修改block划分，对负载高的rank所在block进行划分，添加到空rank，空rank是平移得到的

Rank [  20]: Calc 0.000003 s

Rank [  35]: Calc 0.000003 s

Rank [  10]: Calc 0.000010 s

Rank [  17]: Calc 0.000005 s

Rank [  32]: Calc 0.000003 s，而且rank不固定，一般就是rank 10 和 rank 52；
但尽管有很多，比前者时间还是少很多
对于第二个Surf_Wave 调用 Interp_points，我发现前四个rank时间最长，比较固定，就是下面四个rank

Rank [  27]: Calc 0.331978 s

Rank [  35]: Calc 0.242219 s

Rank [  28]: Calc 0.242132 s

Rank [  36]: Calc 0.197024 s
因此下面surf_wave是核心

AMSS_NCKU_Input.py

@@ -16,7 +16,7 @@ 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    = 48                             ## number of mpi processes used in the simulation
+MPI_processes    = 64                             ## number of mpi processes used in the simulation
 
 GPU_Calculation  = "no"                          ## Use GPU or not 
                                                  ## (prefer "no" in the current version, because the GPU part may have bugs when integrated in this Python interface)

AMSS_NCKU_Program.py

@@ -8,6 +8,14 @@
 ##
 ##################################################################
 
+## Guard against re-execution by multiprocessing child processes.
+## Without this, using 'spawn' or 'forkserver' context would cause every
+## worker to re-run the entire script, spawning exponentially more
+## workers (fork bomb).
+if __name__ != '__main__':
+    import sys as _sys
+    _sys.exit(0)
+
 
 ##################################################################
 
@@ -58,7 +66,8 @@ if os.path.exists(File_directory):
     ## Prompt whether to overwrite the existing directory
     while True:
         try:
-            inputvalue = input()
+            ## inputvalue = input()
+            inputvalue = "continue"
             ## If the user agrees to overwrite, proceed and remove the existing directory
             if ( inputvalue == "continue" ):
                 print( " Continue the calculation !!! " )
@@ -424,26 +433,31 @@ print(
 
 import plot_xiaoqu
 import plot_GW_strain_amplitude_xiaoqu
+from parallel_plot_helper import run_plot_tasks_parallel
+
+plot_tasks = []
 
 ## Plot black hole trajectory
-plot_xiaoqu.generate_puncture_orbit_plot(   binary_results_directory, figure_directory )
-plot_xiaoqu.generate_puncture_orbit_plot3D( binary_results_directory, figure_directory )
+plot_tasks.append( ( plot_xiaoqu.generate_puncture_orbit_plot,   (binary_results_directory, figure_directory) ) )
+plot_tasks.append( ( plot_xiaoqu.generate_puncture_orbit_plot3D, (binary_results_directory, figure_directory) ) )
 
 ## Plot black hole separation vs. time
-plot_xiaoqu.generate_puncture_distence_plot( binary_results_directory, figure_directory )
+plot_tasks.append( ( plot_xiaoqu.generate_puncture_distence_plot, (binary_results_directory, figure_directory) ) )
 
 ## Plot gravitational waveforms (psi4 and strain amplitude)
 for i in range(input_data.Detector_Number):
-    plot_xiaoqu.generate_gravitational_wave_psi4_plot( binary_results_directory, figure_directory, i )
-    plot_GW_strain_amplitude_xiaoqu.generate_gravitational_wave_amplitude_plot( binary_results_directory, figure_directory, i )
+    plot_tasks.append( ( plot_xiaoqu.generate_gravitational_wave_psi4_plot, (binary_results_directory, figure_directory, i) ) )
+    plot_tasks.append( ( plot_GW_strain_amplitude_xiaoqu.generate_gravitational_wave_amplitude_plot, (binary_results_directory, figure_directory, i) ) )
 
 ## Plot ADM mass evolution
 for i in range(input_data.Detector_Number):
-    plot_xiaoqu.generate_ADMmass_plot( binary_results_directory, figure_directory, i )
+    plot_tasks.append( ( plot_xiaoqu.generate_ADMmass_plot, (binary_results_directory, figure_directory, i) ) )
 
 ## Plot Hamiltonian constraint violation over time
 for i in range(input_data.grid_level):
-    plot_xiaoqu.generate_constraint_check_plot( binary_results_directory, figure_directory, i )
+    plot_tasks.append( ( plot_xiaoqu.generate_constraint_check_plot, (binary_results_directory, figure_directory, i) ) )
+
+run_plot_tasks_parallel(plot_tasks)
 
 ## Plot stored binary data
 plot_xiaoqu.generate_binary_data_plot( binary_results_directory, figure_directory )

AMSS_NCKU_source/MPatch.h

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

AMSS_NCKU_source/NullShellPatch.h

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

AMSS_NCKU_source/Parallel.h

@@ -1,167 +1,235 @@
-
-#ifndef PARALLEL_H
-#define PARALLEL_H
-
-#include <iostream>
-#include <iomanip>
-#include <fstream>
-#include <cstdlib>
-#include <cstdio>
-#include <string>
-#include <cmath>
-#include <new>
-using namespace std;
-
-#include "Parallel_bam.h"
-#include "var.h"
-#include "MPatch.h"
-#include "Block.h"
-#include "MyList.h"
-#include "macrodef.h" //need dim; ghost_width; CONTRACT
-namespace Parallel
-{
-  struct gridseg
-  {
-    double llb[dim];
-    double uub[dim];
-    int shape[dim];
-    double illb[dim], iuub[dim]; // only use for OutBdLow2Hi
-    Block *Bg;
-  };
-  int partition1(int &nx, int split_size, int min_width, int cpusize, int shape);    // special for 1 diemnsion
-  int partition2(int *nxy, int split_size, int *min_width, int cpusize, int *shape); // special for 2 diemnsions
-  int partition3(int *nxyz, int split_size, int *min_width, int cpusize, int *shape);
-  MyList<Block> *distribute(MyList<Patch> *PatchLIST, int cpusize, int ingfsi, int fngfs, bool periodic, int nodes = 0); // produce corresponding Blocks
-  void KillBlocks(MyList<Patch> *PatchLIST);
-
-  void setfunction(MyList<Block> *BlL, var *vn, double func(double x, double y, double z));
-  void setfunction(int rank, MyList<Block> *BlL, var *vn, double func(double x, double y, double z));
-  void writefile(double time, int nx, int ny, int nz, double xmin, double xmax, double ymin, double ymax,
-                 double zmin, double zmax, char *filename, double *data_out);
-  void writefile(double time, int nx, int ny, double xmin, double xmax, double ymin, double ymax,
-                 char *filename, double *datain);
-  void getarrayindex(int DIM, int *shape, int *index, int n);
-  int getarraylocation(int DIM, int *shape, int *index);
-  void copy(int DIM, double *llbout, double *uubout, int *Dshape, double *DD, double *llbin, double *uubin,
-            int *shape, double *datain, double *llb, double *uub);
-  void Dump_CPU_Data(MyList<Block> *BlL, MyList<var> *DumpList, char *tag, double time, double dT);
-  void Dump_Data(MyList<Patch> *PL, MyList<var> *DumpList, char *tag, double time, double dT);
-  void Dump_Data(Patch *PP, MyList<var> *DumpList, char *tag, double time, double dT, int grd);
-  double *Collect_Data(Patch *PP, var *VP);
-  void d2Dump_Data(MyList<Patch> *PL, MyList<var> *DumpList, char *tag, double time, double dT);
-  void d2Dump_Data(Patch *PP, MyList<var> *DumpList, char *tag, double time, double dT, int grd);
-  void Dump_Data0(Patch *PP, MyList<var> *DumpList, char *tag, double time, double dT);
-  double global_interp(int DIM, int *ext, double **CoX, double *datain,
-                       double *poX, int ordn, double *SoA, int Symmetry);
-  double global_interp(int DIM, int *ext, double **CoX, double *datain,
-                       double *poX, int ordn);
-  double Lagrangian_Int(double x, int npts, double *xpts, double *funcvals);
-  double LagrangePoly(double x, int pt, int npts, double *xpts);
-  MyList<gridseg> *build_complete_gsl(Patch *Pat);
-  MyList<gridseg> *build_complete_gsl(MyList<Patch> *PatL);
-  MyList<gridseg> *build_complete_gsl_virtual(MyList<Patch> *PatL);
-  MyList<gridseg> *build_complete_gsl_virtual2(MyList<Patch> *PatL);        // - buffer
-  MyList<gridseg> *build_owned_gsl0(Patch *Pat, int rank_in);               // - ghost without extension, special for Sync usage
-  MyList<gridseg> *build_owned_gsl1(Patch *Pat, int rank_in);               // - ghost, similar to build_owned_gsl0 but extend one point on left side for vertex grid
-  MyList<gridseg> *build_owned_gsl2(Patch *Pat, int rank_in);               // - buffer - ghost
-  MyList<gridseg> *build_owned_gsl3(Patch *Pat, int rank_in, int Symmetry); // - ghost - BD ghost
-  MyList<gridseg> *build_owned_gsl4(Patch *Pat, int rank_in, int Symmetry); // - buffer - ghost - BD ghost
-  MyList<gridseg> *build_owned_gsl5(Patch *Pat, int rank_in);               // similar to build_owned_gsl2 but no extension
-  MyList<gridseg> *build_owned_gsl(MyList<Patch> *PatL, int rank_in, int type, int Symmetry);
-  void build_gstl(MyList<gridseg> *srci, MyList<gridseg> *dsti, MyList<gridseg> **out_src, MyList<gridseg> **out_dst);
-  int data_packer(double *data, MyList<gridseg> *src, MyList<gridseg> *dst, int rank_in, int dir,
-                  MyList<var> *VarLists, MyList<var> *VarListd, int Symmetry);
-  void transfer(MyList<gridseg> **src, MyList<gridseg> **dst,
-                MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /*target */,
-                int Symmetry);
-  int data_packermix(double *data, MyList<gridseg> *src, MyList<gridseg> *dst, int rank_in, int dir,
-                     MyList<var> *VarLists, MyList<var> *VarListd, int Symmetry);
-  void transfermix(MyList<gridseg> **src, MyList<gridseg> **dst,
-                   MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /*target */,
-                   int Symmetry);
-  void Sync(Patch *Pat, MyList<var> *VarList, int Symmetry);
-  void Sync(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
-  void OutBdLow2Hi(Patch *Patc, Patch *Patf,
-                   MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
-                   int Symmetry);
-  void OutBdLow2Hi(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
-                   MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
-                   int Symmetry);
-  void OutBdLow2Himix(Patch *Patc, Patch *Patf,
-                      MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
-                      int Symmetry);
-  void OutBdLow2Himix(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
-                      MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
-                      int Symmetry);
-  void Prolong(Patch *Patc, Patch *Patf,
-               MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
-               int Symmetry);
-  void Prolongint(Patch *Patc, Patch *Patf,
-                  MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
-                  int Symmetry);
-  void Restrict(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
-                MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
-                int Symmetry);
-  void Restrict_after(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
-                      MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
-                      int Symmetry); // for -ghost - BDghost
-  MyList<Parallel::gridseg> *build_PhysBD_gsl(Patch *Pat);
-  MyList<Parallel::gridseg> *build_ghost_gsl(MyList<Patch> *PatL);
-  MyList<Parallel::gridseg> *build_ghost_gsl(Patch *Pat);
-  MyList<Parallel::gridseg> *build_buffer_gsl(Patch *Pat);
-  MyList<Parallel::gridseg> *build_buffer_gsl(MyList<Patch> *PatL);
-  MyList<Parallel::gridseg> *gsl_subtract(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
-  MyList<Parallel::gridseg> *gs_subtract(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
-  MyList<Parallel::gridseg> *gsl_and(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
-  MyList<Parallel::gridseg> *gs_and(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
-  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 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,
-                                MyList<var> *VarList1 /* source (t+dt) */, MyList<var> *VarList2 /* source (t) */,
-                                MyList<var> *VarList3 /* target (t+a*dt) */, int tindex);
-  void prepare_inter_time_level(Patch *Pat,
-                                MyList<var> *VarList1 /* source (t+dt) */, MyList<var> *VarList2 /* source (t) */,
-                                MyList<var> *VarList3 /* source (t-dt) */, MyList<var> *VarList4 /* target (t+a*dt) */, int tindex);
-  void prepare_inter_time_level(MyList<Patch> *PatL,
-                                MyList<var> *VarList1 /* source (t+dt) */, MyList<var> *VarList2 /* source (t) */,
-                                MyList<var> *VarList3 /* target (t+a*dt) */, int tindex);
-  void prepare_inter_time_level(MyList<Patch> *Pat,
-                                MyList<var> *VarList1 /* source (t+dt) */, MyList<var> *VarList2 /* source (t) */,
-                                MyList<var> *VarList3 /* source (t-dt) */, MyList<var> *VarList4 /* target (t+a*dt) */, int tindex);
-  void merge_gsl(MyList<gridseg> *&A, const double ratio);
-  bool merge_gs(MyList<gridseg> *D, MyList<gridseg> *B, MyList<gridseg> *&C, const double ratio);
-  // Add ghost region to tangent plane
-  // we assume the grids have the same resolution
-  void add_ghost_touch(MyList<gridseg> *&A);
-  void cut_gsl(MyList<gridseg> *&A);
-  bool cut_gs(MyList<gridseg> *D, MyList<gridseg> *B, MyList<gridseg> *&C);
-  MyList<Parallel::gridseg> *gs_subtract_virtual(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
-  void fill_level_data(MyList<Patch> *PatLd, MyList<Patch> *PatLs, MyList<Patch> *PatcL,
-                       MyList<var> *OldList, MyList<var> *StateList, MyList<var> *FutureList,
-                       MyList<var> *tmList, int Symmetry, bool BB, bool CC);
-  bool PatList_Interp_Points(MyList<Patch> *PatL, MyList<var> *VarList,
-                             int NN, double **XX,
-                             double *Shellf, int Symmetry);
-  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);
-  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);
-#endif
-}
-#endif /*PARALLEL_H */
+
+#ifndef PARALLEL_H
+#define PARALLEL_H
+
+#include <iostream>
+#include <iomanip>
+#include <fstream>
+#include <cstdlib>
+#include <cstdio>
+#include <string>
+#include <cmath>
+#include <new>
+using namespace std;
+#include <memory.h>
+#include "Parallel_bam.h"
+#include "var.h"
+#include "MPatch.h"
+#include "Block.h"
+#include "MyList.h"
+#include "macrodef.h" //need dim; ghost_width; CONTRACT
+namespace Parallel
+{
+  struct gridseg
+  {
+    double llb[dim];
+    double uub[dim];
+    int shape[dim];
+    double illb[dim], iuub[dim]; // only use for OutBdLow2Hi
+    Block *Bg;
+  };
+  int partition1(int &nx, int split_size, int min_width, int cpusize, int shape);    // special for 1 diemnsion
+  int partition2(int *nxy, int split_size, int *min_width, int cpusize, int *shape); // special for 2 diemnsions
+  int partition3(int *nxyz, int split_size, int *min_width, int cpusize, int *shape);
+  MyList<Block> *distribute(MyList<Patch> *PatchLIST, int cpusize, int ingfsi, int fngfs, bool periodic, int nodes = 0); // produce corresponding Blocks
+  MyList<Block> *distribute_hard(MyList<Patch> *PatchLIST, int cpusize, int ingfsi, int fngfs, bool periodic, int nodes = 0); // produce corresponding Blocks
+  Block* splitHotspotBlock(MyList<Block>* &BlL, int _dim, 
+                                 int ib0_orig, int ib3_orig, 
+                                 int jb1_orig, int jb4_orig, 
+                                 int kb2_orig, int kb5_orig, 
+                                 Patch* PP, int r_left, int r_right, 
+                                 int ingfsi, int fngfsi, bool periodic,
+                                 Block* &split_first_block, Block* &split_last_block);
+  Block* createMappedBlock(MyList<Block>* &BlL, int _dim, int* shape, double* bbox,
+                        int block_id, int ingfsi, int fngfsi, int lev); 
+  void KillBlocks(MyList<Patch> *PatchLIST);
+
+  void setfunction(MyList<Block> *BlL, var *vn, double func(double x, double y, double z));
+  void setfunction(int rank, MyList<Block> *BlL, var *vn, double func(double x, double y, double z));
+  void writefile(double time, int nx, int ny, int nz, double xmin, double xmax, double ymin, double ymax,
+                 double zmin, double zmax, char *filename, double *data_out);
+  void writefile(double time, int nx, int ny, double xmin, double xmax, double ymin, double ymax,
+                 char *filename, double *datain);
+  void getarrayindex(int DIM, int *shape, int *index, int n);
+  int getarraylocation(int DIM, int *shape, int *index);
+  void copy(int DIM, double *llbout, double *uubout, int *Dshape, double *DD, double *llbin, double *uubin,
+            int *shape, double *datain, double *llb, double *uub);
+  void Dump_CPU_Data(MyList<Block> *BlL, MyList<var> *DumpList, char *tag, double time, double dT);
+  void Dump_Data(MyList<Patch> *PL, MyList<var> *DumpList, char *tag, double time, double dT);
+  void Dump_Data(Patch *PP, MyList<var> *DumpList, char *tag, double time, double dT, int grd);
+  double *Collect_Data(Patch *PP, var *VP);
+  void d2Dump_Data(MyList<Patch> *PL, MyList<var> *DumpList, char *tag, double time, double dT);
+  void d2Dump_Data(Patch *PP, MyList<var> *DumpList, char *tag, double time, double dT, int grd);
+  void Dump_Data0(Patch *PP, MyList<var> *DumpList, char *tag, double time, double dT);
+  double global_interp(int DIM, int *ext, double **CoX, double *datain,
+                       double *poX, int ordn, double *SoA, int Symmetry);
+  double global_interp(int DIM, int *ext, double **CoX, double *datain,
+                       double *poX, int ordn);
+  double Lagrangian_Int(double x, int npts, double *xpts, double *funcvals);
+  double LagrangePoly(double x, int pt, int npts, double *xpts);
+  MyList<gridseg> *build_complete_gsl(Patch *Pat);
+  MyList<gridseg> *build_complete_gsl(MyList<Patch> *PatL);
+  MyList<gridseg> *build_complete_gsl_virtual(MyList<Patch> *PatL);
+  MyList<gridseg> *build_complete_gsl_virtual2(MyList<Patch> *PatL);        // - buffer
+  MyList<gridseg> *build_owned_gsl0(Patch *Pat, int rank_in);               // - ghost without extension, special for Sync usage
+  MyList<gridseg> *build_owned_gsl1(Patch *Pat, int rank_in);               // - ghost, similar to build_owned_gsl0 but extend one point on left side for vertex grid
+  MyList<gridseg> *build_owned_gsl2(Patch *Pat, int rank_in);               // - buffer - ghost
+  MyList<gridseg> *build_owned_gsl3(Patch *Pat, int rank_in, int Symmetry); // - ghost - BD ghost
+  MyList<gridseg> *build_owned_gsl4(Patch *Pat, int rank_in, int Symmetry); // - buffer - ghost - BD ghost
+  MyList<gridseg> *build_owned_gsl5(Patch *Pat, int rank_in);               // similar to build_owned_gsl2 but no extension
+  MyList<gridseg> *build_owned_gsl(MyList<Patch> *PatL, int rank_in, int type, int Symmetry);
+  void build_gstl(MyList<gridseg> *srci, MyList<gridseg> *dsti, MyList<gridseg> **out_src, MyList<gridseg> **out_dst);
+  int data_packer(double *data, MyList<gridseg> *src, MyList<gridseg> *dst, int rank_in, int dir,
+                  MyList<var> *VarLists, MyList<var> *VarListd, int Symmetry);
+  void transfer(MyList<gridseg> **src, MyList<gridseg> **dst,
+                MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /*target */,
+                int Symmetry);
+  int data_packermix(double *data, MyList<gridseg> *src, MyList<gridseg> *dst, int rank_in, int dir,
+                     MyList<var> *VarLists, MyList<var> *VarListd, int Symmetry);
+  void transfermix(MyList<gridseg> **src, MyList<gridseg> **dst,
+                   MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /*target */,
+                   int Symmetry);
+  void Sync(Patch *Pat, MyList<var> *VarList, int Symmetry);
+  void Sync(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
+  void Sync_merged(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
+
+  struct SyncCache {
+    bool valid;
+    int cpusize;
+    MyList<gridseg> **combined_src;
+    MyList<gridseg> **combined_dst;
+    int *send_lengths;
+    int *recv_lengths;
+    double **send_bufs;
+    double **recv_bufs;
+    int *send_buf_caps;
+    int *recv_buf_caps;
+    MPI_Request *reqs;
+    MPI_Status *stats;
+    int max_reqs;
+    bool lengths_valid;
+    SyncCache();
+    void invalidate();
+    void destroy();
+  };
+
+  void Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, SyncCache &cache);
+  void transfer_cached(MyList<gridseg> **src, MyList<gridseg> **dst,
+                       MyList<var> *VarList1, MyList<var> *VarList2,
+                       int Symmetry, SyncCache &cache);
+
+  struct AsyncSyncState {
+    int req_no;
+    bool active;
+    AsyncSyncState() : req_no(0), active(false) {}
+  };
+
+  void Sync_start(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry,
+                  SyncCache &cache, AsyncSyncState &state);
+  void Sync_finish(SyncCache &cache, AsyncSyncState &state,
+                   MyList<var> *VarList, int Symmetry);
+  void OutBdLow2Hi(Patch *Patc, Patch *Patf,
+                   MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
+                   int Symmetry);
+  void OutBdLow2Hi(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                   MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
+                   int Symmetry);
+  void OutBdLow2Himix(Patch *Patc, Patch *Patf,
+                      MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
+                      int Symmetry);
+  void OutBdLow2Himix(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                      MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
+                      int Symmetry);
+  void Restrict_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                       MyList<var> *VarList1, MyList<var> *VarList2,
+                       int Symmetry, SyncCache &cache);
+  void OutBdLow2Hi_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                          MyList<var> *VarList1, MyList<var> *VarList2,
+                          int Symmetry, SyncCache &cache);
+  void OutBdLow2Himix_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                             MyList<var> *VarList1, MyList<var> *VarList2,
+                             int Symmetry, SyncCache &cache);
+  void Prolong(Patch *Patc, Patch *Patf,
+               MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
+               int Symmetry);
+  void Prolongint(Patch *Patc, Patch *Patf,
+                  MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
+                  int Symmetry);
+  void Restrict(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
+                int Symmetry);
+  void Restrict_after(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
+                      MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
+                      int Symmetry); // for -ghost - BDghost
+  MyList<Parallel::gridseg> *build_PhysBD_gsl(Patch *Pat);
+  MyList<Parallel::gridseg> *build_ghost_gsl(MyList<Patch> *PatL);
+  MyList<Parallel::gridseg> *build_ghost_gsl(Patch *Pat);
+  MyList<Parallel::gridseg> *build_buffer_gsl(Patch *Pat);
+  MyList<Parallel::gridseg> *build_buffer_gsl(MyList<Patch> *PatL);
+  MyList<Parallel::gridseg> *gsl_subtract(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
+  MyList<Parallel::gridseg> *gs_subtract(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
+  MyList<Parallel::gridseg> *gsl_and(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
+  MyList<Parallel::gridseg> *gs_and(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
+  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 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,
+                                MyList<var> *VarList1 /* source (t+dt) */, MyList<var> *VarList2 /* source (t) */,
+                                MyList<var> *VarList3 /* target (t+a*dt) */, int tindex);
+  void prepare_inter_time_level(Patch *Pat,
+                                MyList<var> *VarList1 /* source (t+dt) */, MyList<var> *VarList2 /* source (t) */,
+                                MyList<var> *VarList3 /* source (t-dt) */, MyList<var> *VarList4 /* target (t+a*dt) */, int tindex);
+  void prepare_inter_time_level(MyList<Patch> *PatL,
+                                MyList<var> *VarList1 /* source (t+dt) */, MyList<var> *VarList2 /* source (t) */,
+                                MyList<var> *VarList3 /* target (t+a*dt) */, int tindex);
+  void prepare_inter_time_level(MyList<Patch> *Pat,
+                                MyList<var> *VarList1 /* source (t+dt) */, MyList<var> *VarList2 /* source (t) */,
+                                MyList<var> *VarList3 /* source (t-dt) */, MyList<var> *VarList4 /* target (t+a*dt) */, int tindex);
+  void merge_gsl(MyList<gridseg> *&A, const double ratio);
+  bool merge_gs(MyList<gridseg> *D, MyList<gridseg> *B, MyList<gridseg> *&C, const double ratio);
+  // Add ghost region to tangent plane
+  // we assume the grids have the same resolution
+  void add_ghost_touch(MyList<gridseg> *&A);
+  void cut_gsl(MyList<gridseg> *&A);
+  bool cut_gs(MyList<gridseg> *D, MyList<gridseg> *B, MyList<gridseg> *&C);
+  MyList<Parallel::gridseg> *gs_subtract_virtual(MyList<Parallel::gridseg> *A, MyList<Parallel::gridseg> *B);
+  void fill_level_data(MyList<Patch> *PatLd, MyList<Patch> *PatLs, MyList<Patch> *PatcL,
+                       MyList<var> *OldList, MyList<var> *StateList, MyList<var> *FutureList,
+                       MyList<var> *tmList, int Symmetry, bool BB, bool CC);
+  bool PatList_Interp_Points(MyList<Patch> *PatL, MyList<var> *VarList,
+                             int NN, double **XX,
+                             double *Shellf, int Symmetry);
+  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);
+  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);
+  
+  // Redistribute blocks with time statistics for load balancing
+  MyList<Block> *distribute(MyList<Patch> *PatchLIST, MyList<Block> *OldBlockL,
+                            int cpusize, int ingfsi, int fngfsi,
+                            bool periodic, int start_rank, int end_rank, int nodes = 0);
+#endif
+
+  // Dynamic load balancing: split blocks for heavy ranks
+    void split_heavy_blocks(MyList<Patch> *PatL, int *heavy_ranks, int num_heavy,
+                            int split_factor, int cpusize, int ingfsi, int fngfsi);
+
+    // Check if load balancing is needed based on interpolation times
+    bool check_load_balance_need(double *rank_times, int nprocs, int &num_heavy, int *heavy_ranks);
+  }
+  #endif /*PARALLEL_H */

AMSS_NCKU_source/TwoPunctures.h

@@ -1,7 +1,8 @@
-
 #ifndef TWO_PUNCTURES_H
 #define TWO_PUNCTURES_H
 
+#include <omp.h>
+
 #define StencilSize 19
 #define N_PlaneRelax 1
 #define NRELAX 200
@@ -32,7 +33,7 @@ private:
        int npoints_A, npoints_B, npoints_phi;
 
        double target_M_plus, target_M_minus;
-       
+
        double admMass;
 
        double adm_tol;
@@ -42,6 +43,18 @@ private:
 
        int ntotal;
 
+       // ===== Precomputed spectral derivative matrices =====
+       double *D1_A, *D2_A;
+       double *D1_B, *D2_B;
+       double *DF1_phi, *DF2_phi;
+
+       // ===== Pre-allocated workspace for LineRelax (per-thread) =====
+       int max_threads;
+       double **ws_diag_be, **ws_e_be, **ws_f_be, **ws_b_be, **ws_x_be;
+       double **ws_l_be, **ws_u_be, **ws_d_be, **ws_y_be;
+       double **ws_diag_al, **ws_e_al, **ws_f_al, **ws_b_al, **ws_x_al;
+       double **ws_l_al, **ws_u_al, **ws_d_al, **ws_y_al;
+
        struct parameters
        {
               int nvar, n1, n2, n3;
@@ -58,6 +71,28 @@ public:
                     int Newtonmaxit);
        ~TwoPunctures();
 
+       // 02/07: New/modified methods
+       void allocate_workspace();
+       void free_workspace();
+       void precompute_derivative_matrices();
+       void build_cheb_deriv_matrices(int n, double *D1, double *D2);
+       void build_fourier_deriv_matrices(int N, double *DF1, double *DF2);
+       void Derivatives_AB3_MatMul(int nvar, int n1, int n2, int n3, derivs v);
+       void ThomasAlgorithm_ws(int N, double *b, double *a, double *c, double *x, double *q,
+                                double *l, double *u_ws, double *d, double *y);
+       void LineRelax_be_omp(double *dv,
+                         int const i, int const k, int const nvar,
+                         int const n1, int const n2, int const n3,
+                         double const *rhs, int const *ncols, int **cols,
+                         double **JFD, int tid);
+       void LineRelax_al_omp(double *dv,
+                         int const j, int const k, int const nvar,
+                         int const n1, int const n2, int const n3,
+                         double const *rhs, int const *ncols,
+                         int **cols, double **JFD, int tid);
+       void relax_omp(double *dv, int const nvar, int const n1, int const n2, int const n3,
+                  double const *rhs, int const *ncols, int **cols, double **JFD);
+
        void Solve();
        void set_initial_guess(derivs v);
        int index(int i, int j, int k, int l, int a, int b, int c, int d);
@@ -116,23 +151,11 @@ public:
        double BY_KKofxyz(double x, double y, double z);
        void SetMatrix_JFD(int nvar, int n1, int n2, int n3, derivs u, int *ncols, int **cols, double **Matrix);
        void J_times_dv(int nvar, int n1, int n2, int n3, derivs dv, double *Jdv, derivs u);
-       void relax(double *dv, int const nvar, int const n1, int const n2, int const n3,
-                  double const *rhs, int const *ncols, int **cols, double **JFD);
-       void LineRelax_be(double *dv,
-                         int const i, int const k, int const nvar,
-                         int const n1, int const n2, int const n3,
-                         double const *rhs, int const *ncols, int **cols,
-                         double **JFD);
        void JFD_times_dv(int i, int j, int k, int nvar, int n1, int n2,
                          int n3, derivs dv, derivs u, double *values);
        void LinEquations(double A, double B, double X, double R,
                          double x, double r, double phi,
                          double y, double z, derivs dU, derivs U, double *values);
-       void LineRelax_al(double *dv,
-                         int const j, int const k, int const nvar,
-                         int const n1, int const n2, int const n3,
-                         double const *rhs, int const *ncols,
-                         int **cols, double **JFD);
        void ThomasAlgorithm(int N, double *b, double *a, double *c, double *x, double *q);
        void Save(char *fname);
        // provided by Vasileios Paschalidis (vpaschal@illinois.edu)
@@ -141,4 +164,4 @@ public:
        void SpecCoef(parameters par, int ivar, double *v, double *cf);
 };
 
-#endif /* TWO_PUNCTURES_H */
+#endif /* TWO_PUNCTURES_H */

AMSS_NCKU_source/bssn_class.C

@@ -730,6 +730,12 @@ void bssn_class::Initialize()
     PhysTime = StartTime;
     Setup_Black_Hole_position();
   }
+
+  // Initialize sync caches (per-level, for predictor and corrector)
+  sync_cache_pre = new Parallel::SyncCache[GH->levels];
+  sync_cache_cor = new Parallel::SyncCache[GH->levels];
+  sync_cache_rp_coarse = new Parallel::SyncCache[GH->levels];
+  sync_cache_rp_fine = new Parallel::SyncCache[GH->levels];
 }
 
 //================================================================================================
@@ -981,6 +987,32 @@ bssn_class::~bssn_class()
   delete Azzz;
 #endif
 
+  // Destroy sync caches before GH
+  if (sync_cache_pre)
+  {
+    for (int i = 0; i < GH->levels; i++)
+      sync_cache_pre[i].destroy();
+    delete[] sync_cache_pre;
+  }
+  if (sync_cache_cor)
+  {
+    for (int i = 0; i < GH->levels; i++)
+      sync_cache_cor[i].destroy();
+    delete[] sync_cache_cor;
+  }
+  if (sync_cache_rp_coarse)
+  {
+    for (int i = 0; i < GH->levels; i++)
+      sync_cache_rp_coarse[i].destroy();
+    delete[] sync_cache_rp_coarse;
+  }
+  if (sync_cache_rp_fine)
+  {
+    for (int i = 0; i < GH->levels; i++)
+      sync_cache_rp_fine[i].destroy();
+    delete[] sync_cache_rp_fine;
+  }
+
   delete GH;
 #ifdef WithShell
   delete SH;
@@ -2181,6 +2213,7 @@ void bssn_class::Evolve(int Steps)
     GH->Regrid(Symmetry, BH_num, Porgbr, Porg0,
                SynchList_cor, OldStateList, StateList, SynchList_pre,
                fgt(PhysTime - dT_mon, StartTime, dT_mon / 2), ErrorMonitor);
+    for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 
 #if (REGLEV == 0 && (PSTR == 1 || PSTR == 2))
@@ -2396,6 +2429,7 @@ void bssn_class::RecursiveStep(int lev)
   GH->Regrid_Onelevel(lev, Symmetry, BH_num, Porgbr, Porg0,
                       SynchList_cor, OldStateList, StateList, SynchList_pre,
                       fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
+  for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 }
 
@@ -2574,6 +2608,7 @@ void bssn_class::ParallelStep()
   GH->Regrid_Onelevel(GH->mylev, Symmetry, BH_num, Porgbr, Porg0,
                       SynchList_cor, OldStateList, StateList, SynchList_pre,
                       fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
+  for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 #endif
 }
 
@@ -2740,6 +2775,7 @@ void bssn_class::ParallelStep()
         GH->Regrid_Onelevel(lev + 1, Symmetry, BH_num, Porgbr, Porg0,
                             SynchList_cor, OldStateList, StateList, SynchList_pre,
                             fgt(PhysTime - dT_levp1, StartTime, dT_levp1 / 2), ErrorMonitor);
+        for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
         //               a_stream.clear();
         //               a_stream.str("");
@@ -2754,6 +2790,7 @@ void bssn_class::ParallelStep()
       GH->Regrid_Onelevel(lev, Symmetry, BH_num, Porgbr, Porg0,
                           SynchList_cor, OldStateList, StateList, SynchList_pre,
                           fgt(PhysTime - dT_lev, StartTime, dT_lev / 2), ErrorMonitor);
+      for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
       //               a_stream.clear();
       //               a_stream.str("");
@@ -2772,6 +2809,7 @@ void bssn_class::ParallelStep()
           GH->Regrid_Onelevel(lev - 1, Symmetry, BH_num, Porgbr, Porg0,
                               SynchList_cor, OldStateList, StateList, SynchList_pre,
                               fgt(PhysTime - dT_lev, StartTime, dT_levm1 / 2), ErrorMonitor);
+          for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
           //               a_stream.clear();
           //               a_stream.str("");
@@ -2787,6 +2825,7 @@ void bssn_class::ParallelStep()
           GH->Regrid_Onelevel(lev - 1, Symmetry, BH_num, Porgbr, Porg0,
                               SynchList_cor, OldStateList, StateList, SynchList_pre,
                               fgt(PhysTime - dT_lev, StartTime, dT_levm1 / 2), ErrorMonitor);
+          for (int il = 0; il < GH->levels; il++) { sync_cache_pre[il].invalidate(); sync_cache_cor[il].invalidate(); sync_cache_rp_coarse[il].invalidate(); sync_cache_rp_fine[il].invalidate(); }
 
           //               a_stream.clear();
           //               a_stream.str("");
@@ -3158,21 +3197,7 @@ void bssn_class::Step(int lev, int YN)
     }
     Pp = Pp->next;
   }
-  // check error information
-  {
-    int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-  if (ERROR)
-  {
-    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
+  // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
 
 #ifdef WithShell
   // evolve Shell Patches
@@ -3190,9 +3215,9 @@ void bssn_class::Step(int lev, int YN)
         {
 #if (AGM == 0)
           f_enforce_ga(cg->shape,
-                       cg->fgfs[gxx0->sgfn], cg->fgfs[gxy0->sgfn], cg->fgfs[gxz0->sgfn], 
+                       cg->fgfs[gxx0->sgfn], cg->fgfs[gxy0->sgfn], cg->fgfs[gxz0->sgfn],
                        cg->fgfs[gyy0->sgfn], cg->fgfs[gyz0->sgfn], cg->fgfs[gzz0->sgfn],
-                       cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn], 
+                       cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn],
                        cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn]);
 #endif
 
@@ -3316,25 +3341,16 @@ void bssn_class::Step(int lev, int YN)
 #endif
   }
 
-  // check error information
+  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-
-  if (ERROR)
-  {
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
   }
 #endif
 
-  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
+  Parallel::AsyncSyncState async_pre;
+  Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
 
 #ifdef WithShell
   if (lev == 0)
@@ -3347,12 +3363,29 @@ void bssn_class::Step(int lev, int YN)
     {
       prev_clock = curr_clock;
       curr_clock = clock();
-      cout << " Shell stuff synchronization used " 
-           << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
+      cout << " Shell stuff synchronization used "
+           << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC)
            << " seconds! " << endl;
     }
   }
 #endif
+  Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
+
+#ifdef WithShell
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
+  if (ERROR)
+  {
+    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
+#endif
 
 #if (MAPBH == 0)
   // for black hole position
@@ -3528,24 +3561,7 @@ void bssn_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // check error information
-    {
-      int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-
-    if (ERROR)
-    {
-      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime 
-                                << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
+    // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
 
 #ifdef WithShell
     // evolve Shell Patches
@@ -3563,9 +3579,9 @@ void bssn_class::Step(int lev, int YN)
           {
 #if (AGM == 0)
             f_enforce_ga(cg->shape,
-                         cg->fgfs[gxx->sgfn], cg->fgfs[gxy->sgfn], cg->fgfs[gxz->sgfn], 
+                         cg->fgfs[gxx->sgfn], cg->fgfs[gxy->sgfn], cg->fgfs[gxz->sgfn],
                          cg->fgfs[gyy->sgfn], cg->fgfs[gyz->sgfn], cg->fgfs[gzz->sgfn],
-                         cg->fgfs[Axx->sgfn], cg->fgfs[Axy->sgfn], cg->fgfs[Axz->sgfn], 
+                         cg->fgfs[Axx->sgfn], cg->fgfs[Axy->sgfn], cg->fgfs[Axz->sgfn],
                          cg->fgfs[Ayy->sgfn], cg->fgfs[Ayz->sgfn], cg->fgfs[Azz->sgfn]);
 #elif (AGM == 1)
             if (iter_count == 3)
@@ -3685,26 +3701,16 @@ void bssn_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // check error information
+    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-    if (ERROR)
-    {
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" 
-                                << iter_count << " variables at t = " 
-                                << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
     }
 #endif
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::AsyncSyncState async_cor;
+    Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
 
 #ifdef WithShell
     if (lev == 0)
@@ -3717,12 +3723,31 @@ void bssn_class::Step(int lev, int YN)
       {
         prev_clock = curr_clock;
         curr_clock = clock();
-        cout << " Shell stuff synchronization used " 
-             << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
+        cout << " Shell stuff synchronization used "
+             << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC)
              << " seconds! " << endl;
       }
     }
 #endif
+    Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
+
+#ifdef WithShell
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
+    if (ERROR)
+    {
+      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
+#endif
 
 #if (MAPBH == 0)
     // for black hole position
@@ -4034,22 +4059,7 @@ void bssn_class::Step(int lev, int YN)
     }
     Pp = Pp->next;
   }
-  // check error information
-  {
-    int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-  if (ERROR)
-  {
-    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime 
-                              << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
+  // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
 
 #ifdef WithShell
   // evolve Shell Patches
@@ -4067,15 +4077,15 @@ void bssn_class::Step(int lev, int YN)
         {
 #if (AGM == 0)
           f_enforce_ga(cg->shape,
-                       cg->fgfs[gxx0->sgfn], cg->fgfs[gxy0->sgfn], cg->fgfs[gxz0->sgfn], 
+                       cg->fgfs[gxx0->sgfn], cg->fgfs[gxy0->sgfn], cg->fgfs[gxz0->sgfn],
                        cg->fgfs[gyy0->sgfn], cg->fgfs[gyz0->sgfn], cg->fgfs[gzz0->sgfn],
-                       cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn], 
+                       cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn],
                        cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn]);
 #endif
 
           if (f_compute_rhs_bssn_ss(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
-                                    cg->fgfs[fngfs + ShellPatch::gx], 
-                                    cg->fgfs[fngfs + ShellPatch::gy], 
+                                    cg->fgfs[fngfs + ShellPatch::gx],
+                                    cg->fgfs[fngfs + ShellPatch::gy],
                                     cg->fgfs[fngfs + ShellPatch::gz],
                                     cg->fgfs[fngfs + ShellPatch::drhodx], 
                                     cg->fgfs[fngfs + ShellPatch::drhody], 
@@ -4190,25 +4200,16 @@ void bssn_class::Step(int lev, int YN)
   }
 #endif
   }
-  // check error information
+  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-  if (ERROR)
-  {
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " 
-                              << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
   }
 #endif
 
-  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
+  Parallel::AsyncSyncState async_pre;
+  Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
 
 #ifdef WithShell
   if (lev == 0)
@@ -4221,9 +4222,27 @@ void bssn_class::Step(int lev, int YN)
     {
       prev_clock = curr_clock;
       curr_clock = clock();
-      cout << " Shell stuff synchronization used " 
-      << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
-      << " seconds! " << endl;
+      cout << " Shell stuff synchronization used "
+           << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC)
+           << " seconds! " << endl;
+    }
+  }
+#endif
+  Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
+
+#ifdef WithShell
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
+  if (ERROR)
+  {
+    Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime
+                              << ", lev = " << lev << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
     }
   }
 #endif
@@ -4386,23 +4405,7 @@ void bssn_class::Step(int lev, int YN)
       Pp = Pp->next;
     }
 
-    // check error information
-    {
-      int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-    if (ERROR)
-    {
-      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime 
-                                << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
+    // NOTE: error check deferred to after Shell Patch computation to reduce MPI_Allreduce calls
 
 #ifdef WithShell
     // evolve Shell Patches
@@ -4420,9 +4423,9 @@ void bssn_class::Step(int lev, int YN)
           {
 #if (AGM == 0)
             f_enforce_ga(cg->shape,
-                         cg->fgfs[gxx->sgfn], cg->fgfs[gxy->sgfn], cg->fgfs[gxz->sgfn], 
+                         cg->fgfs[gxx->sgfn], cg->fgfs[gxy->sgfn], cg->fgfs[gxz->sgfn],
                          cg->fgfs[gyy->sgfn], cg->fgfs[gyz->sgfn], cg->fgfs[gzz->sgfn],
-                         cg->fgfs[Axx->sgfn], cg->fgfs[Axy->sgfn], cg->fgfs[Axz->sgfn], 
+                         cg->fgfs[Axx->sgfn], cg->fgfs[Axy->sgfn], cg->fgfs[Axz->sgfn],
                          cg->fgfs[Ayy->sgfn], cg->fgfs[Ayz->sgfn], cg->fgfs[Azz->sgfn]);
 #elif (AGM == 1)
             if (iter_count == 3)
@@ -4542,25 +4545,16 @@ void bssn_class::Step(int lev, int YN)
         sPp = sPp->next;
       }
     }
-    // check error information
+    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-    if (ERROR)
-    {
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
     }
 #endif
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::AsyncSyncState async_cor;
+    Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
 
 #ifdef WithShell
     if (lev == 0)
@@ -4573,11 +4567,30 @@ void bssn_class::Step(int lev, int YN)
       {
         prev_clock = curr_clock;
         curr_clock = clock();
-        cout << " Shell stuff synchronization used " 
-             << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
+        cout << " Shell stuff synchronization used "
+             << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC)
              << " seconds! " << endl;
       }
     }
+#endif
+    Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
+
+#ifdef WithShell
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
+    if (ERROR)
+    {
+      Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
 #endif
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
@@ -4943,11 +4956,19 @@ void bssn_class::Step(int lev, int YN)
 
   //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Predictor rhs calculation");
 
-  // check error information
+  // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev]);
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev], &err_req);
   }
+
+  //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor sync");
+
+  Parallel::Sync_cached(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev]);
+
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
   if (ERROR)
   {
     Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
@@ -4959,10 +4980,6 @@ void bssn_class::Step(int lev, int YN)
     }
   }
 
-  //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor sync");
-
-  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
-
 #if (MAPBH == 0)
   // for black hole position
   if (BH_num > 0 && lev == GH->levels - 1)
@@ -5140,30 +5157,34 @@ void bssn_class::Step(int lev, int YN)
 
     //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector error check");
 
-    // check error information
+    // Non-blocking error reduction overlapped with Sync to hide Allreduce latency
+    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev]);
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, GH->Commlev[lev], &err_req_cor);
     }
+
+    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector sync");
+
+    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev]);
+
+    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Corrector sync");
+
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
     if (ERROR)
     {
       Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
       if (myrank == 0)
       {
         if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime 
+          ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime
                                 << ", lev = " << lev << endl;
         MPI_Abort(MPI_COMM_WORLD, 1);
       }
     }
 
-    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Corrector sync");
-
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
-
-    //    misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"after Corrector sync");
-
 #if (MAPBH == 0)
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
@@ -5447,21 +5468,11 @@ void bssn_class::SHStep()
 #if (PSTR == 1 || PSTR == 2)
 //   misc::tillherecheck(GH->Commlev[lev],GH->start_rank[lev],"before Predictor's error check");
 #endif
-  // check error information
+  // Non-blocking error reduction overlapped with Synch to hide Allreduce latency
+  MPI_Request err_req;
   {
     int erh = ERROR;
-    MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-  }
-
-  if (ERROR)
-  {
-    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
-    if (myrank == 0)
-    {
-      if (ErrorMonitor->outfile)
-        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
   }
 
   {
@@ -5473,12 +5484,25 @@ void bssn_class::SHStep()
     {
       prev_clock = curr_clock;
       curr_clock = clock();
-      cout << " Shell stuff synchronization used " 
-           << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
+      cout << " Shell stuff synchronization used "
+           << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC)
            << " seconds! " << endl;
     }
   }
 
+  // Complete non-blocking error reduction and check
+  MPI_Wait(&err_req, MPI_STATUS_IGNORE);
+  if (ERROR)
+  {
+    SH->Dump_Data(StateList, 0, PhysTime, dT_lev);
+    if (myrank == 0)
+    {
+      if (ErrorMonitor->outfile)
+        ErrorMonitor->outfile << "find NaN in state variables on Shell Patches at t = " << PhysTime << endl;
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
+
   // corrector
   for (iter_count = 1; iter_count < 4; iter_count++)
   {
@@ -5621,21 +5645,11 @@ void bssn_class::SHStep()
         sPp = sPp->next;
       }
     }
-    // check error information
+    // Non-blocking error reduction overlapped with Synch to hide Allreduce latency
+    MPI_Request err_req_cor;
     {
       int erh = ERROR;
-      MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
-    }
-    if (ERROR)
-    {
-      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
-      if (myrank == 0)
-      {
-        if (ErrorMonitor->outfile)
-          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count 
-                                << " variables at t = " << PhysTime << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
     }
 
     {
@@ -5647,12 +5661,26 @@ void bssn_class::SHStep()
       {
         prev_clock = curr_clock;
         curr_clock = clock();
-        cout << " Shell stuff synchronization used " 
-             << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
+        cout << " Shell stuff synchronization used "
+             << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC)
              << " seconds! " << endl;
       }
     }
 
+    // Complete non-blocking error reduction and check
+    MPI_Wait(&err_req_cor, MPI_STATUS_IGNORE);
+    if (ERROR)
+    {
+      SH->Dump_Data(SynchList_pre, 0, PhysTime, dT_lev);
+      if (myrank == 0)
+      {
+        if (ErrorMonitor->outfile)
+          ErrorMonitor->outfile << "find NaN on Shell Patches in RK4 substep#" << iter_count
+                                << " variables at t = " << PhysTime << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
+
     sPp = SH->PatL;
     while (sPp)
     {
@@ -5781,7 +5809,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 //       misc::tillherecheck(GH->Commlev[GH->mylev],GH->start_rank[GH->mylev],a_stream.str());
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (PSTR == 1 || PSTR == 2)
 //       a_stream.clear();
@@ -5791,21 +5819,11 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, GH->bdsul[lev], Symmetry);
@@ -5842,7 +5860,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 //       misc::tillherecheck(GH->Commlev[GH->mylev],GH->start_rank[GH->mylev],a_stream.str());
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SL, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SL, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (PSTR == 1 || PSTR == 2)
 //       a_stream.clear();
@@ -5852,21 +5870,11 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SL, SL, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SL, SL, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SL,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->bdsul[lev], Symmetry);
@@ -5880,7 +5888,7 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB,
 #endif
     }
 
-    Parallel::Sync(GH->PatL[lev], SL, Symmetry);
+    Parallel::Sync_cached(GH->PatL[lev], SL, Symmetry, sync_cache_rp_fine[lev]);
 
 #if (PSTR == 1 || PSTR == 2)
 //    a_stream.clear();
@@ -5938,24 +5946,14 @@ void bssn_class::RestrictProlong_aux(int lev, int YN, bool BB,
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SynchList_pre, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SL, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SL, GH->bdsul[lev], Symmetry);
@@ -5970,31 +5968,21 @@ void bssn_class::RestrictProlong_aux(int lev, int YN, bool BB,
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SL, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SL, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SL, SL, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SL, SL, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SL,SL,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SL, SL, GH->bdsul[lev], Symmetry);
 #endif
     }
 
-    Parallel::Sync(GH->PatL[lev], SL, Symmetry);
+    Parallel::Sync_cached(GH->PatL[lev], SL, Symmetry, sync_cache_rp_fine[lev]);
   }
 }
 
@@ -6045,24 +6033,14 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB)
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, SynchList_pre, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], SynchList_pre, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], SynchList_pre, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6079,31 +6057,21 @@ void bssn_class::RestrictProlong(int lev, int YN, bool BB)
       Parallel::Restrict_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, StateList, GH->rsul[lev], Symmetry);
 #endif
 
-      Parallel::Sync(GH->PatL[lev - 1], StateList, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], StateList, Symmetry, sync_cache_rp_coarse[lev]);
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],StateList,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, GH->bdsul[lev], Symmetry);
 #endif
     }
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_rp_fine[lev]);
   }
 }
 
@@ -6133,21 +6101,11 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
       }
 
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, SynchList_pre, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],SynchList_pre,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], SynchList_pre, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6156,21 +6114,11 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
     else // no time refinement levels and for all same time levels
     {
 #if (RPB == 0)
-      Ppc = GH->PatL[lev - 1];
-      while (Ppc)
-      {
-        Pp = GH->PatL[lev];
-        while (Pp)
-        {
 #if (MIXOUTB == 0)
-          Parallel::OutBdLow2Hi(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Hi(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
 #elif (MIXOUTB == 1)
-          Parallel::OutBdLow2Himix(Ppc->data, Pp->data, StateList, SynchList_cor, Symmetry);
+      Parallel::OutBdLow2Himix(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, Symmetry);
 #endif
-          Pp = Pp->next;
-        }
-        Ppc = Ppc->next;
-      }
 #elif (RPB == 1)
       //       Parallel::OutBdLow2Hi_bam(GH->PatL[lev-1],GH->PatL[lev],StateList,SynchList_cor,Symmetry);
       Parallel::OutBdLow2Hi_bam(GH->PatL[lev - 1], GH->PatL[lev], StateList, SynchList_cor, GH->bdsul[lev], Symmetry);
@@ -6186,10 +6134,10 @@ void bssn_class::ProlongRestrict(int lev, int YN, bool BB)
 #else
       Parallel::Restrict_after(GH->PatL[lev - 1], GH->PatL[lev], SynchList_cor, StateList, Symmetry);
 #endif
-      Parallel::Sync(GH->PatL[lev - 1], StateList, Symmetry);
+      Parallel::Sync_cached(GH->PatL[lev - 1], StateList, Symmetry, sync_cache_rp_coarse[lev]);
     }
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    Parallel::Sync_cached(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_rp_fine[lev]);
   }
 }
 #undef MIXOUTB

AMSS_NCKU_source/bssn_class.h

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

AMSS_NCKU_source/cgh.h

@@ -1,92 +1,107 @@
-
-#ifndef CGH_H
-#define CGH_H
-
-#include <mpi.h>
-#include "MyList.h"
-#include "MPatch.h"
-#include "macrodef.h"
-#include "monitor.h"
-#include "Parallel.h"
-
-class cgh
-{
-
-public:
-   int levels, movls, BH_num_in;
-   // information of boxes
-   int *grids;
-   double ***bbox;
-   int ***shape;
-   double ***handle;
-   double ***Porgls;
-   double *Lt;
-
-   // information of Patch list
-   MyList<Patch> **PatL;
-
-// information of OutBdLow2Hi point list and Restrict point list
-#if (RPB == 1)
-   MyList<Parallel::pointstru_bam> **bdsul, **rsul;
-#endif
-
-#if (PSTR == 1 || PSTR == 2 || PSTR == 3)
-   int mylev;
-   int *start_rank, *end_rank;
-   MPI_Comm *Commlev;
-#endif
-
-protected:
-   int ingfs, fngfs;
-   static constexpr double ratio = 0.75;
-   int trfls;
-
-public:
-   cgh(int ingfsi, int fngfsi, int Symmetry, char *filename, int checkrun, monitor *ErrorMonitor);
-
-   ~cgh();
-
-   void compose_cgh(int nprocs);
-   void sethandle(monitor *ErrorMonitor);
-   void checkPatchList(MyList<Patch> *PatL, bool buflog);
-   void Regrid(int Symmetry, int BH_num, double **Porgbr, double **Porg0,
-               MyList<var> *OldList, MyList<var> *StateList,
-               MyList<var> *FutureList, MyList<var> *tmList, bool BB,
-               monitor *ErrorMonitor);
-   void Regrid_fake(int Symmetry, int BH_num, double **Porgbr, double **Porg0,
-                    MyList<var> *OldList, MyList<var> *StateList,
-                    MyList<var> *FutureList, MyList<var> *tmList, bool BB,
-                    monitor *ErrorMonitor);
-   void recompose_cgh(int nprocs, bool *lev_flag,
-                      MyList<var> *OldList, MyList<var> *StateList,
-                      MyList<var> *FutureList, MyList<var> *tmList,
-                      int Symmetry, bool BB);
-   void recompose_cgh_fake(int nprocs, bool *lev_flag,
-                           MyList<var> *OldList, MyList<var> *StateList,
-                           MyList<var> *FutureList, MyList<var> *tmList,
-                           int Symmetry, bool BB);
-   void read_bbox(int Symmetry, char *filename);
-   MyList<Patch> *construct_patchlist(int lev, int Symmetry);
-   bool Interp_One_Point(MyList<var> *VarList,
-                         double *XX, /*input global Cartesian coordinate*/
-                         double *Shellf, int Symmetry);
-   void recompose_cgh_Onelevel(int nprocs, int lev,
-                               MyList<var> *OldList, MyList<var> *StateList,
-                               MyList<var> *FutureList, MyList<var> *tmList,
-                               int Symmetry, bool BB);
-   void Regrid_Onelevel(int lev, int Symmetry, int BH_num, double **Porgbr, double **Porg0,
-                        MyList<var> *OldList, MyList<var> *StateList,
-                        MyList<var> *FutureList, MyList<var> *tmList, bool BB,
-                        monitor *ErrorMonitor);
-   void Regrid_Onelevel_aux(int lev, int Symmetry, int BH_num, double **Porgbr, double **Porg0,
-                            MyList<var> *OldList, MyList<var> *StateList,
-                            MyList<var> *FutureList, MyList<var> *tmList, bool BB,
-                            monitor *ErrorMonitor);
-   void settrfls(const int lev);
-
-#if (PSTR == 1 || PSTR == 2 || PSTR == 3)
-   void construct_mylev(int nprocs);
-#endif
-};
-
-#endif /* CGH_H */
+
+#ifndef CGH_H
+#define CGH_H
+
+#include <mpi.h>
+#include "MyList.h"
+#include "MPatch.h"
+#include "macrodef.h"
+#include "monitor.h"
+#include "Parallel.h"
+
+class cgh
+{
+
+public:
+   int levels, movls, BH_num_in;
+   // information of boxes
+   int *grids;
+   double ***bbox;
+   int ***shape;
+   double ***handle;
+   double ***Porgls;
+   double *Lt;
+
+   // information of Patch list
+   MyList<Patch> **PatL;
+
+// information of OutBdLow2Hi point list and Restrict point list
+#if (RPB == 1)
+   MyList<Parallel::pointstru_bam> **bdsul, **rsul;
+#endif
+
+#if (PSTR == 1 || PSTR == 2 || PSTR == 3)
+   int mylev;
+   int *start_rank, *end_rank;
+   MPI_Comm *Commlev;
+#endif
+
+protected:
+   int ingfs, fngfs;
+   static constexpr double ratio = 0.75;
+   int trfls;
+
+public:
+   cgh(int ingfsi, int fngfsi, int Symmetry, char *filename, int checkrun, monitor *ErrorMonitor);
+
+   ~cgh();
+
+   void compose_cgh(int nprocs);
+   void sethandle(monitor *ErrorMonitor);
+   void checkPatchList(MyList<Patch> *PatL, bool buflog);
+   void Regrid(int Symmetry, int BH_num, double **Porgbr, double **Porg0,
+               MyList<var> *OldList, MyList<var> *StateList,
+               MyList<var> *FutureList, MyList<var> *tmList, bool BB,
+               monitor *ErrorMonitor);
+   void Regrid_fake(int Symmetry, int BH_num, double **Porgbr, double **Porg0,
+                    MyList<var> *OldList, MyList<var> *StateList,
+                    MyList<var> *FutureList, MyList<var> *tmList, bool BB,
+                    monitor *ErrorMonitor);
+   void recompose_cgh(int nprocs, bool *lev_flag,
+                      MyList<var> *OldList, MyList<var> *StateList,
+                      MyList<var> *FutureList, MyList<var> *tmList,
+                      int Symmetry, bool BB);
+   void recompose_cgh_fake(int nprocs, bool *lev_flag,
+                           MyList<var> *OldList, MyList<var> *StateList,
+                           MyList<var> *FutureList, MyList<var> *tmList,
+                           int Symmetry, bool BB);
+   void read_bbox(int Symmetry, char *filename);
+   MyList<Patch> *construct_patchlist(int lev, int Symmetry);
+   bool Interp_One_Point(MyList<var> *VarList,
+                         double *XX, /*input global Cartesian coordinate*/
+                         double *Shellf, int Symmetry);
+   void recompose_cgh_Onelevel(int nprocs, int lev,
+                               MyList<var> *OldList, MyList<var> *StateList,
+                               MyList<var> *FutureList, MyList<var> *tmList,
+                               int Symmetry, bool BB);
+   void Regrid_Onelevel(int lev, int Symmetry, int BH_num, double **Porgbr, double **Porg0,
+                        MyList<var> *OldList, MyList<var> *StateList,
+                        MyList<var> *FutureList, MyList<var> *tmList, bool BB,
+                        monitor *ErrorMonitor);
+   void Regrid_Onelevel_aux(int lev, int Symmetry, int BH_num, double **Porgbr, double **Porg0,
+                            MyList<var> *OldList, MyList<var> *StateList,
+                            MyList<var> *FutureList, MyList<var> *tmList, bool BB,
+                            monitor *ErrorMonitor);
+   void settrfls(const int lev);
+
+#if (PSTR == 1 || PSTR == 2 || PSTR == 3)
+   void construct_mylev(int nprocs);
+#endif
+
+   // Load balancing support
+   bool enable_load_balance;         // Enable load balancing
+   int load_balance_check_interval;  // Check interval (in time steps)
+   int current_time_step;            // Current time step counter
+   double *rank_interp_times;        // Store interpolation times for each rank
+   int *heavy_ranks;                 // Store heavy rank numbers
+   int num_heavy_ranks;              // Number of heavy ranks
+
+   void init_load_balance(int nprocs);
+   void update_interp_time(int rank, double time);
+   bool check_and_rebalance(int nprocs, int lev,
+                           MyList<var> *OldList, MyList<var> *StateList,
+                           MyList<var> *FutureList, MyList<var> *tmList,
+                           int Symmetry, bool BB);
+};
+
+#endif /* CGH_H */

AMSS_NCKU_source/diff_new.f90

@@ -69,10 +69,12 @@
   fy = ZEO
   fz = ZEO
 
+!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
+!DIR$ UNROLL PARTIAL(4)
   do k=1,ex(3)-1
   do j=1,ex(2)-1
   do i=1,ex(1)-1
-! x direction   
+! x direction
         if(i+1 <= imax .and. i-1 >= imin)then
 !
 !              - f(i-1) + f(i+1)
@@ -371,6 +373,8 @@
   fxz = ZEO
   fyz = ZEO
 
+!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
+!DIR$ UNROLL PARTIAL(4)
   do k=1,ex(3)-1
   do j=1,ex(2)-1
   do i=1,ex(1)-1

AMSS_NCKU_source/enforce_algebra.f90

@@ -18,49 +18,61 @@
   real*8, dimension(ex(1),ex(2),ex(3)), intent(inout) :: Ayy,Ayz,Azz
 
 !~~~~~~~> Local variable:
-  
-  real*8, dimension(ex(1),ex(2),ex(3)) :: trA,detg
-  real*8, dimension(ex(1),ex(2),ex(3)) :: gxx,gyy,gzz 
-  real*8, dimension(ex(1),ex(2),ex(3)) :: gupxx,gupxy,gupxz,gupyy,gupyz,gupzz
+
+  integer :: i,j,k
+  real*8 :: lgxx,lgyy,lgzz,ldetg
+  real*8 :: lgupxx,lgupxy,lgupxz,lgupyy,lgupyz,lgupzz
+  real*8 :: ltrA,lscale
   real*8, parameter :: F1o3 = 1.D0 / 3.D0, ONE = 1.D0, TWO = 2.D0
 
 !~~~~~~>
 
-  gxx = dxx + ONE
-  gyy = dyy + ONE
-  gzz = dzz + ONE
+  do k=1,ex(3)
+  do j=1,ex(2)
+  do i=1,ex(1)
 
-  detg =  gxx * gyy * gzz + gxy * gyz * gxz + gxz * gxy * gyz - &
-          gxz * gyy * gxz - gxy * gxy * gzz - gxx * gyz * gyz
-  gupxx =   ( gyy * gzz - gyz * gyz ) / detg
-  gupxy = - ( gxy * gzz - gyz * gxz ) / detg
-  gupxz =   ( gxy * gyz - gyy * gxz ) / detg
-  gupyy =   ( gxx * gzz - gxz * gxz ) / detg
-  gupyz = - ( gxx * gyz - gxy * gxz ) / detg
-  gupzz =   ( gxx * gyy - gxy * gxy ) / detg
+    lgxx = dxx(i,j,k) + ONE
+    lgyy = dyy(i,j,k) + ONE
+    lgzz = dzz(i,j,k) + ONE
 
-  trA =         gupxx * Axx + gupyy * Ayy + gupzz * Azz &
-       + TWO * (gupxy * Axy + gupxz * Axz + gupyz * Ayz)
+    ldetg =  lgxx * lgyy * lgzz &
+           + gxy(i,j,k) * gyz(i,j,k) * gxz(i,j,k) &
+           + gxz(i,j,k) * gxy(i,j,k) * gyz(i,j,k) &
+           - gxz(i,j,k) * lgyy * gxz(i,j,k) &
+           - gxy(i,j,k) * gxy(i,j,k) * lgzz &
+           - lgxx * gyz(i,j,k) * gyz(i,j,k)
 
-  Axx = Axx - F1o3 * gxx * trA
-  Axy = Axy - F1o3 * gxy * trA
-  Axz = Axz - F1o3 * gxz * trA
-  Ayy = Ayy - F1o3 * gyy * trA
-  Ayz = Ayz - F1o3 * gyz * trA
-  Azz = Azz - F1o3 * gzz * trA
+    lgupxx =   ( lgyy * lgzz - gyz(i,j,k) * gyz(i,j,k) ) / ldetg
+    lgupxy = - ( gxy(i,j,k) * lgzz - gyz(i,j,k) * gxz(i,j,k) ) / ldetg
+    lgupxz =   ( gxy(i,j,k) * gyz(i,j,k) - lgyy * gxz(i,j,k) ) / ldetg
+    lgupyy =   ( lgxx * lgzz - gxz(i,j,k) * gxz(i,j,k) ) / ldetg
+    lgupyz = - ( lgxx * gyz(i,j,k) - gxy(i,j,k) * gxz(i,j,k) ) / ldetg
+    lgupzz =   ( lgxx * lgyy - gxy(i,j,k) * gxy(i,j,k) ) / ldetg
 
-  detg = ONE / ( detg ** F1o3 ) 
-  
-  gxx = gxx * detg
-  gxy = gxy * detg
-  gxz = gxz * detg
-  gyy = gyy * detg
-  gyz = gyz * detg
-  gzz = gzz * detg
+    ltrA =         lgupxx * Axx(i,j,k) + lgupyy * Ayy(i,j,k) &
+                 + lgupzz * Azz(i,j,k) &
+         + TWO * (lgupxy * Axy(i,j,k) + lgupxz * Axz(i,j,k) &
+                 + lgupyz * Ayz(i,j,k))
 
-  dxx = gxx - ONE
-  dyy = gyy - ONE
-  dzz = gzz - ONE
+    Axx(i,j,k) = Axx(i,j,k) - F1o3 * lgxx * ltrA
+    Axy(i,j,k) = Axy(i,j,k) - F1o3 * gxy(i,j,k) * ltrA
+    Axz(i,j,k) = Axz(i,j,k) - F1o3 * gxz(i,j,k) * ltrA
+    Ayy(i,j,k) = Ayy(i,j,k) - F1o3 * lgyy * ltrA
+    Ayz(i,j,k) = Ayz(i,j,k) - F1o3 * gyz(i,j,k) * ltrA
+    Azz(i,j,k) = Azz(i,j,k) - F1o3 * lgzz * ltrA
+
+    lscale = ONE / ( ldetg ** F1o3 )
+
+    dxx(i,j,k) = lgxx * lscale - ONE
+    gxy(i,j,k) = gxy(i,j,k) * lscale
+    gxz(i,j,k) = gxz(i,j,k) * lscale
+    dyy(i,j,k) = lgyy * lscale - ONE
+    gyz(i,j,k) = gyz(i,j,k) * lscale
+    dzz(i,j,k) = lgzz * lscale - ONE
+
+  enddo
+  enddo
+  enddo
 
   return
 
@@ -82,51 +94,71 @@
   real*8, dimension(ex(1),ex(2),ex(3)), intent(inout) :: Ayy,Ayz,Azz
 
 !~~~~~~~> Local variable:
-  
-  real*8, dimension(ex(1),ex(2),ex(3)) :: trA
-  real*8, dimension(ex(1),ex(2),ex(3)) :: gxx,gyy,gzz 
-  real*8, dimension(ex(1),ex(2),ex(3)) :: gupxx,gupxy,gupxz,gupyy,gupyz,gupzz
+
+  integer :: i,j,k
+  real*8 :: lgxx,lgyy,lgzz,lscale
+  real*8 :: lgxy,lgxz,lgyz
+  real*8 :: lgupxx,lgupxy,lgupxz,lgupyy,lgupyz,lgupzz
+  real*8 :: ltrA
   real*8, parameter :: F1o3 = 1.D0 / 3.D0, ONE = 1.D0, TWO = 2.D0
 
 !~~~~~~>
 
-  gxx = dxx + ONE
-  gyy = dyy + ONE
-  gzz = dzz + ONE
-! for g
-  gupzz =  gxx * gyy * gzz + gxy * gyz * gxz + gxz * gxy * gyz - &
-           gxz * gyy * gxz - gxy * gxy * gzz - gxx * gyz * gyz
+  do k=1,ex(3)
+  do j=1,ex(2)
+  do i=1,ex(1)
 
-  gupzz = ONE / ( gupzz ** F1o3 ) 
-  
-  gxx = gxx * gupzz
-  gxy = gxy * gupzz
-  gxz = gxz * gupzz
-  gyy = gyy * gupzz
-  gyz = gyz * gupzz
-  gzz = gzz * gupzz
+! for g: normalize determinant first
+    lgxx = dxx(i,j,k) + ONE
+    lgyy = dyy(i,j,k) + ONE
+    lgzz = dzz(i,j,k) + ONE
+    lgxy = gxy(i,j,k)
+    lgxz = gxz(i,j,k)
+    lgyz = gyz(i,j,k)
 
-  dxx = gxx - ONE
-  dyy = gyy - ONE
-  dzz = gzz - ONE
-! for A  
+    lscale =  lgxx * lgyy * lgzz + lgxy * lgyz * lgxz &
+            + lgxz * lgxy * lgyz - lgxz * lgyy * lgxz &
+            - lgxy * lgxy * lgzz - lgxx * lgyz * lgyz
 
-  gupxx =   ( gyy * gzz - gyz * gyz )
-  gupxy = - ( gxy * gzz - gyz * gxz )
-  gupxz =   ( gxy * gyz - gyy * gxz )
-  gupyy =   ( gxx * gzz - gxz * gxz )
-  gupyz = - ( gxx * gyz - gxy * gxz )
-  gupzz =   ( gxx * gyy - gxy * gxy )
+    lscale = ONE / ( lscale ** F1o3 )
 
-  trA =         gupxx * Axx + gupyy * Ayy + gupzz * Azz &
-       + TWO * (gupxy * Axy + gupxz * Axz + gupyz * Ayz)
+    lgxx = lgxx * lscale
+    lgxy = lgxy * lscale
+    lgxz = lgxz * lscale
+    lgyy = lgyy * lscale
+    lgyz = lgyz * lscale
+    lgzz = lgzz * lscale
 
-  Axx = Axx - F1o3 * gxx * trA
-  Axy = Axy - F1o3 * gxy * trA
-  Axz = Axz - F1o3 * gxz * trA
-  Ayy = Ayy - F1o3 * gyy * trA
-  Ayz = Ayz - F1o3 * gyz * trA
-  Azz = Azz - F1o3 * gzz * trA
+    dxx(i,j,k) = lgxx - ONE
+    gxy(i,j,k) = lgxy
+    gxz(i,j,k) = lgxz
+    dyy(i,j,k) = lgyy - ONE
+    gyz(i,j,k) = lgyz
+    dzz(i,j,k) = lgzz - ONE
+
+! for A: trace-free using normalized metric (det=1, no division needed)
+    lgupxx =   ( lgyy * lgzz - lgyz * lgyz )
+    lgupxy = - ( lgxy * lgzz - lgyz * lgxz )
+    lgupxz =   ( lgxy * lgyz - lgyy * lgxz )
+    lgupyy =   ( lgxx * lgzz - lgxz * lgxz )
+    lgupyz = - ( lgxx * lgyz - lgxy * lgxz )
+    lgupzz =   ( lgxx * lgyy - lgxy * lgxy )
+
+    ltrA =         lgupxx * Axx(i,j,k) + lgupyy * Ayy(i,j,k) &
+                 + lgupzz * Azz(i,j,k) &
+         + TWO * (lgupxy * Axy(i,j,k) + lgupxz * Axz(i,j,k) &
+                 + lgupyz * Ayz(i,j,k))
+
+    Axx(i,j,k) = Axx(i,j,k) - F1o3 * lgxx * ltrA
+    Axy(i,j,k) = Axy(i,j,k) - F1o3 * lgxy * ltrA
+    Axz(i,j,k) = Axz(i,j,k) - F1o3 * lgxz * ltrA
+    Ayy(i,j,k) = Ayy(i,j,k) - F1o3 * lgyy * ltrA
+    Ayz(i,j,k) = Ayz(i,j,k) - F1o3 * lgyz * ltrA
+    Azz(i,j,k) = Azz(i,j,k) - F1o3 * lgzz * ltrA
+
+  enddo
+  enddo
+  enddo
 
   return
 

AMSS_NCKU_source/fmisc.f90

@@ -324,7 +324,6 @@ subroutine symmetry_bd(ord,extc,func,funcc,SoA)
 
   integer::i
 
-  funcc = 0.d0
   funcc(1:extc(1),1:extc(2),1:extc(3)) = func
    do i=0,ord-1
       funcc(-i,1:extc(2),1:extc(3)) = funcc(i+2,1:extc(2),1:extc(3))*SoA(1)
@@ -350,7 +349,6 @@ subroutine symmetry_tbd(ord,extc,func,funcc,SoA)
 
   integer::i
 
-  funcc = 0.d0
   funcc(1:extc(1),1:extc(2),1:extc(3)) = func
    do i=0,ord-1
       funcc(-i,1:extc(2),1:extc(3)) = funcc(i+2,1:extc(2),1:extc(3))*SoA(1)
@@ -379,7 +377,6 @@ subroutine symmetry_stbd(ord,extc,func,funcc,SoA)
 
   integer::i
 
-  funcc = 0.d0
   funcc(1:extc(1),1:extc(2),1:extc(3)) = func
    do i=0,ord-1
       funcc(-i,1:extc(2),1:extc(3)) = funcc(i+2,1:extc(2),1:extc(3))*SoA(1)
@@ -886,14 +883,17 @@ subroutine symmetry_bd(ord,extc,func,funcc,SoA)
 
   integer::i
 
-  funcc = 0.d0
+!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
   funcc(1:extc(1),1:extc(2),1:extc(3)) = func
+!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
    do i=0,ord-1
       funcc(-i,1:extc(2),1:extc(3)) = funcc(i+1,1:extc(2),1:extc(3))*SoA(1)
    enddo
+!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
    do i=0,ord-1
       funcc(:,-i,1:extc(3)) = funcc(:,i+1,1:extc(3))*SoA(2)
    enddo
+!DIR$ SIMD VECTORLENGTHFOR(KNOWN_INTEGER=8)
    do i=0,ord-1
       funcc(:,:,-i) = funcc(:,:,i+1)*SoA(3)
    enddo
@@ -912,7 +912,6 @@ subroutine symmetry_tbd(ord,extc,func,funcc,SoA)
 
   integer::i
 
-  funcc = 0.d0
   funcc(1:extc(1),1:extc(2),1:extc(3)) = func
    do i=0,ord-1
       funcc(-i,1:extc(2),1:extc(3)) = funcc(i+1,1:extc(2),1:extc(3))*SoA(1)
@@ -941,7 +940,6 @@ subroutine symmetry_stbd(ord,extc,func,funcc,SoA)
 
   integer::i
 
-  funcc = 0.d0
   funcc(1:extc(1),1:extc(2),1:extc(3)) = func
    do i=0,ord-1
       funcc(-i,1:extc(2),1:extc(3)) = funcc(i+1,1:extc(2),1:extc(3))*SoA(1)
@@ -1118,64 +1116,66 @@ end subroutine d2dump
 ! Lagrangian polynomial interpolation
 !------------------------------------------------------------------------------
 
-  subroutine polint(xa,ya,x,y,dy,ordn)
-
+!DIR$ ATTRIBUTES FORCEINLINE :: polint
+  subroutine polint(xa, ya, x, y, dy, ordn)
   implicit none
 
-!~~~~~~> Input Parameter:
-  integer,intent(in) :: ordn
-  real*8, dimension(ordn), intent(in) :: xa,ya
+  integer, intent(in) :: ordn
+  real*8, dimension(ordn), intent(in) :: xa, ya
   real*8, intent(in) :: x
-  real*8, intent(out) :: y,dy
+  real*8, intent(out) :: y, dy
 
-!~~~~~~> Other parameter:
+  integer :: i, m, ns, n_m
+  real*8, dimension(ordn) :: c, d, ho
+  real*8 :: dif, dift, hp, h, den_val
 
-  integer :: m,n,ns
-  real*8, dimension(ordn) :: c,d,den,ho
-  real*8 :: dif,dift
+  c = ya
+  d = ya
+  ho = xa - x
 
-!~~~~~~>
+  ns = 1
+  dif = abs(x - xa(1))
 
-  n=ordn
-  m=ordn
-
-  c=ya
-  d=ya
-  ho=xa-x
-
-  ns=1
-  dif=abs(x-xa(1))
-  do m=1,n
-   dift=abs(x-xa(m))
-   if(dift < dif) then
-    ns=m
-    dif=dift
-   end if
+  do i = 2, ordn
+    dift = abs(x - xa(i))
+    if (dift < dif) then
+      ns = i
+      dif = dift
+    end if
   end do
 
-  y=ya(ns)
-  ns=ns-1
-  do m=1,n-1
-    den(1:n-m)=ho(1:n-m)-ho(1+m:n)
-    if (any(den(1:n-m) == 0.0))then
-      write(*,*) 'failure in polint for point',x
-      write(*,*) 'with input points: ',xa
-      stop
-    endif
-    den(1:n-m)=(c(2:n-m+1)-d(1:n-m))/den(1:n-m)
-    d(1:n-m)=ho(1+m:n)*den(1:n-m)
-    c(1:n-m)=ho(1:n-m)*den(1:n-m)
-    if (2*ns < n-m) then
-      dy=c(ns+1)
+  y = ya(ns)
+  ns = ns - 1
+
+  do m = 1, ordn - 1
+    n_m = ordn - m
+    do i = 1, n_m
+      hp = ho(i)
+      h  = ho(i+m)
+      den_val = hp - h
+
+      if (den_val == 0.0d0) then
+        write(*,*) 'failure in polint for point',x
+        write(*,*) 'with input points: ',xa
+        stop
+      end if
+
+      den_val = (c(i+1) - d(i)) / den_val
+
+      d(i) = h * den_val
+      c(i) = hp * den_val
+    end do
+
+    if (2 * ns < n_m) then
+      dy = c(ns + 1)
     else
-      dy=d(ns)
-      ns=ns-1
+      dy = d(ns)
+      ns = ns - 1
     end if
-    y=y+dy
+    y = y + dy
   end do
 
   return
-
   end subroutine polint
 !------------------------------------------------------------------------------
 !
@@ -1183,35 +1183,37 @@ end subroutine d2dump
 !
 !------------------------------------------------------------------------------
   subroutine polin2(x1a,x2a,ya,x1,x2,y,dy,ordn)
-
   implicit none
 
-!~~~~~~> Input parameters:
   integer,intent(in) :: ordn
   real*8, dimension(1:ordn), intent(in) :: x1a,x2a
   real*8, dimension(1:ordn,1:ordn), intent(in) :: ya
   real*8, intent(in) :: x1,x2
   real*8, intent(out) :: y,dy
 
-!~~~~~~> Other parameters:
-
+#ifdef POLINT_LEGACY_ORDER
   integer  :: i,m
   real*8, dimension(ordn) :: ymtmp
   real*8, dimension(ordn) :: yntmp
 
   m=size(x1a)
-  
   do i=1,m
-
     yntmp=ya(i,:)
     call polint(x2a,yntmp,x2,ymtmp(i),dy,ordn)
-
   end do
-
   call polint(x1a,ymtmp,x1,y,dy,ordn)
+#else
+  integer  :: j
+  real*8, dimension(ordn) :: ymtmp
+  real*8 :: dy_temp
+
+  do j=1,ordn
+    call polint(x1a, ya(:,j), x1, ymtmp(j), dy_temp, ordn)
+  end do
+  call polint(x2a, ymtmp, x2, y, dy, ordn)
+#endif
 
   return
-
   end subroutine polin2
 !------------------------------------------------------------------------------
 !
@@ -1219,18 +1221,15 @@ end subroutine d2dump
 !
 !------------------------------------------------------------------------------
   subroutine polin3(x1a,x2a,x3a,ya,x1,x2,x3,y,dy,ordn)
-
   implicit none
 
-!~~~~~~> Input parameters:
   integer,intent(in) :: ordn
   real*8, dimension(1:ordn), intent(in) :: x1a,x2a,x3a
   real*8, dimension(1:ordn,1:ordn,1:ordn), intent(in) :: ya
   real*8, intent(in) :: x1,x2,x3
   real*8, intent(out) :: y,dy
 
-!~~~~~~> Other parameters:
-
+#ifdef POLINT_LEGACY_ORDER
   integer  :: i,j,m,n
   real*8, dimension(ordn,ordn) :: yatmp
   real*8, dimension(ordn) :: ymtmp
@@ -1239,24 +1238,33 @@ end subroutine d2dump
 
   m=size(x1a)
   n=size(x2a)
-  
   do i=1,m
    do j=1,n
-
     yqtmp=ya(i,j,:)
     call polint(x3a,yqtmp,x3,yatmp(i,j),dy,ordn)
-
    end do
-
     yntmp=yatmp(i,:)
     call polint(x2a,yntmp,x2,ymtmp(i),dy,ordn)
-
   end do
-
   call polint(x1a,ymtmp,x1,y,dy,ordn)
+#else
+  integer  :: j, k
+  real*8, dimension(ordn,ordn) :: yatmp
+  real*8, dimension(ordn) :: ymtmp
+  real*8 :: dy_temp
+
+  do k=1,ordn
+    do j=1,ordn
+      call polint(x1a, ya(:,j,k), x1, yatmp(j,k), dy_temp, ordn)
+    end do
+  end do
+  do k=1,ordn
+    call polint(x2a, yatmp(:,k), x2, ymtmp(k), dy_temp, ordn)
+  end do
+  call polint(x3a, ymtmp, x3, y, dy, ordn)
+#endif
 
   return
-
   end subroutine polin3
 !--------------------------------------------------------------------------------------
 ! calculate L2norm

AMSS_NCKU_source/kodiss.f90

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

AMSS_NCKU_source/lopsidediff.f90

@@ -487,6 +487,201 @@ subroutine lopsided(ex,X,Y,Z,f,f_rhs,Sfx,Sfy,Sfz,Symmetry,SoA)
 
   end subroutine lopsided
 
+!-----------------------------------------------------------------------------
+! Combined advection (lopsided) + Kreiss-Oliger dissipation (kodis)
+! Shares the symmetry_bd buffer fh, eliminating one full-grid copy per call.
+! Mathematically identical to calling lopsided then kodis separately.
+!-----------------------------------------------------------------------------
+subroutine lopsided_kodis(ex,X,Y,Z,f,f_rhs,Sfx,Sfy,Sfz,Symmetry,SoA,eps)
+  implicit none
+
+!~~~~~~> Input parameters:
+
+  integer, intent(in)  :: ex(1:3),Symmetry
+  real*8,  intent(in)  :: X(1:ex(1)),Y(1:ex(2)),Z(1:ex(3))
+  real*8,dimension(ex(1),ex(2),ex(3)),intent(in)   :: f,Sfx,Sfy,Sfz
+
+  real*8,dimension(ex(1),ex(2),ex(3)),intent(inout):: f_rhs
+  real*8,dimension(3),intent(in) ::SoA
+  real*8,intent(in) :: eps
+
+!~~~~~~> local variables:
+! note index -2,-1,0, so we have 3 extra points
+  real*8,dimension(-2:ex(1),-2:ex(2),-2:ex(3))   :: fh
+  integer :: imin,jmin,kmin,imax,jmax,kmax,i,j,k
+  real*8 :: dX,dY,dZ
+  real*8 :: d12dx,d12dy,d12dz,d2dx,d2dy,d2dz
+  real*8,  parameter :: ZEO=0.d0,ONE=1.d0, F3=3.d0
+  real*8,  parameter :: TWO=2.d0,F6=6.0d0,F18=1.8d1
+  real*8,  parameter :: F12=1.2d1, F10=1.d1,EIT=8.d0
+  integer, parameter :: NO_SYMM = 0, EQ_SYMM = 1, OCTANT = 2
+! kodis parameters
+  real*8, parameter :: SIX=6.d0,FIT=1.5d1,TWT=2.d1
+  real*8, parameter :: cof=6.4d1   ! 2^6
+
+  dX = X(2)-X(1)
+  dY = Y(2)-Y(1)
+  dZ = Z(2)-Z(1)
+
+  d12dx = ONE/F12/dX
+  d12dy = ONE/F12/dY
+  d12dz = ONE/F12/dZ
+
+  d2dx = ONE/TWO/dX
+  d2dy = ONE/TWO/dY
+  d2dz = ONE/TWO/dZ
+
+  imax = ex(1)
+  jmax = ex(2)
+  kmax = ex(3)
+
+  imin = 1
+  jmin = 1
+  kmin = 1
+  if(Symmetry > NO_SYMM .and. dabs(Z(1)) < dZ) kmin = -2
+  if(Symmetry > EQ_SYMM .and. dabs(X(1)) < dX) imin = -2
+  if(Symmetry > EQ_SYMM .and. dabs(Y(1)) < dY) jmin = -2
+
+! Single symmetry_bd call shared by both advection and dissipation
+  call symmetry_bd(3,ex,f,fh,SoA)
+
+! ---- Advection (lopsided) loop ----
+! upper bound set ex-1 only for efficiency, 
+! the loop body will set ex 0 also
+  do k=1,ex(3)-1
+  do j=1,ex(2)-1
+  do i=1,ex(1)-1
+! x direction   
+    if(Sfx(i,j,k) > ZEO)then
+      if(i+3 <= imax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                   &
+                  Sfx(i,j,k)*d12dx*(-F3*fh(i-1,j,k)-F10*fh(i,j,k)+F18*fh(i+1,j,k) &
+                                    -F6*fh(i+2,j,k)+    fh(i+3,j,k))
+     elseif(i+2 <= imax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                           &
+                  Sfx(i,j,k)*d12dx*(fh(i-2,j,k)-EIT*fh(i-1,j,k)+EIT*fh(i+1,j,k)-fh(i+2,j,k))
+
+     elseif(i+1 <= imax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)-                                                   &
+                  Sfx(i,j,k)*d12dx*(-F3*fh(i+1,j,k)-F10*fh(i,j,k)+F18*fh(i-1,j,k) &
+                                    -F6*fh(i-2,j,k)+    fh(i-3,j,k))
+     endif
+   elseif(Sfx(i,j,k) < ZEO)then
+      if(i-3 >= imin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)-                                                   &
+                  Sfx(i,j,k)*d12dx*(-F3*fh(i+1,j,k)-F10*fh(i,j,k)+F18*fh(i-1,j,k) &
+                                    -F6*fh(i-2,j,k)+    fh(i-3,j,k))
+     elseif(i-2 >= imin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                           &
+                  Sfx(i,j,k)*d12dx*(fh(i-2,j,k)-EIT*fh(i-1,j,k)+EIT*fh(i+1,j,k)-fh(i+2,j,k))
+
+     elseif(i-1 >= imin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                   &
+                  Sfx(i,j,k)*d12dx*(-F3*fh(i-1,j,k)-F10*fh(i,j,k)+F18*fh(i+1,j,k) &
+                                    -F6*fh(i+2,j,k)+    fh(i+3,j,k))
+     endif
+   endif
+
+! y direction   
+    if(Sfy(i,j,k) > ZEO)then
+      if(j+3 <= jmax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                   &
+                  Sfy(i,j,k)*d12dy*(-F3*fh(i,j-1,k)-F10*fh(i,j,k)+F18*fh(i,j+1,k) &
+                                    -F6*fh(i,j+2,k)+    fh(i,j+3,k))
+     elseif(j+2 <= jmax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                           &
+                  Sfy(i,j,k)*d12dy*(fh(i,j-2,k)-EIT*fh(i,j-1,k)+EIT*fh(i,j+1,k)-fh(i,j+2,k))
+
+     elseif(j+1 <= jmax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)-                                                   &
+                  Sfy(i,j,k)*d12dy*(-F3*fh(i,j+1,k)-F10*fh(i,j,k)+F18*fh(i,j-1,k) &
+                                    -F6*fh(i,j-2,k)+    fh(i,j-3,k))
+     endif
+   elseif(Sfy(i,j,k) < ZEO)then
+      if(j-3 >= jmin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)-                                                   &
+                  Sfy(i,j,k)*d12dy*(-F3*fh(i,j+1,k)-F10*fh(i,j,k)+F18*fh(i,j-1,k) &
+                                    -F6*fh(i,j-2,k)+    fh(i,j-3,k))
+     elseif(j-2 >= jmin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                           &
+                  Sfy(i,j,k)*d12dy*(fh(i,j-2,k)-EIT*fh(i,j-1,k)+EIT*fh(i,j+1,k)-fh(i,j+2,k))
+
+     elseif(j-1 >= jmin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                   &
+                  Sfy(i,j,k)*d12dy*(-F3*fh(i,j-1,k)-F10*fh(i,j,k)+F18*fh(i,j+1,k) &
+                                    -F6*fh(i,j+2,k)+    fh(i,j+3,k))
+     endif
+   endif
+
+! z direction   
+    if(Sfz(i,j,k) > ZEO)then
+      if(k+3 <= kmax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                   &
+                  Sfz(i,j,k)*d12dz*(-F3*fh(i,j,k-1)-F10*fh(i,j,k)+F18*fh(i,j,k+1) &
+                                    -F6*fh(i,j,k+2)+    fh(i,j,k+3))
+     elseif(k+2 <= kmax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                           &
+                  Sfz(i,j,k)*d12dz*(fh(i,j,k-2)-EIT*fh(i,j,k-1)+EIT*fh(i,j,k+1)-fh(i,j,k+2))
+
+     elseif(k+1 <= kmax)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)-                                                   &
+                  Sfz(i,j,k)*d12dz*(-F3*fh(i,j,k+1)-F10*fh(i,j,k)+F18*fh(i,j,k-1) &
+                                    -F6*fh(i,j,k-2)+    fh(i,j,k-3))
+     endif
+   elseif(Sfz(i,j,k) < ZEO)then
+      if(k-3 >= kmin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)-                                                   &
+                  Sfz(i,j,k)*d12dz*(-F3*fh(i,j,k+1)-F10*fh(i,j,k)+F18*fh(i,j,k-1) &
+                                    -F6*fh(i,j,k-2)+    fh(i,j,k-3))
+     elseif(k-2 >= kmin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                           &
+                  Sfz(i,j,k)*d12dz*(fh(i,j,k-2)-EIT*fh(i,j,k-1)+EIT*fh(i,j,k+1)-fh(i,j,k+2))
+
+     elseif(k-1 >= kmin)then
+     f_rhs(i,j,k)=f_rhs(i,j,k)+                                                   &
+                  Sfz(i,j,k)*d12dz*(-F3*fh(i,j,k-1)-F10*fh(i,j,k)+F18*fh(i,j,k+1) &
+                                    -F6*fh(i,j,k+2)+    fh(i,j,k+3))
+     endif
+   endif
+  enddo
+  enddo
+  enddo
+
+! ---- Dissipation (kodis) loop ----
+  if(eps > ZEO) then
+  do k=1,ex(3)
+  do j=1,ex(2)
+  do i=1,ex(1)
+
+  if(i-3 >= imin .and. i+3 <= imax .and. &
+     j-3 >= jmin .and. j+3 <= jmax .and. &
+     k-3 >= kmin .and. k+3 <= kmax) then
+   f_rhs(i,j,k)       = f_rhs(i,j,k) + eps/cof *( (     &
+                              (fh(i-3,j,k)+fh(i+3,j,k)) - &
+                          SIX*(fh(i-2,j,k)+fh(i+2,j,k)) + &
+                          FIT*(fh(i-1,j,k)+fh(i+1,j,k)) - &
+                          TWT* fh(i,j,k)            )/dX + &
+                                                  (     &
+                              (fh(i,j-3,k)+fh(i,j+3,k)) - &
+                          SIX*(fh(i,j-2,k)+fh(i,j+2,k)) + &
+                          FIT*(fh(i,j-1,k)+fh(i,j+1,k)) - &
+                          TWT* fh(i,j,k)            )/dY + &
+                                                  (     &
+                              (fh(i,j,k-3)+fh(i,j,k+3)) - &
+                          SIX*(fh(i,j,k-2)+fh(i,j,k+2)) + &
+                          FIT*(fh(i,j,k-1)+fh(i,j,k+1)) - &
+                          TWT* fh(i,j,k)            )/dZ )
+  endif
+
+  enddo
+  enddo
+  enddo
+  endif
+
+  return
+
+  end subroutine lopsided_kodis
+
 #elif (ghost_width == 4)
 ! sixth order code
 ! Compute advection terms in right hand sides of field equations

AMSS_NCKU_source/makefile

@@ -16,6 +16,12 @@ include makefile.inc
 .cu.o:
 	$(Cu) $(CUDA_APP_FLAGS) -c $< -o $@ $(CUDA_LIB_PATH)
 
+TwoPunctures.o: TwoPunctures.C
+	${CXX} $(CXXAPPFLAGS) -qopenmp -c $< -o $@
+
+TwoPunctureABE.o: TwoPunctureABE.C
+	${CXX} $(CXXAPPFLAGS) -qopenmp -c $< -o $@
+
 # Input files
 C++FILES = ABE.o Ansorg.o Block.o misc.o monitor.o Parallel.o MPatch.o var.o\
            cgh.o bssn_class.o surface_integral.o ShellPatch.o\
@@ -34,7 +40,7 @@ C++FILES_GPU = ABE.o Ansorg.o Block.o misc.o monitor.o Parallel.o MPatch.o var.o
 
 F90FILES = enforce_algebra.o fmisc.o initial_puncture.o prolongrestrict.o\
 	   prolongrestrict_cell.o prolongrestrict_vertex.o\
-	   rungekutta4_rout.o bssn_rhs_opt.o bssn_rhs.o bssn_rhs_legacy.o diff_new.o kodiss.o kodiss_sh.o\
+	   rungekutta4_rout.o bssn_rhs.o diff_new.o kodiss.o kodiss_sh.o\
 	   lopsidediff.o sommerfeld_rout.o getnp4.o diff_new_sh.o\
 	   shellfunctions.o bssn_rhs_ss.o Set_Rho_ADM.o\
            getnp4EScalar.o bssnEScalar_rhs.o bssn_constraint.o ricci_gamma.o\
@@ -96,7 +102,7 @@ ABEGPU: $(C++FILES_GPU) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES)
 	$(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES_GPU) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES) $(LDLIBS)
 
 TwoPunctureABE: $(TwoPunctureFILES)
-	$(CLINKER) $(CXXAPPFLAGS) -o $@ $(TwoPunctureFILES) $(LDLIBS)
+	$(CLINKER) $(CXXAPPFLAGS) -qopenmp -o $@ $(TwoPunctureFILES) $(LDLIBS)
 
 clean:
 	rm *.o ABE ABEGPU TwoPunctureABE make.log -f

AMSS_NCKU_source/makefile.inc

@@ -7,20 +7,19 @@
 filein  = -I/usr/include/ -I${MKLROOT}/include
 
 ## Using sequential MKL (OpenMP disabled for better single-threaded performance)
-LDLIBS  = -L/usr/lib/x86_64-linux-gnu -L/usr/lib64 -lifcore -limf -lmpi \
-          -L${MKLROOT}/lib -lmkl_intel_lp64 -lmkl_sequential -lmkl_core \
-          -lpthread -lm -ldl
+## Added -lifcore for Intel Fortran runtime and -limf for Intel math library
+LDLIBS  = -L${MKLROOT}/lib -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lifcore -limf -lpthread -lm -ldl
 
-## Aggressive optimization flags:
-## -O3: Maximum optimization
-## -xHost: Optimize for the host CPU architecture (Intel/AMD compatible)
-## -fp-model fast=2: Aggressive floating-point optimizations
-## -fma: Enable fused multiply-add instructions
-## Note: OpenMP has been disabled (-qopenmp removed) due to performance issues
-CXXAPPFLAGS  = -O3 -xHost -fp-model fast=2 -fma \
+## Aggressive optimization flags + PGO Phase 2 (profile-guided optimization)
+## -fprofile-instr-use: use collected profile data to guide optimization decisions
+##   (branch prediction, basic block layout, inlining, loop unrolling)
+PROFDATA     = ../../pgo_profile/default.profdata
+CXXAPPFLAGS  = -O3 -xHost -fp-model fast=2 -fma -ipo \
+               -fprofile-instr-use=$(PROFDATA) \
                -Dfortran3 -Dnewc -I${MKLROOT}/include
-f90appflags  = -O3 -xHost -fp-model fast=2 -fma \
-               -fpp -I${MKLROOT}/include
+f90appflags  = -O3 -xHost -fp-model fast=2 -fma -ipo \
+               -fprofile-instr-use=$(PROFDATA) \
+               -align array64byte -fpp -I${MKLROOT}/include
 f90          = ifx
 f77          = ifx
 CXX          = icpx

makefile_and_run.py

@@ -10,17 +10,18 @@
 
 import AMSS_NCKU_Input as input_data
 import subprocess
-
+import time
 ## CPU core binding configuration using taskset
 ## taskset ensures all child processes inherit the CPU affinity mask
 ## This forces make and all compiler processes to use only nohz_full cores (4-55, 60-111)
 ## Format: taskset -c 4-55,60-111 ensures processes only run on these cores
-NUMACTL_CPU_BIND = "taskset -c 4-55,60-111"
+#NUMACTL_CPU_BIND = "taskset -c 0-111"
+NUMACTL_CPU_BIND = "taskset -c 16-47,64-95"
 
 ## Build parallelism configuration
 ## Use nohz_full cores (4-55, 60-111) for compilation: 52 + 52 = 104 cores
 ## Set make -j to utilize available cores for faster builds
-BUILD_JOBS = 104
+BUILD_JOBS = 96
 
 
 ##################################################################
@@ -117,6 +118,7 @@ def run_ABE():
     
     if (input_data.GPU_Calculation == "no"):
         mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
+        #mpi_command         = " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
         mpi_command_outfile = "ABE_out.log"
     elif (input_data.GPU_Calculation == "yes"):
         mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABEGPU"
@@ -152,13 +154,14 @@ def run_ABE():
 ## Run the AMSS-NCKU TwoPuncture program TwoPunctureABE
 
 def run_TwoPunctureABE():
-
+    tp_time1=time.time()
     print(                                                          )
     print( " Running the AMSS-NCKU executable file TwoPunctureABE " ) 
     print(                                                          )
     
     ## Define the command to run
-    TwoPuncture_command         = NUMACTL_CPU_BIND + " ./TwoPunctureABE"
+    #TwoPuncture_command         = NUMACTL_CPU_BIND + " ./TwoPunctureABE"
+    TwoPuncture_command         = " ./TwoPunctureABE"
     TwoPuncture_command_outfile = "TwoPunctureABE_out.log"
 
     ## Execute the command with subprocess.Popen and stream output
@@ -179,7 +182,9 @@ def run_TwoPunctureABE():
     print(                                               )
     print( " The TwoPunctureABE simulation is finished " ) 
     print(                                               )
-    
+    tp_time2=time.time()
+    et=tp_time2-tp_time1
+    print(f"Used time: {et}")
     return
 
 ##################################################################

parallel_plot_helper.py

@@ -0,0 +1,29 @@
+import multiprocessing
+
+def run_plot_task(task):
+    """Execute a single plotting task.
+    
+    Parameters
+    ----------
+    task : tuple
+        A tuple of (function, args_tuple) where function is a callable
+        plotting function and args_tuple contains its arguments.
+    """
+    func, args = task
+    return func(*args)
+
+
+def run_plot_tasks_parallel(plot_tasks):
+    """Execute a list of independent plotting tasks in parallel.
+
+    Uses the 'fork' context to create worker processes so that the main
+    script is NOT re-imported/re-executed in child processes.
+
+    Parameters
+    ----------
+    plot_tasks : list of tuples
+        Each element is (function, args_tuple).
+    """
+    ctx = multiprocessing.get_context('fork')
+    with ctx.Pool() as pool:
+        pool.map(run_plot_task, plot_tasks)

pgo_profile/PGO_Profile_Analysis.md

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

plot_GW_strain_amplitude_xiaoqu.py

@@ -11,6 +11,8 @@
 import numpy                               ## numpy for array operations
 import scipy                               ## scipy for interpolation and signal processing
 import math
+import matplotlib
+matplotlib.use('Agg')                      ## use non-interactive backend for multiprocessing safety
 import matplotlib.pyplot    as     plt     ## matplotlib for plotting
 import os                                  ## os for system/file operations
 

plot_binary_data.py

@@ -8,16 +8,23 @@
 ##
 #################################################
 
+## Restrict OpenMP to one thread per process so that running
+## many workers in parallel does not create an O(workers * BLAS_threads)
+## thread explosion.  The variable MUST be set before numpy/scipy
+## are imported, because the BLAS library reads them only at load time.
+import os
+os.environ.setdefault("OMP_NUM_THREADS",        "1")
+
 import numpy
 import scipy
+import matplotlib
+matplotlib.use('Agg')                      ## use non-interactive backend for multiprocessing safety
 import matplotlib.pyplot    as     plt
 from   matplotlib.colors    import LogNorm
 from   mpl_toolkits.mplot3d import Axes3D
 ## import torch
 import AMSS_NCKU_Input      as input_data
 
-import os
-
 
 #########################################################################################
 
@@ -192,3 +199,19 @@ def get_data_xy( Rmin, Rmax, n, data0, time, figure_title, figure_outdir ):
 
 ####################################################################################
 
+
+####################################################################################
+## Allow this module to be run as a standalone script so that each
+## binary-data plot can be executed in a fresh subprocess whose BLAS
+## environment variables (set above) take effect before numpy loads.
+##
+## Usage:  python3 plot_binary_data.py <filename> <binary_outdir> <figure_outdir>
+####################################################################################
+
+if __name__ == '__main__':
+    import sys
+    if len(sys.argv) != 4:
+        print(f"Usage: {sys.argv[0]} <filename> <binary_outdir> <figure_outdir>")
+        sys.exit(1)
+    plot_binary_data(sys.argv[1], sys.argv[2], sys.argv[3])
+

plot_xiaoqu.py

@@ -8,6 +8,8 @@
 #################################################
 
 import numpy                               ## numpy for array operations
+import matplotlib
+matplotlib.use('Agg')                      ## use non-interactive backend for multiprocessing safety
 import matplotlib.pyplot    as     plt     ## matplotlib for plotting
 from   mpl_toolkits.mplot3d import Axes3D  ## needed for 3D plots
 import glob
@@ -15,6 +17,9 @@ import os                                  ## operating system utilities
 
 import plot_binary_data
 import AMSS_NCKU_Input as input_data
+import subprocess
+import sys
+import multiprocessing
 
 # plt.rcParams['text.usetex'] = True  ## enable LaTeX fonts in plots
 
@@ -50,10 +55,40 @@ def generate_binary_data_plot( binary_outdir, figure_outdir ):
         file_list.append(x)
         print(x)
 
-    ## Plot each file in the list
+    ## Plot each file in parallel using subprocesses.
+    ## Each subprocess is a fresh Python process where the BLAS thread-count
+    ## environment variables (set at the top of plot_binary_data.py) take
+    ## effect before numpy is imported.  This avoids the thread explosion
+    ## that occurs when multiprocessing.Pool with 'fork' context inherits
+    ## already-initialized multi-threaded BLAS from the parent.
+    script = os.path.join( os.path.dirname(__file__), "plot_binary_data.py" )
+    max_workers = min( multiprocessing.cpu_count(), len(file_list) ) if file_list else 0
+
+    running = []
+    failed  = []
     for filename in file_list:
         print(filename)
-        plot_binary_data.plot_binary_data(filename, binary_outdir, figure_outdir)
+        proc = subprocess.Popen(
+            [sys.executable, script, filename, binary_outdir, figure_outdir],
+        )
+        running.append( (proc, filename) )
+        ## Keep at most max_workers subprocesses active at a time
+        if len(running) >= max_workers:
+            p, fn = running.pop(0)
+            p.wait()
+            if p.returncode != 0:
+                failed.append(fn)
+
+    ## Wait for all remaining subprocesses to finish
+    for p, fn in running:
+        p.wait()
+        if p.returncode != 0:
+            failed.append(fn)
+
+    if failed:
+        print( " WARNING: the following binary data plots failed:" )
+        for fn in failed:
+            print( "   ", fn )
 
     print(                        )
     print( " Binary Data Plot Has been Finished " )