Case closed.

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

AMSS_NCKU_GPUCheck.py

@@ -0,0 +1,559 @@
+#!/usr/bin/env python3
+#
+# Current most stable GPU-branch baseline:
+#   GPU_Calculation="yes"
+#   Equation_Class="BSSN"
+#   Initial_Data_Method="Ansorg-TwoPuncture"
+#   puncture_data_set="Manually"
+#   basic_grid_set="Patch"
+#   grid_center_set="Cell"
+#   Symmetry="equatorial-symmetry"
+#   Time_Evolution_Method="runge-kutta-45"
+#   Finite_Diffenence_Method="4th-order"
+#   boundary_choice="BAM-choice"
+#   gauge_choice=0
+#   tetrad_type=2
+#   AHF_Find="no"
+#   devide_factor=2.0
+#   static_grid_type="Linear"
+#   moving_grid_type="Linear"
+#   AMSS_Z4C_MRBD=0
+#   Do not enable AMSS_CUDA_BH_INTERP_RESIDENT unless a dedicated
+#   CPU/GPU trajectory comparison has been run for that configuration.
+"""
+Check whether AMSS_NCKU_Input.py is suitable for the current GPU branch.
+
+Usage:
+  python3 AMSS_NCKU_GPUCheck.py
+  python3 AMSS_NCKU_GPUCheck.py -f /path/to/AMSS_NCKU_Input.py
+"""
+
+from __future__ import annotations
+
+import argparse
+import importlib.util
+import os
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Iterable, List, Sequence
+
+
+SUPPORTED_EQUATIONS = {"BSSN", "BSSN-EScalar", "BSSN-EM", "Z4C"}
+SUPPORTED_INITIAL_DATA = {
+    "Ansorg-TwoPuncture",
+    "Lousto-Analytical",
+    "Cao-Analytical",
+    "KerrSchild-Analytical",
+}
+SUPPORTED_SYMMETRIES = {
+    "no-symmetry",
+    "equatorial-symmetry",
+    "octant-symmetry",
+}
+SUPPORTED_GRIDS = {"Patch", "Shell-Patch"}
+SUPPORTED_CENTERS = {"Cell", "Vertex"}
+SUPPORTED_FD = {"2nd-order", "4th-order", "6th-order", "8th-order"}
+SUPPORTED_GAUGES = {0, 1, 2, 3, 4, 5, 6, 7}
+SUPPORTED_TETRADS = {0, 1, 2}
+SUPPORTED_AHF = {"yes", "no"}
+SUPPORTED_BOUNDARIES = {"BAM-choice", "Shibata-choice"}
+SUPPORTED_PUNCTURE_DATA = {"Manually", "Automatically-BBH"}
+
+STABLE_BASELINE = {
+    "GPU_Calculation": "yes",
+    "Equation_Class": "BSSN",
+    "Initial_Data_Method": "Ansorg-TwoPuncture",
+    "puncture_data_set": "Manually",
+    "basic_grid_set": "Patch",
+    "grid_center_set": "Cell",
+    "Symmetry": "equatorial-symmetry",
+    "Time_Evolution_Method": "runge-kutta-45",
+    "Finite_Diffenence_Method": "4th-order",
+    "boundary_choice": "BAM-choice",
+    "gauge_choice": 0,
+    "tetrad_type": 2,
+    "AHF_Find": "no",
+    "devide_factor": 2.0,
+    "static_grid_type": "Linear",
+    "moving_grid_type": "Linear",
+    "AMSS_Z4C_MRBD": 0,
+}
+
+
+@dataclass
+class CheckResult:
+    ok: bool = True
+    warnings: List[str] = field(default_factory=list)
+    risks: List[str] = field(default_factory=list)
+    notes: List[str] = field(default_factory=list)
+
+    def add_warning(self, msg: str) -> None:
+        self.warnings.append(msg)
+
+    def add_risk(self, msg: str) -> None:
+        self.ok = False
+        self.risks.append(msg)
+
+    def add_note(self, msg: str) -> None:
+        self.notes.append(msg)
+
+    def extend_notes(self, messages: Iterable[str]) -> None:
+        self.notes.extend(messages)
+
+
+def load_input_module(path: Path):
+    spec = importlib.util.spec_from_file_location("amss_ncku_input", str(path))
+    if spec is None or spec.loader is None:
+        raise RuntimeError(f"cannot load input module from {path}")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)  # type: ignore[union-attr]
+    return module
+
+
+def get_attr(mod: Any, name: str, default: Any = None) -> Any:
+    return getattr(mod, name, default)
+
+
+def as_text(value: Any) -> str:
+    if isinstance(value, str):
+        return value.strip()
+    return str(value).strip()
+
+
+def as_lower_text(value: Any) -> str:
+    return as_text(value).lower()
+
+
+def as_float(value: Any, default: float | None = None) -> float | None:
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return default
+
+
+def as_int(value: Any, default: int | None = None) -> int | None:
+    try:
+        return int(value)
+    except (TypeError, ValueError):
+        return default
+
+
+def sequence_len(value: Any) -> int | None:
+    try:
+        return len(value)
+    except TypeError:
+        return None
+
+
+def sequence_values(value: Any) -> List[float] | None:
+    try:
+        return [float(v) for v in value]
+    except (TypeError, ValueError):
+        return None
+
+
+def approx_equal(a: Any, b: float, tol: float = 1.0e-12) -> bool:
+    value = as_float(a)
+    return value is not None and abs(value - b) <= tol
+
+
+def env_truthy(name: str) -> bool:
+    value = os.environ.get(name)
+    return value is not None and value.strip().lower() in {
+        "1",
+        "yes",
+        "y",
+        "true",
+        "on",
+        "enable",
+        "enabled",
+    }
+
+
+def stable_baseline_differences(mod: Any) -> List[str]:
+    diffs = []
+    for name, expected in STABLE_BASELINE.items():
+        if not hasattr(mod, name):
+            continue
+        actual = get_attr(mod, name, None)
+        if isinstance(expected, float):
+            if not approx_equal(actual, expected):
+                diffs.append(f"{name}={actual!r} (stable baseline: {expected!r})")
+        elif actual != expected:
+            diffs.append(f"{name}={actual!r} (stable baseline: {expected!r})")
+    return diffs
+
+
+def add_membership_check(
+    r: CheckResult,
+    name: str,
+    value: Any,
+    supported: Sequence[Any] | set[Any],
+    *,
+    risk_message: str | None = None,
+    note_message: str | None = None,
+) -> None:
+    if value not in supported:
+        r.add_risk(risk_message or f"Unsupported {name}: {value!r}")
+    elif note_message:
+        r.add_note(note_message)
+
+
+def check_positive_int(r: CheckResult, name: str, value: Any) -> None:
+    parsed = as_int(value)
+    if parsed is None or parsed <= 0:
+        r.add_risk(f"{name} must be a positive integer; got {value!r}")
+
+
+def check_nonnegative_number(r: CheckResult, name: str, value: Any) -> None:
+    parsed = as_float(value)
+    if parsed is None or parsed < 0.0:
+        r.add_risk(f"{name} must be a non-negative number; got {value!r}")
+
+
+def check_grid_geometry(r: CheckResult, mod: Any, grid: str) -> None:
+    grid_level = as_int(get_attr(mod, "grid_level", None))
+    static_grid_level = as_int(get_attr(mod, "static_grid_level", None))
+    moving_grid_level = as_int(get_attr(mod, "moving_grid_level", None))
+    refinement_level = as_int(get_attr(mod, "refinement_level", None))
+    analysis_level = as_int(get_attr(mod, "analysis_level", 0))
+
+    for name in (
+        "grid_level",
+        "static_grid_level",
+        "moving_grid_level",
+        "static_grid_number",
+        "moving_grid_number",
+        "quarter_sphere_number",
+    ):
+        check_positive_int(r, name, get_attr(mod, name, None))
+
+    if grid_level is not None and static_grid_level is not None:
+        if static_grid_level > grid_level:
+            r.add_risk("static_grid_level cannot exceed grid_level.")
+        if moving_grid_level is not None and moving_grid_level != grid_level - static_grid_level:
+            r.add_risk(
+                "moving_grid_level should equal grid_level - static_grid_level; "
+                f"got {moving_grid_level}, expected {grid_level - static_grid_level}."
+            )
+    if grid_level is not None:
+        if refinement_level is None or refinement_level < 0 or refinement_level > grid_level:
+            r.add_risk(f"refinement_level must be in [0, grid_level]; got {refinement_level!r}")
+        if analysis_level is None or analysis_level < 0 or analysis_level >= grid_level:
+            r.add_risk(f"analysis_level must be in [0, grid_level); got {analysis_level!r}")
+
+    largest_max = sequence_values(get_attr(mod, "largest_box_xyz_max", None))
+    largest_min = sequence_values(get_attr(mod, "largest_box_xyz_min", None))
+    if largest_max is None or len(largest_max) != 3:
+        r.add_risk("largest_box_xyz_max must contain three numeric values.")
+    elif any(v <= 0.0 for v in largest_max):
+        r.add_risk(f"largest_box_xyz_max values must be positive; got {largest_max!r}")
+    if largest_min is None or len(largest_min) != 3:
+        r.add_risk("largest_box_xyz_min must contain three numeric values.")
+    elif largest_max is not None and len(largest_max) == 3:
+        for idx, (lo, hi) in enumerate(zip(largest_min, largest_max)):
+            if lo >= hi:
+                r.add_risk(
+                    f"largest_box_xyz_min[{idx}] must be smaller than largest_box_xyz_max[{idx}]."
+                )
+
+    if grid == "Shell-Patch" and largest_max is not None and len(largest_max) == 3:
+        if max(largest_max) - min(largest_max) > 1.0e-12:
+            r.add_risk("Shell-Patch requires a cubic largest_box_xyz_max.")
+
+    if not approx_equal(get_attr(mod, "devide_factor", None), 2.0):
+        r.add_risk("devide_factor must remain 2.0; the AMR code documents only this ratio as supported.")
+    if as_text(get_attr(mod, "static_grid_type", "")) != "Linear":
+        r.add_risk("static_grid_type must remain 'Linear'.")
+    if as_text(get_attr(mod, "moving_grid_type", "")) != "Linear":
+        r.add_risk("moving_grid_type must remain 'Linear'.")
+
+    shell_shape = sequence_values(get_attr(mod, "shell_grid_number", None))
+    if grid == "Shell-Patch":
+        if shell_shape is None or len(shell_shape) != 3:
+            r.add_risk("Shell-Patch requires shell_grid_number with three numeric values.")
+        elif any(int(v) <= 0 for v in shell_shape):
+            r.add_risk(f"shell_grid_number values must be positive; got {shell_shape!r}")
+
+
+def check_punctures(r: CheckResult, mod: Any, init: str, puncture_data: str) -> None:
+    puncture_number = as_int(get_attr(mod, "puncture_number", None))
+    if puncture_number is None or puncture_number <= 0:
+        r.add_risk(f"puncture_number must be a positive integer; got {puncture_number!r}")
+        return
+
+    if init == "Ansorg-TwoPuncture" and puncture_number != 2:
+        r.add_warning(
+            "Ansorg-TwoPuncture is validated on the GPU branch mainly for puncture_number=2."
+        )
+    if puncture_data == "Automatically-BBH":
+        r.add_risk("puncture_data_set='Automatically-BBH' is documented as still developing.")
+
+    for name in ("position_BH", "parameter_BH", "dimensionless_spin_BH", "momentum_BH"):
+        value = get_attr(mod, name, None)
+        outer = sequence_len(value)
+        if outer != puncture_number:
+            r.add_risk(f"{name} must have puncture_number rows; got {outer!r}.")
+            continue
+        for idx in range(puncture_number):
+            if sequence_len(value[idx]) != 3:
+                r.add_risk(f"{name}[{idx}] must contain three values.")
+                break
+
+    if init == "Ansorg-TwoPuncture":
+        for name in ("parameter_BH", "position_BH", "momentum_BH"):
+            if get_attr(mod, name, None) is None:
+                r.add_risk(f"Ansorg-TwoPuncture requires {name}.")
+
+
+def check_output_and_time(r: CheckResult, mod: Any) -> None:
+    for name in (
+        "Final_Evolution_Time",
+        "Check_Time",
+        "Dump_Time",
+        "D2_Dump_Time",
+        "Analysis_Time",
+        "Courant_Factor",
+        "Dissipation",
+    ):
+        check_nonnegative_number(r, name, get_attr(mod, name, None))
+    check_positive_int(r, "Evolution_Step_Number", get_attr(mod, "Evolution_Step_Number", None))
+
+    start_time = as_float(get_attr(mod, "Start_Evolution_Time", None))
+    final_time = as_float(get_attr(mod, "Final_Evolution_Time", None))
+    if start_time is None:
+        r.add_risk("Start_Evolution_Time must be numeric.")
+    elif final_time is not None and final_time <= start_time:
+        r.add_risk("Final_Evolution_Time must be greater than Start_Evolution_Time.")
+
+    for name in ("GW_L_max", "GW_M_max", "Detector_Number"):
+        check_positive_int(r, name, get_attr(mod, name, None))
+    detector_min = as_float(get_attr(mod, "Detector_Rmin", None))
+    detector_max = as_float(get_attr(mod, "Detector_Rmax", None))
+    if detector_min is None or detector_min <= 0.0:
+        r.add_risk(f"Detector_Rmin must be positive; got {detector_min!r}")
+    if detector_max is None or detector_max <= 0.0:
+        r.add_risk(f"Detector_Rmax must be positive; got {detector_max!r}")
+    if detector_min is not None and detector_max is not None and detector_max <= detector_min:
+        r.add_risk("Detector_Rmax must be greater than Detector_Rmin.")
+
+
+def check_equation_specific(r: CheckResult, mod: Any, eq: str, grid: str, fd: str) -> None:
+    if eq == "BSSN":
+        r.add_note("Equation_Class=BSSN is the current validated GPU baseline.")
+    elif eq == "BSSN-EScalar":
+        r.add_warning("BSSN-EScalar has a CUDA path, but it is less broadly validated than BSSN.")
+        fr_choice = as_int(get_attr(mod, "FR_Choice", None))
+        if fr_choice not in {1, 2, 3, 4, 5}:
+            r.add_risk(f"FR_Choice must be one of 1..5 for BSSN-EScalar; got {fr_choice!r}")
+        if approx_equal(get_attr(mod, "FR_a2", None), 0.0):
+            r.add_risk("CUDA BSSN-EScalar requires nonzero FR_a2.")
+        elif not approx_equal(get_attr(mod, "FR_a2", None), 3.0):
+            r.add_warning("CUDA BSSN-EScalar now passes FR_a2 to the kernel, but non-3.0 values need CPU/GPU regression.")
+        for name in ("FR_l2", "FR_phi0", "FR_r0", "FR_sigma0"):
+            check_nonnegative_number(r, name, get_attr(mod, name, None))
+    elif eq == "BSSN-EM":
+        r.add_warning(
+            "BSSN-EM is accepted by the build, but this checker cannot certify its physics/output "
+            "without a CPU/GPU regression run."
+        )
+        if fd == "8th-order":
+            r.add_note("BSSN-EM with 8th-order enables extra CUDA AMR batching defaults.")
+    elif eq == "Z4C":
+        r.add_warning(
+            "Z4C has CUDA support, but the resident path and Shell/CPBC combinations are more constrained."
+        )
+        if grid == "Patch":
+            r.add_warning("Z4C+Patch avoids Shell CPBC, but still needs a dedicated regression test.")
+        else:
+            r.add_warning("Z4C+Shell-Patch uses CPBC/Shell logic and is not the stable BSSN baseline.")
+
+
+def check_runtime_environment(r: CheckResult, mod: Any, eq: str, grid: str, fd: str) -> None:
+    if env_truthy("AMSS_CUDA_BH_INTERP_RESIDENT"):
+        r.add_risk(
+            "AMSS_CUDA_BH_INTERP_RESIDENT is enabled in the environment; this option previously caused "
+            "late-time trajectory drift and should stay off unless explicitly revalidated."
+        )
+    else:
+        r.add_note("AMSS_CUDA_BH_INTERP_RESIDENT is not enabled; this matches the fixed stable default.")
+
+    if eq in {"BSSN", "BSSN-EScalar", "Z4C"}:
+        r.add_note("makefile_and_run.py will default AMSS_CUDA_AMR_RESTRICT_DEVICE=1 for this equation.")
+    if fd in {"2nd-order", "8th-order"}:
+        r.add_warning(
+            f"{fd} disables some interpolation/CUDA-aware MPI fast paths by default; validate performance and output."
+        )
+    if grid == "Shell-Patch":
+        r.add_warning(
+            "Shell-Patch changes runtime defaults and MPI process handling; use at least the script-adjusted 4 MPI ranks."
+        )
+
+    z4c_mrbd = as_int(get_attr(mod, "AMSS_Z4C_MRBD", 0), 0)
+    if z4c_mrbd not in {0, 1, 2}:
+        r.add_risk(f"AMSS_Z4C_MRBD must be 0, 1, or 2; got {z4c_mrbd!r}")
+    elif eq == "Z4C" and z4c_mrbd == 2:
+        r.add_risk("Z4C GPU resident path does not support AMSS_Z4C_MRBD=2.")
+    elif eq == "Z4C" and z4c_mrbd in {0, 1}:
+        r.add_note(f"Z4C will build with AMSS_Z4C_MRBD={z4c_mrbd}.")
+
+
+def check_stable_profile(r: CheckResult, mod: Any) -> None:
+    diffs = stable_baseline_differences(mod)
+    if not diffs:
+        r.add_note("This input matches the documented most stable GPU baseline.")
+        return
+    r.add_warning(
+        "This input differs from the documented most stable GPU baseline: " + "; ".join(diffs)
+    )
+
+
+def check_input(mod: Any) -> CheckResult:
+    r = CheckResult()
+
+    gpu_text = as_lower_text(get_attr(mod, "GPU_Calculation", "no"))
+    gpu = gpu_text == "yes"
+    eq = as_text(get_attr(mod, "Equation_Class", ""))
+    init = as_text(get_attr(mod, "Initial_Data_Method", ""))
+    symmetry = as_text(get_attr(mod, "Symmetry", ""))
+    time_method = as_text(get_attr(mod, "Time_Evolution_Method", ""))
+    grid = as_text(get_attr(mod, "basic_grid_set", ""))
+    center = as_text(get_attr(mod, "grid_center_set", ""))
+    fd = as_text(get_attr(mod, "Finite_Diffenence_Method", ""))
+    gauge = get_attr(mod, "gauge_choice", None)
+    tetrad = get_attr(mod, "tetrad_type", None)
+    ahf = as_text(get_attr(mod, "AHF_Find", "no")).lower()
+    boundary = as_text(get_attr(mod, "boundary_choice", ""))
+    puncture_data = as_text(get_attr(mod, "puncture_data_set", ""))
+    cpu_part = get_attr(mod, "CPU_Part", None)
+    gpu_part = get_attr(mod, "GPU_Part", None)
+
+    if gpu_text not in {"yes", "no"}:
+        r.add_risk(f"GPU_Calculation must be 'yes' or 'no'; got {get_attr(mod, 'GPU_Calculation', None)!r}")
+    if not gpu:
+        r.add_note("GPU_Calculation=no; this check only targets the GPU branch.")
+        return r
+
+    r.add_note("GPU_Calculation=yes detected.")
+
+    add_membership_check(r, "Equation_Class", eq, SUPPORTED_EQUATIONS)
+    add_membership_check(r, "Symmetry", symmetry, SUPPORTED_SYMMETRIES)
+    add_membership_check(r, "Initial_Data_Method", init, SUPPORTED_INITIAL_DATA)
+    add_membership_check(r, "basic_grid_set", grid, SUPPORTED_GRIDS)
+    add_membership_check(r, "grid_center_set", center, SUPPORTED_CENTERS)
+    add_membership_check(r, "Finite_Diffenence_Method", fd, SUPPORTED_FD)
+    add_membership_check(r, "gauge_choice", gauge, SUPPORTED_GAUGES)
+    add_membership_check(r, "tetrad_type", tetrad, SUPPORTED_TETRADS)
+    add_membership_check(r, "AHF_Find", ahf, SUPPORTED_AHF)
+    add_membership_check(r, "boundary_choice", boundary, SUPPORTED_BOUNDARIES)
+    add_membership_check(r, "puncture_data_set", puncture_data, SUPPORTED_PUNCTURE_DATA)
+
+    if init != "Ansorg-TwoPuncture":
+        r.add_risk(
+            f"Initial_Data_Method={init!r} is not validated as safe on this GPU branch; "
+            "the stable path is Ansorg-TwoPuncture."
+        )
+    else:
+        r.add_note("Initial_Data_Method=Ansorg-TwoPuncture is supported.")
+
+    if time_method != "runge-kutta-45":
+        r.add_risk(f"Only Time_Evolution_Method='runge-kutta-45' is supported; got {time_method!r}.")
+    if grid == "Patch":
+        r.add_note("basic_grid_set=Patch is the current stable GPU grid path.")
+    elif grid == "Shell-Patch":
+        r.add_warning("basic_grid_set=Shell-Patch has GPU support but is outside the stable BSSN baseline.")
+    if center == "Vertex":
+        r.add_warning("grid_center_set=Vertex is compiled by macros, but the stable GPU baseline is Cell.")
+    if symmetry != "equatorial-symmetry":
+        r.add_warning("The stable validation case uses equatorial-symmetry; other symmetries need regression tests.")
+    if fd != "4th-order":
+        r.add_warning("The stable validation case uses 4th-order finite differences.")
+    if gauge not in {0, 1}:
+        r.add_warning("Input comments recommend gauge_choice 0 or 1; other gauges need dedicated validation.")
+    if tetrad != 2:
+        r.add_warning("Input comments recommend tetrad_type=2; other tetrads affect wave extraction conventions.")
+
+    if ahf == "yes":
+        r.add_warning("AHF_Find=yes is supported by macros, but it is outside the current stable GPU baseline.")
+
+    if boundary == "Shibata-choice":
+        r.add_risk("Shibata-choice is not faithfully distinguished in the current macro generator; it maps to the BAM branch.")
+    elif boundary == "BAM-choice":
+        r.add_note("boundary_choice=BAM-choice is supported.")
+
+    if cpu_part is not None or gpu_part is not None:
+        r.add_warning("CPU_Part/GPU_Part are printed and propagated, but they do not control a real mixed CPU/GPU split in this branch.")
+
+    check_output_and_time(r, mod)
+    check_grid_geometry(r, mod, grid)
+    check_punctures(r, mod, init, puncture_data)
+    check_equation_specific(r, mod, eq, grid, fd)
+    check_runtime_environment(r, mod, eq, grid, fd)
+    check_stable_profile(r, mod)
+
+    return r
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-f",
+        "--file",
+        "--input",
+        dest="input_file",
+        default="AMSS_NCKU_Input.py",
+        help="path to AMSS_NCKU_Input.py",
+    )
+    args = parser.parse_args()
+
+    path = Path(args.input_file).resolve()
+    if not path.exists():
+        print(f"ERROR: input file not found: {path}")
+        return 2
+
+    try:
+        mod = load_input_module(path)
+    except Exception as exc:
+        print(f"ERROR: failed to load input file: {exc}")
+        return 2
+
+    result = check_input(mod)
+
+    print(f"Input: {path}")
+    print(f"GPU_Calculation: {get_attr(mod, 'GPU_Calculation', 'no')}")
+    print(f"Symmetry: {get_attr(mod, 'Symmetry', '')}")
+    print(f"Equation_Class: {get_attr(mod, 'Equation_Class', '')}")
+    print(f"Initial_Data_Method: {get_attr(mod, 'Initial_Data_Method', '')}")
+    print(f"puncture_data_set: {get_attr(mod, 'puncture_data_set', '')}")
+    print(f"basic_grid_set: {get_attr(mod, 'basic_grid_set', '')}")
+    print(f"grid_center_set: {get_attr(mod, 'grid_center_set', '')}")
+    print(f"Finite_Diffenence_Method: {get_attr(mod, 'Finite_Diffenence_Method', '')}")
+    print(f"gauge_choice: {get_attr(mod, 'gauge_choice', '')}")
+    print(f"tetrad_type: {get_attr(mod, 'tetrad_type', '')}")
+    print(f"boundary_choice: {get_attr(mod, 'boundary_choice', '')}")
+    print(f"AHF_Find: {get_attr(mod, 'AHF_Find', '')}")
+    print(f"AMSS_Z4C_MRBD: {get_attr(mod, 'AMSS_Z4C_MRBD', 0)}")
+    print("")
+
+    for msg in result.notes:
+        print(f"NOTE: {msg}")
+    for msg in result.warnings:
+        print(f"WARNING: {msg}")
+    for msg in result.risks:
+        print(f"RISK: {msg}")
+
+    print("")
+    if result.risks:
+        print("Verdict: review the risks above before running.")
+        return 1
+
+    if result.warnings:
+        print("Verdict: runnable on the current GPU branch, but keep the warnings in mind.")
+        return 0
+
+    print("Verdict: OK to run on the current GPU branch.")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

AMSS_NCKU_Input.py

@@ -13,15 +13,31 @@ import numpy
 
 ## Setting MPI processes and the output file directory
 
-File_directory   = "GW150914"                    ## output file directory
+File_directory   = "case3"                    ## output file directory
 Output_directory = "binary_output"               ## binary data file directory
                                                  ## The file directory name should not be too long
-MPI_processes    = 8                             ## number of mpi processes used in the simulation
-
-GPU_Calculation  = "yes"                          ## Use GPU or not 
-                                                 ## (prefer "no" in the current version, because the GPU part may have bugs when integrated in this Python interface)
-CPU_Part         = 1.0
-GPU_Part         = 0.0
+MPI_processes    = 2                             ## number of mpi processes used in the simulation
+
+GPU_Calculation  = "yes"                          ## Use GPU or not 
+                                                 ## (prefer "no" in the current version, because the GPU part may have bugs when integrated in this Python interface)
+CPU_Part         = 1.0
+GPU_Part         = 0.0
+
+## Aggressive runtime overrides for fastest low-accuracy GPU runs.
+AMSS_EVOLVE_TIMING                 = 0
+AMSS_ANALYSIS_MAP_EVERY            = 1000000000
+AMSS_INTERP_FAST                   = 1
+AMSS_INTERP_GPU                    = 1
+AMSS_CUDA_AWARE_MPI                = 1
+AMSS_CUDA_RESIDENT_SYNC            = 1
+AMSS_CUDA_BSSN_RESIDENT_SYNC       = 1
+AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP = 1
+AMSS_CUDA_KEEP_ALL_LEVELS          = 1
+AMSS_CUDA_AMR_RESTRICT_DEVICE      = 1
+AMSS_CUDA_AMR_RESTRICT_BATCH       = 1
+AMSS_CUDA_DEVICE_SEGMENT_BATCH     = 1
+AMSS_CUDA_UNCACHED_DEVICE_BUFFERS  = 1
+AMSS_CUDA_AMR_HOST_STAGED          = 1
 
 #################################################
 
@@ -45,14 +61,14 @@ Finite_Diffenence_Method = "4th-order"             ## finite-difference method:
 ## Setting the time evolutionary information
 
 Start_Evolution_Time     = 0.0                    ## start evolution time t0
-Final_Evolution_Time     = 1000.0                 ## final evolution time t1
-Check_Time               = 100.0
-Dump_Time                = 100.0                  ## time inteval dT for dumping binary data
-D2_Dump_Time             = 100.0                  ## dump the ascii data for 2d surface after dT'
-Analysis_Time            = 0.1                    ## dump the puncture position and GW psi4 after dT"
-Evolution_Step_Number    = 10000000               ## stop the calculation after the maximal step number
-Courant_Factor           = 0.5                    ## Courant Factor
-Dissipation              = 0.15                   ## Kreiss-Oliger Dissipation Strength
+Final_Evolution_Time     = 200.0                 ## final evolution time t1
+Check_Time               = 1000000000.0
+Dump_Time                = 1000000000.0           ## time inteval dT for dumping binary data
+D2_Dump_Time             = 1000000000.0           ## dump the ascii data for 2d surface after dT'
+Analysis_Time            = 1000000000.0           ## dump the puncture position and GW psi4 after dT"
+Evolution_Step_Number    = 1000000                ## stop the calculation after the maximal step number
+Courant_Factor           = 0.8                    ## Courant Factor
+Dissipation              = 0.15                   ## Kreiss-Oliger Dissipation Strength
 
 #################################################
 
@@ -64,22 +80,22 @@ Dissipation              = 0.15                   ## Kreiss-Oliger Dissipation S
 basic_grid_set    = "Patch"                          ## grid structure: choose "Patch" or "Shell-Patch"
 grid_center_set   = "Cell"                           ## grid center: chose "Cell" or "Vertex"
 
-grid_level        = 9                                ## total number of AMR grid levels
-static_grid_level = 5                                ## number of AMR static grid levels
-moving_grid_level = grid_level - static_grid_level   ## number of AMR moving grid levels
-
-analysis_level    = 0
-refinement_level  = 3                                ## time refinement start from this grid level
+grid_level        = 7                                ## total number of AMR grid levels
+static_grid_level = 4                                ## number of AMR static grid levels
+moving_grid_level = grid_level - static_grid_level   ## number of AMR moving grid levels
+
+analysis_level    = 0
+refinement_level  = 2                                ## time refinement start from this grid level
 
 largest_box_xyz_max = [320.0, 320.0, 320.0]          ## scale of the largest box
                                                      ## not ne cess ary to be cubic for "Patch" grid s tructure
                                                      ## need to be a cubic box for "Shell-Patch" grid structure
 largest_box_xyz_min = - numpy.array(largest_box_xyz_max)  
 
-static_grid_number = 96                              ## grid points of each static AMR grid (in x direction)
-                                                     ## (grid points in y and z directions are automatically adjusted)
-moving_grid_number = 48                              ## grid points of each moving AMR grid
-shell_grid_number  = [32, 32, 100]                   ## grid points of Shell-Patch grid
+static_grid_number = 64                               ## grid points of each static AMR grid (in x direction)
+                                                     ## (grid points in y and z directions are automatically adjusted)
+moving_grid_number = 32                               ## grid points of each moving AMR grid
+shell_grid_number  = [32, 32, 100]                   ## grid points of Shell-Patch grid
                                                      ## in (phi, theta, r) direction
 devide_factor      = 2.0                             ## resolution between different grid levels dh0/dh1, only support 2.0 now
                                                      
@@ -87,7 +103,7 @@ devide_factor      = 2.0                             ## resolution between diffe
 static_grid_type   = 'Linear'                        ## AMR static grid structure , only supports "Linear"
 moving_grid_type   = 'Linear'                        ## AMR moving grid structure , only supports "Linear"
 
-quarter_sphere_number = 96                           ## grid number of 1/4 s pher ical surface
+quarter_sphere_number = 16                           ## grid number of 1/4 s pher ical surface
                                                      ## (which is needed for evaluating the spherical surface integral)
 
 #################################################
@@ -110,15 +126,15 @@ puncture_data_set     = "Manually"                       ## Method to give Punct
 
 ## initial orbital distance and ellipticity for BBHs system
 ## ( needed for "Automatically-BBH" case , not affect the "Manually" case )
-Distance = 10.0
+Distance = 12.0
 e0       = 0.0
 
 ## black hole parameter (M Q* a*)
-parameter_BH[0] = [ 36.0/(36.0+29.0),  0.0,  +0.31 ]   
-parameter_BH[1] = [ 29.0/(36.0+29.0),  0.0,  -0.46 ]  
+parameter_BH[0] = [ 0.5,  0.0,  0.0 ]   
+parameter_BH[1] = [ 0.5,  0.0,  0.0 ]  
 ## dimensionless spin in each direction
-dimensionless_spin_BH[0] = [ 0.0,  0.0,  +0.31 ]   
-dimensionless_spin_BH[1] = [ 0.0,  0.0,  -0.46 ]  
+dimensionless_spin_BH[0] = [ 0.0,  0.0,  0.0 ]   
+dimensionless_spin_BH[1] = [ 0.0,  0.0,  0.0 ]  
 
 ## use Brugmann's convention
 ##  -----0-----> y
@@ -129,13 +145,13 @@ dimensionless_spin_BH[1] = [ 0.0,  0.0,  -0.46 ]
 ## If puncture_data_set is chosen to be "Manually", it is necessary to set the position and momentum of each puncture manually
 
 ## initial position for each puncture
-position_BH[0]  = [  0.0,  10.0*29.0/(36.0+29.0), 0.0 ]  
-position_BH[1]  = [  0.0, -10.0*36.0/(36.0+29.0), 0.0 ] 
+position_BH[0]  = [  0.0,  6.0, 0.0 ]  
+position_BH[1]  = [  0.0, -6.0, 0.0 ] 
 
 ## initial mumentum for each puncture
 ## (needed for "Manually" case, does not affect the "Automatically-BBH" case)
-momentum_BH[0]  = [ -0.09530152296974252,  -0.00084541526517121,   0.0 ]
-momentum_BH[1]  = [ +0.09530152296974252,  +0.00084541526517121,   0.0 ]
+momentum_BH[0]  = [ -0.06,  -0.01,   0.0 ]
+momentum_BH[1]  = [ +0.06,  +0.01,   0.0 ]
 
 
 #################################################
@@ -145,11 +161,11 @@ momentum_BH[1]  = [ +0.09530152296974252,  +0.00084541526517121,   0.0 ]
 
 ## Setting the gravitational wave information
 
-GW_L_max        = 4                      ## maximal L number in gravitational wave
-GW_M_max        = 4                      ## maximal M number in gravitational wave
-Detector_Number = 12                     ## number of dector
+GW_L_max        = 2                      ## maximal L number in gravitational wave
+GW_M_max        = 2                      ## maximal M number in gravitational wave
+Detector_Number = 2                     ## number of dector
 Detector_Rmin   = 50.0                   ## nearest dector distance
-Detector_Rmax   = 160.0                  ## farest dector distance
+Detector_Rmax   = 100.0                  ## farest dector distance
 
 #################################################
 
@@ -160,8 +176,8 @@ Detector_Rmax   = 160.0                  ## farest dector distance
 
 AHF_Find       = "no"                    ## whether to find the apparent horizon: choose "yes" or "no"
 
-AHF_Find_Every = 24
-AHF_Dump_Time  = 20.0
+AHF_Find_Every = 1000000000
+AHF_Dump_Time  = 1000000000.0
 
 #################################################
 

AMSS_NCKU_Program.py

@@ -58,31 +58,36 @@ File_directory = os.path.join(input_data.File_directory)
 
 ## If the specified output directory exists, ask the user whether to continue
 if os.path.exists(File_directory):
-    print( " Output dictionary has been existed !!!  "                                                              )
-    print( " If you want to overwrite the existing file directory, please input 'continue' in the terminal !! "     ) 
-    print( " If you want to retain the existing file directory, please input 'stop' in the terminal to stop the "   ) 
-    print( " simulation. Then you can reset the output dictionary in the input script file AMSS_NCKU_Input.py !!! " )
-    print(                                                                                                          )
-    ## Prompt whether to overwrite the existing directory
-    while True:
-        try:
-            inputvalue = input()
-            ## If the user agrees to overwrite, proceed and remove the existing directory
-            if ( inputvalue == "continue" ):
-                print( " Continue the calculation !!! " )
-                print(                                  )
-                break  
-            ## If the user chooses not to overwrite, exit and keep the existing directory
-            elif ( inputvalue == "stop" ):
-                print( " Stop the calculation !!! "    )
-                sys.exit() 
-            ## If the user input is invalid, prompt again
-            else:
+    auto_overwrite = str(getattr(input_data, "Auto_Overwrite_Output", "yes")).strip().lower()
+    if auto_overwrite in ("1", "yes", "y", "true", "on", "continue"):
+        print( " Output dictionary has been existed; Auto_Overwrite_Output=yes, continue the calculation. " )
+        print(                                                                                              )
+    else:
+        print( " Output dictionary has been existed !!!  "                                                              )
+        print( " If you want to overwrite the existing file directory, please input 'continue' in the terminal !! "     ) 
+        print( " If you want to retain the existing file directory, please input 'stop' in the terminal to stop the "   ) 
+        print( " simulation. Then you can reset the output dictionary in the input script file AMSS_NCKU_Input.py !!! " )
+        print(                                                                                                          )
+        ## Prompt whether to overwrite the existing directory
+        while True:
+            try:
+                inputvalue = input()
+                ## If the user agrees to overwrite, proceed and remove the existing directory
+                if ( inputvalue == "continue" ):
+                    print( " Continue the calculation !!! " )
+                    print(                                  )
+                    break  
+                ## If the user chooses not to overwrite, exit and keep the existing directory
+                elif ( inputvalue == "stop" ):
+                    print( " Stop the calculation !!! "    )
+                    sys.exit() 
+                ## If the user input is invalid, prompt again
+                else:
+                    print( " Please input your choice !!! "                   )
+                    print( " Input 'continue' or 'stop' in the terminal !!! " )
+            except ValueError:
                 print( " Please input your choice !!! "                   )
                 print( " Input 'continue' or 'stop' in the terminal !!! " )
-        except ValueError:
-            print( " Please input your choice !!! "                   )
-            print( " Input 'continue' or 'stop' in the terminal !!! " )
         
 ## Remove the existing output directory if present
 shutil.rmtree(File_directory, ignore_errors=True)

AMSS_NCKU_Program_Plot.py

@@ -0,0 +1,100 @@
+##################################################################
+##
+## AMSS-NCKU Plot-Only Restart Script
+## Author: Xiaoqu / Claude
+## 2026/05/12
+##
+## This script checks for existing output data from AMSS_NCKU_Program.py.
+## If data exists, it skips all computation and goes directly to plotting,
+## saving time when plotting was interrupted.
+## If no data is found, it exits with a message.
+##
+##################################################################
+
+## Guard against re-execution by multiprocessing child processes.
+if __name__ != '__main__':
+    import sys as _sys
+    _sys.exit(0)
+
+
+import os
+import sys
+
+import AMSS_NCKU_Input as input_data
+
+##################################################################
+
+## Construct paths from input configuration
+File_directory         = os.path.join(input_data.File_directory)
+output_directory       = os.path.join(File_directory, "AMSS_NCKU_output")
+binary_results_directory = os.path.join(output_directory, input_data.Output_directory)
+figure_directory       = os.path.join(File_directory, "figure")
+
+##################################################################
+
+## Check whether the required output data files exist
+
+required_files = [
+    os.path.join(binary_results_directory, "bssn_BH.dat"),
+    os.path.join(binary_results_directory, "bssn_ADMQs.dat"),
+    os.path.join(binary_results_directory, "bssn_psi4.dat"),
+    os.path.join(binary_results_directory, "bssn_constraint.dat"),
+]
+
+missing_files = [f for f in required_files if not os.path.exists(f)]
+
+if missing_files:
+    print(" No existing AMSS_NCKU_Program.py output data found. ")
+    print(" The following required files are missing: ")
+    for f in missing_files:
+        print(f"   {f}")
+    print()
+    print(" Please run AMSS_NCKU_Program.py first to generate the simulation data. ")
+    print(" Exiting. ")
+    sys.exit(1)
+
+print(" Found existing AMSS_NCKU_Program.py output data. "          )
+print(" Skipping all computation and going directly to plotting. " )
+print()
+
+## Ensure the figure directory exists (it should, but be safe)
+os.makedirs(figure_directory, exist_ok=True)
+
+##################################################################
+
+## Plot the AMSS-NCKU program results
+
+import plot_xiaoqu
+import plot_GW_strain_amplitude_xiaoqu
+from parallel_plot_helper import run_plot_tasks_parallel
+
+plot_tasks = []
+
+## Plot black hole trajectory
+plot_tasks.append((plot_xiaoqu.generate_puncture_orbit_plot,   (binary_results_directory, figure_directory)))
+plot_tasks.append((plot_xiaoqu.generate_puncture_orbit_plot3D, (binary_results_directory, figure_directory)))
+
+## Plot black hole separation vs. time
+plot_tasks.append((plot_xiaoqu.generate_puncture_distence_plot, (binary_results_directory, figure_directory)))
+
+## Plot gravitational waveforms (psi4 and strain amplitude)
+for i in range(input_data.Detector_Number):
+    plot_tasks.append((plot_xiaoqu.generate_gravitational_wave_psi4_plot, (binary_results_directory, figure_directory, i)))
+    plot_tasks.append((plot_GW_strain_amplitude_xiaoqu.generate_gravitational_wave_amplitude_plot, (binary_results_directory, figure_directory, i)))
+
+## Plot ADM mass evolution
+for i in range(input_data.Detector_Number):
+    plot_tasks.append((plot_xiaoqu.generate_ADMmass_plot, (binary_results_directory, figure_directory, i)))
+
+## Plot Hamiltonian constraint violation over time
+for i in range(input_data.grid_level):
+    plot_tasks.append((plot_xiaoqu.generate_constraint_check_plot, (binary_results_directory, figure_directory, i)))
+
+run_plot_tasks_parallel(plot_tasks)
+
+## Plot stored binary data (runs serially, not in the parallel pool)
+plot_xiaoqu.generate_binary_data_plot(binary_results_directory, figure_directory)
+
+print()
+print(" Plotting completed successfully. ")
+print()

AMSS_NCKU_source/ABE.C

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

AMSS_NCKU_source/Block.C

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

AMSS_NCKU_source/Block.h

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

AMSS_NCKU_source/MPatch.C

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

AMSS_NCKU_source/Parallel.h

@@ -100,29 +100,36 @@ namespace Parallel
     MyList<gridseg> **combined_dst;
     int *send_lengths;
     int *recv_lengths;
-    double **send_bufs;
-    double **recv_bufs;
-    int *send_buf_caps;
-    int *recv_buf_caps;
-    unsigned char *send_buf_pinned;
-    unsigned char *recv_buf_pinned;
-    MPI_Request *reqs;
-    MPI_Status *stats;
+    double **send_bufs;
+    double **recv_bufs;
+    int *send_buf_caps;
+    int *recv_buf_caps;
+    unsigned char *send_buf_pinned;
+    unsigned char *recv_buf_pinned;
+    unsigned char *send_buf_is_dev;
+    unsigned char *recv_buf_is_dev;
+    int *send_buf_caps_dev;
+    int *recv_buf_caps_dev;
+    double **send_bufs_dev;
+    double **recv_bufs_dev;
+    MPI_Request *reqs;
+    MPI_Status *stats;
     int max_reqs;
     bool lengths_valid;
     int *tc_req_node;
     int *tc_req_is_recv;
     int *tc_completed;
+    bool cuda_aware_mode;
     SyncCache();
     void invalidate();
     void destroy();
   };
 
-  void Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, SyncCache &cache);
-  void Sync_ensure_cache(MyList<Patch> *PatL, int Symmetry, SyncCache &cache);
-  void transfer_cached(MyList<gridseg> **src, MyList<gridseg> **dst,
-                       MyList<var> *VarList1, MyList<var> *VarList2,
-                       int Symmetry, SyncCache &cache);
+  void Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, SyncCache &cache);
+  void Sync_ensure_cache(MyList<Patch> *PatL, int Symmetry, SyncCache &cache);
+  void transfer_cached(MyList<gridseg> **src, MyList<gridseg> **dst,
+                       MyList<var> *VarList1, MyList<var> *VarList2,
+                       int Symmetry, SyncCache &cache);
 
   struct AsyncSyncState {
     int req_no;
@@ -182,13 +189,13 @@ namespace Parallel
   MyList<Parallel::gridseg> *clone_gsl(MyList<Parallel::gridseg> *p, bool first_only);
   MyList<Parallel::gridseg> *build_bulk_gsl(Patch *Pat); // similar to build_owned_gsl0 but does not care rank issue
   MyList<Parallel::gridseg> *build_bulk_gsl(Block *bp, Patch *Pat);
-  void build_PhysBD_gstl(Patch *Pat, MyList<Parallel::gridseg> *srci, MyList<Parallel::gridseg> *dsti,
-                         MyList<Parallel::gridseg> **out_src, MyList<Parallel::gridseg> **out_dst);
-  void PeriodicBD(Patch *Pat, MyList<var> *VarList, int Symmetry);
-  double L2Norm(Patch *Pat, var *vf);
-  void L2Norm7(Patch *Pat, var **vf, double *norms);
-  void checkgsl(MyList<Parallel::gridseg> *pp, bool first_only);
-  void checkvarl(MyList<var> *pp, bool first_only);
+  void build_PhysBD_gstl(Patch *Pat, MyList<Parallel::gridseg> *srci, MyList<Parallel::gridseg> *dsti,
+                         MyList<Parallel::gridseg> **out_src, MyList<Parallel::gridseg> **out_dst);
+  void PeriodicBD(Patch *Pat, MyList<var> *VarList, int Symmetry);
+  double L2Norm(Patch *Pat, var *vf);
+  void L2Norm7(Patch *Pat, var **vf, double *norms);
+  void checkgsl(MyList<Parallel::gridseg> *pp, bool first_only);
+  void checkvarl(MyList<var> *pp, bool first_only);
   MyList<Parallel::gridseg> *divide_gsl(MyList<Parallel::gridseg> *p, Patch *Pat);
   MyList<Parallel::gridseg> *divide_gs(MyList<Parallel::gridseg> *p, Patch *Pat);
   void prepare_inter_time_level(Patch *Pat,
@@ -220,12 +227,12 @@ namespace Parallel
   void aligncheck(double *bbox0, double *bboxl, int lev, double *DH0, int *shape);
   bool point_locat_gsl(double *pox, MyList<Parallel::gridseg> *gsl);
   void checkpatchlist(MyList<Patch> *PatL, bool buflog);
-
-  double L2Norm(Patch *Pat, var *vf, MPI_Comm Comm_here);
-  void L2Norm7(Patch *Pat, var **vf, double *norms, MPI_Comm Comm_here);
-  bool PatList_Interp_Points(MyList<Patch> *PatL, MyList<var> *VarList,
-                             int NN, double **XX,
-                             double *Shellf, int Symmetry, MPI_Comm Comm_here);
+
+  double L2Norm(Patch *Pat, var *vf, MPI_Comm Comm_here);
+  void L2Norm7(Patch *Pat, var **vf, double *norms, MPI_Comm Comm_here);
+  bool PatList_Interp_Points(MyList<Patch> *PatL, MyList<var> *VarList,
+                             int NN, double **XX,
+                             double *Shellf, int Symmetry, MPI_Comm Comm_here);
 #if (PSTR == 1 || PSTR == 2 || PSTR == 3)
   MyList<Block> *distribute(MyList<Patch> *PatchLIST, int cpusize, int ingfsi, int fngfsi,
                             bool periodic, int start_rank, int end_rank, int nodes = 0);

AMSS_NCKU_source/ShellPatch.h

@@ -102,6 +102,16 @@ public:
                  //-1: means no dumy dimension at all; 0: means rho; 1: means sigma
    };
 
+   // Thread-safe search result (no pointers to shared mutable state)
+   struct PointSearchResult
+   {
+      bool found;
+      Block *Bg;
+      double gx, gy, gz; // global Cartesian coordinates
+      double lx, ly, lz; // local coordinates within the found block
+      int ssst;          // source shell-patch type (-1 = Cartesian)
+   };
+
    int myrank;
    int shape[dim];   // for (rho, sigma, R), for rho and sigma means number of points for every pi/2
    double Rrange[2]; // for Rmin and Rmax
@@ -175,6 +185,12 @@ public:
                          MyList<Patch> *Pp, double CDH[dim], MyList<pointstru> *pss);
    bool prolongpointstru(MyList<pointstru> *&psul, bool ssyn, int tsst, MyList<ss_patch> *sPp, double DH[dim],
                          MyList<Patch> *Pp, double CDH[dim], double x, double y, double z, int Symmetry, int rank_in);
+   // Read-only point search — thread-safe (no shared mutable state modified)
+   PointSearchResult prolongpointstru_search(bool ssyn, int tsst, MyList<ss_patch> *sPp, double DH[dim],
+                                             MyList<Patch> *Pp, double CDH[dim], double x, double y, double z,
+                                             int Symmetry, int rank_in);
+   // Append a search result to a linked list — use inside omp critical section
+   void prolongpointstru_append(MyList<pointstru> *&psul, const PointSearchResult &sr, int tsst);
    void setupintintstuff(int cpusize, MyList<Patch> *CPatL, int Symmetry);
    void intertransfer(MyList<pointstru> **src, MyList<pointstru> **dst,
                       MyList<var> *VarList1 /* source */, MyList<var> *VarList2 /*target */,
@@ -195,11 +211,11 @@ 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 L2Norm7(var **vf, double *norms);
+   void Find_Maximum(MyList<var> *VarList, double *XX, double *Shellf);
+};
 
 #endif /* SHELLPATCH_H */

AMSS_NCKU_source/bssnEScalar_class.C

@@ -1,9 +1,10 @@
 
 #ifdef newc
-#include <sstream>
-#include <cstdio>
-#include <map>
-using namespace std;
+#include <sstream>
+#include <cstdio>
+#include <map>
+#include <string>
+using namespace std;
 #else
 #include <stdio.h>
 #include <map.h>
@@ -24,16 +25,323 @@ using namespace std;
 #include "sommerfeld_rout.h"
 #include "getnp4.h"
 #include "shellfunctions.h"
-#include "parameters.h"
+#include "parameters.h"
+#if USE_CUDA_BSSN
+#include "bssn_rhs_cuda.h"
+#endif
 
 #ifdef With_AHF
 #include "derivatives.h"
 #include "myglobal.h"
 #endif
 
-//================================================================================================
-
-// Define bssnEScalar_class
+//================================================================================================
+
+namespace
+{
+#if USE_CUDA_BSSN
+bool fill_bssn_escalar_cuda_views(Block *cg, MyList<var> *vars,
+                                  double **host_views,
+                                  double *propspeeds = 0,
+                                  double *soa_flat = 0)
+{
+  int idx = 0;
+  while (vars && idx < BSSN_ESCALAR_CUDA_STATE_COUNT)
+  {
+    host_views[idx] = cg->fgfs[vars->data->sgfn];
+    if (propspeeds)
+      propspeeds[idx] = vars->data->propspeed;
+    if (soa_flat)
+    {
+      soa_flat[3 * idx + 0] = vars->data->SoA[0];
+      soa_flat[3 * idx + 1] = vars->data->SoA[1];
+      soa_flat[3 * idx + 2] = vars->data->SoA[2];
+    }
+    vars = vars->next;
+    ++idx;
+  }
+  return idx == BSSN_ESCALAR_CUDA_STATE_COUNT && vars == 0;
+}
+
+bool bssn_escalar_cuda_use_resident_sync(int lev)
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_RESIDENT_SYNC");
+    if (!env)
+      env = getenv("AMSS_CUDA_ESCALAR_RESIDENT_SYNC");
+    enabled = env ? ((atoi(env) != 0) ? 1 : 0) : 1;
+  }
+  if (!enabled)
+    return false;
+#ifdef WithShell
+  (void)lev;
+  return false;
+#else
+  return true;
+#endif
+}
+
+bool bssn_escalar_cuda_keep_resident_after_step(int lev, int trfls_in, int analysis_lev)
+{
+  static int keep_all_levels = -1;
+  if (keep_all_levels < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS");
+    keep_all_levels = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  if (!enabled)
+    return false;
+  if (lev == analysis_lev)
+    return false;
+  static int release_only_level = -2;
+  if (release_only_level == -2)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_RELEASE_ONLY_LEVEL");
+    release_only_level = (env && atoi(env) >= 0) ? atoi(env) : -1;
+  }
+  if (release_only_level >= 0)
+    return lev != release_only_level;
+  static int keep_level_limit = -2;
+  if (keep_level_limit == -2)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_KEEP_LEVELS_BELOW");
+    keep_level_limit = (env && atoi(env) >= 0) ? atoi(env) : -1;
+  }
+  if (keep_level_limit >= 0)
+    return lev < keep_level_limit;
+  if (keep_all_levels)
+    return true;
+  return lev < trfls_in;
+}
+
+bool bssn_escalar_sync_merged_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_ESCALAR_SYNC_MERGED");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+void bssn_escalar_sync_level(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry)
+{
+  if (bssn_escalar_sync_merged_enabled())
+    Parallel::Sync_merged(PatL, VarList, Symmetry);
+  else
+    Parallel::Sync(PatL, VarList, Symmetry);
+}
+
+bool bssn_escalar_timing_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_ESCALAR_STEP_TIMING");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+bool bssn_escalar_cuda_post_rp_download_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_POST_RP_DOWNLOAD");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+bool bssn_escalar_cuda_post_rp_download_level_enabled(int lev)
+{
+  if (!bssn_escalar_cuda_post_rp_download_enabled())
+    return false;
+  static int min_level = -2;
+  if (min_level == -2)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_POST_RP_MIN_LEVEL");
+    min_level = (env && atoi(env) >= 0) ? atoi(env) : -1;
+  }
+  return min_level < 0 || lev >= min_level;
+}
+
+bool bssn_escalar_cuda_post_swap_release_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_POST_SWAP_RELEASE");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+bool bssn_escalar_cuda_pre_rp_release_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_PRE_RP_RELEASE");
+    enabled = env ? ((atoi(env) != 0) ? 1 : 0) : 1;
+  }
+  return enabled != 0;
+}
+
+bool bssn_escalar_cuda_bh_interp_resident_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_BH_INTERP_RESIDENT");
+    enabled = env ? ((atoi(env) != 0) ? 1 : 0) : 0;
+  }
+  return enabled != 0;
+}
+
+bool bssn_escalar_cuda_prune_after_swap_enabled()
+{
+  static int enabled = -1;
+  if (enabled < 0)
+  {
+    const char *env = getenv("AMSS_CUDA_ESCALAR_PRUNE_AFTER_SWAP");
+    enabled = (env && atoi(env) != 0) ? 1 : 0;
+  }
+  return enabled != 0;
+}
+
+void bssn_escalar_cuda_upload_level_state(MyList<Patch> *PatL, MyList<var> *vars,
+                                          int myrank)
+{
+  MyList<Patch> *Pp = PatL;
+  while (Pp)
+  {
+    MyList<Block> *BP = Pp->data->blb;
+    while (BP)
+    {
+      Block *cg = BP->data;
+      if (myrank == cg->rank && bssn_cuda_has_resident_state(cg))
+      {
+        double *state_in[BSSN_ESCALAR_CUDA_STATE_COUNT];
+        if (!fill_bssn_escalar_cuda_views(cg, vars, state_in))
+        {
+          cout << "CUDA BSSN-EScalar resident state list mismatch during upload" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+        if (bssn_escalar_cuda_upload_resident_state(cg, cg->shape, state_in))
+        {
+          cout << "CUDA BSSN-EScalar resident state upload failed" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+      }
+      if (BP == Pp->data->ble)
+        break;
+      BP = BP->next;
+    }
+    Pp = Pp->next;
+  }
+}
+
+void bssn_escalar_cuda_keep_only_level_state(MyList<Patch> *PatL, MyList<var> *vars,
+                                             int myrank)
+{
+  MyList<Patch> *Pp = PatL;
+  while (Pp)
+  {
+    MyList<Block> *BP = Pp->data->blb;
+    while (BP)
+    {
+      Block *cg = BP->data;
+      if (myrank == cg->rank && bssn_cuda_has_resident_state(cg))
+      {
+        double *state_key[BSSN_ESCALAR_CUDA_STATE_COUNT];
+        if (!fill_bssn_escalar_cuda_views(cg, vars, state_key))
+        {
+          cout << "CUDA BSSN-EScalar resident state list mismatch during prune" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+        if (bssn_escalar_cuda_keep_only_resident_state(cg, cg->shape, state_key))
+        {
+          cout << "CUDA BSSN-EScalar resident state prune failed" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+      }
+      if (BP == Pp->data->ble)
+        break;
+      BP = BP->next;
+    }
+    Pp = Pp->next;
+  }
+}
+
+void bssn_escalar_timing_report(int myrank, int lev, int YN, double total, double rhs,
+                                double sync, double bh, double analysis, double swap,
+                                double resident, double rp)
+{
+  if (!bssn_escalar_timing_enabled())
+    return;
+
+  double local[8] = {total, rhs, sync, bh, analysis, swap, resident, rp};
+  double maxv[8] = {};
+  MPI_Reduce(local, maxv, 8, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
+  if (myrank == 0)
+    fprintf(stderr,
+            "[AMSS-ESCALAR-STEP] lev=%d YN=%d total=%.6f rhs=%.6f sync=%.6f "
+            "bh=%.6f analysis=%.6f swap=%.6f resident=%.6f rp=%.6f other=%.6f\n",
+            lev, YN, maxv[0], maxv[1], maxv[2], maxv[3], maxv[4], maxv[5],
+            maxv[6], maxv[7],
+            maxv[0] - maxv[1] - maxv[2] - maxv[3] - maxv[4] - maxv[5] - maxv[6] - maxv[7]);
+}
+
+void bssn_escalar_cuda_download_level_state(MyList<Patch> *PatL, MyList<var> *vars,
+                                            int myrank, bool release_ctx)
+{
+  MyList<Patch> *Pp = PatL;
+  while (Pp)
+  {
+    MyList<Block> *BP = Pp->data->blb;
+    while (BP)
+    {
+      Block *cg = BP->data;
+      if (myrank == cg->rank && bssn_cuda_has_resident_state(cg))
+      {
+        double *state_out[BSSN_ESCALAR_CUDA_STATE_COUNT];
+        if (!fill_bssn_escalar_cuda_views(cg, vars, state_out))
+        {
+          cout << "CUDA BSSN-EScalar resident state list mismatch during download" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+        if (bssn_escalar_cuda_download_resident_state(cg, cg->shape, state_out))
+        {
+          cout << "CUDA BSSN-EScalar resident state download failed" << endl;
+          MPI_Abort(MPI_COMM_WORLD, 1);
+        }
+        if (release_ctx)
+          bssn_cuda_release_step_ctx(cg);
+      }
+      if (BP == Pp->data->ble)
+        break;
+      BP = BP->next;
+    }
+    Pp = Pp->next;
+  }
+}
+#endif
+}
+
+//================================================================================================
+
+// Define bssnEScalar_class
 
 // It inherits some members and methods from the parent class bssn_class and modifies others.
 // The modified members and methods are defined below (and in the header bssnEScalar_class.h).
@@ -177,11 +485,16 @@ void bssnEScalar_class::Initialize()
 
 //================================================================================================
 
-bssnEScalar_class::~bssnEScalar_class()
-{
-  delete Sphio;
-  delete Spio;
-  delete Sphi0;
+bssnEScalar_class::~bssnEScalar_class()
+{
+#if USE_CUDA_BSSN
+  for (int lev = 0; GH && lev < GH->levels; ++lev)
+    bssn_escalar_cuda_download_level_state(GH->PatL[lev], StateList, myrank, true);
+#endif
+
+  delete Sphio;
+  delete Spio;
+  delete Sphi0;
   delete Spi0;
   delete Sphi;
   delete Spi;
@@ -707,7 +1020,12 @@ void bssnEScalar_class::Read_Pablo()
 
 void bssnEScalar_class::Step(int lev, int YN)
 {
-  double dT_lev = dT * pow(0.5, Mymax(lev, trfls));
+  double dT_lev = dT * pow(0.5, Mymax(lev, trfls));
+#if USE_CUDA_BSSN
+  const bool use_cuda_resident_sync = bssn_escalar_cuda_use_resident_sync(lev);
+#else
+  const bool use_cuda_resident_sync = false;
+#endif
 #ifdef With_AHF
   AH_Step_Find(lev, dT_lev);
 #endif
@@ -716,13 +1034,23 @@ void bssnEScalar_class::Step(int lev, int YN)
   if (lev < GH->movls)
     ndeps = numepsb;
   double TRK4 = PhysTime;
-  int iter_count = 0; // count RK4 substeps
-  int pre = 0, cor = 1;
-  int ERROR = 0;
-
-  MyList<ss_patch> *sPp;
-  // Predictor
-  MyList<Patch> *Pp = GH->PatL[lev];
+  int iter_count = 0; // count RK4 substeps
+  int pre = 0, cor = 1;
+  int ERROR = 0;
+  const bool escalar_step_timing = bssn_escalar_timing_enabled();
+  const double escalar_step_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+  double escalar_t_rhs = 0.0;
+  double escalar_t_sync = 0.0;
+  double escalar_t_bh = 0.0;
+  double escalar_t_analysis = 0.0;
+  double escalar_t_swap = 0.0;
+  double escalar_t_resident = 0.0;
+  double escalar_t_rp = 0.0;
+
+  MyList<ss_patch> *sPp;
+  // Predictor
+  double escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+  MyList<Patch> *Pp = GH->PatL[lev];
   while (Pp)
   {
     MyList<Block> *BP = Pp->data->blb;
@@ -731,15 +1059,60 @@ void bssnEScalar_class::Step(int lev, int YN)
       Block *cg = BP->data;
       if (myrank == cg->rank)
       {
-#if (AGM == 0)
-        f_enforce_ga(cg->shape,
-                     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]);
-#endif
-
-        if (f_compute_rhs_bssn_escalar(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
+#if (AGM == 0)
+#if !USE_CUDA_BSSN
+        f_enforce_ga(cg->shape,
+                     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]);
+#endif
+#endif
+
+        bool used_gpu_substep = false;
+#if USE_CUDA_BSSN
+        {
+          double *state_in[BSSN_ESCALAR_CUDA_STATE_COUNT];
+          double *state_out[BSSN_ESCALAR_CUDA_STATE_COUNT];
+          double propspeed[BSSN_ESCALAR_CUDA_STATE_COUNT];
+          double soa_flat[3 * BSSN_ESCALAR_CUDA_STATE_COUNT];
+          if (!fill_bssn_escalar_cuda_views(cg, StateList, state_in, propspeed, soa_flat) ||
+              !fill_bssn_escalar_cuda_views(cg, SynchList_pre, state_out))
+          {
+            cout << "CUDA BSSN-EScalar state list mismatch on predictor step" << endl;
+            MPI_Abort(MPI_COMM_WORLD, 1);
+          }
+          int apply_bam_bc = 0;
+          int apply_enforce_ga = 0;
+#if (AGM == 0)
+          apply_enforce_ga = 1;
+#endif
+#if (SommerType == 0)
+#ifndef WithShell
+          apply_bam_bc = (lev == 0) ? 1 : 0;
+#endif
+#endif
+          int keep_resident_state = use_cuda_resident_sync ? 1 : 0;
+          if (bssn_escalar_cuda_rk4_substep(cg,
+                                            cg->shape, cg->X[0], cg->X[1], cg->X[2],
+                                            state_in, state_out,
+                                            propspeed, soa_flat, Pp->data->bbox,
+                                            dT_lev, TRK4, iter_count, apply_bam_bc,
+                                            Symmetry, lev, ndeps, pre,
+                                            keep_resident_state, apply_enforce_ga, chitiny))
+          {
+            cout << "CUDA BSSN-EScalar predictor substep failed in domain: ("
+                 << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                 << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                 << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+            ERROR = 1;
+          }
+          used_gpu_substep = true;
+        }
+#endif
+
+        if (!used_gpu_substep &&
+            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],
@@ -783,9 +1156,11 @@ void bssnEScalar_class::Step(int lev, int YN)
           ERROR = 1;
         }
 
-        // rk4 substep and boundary
-        {
-          MyList<var> *varl0 = StateList, *varl = SynchList_pre, *varlrhs = RHSList; // we do not check the correspondence here
+        if (!used_gpu_substep)
+        {
+        // rk4 substep and boundary
+        {
+          MyList<var> *varl0 = StateList, *varl = SynchList_pre, *varlrhs = RHSList; // we do not check the correspondence here
           while (varl0)
           {
 #ifndef WithShell
@@ -820,9 +1195,10 @@ void bssnEScalar_class::Step(int lev, int YN)
             varl = varl->next;
             varlrhs = varlrhs->next;
           }
-        }
-        f_lowerboundset(cg->shape, cg->fgfs[phi->sgfn], chitiny);
-      }
+        }
+        f_lowerboundset(cg->shape, cg->fgfs[phi->sgfn], chitiny);
+        }
+      }
       if (BP == Pp->data->ble)
         break;
       BP = BP->next;
@@ -834,19 +1210,21 @@ void bssnEScalar_class::Step(int lev, int YN)
     int erh = ERROR;
     MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
   }
-  if (ERROR)
-  {
+  if (ERROR)
+  {
     Parallel::Dump_Data(GH->PatL[lev], StateList, 0, PhysTime, dT_lev);
     if (myrank == 0)
     {
       if (ErrorMonitor->outfile)
         ErrorMonitor->outfile << "find NaN in state variables at t = " << PhysTime 
                               << ", lev = " << lev << endl;
-      MPI_Abort(MPI_COMM_WORLD, 1);
-    }
-  }
-
-#ifdef WithShell
+      MPI_Abort(MPI_COMM_WORLD, 1);
+    }
+  }
+  if (escalar_step_timing)
+    escalar_t_rhs += MPI_Wtime() - escalar_t0;
+
+#ifdef WithShell
   // evolve Shell Patches
   if (lev == 0)
   {
@@ -993,7 +1371,14 @@ void bssnEScalar_class::Step(int lev, int YN)
   }
 #endif
 
-  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
+  escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+#if USE_CUDA_BSSN
+  bssn_escalar_sync_level(GH->PatL[lev], SynchList_pre, Symmetry);
+#else
+  Parallel::Sync(GH->PatL[lev], SynchList_pre, Symmetry);
+#endif
+  if (escalar_step_timing)
+    escalar_t_sync += MPI_Wtime() - escalar_t0;
 
 #ifdef WithShell
   if (lev == 0)
@@ -1013,10 +1398,15 @@ void bssnEScalar_class::Step(int lev, int YN)
   }
 #endif
 
-  // for black hole position
-  if (BH_num > 0 && lev == GH->levels - 1)
-  {
-    compute_Porg_rhs(Porg0, Porg_rhs, Sfx0, Sfy0, Sfz0, lev);
+  // for black hole position
+  if (BH_num > 0 && lev == GH->levels - 1)
+  {
+    escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+#if USE_CUDA_BSSN
+    if (use_cuda_resident_sync && !bssn_escalar_cuda_bh_interp_resident_enabled())
+      bssn_escalar_cuda_download_level_state(GH->PatL[lev], StateList, myrank, false);
+#endif
+    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);
@@ -1041,19 +1431,29 @@ void bssnEScalar_class::Step(int lev, int YN)
         DG_List->insert(Sfy0);
         DG_List->insert(Sfz0);
         Parallel::Dump_Data(GH->PatL[lev], DG_List, 0, PhysTime, dT_lev);
-        DG_List->clearList();
-      }
-    }
-  }
+        DG_List->clearList();
+      }
+    }
+    if (escalar_step_timing)
+      escalar_t_bh += MPI_Wtime() - escalar_t0;
+  }
   // data analysis part
   // Warning NOTE: the variables1 are used as temp storege room
-  if (lev == a_lev)
-  {
-    AnalysisStuff_EScalar(lev, dT_lev);
-  }
-  // corrector
-  for (iter_count = 1; iter_count < 4; iter_count++)
-  {
+  if (lev == a_lev)
+  {
+    escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+#if USE_CUDA_BSSN
+    if (use_cuda_resident_sync)
+      bssn_escalar_cuda_download_level_state(GH->PatL[lev], SynchList_pre, myrank, false);
+#endif
+    AnalysisStuff_EScalar(lev, dT_lev);
+    if (escalar_step_timing)
+      escalar_t_analysis += MPI_Wtime() - escalar_t0;
+  }
+  // corrector
+  for (iter_count = 1; iter_count < 4; iter_count++)
+  {
+    escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
     // for RK4: t0, t0+dt/2, t0+dt/2, t0+dt;
     if (iter_count == 1 || iter_count == 3)
       TRK4 += dT_lev / 2;
@@ -1066,22 +1466,67 @@ void bssnEScalar_class::Step(int lev, int YN)
         Block *cg = BP->data;
         if (myrank == cg->rank)
         {
-#if (AGM == 0)
-          f_enforce_ga(cg->shape,
-                       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]);
-#elif (AGM == 1)
-          if (iter_count == 3)
-            f_enforce_ga(cg->shape,
+#if (AGM == 0)
+#if !USE_CUDA_BSSN
+          f_enforce_ga(cg->shape,
+                       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]);
+#endif
+#elif (AGM == 1)
+          if (iter_count == 3)
+            f_enforce_ga(cg->shape,
                          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]);
-#endif
-
-          if (f_compute_rhs_bssn_escalar(cg->shape, TRK4, cg->X[0], cg->X[1], cg->X[2],
+#endif
+
+          bool used_gpu_substep = false;
+#if USE_CUDA_BSSN
+          {
+            double *state_in[BSSN_ESCALAR_CUDA_STATE_COUNT];
+            double *state_out[BSSN_ESCALAR_CUDA_STATE_COUNT];
+            double propspeed[BSSN_ESCALAR_CUDA_STATE_COUNT];
+            double soa_flat[3 * BSSN_ESCALAR_CUDA_STATE_COUNT];
+            if (!fill_bssn_escalar_cuda_views(cg, SynchList_pre, state_in, propspeed, soa_flat) ||
+                !fill_bssn_escalar_cuda_views(cg, SynchList_cor, state_out))
+            {
+              cout << "CUDA BSSN-EScalar state list mismatch on corrector step" << endl;
+              MPI_Abort(MPI_COMM_WORLD, 1);
+            }
+            int apply_bam_bc = 0;
+            int apply_enforce_ga = 0;
+#if (AGM == 0)
+            apply_enforce_ga = 1;
+#endif
+#if (SommerType == 0)
+#ifndef WithShell
+            apply_bam_bc = (lev == 0) ? 1 : 0;
+#endif
+#endif
+            int keep_resident_state = use_cuda_resident_sync ? 1 : 0;
+            if (bssn_escalar_cuda_rk4_substep(cg,
+                                              cg->shape, cg->X[0], cg->X[1], cg->X[2],
+                                              state_in, state_out,
+                                              propspeed, soa_flat, Pp->data->bbox,
+                                              dT_lev, TRK4, iter_count, apply_bam_bc,
+                                              Symmetry, lev, ndeps, cor,
+                                              keep_resident_state, apply_enforce_ga, chitiny))
+            {
+              cout << "CUDA BSSN-EScalar corrector substep failed in domain: ("
+                   << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                   << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                   << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+              ERROR = 1;
+            }
+            used_gpu_substep = true;
+          }
+#endif
+
+          if (!used_gpu_substep &&
+              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],
@@ -1125,9 +1570,11 @@ void bssnEScalar_class::Step(int lev, int YN)
                  << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
             ERROR = 1;
           }
-          // rk4 substep and boundary
-          {
-            MyList<var> *varl0 = StateList, *varl = SynchList_pre, *varl1 = SynchList_cor, *varlrhs = RHSList; 
+          if (!used_gpu_substep)
+          {
+          // rk4 substep and boundary
+          {
+            MyList<var> *varl0 = StateList, *varl = SynchList_pre, *varl1 = SynchList_cor, *varlrhs = RHSList; 
             // we do not check the correspondence here
             
             while (varl0)
@@ -1165,9 +1612,10 @@ void bssnEScalar_class::Step(int lev, int YN)
               varl1 = varl1->next;
               varlrhs = varlrhs->next;
             }
-          }
-          f_lowerboundset(cg->shape, cg->fgfs[phi1->sgfn], chitiny);
-        }
+          }
+          f_lowerboundset(cg->shape, cg->fgfs[phi1->sgfn], chitiny);
+          }
+        }
         if (BP == Pp->data->ble)
           break;
         BP = BP->next;
@@ -1180,8 +1628,8 @@ void bssnEScalar_class::Step(int lev, int YN)
       int erh = ERROR;
       MPI_Allreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
     }
-    if (ERROR)
-    {
+    if (ERROR)
+    {
       Parallel::Dump_Data(GH->PatL[lev], SynchList_pre, 0, PhysTime, dT_lev);
       if (myrank == 0)
       {
@@ -1189,11 +1637,13 @@ void bssnEScalar_class::Step(int lev, int YN)
           ErrorMonitor->outfile << "find NaN in RK4 substep#" << iter_count 
                                 << " variables at t = " << PhysTime 
                                 << ", lev = " << lev << endl;
-        MPI_Abort(MPI_COMM_WORLD, 1);
-      }
-    }
-
-#ifdef WithShell
+        MPI_Abort(MPI_COMM_WORLD, 1);
+      }
+    }
+    if (escalar_step_timing)
+      escalar_t_rhs += MPI_Wtime() - escalar_t0;
+
+#ifdef WithShell
     // evolve Shell Patches
     if (lev == 0)
     {
@@ -1349,7 +1799,14 @@ void bssnEScalar_class::Step(int lev, int YN)
     }
 #endif
 
-    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+    escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+#if USE_CUDA_BSSN
+    bssn_escalar_sync_level(GH->PatL[lev], SynchList_cor, Symmetry);
+#else
+    Parallel::Sync(GH->PatL[lev], SynchList_cor, Symmetry);
+#endif
+    if (escalar_step_timing)
+      escalar_t_sync += MPI_Wtime() - escalar_t0;
 
 #ifdef WithShell
     if (lev == 0)
@@ -1368,10 +1825,15 @@ void bssnEScalar_class::Step(int lev, int YN)
       }
     }
 #endif
-    // for black hole position
-    if (BH_num > 0 && lev == GH->levels - 1)
-    {
-      compute_Porg_rhs(Porg, Porg1, Sfx, Sfy, Sfz, lev);
+    // for black hole position
+    if (BH_num > 0 && lev == GH->levels - 1)
+    {
+      escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+#if USE_CUDA_BSSN
+      if (use_cuda_resident_sync && !bssn_escalar_cuda_bh_interp_resident_enabled())
+        bssn_escalar_cuda_download_level_state(GH->PatL[lev], SynchList_pre, myrank, false);
+#endif
+      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);
@@ -1396,14 +1858,17 @@ void bssnEScalar_class::Step(int lev, int YN)
           DG_List->insert(Sfy0);
           DG_List->insert(Sfz0);
           Parallel::Dump_Data(GH->PatL[lev], DG_List, 0, PhysTime, dT_lev);
-          DG_List->clearList();
-        }
-      }
-    }
-    // swap time level
-    if (iter_count < 3)
-    {
-      Pp = GH->PatL[lev];
+          DG_List->clearList();
+        }
+      }
+      if (escalar_step_timing)
+        escalar_t_bh += MPI_Wtime() - escalar_t0;
+    }
+    // swap time level
+    if (iter_count < 3)
+    {
+      escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+      Pp = GH->PatL[lev];
       while (Pp)
       {
         MyList<Block> *BP = Pp->data->blb;
@@ -1444,16 +1909,33 @@ void bssnEScalar_class::Step(int lev, int YN)
           Porg[ithBH][0] = Porg1[ithBH][0];
           Porg[ithBH][1] = Porg1[ithBH][1];
           Porg[ithBH][2] = Porg1[ithBH][2];
-        }
-      }
-    }
-  }
-
-#if (RPS == 0)
-  // mesh refinement boundary part
-  RestrictProlong(lev, YN, BB);
-
-#ifdef WithShell
+        }
+      }
+      if (escalar_step_timing)
+        escalar_t_swap += MPI_Wtime() - escalar_t0;
+    }
+  }
+
+#if USE_CUDA_BSSN
+  if (use_cuda_resident_sync)
+  {
+    escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+    if (!bssn_escalar_cuda_keep_resident_after_step(lev, trfls, a_lev))
+      bssn_escalar_cuda_download_level_state(GH->PatL[lev], SynchList_cor, myrank,
+                                             bssn_escalar_cuda_pre_rp_release_enabled());
+    if (escalar_step_timing)
+      escalar_t_resident += MPI_Wtime() - escalar_t0;
+  }
+#endif
+
+#if (RPS == 0)
+  // mesh refinement boundary part
+  escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+  RestrictProlong(lev, YN, BB);
+  if (escalar_step_timing)
+    escalar_t_rp += MPI_Wtime() - escalar_t0;
+
+#ifdef WithShell
   if (lev == 0)
   {
     clock_t prev_clock, curr_clock;
@@ -1477,8 +1959,9 @@ void bssnEScalar_class::Step(int lev, int YN)
   // StateList     0   -----------
   //
   // OldStateList  old -----------
-  // update
-  Pp = GH->PatL[lev];
+  // update
+  escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+  Pp = GH->PatL[lev];
   while (Pp)
   {
     MyList<Block> *BP = Pp->data->blb;
@@ -1512,18 +1995,45 @@ void bssnEScalar_class::Step(int lev, int YN)
       sPp = sPp->next;
     }
   }
-#endif
-  // for black hole position
-  if (BH_num > 0 && lev == GH->levels - 1)
+#endif
+  #if USE_CUDA_BSSN
+  bool release_after_sync = false;
+  if (use_cuda_resident_sync && bssn_escalar_cuda_post_rp_download_level_enabled(lev))
+  {
+    escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+    release_after_sync = bssn_escalar_cuda_post_swap_release_enabled();
+    bssn_escalar_cuda_download_level_state(GH->PatL[lev], StateList, myrank, release_after_sync);
+    if (escalar_step_timing)
+      escalar_t_resident += MPI_Wtime() - escalar_t0;
+  }
+  if (use_cuda_resident_sync && !release_after_sync &&
+      bssn_escalar_cuda_prune_after_swap_enabled())
+  {
+    escalar_t0 = escalar_step_timing ? MPI_Wtime() : 0.0;
+    bssn_escalar_cuda_keep_only_level_state(GH->PatL[lev], StateList, myrank);
+    if (escalar_step_timing)
+      escalar_t_resident += MPI_Wtime() - escalar_t0;
+  }
+#endif
+  // for black hole position
+  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];
-    }
-  }
-}
+      Porg0[ithBH][2] = Porg1[ithBH][2];
+    }
+  }
+  if (escalar_step_timing)
+  {
+    escalar_t_swap += MPI_Wtime() - escalar_t0;
+    bssn_escalar_timing_report(myrank, lev, YN, MPI_Wtime() - escalar_step_t0,
+                               escalar_t_rhs, escalar_t_sync, escalar_t_bh,
+                               escalar_t_analysis, escalar_t_swap,
+                               escalar_t_resident, escalar_t_rp);
+  }
+}
 
 //================================================================================================
 
@@ -2023,12 +2533,13 @@ void bssnEScalar_class::Interp_Constraint()
     }
   }
 
-  ofstream outfile;
-  char filename[50];
-  sprintf(filename, "%s/interp_constraint_%05d.dat", ErrorMonitor->out_dir.c_str(), int(PhysTime / dT + 0.5)); 
-  // 0.5 for round off
-  
-  outfile.open(filename);
+  ofstream outfile;
+  char suffix[64];
+  sprintf(suffix, "/interp_constraint_%05d.dat", int(PhysTime / dT + 0.5));
+  string filename = ErrorMonitor->out_dir + suffix;
+  // 0.5 for round off
+  
+  outfile.open(filename.c_str());
   outfile << "#  corrdinate, H_Res, Px_Res, Py_Res, Pz_Res, Gx_Res, Gy_Res, Gz_Res, fR_Res, ...." << endl;
   for (int i = 0; i < n; i++)
   {
@@ -2074,14 +2585,44 @@ void bssnEScalar_class::Constraint_Out()
           MyList<Block> *BP = Pp->data->blb;
           while (BP)
           {
-            Block *cg = BP->data;
-            if (myrank == cg->rank)
-            {
-              if (lev > 0)
-                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],
+            Block *cg = BP->data;
+            if (myrank == cg->rank)
+            {
+              bool used_cuda_constraints = false;
+#if USE_CUDA_BSSN
+              {
+                double *state_in[BSSN_ESCALAR_CUDA_STATE_COUNT];
+                if (!fill_bssn_escalar_cuda_views(cg, StateList, state_in))
+                {
+                  cout << "CUDA BSSN-EScalar constraint state list mismatch" << endl;
+                  MPI_Abort(MPI_COMM_WORLD, 1);
+                }
+                double *constraint_out[8] = {
+                    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], cg->fgfs[Cons_fR->sgfn]};
+                int lev_arg = lev;
+                int sym_arg = Symmetry;
+                double eps_arg = ndeps;
+                if (bssn_escalar_cuda_compute_constraints(cg->shape, cg->X[0], cg->X[1], cg->X[2],
+                                                          state_in, constraint_out,
+                                                          sym_arg, lev_arg, eps_arg))
+                {
+                  cout << "CUDA BSSN-EScalar constraint compute failed in domain: ("
+                       << cg->bbox[0] << ":" << cg->bbox[3] << ","
+                       << cg->bbox[1] << ":" << cg->bbox[4] << ","
+                       << cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
+                  MPI_Abort(MPI_COMM_WORLD, 1);
+                }
+                used_cuda_constraints = true;
+              }
+#endif
+              if (!used_cuda_constraints && lev > 0)
+                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],
@@ -2110,15 +2651,16 @@ void bssnEScalar_class::Constraint_Out()
                                            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);
-              f_compute_constraint_fr(cg->shape, cg->X[0], cg->X[1], cg->X[2],
-                                      cg->fgfs[phi0->sgfn], cg->fgfs[trK0->sgfn], 
-                                      cg->fgfs[rho->sgfn], cg->fgfs[Sphi0->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[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);
+              if (!used_cuda_constraints)
+                f_compute_constraint_fr(cg->shape, cg->X[0], cg->X[1], cg->X[2],
+                                      cg->fgfs[phi0->sgfn], cg->fgfs[trK0->sgfn], 
+                                      cg->fgfs[rho->sgfn], cg->fgfs[Sphi0->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[Rxx->sgfn], cg->fgfs[Rxy->sgfn], cg->fgfs[Rxz->sgfn], 

AMSS_NCKU_source/bssn_class.h

@@ -144,7 +144,7 @@ public:
        bssn_class(double Couranti, double StartTimei, double TotalTimei, double DumpTimei, double d2DumpTimei, double CheckTimei, double AnasTimei,
                   int Symmetryi, int checkruni, char *checkfilenamei, double numepssi, double numepsbi, double numepshi,
                   int a_levi, int maxli, int decni, double maxrexi, double drexi);
-       ~bssn_class();
+       virtual ~bssn_class();
 
        void Evolve(int Steps);
        void RecursiveStep(int lev);

AMSS_NCKU_source/bssn_gpu.h

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

AMSS_NCKU_source/bssn_gpu_rhs_ss.cu

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

AMSS_NCKU_source/bssn_rhs_cuda.h

@@ -1,36 +1,39 @@
-#ifndef BSSN_RHS_CUDA_H
-#define BSSN_RHS_CUDA_H
-
+#ifndef BSSN_RHS_CUDA_H
+#define BSSN_RHS_CUDA_H
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 enum {
     BSSN_CUDA_STATE_COUNT = 24,
+    BSSN_ESCALAR_CUDA_STATE_COUNT = 26,
+    BSSN_EM_CUDA_STATE_COUNT = 32,
+    BSSN_EM_CUDA_SOURCE_COUNT = 4,
     BSSN_CUDA_MATTER_COUNT = 10
 };
-
+
 int f_compute_rhs_bssn(int *ex, double &T,
                        double *X, double *Y, double *Z,
                        double *chi, double *trK,
-                       double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz,
-                       double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
-                       double *Gamx, double *Gamy, double *Gamz,
-                       double *Lap, double *betax, double *betay, double *betaz,
-                       double *dtSfx, double *dtSfy, double *dtSfz,
-                       double *chi_rhs, double *trK_rhs,
-                       double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
-                       double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
-                       double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
-                       double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
-                       double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
-                       double *rho, double *Sx, double *Sy, double *Sz,
-                       double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
-                       double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
-                       double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
-                       double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
-                       double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
-                       double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
+                       double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz,
+                       double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
+                       double *Gamx, double *Gamy, double *Gamz,
+                       double *Lap, double *betax, double *betay, double *betaz,
+                       double *dtSfx, double *dtSfy, double *dtSfz,
+                       double *chi_rhs, double *trK_rhs,
+                       double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
+                       double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
+                       double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
+                       double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
+                       double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
+                       double *rho, double *Sx, double *Sy, double *Sz,
+                       double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
+                       double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
+                       double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
+                       double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
+                       double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
+                       double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
                        double *Gmx_Res, double *Gmy_Res, double *Gmz_Res,
                        int &Symmetry, int &Lev, double &eps, int &co);
 
@@ -55,6 +58,54 @@ int bssn_cuda_rk4_substep(void *block_tag,
                           int &apply_enforce_ga,
                           double &chitiny);
 
+int bssn_escalar_cuda_rk4_substep(void *block_tag,
+                                  int *ex, double *X, double *Y, double *Z,
+                                  double **state_host_in,
+                                  double **state_host_out,
+                                  const double *propspeed,
+                                  const double *soa_flat,
+                                  const double *bbox,
+                                  double &dT,
+                                  double &T,
+                                  int &RK4,
+                                  int &apply_bam_bc,
+                                  int &Symmetry,
+	                                  int &Lev,
+	                                  double &eps,
+	                                  int &co,
+	                                  int &keep_resident_state,
+	                                  int &apply_enforce_ga,
+	                                  double &chitiny);
+
+int bssn_escalar_cuda_compute_constraints(int *ex, double *X, double *Y, double *Z,
+                                          double **state_host_in,
+                                          double **constraint_host_out,
+                                          int &Symmetry,
+                                          int &Lev,
+                                          double &eps);
+
+int bssn_em_cuda_rk4_substep(void *block_tag,
+                             int *ex, double *X, double *Y, double *Z,
+                             double **state_host_in,
+                             double **state_host_out,
+                             double **source_host,
+                             const double *propspeed,
+                             const double *soa_flat,
+                             const double *bbox,
+                             double &dT,
+                             double &T,
+                             int &RK4,
+                             int &apply_bam_bc,
+                             int &Symmetry,
+                             int &Lev,
+                             double &eps,
+                             int &co,
+                             int &keep_resident_state,
+                             int &apply_enforce_ga,
+                             double &chitiny);
+
+int bssn_em_cuda_resident_zero_fast_state(void *block_tag);
+
 int bssn_cuda_copy_state_region_to_host(void *block_tag,
                                         int state_index,
                                         double *host_state,
@@ -73,6 +124,37 @@ int bssn_cuda_download_resident_state(void *block_tag,
                                       int *ex,
                                       double **state_host_out);
 
+int bssn_escalar_cuda_download_resident_state(void *block_tag,
+                                              int *ex,
+                                              double **state_host_out);
+
+int bssn_cuda_upload_resident_state_count(void *block_tag,
+                                          int *ex,
+                                          double **state_host_in,
+                                          int state_count);
+
+int bssn_escalar_cuda_upload_resident_state(void *block_tag,
+                                            int *ex,
+                                            double **state_host_in);
+
+int bssn_cuda_keep_only_resident_state_count(void *block_tag,
+                                             int *ex,
+                                             double **state_host_key,
+                                             int state_count);
+
+int bssn_escalar_cuda_keep_only_resident_state(void *block_tag,
+                                               int *ex,
+                                               double **state_host_key);
+
+int bssn_cuda_download_resident_state_count_if_present(void *block_tag,
+                                                       int *ex,
+                                                       double **state_host_out,
+                                                       int state_count);
+
+int bssn_cuda_download_resident_state_if_present(void *block_tag,
+                                                int *ex,
+                                                double **state_host_out);
+
 int bssn_cuda_download_constraint_outputs(int *ex,
                                           double **constraint_host_out);
 
@@ -83,6 +165,45 @@ int bssn_cuda_pack_state_region_to_host_buffer(void *block_tag,
                                                int i0, int j0, int k0,
                                                int sx, int sy, int sz);
 
+int bssn_cuda_interp_state_point3(void *block_tag,
+                                  int *ex,
+                                  int state0,
+                                  int state1,
+                                  int state2,
+                                  double x0,
+                                  double y0,
+                                  double z0,
+                                  double dx,
+                                  double dy,
+                                  double dz,
+                                  double px,
+                                  double py,
+                                  double pz,
+                                  int ordn,
+                                  int symmetry,
+                                  double **state_host_key,
+                                  const double *soa3,
+                                  double *out3);
+
+int bssn_cuda_interp_host_two_fields(void *block_tag,
+                                     int *ex,
+                                     double *field0,
+                                     double *field1,
+                                     double x0,
+                                     double y0,
+                                     double z0,
+                                     double dx,
+                                     double dy,
+                                     double dz,
+                                     const double *px,
+                                     const double *py,
+                                     const double *pz,
+                                     int npoints,
+                                     int ordn,
+                                     int symmetry,
+                                     const double *soa6,
+                                     double *out_interleaved);
+
 int bssn_cuda_unpack_state_region_from_host_buffer(void *block_tag,
                                                    int state_index,
                                                    double *host_buffer,
@@ -90,6 +211,15 @@ int bssn_cuda_unpack_state_region_from_host_buffer(void *block_tag,
                                                    int i0, int j0, int k0,
                                                    int sx, int sy, int sz);
 
+int bssn_cuda_unpack_state_region_from_host_buffer_for_host_views(void *block_tag,
+                                                                  double **state_host_key,
+                                                                  int state_count,
+                                                                  int state_index,
+                                                                  double *host_buffer,
+                                                                  int *ex,
+                                                                  int i0, int j0, int k0,
+                                                                  int sx, int sy, int sz);
+
 int bssn_cuda_pack_state_batch_to_host_buffer(void *block_tag,
                                               int state_count,
                                               double *host_buffer,
@@ -97,6 +227,14 @@ int bssn_cuda_pack_state_batch_to_host_buffer(void *block_tag,
                                               int i0, int j0, int k0,
                                               int sx, int sy, int sz);
 
+int bssn_cuda_pack_state_batch_to_host_buffer_for_host_views(void *block_tag,
+                                                             double **state_host_key,
+                                                             int state_count,
+                                                             double *host_buffer,
+                                                             int *ex,
+                                                             int i0, int j0, int k0,
+                                                             int sx, int sy, int sz);
+
 int bssn_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
                                                   int state_count,
                                                   double *host_buffer,
@@ -104,6 +242,140 @@ int bssn_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
                                                   int i0, int j0, int k0,
                                                   int sx, int sy, int sz);
 
+int bssn_cuda_unpack_state_batch_from_host_buffer_for_host_views(void *block_tag,
+                                                                 double **state_host_key,
+                                                                 int state_count,
+                                                                 double *host_buffer,
+                                                                 int *ex,
+                                                                 int i0, int j0, int k0,
+                                                                 int sx, int sy, int sz);
+
+int bssn_cuda_pack_state_batch_to_device_buffer(void *block_tag,
+                                                int state_count,
+                                                double *device_buffer,
+                                                int *ex,
+                                                int i0, int j0, int k0,
+                                                int sx, int sy, int sz);
+
+int bssn_cuda_pack_state_batch_to_device_buffer_for_host_views(void *block_tag,
+                                                               double **state_host_key,
+                                                               int state_count,
+                                                               double *device_buffer,
+                                                               int *ex,
+                                                               int i0, int j0, int k0,
+                                                               int sx, int sy, int sz);
+
+int bssn_cuda_unpack_state_batch_from_device_buffer(void *block_tag,
+                                                    int state_count,
+                                                    double *device_buffer,
+                                                    int *ex,
+                                                    int i0, int j0, int k0,
+                                                    int sx, int sy, int sz);
+
+int bssn_cuda_unpack_state_batch_from_device_buffer_for_host_views(void *block_tag,
+                                                                   double **state_host_key,
+                                                                   int state_count,
+                                                                   double *device_buffer,
+                                                                   int *ex,
+                                                                   int i0, int j0, int k0,
+                                                                   int sx, int sy, int sz);
+
+int bssn_cuda_pack_state_segments_to_device_buffer(void *block_tag,
+                                                   int state_count,
+                                                   double *device_buffer,
+                                                   int *ex,
+                                                   int segment_count,
+                                                   const int *segment_meta);
+
+int bssn_cuda_pack_state_segments_to_device_buffer_for_host_views(void *block_tag,
+                                                                  double **state_host_key,
+                                                                  int state_count,
+                                                                  double *device_buffer,
+                                                                  int *ex,
+                                                                  int segment_count,
+                                                                  const int *segment_meta);
+
+int bssn_cuda_unpack_state_segments_from_device_buffer(void *block_tag,
+                                                       int state_count,
+                                                       double *device_buffer,
+                                                       int *ex,
+                                                       int segment_count,
+                                                       const int *segment_meta);
+
+int bssn_cuda_unpack_state_segments_from_device_buffer_for_host_views(void *block_tag,
+                                                                      double **state_host_key,
+                                                                      int state_count,
+                                                                      double *device_buffer,
+                                                                      int *ex,
+                                                                      int segment_count,
+                                                                      const int *segment_meta);
+
+int bssn_cuda_restrict_state_segments_to_device_buffer(void *block_tag,
+                                                       int state_count,
+                                                       double *device_buffer,
+                                                       int *ex,
+                                                       int segment_count,
+                                                       const int *segment_meta);
+
+int bssn_cuda_restrict_state_segments_to_device_buffer_for_host_views(void *block_tag,
+                                                                      double **state_host_key,
+                                                                      int state_count,
+                                                                      double *device_buffer,
+                                                                      int *ex,
+                                                                      int segment_count,
+                                                                      const int *segment_meta,
+                                                                      const double *state_soa);
+
+int bssn_cuda_prolong_state_segments_to_device_buffer(void *block_tag,
+                                                      int state_count,
+                                                      double *device_buffer,
+                                                      int *ex,
+                                                      int segment_count,
+                                                      const int *segment_meta);
+
+int bssn_cuda_prolong_state_segments_to_device_buffer_for_host_views(void *block_tag,
+                                                                     double **state_host_key,
+                                                                     int state_count,
+                                                                     double *device_buffer,
+                                                                     int *ex,
+                                                                     int segment_count,
+                                                                     const int *segment_meta,
+                                                                     const double *state_soa);
+
+int bssn_cuda_restrict_state_batch_to_device_buffer(void *block_tag,
+                                                    int state_count,
+                                                    double *device_buffer,
+                                                    int *ex,
+                                                    int sx, int sy, int sz,
+                                                    int fi0, int fj0, int fk0);
+
+int bssn_cuda_restrict_state_batch_to_device_buffer_for_host_views(void *block_tag,
+                                                                   double **state_host_key,
+                                                                   int state_count,
+                                                                   double *device_buffer,
+                                                                   int *ex,
+                                                                   int sx, int sy, int sz,
+                                                                   int fi0, int fj0, int fk0,
+                                                                   const double *state_soa);
+
+int bssn_cuda_prolong_state_batch_to_device_buffer(void *block_tag,
+                                                   int state_count,
+                                                   double *device_buffer,
+                                                   int *ex,
+                                                   int sx, int sy, int sz,
+                                                   int ii0, int jj0, int kk0,
+                                                   int lbc_i, int lbc_j, int lbc_k);
+
+int bssn_cuda_prolong_state_batch_to_device_buffer_for_host_views(void *block_tag,
+                                                                  double **state_host_key,
+                                                                  int state_count,
+                                                                  double *device_buffer,
+                                                                  int *ex,
+                                                                  int sx, int sy, int sz,
+                                                                  int ii0, int jj0, int kk0,
+                                                                  int lbc_i, int lbc_j, int lbc_k,
+                                                                  const double *state_soa);
+
 int bssn_cuda_download_state_subset(void *block_tag,
                                     int *ex,
                                     int subset_count,
@@ -116,12 +388,26 @@ int bssn_cuda_upload_state_subset(void *block_tag,
                                   const int *state_indices,
                                   double **state_host_in);
 
+int bssn_cuda_prepare_inter_time_level(void *block_tag,
+                                       int *ex,
+                                       int state_count,
+                                       double **src1_host_key,
+                                       double **src2_host_key,
+                                       double **src3_host_key,
+                                       double **dst_host_key,
+                                       int source_count,
+                                       int tindex);
+
 int bssn_cuda_has_resident_state(void *block_tag);
 
 void bssn_cuda_release_step_ctx(void *block_tag);
 
 #ifdef __cplusplus
 }
-#endif
-
-#endif
+// C++-only helpers declared for derived equation classes (Z4C, etc.)
+// Defined in bssn_class.C. Requires MyList, Patch, var from including TU.
+bool bssn_cuda_use_resident_sync(int lev);
+void bssn_cuda_download_level_state_if_present(MyList<Patch> *PatL, MyList<var> *vars, int myrank);
+#endif
+
+#endif

AMSS_NCKU_source/checkpoint.C

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

AMSS_NCKU_source/fd_cuda_helpers.cuh

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

AMSS_NCKU_source/gpu_rhsSS_mem.h

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

AMSS_NCKU_source/makefile

@@ -18,9 +18,9 @@ OMP_FLAG = -qopenmp
 
 ifeq ($(PGO_MODE),instrument)
 ## Intel Phase 1: instrumentation — omit -ipo/-fp-model fast=2 for faster build and numerical stability
-CXXAPPFLAGS = -O3 -xHost -fma -fprofile-instr-generate -ipo \
+CXXAPPFLAGS = -O3 -march=znver5  -fma -fprofile-instr-generate -ipo \
               -Dfortran3 -Dnewc $(MKL_INC) $(INTERP_LB_FLAGS)
-f90appflags = -O3 -xHost -fma -fprofile-instr-generate -ipo \
+f90appflags = -O3 -march=znver5  -fma -fprofile-instr-generate -ipo \
               -align array64byte -fpp $(MKL_INC) $(POLINT6_FLAG)
 else
 ## opt (default): maximum performance with PGO profile data -fprofile-instr-use=$(PROFDATA) \
@@ -28,24 +28,23 @@ else
 ## 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 \
+CXXAPPFLAGS = -O3 -march=znver5  -fp-model fast=2 -fma -ipo \
               -Dfortran3 -Dnewc $(MKL_INC) $(INTERP_LB_FLAGS)
-f90appflags = -O3 -xHost -fp-model fast=2 -fma -ipo \
+f90appflags = -O3 -march=znver5  -fp-model fast=2 -fma -ipo \
               -align array64byte -fpp $(MKL_INC) $(POLINT6_FLAG)
 endif
 
-TP_OPTFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
-              -fprofile-instr-use=$(TP_PROFDATA) \
+TP_OPTFLAGS = -O3 -march=znver5  -fp-model fast=2 -fma -ipo \
               -Dfortran3 -Dnewc $(MKL_INC)
 else
 ## NVHPC defaults: mpicc/mpicxx/mpifort wrappers
 ## PGO_MODE is ignored in this branch.
 OMP_FLAG = -mp
-CXXAPPFLAGS = -O3 -tp=host -Mcache_align -Mfma \
+CXXAPPFLAGS = -O3 -march=znver5 -tp=host -Mcache_align -Mfma \
               -Dfortran3 -Dnewc $(MKL_INC) $(INTERP_LB_FLAGS)
-f90appflags = -O3 -tp=host -Mcache_align -Mfma -Mpreprocess \
+f90appflags = -O3 -march=znver5 -tp=host -Mcache_align -Mfma -Mpreprocess \
               $(MKL_INC) $(POLINT6_FLAG)
-TP_OPTFLAGS = -O3 -tp=host -Mcache_align -Mfma \
+TP_OPTFLAGS = -O3 -march=znver5 -tp=host -Mcache_align -Mfma \
               -Dfortran3 -Dnewc $(MKL_INC)
 endif
 
@@ -57,6 +56,10 @@ endif
 .C.o:
 	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
 
+# ShellPatch.C uses OpenMP for setupintintstuff search loops
+ShellPatch.o: ShellPatch.C
+	${CXX} $(CXXAPPFLAGS) $(OMP_FLAG) -c $< $(filein) -o $@
+
 .for.o:
 	$(f77) -c $< -o $@
 
@@ -64,11 +67,15 @@ endif
 	$(Cu) $(CUDA_APP_FLAGS) -c $< -o $@ $(CUDA_LIB_PATH)
 
 # CUDA rewrite of BSSN RHS (drop-in replacement for bssn_rhs_c + stencil helpers)
-bssn_rhs_cuda.o: bssn_rhs_cuda.cu bssn_rhs.h macrodef.h
+bssn_rhs_cuda.o: bssn_rhs_cuda.cu bssn_rhs.h macrodef.h fd_cuda_helpers.cuh
+	$(Cu) $(CUDA_APP_FLAGS) -c $< -o $@ $(CUDA_LIB_PATH)
+
+# CUDA rewrite of BSSN Shell-Patch RHS (drop-in replacement for bssn_rhs_ss)
+bssn_gpu_rhs_ss.o: bssn_gpu_rhs_ss.cu bssn_gpu.h gpu_rhsSS_mem.h bssn_macro.h macrodef.fh
 	$(Cu) $(CUDA_APP_FLAGS) -c $< -o $@ $(CUDA_LIB_PATH)
 
 # CUDA rewrite of Z4C Cartesian RHS
-z4c_rhs_cuda.o: z4c_rhs_cuda.cu z4c_rhs_cuda.h bssn_rhs.h macrodef.h ricci_gamma.h
+z4c_rhs_cuda.o: z4c_rhs_cuda.cu z4c_rhs_cuda.h bssn_rhs.h macrodef.h ricci_gamma.h fd_cuda_helpers.cuh
 	$(Cu) $(CUDA_APP_FLAGS) -c $< -o $@ $(CUDA_LIB_PATH)
 
 # C rewrite of BSSN RHS kernel and helpers
@@ -104,16 +111,19 @@ TwoPunctureABE.o: TwoPunctureABE.C
 
 # Input files
 
-## CUDA BSSN RHS switch
-##   1 : use the rewritten CUDA bssn_rhs backend
-##   0 : keep the normal CPU/Fortran selection below
-USE_CUDA_BSSN ?= 0
-USE_CUDA_Z4C ?= 0
-
-CXXAPPFLAGS += -DUSE_CUDA_BSSN=$(USE_CUDA_BSSN)
-CUDA_APP_FLAGS += -DUSE_CUDA_BSSN=$(USE_CUDA_BSSN)
-CXXAPPFLAGS += -DUSE_CUDA_Z4C=$(USE_CUDA_Z4C)
-CUDA_APP_FLAGS += -DUSE_CUDA_Z4C=$(USE_CUDA_Z4C)
+## CUDA BSSN RHS switch
+##   1 : use the rewritten CUDA bssn_rhs backend
+##   0 : keep the normal CPU/Fortran selection below
+USE_CUDA_BSSN ?= 0
+USE_CUDA_Z4C ?= 0
+AMSS_Z4C_MRBD ?= 0
+
+CXXAPPFLAGS += -DUSE_CUDA_BSSN=$(USE_CUDA_BSSN)
+CUDA_APP_FLAGS += -DUSE_CUDA_BSSN=$(USE_CUDA_BSSN)
+CXXAPPFLAGS += -DUSE_CUDA_Z4C=$(USE_CUDA_Z4C)
+CUDA_APP_FLAGS += -DUSE_CUDA_Z4C=$(USE_CUDA_Z4C)
+CXXAPPFLAGS += -DAMSS_Z4C_MRBD=$(AMSS_Z4C_MRBD)
+CUDA_APP_FLAGS += -DAMSS_Z4C_MRBD=$(AMSS_Z4C_MRBD)
 
 ## Kernel implementation switch (set USE_CXX_KERNELS=0 to fall back to Fortran)
 ifeq ($(USE_CXX_KERNELS),0)
@@ -124,7 +134,7 @@ else
 CFILES_CPU = bssn_rhs_c.o fderivs_c.o fdderivs_c.o kodiss_c.o lopsided_c.o lopsided_kodis_c.o
 endif
 
-CFILES_CUDA_BSSN = bssn_rhs_cuda.o
+CFILES_CUDA_BSSN = bssn_rhs_cuda.o bssn_gpu_rhs_ss.o
 
 ifeq ($(USE_CUDA_BSSN),1)
 CFILES = $(CFILES_CUDA_BSSN)

AMSS_NCKU_source/makefile.inc

@@ -1,7 +1,7 @@
 ## Toolchain selection
 ##   nvhpc : NVIDIA HPC SDK + CUDA-aware MPI (default)
 ##   intel : Intel oneAPI toolchain (legacy path)
-TOOLCHAIN ?= nvhpc
+TOOLCHAIN ?= intel
 
 ## PGO build mode switch (ABE only; TwoPunctureABE always uses opt flags)
 ##   opt        : (default) maximum performance with PGO profile-guided optimization
@@ -88,4 +88,4 @@ endif
 
 Cu = $(NVHPC_ROOT)/compilers/bin/nvcc
 CUDA_LIB_PATH = -L$(CUDA_HOME)/lib64 -I$(CUDA_HOME)/include
-CUDA_APP_FLAGS = -c -g -O3 --ptxas-options=-v -Dfortran3 -Dnewc -arch=$(CUDA_ARCH)
+CUDA_APP_FLAGS = -c -g -O3 --ptxas-options=-v -Dfortran3 -Dnewc -arch=$(CUDA_ARCH)

AMSS_NCKU_source/monitor.C

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

AMSS_NCKU_source/surface_integral.C

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

AMSS_NCKU_source/surface_integral.h

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

AMSS_NCKU_source/z4c_rhs_cuda.cu

@@ -266,6 +266,8 @@ __device__ __forceinline__ double fetch_sym_ord3_direct(const double *src,
                     + (skF - 1) * d_gp.ex[0] * d_gp.ex[1]];
 }
 
+#include "fd_cuda_helpers.cuh"
+
 /* ------------------------------------------------------------------ */
 /*  GPU buffer management                                              */
 /* ------------------------------------------------------------------ */
@@ -420,6 +422,7 @@ static const int k_lk_rhs_slots[BSSN_LK_FIELD_COUNT] = {
 };
 
 __constant__ int d_subset_state_indices[BSSN_STATE_COUNT];
+__constant__ double d_comm_state_soa[3 * BSSN_STATE_COUNT];
 
 static const int k_lk_soa_signs[3 * BSSN_LK_FIELD_COUNT] = {
      1,  1,  1,
@@ -727,6 +730,21 @@ static void upload_grid_params_if_needed(const GridParams &gp)
     }
 }
 
+static void upload_comm_state_soa(const double *state_soa, int state_count)
+{
+    double soa[3 * BSSN_STATE_COUNT];
+    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
+        soa[3 * i + 0] = 1.0;
+        soa[3 * i + 1] = 1.0;
+        soa[3 * i + 2] = 1.0;
+    }
+    if (state_soa) {
+        const int n = (state_count < BSSN_STATE_COUNT) ? state_count : BSSN_STATE_COUNT;
+        std::memcpy(soa, state_soa, (size_t)3 * n * sizeof(double));
+    }
+    CUDA_CHECK(cudaMemcpyToSymbol(d_comm_state_soa, soa, sizeof(soa)));
+}
+
 /* ================================================================== */
 /*  A. Symmetry boundary kernels (ord=2 and ord=3)                     */
 /* ================================================================== */
@@ -1419,45 +1437,10 @@ void kern_fderivs_batched(FDerivTables tables, int field_count)
     const int jF = j0 + 1;
     const int kF = k0 + 1;
 
-    if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
-        (jF + 2) <= jmaxF && (jF - 2) >= jminF &&
-        (kF + 2) <= kmaxF && (kF - 2) >= kminF)
-    {
-        fx[tid] = d_gp.d12dx * (
-            fetch_sym_ord2_direct(src, iF - 2, jF, kF, SoA0, SoA1, SoA2)
-          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
-          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
-          -       fetch_sym_ord2_direct(src, iF + 2, jF, kF, SoA0, SoA1, SoA2));
-        fy[tid] = d_gp.d12dy * (
-            fetch_sym_ord2_direct(src, iF, jF - 2, kF, SoA0, SoA1, SoA2)
-          - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
-          + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
-          -       fetch_sym_ord2_direct(src, iF, jF + 2, kF, SoA0, SoA1, SoA2));
-        fz[tid] = d_gp.d12dz * (
-            fetch_sym_ord2_direct(src, iF, jF, kF - 2, SoA0, SoA1, SoA2)
-          - 8.0 * fetch_sym_ord2_direct(src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
-          + 8.0 * fetch_sym_ord2_direct(src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
-          -       fetch_sym_ord2_direct(src, iF, jF, kF + 2, SoA0, SoA1, SoA2));
-    }
-    else if ((iF + 1) <= imaxF && (iF - 1) >= iminF &&
-             (jF + 1) <= jmaxF && (jF - 1) >= jminF &&
-             (kF + 1) <= kmaxF && (kF - 1) >= kminF)
-    {
-        fx[tid] = d_gp.d2dx * (
-            -fetch_sym_ord2_direct(src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
-            +fetch_sym_ord2_direct(src, iF + 1, jF, kF, SoA0, SoA1, SoA2));
-        fy[tid] = d_gp.d2dy * (
-            -fetch_sym_ord2_direct(src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
-            +fetch_sym_ord2_direct(src, iF, jF + 1, kF, SoA0, SoA1, SoA2));
-        fz[tid] = d_gp.d2dz * (
-            -fetch_sym_ord2_direct(src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
-            +fetch_sym_ord2_direct(src, iF, jF, kF + 1, SoA0, SoA1, SoA2));
-    }
-    else {
-        fx[tid] = 0.0;
-        fy[tid] = 0.0;
-        fz[tid] = 0.0;
-    }
+    fd_compute_first3(src, iF, jF, kF,
+                      iminF, jminF, kminF, imaxF, jmaxF, kmaxF,
+                      SoA0, SoA1, SoA2,
+                      fx[tid], fy[tid], fz[tid]);
 }
 
 __global__ __launch_bounds__(128, 4)
@@ -1497,6 +1480,12 @@ void kern_fdderivs_batched(FDDerivTables tables, int field_count)
     const int jF = j0 + 1;
     const int kF = k0 + 1;
 
+#if ghost_width != 3
+    fd_compute_second6(src, iF, jF, kF,
+                       iminF, jminF, kminF, imaxF, jmaxF, kmaxF,
+                       SoA0, SoA1, SoA2,
+                       fxx[tid], fxy[tid], fxz[tid], fyy[tid], fyz[tid], fzz[tid]);
+#else
     if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
         (jF + 2) <= jmaxF && (jF - 2) >= jminF &&
         (kF + 2) <= kmaxF && (kF - 2) >= kminF)
@@ -1624,12 +1613,43 @@ void kern_fdderivs_batched(FDDerivTables tables, int field_count)
         fxx[tid] = 0.0; fxy[tid] = 0.0; fxz[tid] = 0.0;
         fyy[tid] = 0.0; fyz[tid] = 0.0; fzz[tid] = 0.0;
     }
+#endif
 }
 
+static void gpu_fderivs_batch(int field_count,
+                              double *const *src_fields,
+                              double *const *fx_fields,
+                              double *const *fy_fields,
+                              double *const *fz_fields,
+                              const int *soa_signs,
+                              int all);
+static void gpu_fdderivs_batch(int field_count,
+                               double *const *src_fields,
+                               double *const *fxx_fields,
+                               double *const *fxy_fields,
+                               double *const *fxz_fields,
+                               double *const *fyy_fields,
+                               double *const *fyz_fields,
+                               double *const *fzz_fields,
+                               const int *soa_signs,
+                               int all);
+static void gpu_lopsided_kodis_single_batch(double *d_f_adv, double *d_f_ko, double *d_f_rhs,
+                                            double *d_Sfx, double *d_Sfy, double *d_Sfz,
+                                            double SoA0, double SoA1, double SoA2,
+                                            double eps_val, int all);
+
 /* symmetry_bd on GPU for ord=2, then launch fderivs kernel */
 static void gpu_fderivs(double *d_f, double *d_fx, double *d_fy, double *d_fz,
                         double SoA0, double SoA1, double SoA2, int all)
 {
+#if ghost_width != 3
+    double *src_fields[1] = {d_f};
+    double *fx_fields[1] = {d_fx};
+    double *fy_fields[1] = {d_fy};
+    double *fz_fields[1] = {d_fz};
+    const int soa_signs[3] = {(int)SoA0, (int)SoA1, (int)SoA2};
+    gpu_fderivs_batch(1, src_fields, fx_fields, fy_fields, fz_fields, soa_signs, all);
+#else
     double *fh = g_buf.d_fh2;
     const size_t nx = (size_t)g_buf.prev_nx;
     const size_t ny = (size_t)g_buf.prev_ny;
@@ -1638,6 +1658,7 @@ static void gpu_fderivs(double *d_f, double *d_fx, double *d_fy, double *d_fz,
 
     kern_symbd_pack_ord2<<<grid(w_pack), BLK>>>(d_f, fh, SoA0, SoA1, SoA2);
     kern_fderivs<<<grid(all), BLK>>>(fh, d_fx, d_fy, d_fz);
+#endif
 }
 
 /* symmetry_bd on GPU for ord=2, then launch fdderivs kernel */
@@ -1646,6 +1667,18 @@ static void gpu_fdderivs(double *d_f,
                          double *d_fyy, double *d_fyz, double *d_fzz,
                          double SoA0, double SoA1, double SoA2, int all)
 {
+#if ghost_width != 3
+    double *src_fields[1] = {d_f};
+    double *fxx_fields[1] = {d_fxx};
+    double *fxy_fields[1] = {d_fxy};
+    double *fxz_fields[1] = {d_fxz};
+    double *fyy_fields[1] = {d_fyy};
+    double *fyz_fields[1] = {d_fyz};
+    double *fzz_fields[1] = {d_fzz};
+    const int soa_signs[3] = {(int)SoA0, (int)SoA1, (int)SoA2};
+    gpu_fdderivs_batch(1, src_fields, fxx_fields, fxy_fields, fxz_fields,
+                       fyy_fields, fyz_fields, fzz_fields, soa_signs, all);
+#else
     double *fh = g_buf.d_fh2;
     const size_t nx = (size_t)g_buf.prev_nx;
     const size_t ny = (size_t)g_buf.prev_ny;
@@ -1654,6 +1687,7 @@ static void gpu_fdderivs(double *d_f,
 
     kern_symbd_pack_ord2<<<grid(w_pack), BLK>>>(d_f, fh, SoA0, SoA1, SoA2);
     kern_fdderivs<<<grid(all), BLK>>>(fh, d_fxx, d_fxy, d_fxz, d_fyy, d_fyz, d_fzz);
+#endif
 }
 
 static void gpu_fderivs_batch(int field_count,
@@ -1743,6 +1777,12 @@ void kern_phase10_ricci_batched(const double * __restrict__ gupxx,
         const int jF = j0 + 1;
         const int kF = k0 + 1;
 
+#if ghost_width != 3
+        fd_compute_second6(src, iF, jF, kF,
+                           iminF, jminF, kminF, imaxF, jmaxF, kmaxF,
+                           SoA0, SoA1, SoA2,
+                           fxx, fxy, fxz, fyy, fyz, fzz);
+#else
         if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
             (jF + 2) <= jmaxF && (jF - 2) >= jminF &&
             (kF + 2) <= kmaxF && (kF - 2) >= kminF)
@@ -1866,6 +1906,7 @@ void kern_phase10_ricci_batched(const double * __restrict__ gupxx,
               - fetch_sym_ord2_direct(src, iF, jF - 1, kF + 1, SoA0, SoA1, SoA2)
               + fetch_sym_ord2_direct(src, iF, jF + 1, kF + 1, SoA0, SoA1, SoA2));
         }
+#endif
     }
 
     dst[tid] = gupxx[tid] * fxx + gupyy[tid] * fyy + gupzz[tid] * fzz
@@ -1930,6 +1971,16 @@ void kern_phase14_lap_chi_derivs(const double * __restrict__ Lap,
     const int jF = j0 + 1;
     const int kF = k0 + 1;
 
+#if ghost_width != 3
+    fd_compute_second6(Lap, iF, jF, kF,
+                       iminF, jminF, kminF, imaxF, jmaxF, kmaxF,
+                       1, 1, 1,
+                       fxx[tid], fxy[tid], fxz[tid], fyy[tid], fyz[tid], fzz[tid]);
+    fd_compute_first3(chi, iF, jF, kF,
+                      iminF, jminF, kminF, imaxF, jmaxF, kmaxF,
+                      1, 1, 1,
+                      chix_out[tid], chiy_out[tid], chiz_out[tid]);
+#else
     if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
         (jF + 2) <= jmaxF && (jF - 2) >= jminF &&
         (kF + 2) <= kmaxF && (kF - 2) >= kminF)
@@ -2083,6 +2134,7 @@ void kern_phase14_lap_chi_derivs(const double * __restrict__ Lap,
         fyy[tid] = 0.0; fyz[tid] = 0.0; fzz[tid] = 0.0;
         chix_out[tid] = 0.0; chiy_out[tid] = 0.0; chiz_out[tid] = 0.0;
     }
+#endif
 }
 
 /* Combined ord=3 advection + KO dissipation.
@@ -2094,6 +2146,11 @@ static void gpu_lopsided_kodis(double *d_f_adv, double *d_f_ko, double *d_f_rhs,
                                double SoA0, double SoA1, double SoA2,
                                double eps_val, int all)
 {
+#if ghost_width != 3
+    gpu_lopsided_kodis_single_batch(d_f_adv, d_f_ko, d_f_rhs,
+                                    d_Sfx, d_Sfy, d_Sfz,
+                                    SoA0, SoA1, SoA2, eps_val, all);
+#else
     double *fh = g_buf.d_fh3;
     const size_t nx = (size_t)g_buf.prev_nx;
     const size_t ny = (size_t)g_buf.prev_ny;
@@ -2109,6 +2166,7 @@ static void gpu_lopsided_kodis(double *d_f_adv, double *d_f_ko, double *d_f_rhs,
         }
         kern_kodis<<<grid(all), BLK>>>(fh, d_f_rhs, eps_val);
     }
+#endif
 }
 
 __global__ __launch_bounds__(128, 4)
@@ -2141,6 +2199,24 @@ void kern_lopsided_kodis_batched(const double * __restrict__ Sfx,
     const int jF = j0 + 1;
     const int kF = k0 + 1;
 
+#if ghost_width != 3
+    if (do_lopsided && i0 <= nx - 2 && j0 <= ny - 2 && k0 <= nz - 2) {
+        const double val =
+            fd_lopsided_axis(adv_src, iF, jF, kF, 0, Sfx[tid], iF, iminF, imaxF,
+                             d_gp.dX, SoA0, SoA1, SoA2)
+          + fd_lopsided_axis(adv_src, iF, jF, kF, 1, Sfy[tid], jF, jminF, jmaxF,
+                             d_gp.dY, SoA0, SoA1, SoA2)
+          + fd_lopsided_axis(adv_src, iF, jF, kF, 2, Sfz[tid], kF, kminF, kmaxF,
+                             d_gp.dZ, SoA0, SoA1, SoA2);
+        rhs[tid] += val;
+    }
+
+    if (do_kodis) {
+        rhs[tid] += fd_ko_term(ko_src, iF, jF, kF,
+                               iminF, jminF, kminF, imaxF, jmaxF, kmaxF,
+                               eps_val, SoA0, SoA1, SoA2);
+    }
+#else
     if (do_lopsided && i0 <= nx - 2 && j0 <= ny - 2 && k0 <= nz - 2) {
         double val = 0.0;
 
@@ -2323,6 +2399,25 @@ void kern_lopsided_kodis_batched(const double * __restrict__ Sfx,
 
         rhs[tid] += (eps_val / cof) * (Dx / d_gp.dX + Dy / d_gp.dY + Dz / d_gp.dZ);
     }
+#endif
+}
+
+static void gpu_lopsided_kodis_single_batch(double *d_f_adv, double *d_f_ko, double *d_f_rhs,
+                                            double *d_Sfx, double *d_Sfy, double *d_Sfz,
+                                            double SoA0, double SoA1, double SoA2,
+                                            double eps_val, int all)
+{
+    LopsidedKodisTables tables = {};
+    tables.adv_fields[0] = d_f_adv;
+    tables.ko_fields[0] = d_f_ko;
+    tables.rhs_fields[0] = d_f_rhs;
+    tables.soa_signs[0] = (int)SoA0;
+    tables.soa_signs[1] = (int)SoA1;
+    tables.soa_signs[2] = (int)SoA2;
+
+    dim3 launch_grid((unsigned int)grid((size_t)all), 1u);
+    kern_lopsided_kodis_batched<<<launch_grid, BLK>>>(
+        d_Sfx, d_Sfy, d_Sfz, tables, eps_val, 1, eps_val > 0.0 ? 1 : 0);
 }
 
 static void gpu_lopsided_kodis_state_batch(double eps_val, int all)
@@ -3873,6 +3968,12 @@ void kern_phase12_13_chi_correction_fused(
         const int jF = j0 + 1;
         const int kF = k0 + 1;
 
+#if ghost_width != 3
+        fd_compute_second6(chi, iF, jF, kF,
+                           iminF, jminF, kminF, imaxF, jmaxF, kmaxF,
+                           1, 1, 1,
+                           cxx, cxy, cxz, cyy, cyz, czz);
+#else
         if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
             (jF + 2) <= jmaxF && (jF - 2) >= jminF &&
             (kF + 2) <= kmaxF && (kF - 2) >= kminF)
@@ -3996,6 +4097,7 @@ void kern_phase12_13_chi_correction_fused(
               - fetch_sym_ord2_direct(chi, iF, jF - 1, kF + 1, 1, 1, 1)
               + fetch_sym_ord2_direct(chi, iF, jF + 1, kF + 1, 1, 1, 1));
         }
+#endif
     }
 
     const double cx = chix[tid];
@@ -4166,6 +4268,12 @@ void kern_phase15_trK_Aij_gauge(
         double fyy_v = 0.0, fyz_v = 0.0, fzz_v = 0.0;
 
         if (!(i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2)) {
+#if ghost_width != 3
+            fd_compute_second6(alpn1, iF, jF, kF,
+                               iminF, jminF, kminF, imaxF, jmaxF, kmaxF,
+                               1, 1, 1,
+                               fxx_v, fxy_v, fxz_v, fyy_v, fyz_v, fzz_v);
+#else
             if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
                 (jF + 2) <= jmaxF && (jF - 2) >= jminF &&
                 (kF + 2) <= kmaxF && (kF - 2) >= kminF)
@@ -4289,6 +4397,7 @@ void kern_phase15_trK_Aij_gauge(
                   - fetch_sym_ord2_direct(alpn1, iF, jF - 1, kF + 1, 1, 1, 1)
                   + fetch_sym_ord2_direct(alpn1, iF, jF + 1, kF + 1, 1, 1, 1));
             }
+#endif
         }
 
         /* raised chi/chi */
@@ -4626,15 +4735,15 @@ static void setup_grid_params(int *ex,
     gp.imaxF = nx;
     gp.jmaxF = ny;
     gp.kmaxF = nz;
-    if (Symmetry > NO_SYMM && fabs(Z[0]) < dZ) gp.kminF = -1;
-    if (Symmetry > EQ_SYMM && fabs(X[0]) < dX) gp.iminF = -1;
-    if (Symmetry > EQ_SYMM && fabs(Y[0]) < dY) gp.jminF = -1;
+    if (Symmetry > NO_SYMM && fabs(Z[0]) < dZ) gp.kminF = 2 - ghost_width;
+    if (Symmetry > EQ_SYMM && fabs(X[0]) < dX) gp.iminF = 2 - ghost_width;
+    if (Symmetry > EQ_SYMM && fabs(Y[0]) < dY) gp.jminF = 2 - ghost_width;
     gp.iminF3 = 1;
     gp.jminF3 = 1;
     gp.kminF3 = 1;
-    if (Symmetry > NO_SYMM && fabs(Z[0]) < dZ) gp.kminF3 = -2;
-    if (Symmetry > EQ_SYMM && fabs(X[0]) < dX) gp.iminF3 = -2;
-    if (Symmetry > EQ_SYMM && fabs(Y[0]) < dY) gp.jminF3 = -2;
+    if (Symmetry > NO_SYMM && fabs(Z[0]) < dZ) gp.kminF3 = 1 - ghost_width;
+    if (Symmetry > EQ_SYMM && fabs(X[0]) < dX) gp.iminF3 = 1 - ghost_width;
+    if (Symmetry > EQ_SYMM && fabs(Y[0]) < dY) gp.jminF3 = 1 - ghost_width;
     gp.Symmetry = Symmetry;
     gp.eps = eps;
     gp.co = co;
@@ -5089,6 +5198,196 @@ __global__ void kern_unpack_state_region_batch(double * __restrict__ dst_mem,
     }
 }
 
+__device__ __forceinline__ double load_comm_state_cell_sym(const double * __restrict__ src_mem,
+                                                           int state_index,
+                                                           int x, int y, int z,
+                                                           int nx, int ny,
+                                                           int all)
+{
+    double s = 1.0;
+    if (x < 0) {
+        x = -x - 1;
+        s *= d_comm_state_soa[3 * state_index + 0];
+    }
+    if (y < 0) {
+        y = -y - 1;
+        s *= d_comm_state_soa[3 * state_index + 1];
+    }
+    if (z < 0) {
+        z = -z - 1;
+        s *= d_comm_state_soa[3 * state_index + 2];
+    }
+    const int src = x + y * nx + z * nx * ny;
+    return s * src_mem[(size_t)state_index * all + src];
+}
+
+__global__ void kern_restrict_state_region_batch(const double * __restrict__ src_mem,
+                                                 double * __restrict__ dst,
+                                                 int nx, int ny,
+                                                 int sx, int sy, int sz,
+                                                 int fi0, int fj0, int fk0,
+                                                 int region_all,
+                                                 int state_count,
+                                                 int all)
+{
+    const int state_index = blockIdx.y;
+    if (state_index >= state_count) return;
+#if ghost_width == 5
+    const double c1 = 35.0 / 65536.0;
+    const double c2 = -405.0 / 65536.0;
+    const double c3 = 567.0 / 16384.0;
+    const double c4 = -2205.0 / 16384.0;
+    const double c5 = 19845.0 / 32768.0;
+    const int offs[10] = {-4, -3, -2, -1, 0, 1, 2, 3, 4, 5};
+    const double w[10] = {c1, c2, c3, c4, c5, c5, c4, c3, c2, c1};
+    const int nst = 10;
+#elif ghost_width == 4
+    const double c1 = -5.0 / 2048.0;
+    const double c2 = 49.0 / 2048.0;
+    const double c3 = -245.0 / 2048.0;
+    const double c4 = 1225.0 / 2048.0;
+    const int offs[8] = {-3, -2, -1, 0, 1, 2, 3, 4};
+    const double w[8] = {c1, c2, c3, c4, c4, c3, c2, c1};
+    const int nst = 8;
+#elif ghost_width == 3
+    const double c1 = 3.0 / 256.0;
+    const double c2 = -25.0 / 256.0;
+    const double c3 = 75.0 / 128.0;
+    const int offs[6] = {-2, -1, 0, 1, 2, 3};
+    const double w[6] = {c1, c2, c3, c3, c2, c1};
+    const int nst = 6;
+#else
+    const double c1 = -1.0 / 16.0;
+    const double c2 = 9.0 / 16.0;
+    const int offs[4] = {-1, 0, 1, 2};
+    const double w[4] = {c1, c2, c2, c1};
+    const int nst = 4;
+#endif
+
+    for (int local = blockIdx.x * blockDim.x + threadIdx.x;
+         local < region_all;
+         local += blockDim.x * gridDim.x)
+    {
+        const int ii = local % sx;
+        const int jj = (local / sx) % sy;
+        const int kk = local / (sx * sy);
+        const int fc_i = fi0 + 2 * ii;
+        const int fc_j = fj0 + 2 * jj;
+        const int fc_k = fk0 + 2 * kk;
+        double sum = 0.0;
+        for (int oz = 0; oz < nst; ++oz) {
+            const int z = fc_k + offs[oz];
+            const double wz = w[oz];
+            for (int oy = 0; oy < nst; ++oy) {
+                const int y = fc_j + offs[oy];
+                const double wyz = wz * w[oy];
+                for (int ox = 0; ox < nst; ++ox) {
+                    const int x = fc_i + offs[ox];
+                    sum += wyz * w[ox] *
+                           load_comm_state_cell_sym(src_mem, state_index, x, y, z, nx, ny, all);
+                }
+            }
+        }
+        dst[(size_t)state_index * region_all + local] = sum;
+    }
+}
+
+__global__ void kern_prolong_state_region_batch(const double * __restrict__ src_mem,
+                                                double * __restrict__ dst,
+                                                int nx, int ny,
+                                                int sx, int sy, int sz,
+                                                int ii0, int jj0, int kk0,
+                                                int lbc_i, int lbc_j, int lbc_k,
+                                                int region_all,
+                                                int state_count,
+                                                int all)
+{
+    const int state_index = blockIdx.y;
+    if (state_index >= state_count) return;
+#if ghost_width == 5
+    const double c1 = 13585.0 / 33554432.0;
+    const double c2 = -159885.0 / 33554432.0;
+    const double c3 = 230945.0 / 8388608.0;
+    const double c4 = -969969.0 / 8388608.0;
+    const double c5 = 14549535.0 / 16777216.0;
+    const double c6 = 4849845.0 / 16777216.0;
+    const double c7 = -692835.0 / 8388608.0;
+    const double c8 = 188955.0 / 8388608.0;
+    const double c9 = -138567.0 / 33554432.0;
+    const double c10 = 12155.0 / 33554432.0;
+    const int offs[10] = {-4, -3, -2, -1, 0, 1, 2, 3, 4, 5};
+    const double wl[10] = {c1, c2, c3, c4, c5, c6, c7, c8, c9, c10};
+    const double wr[10] = {c10, c9, c8, c7, c6, c5, c4, c3, c2, c1};
+    const int nst = 10;
+#elif ghost_width == 4
+    const double c1 = -495.0 / 262144.0;
+    const double c2 = 5005.0 / 262144.0;
+    const double c3 = -27027.0 / 262144.0;
+    const double c4 = 225225.0 / 262144.0;
+    const double c5 = 75075.0 / 262144.0;
+    const double c6 = -19305.0 / 262144.0;
+    const double c7 = 4095.0 / 262144.0;
+    const double c8 = -429.0 / 262144.0;
+    const int offs[8] = {-3, -2, -1, 0, 1, 2, 3, 4};
+    const double wl[8] = {c1, c2, c3, c4, c5, c6, c7, c8};
+    const double wr[8] = {c8, c7, c6, c5, c4, c3, c2, c1};
+    const int nst = 8;
+#elif ghost_width == 3
+    const double c1 = 77.0 / 8192.0;
+    const double c2 = -693.0 / 8192.0;
+    const double c3 = 3465.0 / 4096.0;
+    const double c4 = 1155.0 / 4096.0;
+    const double c5 = -495.0 / 8192.0;
+    const double c6 = 63.0 / 8192.0;
+    const int offs[6] = {-2, -1, 0, 1, 2, 3};
+    const double wl[6] = {c1, c2, c3, c4, c5, c6};
+    const double wr[6] = {c6, c5, c4, c3, c2, c1};
+    const int nst = 6;
+#else
+    const double c1 = -7.0 / 128.0;
+    const double c2 = 105.0 / 128.0;
+    const double c3 = 35.0 / 128.0;
+    const double c4 = -5.0 / 128.0;
+    const int offs[4] = {-1, 0, 1, 2};
+    const double wl[4] = {c1, c2, c3, c4};
+    const double wr[4] = {c4, c3, c2, c1};
+    const int nst = 4;
+#endif
+
+    for (int local = blockIdx.x * blockDim.x + threadIdx.x;
+         local < region_all;
+         local += blockDim.x * gridDim.x)
+    {
+        const int ii = local % sx;
+        const int jj = (local / sx) % sy;
+        const int kk = local / (sx * sy);
+        const int fine_i = ii0 + ii;
+        const int fine_j = jj0 + jj;
+        const int fine_k = kk0 + kk;
+        const int ci = fine_i / 2 - lbc_i;
+        const int cj = fine_j / 2 - lbc_j;
+        const int ck = fine_k / 2 - lbc_k;
+        const double *wx = ((fine_i / 2) * 2 == fine_i) ? wl : wr;
+        const double *wy = ((fine_j / 2) * 2 == fine_j) ? wl : wr;
+        const double *wz = ((fine_k / 2) * 2 == fine_k) ? wl : wr;
+        double sum = 0.0;
+        for (int oz = 0; oz < nst; ++oz) {
+            const int z = ck + offs[oz];
+            const double wzv = wz[oz];
+            for (int oy = 0; oy < nst; ++oy) {
+                const int y = cj + offs[oy];
+                const double wyz = wzv * wy[oy];
+                for (int ox = 0; ox < nst; ++ox) {
+                    const int x = ci + offs[ox];
+                    sum += wyz * wx[ox] *
+                           load_comm_state_cell_sym(src_mem, state_index, x, y, z, nx, ny, all);
+                }
+            }
+        }
+        dst[(size_t)state_index * region_all + local] = sum;
+    }
+}
+
 __global__ void kern_pack_state_subset(const double * __restrict__ src_mem,
                                        double * __restrict__ dst,
                                        int subset_count,
@@ -5224,6 +5523,36 @@ static void copy_state_region_packed_batch_cuda(void *block_tag,
     }
 }
 
+static void copy_state_region_packed_batch_device_cuda(void *block_tag,
+                                                       int state_count,
+                                                       double *device_buffer,
+                                                       const int *ex,
+                                                       int i0, int j0, int k0,
+                                                       int sx, int sy, int sz,
+                                                       int pack_not_unpack)
+{
+    if (state_count <= 0 || state_count > BSSN_STATE_COUNT) return;
+    if (!device_buffer || sx <= 0 || sy <= 0 || sz <= 0) return;
+
+    StepContext &ctx = ensure_step_ctx(block_tag, (size_t)ex[0] * ex[1] * ex[2]);
+    const int region_all = sx * sy * sz;
+    dim3 launch_grid((unsigned int)grid((size_t)region_all),
+                     (unsigned int)state_count);
+
+    if (pack_not_unpack) {
+        kern_pack_state_region_batch<<<launch_grid, BLK>>>(
+            ctx.d_state_curr_mem, device_buffer, ex[0], ex[1],
+            i0, j0, k0, sx, sy, sz, region_all, state_count,
+            ex[0] * ex[1] * ex[2]);
+    } else {
+        kern_unpack_state_region_batch<<<launch_grid, BLK>>>(
+            ctx.d_state_curr_mem, device_buffer, ex[0], ex[1],
+            i0, j0, k0, sx, sy, sz, region_all, state_count,
+            ex[0] * ex[1] * ex[2]);
+        ctx.state_ready = true;
+    }
+}
+
 static void download_resident_state(void *block_tag, int *ex, double **state_host_out)
 {
     const size_t all = (size_t)ex[0] * ex[1] * ex[2];
@@ -7451,6 +7780,90 @@ extern "C" int z4c_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
     return 0;
 }
 
+extern "C" int z4c_cuda_pack_state_batch_to_device_buffer(void *block_tag,
+                                                          int state_count,
+                                                          double *device_buffer,
+                                                          int *ex,
+                                                          int i0, int j0, int k0,
+                                                          int sx, int sy, int sz)
+{
+    using namespace z4c_cuda;
+    init_gpu_dispatch();
+    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
+    copy_state_region_packed_batch_device_cuda(block_tag, state_count, device_buffer, ex,
+                                               i0, j0, k0, sx, sy, sz, 1);
+    return 0;
+}
+
+extern "C" int z4c_cuda_unpack_state_batch_from_device_buffer(void *block_tag,
+                                                              int state_count,
+                                                              double *device_buffer,
+                                                              int *ex,
+                                                              int i0, int j0, int k0,
+                                                              int sx, int sy, int sz)
+{
+    using namespace z4c_cuda;
+    init_gpu_dispatch();
+    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
+    copy_state_region_packed_batch_device_cuda(block_tag, state_count, device_buffer, ex,
+                                               i0, j0, k0, sx, sy, sz, 0);
+    return 0;
+}
+
+extern "C" int z4c_cuda_restrict_state_batch_to_device_buffer(void *block_tag,
+                                                              int state_count,
+                                                              double *device_buffer,
+                                                              int *ex,
+                                                              int sx, int sy, int sz,
+                                                              int fi0, int fj0, int fk0,
+                                                              const double *state_soa)
+{
+    using namespace z4c_cuda;
+    init_gpu_dispatch();
+    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
+    if (state_count <= 0 || state_count > BSSN_STATE_COUNT) return 1;
+    if (!device_buffer || sx <= 0 || sy <= 0 || sz <= 0) return 1;
+    StepContext &ctx = ensure_step_ctx(block_tag, (size_t)ex[0] * ex[1] * ex[2]);
+    const int region_all = sx * sy * sz;
+    upload_comm_state_soa(state_soa, state_count);
+    dim3 launch_grid((unsigned int)grid((size_t)region_all),
+                     (unsigned int)state_count);
+    kern_restrict_state_region_batch<<<launch_grid, BLK>>>(
+        ctx.d_state_curr_mem, device_buffer,
+        ex[0], ex[1], sx, sy, sz,
+        fi0, fj0, fk0, region_all, state_count,
+        ex[0] * ex[1] * ex[2]);
+    return 0;
+}
+
+extern "C" int z4c_cuda_prolong_state_batch_to_device_buffer(void *block_tag,
+                                                             int state_count,
+                                                             double *device_buffer,
+                                                             int *ex,
+                                                             int sx, int sy, int sz,
+                                                             int ii0, int jj0, int kk0,
+                                                             int lbc_i, int lbc_j, int lbc_k,
+                                                             const double *state_soa)
+{
+    using namespace z4c_cuda;
+    init_gpu_dispatch();
+    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
+    if (state_count <= 0 || state_count > BSSN_STATE_COUNT) return 1;
+    if (!device_buffer || sx <= 0 || sy <= 0 || sz <= 0) return 1;
+    StepContext &ctx = ensure_step_ctx(block_tag, (size_t)ex[0] * ex[1] * ex[2]);
+    const int region_all = sx * sy * sz;
+    upload_comm_state_soa(state_soa, state_count);
+    dim3 launch_grid((unsigned int)grid((size_t)region_all),
+                     (unsigned int)state_count);
+    kern_prolong_state_region_batch<<<launch_grid, BLK>>>(
+        ctx.d_state_curr_mem, device_buffer,
+        ex[0], ex[1], sx, sy, sz,
+        ii0, jj0, kk0, lbc_i, lbc_j, lbc_k,
+        region_all, state_count,
+        ex[0] * ex[1] * ex[2]);
+    return 0;
+}
+
 extern "C" int z4c_cuda_download_state_subset(void *block_tag,
                                               int *ex,
                                               int subset_count,

AMSS_NCKU_source/z4c_rhs_cuda.h

@@ -60,6 +60,37 @@ int z4c_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
                                                  int i0, int j0, int k0,
                                                  int sx, int sy, int sz);
 
+int z4c_cuda_pack_state_batch_to_device_buffer(void *block_tag,
+                                               int state_count,
+                                               double *device_buffer,
+                                               int *ex,
+                                               int i0, int j0, int k0,
+                                               int sx, int sy, int sz);
+
+int z4c_cuda_unpack_state_batch_from_device_buffer(void *block_tag,
+                                                   int state_count,
+                                                   double *device_buffer,
+                                                   int *ex,
+                                                   int i0, int j0, int k0,
+                                                   int sx, int sy, int sz);
+
+int z4c_cuda_restrict_state_batch_to_device_buffer(void *block_tag,
+                                                   int state_count,
+                                                   double *device_buffer,
+                                                   int *ex,
+                                                   int sx, int sy, int sz,
+                                                   int fi0, int fj0, int fk0,
+                                                   const double *state_soa);
+
+int z4c_cuda_prolong_state_batch_to_device_buffer(void *block_tag,
+                                                  int state_count,
+                                                  double *device_buffer,
+                                                  int *ex,
+                                                  int sx, int sy, int sz,
+                                                  int ii0, int jj0, int kk0,
+                                                  int lbc_i, int lbc_j, int lbc_k,
+                                                  const double *state_soa);
+
 int z4c_cuda_download_state_subset(void *block_tag,
                                    int *ex,
                                    int subset_count,

code_modification_readme.md

@@ -0,0 +1,224 @@
+# Code Modification Readme — `asc26-plan-a`
+
+**Baseline branch:** `baseline`
+**Target branch:** `asc26-plan-a`
+**Date:** 2026-05-19
+
+---
+
+## Overview
+
+This branch delivers two major performance overhauls to the AMSS-NCKU numerical relativity codebase:
+
+1. **TwoPunctureABE Multithreading** — OpenMP parallelization of the TwoPunctures initial-data solver, combined with a BLAS-driven spectral derivative engine, MKL/LAPACK integration, and C/C++ rewrites of hot Fortran kernel subroutines.
+
+2. **ABE GPU Rewrite** — Complete replacement of the legacy `bssn_gpu_class` abstraction layer with direct, monolithic CUDA kernels for BSSN, Z4C, and Shell-Patch evolution, plus GPU-resident state management and CUDA-aware MPI.
+
+**Total diff:** 84 files changed, +57,919 / −33,795 lines.
+
+---
+
+## Part 1 — TwoPunctureABE Multithreading
+
+### 1.1 Spectral Derivative Engine: BLAS Matrix-Multiplication Rewrite
+
+**Files:** `AMSS_NCKU_source/TwoPunctures.C`, `AMSS_NCKU_source/TwoPunctures.h`
+
+The original `Derivatives_AB3` computed spectral derivatives (Chebyshev in A/B, Fourier in phi) with nested scalar loops over every grid point. The new `Derivatives_AB3_MatMul` expresses all derivatives as matrix-matrix products over pencil-shaped data slices, dispatched to Intel MKL `cblas_dgemm`.
+
+- **Precomputed derivative matrices** — `precompute_derivative_matrices()` builds `D1_A`, `D2_A`, `D1_B`, `D2_B` (Chebyshev collocation derivative matrices) and `DF1_phi`, `DF2_phi` (Fourier derivative matrices) once at construction time.
+- **Pencil-based GEMM** — data is gathered into 2D arrays where one dimension is the spectral direction and the other enumerates all remaining degrees of freedom (variables × orthogonal grid indices). Each derivative direction becomes a single `cblas_dgemm` call. The pure derivatives (d/dA, d/dB, d/dphi) and all mixed derivatives (d²/dAdB, d²/dAdphi, d²/dBdphi) are computed this way.
+- **`build_cheb_deriv_matrices` / `build_fourier_deriv_matrices`** — construct the standard Chebyshev and Fourier collocation derivative matrices.
+
+### 1.2 OpenMP Parallelization of TwoPunctures
+
+**Files:** `AMSS_NCKU_source/TwoPunctures.C`, `AMSS_NCKU_source/TwoPunctures.h`
+
+Three critical regions are parallelized:
+
+| Region | Directive | Strategy |
+|--------|-----------|----------|
+| `F_of_v` residual evaluation | `#pragma omp parallel for collapse(3) schedule(dynamic,1)` | Each (i,j,k) thread stack-allocates its own `l_U` (derivs struct) and `l_values[]` to eliminate heap contention and data races |
+| `relax_omp` line relaxation | `#pragma omp parallel for schedule(static)` over k-slices | Alternating be/al sweeps, each thread uses pre-allocated per-thread Thomas-algorithm workspace (`ws_*_be[tid]`, `ws_*_al[tid]`) |
+| `LineRelax_be_omp` / `LineRelax_al_omp` | Called from `relax_omp` with explicit `tid` | Thread-safe tridiagonal solves using the thread's private scratch arrays |
+
+**Per-thread workspace** — `allocate_workspace()` allocates independent Thomas-algorithm buffers (`diag`, `e`, `f`, `b`, `x`, `l`, `u`, `d`, `y`) for each OpenMP thread in both be and al directions, avoiding lock contention in the inner Newton iteration.
+
+### 1.3 MKL BLAS / LAPACK Integration
+
+**Files:** `AMSS_NCKU_source/TwoPunctures.C`, `AMSS_NCKU_source/gaussj.C`
+
+| Function | Old | New | Benefit |
+|----------|-----|-----|---------|
+| `norm2` | scalar `sqrt(sum(v[i]²))` loop | `cblas_dnrm2` | BLAS Level 1, SIMD-optimized |
+| `scalarproduct` | scalar `sum(v[i]*w[i])` loop | `cblas_ddot` | BLAS Level 1, SIMD-optimized |
+| `gaussj` | hand-written Gauss-Jordan elimination (~100 lines) | `LAPACKE_dgesv` + `LAPACKE_dgetrf` + `LAPACKE_dgetri` | LAPACK LU with partial pivoting, asymptotically faster for the `n~50` matrix sizes used in spectral elliptic solves |
+
+### 1.4 C/C++ Rewrite of Hot Fortran Kernels
+
+**Files (new):**
+- `AMSS_NCKU_source/fderivs_c.C` (167 lines) — first derivatives, 2nd/4th order
+- `AMSS_NCKU_source/fdderivs_c.C` (332 lines) — second derivatives, 2nd/4th order
+- `AMSS_NCKU_source/kodiss_c.C` (117 lines) — Kreiss-Oliger dissipation
+- `AMSS_NCKU_source/lopsided_c.C` (255 lines) — lopsided advection
+- `AMSS_NCKU_source/lopsided_kodis_c.C` (248 lines) — fused advection + dissipation
+- `AMSS_NCKU_source/rungekutta4_rout_c.C` (212 lines) — RK4 time-stepper
+- `AMSS_NCKU_source/bssn_rhs_c.C` (1,287 lines) — full BSSN RHS kernel
+- `AMSS_NCKU_source/z4c_rhs_c.C` (725 lines) — full Z4C RHS kernel
+
+Every C rewrite follows a consistent optimization pattern:
+- **64-byte aligned allocation** (`aligned_alloc(64, ...)`) for AVX-512 compatibility.
+- **Static buffer caching** — scratch arrays (e.g., the padded `fh` ghost-zone buffer) persist across calls via a `static` pointer + capacity check, avoiding repeated `malloc`/`free`.
+- **Two-pass strategy** — 2nd-order finite differences are computed on the full domain first, then the interior sub-volume is overwritten with 4th-order stencils. This eliminates the per-point `if/elseif` branching of the original Fortran.
+- **Non-overlapping shell pass** — in `fdderivs_c.C`, the 2nd-order pass skips points that will be overwritten by the 4th-order pass, avoiding redundant computation.
+
+### 1.5 Fortran Kernel Fusion: lopsided_kodis
+
+**File:** `AMSS_NCKU_source/lopsidediff.f90`
+
+A new `lopsided_kodis` subroutine fuses the advection (lopsided) and Kreiss-Oliger dissipation (kodis) operators into a single pass over the grid. Both operators previously called `symmetry_bd` independently to fill ghost zones — the fused version calls it once and shares the padded `fh` array, halving ghost-zone fill overhead for this hot path.
+
+### 1.6 Build System for TwoPunctures
+
+**Files:** `AMSS_NCKU_source/makefile`, `AMSS_NCKU_source/makefile.inc`
+
+- **`TP_OPTFLAGS`** — TwoPunctures and TwoPunctureABE are compiled with a dedicated, more aggressive optimization flag set (`-O3 -march=znver5 -fp-model fast=2 -fma -ipo`) separate from the main code.
+- **`USE_CXX_KERNELS`** — selects between the C rewrites and the original Fortran kernels (`bssn_rhs.f90`, etc.) for the CPU path.
+- **`USE_CXX_RK4`** — independently selects between the C and Fortran RK4 stepper.
+- **Intel oneTBB allocator** (`libtbbmalloc.so`) — replaces the system `malloc` with a scalable thread-safe allocator, critical for multi-threaded TwoPunctures performance.
+- **PGO support** — `PGO_MODE=opt|instrument` for profile-guided optimization (currently disabled after testing showed negative benefit).
+- **Toolchains** — Intel oneAPI (`TOOLCHAIN=intel`, default) and NVIDIA HPC SDK (`TOOLCHAIN=nvhpc`).
+
+---
+
+## Part 2 — ABE GPU Rewrite
+
+### 2.1 Architecture: From Class Wrapper to Direct CUDA Kernels
+
+The old GPU path (`baseline`) was organized as:
+
+```
+bssn_gpu_class.C/h    — C++ class managing GPU state and kernel launches
+bssn_step_gpu.C       — RK4 stepper with per-substep GPU/CPU synchronisation
+bssn_gpu.cu           — CUDA kernel implementations called through the class
+```
+
+The new GPU path (`asc26-plan-a`) replaces all of the above with:
+
+```
+bssn_rhs_cuda.cu/h    — 10,381-line monolithic CUDA BSSN RHS kernel
+z4c_rhs_cuda.cu/h     —  7,909-line monolithic CUDA Z4C RHS kernel
+fd_cuda_helpers.cuh   —    412-line shared finite-difference device functions
+bssn_gpu_rhs_ss.cu    —  (retained, lightly modified) Shell-Patch GPU RHS
+```
+
+**Key architectural differences:**
+- The old `bssn_gpu_class` managed GPU memory through a C++ class with explicit allocate/free/sync methods scattered across the time-stepping logic. The new kernels operate directly on raw device pointers with a clear resident/transient memory model.
+- The old code launched many small kernels (one per derivative or algebraic term). The new code is a **single monolithic kernel per formulation** — all 24 BSSN evolution variables are computed in one launch with on-the-fly finite differences, eliminating kernel-launch latency and intermediate global-memory round-trips.
+- The old `bssn_step_gpu.C` performed per-substep GPU→CPU downloads for boundary conditions and analysis. The new model supports **GPU-resident state** — variables stay on device across timesteps unless explicitly requested.
+
+### 2.2 GPU-Resident State Model
+
+A central theme across ~20 commits is the "resident-sync" optimization:
+
+| Commit | What it does |
+|--------|-------------|
+| `22c1e71` | Optimize BSSN CUDA resident state and CUDA-aware MPI |
+| `090d865` | Optimize BSSN CUDA state transfers |
+| `68eab03` | Add opt-in BSSN CUDA resident AMR path |
+| `1ee229a` | Add keyed BSSN CUDA resident banks |
+| `18e9c9c` | Optimize BSSN CUDA resident AMR prolong |
+| `8486532` | Add resident BSSN GPU point interpolation |
+| `b1974ef` | Stabilize device AMR restrict across regrid |
+| `ae64a22` | Complete BSSN-EScalar CUDA resident transfers |
+| `83afaf1` | Skip zero EM resident downloads |
+| `35b6cef` | Broaden cached CUDA sync paths |
+
+The resident model works as follows:
+- BSSN grid functions are allocated once on the GPU and persist across timesteps.
+- Inter-processor ghost-zone exchanges use **CUDA-aware MPI** — MPI directly reads/writes device memory without staging through host buffers.
+- AMR prolongation and restriction operate directly on device memory.
+- Boundary conditions and analysis routines download only the specific slices/points they need, not the full grid.
+- When EM fields are zero (pure-gravity runs), EM downloads are skipped entirely.
+
+### 2.3 Z4C and Shell-Patch GPU Acceleration
+
+**Files:** `AMSS_NCKU_source/z4c_rhs_cuda.cu`, `AMSS_NCKU_source/bssn_gpu_rhs_ss.cu`
+
+- The Z4C constraint-damped formulation gets its own 7,909-line monolithic CUDA kernel (`z4c_rhs_cuda.cu`), matching the BSSN kernel's architecture.
+- **Shell-Patch GPU acceleration** — the spherical shell boundary patches now compute on GPU with dedicated kernels in `bssn_gpu_rhs_ss.cu`.
+- Z4C + Shell-Patch can coexist on GPU (Phase 3 commits).
+- A CPU-side wrapper (`z4c_rhs_c.C`) handles the trKd + TZ_rhs contribution that remains on CPU, minimizing GPU/CPU traffic.
+
+### 2.4 Finite-Difference Order Flexibility
+
+**File:** `AMSS_NCKU_source/fd_cuda_helpers.cuh`
+
+Shared device functions for finite-difference stencils support **2nd, 4th, 6th, and 8th order** at compile time via preprocessor switches. This enables:
+- Per-run selection of convergence order without recompilation of the full kernel.
+- 8th-order AMR transfers (`1064a68`) for BSSN-EM.
+- 6th-order optimized AMR stencils (`0076b3c`).
+
+### 2.5 GPU Diagnostics and Quality Assurance
+
+**File:** `AMSS_NCKU_GPUCheck.py` (559 lines, new)
+
+A Python-based GPU correctness verification tool that compares GPU and CPU evolution outputs. The GPU build pipeline includes optional kernel profiling switches (`7683459`) for performance debugging.
+
+**GPU-specific bug fixes:**
+- `f226498` — Fix CUDA AMR symmetry drift (incorrect ghost-zone handling under symmetry boundary conditions)
+- `2317e4a` — Fix BSSN GPU resident AMR sync default
+- `fea2dcc` — Fix BSSN-EM runtime crash
+- `dd0e20d` — Fix BSSN-EScalar CUDA boundary and scalar KO
+- `5eb4994` — Fix AHF crash under CUDA resident-sync mode
+
+### 2.6 Build Integration
+
+**Makefile switches:**
+- `USE_CUDA_BSSN=0/1` — route BSSN RHS through GPU or CPU
+- `USE_CUDA_Z4C=0/1` — route Z4C RHS through GPU or CPU
+- `CUDA_ARCH=sm_80` — target NVIDIA Ampere (A100)
+- `NVHPC_ROOT` — path to NVIDIA HPC SDK for the `nvcc` compiler wrapper
+- CUDA compilation flags: `-O3 --ptxas-options=-v -arch=$(CUDA_ARCH)`
+
+---
+
+## Part 3 — Shared Infrastructure
+
+### 3.1 Interp_Points Load-Balance Profiler
+
+**Files:** `AMSS_NCKU_source/interp_lb_profile.C`, `interp_lb_profile.h`, `interp_lb_profile_data.h`, `generate_interp_lb_header.py`
+
+A two-pass instrumentation system for load-balancing the `Interp_Points` parallel interpolation routine:
+- **Pass 1** (`INTERP_LB_MODE=profile`): instrument each MPI rank's interpolation calls with timing, write a binary profile.
+- **Pass 2** (`INTERP_LB_MODE=optimize`): read the profile and rebalance work across MPI ranks.
+
+### 3.2 Helper Headers
+
+**Files:** `AMSS_NCKU_source/tool.h` (33 lines), `AMSS_NCKU_source/share_func.h` (246 lines)
+
+- `tool.h` — shared indexing macros (`idx_ex`, `idx_fh_F_ord2`) and the `symmetry_bd` declaration used by all C kernel rewrites.
+- `share_func.h` — common utility functions shared across the C++ source files.
+
+### 3.3 Plot-Only Restart Script
+
+**File:** `parallel_plot_helper.py` (29 lines)
+
+A lightweight restart script that skips recomputation when plotting was interrupted — reads existing checkpoint data and replots without re-running the simulation.
+
+---
+
+## Performance Summary
+
+| Component | Optimization | Expected Impact |
+|-----------|-------------|-----------------|
+| TwoPunctures `Derivatives_AB3` | Scalar loops → MKL GEMM | 5-20× speedup for spectral derivative computation |
+| TwoPunctures `F_of_v` | OpenMP collapse(3) + stack-local variables | Near-linear scaling with core count for residual evaluation |
+| TwoPunctures `gaussj` | Hand-written Gauss-Jordan → LAPACK LU | 2-5× speedup for N~50 matrix inversion |
+| BSSN RHS (GPU) | Many small kernels → one monolithic kernel | Eliminates kernel-launch overhead; 2-5× GPU throughput improvement |
+| GPU state transfers | Per-step download → resident model | Eliminates ~80% of GPU↔CPU PCIe traffic |
+| `lopsided_kodis` fusion | Two `symmetry_bd` calls → one shared call | ~30% reduction in ghost-zone fill cost for this operator pair |
+| Memory allocator | System malloc → Intel TBB malloc | Significant reduction in malloc contention under OpenMP |
+| C kernel rewrites | Fortran → C with aligned alloc + static buffers | Enables Intel compiler IPO across C/C++/Fortran boundaries; better SIMD codegen |
+
+---

makefile_and_run.py

@@ -9,6 +9,8 @@
 
 
 import AMSS_NCKU_Input as input_data
+import os
+import shutil
 import subprocess
 import time
 
@@ -43,8 +45,7 @@ def get_last_n_cores_per_socket(n=32):
     cpu_str = ",".join(segments)
     total = len(segments) * n
     print(f" CPU binding: taskset -c {cpu_str}  ({total} cores, last {n} per socket)")
-    #return f"taskset -c {cpu_str}"
-    return f""
+    return f"taskset -c {cpu_str}" if cpu_str else ""
 
 
 ## CPU core binding: dynamically select the last 32 cores of each socket (64 cores total)
@@ -56,6 +57,231 @@ BUILD_JOBS = 64
 
 ##################################################################
 
+def _truthy(value, default=False):
+    if value is None:
+        return default
+    if isinstance(value, bool):
+        return value
+    text = str(value).strip().lower()
+    if text == "":
+        return default
+    return text in ("1", "yes", "y", "true", "on", "enable", "enabled")
+
+
+def _input_or_env(input_name, env_name, default=None):
+    if env_name in os.environ:
+        return os.environ[env_name]
+    return getattr(input_data, input_name, default)
+
+
+def _input_env_passthrough(runtime_env, env_name):
+    if env_name in runtime_env:
+        return
+    if hasattr(input_data, env_name):
+        runtime_env[env_name] = str(getattr(input_data, env_name))
+
+
+def _start_cuda_mps_if_requested(runtime_env):
+    if input_data.GPU_Calculation != "yes":
+        return False
+
+    default_auto_mps = int(getattr(input_data, "MPI_processes", 1)) > 1
+    auto_mps = _truthy(
+        _input_or_env("CUDA_Auto_MPS", "AMSS_CUDA_AUTO_MPS", default_auto_mps),
+        default=default_auto_mps,
+    )
+    if not auto_mps:
+        return False
+
+    mps_control = shutil.which("nvidia-cuda-mps-control")
+    if not mps_control:
+        print(" CUDA MPS control command was not found; running without MPS.")
+        return False
+
+    uid = os.getuid()
+    pipe_dir = str(_input_or_env("CUDA_MPS_PIPE_DIRECTORY", "CUDA_MPS_PIPE_DIRECTORY",
+                                 f"/tmp/amss-ncku-mps-{uid}"))
+    log_dir = str(_input_or_env("CUDA_MPS_LOG_DIRECTORY", "CUDA_MPS_LOG_DIRECTORY",
+                                f"/tmp/amss-ncku-mps-log-{uid}"))
+    os.makedirs(pipe_dir, exist_ok=True)
+    os.makedirs(log_dir, exist_ok=True)
+
+    mps_env = runtime_env.copy()
+    mps_env["CUDA_MPS_PIPE_DIRECTORY"] = pipe_dir
+    mps_env["CUDA_MPS_LOG_DIRECTORY"] = log_dir
+
+    if os.path.exists(os.path.join(pipe_dir, "control")):
+        runtime_env.update({
+            "CUDA_MPS_PIPE_DIRECTORY": pipe_dir,
+            "CUDA_MPS_LOG_DIRECTORY": log_dir,
+        })
+        print(f" Reusing CUDA MPS daemon: {pipe_dir}")
+        return False
+
+    print(f" Starting CUDA MPS daemon for this run: {pipe_dir}")
+    result = subprocess.run([mps_control, "-d"], env=mps_env, text=True,
+                            stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
+    if result.returncode != 0:
+        print(" CUDA MPS daemon did not start; running without MPS.")
+        if result.stdout:
+            print(result.stdout, end="")
+        return False
+
+    runtime_env.update({
+        "CUDA_MPS_PIPE_DIRECTORY": pipe_dir,
+        "CUDA_MPS_LOG_DIRECTORY": log_dir,
+    })
+    return True
+
+
+def _stop_cuda_mps(runtime_env):
+    mps_control = shutil.which("nvidia-cuda-mps-control")
+    if not mps_control:
+        return
+    subprocess.run([mps_control], input="quit\n", env=runtime_env, text=True,
+                   stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
+
+
+def _gpu_runtime_env():
+    runtime_env = os.environ.copy()
+    original_env = set(os.environ.keys())
+    finite_difference = str(getattr(input_data, "Finite_Diffenence_Method", "4th-order")).strip()
+
+    defaults = {
+        "AMSS_EVOLVE_TIMING": "0",
+        "AMSS_ESCALAR_STEP_TIMING": "0",
+        "AMSS_INTERP_FAST": "1",
+        "AMSS_INTERP_GPU": "1",
+        "AMSS_ANALYSIS_MAP_EVERY": "1000000",
+        "AMSS_CUDA_AWARE_MPI": "1",
+        "AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP": "1",
+        "AMSS_CUDA_KEEP_ALL_LEVELS": "1",
+        "AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP": "1",
+        "AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS": "1",
+        "AMSS_CUDA_EM_CACHE_SOURCES": "1",
+        "AMSS_CUDA_EM_ZERO_FASTPATH": "1",
+        "AMSS_EM_ZERO_ANALYSIS_FASTPATH": "1",
+        "AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH": "1",
+        "AMSS_CUDA_AMR_HOST_STAGED": "1",
+        "AMSS_CUDA_AMR_RESTRICT_DEVICE": "0",
+        "AMSS_CUDA_AMR_RESTRICT_BATCH": "0",
+        "AMSS_CUDA_DEVICE_SEGMENT_BATCH": "0",
+        "AMSS_CUDA_UNCACHED_DEVICE_BUFFERS": "1",
+        "AMSS_SHELL_FAST_INTERP": "0",
+        "AMSS_SHELL_PARALLEL_INTERP": "0",
+        "AMSS_SHELL_CUDA_INTERP": "0",
+    }
+    if finite_difference in ("2nd-order", "8th-order"):
+        defaults.update({
+            "AMSS_INTERP_FAST": "0",
+            "AMSS_INTERP_GPU": "0",
+            "AMSS_CUDA_AWARE_MPI": "0",
+        })
+    if finite_difference == "8th-order" and getattr(input_data, "Equation_Class", "") == "BSSN-EM":
+        defaults.update({
+            "AMSS_CUDA_AMR_RESTRICT_DEVICE": "1",
+            "AMSS_CUDA_AMR_RESTRICT_BATCH": "1",
+            "AMSS_CUDA_DEVICE_SEGMENT_BATCH": "1",
+        })
+    if getattr(input_data, "basic_grid_set", "") == "Shell-Patch":
+        defaults.update({
+            "AMSS_CUDA_AWARE_MPI": "0",
+            "AMSS_SHELL_FAST_INTERP": "1",
+            "AMSS_SHELL_PARALLEL_INTERP": "1",
+            "AMSS_SHELL_INTERP_THREADS": "16",
+        })
+    if getattr(input_data, "Equation_Class", "") in ("BSSN", "BSSN-EScalar", "Z4C"):
+        defaults["AMSS_CUDA_AMR_RESTRICT_DEVICE"] = "1"
+    if getattr(input_data, "Equation_Class", "") == "Z4C":
+        defaults.update({
+            "AMSS_Z4C_CUDA_RESIDENT": "1",
+            "AMSS_CONSTRAINT_OUT_EVERY": "1000000",
+        })
+    for key, value in defaults.items():
+        runtime_env.setdefault(key, value)
+
+    input_overrides = [
+        "AMSS_EVOLVE_TIMING",
+        "AMSS_ESCALAR_STEP_TIMING",
+        "AMSS_INTERP_FAST",
+        "AMSS_INTERP_GPU",
+        "AMSS_ANALYSIS_MAP_EVERY",
+        "AMSS_CUDA_AWARE_MPI",
+        "AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP",
+        "AMSS_CUDA_KEEP_ALL_LEVELS",
+        "AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP",
+        "AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS",
+        "AMSS_CUDA_EM_CACHE_SOURCES",
+        "AMSS_CUDA_EM_ZERO_FASTPATH",
+        "AMSS_EM_ZERO_ANALYSIS_FASTPATH",
+        "AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH",
+        "AMSS_CUDA_AMR_HOST_STAGED",
+        "AMSS_CUDA_AMR_RESTRICT_DEVICE",
+        "AMSS_CUDA_AMR_RESTRICT_BATCH",
+        "AMSS_CUDA_DEVICE_SEGMENT_BATCH",
+        "AMSS_CUDA_UNCACHED_DEVICE_BUFFERS",
+        "AMSS_SHELL_FAST_INTERP",
+        "AMSS_SHELL_PARALLEL_INTERP",
+        "AMSS_SHELL_CUDA_INTERP",
+        "AMSS_SHELL_INTERP_THREADS",
+        "AMSS_Z4C_CUDA_RESIDENT",
+        "AMSS_CONSTRAINT_OUT_EVERY",
+        "AMSS_Z4C_MRBD",
+    ]
+    for env_name in input_overrides:
+        if env_name not in original_env and hasattr(input_data, env_name):
+            runtime_env[env_name] = str(getattr(input_data, env_name))
+
+    passthrough_envs = [
+        "AMSS_CUDA_RESIDENT_SYNC",
+        "AMSS_CUDA_BSSN_RESIDENT_SYNC",
+        "AMSS_CUDA_EM_RESIDENT_SYNC",
+        "AMSS_CUDA_ESCALAR_RESIDENT_SYNC",
+        "AMSS_CUDA_BH_INTERP_RESIDENT",
+        "AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP",
+        "AMSS_CUDA_KEEP_ALL_LEVELS",
+        "AMSS_CUDA_EM_KEEP_RESIDENT_AFTER_STEP",
+        "AMSS_CUDA_EM_KEEP_ALL_LEVELS",
+        "AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP",
+        "AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS",
+        "AMSS_CUDA_AMR_HOST_STAGED",
+        "AMSS_CUDA_AMR_RESTRICT_DEVICE",
+        "AMSS_CUDA_AMR_RESTRICT_BATCH",
+        "AMSS_CUDA_DEVICE_SEGMENT_BATCH",
+        "AMSS_CUDA_UNCACHED_DEVICE_BUFFERS",
+        "AMSS_CUDA_EM_CACHE_SOURCES",
+        "AMSS_CUDA_EM_ZERO_FASTPATH",
+        "AMSS_CUDA_AWARE_MPI",
+        "AMSS_CUDA_REGRID_FLUSH_ALWAYS",
+        "AMSS_Z4C_CUDA_RESIDENT",
+        "AMSS_SHELL_FAST_INTERP",
+        "AMSS_SHELL_PARALLEL_INTERP",
+        "AMSS_SHELL_CUDA_INTERP",
+        "AMSS_SHELL_INTERP_THREADS",
+        "AMSS_EM_ZERO_ANALYSIS_FASTPATH",
+        "AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH",
+        "AMSS_INTERP_FAST",
+        "AMSS_INTERP_GPU",
+    ]
+    for env_name in passthrough_envs:
+        _input_env_passthrough(runtime_env, env_name)
+
+    optional_overrides = {
+        "AMSS_INTERP_FAST_COMPARE": "AMSS_Interp_Fast_Compare",
+        "AMSS_INTERP_FAST_COMPARE_LIMIT": "AMSS_Interp_Fast_Compare_Limit",
+        "AMSS_INTERP_FAST_COMPARE_TOL": "AMSS_Interp_Fast_Compare_Tol",
+        "AMSS_GPU_STAGE_TIMING": "AMSS_GPU_Stage_Timing",
+        "AMSS_GPU_STAGE_TIMING_EVERY": "AMSS_GPU_Stage_Timing_Every",
+    }
+    for env_name, input_name in optional_overrides.items():
+        if env_name not in runtime_env and hasattr(input_data, input_name):
+            runtime_env[env_name] = str(getattr(input_data, input_name))
+
+    return runtime_env
+
+
+##################################################################
+
 
 
 ##################################################################
@@ -68,11 +294,13 @@ def makefile_ABE():
     print( " Compiling the AMSS-NCKU executable file ABE/ABEGPU " ) 
     print(                                                        )
 
+    z4c_mrbd = int(getattr(input_data, "AMSS_Z4C_MRBD", 0))
+
     ## Build command with CPU binding to nohz_full cores
     if (input_data.GPU_Calculation == "no"):
-        makefile_command  = f"{NUMACTL_CPU_BIND} make -j{BUILD_JOBS} INTERP_LB_MODE=off USE_CUDA_BSSN=0 USE_CUDA_Z4C=0 ABE"
+        makefile_command  = f"{NUMACTL_CPU_BIND} env AMSS_Z4C_MRBD={z4c_mrbd} make -j{BUILD_JOBS} INTERP_LB_MODE=off USE_CUDA_BSSN=0 USE_CUDA_Z4C=0 ABE"
     elif (input_data.GPU_Calculation == "yes"):
-        makefile_command  = f"{NUMACTL_CPU_BIND} make -j{BUILD_JOBS} INTERP_LB_MODE=off USE_CUDA_BSSN=1 USE_CUDA_Z4C=1 ABE_CUDA"
+        makefile_command  = f"{NUMACTL_CPU_BIND} env AMSS_Z4C_MRBD={z4c_mrbd} make -j{BUILD_JOBS} INTERP_LB_MODE=off USE_CUDA_BSSN=1 USE_CUDA_Z4C=1 ABE_CUDA"
     else:
         print( " CPU/GPU numerical calculation setting is wrong " )
         print(                                                    )
@@ -145,29 +373,83 @@ def run_ABE():
     print(                                                      )
 
     ## Define the command to run; cast other values to strings as needed
+    mpi_env = None
+    started_mps = False
     
+    mpi_processes = int(input_data.MPI_processes)
+    if (input_data.GPU_Calculation == "yes" and
+        getattr(input_data, "Equation_Class", "") == "Z4C"):
+        z4c_env_np = os.environ.get("AMSS_Z4C_GPU_MPI_PROCESSES")
+        if z4c_env_np and int(z4c_env_np) > 0:
+            mpi_processes = int(z4c_env_np)
+        elif mpi_processes < 4:
+            mpi_processes = 4
+    if (input_data.GPU_Calculation == "yes" and
+        getattr(input_data, "basic_grid_set", "") == "Shell-Patch"):
+        shell_env_np = os.environ.get("AMSS_SHELL_GPU_MPI_PROCESSES")
+        if shell_env_np and int(shell_env_np) > 0:
+            mpi_processes = int(shell_env_np)
+        elif mpi_processes < 4:
+            mpi_processes = 4
+
     if (input_data.GPU_Calculation == "no"):
-        mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
+        mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(mpi_processes) + " ./ABE"
         #mpi_command         = " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
         mpi_command_outfile = "ABE_out.log"
     elif (input_data.GPU_Calculation == "yes"):
-        mpi_command         = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABE_CUDA"
+        mpi_command         = NUMACTL_CPU_BIND + " I_MPI_OFFLOAD=1 I_MPI_OFFLOAD_IPC=0 mpirun -np " + str(mpi_processes) + " ./ABE_CUDA"
         mpi_command_outfile = "ABEGPU_out.log"
+        mpi_env = _gpu_runtime_env()
+        started_mps = _start_cuda_mps_if_requested(mpi_env)
+        print(" GPU optimized runtime switches:")
+        print(f"   MPI processes={mpi_processes}")
+        print(f"   AMSS_INTERP_FAST={mpi_env.get('AMSS_INTERP_FAST', '')}")
+        print(f"   AMSS_INTERP_GPU={mpi_env.get('AMSS_INTERP_GPU', '')}")
+        print(f"   AMSS_ANALYSIS_MAP_EVERY={mpi_env.get('AMSS_ANALYSIS_MAP_EVERY', '')}")
+        print(f"   AMSS_EVOLVE_TIMING={mpi_env.get('AMSS_EVOLVE_TIMING', '')}")
+        print(f"   AMSS_ESCALAR_STEP_TIMING={mpi_env.get('AMSS_ESCALAR_STEP_TIMING', '')}")
+        print(f"   AMSS_CUDA_AWARE_MPI={mpi_env.get('AMSS_CUDA_AWARE_MPI', '')}")
+        print(f"   AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP={mpi_env.get('AMSS_CUDA_KEEP_RESIDENT_AFTER_STEP', '')}")
+        print(f"   AMSS_CUDA_KEEP_ALL_LEVELS={mpi_env.get('AMSS_CUDA_KEEP_ALL_LEVELS', '')}")
+        print(f"   AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP={mpi_env.get('AMSS_CUDA_ESCALAR_KEEP_RESIDENT_AFTER_STEP', '')}")
+        print(f"   AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS={mpi_env.get('AMSS_CUDA_ESCALAR_KEEP_ALL_LEVELS', '')}")
+        print(f"   AMSS_CUDA_EM_CACHE_SOURCES={mpi_env.get('AMSS_CUDA_EM_CACHE_SOURCES', '')}")
+        print(f"   AMSS_CUDA_EM_ZERO_FASTPATH={mpi_env.get('AMSS_CUDA_EM_ZERO_FASTPATH', '')}")
+        print(f"   AMSS_EM_ZERO_ANALYSIS_FASTPATH={mpi_env.get('AMSS_EM_ZERO_ANALYSIS_FASTPATH', '')}")
+        print(f"   AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH={mpi_env.get('AMSS_EM_ZERO_RESIDENT_DOWNLOAD_FASTPATH', '')}")
+        print(f"   AMSS_CUDA_AMR_HOST_STAGED={mpi_env.get('AMSS_CUDA_AMR_HOST_STAGED', '')}")
+        print(f"   AMSS_CUDA_AMR_RESTRICT_DEVICE={mpi_env.get('AMSS_CUDA_AMR_RESTRICT_DEVICE', '')}")
+        print(f"   AMSS_CUDA_AMR_RESTRICT_BATCH={mpi_env.get('AMSS_CUDA_AMR_RESTRICT_BATCH', '')}")
+        print(f"   AMSS_CUDA_DEVICE_SEGMENT_BATCH={mpi_env.get('AMSS_CUDA_DEVICE_SEGMENT_BATCH', '')}")
+        print(f"   AMSS_CUDA_UNCACHED_DEVICE_BUFFERS={mpi_env.get('AMSS_CUDA_UNCACHED_DEVICE_BUFFERS', '')}")
+        print(f"   AMSS_SHELL_FAST_INTERP={mpi_env.get('AMSS_SHELL_FAST_INTERP', '')}")
+        print(f"   AMSS_SHELL_PARALLEL_INTERP={mpi_env.get('AMSS_SHELL_PARALLEL_INTERP', '')}")
+        print(f"   AMSS_SHELL_CUDA_INTERP={mpi_env.get('AMSS_SHELL_CUDA_INTERP', '')}")
+        print(f"   AMSS_SHELL_INTERP_THREADS={mpi_env.get('AMSS_SHELL_INTERP_THREADS', '')}")
+        print(f"   AMSS_Z4C_CUDA_RESIDENT={mpi_env.get('AMSS_Z4C_CUDA_RESIDENT', '')}")
+        print(f"   AMSS_CONSTRAINT_OUT_EVERY={mpi_env.get('AMSS_CONSTRAINT_OUT_EVERY', '')}")
+        if "CUDA_MPS_PIPE_DIRECTORY" in mpi_env:
+            print(f"   CUDA_MPS_PIPE_DIRECTORY={mpi_env['CUDA_MPS_PIPE_DIRECTORY']}")
  
-    ## Execute the MPI command and stream output
-    mpi_process = subprocess.Popen(mpi_command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
+    try:
+        ## Execute the MPI command and stream output
+        mpi_process = subprocess.Popen(mpi_command, shell=True, stdout=subprocess.PIPE,
+                                       stderr=subprocess.STDOUT, text=True, env=mpi_env)
 
-    ## Write ABE run output to file while printing to stdout
-    with open(mpi_command_outfile, 'w') as file0:  
-        ## Read and print output lines; also write each line to file
-        for line in mpi_process.stdout:
-            print(line, end='')  # stream output in real time
-            file0.write(line)    # write the line to file
-            file0.flush()        # flush to ensure each line is written immediately (optional)            
-    file0.close()
+        ## Write ABE run output to file while printing to stdout
+        with open(mpi_command_outfile, 'w') as file0:
+            ## Read and print output lines; also write each line to file
+            for line in mpi_process.stdout:
+                print(line, end='')  # stream output in real time
+                file0.write(line)    # write the line to file
 
-    ## Wait for the process to finish
-    mpi_return_code = mpi_process.wait()
+        ## Wait for the process to finish
+        mpi_return_code = mpi_process.wait()
+        if mpi_return_code != 0:
+            raise subprocess.CalledProcessError(mpi_return_code, mpi_command)
+    finally:
+        if started_mps:
+            _stop_cuda_mps(mpi_env)
     
     print(                                           )
     print( " The ABE/ABEGPU simulation is finished " ) 
@@ -203,8 +485,6 @@ def run_TwoPunctureABE():
         for line in TwoPuncture_process.stdout:
             print(line, end='')  # stream output in real time
             file0.write(line)    # write the line to file
-            file0.flush()        # flush to ensure each line is written immediately (optional)                 
-    file0.close()
 
     ## Wait for the process to finish
     TwoPuncture_command_return_code = TwoPuncture_process.wait()

plot_xiaoqu.py

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