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH": "1",
+        "AMSS_CUDA_AMR_HOST_STAGED": "1",
+        "AMSS_CUDA_AMR_RESTRICT_DEVICE": "0",
+        "AMSS_CUDA_AMR_RESTRICT_BATCH": "0",
+        "AMSS_CUDA_DEVICE_SEGMENT_BATCH": "0",
+        "AMSS_CUDA_UNCACHED_DEVICE_BUFFERS": "1",
+        "AMSS_SHELL_FAST_INTERP": "0",
+        "AMSS_SHELL_PARALLEL_INTERP": "0",
+        "AMSS_SHELL_CUDA_INTERP": "0",
+    }
+    if finite_difference in ("2nd-order", "8th-order"):
+        defaults.update({
+            "AMSS_INTERP_FAST": "0",
+            "AMSS_INTERP_GPU": "0",
+            "AMSS_CUDA_AWARE_MPI": "0",
+        })
+    if finite_difference == "8th-order" and getattr(input_data, "Equation_Class", "") == "BSSN-EM":
+        defaults.update({
+            "AMSS_CUDA_AMR_RESTRICT_DEVICE": "1",
+            "AMSS_CUDA_AMR_RESTRICT_BATCH": "1",
+            "AMSS_CUDA_DEVICE_SEGMENT_BATCH": "1",
+        })
+    if getattr(input_data, "basic_grid_set", "") == "Shell-Patch":
+        defaults.update({
+            "AMSS_CUDA_AWARE_MPI": "0",
+            "AMSS_SHELL_FAST_INTERP": "1",
+            "AMSS_SHELL_PARALLEL_INTERP": "1",
+            "AMSS_SHELL_INTERP_THREADS": "16",
+        })
+    if getattr(input_data, "Equation_Class", "") in ("BSSN", "BSSN-EScalar", "Z4C"):
+        defaults["AMSS_CUDA_AMR_RESTRICT_DEVICE"] = "1"
+    if getattr(input_data, "Equation_Class", "") == "Z4C":
+        defaults.update({
+            "AMSS_Z4C_CUDA_RESIDENT": "1",
+            "AMSS_CONSTRAINT_OUT_EVERY": "1000000",
+        })
+    for key, value in defaults.items():
+        runtime_env.setdefault(key, value)
+
+    optional_overrides = {
+        "AMSS_INTERP_FAST_COMPARE": "AMSS_Interp_Fast_Compare",
+        "AMSS_INTERP_FAST_COMPARE_LIMIT": "AMSS_Interp_Fast_Compare_Limit",
+        "AMSS_INTERP_FAST_COMPARE_TOL": "AMSS_Interp_Fast_Compare_Tol",
+        "AMSS_GPU_STAGE_TIMING": "AMSS_GPU_Stage_Timing",
+        "AMSS_GPU_STAGE_TIMING_EVERY": "AMSS_GPU_Stage_Timing_Every",
+    }
+    for env_name, input_name in optional_overrides.items():
+        if env_name not in runtime_env and hasattr(input_data, input_name):
+            runtime_env[env_name] = str(getattr(input_data, input_name))
+
+    return runtime_env
+
+
+##################################################################
+
 
 
 ##################################################################
@@ -70,9 +223,9 @@ def makefile_ABE():
 
     ## Build command with CPU binding to nohz_full cores
     if (input_data.GPU_Calculation == "no"):
-        makefile_command  = f"{NUMACTL_CPU_BIND} make -j{BUILD_JOBS} INTERP_LB_MODE=off ABE"
+        makefile_command  = f"{NUMACTL_CPU_BIND} make -j{BUILD_JOBS} INTERP_LB_MODE=off USE_CUDA_BSSN=0 USE_CUDA_Z4C=0 ABE"
     elif (input_data.GPU_Calculation == "yes"):
-        makefile_command  = f"{NUMACTL_CPU_BIND} make -j{BUILD_JOBS} ABEGPU"
+        makefile_command  = f"{NUMACTL_CPU_BIND} make -j{BUILD_JOBS} INTERP_LB_MODE=off USE_CUDA_BSSN=1 USE_CUDA_Z4C=1 ABE_CUDA"
     else:
         print( " CPU/GPU numerical calculation setting is wrong " )
         print(                                                    )
@@ -145,29 +298,84 @@ def run_ABE():
     print(                                                      )
 
     ## Define the command to run; cast other values to strings as needed
+    mpi_env = None
+    started_mps = False
     
+    mpi_processes = int(input_data.MPI_processes)
+    if (input_data.GPU_Calculation == "yes" and
+        getattr(input_data, "Equation_Class", "") == "Z4C"):
+        z4c_env_np = os.environ.get("AMSS_Z4C_GPU_MPI_PROCESSES")
+        if z4c_env_np and int(z4c_env_np) > 0:
+            mpi_processes = int(z4c_env_np)
+        elif mpi_processes < 4:
+            mpi_processes = 4
+    if (input_data.GPU_Calculation == "yes" and
+        getattr(input_data, "basic_grid_set", "") == "Shell-Patch"):
+        shell_env_np = os.environ.get("AMSS_SHELL_GPU_MPI_PROCESSES")
+        if shell_env_np and int(shell_env_np) > 0:
+            mpi_processes = int(shell_env_np)
+        elif mpi_processes < 4:
+            mpi_processes = 4
+
     if (input_data.GPU_Calculation == "no"):
-        mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
+        mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(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"
+        mpi_command         = NUMACTL_CPU_BIND + " I_MPI_OFFLOAD=1 I_MPI_OFFLOAD_IPC=0 mpirun -np " + str(mpi_processes) + " ./ABE_CUDA"
         mpi_command_outfile = "ABEGPU_out.log"
+        mpi_env = _gpu_runtime_env()
+        started_mps = _start_cuda_mps_if_requested(mpi_env)
+        print(" GPU optimized runtime switches:")
+        print(f"   MPI processes={mpi_processes}")
+        print(f"   AMSS_INTERP_FAST={mpi_env.get('AMSS_INTERP_FAST', '')}")
+        print(f"   AMSS_INTERP_GPU={mpi_env.get('AMSS_INTERP_GPU', '')}")
+        print(f"   AMSS_ANALYSIS_MAP_EVERY={mpi_env.get('AMSS_ANALYSIS_MAP_EVERY', '')}")
+        print(f"   AMSS_EVOLVE_TIMING={mpi_env.get('AMSS_EVOLVE_TIMING', '')}")
+        print(f"   AMSS_ESCALAR_STEP_TIMING={mpi_env.get('AMSS_ESCALAR_STEP_TIMING', '')}")
+        print(f"   AMSS_CUDA_AWARE_MPI={mpi_env.get('AMSS_CUDA_AWARE_MPI', '')}")
+        print(f"   AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP={mpi_env.get('AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP', '')}")
+        print(f"   AMSS_CUDA_KEEP_ALL_LEVELS={mpi_env.get('AMSS_CUDA_KEEP_ALL_LEVELS', '')}")
+        print(f"   AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP={mpi_env.get('AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP', '')}")
+        print(f"   AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS={mpi_env.get('AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS', '')}")
+        print(f"   AMSS_CUDA_EM_CACHE_SOURCES={mpi_env.get('AMSS_CUDA_EM_CACHE_SOURCES', '')}")
+        print(f"   AMSS_CUDA_EM_ZERO_FASTPATH={mpi_env.get('AMSS_CUDA_EM_ZERO_FASTPATH', '')}")
+        print(f"   AMSS_EM_ZERO_ANALYSIS_FASTPATH={mpi_env.get('AMSS_EM_ZERO_ANALYSIS_FASTPATH', '')}")
+        print(f"   AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH={mpi_env.get('AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH', '')}")
+        print(f"   AMSS_CUDA_AMR_HOST_STAGED={mpi_env.get('AMSS_CUDA_AMR_HOST_STAGED', '')}")
+        print(f"   AMSS_CUDA_AMR_RESTRICT_DEVICE={mpi_env.get('AMSS_CUDA_AMR_RESTRICT_DEVICE', '')}")
+        print(f"   AMSS_CUDA_AMR_RESTRICT_BATCH={mpi_env.get('AMSS_CUDA_AMR_RESTRICT_BATCH', '')}")
+        print(f"   AMSS_CUDA_DEVICE_SEGMENT_BATCH={mpi_env.get('AMSS_CUDA_DEVICE_SEGMENT_BATCH', '')}")
+        print(f"   AMSS_CUDA_UNCACHED_DEVICE_BUFFERS={mpi_env.get('AMSS_CUDA_UNCACHED_DEVICE_BUFFERS', '')}")
+        print(f"   AMSS_SHELL_FAST_INTERP={mpi_env.get('AMSS_SHELL_FAST_INTERP', '')}")
+        print(f"   AMSS_SHELL_PARALLEL_INTERP={mpi_env.get('AMSS_SHELL_PARALLEL_INTERP', '')}")
+        print(f"   AMSS_SHELL_CUDA_INTERP={mpi_env.get('AMSS_SHELL_CUDA_INTERP', '')}")
+        print(f"   AMSS_SHELL_INTERP_THREADS={mpi_env.get('AMSS_SHELL_INTERP_THREADS', '')}")
+        print(f"   AMSS_Z4C_CUDA_RESIDENT={mpi_env.get('AMSS_Z4C_CUDA_RESIDENT', '')}")
+        print(f"   AMSS_CONSTRAINT_OUT_EVERY={mpi_env.get('AMSS_CONSTRAINT_OUT_EVERY', '')}")
+        if "CUDA_MPS_PIPE_DIRECTORY" in mpi_env:
+            print(f"   CUDA_MPS_PIPE_DIRECTORY={mpi_env['CUDA_MPS_PIPE_DIRECTORY']}")
  
-    ## Execute the MPI command and stream output
-    mpi_process = subprocess.Popen(mpi_command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
+    try:
+        ## Execute the MPI command and stream output
+        mpi_process = subprocess.Popen(mpi_command, shell=True, stdout=subprocess.PIPE,
+                                       stderr=subprocess.STDOUT, text=True, env=mpi_env)
 
-    ## Write ABE run output to file while printing to stdout
-    with open(mpi_command_outfile, 'w') as file0:  
-        ## Read and print output lines; also write each line to file
-        for line in mpi_process.stdout:
-            print(line, end='')  # stream output in real time
-            file0.write(line)    # write the line to file
-            file0.flush()        # flush to ensure each line is written immediately (optional)            
-    file0.close()
+        ## Write ABE run output to file while printing to stdout
+        with open(mpi_command_outfile, 'w') as file0:
+            ## Read and print output lines; also write each line to file
+            for line in mpi_process.stdout:
+                print(line, end='')  # stream output in real time
+                file0.write(line)    # write the line to file
+                file0.flush()        # flush to ensure each line is written immediately (optional)
 
-    ## Wait for the process to finish
-    mpi_return_code = mpi_process.wait()
+        ## Wait for the process to finish
+        mpi_return_code = mpi_process.wait()
+        if mpi_return_code != 0:
+            raise subprocess.CalledProcessError(mpi_return_code, mpi_command)
+    finally:
+        if started_mps:
+            _stop_cuda_mps(mpi_env)
     
     print(                                           )
     print( " The ABE/ABEGPU simulation is finished " ) 

plot_xiaoqu.py

@@ -808,10 +808,10 @@ def generate_ADMmass_plot( outdir, figure_outdir, detector_number_i ):
 
 ## Plot constraint violation for each grid level
 
-def generate_constraint_check_plot( outdir, figure_outdir, input_level_number ):
-
-    # path to data file
-    file0 = os.path.join(outdir, "bssn_constraint.dat")
+def generate_constraint_check_plot( outdir, figure_outdir, input_level_number ):
+
+    # path to data file
+    file0 = os.path.join(outdir, "bssn_constraint.dat")
 
     if ( input_level_number == 0 ):
         print(                                                   )
@@ -819,13 +819,26 @@ def generate_constraint_check_plot( outdir, figure_outdir, input_level_number ):
         print(                                                          )
         print( " corresponding data file = ", file0 )
         print(                                      )
-
-    print( " Begin the constraint violation plot for grid level number =  ", input_level_number )
-    
-    # load the full data file (assumed whitespace-separated floats)
-    data = numpy.loadtxt(file0)
-    
-    # extract columns from the constraint data file
+
+    print( " Begin the constraint violation plot for grid level number =  ", input_level_number )
+
+    if (not os.path.exists(file0)) or os.path.getsize(file0) == 0:
+        if ( input_level_number == 0 ):
+            print( " Constraint data file is empty; skip constraint violation plots" )
+            print(                                                            )
+        return
+
+    # load the full data file (assumed whitespace-separated floats)
+    data = numpy.loadtxt(file0)
+    data = numpy.atleast_2d(data)
+
+    if data.shape[1] < 8:
+        if ( input_level_number == 0 ):
+            print( " Constraint data file has insufficient columns; skip constraint violation plots" )
+            print(                                                                                )
+        return
+
+    # extract columns from the constraint data file
     time          = data[:,0]
     Constraint_H  = data[:,1]
     Constraint_Px = data[:,2]