Cache GPU RHS symbols and zero vacuum sources once

.gitignore

@@ -1,6 +1,6 @@
 __pycache__
 GW150914
-GW150914*
+GW150914-origin
 docs
 *.tmp
-.codex
+

AMSS_NCKU_Program.py

@@ -174,14 +174,11 @@ import generate_macrodef
 generate_macrodef.generate_macrodef_h()
 print( " AMSS-NCKU macro file macrodef.h has been generated. " )
      
-generate_macrodef.generate_macrodef_fh()
-print( " AMSS-NCKU macro file macrodef.fh has been generated. " )
-
-generate_macrodef.generate_build_config()
-print( " AMSS-NCKU build config AMSS_NCKU_build.mk has been generated. " )
-
-
-##################################################################
+generate_macrodef.generate_macrodef_fh()
+print( " AMSS-NCKU macro file macrodef.fh has been generated. " )
+
+
+##################################################################
 
 # Compile the AMSS-NCKU program according to user requirements
 
@@ -220,13 +217,11 @@ shutil.copytree(AMSS_NCKU_source_path, AMSS_NCKU_source_copy)
 
 # Copy the generated macro files into the AMSS_NCKU source folder
 
-macrodef_h_path  = os.path.join(File_directory, "macrodef.h")
-macrodef_fh_path = os.path.join(File_directory, "macrodef.fh")
-build_config_path = os.path.join(File_directory, "AMSS_NCKU_build.mk")
-
-shutil.copy2(macrodef_h_path,  AMSS_NCKU_source_copy)
-shutil.copy2(macrodef_fh_path, AMSS_NCKU_source_copy)
-shutil.copy2(build_config_path, AMSS_NCKU_source_copy)
+macrodef_h_path  = os.path.join(File_directory, "macrodef.h") 
+macrodef_fh_path = os.path.join(File_directory, "macrodef.fh") 
+
+shutil.copy2(macrodef_h_path,  AMSS_NCKU_source_copy)
+shutil.copy2(macrodef_fh_path, AMSS_NCKU_source_copy)
 
 # Notes on copying files:
 # shutil.copy2 preserves file metadata such as modification time.

AMSS_NCKU_Verify_ASC26.py

@@ -2,18 +2,13 @@
 """
 AMSS-NCKU GW150914 Simulation Regression Test Script (Comprehensive Version)
 
-Verification Requirements:
-1. RMS errors < 1% for:
-   - 3D Vector Total RMS
-   - X Component RMS
-   - Y Component RMS
-   - Z Component RMS
-2. ADM constraint violation < 2 (Grid Level 0)
-3. The following figure PDFs must match GW150914-origin exactly after rasterization:
-   - ADM_Constraint_Grid_Level_0.pdf
-   - BH_Trajectory_21_XY.pdf
-   - BH_Trajectory_XY.pdf
-   The script also reports the percentage of differing pixels for each figure.
+Verification Requirements:
+1. RMS errors < 1% for:
+   - 3D Vector Total RMS
+   - X Component RMS
+   - Y Component RMS
+   - Z Component RMS
+2. ADM constraint violation < 2 (Grid Level 0)
 
 RMS Calculation Method:
 - Computes trajectory deviation on the XY plane independently for BH1 and BH2
@@ -25,13 +20,9 @@ Default: output_dir = GW150914/AMSS_NCKU_output
 Reference: GW150914-origin (baseline simulation)
 """
 
-import numpy as np
-import sys
-import os
-import shutil
-import subprocess
-import tempfile
-from PIL import Image
+import numpy as np
+import sys
+import os
 
 # ANSI Color Codes
 class Color:
@@ -58,143 +49,17 @@ def load_bh_trajectory(filepath):
     }
 
 
-def load_constraint_data(filepath):
-    """Load constraint violation data"""
-    data = []
+def load_constraint_data(filepath):
+    """Load constraint violation data"""
+    data = []
     with open(filepath, 'r') as f:
         for line in f:
             if line.startswith('#'):
                 continue
             parts = line.split()
             if len(parts) >= 8:
-                data.append([float(x) for x in parts[:8]])
-    return np.array(data)
-
-
-def resolve_figure_dir(path):
-    """Resolve the sibling figure directory from an output or figure path."""
-    normalized = os.path.normpath(path)
-    if os.path.basename(normalized) == "figure":
-        return normalized
-    return os.path.join(os.path.dirname(normalized), "figure")
-
-
-def render_pdf_to_images(pdf_path, dpi=150):
-    """Render a PDF to RGB images using Ghostscript."""
-    gs_path = shutil.which("gs")
-    if gs_path is None:
-        raise RuntimeError("Ghostscript executable 'gs' was not found in PATH")
-
-    with tempfile.TemporaryDirectory(prefix="amss_verify_pdf_") as temp_dir:
-        output_pattern = os.path.join(temp_dir, "page-%03d.ppm")
-        cmd = [
-            gs_path,
-            "-q",
-            "-dSAFER",
-            "-dBATCH",
-            "-dNOPAUSE",
-            "-sDEVICE=ppmraw",
-            f"-r{dpi}",
-            f"-o{output_pattern}",
-            pdf_path
-        ]
-
-        try:
-            subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.PIPE, text=True)
-        except subprocess.CalledProcessError as exc:
-            message = exc.stderr.strip() or str(exc)
-            raise RuntimeError(f"Failed to render PDF '{pdf_path}': {message}") from exc
-
-        ppm_files = sorted(
-            os.path.join(temp_dir, filename)
-            for filename in os.listdir(temp_dir)
-            if filename.endswith(".ppm")
-        )
-
-        if not ppm_files:
-            raise RuntimeError(f"No rendered pages were produced for '{pdf_path}'")
-
-        images = []
-        for ppm_file in ppm_files:
-            with Image.open(ppm_file) as img:
-                images.append(np.array(img.convert("RGB"), dtype=np.uint8))
-
-        return images
-
-
-def compare_rendered_pages(ref_img, target_img):
-    """Return (different_pixels, total_pixels) for two rendered RGB pages."""
-    ref_h, ref_w = ref_img.shape[:2]
-    tgt_h, tgt_w = target_img.shape[:2]
-    total_pixels = max(ref_h, tgt_h) * max(ref_w, tgt_w)
-
-    if ref_h == tgt_h and ref_w == tgt_w:
-        different_pixels = int(np.count_nonzero(np.any(ref_img != target_img, axis=2)))
-        return different_pixels, total_pixels
-
-    diff_mask = np.ones((max(ref_h, tgt_h), max(ref_w, tgt_w)), dtype=bool)
-    overlap_h = min(ref_h, tgt_h)
-    overlap_w = min(ref_w, tgt_w)
-    overlap_diff = np.any(ref_img[:overlap_h, :overlap_w] != target_img[:overlap_h, :overlap_w], axis=2)
-    diff_mask[:overlap_h, :overlap_w] = overlap_diff
-    different_pixels = int(np.count_nonzero(diff_mask))
-    return different_pixels, total_pixels
-
-
-def compare_pdf_images(ref_pdf, target_pdf, dpi=150, threshold_percent=0.001):
-    """Compare two PDFs by rasterizing them and counting differing pixels."""
-    ref_pages = render_pdf_to_images(ref_pdf, dpi=dpi)
-    target_pages = render_pdf_to_images(target_pdf, dpi=dpi)
-
-    total_pixels = 0
-    different_pixels = 0
-    max_pages = max(len(ref_pages), len(target_pages))
-
-    for page_idx in range(max_pages):
-        if page_idx < len(ref_pages) and page_idx < len(target_pages):
-            page_diff, page_total = compare_rendered_pages(ref_pages[page_idx], target_pages[page_idx])
-        else:
-            existing_page = ref_pages[page_idx] if page_idx < len(ref_pages) else target_pages[page_idx]
-            page_total = existing_page.shape[0] * existing_page.shape[1]
-            page_diff = page_total
-
-        total_pixels += page_total
-        different_pixels += page_diff
-
-    diff_percent = (different_pixels / total_pixels * 100.0) if total_pixels else 0.0
-    return {
-        "different_pixels": different_pixels,
-        "total_pixels": total_pixels,
-        "diff_percent": diff_percent,
-        "pages_ref": len(ref_pages),
-        "pages_target": len(target_pages),
-        "passed": diff_percent < threshold_percent
-    }
-
-
-def compare_required_figures(reference_figure_dir, target_figure_dir):
-    """Compare the required GW150914 figure PDFs."""
-    figure_names = [
-        "ADM_Constraint_Grid_Level_0.pdf",
-        "BH_Trajectory_21_XY.pdf",
-        "BH_Trajectory_XY.pdf"
-    ]
-
-    results = []
-    for figure_name in figure_names:
-        ref_pdf = os.path.join(reference_figure_dir, figure_name)
-        target_pdf = os.path.join(target_figure_dir, figure_name)
-
-        if not os.path.exists(ref_pdf):
-            raise FileNotFoundError(f"Reference figure not found: {ref_pdf}")
-        if not os.path.exists(target_pdf):
-            raise FileNotFoundError(f"Target figure not found: {target_pdf}")
-
-        comparison = compare_pdf_images(ref_pdf, target_pdf)
-        comparison["name"] = figure_name
-        results.append(comparison)
-
-    return results
+                data.append([float(x) for x in parts[:8]])
+    return np.array(data)
 
 def calculate_all_rms_errors(bh_data_ref, bh_data_target):
     """
@@ -300,7 +165,7 @@ def print_rms_results(rms_dict, error, threshold=1.0):
 
     return all_passed
 
-def print_constraint_results(results, threshold=2.0):
+def print_constraint_results(results, threshold=2.0):
     print(f"\n{Color.BOLD}2. ADM Constraint Violation Analysis (Grid Level 0){Color.RESET}")
     print("-" * 65)
 
@@ -315,49 +180,22 @@ def print_constraint_results(results, threshold=2.0):
     print(f"\n   Maximum violation:  {results['max_violation']:.6f}")
     print(f"   Requirement:        < {threshold}")
     print(f"   Status:             {get_status_text(passed)}")
-
-    return passed
-
-
-def print_figure_results(results, threshold_percent=0.001):
-    print(f"\n{Color.BOLD}3. Figure Pixel Comparison (PDF Rasterization){Color.RESET}")
-    print("-" * 65)
-    print(f"   Requirement: < {threshold_percent:.3f}% differing pixels\n")
-
-    all_passed = True
-    for result in results:
-        passed = result["passed"]
-        all_passed = all_passed and passed
-        status = get_status_text(passed)
-        print(f"   {result['name']:32}: {result['diff_percent']:10.6f}%   |   Status: {status}")
-
-        if result["pages_ref"] != result["pages_target"]:
-            print(f"   {'':32}  pages(ref/target): {result['pages_ref']}/{result['pages_target']}")
-
-    return all_passed
-
-
-def print_figure_error(error_message):
-    print(f"\n{Color.BOLD}3. Figure Pixel Comparison (PDF Rasterization){Color.RESET}")
-    print("-" * 65)
-    print(f"   {Color.RED}Error: {error_message}{Color.RESET}")
-    return False
-
-
-def print_summary(rms_passed, constraint_passed, figure_passed):
-    print("\n" + Color.BLUE + Color.BOLD + "=" * 65 + Color.RESET)
-    print(Color.BOLD + "Verification Summary" + Color.RESET)
-    print(Color.BLUE + Color.BOLD + "=" * 65 + Color.RESET)
-
-    all_passed = rms_passed and constraint_passed and figure_passed
-    
-    res_rms = get_status_text(rms_passed)
-    res_con = get_status_text(constraint_passed)
-    res_fig = get_status_text(figure_passed)
-
-    print(f"   [1] Comprehensive RMS check:      {res_rms}")
-    print(f"   [2] ADM constraint check:         {res_con}")
-    print(f"   [3] Figure pixel comparison:      {res_fig}")
+
+    return passed
+
+
+def print_summary(rms_passed, constraint_passed):
+    print("\n" + Color.BLUE + Color.BOLD + "=" * 65 + Color.RESET)
+    print(Color.BOLD + "Verification Summary" + Color.RESET)
+    print(Color.BLUE + Color.BOLD + "=" * 65 + Color.RESET)
+
+    all_passed = rms_passed and constraint_passed
+    
+    res_rms = get_status_text(rms_passed)
+    res_con = get_status_text(constraint_passed)
+
+    print(f"   [1] Comprehensive RMS check:      {res_rms}")
+    print(f"   [2] ADM constraint check:         {res_con}")
     
     final_status = f"{Color.GREEN}{Color.BOLD}ALL CHECKS PASSED{Color.RESET}" if all_passed else f"{Color.RED}{Color.BOLD}SOME CHECKS FAILED{Color.RESET}"
     print(f"\n   Overall result: {final_status}")
@@ -372,14 +210,12 @@ def main():
         script_dir = os.path.dirname(os.path.abspath(__file__))
         target_dir = os.path.join(script_dir, "GW150914/AMSS_NCKU_output")
 
-    script_dir = os.path.dirname(os.path.abspath(__file__))
-    reference_dir = os.path.join(script_dir, "GW150914-origin/AMSS_NCKU_output")
-    target_figure_dir = resolve_figure_dir(target_dir)
-    reference_figure_dir = os.path.join(script_dir, "GW150914-origin/figure")
-
-    bh_file_ref = os.path.join(reference_dir, "bssn_BH.dat")
-    bh_file_target = os.path.join(target_dir, "bssn_BH.dat")
-    constraint_file = os.path.join(target_dir, "bssn_constraint.dat")
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    reference_dir = os.path.join(script_dir, "GW150914-origin/AMSS_NCKU_output")
+
+    bh_file_ref = os.path.join(reference_dir, "bssn_BH.dat")
+    bh_file_target = os.path.join(target_dir, "bssn_BH.dat")
+    constraint_file = os.path.join(target_dir, "bssn_constraint.dat")
 
     if not os.path.exists(bh_file_ref):
         print(f"{Color.RED}{Color.BOLD}Error:{Color.RESET} Baseline trajectory file not found: {bh_file_ref}")
@@ -391,11 +227,9 @@ def main():
         print(f"{Color.RED}{Color.BOLD}Error:{Color.RESET} Constraint data file not found: {constraint_file}")
         sys.exit(1)
 
-    print_header()
-    print(f"\n{Color.BOLD}Reference (Baseline):{Color.RESET} {Color.BLUE}{reference_dir}{Color.RESET}")
-    print(f"{Color.BOLD}Target (Optimized):  {Color.RESET} {Color.BLUE}{target_dir}{Color.RESET}")
-    print(f"{Color.BOLD}Reference Figures:   {Color.RESET} {Color.BLUE}{reference_figure_dir}{Color.RESET}")
-    print(f"{Color.BOLD}Target Figures:      {Color.RESET} {Color.BLUE}{target_figure_dir}{Color.RESET}")
+    print_header()
+    print(f"\n{Color.BOLD}Reference (Baseline):{Color.RESET} {Color.BLUE}{reference_dir}{Color.RESET}")
+    print(f"{Color.BOLD}Target (Optimized):  {Color.RESET} {Color.BLUE}{target_dir}{Color.RESET}")
 
     bh_data_ref = load_bh_trajectory(bh_file_ref)
     bh_data_target = load_bh_trajectory(bh_file_target)
@@ -405,18 +239,12 @@ def main():
     rms_dict, error = calculate_all_rms_errors(bh_data_ref, bh_data_target)
     rms_passed = print_rms_results(rms_dict, error)
 
-    # Output constraint results
-    constraint_results = analyze_constraint_violation(constraint_data)
-    constraint_passed = print_constraint_results(constraint_results)
-
-    try:
-        figure_results = compare_required_figures(reference_figure_dir, target_figure_dir)
-        figure_passed = print_figure_results(figure_results)
-    except (FileNotFoundError, RuntimeError) as exc:
-        figure_passed = print_figure_error(str(exc))
-
-    all_passed = print_summary(rms_passed, constraint_passed, figure_passed)
-    sys.exit(0 if all_passed else 1)
+    # Output constraint results
+    constraint_results = analyze_constraint_violation(constraint_data)
+    constraint_passed = print_constraint_results(constraint_results)
+
+    all_passed = print_summary(rms_passed, constraint_passed)
+    sys.exit(0 if all_passed else 1)
 
 if __name__ == "__main__":
     main()

AMSS_NCKU_source/ABE.C

@@ -23,22 +23,20 @@ using namespace std;
 #include <mpi.h>
 
 #include "misc.h"
-#include "macrodef.h"
+#include "macrodef.h"
+#ifdef USE_GPU
+extern void bssn_cuda_dump_stage_profile();
+#endif
 
 #ifndef ABEtype
 #error "not define ABEtype"
 #endif
 
-#if (ABEtype == 0)
-
-#ifdef USE_GPU
-#include "bssn_gpu_class.h"
-#else
-#include "bssn_class.h"
-#endif
-
-#elif (ABEtype == 1)
-#include "bssnEScalar_class.h"
+#if (ABEtype == 0)
+#include "bssn_class.h"
+
+#elif (ABEtype == 1)
+#include "bssnEScalar_class.h"
 
 #elif (ABEtype == 2)
 #include "Z4c_class.h"
@@ -474,10 +472,13 @@ int main(int argc, char *argv[])
             cout << endl;
       }
 
-      ADM->Evolve(Steps);
-
-      if (myrank == 0)
-      {
+	      ADM->Evolve(Steps);
+#ifdef USE_GPU
+	      bssn_cuda_dump_stage_profile();
+#endif
+	
+	      if (myrank == 0)
+	      {
             cout << endl;
             cout << " Total Evolve Time: "  << MPI_Wtime() - End_clock   << " seconds!" << endl;
             cout << " Total Running Time: " << MPI_Wtime() - Begin_clock << " seconds!" << endl;

AMSS_NCKU_source/Block.C

@@ -9,8 +9,12 @@
 #include <new>
 using namespace std;
 
-#include "Block.h"
-#include "misc.h"
+#include "Block.h"
+#include "misc.h"
+#ifdef USE_GPU
+#include "bssn_gpu.h"
+#include "bssn_cuda_ops.h"
+#endif
 
 Block::Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fngfsi, int levi, const int cgpui) : rank(ranki), ingfs(ingfsi), fngfs(fngfsi), lev(levi), cgpu(cgpui)
 {
@@ -95,14 +99,19 @@ Block::Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fng
   }
 #endif
 }
-Block::~Block()
-{
-  int myrank;
-  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
-  if (myrank == rank)
-  {
-    for (int i = 0; i < dim; i++)
-      delete[] X[i];
+Block::~Block()
+{
+  int myrank;
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+  if (myrank == rank)
+  {
+#ifdef USE_GPU
+    bssn_gpu_clear_cached_device_buffers();
+    bssn_cuda_release_rk4_caches();
+    bssn_cuda_release_interp_caches();
+#endif
+    for (int i = 0; i < dim; i++)
+      delete[] X[i];
     for (int i = 0; i < ingfs; i++)
       free(igfs[i]);
     delete[] igfs;

AMSS_NCKU_source/MPatch.C

@@ -2,29 +2,100 @@
 #include <iostream>
 #include <iomanip>
 #include <fstream>
-#include <cstdlib>
-#include <cstdio>
-#include <string>
-#include <cmath>
-#include <new>
-#include <vector>
-using namespace std;
+#include <cstdlib>
+#include <cstdio>
+#include <string>
+#include <cmath>
+#include <new>
+#include <map>
+#include <vector>
+using namespace std;
 
-#include "misc.h"
-#include "MPatch.h"
-#include "Parallel.h"
-#include "fmisc.h"
-#ifdef INTERP_LB_PROFILE
-#include "interp_lb_profile.h"
-#endif
-
-namespace
-{
-struct InterpBlockView
-{
-  Block *bp;
-  double llb[dim];
-  double uub[dim];
+#include "misc.h"
+#include "MPatch.h"
+#include "Parallel.h"
+#include "fmisc.h"
+#include "bssn_cuda_ops.h"
+#ifdef INTERP_LB_PROFILE
+#include "interp_lb_profile.h"
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+extern int bssn_cuda_interp_points_batch(const int *ex,
+                                         const double *X, const double *Y, const double *Z,
+                                         const double *const *fields,
+                                         const double *soa_flat,
+                                         int num_var,
+                                         const double *px, const double *py, const double *pz,
+                                         int num_points,
+                                         int ordn,
+                                         int symmetry,
+                                         double *out) __attribute__((weak));
+#endif
+
+namespace
+{
+struct InterpVarDesc
+{
+  int sgfn;
+  double soa[dim];
+};
+
+struct InterpPlanKey
+{
+  const Patch *patch;
+  const double *x;
+  const double *y;
+  const double *z;
+  int NN;
+  int Symmetry;
+  int myrank;
+};
+
+struct InterpPlanKeyLess
+{
+  bool operator()(const InterpPlanKey &lhs, const InterpPlanKey &rhs) const
+  {
+    if (lhs.patch != rhs.patch) return lhs.patch < rhs.patch;
+    if (lhs.x != rhs.x) return lhs.x < rhs.x;
+    if (lhs.y != rhs.y) return lhs.y < rhs.y;
+    if (lhs.z != rhs.z) return lhs.z < rhs.z;
+    if (lhs.NN != rhs.NN) return lhs.NN < rhs.NN;
+    if (lhs.Symmetry != rhs.Symmetry) return lhs.Symmetry < rhs.Symmetry;
+    return lhs.myrank < rhs.myrank;
+  }
+};
+
+struct CachedInterpPlan
+{
+  int nblocks;
+  vector<int> owner_rank;
+  vector<int> owner_block;
+  vector<vector<int> > block_points;
+  vector<vector<double> > block_px;
+  vector<vector<double> > block_py;
+  vector<vector<double> > block_pz;
+
+  CachedInterpPlan() : nblocks(0) {}
+};
+
+struct CachedInterpPlanEntry
+{
+  bool valid;
+  InterpPlanKey key;
+  vector<double> xvals;
+  vector<double> yvals;
+  vector<double> zvals;
+  CachedInterpPlan plan;
+
+  CachedInterpPlanEntry() : valid(false) {}
+};
+
+struct InterpBlockView
+{
+  Block *bp;
+  double llb[dim];
+  double uub[dim];
 };
 
 struct BlockBinIndex
@@ -154,10 +225,10 @@ void build_block_bin_index(Patch *patch, const double *DH, BlockBinIndex &index)
   index.valid = true;
 }
 
-int find_block_index_for_point(const BlockBinIndex &index, const double *pox, const double *DH)
-{
-  if (!index.valid)
-    return -1;
+int find_block_index_for_point(const BlockBinIndex &index, const double *pox, const double *DH)
+{
+  if (!index.valid)
+    return -1;
 
   const int bx = coord_to_bin(pox[0], index.lo[0], index.inv[0], index.bins[0]);
   const int by = coord_to_bin(pox[1], index.lo[1], index.inv[1], index.bins[1]);
@@ -175,13 +246,314 @@ int find_block_index_for_point(const BlockBinIndex &index, const double *pox, co
   for (size_t bi = 0; bi < index.views.size(); bi++)
     if (point_in_block_view(index.views[bi], pox, DH))
       return int(bi);
-
-  return -1;
-}
-} // namespace
-
-Patch::Patch(int DIM, int *shapei, double *bboxi, int levi, bool buflog, int Symmetry) : lev(levi)
-{
+
+  return -1;
+}
+
+void collect_interp_vars(MyList<var> *VarList, vector<InterpVarDesc> &vars)
+{
+  vars.clear();
+  MyList<var> *varl = VarList;
+  while (varl)
+  {
+    InterpVarDesc desc;
+    desc.sgfn = varl->data->sgfn;
+    for (int d = 0; d < dim; ++d)
+      desc.soa[d] = varl->data->SoA[d];
+    vars.push_back(desc);
+    varl = varl->next;
+  }
+}
+
+bool should_try_cuda_interp(int ordn, int num_points, int num_var)
+{
+#if defined(__GNUC__) || defined(__clang__)
+  if (!bssn_cuda_interp_points_batch)
+    return false;
+#else
+  return false;
+#endif
+  if (ordn != 6)
+    return false;
+  if (num_points < 32)
+    return false;
+  return num_points * num_var >= 256;
+}
+
+CachedInterpPlanEntry &interp_plan_cache_entry()
+{
+  static CachedInterpPlanEntry cache;
+  return cache;
+}
+
+bool same_interp_plan_key(const InterpPlanKey &lhs, const InterpPlanKey &rhs)
+{
+  return lhs.patch == rhs.patch &&
+         lhs.NN == rhs.NN &&
+         lhs.Symmetry == rhs.Symmetry &&
+         lhs.myrank == rhs.myrank;
+}
+
+bool same_interp_plan_points(const CachedInterpPlanEntry &cache, int NN, double **XX)
+{
+  if (static_cast<int>(cache.xvals.size()) != NN ||
+      static_cast<int>(cache.yvals.size()) != NN ||
+      static_cast<int>(cache.zvals.size()) != NN)
+    return false;
+
+  for (int j = 0; j < NN; ++j)
+  {
+    if (cache.xvals[j] != XX[0][j] ||
+        cache.yvals[j] != XX[1][j] ||
+        cache.zvals[j] != XX[2][j])
+      return false;
+  }
+  return true;
+}
+
+CachedInterpPlan &get_cached_interp_plan(Patch *patch,
+                                         int NN, double **XX,
+                                         int Symmetry, int myrank,
+                                         const double *DH,
+                                         const BlockBinIndex &block_index,
+                                         bool report_bounds_here,
+                                         bool allow_missing_points)
+{
+  InterpPlanKey key;
+  key.patch = patch;
+  key.x = XX[0];
+  key.y = XX[1];
+  key.z = XX[2];
+  key.NN = NN;
+  key.Symmetry = Symmetry;
+  key.myrank = myrank;
+
+  CachedInterpPlanEntry &cache = interp_plan_cache_entry();
+  if (cache.valid &&
+      same_interp_plan_key(cache.key, key) &&
+      same_interp_plan_points(cache, NN, XX) &&
+      cache.plan.nblocks == static_cast<int>(block_index.views.size()))
+    return cache.plan;
+
+  cache.valid = true;
+  cache.key = key;
+  cache.xvals.assign(XX[0], XX[0] + NN);
+  cache.yvals.assign(XX[1], XX[1] + NN);
+  cache.zvals.assign(XX[2], XX[2] + NN);
+  cache.plan = CachedInterpPlan();
+  CachedInterpPlan &plan = cache.plan;
+  plan.nblocks = static_cast<int>(block_index.views.size());
+  plan.owner_rank.assign(NN, -1);
+  plan.owner_block.assign(NN, -1);
+  plan.block_points.resize(plan.nblocks);
+  plan.block_px.resize(plan.nblocks);
+  plan.block_py.resize(plan.nblocks);
+  plan.block_pz.resize(plan.nblocks);
+
+  for (int j = 0; j < NN; ++j)
+  {
+    double pox[dim];
+    for (int i = 0; i < dim; ++i)
+    {
+      pox[i] = XX[i][j];
+      if (report_bounds_here &&
+          (XX[i][j] < patch->bbox[i] + patch->lli[i] * DH[i] ||
+           XX[i][j] > patch->bbox[dim + i] - patch->uui[i] * DH[i]))
+      {
+        cout << "Patch::Interp_Points: point (";
+        for (int k = 0; k < dim; ++k)
+        {
+          cout << XX[k][j];
+          if (k < dim - 1)
+            cout << ",";
+          else
+            cout << ") is out of current Patch." << endl;
+        }
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
+
+    const int block_i = find_block_index_for_point(block_index, pox, DH);
+    if (block_i >= 0)
+    {
+      Block *BP = block_index.views[block_i].bp;
+      plan.owner_rank[j] = BP->rank;
+      plan.owner_block[j] = block_i;
+      if (BP->rank == myrank)
+      {
+        plan.block_points[block_i].push_back(j);
+        plan.block_px[block_i].push_back(XX[0][j]);
+        plan.block_py[block_i].push_back(XX[1][j]);
+        plan.block_pz[block_i].push_back(XX[2][j]);
+      }
+    }
+  }
+
+  if (!allow_missing_points && report_bounds_here)
+  {
+    for (int j = 0; j < NN; ++j)
+    {
+      if (plan.owner_rank[j] >= 0)
+        continue;
+      cout << "ERROR: Patch::Interp_Points fails to find point (";
+      for (int d = 0; d < dim; ++d)
+      {
+        cout << XX[d][j];
+        if (d < dim - 1)
+          cout << ",";
+        else
+          cout << ")";
+      }
+      cout << " on Patch (";
+      for (int d = 0; d < dim; ++d)
+      {
+        cout << patch->bbox[d] << "+" << patch->lli[d] * DH[d];
+        if (d < dim - 1)
+          cout << ",";
+        else
+          cout << ")--";
+      }
+      cout << "(";
+      for (int d = 0; d < dim; ++d)
+      {
+        cout << patch->bbox[dim + d] << "-" << patch->uui[d] * DH[d];
+        if (d < dim - 1)
+          cout << ",";
+        else
+          cout << ")" << endl;
+      }
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
+
+  return plan;
+}
+
+void release_interp_plan_cache_internal()
+{
+  CachedInterpPlanEntry &cache = interp_plan_cache_entry();
+  cache.valid = false;
+  cache.xvals.clear();
+  cache.yvals.clear();
+  cache.zvals.clear();
+  cache.plan = CachedInterpPlan();
+}
+
+bool run_cuda_interp_for_block(Block *BP,
+                               const vector<InterpVarDesc> &vars,
+                               const vector<int> &point_ids,
+                               const vector<double> &px,
+                               const vector<double> &py,
+                               const vector<double> &pz,
+                               double *Shellf,
+                               int num_var,
+                               int ordn,
+                               int Symmetry)
+{
+  if (!should_try_cuda_interp(ordn, static_cast<int>(point_ids.size()), num_var))
+    return false;
+
+  vector<const double *> field_ptrs(num_var);
+  vector<double> soa_flat(3 * num_var);
+  for (int v = 0; v < num_var; ++v)
+  {
+    field_ptrs[v] = BP->fgfs[vars[v].sgfn];
+    for (int d = 0; d < dim; ++d)
+      soa_flat[3 * v + d] = vars[v].soa[d];
+  }
+
+  const int npts = static_cast<int>(point_ids.size());
+  vector<double> out(static_cast<size_t>(npts) * static_cast<size_t>(num_var));
+  if (bssn_cuda_interp_points_batch(BP->shape,
+                                    BP->X[0], BP->X[1], BP->X[2],
+                                    field_ptrs.data(),
+                                    soa_flat.data(),
+                                    num_var,
+                                    px.data(), py.data(), pz.data(),
+                                    npts,
+                                    ordn,
+                                    Symmetry,
+                                    out.data()) != 0)
+  {
+    return false;
+  }
+
+  for (int p = 0; p < npts; ++p)
+  {
+    const int j = point_ids[p];
+    memcpy(Shellf + j * num_var, out.data() + p * num_var, sizeof(double) * num_var);
+  }
+  return true;
+}
+
+void run_cpu_interp_for_block(Block *BP,
+                              const vector<InterpVarDesc> &vars,
+                              const vector<int> &point_ids,
+                              const vector<double> &px,
+                              const vector<double> &py,
+                              const vector<double> &pz,
+                              double *Shellf,
+                              int num_var,
+                              int ordn,
+                              int Symmetry)
+{
+  for (size_t p = 0; p < point_ids.size(); ++p)
+  {
+    const int j = point_ids[p];
+    double x = px[p];
+    double y = py[p];
+    double z = pz[p];
+    int ordn_local = ordn;
+    int symmetry_local = Symmetry;
+    for (int v = 0; v < num_var; ++v)
+    {
+      f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2],
+                      BP->fgfs[vars[v].sgfn], Shellf[j * num_var + v],
+                      x, y, z, ordn_local, const_cast<double *>(vars[v].soa), symmetry_local);
+    }
+  }
+}
+
+void interpolate_owned_points(MyList<var> *VarList,
+                              double *Shellf, int Symmetry,
+                              int ordn,
+                              const BlockBinIndex &block_index,
+                              const CachedInterpPlan &plan)
+{
+  vector<InterpVarDesc> vars;
+  collect_interp_vars(VarList, vars);
+  const int num_var = static_cast<int>(vars.size());
+
+  for (size_t bi = 0; bi < plan.block_points.size(); ++bi)
+  {
+    if (plan.block_points[bi].empty())
+      continue;
+
+    Block *BP = block_index.views[bi].bp;
+    bool done = run_cuda_interp_for_block(BP, vars,
+                                          plan.block_points[bi],
+                                          plan.block_px[bi],
+                                          plan.block_py[bi],
+                                          plan.block_pz[bi],
+                                          Shellf, num_var, ordn, Symmetry);
+    if (!done)
+      run_cpu_interp_for_block(BP, vars,
+                               plan.block_points[bi],
+                               plan.block_px[bi],
+                               plan.block_py[bi],
+                               plan.block_pz[bi],
+                               Shellf, num_var, ordn, Symmetry);
+  }
+}
+} // namespace
+
+void patch_release_interp_plan_cache()
+{
+  release_interp_plan_cache_internal();
+}
+
+Patch::Patch(int DIM, int *shapei, double *bboxi, int levi, bool buflog, int Symmetry) : lev(levi)
+{
 
   int hbuffer_width = buffer_width;
   if (lev == 0)
@@ -523,60 +895,15 @@ void Patch::Interp_Points(MyList<var> *VarList,
 
   memset(Shellf, 0, sizeof(double) * NN * num_var);
 
-  // owner_rank[j] records which MPI rank owns point j
-  // All ranks traverse the same block list so they all agree on ownership
-  int *owner_rank;
-  owner_rank = new int[NN];
-  for (int j = 0; j < NN; j++)
-    owner_rank[j] = -1;
-
-  double DH[dim];
-  for (int i = 0; i < dim; i++)
-    DH[i] = getdX(i);
-  BlockBinIndex block_index;
-  build_block_bin_index(this, DH, block_index);
-
-  for (int j = 0; j < NN; j++) // run along points
-  {
-    double pox[dim];
-    for (int i = 0; i < dim; i++)
-    {
-      pox[i] = XX[i][j];
-      if (myrank == 0 && (XX[i][j] < bbox[i] + lli[i] * DH[i] || XX[i][j] > bbox[dim + i] - uui[i] * DH[i]))
-      {
-        cout << "Patch::Interp_Points: point (";
-        for (int k = 0; k < dim; k++)
-        {
-          cout << XX[k][j];
-          if (k < dim - 1)
-            cout << ",";
-          else
-            cout << ") is out of current Patch." << endl;
-        }
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
-
-    const int block_i = find_block_index_for_point(block_index, pox, DH);
-    if (block_i >= 0)
-    {
-      Block *BP = block_index.views[block_i].bp;
-      owner_rank[j] = BP->rank;
-      if (myrank == BP->rank)
-      {
-        //---> interpolation
-        varl = VarList;
-        int k = 0;
-        while (varl) // run along variables
-        {
-          f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
-                          pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
-          varl = varl->next;
-          k++;
-        }
-      }
-    }
-  }
+  double DH[dim];
+  for (int i = 0; i < dim; i++)
+    DH[i] = getdX(i);
+  BlockBinIndex block_index;
+  build_block_bin_index(this, DH, block_index);
+  CachedInterpPlan &plan = get_cached_interp_plan(this, NN, XX, Symmetry, myrank, DH, block_index, myrank == 0, false);
+  const int *owner_rank = plan.owner_rank.data();
+
+  interpolate_owned_points(VarList, Shellf, Symmetry, ordn, block_index, plan);
 
   // Replace MPI_Allreduce with per-owner MPI_Bcast:
   // Group consecutive points by owner rank and broadcast each group.
@@ -631,9 +958,8 @@ void Patch::Interp_Points(MyList<var> *VarList,
       MPI_Bcast(Shellf + jstart * num_var, count, MPI_DOUBLE, cur_owner, MPI_COMM_WORLD);
     }
   }
-
-  delete[] owner_rank;
-}
+
+}
 void Patch::Interp_Points(MyList<var> *VarList,
                           int NN, double **XX,
                           double *Shellf, int Symmetry,
@@ -661,102 +987,22 @@ void Patch::Interp_Points(MyList<var> *VarList,
 
   memset(Shellf, 0, sizeof(double) * NN * num_var);
 
-  // owner_rank[j] records which MPI rank owns point j
-  int *owner_rank;
-  owner_rank = new int[NN];
-  for (int j = 0; j < NN; j++)
-    owner_rank[j] = -1;
-
-  double DH[dim];
-  for (int i = 0; i < dim; i++)
-    DH[i] = getdX(i);
-  BlockBinIndex block_index;
-  build_block_bin_index(this, DH, block_index);
-
-  // --- Interpolation phase (identical to original) ---
-  for (int j = 0; j < NN; j++)
-  {
-    double pox[dim];
-    for (int i = 0; i < dim; i++)
-    {
-      pox[i] = XX[i][j];
-      if (myrank == 0 && (XX[i][j] < bbox[i] + lli[i] * DH[i] || XX[i][j] > bbox[dim + i] - uui[i] * DH[i]))
-      {
-        cout << "Patch::Interp_Points: point (";
-        for (int k = 0; k < dim; k++)
-        {
-          cout << XX[k][j];
-          if (k < dim - 1)
-            cout << ",";
-          else
-            cout << ") is out of current Patch." << endl;
-        }
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
-
-    const int block_i = find_block_index_for_point(block_index, pox, DH);
-    if (block_i >= 0)
-    {
-      Block *BP = block_index.views[block_i].bp;
-      owner_rank[j] = BP->rank;
-      if (myrank == BP->rank)
-      {
-        varl = VarList;
-        int k = 0;
-        while (varl)
-        {
-          f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
-                          pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
-          varl = varl->next;
-          k++;
-        }
-      }
-    }
-  }
+  double DH[dim];
+  for (int i = 0; i < dim; i++)
+    DH[i] = getdX(i);
+  BlockBinIndex block_index;
+  build_block_bin_index(this, DH, block_index);
+  CachedInterpPlan &plan = get_cached_interp_plan(this, NN, XX, Symmetry, myrank, DH, block_index, myrank == 0, false);
+  const int *owner_rank = plan.owner_rank.data();
+
+  interpolate_owned_points(VarList, Shellf, Symmetry, ordn, block_index, plan);
 
 #ifdef INTERP_LB_PROFILE
   double t_interp_end = MPI_Wtime();
   double t_interp_local = t_interp_end - t_interp_start;
 #endif
 
-  // --- Error check for unfound points ---
-  for (int j = 0; j < NN; j++)
-  {
-    if (owner_rank[j] < 0 && myrank == 0)
-    {
-      cout << "ERROR: Patch::Interp_Points fails to find point (";
-      for (int d = 0; d < dim; d++)
-      {
-        cout << XX[d][j];
-        if (d < dim - 1)
-          cout << ",";
-        else
-          cout << ")";
-      }
-      cout << " on Patch (";
-      for (int d = 0; d < dim; d++)
-      {
-        cout << bbox[d] << "+" << lli[d] * DH[d];
-        if (d < dim - 1)
-          cout << ",";
-        else
-          cout << ")--";
-      }
-      cout << "(";
-      for (int d = 0; d < dim; d++)
-      {
-        cout << bbox[dim + d] << "-" << uui[d] * DH[d];
-        if (d < dim - 1)
-          cout << ",";
-        else
-          cout << ")" << endl;
-      }
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
-
-  // --- Targeted point-to-point communication phase ---
+  // --- Targeted point-to-point communication phase ---
   // Compute consumer_rank[j] using the same deterministic formula as surface_integral
   int *consumer_rank = new int[NN];
   {
@@ -873,9 +1119,8 @@ void Patch::Interp_Points(MyList<var> *VarList,
   delete[] send_offset;
   delete[] recv_offset;
   delete[] send_count;
-  delete[] recv_count;
-  delete[] consumer_rank;
-  delete[] owner_rank;
+  delete[] recv_count;
+  delete[] consumer_rank;
 
 #ifdef INTERP_LB_PROFILE
   {
@@ -923,64 +1168,20 @@ void Patch::Interp_Points(MyList<var> *VarList,
 
   memset(Shellf, 0, sizeof(double) * NN * num_var);
 
-  // owner_rank[j] stores the global rank that owns point j
-  int *owner_rank;
-  owner_rank = new int[NN];
-  for (int j = 0; j < NN; j++)
-    owner_rank[j] = -1;
+  // Build global-to-local rank translation for Comm_here
+  MPI_Group world_group, local_group;
+  MPI_Comm_group(MPI_COMM_WORLD, &world_group);
+  MPI_Comm_group(Comm_here, &local_group);
 
-  // Build global-to-local rank translation for Comm_here
-  MPI_Group world_group, local_group;
-  MPI_Comm_group(MPI_COMM_WORLD, &world_group);
-  MPI_Comm_group(Comm_here, &local_group);
-
-  double DH[dim];
-  for (int i = 0; i < dim; i++)
-    DH[i] = getdX(i);
-  BlockBinIndex block_index;
-  build_block_bin_index(this, DH, block_index);
-
-  for (int j = 0; j < NN; j++) // run along points
-  {
-    double pox[dim];
-    for (int i = 0; i < dim; i++)
-    {
-      pox[i] = XX[i][j];
-      if (lmyrank == 0 && (XX[i][j] < bbox[i] + lli[i] * DH[i] || XX[i][j] > bbox[dim + i] - uui[i] * DH[i]))
-      {
-        cout << "Patch::Interp_Points: point (";
-        for (int k = 0; k < dim; k++)
-        {
-          cout << XX[k][j];
-          if (k < dim - 1)
-            cout << ",";
-          else
-            cout << ") is out of current Patch." << endl;
-        }
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
-
-    const int block_i = find_block_index_for_point(block_index, pox, DH);
-    if (block_i >= 0)
-    {
-      Block *BP = block_index.views[block_i].bp;
-      owner_rank[j] = BP->rank;
-      if (myrank == BP->rank)
-      {
-        //---> interpolation
-        varl = VarList;
-        int k = 0;
-        while (varl) // run along variables
-        {
-          f_global_interp(BP->shape, BP->X[0], BP->X[1], BP->X[2], BP->fgfs[varl->data->sgfn], Shellf[j * num_var + k],
-                          pox[0], pox[1], pox[2], ordn, varl->data->SoA, Symmetry);
-          varl = varl->next;
-          k++;
-        }
-      }
-    }
-  }
+  double DH[dim];
+  for (int i = 0; i < dim; i++)
+    DH[i] = getdX(i);
+  BlockBinIndex block_index;
+  build_block_bin_index(this, DH, block_index);
+  CachedInterpPlan &plan = get_cached_interp_plan(this, NN, XX, Symmetry, myrank, DH, block_index, lmyrank == 0, true);
+  const int *owner_rank = plan.owner_rank.data();
+
+  interpolate_owned_points(VarList, Shellf, Symmetry, ordn, block_index, plan);
 
   // Collect unique global owner ranks and translate to local ranks in Comm_here
   // Then broadcast each owner's points via MPI_Bcast on Comm_here
@@ -1008,10 +1209,9 @@ void Patch::Interp_Points(MyList<var> *VarList,
     }
   }
 
-  MPI_Group_free(&world_group);
-  MPI_Group_free(&local_group);
-  delete[] owner_rank;
-}
+  MPI_Group_free(&world_group);
+  MPI_Group_free(&local_group);
+}
 void Patch::checkBlock()
 {
   int myrank;

AMSS_NCKU_source/MPatch.h

@@ -8,7 +8,7 @@
 #include "var.h"
 #include "macrodef.h" //need dim here; Vertex or Cell; ghost_width
 
-class Patch
+class Patch
 {
 
 public:
@@ -50,6 +50,8 @@ public:
                          double *Shellf, int Symmetry, MPI_Comm Comm_here);
    void Find_Maximum(MyList<var> *VarList, double *XX,
                      double *Shellf, MPI_Comm Comm_here);
-};
-
-#endif /* PATCH_H */
+};
+
+void patch_release_interp_plan_cache();
+
+#endif /* PATCH_H */

AMSS_NCKU_source/Parallel.h

@@ -89,9 +89,12 @@ namespace Parallel
   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);
+  void Sync(Patch *Pat, MyList<var> *VarList, int Symmetry);
+  void Sync(Patch *Pat, MyList<var> *VarList, int Symmetry, const char *context);
+  void Sync(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
+  void Sync(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, const char *context);
+  void Sync_merged(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry);
+  void Sync_merged(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, const char *context);
 
   struct SyncCache {
     bool valid;
@@ -105,12 +108,13 @@ namespace Parallel
     int *send_buf_caps;
     int *recv_buf_caps;
     MPI_Request *reqs;
-    MPI_Status *stats;
-    int max_reqs;
-    bool lengths_valid;
-    int *tc_req_node;
-    int *tc_req_is_recv;
-    int *tc_completed;
+    MPI_Status *stats;
+    int max_reqs;
+    bool lengths_valid;
+    int lengths_var_count;
+    int *tc_req_node;
+    int *tc_req_is_recv;
+    int *tc_completed;
     SyncCache();
     void invalidate();
     void destroy();
@@ -121,19 +125,20 @@ namespace Parallel
                        MyList<var> *VarList1, MyList<var> *VarList2,
                        int Symmetry, SyncCache &cache);
 
-  struct AsyncSyncState {
-    int req_no;
-    bool active;
-    int *req_node;
-    int *req_is_recv;
-    int pending_recv;
-    AsyncSyncState() : req_no(0), active(false), req_node(0), req_is_recv(0), pending_recv(0) {}
-  };
+  struct AsyncSyncState {
+    int req_no;
+    bool active;
+    int mpi_tag;
+    int *req_node;
+    int *req_is_recv;
+    int pending_recv;
+    AsyncSyncState() : req_no(0), active(false), mpi_tag(0), req_node(0), req_is_recv(0), pending_recv(0) {}
+  };
 
-  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 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, bool unpack_to_host = true);
   void OutBdLow2Hi(Patch *Patc, Patch *Patf,
                    MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /* target */,
                    int Symmetry);
@@ -179,13 +184,12 @@ namespace Parallel
   MyList<Parallel::gridseg> *clone_gsl(MyList<Parallel::gridseg> *p, bool first_only);
   MyList<Parallel::gridseg> *build_bulk_gsl(Patch *Pat); // similar to build_owned_gsl0 but does not care rank issue
   MyList<Parallel::gridseg> *build_bulk_gsl(Block *bp, Patch *Pat);
-  void build_PhysBD_gstl(Patch *Pat, MyList<Parallel::gridseg> *srci, MyList<Parallel::gridseg> *dsti,
-                         MyList<Parallel::gridseg> **out_src, MyList<Parallel::gridseg> **out_dst);
-  void PeriodicBD(Patch *Pat, MyList<var> *VarList, int Symmetry);
-  double L2Norm(Patch *Pat, var *vf);
-  void L2Norm7(Patch *Pat, var **vf, double *norms);
-  void checkgsl(MyList<Parallel::gridseg> *pp, bool first_only);
-  void checkvarl(MyList<var> *pp, bool first_only);
+  void build_PhysBD_gstl(Patch *Pat, MyList<Parallel::gridseg> *srci, MyList<Parallel::gridseg> *dsti,
+                         MyList<Parallel::gridseg> **out_src, MyList<Parallel::gridseg> **out_dst);
+  void PeriodicBD(Patch *Pat, MyList<var> *VarList, int Symmetry);
+  double L2Norm(Patch *Pat, var *vf);
+  void checkgsl(MyList<Parallel::gridseg> *pp, bool first_only);
+  void checkvarl(MyList<var> *pp, bool first_only);
   MyList<Parallel::gridseg> *divide_gsl(MyList<Parallel::gridseg> *p, Patch *Pat);
   MyList<Parallel::gridseg> *divide_gs(MyList<Parallel::gridseg> *p, Patch *Pat);
   void prepare_inter_time_level(Patch *Pat,
@@ -217,12 +221,11 @@ namespace Parallel
   void aligncheck(double *bbox0, double *bboxl, int lev, double *DH0, int *shape);
   bool point_locat_gsl(double *pox, MyList<Parallel::gridseg> *gsl);
   void checkpatchlist(MyList<Patch> *PatL, bool buflog);
-
-  double L2Norm(Patch *Pat, var *vf, MPI_Comm Comm_here);
-  void L2Norm7(Patch *Pat, var **vf, double *norms, MPI_Comm Comm_here);
-  bool PatList_Interp_Points(MyList<Patch> *PatL, MyList<var> *VarList,
-                             int NN, double **XX,
-                             double *Shellf, int Symmetry, MPI_Comm Comm_here);
+
+  double L2Norm(Patch *Pat, var *vf, MPI_Comm Comm_here);
+  bool PatList_Interp_Points(MyList<Patch> *PatL, MyList<var> *VarList,
+                             int NN, double **XX,
+                             double *Shellf, int Symmetry, MPI_Comm Comm_here);
 #if (PSTR == 1 || PSTR == 2 || PSTR == 3)
   MyList<Block> *distribute(MyList<Patch> *PatchLIST, int cpusize, int ingfsi, int fngfsi,
                             bool periodic, int start_rank, int end_rank, int nodes = 0);

AMSS_NCKU_source/ShellPatch.C

@@ -3439,10 +3439,10 @@ void ShellPatch::write_Pablo_file_ss(int *ext, double xmin, double xmax, double
   delete[] Z;
 }
 
-double ShellPatch::L2Norm(var *vf)
-{
-  double tvf, dtvf = 0;
-  int BDW = overghost;
+double ShellPatch::L2Norm(var *vf)
+{
+  double tvf, dtvf = 0;
+  int BDW = overghost;
 
   MyList<ss_patch> *sPp = PatL;
   while (sPp)
@@ -3469,50 +3469,13 @@ double ShellPatch::L2Norm(var *vf)
   MPI_Allreduce(&dtvf, &tvf, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
 
   tvf = sqrt(tvf);
-
-  return tvf;
-}
-void ShellPatch::L2Norm7(var **vf, double *norms)
-{
-  double tvf[7], dtvf[7];
-  int BDW = overghost;
-  for (int i = 0; i < 7; i++)
-    dtvf[i] = 0;
-
-  MyList<ss_patch> *sPp = PatL;
-  while (sPp)
-  {
-    MyList<Block> *Bp = sPp->data->blb;
-    while (Bp)
-    {
-      Block *cg = Bp->data;
-      if (myrank == cg->rank)
-      {
-        f_l2normhelper7(cg->shape, cg->X[0], cg->X[1], cg->X[2],
-                        sPp->data->bbox[0], sPp->data->bbox[1], sPp->data->bbox[2],
-                        sPp->data->bbox[3], sPp->data->bbox[4], sPp->data->bbox[5],
-                        cg->fgfs[vf[0]->sgfn], cg->fgfs[vf[1]->sgfn], cg->fgfs[vf[2]->sgfn],
-                        cg->fgfs[vf[3]->sgfn], cg->fgfs[vf[4]->sgfn], cg->fgfs[vf[5]->sgfn],
-                        cg->fgfs[vf[6]->sgfn], tvf, BDW);
-        for (int i = 0; i < 7; i++)
-          dtvf[i] += tvf[i];
-      }
-      if (Bp == sPp->data->ble)
-        break;
-      Bp = Bp->next;
-    }
-    sPp = sPp->next;
-  }
-
-  MPI_Allreduce(dtvf, tvf, 7, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
-
-  for (int i = 0; i < 7; i++)
-    norms[i] = sqrt(tvf[i]);
-}
-
-// find maximum of abstract value, XX store position for maximum, Shellf store maximum themselvs
-void ShellPatch::Find_Maximum(MyList<var> *VarList, double *XX,
-                              double *Shellf)
+
+  return tvf;
+}
+
+// find maximum of abstract value, XX store position for maximum, Shellf store maximum themselvs
+void ShellPatch::Find_Maximum(MyList<var> *VarList, double *XX,
+                              double *Shellf)
 {
   MyList<var> *varl;
   int num_var = 0;

AMSS_NCKU_source/ShellPatch.h

@@ -195,11 +195,10 @@ public:
    bool Interp_One_Point(MyList<var> *VarList,
                          double *XX, /*input global Cartesian coordinate*/
                          double *Shellf, int Symmetry);
-   void write_Pablo_file_ss(int *ext, double xmin, double xmax, double ymin, double ymax, double zmin, double zmax,
-                            char *filename, int sst);
-   double L2Norm(var *vf);
-   void L2Norm7(var **vf, double *norms);
-   void Find_Maximum(MyList<var> *VarList, double *XX, double *Shellf);
-};
+   void write_Pablo_file_ss(int *ext, double xmin, double xmax, double ymin, double ymax, double zmin, double zmax,
+                            char *filename, int sst);
+   double L2Norm(var *vf);
+   void Find_Maximum(MyList<var> *VarList, double *XX, double *Shellf);
+};
 
 #endif /* SHELLPATCH_H */

AMSS_NCKU_source/TwoPunctures.C

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

AMSS_NCKU_source/bssnEM_class.C

@@ -258,8 +258,6 @@ void bssnEM_class::Initialize()
     PhysTime = StartTime;
     Setup_Black_Hole_position();
   }
-
-  setup_transfer_caches();
 }
 
 //================================================================================================

AMSS_NCKU_source/bssnEScalar_class.C

@@ -23,14 +23,8 @@ using namespace std;
 #include "rungekutta4_rout.h"
 #include "sommerfeld_rout.h"
 #include "getnp4.h"
-#include "shellfunctions.h"
-#include "parameters.h"
-
-#if BSSN_USE_ESCALAR_C_KERNEL
-#define BSSN_ESCALAR_RHS f_compute_rhs_bssn_escalar_c
-#else
-#define BSSN_ESCALAR_RHS f_compute_rhs_bssn_escalar
-#endif
+#include "shellfunctions.h"
+#include "parameters.h"
 
 #ifdef With_AHF
 #include "derivatives.h"
@@ -80,8 +74,8 @@ bssnEScalar_class::bssnEScalar_class(double Couranti, double StartTimei, double
 
 //================================================================================================
 
-void bssnEScalar_class::Initialize()
-{
+void bssnEScalar_class::Initialize()
+{
   Sphio = new var("Sphio", ngfs++, 1, 1, 1);
   Spio = new var("Spio", ngfs++, 1, 1, 1);
   Sphi0 = new var("Sphi0", ngfs++, 1, 1, 1);
@@ -138,14 +132,11 @@ void bssnEScalar_class::Initialize()
       }
     } 
    
-  GH = new cgh(0, ngfs, Symmetry, pname, checkrun, ErrorMonitor);
-  ConstraintRefreshLevels = new int[GH->levels];
-  for (int il = 0; il < GH->levels; il++)
-    ConstraintRefreshLevels[il] = 0;
-  if (checkrun)
-    CheckPoint->readcheck_cgh(PhysTime, GH, myrank, nprocs, Symmetry);
-  else
-    GH->compose_cgh(nprocs);
+  GH = new cgh(0, ngfs, Symmetry, pname, checkrun, ErrorMonitor);
+  if (checkrun)
+    CheckPoint->readcheck_cgh(PhysTime, GH, myrank, nprocs, Symmetry);
+  else
+    GH->compose_cgh(nprocs);
 
 #ifdef WithShell
   SH = new ShellPatch(0, ngfs, pname, Symmetry, myrank, ErrorMonitor);
@@ -169,14 +160,12 @@ void bssnEScalar_class::Initialize()
   {
     CheckPoint->read_Black_Hole_position(BH_num_input, BH_num, Porg0, Pmom, Spin, Mass, Porgbr, Porg, Porg1, Porg_rhs);
   }
-  else
-  {
-    PhysTime = StartTime;
-    Setup_Black_Hole_position();
-  }
-
-  setup_transfer_caches();
-}
+  else
+  {
+    PhysTime = StartTime;
+    Setup_Black_Hole_position();
+  }
+}
 
 //================================================================================================
 
@@ -218,10 +207,10 @@ bssnEScalar_class::~bssnEScalar_class()
 
 // Read initial data solved by Ansorg, PRD 70, 064011 (2004)
 
-void bssnEScalar_class::Read_Ansorg()
-{
-  if (!checkrun)
-  {
+void bssnEScalar_class::Read_Ansorg()
+{
+  if (!checkrun)
+  {
     if (myrank == 0)
       cout << "Read initial data from Ansorg's solver,"
            << " please be sure the input parameters for black holes are puncture parameters!!" 
@@ -238,12 +227,9 @@ void bssnEScalar_class::Read_Ansorg()
         cout << "Error inputpar" << endl;
         exit(0);
       }
-    }
-    int BH_NM;
-    double *Porg_here;
-    double *pmom_local;
-    double *spin_local;
-    double *mass_local;
+    }
+    int BH_NM;
+    double *Porg_here;
     // read parameter from file
     {
       const int LEN = 256;
@@ -283,11 +269,11 @@ void bssnEScalar_class::Read_Ansorg()
       }
       inf.close();
     }
-
-    Porg_here = new double[3 * BH_NM];
-    pmom_local = new double[3 * BH_NM];
-    spin_local = new double[3 * BH_NM];
-    mass_local = new double[BH_NM];
+
+    Porg_here = new double[3 * BH_NM];
+    Pmom = new double[3 * BH_NM];
+    Spin = new double[3 * BH_NM];
+    Mass = new double[BH_NM];
     // read parameter from file
     {
       const int LEN = 256;
@@ -319,37 +305,37 @@ void bssnEScalar_class::Read_Ansorg()
         else if (status == 0)
           continue;
 
-        if (sgrp == "BSSN" && sind < BH_NM)
-        {
-          if (skey == "Mass")
-            mass_local[sind] = atof(sval.c_str());
-          else if (skey == "Porgx")
-            Porg_here[sind * 3] = atof(sval.c_str());
-          else if (skey == "Porgy")
-            Porg_here[sind * 3 + 1] = atof(sval.c_str());
-          else if (skey == "Porgz")
-            Porg_here[sind * 3 + 2] = atof(sval.c_str());
-          else if (skey == "Spinx")
-            spin_local[sind * 3] = atof(sval.c_str());
-          else if (skey == "Spiny")
-            spin_local[sind * 3 + 1] = atof(sval.c_str());
-          else if (skey == "Spinz")
-            spin_local[sind * 3 + 2] = atof(sval.c_str());
-          else if (skey == "Pmomx")
-            pmom_local[sind * 3] = atof(sval.c_str());
-          else if (skey == "Pmomy")
-            pmom_local[sind * 3 + 1] = atof(sval.c_str());
-          else if (skey == "Pmomz")
-            pmom_local[sind * 3 + 2] = atof(sval.c_str());
-        }
-      }
-      inf.close();
+        if (sgrp == "BSSN" && sind < BH_NM)
+        {
+          if (skey == "Mass")
+            Mass[sind] = atof(sval.c_str());
+          else if (skey == "Porgx")
+            Porg_here[sind * 3] = atof(sval.c_str());
+          else if (skey == "Porgy")
+            Porg_here[sind * 3 + 1] = atof(sval.c_str());
+          else if (skey == "Porgz")
+            Porg_here[sind * 3 + 2] = atof(sval.c_str());
+          else if (skey == "Spinx")
+            Spin[sind * 3] = atof(sval.c_str());
+          else if (skey == "Spiny")
+            Spin[sind * 3 + 1] = atof(sval.c_str());
+          else if (skey == "Spinz")
+            Spin[sind * 3 + 2] = atof(sval.c_str());
+          else if (skey == "Pmomx")
+            Pmom[sind * 3] = atof(sval.c_str());
+          else if (skey == "Pmomy")
+            Pmom[sind * 3 + 1] = atof(sval.c_str());
+          else if (skey == "Pmomz")
+            Pmom[sind * 3 + 2] = atof(sval.c_str());
+        }
+      }
+      inf.close();
     }
-    int order = 6;
-    Ansorg read_ansorg("Ansorg.psid", order);
-    // set initial data
-    for (int lev = 0; lev < GH->levels; lev++)
-    {
+    int order = 6;
+    Ansorg read_ansorg("Ansorg.psid", order);
+    // set initial data
+    for (int lev = 0; lev < GH->levels; lev++)
+    {
       MyList<Patch> *Pp = GH->PatL[lev];
       while (Pp)
       {
@@ -372,21 +358,21 @@ void bssnEScalar_class::Read_Ansorg()
                                       cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn], 
                                       cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn],
                                       cg->fgfs[Gmx0->sgfn], cg->fgfs[Gmy0->sgfn], cg->fgfs[Gmz0->sgfn],
-                                      cg->fgfs[Lap0->sgfn], 
-                                      cg->fgfs[Sfx0->sgfn], cg->fgfs[Sfy0->sgfn], cg->fgfs[Sfz0->sgfn],
-                                      cg->fgfs[dtSfx0->sgfn], cg->fgfs[dtSfy0->sgfn], cg->fgfs[dtSfz0->sgfn],
-                                      cg->fgfs[Sphi0->sgfn], cg->fgfs[Spi0->sgfn],
-                                      mass_local, Porg_here, pmom_local, spin_local, BH_NM);
+                                      cg->fgfs[Lap0->sgfn], 
+                                      cg->fgfs[Sfx0->sgfn], cg->fgfs[Sfy0->sgfn], cg->fgfs[Sfz0->sgfn],
+                                      cg->fgfs[dtSfx0->sgfn], cg->fgfs[dtSfy0->sgfn], cg->fgfs[dtSfz0->sgfn],
+                                      cg->fgfs[Sphi0->sgfn], cg->fgfs[Spi0->sgfn],
+                                      Mass, Porg_here, Pmom, Spin, BH_NM);
           }
           if (BL == Pp->data->ble)
             break;
           BL = BL->next;
-        }
-        Pp = Pp->next;
-      }
-    }
-#ifdef WithShell
-    // ShellPatch part
+        }
+        Pp = Pp->next;
+      }
+    }
+#ifdef WithShell
+    // ShellPatch part
     MyList<ss_patch> *Pp = SH->PatL;
     while (Pp)
     {
@@ -414,28 +400,25 @@ void bssnEScalar_class::Read_Ansorg()
                                        cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn], 
                                        cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn],
                                        cg->fgfs[Gmx0->sgfn], cg->fgfs[Gmy0->sgfn], cg->fgfs[Gmz0->sgfn],
-                                       cg->fgfs[Lap0->sgfn], 
-                                       cg->fgfs[Sfx0->sgfn], cg->fgfs[Sfy0->sgfn], cg->fgfs[Sfz0->sgfn],
-                                       cg->fgfs[dtSfx0->sgfn], cg->fgfs[dtSfy0->sgfn], cg->fgfs[dtSfz0->sgfn],
-                                       cg->fgfs[Sphi0->sgfn], cg->fgfs[Spi0->sgfn],
-                                       mass_local, Porg_here, pmom_local, spin_local, BH_NM);
+                                       cg->fgfs[Lap0->sgfn], 
+                                       cg->fgfs[Sfx0->sgfn], cg->fgfs[Sfy0->sgfn], cg->fgfs[Sfz0->sgfn],
+                                       cg->fgfs[dtSfx0->sgfn], cg->fgfs[dtSfy0->sgfn], cg->fgfs[dtSfz0->sgfn],
+                                       cg->fgfs[Sphi0->sgfn], cg->fgfs[Spi0->sgfn],
+                                       Mass, Porg_here, Pmom, Spin, BH_NM);
         }
         if (BL == Pp->data->ble)
           break;
         BL = BL->next;
-      }
-      Pp = Pp->next;
-    }
-#endif
-
-    delete[] Porg_here;
-    delete[] pmom_local;
-    delete[] spin_local;
-    delete[] mass_local;
-    // dump read_in initial data
-    //   for(int lev=0;lev<GH->levels;lev++) Parallel::Dump_Data(GH->PatL[lev],StateList,0,PhysTime,dT);
-  }
-}
+      }
+      Pp = Pp->next;
+    }
+#endif
+
+    delete[] Porg_here;
+    // dump read_in initial data
+    //   for(int lev=0;lev<GH->levels;lev++) Parallel::Dump_Data(GH->PatL[lev],StateList,0,PhysTime,dT);
+  }
+}
 
 //================================================================================================
 
@@ -449,10 +432,10 @@ void bssnEScalar_class::Read_Ansorg()
 
 // Read initial data solved by Pablo's Olliptic Phys.Rev.D 82 024005 (2010)
 
-void bssnEScalar_class::Read_Pablo()
-{
-  if (!checkrun)
-  {
+void bssnEScalar_class::Read_Pablo()
+{
+  if (!checkrun)
+  {
     if (myrank == 0)
       cout << "Read initial data from Pablo's solver,"
            << " please be sure the input parameters for black holes are puncture parameters!!" 
@@ -469,12 +452,9 @@ void bssnEScalar_class::Read_Pablo()
         cout << "Error inputpar" << endl;
         exit(0);
       }
-    }
-    int BH_NM;
-    double *Porg_here;
-    double *pmom_local;
-    double *spin_local;
-    double *mass_local;
+    }
+    int BH_NM;
+    double *Porg_here;
     // read parameter from file
     {
       const int LEN = 256;
@@ -514,11 +494,11 @@ void bssnEScalar_class::Read_Pablo()
       }
       inf.close();
     }
-
-    Porg_here = new double[3 * BH_NM];
-    pmom_local = new double[3 * BH_NM];
-    spin_local = new double[3 * BH_NM];
-    mass_local = new double[BH_NM];
+
+    Porg_here = new double[3 * BH_NM];
+    Pmom = new double[3 * BH_NM];
+    Spin = new double[3 * BH_NM];
+    Mass = new double[BH_NM];
     // read parameter from file
     {
       const int LEN = 256;
@@ -550,31 +530,31 @@ void bssnEScalar_class::Read_Pablo()
         else if (status == 0)
           continue;
 
-        if (sgrp == "BSSN" && sind < BH_NM)
-        {
-          if (skey == "Mass")
-            mass_local[sind] = atof(sval.c_str());
-          else if (skey == "Porgx")
-            Porg_here[sind * 3] = atof(sval.c_str());
-          else if (skey == "Porgy")
-            Porg_here[sind * 3 + 1] = atof(sval.c_str());
-          else if (skey == "Porgz")
-            Porg_here[sind * 3 + 2] = atof(sval.c_str());
-          else if (skey == "Spinx")
-            spin_local[sind * 3] = atof(sval.c_str());
-          else if (skey == "Spiny")
-            spin_local[sind * 3 + 1] = atof(sval.c_str());
-          else if (skey == "Spinz")
-            spin_local[sind * 3 + 2] = atof(sval.c_str());
-          else if (skey == "Pmomx")
-            pmom_local[sind * 3] = atof(sval.c_str());
-          else if (skey == "Pmomy")
-            pmom_local[sind * 3 + 1] = atof(sval.c_str());
-          else if (skey == "Pmomz")
-            pmom_local[sind * 3 + 2] = atof(sval.c_str());
-        }
-      }
-      inf.close();
+        if (sgrp == "BSSN" && sind < BH_NM)
+        {
+          if (skey == "Mass")
+            Mass[sind] = atof(sval.c_str());
+          else if (skey == "Porgx")
+            Porg_here[sind * 3] = atof(sval.c_str());
+          else if (skey == "Porgy")
+            Porg_here[sind * 3 + 1] = atof(sval.c_str());
+          else if (skey == "Porgz")
+            Porg_here[sind * 3 + 2] = atof(sval.c_str());
+          else if (skey == "Spinx")
+            Spin[sind * 3] = atof(sval.c_str());
+          else if (skey == "Spiny")
+            Spin[sind * 3 + 1] = atof(sval.c_str());
+          else if (skey == "Spinz")
+            Spin[sind * 3 + 2] = atof(sval.c_str());
+          else if (skey == "Pmomx")
+            Pmom[sind * 3] = atof(sval.c_str());
+          else if (skey == "Pmomy")
+            Pmom[sind * 3 + 1] = atof(sval.c_str());
+          else if (skey == "Pmomz")
+            Pmom[sind * 3 + 2] = atof(sval.c_str());
+        }
+      }
+      inf.close();
     }
     bool flag = false;
     int DIM = dim;
@@ -614,11 +594,11 @@ void bssnEScalar_class::Read_Pablo()
                                         cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn], 
                                         cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn],
                                         cg->fgfs[Gmx0->sgfn], cg->fgfs[Gmy0->sgfn], cg->fgfs[Gmz0->sgfn],
-                                        cg->fgfs[Lap0->sgfn], 
-                                        cg->fgfs[Sfx0->sgfn], cg->fgfs[Sfy0->sgfn], cg->fgfs[Sfz0->sgfn],
-                                        cg->fgfs[dtSfx0->sgfn], cg->fgfs[dtSfy0->sgfn], cg->fgfs[dtSfz0->sgfn],
-                                        cg->fgfs[Sphi0->sgfn], cg->fgfs[Spi0->sgfn],
-                                        mass_local, Porg_here, pmom_local, spin_local, BH_NM);
+                                        cg->fgfs[Lap0->sgfn], 
+                                        cg->fgfs[Sfx0->sgfn], cg->fgfs[Sfy0->sgfn], cg->fgfs[Sfz0->sgfn],
+                                        cg->fgfs[dtSfx0->sgfn], cg->fgfs[dtSfy0->sgfn], cg->fgfs[dtSfz0->sgfn],
+                                        cg->fgfs[Sphi0->sgfn], cg->fgfs[Spi0->sgfn],
+                                        Mass, Porg_here, Pmom, Spin, BH_NM);
             }
             if (BL == Pp->data->ble)
               break;
@@ -678,11 +658,11 @@ void bssnEScalar_class::Read_Pablo()
                                          cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn], 
                                          cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn],
                                          cg->fgfs[Gmx0->sgfn], cg->fgfs[Gmy0->sgfn], cg->fgfs[Gmz0->sgfn],
-                                         cg->fgfs[Lap0->sgfn], 
-                                         cg->fgfs[Sfx0->sgfn], cg->fgfs[Sfy0->sgfn], cg->fgfs[Sfz0->sgfn],
-                                         cg->fgfs[dtSfx0->sgfn], cg->fgfs[dtSfy0->sgfn], cg->fgfs[dtSfz0->sgfn],
-                                         cg->fgfs[Sphi0->sgfn], cg->fgfs[Spi0->sgfn],
-                                         mass_local, Porg_here, pmom_local, spin_local, BH_NM);
+                                         cg->fgfs[Lap0->sgfn], 
+                                         cg->fgfs[Sfx0->sgfn], cg->fgfs[Sfy0->sgfn], cg->fgfs[Sfz0->sgfn],
+                                         cg->fgfs[dtSfx0->sgfn], cg->fgfs[dtSfy0->sgfn], cg->fgfs[dtSfz0->sgfn],
+                                         cg->fgfs[Sphi0->sgfn], cg->fgfs[Spi0->sgfn],
+                                         Mass, Porg_here, Pmom, Spin, BH_NM);
           }
           if (BL == Pp->data->ble)
             break;
@@ -704,13 +684,10 @@ void bssnEScalar_class::Read_Pablo()
       Pp = Pp->next;
     }
 #endif
-
-    delete[] Porg_here;
-    delete[] pmom_local;
-    delete[] spin_local;
-    delete[] mass_local;
-    if (flag && myrank == 0)
-      MPI_Abort(MPI_COMM_WORLD, 1);
+
+    delete[] Porg_here;
+    if (flag && myrank == 0)
+      MPI_Abort(MPI_COMM_WORLD, 1);
     // dump read_in initial data
     for (int lev = 0; lev < GH->levels; lev++)
       Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT);
@@ -762,10 +739,10 @@ void bssnEScalar_class::Step(int lev, int YN)
                      cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn]);
 #endif
 
-        if (BSSN_ESCALAR_RHS(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
-                                       cg->fgfs[phi0->sgfn], cg->fgfs[trK0->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],
+        if (f_compute_rhs_bssn_escalar(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
+                                       cg->fgfs[phi0->sgfn], cg->fgfs[trK0->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[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn],
                                        cg->fgfs[Gmx0->sgfn], cg->fgfs[Gmy0->sgfn], cg->fgfs[Gmz0->sgfn],
@@ -1016,12 +993,11 @@ void bssnEScalar_class::Step(int lev, int YN)
   }
 #endif
 
-  Parallel::AsyncSyncState async_pre;
-  sync_predictor_start(lev, SynchList_pre, async_pre);
+  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
 
-#ifdef WithShell
-  if (lev == 0)
-  {
+#ifdef WithShell
+  if (lev == 0)
+  {
     clock_t prev_clock, curr_clock;
     if (myrank == 0)
       curr_clock = clock();
@@ -1033,10 +1009,9 @@ void bssnEScalar_class::Step(int lev, int YN)
       cout << " Shell stuff synchronization used " 
            << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
            << " seconds! " << endl;
-    }
-  }
-#endif
-  sync_predictor_finish(lev, async_pre, SynchList_pre);
+    }
+  }
+#endif
 
   // for black hole position
   if (BH_num > 0 && lev == GH->levels - 1)
@@ -1106,10 +1081,10 @@ void bssnEScalar_class::Step(int lev, int YN)
                          cg->fgfs[Ayy->sgfn], cg->fgfs[Ayz->sgfn], cg->fgfs[Azz->sgfn]);
 #endif
 
-          if (BSSN_ESCALAR_RHS(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
-                                         cg->fgfs[phi->sgfn], cg->fgfs[trK->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],
+          if (f_compute_rhs_bssn_escalar(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
+                                         cg->fgfs[phi->sgfn], cg->fgfs[trK->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[Ayy->sgfn], cg->fgfs[Ayz->sgfn], cg->fgfs[Azz->sgfn],
                                          cg->fgfs[Gmx->sgfn], cg->fgfs[Gmy->sgfn], cg->fgfs[Gmz->sgfn],
@@ -1374,12 +1349,11 @@ void bssnEScalar_class::Step(int lev, int YN)
     }
 #endif
 
-    Parallel::AsyncSyncState async_cor;
-    sync_corrector_start(lev, SynchList_cor, async_cor);
+    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
 
-#ifdef WithShell
-    if (lev == 0)
-    {
+#ifdef WithShell
+    if (lev == 0)
+    {
       clock_t prev_clock, curr_clock;
       if (myrank == 0)
         curr_clock = clock();
@@ -1391,10 +1365,9 @@ void bssnEScalar_class::Step(int lev, int YN)
         cout << " Shell stuff synchronization used " 
              << (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC) 
              << " seconds! " << endl;
-      }
-    }
-#endif
-    sync_corrector_finish(lev, async_cor, SynchList_cor);
+      }
+    }
+#endif
     // for black hole position
     if (BH_num > 0 && lev == GH->levels - 1)
     {
@@ -1862,14 +1835,11 @@ void bssnEScalar_class::AnalysisStuff_EScalar(int lev, double dT_lev)
 
 //================================================================================================
 
-void bssnEScalar_class::Interp_Constraint(bool infg)
-{
-  if (!infg)
-    return;
-
-  // we do not support a_lev != 0 yet.
-  if (a_lev > 0)
-    return;
+void bssnEScalar_class::Interp_Constraint()
+{
+  // we do not support a_lev != 0 yet.
+  if (a_lev > 0)
+    return;
 
   for (int lev = 0; lev < GH->levels; lev++)
   {
@@ -1888,10 +1858,10 @@ void bssnEScalar_class::Interp_Constraint(bool infg)
           if (myrank == cg->rank)
           {
             if (lev > 0)
-              BSSN_ESCALAR_RHS(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
-                                         cg->fgfs[phi0->sgfn], cg->fgfs[trK0->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],
+              f_compute_rhs_bssn_escalar(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
+                                         cg->fgfs[phi0->sgfn], cg->fgfs[trK0->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[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn],
                                          cg->fgfs[Gmx0->sgfn], cg->fgfs[Gmy0->sgfn], cg->fgfs[Gmz0->sgfn],
@@ -2108,10 +2078,10 @@ void bssnEScalar_class::Constraint_Out()
             if (myrank == cg->rank)
             {
               if (lev > 0)
-                BSSN_ESCALAR_RHS(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
-                                           cg->fgfs[phi0->sgfn], cg->fgfs[trK0->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],
+                f_compute_rhs_bssn_escalar(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
+                                           cg->fgfs[phi0->sgfn], cg->fgfs[trK0->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[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn],
                                            cg->fgfs[Gmx0->sgfn], cg->fgfs[Gmy0->sgfn], cg->fgfs[Gmz0->sgfn],

AMSS_NCKU_source/bssnEScalar_class.h

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

AMSS_NCKU_source/bssn_class.h

@@ -31,19 +31,11 @@ using namespace std;
 #include "surface_integral.h"
 #include "checkpoint.h"
 
-extern void setpbh(int iBHN, double **iPBH, double *iMass, int rBHN);
-
-#ifndef BSSN_USE_TRANSFER_CACHE
-#define BSSN_USE_TRANSFER_CACHE 1
-#endif
-
-#ifndef BSSN_USE_ESCALAR_C_KERNEL
-#define BSSN_USE_ESCALAR_C_KERNEL 1
-#endif
-
-class bssn_class
-{
-public:
+extern void setpbh(int iBHN, double **iPBH, double *iMass, int rBHN);
+
+class bssn_class
+{
+public:
        int ngfs;
        int nprocs, myrank;
        cgh *GH;
@@ -53,11 +45,10 @@ public:
        int checkrun;
        char checkfilename[50];
        int Steps;
-       double StartTime, TotalTime;
-       double AnasTime, DumpTime, d2DumpTime, CheckTime;
-       double LastAnas, LastConsOut;
-       int *ConstraintRefreshLevels;
-       double Courant;
+       double StartTime, TotalTime;
+       double AnasTime, DumpTime, d2DumpTime, CheckTime;
+       double LastAnas, LastConsOut;
+       double Courant;
        double numepss, numepsb, numepsh;
        int Symmetry;
        int maxl, decn;
@@ -137,14 +128,15 @@ public:
 
        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]
-       Parallel::SyncCache *sync_cache_restrict;   // cached Restrict in RestrictProlong
-       Parallel::SyncCache *sync_cache_outbd;      // cached OutBdLow2Hi in RestrictProlong
+       Parallel::SyncCache *sync_cache_rp_coarse;  // RestrictProlong sync on PatL[lev-1]
+       Parallel::SyncCache *sync_cache_rp_fine;    // RestrictProlong sync on PatL[lev]
+       Parallel::SyncCache *sync_cache_restrict;   // cached Restrict in RestrictProlong
+       Parallel::SyncCache *sync_cache_outbd;      // cached OutBdLow2Hi in RestrictProlong
+       Parallel::SyncCache *sync_cache_psi4;       // cached Psi4 sync on PatL[lev]
 
-       monitor *ErrorMonitor, *Psi4Monitor, *BHMonitor, *MAPMonitor;
-       monitor *ConVMonitor, *TimingMonitor;
-       surface_integral *Waveshell;
+       monitor *ErrorMonitor, *Psi4Monitor, *BHMonitor, *MAPMonitor;
+       monitor *ConVMonitor;
+       surface_integral *Waveshell;
        checkpoint *CheckPoint;
 
 public:
@@ -175,32 +167,25 @@ public:
        void Setup_KerrSchild();
        void Enforce_algcon(int lev, int fg);
 
-       void testRestrict();
-       void testOutBd();
-       
-       bool check_Stdin_Abort(); 
-       bool use_transfer_cache() const;
-       void setup_transfer_caches();
-       void invalidate_transfer_caches();
-       void destroy_transfer_caches();
-       void sync_predictor_start(int lev, MyList<var> *VarList, Parallel::AsyncSyncState &async_state);
-       void sync_predictor_finish(int lev, Parallel::AsyncSyncState &async_state, MyList<var> *VarList);
-       void sync_corrector_start(int lev, MyList<var> *VarList, Parallel::AsyncSyncState &async_state);
-       void sync_corrector_finish(int lev, Parallel::AsyncSyncState &async_state, MyList<var> *VarList);
-       void sync_evolution(int lev, MyList<var> *VarList, Parallel::SyncCache *cache_array = 0);
-       void restrict_evolution(int lev, MyList<var> *src_var_list, MyList<var> *dst_var_list);
-       void outbdlow2hi_evolution(int lev, MyList<var> *src_var_list, MyList<var> *dst_var_list);
-
+       void testRestrict();
+       void testOutBd();
+       
+       bool check_Stdin_Abort(); 
+
        virtual void Setup_Initial_Data_Cao();
        virtual void Setup_Initial_Data_Lousto();
        virtual void Initialize();
-       virtual void Read_Ansorg();
-       virtual void Read_Pablo() {};
-       virtual void Compute_Psi4(int lev);
-       virtual void Step(int lev, int YN);
-       virtual void Interp_Constraint(bool infg);
-       virtual void Constraint_Out();
-       virtual void Compute_Constraint();
+       virtual void Read_Ansorg();
+       virtual void Read_Pablo() {};
+       void InvalidateSyncCaches();
+       virtual void Compute_Psi4(int lev);
+       virtual void Step(int lev, int YN);
+#ifdef USE_GPU
+       void Step_MainPath_GPU(int lev, int YN);
+#endif
+       virtual void Interp_Constraint(bool infg);
+       virtual void Constraint_Out();
+       virtual void Compute_Constraint();
 
 #ifdef With_AHF
 protected:

AMSS_NCKU_source/bssn_cuda_ops.h

@@ -0,0 +1,68 @@
+#ifndef BSSN_CUDA_OPS_H
+#define BSSN_CUDA_OPS_H
+
+int bssn_cuda_enforce_ga(int *ex,
+                         double *dxx, double *gxy, double *gxz,
+                         double *dyy, double *gyz, double *dzz,
+                         double *Axx, double *Axy, double *Axz,
+                         double *Ayy, double *Ayz, double *Azz);
+
+int bssn_cuda_rk4_boundary_var(int *ex, double dT,
+                               const double *X, const double *Y, const double *Z,
+                               double xmin, double ymin, double zmin,
+                               double xmax, double ymax, double zmax,
+                               const double *state0,
+                               const double *phi_field,
+                               const double *lap_field,
+                               const double *boundary_src,
+                               double *stage_data,
+                               double *rhs_accum,
+                               double propspeed,
+                               const double SoA[3],
+                               int symmetry,
+                               int lev,
+                               int rk_stage,
+                               bool force_host_boundary_fix,
+                               bool download_to_host = true);
+
+int bssn_cuda_rk4_boundary_batch(int *ex, double dT,
+                                 const double *X, const double *Y, const double *Z,
+                                 double xmin, double ymin, double zmin,
+                                 double xmax, double ymax, double zmax,
+                                 int symmetry,
+                                 const double *const *state0_list,
+                                 double *const *stage_data_list,
+                                 double *const *rhs_accum_list,
+                                 int num_var,
+                                 int rk_stage,
+                                 bool download_to_host = false);
+
+int bssn_cuda_lowerbound(int *ex, double *chi, double tinny, bool download_to_host = true);
+int bssn_cuda_download_buffer(int *ex, double *host_ptr);
+void bssn_cuda_release_rk4_caches();
+void bssn_cuda_release_interp_caches();
+
+int bssn_cuda_prolong3_pack(int wei,
+                            const double *llbc, const double *uubc, const int *extc, const double *func,
+                            const double *llbf, const double *uubf, const int *extf, double *funf,
+                            const double *llbp, const double *uubp,
+                            const double *SoA, int symmetry);
+
+int bssn_cuda_restrict3_pack(int wei,
+                             const double *llbc, const double *uubc, const int *extc, double *func,
+                             const double *llbf, const double *uubf, const int *extf, const double *funf,
+                             const double *llbr, const double *uubr,
+                             const double *SoA, int symmetry);
+
+int bssn_cuda_interp_points_batch(const int *ex,
+                                  const double *X, const double *Y, const double *Z,
+                                  const double *const *fields,
+                                  const double *soa_flat,
+                                  int num_var,
+                                  const double *px, const double *py, const double *pz,
+                                  int num_points,
+                                  int ordn,
+                                  int symmetry,
+                                  double *out);
+
+#endif

AMSS_NCKU_source/bssn_cuda_step.C

@@ -0,0 +1,936 @@
+#include "macrodef.h"
+
+#ifdef USE_GPU
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <cstdlib>
+#include <iomanip>
+#include <vector>
+
+#include "bssn_class.h"
+#include "bssn_cuda_ops.h"
+#include "bssn_gpu.h"
+#include "bssn_macro.h"
+
+namespace
+{
+enum StageProfileMetric
+{
+  STAGE_PROFILE_TOTAL = 0,
+  STAGE_PROFILE_RHS,
+  STAGE_PROFILE_RUN_STAGE,
+  STAGE_PROFILE_RUN_STAGE_DEVICE,
+  STAGE_PROFILE_RUN_STAGE_HOST_FIX,
+  STAGE_PROFILE_LOWERBOUND,
+  STAGE_PROFILE_ENSURE,
+  STAGE_PROFILE_DOWNLOAD,
+  STAGE_PROFILE_CLEAR_CACHE,
+  STAGE_PROFILE_SYNC_START,
+  STAGE_PROFILE_SYNC_FINISH,
+  STAGE_PROFILE_REFRESH,
+  STAGE_PROFILE_COUNT
+};
+
+static const int kStageProfileMaxLevels = 32;
+
+struct StageProfileStore
+{
+  bool env_checked;
+  bool enabled;
+  int calls[kStageProfileMaxLevels];
+  double metric[kStageProfileMaxLevels][STAGE_PROFILE_COUNT];
+};
+
+StageProfileStore &stage_profile_store()
+{
+  static StageProfileStore store = {};
+  return store;
+}
+
+bool stage_profile_enabled()
+{
+  StageProfileStore &store = stage_profile_store();
+  if (!store.env_checked)
+  {
+    const char *env = getenv("AMSS_GPU_STAGE_TIMING");
+    store.enabled = (env && env[0] && strcmp(env, "0") != 0);
+    store.env_checked = true;
+  }
+  return store.enabled;
+}
+
+void stage_profile_note_call(int lev)
+{
+  if (lev >= 0 && lev < kStageProfileMaxLevels)
+    stage_profile_store().calls[lev]++;
+}
+
+void stage_profile_add(int lev, StageProfileMetric metric, double seconds)
+{
+  if (lev >= 0 && lev < kStageProfileMaxLevels)
+    stage_profile_store().metric[lev][metric] += seconds;
+}
+
+const char *stage_profile_metric_name(StageProfileMetric metric)
+{
+  switch (metric)
+  {
+  case STAGE_PROFILE_TOTAL:
+    return "total";
+  case STAGE_PROFILE_RHS:
+    return "rhs";
+  case STAGE_PROFILE_RUN_STAGE:
+    return "run_stage";
+  case STAGE_PROFILE_RUN_STAGE_DEVICE:
+    return "run_stage_dev";
+  case STAGE_PROFILE_RUN_STAGE_HOST_FIX:
+    return "run_stage_host";
+  case STAGE_PROFILE_LOWERBOUND:
+    return "lower";
+  case STAGE_PROFILE_ENSURE:
+    return "ensure";
+  case STAGE_PROFILE_DOWNLOAD:
+    return "download";
+  case STAGE_PROFILE_CLEAR_CACHE:
+    return "clear_cache";
+  case STAGE_PROFILE_SYNC_START:
+    return "sync_start";
+  case STAGE_PROFILE_SYNC_FINISH:
+    return "sync_finish";
+  case STAGE_PROFILE_REFRESH:
+    return "refresh";
+  default:
+    return "unknown";
+  }
+}
+} // namespace
+
+void bssn_cuda_dump_stage_profile()
+{
+  if (!stage_profile_enabled())
+    return;
+
+  int myrank = 0;
+  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
+
+  StageProfileStore &store = stage_profile_store();
+  int global_calls_sum[kStageProfileMaxLevels] = {};
+  double global_metric_sum[kStageProfileMaxLevels][STAGE_PROFILE_COUNT] = {};
+  double global_metric_max[kStageProfileMaxLevels][STAGE_PROFILE_COUNT] = {};
+
+  MPI_Reduce(store.calls, global_calls_sum, kStageProfileMaxLevels, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
+  MPI_Reduce(store.metric[0], global_metric_sum[0],
+             kStageProfileMaxLevels * STAGE_PROFILE_COUNT,
+             MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
+  MPI_Reduce(store.metric[0], global_metric_max[0],
+             kStageProfileMaxLevels * STAGE_PROFILE_COUNT,
+             MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
+
+  if (myrank != 0)
+    return;
+
+  cout << endl;
+  cout << " GPU stage timing summary (sum/max over MPI ranks) " << endl;
+  cout << " lev  calls";
+  for (int metric = 0; metric < STAGE_PROFILE_COUNT; ++metric)
+    cout << "  " << setw(22) << stage_profile_metric_name(static_cast<StageProfileMetric>(metric));
+  cout << endl;
+
+  for (int lev = 0; lev < kStageProfileMaxLevels; ++lev)
+  {
+    if (global_calls_sum[lev] == 0)
+      continue;
+
+    cout << setw(4) << lev << "  " << setw(5) << global_calls_sum[lev];
+    for (int metric = 0; metric < STAGE_PROFILE_COUNT; ++metric)
+    {
+      cout << "  "
+           << setw(10) << setprecision(6) << fixed << global_metric_sum[lev][metric]
+           << "/"
+           << setw(10) << setprecision(6) << fixed << global_metric_max[lev][metric];
+    }
+    cout << endl;
+  }
+  cout << endl;
+}
+
+void bssn_class::Step_MainPath_GPU(int lev, int YN)
+{
+#ifdef WithShell
+#error "Step_MainPath_GPU currently supports Patch grids only."
+#endif
+
+  const bool profile_enabled = stage_profile_enabled();
+  const double step_total_begin = profile_enabled ? MPI_Wtime() : 0.0;
+  if (profile_enabled)
+    stage_profile_note_call(lev);
+
+  if (bssn_gpu_bind_process_device(myrank))
+  {
+    cerr << "GPU device bind failure on MPI rank " << myrank << endl;
+    MPI_Abort(MPI_COMM_WORLD, 1);
+  }
+  if (profile_enabled)
+  {
+    const double t0 = MPI_Wtime();
+    bssn_gpu_clear_cached_device_buffers();
+    stage_profile_add(lev, STAGE_PROFILE_CLEAR_CACHE, MPI_Wtime() - t0);
+  }
+  else
+    bssn_gpu_clear_cached_device_buffers();
+
+  setpbh(BH_num, Porg0, Mass, BH_num_input);
+
+  const double dT_lev = dT * pow(0.5, Mymax(lev, trfls));
+
+#if (MAPBH == 1)
+  if (BH_num > 0 && lev == GH->levels - 1)
+  {
+    compute_Porg_rhs(Porg0, Porg_rhs, Sfx0, Sfy0, Sfz0, lev);
+    for (int ithBH = 0; ithBH < BH_num; ithBH++)
+    {
+      for (int ith = 0; ith < 3; ith++)
+        Porg1[ithBH][ith] = Porg0[ithBH][ith] + Porg_rhs[ithBH][ith] * dT_lev;
+      if (Symmetry > 0)
+        Porg1[ithBH][2] = fabs(Porg1[ithBH][2]);
+      if (Symmetry == 2)
+      {
+        Porg1[ithBH][0] = fabs(Porg1[ithBH][0]);
+        Porg1[ithBH][1] = fabs(Porg1[ithBH][1]);
+      }
+    }
+  }
+
+  if (lev == a_lev)
+    AnalysisStuff(lev, dT_lev);
+#endif
+
+#ifdef With_AHF
+  AH_Step_Find(lev, dT_lev);
+#endif
+
+  const bool BB = fgt(PhysTime, StartTime, dT_lev / 2);
+  (void)BB;
+  double ndeps = (lev < GH->movls) ? numepsb : numepss;
+  double TRK4 = PhysTime;
+  int iter_count = 0;
+  int pre = 0, cor = 1;
+  int ERROR = 0;
+  const bool keep_stage_sync_on_device = (RPS == 1) && (MAPBH == 1) && (REGLEV == 0);
+
+  auto run_stage_on_block =
+      [&](Block *cg, Patch *patch, MyList<var> *state0_list,
+          MyList<var> *boundary_src_list, MyList<var> *stage_data_list,
+          MyList<var> *rhs_list, int rk_stage) {
+        MyList<var> *varl0 = state0_list;
+        MyList<var> *varlb = boundary_src_list;
+        MyList<var> *varls = stage_data_list;
+        MyList<var> *varlr = rhs_list;
+        std::vector<const double *> batch_state0;
+        std::vector<double *> batch_stage;
+        std::vector<double *> batch_rhs;
+
+        while (varl0)
+        {
+          const bool force_host_boundary_fix = false;
+          const bool can_batch_device_path = (lev > 0) && !force_host_boundary_fix;
+          if (can_batch_device_path)
+          {
+            batch_state0.push_back(cg->fgfs[varl0->data->sgfn]);
+            batch_stage.push_back(cg->fgfs[varls->data->sgfn]);
+            batch_rhs.push_back(cg->fgfs[varlr->data->sgfn]);
+            varl0 = varl0->next;
+            varlb = varlb->next;
+            varls = varls->next;
+            varlr = varlr->next;
+            continue;
+          }
+
+          const double var_begin = profile_enabled ? MPI_Wtime() : 0.0;
+          if (bssn_cuda_rk4_boundary_var(cg->shape, dT_lev,
+                                         cg->X[0], cg->X[1], cg->X[2],
+                                         patch->bbox[0], patch->bbox[1], patch->bbox[2],
+                                         patch->bbox[3], patch->bbox[4], patch->bbox[5],
+                                         cg->fgfs[varl0->data->sgfn],
+                                         cg->fgfs[phi0->sgfn],
+                                         cg->fgfs[Lap0->sgfn],
+                                         cg->fgfs[varlb->data->sgfn],
+                                         cg->fgfs[varls->data->sgfn],
+                                         cg->fgfs[varlr->data->sgfn],
+                                         varl0->data->propspeed,
+                                         varl0->data->SoA,
+                                         Symmetry, lev, rk_stage,
+                                         force_host_boundary_fix, false))
+          {
+            cerr << "GPU rk4/boundary failure: lev=" << lev
+                 << " rk_stage=" << rk_stage
+                 << " var=" << varl0->data->name
+                 << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                 << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                 << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+            ERROR = 1;
+            break;
+          }
+          if (profile_enabled)
+          {
+            stage_profile_add(lev,
+                              force_host_boundary_fix ? STAGE_PROFILE_RUN_STAGE_HOST_FIX
+                                                      : STAGE_PROFILE_RUN_STAGE_DEVICE,
+                              MPI_Wtime() - var_begin);
+          }
+          varl0 = varl0->next;
+          varlb = varlb->next;
+          varls = varls->next;
+          varlr = varlr->next;
+        }
+
+        if (!ERROR && !batch_state0.empty())
+        {
+          const double batch_begin = profile_enabled ? MPI_Wtime() : 0.0;
+          if (bssn_cuda_rk4_boundary_batch(cg->shape, dT_lev,
+                                           cg->X[0], cg->X[1], cg->X[2],
+                                           patch->bbox[0], patch->bbox[1], patch->bbox[2],
+                                           patch->bbox[3], patch->bbox[4], patch->bbox[5],
+                                           Symmetry,
+                                           &batch_state0[0],
+                                           &batch_stage[0],
+                                           &batch_rhs[0],
+                                           static_cast<int>(batch_state0.size()),
+                                           rk_stage, false))
+          {
+            cerr << "GPU rk4/boundary batch failure: lev=" << lev
+                 << " rk_stage=" << rk_stage
+                 << " vars=" << batch_state0.size()
+                 << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                 << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                 << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+            ERROR = 1;
+          }
+          else if (profile_enabled)
+          {
+            stage_profile_add(lev, STAGE_PROFILE_RUN_STAGE_DEVICE, MPI_Wtime() - batch_begin);
+          }
+        }
+      };
+
+  auto stage_download_var_list =
+      [&](Block *cg, MyList<var> *var_list, bool skip_unmapped) {
+        std::vector<double *> batch_host_ptrs;
+        std::vector<MyList<var> *> batch_vars;
+        while (var_list)
+        {
+          double *host_ptr = cg->fgfs[var_list->data->sgfn];
+          if (skip_unmapped && !bssn_gpu_find_device_buffer(host_ptr))
+          {
+            var_list = var_list->next;
+            continue;
+          }
+          batch_host_ptrs.push_back(host_ptr);
+          batch_vars.push_back(var_list);
+          var_list = var_list->next;
+        }
+        if (!batch_host_ptrs.empty() &&
+            bssn_gpu_download_buffer_batch(cg->shape, &batch_host_ptrs[0],
+                                           static_cast<int>(batch_host_ptrs.size())))
+        {
+          for (size_t i = 0; i < batch_host_ptrs.size(); ++i)
+          {
+            if (bssn_cuda_download_buffer(cg->shape, batch_host_ptrs[i]))
+            {
+              cerr << "GPU stage download failure: lev=" << lev
+                   << " var=" << batch_vars[i]->data->name
+                   << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                   << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                   << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+              ERROR = 1;
+              break;
+            }
+          }
+        }
+      };
+
+  auto stage_download_patch_list =
+      [&](MyList<var> *var_list, bool skip_unmapped) {
+        MyList<Patch> *patch_it = GH->PatL[lev];
+        while (patch_it)
+        {
+          MyList<Block> *block_it = patch_it->data->blb;
+          while (block_it)
+          {
+            Block *cg = block_it->data;
+            if (myrank == cg->rank)
+              stage_download_var_list(cg, var_list, skip_unmapped);
+
+            if (block_it == patch_it->data->ble)
+              break;
+            block_it = block_it->next;
+          }
+          if (ERROR)
+            break;
+          patch_it = patch_it->next;
+        }
+      };
+
+  auto ensure_stage_device_var_list =
+      [&](Block *cg, MyList<var> *var_list) {
+        const int n = cg->shape[0] * cg->shape[1] * cg->shape[2];
+        while (var_list)
+        {
+          double *host_ptr = cg->fgfs[var_list->data->sgfn];
+          if (!bssn_gpu_find_device_buffer(host_ptr) &&
+              bssn_gpu_stage_upload_buffer(host_ptr, n))
+          {
+            cerr << "GPU state ensure failure: lev=" << lev
+                 << " var=" << var_list->data->name
+                 << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                 << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                 << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+            ERROR = 1;
+            break;
+          }
+          var_list = var_list->next;
+        }
+      };
+
+  auto refresh_synced_device_regions =
+      [&](Block *cg, MyList<var> *var_list, Parallel::SyncCache &cache) {
+        std::vector<Parallel::gridseg *> local_segments;
+        for (int node = 0; node < cache.cpusize; ++node)
+        {
+          MyList<Parallel::gridseg> *seg = cache.combined_dst[node];
+          while (seg)
+          {
+            if (seg->data && seg->data->Bg == cg)
+              local_segments.push_back(seg->data);
+            seg = seg->next;
+          }
+        }
+
+        if (local_segments.empty())
+          return;
+
+        const int n = cg->shape[0] * cg->shape[1] * cg->shape[2];
+        while (var_list)
+        {
+          double *host_ptr = cg->fgfs[var_list->data->sgfn];
+          if (!bssn_gpu_find_device_buffer(host_ptr))
+          {
+            if (bssn_gpu_stage_upload_buffer(host_ptr, n))
+            {
+              cerr << "GPU sync refresh upload failure: lev=" << lev
+                   << " var=" << var_list->data->name
+                   << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                   << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                   << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+              ERROR = 1;
+              break;
+            }
+          }
+          else
+          {
+            for (size_t i = 0; i < local_segments.size(); ++i)
+            {
+              Parallel::gridseg *seg = local_segments[i];
+              if (bssn_gpu_stage_upload_region(host_ptr,
+                                               cg->shape,
+                                               cg->bbox,
+                                               cg->bbox + dim,
+                                               seg->shape,
+                                               seg->llb))
+              {
+                cerr << "GPU sync region refresh failure: lev=" << lev
+                     << " var=" << var_list->data->name
+                     << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                     << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                     << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+                ERROR = 1;
+                break;
+              }
+            }
+            if (ERROR)
+              break;
+          }
+          var_list = var_list->next;
+        }
+      };
+
+  auto refresh_stage_device_after_sync =
+      [&](MyList<var> *var_list, Parallel::SyncCache &cache) {
+        MyList<Patch> *patch_it = GH->PatL[lev];
+        while (patch_it)
+        {
+          MyList<Block> *block_it = patch_it->data->blb;
+          while (block_it)
+          {
+            Block *cg = block_it->data;
+            if (myrank == cg->rank)
+              refresh_synced_device_regions(cg, var_list, cache);
+
+            if (block_it == patch_it->data->ble)
+              break;
+            block_it = block_it->next;
+          }
+          if (ERROR)
+            break;
+          patch_it = patch_it->next;
+        }
+      };
+
+  auto refresh_stage_host_before_sync =
+      [&](MyList<var> *var_list, Parallel::SyncCache &cache) -> bool {
+        if (!cache.valid || !cache.combined_src || myrank < 0 || myrank >= cache.cpusize)
+          return false;
+
+        MyList<Patch> *patch_it = GH->PatL[lev];
+        while (patch_it)
+        {
+          MyList<Block> *block_it = patch_it->data->blb;
+          while (block_it)
+          {
+            Block *cg = block_it->data;
+            if (myrank == cg->rank)
+            {
+              std::vector<Parallel::gridseg *> local_segments;
+              MyList<Parallel::gridseg> *seg = cache.combined_src[myrank];
+              while (seg)
+              {
+                if (seg->data && seg->data->Bg == cg)
+                  local_segments.push_back(seg->data);
+                seg = seg->next;
+              }
+
+              if (!local_segments.empty())
+              {
+                MyList<var> *var_it = var_list;
+                while (var_it)
+                {
+                  double *host_ptr = cg->fgfs[var_it->data->sgfn];
+                  for (size_t i = 0; i < local_segments.size(); ++i)
+                  {
+                    Parallel::gridseg *src_seg = local_segments[i];
+                    if (bssn_gpu_stage_download_region(host_ptr,
+                                                       cg->shape,
+                                                       cg->bbox,
+                                                       cg->bbox + dim,
+                                                       src_seg->shape,
+                                                       src_seg->llb))
+                    {
+                      cerr << "GPU sync region download failure: lev=" << lev
+                           << " var=" << var_it->data->name
+                           << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                           << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                           << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+                      ERROR = 1;
+                      return true;
+                    }
+                  }
+                  var_it = var_it->next;
+                }
+              }
+            }
+
+            if (block_it == patch_it->data->ble)
+              break;
+            block_it = block_it->next;
+          }
+          patch_it = patch_it->next;
+        }
+
+        return true;
+      };
+
+  auto can_pack_sync_from_device =
+      [&](MyList<var> *var_list, Parallel::SyncCache &cache) -> bool {
+        if (!cache.valid || !cache.combined_src || myrank < 0 || myrank >= cache.cpusize)
+          return false;
+
+        MyList<Parallel::gridseg> *seg = cache.combined_src[myrank];
+        while (seg)
+        {
+          MyList<var> *var_it = var_list;
+          while (var_it)
+          {
+            if (!bssn_gpu_find_device_buffer(seg->data->Bg->fgfs[var_it->data->sgfn]))
+              return false;
+            var_it = var_it->next;
+          }
+          seg = seg->next;
+        }
+        return true;
+      };
+
+  MyList<Patch> *Pp = GH->PatL[lev];
+  while (Pp)
+  {
+    MyList<Block> *BP = Pp->data->blb;
+    while (BP)
+    {
+      Block *cg = BP->data;
+      if (myrank == cg->rank)
+      {
+        double t0 = 0.0;
+        if (profile_enabled)
+          t0 = MPI_Wtime();
+        if (gpu_rhs(CALLED_BY_STEP, myrank, RHS_PARA_CALLED_FIRST_TIME))
+          ERROR = 1;
+        if (profile_enabled)
+          stage_profile_add(lev, STAGE_PROFILE_RHS, MPI_Wtime() - t0);
+
+        if (profile_enabled)
+          t0 = MPI_Wtime();
+        run_stage_on_block(cg, Pp->data, StateList, StateList, SynchList_pre, RHSList, iter_count);
+        if (profile_enabled)
+          stage_profile_add(lev, STAGE_PROFILE_RUN_STAGE, MPI_Wtime() - t0);
+
+        if (profile_enabled)
+          t0 = MPI_Wtime();
+        if (bssn_cuda_lowerbound(cg->shape, cg->fgfs[phi->sgfn], chitiny, false))
+        {
+          cerr << "GPU lowerbound failure: lev=" << lev
+               << " rk_stage=" << iter_count
+               << " var=" << phi->name
+               << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+               << cg->bbox[1] << ":" << cg->bbox[4] << ","
+               << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+          ERROR = 1;
+        }
+        if (profile_enabled)
+          stage_profile_add(lev, STAGE_PROFILE_LOWERBOUND, MPI_Wtime() - t0);
+      }
+      if (BP == Pp->data->ble)
+        break;
+      BP = BP->next;
+    }
+    Pp = Pp->next;
+  }
+
+  if (!ERROR)
+  {
+    if (!keep_stage_sync_on_device)
+    {
+      double t0 = 0.0;
+      if (profile_enabled)
+        t0 = MPI_Wtime();
+      stage_download_patch_list(SynchList_pre, false);
+      if (profile_enabled)
+        stage_profile_add(lev, STAGE_PROFILE_DOWNLOAD, MPI_Wtime() - t0);
+      if (!ERROR)
+      {
+        if (profile_enabled)
+          t0 = MPI_Wtime();
+        bssn_gpu_clear_cached_device_buffers();
+        if (profile_enabled)
+          stage_profile_add(lev, STAGE_PROFILE_CLEAR_CACHE, MPI_Wtime() - t0);
+      }
+    }
+  }
+
+  MPI_Request err_req_pre;
+  {
+    int erh = ERROR;
+    MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_pre);
+  }
+
+  Parallel::AsyncSyncState async_pre;
+  if (profile_enabled)
+  {
+    const double t0 = MPI_Wtime();
+    Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
+    stage_profile_add(lev, STAGE_PROFILE_SYNC_START, MPI_Wtime() - t0);
+  }
+  else
+    Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
+  if (profile_enabled)
+  {
+    const double t0 = MPI_Wtime();
+    Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry,
+                          !keep_stage_sync_on_device);
+    stage_profile_add(lev, STAGE_PROFILE_SYNC_FINISH, MPI_Wtime() - t0);
+  }
+  else
+    Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry,
+                          !keep_stage_sync_on_device);
+  if (!ERROR && !keep_stage_sync_on_device)
+  {
+    if (profile_enabled)
+    {
+      const double t0 = MPI_Wtime();
+      refresh_stage_device_after_sync(SynchList_pre, sync_cache_pre[lev]);
+      stage_profile_add(lev, STAGE_PROFILE_REFRESH, MPI_Wtime() - t0);
+    }
+    else
+      refresh_stage_device_after_sync(SynchList_pre, sync_cache_pre[lev]);
+  }
+
+  MPI_Wait(&err_req_pre, MPI_STATUS_IGNORE);
+  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);
+    }
+  }
+
+#if (MAPBH == 0)
+  if (BH_num > 0 && lev == GH->levels - 1)
+  {
+    compute_Porg_rhs(Porg0, Porg_rhs, Sfx0, Sfy0, Sfz0, lev);
+    for (int ithBH = 0; ithBH < BH_num; ithBH++)
+    {
+      f_rungekutta4_scalar(dT_lev, Porg0[ithBH][0], Porg[ithBH][0], Porg_rhs[ithBH][0], iter_count);
+      f_rungekutta4_scalar(dT_lev, Porg0[ithBH][1], Porg[ithBH][1], Porg_rhs[ithBH][1], iter_count);
+      f_rungekutta4_scalar(dT_lev, Porg0[ithBH][2], Porg[ithBH][2], Porg_rhs[ithBH][2], iter_count);
+      if (Symmetry > 0)
+        Porg[ithBH][2] = fabs(Porg[ithBH][2]);
+      if (Symmetry == 2)
+      {
+        Porg[ithBH][0] = fabs(Porg[ithBH][0]);
+        Porg[ithBH][1] = fabs(Porg[ithBH][1]);
+      }
+    }
+  }
+
+  if (lev == a_lev)
+    AnalysisStuff(lev, dT_lev);
+#endif
+
+  for (iter_count = 1; iter_count < 4; iter_count++)
+  {
+    if (iter_count == 1 || iter_count == 3)
+      TRK4 += dT_lev / 2;
+
+    Pp = GH->PatL[lev];
+    while (Pp)
+    {
+      MyList<Block> *BP = Pp->data->blb;
+      while (BP)
+      {
+        Block *cg = BP->data;
+        if (myrank == cg->rank)
+        {
+          double t0 = 0.0;
+          if (profile_enabled)
+            t0 = MPI_Wtime();
+          ensure_stage_device_var_list(cg, SynchList_pre);
+          if (profile_enabled)
+            stage_profile_add(lev, STAGE_PROFILE_ENSURE, MPI_Wtime() - t0);
+
+          if (profile_enabled)
+            t0 = MPI_Wtime();
+          if (gpu_rhs(CALLED_BY_STEP, myrank, RHS_PARA_CALLED_THEN))
+            ERROR = 1;
+          if (profile_enabled)
+            stage_profile_add(lev, STAGE_PROFILE_RHS, MPI_Wtime() - t0);
+
+          if (profile_enabled)
+            t0 = MPI_Wtime();
+          run_stage_on_block(cg, Pp->data, StateList, SynchList_pre, SynchList_cor, RHSList, iter_count);
+          if (profile_enabled)
+            stage_profile_add(lev, STAGE_PROFILE_RUN_STAGE, MPI_Wtime() - t0);
+
+          if (profile_enabled)
+            t0 = MPI_Wtime();
+          if (bssn_cuda_lowerbound(cg->shape, cg->fgfs[phi1->sgfn], chitiny, false))
+          {
+          cerr << "GPU lowerbound failure: lev=" << lev
+               << " rk_stage=" << iter_count
+               << " var=" << phi1->name
+                 << " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                 << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                 << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+            ERROR = 1;
+          }
+          if (profile_enabled)
+            stage_profile_add(lev, STAGE_PROFILE_LOWERBOUND, MPI_Wtime() - t0);
+        }
+
+        if (BP == Pp->data->ble)
+          break;
+        BP = BP->next;
+      }
+      Pp = Pp->next;
+    }
+
+    if (!ERROR)
+    {
+      if (!keep_stage_sync_on_device)
+      {
+        double t0 = 0.0;
+        if (profile_enabled)
+          t0 = MPI_Wtime();
+        stage_download_patch_list(SynchList_cor, false);
+        if (profile_enabled)
+          stage_profile_add(lev, STAGE_PROFILE_DOWNLOAD, MPI_Wtime() - t0);
+        if (!ERROR)
+        {
+          if (profile_enabled)
+            t0 = MPI_Wtime();
+          bssn_gpu_clear_cached_device_buffers();
+          if (profile_enabled)
+            stage_profile_add(lev, STAGE_PROFILE_CLEAR_CACHE, MPI_Wtime() - t0);
+        }
+      }
+    }
+
+    MPI_Request err_req_cor;
+    {
+      int erh = ERROR;
+      MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
+    }
+
+    Parallel::AsyncSyncState async_cor;
+    if (profile_enabled)
+    {
+      const double t0 = MPI_Wtime();
+      Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
+      stage_profile_add(lev, STAGE_PROFILE_SYNC_START, MPI_Wtime() - t0);
+    }
+    else
+      Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
+    if (profile_enabled)
+    {
+      const double t0 = MPI_Wtime();
+      Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry,
+                            !keep_stage_sync_on_device);
+      stage_profile_add(lev, STAGE_PROFILE_SYNC_FINISH, MPI_Wtime() - t0);
+    }
+    else
+      Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry,
+                            !keep_stage_sync_on_device);
+    if (!ERROR && !keep_stage_sync_on_device && iter_count < 3)
+    {
+      if (profile_enabled)
+      {
+        const double t0 = MPI_Wtime();
+        refresh_stage_device_after_sync(SynchList_cor, sync_cache_cor[lev]);
+        stage_profile_add(lev, STAGE_PROFILE_REFRESH, MPI_Wtime() - t0);
+      }
+      else
+        refresh_stage_device_after_sync(SynchList_cor, sync_cache_cor[lev]);
+    }
+
+    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
+                                << ", lev = " << lev << endl;
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
+
+#if (MAPBH == 0)
+    if (BH_num > 0 && lev == GH->levels - 1)
+    {
+      compute_Porg_rhs(Porg, Porg1, Sfx, Sfy, Sfz, lev);
+      for (int ithBH = 0; ithBH < BH_num; ithBH++)
+      {
+        f_rungekutta4_scalar(dT_lev, Porg0[ithBH][0], Porg1[ithBH][0], Porg_rhs[ithBH][0], iter_count);
+        f_rungekutta4_scalar(dT_lev, Porg0[ithBH][1], Porg1[ithBH][1], Porg_rhs[ithBH][1], iter_count);
+        f_rungekutta4_scalar(dT_lev, Porg0[ithBH][2], Porg1[ithBH][2], Porg_rhs[ithBH][2], iter_count);
+        if (Symmetry > 0)
+          Porg1[ithBH][2] = fabs(Porg1[ithBH][2]);
+        if (Symmetry == 2)
+        {
+          Porg1[ithBH][0] = fabs(Porg1[ithBH][0]);
+          Porg1[ithBH][1] = fabs(Porg1[ithBH][1]);
+        }
+      }
+    }
+#endif
+
+    if (iter_count < 3)
+    {
+      Pp = GH->PatL[lev];
+      while (Pp)
+      {
+        MyList<Block> *BP = Pp->data->blb;
+        while (BP)
+        {
+          BP->data->swapList(SynchList_pre, SynchList_cor, myrank);
+          if (BP == Pp->data->ble)
+            break;
+          BP = BP->next;
+        }
+        Pp = Pp->next;
+      }
+
+#if (MAPBH == 0)
+      if (BH_num > 0 && lev == GH->levels - 1)
+      {
+        for (int ithBH = 0; ithBH < BH_num; ithBH++)
+        {
+          Porg[ithBH][0] = Porg1[ithBH][0];
+          Porg[ithBH][1] = Porg1[ithBH][1];
+          Porg[ithBH][2] = Porg1[ithBH][2];
+        }
+      }
+#endif
+    }
+  }
+
+#if (RPS == 0)
+  RestrictProlong(lev, YN, BB);
+#endif
+
+  Pp = GH->PatL[lev];
+  while (Pp)
+  {
+    MyList<Block> *BP = Pp->data->blb;
+    while (BP)
+    {
+      Block *cg = BP->data;
+      cg->swapList(StateList, SynchList_cor, myrank);
+      cg->swapList(OldStateList, SynchList_cor, myrank);
+      if (BP == Pp->data->ble)
+        break;
+      BP = BP->next;
+    }
+    Pp = Pp->next;
+  }
+
+  if (!ERROR && keep_stage_sync_on_device)
+  {
+    // After the swaps above, only StateList points at arrays updated during this step.
+    // OldStateList/SynchList_cor remain valid on host because their backing arrays were
+    // read-only during the RK step, and SynchList_pre is reused only as scratch later.
+    const double t0 = profile_enabled ? MPI_Wtime() : 0.0;
+    stage_download_patch_list(StateList, true);
+    if (profile_enabled)
+      stage_profile_add(lev, STAGE_PROFILE_DOWNLOAD, MPI_Wtime() - t0);
+  }
+
+  if (profile_enabled)
+  {
+    const double t0 = MPI_Wtime();
+    bssn_gpu_clear_cached_device_buffers();
+    stage_profile_add(lev, STAGE_PROFILE_CLEAR_CACHE, MPI_Wtime() - t0);
+  }
+  else
+    bssn_gpu_clear_cached_device_buffers();
+
+  if (BH_num > 0 && lev == GH->levels - 1)
+  {
+    for (int ithBH = 0; ithBH < BH_num; ithBH++)
+    {
+      Porg0[ithBH][0] = Porg1[ithBH][0];
+      Porg0[ithBH][1] = Porg1[ithBH][1];
+      Porg0[ithBH][2] = Porg1[ithBH][2];
+    }
+  }
+
+  if (profile_enabled)
+    stage_profile_add(lev, STAGE_PROFILE_TOTAL, MPI_Wtime() - step_total_begin);
+}
+
+#endif

AMSS_NCKU_source/bssn_escalar_rhs_c.C

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

AMSS_NCKU_source/bssn_gpu.h

@@ -4,10 +4,8 @@
 #include "bssn_macro.h"
 #include "macrodef.fh"
 
-#define DEVICE_ID 0
-// #define DEVICE_ID_BY_MPI_RANK
-#define GRID_DIM 256
-#define BLOCK_DIM 128
+#define GRID_DIM 256
+#define BLOCK_DIM 128
 
 #define _FH2_(i, j, k) fh[(i) + (j) * _1D_SIZE[2] + (k) * _2D_SIZE[2]]
 #define _FH3_(i, j, k) fh[(i) + (j) * _1D_SIZE[3] + (k) * _2D_SIZE[3]]
@@ -65,9 +63,45 @@ int gpu_rhs(int calledby, int mpi_rank, int *ex, double &T,
             double *Gmx_Res, double *Gmy_Res, double *Gmz_Res,
             int &Symmetry, int &Lev, double &eps, int &co);
 
-int gpu_rhs_ss(RHS_SS_PARA);
-
-/** Init GPU side data in GPUMeta. */
-// void init_fluid_meta_gpu(GPUMeta *gpu_meta);
+int gpu_rhs_ss(RHS_SS_PARA);
+
+int bssn_gpu_bind_process_device(int mpi_rank);
+void bssn_gpu_clear_cached_device_buffers();
+void bssn_gpu_release_pinned_host_buffers();
+const double *bssn_gpu_find_device_buffer(const double *host_ptr);
+void bssn_gpu_register_device_buffer(const double *host_ptr, const double *device_ptr);
+void bssn_gpu_prepare_host_buffer(const double *host_ptr, int count);
+int bssn_gpu_stage_upload_buffer(const double *host_ptr, int count);
+int bssn_gpu_stage_zero_buffer(const double *host_ptr, int count);
+int bssn_gpu_stage_upload_region(const double *host_ptr,
+                                 const int *full_shape,
+                                 const double *full_llb,
+                                 const double *full_uub,
+                                 const int *region_shape,
+                                 const double *region_llb);
+int bssn_gpu_stage_download_region(double *host_ptr,
+                                   const int *full_shape,
+                                   const double *full_llb,
+                                   const double *full_uub,
+                                   const int *region_shape,
+                                   const double *region_llb);
+int bssn_gpu_stage_download_region_to_buffer(const double *host_src_ptr,
+                                             const int *full_shape,
+                                             const double *full_llb,
+                                             const double *full_uub,
+                                             const int *region_shape,
+                                             const double *region_llb,
+                                             double *host_dst_ptr);
+int bssn_gpu_stage_upload_buffer_to_region(const double *host_src_ptr,
+                                           double *host_dst_ptr,
+                                           const int *full_shape,
+                                           const double *full_llb,
+                                           const double *full_uub,
+                                           const int *region_shape,
+                                           const double *region_llb);
+int bssn_gpu_download_buffer_batch(const int *ex, double **host_ptrs, int num_buffers);
+
+/** Init GPU side data in GPUMeta. */
+// void init_fluid_meta_gpu(GPUMeta *gpu_meta);
 
 #endif

AMSS_NCKU_source/bssn_macro.h

@@ -65,9 +65,10 @@ if(TIME_COUNT_EACH_RANK == 1){\
      }\
 }
 
-//3---------------------GPU---------------------
-#define CALLED_BY_STEP 0
-#define CALLED_BY_CONSTRAINT 1
+//3---------------------GPU---------------------
+#define CALLED_BY_STEP 0
+#define CALLED_BY_CONSTRAINT 1
+#define CALLED_BY_CONSTRAINT_CONS_ONLY 2
 
 #define RHS_PARA_CALLED_FIRST_TIME cg->shape,TRK4,cg->X[0],cg->X[1],cg->X[2],cg->fgfs[phi0->sgfn],cg->fgfs[trK0->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[Ayy0->sgfn],cg->fgfs[Ayz0->sgfn],cg->fgfs[Azz0->sgfn],cg->fgfs[Gmx0->sgfn],cg->fgfs[Gmy0->sgfn],cg->fgfs[Gmz0->sgfn],cg->fgfs[Lap0->sgfn],cg->fgfs[Sfx0->sgfn],cg->fgfs[Sfy0->sgfn],cg->fgfs[Sfz0->sgfn],cg->fgfs[dtSfx0->sgfn],cg->fgfs[dtSfy0->sgfn],cg->fgfs[dtSfz0->sgfn],cg->fgfs[phi_rhs->sgfn],cg->fgfs[trK_rhs->sgfn],cg->fgfs[gxx_rhs->sgfn],cg->fgfs[gxy_rhs->sgfn],cg->fgfs[gxz_rhs->sgfn],cg->fgfs[gyy_rhs->sgfn],cg->fgfs[gyz_rhs->sgfn],cg->fgfs[gzz_rhs->sgfn],cg->fgfs[Axx_rhs->sgfn],cg->fgfs[Axy_rhs->sgfn],cg->fgfs[Axz_rhs->sgfn],cg->fgfs[Ayy_rhs->sgfn],cg->fgfs[Ayz_rhs->sgfn],cg->fgfs[Azz_rhs->sgfn],cg->fgfs[Gmx_rhs->sgfn],cg->fgfs[Gmy_rhs->sgfn],cg->fgfs[Gmz_rhs->sgfn],cg->fgfs[Lap_rhs->sgfn],cg->fgfs[Sfx_rhs->sgfn],cg->fgfs[Sfy_rhs->sgfn],cg->fgfs[Sfz_rhs->sgfn],cg->fgfs[dtSfx_rhs->sgfn],cg->fgfs[dtSfy_rhs->sgfn],cg->fgfs[dtSfz_rhs->sgfn],cg->fgfs[rho->sgfn],cg->fgfs[Sx->sgfn],cg->fgfs[Sy->sgfn],cg->fgfs[Sz->sgfn],cg->fgfs[Sxx->sgfn],cg->fgfs[Sxy->sgfn],cg->fgfs[Sxz->sgfn],cg->fgfs[Syy->sgfn],cg->fgfs[Syz->sgfn],cg->fgfs[Szz->sgfn],cg->fgfs[Gamxxx->sgfn],cg->fgfs[Gamxxy->sgfn],cg->fgfs[Gamxxz->sgfn],cg->fgfs[Gamxyy->sgfn],cg->fgfs[Gamxyz->sgfn],cg->fgfs[Gamxzz->sgfn],cg->fgfs[Gamyxx->sgfn],cg->fgfs[Gamyxy->sgfn],cg->fgfs[Gamyxz->sgfn],cg->fgfs[Gamyyy->sgfn],cg->fgfs[Gamyyz->sgfn],cg->fgfs[Gamyzz->sgfn],cg->fgfs[Gamzxx->sgfn],cg->fgfs[Gamzxy->sgfn],cg->fgfs[Gamzxz->sgfn],cg->fgfs[Gamzyy->sgfn],cg->fgfs[Gamzyz->sgfn],cg->fgfs[Gamzzz->sgfn],cg->fgfs[Rxx->sgfn],cg->fgfs[Rxy->sgfn],cg->fgfs[Rxz->sgfn],cg->fgfs[Ryy->sgfn],cg->fgfs[Ryz->sgfn],cg->fgfs[Rzz->sgfn],cg->fgfs[Cons_Ham->sgfn],cg->fgfs[Cons_Px->sgfn],cg->fgfs[Cons_Py->sgfn],cg->fgfs[Cons_Pz->sgfn],cg->fgfs[Cons_Gx->sgfn],cg->fgfs[Cons_Gy->sgfn],cg->fgfs[Cons_Gz->sgfn],Symmetry,lev,ndeps,pre
 

AMSS_NCKU_source/bssn_rhs.h

@@ -22,32 +22,19 @@
 #define f_compute_rhs_Z4c_ss COMPUTE_RHS_Z4C_SS
 #define f_compute_constraint_fr COMPUTE_CONSTRAINT_FR
 #endif
-#ifdef fortran3
-#define f_compute_rhs_bssn compute_rhs_bssn_
+#ifdef fortran3
+#define f_compute_rhs_bssn compute_rhs_bssn_
 #define f_compute_rhs_bssn_ss compute_rhs_bssn_ss_
 #define f_compute_rhs_bssn_escalar compute_rhs_bssn_escalar_
 #define f_compute_rhs_bssn_escalar_ss compute_rhs_bssn_escalar_ss_
 #define f_compute_rhs_Z4c compute_rhs_z4c_
 #define f_compute_rhs_Z4cnot compute_rhs_z4cnot_
 #define f_compute_rhs_Z4c_ss compute_rhs_z4c_ss_
-#define f_compute_constraint_fr compute_constraint_fr_
-#endif
-
-#ifdef __cplusplus
-extern "C"
-{
-#endif
-        void f_bssn_rhs_kernel_timing_reset();
-        int f_bssn_rhs_kernel_timing_bucket_count();
-        const double *f_bssn_rhs_kernel_timing_local_seconds();
-        const char *f_bssn_rhs_kernel_timing_label(int);
-#ifdef __cplusplus
-}
-#endif
-
-extern "C"
-{
-        int f_compute_rhs_bssn(int *, double &, double *, double *, double *,                                                      // ex,T,X,Y,Z
+#define f_compute_constraint_fr compute_constraint_fr_
+#endif
+extern "C"
+{
+        int f_compute_rhs_bssn(int *, double &, double *, double *, double *,                                                      // ex,T,X,Y,Z
                                double *, double *,                                                                                 // chi, trK
                                double *, double *, double *, double *, double *, double *,                                         // gij
                                double *, double *, double *, double *, double *, double *,                                         // Aij
@@ -63,34 +50,13 @@ extern "C"
                                double *, double *, double *, double *, double *, double *,                                         // Christoffel
                                double *, double *, double *, double *, double *, double *,                                         // Christoffel
                                double *, double *, double *, double *, double *, double *,                                         // Ricci
-                               double *, double *, double *, double *, double *, double *, double *,                               // constraint violation
-                               int &, int &, double &, int &);
-}
-
-int f_compute_rhs_bssn_escalar_c(int *, double &, double *, double *, double *,                                                      // ex,T,X,Y,Z
-                                 double *, double *,                                                                                 // chi, trK
-                                 double *, double *, double *, double *, double *, double *,                                         // gij
-                                 double *, double *, double *, double *, double *, double *,                                         // Aij
-                                 double *, double *, double *,                                                                       // Gam
-                                 double *, double *, double *, double *, double *, double *, double *,                               // Gauge
-                                 double *, double *,                                                                                 // Sphi, Spi
-                                 double *, double *,                                                                                 // chi, trK
-                                 double *, double *, double *, double *, double *, double *,                                         // gij
-                                 double *, double *, double *, double *, double *, double *,                                         // Aij
-                                 double *, double *, double *,                                                                       // Gam
-                                 double *, double *, double *, double *, double *, double *, double *,                               // Gauge
-                                 double *, double *,                                                                                 // Sphi, Spi
-                                 double *, double *, double *, double *, double *, double *, double *, double *, double *, double *, // stress-energy
-                                 double *, double *, double *, double *, double *, double *,                                         // Christoffel
-                                 double *, double *, double *, double *, double *, double *,                                         // Christoffel
-                                 double *, double *, double *, double *, double *, double *,                                         // Christoffel
-                                 double *, double *, double *, double *, double *, double *,                                         // Ricci
-                                 double *, double *, double *, double *, double *, double *, double *,                               // constraint violation
-                                 int &, int &, double &, int &);
-
-extern "C"
-{
-        int f_compute_rhs_bssn_ss(int *, double &, double *, double *, double *,                                                      // ex,T,rho,sigma,R
+                               double *, double *, double *, double *, double *, double *, double *,                               // constraint violation
+                               int &, int &, double &, int &);
+}
+
+extern "C"
+{
+        int f_compute_rhs_bssn_ss(int *, double &, double *, double *, double *,                                                      // ex,T,rho,sigma,R
                                   double *, double *, double *,                                                                       // X,Y,Z
                                   double *, double *, double *,                                                                       // drhodx,drhody,drhodz
                                   double *, double *, double *,                                                                       // dsigmadx,dsigmady,dsigmadz
@@ -117,10 +83,10 @@ extern "C"
                                   int &, int &, double &, int &, int &);
 }
 
-extern "C"
-{
-        int f_compute_rhs_bssn_escalar(int *, double &, double *, double *, double *,                                                      // ex,T,X,Y,Z
-                                       double *, double *,                                                                                 // chi, trK
+extern "C"
+{
+        int f_compute_rhs_bssn_escalar(int *, double &, double *, double *, double *,                                                      // ex,T,X,Y,Z
+                                       double *, double *,                                                                                 // chi, trK
                                        double *, double *, double *, double *, double *, double *,                                         // gij
                                        double *, double *, double *, double *, double *, double *,                                         // Aij
                                        double *, double *, double *,                                                                       // Gam
@@ -137,14 +103,14 @@ extern "C"
                                        double *, double *, double *, double *, double *, double *,                                         // Christoffel
                                        double *, double *, double *, double *, double *, double *,                                         // Christoffel
                                        double *, double *, double *, double *, double *, double *,                                         // Ricci
-                                       double *, double *, double *, double *, double *, double *, double *,                               // constraint violation
-                                       int &, int &, double &, int &);
-}
-
-extern "C"
-{
-        int f_compute_rhs_bssn_escalar_ss(int *, double &, double *, double *, double *,                                                      // ex,T,rho,sigma,R
-                                          double *, double *, double *,                                                                       // X,Y,Z
+                                       double *, double *, double *, double *, double *, double *, double *,                               // constraint violation
+                                       int &, int &, double &, int &);
+}
+
+extern "C"
+{
+        int f_compute_rhs_bssn_escalar_ss(int *, double &, double *, double *, double *,                                                      // ex,T,rho,sigma,R
+                                          double *, double *, double *,                                                                       // X,Y,Z
                                           double *, double *, double *,                                                                       // drhodx,drhody,drhodz
                                           double *, double *, double *,                                                                       // dsigmadx,dsigmady,dsigmadz
                                           double *, double *, double *,                                                                       // dRdx,dRdy,dRdz

AMSS_NCKU_source/bssn_rhs_c.C

@@ -2,88 +2,12 @@
 #include "bssn_rhs.h"
 #include "share_func.h"
 #include "tool.h"
-#include <time.h>
 // 0-based i,j,k
 // #define IDX_F(i,j,k,nx,ny) ((i) + (j)*(nx) + (k)*(nx)*(ny))
 // ex(1)=nx, ex(2)=ny, ex(3)=nz
 
 // 用法：a[ IDX_F(i,j,k,nx,ny) ]
 
-#ifndef BSSN_KERNEL_FINE_TIMING
-#define BSSN_KERNEL_FINE_TIMING 0
-#endif
-
-#if BSSN_KERNEL_FINE_TIMING
-namespace rhs_kernel_timing
-{
-    enum Bucket
-    {
-        KB_SETUP_DERIVS = 0,
-        KB_GEOM_GAMMA,
-        KB_RICCI_METRIC,
-        KB_CHI_LAPSE,
-        KB_AIJ_TRK_GAUGE,
-        KB_KO_CONSTRAINT,
-        KB_COUNT
-    };
-
-    static double local_bucket_seconds[KB_COUNT];
-
-    static const char *bucket_labels[KB_COUNT] =
-    {
-        "setup_derivs",
-        "geom_gamma",
-        "ricci_metric",
-        "chi_lapse",
-        "aij_trk_gauge",
-        "ko_constraint"
-    };
-
-    static inline double now_seconds()
-    {
-        struct timespec ts;
-        clock_gettime(CLOCK_MONOTONIC, &ts);
-        return double(ts.tv_sec) + 1.0e-9 * double(ts.tv_nsec);
-    }
-}
-
-extern "C" void f_bssn_rhs_kernel_timing_reset()
-{
-    for (int i = 0; i < rhs_kernel_timing::KB_COUNT; ++i)
-        rhs_kernel_timing::local_bucket_seconds[i] = 0.0;
-}
-
-extern "C" int f_bssn_rhs_kernel_timing_bucket_count()
-{
-    return rhs_kernel_timing::KB_COUNT;
-}
-
-extern "C" const double *f_bssn_rhs_kernel_timing_local_seconds()
-{
-    return rhs_kernel_timing::local_bucket_seconds;
-}
-
-extern "C" const char *f_bssn_rhs_kernel_timing_label(int bucket_index)
-{
-    if (bucket_index < 0 || bucket_index >= rhs_kernel_timing::KB_COUNT)
-        return "unknown";
-    return rhs_kernel_timing::bucket_labels[bucket_index];
-}
-
-#define RHS_KERNEL_TIMER_DECL(var_name) const double var_name = rhs_kernel_timing::now_seconds()
-#define RHS_KERNEL_TIMER_ADD(bucket_name, var_name) \
-    rhs_kernel_timing::local_bucket_seconds[int(rhs_kernel_timing::bucket_name)] += \
-        rhs_kernel_timing::now_seconds() - (var_name)
-#else
-extern "C" void f_bssn_rhs_kernel_timing_reset() {}
-extern "C" int f_bssn_rhs_kernel_timing_bucket_count() { return 0; }
-extern "C" const double *f_bssn_rhs_kernel_timing_local_seconds() { return 0; }
-extern "C" const char *f_bssn_rhs_kernel_timing_label(int) { return "disabled"; }
-
-#define RHS_KERNEL_TIMER_DECL(var_name)
-#define RHS_KERNEL_TIMER_ADD(bucket_name, var_name)
-#endif
-
 // C function that calculates the right-hand side for BSSN equations
 int f_compute_rhs_bssn(int *ex, double &T, 
                        double *X, double *Y, double *Z,
@@ -178,7 +102,6 @@ int f_compute_rhs_bssn(int *ex, double &T,
     dY = Y[1] - Y[0];
     dZ = Z[1] - Z[0];
 
-        RHS_KERNEL_TIMER_DECL(timer_setup_derivs);
         // 1ms //
         for(int i=0;i<all;i+=1){
             alpn1[i] = Lap[i] + 1.0;
@@ -218,8 +141,6 @@ int f_compute_rhs_bssn(int *ex, double &T,
                         (dxx[i] + ONE) * betaxz[i] + gxy[i] * betayz[i] + gyz[i] * betayx[i] 
                         + (dzz[i] + ONE) * betazx[i] - gxz[i] * betayy[i];                  
         }
-        RHS_KERNEL_TIMER_ADD(KB_SETUP_DERIVS, timer_setup_derivs);
-        RHS_KERNEL_TIMER_DECL(timer_geom_gamma);
         // Fused: inverse metric + Gamma constraint + Christoffel (3 loops -> 1)
         for(int i=0;i<all;i+=1){
             double det = (dxx[i] + ONE) * (dyy[i] + ONE) * (dzz[i] + ONE) + gxy[i] * gyz[i] * gxz[i] + gxz[i] * gxy[i] * gyz[i] -
@@ -362,6 +283,9 @@ int f_compute_rhs_bssn(int *ex, double &T,
                 + ( gupxy[i]*gupyz[i] + gupyy[i]*gupxz[i] ) * Axy[i]
                 + ( gupxy[i]*gupzz[i] + gupyz[i]*gupxz[i] ) * Axz[i]
                 + ( gupyy[i]*gupzz[i] + gupyz[i]*gupyz[i] ) * Ayz[i];
+            Rxx[i] = axx; Ryy[i] = ayy; Rzz[i] = azz;
+            Rxy[i] = axy; Rxz[i] = axz; Ryz[i] = ayz;
+
             Gamx_rhs[i] = - TWO * ( Lapx[i]*axx + Lapy[i]*axy + Lapz[i]*axz ) +
                 TWO * alpn1[i] * (
                 -F3o2/chin1[i] * ( chix[i]*axx + chiy[i]*axy + chiz[i]*axz ) -
@@ -391,8 +315,6 @@ int f_compute_rhs_bssn(int *ex, double &T,
                             + TWO * ( Gamzxy[i]*axy + Gamzxz[i]*axz + Gamzyz[i]*ayz )
                         );
         }
-        RHS_KERNEL_TIMER_ADD(KB_GEOM_GAMMA, timer_geom_gamma);
-        RHS_KERNEL_TIMER_DECL(timer_ricci_metric);
         // 22.3ms //
         fdderivs(ex,betax,gxxx,gxyx,gxzx,gyyx,gyzx,gzzx,
         X,Y,Z,ANTI,SYM, SYM ,Symmetry,Lev);
@@ -410,6 +332,7 @@ int f_compute_rhs_bssn(int *ex, double &T,
             double lfxx = gxxx[i] + gxyy[i] + gxzz[i];
             double lfxy = gxyx[i] + gyyy[i] + gyzz[i];
             double lfxz = gxzx[i] + gyzy[i] + gzzz[i];
+            fxx[i] = lfxx; fxy[i] = lfxy; fxz[i] = lfxz;
 
             double gxa = gupxx[i]*Gamxxx[i] + gupyy[i]*Gamxyy[i] + gupzz[i]*Gamxzz[i]
                     + TWO * ( gupxy[i]*Gamxxy[i] + gupxz[i]*Gamxxz[i] + gupyz[i]*Gamxyz[i] );
@@ -763,74 +686,69 @@ int f_compute_rhs_bssn(int *ex, double &T,
                     + Gamxyz[i] * gzzx[i] + Gamyyz[i] * gzzy[i] + Gamzyz[i] * gzzz[i]
                 );
         }
-        RHS_KERNEL_TIMER_ADD(KB_RICCI_METRIC, timer_ricci_metric);
 
-        RHS_KERNEL_TIMER_DECL(timer_chi_lapse);
         // 22.3ms //
             fdderivs(ex,chi,fxx,fxy,fxz,fyy,fyz,fzz,X,Y,Z,SYM,SYM,SYM,Symmetry,Lev);
 
         // 7ms //
         for (int i=0;i<all;i+=1) {
-            const double inv_chin1 = ONE / chin1[i];
-            const double half_inv_chin1 = HALF * inv_chin1;
-            const double scaled_inv = F3o2 * inv_chin1;
-            const double cxx = fxx[i] - Gamxxx[i] * chix[i] - Gamyxx[i] * chiy[i] - Gamzxx[i] * chiz[i];
-            const double cxy = fxy[i] - Gamxxy[i] * chix[i] - Gamyxy[i] * chiy[i] - Gamzxy[i] * chiz[i];
-            const double cxz = fxz[i] - Gamxxz[i] * chix[i] - Gamyxz[i] * chiy[i] - Gamzxz[i] * chiz[i];
-            const double cyy = fyy[i] - Gamxyy[i] * chix[i] - Gamyyy[i] * chiy[i] - Gamzyy[i] * chiz[i];
-            const double cyz = fyz[i] - Gamxyz[i] * chix[i] - Gamyyz[i] * chiy[i] - Gamzyz[i] * chiz[i];
-            const double czz = fzz[i] - Gamxzz[i] * chix[i] - Gamyzz[i] * chiy[i] - Gamzzz[i] * chiz[i];
-            const double ricci_chi =
-                gupxx[i] * (cxx - scaled_inv * chix[i] * chix[i])
-                + gupyy[i] * (cyy - scaled_inv * chiy[i] * chiy[i])
-                + gupzz[i] * (czz - scaled_inv * chiz[i] * chiz[i])
-                + TWO * gupxy[i] * (cxy - scaled_inv * chix[i] * chiy[i])
-                + TWO * gupxz[i] * (cxz - scaled_inv * chix[i] * chiz[i])
-                + TWO * gupyz[i] * (cyz - scaled_inv * chiy[i] * chiz[i]);
-            f[i] = ricci_chi;
-            Rxx[i] = Rxx[i] + ( cxx - half_inv_chin1 * chix[i] * chix[i] + (dxx[i] + ONE) * ricci_chi ) * half_inv_chin1;
-            Ryy[i] = Ryy[i] + ( cyy - half_inv_chin1 * chiy[i] * chiy[i] + (dyy[i] + ONE) * ricci_chi ) * half_inv_chin1;
-            Rzz[i] = Rzz[i] + ( czz - half_inv_chin1 * chiz[i] * chiz[i] + (dzz[i] + ONE) * ricci_chi ) * half_inv_chin1;
+            fxx[i] = fxx[i] - Gamxxx[i] * chix[i] - Gamyxx[i] * chiy[i] - Gamzxx[i] * chiz[i];
+            fxy[i] = fxy[i] - Gamxxy[i] * chix[i] - Gamyxy[i] * chiy[i] - Gamzxy[i] * chiz[i];
+            fxz[i] = fxz[i] - Gamxxz[i] * chix[i] - Gamyxz[i] * chiy[i] - Gamzxz[i] * chiz[i];
+            fyy[i] = fyy[i] - Gamxyy[i] * chix[i] - Gamyyy[i] * chiy[i] - Gamzyy[i] * chiz[i];
+            fyz[i] = fyz[i] - Gamxyz[i] * chix[i] - Gamyyz[i] * chiy[i] - Gamzyz[i] * chiz[i];
+            fzz[i] = fzz[i] - Gamxzz[i] * chix[i] - Gamyzz[i] * chiy[i] - Gamzzz[i] * chiz[i];
+            f[i] =
+                gupxx[i] * (fxx[i] - (F3o2 / chin1[i]) * chix[i] * chix[i])
+                + gupyy[i] * (fyy[i] - (F3o2 / chin1[i]) * chiy[i] * chiy[i])
+                + gupzz[i] * (fzz[i] - (F3o2 / chin1[i]) * chiz[i] * chiz[i])
+                + TWO * gupxy[i] * (fxy[i] - (F3o2 / chin1[i]) * chix[i] * chiy[i])
+                + TWO * gupxz[i] * (fxz[i] - (F3o2 / chin1[i]) * chix[i] * chiz[i])
+                + TWO * gupyz[i] * (fyz[i] - (F3o2 / chin1[i]) * chiy[i] * chiz[i]);
+            Rxx[i] = Rxx[i] + ( fxx[i] - (chix[i] * chix[i]) / (chin1[i] * TWO) + (dxx[i] + ONE) * f[i] ) / (chin1[i] * TWO);
+            Ryy[i] = Ryy[i] + ( fyy[i] - (chiy[i] * chiy[i]) / (chin1[i] * TWO) + (dyy[i] + ONE) * f[i] ) / (chin1[i] * TWO);
+            Rzz[i] = Rzz[i] + ( fzz[i] - (chiz[i] * chiz[i]) / (chin1[i] * TWO) + (dzz[i] + ONE) * f[i] ) / (chin1[i] * TWO);
 
-            Rxy[i] = Rxy[i] + ( cxy - half_inv_chin1 * chix[i] * chiy[i] + gxy[i] * ricci_chi ) * half_inv_chin1;
-            Rxz[i] = Rxz[i] + ( cxz - half_inv_chin1 * chix[i] * chiz[i] + gxz[i] * ricci_chi ) * half_inv_chin1;
-            Ryz[i] = Ryz[i] + ( cyz - half_inv_chin1 * chiy[i] * chiz[i] + gyz[i] * ricci_chi ) * half_inv_chin1;
+            Rxy[i] = Rxy[i] + ( fxy[i] - (chix[i] * chiy[i]) / (chin1[i] * TWO) + gxy[i] * f[i] ) / (chin1[i] * TWO);
+            Rxz[i] = Rxz[i] + ( fxz[i] - (chix[i] * chiz[i]) / (chin1[i] * TWO) + gxz[i] * f[i] ) / (chin1[i] * TWO);
+            Ryz[i] = Ryz[i] + ( fyz[i] - (chiy[i] * chiz[i]) / (chin1[i] * TWO) + gyz[i] * f[i] ) / (chin1[i] * TWO);
         }
 
         // 24ms //
         fdderivs(ex,Lap,fxx,fxy,fxz,fyy,fyz,fzz,X,Y,Z,SYM,SYM,SYM,Symmetry,Lev);
+        fderivs(ex,chi,dtSfx_rhs,dtSfy_rhs,dtSfz_rhs,X,Y,Z,SYM,SYM,SYM,Symmetry,Lev);
 
         // 6ms //
         for (int i=0;i<all;i+=1) {
-            const double inv_chin1 = ONE / chin1[i];
-            const double gchi_x = (gupxx[i] * chix[i] + gupxy[i] * chiy[i] + gupxz[i] * chiz[i]) * inv_chin1;
-            const double gchi_y = (gupxy[i] * chix[i] + gupyy[i] * chiy[i] + gupyz[i] * chiz[i]) * inv_chin1;
-            const double gchi_z = (gupxz[i] * chix[i] + gupyz[i] * chiy[i] + gupzz[i] * chiz[i]) * inv_chin1;
+            /* gxxx,gxxy,gxxz (这里是“升指标后的chi导数/chi”那类量，你沿用原变量名即可) */
+            gxxx[i] = (gupxx[i] * chix[i] + gupxy[i] * chiy[i] + gupxz[i] * chiz[i]) / chin1[i];
+            gxxy[i] = (gupxy[i] * chix[i] + gupyy[i] * chiy[i] + gupyz[i] * chiz[i]) / chin1[i];
+            gxxz[i] = (gupxz[i] * chix[i] + gupyz[i] * chiy[i] + gupzz[i] * chiz[i]) / chin1[i];
 
             /* Christoffel 修正项 */
-            Gamxxx[i] = Gamxxx[i] - ( ((chix[i] + chix[i]) * inv_chin1) - (dxx[i] + ONE) * gchi_x ) * HALF;
-            Gamyxx[i] = Gamyxx[i] - ( 0.0           - (dxx[i] + ONE) * gchi_y ) * HALF;  /* 原式只有 -gxx*gxxy */
-            Gamzxx[i] = Gamzxx[i] - ( 0.0           - (dxx[i] + ONE) * gchi_z ) * HALF;
+            Gamxxx[i] = Gamxxx[i] - ( ((chix[i] + chix[i]) / chin1[i]) - (dxx[i] + ONE) * gxxx[i] ) * HALF;
+            Gamyxx[i] = Gamyxx[i] - ( 0.0           - (dxx[i] + ONE) * gxxy[i] ) * HALF;  /* 原式只有 -gxx*gxxy */
+            Gamzxx[i] = Gamzxx[i] - ( 0.0           - (dxx[i] + ONE) * gxxz[i] ) * HALF;
 
-            Gamxyy[i] = Gamxyy[i] - ( 0.0           - (dyy[i] + ONE) * gchi_x ) * HALF;
-            Gamyyy[i] = Gamyyy[i] - ( ((chiy[i] + chiy[i]) * inv_chin1) - (dyy[i] + ONE) * gchi_y ) * HALF;
-            Gamzyy[i] = Gamzyy[i] - ( 0.0           - (dyy[i] + ONE) * gchi_z ) * HALF;
+            Gamxyy[i] = Gamxyy[i] - ( 0.0           - (dyy[i] + ONE) * gxxx[i] ) * HALF;
+            Gamyyy[i] = Gamyyy[i] - ( ((chiy[i] + chiy[i]) / chin1[i]) - (dyy[i] + ONE) * gxxy[i] ) * HALF;
+            Gamzyy[i] = Gamzyy[i] - ( 0.0           - (dyy[i] + ONE) * gxxz[i] ) * HALF;
 
-            Gamxzz[i] = Gamxzz[i] - ( 0.0           - (dzz[i] + ONE) * gchi_x ) * HALF;
-            Gamyzz[i] = Gamyzz[i] - ( 0.0           - (dzz[i] + ONE) * gchi_y ) * HALF;
-            Gamzzz[i] = Gamzzz[i] - ( ((chiz[i] + chiz[i]) * inv_chin1) - (dzz[i] + ONE) * gchi_z ) * HALF;
+            Gamxzz[i] = Gamxzz[i] - ( 0.0           - (dzz[i] + ONE) * gxxx[i] ) * HALF;
+            Gamyzz[i] = Gamyzz[i] - ( 0.0           - (dzz[i] + ONE) * gxxy[i] ) * HALF;
+            Gamzzz[i] = Gamzzz[i] - ( ((chiz[i] + chiz[i]) / chin1[i]) - (dzz[i] + ONE) * gxxz[i] ) * HALF;
 
-            Gamxxy[i] = Gamxxy[i] - ( ( chiy[i] * inv_chin1)      - gxy[i] * gchi_x ) * HALF;
-            Gamyxy[i] = Gamyxy[i] - ( ( chix[i] * inv_chin1)      - gxy[i] * gchi_y ) * HALF;
-            Gamzxy[i] = Gamzxy[i] - ( 0.0           - gxy[i] * gchi_z ) * HALF;
+            Gamxxy[i] = Gamxxy[i] - ( ( chiy[i] / chin1[i])      - gxy[i] * gxxx[i] ) * HALF;
+            Gamyxy[i] = Gamyxy[i] - ( ( chix[i] / chin1[i])      - gxy[i] * gxxy[i] ) * HALF;
+            Gamzxy[i] = Gamzxy[i] - ( 0.0           - gxy[i] * gxxz[i] ) * HALF;
 
-            Gamxxz[i] = Gamxxz[i] - ( ( chiz[i] * inv_chin1)      - gxz[i] * gchi_x ) * HALF;
-            Gamyxz[i] = Gamyxz[i] - ( 0.0           - gxz[i] * gchi_y ) * HALF;
-            Gamzxz[i] = Gamzxz[i] - ( ( chix[i] * inv_chin1)      - gxz[i] * gchi_z ) * HALF;
+            Gamxxz[i] = Gamxxz[i] - ( ( chiz[i] / chin1[i])      - gxz[i] * gxxx[i] ) * HALF;
+            Gamyxz[i] = Gamyxz[i] - ( 0.0           - gxz[i] * gxxy[i] ) * HALF;
+            Gamzxz[i] = Gamzxz[i] - ( ( chix[i] / chin1[i])      - gxz[i] * gxxz[i] ) * HALF;
 
-            Gamxyz[i] = Gamxyz[i] - ( 0.0           - gyz[i] * gchi_x ) * HALF;
-            Gamyyz[i] = Gamyyz[i] - ( ( chiz[i] * inv_chin1)      - gyz[i] * gchi_y ) * HALF;
-            Gamzyz[i] = Gamzyz[i] - ( ( chiy[i] * inv_chin1)      - gyz[i] * gchi_z ) * HALF;
+            Gamxyz[i] = Gamxyz[i] - ( 0.0           - gyz[i] * gxxx[i] ) * HALF;
+            Gamyyz[i] = Gamyyz[i] - ( ( chiz[i] / chin1[i])      - gyz[i] * gxxy[i] ) * HALF;
+            Gamzyz[i] = Gamzyz[i] - ( ( chiy[i] / chin1[i])      - gyz[i] * gxxz[i] ) * HALF;
 
             /* fxx..fyz 修正：减去 Γ * ∂Lap */
             fxx[i] = fxx[i] - Gamxxx[i] * Lapx[i] - Gamyxx[i] * Lapy[i] - Gamzxx[i] * Lapz[i];
@@ -844,8 +762,6 @@ int f_compute_rhs_bssn(int *ex, double &T,
             trK_rhs[i] =  gupxx[i] * fxx[i] + gupyy[i] * fyy[i] + gupzz[i] * fzz[i]
                     +  TWO * ( gupxy[i] * fxy[i] + gupxz[i] * fxz[i] + gupyz[i] * fyz[i] );
         }
-        RHS_KERNEL_TIMER_ADD(KB_CHI_LAPSE, timer_chi_lapse);
-        RHS_KERNEL_TIMER_DECL(timer_aij_trk_gauge);
         // 2.5ms //
         for (int i=0;i<all;i+=1) {
             const double divb = betaxx[i] + betayy[i] + betazz[i];
@@ -1106,9 +1022,16 @@ int f_compute_rhs_bssn(int *ex, double &T,
                         + gupyz[i] * dtSfy_rhs[i] * dtSfz_rhs[i] );
 
             #if (GAUGE == 2)
-            reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / pow( (ONE - sqrt(chin1[i])), 2.0 );
+            {
+                const double chi_sqrt = sqrt(chin1[i]);
+                const double damping = ONE - chi_sqrt;
+                reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / (damping * damping);
+            }
             #else
-            reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / pow( (ONE - chin1[i]), 2.0 );
+            {
+                const double damping = ONE - chin1[i];
+                reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / (damping * damping);
+            }
             #endif
 
             dtSfx_rhs[i] = Gamx_rhs[i] - reta[i] * dtSfx[i];
@@ -1124,9 +1047,16 @@ int f_compute_rhs_bssn(int *ex, double &T,
                         + gupyz[i] * dtSfy_rhs[i] * dtSfz_rhs[i] );
 
             #if (GAUGE == 4)
-            reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / pow( (ONE - sqrt(chin1[i])), 2.0 );
+            {
+                const double chi_sqrt = sqrt(chin1[i]);
+                const double damping = ONE - chi_sqrt;
+                reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / (damping * damping);
+            }
             #else
-            reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / pow( (ONE - chin1[i]), 2.0 );
+            {
+                const double damping = ONE - chin1[i];
+                reta[i] = 1.31 / 2.0 * sqrt( reta[i] / chin1[i] ) / (damping * damping);
+            }
             #endif
 
             betax_rhs[i] = FF * Gamx[i] - reta[i] * betax[i];
@@ -1146,8 +1076,6 @@ int f_compute_rhs_bssn(int *ex, double &T,
             dtSfz_rhs[i] = Gamz_rhs[i] - reta[i] * dtSfz[i];
             #endif
         }
-        RHS_KERNEL_TIMER_ADD(KB_AIJ_TRK_GAUGE, timer_aij_trk_gauge);
-        RHS_KERNEL_TIMER_DECL(timer_ko_constraint);
         // advection + KO dissipation with shared symmetry buffer
         lopsided_kodis(ex,X,Y,Z,dxx,gxx_rhs,betax,betay,betaz,Symmetry,SSS,eps);
         lopsided_kodis(ex,X,Y,Z,Gamz,Gamz_rhs,betax,betay,betaz,Symmetry,SSA,eps);
@@ -1279,7 +1207,6 @@ int f_compute_rhs_bssn(int *ex, double &T,
                 movz_Res[i] = movz_Res[i] - F2o3*Kz[i] - F8*PI*Sz[i];
             }
         }
-        RHS_KERNEL_TIMER_ADD(KB_KO_CONSTRAINT, timer_ko_constraint);
 
 
             

AMSS_NCKU_source/cgh.C

@@ -23,10 +23,14 @@ using namespace std;
 #include <mpi.h>
 
 #include "macrodef.h"
-#include "misc.h"
-#include "cgh.h"
-#include "Parallel.h"
-#include "parameters.h"
+#include "misc.h"
+#include "cgh.h"
+#include "Parallel.h"
+#include "parameters.h"
+#ifdef USE_GPU
+#include "bssn_gpu.h"
+#include "bssn_cuda_ops.h"
+#endif
 
 //================================================================================================
 
@@ -881,13 +885,20 @@ void cgh::recompose_cgh(int nprocs, bool *lev_flag,
       tmPat = construct_patchlist(lev, Symmetry);
       // tmPat construction completes
       Parallel::distribute(tmPat, nprocs, ingfs, fngfs, false);
-      //    checkPatchList(tmPat,true);
-      bool CC = (lev > trfls);
-      Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
-
-      Parallel::KillBlocks(PatL[lev]);
-      PatL[lev]->destroyList();
-      PatL[lev] = tmPat;
+      //    checkPatchList(tmPat,true);
+      bool CC = (lev > trfls);
+      Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
+
+#ifdef USE_GPU
+      bssn_gpu_clear_cached_device_buffers();
+      bssn_gpu_release_pinned_host_buffers();
+      bssn_cuda_release_rk4_caches();
+      bssn_cuda_release_interp_caches();
+      patch_release_interp_plan_cache();
+#endif
+      Parallel::KillBlocks(PatL[lev]);
+      PatL[lev]->destroyList();
+      PatL[lev] = tmPat;
 #if (RPB == 1)
       Parallel::destroypsuList_bam(bdsul[lev]);
       Parallel::destroypsuList_bam(rsul[lev]);
@@ -910,13 +921,20 @@ void cgh::recompose_cgh(int nprocs, bool *lev_flag,
       tmPat = construct_patchlist(lev, Symmetry);
       // tmPat construction completes
       Parallel::distribute(tmPat, end_rank[lev] - start_rank[lev] + 1, ingfs, fngfs, false, start_rank[lev], end_rank[lev]);
-      //    checkPatchList(tmPat,true);
-      bool CC = (lev > trfls);
-      Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
-
-      Parallel::KillBlocks(PatL[lev]);
-      PatL[lev]->destroyList();
-      PatL[lev] = tmPat;
+      //    checkPatchList(tmPat,true);
+      bool CC = (lev > trfls);
+      Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
+
+#ifdef USE_GPU
+      bssn_gpu_clear_cached_device_buffers();
+      bssn_gpu_release_pinned_host_buffers();
+      bssn_cuda_release_rk4_caches();
+      bssn_cuda_release_interp_caches();
+      patch_release_interp_plan_cache();
+#endif
+      Parallel::KillBlocks(PatL[lev]);
+      PatL[lev]->destroyList();
+      PatL[lev] = tmPat;
 #if (RPB == 1)
 #error "not support yet"
 #endif
@@ -1518,13 +1536,20 @@ void cgh::recompose_cgh_Onelevel(int nprocs, int lev,
   tmPat = construct_patchlist(lev, Symmetry);
   // tmPat construction completes
   Parallel::distribute(tmPat, nprocs, ingfs, fngfs, false);
-  //    checkPatchList(tmPat,true);
-  bool CC = (lev > trfls);
-  Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
-
-  Parallel::KillBlocks(PatL[lev]);
-  PatL[lev]->destroyList();
-  PatL[lev] = tmPat;
+  //    checkPatchList(tmPat,true);
+  bool CC = (lev > trfls);
+  Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
+
+#ifdef USE_GPU
+  bssn_gpu_clear_cached_device_buffers();
+  bssn_gpu_release_pinned_host_buffers();
+  bssn_cuda_release_rk4_caches();
+  bssn_cuda_release_interp_caches();
+  patch_release_interp_plan_cache();
+#endif
+  Parallel::KillBlocks(PatL[lev]);
+  PatL[lev]->destroyList();
+  PatL[lev] = tmPat;
 }
 #elif (PSTR == 1 || PSTR == 2 || PSTR == 3)
 #warning "recompose_cgh_Onelevel is not implimented yet"
@@ -1540,14 +1565,21 @@ void cgh::recompose_cgh_Onelevel(int nprocs, int lev,
   // tmPat construction completes
   Parallel::distribute(tmPat, end_rank[lev] - start_rank[lev] + 1, ingfs, fngfs, false, start_rank[lev], end_rank[lev]);
   misc::tillherecheck(Commlev[lev], start_rank[lev], "after distribute");
-  //    checkPatchList(tmPat,true);
-  bool CC = (lev > trfls);
-  Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
-  misc::tillherecheck(Commlev[lev], start_rank[lev], "after fill_level_data");
-
-  Parallel::KillBlocks(PatL[lev]);
-  PatL[lev]->destroyList();
-  PatL[lev] = tmPat;
+  //    checkPatchList(tmPat,true);
+  bool CC = (lev > trfls);
+  Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
+  misc::tillherecheck(Commlev[lev], start_rank[lev], "after fill_level_data");
+
+#ifdef USE_GPU
+  bssn_gpu_clear_cached_device_buffers();
+  bssn_gpu_release_pinned_host_buffers();
+  bssn_cuda_release_rk4_caches();
+  bssn_cuda_release_interp_caches();
+  patch_release_interp_plan_cache();
+#endif
+  Parallel::KillBlocks(PatL[lev]);
+  PatL[lev]->destroyList();
+  PatL[lev] = tmPat;
 }
 
 

AMSS_NCKU_source/fmisc.f90

@@ -1511,88 +1511,13 @@ deallocate(f_flat)
 
 f_out = f_out*dX*dY*dZ
 
-  return
-
-  end subroutine l2normhelper
-!--------------------------------------------------------------------------------------
-  subroutine l2normhelper7(ex, X, Y, Z,xmin,ymin,zmin,xmax,ymax,zmax,&
-                           f1,f2,f3,f4,f5,f6,f7,f_out,gw)
-
-  implicit none
-!~~~~~~> Input parameters:
-  integer,intent(in ):: ex(1:3)
-  real*8, intent(in ):: X(1:ex(1)),Y(1:ex(2)),Z(1:ex(3)),xmin,ymin,zmin,xmax,ymax,zmax
-  integer,intent(in)::gw
-  real*8, dimension(ex(1),ex(2),ex(3)),intent(in) :: f1,f2,f3,f4,f5,f6,f7
-  real*8, intent(out) :: f_out(7)
-!~~~~~~> Other variables:
-
-  real*8            :: dX, dY, dZ
-  integer::imin,jmin,kmin
-  integer::imax,jmax,kmax
-  integer::i,j,k
-  real*8 :: s1,s2,s3,s4,s5,s6,s7
-
-  dX = X(2) - X(1)
-  dY = Y(2) - Y(1)
-  dZ = Z(2) - Z(1)
-
-! for ghost zone
-   imin = gw+1
-   jmin = gw+1
-   kmin = gw+1
-
-   imax = ex(1) - gw
-   jmax = ex(2) - gw
-   kmax = ex(3) - gw
-
-!for patch boundary (i.e., not ghost boundary)
-
-if(dabs(X(ex(1))-xmax) < dX) imax = ex(1)
-if(dabs(Y(ex(2))-ymax) < dY) jmax = ex(2)
-if(dabs(Z(ex(3))-zmax) < dZ) kmax = ex(3)
-if(dabs(X(1)-xmin) < dX) imin = 1
-if(dabs(Y(1)-ymin) < dY) jmin = 1
-if(dabs(Z(1)-zmin) < dZ) kmin = 1
-
-  s1 = 0.d0
-  s2 = 0.d0
-  s3 = 0.d0
-  s4 = 0.d0
-  s5 = 0.d0
-  s6 = 0.d0
-  s7 = 0.d0
-
-  do k=kmin,kmax
-    do j=jmin,jmax
-!DIR$ SIMD REDUCTION(+:s1,s2,s3,s4,s5,s6,s7)
-      do i=imin,imax
-        s1 = s1 + f1(i,j,k)*f1(i,j,k)
-        s2 = s2 + f2(i,j,k)*f2(i,j,k)
-        s3 = s3 + f3(i,j,k)*f3(i,j,k)
-        s4 = s4 + f4(i,j,k)*f4(i,j,k)
-        s5 = s5 + f5(i,j,k)*f5(i,j,k)
-        s6 = s6 + f6(i,j,k)*f6(i,j,k)
-        s7 = s7 + f7(i,j,k)*f7(i,j,k)
-      enddo
-    enddo
-  enddo
-
-  f_out(1) = s1*dX*dY*dZ
-  f_out(2) = s2*dX*dY*dZ
-  f_out(3) = s3*dX*dY*dZ
-  f_out(4) = s4*dX*dY*dZ
-  f_out(5) = s5*dX*dY*dZ
-  f_out(6) = s6*dX*dY*dZ
-  f_out(7) = s7*dX*dY*dZ
-
-  return
-
-  end subroutine l2normhelper7
-!--------------------------------------------------------------------------------------
-! calculate L2norm especially for shell Blocks
-  subroutine l2normhelper_sh(ex, X, Y, Z,xmin,ymin,zmin,xmax,ymax,zmax,&
-                          f,f_out,gw,ogw,Symmetry)
+  return
+
+  end subroutine l2normhelper
+!--------------------------------------------------------------------------------------
+! calculate L2norm especially for shell Blocks
+  subroutine l2normhelper_sh(ex, X, Y, Z,xmin,ymin,zmin,xmax,ymax,zmax,&
+                          f,f_out,gw,ogw,Symmetry)
 
   implicit none
 !~~~~~~> Input parameters:

AMSS_NCKU_source/fmisc.h

@@ -12,10 +12,9 @@
 #define f_global_interpind global_interpind
 #define f_global_interpind2d global_interpind2d
 #define f_global_interpind1d global_interpind1d
-#define f_l2normhelper l2normhelper
-#define f_l2normhelper7 l2normhelper7
-#define f_l2normhelper_sh l2normhelper_sh
-#define f_l2normhelper_sh_rms l2normhelper_sh_rms
+#define f_l2normhelper l2normhelper
+#define f_l2normhelper_sh l2normhelper_sh
+#define f_l2normhelper_sh_rms l2normhelper_sh_rms
 #define f_average average
 #define f_average3 average3
 #define f_average2 average2
@@ -42,10 +41,9 @@
 #define f_global_interpind GLOBAL_INTERPIND
 #define f_global_interpind2d GLOBAL_INTERPIND2D
 #define f_global_interpind1d GLOBAL_INTERPIND1D
-#define f_l2normhelper L2NORMHELPER
-#define f_l2normhelper7 L2NORMHELPER7
-#define f_l2normhelper_sh L2NORMHELPER_SH
-#define f_l2normhelper_sh_rms L2NORMHELPER_SH_RMS
+#define f_l2normhelper L2NORMHELPER
+#define f_l2normhelper_sh L2NORMHELPER_SH
+#define f_l2normhelper_sh_rms L2NORMHELPER_SH_RMS
 #define f_average AVERAGE
 #define f_average3 AVERAGE3
 #define f_average2 AVERAGE2
@@ -72,10 +70,9 @@
 #define f_global_interpind global_interpind_
 #define f_global_interpind2d global_interpind2d_
 #define f_global_interpind1d global_interpind1d_
-#define f_l2normhelper l2normhelper_
-#define f_l2normhelper7 l2normhelper7_
-#define f_l2normhelper_sh l2normhelper_sh_
-#define f_l2normhelper_sh_rms l2normhelper_sh_rms_
+#define f_l2normhelper l2normhelper_
+#define f_l2normhelper_sh l2normhelper_sh_
+#define f_l2normhelper_sh_rms l2normhelper_sh_rms_
 #define f_average average_
 #define f_average3 average3_
 #define f_average2 average2_
@@ -159,29 +156,20 @@ extern "C"
 							  int *, double *, int &, int &);
 }
 
-extern "C"
-{
-	void f_l2normhelper(int *, double *, double *, double *,
-						double &, double &, double &,
-						double &, double &, double &,
-						double *, double &, int &);
-}
-
-extern "C"
-{
-	void f_l2normhelper7(int *, double *, double *, double *,
-						 double &, double &, double &,
-						 double &, double &, double &,
-						 double *, double *, double *, double *,
-						 double *, double *, double *, double *, int &);
-}
-
-extern "C"
-{
-	void f_l2normhelper_sh(int *, double *, double *, double *,
-						   double &, double &, double &,
-						   double &, double &, double &,
-						   double *, double &, int &, int &, int &);
+extern "C"
+{
+	void f_l2normhelper(int *, double *, double *, double *,
+						double &, double &, double &,
+						double &, double &, double &,
+						double *, double &, int &);
+}
+
+extern "C"
+{
+	void f_l2normhelper_sh(int *, double *, double *, double *,
+						   double &, double &, double &,
+						   double &, double &, double &,
+						   double *, double &, int &, int &, int &);
 }
 
 extern "C"

AMSS_NCKU_source/gaussj.C

@@ -18,7 +18,7 @@ using namespace std;
 #endif
 
 // Intel oneMKL LAPACK interface
-#include <lapacke.h>
+#include <mkl_lapacke.h>
 /* Linear equation solution using Intel oneMKL LAPACK.
 a[0..n-1][0..n-1] is the input matrix. b[0..n-1] is input
 containing the right-hand side vectors. On output a is

AMSS_NCKU_source/macrodef.h

@@ -29,16 +29,6 @@
 
 #define REGLEV 0
 
-#define BSSN_FINE_TIMING 0
-
-#define BSSN_FINE_TIMING_EVERY 1
-
-#define BSSN_FINE_TIMING_TOPN 8
-
-#define BSSN_KERNEL_FINE_TIMING 0
-
-#define BSSN_ENABLE_STDIN_ABORT_POLL 0
-
 //#define USE_GPU
 
 //#define CHECKDETAIL
@@ -98,21 +88,6 @@
 //     0: for every level;
 //     1: for all
 //
-// define BSSN_FINE_TIMING
-//     enable fine-grained per-timestep timing monitor
-//
-// define BSSN_FINE_TIMING_EVERY
-//     report timing every N coarse timesteps
-//
-// define BSSN_FINE_TIMING_TOPN
-//     number of hottest timing buckets shown in stdout
-//
-// define BSSN_KERNEL_FINE_TIMING
-//     enable split timing inside compute_rhs_bssn
-//
-// define BSSN_ENABLE_STDIN_ABORT_POLL
-//     poll stdin and broadcast abort flag every coarse step
-//
 // define USE_GPU
 //     use gpu or not
 //
@@ -167,3 +142,4 @@
 #define TINY 1e-10
 
 #endif   /* MICRODEF_H */
+

AMSS_NCKU_source/makefile

@@ -1,79 +1,63 @@
-
-
+
+
 include makefile.inc
 
--include AMSS_NCKU_build.mk
-
-ABE_TYPE ?= $(shell awk '/^[[:space:]]*\#define[[:space:]]+ABEtype/ {print $$3; exit}' macrodef.h 2>/dev/null)
-
-ifeq ($(USE_TRANSFER_CACHE),auto)
-ifeq ($(ABE_TYPE),0)
-EFFECTIVE_USE_TRANSFER_CACHE = 1
-else
-EFFECTIVE_USE_TRANSFER_CACHE = 0
-endif
-else
-EFFECTIVE_USE_TRANSFER_CACHE = $(USE_TRANSFER_CACHE)
-endif
-
-ifeq ($(USE_CXX_ESCALAR_KERNEL),1)
-ifeq ($(ABE_TYPE),1)
-EFFECTIVE_USE_CXX_ESCALAR_KERNEL = 1
-else
-EFFECTIVE_USE_CXX_ESCALAR_KERNEL = 0
-endif
-else
-EFFECTIVE_USE_CXX_ESCALAR_KERNEL = 0
-endif
-
-ifeq ($(EFFECTIVE_USE_CXX_ESCALAR_KERNEL),1)
-ifeq ($(USE_CXX_KERNELS),0)
-$(error USE_CXX_ESCALAR_KERNEL=1 requires USE_CXX_KERNELS=1 because bssn_escalar_rhs_c.C reuses the C BSSN kernel)
-endif
-endif
-
 ## polint(ordn=6) kernel selector:
 ##   1 (default): barycentric fast path
 ##   0          : fallback to Neville path
 POLINT6_USE_BARY ?= 1
 POLINT6_FLAG = -DPOLINT6_USE_BARYCENTRIC=$(POLINT6_USE_BARY)
-TRANSFER_CACHE_FLAG = -DBSSN_USE_TRANSFER_CACHE=$(EFFECTIVE_USE_TRANSFER_CACHE)
-ESCALAR_KERNEL_FLAG = -DBSSN_USE_ESCALAR_C_KERNEL=$(EFFECTIVE_USE_CXX_ESCALAR_KERNEL)
 
-## AMD AOCC build flags optimized for EPYC Zen 4 (-march=znver4)
-CXXAPPFLAGS = -O3 -march=znver4 -ffast-math -flto \
-              -Dfortran3 -Dnewc -I$(AOCL_ROOT)/include $(INTERP_LB_FLAGS) \
-              $(TRANSFER_CACHE_FLAG) $(ESCALAR_KERNEL_FLAG)
-f90appflags = -O3 -march=znver4 -ffast-math -flto \
-              -cpp -I$(AOCL_ROOT)/include $(POLINT6_FLAG)
+## ABE build flags selected by PGO_MODE (set in makefile.inc, default: opt)
+##   make                        -> opt  (PGO-guided, maximum performance)
+##   make PGO_MODE=instrument    -> instrument (Phase 1: collect fresh profile data)
+PROFDATA = /home/$(shell whoami)/AMSS-NCKU/pgo_profile/default.profdata
+
+ifeq ($(PGO_MODE),instrument)
+## Phase 1: instrumentation — omit -ipo/-fp-model fast=2 for faster build and numerical stability
+CXXAPPFLAGS = -O3 -xHost -fma -fprofile-instr-generate -ipo \
+              -Dfortran3 -Dnewc -I${MKLROOT}/include $(INTERP_LB_FLAGS)
+f90appflags = -O3 -xHost -fma -fprofile-instr-generate -ipo \
+              -align array64byte -fpp -I${MKLROOT}/include $(POLINT6_FLAG)
+else
+## opt (default): maximum performance with PGO profile data -fprofile-instr-use=$(PROFDATA) \
+## PGO has been turned off, now tested and found to be negative optimization
+## INTERP_LB_FLAGS has been turned off too, now tested and found to be negative optimization
+
+
+CXXAPPFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
+              -Dfortran3 -Dnewc -I${MKLROOT}/include $(INTERP_LB_FLAGS)
+f90appflags = -O3 -xHost -fp-model fast=2 -fma -ipo \
+              -align array64byte -fpp -I${MKLROOT}/include $(POLINT6_FLAG)
+endif
 
 .SUFFIXES: .o .f90 .C .for .cu
-
-.f90.o:
-	$(f90) $(f90appflags) -c $< -o $@
-
-.C.o:
-	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
-
-.for.o:
-	$(f77) -c $< -o $@
-
-.cu.o:
-	$(Cu) $(CUDA_APP_FLAGS) -c $< -o $@ $(CUDA_LIB_PATH)
-
-# C rewrite of BSSN RHS kernel and helpers
-bssn_rhs_c.o: bssn_rhs_c.C
-	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
-
-fderivs_c.o: fderivs_c.C
-	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
-
-fdderivs_c.o: fdderivs_c.C
-	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
-
-kodiss_c.o: kodiss_c.C
-	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
-
+
+.f90.o:
+	$(f90) $(f90appflags) -c $< -o $@
+
+.C.o:
+	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
+
+.for.o:
+	$(f77) -c $< -o $@
+
+.cu.o:
+	$(Cu) $(CUDA_APP_FLAGS) -c $< -o $@ $(CUDA_LIB_PATH)
+
+# C rewrite of BSSN RHS kernel and helpers
+bssn_rhs_c.o: bssn_rhs_c.C
+	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
+
+fderivs_c.o: fderivs_c.C
+	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
+
+fdderivs_c.o: fdderivs_c.C
+	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
+
+kodiss_c.o: kodiss_c.C
+	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
+
 lopsided_c.o: lopsided_c.C
 	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
 
@@ -82,29 +66,28 @@ lopsided_kodis_c.o: lopsided_kodis_c.C
 
 #interp_lb_profile.o: interp_lb_profile.C interp_lb_profile.h
 #	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
-
-## TwoPunctureABE uses fixed optimal flags (AMD AOCC, no PGO)
-TP_OPTFLAGS = -O3 -march=znver4 -ffast-math -flto \
-              -Dfortran3 -Dnewc -I$(AOCL_ROOT)/include
-
-TwoPunctures.o: TwoPunctures.C
-	${CXX} $(TP_OPTFLAGS) -fopenmp -c $< -o $@
-
-TwoPunctureABE.o: TwoPunctureABE.C
-	${CXX} $(TP_OPTFLAGS) -fopenmp -c $< -o $@
-
-# Input files
-
-## Kernel implementation switch (set USE_CXX_KERNELS=0 to fall back to Fortran)
+
+## TwoPunctureABE uses fixed optimal flags with its own PGO profile, independent of CXXAPPFLAGS
+TP_PROFDATA = /home/$(shell whoami)/AMSS-NCKU/pgo_profile/TwoPunctureABE.profdata
+TP_OPTFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
+              -fprofile-instr-use=$(TP_PROFDATA) \
+              -Dfortran3 -Dnewc -I${MKLROOT}/include
+
+TwoPunctures.o: TwoPunctures.C
+	${CXX} $(TP_OPTFLAGS) -qopenmp -c $< -o $@
+
+TwoPunctureABE.o: TwoPunctureABE.C
+	${CXX} $(TP_OPTFLAGS) -qopenmp -c $< -o $@
+
+# Input files
+
+## Kernel implementation switch (set USE_CXX_KERNELS=0 to fall back to Fortran)
 ifeq ($(USE_CXX_KERNELS),0)
 # Fortran mode: no C rewrite files; bssn_rhs.o is included via F90FILES below
 CFILES =
 else
-# C++ mode (default): C rewrite of bssn/bssn-escalar rhs and helper kernels
+# C++ mode (default): C rewrite of bssn_rhs and helper kernels
 CFILES = bssn_rhs_c.o fderivs_c.o fdderivs_c.o kodiss_c.o lopsided_c.o lopsided_kodis_c.o
-ifeq ($(EFFECTIVE_USE_CXX_ESCALAR_KERNEL),1)
-CFILES += bssn_escalar_rhs_c.o
-endif
 endif
 
 ## RK4 kernel switch (independent from USE_CXX_KERNELS)
@@ -114,95 +97,94 @@ RK4_F90_OBJ =
 else
 RK4_F90_OBJ = rungekutta4_rout.o
 endif
-
-C++FILES = ABE.o Ansorg.o Block.o misc.o monitor.o Parallel.o MPatch.o var.o\
+
+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\
+	   bssnEScalar_class.o perf.o Z4c_class.o NullShellPatch.o\
+	   bssnEM_class.o cpbc_util.o z4c_rhs_point.o checkpoint.o\
+           Parallel_bam.o scalar_class.o transpbh.o NullShellPatch2.o\
+	   NullShellPatch2_Evo.o writefile_f.o interp_lb_profile.o
+	   
+C++FILES_GPU = 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\
 	   bssnEScalar_class.o perf.o Z4c_class.o NullShellPatch.o\
 	   bssnEM_class.o cpbc_util.o z4c_rhs_point.o checkpoint.o\
            Parallel_bam.o scalar_class.o transpbh.o NullShellPatch2.o\
-	   NullShellPatch2_Evo.o writefile_f.o interp_lb_profile.o
-	   
-C++FILES_GPU = ABE.o Ansorg.o Block.o misc.o monitor.o Parallel.o MPatch.o var.o\
-           cgh.o surface_integral.o ShellPatch.o\
-	   bssnEScalar_class.o perf.o Z4c_class.o NullShellPatch.o\
-	   bssnEM_class.o cpbc_util.o z4c_rhs_point.o checkpoint.o\
-           Parallel_bam.o scalar_class.o transpbh.o NullShellPatch2.o\
-	   NullShellPatch2_Evo.o \
-	   bssn_gpu_class.o bssn_step_gpu.o bssn_macro.o writefile_f.o
-
+	   NullShellPatch2_Evo.o bssn_cuda_step.o writefile_f.o
+
 F90FILES_BASE = enforce_algebra.o fmisc.o initial_puncture.o prolongrestrict.o\
 	   prolongrestrict_cell.o prolongrestrict_vertex.o\
 	   $(RK4_F90_OBJ) diff_new.o kodiss.o kodiss_sh.o\
 	   lopsidediff.o sommerfeld_rout.o getnp4.o diff_new_sh.o\
 	   shellfunctions.o bssn_rhs_ss.o Set_Rho_ADM.o\
            getnp4EScalar.o bssnEScalar_rhs.o bssn_constraint.o ricci_gamma.o\
-           fadmquantites_bssn.o Z4c_rhs.o Z4c_rhs_ss.o point_diff_new_sh.o\
-	   cpbc.o getnp4old.o NullEvol.o initial_null.o initial_maxwell.o\
-	   getnpem2.o empart.o NullNews.o fourdcurvature.o\
-	   bssn2adm.o adm_constraint.o adm_ricci_gamma.o\
-	   scalar_rhs.o initial_scalar.o NullEvol2.o initial_null2.o\
-	   NullNews2.o tool_f.o
-
-ifeq ($(USE_CXX_KERNELS),0)
-# Fortran mode: include original bssn_rhs.o
-F90FILES = $(F90FILES_BASE) bssn_rhs.o
-else
-# C++ mode (default): bssn_rhs.o replaced by C++ kernel
-F90FILES = $(F90FILES_BASE)
-endif
-
-F77FILES = zbesh.o
-
-AHFDOBJS = expansion.o expansion_Jacobian.o patch.o coords.o patch_info.o patch_interp.o patch_system.o \
-tgrid.o fd_grid.o ghost_zone.o array.o round.o norm.o fuzzy.o error_exit.o miscfp.o \
-linear_map.o cpm_map.o BH_diagnostics.o setup.o horizon_sequence.o find_horizons.o \
-initial_guess.o Newton.o Jacobian.o ilucg.o IntPnts0.o IntPnts.o
-
-TwoPunctureFILES = TwoPunctureABE.o TwoPunctures.o 
-
-CUDAFILES = bssn_gpu.o bssn_gpu_rhs_ss.o
-
-# file dependences
-$(C++FILES) $(C++FILES_GPU) $(F90FILES) $(CFILES) $(AHFDOBJS) $(CUDAFILES): macrodef.fh
-
-$(C++FILES): Block.h enforce_algebra.h fmisc.h initial_puncture.h macrodef.h\
-	     misc.h monitor.h MyList.h Parallel.h MPatch.h prolongrestrict.h\
-	     rungekutta4_rout.h var.h bssn_class.h bssn_rhs.h sommerfeld_rout.h\
-	     cgh.h surface_integral.h ShellPatch.h shellfunctions.h perf.h\
-             fadmquantites_bssn.h cpbc.h getnp4.h initial_null.h NullEvol.h\
-	     NullShellPatch.h initial_maxwell.h bssnEM_class.h getnpem2.h\
-	     empart.h NullNews.h kodiss.h Parallel_bam.h ricci_gamma.h\
-             initial_null2.h NullShellPatch2.h 
-             
-$(C++FILES_GPU): Block.h enforce_algebra.h fmisc.h initial_puncture.h macrodef.h\
-	     misc.h monitor.h MyList.h Parallel.h MPatch.h prolongrestrict.h\
-	     rungekutta4_rout.h var.h bssn_rhs.h sommerfeld_rout.h\
-	     cgh.h surface_integral.h ShellPatch.h shellfunctions.h perf.h\
-             fadmquantites_bssn.h cpbc.h getnp4.h initial_null.h NullEvol.h\
-	     NullShellPatch.h initial_maxwell.h bssnEM_class.h getnpem2.h\
-	     empart.h NullNews.h kodiss.h Parallel_bam.h ricci_gamma.h\
-             initial_null2.h NullShellPatch2.h \
-             bssn_gpu_class.h bssn_macro.h
-             
-$(AHFDOBJS): cctk.h cctk_Config.h cctk_Types.h cctk_Constants.h myglobal.h
-
-$(C++FILES) $(C++FILES_GPU) $(CFILES) $(AHFDOBJS) $(CUDAFILES): macrodef.h
-
-TwoPunctureFILES: TwoPunctures.h
-
-$(CUDAFILES): bssn_gpu.h gpu_mem.h gpu_rhsSS_mem.h
-
-misc.o : zbesh.o
-
-# projects
-ABE: $(C++FILES) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS)
-	$(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(LDLIBS)
-	
-ABEGPU: $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES)
-	$(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES) $(LDLIBS)
-
-TwoPunctureABE: $(TwoPunctureFILES)
-	$(CLINKER) $(TP_OPTFLAGS) -fopenmp -o $@ $(TwoPunctureFILES) $(LDLIBS)
-
-clean:
-	rm *.o ABE ABEGPU TwoPunctureABE make.log -f
+           fadmquantites_bssn.o Z4c_rhs.o Z4c_rhs_ss.o point_diff_new_sh.o\
+	   cpbc.o getnp4old.o NullEvol.o initial_null.o initial_maxwell.o\
+	   getnpem2.o empart.o NullNews.o fourdcurvature.o\
+	   bssn2adm.o adm_constraint.o adm_ricci_gamma.o\
+	   scalar_rhs.o initial_scalar.o NullEvol2.o initial_null2.o\
+	   NullNews2.o tool_f.o
+
+ifeq ($(USE_CXX_KERNELS),0)
+# Fortran mode: include original bssn_rhs.o
+F90FILES = $(F90FILES_BASE) bssn_rhs.o
+else
+# C++ mode (default): bssn_rhs.o replaced by C++ kernel
+F90FILES = $(F90FILES_BASE)
+endif
+
+F77FILES = zbesh.o
+
+AHFDOBJS = expansion.o expansion_Jacobian.o patch.o coords.o patch_info.o patch_interp.o patch_system.o \
+tgrid.o fd_grid.o ghost_zone.o array.o round.o norm.o fuzzy.o error_exit.o miscfp.o \
+linear_map.o cpm_map.o BH_diagnostics.o setup.o horizon_sequence.o find_horizons.o \
+initial_guess.o Newton.o Jacobian.o ilucg.o IntPnts0.o IntPnts.o
+
+TwoPunctureFILES = TwoPunctureABE.o TwoPunctures.o 
+
+CUDAFILES = bssn_gpu.o bssn_cuda_ops.o
+
+# file dependences
+$(C++FILES) $(C++FILES_GPU) $(F90FILES) $(CFILES) $(AHFDOBJS) $(CUDAFILES): macrodef.fh
+
+$(C++FILES): Block.h enforce_algebra.h fmisc.h initial_puncture.h macrodef.h\
+	     misc.h monitor.h MyList.h Parallel.h MPatch.h prolongrestrict.h\
+	     rungekutta4_rout.h var.h bssn_class.h bssn_rhs.h sommerfeld_rout.h\
+	     cgh.h surface_integral.h ShellPatch.h shellfunctions.h perf.h\
+             fadmquantites_bssn.h cpbc.h getnp4.h initial_null.h NullEvol.h\
+	     NullShellPatch.h initial_maxwell.h bssnEM_class.h getnpem2.h\
+	     empart.h NullNews.h kodiss.h Parallel_bam.h ricci_gamma.h\
+             initial_null2.h NullShellPatch2.h 
+             
+$(C++FILES_GPU): Block.h enforce_algebra.h fmisc.h initial_puncture.h macrodef.h\
+	     misc.h monitor.h MyList.h Parallel.h MPatch.h prolongrestrict.h\
+	     rungekutta4_rout.h var.h bssn_rhs.h sommerfeld_rout.h\
+	     cgh.h surface_integral.h ShellPatch.h shellfunctions.h perf.h\
+             fadmquantites_bssn.h cpbc.h getnp4.h initial_null.h NullEvol.h\
+	     NullShellPatch.h initial_maxwell.h bssnEM_class.h getnpem2.h\
+	     empart.h NullNews.h kodiss.h Parallel_bam.h ricci_gamma.h\
+             initial_null2.h NullShellPatch2.h \
+             bssn_gpu_class.h bssn_macro.h
+             
+$(AHFDOBJS): cctk.h cctk_Config.h cctk_Types.h cctk_Constants.h myglobal.h
+
+$(C++FILES) $(C++FILES_GPU) $(CFILES) $(AHFDOBJS) $(CUDAFILES): macrodef.h
+
+TwoPunctureFILES: TwoPunctures.h
+
+$(CUDAFILES): bssn_gpu.h gpu_mem.h gpu_rhsSS_mem.h
+
+misc.o : zbesh.o
+
+# projects
+ABE: $(C++FILES) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS)
+	$(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(LDLIBS)
+	
+ABEGPU: $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES)
+	$(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES) $(LDLIBS)
+
+TwoPunctureABE: $(TwoPunctureFILES)
+	$(CLINKER) $(TP_OPTFLAGS) -qopenmp -o $@ $(TwoPunctureFILES) $(LDLIBS)
+
+clean:
+	rm *.o ABE ABEGPU TwoPunctureABE make.log -f

AMSS_NCKU_source/makefile.inc

@@ -1,17 +1,36 @@
-## AMD AOCC version with AOCL (Optimized for AMD EPYC Zen 4)
+## GCC version (commented out)
+## filein  = -I/usr/include -I/usr/lib/x86_64-linux-gnu/mpich/include -I/usr/lib/x86_64-linux-gnu/openmpi/lib/ -I/usr/lib/gcc/x86_64-linux-gnu/11/ -I/usr/include/c++/11/
+## filein  = -I/usr/include/ -I/usr/include/openmpi-x86_64/ -I/usr/lib/x86_64-linux-gnu/openmpi/include/ -I/usr/lib/x86_64-linux-gnu/openmpi/lib/ -I/usr/lib/gcc/x86_64-linux-gnu/11/ -I/usr/include/c++/11/
+## LDLIBS  = -L/usr/lib/x86_64-linux-gnu -L/usr/lib64 -L/usr/lib/gcc/x86_64-linux-gnu/11 -lgfortran -lmpi -lgfortran
 
-## AOCL root path for includes and libraries
-AOCL_ROOT ?= /home/gh0s7/AOCC/aocl/5.2.0/aocc
+## Intel oneAPI version with oneMKL (Optimized for performance)
+filein  = -I/usr/include/ -I${MKLROOT}/include
 
-## AOCC-built OpenMPI prefix
-OMPI_PREFIX ?= /home/gh0s7/AOCC/aocc-openmpi
+## Using sequential MKL (OpenMP disabled for better single-threaded performance)
+## 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 -liomp5
+CUDA_LDLIBS = -L/usr/local/cuda-12.9/targets/x86_64-linux/lib -lcudart
 
-filein  = -I/usr/include/ -I$(AOCL_ROOT)/include
+## Memory allocator switch
+##   1 (default) : link Intel oneTBB allocator (libtbbmalloc)
+##   0           : use system default allocator (ptmalloc)
+USE_TBBMALLOC ?= 1
+TBBMALLOC_SO ?= /home/intel/oneapi/2025.3/lib/libtbbmalloc.so
+ifneq ($(wildcard $(TBBMALLOC_SO)),)
+TBBMALLOC_LIBS = -Wl,--no-as-needed $(TBBMALLOC_SO) -Wl,--as-needed
+else
+TBBMALLOC_LIBS = -Wl,--no-as-needed -ltbbmalloc -Wl,--as-needed
+endif
+ifeq ($(USE_TBBMALLOC),1)
+LDLIBS := $(TBBMALLOC_LIBS) $(LDLIBS)
+endif
 
-## Using AOCL BLIS + libFLAME for BLAS/LAPACK
-## AOCC Fortran runtime: -lflang (includes FortranRuntime)
-## AOCC OpenMP runtime: -lomp (LLVM OpenMP)
-LDLIBS  = -L$(AOCL_ROOT)/lib -lblis -lflame -lamdlibm -lflang -lpgmath -lpthread -lm -ldl -lomp
+LDLIBS := $(CUDA_LDLIBS) $(LDLIBS)
+
+## PGO build mode switch (ABE only; TwoPunctureABE always uses opt flags)
+##   opt        : (default) maximum performance with PGO profile-guided optimization
+##   instrument : PGO Phase 1 instrumentation to collect fresh profile data
+PGO_MODE ?= opt
 
 ## Interp_Points load balance profiling mode
 ##   off        : (default) no load balance instrumentation
@@ -32,28 +51,16 @@ endif
 ##   0           : fall back to original Fortran kernels
 USE_CXX_KERNELS ?= 1
 
-## BSSN-EScalar RHS switch
-##   1 (default) : use BSSN-EScalar C wrapper on the normal patch path
-##   0           : keep the original Fortran BSSN-EScalar RHS for precision-safe runs
-## Note: this requires USE_CXX_KERNELS=1 because the wrapper reuses the C BSSN kernel.
-USE_CXX_ESCALAR_KERNEL ?= 1
-
-## Cached transfer switch
-##   auto (default): enable for BSSN vacuum, keep other paths on the safe uncached path
-##   1             : force cached Sync/Restrict/OutBd transfer on evolution hot paths
-##   0             : force the original uncached transfer path
-USE_TRANSFER_CACHE ?= auto
-
 ## RK4 kernel implementation switch
 ##   1 (default) : use C/C++ rewrite of rungekutta4_rout (for optimization experiments)
 ##   0           : use original Fortran rungekutta4_rout.o
 USE_CXX_RK4 ?= 1
 
-f90          = flang
-f77          = flang
-CXX          = clang++
-CC           = clang
-CLINKER      = $(OMPI_PREFIX)/bin/mpicxx
+f90          = ifx
+f77          = ifx
+CXX          = icpx
+CC           = icx
+CLINKER      = mpiicpx
 
 Cu = nvcc
 CUDA_LIB_PATH = -L/usr/lib/cuda/lib64 -I/usr/include -I/usr/lib/cuda/include

AMSS_NCKU_source/surface_integral.h

@@ -20,14 +20,24 @@ using namespace std;
 #include "cgh.h"
 #include "ShellPatch.h"
 #include "NullShellPatch.h"
-#include "NullShellPatch2.h"
-#include "var.h"
-#include "monitor.h"
+#include "NullShellPatch2.h"
+#include "var.h"
+#include "monitor.h"
+#include <map>
 
 class surface_integral
 {
 
 private:
+	struct SpherePointCache
+	{
+		double *pox[3];
+		SpherePointCache()
+		{
+			pox[0] = pox[1] = pox[2] = 0;
+		}
+	};
+
 	int Symmetry, factor;
 	int N_theta, N_phi; // Number of points in Theta & Phi directions
 	double dphi, dcostheta;
@@ -36,15 +46,16 @@ private:
 
 	double *nx_g, *ny_g, *nz_g; // global list of unit normals
 	int myrank, cpusize;
-	int wave_cache_spinw, wave_cache_maxl, wave_cache_modes;
-	double *wave_theta_pos, *wave_theta_neg;
-	double *wave_phi_cos, *wave_phi_sin;
-	void clear_wave_cache();
-	void build_wave_cache(int spinw, int maxl);
+	map<double, SpherePointCache> sphere_point_cache;
+	map<int, double *> shellf_cache;
+
+	void get_surface_points(double rex, double **pox);
+	double *get_shellf_buffer(int num_var);
+	void release_cached_buffers();
 
 public:
 	surface_integral(int iSymmetry);
-	~surface_integral();
+	~surface_integral();
 
 	void surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *Ipsi4,
 				   int spinw, int maxl, int NN, double *RP, double *IP,
@@ -82,37 +93,21 @@ public:
 								 double &, double &, double &, double &, double &, double &, double &,
 								 double &, double &, double &, double &, double &, double &,
 								 double &, double &)); // NN is the length of RP and IP
-	void surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var *trK,
-					   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
-					   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
-					   var *Gmx, var *Gmy, var *Gmz,
-					   var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs,
-					   double *Rout, monitor *Monitor, bool refresh_mass_fields = true);
-	void surf_MassPAng(double rex, int lev, ShellPatch *GH, var *chi, var *trK,
-					   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
-					   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
-					   var *Gmx, var *Gmy, var *Gmz,
-					   var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs,
-					   double *Rout, monitor *Monitor, bool refresh_mass_fields = true);
-	void surf_WaveMassPAng(double rex, int lev, cgh *GH,
-					   var *Rpsi4, var *Ipsi4, int spinw, int maxl, int NN, double *RP, double *IP,
-					   var *chi, var *trK,
-					   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
-					   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
-					   var *Gmx, var *Gmy, var *Gmz,
-					   var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs,
-					   double *Rout, monitor *Monitor, bool refresh_mass_fields = true);
-	void surf_WaveMassPAng(double rex, int lev, ShellPatch *GH,
-					   var *Rpsi4, var *Ipsi4, int spinw, int maxl, int NN, double *RP, double *IP,
-					   var *chi, var *trK,
-					   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
-					   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
-					   var *Gmx, var *Gmy, var *Gmz,
-					   var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs,
-					   double *Rout, monitor *Monitor, bool refresh_mass_fields = true);
-	void surf_Wave(double rex, cgh *GH, ShellPatch *SH,
-				   var *chi, var *trK,
-				   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
+	void surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var *trK,
+					   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
+					   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
+					   var *Gmx, var *Gmy, var *Gmz,
+					   var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs,
+					   double *Rout, monitor *Monitor);
+	void surf_MassPAng(double rex, int lev, ShellPatch *GH, var *chi, var *trK,
+					   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
+					   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
+					   var *Gmx, var *Gmy, var *Gmz,
+					   var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs,
+					   double *Rout, monitor *Monitor);
+	void surf_Wave(double rex, cgh *GH, ShellPatch *SH,
+				   var *chi, var *trK,
+				   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
 				   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
 				   var *chix, var *chiy, var *chiz,
 				   var *trKx, var *trKy, var *trKz,
@@ -131,12 +126,12 @@ public:
 	bool SR_Interp_Points(MyList<var> *VarList, cgh *GH, ShellPatch *SH,
 						  int NN, double **XX, double *Shellf);
 
-	void surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var *trK,
-					   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
-					   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
-					   var *Gmx, var *Gmy, var *Gmz,
-					   var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs, // temparay memory for mass^i
-					   double *Rout, monitor *Monitor, MPI_Comm Comm_here, bool refresh_mass_fields = true);
+	void surf_MassPAng(double rex, int lev, cgh *GH, var *chi, var *trK,
+					   var *gxx, var *gxy, var *gxz, var *gyy, var *gyz, var *gzz,
+					   var *Axx, var *Axy, var *Axz, var *Ayy, var *Ayz, var *Azz,
+					   var *Gmx, var *Gmy, var *Gmz,
+					   var *Sfx_rhs, var *Sfy_rhs, var *Sfz_rhs, // temparay memory for mass^i
+					   double *Rout, monitor *Monitor, MPI_Comm Comm_here);
 	void surf_Wave(double rex, int lev, cgh *GH, var *Rpsi4, var *Ipsi4,
 				   int spinw, int maxl, int NN, double *RP, double *IP,
 				   monitor *Monitor, MPI_Comm Comm_here);

BSSN_BUILD_CONFIG_MIGRATION.md

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

generate_macrodef.py

@@ -12,37 +12,6 @@ import os
 import AMSS_NCKU_Input as input_data          ## import program input file
 
 
-##################################################################
-
-def get_abe_type():
-    if ( input_data.Equation_Class == "BSSN" ):
-        return 0
-    elif ( input_data.Equation_Class == "BSSN-EScalar" ):
-        return 1
-    elif ( input_data.Equation_Class == "BSSN-EM" ):
-        return 3
-    elif ( input_data.Equation_Class == "Z4C" ):
-        return 2
-    else:
-        raise ValueError("Equation_Class setting error!!!")
-
-
-##################################################################
-
-## Generate the makefile fragment used by the copied source tree.
-## The source-tree macrodef.h is not authoritative because macro files
-## are regenerated under File_directory for each run.
-
-def generate_build_config():
-
-    file1 = open( os.path.join(input_data.File_directory, "AMSS_NCKU_build.mk"), "w")
-
-    print( "# Generated by generate_macrodef.py; do not edit manually.", file=file1 )
-    print( f"ABE_TYPE := {get_abe_type()}",                             file=file1 )
-
-    file1.close()
-
-
 ##################################################################
 
 ## Generate the macro file macrodef.h according to user settings
@@ -89,10 +58,19 @@ def generate_macrodef_h():
     # 2: Z4c vacuum
     # 3: coupled to Maxwell field
 
-    try:
-        print( f"#define ABEtype {get_abe_type()}", file=file1 )
-        print(                                      file=file1 )
-    except ValueError:
+    if ( input_data.Equation_Class == "BSSN" ):
+        print( "#define ABEtype 0", file=file1 )
+        print(                      file=file1 )
+    elif ( input_data.Equation_Class == "BSSN-EScalar" ):
+        print( "#define ABEtype 1", file=file1 )
+        print(                      file=file1 )
+    elif ( input_data.Equation_Class == "BSSN-EM" ):
+        print( "#define ABEtype 3", file=file1 )
+        print(                      file=file1 )
+    elif ( input_data.Equation_Class == "Z4C" ):
+        print( "#define ABEtype 2", file=file1 )
+        print(                      file=file1 )
+    else:
         print( "Equation_Class setting error!!!"                )
         print()
         print( "# Equation type #define ABEtype setting error!!!", file=file1 )
@@ -166,62 +144,6 @@ def generate_macrodef_h():
     print( "#define REGLEV 0",      file=file1 )
     print(                          file=file1 )
 
-    # Define fine-grained timing/debug macros.
-    # All of them default to OFF so production builds do not pay profiling overhead.
-
-    fine_timing = getattr(input_data, "Fine_Timing",
-                  getattr(input_data, "Finegrained_Timing", "no"))
-    kernel_fine_timing = getattr(input_data, "Kernel_Fine_Timing",
-                          getattr(input_data, "BSSN_Kernel_Fine_Timing", "no"))
-    stdin_abort_poll = getattr(input_data, "Enable_Stdin_Abort_Poll",
-                       getattr(input_data, "Stdin_Abort_Poll", "no"))
-    timing_report_every = max(1, int(getattr(
-        input_data, "Timing_Every_Steps",
-        getattr(input_data, "Timing_Report_Every", 1))))
-    timing_top_hotspots = max(1, int(getattr(
-        input_data, "Timing_Top_Hotspots", 8)))
-
-    if ( fine_timing == "yes" ):
-        print( "#define BSSN_FINE_TIMING 1", file=file1 )
-        print(                               file=file1 )
-    elif ( fine_timing == "no" ):
-        print( "#define BSSN_FINE_TIMING 0", file=file1 )
-        print(                               file=file1 )
-    else:
-        print( "Fine_Timing setting error!!!" )
-        print()
-        print( "# Fine_Timing setting error!!!", file=file1 )
-        print(                                   file=file1 )
-
-    print( f"#define BSSN_FINE_TIMING_EVERY {timing_report_every}", file=file1 )
-    print(                                                          file=file1 )
-    print( f"#define BSSN_FINE_TIMING_TOPN {timing_top_hotspots}",  file=file1 )
-    print(                                                          file=file1 )
-
-    if ( kernel_fine_timing == "yes" ):
-        print( "#define BSSN_KERNEL_FINE_TIMING 1", file=file1 )
-        print(                                      file=file1 )
-    elif ( kernel_fine_timing == "no" ):
-        print( "#define BSSN_KERNEL_FINE_TIMING 0", file=file1 )
-        print(                                      file=file1 )
-    else:
-        print( "Kernel_Fine_Timing setting error!!!" )
-        print()
-        print( "# Kernel_Fine_Timing setting error!!!", file=file1 )
-        print(                                          file=file1 )
-
-    if ( stdin_abort_poll == "yes" ):
-        print( "#define BSSN_ENABLE_STDIN_ABORT_POLL 1", file=file1 )
-        print(                                           file=file1 )
-    elif ( stdin_abort_poll == "no" ):
-        print( "#define BSSN_ENABLE_STDIN_ABORT_POLL 0", file=file1 )
-        print(                                           file=file1 )
-    else:
-        print( "Enable_Stdin_Abort_Poll setting error!!!" )
-        print()
-        print( "# Enable_Stdin_Abort_Poll setting error!!!", file=file1 )
-        print(                                               file=file1 )
-
     # Define macro USE_GPU
     # use GPU or not
     
@@ -302,21 +224,6 @@ def generate_macrodef_h():
     print( "//     0: for every level;",                                                 file=file1 ) 
     print( "//     1: for all",                                                          file=file1 )
     print( "//",                                                                         file=file1 )
-    print( "// define BSSN_FINE_TIMING",                                                 file=file1 )
-    print( "//     enable fine-grained per-timestep timing monitor",                     file=file1 )
-    print( "//",                                                                         file=file1 )
-    print( "// define BSSN_FINE_TIMING_EVERY",                                           file=file1 )
-    print( "//     report timing every N coarse timesteps",                              file=file1 )
-    print( "//",                                                                         file=file1 )
-    print( "// define BSSN_FINE_TIMING_TOPN",                                            file=file1 )
-    print( "//     number of hottest timing buckets shown in stdout",                    file=file1 )
-    print( "//",                                                                         file=file1 )
-    print( "// define BSSN_KERNEL_FINE_TIMING",                                          file=file1 )
-    print( "//     enable split timing inside compute_rhs_bssn",                         file=file1 )
-    print( "//",                                                                         file=file1 )
-    print( "// define BSSN_ENABLE_STDIN_ABORT_POLL",                                     file=file1 )
-    print( "//     poll stdin and broadcast abort flag every coarse step",               file=file1 )
-    print( "//",                                                                         file=file1 )
     print( "// define USE_GPU",                                                          file=file1 )
     print( "//     use gpu or not",                                                      file=file1 )
     print( "//",                                                                         file=file1 )

makefile_and_run.py

@@ -9,6 +9,7 @@
 
 
 import AMSS_NCKU_Input as input_data
+import os
 import subprocess
 import time
 
@@ -57,6 +58,48 @@ BUILD_JOBS = 64
 ##################################################################
 
 
+##################################################################
+
+def prepare_gpu_runtime_env():
+    """
+    Create a user-private CUDA MPS environment for GPU runs.
+
+    On shared machines another user's daemon may already occupy the default
+    /tmp/nvidia-mps pipe directory, which makes plain cudaSetDevice/cudaMalloc
+    fail with cudaErrorMpsConnectionFailed.  Binding AMSS-NCKU to a private
+    pipe directory avoids cross-user interference.
+    """
+    env = os.environ.copy()
+
+    pipe_dir = env.get("CUDA_MPS_PIPE_DIRECTORY", f"/tmp/amss-ncku-mps-{os.getuid()}")
+    log_dir  = env.get("CUDA_MPS_LOG_DIRECTORY",  f"/tmp/amss-ncku-mps-log-{os.getuid()}")
+
+    os.makedirs(pipe_dir, exist_ok=True)
+    os.makedirs(log_dir, exist_ok=True)
+
+    env["CUDA_MPS_PIPE_DIRECTORY"] = pipe_dir
+    env["CUDA_MPS_LOG_DIRECTORY"] = log_dir
+
+    control_socket = os.path.join(pipe_dir, "control")
+    if not os.path.exists(control_socket):
+        start = subprocess.run(
+            ["nvidia-cuda-mps-control", "-d"],
+            env=env,
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.DEVNULL,
+        )
+        if start.returncode != 0:
+            print(f" Warning: failed to start private CUDA MPS daemon in {pipe_dir}")
+        else:
+            print(f" Using private CUDA MPS pipe directory: {pipe_dir}")
+    else:
+        print(f" Using existing private CUDA MPS pipe directory: {pipe_dir}")
+
+    return env
+
+##################################################################
+
+
 
 ##################################################################
 
@@ -146,16 +189,29 @@ def run_ABE():
 
     ## Define the command to run; cast other values to strings as needed
     
+    run_env = None
+
     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"
+        run_env = prepare_gpu_runtime_env()
+        if int(input_data.MPI_processes) == 1:
+            mpi_command = "./ABEGPU"
+        else:
+            mpi_command = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABEGPU"
         mpi_command_outfile = "ABEGPU_out.log"
  
     ## Execute the MPI command and stream output
-    mpi_process = subprocess.Popen(mpi_command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
+    mpi_process = subprocess.Popen(
+        mpi_command,
+        shell=True,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+        env=run_env,
+    )
 
     ## Write ABE run output to file while printing to stdout
     with open(mpi_command_outfile, 'w') as file0: