/*
 * bssn_rhs_cuda.cu — GPU implementation of f_compute_rhs_bssn
 *
 * Drop-in replacement for bssn_rhs_c.C.
 * Compile with nvcc, link bssn_rhs_cuda.o in place of bssn_rhs_c.o.
 */

#include <array>
#include <chrono>
#include <cstdio>
#include <cstdlib>
#include <cmath>
#include <cstring>
#include <unordered_map>
#include <vector>
#include <cuda_runtime.h>
#include "macrodef.h"
#include "bssn_rhs.h"

/* ------------------------------------------------------------------ */
/*  Multi-GPU dispatch: distribute ranks across available GPUs          */
/* ------------------------------------------------------------------ */
static struct {
    int  num_gpus;
    int  my_rank;
    int  my_local_rank;
    int  my_device;
    bool inited;
} g_dispatch = {0, -1, -1, -1, false};

static int env_to_int(const char *name, int fallback = -1) {
    const char *v = getenv(name);
    if (!v || !*v) return fallback;
    return atoi(v);
}

static void init_gpu_dispatch() {
    if (g_dispatch.inited) return;
    cudaError_t err = cudaGetDeviceCount(&g_dispatch.num_gpus);
    if (err != cudaSuccess) g_dispatch.num_gpus = 1;
    if (g_dispatch.num_gpus < 1) g_dispatch.num_gpus = 1;

    /* Get MPI rank from environment (set by mpirun/mpiexec). */
    g_dispatch.my_rank = env_to_int("PMI_RANK",
                         env_to_int("OMPI_COMM_WORLD_RANK",
                         env_to_int("MV2_COMM_WORLD_RANK",
                         env_to_int("SLURM_PROCID", 0))));

    /* Prefer local rank for per-node GPU mapping (avoids cross-node skew). */
    g_dispatch.my_local_rank = env_to_int("OMPI_COMM_WORLD_LOCAL_RANK",
                               env_to_int("MV2_COMM_WORLD_LOCAL_RANK",
                               env_to_int("MPI_LOCALRANKID",
                               env_to_int("SLURM_LOCALID", -1))));

    const int rank_for_map = (g_dispatch.my_local_rank >= 0)
                           ? g_dispatch.my_local_rank : g_dispatch.my_rank;
    g_dispatch.my_device = rank_for_map % g_dispatch.num_gpus;
    cudaSetDevice(g_dispatch.my_device);

    if (g_dispatch.my_rank == 0) {
        printf("[AMSS-GPU] %d GPU(s) detected, device map uses %s rank\n",
               g_dispatch.num_gpus,
               (g_dispatch.my_local_rank >= 0) ? "local" : "global");
    }
    g_dispatch.inited = true;
}

struct CudaProfileStats {
    long long calls;
    double total_ms;
    double state_ms;
    double matter_ms;
    double rhs_ms;
    double bc_ms;
    double finalize_ms;
    double output_ms;
    long long upload_calls;
    long long resident_download_calls;
    double upload_ms;
    double resident_download_ms;
    double upload_gb;
    double resident_download_gb;
};

enum RhsStageId {
    RHS_STAGE_PREP = 0,
    RHS_STAGE_DERIV1,
    RHS_STAGE_METRIC,
    RHS_STAGE_GAUGE_DERIV,
    RHS_STAGE_GAMMA_CONTRACT,
    RHS_STAGE_RICCI_DIFF,
    RHS_STAGE_RICCI_FUSED,
    RHS_STAGE_CHI,
    RHS_STAGE_GAUGE_RHS,
    RHS_STAGE_KODIS,
    RHS_STAGE_CONSTRAINTS,
    RHS_STAGE_COUNT
};

struct RhsStageProfileStats {
    long long calls;
    double ms[RHS_STAGE_COUNT];
};

static CudaProfileStats &cuda_profile_stats() {
    static CudaProfileStats stats = {
        0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
        0, 0, 0.0, 0.0, 0.0, 0.0
    };
    return stats;
}

static RhsStageProfileStats &rhs_stage_profile_stats() {
    static RhsStageProfileStats stats = {};
    return stats;
}

static bool cuda_profile_enabled() {
    static int enabled = -1;
    if (enabled < 0) {
        const char *env = getenv("AMSS_PROFILE_CUDA");
        enabled = (env && atoi(env) != 0) ? 1 : 0;
    }
    return enabled != 0;
}

static int cuda_profile_every() {
    static int every = -1;
    if (every < 0) {
        const char *env = getenv("AMSS_PROFILE_CUDA_EVERY");
        every = (env && atoi(env) > 0) ? atoi(env) : 100;
    }
    return every;
}

static bool rhs_stage_timing_enabled() {
    static int enabled = -1;
    if (enabled < 0) {
        const char *env = getenv("AMSS_GPU_STAGE_TIMING");
        enabled = (env && atoi(env) != 0) ? 1 : 0;
    }
    return enabled != 0;
}

static int rhs_stage_timing_every() {
    static int every = -1;
    if (every < 0) {
        const char *env = getenv("AMSS_GPU_STAGE_TIMING_EVERY");
        every = (env && atoi(env) > 0) ? atoi(env) : cuda_profile_every();
    }
    return every;
}

static double cuda_profile_now_ms() {
    using clock = std::chrono::steady_clock;
    return std::chrono::duration<double, std::milli>(
               clock::now().time_since_epoch()).count();
}

static void cuda_profile_sync() {
    cudaError_t err = cudaDeviceSynchronize();
    if (err != cudaSuccess) {
        fprintf(stderr, "CUDA error %s:%d: %s\n",
                __FILE__, __LINE__, cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }
}

static void cuda_profile_maybe_log() {
    if (!cuda_profile_enabled()) return;
    CudaProfileStats &stats = cuda_profile_stats();
    if (stats.calls <= 0 || stats.calls % cuda_profile_every() != 0) return;
    fprintf(stderr,
            "[AMSS-CUDA][rank %d][dev %d] calls=%lld avg_total=%.3f ms avg_state=%.3f ms avg_matter=%.3f ms avg_rhs=%.3f ms avg_bc=%.3f ms avg_finalize=%.3f ms avg_output=%.3f ms"
            " uploads=%lld avg_upload=%.3f ms upload_GB=%.3f resident_downloads=%lld avg_resident_download=%.3f ms resident_download_GB=%.3f\n",
            g_dispatch.my_rank, g_dispatch.my_device, stats.calls,
            stats.total_ms / (double)stats.calls,
            stats.state_ms / (double)stats.calls,
            stats.matter_ms / (double)stats.calls,
            stats.rhs_ms / (double)stats.calls,
            stats.bc_ms / (double)stats.calls,
            stats.finalize_ms / (double)stats.calls,
            stats.output_ms / (double)stats.calls,
            stats.upload_calls,
            stats.upload_calls ? stats.upload_ms / (double)stats.upload_calls : 0.0,
            stats.upload_gb,
            stats.resident_download_calls,
            stats.resident_download_calls ? stats.resident_download_ms / (double)stats.resident_download_calls : 0.0,
            stats.resident_download_gb);
    fflush(stderr);
}

static void rhs_stage_profile_accumulate(const double *stage_ms) {
    if (!rhs_stage_timing_enabled()) return;

    RhsStageProfileStats &stats = rhs_stage_profile_stats();
    stats.calls++;
    for (int i = 0; i < RHS_STAGE_COUNT; ++i) {
        stats.ms[i] += stage_ms[i];
    }
    if (stats.calls <= 0 || stats.calls % rhs_stage_timing_every() != 0) return;

    fprintf(stderr,
            "[AMSS-CUDA-STAGE][rank %d][dev %d] calls=%lld"
            " prep=%.3f deriv1=%.3f metric=%.3f gauge_deriv=%.3f"
            " gamma_contract=%.3f ricci_diff=%.3f ricci_fused=%.3f"
            " chi=%.3f gauge_rhs=%.3f kodis=%.3f constraints=%.3f ms\n",
            g_dispatch.my_rank, g_dispatch.my_device, stats.calls,
            stats.ms[RHS_STAGE_PREP] / (double)stats.calls,
            stats.ms[RHS_STAGE_DERIV1] / (double)stats.calls,
            stats.ms[RHS_STAGE_METRIC] / (double)stats.calls,
            stats.ms[RHS_STAGE_GAUGE_DERIV] / (double)stats.calls,
            stats.ms[RHS_STAGE_GAMMA_CONTRACT] / (double)stats.calls,
            stats.ms[RHS_STAGE_RICCI_DIFF] / (double)stats.calls,
            stats.ms[RHS_STAGE_RICCI_FUSED] / (double)stats.calls,
            stats.ms[RHS_STAGE_CHI] / (double)stats.calls,
            stats.ms[RHS_STAGE_GAUGE_RHS] / (double)stats.calls,
            stats.ms[RHS_STAGE_KODIS] / (double)stats.calls,
            stats.ms[RHS_STAGE_CONSTRAINTS] / (double)stats.calls);
    fflush(stderr);
}

/* ------------------------------------------------------------------ */
/*  Error checking                                                     */
/* ------------------------------------------------------------------ */
#define CUDA_CHECK(call) do {                                         \
    cudaError_t err = (call);                                         \
    if (err != cudaSuccess) {                                         \
        fprintf(stderr, "CUDA error %s:%d: %s\n",                    \
                __FILE__, __LINE__, cudaGetErrorString(err));         \
        exit(EXIT_FAILURE);                                           \
    }                                                                 \
} while(0)

/* ------------------------------------------------------------------ */
/*  Physical / gauge constants (matching bssn_rhs_c.C)                 */
/* ------------------------------------------------------------------ */
static const double PI_VAL  = 3.14159265358979323846;
static const double FF_VAL  = 0.75;
static const double ETA_VAL = 2.0;

/* ------------------------------------------------------------------ */
/*  Constant memory for grid parameters and stencil coefficients       */
/* ------------------------------------------------------------------ */
struct GridParams {
    int    ex[3];          /* nx, ny, nz */
    int    all;            /* nx*ny*nz   */
    double dX, dY, dZ;
    /* fderivs coefficients */
    double d12dx, d12dy, d12dz;   /* 1/(12*dX) etc */
    double d2dx,  d2dy,  d2dz;   /* 1/(2*dX)  etc */
    /* fdderivs coefficients */
    double Fdxdx, Fdydy, Fdzdz;  /* 1/(12*dX^2) etc */
    double Sdxdx, Sdydy, Sdzdz;  /* 1/(dX^2)    etc */
    double Fdxdy, Fdxdz, Fdydz;  /* 1/(144*dX*dY) etc */
    double Sdxdy, Sdxdz, Sdydz;  /* 1/(4*dX*dY)   etc */
    /* symmetry bounds (Fortran 1-based) */
    int iminF, jminF, kminF;
    int imaxF, jmaxF, kmaxF;
    /* symmetry bounds for ord=3 (lopsided/kodis) */
    int iminF3, jminF3, kminF3;
    int Symmetry;
    double eps;
    int co;
    /* padded sizes */
    int fh2_nx, fh2_ny, fh2_nz;  /* (nx+2), (ny+2), (nz+2) for ord=2 */
    int fh3_nx, fh3_ny, fh3_nz;  /* (nx+3), (ny+3), (nz+3) for ord=3 */
};

__constant__ GridParams d_gp;
static GridParams g_gp_host_cache = {};
static bool g_gp_host_cache_valid = false;

/* ------------------------------------------------------------------ */
/*  Device indexing helpers                                             */
/* ------------------------------------------------------------------ */
__device__ __forceinline__ int idx_ex_d(int i0, int j0, int k0) {
    return i0 + j0 * d_gp.ex[0] + k0 * d_gp.ex[0] * d_gp.ex[1];
}

__device__ __forceinline__ double fetch_sym_ord2_direct(const double *src,
                                                        int iF, int jF, int kF,
                                                        int SoA0, int SoA1, int SoA2)
{
    int siF = iF;
    int sjF = jF;
    int skF = kF;
    double sign = 1.0;

    if (iF <= 0) {
        siF = 1 - iF;
        sign *= (double)SoA0;
    }
    if (jF <= 0) {
        sjF = 1 - jF;
        sign *= (double)SoA1;
    }
    if (kF <= 0) {
        skF = 1 - kF;
        sign *= (double)SoA2;
    }

    return sign * src[(siF - 1)
                    + (sjF - 1) * d_gp.ex[0]
                    + (skF - 1) * d_gp.ex[0] * d_gp.ex[1]];
}

/* ord=2 ghost-padded: Fortran index iF -> flat index */
__device__ __forceinline__ int idx_fh2(int iF, int jF, int kF) {
    return (iF + 1) + (jF + 1) * d_gp.fh2_nx + (kF + 1) * d_gp.fh2_nx * d_gp.fh2_ny;
}

/* ord=3 ghost-padded: Fortran index iF -> flat index */
__device__ __forceinline__ int idx_fh3(int iF, int jF, int kF) {
    return (iF + 2) + (jF + 2) * d_gp.fh3_nx + (kF + 2) * d_gp.fh3_nx * d_gp.fh3_ny;
}

__device__ __forceinline__ double fetch_sym_ord3_direct(const double *src,
                                                        int iF, int jF, int kF,
                                                        int SoA0, int SoA1, int SoA2)
{
    int siF = iF;
    int sjF = jF;
    int skF = kF;
    double sign = 1.0;

    if (iF <= 0) {
        siF = 1 - iF;
        sign *= (double)SoA0;
    }
    if (jF <= 0) {
        sjF = 1 - jF;
        sign *= (double)SoA1;
    }
    if (kF <= 0) {
        skF = 1 - kF;
        sign *= (double)SoA2;
    }

    return sign * src[(siF - 1)
                    + (sjF - 1) * d_gp.ex[0]
                    + (skF - 1) * d_gp.ex[0] * d_gp.ex[1]];
}

/* ------------------------------------------------------------------ */
/*  GPU buffer management                                              */
/* ------------------------------------------------------------------ */
/*
 * Array slot indices — all arrays live in one big cudaMalloc block.
 * INPUT arrays (H2D): 39 slots
 * OUTPUT arrays (D2H): 52 slots
 * TEMPORARY arrays (GPU-only): ~65 slots
 * Plus 2 extended arrays for ghost-padded stencils (fh_ord2, fh_ord3)
 */

/* Total number of "all"-sized slots */
#define NUM_SLOTS 160

struct GpuBuffers {
    double *d_mem;          /* single big allocation */
    double *d_fh2;          /* ghost-padded ord=2: (nx+2)*(ny+2)*(nz+2) */
    double *d_fh3;          /* ghost-padded ord=3: (nx+3)*(ny+3)*(nz+3) */
    double *h_stage;        /* host staging buffer for bulk H2D/D2H */
    bool    h_stage_pinned; /* true if allocated by cudaMallocHost */
    double *slot[NUM_SLOTS]; /* pointers into d_mem */
    size_t  cap_all;
    size_t  cap_fh2_size;
    size_t  cap_fh3_size;
    int     prev_nx, prev_ny, prev_nz;
    bool    initialized;
};

static GpuBuffers g_buf = {
    nullptr, nullptr, nullptr, nullptr, false, {},
    0, 0, 0, 0, 0, 0, false
};

/* Slot assignments — INPUT (H2D) */
enum {
    S_chi=0, S_trK, S_dxx, S_gxy, S_gxz, S_dyy, S_gyz, S_dzz,
    S_Axx, S_Axy, S_Axz, S_Ayy, S_Ayz, S_Azz,
    S_Gamx, S_Gamy, S_Gamz,
    S_Lap, S_betax, S_betay, S_betaz,
    S_dtSfx, S_dtSfy, S_dtSfz,
    S_rho, S_Sx, S_Sy, S_Sz,
    S_Sxx, S_Sxy, S_Sxz, S_Syy, S_Syz, S_Szz,
    S_X, S_Y, S_Z,                /* coordinate arrays — only nx/ny/nz long */
    /* 37 input slots so far; X/Y/Z are special-sized */

    /* OUTPUT (D2H) */
    S_chi_rhs, S_trK_rhs,
    S_gxx_rhs, S_gxy_rhs, S_gxz_rhs, S_gyy_rhs, S_gyz_rhs, S_gzz_rhs,
    S_Axx_rhs, S_Axy_rhs, S_Axz_rhs, S_Ayy_rhs, S_Ayz_rhs, S_Azz_rhs,
    S_Gamx_rhs, S_Gamy_rhs, S_Gamz_rhs,
    S_Lap_rhs, S_betax_rhs, S_betay_rhs, S_betaz_rhs,
    S_dtSfx_rhs, S_dtSfy_rhs, S_dtSfz_rhs,
    S_Gamxxx, S_Gamxxy, S_Gamxxz, S_Gamxyy, S_Gamxyz, S_Gamxzz,
    S_Gamyxx, S_Gamyxy, S_Gamyxz, S_Gamyyy, S_Gamyyz, S_Gamyzz,
    S_Gamzxx, S_Gamzxy, S_Gamzxz, S_Gamzyy, S_Gamzyz, S_Gamzzz,
    S_Rxx, S_Rxy, S_Rxz, S_Ryy, S_Ryz, S_Rzz,
    S_ham_Res, S_movx_Res, S_movy_Res, S_movz_Res,
    S_Gmx_Res, S_Gmy_Res, S_Gmz_Res,

    /* TEMPORARY (GPU-only) */
    S_gxx, S_gyy, S_gzz,          /* physical metric = dxx+1 etc */
    S_alpn1, S_chin1,
    S_chix, S_chiy, S_chiz,
    S_gxxx, S_gxyx, S_gxzx, S_gyyx, S_gyzx, S_gzzx,
    S_gxxy, S_gxyy, S_gxzy, S_gyyy, S_gyzy, S_gzzy,
    S_gxxz, S_gxyz, S_gxzz, S_gyyz, S_gyzz, S_gzzz,
    S_Lapx, S_Lapy, S_Lapz,
    S_betaxx, S_betaxy, S_betaxz,
    S_betayx, S_betayy, S_betayz,
    S_betazx, S_betazy, S_betazz,
    S_Gamxx, S_Gamxy, S_Gamxz,
    S_Gamyx, S_Gamyy_t, S_Gamyz_t,
    S_Gamzx, S_Gamzy, S_Gamzz_t,
    S_Kx, S_Ky, S_Kz,
    S_S_arr, S_f_arr,
    S_fxx, S_fxy, S_fxz, S_fyy, S_fyz, S_fzz,
    S_Gamxa, S_Gamya, S_Gamza,
    S_gupxx, S_gupxy, S_gupxz,
    S_gupyy, S_gupyz, S_gupzz,
    NUM_USED_SLOTS
};

static_assert(NUM_USED_SLOTS <= NUM_SLOTS, "Increase NUM_SLOTS");

static const int H2D_INPUT_SLOT_COUNT = (S_Szz - S_chi + 1);
static const int D2H_BASE_SLOT_COUNT = (S_Rzz - S_chi_rhs + 1);
static const int D2H_CONSTRAINT_SLOT_COUNT = (S_Gmz_Res - S_ham_Res + 1);
static const int STAGE_SLOT_COUNT =
    (H2D_INPUT_SLOT_COUNT > (D2H_BASE_SLOT_COUNT + D2H_CONSTRAINT_SLOT_COUNT))
        ? H2D_INPUT_SLOT_COUNT
        : (D2H_BASE_SLOT_COUNT + D2H_CONSTRAINT_SLOT_COUNT);

static constexpr int BSSN_STATE_COUNT = 24;
static constexpr int BSSN_MATTER_COUNT = 10;
static constexpr int BSSN_LK_FIELD_COUNT = 24;

static const int k_state_input_slots[BSSN_STATE_COUNT] = {
    S_chi, S_trK, S_dxx, S_gxy, S_gxz, S_dyy, S_gyz, S_dzz,
    S_Axx, S_Axy, S_Axz, S_Ayy, S_Ayz, S_Azz,
    S_Gamx, S_Gamy, S_Gamz,
    S_Lap, S_betax, S_betay, S_betaz,
    S_dtSfx, S_dtSfy, S_dtSfz
};

static const int k_state_rhs_slots[BSSN_STATE_COUNT] = {
    S_chi_rhs, S_trK_rhs,
    S_gxx_rhs, S_gxy_rhs, S_gxz_rhs, S_gyy_rhs, S_gyz_rhs, S_gzz_rhs,
    S_Axx_rhs, S_Axy_rhs, S_Axz_rhs, S_Ayy_rhs, S_Ayz_rhs, S_Azz_rhs,
    S_Gamx_rhs, S_Gamy_rhs, S_Gamz_rhs,
    S_Lap_rhs, S_betax_rhs, S_betay_rhs, S_betaz_rhs,
    S_dtSfx_rhs, S_dtSfy_rhs, S_dtSfz_rhs
};

static const int k_matter_slots[BSSN_MATTER_COUNT] = {
    S_rho, S_Sx, S_Sy, S_Sz, S_Sxx, S_Sxy, S_Sxz, S_Syy, S_Syz, S_Szz
};

static const int k_lk_adv_slots[BSSN_LK_FIELD_COUNT] = {
    S_gxx, S_Gamz, S_gxy, S_Lap, S_gxz, S_betax, S_gyy, S_betay,
    S_gyz, S_betaz, S_gzz, S_dtSfx, S_Axx, S_dtSfy, S_Axy, S_dtSfz,
    S_Axz, S_Ayy, S_Ayz, S_Azz, S_chi, S_trK, S_Gamx, S_Gamy
};

static const int k_lk_ko_slots[BSSN_LK_FIELD_COUNT] = {
    S_dxx, S_Gamz, S_gxy, S_Lap, S_gxz, S_betax, S_dyy, S_betay,
    S_gyz, S_betaz, S_dzz, S_dtSfx, S_Axx, S_dtSfy, S_Axy, S_dtSfz,
    S_Axz, S_Ayy, S_Ayz, S_Azz, S_chi, S_trK, S_Gamx, S_Gamy
};

static const int k_lk_rhs_slots[BSSN_LK_FIELD_COUNT] = {
    S_gxx_rhs, S_Gamz_rhs, S_gxy_rhs, S_Lap_rhs, S_gxz_rhs, S_betax_rhs,
    S_gyy_rhs, S_betay_rhs, S_gyz_rhs, S_betaz_rhs, S_gzz_rhs, S_dtSfx_rhs,
    S_Axx_rhs, S_dtSfy_rhs, S_Axy_rhs, S_dtSfz_rhs, S_Axz_rhs, S_Ayy_rhs,
    S_Ayz_rhs, S_Azz_rhs, S_chi_rhs, S_trK_rhs, S_Gamx_rhs, S_Gamy_rhs
};

__constant__ int d_subset_state_indices[BSSN_STATE_COUNT];

static const int k_lk_soa_signs[3 * BSSN_LK_FIELD_COUNT] = {
     1,  1,  1,
     1,  1, -1,
    -1, -1,  1,
     1,  1,  1,
    -1,  1, -1,
    -1,  1,  1,
     1,  1,  1,
     1, -1,  1,
     1, -1, -1,
     1,  1, -1,
     1,  1,  1,
    -1,  1,  1,
     1,  1,  1,
     1, -1,  1,
    -1, -1,  1,
     1,  1, -1,
    -1,  1, -1,
     1,  1,  1,
     1, -1, -1,
     1,  1,  1,
     1,  1,  1,
     1,  1,  1,
    -1,  1,  1,
     1, -1,  1
};

struct StepContext {
    double *d_state0_mem;
    double *d_accum_mem;
    double *d_state_curr_mem;
    double *d_state_next_mem;
    double *d_matter_mem;
    double *d_comm_mem;
    double *h_comm_mem;
    std::array<double *, BSSN_STATE_COUNT> d_state0;
    std::array<double *, BSSN_STATE_COUNT> d_accum;
    std::array<double *, BSSN_STATE_COUNT> d_state_curr;
    std::array<double *, BSSN_STATE_COUNT> d_state_next;
    std::array<double *, BSSN_MATTER_COUNT> d_matter;
    size_t cap_all;
    size_t cap_comm;
    bool h_comm_pinned;
    size_t cap_h_comm;
    bool matter_ready;
    bool state_ready;

    StepContext()
        : d_state0_mem(nullptr), d_accum_mem(nullptr),
          d_state_curr_mem(nullptr), d_state_next_mem(nullptr),
          d_matter_mem(nullptr), d_comm_mem(nullptr), h_comm_mem(nullptr),
          cap_all(0), cap_comm(0), h_comm_pinned(false), cap_h_comm(0),
          matter_ready(false), state_ready(false)
    {
        d_state0.fill(nullptr);
        d_accum.fill(nullptr);
        d_state_curr.fill(nullptr);
        d_state_next.fill(nullptr);
        d_matter.fill(nullptr);
    }
};

struct StepAllocation {
    double *d_state0_mem;
    double *d_accum_mem;
    double *d_state_curr_mem;
    double *d_state_next_mem;
    double *d_matter_mem;
    double *d_comm_mem;
    double *h_comm_mem;
    size_t cap_all;
    size_t cap_comm;
    bool h_comm_pinned;
    size_t cap_h_comm;
};

static std::unordered_map<void *, StepContext> g_step_ctx;
static std::vector<StepAllocation> g_step_pool;
static int *g_comm_segment_meta = nullptr;
static size_t g_comm_segment_meta_cap = 0;

static StepAllocation empty_step_allocation()
{
    StepAllocation alloc = {
        nullptr, nullptr, nullptr, nullptr, nullptr,
        nullptr, nullptr,
        0, 0, false, 0
    };
    return alloc;
}

static bool has_step_allocation(const StepAllocation &alloc)
{
    return alloc.cap_all != 0;
}

static StepAllocation detach_step_allocation(StepContext &ctx)
{
    StepAllocation alloc = {
        ctx.d_state0_mem, ctx.d_accum_mem, ctx.d_state_curr_mem,
        ctx.d_state_next_mem, ctx.d_matter_mem,
        ctx.d_comm_mem, ctx.h_comm_mem,
        ctx.cap_all, ctx.cap_comm, ctx.h_comm_pinned, ctx.cap_h_comm
    };
    ctx.d_state0_mem = nullptr;
    ctx.d_accum_mem = nullptr;
    ctx.d_state_curr_mem = nullptr;
    ctx.d_state_next_mem = nullptr;
    ctx.d_matter_mem = nullptr;
    ctx.d_comm_mem = nullptr;
    ctx.h_comm_mem = nullptr;
    ctx.cap_all = 0;
    ctx.cap_comm = 0;
    ctx.h_comm_pinned = false;
    ctx.cap_h_comm = 0;
    ctx.matter_ready = false;
    ctx.state_ready = false;
    ctx.d_state0.fill(nullptr);
    ctx.d_accum.fill(nullptr);
    ctx.d_state_curr.fill(nullptr);
    ctx.d_state_next.fill(nullptr);
    ctx.d_matter.fill(nullptr);
    return alloc;
}

static void attach_step_allocation(StepContext &ctx, const StepAllocation &alloc)
{
    ctx.d_state0_mem = alloc.d_state0_mem;
    ctx.d_accum_mem = alloc.d_accum_mem;
    ctx.d_state_curr_mem = alloc.d_state_curr_mem;
    ctx.d_state_next_mem = alloc.d_state_next_mem;
    ctx.d_matter_mem = alloc.d_matter_mem;
    ctx.d_comm_mem = alloc.d_comm_mem;
    ctx.h_comm_mem = alloc.h_comm_mem;
    ctx.cap_all = alloc.cap_all;
    ctx.cap_comm = alloc.cap_comm;
    ctx.h_comm_pinned = alloc.h_comm_pinned;
    ctx.cap_h_comm = alloc.cap_h_comm;
    ctx.matter_ready = false;
    ctx.state_ready = false;
}

static void recycle_step_allocation(StepAllocation &alloc)
{
    if (!has_step_allocation(alloc)) return;
    g_step_pool.push_back(alloc);
    alloc = empty_step_allocation();
}

static StepAllocation acquire_step_allocation(size_t all)
{
    size_t best = g_step_pool.size();
    for (size_t i = 0; i < g_step_pool.size(); ++i) {
        if (g_step_pool[i].cap_all < all) continue;
        if (best == g_step_pool.size() || g_step_pool[i].cap_all < g_step_pool[best].cap_all)
            best = i;
    }
    if (best == g_step_pool.size())
        return empty_step_allocation();

    StepAllocation alloc = g_step_pool[best];
    g_step_pool[best] = g_step_pool.back();
    g_step_pool.pop_back();
    return alloc;
}

static void ensure_gpu_buffers(int nx, int ny, int nz) {
    size_t all = (size_t)nx * ny * nz;
    size_t fh2_size = (size_t)(nx+2) * (ny+2) * (nz+2);
    size_t fh3_size = (size_t)(nx+3) * (ny+3) * (nz+3);
    const bool need_grow = (!g_buf.initialized)
                        || (all > g_buf.cap_all)
                        || (fh2_size > g_buf.cap_fh2_size)
                        || (fh3_size > g_buf.cap_fh3_size);

    if (need_grow) {
        if (g_buf.d_mem) { cudaFree(g_buf.d_mem); g_buf.d_mem = nullptr; }
        if (g_buf.d_fh2) { cudaFree(g_buf.d_fh2); g_buf.d_fh2 = nullptr; }
        if (g_buf.d_fh3) { cudaFree(g_buf.d_fh3); g_buf.d_fh3 = nullptr; }
        if (g_buf.h_stage) {
            if (g_buf.h_stage_pinned) cudaFreeHost(g_buf.h_stage);
            else free(g_buf.h_stage);
            g_buf.h_stage = nullptr;
            g_buf.h_stage_pinned = false;
        }

        CUDA_CHECK(cudaMalloc(&g_buf.d_mem, NUM_USED_SLOTS * all * sizeof(double)));
        CUDA_CHECK(cudaMalloc(&g_buf.d_fh2, fh2_size * sizeof(double)));
        CUDA_CHECK(cudaMalloc(&g_buf.d_fh3, fh3_size * sizeof(double)));

        const size_t stage_bytes = (size_t)STAGE_SLOT_COUNT * all * sizeof(double);
        cudaError_t stage_err = cudaMallocHost((void**)&g_buf.h_stage, stage_bytes);
        if (stage_err == cudaSuccess) {
            g_buf.h_stage_pinned = true;
        } else {
            g_buf.h_stage = (double *)malloc(stage_bytes);
            g_buf.h_stage_pinned = false;
            if (!g_buf.h_stage) {
                fprintf(stderr, "Host stage allocation failed (%zu bytes)\n", stage_bytes);
                exit(EXIT_FAILURE);
            }
        }

        g_buf.cap_all = all;
        g_buf.cap_fh2_size = fh2_size;
        g_buf.cap_fh3_size = fh3_size;
        g_buf.initialized = true;
    }

    for (int s = 0; s < NUM_USED_SLOTS; ++s)
        g_buf.slot[s] = g_buf.d_mem + s * all;

    g_buf.prev_nx = nx;
    g_buf.prev_ny = ny;
    g_buf.prev_nz = nz;
}

static StepContext &ensure_step_ctx(void *block_tag, size_t all)
{
    StepContext &ctx = g_step_ctx[block_tag];
    if (ctx.cap_all < all) {
        StepAllocation old_alloc = detach_step_allocation(ctx);
        recycle_step_allocation(old_alloc);

        StepAllocation alloc = acquire_step_allocation(all);
        if (!has_step_allocation(alloc)) {
            CUDA_CHECK(cudaMalloc(&alloc.d_state0_mem, BSSN_STATE_COUNT * all * sizeof(double)));
            CUDA_CHECK(cudaMalloc(&alloc.d_accum_mem, BSSN_STATE_COUNT * all * sizeof(double)));
            CUDA_CHECK(cudaMalloc(&alloc.d_state_curr_mem, BSSN_STATE_COUNT * all * sizeof(double)));
            CUDA_CHECK(cudaMalloc(&alloc.d_state_next_mem, BSSN_STATE_COUNT * all * sizeof(double)));
            CUDA_CHECK(cudaMalloc(&alloc.d_matter_mem, BSSN_MATTER_COUNT * all * sizeof(double)));
            alloc.cap_all = all;
        }
        attach_step_allocation(ctx, alloc);
    }
    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
        ctx.d_state0[i] = ctx.d_state0_mem + (size_t)i * all;
        ctx.d_accum[i] = ctx.d_accum_mem + (size_t)i * all;
        ctx.d_state_curr[i] = ctx.d_state_curr_mem + (size_t)i * all;
        ctx.d_state_next[i] = ctx.d_state_next_mem + (size_t)i * all;
    }
    for (int i = 0; i < BSSN_MATTER_COUNT; ++i) {
        ctx.d_matter[i] = ctx.d_matter_mem + (size_t)i * all;
    }
    return ctx;
}

static void release_step_ctx(void *block_tag)
{
    auto it = g_step_ctx.find(block_tag);
    if (it == g_step_ctx.end()) return;
    StepAllocation alloc = detach_step_allocation(it->second);
    recycle_step_allocation(alloc);
    g_step_ctx.erase(it);
}

static double *ensure_step_comm_buffer(StepContext &ctx, size_t needed_doubles)
{
    if (needed_doubles == 0) return nullptr;
    if (ctx.cap_comm < needed_doubles) {
        if (ctx.d_comm_mem) {
            CUDA_CHECK(cudaFree(ctx.d_comm_mem));
            ctx.d_comm_mem = nullptr;
        }
        CUDA_CHECK(cudaMalloc(&ctx.d_comm_mem, needed_doubles * sizeof(double)));
        ctx.cap_comm = needed_doubles;
    }
    return ctx.d_comm_mem;
}

static double *ensure_step_host_comm_buffer(StepContext &ctx, size_t needed_doubles)
{
    if (needed_doubles == 0) return nullptr;
    if (ctx.cap_h_comm < needed_doubles) {
        if (ctx.h_comm_mem) {
            if (ctx.h_comm_pinned) cudaFreeHost(ctx.h_comm_mem);
            else free(ctx.h_comm_mem);
            ctx.h_comm_mem = nullptr;
            ctx.h_comm_pinned = false;
        }

        const size_t bytes = needed_doubles * sizeof(double);
        cudaError_t err = cudaMallocHost((void **)&ctx.h_comm_mem, bytes);
        if (err == cudaSuccess) {
            ctx.h_comm_pinned = true;
        } else {
            ctx.h_comm_mem = (double *)malloc(bytes);
            ctx.h_comm_pinned = false;
            if (!ctx.h_comm_mem) {
                fprintf(stderr, "Host comm allocation failed (%zu bytes)\n", bytes);
                exit(EXIT_FAILURE);
            }
        }
        ctx.cap_h_comm = needed_doubles;
    }
    return ctx.h_comm_mem;
}

static int *ensure_comm_segment_meta_buffer(size_t needed_ints)
{
    if (needed_ints == 0) return nullptr;
    if (g_comm_segment_meta_cap < needed_ints) {
        if (g_comm_segment_meta) {
            CUDA_CHECK(cudaFree(g_comm_segment_meta));
            g_comm_segment_meta = nullptr;
        }
        CUDA_CHECK(cudaMalloc(&g_comm_segment_meta, needed_ints * sizeof(int)));
        g_comm_segment_meta_cap = needed_ints;
    }
    return g_comm_segment_meta;
}

static void upload_grid_params_if_needed(const GridParams &gp)
{
    if (!g_gp_host_cache_valid ||
        std::memcmp(&g_gp_host_cache, &gp, sizeof(GridParams)) != 0) {
        CUDA_CHECK(cudaMemcpyToSymbol(d_gp, &gp, sizeof(GridParams)));
        g_gp_host_cache = gp;
        g_gp_host_cache_valid = true;
    }
}

/* ================================================================== */
/*  A. Symmetry boundary kernels (ord=2 and ord=3)                     */
/* ================================================================== */

/* Step 1: Copy interior into ghost-padded array */
__global__ void kern_symbd_copy_interior_ord2(const double * __restrict__ func,
                                               double * __restrict__ fh,
                                               double SoA0, double SoA1, double SoA2)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int fnx = d_gp.fh2_nx, fny = d_gp.fh2_ny;
    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < d_gp.all;
         tid += blockDim.x * gridDim.x)
    {
        int i0 = tid % nx;
        int j0 = (tid / nx) % ny;
        int k0 = tid / (nx * ny);
        int iF = i0 + 1, jF = j0 + 1, kF = k0 + 1;
        fh[(iF+1) + (jF+1)*fnx + (kF+1)*fnx*fny] = func[tid];
    }
}

/* Fused symmetry pack (ord=2): fill full fh from interior func in one pass. */
__global__ void kern_symbd_pack_ord2(const double * __restrict__ func,
                                     double * __restrict__ fh,
                                     double SoA0, double SoA1, double SoA2)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1];
    const int fnx = d_gp.fh2_nx, fny = d_gp.fh2_ny, fnz = d_gp.fh2_nz;
    const int total = fnx * fny * fnz;

    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < total;
         tid += blockDim.x * gridDim.x)
    {
        int ii = tid % fnx;
        int jj = (tid / fnx) % fny;
        int kk = tid / (fnx * fny);

        int iF = ii - 1;   /* -1 .. nx */
        int jF = jj - 1;   /* -1 .. ny */
        int kF = kk - 1;   /* -1 .. nz */

        int siF = (iF <= 0) ? (1 - iF) : iF;  /* 1..nx */
        int sjF = (jF <= 0) ? (1 - jF) : jF;  /* 1..ny */
        int skF = (kF <= 0) ? (1 - kF) : kF;  /* 1..nz */

        double sign = 1.0;
        if (iF <= 0) sign *= SoA0;
        if (jF <= 0) sign *= SoA1;
        if (kF <= 0) sign *= SoA2;

        int src = (siF - 1) + (sjF - 1) * nx + (skF - 1) * nx * ny;
        fh[tid] = sign * func[src];
    }
}

/* Step 2: Fill i-ghosts (x-direction symmetry) */
__global__ void kern_symbd_ighost_ord2(double * __restrict__ fh, double SoA0)
{
    const int ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int fnx = d_gp.fh2_nx, fny = d_gp.fh2_ny;
    /* ord=2: fill iF=0 and iF=-1, i.e. ghost layers ii=0 from ii=2, ii=1 from ii=1 */
    /* Fortran: do ii=0,ord-1: funcc(-ii,jF,kF) = funcc(ii+1,jF,kF)*SoA[0] */
    int total = ny * nz;  /* jF=1..ny, kF=1..nz */
    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < total * 2;  /* 2 ghost layers */
         tid += blockDim.x * gridDim.x)
    {
        int ii = tid / total;  /* 0 or 1 */
        int rem = tid % total;
        int j0 = rem % ny;
        int k0 = rem / ny;
        int jF = j0 + 1, kF = k0 + 1;
        int iF_dst = -ii;     /* 0, -1 */
        int iF_src = ii + 1;  /* 1, 2  */
        fh[(iF_dst+1) + (jF+1)*fnx + (kF+1)*fnx*fny] =
            fh[(iF_src+1) + (jF+1)*fnx + (kF+1)*fnx*fny] * SoA0;
    }
}

/* Step 3: Fill j-ghosts (y-direction symmetry) */
__global__ void kern_symbd_jghost_ord2(double * __restrict__ fh, double SoA1)
{
    const int nx = d_gp.ex[0], nz = d_gp.ex[2];
    const int fnx = d_gp.fh2_nx, fny = d_gp.fh2_ny;
    /* iF ranges from -1 to nx (i.e. -ord+1 to ex1), total = nx+2 */
    int irange = nx + 2;
    int total = irange * nz;
    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < total * 2;
         tid += blockDim.x * gridDim.x)
    {
        int jj = tid / total;
        int rem = tid % total;
        int ii = rem % irange;
        int k0 = rem / irange;
        int iF = ii - 1;  /* -1 .. nx */
        int kF = k0 + 1;
        int jF_dst = -jj;
        int jF_src = jj + 1;
        fh[(iF+1) + (jF_dst+1)*fnx + (kF+1)*fnx*fny] =
            fh[(iF+1) + (jF_src+1)*fnx + (kF+1)*fnx*fny] * SoA1;
    }
}

/* Step 4: Fill k-ghosts (z-direction symmetry) */
__global__ void kern_symbd_kghost_ord2(double * __restrict__ fh, double SoA2)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1];
    const int fnx = d_gp.fh2_nx, fny = d_gp.fh2_ny;
    int irange = nx + 2;
    int jrange = ny + 2;
    int total = irange * jrange;
    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < total * 2;
         tid += blockDim.x * gridDim.x)
    {
        int kk = tid / total;
        int rem = tid % total;
        int ii = rem % irange;
        int jj = rem / irange;
        int iF = ii - 1;
        int jF = jj - 1;
        int kF_dst = -kk;
        int kF_src = kk + 1;
        fh[(iF+1) + (jF+1)*fnx + (kF_dst+1)*fnx*fny] =
            fh[(iF+1) + (jF+1)*fnx + (kF_src+1)*fnx*fny] * SoA2;
    }
}

/* ---- ord=3 variants (for lopsided / kodis) ---- */

__global__ void kern_symbd_copy_interior_ord3(const double * __restrict__ func,
                                               double * __restrict__ fh)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int fnx = d_gp.fh3_nx, fny = d_gp.fh3_ny;
    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < d_gp.all;
         tid += blockDim.x * gridDim.x)
    {
        int i0 = tid % nx;
        int j0 = (tid / nx) % ny;
        int k0 = tid / (nx * ny);
        int iF = i0 + 1, jF = j0 + 1, kF = k0 + 1;
        fh[(iF+2) + (jF+2)*fnx + (kF+2)*fnx*fny] = func[tid];
    }
}

/* Fused symmetry pack (ord=3): fill full fh from interior func in one pass. */
__global__ void kern_symbd_pack_ord3(const double * __restrict__ func,
                                     double * __restrict__ fh,
                                     double SoA0, double SoA1, double SoA2)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1];
    const int fnx = d_gp.fh3_nx, fny = d_gp.fh3_ny, fnz = d_gp.fh3_nz;
    const int total = fnx * fny * fnz;

    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < total;
         tid += blockDim.x * gridDim.x)
    {
        int ii = tid % fnx;
        int jj = (tid / fnx) % fny;
        int kk = tid / (fnx * fny);

        int iF = ii - 2;   /* -2 .. nx */
        int jF = jj - 2;   /* -2 .. ny */
        int kF = kk - 2;   /* -2 .. nz */

        int siF = (iF <= 0) ? (1 - iF) : iF;  /* 1..nx */
        int sjF = (jF <= 0) ? (1 - jF) : jF;  /* 1..ny */
        int skF = (kF <= 0) ? (1 - kF) : kF;  /* 1..nz */

        double sign = 1.0;
        if (iF <= 0) sign *= SoA0;
        if (jF <= 0) sign *= SoA1;
        if (kF <= 0) sign *= SoA2;

        int src = (siF - 1) + (sjF - 1) * nx + (skF - 1) * nx * ny;
        fh[tid] = sign * func[src];
    }
}

__global__ void kern_symbd_ighost_ord3(double * __restrict__ fh, double SoA0)
{
    const int ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int fnx = d_gp.fh3_nx, fny = d_gp.fh3_ny;
    int total = ny * nz;
    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < total * 3;
         tid += blockDim.x * gridDim.x)
    {
        int ii = tid / total;
        int rem = tid % total;
        int j0 = rem % ny;
        int k0 = rem / ny;
        int jF = j0 + 1, kF = k0 + 1;
        int iF_dst = -ii;
        int iF_src = ii + 1;
        fh[(iF_dst+2) + (jF+2)*fnx + (kF+2)*fnx*fny] =
            fh[(iF_src+2) + (jF+2)*fnx + (kF+2)*fnx*fny] * SoA0;
    }
}

__global__ void kern_symbd_jghost_ord3(double * __restrict__ fh, double SoA1)
{
    const int nx = d_gp.ex[0], nz = d_gp.ex[2];
    const int fnx = d_gp.fh3_nx, fny = d_gp.fh3_ny;
    int irange = nx + 3;
    int total = irange * nz;
    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < total * 3;
         tid += blockDim.x * gridDim.x)
    {
        int jj = tid / total;
        int rem = tid % total;
        int ii = rem % irange;
        int k0 = rem / irange;
        int iF = ii - 2;
        int kF = k0 + 1;
        int jF_dst = -jj;
        int jF_src = jj + 1;
        fh[(iF+2) + (jF_dst+2)*fnx + (kF+2)*fnx*fny] =
            fh[(iF+2) + (jF_src+2)*fnx + (kF+2)*fnx*fny] * SoA1;
    }
}

__global__ void kern_symbd_kghost_ord3(double * __restrict__ fh, double SoA2)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1];
    const int fnx = d_gp.fh3_nx, fny = d_gp.fh3_ny;
    int irange = nx + 3;
    int jrange = ny + 3;
    int total = irange * jrange;
    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < total * 3;
         tid += blockDim.x * gridDim.x)
    {
        int kk = tid / total;
        int rem = tid % total;
        int ii = rem % irange;
        int jj = rem / irange;
        int iF = ii - 2;
        int jF = jj - 2;
        int kF_dst = -kk;
        int kF_src = kk + 1;
        fh[(iF+2) + (jF+2)*fnx + (kF_dst+2)*fnx*fny] =
            fh[(iF+2) + (jF+2)*fnx + (kF_src+2)*fnx*fny] * SoA2;
    }
}

/* ================================================================== */
/*  B. Stencil kernels                                                 */
/* ================================================================== */

/* ---- First derivatives (ord=2, 4th/2nd order) ---- */
__global__ __launch_bounds__(128, 4)
void kern_fderivs(const double * __restrict__ fh,
                  double * __restrict__ fx,
                  double * __restrict__ fy,
                  double * __restrict__ fz)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const int iminF = d_gp.iminF, jminF = d_gp.jminF, kminF = d_gp.kminF;

    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < d_gp.all;
         tid += blockDim.x * gridDim.x)
    {
        int i0 = tid % nx;
        int j0 = (tid / nx) % ny;
        int k0 = tid / (nx * ny);

        /* boundary points: leave as zero */
        if (i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2) {
            fx[tid] = 0.0; fy[tid] = 0.0; fz[tid] = 0.0;
            continue;
        }

        int iF = i0 + 1, jF = j0 + 1, 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 * (
                fh[idx_fh2(iF-2,jF,kF)] - 8.0*fh[idx_fh2(iF-1,jF,kF)]
              + 8.0*fh[idx_fh2(iF+1,jF,kF)] - fh[idx_fh2(iF+2,jF,kF)]);
            fy[tid] = d_gp.d12dy * (
                fh[idx_fh2(iF,jF-2,kF)] - 8.0*fh[idx_fh2(iF,jF-1,kF)]
              + 8.0*fh[idx_fh2(iF,jF+1,kF)] - fh[idx_fh2(iF,jF+2,kF)]);
            fz[tid] = d_gp.d12dz * (
                fh[idx_fh2(iF,jF,kF-2)] - 8.0*fh[idx_fh2(iF,jF,kF-1)]
              + 8.0*fh[idx_fh2(iF,jF,kF+1)] - fh[idx_fh2(iF,jF,kF+2)]);
        }
        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 * (
                -fh[idx_fh2(iF-1,jF,kF)] + fh[idx_fh2(iF+1,jF,kF)]);
            fy[tid] = d_gp.d2dy * (
                -fh[idx_fh2(iF,jF-1,kF)] + fh[idx_fh2(iF,jF+1,kF)]);
            fz[tid] = d_gp.d2dz * (
                -fh[idx_fh2(iF,jF,kF-1)] + fh[idx_fh2(iF,jF,kF+1)]);
        }
        else {
            fx[tid] = 0.0; fy[tid] = 0.0; fz[tid] = 0.0;
        }
    }
}

/* ---- Second derivatives (ord=2, 4th/2nd order) ---- */
__global__ __launch_bounds__(128, 4)
void kern_fdderivs(const double * __restrict__ fh,
                   double * __restrict__ fxx, double * __restrict__ fxy,
                   double * __restrict__ fxz, double * __restrict__ fyy,
                   double * __restrict__ fyz, double * __restrict__ fzz)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const int iminF = d_gp.iminF, jminF = d_gp.jminF, kminF = d_gp.kminF;

    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < d_gp.all;
         tid += blockDim.x * gridDim.x)
    {
        int i0 = tid % nx;
        int j0 = (tid / nx) % ny;
        int k0 = tid / (nx * ny);

        if (i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2) {
            fxx[tid]=0; fxy[tid]=0; fxz[tid]=0;
            fyy[tid]=0; fyz[tid]=0; fzz[tid]=0;
            continue;
        }

        int iF = i0+1, jF = j0+1, kF = k0+1;

        if ((iF+2)<=imaxF && (iF-2)>=iminF &&
            (jF+2)<=jmaxF && (jF-2)>=jminF &&
            (kF+2)<=kmaxF && (kF-2)>=kminF)
        {
            /* 4th-order diagonal */
            double c = fh[idx_fh2(iF,jF,kF)];
            fxx[tid] = d_gp.Fdxdx*(
                -fh[idx_fh2(iF-2,jF,kF)] + 16.0*fh[idx_fh2(iF-1,jF,kF)]
                -30.0*c + 16.0*fh[idx_fh2(iF+1,jF,kF)] - fh[idx_fh2(iF+2,jF,kF)]);
            fyy[tid] = d_gp.Fdydy*(
                -fh[idx_fh2(iF,jF-2,kF)] + 16.0*fh[idx_fh2(iF,jF-1,kF)]
                -30.0*c + 16.0*fh[idx_fh2(iF,jF+1,kF)] - fh[idx_fh2(iF,jF+2,kF)]);
            fzz[tid] = d_gp.Fdzdz*(
                -fh[idx_fh2(iF,jF,kF-2)] + 16.0*fh[idx_fh2(iF,jF,kF-1)]
                -30.0*c + 16.0*fh[idx_fh2(iF,jF,kF+1)] - fh[idx_fh2(iF,jF,kF+2)]);

            /* 4th-order cross: fxy */
            {
                double t_jm2 = fh[idx_fh2(iF-2,jF-2,kF)] - 8.0*fh[idx_fh2(iF-1,jF-2,kF)]
                             + 8.0*fh[idx_fh2(iF+1,jF-2,kF)] - fh[idx_fh2(iF+2,jF-2,kF)];
                double t_jm1 = fh[idx_fh2(iF-2,jF-1,kF)] - 8.0*fh[idx_fh2(iF-1,jF-1,kF)]
                             + 8.0*fh[idx_fh2(iF+1,jF-1,kF)] - fh[idx_fh2(iF+2,jF-1,kF)];
                double t_jp1 = fh[idx_fh2(iF-2,jF+1,kF)] - 8.0*fh[idx_fh2(iF-1,jF+1,kF)]
                             + 8.0*fh[idx_fh2(iF+1,jF+1,kF)] - fh[idx_fh2(iF+2,jF+1,kF)];
                double t_jp2 = fh[idx_fh2(iF-2,jF+2,kF)] - 8.0*fh[idx_fh2(iF-1,jF+2,kF)]
                             + 8.0*fh[idx_fh2(iF+1,jF+2,kF)] - fh[idx_fh2(iF+2,jF+2,kF)];
                fxy[tid] = d_gp.Fdxdy*(t_jm2 - 8.0*t_jm1 + 8.0*t_jp1 - t_jp2);
            }
            /* 4th-order cross: fxz */
            {
                double t_km2 = fh[idx_fh2(iF-2,jF,kF-2)] - 8.0*fh[idx_fh2(iF-1,jF,kF-2)]
                             + 8.0*fh[idx_fh2(iF+1,jF,kF-2)] - fh[idx_fh2(iF+2,jF,kF-2)];
                double t_km1 = fh[idx_fh2(iF-2,jF,kF-1)] - 8.0*fh[idx_fh2(iF-1,jF,kF-1)]
                             + 8.0*fh[idx_fh2(iF+1,jF,kF-1)] - fh[idx_fh2(iF+2,jF,kF-1)];
                double t_kp1 = fh[idx_fh2(iF-2,jF,kF+1)] - 8.0*fh[idx_fh2(iF-1,jF,kF+1)]
                             + 8.0*fh[idx_fh2(iF+1,jF,kF+1)] - fh[idx_fh2(iF+2,jF,kF+1)];
                double t_kp2 = fh[idx_fh2(iF-2,jF,kF+2)] - 8.0*fh[idx_fh2(iF-1,jF,kF+2)]
                             + 8.0*fh[idx_fh2(iF+1,jF,kF+2)] - fh[idx_fh2(iF+2,jF,kF+2)];
                fxz[tid] = d_gp.Fdxdz*(t_km2 - 8.0*t_km1 + 8.0*t_kp1 - t_kp2);
            }
            /* 4th-order cross: fyz */
            {
                double t_km2 = fh[idx_fh2(iF,jF-2,kF-2)] - 8.0*fh[idx_fh2(iF,jF-1,kF-2)]
                             + 8.0*fh[idx_fh2(iF,jF+1,kF-2)] - fh[idx_fh2(iF,jF+2,kF-2)];
                double t_km1 = fh[idx_fh2(iF,jF-2,kF-1)] - 8.0*fh[idx_fh2(iF,jF-1,kF-1)]
                             + 8.0*fh[idx_fh2(iF,jF+1,kF-1)] - fh[idx_fh2(iF,jF+2,kF-1)];
                double t_kp1 = fh[idx_fh2(iF,jF-2,kF+1)] - 8.0*fh[idx_fh2(iF,jF-1,kF+1)]
                             + 8.0*fh[idx_fh2(iF,jF+1,kF+1)] - fh[idx_fh2(iF,jF+2,kF+1)];
                double t_kp2 = fh[idx_fh2(iF,jF-2,kF+2)] - 8.0*fh[idx_fh2(iF,jF-1,kF+2)]
                             + 8.0*fh[idx_fh2(iF,jF+1,kF+2)] - fh[idx_fh2(iF,jF+2,kF+2)];
                fyz[tid] = d_gp.Fdydz*(t_km2 - 8.0*t_km1 + 8.0*t_kp1 - t_kp2);
            }
        }
        else if ((iF+1)<=imaxF && (iF-1)>=iminF &&
                 (jF+1)<=jmaxF && (jF-1)>=jminF &&
                 (kF+1)<=kmaxF && (kF-1)>=kminF)
        {
            double c = fh[idx_fh2(iF,jF,kF)];
            fxx[tid] = d_gp.Sdxdx*(fh[idx_fh2(iF-1,jF,kF)] - 2.0*c + fh[idx_fh2(iF+1,jF,kF)]);
            fyy[tid] = d_gp.Sdydy*(fh[idx_fh2(iF,jF-1,kF)] - 2.0*c + fh[idx_fh2(iF,jF+1,kF)]);
            fzz[tid] = d_gp.Sdzdz*(fh[idx_fh2(iF,jF,kF-1)] - 2.0*c + fh[idx_fh2(iF,jF,kF+1)]);
            fxy[tid] = d_gp.Sdxdy*(fh[idx_fh2(iF-1,jF-1,kF)] - fh[idx_fh2(iF+1,jF-1,kF)]
                                   -fh[idx_fh2(iF-1,jF+1,kF)] + fh[idx_fh2(iF+1,jF+1,kF)]);
            fxz[tid] = d_gp.Sdxdz*(fh[idx_fh2(iF-1,jF,kF-1)] - fh[idx_fh2(iF+1,jF,kF-1)]
                                   -fh[idx_fh2(iF-1,jF,kF+1)] + fh[idx_fh2(iF+1,jF,kF+1)]);
            fyz[tid] = d_gp.Sdydz*(fh[idx_fh2(iF,jF-1,kF-1)] - fh[idx_fh2(iF,jF+1,kF-1)]
                                   -fh[idx_fh2(iF,jF-1,kF+1)] + fh[idx_fh2(iF,jF+1,kF+1)]);
        }
        else {
            fxx[tid]=0; fxy[tid]=0; fxz[tid]=0;
            fyy[tid]=0; fyz[tid]=0; fzz[tid]=0;
        }
    }
}

/* ---- Lopsided (upwind advection) kernel ---- */
__global__ __launch_bounds__(128, 4)
void kern_lopsided(const double * __restrict__ fh,
                   double * __restrict__ f_rhs,
                   const double * __restrict__ Sfx,
                   const double * __restrict__ Sfy,
                   const double * __restrict__ Sfz)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int iminF = d_gp.iminF3, jminF = d_gp.jminF3, kminF = d_gp.kminF3;

    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < d_gp.all;
         tid += blockDim.x * gridDim.x)
    {
        int i0 = tid % nx;
        int j0 = (tid / nx) % ny;
        int k0 = tid / (nx * ny);

        if (i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2) continue;

        int iF = i0 + 1, jF = j0 + 1, kF = k0 + 1;
        double val = 0.0;

        /* --- x direction --- */
        double sfx = Sfx[tid];
        if (sfx > 0.0) {
            if (i0 <= nx - 4) {
                val += sfx * d_gp.d12dx * (
                    -3.0*fh[idx_fh3(iF-1,jF,kF)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF+1,jF,kF)] - 6.0*fh[idx_fh3(iF+2,jF,kF)]
                    + fh[idx_fh3(iF+3,jF,kF)]);
            } else if (i0 <= nx - 3) {
                val += sfx * d_gp.d12dx * (
                    fh[idx_fh3(iF-2,jF,kF)] - 8.0*fh[idx_fh3(iF-1,jF,kF)]
                    +8.0*fh[idx_fh3(iF+1,jF,kF)] - fh[idx_fh3(iF+2,jF,kF)]);
            } else if (i0 <= nx - 2) {
                val -= sfx * d_gp.d12dx * (
                    -3.0*fh[idx_fh3(iF+1,jF,kF)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF-1,jF,kF)] - 6.0*fh[idx_fh3(iF-2,jF,kF)]
                    + fh[idx_fh3(iF-3,jF,kF)]);
            }
        } else if (sfx < 0.0) {
            if ((i0 - 2) >= iminF) {
                val -= sfx * d_gp.d12dx * (
                    -3.0*fh[idx_fh3(iF+1,jF,kF)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF-1,jF,kF)] - 6.0*fh[idx_fh3(iF-2,jF,kF)]
                    + fh[idx_fh3(iF-3,jF,kF)]);
            } else if ((i0 - 1) >= iminF) {
                val += sfx * d_gp.d12dx * (
                    fh[idx_fh3(iF-2,jF,kF)] - 8.0*fh[idx_fh3(iF-1,jF,kF)]
                    +8.0*fh[idx_fh3(iF+1,jF,kF)] - fh[idx_fh3(iF+2,jF,kF)]);
            } else if (i0 >= iminF) {
                val += sfx * d_gp.d12dx * (
                    -3.0*fh[idx_fh3(iF-1,jF,kF)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF+1,jF,kF)] - 6.0*fh[idx_fh3(iF+2,jF,kF)]
                    + fh[idx_fh3(iF+3,jF,kF)]);
            }
        }

        /* --- y direction --- */
        double sfy = Sfy[tid];
        if (sfy > 0.0) {
            if (j0 <= ny - 4) {
                val += sfy * d_gp.d12dy * (
                    -3.0*fh[idx_fh3(iF,jF-1,kF)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF,jF+1,kF)] - 6.0*fh[idx_fh3(iF,jF+2,kF)]
                    + fh[idx_fh3(iF,jF+3,kF)]);
            } else if (j0 <= ny - 3) {
                val += sfy * d_gp.d12dy * (
                    fh[idx_fh3(iF,jF-2,kF)] - 8.0*fh[idx_fh3(iF,jF-1,kF)]
                    +8.0*fh[idx_fh3(iF,jF+1,kF)] - fh[idx_fh3(iF,jF+2,kF)]);
            } else if (j0 <= ny - 2) {
                val -= sfy * d_gp.d12dy * (
                    -3.0*fh[idx_fh3(iF,jF+1,kF)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF,jF-1,kF)] - 6.0*fh[idx_fh3(iF,jF-2,kF)]
                    + fh[idx_fh3(iF,jF-3,kF)]);
            }
        } else if (sfy < 0.0) {
            if ((j0 - 2) >= jminF) {
                val -= sfy * d_gp.d12dy * (
                    -3.0*fh[idx_fh3(iF,jF+1,kF)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF,jF-1,kF)] - 6.0*fh[idx_fh3(iF,jF-2,kF)]
                    + fh[idx_fh3(iF,jF-3,kF)]);
            } else if ((j0 - 1) >= jminF) {
                val += sfy * d_gp.d12dy * (
                    fh[idx_fh3(iF,jF-2,kF)] - 8.0*fh[idx_fh3(iF,jF-1,kF)]
                    +8.0*fh[idx_fh3(iF,jF+1,kF)] - fh[idx_fh3(iF,jF+2,kF)]);
            } else if (j0 >= jminF) {
                val += sfy * d_gp.d12dy * (
                    -3.0*fh[idx_fh3(iF,jF-1,kF)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF,jF+1,kF)] - 6.0*fh[idx_fh3(iF,jF+2,kF)]
                    + fh[idx_fh3(iF,jF+3,kF)]);
            }
        }

        /* --- z direction --- */
        double sfz = Sfz[tid];
        if (sfz > 0.0) {
            if (k0 <= nz - 4) {
                val += sfz * d_gp.d12dz * (
                    -3.0*fh[idx_fh3(iF,jF,kF-1)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF,jF,kF+1)] - 6.0*fh[idx_fh3(iF,jF,kF+2)]
                    + fh[idx_fh3(iF,jF,kF+3)]);
            } else if (k0 <= nz - 3) {
                val += sfz * d_gp.d12dz * (
                    fh[idx_fh3(iF,jF,kF-2)] - 8.0*fh[idx_fh3(iF,jF,kF-1)]
                    +8.0*fh[idx_fh3(iF,jF,kF+1)] - fh[idx_fh3(iF,jF,kF+2)]);
            } else if (k0 <= nz - 2) {
                val -= sfz * d_gp.d12dz * (
                    -3.0*fh[idx_fh3(iF,jF,kF+1)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF,jF,kF-1)] - 6.0*fh[idx_fh3(iF,jF,kF-2)]
                    + fh[idx_fh3(iF,jF,kF-3)]);
            }
        } else if (sfz < 0.0) {
            if ((k0 - 2) >= kminF) {
                val -= sfz * d_gp.d12dz * (
                    -3.0*fh[idx_fh3(iF,jF,kF+1)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF,jF,kF-1)] - 6.0*fh[idx_fh3(iF,jF,kF-2)]
                    + fh[idx_fh3(iF,jF,kF-3)]);
            } else if ((k0 - 1) >= kminF) {
                val += sfz * d_gp.d12dz * (
                    fh[idx_fh3(iF,jF,kF-2)] - 8.0*fh[idx_fh3(iF,jF,kF-1)]
                    +8.0*fh[idx_fh3(iF,jF,kF+1)] - fh[idx_fh3(iF,jF,kF+2)]);
            } else if (k0 >= kminF) {
                val += sfz * d_gp.d12dz * (
                    -3.0*fh[idx_fh3(iF,jF,kF-1)] - 10.0*fh[idx_fh3(iF,jF,kF)]
                    +18.0*fh[idx_fh3(iF,jF,kF+1)] - 6.0*fh[idx_fh3(iF,jF,kF+2)]
                    + fh[idx_fh3(iF,jF,kF+3)]);
            }
        }

        f_rhs[tid] += val;
    }
}

/* ---- KO dissipation kernel (ord=3, 6th-order) ---- */
__global__ __launch_bounds__(128, 4)
void kern_kodis(const double * __restrict__ fh,
                double * __restrict__ f_rhs,
                double eps_val)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int iminF = d_gp.iminF3, jminF = d_gp.jminF3, kminF = d_gp.kminF3;
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const double cof = 64.0;

    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < d_gp.all;
         tid += blockDim.x * gridDim.x)
    {
        int i0 = tid % nx;
        int j0 = (tid / nx) % ny;
        int k0 = tid / (nx * ny);
        int iF = i0 + 1, jF = j0 + 1, kF = k0 + 1;

        if ((iF-3) >= iminF && (iF+3) <= imaxF &&
            (jF-3) >= jminF && (jF+3) <= jmaxF &&
            (kF-3) >= kminF && (kF+3) <= kmaxF)
        {
            double Dx = (fh[idx_fh3(iF-3,jF,kF)] + fh[idx_fh3(iF+3,jF,kF)])
                - 6.0*(fh[idx_fh3(iF-2,jF,kF)] + fh[idx_fh3(iF+2,jF,kF)])
                +15.0*(fh[idx_fh3(iF-1,jF,kF)] + fh[idx_fh3(iF+1,jF,kF)])
                -20.0* fh[idx_fh3(iF,jF,kF)];
            Dx /= d_gp.dX;

            double Dy = (fh[idx_fh3(iF,jF-3,kF)] + fh[idx_fh3(iF,jF+3,kF)])
                - 6.0*(fh[idx_fh3(iF,jF-2,kF)] + fh[idx_fh3(iF,jF+2,kF)])
                +15.0*(fh[idx_fh3(iF,jF-1,kF)] + fh[idx_fh3(iF,jF+1,kF)])
                -20.0* fh[idx_fh3(iF,jF,kF)];
            Dy /= d_gp.dY;

            double Dz = (fh[idx_fh3(iF,jF,kF-3)] + fh[idx_fh3(iF,jF,kF+3)])
                - 6.0*(fh[idx_fh3(iF,jF,kF-2)] + fh[idx_fh3(iF,jF,kF+2)])
                +15.0*(fh[idx_fh3(iF,jF,kF-1)] + fh[idx_fh3(iF,jF,kF+1)])
                -20.0* fh[idx_fh3(iF,jF,kF)];
            Dz /= d_gp.dZ;

            f_rhs[tid] += (eps_val / cof) * (Dx + Dy + Dz);
        }
    }
}

/* ================================================================== */
/*  Host wrapper helpers                                               */
/* ================================================================== */
struct LopsidedKodisTables {
    const double *adv_fields[BSSN_LK_FIELD_COUNT];
    const double *ko_fields[BSSN_LK_FIELD_COUNT];
    double *rhs_fields[BSSN_LK_FIELD_COUNT];
    int soa_signs[3 * BSSN_LK_FIELD_COUNT];
};

struct FDerivTables {
    const double *src_fields[BSSN_STATE_COUNT];
    double *fx_fields[BSSN_STATE_COUNT];
    double *fy_fields[BSSN_STATE_COUNT];
    double *fz_fields[BSSN_STATE_COUNT];
    int soa_signs[3 * BSSN_STATE_COUNT];
};

struct FDDerivTables {
    const double *src_fields[BSSN_STATE_COUNT];
    double *fxx_fields[BSSN_STATE_COUNT];
    double *fxy_fields[BSSN_STATE_COUNT];
    double *fxz_fields[BSSN_STATE_COUNT];
    double *fyy_fields[BSSN_STATE_COUNT];
    double *fyz_fields[BSSN_STATE_COUNT];
    double *fzz_fields[BSSN_STATE_COUNT];
    int soa_signs[3 * BSSN_STATE_COUNT];
};

static constexpr int PHASE10_METRIC_FIELD_COUNT = 6;

struct Phase10RicciTables {
    const double *src_fields[PHASE10_METRIC_FIELD_COUNT];
    double *dst_fields[PHASE10_METRIC_FIELD_COUNT];
    int soa_signs[3 * PHASE10_METRIC_FIELD_COUNT];
};

struct Rk4FinalizeTables {
    const double *f0_fields[BSSN_STATE_COUNT];
    double *rhs_fields[BSSN_STATE_COUNT];
    double *accum_fields[BSSN_STATE_COUNT];
};

struct PatchBoundaryTables {
    const double *src_fields[BSSN_STATE_COUNT];
    double *dst_fields[BSSN_STATE_COUNT];
};

static const int BLK = 128;
static inline int grid(size_t n) {
    if (n == 0) return 1;
    size_t g = (n + BLK - 1) / BLK;
    if (g > 2147483647u) g = 2147483647u;
    return (int)g;
}

__global__ __launch_bounds__(128, 4)
void kern_fderivs_batched(FDerivTables tables, int field_count)
{
    const int field = blockIdx.y;
    if (field >= field_count) return;

    const double *src = tables.src_fields[field];
    double *fx = tables.fx_fields[field];
    double *fy = tables.fy_fields[field];
    double *fz = tables.fz_fields[field];
    const int SoA0 = tables.soa_signs[3 * field + 0];
    const int SoA1 = tables.soa_signs[3 * field + 1];
    const int SoA2 = tables.soa_signs[3 * field + 2];
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const int iminF = d_gp.iminF, jminF = d_gp.jminF, kminF = d_gp.kminF;

    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid >= d_gp.all) return;

    const int i0 = tid % nx;
    const int j0 = (tid / nx) % ny;
    const int k0 = tid / (nx * ny);

    if (i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2) {
        fx[tid] = 0.0;
        fy[tid] = 0.0;
        fz[tid] = 0.0;
        return;
    }

    const int iF = i0 + 1;
    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;
    }
}

__global__ __launch_bounds__(128, 4)
void kern_fdderivs_batched(FDDerivTables tables, int field_count)
{
    const int field = blockIdx.y;
    if (field >= field_count) return;

    const double *src = tables.src_fields[field];
    double *fxx = tables.fxx_fields[field];
    double *fxy = tables.fxy_fields[field];
    double *fxz = tables.fxz_fields[field];
    double *fyy = tables.fyy_fields[field];
    double *fyz = tables.fyz_fields[field];
    double *fzz = tables.fzz_fields[field];
    const int SoA0 = tables.soa_signs[3 * field + 0];
    const int SoA1 = tables.soa_signs[3 * field + 1];
    const int SoA2 = tables.soa_signs[3 * field + 2];
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const int iminF = d_gp.iminF, jminF = d_gp.jminF, kminF = d_gp.kminF;

    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid >= d_gp.all) return;

    const int i0 = tid % nx;
    const int j0 = (tid / nx) % ny;
    const int k0 = tid / (nx * ny);

    if (i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2) {
        fxx[tid] = 0.0; fxy[tid] = 0.0; fxz[tid] = 0.0;
        fyy[tid] = 0.0; fyz[tid] = 0.0; fzz[tid] = 0.0;
        return;
    }

    const int iF = i0 + 1;
    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)
    {
        const double c = fetch_sym_ord2_direct(src, iF, jF, kF, SoA0, SoA1, SoA2);
        fxx[tid] = d_gp.Fdxdx * (
            -fetch_sym_ord2_direct(src, iF - 2, jF, kF, SoA0, SoA1, SoA2)
            +16.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
            -30.0 * c
            +16.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
            -       fetch_sym_ord2_direct(src, iF + 2, jF, kF, SoA0, SoA1, SoA2));
        fyy[tid] = d_gp.Fdydy * (
            -fetch_sym_ord2_direct(src, iF, jF - 2, kF, SoA0, SoA1, SoA2)
            +16.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
            -30.0 * c
            +16.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
            -       fetch_sym_ord2_direct(src, iF, jF + 2, kF, SoA0, SoA1, SoA2));
        fzz[tid] = d_gp.Fdzdz * (
            -fetch_sym_ord2_direct(src, iF, jF, kF - 2, SoA0, SoA1, SoA2)
            +16.0 * fetch_sym_ord2_direct(src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
            -30.0 * c
            +16.0 * fetch_sym_ord2_direct(src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
            -       fetch_sym_ord2_direct(src, iF, jF, kF + 2, SoA0, SoA1, SoA2));

        const double t_jm2 =
            fetch_sym_ord2_direct(src, iF - 2, jF - 2, kF, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF - 2, kF, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF - 2, kF, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF + 2, jF - 2, kF, SoA0, SoA1, SoA2);
        const double t_jm1 =
            fetch_sym_ord2_direct(src, iF - 2, jF - 1, kF, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF - 1, kF, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF - 1, kF, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF + 2, jF - 1, kF, SoA0, SoA1, SoA2);
        const double t_jp1 =
            fetch_sym_ord2_direct(src, iF - 2, jF + 1, kF, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF + 1, kF, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF + 1, kF, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF + 2, jF + 1, kF, SoA0, SoA1, SoA2);
        const double t_jp2 =
            fetch_sym_ord2_direct(src, iF - 2, jF + 2, kF, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF + 2, kF, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF + 2, kF, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF + 2, jF + 2, kF, SoA0, SoA1, SoA2);
        fxy[tid] = d_gp.Fdxdy * (t_jm2 - 8.0 * t_jm1 + 8.0 * t_jp1 - t_jp2);

        const double t_km2_x =
            fetch_sym_ord2_direct(src, iF - 2, jF, kF - 2, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF - 2, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF - 2, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF + 2, jF, kF - 2, SoA0, SoA1, SoA2);
        const double t_km1_x =
            fetch_sym_ord2_direct(src, iF - 2, jF, kF - 1, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF - 1, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF - 1, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF + 2, jF, kF - 1, SoA0, SoA1, SoA2);
        const double t_kp1_x =
            fetch_sym_ord2_direct(src, iF - 2, jF, kF + 1, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF + 1, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF + 1, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF + 2, jF, kF + 1, SoA0, SoA1, SoA2);
        const double t_kp2_x =
            fetch_sym_ord2_direct(src, iF - 2, jF, kF + 2, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF + 2, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF + 2, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF + 2, jF, kF + 2, SoA0, SoA1, SoA2);
        fxz[tid] = d_gp.Fdxdz * (t_km2_x - 8.0 * t_km1_x + 8.0 * t_kp1_x - t_kp2_x);

        const double t_km2_y =
            fetch_sym_ord2_direct(src, iF, jF - 2, kF - 2, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF - 2, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF - 2, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF, jF + 2, kF - 2, SoA0, SoA1, SoA2);
        const double t_km1_y =
            fetch_sym_ord2_direct(src, iF, jF - 2, kF - 1, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF - 1, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF - 1, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF, jF + 2, kF - 1, SoA0, SoA1, SoA2);
        const double t_kp1_y =
            fetch_sym_ord2_direct(src, iF, jF - 2, kF + 1, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF + 1, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF + 1, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF, jF + 2, kF + 1, SoA0, SoA1, SoA2);
        const double t_kp2_y =
            fetch_sym_ord2_direct(src, iF, jF - 2, kF + 2, SoA0, SoA1, SoA2)
          - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF + 2, SoA0, SoA1, SoA2)
          + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF + 2, SoA0, SoA1, SoA2)
          -       fetch_sym_ord2_direct(src, iF, jF + 2, kF + 2, SoA0, SoA1, SoA2);
        fyz[tid] = d_gp.Fdydz * (t_km2_y - 8.0 * t_km1_y + 8.0 * t_kp1_y - t_kp2_y);
    }
    else if ((iF + 1) <= imaxF && (iF - 1) >= iminF &&
             (jF + 1) <= jmaxF && (jF - 1) >= jminF &&
             (kF + 1) <= kmaxF && (kF - 1) >= kminF)
    {
        const double c = fetch_sym_ord2_direct(src, iF, jF, kF, SoA0, SoA1, SoA2);
        fxx[tid] = d_gp.Sdxdx * (
            fetch_sym_ord2_direct(src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
          - 2.0 * c
          + fetch_sym_ord2_direct(src, iF + 1, jF, kF, SoA0, SoA1, SoA2));
        fyy[tid] = d_gp.Sdydy * (
            fetch_sym_ord2_direct(src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
          - 2.0 * c
          + fetch_sym_ord2_direct(src, iF, jF + 1, kF, SoA0, SoA1, SoA2));
        fzz[tid] = d_gp.Sdzdz * (
            fetch_sym_ord2_direct(src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
          - 2.0 * c
          + fetch_sym_ord2_direct(src, iF, jF, kF + 1, SoA0, SoA1, SoA2));
        fxy[tid] = d_gp.Sdxdy * (
            fetch_sym_ord2_direct(src, iF - 1, jF - 1, kF, SoA0, SoA1, SoA2)
          - fetch_sym_ord2_direct(src, iF + 1, jF - 1, kF, SoA0, SoA1, SoA2)
          - fetch_sym_ord2_direct(src, iF - 1, jF + 1, kF, SoA0, SoA1, SoA2)
          + fetch_sym_ord2_direct(src, iF + 1, jF + 1, kF, SoA0, SoA1, SoA2));
        fxz[tid] = d_gp.Sdxdz * (
            fetch_sym_ord2_direct(src, iF - 1, jF, kF - 1, SoA0, SoA1, SoA2)
          - fetch_sym_ord2_direct(src, iF + 1, jF, kF - 1, SoA0, SoA1, SoA2)
          - fetch_sym_ord2_direct(src, iF - 1, jF, kF + 1, SoA0, SoA1, SoA2)
          + fetch_sym_ord2_direct(src, iF + 1, jF, kF + 1, SoA0, SoA1, SoA2));
        fyz[tid] = d_gp.Sdydz * (
            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)
          - 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));
    }
    else {
        fxx[tid] = 0.0; fxy[tid] = 0.0; fxz[tid] = 0.0;
        fyy[tid] = 0.0; fyz[tid] = 0.0; fzz[tid] = 0.0;
    }
}

/* 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)
{
    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;
    const size_t nz = (size_t)g_buf.prev_nz;
    const size_t w_pack = (nx + 2ull) * (ny + 2ull) * (nz + 2ull);

    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);
}

/* symmetry_bd on GPU for ord=2, then launch fdderivs kernel */
static void gpu_fdderivs(double *d_f,
                         double *d_fxx, double *d_fxy, double *d_fxz,
                         double *d_fyy, double *d_fyz, double *d_fzz,
                         double SoA0, double SoA1, double SoA2, int all)
{
    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;
    const size_t nz = (size_t)g_buf.prev_nz;
    const size_t w_pack = (nx + 2ull) * (ny + 2ull) * (nz + 2ull);

    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);
}

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)
{
    if (field_count <= 0) return;
    FDerivTables tables = {};
    for (int i = 0; i < field_count; ++i) {
        tables.src_fields[i] = src_fields[i];
        tables.fx_fields[i] = fx_fields[i];
        tables.fy_fields[i] = fy_fields[i];
        tables.fz_fields[i] = fz_fields[i];
        tables.soa_signs[3 * i + 0] = soa_signs[3 * i + 0];
        tables.soa_signs[3 * i + 1] = soa_signs[3 * i + 1];
        tables.soa_signs[3 * i + 2] = soa_signs[3 * i + 2];
    }
    dim3 launch_grid((unsigned int)grid((size_t)all), (unsigned int)field_count);
    kern_fderivs_batched<<<launch_grid, BLK>>>(tables, field_count);
}

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)
{
    if (field_count <= 0) return;
    FDDerivTables tables = {};
    for (int i = 0; i < field_count; ++i) {
        tables.src_fields[i] = src_fields[i];
        tables.fxx_fields[i] = fxx_fields[i];
        tables.fxy_fields[i] = fxy_fields[i];
        tables.fxz_fields[i] = fxz_fields[i];
        tables.fyy_fields[i] = fyy_fields[i];
        tables.fyz_fields[i] = fyz_fields[i];
        tables.fzz_fields[i] = fzz_fields[i];
        tables.soa_signs[3 * i + 0] = soa_signs[3 * i + 0];
        tables.soa_signs[3 * i + 1] = soa_signs[3 * i + 1];
        tables.soa_signs[3 * i + 2] = soa_signs[3 * i + 2];
    }
    dim3 launch_grid((unsigned int)grid((size_t)all), (unsigned int)field_count);
    kern_fdderivs_batched<<<launch_grid, BLK>>>(tables, field_count);
}

__global__ __launch_bounds__(128, 4)
void kern_phase10_ricci_batched(const double * __restrict__ gupxx,
                                const double * __restrict__ gupxy,
                                const double * __restrict__ gupxz,
                                const double * __restrict__ gupyy,
                                const double * __restrict__ gupyz,
                                const double * __restrict__ gupzz,
                                Phase10RicciTables tables)
{
    const int field = blockIdx.y;
    if (field >= PHASE10_METRIC_FIELD_COUNT) return;

    const double *src = tables.src_fields[field];
    double *dst = tables.dst_fields[field];
    const int SoA0 = tables.soa_signs[3 * field + 0];
    const int SoA1 = tables.soa_signs[3 * field + 1];
    const int SoA2 = tables.soa_signs[3 * field + 2];
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const int iminF = d_gp.iminF, jminF = d_gp.jminF, kminF = d_gp.kminF;
    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid >= d_gp.all) return;

    const int i0 = tid % nx;
    const int j0 = (tid / nx) % ny;
    const int k0 = tid / (nx * ny);

    double fxx = 0.0, fxy = 0.0, fxz = 0.0;
    double fyy = 0.0, fyz = 0.0, fzz = 0.0;

    if (!(i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2)) {
        const int iF = i0 + 1;
        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)
        {
            const double c = fetch_sym_ord2_direct(src, iF, jF, kF, SoA0, SoA1, SoA2);
            fxx = d_gp.Fdxdx * (
                -fetch_sym_ord2_direct(src, iF - 2, jF, kF, SoA0, SoA1, SoA2)
                +16.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
                -30.0 * c
                +16.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
                -       fetch_sym_ord2_direct(src, iF + 2, jF, kF, SoA0, SoA1, SoA2));
            fyy = d_gp.Fdydy * (
                -fetch_sym_ord2_direct(src, iF, jF - 2, kF, SoA0, SoA1, SoA2)
                +16.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
                -30.0 * c
                +16.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
                -       fetch_sym_ord2_direct(src, iF, jF + 2, kF, SoA0, SoA1, SoA2));
            fzz = d_gp.Fdzdz * (
                -fetch_sym_ord2_direct(src, iF, jF, kF - 2, SoA0, SoA1, SoA2)
                +16.0 * fetch_sym_ord2_direct(src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
                -30.0 * c
                +16.0 * fetch_sym_ord2_direct(src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
                -       fetch_sym_ord2_direct(src, iF, jF, kF + 2, SoA0, SoA1, SoA2));

            const double t_jm2 =
                fetch_sym_ord2_direct(src, iF - 2, jF - 2, kF, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF - 2, kF, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF - 2, kF, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF + 2, jF - 2, kF, SoA0, SoA1, SoA2);
            const double t_jm1 =
                fetch_sym_ord2_direct(src, iF - 2, jF - 1, kF, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF - 1, kF, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF - 1, kF, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF + 2, jF - 1, kF, SoA0, SoA1, SoA2);
            const double t_jp1 =
                fetch_sym_ord2_direct(src, iF - 2, jF + 1, kF, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF + 1, kF, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF + 1, kF, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF + 2, jF + 1, kF, SoA0, SoA1, SoA2);
            const double t_jp2 =
                fetch_sym_ord2_direct(src, iF - 2, jF + 2, kF, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF + 2, kF, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF + 2, kF, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF + 2, jF + 2, kF, SoA0, SoA1, SoA2);
            fxy = d_gp.Fdxdy * (t_jm2 - 8.0 * t_jm1 + 8.0 * t_jp1 - t_jp2);

            const double t_km2_x =
                fetch_sym_ord2_direct(src, iF - 2, jF, kF - 2, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF - 2, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF - 2, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF + 2, jF, kF - 2, SoA0, SoA1, SoA2);
            const double t_km1_x =
                fetch_sym_ord2_direct(src, iF - 2, jF, kF - 1, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF - 1, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF - 1, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF + 2, jF, kF - 1, SoA0, SoA1, SoA2);
            const double t_kp1_x =
                fetch_sym_ord2_direct(src, iF - 2, jF, kF + 1, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF + 1, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF + 1, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF + 2, jF, kF + 1, SoA0, SoA1, SoA2);
            const double t_kp2_x =
                fetch_sym_ord2_direct(src, iF - 2, jF, kF + 2, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF - 1, jF, kF + 2, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF + 1, jF, kF + 2, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF + 2, jF, kF + 2, SoA0, SoA1, SoA2);
            fxz = d_gp.Fdxdz * (t_km2_x - 8.0 * t_km1_x + 8.0 * t_kp1_x - t_kp2_x);

            const double t_km2_y =
                fetch_sym_ord2_direct(src, iF, jF - 2, kF - 2, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF - 2, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF - 2, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF, jF + 2, kF - 2, SoA0, SoA1, SoA2);
            const double t_km1_y =
                fetch_sym_ord2_direct(src, iF, jF - 2, kF - 1, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF - 1, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF - 1, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF, jF + 2, kF - 1, SoA0, SoA1, SoA2);
            const double t_kp1_y =
                fetch_sym_ord2_direct(src, iF, jF - 2, kF + 1, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF + 1, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF + 1, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF, jF + 2, kF + 1, SoA0, SoA1, SoA2);
            const double t_kp2_y =
                fetch_sym_ord2_direct(src, iF, jF - 2, kF + 2, SoA0, SoA1, SoA2)
              - 8.0 * fetch_sym_ord2_direct(src, iF, jF - 1, kF + 2, SoA0, SoA1, SoA2)
              + 8.0 * fetch_sym_ord2_direct(src, iF, jF + 1, kF + 2, SoA0, SoA1, SoA2)
              -       fetch_sym_ord2_direct(src, iF, jF + 2, kF + 2, SoA0, SoA1, SoA2);
            fyz = d_gp.Fdydz * (t_km2_y - 8.0 * t_km1_y + 8.0 * t_kp1_y - t_kp2_y);
        }
        else if ((iF + 1) <= imaxF && (iF - 1) >= iminF &&
                 (jF + 1) <= jmaxF && (jF - 1) >= jminF &&
                 (kF + 1) <= kmaxF && (kF - 1) >= kminF)
        {
            const double c = fetch_sym_ord2_direct(src, iF, jF, kF, SoA0, SoA1, SoA2);
            fxx = d_gp.Sdxdx * (
                fetch_sym_ord2_direct(src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
              - 2.0 * c
              + fetch_sym_ord2_direct(src, iF + 1, jF, kF, SoA0, SoA1, SoA2));
            fyy = d_gp.Sdydy * (
                fetch_sym_ord2_direct(src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
              - 2.0 * c
              + fetch_sym_ord2_direct(src, iF, jF + 1, kF, SoA0, SoA1, SoA2));
            fzz = d_gp.Sdzdz * (
                fetch_sym_ord2_direct(src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
              - 2.0 * c
              + fetch_sym_ord2_direct(src, iF, jF, kF + 1, SoA0, SoA1, SoA2));
            fxy = d_gp.Sdxdy * (
                fetch_sym_ord2_direct(src, iF - 1, jF - 1, kF, SoA0, SoA1, SoA2)
              - fetch_sym_ord2_direct(src, iF + 1, jF - 1, kF, SoA0, SoA1, SoA2)
              - fetch_sym_ord2_direct(src, iF - 1, jF + 1, kF, SoA0, SoA1, SoA2)
              + fetch_sym_ord2_direct(src, iF + 1, jF + 1, kF, SoA0, SoA1, SoA2));
            fxz = d_gp.Sdxdz * (
                fetch_sym_ord2_direct(src, iF - 1, jF, kF - 1, SoA0, SoA1, SoA2)
              - fetch_sym_ord2_direct(src, iF + 1, jF, kF - 1, SoA0, SoA1, SoA2)
              - fetch_sym_ord2_direct(src, iF - 1, jF, kF + 1, SoA0, SoA1, SoA2)
              + fetch_sym_ord2_direct(src, iF + 1, jF, kF + 1, SoA0, SoA1, SoA2));
            fyz = d_gp.Sdydz * (
                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)
              - 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));
        }
    }

    dst[tid] = gupxx[tid] * fxx + gupyy[tid] * fyy + gupzz[tid] * fzz
             + 2.0 * (gupxy[tid] * fxy + gupxz[tid] * fxz + gupyz[tid] * fyz);
}

static void gpu_phase10_ricci_batch(const double *gupxx,
                                    const double *gupxy,
                                    const double *gupxz,
                                    const double *gupyy,
                                    const double *gupyz,
                                    const double *gupzz,
                                    double *const *src_fields,
                                    double *const *dst_fields,
                                    const int *soa_signs,
                                    int all)
{
    Phase10RicciTables tables = {};
    for (int i = 0; i < PHASE10_METRIC_FIELD_COUNT; ++i) {
        tables.src_fields[i] = src_fields[i];
        tables.dst_fields[i] = dst_fields[i];
        tables.soa_signs[3 * i + 0] = soa_signs[3 * i + 0];
        tables.soa_signs[3 * i + 1] = soa_signs[3 * i + 1];
        tables.soa_signs[3 * i + 2] = soa_signs[3 * i + 2];
    }
    dim3 launch_grid((unsigned int)grid((size_t)all), (unsigned int)PHASE10_METRIC_FIELD_COUNT);
    kern_phase10_ricci_batched<<<launch_grid, BLK>>>(
        gupxx, gupxy, gupxz, gupyy, gupyz, gupzz, tables);
}

__global__ __launch_bounds__(128, 4)
void kern_phase14_lap_chi_derivs(const double * __restrict__ Lap,
                                 const double * __restrict__ chi,
                                 double * __restrict__ fxx,
                                 double * __restrict__ fxy,
                                 double * __restrict__ fxz,
                                 double * __restrict__ fyy,
                                 double * __restrict__ fyz,
                                 double * __restrict__ fzz,
                                 double * __restrict__ chix_out,
                                 double * __restrict__ chiy_out,
                                 double * __restrict__ chiz_out)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const int iminF = d_gp.iminF, jminF = d_gp.jminF, kminF = d_gp.kminF;
    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid >= d_gp.all) return;

    const int i0 = tid % nx;
    const int j0 = (tid / nx) % ny;
    const int k0 = tid / (nx * ny);

    if (i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2) {
        fxx[tid] = 0.0; fxy[tid] = 0.0; fxz[tid] = 0.0;
        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;
        return;
    }

    const int iF = i0 + 1;
    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)
    {
        const double lap_c = fetch_sym_ord2_direct(Lap, iF, jF, kF, 1, 1, 1);
        fxx[tid] = d_gp.Fdxdx * (
            -fetch_sym_ord2_direct(Lap, iF - 2, jF, kF, 1, 1, 1)
            +16.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF, kF, 1, 1, 1)
            -30.0 * lap_c
            +16.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF, kF, 1, 1, 1)
            -       fetch_sym_ord2_direct(Lap, iF + 2, jF, kF, 1, 1, 1));
        fyy[tid] = d_gp.Fdydy * (
            -fetch_sym_ord2_direct(Lap, iF, jF - 2, kF, 1, 1, 1)
            +16.0 * fetch_sym_ord2_direct(Lap, iF, jF - 1, kF, 1, 1, 1)
            -30.0 * lap_c
            +16.0 * fetch_sym_ord2_direct(Lap, iF, jF + 1, kF, 1, 1, 1)
            -       fetch_sym_ord2_direct(Lap, iF, jF + 2, kF, 1, 1, 1));
        fzz[tid] = d_gp.Fdzdz * (
            -fetch_sym_ord2_direct(Lap, iF, jF, kF - 2, 1, 1, 1)
            +16.0 * fetch_sym_ord2_direct(Lap, iF, jF, kF - 1, 1, 1, 1)
            -30.0 * lap_c
            +16.0 * fetch_sym_ord2_direct(Lap, iF, jF, kF + 1, 1, 1, 1)
            -       fetch_sym_ord2_direct(Lap, iF, jF, kF + 2, 1, 1, 1));

        const double t_jm2 =
            fetch_sym_ord2_direct(Lap, iF - 2, jF - 2, kF, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF - 2, kF, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF - 2, kF, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF + 2, jF - 2, kF, 1, 1, 1);
        const double t_jm1 =
            fetch_sym_ord2_direct(Lap, iF - 2, jF - 1, kF, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF - 1, kF, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF - 1, kF, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF + 2, jF - 1, kF, 1, 1, 1);
        const double t_jp1 =
            fetch_sym_ord2_direct(Lap, iF - 2, jF + 1, kF, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF + 1, kF, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF + 1, kF, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF + 2, jF + 1, kF, 1, 1, 1);
        const double t_jp2 =
            fetch_sym_ord2_direct(Lap, iF - 2, jF + 2, kF, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF + 2, kF, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF + 2, kF, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF + 2, jF + 2, kF, 1, 1, 1);
        fxy[tid] = d_gp.Fdxdy * (t_jm2 - 8.0 * t_jm1 + 8.0 * t_jp1 - t_jp2);

        const double t_km2_x =
            fetch_sym_ord2_direct(Lap, iF - 2, jF, kF - 2, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF, kF - 2, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF, kF - 2, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF + 2, jF, kF - 2, 1, 1, 1);
        const double t_km1_x =
            fetch_sym_ord2_direct(Lap, iF - 2, jF, kF - 1, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF, kF - 1, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF, kF - 1, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF + 2, jF, kF - 1, 1, 1, 1);
        const double t_kp1_x =
            fetch_sym_ord2_direct(Lap, iF - 2, jF, kF + 1, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF, kF + 1, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF, kF + 1, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF + 2, jF, kF + 1, 1, 1, 1);
        const double t_kp2_x =
            fetch_sym_ord2_direct(Lap, iF - 2, jF, kF + 2, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF - 1, jF, kF + 2, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF + 1, jF, kF + 2, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF + 2, jF, kF + 2, 1, 1, 1);
        fxz[tid] = d_gp.Fdxdz * (t_km2_x - 8.0 * t_km1_x + 8.0 * t_kp1_x - t_kp2_x);

        const double t_km2_y =
            fetch_sym_ord2_direct(Lap, iF, jF - 2, kF - 2, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF, jF - 1, kF - 2, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF, jF + 1, kF - 2, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF, jF + 2, kF - 2, 1, 1, 1);
        const double t_km1_y =
            fetch_sym_ord2_direct(Lap, iF, jF - 2, kF - 1, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF, jF - 1, kF - 1, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF, jF + 1, kF - 1, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF, jF + 2, kF - 1, 1, 1, 1);
        const double t_kp1_y =
            fetch_sym_ord2_direct(Lap, iF, jF - 2, kF + 1, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF, jF - 1, kF + 1, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF, jF + 1, kF + 1, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF, jF + 2, kF + 1, 1, 1, 1);
        const double t_kp2_y =
            fetch_sym_ord2_direct(Lap, iF, jF - 2, kF + 2, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(Lap, iF, jF - 1, kF + 2, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(Lap, iF, jF + 1, kF + 2, 1, 1, 1)
          -       fetch_sym_ord2_direct(Lap, iF, jF + 2, kF + 2, 1, 1, 1);
        fyz[tid] = d_gp.Fdydz * (t_km2_y - 8.0 * t_km1_y + 8.0 * t_kp1_y - t_kp2_y);

        chix_out[tid] = d_gp.d12dx * (
            fetch_sym_ord2_direct(chi, iF - 2, jF, kF, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF, kF, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF, kF, 1, 1, 1)
          -       fetch_sym_ord2_direct(chi, iF + 2, jF, kF, 1, 1, 1));
        chiy_out[tid] = d_gp.d12dy * (
            fetch_sym_ord2_direct(chi, iF, jF - 2, kF, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(chi, iF, jF - 1, kF, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(chi, iF, jF + 1, kF, 1, 1, 1)
          -       fetch_sym_ord2_direct(chi, iF, jF + 2, kF, 1, 1, 1));
        chiz_out[tid] = d_gp.d12dz * (
            fetch_sym_ord2_direct(chi, iF, jF, kF - 2, 1, 1, 1)
          - 8.0 * fetch_sym_ord2_direct(chi, iF, jF, kF - 1, 1, 1, 1)
          + 8.0 * fetch_sym_ord2_direct(chi, iF, jF, kF + 1, 1, 1, 1)
          -       fetch_sym_ord2_direct(chi, iF, jF, kF + 2, 1, 1, 1));
    }
    else if ((iF + 1) <= imaxF && (iF - 1) >= iminF &&
             (jF + 1) <= jmaxF && (jF - 1) >= jminF &&
             (kF + 1) <= kmaxF && (kF - 1) >= kminF)
    {
        const double lap_c = fetch_sym_ord2_direct(Lap, iF, jF, kF, 1, 1, 1);
        fxx[tid] = d_gp.Sdxdx * (
            fetch_sym_ord2_direct(Lap, iF - 1, jF, kF, 1, 1, 1)
          - 2.0 * lap_c
          + fetch_sym_ord2_direct(Lap, iF + 1, jF, kF, 1, 1, 1));
        fyy[tid] = d_gp.Sdydy * (
            fetch_sym_ord2_direct(Lap, iF, jF - 1, kF, 1, 1, 1)
          - 2.0 * lap_c
          + fetch_sym_ord2_direct(Lap, iF, jF + 1, kF, 1, 1, 1));
        fzz[tid] = d_gp.Sdzdz * (
            fetch_sym_ord2_direct(Lap, iF, jF, kF - 1, 1, 1, 1)
          - 2.0 * lap_c
          + fetch_sym_ord2_direct(Lap, iF, jF, kF + 1, 1, 1, 1));
        fxy[tid] = d_gp.Sdxdy * (
            fetch_sym_ord2_direct(Lap, iF - 1, jF - 1, kF, 1, 1, 1)
          - fetch_sym_ord2_direct(Lap, iF + 1, jF - 1, kF, 1, 1, 1)
          - fetch_sym_ord2_direct(Lap, iF - 1, jF + 1, kF, 1, 1, 1)
          + fetch_sym_ord2_direct(Lap, iF + 1, jF + 1, kF, 1, 1, 1));
        fxz[tid] = d_gp.Sdxdz * (
            fetch_sym_ord2_direct(Lap, iF - 1, jF, kF - 1, 1, 1, 1)
          - fetch_sym_ord2_direct(Lap, iF + 1, jF, kF - 1, 1, 1, 1)
          - fetch_sym_ord2_direct(Lap, iF - 1, jF, kF + 1, 1, 1, 1)
          + fetch_sym_ord2_direct(Lap, iF + 1, jF, kF + 1, 1, 1, 1));
        fyz[tid] = d_gp.Sdydz * (
            fetch_sym_ord2_direct(Lap, iF, jF - 1, kF - 1, 1, 1, 1)
          - fetch_sym_ord2_direct(Lap, iF, jF + 1, kF - 1, 1, 1, 1)
          - fetch_sym_ord2_direct(Lap, iF, jF - 1, kF + 1, 1, 1, 1)
          + fetch_sym_ord2_direct(Lap, iF, jF + 1, kF + 1, 1, 1, 1));
        chix_out[tid] = d_gp.d2dx * (
            -fetch_sym_ord2_direct(chi, iF - 1, jF, kF, 1, 1, 1)
            +fetch_sym_ord2_direct(chi, iF + 1, jF, kF, 1, 1, 1));
        chiy_out[tid] = d_gp.d2dy * (
            -fetch_sym_ord2_direct(chi, iF, jF - 1, kF, 1, 1, 1)
            +fetch_sym_ord2_direct(chi, iF, jF + 1, kF, 1, 1, 1));
        chiz_out[tid] = d_gp.d2dz * (
            -fetch_sym_ord2_direct(chi, iF, jF, kF - 1, 1, 1, 1)
            +fetch_sym_ord2_direct(chi, iF, jF, kF + 1, 1, 1, 1));
    }
    else {
        fxx[tid] = 0.0; fxy[tid] = 0.0; fxz[tid] = 0.0;
        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;
    }
}

/* Combined ord=3 advection + KO dissipation.
 * When advection and KO use the same source field, symmetry packing is shared.
 * If they differ (e.g. gxx advection + dxx KO), only KO repacks.
 */
static void gpu_lopsided_kodis(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)
{
    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;
    const size_t nz = (size_t)g_buf.prev_nz;
    const size_t w_pack = (nx + 3ull) * (ny + 3ull) * (nz + 3ull);

    kern_symbd_pack_ord3<<<grid(w_pack), BLK>>>(d_f_adv, fh, SoA0, SoA1, SoA2);
    kern_lopsided<<<grid(all), BLK>>>(fh, d_f_rhs, d_Sfx, d_Sfy, d_Sfz);

    if (eps_val > 0.0) {
        if (d_f_ko != d_f_adv) {
            kern_symbd_pack_ord3<<<grid(w_pack), BLK>>>(d_f_ko, fh, SoA0, SoA1, SoA2);
        }
        kern_kodis<<<grid(all), BLK>>>(fh, d_f_rhs, eps_val);
    }
}

__global__ __launch_bounds__(128, 4)
void kern_lopsided_kodis_batched(const double * __restrict__ Sfx,
                                 const double * __restrict__ Sfy,
                                 const double * __restrict__ Sfz,
                                 LopsidedKodisTables tables,
                                 double eps_val)
{
    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid >= d_gp.all) return;

    const int field = blockIdx.y;
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int iminF = d_gp.iminF3, jminF = d_gp.jminF3, kminF = d_gp.kminF3;
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const int SoA0 = tables.soa_signs[3 * field + 0];
    const int SoA1 = tables.soa_signs[3 * field + 1];
    const int SoA2 = tables.soa_signs[3 * field + 2];
    const double *adv_src = tables.adv_fields[field];
    const double *ko_src = tables.ko_fields[field];
    double *rhs = tables.rhs_fields[field];

    const int i0 = tid % nx;
    const int j0 = (tid / nx) % ny;
    const int k0 = tid / (nx * ny);
    const int iF = i0 + 1;
    const int jF = j0 + 1;
    const int kF = k0 + 1;

    if (i0 <= nx - 2 && j0 <= ny - 2 && k0 <= nz - 2) {
        double val = 0.0;

        const double sfx = Sfx[tid];
        if (sfx > 0.0) {
            if (i0 <= nx - 4) {
                val += sfx * d_gp.d12dx * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF + 2, jF, kF, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF + 3, jF, kF, SoA0, SoA1, SoA2));
            } else if (i0 <= nx - 3) {
                val += sfx * d_gp.d12dx * (
                    fetch_sym_ord3_direct(adv_src, iF - 2, jF, kF, SoA0, SoA1, SoA2)
                    - 8.0 * fetch_sym_ord3_direct(adv_src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
                    + 8.0 * fetch_sym_ord3_direct(adv_src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
                    -       fetch_sym_ord3_direct(adv_src, iF + 2, jF, kF, SoA0, SoA1, SoA2));
            } else {
                val -= sfx * d_gp.d12dx * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF - 2, jF, kF, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF - 3, jF, kF, SoA0, SoA1, SoA2));
            }
        } else if (sfx < 0.0) {
            if ((i0 - 2) >= iminF) {
                val -= sfx * d_gp.d12dx * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF - 2, jF, kF, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF - 3, jF, kF, SoA0, SoA1, SoA2));
            } else if ((i0 - 1) >= iminF) {
                val += sfx * d_gp.d12dx * (
                    fetch_sym_ord3_direct(adv_src, iF - 2, jF, kF, SoA0, SoA1, SoA2)
                    - 8.0 * fetch_sym_ord3_direct(adv_src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
                    + 8.0 * fetch_sym_ord3_direct(adv_src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
                    -       fetch_sym_ord3_direct(adv_src, iF + 2, jF, kF, SoA0, SoA1, SoA2));
            } else if (i0 >= iminF) {
                val += sfx * d_gp.d12dx * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF + 1, jF, kF, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF + 2, jF, kF, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF + 3, jF, kF, SoA0, SoA1, SoA2));
            }
        }

        const double sfy = Sfy[tid];
        if (sfy > 0.0) {
            if (j0 <= ny - 4) {
                val += sfy * d_gp.d12dy * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF, jF + 2, kF, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF, jF + 3, kF, SoA0, SoA1, SoA2));
            } else if (j0 <= ny - 3) {
                val += sfy * d_gp.d12dy * (
                    fetch_sym_ord3_direct(adv_src, iF, jF - 2, kF, SoA0, SoA1, SoA2)
                    - 8.0 * fetch_sym_ord3_direct(adv_src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
                    + 8.0 * fetch_sym_ord3_direct(adv_src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
                    -       fetch_sym_ord3_direct(adv_src, iF, jF + 2, kF, SoA0, SoA1, SoA2));
            } else {
                val -= sfy * d_gp.d12dy * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF, jF - 2, kF, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF, jF - 3, kF, SoA0, SoA1, SoA2));
            }
        } else if (sfy < 0.0) {
            if ((j0 - 2) >= jminF) {
                val -= sfy * d_gp.d12dy * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF, jF - 2, kF, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF, jF - 3, kF, SoA0, SoA1, SoA2));
            } else if ((j0 - 1) >= jminF) {
                val += sfy * d_gp.d12dy * (
                    fetch_sym_ord3_direct(adv_src, iF, jF - 2, kF, SoA0, SoA1, SoA2)
                    - 8.0 * fetch_sym_ord3_direct(adv_src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
                    + 8.0 * fetch_sym_ord3_direct(adv_src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
                    -       fetch_sym_ord3_direct(adv_src, iF, jF + 2, kF, SoA0, SoA1, SoA2));
            } else if (j0 >= jminF) {
                val += sfy * d_gp.d12dy * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF, jF + 1, kF, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF, jF + 2, kF, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF, jF + 3, kF, SoA0, SoA1, SoA2));
            }
        }

        const double sfz = Sfz[tid];
        if (sfz > 0.0) {
            if (k0 <= nz - 4) {
                val += sfz * d_gp.d12dz * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF + 2, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF, jF, kF + 3, SoA0, SoA1, SoA2));
            } else if (k0 <= nz - 3) {
                val += sfz * d_gp.d12dz * (
                    fetch_sym_ord3_direct(adv_src, iF, jF, kF - 2, SoA0, SoA1, SoA2)
                    - 8.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
                    + 8.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
                    -       fetch_sym_ord3_direct(adv_src, iF, jF, kF + 2, SoA0, SoA1, SoA2));
            } else {
                val -= sfz * d_gp.d12dz * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF - 2, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF, jF, kF - 3, SoA0, SoA1, SoA2));
            }
        } else if (sfz < 0.0) {
            if ((k0 - 2) >= kminF) {
                val -= sfz * d_gp.d12dz * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF - 2, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF, jF, kF - 3, SoA0, SoA1, SoA2));
            } else if ((k0 - 1) >= kminF) {
                val += sfz * d_gp.d12dz * (
                    fetch_sym_ord3_direct(adv_src, iF, jF, kF - 2, SoA0, SoA1, SoA2)
                    - 8.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
                    + 8.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
                    -       fetch_sym_ord3_direct(adv_src, iF, jF, kF + 2, SoA0, SoA1, SoA2));
            } else if (k0 >= kminF) {
                val += sfz * d_gp.d12dz * (
                    -3.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
                    -10.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF, SoA0, SoA1, SoA2)
                    +18.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF + 1, SoA0, SoA1, SoA2)
                    - 6.0 * fetch_sym_ord3_direct(adv_src, iF, jF, kF + 2, SoA0, SoA1, SoA2)
                    +       fetch_sym_ord3_direct(adv_src, iF, jF, kF + 3, SoA0, SoA1, SoA2));
            }
        }

        rhs[tid] += val;
    }

    if (eps_val > 0.0 &&
        (iF - 3) >= iminF && (iF + 3) <= imaxF &&
        (jF - 3) >= jminF && (jF + 3) <= jmaxF &&
        (kF - 3) >= kminF && (kF + 3) <= kmaxF)
    {
        const double cof = 64.0;
        const double Dx =
            (fetch_sym_ord3_direct(ko_src, iF - 3, jF, kF, SoA0, SoA1, SoA2)
           + fetch_sym_ord3_direct(ko_src, iF + 3, jF, kF, SoA0, SoA1, SoA2))
          - 6.0 * (fetch_sym_ord3_direct(ko_src, iF - 2, jF, kF, SoA0, SoA1, SoA2)
                 + fetch_sym_ord3_direct(ko_src, iF + 2, jF, kF, SoA0, SoA1, SoA2))
          + 15.0 * (fetch_sym_ord3_direct(ko_src, iF - 1, jF, kF, SoA0, SoA1, SoA2)
                  + fetch_sym_ord3_direct(ko_src, iF + 1, jF, kF, SoA0, SoA1, SoA2))
          - 20.0 * fetch_sym_ord3_direct(ko_src, iF, jF, kF, SoA0, SoA1, SoA2);

        const double Dy =
            (fetch_sym_ord3_direct(ko_src, iF, jF - 3, kF, SoA0, SoA1, SoA2)
           + fetch_sym_ord3_direct(ko_src, iF, jF + 3, kF, SoA0, SoA1, SoA2))
          - 6.0 * (fetch_sym_ord3_direct(ko_src, iF, jF - 2, kF, SoA0, SoA1, SoA2)
                 + fetch_sym_ord3_direct(ko_src, iF, jF + 2, kF, SoA0, SoA1, SoA2))
          + 15.0 * (fetch_sym_ord3_direct(ko_src, iF, jF - 1, kF, SoA0, SoA1, SoA2)
                  + fetch_sym_ord3_direct(ko_src, iF, jF + 1, kF, SoA0, SoA1, SoA2))
          - 20.0 * fetch_sym_ord3_direct(ko_src, iF, jF, kF, SoA0, SoA1, SoA2);

        const double Dz =
            (fetch_sym_ord3_direct(ko_src, iF, jF, kF - 3, SoA0, SoA1, SoA2)
           + fetch_sym_ord3_direct(ko_src, iF, jF, kF + 3, SoA0, SoA1, SoA2))
          - 6.0 * (fetch_sym_ord3_direct(ko_src, iF, jF, kF - 2, SoA0, SoA1, SoA2)
                 + fetch_sym_ord3_direct(ko_src, iF, jF, kF + 2, SoA0, SoA1, SoA2))
          + 15.0 * (fetch_sym_ord3_direct(ko_src, iF, jF, kF - 1, SoA0, SoA1, SoA2)
                  + fetch_sym_ord3_direct(ko_src, iF, jF, kF + 1, SoA0, SoA1, SoA2))
          - 20.0 * fetch_sym_ord3_direct(ko_src, iF, jF, kF, SoA0, SoA1, SoA2);

        rhs[tid] += (eps_val / cof) * (Dx / d_gp.dX + Dy / d_gp.dY + Dz / d_gp.dZ);
    }
}

static void gpu_lopsided_kodis_state_batch(double eps_val, int all)
{
    LopsidedKodisTables tables = {};
    for (int i = 0; i < BSSN_LK_FIELD_COUNT; ++i) {
        tables.adv_fields[i] = g_buf.slot[k_lk_adv_slots[i]];
        tables.ko_fields[i] = g_buf.slot[k_lk_ko_slots[i]];
        tables.rhs_fields[i] = g_buf.slot[k_lk_rhs_slots[i]];
    }
    std::memcpy(tables.soa_signs, k_lk_soa_signs, sizeof(k_lk_soa_signs));

    dim3 launch_grid((unsigned int)grid((size_t)all), (unsigned int)BSSN_LK_FIELD_COUNT);
    kern_lopsided_kodis_batched<<<launch_grid, BLK>>>(
        g_buf.slot[S_betax], g_buf.slot[S_betay], g_buf.slot[S_betaz], tables, eps_val);
}

__global__ void kern_rk4_finalize(const double * __restrict__ f0,
                                  double * __restrict__ frhs,
                                  double * __restrict__ accum,
                                  double dT,
                                  int rk4_stage)
{
    for (int i = blockIdx.x * blockDim.x + threadIdx.x;
         i < d_gp.all;
         i += blockDim.x * gridDim.x)
    {
        const double rhs = frhs[i];
        switch (rk4_stage) {
        case 0:
            accum[i] = rhs;
            frhs[i] = f0[i] + 0.5 * dT * rhs;
            break;
        case 1:
            accum[i] += 2.0 * rhs;
            frhs[i] = f0[i] + 0.5 * dT * rhs;
            break;
        case 2:
            accum[i] += 2.0 * rhs;
            frhs[i] = f0[i] + dT * rhs;
            break;
        default:
            frhs[i] = f0[i] + (dT / 6.0) * (accum[i] + rhs);
            break;
        }
    }
}

__global__ __launch_bounds__(128, 4)
void kern_rk4_finalize_batched(Rk4FinalizeTables tables,
                               double dT,
                               int rk4_stage,
                               double chitiny)
{
    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid >= d_gp.all) return;

    const int field = blockIdx.y;
    const double *f0 = tables.f0_fields[field];
    double *frhs = tables.rhs_fields[field];
    double *accum = tables.accum_fields[field];

    const double rhs = frhs[tid];
    switch (rk4_stage) {
    case 0:
        accum[tid] = rhs;
        frhs[tid] = f0[tid] + 0.5 * dT * rhs;
        break;
    case 1:
        accum[tid] += 2.0 * rhs;
        frhs[tid] = f0[tid] + 0.5 * dT * rhs;
        break;
    case 2:
        accum[tid] += 2.0 * rhs;
        frhs[tid] = f0[tid] + dT * rhs;
        break;
    default:
        frhs[tid] = f0[tid] + (dT / 6.0) * (accum[tid] + rhs);
        break;
    }

    if (field == 0 && frhs[tid] < chitiny) frhs[tid] = chitiny;
}

static void gpu_rk4_finalize_batch(const StepContext &ctx,
                                   size_t all,
                                   double dT,
                                   int rk4_stage,
                                   double chitiny)
{
    Rk4FinalizeTables tables = {};
    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
        tables.f0_fields[i] = ctx.d_state0[i];
        tables.rhs_fields[i] = g_buf.slot[k_state_rhs_slots[i]];
        tables.accum_fields[i] = ctx.d_accum[i];
    }

    dim3 launch_grid((unsigned int)grid(all), (unsigned int)BSSN_STATE_COUNT);
    kern_rk4_finalize_batched<<<launch_grid, BLK>>>(tables, dT, rk4_stage, chitiny);
}

__global__ __launch_bounds__(128, 4)
void kern_copy_patch_boundary_batched(PatchBoundaryTables tables,
                                      int touch_xmin, int touch_xmax,
                                      int touch_ymin, int touch_ymax,
                                      int touch_zmin, int touch_zmax)
{
    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid >= d_gp.all) return;

    const int nx = d_gp.ex[0];
    const int ny = d_gp.ex[1];
    const int i0 = tid % nx;
    const int j0 = (tid / nx) % ny;
    const int k0 = tid / (nx * ny);

    const bool on_boundary =
        (touch_xmin && i0 == 0) ||
        (touch_xmax && i0 == nx - 1) ||
        (touch_ymin && j0 == 0) ||
        (touch_ymax && j0 == ny - 1) ||
        (touch_zmin && k0 == 0) ||
        (touch_zmax && k0 == d_gp.ex[2] - 1);
    if (!on_boundary) return;

    const int field = blockIdx.y;
    tables.dst_fields[field][tid] = tables.src_fields[field][tid];
}

static void gpu_copy_patch_boundary_batch(int all,
                                          int touch_xmin, int touch_xmax,
                                          int touch_ymin, int touch_ymax,
                                          int touch_zmin, int touch_zmax)
{
    if (!(touch_xmin || touch_xmax || touch_ymin || touch_ymax || touch_zmin || touch_zmax))
        return;

    PatchBoundaryTables tables = {};
    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
        tables.src_fields[i] = g_buf.slot[k_state_input_slots[i]];
        tables.dst_fields[i] = g_buf.slot[k_state_rhs_slots[i]];
    }

    dim3 launch_grid((unsigned int)grid((size_t)all), (unsigned int)BSSN_STATE_COUNT);
    kern_copy_patch_boundary_batched<<<launch_grid, BLK>>>(
        tables,
        touch_xmin, touch_xmax,
        touch_ymin, touch_ymax,
        touch_zmin, touch_zmax);
}

__global__ void kern_enforce_ga_cuda(double * __restrict__ dxx,
                                     double * __restrict__ gxy,
                                     double * __restrict__ gxz,
                                     double * __restrict__ dyy,
                                     double * __restrict__ gyz,
                                     double * __restrict__ dzz,
                                     double * __restrict__ Axx,
                                     double * __restrict__ Axy,
                                     double * __restrict__ Axz,
                                     double * __restrict__ Ayy,
                                     double * __restrict__ Ayz,
                                     double * __restrict__ Azz)
{
    constexpr double F1O3 = 1.0 / 3.0;
    constexpr double ONE = 1.0;
    constexpr double TWO = 2.0;

    for (int i = blockIdx.x * blockDim.x + threadIdx.x;
         i < d_gp.all;
         i += blockDim.x * gridDim.x)
    {
        double lgxx = dxx[i] + ONE;
        double lgyy = dyy[i] + ONE;
        double lgzz = dzz[i] + ONE;
        double lgxy = gxy[i];
        double lgxz = gxz[i];
        double lgyz = gyz[i];

        double lscale = lgxx * lgyy * lgzz
                      + lgxy * lgyz * lgxz
                      + lgxz * lgxy * lgyz
                      - lgxz * lgyy * lgxz
                      - lgxy * lgxy * lgzz
                      - lgxx * lgyz * lgyz;

        lscale = ONE / cbrt(lscale);

        lgxx *= lscale;
        lgxy *= lscale;
        lgxz *= lscale;
        lgyy *= lscale;
        lgyz *= lscale;
        lgzz *= lscale;

        dxx[i] = lgxx - ONE;
        gxy[i] = lgxy;
        gxz[i] = lgxz;
        dyy[i] = lgyy - ONE;
        gyz[i] = lgyz;
        dzz[i] = lgzz - ONE;

        const double lgupxx =   (lgyy * lgzz - lgyz * lgyz);
        const double lgupxy = - (lgxy * lgzz - lgyz * lgxz);
        const double lgupxz =   (lgxy * lgyz - lgyy * lgxz);
        const double lgupyy =   (lgxx * lgzz - lgxz * lgxz);
        const double lgupyz = - (lgxx * lgyz - lgxy * lgxz);
        const double lgupzz =   (lgxx * lgyy - lgxy * lgxy);

        const double ltrA = lgupxx * Axx[i] + lgupyy * Ayy[i] + lgupzz * Azz[i]
                          + TWO * (lgupxy * Axy[i] + lgupxz * Axz[i] + lgupyz * Ayz[i]);

        Axx[i] -= F1O3 * lgxx * ltrA;
        Axy[i] -= F1O3 * lgxy * ltrA;
        Axz[i] -= F1O3 * lgxz * ltrA;
        Ayy[i] -= F1O3 * lgyy * ltrA;
        Ayz[i] -= F1O3 * lgyz * ltrA;
        Azz[i] -= F1O3 * lgzz * ltrA;
    }
}

__global__ void kern_lowerboundset_cuda(double * __restrict__ chi, double tinny)
{
    for (int i = blockIdx.x * blockDim.x + threadIdx.x;
         i < d_gp.all;
         i += blockDim.x * gridDim.x)
    {
        if (chi[i] < tinny) chi[i] = tinny;
    }
}

enum SommerFace {
    FACE_XMAX = 0,
    FACE_YMAX = 1,
    FACE_ZMAX = 2,
    FACE_XMIN = 3,
    FACE_YMIN = 4,
    FACE_ZMIN = 5
};

__global__ void kern_sommerfeld_face_bam(const double * __restrict__ fh,
                                         double * __restrict__ f_rhs,
                                         int face,
                                         double velocity,
                                         double x0,
                                         double y0,
                                         double z0)
{
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int imin = (d_gp.Symmetry > 1 && fabs(x0) < d_gp.dX) ? 0 : 1;
    const int jmin = (d_gp.Symmetry > 1 && fabs(y0) < d_gp.dY) ? 0 : 1;
    const int kmin = (d_gp.Symmetry > 0 && fabs(z0) < d_gp.dZ) ? 0 : 1;
    const int imax = nx, jmax = ny, kmax = nz;
    const int plane_count =
        (face == FACE_XMAX || face == FACE_XMIN) ? ny * nz :
        (face == FACE_YMAX || face == FACE_YMIN) ? nx * nz :
                                                   nx * ny;

    for (int tid = blockIdx.x * blockDim.x + threadIdx.x;
         tid < plane_count;
         tid += blockDim.x * gridDim.x)
    {
        int i0 = 0, j0 = 0, k0 = 0;
        if (face == FACE_XMAX || face == FACE_XMIN) {
            j0 = tid % ny;
            k0 = tid / ny;
            i0 = (face == FACE_XMAX) ? (nx - 1) : 0;
        } else if (face == FACE_YMAX || face == FACE_YMIN) {
            i0 = tid % nx;
            k0 = tid / nx;
            j0 = (face == FACE_YMAX) ? (ny - 1) : 0;
        } else {
            i0 = tid % nx;
            j0 = tid / nx;
            k0 = (face == FACE_ZMAX) ? (nz - 1) : 0;
        }

        const int iF = i0 + 1;
        const int jF = j0 + 1;
        const int kF = k0 + 1;
        const int p = idx_ex_d(i0, j0, k0);

        const double x = x0 + i0 * d_gp.dX;
        const double y = y0 + j0 * d_gp.dY;
        const double z = z0 + k0 * d_gp.dZ;
        const double r = sqrt(x * x + y * y + z * z);
        if (r == 0.0) continue;

        const double wx = velocity * x / r;
        const double wy = velocity * y / r;
        const double wz = velocity * z / r;

        double fx = 0.0, fy = 0.0, fz = 0.0;
        if (wx > 0.0) {
            if (iF - 2 >= imin) fx = d_gp.d2dx * (3.0 * fh[idx_fh2(iF, jF, kF)] - 4.0 * fh[idx_fh2(iF - 1, jF, kF)] + fh[idx_fh2(iF - 2, jF, kF)]);
            else if (iF - 1 >= imin) fx = d_gp.d2dx * (-fh[idx_fh2(iF - 1, jF, kF)] + fh[idx_fh2(iF + 1, jF, kF)]);
            else fx = d_gp.d2dx * (-fh[idx_fh2(iF + 2, jF, kF)] + 4.0 * fh[idx_fh2(iF + 1, jF, kF)] - 3.0 * fh[idx_fh2(iF, jF, kF)]);
        } else if (wx < 0.0) {
            if (iF + 2 <= imax) fx = d_gp.d2dx * (-fh[idx_fh2(iF + 2, jF, kF)] + 4.0 * fh[idx_fh2(iF + 1, jF, kF)] - 3.0 * fh[idx_fh2(iF, jF, kF)]);
            else if (iF + 1 <= imax) fx = d_gp.d2dx * (-fh[idx_fh2(iF - 1, jF, kF)] + fh[idx_fh2(iF + 1, jF, kF)]);
            else fx = d_gp.d2dx * (3.0 * fh[idx_fh2(iF, jF, kF)] - 4.0 * fh[idx_fh2(iF - 1, jF, kF)] + fh[idx_fh2(iF - 2, jF, kF)]);
        }

        if (wy > 0.0) {
            if (jF - 2 >= jmin) fy = d_gp.d2dy * (3.0 * fh[idx_fh2(iF, jF, kF)] - 4.0 * fh[idx_fh2(iF, jF - 1, kF)] + fh[idx_fh2(iF, jF - 2, kF)]);
            else if (jF - 1 >= jmin) fy = d_gp.d2dy * (-fh[idx_fh2(iF, jF - 1, kF)] + fh[idx_fh2(iF, jF + 1, kF)]);
            else fy = d_gp.d2dy * (-fh[idx_fh2(iF, jF + 2, kF)] + 4.0 * fh[idx_fh2(iF, jF + 1, kF)] - 3.0 * fh[idx_fh2(iF, jF, kF)]);
        } else if (wy < 0.0) {
            if (jF + 2 <= jmax) fy = d_gp.d2dy * (-fh[idx_fh2(iF, jF + 2, kF)] + 4.0 * fh[idx_fh2(iF, jF + 1, kF)] - 3.0 * fh[idx_fh2(iF, jF, kF)]);
            else if (jF + 1 <= jmax) fy = d_gp.d2dy * (-fh[idx_fh2(iF, jF - 1, kF)] + fh[idx_fh2(iF, jF + 1, kF)]);
            else fy = d_gp.d2dy * (3.0 * fh[idx_fh2(iF, jF, kF)] - 4.0 * fh[idx_fh2(iF, jF - 1, kF)] + fh[idx_fh2(iF, jF - 2, kF)]);
        }

        if (wz > 0.0) {
            if (kF - 2 >= kmin) fz = d_gp.d2dz * (3.0 * fh[idx_fh2(iF, jF, kF)] - 4.0 * fh[idx_fh2(iF, jF, kF - 1)] + fh[idx_fh2(iF, jF, kF - 2)]);
            else if (kF - 1 >= kmin) fz = d_gp.d2dz * (-fh[idx_fh2(iF, jF, kF - 1)] + fh[idx_fh2(iF, jF, kF + 1)]);
            else fz = d_gp.d2dz * (-fh[idx_fh2(iF, jF, kF + 2)] + 4.0 * fh[idx_fh2(iF, jF, kF + 1)] - 3.0 * fh[idx_fh2(iF, jF, kF)]);
        } else if (wz < 0.0) {
            if (kF + 2 <= kmax) fz = d_gp.d2dz * (-fh[idx_fh2(iF, jF, kF + 2)] + 4.0 * fh[idx_fh2(iF, jF, kF + 1)] - 3.0 * fh[idx_fh2(iF, jF, kF)]);
            else if (kF + 1 <= kmax) fz = d_gp.d2dz * (-fh[idx_fh2(iF, jF, kF - 1)] + fh[idx_fh2(iF, jF, kF + 1)]);
            else fz = d_gp.d2dz * (3.0 * fh[idx_fh2(iF, jF, kF)] - 4.0 * fh[idx_fh2(iF, jF, kF - 1)] + fh[idx_fh2(iF, jF, kF - 2)]);
        }

        f_rhs[p] = -velocity * (fx * x + fy * y + fz * z + fh[idx_fh2(iF, jF, kF)]) / r;
    }
}

static void gpu_sommerfeld_routbam(double *d_f0, double *d_f_rhs,
                                   double velocity,
                                   double SoA0, double SoA1, double SoA2,
                                   double *X, double *Y, double *Z,
                                   const double *bbox,
                                   int Symmetry)
{
    if (velocity == 0.0) return;
    const int nx = g_buf.prev_nx;
    const int ny = g_buf.prev_ny;
    const int nz = g_buf.prev_nz;
    const double dX = X[1] - X[0];
    const double dY = Y[1] - Y[0];
    const double dZ = Z[1] - Z[0];
    const bool touch_xmax = fabs(X[nx - 1] - bbox[3]) < dX;
    const bool touch_ymax = fabs(Y[ny - 1] - bbox[4]) < dY;
    const bool touch_zmax = fabs(Z[nz - 1] - bbox[5]) < dZ;
    const bool touch_xmin = fabs(X[0] - bbox[0]) < dX &&
                            !(Symmetry == 2 && fabs(bbox[0]) < dX * 0.5);
    const bool touch_ymin = fabs(Y[0] - bbox[1]) < dY &&
                            !(Symmetry == 2 && fabs(bbox[1]) < dY * 0.5);
    const bool touch_zmin = fabs(Z[0] - bbox[2]) < dZ &&
                            !(Symmetry > 0 && fabs(bbox[2]) < dZ * 0.5);

    const size_t w_pack = (size_t)(nx + 2) * (ny + 2) * (nz + 2);
    kern_symbd_pack_ord2<<<grid(w_pack), BLK>>>(d_f0, g_buf.d_fh2, SoA0, SoA1, SoA2);
    if (touch_xmax) kern_sommerfeld_face_bam<<<grid((size_t)ny * nz), BLK>>>(g_buf.d_fh2, d_f_rhs, FACE_XMAX, velocity, X[0], Y[0], Z[0]);
    if (touch_ymax) kern_sommerfeld_face_bam<<<grid((size_t)nx * nz), BLK>>>(g_buf.d_fh2, d_f_rhs, FACE_YMAX, velocity, X[0], Y[0], Z[0]);
    if (touch_zmax) kern_sommerfeld_face_bam<<<grid((size_t)nx * ny), BLK>>>(g_buf.d_fh2, d_f_rhs, FACE_ZMAX, velocity, X[0], Y[0], Z[0]);
    if (touch_xmin) kern_sommerfeld_face_bam<<<grid((size_t)ny * nz), BLK>>>(g_buf.d_fh2, d_f_rhs, FACE_XMIN, velocity, X[0], Y[0], Z[0]);
    if (touch_ymin) kern_sommerfeld_face_bam<<<grid((size_t)nx * nz), BLK>>>(g_buf.d_fh2, d_f_rhs, FACE_YMIN, velocity, X[0], Y[0], Z[0]);
    if (touch_zmin) kern_sommerfeld_face_bam<<<grid((size_t)nx * ny), BLK>>>(g_buf.d_fh2, d_f_rhs, FACE_ZMIN, velocity, X[0], Y[0], Z[0]);
}

/* ================================================================== */
/*  C. Point-wise computation kernels                                  */
/* ================================================================== */

/* Phase 1: alpn1, chin1, gxx=dxx+1, gyy=dyy+1, gzz=dzz+1 */
__global__ void kern_phase1_prep(
    const double* __restrict__ Lap, const double* __restrict__ chi,
    const double* __restrict__ dxx, const double* __restrict__ dyy,
    const double* __restrict__ dzz,
    double* __restrict__ alpn1, double* __restrict__ chin1,
    double* __restrict__ gxx, double* __restrict__ gyy, double* __restrict__ gzz)
{
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        alpn1[i] = Lap[i] + 1.0;
        chin1[i] = chi[i] + 1.0;
        gxx[i] = dxx[i] + 1.0;
        gyy[i] = dyy[i] + 1.0;
        gzz[i] = dzz[i] + 1.0;
    }
}

/* Phase 2a: chi_rhs, gij_rhs */
__global__ void kern_phase2_metric_rhs(
    const double* __restrict__ alpn1, const double* __restrict__ chin1,
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    const double* __restrict__ trK,
    const double* __restrict__ Axx, const double* __restrict__ Axy,
    const double* __restrict__ Axz, const double* __restrict__ Ayy,
    const double* __restrict__ Ayz, const double* __restrict__ Azz,
    const double* __restrict__ betaxx, const double* __restrict__ betaxy,
    const double* __restrict__ betaxz, const double* __restrict__ betayx,
    const double* __restrict__ betayy, const double* __restrict__ betayz,
    const double* __restrict__ betazx, const double* __restrict__ betazy,
    const double* __restrict__ betazz,
    double* __restrict__ chi_rhs, double* __restrict__ gxx_rhs,
    double* __restrict__ gyy_rhs, double* __restrict__ gzz_rhs,
    double* __restrict__ gxy_rhs, double* __restrict__ gyz_rhs,
    double* __restrict__ gxz_rhs)
{
    const double F2o3 = 2.0/3.0, F1o3 = 1.0/3.0, TWO = 2.0;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double db = betaxx[i] + betayy[i] + betazz[i];
        chi_rhs[i] = F2o3 * chin1[i] * (alpn1[i] * trK[i] - db);
        gxx_rhs[i] = -TWO*alpn1[i]*Axx[i] - F2o3*gxx[i]*db
            + TWO*(gxx[i]*betaxx[i] + gxy[i]*betayx[i] + gxz[i]*betazx[i]);
        gyy_rhs[i] = -TWO*alpn1[i]*Ayy[i] - F2o3*gyy[i]*db
            + TWO*(gxy[i]*betaxy[i] + gyy[i]*betayy[i] + gyz[i]*betazy[i]);
        gzz_rhs[i] = -TWO*alpn1[i]*Azz[i] - F2o3*gzz[i]*db
            + TWO*(gxz[i]*betaxz[i] + gyz[i]*betayz[i] + gzz[i]*betazz[i]);
        gxy_rhs[i] = -TWO*alpn1[i]*Axy[i] + F1o3*gxy[i]*db
            + gxx[i]*betaxy[i] + gxz[i]*betazy[i] + gyy[i]*betayx[i]
            + gyz[i]*betazx[i] - gxy[i]*betazz[i];
        gyz_rhs[i] = -TWO*alpn1[i]*Ayz[i] + F1o3*gyz[i]*db
            + gxy[i]*betaxz[i] + gyy[i]*betayz[i] + gxz[i]*betaxy[i]
            + gzz[i]*betazy[i] - gyz[i]*betaxx[i];
        gxz_rhs[i] = -TWO*alpn1[i]*Axz[i] + F1o3*gxz[i]*db
            + gxx[i]*betaxz[i] + gxy[i]*betayz[i] + gyz[i]*betayx[i]
            + gzz[i]*betazx[i] - gxz[i]*betayy[i];
    }
}

/* Phase 2b: metric inverse */
__global__ void kern_phase2_inverse(
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    double* __restrict__ gupxx, double* __restrict__ gupxy,
    double* __restrict__ gupxz, double* __restrict__ gupyy,
    double* __restrict__ gupyz, double* __restrict__ gupzz)
{
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double det = gxx[i]*gyy[i]*gzz[i] + gxy[i]*gyz[i]*gxz[i] + gxz[i]*gxy[i]*gyz[i]
                   - gxz[i]*gyy[i]*gxz[i] - gxy[i]*gxy[i]*gzz[i] - gxx[i]*gyz[i]*gyz[i];
        double inv = 1.0 / det;
        gupxx[i] = (gyy[i]*gzz[i] - gyz[i]*gyz[i]) * inv;
        gupxy[i] = -(gxy[i]*gzz[i] - gyz[i]*gxz[i]) * inv;
        gupxz[i] = (gxy[i]*gyz[i] - gyy[i]*gxz[i]) * inv;
        gupyy[i] = (gxx[i]*gzz[i] - gxz[i]*gxz[i]) * inv;
        gupyz[i] = -(gxx[i]*gyz[i] - gxy[i]*gxz[i]) * inv;
        gupzz[i] = (gxx[i]*gyy[i] - gxy[i]*gxy[i]) * inv;
    }
}

/* Phase 3: Gamma constraint residuals (co==0 only) */
__global__ void kern_phase3_gamma_constraint(
    const double* __restrict__ Gamx, const double* __restrict__ Gamy,
    const double* __restrict__ Gamz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ gxxx, const double* __restrict__ gxyx,
    const double* __restrict__ gxzx, const double* __restrict__ gyyx,
    const double* __restrict__ gyzx, const double* __restrict__ gzzx,
    const double* __restrict__ gxxy, const double* __restrict__ gxyy,
    const double* __restrict__ gxzy, const double* __restrict__ gyyy,
    const double* __restrict__ gyzy, const double* __restrict__ gzzy,
    const double* __restrict__ gxxz, const double* __restrict__ gxyz,
    const double* __restrict__ gxzz, const double* __restrict__ gyyz,
    const double* __restrict__ gyzz, const double* __restrict__ gzzz,
    double* __restrict__ Gmx_Res, double* __restrict__ Gmy_Res,
    double* __restrict__ Gmz_Res)
{
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double uxx=gupxx[i], uxy=gupxy[i], uxz=gupxz[i];
        double uyy=gupyy[i], uyz=gupyz[i], uzz=gupzz[i];

        Gmx_Res[i] = Gamx[i] - (
            uxx*(uxx*gxxx[i]+uxy*gxyx[i]+uxz*gxzx[i]) +
            uxy*(uxx*gxyx[i]+uxy*gyyx[i]+uxz*gyzx[i]) +
            uxz*(uxx*gxzx[i]+uxy*gyzx[i]+uxz*gzzx[i]) +
            uxx*(uxy*gxxy[i]+uyy*gxyy[i]+uyz*gxzy[i]) +
            uxy*(uxy*gxyy[i]+uyy*gyyy[i]+uyz*gyzy[i]) +
            uxz*(uxy*gxzy[i]+uyy*gyzy[i]+uyz*gzzy[i]) +
            uxx*(uxz*gxxz[i]+uyz*gxyz[i]+uzz*gxzz[i]) +
            uxy*(uxz*gxyz[i]+uyz*gyyz[i]+uzz*gyzz[i]) +
            uxz*(uxz*gxzz[i]+uyz*gyzz[i]+uzz*gzzz[i]));

        Gmy_Res[i] = Gamy[i] - (
            uxx*(uxy*gxxx[i]+uyy*gxyx[i]+uyz*gxzx[i]) +
            uxy*(uxy*gxyx[i]+uyy*gyyx[i]+uyz*gyzx[i]) +
            uxz*(uxy*gxzx[i]+uyy*gyzx[i]+uyz*gzzx[i]) +
            uxy*(uxy*gxxy[i]+uyy*gxyy[i]+uyz*gxzy[i]) +
            uyy*(uxy*gxyy[i]+uyy*gyyy[i]+uyz*gyzy[i]) +
            uyz*(uxy*gxzy[i]+uyy*gyzy[i]+uyz*gzzy[i]) +
            uxy*(uxz*gxxz[i]+uyz*gxyz[i]+uzz*gxzz[i]) +
            uyy*(uxz*gxyz[i]+uyz*gyyz[i]+uzz*gyzz[i]) +
            uyz*(uxz*gxzz[i]+uyz*gyzz[i]+uzz*gzzz[i]));

        Gmz_Res[i] = Gamz[i] - (
            uxx*(uxz*gxxx[i]+uyz*gxyx[i]+uzz*gxzx[i]) +
            uxy*(uxz*gxyx[i]+uyz*gyyx[i]+uzz*gyzx[i]) +
            uxz*(uxz*gxzx[i]+uyz*gyzx[i]+uzz*gzzx[i]) +
            uxy*(uxz*gxxy[i]+uyz*gxyy[i]+uzz*gxzy[i]) +
            uyy*(uxz*gxyy[i]+uyz*gyyy[i]+uzz*gyzy[i]) +
            uyz*(uxz*gxzy[i]+uyz*gyzy[i]+uzz*gzzy[i]) +
            uxz*(uxz*gxxz[i]+uyz*gxyz[i]+uzz*gxzz[i]) +
            uyz*(uxz*gxyz[i]+uyz*gyyz[i]+uzz*gyzz[i]) +
            uzz*(uxz*gxzz[i]+uyz*gyzz[i]+uzz*gzzz[i]));
    }
}

/* Phase 4: 18 Christoffel symbols */
__global__ __launch_bounds__(128, 4)
void kern_phase4_christoffel(
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ gxxx, const double* __restrict__ gxyx,
    const double* __restrict__ gxzx, const double* __restrict__ gyyx,
    const double* __restrict__ gyzx, const double* __restrict__ gzzx,
    const double* __restrict__ gxxy, const double* __restrict__ gxyy,
    const double* __restrict__ gxzy, const double* __restrict__ gyyy,
    const double* __restrict__ gyzy, const double* __restrict__ gzzy,
    const double* __restrict__ gxxz, const double* __restrict__ gxyz,
    const double* __restrict__ gxzz, const double* __restrict__ gyyz,
    const double* __restrict__ gyzz, const double* __restrict__ gzzz,
    double* __restrict__ Gxxx, double* __restrict__ Gxxy, double* __restrict__ Gxxz,
    double* __restrict__ Gxyy, double* __restrict__ Gxyz, double* __restrict__ Gxzz,
    double* __restrict__ Gyxx, double* __restrict__ Gyxy, double* __restrict__ Gyxz,
    double* __restrict__ Gyyy, double* __restrict__ Gyyz, double* __restrict__ Gyzz,
    double* __restrict__ Gzxx, double* __restrict__ Gzxy, double* __restrict__ Gzxz,
    double* __restrict__ Gzyy, double* __restrict__ Gzyz, double* __restrict__ Gzzz_o)
{
    const double H = 0.5, TWO = 2.0;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];
        /* Gamma^x_{xx} */
        Gxxx[i]=H*(uxx*gxxx[i]+uxy*(TWO*gxyx[i]-gxxy[i])+uxz*(TWO*gxzx[i]-gxxz[i]));
        Gyxx[i]=H*(uxy*gxxx[i]+uyy*(TWO*gxyx[i]-gxxy[i])+uyz*(TWO*gxzx[i]-gxxz[i]));
        Gzxx[i]=H*(uxz*gxxx[i]+uyz*(TWO*gxyx[i]-gxxy[i])+uzz*(TWO*gxzx[i]-gxxz[i]));
        /* yy */
        Gxyy[i]=H*(uxx*(TWO*gxyy[i]-gyyx[i])+uxy*gyyy[i]+uxz*(TWO*gyzy[i]-gyyz[i]));
        Gyyy[i]=H*(uxy*(TWO*gxyy[i]-gyyx[i])+uyy*gyyy[i]+uyz*(TWO*gyzy[i]-gyyz[i]));
        Gzyy[i]=H*(uxz*(TWO*gxyy[i]-gyyx[i])+uyz*gyyy[i]+uzz*(TWO*gyzy[i]-gyyz[i]));
        /* zz */
        Gxzz[i]=H*(uxx*(TWO*gxzz[i]-gzzx[i])+uxy*(TWO*gyzz[i]-gzzy[i])+uxz*gzzz[i]);
        Gyzz[i]=H*(uxy*(TWO*gxzz[i]-gzzx[i])+uyy*(TWO*gyzz[i]-gzzy[i])+uyz*gzzz[i]);
        Gzzz_o[i]=H*(uxz*(TWO*gxzz[i]-gzzx[i])+uyz*(TWO*gyzz[i]-gzzy[i])+uzz*gzzz[i]);
        /* xy */
        Gxxy[i]=H*(uxx*gxxy[i]+uxy*gyyx[i]+uxz*(gxzy[i]+gyzx[i]-gxyz[i]));
        Gyxy[i]=H*(uxy*gxxy[i]+uyy*gyyx[i]+uyz*(gxzy[i]+gyzx[i]-gxyz[i]));
        Gzxy[i]=H*(uxz*gxxy[i]+uyz*gyyx[i]+uzz*(gxzy[i]+gyzx[i]-gxyz[i]));
        /* xz */
        Gxxz[i]=H*(uxx*gxxz[i]+uxy*(gxyz[i]+gyzx[i]-gxzy[i])+uxz*gzzx[i]);
        Gyxz[i]=H*(uxy*gxxz[i]+uyy*(gxyz[i]+gyzx[i]-gxzy[i])+uyz*gzzx[i]);
        Gzxz[i]=H*(uxz*gxxz[i]+uyz*(gxyz[i]+gyzx[i]-gxzy[i])+uzz*gzzx[i]);
        /* yz */
        Gxyz[i]=H*(uxx*(gxyz[i]+gxzy[i]-gyzx[i])+uxy*gyyz[i]+uxz*gzzy[i]);
        Gyyz[i]=H*(uxy*(gxyz[i]+gxzy[i]-gyzx[i])+uyy*gyyz[i]+uyz*gzzy[i]);
        Gzyz[i]=H*(uxz*(gxyz[i]+gxzy[i]-gyzx[i])+uyz*gyyz[i]+uzz*gzzy[i]);
    }
}

/* Phase 5: A^ij = gup^ia gup^jb A_ab  (stored temporarily in Rxx..Rzz) */
__global__ void kern_phase5_raise_A(
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ Axx, const double* __restrict__ Axy,
    const double* __restrict__ Axz, const double* __restrict__ Ayy,
    const double* __restrict__ Ayz, const double* __restrict__ Azz,
    double* __restrict__ Rxx, double* __restrict__ Rxy, double* __restrict__ Rxz,
    double* __restrict__ Ryy, double* __restrict__ Ryz, double* __restrict__ Rzz)
{
    const double TWO = 2.0;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];
        Rxx[i]=uxx*uxx*Axx[i]+uxy*uxy*Ayy[i]+uxz*uxz*Azz[i]
              +TWO*(uxx*uxy*Axy[i]+uxx*uxz*Axz[i]+uxy*uxz*Ayz[i]);
        Ryy[i]=uxy*uxy*Axx[i]+uyy*uyy*Ayy[i]+uyz*uyz*Azz[i]
              +TWO*(uxy*uyy*Axy[i]+uxy*uyz*Axz[i]+uyy*uyz*Ayz[i]);
        Rzz[i]=uxz*uxz*Axx[i]+uyz*uyz*Ayy[i]+uzz*uzz*Azz[i]
              +TWO*(uxz*uyz*Axy[i]+uxz*uzz*Axz[i]+uyz*uzz*Ayz[i]);
        Rxy[i]=uxx*uxy*Axx[i]+uxy*uyy*Ayy[i]+uxz*uyz*Azz[i]
              +(uxx*uyy+uxy*uxy)*Axy[i]+(uxx*uyz+uxz*uxy)*Axz[i]+(uxy*uyz+uxz*uyy)*Ayz[i];
        Rxz[i]=uxx*uxz*Axx[i]+uxy*uyz*Ayy[i]+uxz*uzz*Azz[i]
              +(uxx*uyz+uxy*uxz)*Axy[i]+(uxx*uzz+uxz*uxz)*Axz[i]+(uxy*uzz+uxz*uyz)*Ayz[i];
        Ryz[i]=uxy*uxz*Axx[i]+uyy*uyz*Ayy[i]+uyz*uzz*Azz[i]
              +(uxy*uyz+uyy*uxz)*Axy[i]+(uxy*uzz+uyz*uxz)*Axz[i]+(uyy*uzz+uyz*uyz)*Ayz[i];
    }
}

/* Phase 6: Gamma_rhs part 1 (before fdderivs(beta) and fderivs(Gamma)) */
__global__ __launch_bounds__(128, 4)
void kern_phase6_gamma_rhs_part1(
    const double* __restrict__ Lapx, const double* __restrict__ Lapy,
    const double* __restrict__ Lapz,
    const double* __restrict__ alpn1, const double* __restrict__ chin1,
    const double* __restrict__ chix, const double* __restrict__ chiy,
    const double* __restrict__ chiz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ Kx, const double* __restrict__ Ky,
    const double* __restrict__ Kz,
    const double* __restrict__ Sx, const double* __restrict__ Sy,
    const double* __restrict__ Sz,
    const double* __restrict__ Rxx, const double* __restrict__ Rxy,
    const double* __restrict__ Rxz, const double* __restrict__ Ryy,
    const double* __restrict__ Ryz, const double* __restrict__ Rzz,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    double* __restrict__ Gamx_rhs, double* __restrict__ Gamy_rhs,
    double* __restrict__ Gamz_rhs)
{
    const double TWO=2.0, F3o2=1.5, F2o3=2.0/3.0, EIGHT=8.0;
    const double PI_V = 3.14159265358979323846;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];
        double lx=Lapx[i],ly=Lapy[i],lz=Lapz[i];
        double a=alpn1[i], c1=chin1[i];
        double cx=chix[i],cy=chiy[i],cz=chiz[i];

        Gamx_rhs[i] = -TWO*(lx*Rxx[i]+ly*Rxy[i]+lz*Rxz[i])
            + TWO*a*(
                -F3o2/c1*(cx*Rxx[i]+cy*Rxy[i]+cz*Rxz[i])
                -uxx*(F2o3*Kx[i]+EIGHT*PI_V*Sx[i])
                -uxy*(F2o3*Ky[i]+EIGHT*PI_V*Sy[i])
                -uxz*(F2o3*Kz[i]+EIGHT*PI_V*Sz[i])
                +Gxxx[i]*Rxx[i]+Gxyy[i]*Ryy[i]+Gxzz[i]*Rzz[i]
                +TWO*(Gxxy[i]*Rxy[i]+Gxxz[i]*Rxz[i]+Gxyz[i]*Ryz[i]));

        Gamy_rhs[i] = -TWO*(lx*Rxy[i]+ly*Ryy[i]+lz*Ryz[i])
            + TWO*a*(
                -F3o2/c1*(cx*Rxy[i]+cy*Ryy[i]+cz*Ryz[i])
                -uxy*(F2o3*Kx[i]+EIGHT*PI_V*Sx[i])
                -uyy*(F2o3*Ky[i]+EIGHT*PI_V*Sy[i])
                -uyz*(F2o3*Kz[i]+EIGHT*PI_V*Sz[i])
                +Gyxx[i]*Rxx[i]+Gyyy[i]*Ryy[i]+Gyzz[i]*Rzz[i]
                +TWO*(Gyxy[i]*Rxy[i]+Gyxz[i]*Rxz[i]+Gyyz[i]*Ryz[i]));

        Gamz_rhs[i] = -TWO*(lx*Rxz[i]+ly*Ryz[i]+lz*Rzz[i])
            + TWO*a*(
                -F3o2/c1*(cx*Rxz[i]+cy*Ryz[i]+cz*Rzz[i])
                -uxz*(F2o3*Kx[i]+EIGHT*PI_V*Sx[i])
                -uyz*(F2o3*Ky[i]+EIGHT*PI_V*Sy[i])
                -uzz*(F2o3*Kz[i]+EIGHT*PI_V*Sz[i])
                +Gzxx[i]*Rxx[i]+Gzyy[i]*Ryy[i]+Gzzz[i]*Rzz[i]
                +TWO*(Gzxy[i]*Rxy[i]+Gzxz[i]*Rxz[i]+Gzyz[i]*Ryz[i]));
    }
}

/* Phase 5+6 fused: raise A^ij in registers, then consume immediately in Gamma_rhs. */
__global__ __launch_bounds__(128, 4)
void kern_phase5_6_gamma_rhs_part1_fused(
    const double* __restrict__ Lapx, const double* __restrict__ Lapy,
    const double* __restrict__ Lapz,
    const double* __restrict__ alpn1, const double* __restrict__ chin1,
    const double* __restrict__ chix, const double* __restrict__ chiy,
    const double* __restrict__ chiz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ Axx, const double* __restrict__ Axy,
    const double* __restrict__ Axz, const double* __restrict__ Ayy,
    const double* __restrict__ Ayz, const double* __restrict__ Azz,
    const double* __restrict__ Kx, const double* __restrict__ Ky,
    const double* __restrict__ Kz,
    const double* __restrict__ Sx, const double* __restrict__ Sy,
    const double* __restrict__ Sz,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    double* __restrict__ Gamx_rhs, double* __restrict__ Gamy_rhs,
    double* __restrict__ Gamz_rhs)
{
    const double TWO = 2.0, F3o2 = 1.5, F2o3 = 2.0 / 3.0, EIGHT = 8.0;
    const double PI_V = 3.14159265358979323846;
    for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < d_gp.all; i += blockDim.x * gridDim.x) {
        const double uxx = gupxx[i], uxy = gupxy[i], uxz = gupxz[i];
        const double uyy = gupyy[i], uyz = gupyz[i], uzz = gupzz[i];
        const double Axx_v = Axx[i], Axy_v = Axy[i], Axz_v = Axz[i];
        const double Ayy_v = Ayy[i], Ayz_v = Ayz[i], Azz_v = Azz[i];

        const double Rxx_v = uxx * uxx * Axx_v + uxy * uxy * Ayy_v + uxz * uxz * Azz_v
                           + TWO * (uxx * uxy * Axy_v + uxx * uxz * Axz_v + uxy * uxz * Ayz_v);
        const double Ryy_v = uxy * uxy * Axx_v + uyy * uyy * Ayy_v + uyz * uyz * Azz_v
                           + TWO * (uxy * uyy * Axy_v + uxy * uyz * Axz_v + uyy * uyz * Ayz_v);
        const double Rzz_v = uxz * uxz * Axx_v + uyz * uyz * Ayy_v + uzz * uzz * Azz_v
                           + TWO * (uxz * uyz * Axy_v + uxz * uzz * Axz_v + uyz * uzz * Ayz_v);
        const double Rxy_v = uxx * uxy * Axx_v + uxy * uyy * Ayy_v + uxz * uyz * Azz_v
                           + (uxx * uyy + uxy * uxy) * Axy_v
                           + (uxx * uyz + uxz * uxy) * Axz_v
                           + (uxy * uyz + uxz * uyy) * Ayz_v;
        const double Rxz_v = uxx * uxz * Axx_v + uxy * uyz * Ayy_v + uxz * uzz * Azz_v
                           + (uxx * uyz + uxy * uxz) * Axy_v
                           + (uxx * uzz + uxz * uxz) * Axz_v
                           + (uxy * uzz + uxz * uyz) * Ayz_v;
        const double Ryz_v = uxy * uxz * Axx_v + uyy * uyz * Ayy_v + uyz * uzz * Azz_v
                           + (uxy * uyz + uyy * uxz) * Axy_v
                           + (uxy * uzz + uyz * uxz) * Axz_v
                           + (uyy * uzz + uyz * uyz) * Ayz_v;

        const double lx = Lapx[i], ly = Lapy[i], lz = Lapz[i];
        const double a = alpn1[i], c1 = chin1[i];
        const double cx = chix[i], cy = chiy[i], cz = chiz[i];

        Gamx_rhs[i] = -TWO * (lx * Rxx_v + ly * Rxy_v + lz * Rxz_v)
            + TWO * a * (
                -F3o2 / c1 * (cx * Rxx_v + cy * Rxy_v + cz * Rxz_v)
                -uxx * (F2o3 * Kx[i] + EIGHT * PI_V * Sx[i])
                -uxy * (F2o3 * Ky[i] + EIGHT * PI_V * Sy[i])
                -uxz * (F2o3 * Kz[i] + EIGHT * PI_V * Sz[i])
                + Gxxx[i] * Rxx_v + Gxyy[i] * Ryy_v + Gxzz[i] * Rzz_v
                + TWO * (Gxxy[i] * Rxy_v + Gxxz[i] * Rxz_v + Gxyz[i] * Ryz_v));

        Gamy_rhs[i] = -TWO * (lx * Rxy_v + ly * Ryy_v + lz * Ryz_v)
            + TWO * a * (
                -F3o2 / c1 * (cx * Rxy_v + cy * Ryy_v + cz * Ryz_v)
                -uxy * (F2o3 * Kx[i] + EIGHT * PI_V * Sx[i])
                -uyy * (F2o3 * Ky[i] + EIGHT * PI_V * Sy[i])
                -uyz * (F2o3 * Kz[i] + EIGHT * PI_V * Sz[i])
                + Gyxx[i] * Rxx_v + Gyyy[i] * Ryy_v + Gyzz[i] * Rzz_v
                + TWO * (Gyxy[i] * Rxy_v + Gyxz[i] * Rxz_v + Gyyz[i] * Ryz_v));

        Gamz_rhs[i] = -TWO * (lx * Rxz_v + ly * Ryz_v + lz * Rzz_v)
            + TWO * a * (
                -F3o2 / c1 * (cx * Rxz_v + cy * Ryz_v + cz * Rzz_v)
                -uxz * (F2o3 * Kx[i] + EIGHT * PI_V * Sx[i])
                -uyz * (F2o3 * Ky[i] + EIGHT * PI_V * Sy[i])
                -uzz * (F2o3 * Kz[i] + EIGHT * PI_V * Sz[i])
                + Gzxx[i] * Rxx_v + Gzyy[i] * Ryy_v + Gzzz[i] * Rzz_v
                + TWO * (Gzxy[i] * Rxy_v + Gzxz[i] * Rxz_v + Gzyz[i] * Ryz_v));
    }
}

/* Phase 8: Gamma_rhs part 2 — after fdderivs(beta) and fderivs(Gamma)
 * Computes: fxx=div(beta_xx), Gamxa, then updates Gamx_rhs etc.
 * Input arrays gxxx..gzzz here hold fdderivs(beta) results,
 * Gamxx..Gamzz hold fderivs(Gamma) results.
 */
__global__ __launch_bounds__(128, 4)
void kern_phase8_gamma_rhs_part2(
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    /* fdderivs(betax) -> gxxx,gxyx,gxzx,gyyx,gyzx,gzzx */
    const double* __restrict__ bxx_xx, const double* __restrict__ bxx_xy,
    const double* __restrict__ bxx_xz, const double* __restrict__ bxx_yy,
    const double* __restrict__ bxx_yz, const double* __restrict__ bxx_zz,
    /* fdderivs(betay) -> gxxy,gxyy,gxzy,gyyy,gyzy,gzzy */
    const double* __restrict__ bxy_xx, const double* __restrict__ bxy_xy,
    const double* __restrict__ bxy_xz, const double* __restrict__ bxy_yy,
    const double* __restrict__ bxy_yz, const double* __restrict__ bxy_zz,
    /* fdderivs(betaz) -> gxxz,gxyz,gxzz,gyyz,gyzz,gzzz */
    const double* __restrict__ bxz_xx, const double* __restrict__ bxz_xy,
    const double* __restrict__ bxz_xz, const double* __restrict__ bxz_yy,
    const double* __restrict__ bxz_yz, const double* __restrict__ bxz_zz,
    /* fderivs(Gamx) -> Gamxx,Gamxy,Gamxz */
    const double* __restrict__ Gamxx, const double* __restrict__ Gamxy,
    const double* __restrict__ Gamxz,
    /* fderivs(Gamy) -> Gamyx,Gamyy,Gamyz */
    const double* __restrict__ Gamyx, const double* __restrict__ Gamyy_d,
    const double* __restrict__ Gamyz_d,
    /* fderivs(Gamz) -> Gamzx,Gamzy,Gamzz */
    const double* __restrict__ Gamzx, const double* __restrict__ Gamzy,
    const double* __restrict__ Gamzz_d,
    /* Christoffel symbols */
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    /* betaij first derivs */
    const double* __restrict__ betaxx, const double* __restrict__ betaxy,
    const double* __restrict__ betaxz, const double* __restrict__ betayx,
    const double* __restrict__ betayy, const double* __restrict__ betayz,
    const double* __restrict__ betazx, const double* __restrict__ betazy,
    const double* __restrict__ betazz,
    double* __restrict__ Gamx_rhs, double* __restrict__ Gamy_rhs,
    double* __restrict__ Gamz_rhs,
    double* __restrict__ Gamxa_out, double* __restrict__ Gamya_out,
    double* __restrict__ Gamza_out)
{
    const double TWO=2.0, F2o3=2.0/3.0, F1o3=1.0/3.0;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];
        /* div(beta_second_derivs) */
        double fxx_v = bxx_xx[i]+bxy_xy[i]+bxz_xz[i];
        double fxy_v = bxx_xy[i]+bxy_yy[i]+bxz_yz[i];
        double fxz_v = bxx_xz[i]+bxy_yz[i]+bxz_zz[i];
        /* Gamma^a contracted */
        double Ga_x = uxx*Gxxx[i]+uyy*Gxyy[i]+uzz*Gxzz[i]
                     +TWO*(uxy*Gxxy[i]+uxz*Gxxz[i]+uyz*Gxyz[i]);
        double Ga_y = uxx*Gyxx[i]+uyy*Gyyy[i]+uzz*Gyzz[i]
                     +TWO*(uxy*Gyxy[i]+uxz*Gyxz[i]+uyz*Gyyz[i]);
        double Ga_z = uxx*Gzxx[i]+uyy*Gzyy[i]+uzz*Gzzz[i]
                     +TWO*(uxy*Gzxy[i]+uxz*Gzxz[i]+uyz*Gzyz[i]);
        Gamxa_out[i]=Ga_x; Gamya_out[i]=Ga_y; Gamza_out[i]=Ga_z;
        double db = betaxx[i] + betayy[i] + betazz[i];
        Gamx_rhs[i] += F2o3*Ga_x*db
            - Ga_x*betaxx[i] - Ga_y*betaxy[i] - Ga_z*betaxz[i]
            + F1o3*(uxx*fxx_v+uxy*fxy_v+uxz*fxz_v)
            + uxx*bxx_xx[i]+uyy*bxx_yy[i]+uzz*bxx_zz[i]
            + TWO*(uxy*bxx_xy[i]+uxz*bxx_xz[i]+uyz*bxx_yz[i]);
        Gamy_rhs[i] += F2o3*Ga_y*db
            - Ga_x*betayx[i] - Ga_y*betayy[i] - Ga_z*betayz[i]
            + F1o3*(uxy*fxx_v+uyy*fxy_v+uyz*fxz_v)
            + uxx*bxy_xx[i]+uyy*bxy_yy[i]+uzz*bxy_zz[i]
            + TWO*(uxy*bxy_xy[i]+uxz*bxy_xz[i]+uyz*bxy_yz[i]);
        Gamz_rhs[i] += F2o3*Ga_z*db
            - Ga_x*betazx[i] - Ga_y*betazy[i] - Ga_z*betazz[i]
            + F1o3*(uxz*fxx_v+uyz*fxy_v+uzz*fxz_v)
            + uxx*bxz_xx[i]+uyy*bxz_yy[i]+uzz*bxz_zz[i]
            + TWO*(uxy*bxz_xy[i]+uxz*bxz_xz[i]+uyz*bxz_yz[i]);
    }
}

/* Phase 9: Christoffel contract — compute g_{ia} Gamma^a_{bc} products
 * Overwrites gxxx..gzzz with lowered Christoffel products needed for Ricci.
 */
__global__ __launch_bounds__(128, 4)
void kern_phase9_christoffel_contract(
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    /* output: lowered products g_{ia} Gamma^a_{bc} */
    double* __restrict__ o_gxxx, double* __restrict__ o_gxyx,
    double* __restrict__ o_gxzx, double* __restrict__ o_gyyx,
    double* __restrict__ o_gyzx, double* __restrict__ o_gzzx,
    double* __restrict__ o_gxxy, double* __restrict__ o_gxyy,
    double* __restrict__ o_gxzy, double* __restrict__ o_gyyy,
    double* __restrict__ o_gyzy, double* __restrict__ o_gzzy,
    double* __restrict__ o_gxxz, double* __restrict__ o_gxyz,
    double* __restrict__ o_gxzz, double* __restrict__ o_gyyz,
    double* __restrict__ o_gyzz, double* __restrict__ o_gzzz)
{
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double g11=gxx[i],g12=gxy[i],g13=gxz[i];
        double g22=gyy[i],g23=gyz[i],g33=gzz[i];
        /* row x: g_{x,a} Gamma^a_{bc} */
        o_gxxx[i]=g11*Gxxx[i]+g12*Gyxx[i]+g13*Gzxx[i];
        o_gxyx[i]=g11*Gxxy[i]+g12*Gyxy[i]+g13*Gzxy[i];
        o_gxzx[i]=g11*Gxxz[i]+g12*Gyxz[i]+g13*Gzxz[i];
        o_gyyx[i]=g11*Gxyy[i]+g12*Gyyy[i]+g13*Gzyy[i];
        o_gyzx[i]=g11*Gxyz[i]+g12*Gyyz[i]+g13*Gzyz[i];
        o_gzzx[i]=g11*Gxzz[i]+g12*Gyzz[i]+g13*Gzzz[i];
        /* row y: g_{y,a} Gamma^a_{bc} */
        o_gxxy[i]=g12*Gxxx[i]+g22*Gyxx[i]+g23*Gzxx[i];
        o_gxyy[i]=g12*Gxxy[i]+g22*Gyxy[i]+g23*Gzxy[i];
        o_gxzy[i]=g12*Gxxz[i]+g22*Gyxz[i]+g23*Gzxz[i];
        o_gyyy[i]=g12*Gxyy[i]+g22*Gyyy[i]+g23*Gzyy[i];
        o_gyzy[i]=g12*Gxyz[i]+g22*Gyyz[i]+g23*Gzyz[i];
        o_gzzy[i]=g12*Gxzz[i]+g22*Gyzz[i]+g23*Gzzz[i];
        /* row z: g_{z,a} Gamma^a_{bc} */
        o_gxxz[i]=g13*Gxxx[i]+g23*Gyxx[i]+g33*Gzxx[i];
        o_gxyz[i]=g13*Gxxy[i]+g23*Gyxy[i]+g33*Gzxy[i];
        o_gxzz[i]=g13*Gxxz[i]+g23*Gyxz[i]+g33*Gzxz[i];
        o_gyyz[i]=g13*Gxyy[i]+g23*Gyyy[i]+g33*Gzyy[i];
        o_gyzz[i]=g13*Gxyz[i]+g23*Gyyz[i]+g33*Gzyz[i];
        o_gzzz[i]=g13*Gxzz[i]+g23*Gyzz[i]+g33*Gzzz[i];
    }
}

/* Phase 8+9 fused: update Gamma rhs, contract Gamma^a, and lower Christoffels in one pass. */
__global__ __launch_bounds__(128, 2)
void kern_phase8_9_gamma_rhs_contract_fused(
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ bxx_xx, const double* __restrict__ bxx_xy,
    const double* __restrict__ bxx_xz, const double* __restrict__ bxx_yy,
    const double* __restrict__ bxx_yz, const double* __restrict__ bxx_zz,
    const double* __restrict__ bxy_xx, const double* __restrict__ bxy_xy,
    const double* __restrict__ bxy_xz, const double* __restrict__ bxy_yy,
    const double* __restrict__ bxy_yz, const double* __restrict__ bxy_zz,
    const double* __restrict__ bxz_xx, const double* __restrict__ bxz_xy,
    const double* __restrict__ bxz_xz, const double* __restrict__ bxz_yy,
    const double* __restrict__ bxz_yz, const double* __restrict__ bxz_zz,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    const double* __restrict__ betaxx, const double* __restrict__ betaxy,
    const double* __restrict__ betaxz, const double* __restrict__ betayx,
    const double* __restrict__ betayy, const double* __restrict__ betayz,
    const double* __restrict__ betazx, const double* __restrict__ betazy,
    const double* __restrict__ betazz,
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    double* __restrict__ Gamx_rhs, double* __restrict__ Gamy_rhs,
    double* __restrict__ Gamz_rhs,
    double* __restrict__ Gamxa_out, double* __restrict__ Gamya_out,
    double* __restrict__ Gamza_out,
    double* __restrict__ o_gxxx, double* __restrict__ o_gxyx,
    double* __restrict__ o_gxzx, double* __restrict__ o_gyyx,
    double* __restrict__ o_gyzx, double* __restrict__ o_gzzx,
    double* __restrict__ o_gxxy, double* __restrict__ o_gxyy,
    double* __restrict__ o_gxzy, double* __restrict__ o_gyyy,
    double* __restrict__ o_gyzy, double* __restrict__ o_gzzy,
    double* __restrict__ o_gxxz, double* __restrict__ o_gxyz,
    double* __restrict__ o_gxzz, double* __restrict__ o_gyyz,
    double* __restrict__ o_gyzz, double* __restrict__ o_gzzz)
{
    const double TWO = 2.0, F2o3 = 2.0 / 3.0, F1o3 = 1.0 / 3.0;
    for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < d_gp.all; i += blockDim.x * gridDim.x) {
        const double uxx = gupxx[i], uxy = gupxy[i], uxz = gupxz[i];
        const double uyy = gupyy[i], uyz = gupyz[i], uzz = gupzz[i];

        const double Gxxx_v = Gxxx[i], Gxxy_v = Gxxy[i], Gxxz_v = Gxxz[i];
        const double Gxyy_v = Gxyy[i], Gxyz_v = Gxyz[i], Gxzz_v = Gxzz[i];
        const double Gyxx_v = Gyxx[i], Gyxy_v = Gyxy[i], Gyxz_v = Gyxz[i];
        const double Gyyy_v = Gyyy[i], Gyyz_v = Gyyz[i], Gyzz_v = Gyzz[i];
        const double Gzxx_v = Gzxx[i], Gzxy_v = Gzxy[i], Gzxz_v = Gzxz[i];
        const double Gzyy_v = Gzyy[i], Gzyz_v = Gzyz[i], Gzzz_v = Gzzz[i];

        const double fxx_v = bxx_xx[i] + bxy_xy[i] + bxz_xz[i];
        const double fxy_v = bxx_xy[i] + bxy_yy[i] + bxz_yz[i];
        const double fxz_v = bxx_xz[i] + bxy_yz[i] + bxz_zz[i];

        const double Ga_x = uxx * Gxxx_v + uyy * Gxyy_v + uzz * Gxzz_v
                          + TWO * (uxy * Gxxy_v + uxz * Gxxz_v + uyz * Gxyz_v);
        const double Ga_y = uxx * Gyxx_v + uyy * Gyyy_v + uzz * Gyzz_v
                          + TWO * (uxy * Gyxy_v + uxz * Gyxz_v + uyz * Gyyz_v);
        const double Ga_z = uxx * Gzxx_v + uyy * Gzyy_v + uzz * Gzzz_v
                          + TWO * (uxy * Gzxy_v + uxz * Gzxz_v + uyz * Gzyz_v);

        Gamxa_out[i] = Ga_x;
        Gamya_out[i] = Ga_y;
        Gamza_out[i] = Ga_z;

        const double betaxx_v = betaxx[i], betaxy_v = betaxy[i], betaxz_v = betaxz[i];
        const double betayx_v = betayx[i], betayy_v = betayy[i], betayz_v = betayz[i];
        const double betazx_v = betazx[i], betazy_v = betazy[i], betazz_v = betazz[i];
        const double db = betaxx_v + betayy_v + betazz_v;

        Gamx_rhs[i] += F2o3 * Ga_x * db
            - Ga_x * betaxx_v - Ga_y * betaxy_v - Ga_z * betaxz_v
            + F1o3 * (uxx * fxx_v + uxy * fxy_v + uxz * fxz_v)
            + uxx * bxx_xx[i] + uyy * bxx_yy[i] + uzz * bxx_zz[i]
            + TWO * (uxy * bxx_xy[i] + uxz * bxx_xz[i] + uyz * bxx_yz[i]);
        Gamy_rhs[i] += F2o3 * Ga_y * db
            - Ga_x * betayx_v - Ga_y * betayy_v - Ga_z * betayz_v
            + F1o3 * (uxy * fxx_v + uyy * fxy_v + uyz * fxz_v)
            + uxx * bxy_xx[i] + uyy * bxy_yy[i] + uzz * bxy_zz[i]
            + TWO * (uxy * bxy_xy[i] + uxz * bxy_xz[i] + uyz * bxy_yz[i]);
        Gamz_rhs[i] += F2o3 * Ga_z * db
            - Ga_x * betazx_v - Ga_y * betazy_v - Ga_z * betazz_v
            + F1o3 * (uxz * fxx_v + uyz * fxy_v + uzz * fxz_v)
            + uxx * bxz_xx[i] + uyy * bxz_yy[i] + uzz * bxz_zz[i]
            + TWO * (uxy * bxz_xy[i] + uxz * bxz_xz[i] + uyz * bxz_yz[i]);

        const double g11 = gxx[i], g12 = gxy[i], g13 = gxz[i];
        const double g22 = gyy[i], g23 = gyz[i], g33 = gzz[i];

        o_gxxx[i] = g11 * Gxxx_v + g12 * Gyxx_v + g13 * Gzxx_v;
        o_gxyx[i] = g11 * Gxxy_v + g12 * Gyxy_v + g13 * Gzxy_v;
        o_gxzx[i] = g11 * Gxxz_v + g12 * Gyxz_v + g13 * Gzxz_v;
        o_gyyx[i] = g11 * Gxyy_v + g12 * Gyyy_v + g13 * Gzyy_v;
        o_gyzx[i] = g11 * Gxyz_v + g12 * Gyyz_v + g13 * Gzyz_v;
        o_gzzx[i] = g11 * Gxzz_v + g12 * Gyzz_v + g13 * Gzzz_v;

        o_gxxy[i] = g12 * Gxxx_v + g22 * Gyxx_v + g23 * Gzxx_v;
        o_gxyy[i] = g12 * Gxxy_v + g22 * Gyxy_v + g23 * Gzxy_v;
        o_gxzy[i] = g12 * Gxxz_v + g22 * Gyxz_v + g23 * Gzxz_v;
        o_gyyy[i] = g12 * Gxyy_v + g22 * Gyyy_v + g23 * Gzyy_v;
        o_gyzy[i] = g12 * Gxyz_v + g22 * Gyyz_v + g23 * Gzyz_v;
        o_gzzy[i] = g12 * Gxzz_v + g22 * Gyzz_v + g23 * Gzzz_v;

        o_gxxz[i] = g13 * Gxxx_v + g23 * Gyxx_v + g33 * Gzxx_v;
        o_gxyz[i] = g13 * Gxxy_v + g23 * Gyxy_v + g33 * Gzxy_v;
        o_gxzz[i] = g13 * Gxxz_v + g23 * Gyxz_v + g33 * Gzxz_v;
        o_gyyz[i] = g13 * Gxyy_v + g23 * Gyyy_v + g33 * Gzyy_v;
        o_gyzz[i] = g13 * Gxyz_v + g23 * Gyyz_v + g33 * Gzyz_v;
        o_gzzz[i] = g13 * Gxzz_v + g23 * Gyzz_v + g33 * Gzzz_v;
    }
}

/* Phase 10: After fdderivs of a metric component, contract with gup^{ij}
 * R_comp = gup^xx*fxx + gup^yy*fyy + gup^zz*fzz + 2*(gup^xy*fxy + gup^xz*fxz + gup^yz*fyz)
 */
__global__ void kern_phase10_ricci_contract(
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ fxx, const double* __restrict__ fxy,
    const double* __restrict__ fxz, const double* __restrict__ fyy,
    const double* __restrict__ fyz, const double* __restrict__ fzz,
    double* __restrict__ R_comp)
{
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        R_comp[i] = gupxx[i]*fxx[i] + gupyy[i]*fyy[i] + gupzz[i]*fzz[i]
                  + 2.0*(gupxy[i]*fxy[i] + gupxz[i]*fxz[i] + gupyz[i]*fyz[i]);
    }
}

/* Phase 11a: Ricci diagonal assembly (Rxx, Ryy, Rzz) */
__global__ __launch_bounds__(128, 4)
void kern_phase11_ricci_diag(
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ Gamxa, const double* __restrict__ Gamya,
    const double* __restrict__ Gamza,
    const double* __restrict__ Gamxx, const double* __restrict__ Gamxy,
    const double* __restrict__ Gamxz,
    const double* __restrict__ Gamyx, const double* __restrict__ Gamyy_d,
    const double* __restrict__ Gamyz_d,
    const double* __restrict__ Gamzx, const double* __restrict__ Gamzy,
    const double* __restrict__ Gamzz_d,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    /* lowered Christoffel products */
    const double* __restrict__ lxxx, const double* __restrict__ lxyx,
    const double* __restrict__ lxzx, const double* __restrict__ lyyx,
    const double* __restrict__ lyzx, const double* __restrict__ lzzx,
    const double* __restrict__ lxxy, const double* __restrict__ lxyy,
    const double* __restrict__ lxzy, const double* __restrict__ lyyy,
    const double* __restrict__ lyzy, const double* __restrict__ lzzy,
    const double* __restrict__ lxxz, const double* __restrict__ lxyz,
    const double* __restrict__ lxzz, const double* __restrict__ lyyz,
    const double* __restrict__ lyzz, const double* __restrict__ lzzz,
    double* __restrict__ Rxx, double* __restrict__ Ryy, double* __restrict__ Rzz)
{
    const double H = 0.5, TWO = 2.0;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];
        /* Rxx */
        Rxx[i] = -H*Rxx[i]
            + gxx[i]*Gamxx[i]+gxy[i]*Gamyx[i]+gxz[i]*Gamzx[i]
            + Gamxa[i]*lxxx[i]+Gamya[i]*lxyx[i]+Gamza[i]*lxzx[i]
            + uxx*(TWO*(Gxxx[i]*lxxx[i]+Gyxx[i]*lxyx[i]+Gzxx[i]*lxzx[i])
                  +(Gxxx[i]*lxxx[i]+Gyxx[i]*lxxy[i]+Gzxx[i]*lxxz[i]))
            + uxy*(TWO*(Gxxx[i]*lxyx[i]+Gyxx[i]*lyyx[i]+Gzxx[i]*lyzx[i]
                       +Gxxy[i]*lxxx[i]+Gyxy[i]*lxyx[i]+Gzxy[i]*lxzx[i])
                  +(Gxxy[i]*lxxx[i]+Gyxy[i]*lxxy[i]+Gzxy[i]*lxxz[i])
                  +(Gxxx[i]*lxyx[i]+Gyxx[i]*lxyy[i]+Gzxx[i]*lxyz[i]))
            + uxz*(TWO*(Gxxx[i]*lxzx[i]+Gyxx[i]*lyzx[i]+Gzxx[i]*lzzx[i]
                       +Gxxz[i]*lxxx[i]+Gyxz[i]*lxyx[i]+Gzxz[i]*lxzx[i])
                  +(Gxxz[i]*lxxx[i]+Gyxz[i]*lxxy[i]+Gzxz[i]*lxxz[i])
                  +(Gxxx[i]*lxzx[i]+Gyxx[i]*lxzy[i]+Gzxx[i]*lxzz[i]))
            + uyy*(TWO*(Gxxy[i]*lxyx[i]+Gyxy[i]*lyyx[i]+Gzxy[i]*lyzx[i])
                  +(Gxxy[i]*lxyx[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxyz[i]))
            + uyz*(TWO*(Gxxy[i]*lxzx[i]+Gyxy[i]*lyzx[i]+Gzxy[i]*lzzx[i]
                       +Gxxz[i]*lxyx[i]+Gyxz[i]*lyyx[i]+Gzxz[i]*lyzx[i])
                  +(Gxxz[i]*lxyx[i]+Gyxz[i]*lxyy[i]+Gzxz[i]*lxyz[i])
                  +(Gxxy[i]*lxzx[i]+Gyxy[i]*lxzy[i]+Gzxy[i]*lxzz[i]))
            + uzz*(TWO*(Gxxz[i]*lxzx[i]+Gyxz[i]*lyzx[i]+Gzxz[i]*lzzx[i])
                  +(Gxxz[i]*lxzx[i]+Gyxz[i]*lxzy[i]+Gzxz[i]*lxzz[i]));

        /* Ryy */
        Ryy[i] = -H*Ryy[i]
            + gxy[i]*Gamxy[i]+gyy[i]*Gamyy_d[i]+gyz[i]*Gamzy[i]
            + Gamxa[i]*lxyy[i]+Gamya[i]*lyyy[i]+Gamza[i]*lyzy[i]
            + uxx*(TWO*(Gxxy[i]*lxxy[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxzy[i])
                  +(Gxxy[i]*lxyx[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxyz[i]))
            + uxy*(TWO*(Gxxy[i]*lxyy[i]+Gyxy[i]*lyyy[i]+Gzxy[i]*lyzy[i]
                       +Gxyy[i]*lxxy[i]+Gyyy[i]*lxyy[i]+Gzyy[i]*lxzy[i])
                  +(Gxyy[i]*lxyx[i]+Gyyy[i]*lxyy[i]+Gzyy[i]*lxyz[i])
                  +(Gxxy[i]*lyyx[i]+Gyxy[i]*lyyy[i]+Gzxy[i]*lyyz[i]))
            + uxz*(TWO*(Gxxy[i]*lxzy[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lzzy[i]
                       +Gxyz[i]*lxxy[i]+Gyyz[i]*lxyy[i]+Gzyz[i]*lxzy[i])
                  +(Gxyz[i]*lxyx[i]+Gyyz[i]*lxyy[i]+Gzyz[i]*lxyz[i])
                  +(Gxxy[i]*lyzx[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lyzz[i]))
            + uyy*(TWO*(Gxyy[i]*lxyy[i]+Gyyy[i]*lyyy[i]+Gzyy[i]*lyzy[i])
                  +(Gxyy[i]*lyyx[i]+Gyyy[i]*lyyy[i]+Gzyy[i]*lyyz[i]))
            + uyz*(TWO*(Gxyy[i]*lxzy[i]+Gyyy[i]*lyzy[i]+Gzyy[i]*lzzy[i]
                       +Gxyz[i]*lxyy[i]+Gyyz[i]*lyyy[i]+Gzyz[i]*lyzy[i])
                  +(Gxyz[i]*lyyx[i]+Gyyz[i]*lyyy[i]+Gzyz[i]*lyyz[i])
                  +(Gxyy[i]*lyzx[i]+Gyyy[i]*lyzy[i]+Gzyy[i]*lyzz[i]))
            + uzz*(TWO*(Gxyz[i]*lxzy[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lzzy[i])
                  +(Gxyz[i]*lyzx[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lyzz[i]));

        /* Rzz */
        Rzz[i] = -H*Rzz[i]
            + gxz[i]*Gamxz[i]+gyz[i]*Gamyz_d[i]+gzz[i]*Gamzz_d[i]
            + Gamxa[i]*lxzz[i]+Gamya[i]*lyzz[i]+Gamza[i]*lzzz[i]
            + uxx*(TWO*(Gxxz[i]*lxxz[i]+Gyxz[i]*lxyz[i]+Gzxz[i]*lxzz[i])
                  +(Gxxz[i]*lxzx[i]+Gyxz[i]*lxzy[i]+Gzxz[i]*lxzz[i]))
            + uxy*(TWO*(Gxxz[i]*lxyz[i]+Gyxz[i]*lyyz[i]+Gzxz[i]*lyzz[i]
                       +Gxyz[i]*lxxz[i]+Gyyz[i]*lxyz[i]+Gzyz[i]*lxzz[i])
                  +(Gxyz[i]*lxzx[i]+Gyyz[i]*lxzy[i]+Gzyz[i]*lxzz[i])
                  +(Gxxz[i]*lyzx[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lyzz[i]))
            + uxz*(TWO*(Gxxz[i]*lxzz[i]+Gyxz[i]*lyzz[i]+Gzxz[i]*lzzz[i]
                       +Gxzz[i]*lxxz[i]+Gyzz[i]*lxyz[i]+Gzzz[i]*lxzz[i])
                  +(Gxzz[i]*lxzx[i]+Gyzz[i]*lxzy[i]+Gzzz[i]*lxzz[i])
                  +(Gxxz[i]*lzzx[i]+Gyxz[i]*lzzy[i]+Gzxz[i]*lzzz[i]))
            + uyy*(TWO*(Gxyz[i]*lxyz[i]+Gyyz[i]*lyyz[i]+Gzyz[i]*lyzz[i])
                  +(Gxyz[i]*lyzx[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lyzz[i]))
            + uyz*(TWO*(Gxyz[i]*lxzz[i]+Gyyz[i]*lyzz[i]+Gzyz[i]*lzzz[i]
                       +Gxzz[i]*lxyz[i]+Gyzz[i]*lyyz[i]+Gzzz[i]*lyzz[i])
                  +(Gxzz[i]*lyzx[i]+Gyzz[i]*lyzy[i]+Gzzz[i]*lyzz[i])
                  +(Gxyz[i]*lzzx[i]+Gyyz[i]*lzzy[i]+Gzyz[i]*lzzz[i]))
            + uzz*(TWO*(Gxzz[i]*lxzz[i]+Gyzz[i]*lyzz[i]+Gzzz[i]*lzzz[i])
                  +(Gxzz[i]*lzzx[i]+Gyzz[i]*lzzy[i]+Gzzz[i]*lzzz[i]));
    }
}

/* Phase 11b: Ricci off-diagonal assembly (Rxy, Rxz, Ryz) */
__global__ __launch_bounds__(128, 4)
void kern_phase11_ricci_offdiag(
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ Gamxa, const double* __restrict__ Gamya,
    const double* __restrict__ Gamza,
    const double* __restrict__ Gamxx, const double* __restrict__ Gamxy,
    const double* __restrict__ Gamxz,
    const double* __restrict__ Gamyx, const double* __restrict__ Gamyy_d,
    const double* __restrict__ Gamyz_d,
    const double* __restrict__ Gamzx, const double* __restrict__ Gamzy,
    const double* __restrict__ Gamzz_d,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    const double* __restrict__ lxxx, const double* __restrict__ lxyx,
    const double* __restrict__ lxzx, const double* __restrict__ lyyx,
    const double* __restrict__ lyzx, const double* __restrict__ lzzx,
    const double* __restrict__ lxxy, const double* __restrict__ lxyy,
    const double* __restrict__ lxzy, const double* __restrict__ lyyy,
    const double* __restrict__ lyzy, const double* __restrict__ lzzy,
    const double* __restrict__ lxxz, const double* __restrict__ lxyz,
    const double* __restrict__ lxzz, const double* __restrict__ lyyz,
    const double* __restrict__ lyzz, const double* __restrict__ lzzz,
    double* __restrict__ Rxy, double* __restrict__ Rxz, double* __restrict__ Ryz)
{
    const double H = 0.5, TWO = 2.0;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];

        /* Rxy */
        Rxy[i] = H*(
            -Rxy[i]
            +gxx[i]*Gamxy[i]+gxy[i]*Gamyy_d[i]+gxz[i]*Gamzy[i]
            +gxy[i]*Gamxx[i]+gyy[i]*Gamyx[i]+gyz[i]*Gamzx[i]
            +Gamxa[i]*lxyx[i]+Gamya[i]*lyyx[i]+Gamza[i]*lyzx[i]
            +Gamxa[i]*lxxy[i]+Gamya[i]*lxyy[i]+Gamza[i]*lxzy[i])
            +uxx*(Gxxx[i]*lxxy[i]+Gyxx[i]*lxyy[i]+Gzxx[i]*lxzy[i]
                  +Gxxy[i]*lxxx[i]+Gyxy[i]*lxyx[i]+Gzxy[i]*lxzx[i]
                  +Gxxx[i]*lxyx[i]+Gyxx[i]*lxyy[i]+Gzxx[i]*lxyz[i])
            +uxy*(Gxxx[i]*lxyy[i]+Gyxx[i]*lyyy[i]+Gzxx[i]*lyzy[i]
                  +Gxxy[i]*lxyx[i]+Gyxy[i]*lyyx[i]+Gzxy[i]*lyzx[i]
                  +Gxxy[i]*lxyx[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxyz[i]
                  +Gxxy[i]*lxxy[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxzy[i]
                  +Gxyy[i]*lxxx[i]+Gyyy[i]*lxyx[i]+Gzyy[i]*lxzx[i]
                  +Gxxx[i]*lyyx[i]+Gyxx[i]*lyyy[i]+Gzxx[i]*lyyz[i])
            +uxz*(Gxxx[i]*lxzy[i]+Gyxx[i]*lyzy[i]+Gzxx[i]*lzzy[i]
                  +Gxxy[i]*lxzx[i]+Gyxy[i]*lyzx[i]+Gzxy[i]*lzzx[i]
                  +Gxxz[i]*lxyx[i]+Gyxz[i]*lxyy[i]+Gzxz[i]*lxyz[i]
                  +Gxxz[i]*lxxy[i]+Gyxz[i]*lxyy[i]+Gzxz[i]*lxzy[i]
                  +Gxyz[i]*lxxx[i]+Gyyz[i]*lxyx[i]+Gzyz[i]*lxzx[i]
                  +Gxxx[i]*lyzx[i]+Gyxx[i]*lyzy[i]+Gzxx[i]*lyzz[i])
            +uyy*(Gxxy[i]*lxyy[i]+Gyxy[i]*lyyy[i]+Gzxy[i]*lyzy[i]
                  +Gxyy[i]*lxyx[i]+Gyyy[i]*lyyx[i]+Gzyy[i]*lyzx[i]
                  +Gxxy[i]*lyyx[i]+Gyxy[i]*lyyy[i]+Gzxy[i]*lyyz[i])
            +uyz*(Gxxy[i]*lxzy[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lzzy[i]
                  +Gxyy[i]*lxzx[i]+Gyyy[i]*lyzx[i]+Gzyy[i]*lzzx[i]
                  +Gxxz[i]*lyyx[i]+Gyxz[i]*lyyy[i]+Gzxz[i]*lyyz[i]
                  +Gxxz[i]*lxyy[i]+Gyxz[i]*lyyy[i]+Gzxz[i]*lyzy[i]
                  +Gxyz[i]*lxyx[i]+Gyyz[i]*lyyx[i]+Gzyz[i]*lyzx[i]
                  +Gxxy[i]*lyzx[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lyzz[i])
            +uzz*(Gxxz[i]*lxzy[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lzzy[i]
                  +Gxyz[i]*lxzx[i]+Gyyz[i]*lyzx[i]+Gzyz[i]*lzzx[i]
                  +Gxxz[i]*lyzx[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lyzz[i]);

        /* Rxz */
        Rxz[i] = H*(
            -Rxz[i]
            +gxx[i]*Gamxz[i]+gxy[i]*Gamyz_d[i]+gxz[i]*Gamzz_d[i]
            +gxz[i]*Gamxx[i]+gyz[i]*Gamyx[i]+gzz[i]*Gamzx[i]
            +Gamxa[i]*lxzx[i]+Gamya[i]*lyzx[i]+Gamza[i]*lzzx[i]
            +Gamxa[i]*lxxz[i]+Gamya[i]*lxyz[i]+Gamza[i]*lxzz[i])
            +uxx*(Gxxx[i]*lxxz[i]+Gyxx[i]*lxyz[i]+Gzxx[i]*lxzz[i]
                  +Gxxz[i]*lxxx[i]+Gyxz[i]*lxyx[i]+Gzxz[i]*lxzx[i]
                  +Gxxx[i]*lxzx[i]+Gyxx[i]*lxzy[i]+Gzxx[i]*lxzz[i])
            +uxy*(Gxxx[i]*lxyz[i]+Gyxx[i]*lyyz[i]+Gzxx[i]*lyzz[i]
                  +Gxxz[i]*lxyx[i]+Gyxz[i]*lyyx[i]+Gzxz[i]*lyzx[i]
                  +Gxxy[i]*lxzx[i]+Gyxy[i]*lxzy[i]+Gzxy[i]*lxzz[i]
                  +Gxxy[i]*lxxz[i]+Gyxy[i]*lxyz[i]+Gzxy[i]*lxzz[i]
                  +Gxyz[i]*lxxx[i]+Gyyz[i]*lxyx[i]+Gzyz[i]*lxzx[i]
                  +Gxxx[i]*lyzx[i]+Gyxx[i]*lyzy[i]+Gzxx[i]*lyzz[i])
            +uxz*(Gxxx[i]*lxzz[i]+Gyxx[i]*lyzz[i]+Gzxx[i]*lzzz[i]
                  +Gxxz[i]*lxzx[i]+Gyxz[i]*lyzx[i]+Gzxz[i]*lzzx[i]
                  +Gxxz[i]*lxzx[i]+Gyxz[i]*lxzy[i]+Gzxz[i]*lxzz[i]
                  +Gxxz[i]*lxxz[i]+Gyxz[i]*lxyz[i]+Gzxz[i]*lxzz[i]
                  +Gxzz[i]*lxxx[i]+Gyzz[i]*lxyx[i]+Gzzz[i]*lxzx[i]
                  +Gxxx[i]*lzzx[i]+Gyxx[i]*lzzy[i]+Gzxx[i]*lzzz[i])
            +uyy*(Gxxy[i]*lxyz[i]+Gyxy[i]*lyyz[i]+Gzxy[i]*lyzz[i]
                  +Gxyz[i]*lxyx[i]+Gyyz[i]*lyyx[i]+Gzyz[i]*lyzx[i]
                  +Gxxy[i]*lyzx[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lyzz[i])
            +uyz*(Gxxy[i]*lxzz[i]+Gyxy[i]*lyzz[i]+Gzxy[i]*lzzz[i]
                  +Gxyz[i]*lxzx[i]+Gyyz[i]*lyzx[i]+Gzyz[i]*lzzx[i]
                  +Gxxz[i]*lyzx[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lyzz[i]
                  +Gxxz[i]*lxyz[i]+Gyxz[i]*lyyz[i]+Gzxz[i]*lyzz[i]
                  +Gxzz[i]*lxyx[i]+Gyzz[i]*lyyx[i]+Gzzz[i]*lyzx[i]
                  +Gxxy[i]*lzzx[i]+Gyxy[i]*lzzy[i]+Gzxy[i]*lzzz[i])
            +uzz*(Gxxz[i]*lxzz[i]+Gyxz[i]*lyzz[i]+Gzxz[i]*lzzz[i]
                  +Gxzz[i]*lxzx[i]+Gyzz[i]*lyzx[i]+Gzzz[i]*lzzx[i]
                  +Gxxz[i]*lzzx[i]+Gyxz[i]*lzzy[i]+Gzxz[i]*lzzz[i]);

        /* Ryz */
        Ryz[i] = H*(
            -Ryz[i]
            +gxy[i]*Gamxz[i]+gyy[i]*Gamyz_d[i]+gyz[i]*Gamzz_d[i]
            +gxz[i]*Gamxy[i]+gyz[i]*Gamyy_d[i]+gzz[i]*Gamzy[i]
            +Gamxa[i]*lxzy[i]+Gamya[i]*lyzy[i]+Gamza[i]*lzzy[i]
            +Gamxa[i]*lxyz[i]+Gamya[i]*lyyz[i]+Gamza[i]*lyzz[i])
            +uxx*(Gxxy[i]*lxxz[i]+Gyxy[i]*lxyz[i]+Gzxy[i]*lxzz[i]
                  +Gxxz[i]*lxxy[i]+Gyxz[i]*lxyy[i]+Gzxz[i]*lxzy[i]
                  +Gxxy[i]*lxzx[i]+Gyxy[i]*lxzy[i]+Gzxy[i]*lxzz[i])
            +uxy*(Gxxy[i]*lxyz[i]+Gyxy[i]*lyyz[i]+Gzxy[i]*lyzz[i]
                  +Gxxz[i]*lxyy[i]+Gyxz[i]*lyyy[i]+Gzxz[i]*lyzy[i]
                  +Gxyy[i]*lxzx[i]+Gyyy[i]*lxzy[i]+Gzyy[i]*lxzz[i]
                  +Gxyy[i]*lxxz[i]+Gyyy[i]*lxyz[i]+Gzyy[i]*lxzz[i]
                  +Gxyz[i]*lxxy[i]+Gyyz[i]*lxyy[i]+Gzyz[i]*lxzy[i]
                  +Gxxy[i]*lyzx[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lyzz[i])
            +uxz*(Gxxy[i]*lxzz[i]+Gyxy[i]*lyzz[i]+Gzxy[i]*lzzz[i]
                  +Gxxz[i]*lxzy[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lzzy[i]
                  +Gxyz[i]*lxzx[i]+Gyyz[i]*lxzy[i]+Gzyz[i]*lxzz[i]
                  +Gxyz[i]*lxxz[i]+Gyyz[i]*lxyz[i]+Gzyz[i]*lxzz[i]
                  +Gxzz[i]*lxxy[i]+Gyzz[i]*lxyy[i]+Gzzz[i]*lxzy[i]
                  +Gxxy[i]*lzzx[i]+Gyxy[i]*lzzy[i]+Gzxy[i]*lzzz[i])
            +uyy*(Gxyy[i]*lxyz[i]+Gyyy[i]*lyyz[i]+Gzyy[i]*lyzz[i]
                  +Gxyz[i]*lxyy[i]+Gyyz[i]*lyyy[i]+Gzyz[i]*lyzy[i]
                  +Gxyy[i]*lyzx[i]+Gyyy[i]*lyzy[i]+Gzyy[i]*lyzz[i])
            +uyz*(Gxyy[i]*lxzz[i]+Gyyy[i]*lyzz[i]+Gzyy[i]*lzzz[i]
                  +Gxyz[i]*lxzy[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lzzy[i]
                  +Gxyz[i]*lyzx[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lyzz[i]
                  +Gxyz[i]*lxyz[i]+Gyyz[i]*lyyz[i]+Gzyz[i]*lyzz[i]
                  +Gxzz[i]*lxyy[i]+Gyzz[i]*lyyy[i]+Gzzz[i]*lyzy[i]
                  +Gxyy[i]*lzzx[i]+Gyyy[i]*lzzy[i]+Gzyy[i]*lzzz[i])
            +uzz*(Gxyz[i]*lxzz[i]+Gyyz[i]*lyzz[i]+Gzyz[i]*lzzz[i]
                  +Gxzz[i]*lxzy[i]+Gyzz[i]*lyzy[i]+Gzzz[i]*lzzy[i]
                  +Gxyz[i]*lzzx[i]+Gyyz[i]*lzzy[i]+Gzyz[i]*lzzz[i]);
    }
}

/* Phase 11: fused Ricci assembly (diag + off-diag) */
__global__ __launch_bounds__(128, 2)
void kern_phase11_ricci_fused(
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ Gamxa, const double* __restrict__ Gamya,
    const double* __restrict__ Gamza,
    const double* __restrict__ Gamxx, const double* __restrict__ Gamxy,
    const double* __restrict__ Gamxz,
    const double* __restrict__ Gamyx, const double* __restrict__ Gamyy_d,
    const double* __restrict__ Gamyz_d,
    const double* __restrict__ Gamzx, const double* __restrict__ Gamzy,
    const double* __restrict__ Gamzz_d,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    const double* __restrict__ lxxx, const double* __restrict__ lxyx,
    const double* __restrict__ lxzx, const double* __restrict__ lyyx,
    const double* __restrict__ lyzx, const double* __restrict__ lzzx,
    const double* __restrict__ lxxy, const double* __restrict__ lxyy,
    const double* __restrict__ lxzy, const double* __restrict__ lyyy,
    const double* __restrict__ lyzy, const double* __restrict__ lzzy,
    const double* __restrict__ lxxz, const double* __restrict__ lxyz,
    const double* __restrict__ lxzz, const double* __restrict__ lyyz,
    const double* __restrict__ lyzz, const double* __restrict__ lzzz,
    double* __restrict__ Rxx, double* __restrict__ Rxy, double* __restrict__ Rxz,
    double* __restrict__ Ryy, double* __restrict__ Ryz, double* __restrict__ Rzz)
{
    const double H = 0.5, TWO = 2.0;
    for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < d_gp.all; i += blockDim.x * gridDim.x) {
        double uxx = gupxx[i], uxy = gupxy[i], uxz = gupxz[i];
        double uyy = gupyy[i], uyz = gupyz[i], uzz = gupzz[i];

        Rxx[i] = -H*Rxx[i]
            + gxx[i]*Gamxx[i]+gxy[i]*Gamyx[i]+gxz[i]*Gamzx[i]
            + Gamxa[i]*lxxx[i]+Gamya[i]*lxyx[i]+Gamza[i]*lxzx[i]
            + uxx*(TWO*(Gxxx[i]*lxxx[i]+Gyxx[i]*lxyx[i]+Gzxx[i]*lxzx[i])
                  +(Gxxx[i]*lxxx[i]+Gyxx[i]*lxxy[i]+Gzxx[i]*lxxz[i]))
            + uxy*(TWO*(Gxxx[i]*lxyx[i]+Gyxx[i]*lyyx[i]+Gzxx[i]*lyzx[i]
                       +Gxxy[i]*lxxx[i]+Gyxy[i]*lxyx[i]+Gzxy[i]*lxzx[i])
                  +(Gxxy[i]*lxxx[i]+Gyxy[i]*lxxy[i]+Gzxy[i]*lxxz[i])
                  +(Gxxx[i]*lxyx[i]+Gyxx[i]*lxyy[i]+Gzxx[i]*lxyz[i]))
            + uxz*(TWO*(Gxxx[i]*lxzx[i]+Gyxx[i]*lyzx[i]+Gzxx[i]*lzzx[i]
                       +Gxxz[i]*lxxx[i]+Gyxz[i]*lxyx[i]+Gzxz[i]*lxzx[i])
                  +(Gxxz[i]*lxxx[i]+Gyxz[i]*lxxy[i]+Gzxz[i]*lxxz[i])
                  +(Gxxx[i]*lxzx[i]+Gyxx[i]*lxzy[i]+Gzxx[i]*lxzz[i]))
            + uyy*(TWO*(Gxxy[i]*lxyx[i]+Gyxy[i]*lyyx[i]+Gzxy[i]*lyzx[i])
                  +(Gxxy[i]*lxyx[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxyz[i]))
            + uyz*(TWO*(Gxxy[i]*lxzx[i]+Gyxy[i]*lyzx[i]+Gzxy[i]*lzzx[i]
                       +Gxxz[i]*lxyx[i]+Gyxz[i]*lyyx[i]+Gzxz[i]*lyzx[i])
                  +(Gxxz[i]*lxyx[i]+Gyxz[i]*lxyy[i]+Gzxz[i]*lxyz[i])
                  +(Gxxy[i]*lxzx[i]+Gyxy[i]*lxzy[i]+Gzxy[i]*lxzz[i]))
            + uzz*(TWO*(Gxxz[i]*lxzx[i]+Gyxz[i]*lyzx[i]+Gzxz[i]*lzzx[i])
                  +(Gxxz[i]*lxzx[i]+Gyxz[i]*lxzy[i]+Gzxz[i]*lxzz[i]));

        Ryy[i] = -H*Ryy[i]
            + gxy[i]*Gamxy[i]+gyy[i]*Gamyy_d[i]+gyz[i]*Gamzy[i]
            + Gamxa[i]*lxyy[i]+Gamya[i]*lyyy[i]+Gamza[i]*lyzy[i]
            + uxx*(TWO*(Gxxy[i]*lxxy[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxzy[i])
                  +(Gxxy[i]*lxyx[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxyz[i]))
            + uxy*(TWO*(Gxxy[i]*lxyy[i]+Gyxy[i]*lyyy[i]+Gzxy[i]*lyzy[i]
                       +Gxyy[i]*lxxy[i]+Gyyy[i]*lxyy[i]+Gzyy[i]*lxzy[i])
                  +(Gxyy[i]*lxyx[i]+Gyyy[i]*lxyy[i]+Gzyy[i]*lxyz[i])
                  +(Gxxy[i]*lyyx[i]+Gyxy[i]*lyyy[i]+Gzxy[i]*lyyz[i]))
            + uxz*(TWO*(Gxxy[i]*lxzy[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lzzy[i]
                       +Gxyz[i]*lxxy[i]+Gyyz[i]*lxyy[i]+Gzyz[i]*lxzy[i])
                  +(Gxyz[i]*lxyx[i]+Gyyz[i]*lxyy[i]+Gzyz[i]*lxyz[i])
                  +(Gxxy[i]*lyzx[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lyzz[i]))
            + uyy*(TWO*(Gxyy[i]*lxyy[i]+Gyyy[i]*lyyy[i]+Gzyy[i]*lyzy[i])
                  +(Gxyy[i]*lyyx[i]+Gyyy[i]*lyyy[i]+Gzyy[i]*lyyz[i]))
            + uyz*(TWO*(Gxyy[i]*lxzy[i]+Gyyy[i]*lyzy[i]+Gzyy[i]*lzzy[i]
                       +Gxyz[i]*lxyy[i]+Gyyz[i]*lyyy[i]+Gzyz[i]*lyzy[i])
                  +(Gxyz[i]*lyyx[i]+Gyyz[i]*lyyy[i]+Gzyz[i]*lyyz[i])
                  +(Gxyy[i]*lyzx[i]+Gyyy[i]*lyzy[i]+Gzyy[i]*lyzz[i]))
            + uzz*(TWO*(Gxyz[i]*lxzy[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lzzy[i])
                  +(Gxyz[i]*lyzx[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lyzz[i]));

        Rzz[i] = -H*Rzz[i]
            + gxz[i]*Gamxz[i]+gyz[i]*Gamyz_d[i]+gzz[i]*Gamzz_d[i]
            + Gamxa[i]*lxzz[i]+Gamya[i]*lyzz[i]+Gamza[i]*lzzz[i]
            + uxx*(TWO*(Gxxz[i]*lxxz[i]+Gyxz[i]*lxyz[i]+Gzxz[i]*lxzz[i])
                  +(Gxxz[i]*lxzx[i]+Gyxz[i]*lxzy[i]+Gzxz[i]*lxzz[i]))
            + uxy*(TWO*(Gxxz[i]*lxyz[i]+Gyxz[i]*lyyz[i]+Gzxz[i]*lyzz[i]
                       +Gxyz[i]*lxxz[i]+Gyyz[i]*lxyz[i]+Gzyz[i]*lxzz[i])
                  +(Gxyz[i]*lxzx[i]+Gyyz[i]*lxzy[i]+Gzyz[i]*lxzz[i])
                  +(Gxxz[i]*lyzx[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lyzz[i]))
            + uxz*(TWO*(Gxxz[i]*lxzz[i]+Gyxz[i]*lyzz[i]+Gzxz[i]*lzzz[i]
                       +Gxzz[i]*lxxz[i]+Gyzz[i]*lxyz[i]+Gzzz[i]*lxzz[i])
                  +(Gxzz[i]*lxzx[i]+Gyzz[i]*lxzy[i]+Gzzz[i]*lxzz[i])
                  +(Gxxz[i]*lzzx[i]+Gyxz[i]*lzzy[i]+Gzxz[i]*lzzz[i]))
            + uyy*(TWO*(Gxyz[i]*lxyz[i]+Gyyz[i]*lyyz[i]+Gzyz[i]*lyzz[i])
                  +(Gxyz[i]*lyzx[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lyzz[i]))
            + uyz*(TWO*(Gxyz[i]*lxzz[i]+Gyyz[i]*lyzz[i]+Gzyz[i]*lzzz[i]
                       +Gxzz[i]*lxyz[i]+Gyzz[i]*lyyz[i]+Gzzz[i]*lyzz[i])
                  +(Gxzz[i]*lyzx[i]+Gyzz[i]*lyzy[i]+Gzzz[i]*lyzz[i])
                  +(Gxyz[i]*lzzx[i]+Gyyz[i]*lzzy[i]+Gzyz[i]*lzzz[i]))
            + uzz*(TWO*(Gxzz[i]*lxzz[i]+Gyzz[i]*lyzz[i]+Gzzz[i]*lzzz[i])
                  +(Gxzz[i]*lzzx[i]+Gyzz[i]*lzzy[i]+Gzzz[i]*lzzz[i]));

        Rxy[i] = H*(
            -Rxy[i]
            +gxx[i]*Gamxy[i]+gxy[i]*Gamyy_d[i]+gxz[i]*Gamzy[i]
            +gxy[i]*Gamxx[i]+gyy[i]*Gamyx[i]+gyz[i]*Gamzx[i]
            +Gamxa[i]*lxyx[i]+Gamya[i]*lyyx[i]+Gamza[i]*lyzx[i]
            +Gamxa[i]*lxxy[i]+Gamya[i]*lxyy[i]+Gamza[i]*lxzy[i])
            +uxx*(Gxxx[i]*lxxy[i]+Gyxx[i]*lxyy[i]+Gzxx[i]*lxzy[i]
                  +Gxxy[i]*lxxx[i]+Gyxy[i]*lxyx[i]+Gzxy[i]*lxzx[i]
                  +Gxxx[i]*lxyx[i]+Gyxx[i]*lxyy[i]+Gzxx[i]*lxyz[i])
            +uxy*(Gxxx[i]*lxyy[i]+Gyxx[i]*lyyy[i]+Gzxx[i]*lyzy[i]
                  +Gxxy[i]*lxyx[i]+Gyxy[i]*lyyx[i]+Gzxy[i]*lyzx[i]
                  +Gxxy[i]*lxyx[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxyz[i]
                  +Gxxy[i]*lxxy[i]+Gyxy[i]*lxyy[i]+Gzxy[i]*lxzy[i]
                  +Gxyy[i]*lxxx[i]+Gyyy[i]*lxyx[i]+Gzyy[i]*lxzx[i]
                  +Gxxx[i]*lyyx[i]+Gyxx[i]*lyyy[i]+Gzxx[i]*lyyz[i])
            +uxz*(Gxxx[i]*lxzy[i]+Gyxx[i]*lyzy[i]+Gzxx[i]*lzzy[i]
                  +Gxxy[i]*lxzx[i]+Gyxy[i]*lyzx[i]+Gzxy[i]*lzzx[i]
                  +Gxxz[i]*lxyx[i]+Gyxz[i]*lxyy[i]+Gzxz[i]*lxyz[i]
                  +Gxxz[i]*lxxy[i]+Gyxz[i]*lxyy[i]+Gzxz[i]*lxzy[i]
                  +Gxyz[i]*lxxx[i]+Gyyz[i]*lxyx[i]+Gzyz[i]*lxzx[i]
                  +Gxxx[i]*lyzx[i]+Gyxx[i]*lyzy[i]+Gzxx[i]*lyzz[i])
            +uyy*(Gxxy[i]*lxyy[i]+Gyxy[i]*lyyy[i]+Gzxy[i]*lyzy[i]
                  +Gxyy[i]*lxyx[i]+Gyyy[i]*lyyx[i]+Gzyy[i]*lyzx[i]
                  +Gxxy[i]*lyyx[i]+Gyxy[i]*lyyy[i]+Gzxy[i]*lyyz[i])
            +uyz*(Gxxy[i]*lxzy[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lzzy[i]
                  +Gxyy[i]*lxzx[i]+Gyyy[i]*lyzx[i]+Gzyy[i]*lzzx[i]
                  +Gxxz[i]*lyyx[i]+Gyxz[i]*lyyy[i]+Gzxz[i]*lyyz[i]
                  +Gxxz[i]*lxyy[i]+Gyxz[i]*lyyy[i]+Gzxz[i]*lyzy[i]
                  +Gxyz[i]*lxyx[i]+Gyyz[i]*lyyx[i]+Gzyz[i]*lyzx[i]
                  +Gxxy[i]*lyzx[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lyzz[i])
            +uzz*(Gxxz[i]*lxzy[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lzzy[i]
                  +Gxyz[i]*lxzx[i]+Gyyz[i]*lyzx[i]+Gzyz[i]*lzzx[i]
                  +Gxxz[i]*lyzx[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lyzz[i]);

        Rxz[i] = H*(
            -Rxz[i]
            +gxx[i]*Gamxz[i]+gxy[i]*Gamyz_d[i]+gxz[i]*Gamzz_d[i]
            +gxz[i]*Gamxx[i]+gyz[i]*Gamyx[i]+gzz[i]*Gamzx[i]
            +Gamxa[i]*lxzx[i]+Gamya[i]*lyzx[i]+Gamza[i]*lzzx[i]
            +Gamxa[i]*lxxz[i]+Gamya[i]*lxyz[i]+Gamza[i]*lxzz[i])
            +uxx*(Gxxx[i]*lxxz[i]+Gyxx[i]*lxyz[i]+Gzxx[i]*lxzz[i]
                  +Gxxz[i]*lxxx[i]+Gyxz[i]*lxyx[i]+Gzxz[i]*lxzx[i]
                  +Gxxx[i]*lxzx[i]+Gyxx[i]*lxzy[i]+Gzxx[i]*lxzz[i])
            +uxy*(Gxxx[i]*lxyz[i]+Gyxx[i]*lyyz[i]+Gzxx[i]*lyzz[i]
                  +Gxxz[i]*lxyx[i]+Gyxz[i]*lyyx[i]+Gzxz[i]*lyzx[i]
                  +Gxxy[i]*lxzx[i]+Gyxy[i]*lxzy[i]+Gzxy[i]*lxzz[i]
                  +Gxxy[i]*lxxz[i]+Gyxy[i]*lxyz[i]+Gzxy[i]*lxzz[i]
                  +Gxyz[i]*lxxx[i]+Gyyz[i]*lxyx[i]+Gzyz[i]*lxzx[i]
                  +Gxxx[i]*lyzx[i]+Gyxx[i]*lyzy[i]+Gzxx[i]*lyzz[i])
            +uxz*(Gxxx[i]*lxzz[i]+Gyxx[i]*lyzz[i]+Gzxx[i]*lzzz[i]
                  +Gxxz[i]*lxzx[i]+Gyxz[i]*lyzx[i]+Gzxz[i]*lzzx[i]
                  +Gxxz[i]*lxzx[i]+Gyxz[i]*lxzy[i]+Gzxz[i]*lxzz[i]
                  +Gxxz[i]*lxxz[i]+Gyxz[i]*lxyz[i]+Gzxz[i]*lxzz[i]
                  +Gxzz[i]*lxxx[i]+Gyzz[i]*lxyx[i]+Gzzz[i]*lxzx[i]
                  +Gxxx[i]*lzzx[i]+Gyxx[i]*lzzy[i]+Gzxx[i]*lzzz[i])
            +uyy*(Gxxy[i]*lxyz[i]+Gyxy[i]*lyyz[i]+Gzxy[i]*lyzz[i]
                  +Gxyz[i]*lxyx[i]+Gyyz[i]*lyyx[i]+Gzyz[i]*lyzx[i]
                  +Gxxy[i]*lyzx[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lyzz[i])
            +uyz*(Gxxy[i]*lxzz[i]+Gyxy[i]*lyzz[i]+Gzxy[i]*lzzz[i]
                  +Gxyz[i]*lxzx[i]+Gyyz[i]*lyzx[i]+Gzyz[i]*lzzx[i]
                  +Gxxz[i]*lyzx[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lyzz[i]
                  +Gxxz[i]*lxyz[i]+Gyxz[i]*lyyz[i]+Gzxz[i]*lyzz[i]
                  +Gxzz[i]*lxyx[i]+Gyzz[i]*lyyx[i]+Gzzz[i]*lyzx[i]
                  +Gxxy[i]*lzzx[i]+Gyxy[i]*lzzy[i]+Gzxy[i]*lzzz[i])
            +uzz*(Gxxz[i]*lxzz[i]+Gyxz[i]*lyzz[i]+Gzxz[i]*lzzz[i]
                  +Gxzz[i]*lxzx[i]+Gyzz[i]*lyzx[i]+Gzzz[i]*lzzx[i]
                  +Gxxz[i]*lzzx[i]+Gyxz[i]*lzzy[i]+Gzxz[i]*lzzz[i]);

        Ryz[i] = H*(
            -Ryz[i]
            +gxy[i]*Gamxz[i]+gyy[i]*Gamyz_d[i]+gyz[i]*Gamzz_d[i]
            +gxz[i]*Gamxy[i]+gyz[i]*Gamyy_d[i]+gzz[i]*Gamzy[i]
            +Gamxa[i]*lxzy[i]+Gamya[i]*lyzy[i]+Gamza[i]*lzzy[i]
            +Gamxa[i]*lxyz[i]+Gamya[i]*lyyz[i]+Gamza[i]*lyzz[i])
            +uxx*(Gxxy[i]*lxxz[i]+Gyxy[i]*lxyz[i]+Gzxy[i]*lxzz[i]
                  +Gxxz[i]*lxxy[i]+Gyxz[i]*lxyy[i]+Gzxz[i]*lxzy[i]
                  +Gxxy[i]*lxzx[i]+Gyxy[i]*lxzy[i]+Gzxy[i]*lxzz[i])
            +uxy*(Gxxy[i]*lxyz[i]+Gyxy[i]*lyyz[i]+Gzxy[i]*lyzz[i]
                  +Gxxz[i]*lxyy[i]+Gyxz[i]*lyyy[i]+Gzxz[i]*lyzy[i]
                  +Gxyy[i]*lxzx[i]+Gyyy[i]*lxzy[i]+Gzyy[i]*lxzz[i]
                  +Gxyy[i]*lxxz[i]+Gyyy[i]*lxyz[i]+Gzyy[i]*lxzz[i]
                  +Gxyz[i]*lxxy[i]+Gyyz[i]*lxyy[i]+Gzyz[i]*lxzy[i]
                  +Gxxy[i]*lyzx[i]+Gyxy[i]*lyzy[i]+Gzxy[i]*lyzz[i])
            +uxz*(Gxxy[i]*lxzz[i]+Gyxy[i]*lyzz[i]+Gzxy[i]*lzzz[i]
                  +Gxxz[i]*lxzy[i]+Gyxz[i]*lyzy[i]+Gzxz[i]*lzzy[i]
                  +Gxyz[i]*lxzx[i]+Gyyz[i]*lxzy[i]+Gzyz[i]*lxzz[i]
                  +Gxyz[i]*lxxz[i]+Gyyz[i]*lxyz[i]+Gzyz[i]*lxzz[i]
                  +Gxzz[i]*lxxy[i]+Gyzz[i]*lxyy[i]+Gzzz[i]*lxzy[i]
                  +Gxxy[i]*lzzx[i]+Gyxy[i]*lzzy[i]+Gzxy[i]*lzzz[i])
            +uyy*(Gxyy[i]*lxyz[i]+Gyyy[i]*lyyz[i]+Gzyy[i]*lyzz[i]
                  +Gxyz[i]*lxyy[i]+Gyyz[i]*lyyy[i]+Gzyz[i]*lyzy[i]
                  +Gxyy[i]*lyzx[i]+Gyyy[i]*lyzy[i]+Gzyy[i]*lyzz[i])
            +uyz*(Gxyy[i]*lxzz[i]+Gyyy[i]*lyzz[i]+Gzyy[i]*lzzz[i]
                  +Gxyz[i]*lxzy[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lzzy[i]
                  +Gxyz[i]*lyzx[i]+Gyyz[i]*lyzy[i]+Gzyz[i]*lyzz[i]
                  +Gxyz[i]*lxyz[i]+Gyyz[i]*lyyz[i]+Gzyz[i]*lyzz[i]
                  +Gxzz[i]*lxyy[i]+Gyzz[i]*lyyy[i]+Gzzz[i]*lyzy[i]
                  +Gxyy[i]*lzzx[i]+Gyyy[i]*lzzy[i]+Gzyy[i]*lzzz[i])
            +uzz*(Gxyz[i]*lxzz[i]+Gyyz[i]*lyzz[i]+Gzyz[i]*lzzz[i]
                  +Gxzz[i]*lxzy[i]+Gyzz[i]*lyzy[i]+Gzzz[i]*lzzy[i]
                  +Gxyz[i]*lzzx[i]+Gyyz[i]*lzzy[i]+Gzyz[i]*lzzz[i]);
    }
}

/* Phase 13: chi correction to Ricci tensor
 * After fdderivs(chi), subtract Christoffel*chi_deriv, compute conformal factor f,
 * then add chi contribution to Rxx..Rzz.
 */
__global__ __launch_bounds__(128, 4)
void kern_phase12_13_chi_correction_fused(
    const double* __restrict__ chi,
    const double* __restrict__ chin1,
    const double* __restrict__ chix, const double* __restrict__ chiy,
    const double* __restrict__ chiz,
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    double* __restrict__ Rxx, double* __restrict__ Rxy,
    double* __restrict__ Rxz, double* __restrict__ Ryy,
    double* __restrict__ Ryz, double* __restrict__ Rzz)
{
    const double TWO = 2.0;
    const double F3o2 = 1.5;
    const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
    const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
    const int iminF = d_gp.iminF, jminF = d_gp.jminF, kminF = d_gp.kminF;
    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid >= d_gp.all) return;

    const int i0 = tid % nx;
    const int j0 = (tid / nx) % ny;
    const int k0 = tid / (nx * ny);

    double cxx = 0.0, cxy = 0.0, cxz = 0.0;
    double cyy = 0.0, cyz = 0.0, czz = 0.0;

    if (!(i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2)) {
        const int iF = i0 + 1;
        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)
        {
            const double c = fetch_sym_ord2_direct(chi, iF, jF, kF, 1, 1, 1);
            cxx = d_gp.Fdxdx * (
                -fetch_sym_ord2_direct(chi, iF - 2, jF, kF, 1, 1, 1)
                +16.0 * fetch_sym_ord2_direct(chi, iF - 1, jF, kF, 1, 1, 1)
                -30.0 * c
                +16.0 * fetch_sym_ord2_direct(chi, iF + 1, jF, kF, 1, 1, 1)
                -       fetch_sym_ord2_direct(chi, iF + 2, jF, kF, 1, 1, 1));
            cyy = d_gp.Fdydy * (
                -fetch_sym_ord2_direct(chi, iF, jF - 2, kF, 1, 1, 1)
                +16.0 * fetch_sym_ord2_direct(chi, iF, jF - 1, kF, 1, 1, 1)
                -30.0 * c
                +16.0 * fetch_sym_ord2_direct(chi, iF, jF + 1, kF, 1, 1, 1)
                -       fetch_sym_ord2_direct(chi, iF, jF + 2, kF, 1, 1, 1));
            czz = d_gp.Fdzdz * (
                -fetch_sym_ord2_direct(chi, iF, jF, kF - 2, 1, 1, 1)
                +16.0 * fetch_sym_ord2_direct(chi, iF, jF, kF - 1, 1, 1, 1)
                -30.0 * c
                +16.0 * fetch_sym_ord2_direct(chi, iF, jF, kF + 1, 1, 1, 1)
                -       fetch_sym_ord2_direct(chi, iF, jF, kF + 2, 1, 1, 1));

            const double t_jm2 =
                fetch_sym_ord2_direct(chi, iF - 2, jF - 2, kF, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF - 2, kF, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF - 2, kF, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF + 2, jF - 2, kF, 1, 1, 1);
            const double t_jm1 =
                fetch_sym_ord2_direct(chi, iF - 2, jF - 1, kF, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF - 1, kF, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF - 1, kF, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF + 2, jF - 1, kF, 1, 1, 1);
            const double t_jp1 =
                fetch_sym_ord2_direct(chi, iF - 2, jF + 1, kF, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF + 1, kF, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF + 1, kF, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF + 2, jF + 1, kF, 1, 1, 1);
            const double t_jp2 =
                fetch_sym_ord2_direct(chi, iF - 2, jF + 2, kF, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF + 2, kF, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF + 2, kF, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF + 2, jF + 2, kF, 1, 1, 1);
            cxy = d_gp.Fdxdy * (t_jm2 - 8.0 * t_jm1 + 8.0 * t_jp1 - t_jp2);

            const double t_km2_x =
                fetch_sym_ord2_direct(chi, iF - 2, jF, kF - 2, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF, kF - 2, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF, kF - 2, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF + 2, jF, kF - 2, 1, 1, 1);
            const double t_km1_x =
                fetch_sym_ord2_direct(chi, iF - 2, jF, kF - 1, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF, kF - 1, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF, kF - 1, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF + 2, jF, kF - 1, 1, 1, 1);
            const double t_kp1_x =
                fetch_sym_ord2_direct(chi, iF - 2, jF, kF + 1, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF, kF + 1, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF, kF + 1, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF + 2, jF, kF + 1, 1, 1, 1);
            const double t_kp2_x =
                fetch_sym_ord2_direct(chi, iF - 2, jF, kF + 2, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF - 1, jF, kF + 2, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF + 1, jF, kF + 2, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF + 2, jF, kF + 2, 1, 1, 1);
            cxz = d_gp.Fdxdz * (t_km2_x - 8.0 * t_km1_x + 8.0 * t_kp1_x - t_kp2_x);

            const double t_km2_y =
                fetch_sym_ord2_direct(chi, iF, jF - 2, kF - 2, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF, jF - 1, kF - 2, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF, jF + 1, kF - 2, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF, jF + 2, kF - 2, 1, 1, 1);
            const double t_km1_y =
                fetch_sym_ord2_direct(chi, iF, jF - 2, kF - 1, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF, jF - 1, kF - 1, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF, jF + 1, kF - 1, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF, jF + 2, kF - 1, 1, 1, 1);
            const double t_kp1_y =
                fetch_sym_ord2_direct(chi, iF, jF - 2, kF + 1, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF, jF - 1, kF + 1, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF, jF + 1, kF + 1, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF, jF + 2, kF + 1, 1, 1, 1);
            const double t_kp2_y =
                fetch_sym_ord2_direct(chi, iF, jF - 2, kF + 2, 1, 1, 1)
              - 8.0 * fetch_sym_ord2_direct(chi, iF, jF - 1, kF + 2, 1, 1, 1)
              + 8.0 * fetch_sym_ord2_direct(chi, iF, jF + 1, kF + 2, 1, 1, 1)
              -       fetch_sym_ord2_direct(chi, iF, jF + 2, kF + 2, 1, 1, 1);
            cyz = d_gp.Fdydz * (t_km2_y - 8.0 * t_km1_y + 8.0 * t_kp1_y - t_kp2_y);
        }
        else if ((iF + 1) <= imaxF && (iF - 1) >= iminF &&
                 (jF + 1) <= jmaxF && (jF - 1) >= jminF &&
                 (kF + 1) <= kmaxF && (kF - 1) >= kminF)
        {
            const double c = fetch_sym_ord2_direct(chi, iF, jF, kF, 1, 1, 1);
            cxx = d_gp.Sdxdx * (
                fetch_sym_ord2_direct(chi, iF - 1, jF, kF, 1, 1, 1)
              - 2.0 * c
              + fetch_sym_ord2_direct(chi, iF + 1, jF, kF, 1, 1, 1));
            cyy = d_gp.Sdydy * (
                fetch_sym_ord2_direct(chi, iF, jF - 1, kF, 1, 1, 1)
              - 2.0 * c
              + fetch_sym_ord2_direct(chi, iF, jF + 1, kF, 1, 1, 1));
            czz = d_gp.Sdzdz * (
                fetch_sym_ord2_direct(chi, iF, jF, kF - 1, 1, 1, 1)
              - 2.0 * c
              + fetch_sym_ord2_direct(chi, iF, jF, kF + 1, 1, 1, 1));
            cxy = d_gp.Sdxdy * (
                fetch_sym_ord2_direct(chi, iF - 1, jF - 1, kF, 1, 1, 1)
              - fetch_sym_ord2_direct(chi, iF + 1, jF - 1, kF, 1, 1, 1)
              - fetch_sym_ord2_direct(chi, iF - 1, jF + 1, kF, 1, 1, 1)
              + fetch_sym_ord2_direct(chi, iF + 1, jF + 1, kF, 1, 1, 1));
            cxz = d_gp.Sdxdz * (
                fetch_sym_ord2_direct(chi, iF - 1, jF, kF - 1, 1, 1, 1)
              - fetch_sym_ord2_direct(chi, iF + 1, jF, kF - 1, 1, 1, 1)
              - fetch_sym_ord2_direct(chi, iF - 1, jF, kF + 1, 1, 1, 1)
              + fetch_sym_ord2_direct(chi, iF + 1, jF, kF + 1, 1, 1, 1));
            cyz = d_gp.Sdydz * (
                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)
              - 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));
        }
    }

    const double cx = chix[tid];
    const double cy = chiy[tid];
    const double cz = chiz[tid];
    const double c1 = chin1[tid];

    cxx -= Gxxx[tid] * cx + Gyxx[tid] * cy + Gzxx[tid] * cz;
    cxy -= Gxxy[tid] * cx + Gyxy[tid] * cy + Gzxy[tid] * cz;
    cxz -= Gxxz[tid] * cx + Gyxz[tid] * cy + Gzxz[tid] * cz;
    cyy -= Gxyy[tid] * cx + Gyyy[tid] * cy + Gzyy[tid] * cz;
    cyz -= Gxyz[tid] * cx + Gyyz[tid] * cy + Gzyz[tid] * cz;
    czz -= Gxzz[tid] * cx + Gyzz[tid] * cy + Gzzz[tid] * cz;

    const double uxx = gupxx[tid], uxy = gupxy[tid], uxz = gupxz[tid];
    const double uyy = gupyy[tid], uyz = gupyz[tid], uzz = gupzz[tid];
    const double f_val = uxx * (cxx - F3o2 / c1 * cx * cx)
                       + uyy * (cyy - F3o2 / c1 * cy * cy)
                       + uzz * (czz - F3o2 / c1 * cz * cz)
                       + TWO * uxy * (cxy - F3o2 / c1 * cx * cy)
                       + TWO * uxz * (cxz - F3o2 / c1 * cx * cz)
                       + TWO * uyz * (cyz - F3o2 / c1 * cy * cz);

    const double inv2c = 1.0 / (c1 * TWO);
    Rxx[tid] += (cxx - cx * cx * inv2c + gxx[tid] * f_val) * inv2c;
    Ryy[tid] += (cyy - cy * cy * inv2c + gyy[tid] * f_val) * inv2c;
    Rzz[tid] += (czz - cz * cz * inv2c + gzz[tid] * f_val) * inv2c;
    Rxy[tid] += (cxy - cx * cy * inv2c + gxy[tid] * f_val) * inv2c;
    Rxz[tid] += (cxz - cx * cz * inv2c + gxz[tid] * f_val) * inv2c;
    Ryz[tid] += (cyz - cy * cz * inv2c + gyz[tid] * f_val) * inv2c;
}

__global__ __launch_bounds__(128, 4)
void kern_phase13_chi_correction(
    const double* __restrict__ chin1,
    const double* __restrict__ chix, const double* __restrict__ chiy,
    const double* __restrict__ chiz,
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    double* __restrict__ fxx, double* __restrict__ fxy,
    double* __restrict__ fxz, double* __restrict__ fyy,
    double* __restrict__ fyz, double* __restrict__ fzz,
    double* __restrict__ Rxx, double* __restrict__ Rxy,
    double* __restrict__ Rxz, double* __restrict__ Ryy,
    double* __restrict__ Ryz, double* __restrict__ Rzz)
{
    const double H=0.5, TWO=2.0, F3o2=1.5;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        double cx=chix[i],cy=chiy[i],cz=chiz[i],c1=chin1[i];
        /* subtract Christoffel * chi_deriv */
        fxx[i] -= Gxxx[i]*cx+Gyxx[i]*cy+Gzxx[i]*cz;
        fxy[i] -= Gxxy[i]*cx+Gyxy[i]*cy+Gzxy[i]*cz;
        fxz[i] -= Gxxz[i]*cx+Gyxz[i]*cy+Gzxz[i]*cz;
        fyy[i] -= Gxyy[i]*cx+Gyyy[i]*cy+Gzyy[i]*cz;
        fyz[i] -= Gxyz[i]*cx+Gyyz[i]*cy+Gzyz[i]*cz;
        fzz[i] -= Gxzz[i]*cx+Gyzz[i]*cy+Gzzz[i]*cz;

        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];
        double f_val = uxx*(fxx[i]-F3o2/c1*cx*cx)
                     + uyy*(fyy[i]-F3o2/c1*cy*cy)
                     + uzz*(fzz[i]-F3o2/c1*cz*cz)
                     + TWO*uxy*(fxy[i]-F3o2/c1*cx*cy)
                     + TWO*uxz*(fxz[i]-F3o2/c1*cx*cz)
                     + TWO*uyz*(fyz[i]-F3o2/c1*cy*cz);

        double inv2c = 1.0/(c1*TWO);
        Rxx[i] += (fxx[i]-cx*cx*inv2c+gxx[i]*f_val)*inv2c;
        Ryy[i] += (fyy[i]-cy*cy*inv2c+gyy[i]*f_val)*inv2c;
        Rzz[i] += (fzz[i]-cz*cz*inv2c+gzz[i]*f_val)*inv2c;
        Rxy[i] += (fxy[i]-cx*cy*inv2c+gxy[i]*f_val)*inv2c;
        Rxz[i] += (fxz[i]-cx*cz*inv2c+gxz[i]*f_val)*inv2c;
        Ryz[i] += (fyz[i]-cy*cz*inv2c+gyz[i]*f_val)*inv2c;
    }
}

/* Phase 15: trK_rhs, Aij_rhs, gauge.
 * Also updates Christoffel with physical chi correction and computes Lap second derivatives on the fly.
 */
__global__ __launch_bounds__(128, 4)
void kern_phase15_trK_Aij_gauge(
    const double* __restrict__ alpn1, const double* __restrict__ chin1,
    const double* __restrict__ chix, const double* __restrict__ chiy,
    const double* __restrict__ chiz,
    const double* __restrict__ gxx, const double* __restrict__ gxy,
    const double* __restrict__ gxz, const double* __restrict__ gyy,
    const double* __restrict__ gyz, const double* __restrict__ gzz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ trK,
    const double* __restrict__ Axx, const double* __restrict__ Axy,
    const double* __restrict__ Axz, const double* __restrict__ Ayy,
    const double* __restrict__ Ayz, const double* __restrict__ Azz,
    const double* __restrict__ Lapx, const double* __restrict__ Lapy,
    const double* __restrict__ Lapz,
    const double* __restrict__ betaxx, const double* __restrict__ betaxy,
    const double* __restrict__ betaxz, const double* __restrict__ betayx,
    const double* __restrict__ betayy, const double* __restrict__ betayz,
    const double* __restrict__ betazx, const double* __restrict__ betazy,
    const double* __restrict__ betazz,
    const double* __restrict__ rho,
    const double* __restrict__ Sx_m, const double* __restrict__ Sy_m,
    const double* __restrict__ Sz_m,
    const double* __restrict__ Sxx_m, const double* __restrict__ Sxy_m,
    const double* __restrict__ Sxz_m, const double* __restrict__ Syy_m,
    const double* __restrict__ Syz_m, const double* __restrict__ Szz_m,
    const double* __restrict__ dtSfx, const double* __restrict__ dtSfy,
    const double* __restrict__ dtSfz,
    const double* __restrict__ Rxx, const double* __restrict__ Rxy,
    const double* __restrict__ Rxz, const double* __restrict__ Ryy,
    const double* __restrict__ Ryz, const double* __restrict__ Rzz,
    double* __restrict__ Gxxx, double* __restrict__ Gxxy,
    double* __restrict__ Gxxz, double* __restrict__ Gxyy,
    double* __restrict__ Gxyz_o, double* __restrict__ Gxzz,
    double* __restrict__ Gyxx, double* __restrict__ Gyxy,
    double* __restrict__ Gyxz, double* __restrict__ Gyyy,
    double* __restrict__ Gyyz, double* __restrict__ Gyzz,
    double* __restrict__ Gzxx, double* __restrict__ Gzxy,
    double* __restrict__ Gzxz, double* __restrict__ Gzyy,
    double* __restrict__ Gzyz, double* __restrict__ Gzzz,
    double* __restrict__ dtSfx_rhs, double* __restrict__ dtSfy_rhs,
    double* __restrict__ dtSfz_rhs,
    double* __restrict__ trK_rhs,
    double* __restrict__ Axx_rhs, double* __restrict__ Axy_rhs,
    double* __restrict__ Axz_rhs, double* __restrict__ Ayy_rhs,
    double* __restrict__ Ayz_rhs, double* __restrict__ Azz_rhs,
    double* __restrict__ Lap_rhs,
    double* __restrict__ betax_rhs, double* __restrict__ betay_rhs,
    double* __restrict__ betaz_rhs,
    double* __restrict__ Gamx_rhs, double* __restrict__ Gamy_rhs,
    double* __restrict__ Gamz_rhs,
    double* __restrict__ f_arr, double* __restrict__ S_arr)
{
    const double TWO=2.0, FOUR=4.0, EIGHT=8.0, H=0.5;
    const double F1o3=1.0/3.0, F2o3=2.0/3.0, F3o2=1.5;
    const double PI_V=3.14159265358979323846;
    const double F16=16.0, F8=8.0;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all; i += blockDim.x*gridDim.x) {
        const int nx = d_gp.ex[0], ny = d_gp.ex[1], nz = d_gp.ex[2];
        const int i0 = i % nx;
        const int j0 = (i / nx) % ny;
        const int k0 = i / (nx * ny);
        const int iF = i0 + 1;
        const int jF = j0 + 1;
        const int kF = k0 + 1;
        const int imaxF = d_gp.imaxF, jmaxF = d_gp.jmaxF, kmaxF = d_gp.kmaxF;
        const int iminF = d_gp.iminF, jminF = d_gp.jminF, kminF = d_gp.kminF;
        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];
        double a=alpn1[i], c1=chin1[i];
        double cx=chix[i],cy=chiy[i],cz=chiz[i];
        double lx=Lapx[i],ly=Lapy[i],lz=Lapz[i];
        double fxx_v = 0.0, fxy_v = 0.0, fxz_v = 0.0;
        double fyy_v = 0.0, fyz_v = 0.0, fzz_v = 0.0;

        if (!(i0 > nx - 2 || j0 > ny - 2 || k0 > nz - 2)) {
            if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
                (jF + 2) <= jmaxF && (jF - 2) >= jminF &&
                (kF + 2) <= kmaxF && (kF - 2) >= kminF)
            {
                const double c = fetch_sym_ord2_direct(alpn1, iF, jF, kF, 1, 1, 1);
                fxx_v = d_gp.Fdxdx * (
                    -fetch_sym_ord2_direct(alpn1, iF - 2, jF, kF, 1, 1, 1)
                    +16.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF, kF, 1, 1, 1)
                    -30.0 * c
                    +16.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF, kF, 1, 1, 1)
                    -       fetch_sym_ord2_direct(alpn1, iF + 2, jF, kF, 1, 1, 1));
                fyy_v = d_gp.Fdydy * (
                    -fetch_sym_ord2_direct(alpn1, iF, jF - 2, kF, 1, 1, 1)
                    +16.0 * fetch_sym_ord2_direct(alpn1, iF, jF - 1, kF, 1, 1, 1)
                    -30.0 * c
                    +16.0 * fetch_sym_ord2_direct(alpn1, iF, jF + 1, kF, 1, 1, 1)
                    -       fetch_sym_ord2_direct(alpn1, iF, jF + 2, kF, 1, 1, 1));
                fzz_v = d_gp.Fdzdz * (
                    -fetch_sym_ord2_direct(alpn1, iF, jF, kF - 2, 1, 1, 1)
                    +16.0 * fetch_sym_ord2_direct(alpn1, iF, jF, kF - 1, 1, 1, 1)
                    -30.0 * c
                    +16.0 * fetch_sym_ord2_direct(alpn1, iF, jF, kF + 1, 1, 1, 1)
                    -       fetch_sym_ord2_direct(alpn1, iF, jF, kF + 2, 1, 1, 1));

                const double t_jm2 =
                    fetch_sym_ord2_direct(alpn1, iF - 2, jF - 2, kF, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF - 2, kF, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF - 2, kF, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF + 2, jF - 2, kF, 1, 1, 1);
                const double t_jm1 =
                    fetch_sym_ord2_direct(alpn1, iF - 2, jF - 1, kF, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF - 1, kF, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF - 1, kF, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF + 2, jF - 1, kF, 1, 1, 1);
                const double t_jp1 =
                    fetch_sym_ord2_direct(alpn1, iF - 2, jF + 1, kF, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF + 1, kF, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF + 1, kF, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF + 2, jF + 1, kF, 1, 1, 1);
                const double t_jp2 =
                    fetch_sym_ord2_direct(alpn1, iF - 2, jF + 2, kF, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF + 2, kF, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF + 2, kF, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF + 2, jF + 2, kF, 1, 1, 1);
                fxy_v = d_gp.Fdxdy * (t_jm2 - 8.0 * t_jm1 + 8.0 * t_jp1 - t_jp2);

                const double t_km2_x =
                    fetch_sym_ord2_direct(alpn1, iF - 2, jF, kF - 2, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF, kF - 2, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF, kF - 2, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF + 2, jF, kF - 2, 1, 1, 1);
                const double t_km1_x =
                    fetch_sym_ord2_direct(alpn1, iF - 2, jF, kF - 1, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF, kF - 1, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF, kF - 1, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF + 2, jF, kF - 1, 1, 1, 1);
                const double t_kp1_x =
                    fetch_sym_ord2_direct(alpn1, iF - 2, jF, kF + 1, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF, kF + 1, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF, kF + 1, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF + 2, jF, kF + 1, 1, 1, 1);
                const double t_kp2_x =
                    fetch_sym_ord2_direct(alpn1, iF - 2, jF, kF + 2, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF - 1, jF, kF + 2, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF + 1, jF, kF + 2, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF + 2, jF, kF + 2, 1, 1, 1);
                fxz_v = d_gp.Fdxdz * (t_km2_x - 8.0 * t_km1_x + 8.0 * t_kp1_x - t_kp2_x);

                const double t_km2_y =
                    fetch_sym_ord2_direct(alpn1, iF, jF - 2, kF - 2, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF, jF - 1, kF - 2, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF, jF + 1, kF - 2, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF, jF + 2, kF - 2, 1, 1, 1);
                const double t_km1_y =
                    fetch_sym_ord2_direct(alpn1, iF, jF - 2, kF - 1, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF, jF - 1, kF - 1, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF, jF + 1, kF - 1, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF, jF + 2, kF - 1, 1, 1, 1);
                const double t_kp1_y =
                    fetch_sym_ord2_direct(alpn1, iF, jF - 2, kF + 1, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF, jF - 1, kF + 1, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF, jF + 1, kF + 1, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF, jF + 2, kF + 1, 1, 1, 1);
                const double t_kp2_y =
                    fetch_sym_ord2_direct(alpn1, iF, jF - 2, kF + 2, 1, 1, 1)
                  - 8.0 * fetch_sym_ord2_direct(alpn1, iF, jF - 1, kF + 2, 1, 1, 1)
                  + 8.0 * fetch_sym_ord2_direct(alpn1, iF, jF + 1, kF + 2, 1, 1, 1)
                  -       fetch_sym_ord2_direct(alpn1, iF, jF + 2, kF + 2, 1, 1, 1);
                fyz_v = d_gp.Fdydz * (t_km2_y - 8.0 * t_km1_y + 8.0 * t_kp1_y - t_kp2_y);
            }
            else if ((iF + 1) <= imaxF && (iF - 1) >= iminF &&
                     (jF + 1) <= jmaxF && (jF - 1) >= jminF &&
                     (kF + 1) <= kmaxF && (kF - 1) >= kminF)
            {
                const double c = fetch_sym_ord2_direct(alpn1, iF, jF, kF, 1, 1, 1);
                fxx_v = d_gp.Sdxdx * (
                    fetch_sym_ord2_direct(alpn1, iF - 1, jF, kF, 1, 1, 1)
                  - 2.0 * c
                  + fetch_sym_ord2_direct(alpn1, iF + 1, jF, kF, 1, 1, 1));
                fyy_v = d_gp.Sdydy * (
                    fetch_sym_ord2_direct(alpn1, iF, jF - 1, kF, 1, 1, 1)
                  - 2.0 * c
                  + fetch_sym_ord2_direct(alpn1, iF, jF + 1, kF, 1, 1, 1));
                fzz_v = d_gp.Sdzdz * (
                    fetch_sym_ord2_direct(alpn1, iF, jF, kF - 1, 1, 1, 1)
                  - 2.0 * c
                  + fetch_sym_ord2_direct(alpn1, iF, jF, kF + 1, 1, 1, 1));
                fxy_v = d_gp.Sdxdy * (
                    fetch_sym_ord2_direct(alpn1, iF - 1, jF - 1, kF, 1, 1, 1)
                  - fetch_sym_ord2_direct(alpn1, iF + 1, jF - 1, kF, 1, 1, 1)
                  - fetch_sym_ord2_direct(alpn1, iF - 1, jF + 1, kF, 1, 1, 1)
                  + fetch_sym_ord2_direct(alpn1, iF + 1, jF + 1, kF, 1, 1, 1));
                fxz_v = d_gp.Sdxdz * (
                    fetch_sym_ord2_direct(alpn1, iF - 1, jF, kF - 1, 1, 1, 1)
                  - fetch_sym_ord2_direct(alpn1, iF + 1, jF, kF - 1, 1, 1, 1)
                  - fetch_sym_ord2_direct(alpn1, iF - 1, jF, kF + 1, 1, 1, 1)
                  + fetch_sym_ord2_direct(alpn1, iF + 1, jF, kF + 1, 1, 1, 1));
                fyz_v = d_gp.Sdydz * (
                    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)
                  - 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));
            }
        }

        /* raised chi/chi */
        double gx=(uxx*cx+uxy*cy+uxz*cz)/c1;
        double gy=(uxy*cx+uyy*cy+uyz*cz)/c1;
        double gz=(uxz*cx+uyz*cy+uzz*cz)/c1;

        /* Christoffel physical correction */
        Gxxx[i]-=((cx+cx)/c1-gxx[i]*gx)*H;
        Gyxx[i]-=(0.0-gxx[i]*gy)*H;
        Gzxx[i]-=(0.0-gxx[i]*gz)*H;
        Gxyy[i]-=(0.0-gyy[i]*gx)*H;
        Gyyy[i]-=((cy+cy)/c1-gyy[i]*gy)*H;
        Gzyy[i]-=(0.0-gyy[i]*gz)*H;
        Gxzz[i]-=(0.0-gzz[i]*gx)*H;
        Gyzz[i]-=(0.0-gzz[i]*gy)*H;
        Gzzz[i]-=((cz+cz)/c1-gzz[i]*gz)*H;
        Gxxy[i]-=(cy/c1-gxy[i]*gx)*H;
        Gyxy[i]-=(cx/c1-gxy[i]*gy)*H;
        Gzxy[i]-=(0.0-gxy[i]*gz)*H;
        Gxxz[i]-=(cz/c1-gxz[i]*gx)*H;
        Gyxz[i]-=(0.0-gxz[i]*gy)*H;
        Gzxz[i]-=(cx/c1-gxz[i]*gz)*H;
        Gxyz_o[i]-=(0.0-gyz[i]*gx)*H;
        Gyyz[i]-=(cz/c1-gyz[i]*gy)*H;
        Gzyz[i]-=(cy/c1-gyz[i]*gz)*H;

        /* Lap second-derivative correction: subtract Gamma*Lap_deriv */
        fxx_v -= Gxxx[i]*lx+Gyxx[i]*ly+Gzxx[i]*lz;
        fyy_v -= Gxyy[i]*lx+Gyyy[i]*ly+Gzyy[i]*lz;
        fzz_v -= Gxzz[i]*lx+Gyzz[i]*ly+Gzzz[i]*lz;
        fxy_v -= Gxxy[i]*lx+Gyxy[i]*ly+Gzxy[i]*lz;
        fxz_v -= Gxxz[i]*lx+Gyxz[i]*ly+Gzxz[i]*lz;
        fyz_v -= Gxyz_o[i]*lx+Gyyz[i]*ly+Gzyz[i]*lz;

        /* D^i D_i alpha */
        double DDA = uxx*fxx_v+uyy*fyy_v+uzz*fzz_v
                   +TWO*(uxy*fxy_v+uxz*fxz_v+uyz*fyz_v);

        /* trace of S_ij (physical) */
        double S_v = c1*(uxx*Sxx_m[i]+uyy*Syy_m[i]+uzz*Szz_m[i]
                    +TWO*(uxy*Sxy_m[i]+uxz*Sxz_m[i]+uyz*Syz_m[i]));

        /* A^ij A_ij */
        double AijAij =
            uxx*(uxx*Axx[i]*Axx[i]+uyy*Axy[i]*Axy[i]+uzz*Axz[i]*Axz[i]
                +TWO*(uxy*Axx[i]*Axy[i]+uxz*Axx[i]*Axz[i]+uyz*Axy[i]*Axz[i]))
           +uyy*(uxx*Axy[i]*Axy[i]+uyy*Ayy[i]*Ayy[i]+uzz*Ayz[i]*Ayz[i]
                +TWO*(uxy*Axy[i]*Ayy[i]+uxz*Axy[i]*Ayz[i]+uyz*Ayy[i]*Ayz[i]))
           +uzz*(uxx*Axz[i]*Axz[i]+uyy*Ayz[i]*Ayz[i]+uzz*Azz[i]*Azz[i]
                +TWO*(uxy*Axz[i]*Ayz[i]+uxz*Axz[i]*Azz[i]+uyz*Ayz[i]*Azz[i]))
           +TWO*(
                uxy*(uxx*Axx[i]*Axy[i]+uyy*Axy[i]*Ayy[i]+uzz*Axz[i]*Ayz[i]
                    +uxy*(Axx[i]*Ayy[i]+Axy[i]*Axy[i])
                    +uxz*(Axx[i]*Ayz[i]+Axz[i]*Axy[i])
                    +uyz*(Axy[i]*Ayz[i]+Axz[i]*Ayy[i]))
               +uxz*(uxx*Axx[i]*Axz[i]+uyy*Axy[i]*Ayz[i]+uzz*Axz[i]*Azz[i]
                    +uxy*(Axx[i]*Ayz[i]+Axy[i]*Axz[i])
                    +uxz*(Axx[i]*Azz[i]+Axz[i]*Axz[i])
                    +uyz*(Axy[i]*Azz[i]+Axz[i]*Ayz[i]))
               +uyz*(uxx*Axy[i]*Axz[i]+uyy*Ayy[i]*Ayz[i]+uzz*Ayz[i]*Azz[i]
                    +uxy*(Axy[i]*Ayz[i]+Ayy[i]*Axz[i])
                    +uxz*(Axy[i]*Azz[i]+Ayz[i]*Axz[i])
                    +uyz*(Ayy[i]*Azz[i]+Ayz[i]*Ayz[i])));

        double trK_v = trK[i];
        double db = betaxx[i] + betayy[i] + betazz[i];

        /* trK_rhs step 1: store D^iD_i alpha * chin1 */
        trK_rhs[i] = c1 * DDA;

        /* f_arr = -(1/3) * (DDA + alpha/chi * (2/3*K^2 - AijAij - 16pi*rho + 8pi*S)) */
        double f_v = F2o3*trK_v*trK_v - AijAij - F16*PI_V*rho[i] + EIGHT*PI_V*S_v;
        f_arr[i] = -F1o3*(uxx*fxx_v+uyy*fyy_v+uzz*fzz_v
                         +TWO*(uxy*fxy_v+uxz*fxz_v+uyz*fyz_v)
                         +(a/c1)*f_v);

        /* fij = alpha*(Rij - 8pi*Sij) - D_iD_j alpha */
        double fij_xx=a*(Rxx[i]-EIGHT*PI_V*Sxx_m[i])-fxx_v;
        double fij_xy=a*(Rxy[i]-EIGHT*PI_V*Sxy_m[i])-fxy_v;
        double fij_xz=a*(Rxz[i]-EIGHT*PI_V*Sxz_m[i])-fxz_v;
        double fij_yy=a*(Ryy[i]-EIGHT*PI_V*Syy_m[i])-fyy_v;
        double fij_yz=a*(Ryz[i]-EIGHT*PI_V*Syz_m[i])-fyz_v;
        double fij_zz=a*(Rzz[i]-EIGHT*PI_V*Szz_m[i])-fzz_v;

        /* Aij_rhs = chi*(fij - gij*f) */
        Axx_rhs[i]=fij_xx-gxx[i]*f_arr[i];
        Ayy_rhs[i]=fij_yy-gyy[i]*f_arr[i];
        Azz_rhs[i]=fij_zz-gzz[i]*f_arr[i];
        Axy_rhs[i]=fij_xy-gxy[i]*f_arr[i];
        Axz_rhs[i]=fij_xz-gxz[i]*f_arr[i];
        Ayz_rhs[i]=fij_yz-gyz[i]*f_arr[i];

        /* A_il A^l_j */
        double AA_xx=uxx*Axx[i]*Axx[i]+uyy*Axy[i]*Axy[i]+uzz*Axz[i]*Axz[i]
                    +TWO*(uxy*Axx[i]*Axy[i]+uxz*Axx[i]*Axz[i]+uyz*Axy[i]*Axz[i]);
        double AA_yy=uxx*Axy[i]*Axy[i]+uyy*Ayy[i]*Ayy[i]+uzz*Ayz[i]*Ayz[i]
                    +TWO*(uxy*Axy[i]*Ayy[i]+uxz*Axy[i]*Ayz[i]+uyz*Ayy[i]*Ayz[i]);
        double AA_zz=uxx*Axz[i]*Axz[i]+uyy*Ayz[i]*Ayz[i]+uzz*Azz[i]*Azz[i]
                    +TWO*(uxy*Axz[i]*Ayz[i]+uxz*Axz[i]*Azz[i]+uyz*Ayz[i]*Azz[i]);
        double AA_xy=uxx*Axx[i]*Axy[i]+uyy*Axy[i]*Ayy[i]+uzz*Axz[i]*Ayz[i]
                    +uxy*(Axx[i]*Ayy[i]+Axy[i]*Axy[i])
                    +uxz*(Axx[i]*Ayz[i]+Axz[i]*Axy[i])
                    +uyz*(Axy[i]*Ayz[i]+Axz[i]*Ayy[i]);
        double AA_xz=uxx*Axx[i]*Axz[i]+uyy*Axy[i]*Ayz[i]+uzz*Axz[i]*Azz[i]
                    +uxy*(Axx[i]*Ayz[i]+Axy[i]*Axz[i])
                    +uxz*(Axx[i]*Azz[i]+Axz[i]*Axz[i])
                    +uyz*(Axy[i]*Azz[i]+Axz[i]*Ayz[i]);
        double AA_yz=uxx*Axy[i]*Axz[i]+uyy*Ayy[i]*Ayz[i]+uzz*Ayz[i]*Azz[i]
                    +uxy*(Axy[i]*Ayz[i]+Ayy[i]*Axz[i])
                    +uxz*(Axy[i]*Azz[i]+Ayz[i]*Axz[i])
                    +uyz*(Ayy[i]*Azz[i]+Ayz[i]*Ayz[i]);

        /* trK_rhs final */
        trK_rhs[i] = -trK_rhs[i]
            + a*(F1o3*trK_v*trK_v
                +uxx*AA_xx+uyy*AA_yy+uzz*AA_zz
                +TWO*(uxy*AA_xy+uxz*AA_xz+uyz*AA_yz)
                +FOUR*PI_V*(rho[i]+S_v));

        /* Aij_rhs final */
        Axx_rhs[i]=c1*Axx_rhs[i]+a*(trK_v*Axx[i]-TWO*AA_xx)
            +TWO*(Axx[i]*betaxx[i]+Axy[i]*betayx[i]+Axz[i]*betazx[i])-F2o3*Axx[i]*db;
        Ayy_rhs[i]=c1*Ayy_rhs[i]+a*(trK_v*Ayy[i]-TWO*AA_yy)
            +TWO*(Axy[i]*betaxy[i]+Ayy[i]*betayy[i]+Ayz[i]*betazy[i])-F2o3*Ayy[i]*db;
        Azz_rhs[i]=c1*Azz_rhs[i]+a*(trK_v*Azz[i]-TWO*AA_zz)
            +TWO*(Axz[i]*betaxz[i]+Ayz[i]*betayz[i]+Azz[i]*betazz[i])-F2o3*Azz[i]*db;
        Axy_rhs[i]=c1*Axy_rhs[i]+a*(trK_v*Axy[i]-TWO*AA_xy)
            +Axx[i]*betaxy[i]+Axz[i]*betazy[i]+Ayy[i]*betayx[i]
            +Ayz[i]*betazx[i]+F1o3*Axy[i]*db-Axy[i]*betazz[i];
        Ayz_rhs[i]=c1*Ayz_rhs[i]+a*(trK_v*Ayz[i]-TWO*AA_yz)
            +Axy[i]*betaxz[i]+Ayy[i]*betayz[i]+Axz[i]*betaxy[i]
            +Azz[i]*betazy[i]+F1o3*Ayz[i]*db-Ayz[i]*betaxx[i];
        Axz_rhs[i]=c1*Axz_rhs[i]+a*(trK_v*Axz[i]-TWO*AA_xz)
            +Axx[i]*betaxz[i]+Axy[i]*betayz[i]+Ayz[i]*betayx[i]
            +Azz[i]*betazx[i]+F1o3*Axz[i]*db-Axz[i]*betayy[i];

        /* gauge */
        Lap_rhs[i] = -TWO*a*trK_v;
        betax_rhs[i] = 0.75*dtSfx[i];
        betay_rhs[i] = 0.75*dtSfy[i];
        betaz_rhs[i] = 0.75*dtSfz[i];
#if (GAUGE == 0)
        dtSfx_rhs[i] = Gamx_rhs[i] - 2.0*dtSfx[i];
        dtSfy_rhs[i] = Gamy_rhs[i] - 2.0*dtSfy[i];
        dtSfz_rhs[i] = Gamz_rhs[i] - 2.0*dtSfz[i];
#endif
    }
}

/* Phase 18: Hamilton & momentum constraints (co==0 only) */
__global__ __launch_bounds__(128, 4)
void kern_phase18_constraints(
    const double* __restrict__ chin1,
    const double* __restrict__ chix, const double* __restrict__ chiy,
    const double* __restrict__ chiz,
    const double* __restrict__ gupxx, const double* __restrict__ gupxy,
    const double* __restrict__ gupxz, const double* __restrict__ gupyy,
    const double* __restrict__ gupyz, const double* __restrict__ gupzz,
    const double* __restrict__ trK,
    const double* __restrict__ Axx, const double* __restrict__ Axy,
    const double* __restrict__ Axz, const double* __restrict__ Ayy,
    const double* __restrict__ Ayz, const double* __restrict__ Azz,
    const double* __restrict__ Rxx, const double* __restrict__ Rxy,
    const double* __restrict__ Rxz, const double* __restrict__ Ryy,
    const double* __restrict__ Ryz, const double* __restrict__ Rzz,
    const double* __restrict__ rho,
    const double* __restrict__ Sx_m, const double* __restrict__ Sy_m,
    const double* __restrict__ Sz_m,
    const double* __restrict__ Kx, const double* __restrict__ Ky,
    const double* __restrict__ Kz,
    const double* __restrict__ Gxxx, const double* __restrict__ Gxxy,
    const double* __restrict__ Gxxz, const double* __restrict__ Gxyy,
    const double* __restrict__ Gxyz, const double* __restrict__ Gxzz,
    const double* __restrict__ Gyxx, const double* __restrict__ Gyxy,
    const double* __restrict__ Gyxz, const double* __restrict__ Gyyy,
    const double* __restrict__ Gyyz, const double* __restrict__ Gyzz,
    const double* __restrict__ Gzxx, const double* __restrict__ Gzxy,
    const double* __restrict__ Gzxz, const double* __restrict__ Gzyy,
    const double* __restrict__ Gzyz, const double* __restrict__ Gzzz,
    /* dA/dx arrays (fderivs of Aij) */
    const double* __restrict__ dAxx_x, const double* __restrict__ dAxx_y,
    const double* __restrict__ dAxx_z,
    const double* __restrict__ dAxy_x, const double* __restrict__ dAxy_y,
    const double* __restrict__ dAxy_z,
    const double* __restrict__ dAxz_x, const double* __restrict__ dAxz_y,
    const double* __restrict__ dAxz_z,
    const double* __restrict__ dAyy_x, const double* __restrict__ dAyy_y,
    const double* __restrict__ dAyy_z,
    const double* __restrict__ dAyz_x, const double* __restrict__ dAyz_y,
    const double* __restrict__ dAyz_z,
    const double* __restrict__ dAzz_x, const double* __restrict__ dAzz_y,
    const double* __restrict__ dAzz_z,
    double* __restrict__ ham_Res,
    double* __restrict__ movx_Res, double* __restrict__ movy_Res,
    double* __restrict__ movz_Res)
{
    const double TWO=2.0, F2o3=2.0/3.0, F8=8.0, F16=16.0;
    const double PI_V=3.14159265358979323846;
    for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < d_gp.all;
         i += blockDim.x*gridDim.x)
    {
        double uxx=gupxx[i],uxy=gupxy[i],uxz=gupxz[i];
        double uyy=gupyy[i],uyz=gupyz[i],uzz=gupzz[i];
        double c1=chin1[i];

        /* Hamiltonian constraint */
        double R_sc = uxx*Rxx[i]+uyy*Ryy[i]+uzz*Rzz[i]
                    +TWO*(uxy*Rxy[i]+uxz*Rxz[i]+uyz*Ryz[i]);
        /* AijAij (same as in phase15) */
        double AijAij =
            uxx*(uxx*Axx[i]*Axx[i]+uyy*Axy[i]*Axy[i]+uzz*Axz[i]*Axz[i]
                +TWO*(uxy*Axx[i]*Axy[i]+uxz*Axx[i]*Axz[i]+uyz*Axy[i]*Axz[i]))
           +uyy*(uxx*Axy[i]*Axy[i]+uyy*Ayy[i]*Ayy[i]+uzz*Ayz[i]*Ayz[i]
                +TWO*(uxy*Axy[i]*Ayy[i]+uxz*Axy[i]*Ayz[i]+uyz*Ayy[i]*Ayz[i]))
           +uzz*(uxx*Axz[i]*Axz[i]+uyy*Ayz[i]*Ayz[i]+uzz*Azz[i]*Azz[i]
                +TWO*(uxy*Axz[i]*Ayz[i]+uxz*Axz[i]*Azz[i]+uyz*Ayz[i]*Azz[i]))
           +TWO*(uxy*(uxx*Axx[i]*Axy[i]+uyy*Axy[i]*Ayy[i]+uzz*Axz[i]*Ayz[i]
                      +uxy*(Axx[i]*Ayy[i]+Axy[i]*Axy[i])
                      +uxz*(Axx[i]*Ayz[i]+Axz[i]*Axy[i])
                      +uyz*(Axy[i]*Ayz[i]+Axz[i]*Ayy[i]))
                 +uxz*(uxx*Axx[i]*Axz[i]+uyy*Axy[i]*Ayz[i]+uzz*Axz[i]*Azz[i]
                      +uxy*(Axx[i]*Ayz[i]+Axy[i]*Axz[i])
                      +uxz*(Axx[i]*Azz[i]+Axz[i]*Axz[i])
                      +uyz*(Axy[i]*Azz[i]+Axz[i]*Ayz[i]))
                 +uyz*(uxx*Axy[i]*Axz[i]+uyy*Ayy[i]*Ayz[i]+uzz*Ayz[i]*Azz[i]
                      +uxy*(Axy[i]*Ayz[i]+Ayy[i]*Axz[i])
                      +uxz*(Axy[i]*Azz[i]+Ayz[i]*Axz[i])
                      +uyz*(Ayy[i]*Azz[i]+Ayz[i]*Ayz[i])));

        ham_Res[i] = c1*R_sc + F2o3*trK[i]*trK[i] - AijAij - F16*PI_V*rho[i];

        /* Momentum constraints: need covariant derivative of A */
        double cx=chix[i],cy=chiy[i],cz=chiz[i];
        /* D_j A^j_x etc — subtract Christoffel and chi terms */
        /* gxxx = dAxx_x - 2*Gxxx*Axx - ... - chix*Axx/chin1 etc */
        double mx_xx = dAxx_x[i]-(Gxxx[i]*Axx[i]+Gyxx[i]*Axy[i]+Gzxx[i]*Axz[i]
                                  +Gxxx[i]*Axx[i]+Gyxx[i]*Axy[i]+Gzxx[i]*Axz[i])-cx*Axx[i]/c1;
        double mx_xy = dAxy_x[i]-(Gxxy[i]*Axx[i]+Gyxy[i]*Axy[i]+Gzxy[i]*Axz[i]
                                  +Gxxx[i]*Axy[i]+Gyxx[i]*Ayy[i]+Gzxx[i]*Ayz[i])-cx*Axy[i]/c1;
        double mx_xz = dAxz_x[i]-(Gxxz[i]*Axx[i]+Gyxz[i]*Axy[i]+Gzxz[i]*Axz[i]
                                  +Gxxx[i]*Axz[i]+Gyxx[i]*Ayz[i]+Gzxx[i]*Azz[i])-cx*Axz[i]/c1;
        double mx_yy = dAyy_x[i]-(Gxxy[i]*Axy[i]+Gyxy[i]*Ayy[i]+Gzxy[i]*Ayz[i]
                                  +Gxxy[i]*Axy[i]+Gyxy[i]*Ayy[i]+Gzxy[i]*Ayz[i])-cx*Ayy[i]/c1;
        double mx_yz = dAyz_x[i]-(Gxxz[i]*Axy[i]+Gyxz[i]*Ayy[i]+Gzxz[i]*Ayz[i]
                                  +Gxxy[i]*Axz[i]+Gyxy[i]*Ayz[i]+Gzxy[i]*Azz[i])-cx*Ayz[i]/c1;
        double mx_zz = dAzz_x[i]-(Gxxz[i]*Axz[i]+Gyxz[i]*Ayz[i]+Gzxz[i]*Azz[i]
                                  +Gxxz[i]*Axz[i]+Gyxz[i]*Ayz[i]+Gzxz[i]*Azz[i])-cx*Azz[i]/c1;

        double my_xx = dAxx_y[i]-(Gxxy[i]*Axx[i]+Gyxy[i]*Axy[i]+Gzxy[i]*Axz[i]
                                  +Gxxy[i]*Axx[i]+Gyxy[i]*Axy[i]+Gzxy[i]*Axz[i])-cy*Axx[i]/c1;
        double my_xy = dAxy_y[i]-(Gxyy[i]*Axx[i]+Gyyy[i]*Axy[i]+Gzyy[i]*Axz[i]
                                  +Gxxy[i]*Axy[i]+Gyxy[i]*Ayy[i]+Gzxy[i]*Ayz[i])-cy*Axy[i]/c1;
        double my_xz = dAxz_y[i]-(Gxyz[i]*Axx[i]+Gyyz[i]*Axy[i]+Gzyz[i]*Axz[i]
                                  +Gxxy[i]*Axz[i]+Gyxy[i]*Ayz[i]+Gzxy[i]*Azz[i])-cy*Axz[i]/c1;
        double my_yy = dAyy_y[i]-(Gxyy[i]*Axy[i]+Gyyy[i]*Ayy[i]+Gzyy[i]*Ayz[i]
                                  +Gxyy[i]*Axy[i]+Gyyy[i]*Ayy[i]+Gzyy[i]*Ayz[i])-cy*Ayy[i]/c1;
        double my_yz = dAyz_y[i]-(Gxyz[i]*Axy[i]+Gyyz[i]*Ayy[i]+Gzyz[i]*Ayz[i]
                                  +Gxyy[i]*Axz[i]+Gyyy[i]*Ayz[i]+Gzyy[i]*Azz[i])-cy*Ayz[i]/c1;
        double my_zz = dAzz_y[i]-(Gxyz[i]*Axz[i]+Gyyz[i]*Ayz[i]+Gzyz[i]*Azz[i]
                                  +Gxyz[i]*Axz[i]+Gyyz[i]*Ayz[i]+Gzyz[i]*Azz[i])-cy*Azz[i]/c1;

        double mz_xx = dAxx_z[i]-(Gxxz[i]*Axx[i]+Gyxz[i]*Axy[i]+Gzxz[i]*Axz[i]
                                  +Gxxz[i]*Axx[i]+Gyxz[i]*Axy[i]+Gzxz[i]*Axz[i])-cz*Axx[i]/c1;
        double mz_xy = dAxy_z[i]-(Gxyz[i]*Axx[i]+Gyyz[i]*Axy[i]+Gzyz[i]*Axz[i]
                                  +Gxxz[i]*Axy[i]+Gyxz[i]*Ayy[i]+Gzxz[i]*Ayz[i])-cz*Axy[i]/c1;
        double mz_xz = dAxz_z[i]-(Gxzz[i]*Axx[i]+Gyzz[i]*Axy[i]+Gzzz[i]*Axz[i]
                                  +Gxxz[i]*Axz[i]+Gyxz[i]*Ayz[i]+Gzxz[i]*Azz[i])-cz*Axz[i]/c1;
        double mz_yy = dAyy_z[i]-(Gxyz[i]*Axy[i]+Gyyz[i]*Ayy[i]+Gzyz[i]*Ayz[i]
                                  +Gxyz[i]*Axy[i]+Gyyz[i]*Ayy[i]+Gzyz[i]*Ayz[i])-cz*Ayy[i]/c1;
        double mz_yz = dAyz_z[i]-(Gxzz[i]*Axy[i]+Gyzz[i]*Ayy[i]+Gzzz[i]*Ayz[i]
                                  +Gxyz[i]*Axz[i]+Gyyz[i]*Ayz[i]+Gzyz[i]*Azz[i])-cz*Ayz[i]/c1;
        double mz_zz = dAzz_z[i]-(Gxzz[i]*Axz[i]+Gyzz[i]*Ayz[i]+Gzzz[i]*Azz[i]
                                  +Gxzz[i]*Axz[i]+Gyzz[i]*Ayz[i]+Gzzz[i]*Azz[i])-cz*Azz[i]/c1;

        movx_Res[i] = uxx*mx_xx+uyy*my_xy+uzz*mz_xz
                     +uxy*mx_xy+uxz*mx_xz+uyz*my_xz
                     +uxy*my_xx+uxz*mz_xx+uyz*mz_xy
                     - F2o3*Kx[i] - F8*PI_V*Sx_m[i];
        movy_Res[i] = uxx*mx_xy+uyy*my_yy+uzz*mz_yz
                     +uxy*mx_yy+uxz*mx_yz+uyz*my_yz
                     +uxy*my_xy+uxz*mz_xy+uyz*mz_yy
                     - F2o3*Ky[i] - F8*PI_V*Sy_m[i];
        movz_Res[i] = uxx*mx_xz+uyy*my_yz+uzz*mz_zz
                     +uxy*mx_yz+uxz*mx_zz+uyz*my_zz
                     +uxy*my_xz+uxz*mz_xz+uyz*mz_yz
                     - F2o3*Kz[i] - F8*PI_V*Sz_m[i];
    }
}

static void setup_grid_params(int *ex,
                              double *X, double *Y, double *Z,
                              int Symmetry, double eps, int co)
{
    const int nx = ex[0];
    const int ny = ex[1];
    const int nz = ex[2];
    const double dX = X[1] - X[0];
    const double dY = Y[1] - Y[0];
    const double dZ = Z[1] - Z[0];
    const int NO_SYMM = 0;
    const int EQ_SYMM = 1;

    ensure_gpu_buffers(nx, ny, nz);

    GridParams gp = {};
    gp.ex[0] = nx;
    gp.ex[1] = ny;
    gp.ex[2] = nz;
    gp.all = nx * ny * nz;
    gp.dX = dX;
    gp.dY = dY;
    gp.dZ = dZ;
    gp.d12dx = 1.0 / (12.0 * dX);
    gp.d12dy = 1.0 / (12.0 * dY);
    gp.d12dz = 1.0 / (12.0 * dZ);
    gp.d2dx = 1.0 / (2.0 * dX);
    gp.d2dy = 1.0 / (2.0 * dY);
    gp.d2dz = 1.0 / (2.0 * dZ);
    gp.Fdxdx = 1.0 / (12.0 * dX * dX);
    gp.Fdydy = 1.0 / (12.0 * dY * dY);
    gp.Fdzdz = 1.0 / (12.0 * dZ * dZ);
    gp.Sdxdx = 1.0 / (dX * dX);
    gp.Sdydy = 1.0 / (dY * dY);
    gp.Sdzdz = 1.0 / (dZ * dZ);
    gp.Fdxdy = 1.0 / (144.0 * dX * dY);
    gp.Fdxdz = 1.0 / (144.0 * dX * dZ);
    gp.Fdydz = 1.0 / (144.0 * dY * dZ);
    gp.Sdxdy = 0.25 / (dX * dY);
    gp.Sdxdz = 0.25 / (dX * dZ);
    gp.Sdydz = 0.25 / (dY * dZ);
    gp.iminF = 1;
    gp.jminF = 1;
    gp.kminF = 1;
    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;
    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;
    gp.Symmetry = Symmetry;
    gp.eps = eps;
    gp.co = co;
    gp.fh2_nx = nx + 2;
    gp.fh2_ny = ny + 2;
    gp.fh2_nz = nz + 2;
    gp.fh3_nx = nx + 3;
    gp.fh3_ny = ny + 3;
    gp.fh3_nz = nz + 3;
    upload_grid_params_if_needed(gp);
}

static void compute_patch_boundary_flags(int *ex,
                                         double *X, double *Y, double *Z,
                                         const double *bbox,
                                         int Symmetry,
                                         int &touch_xmin, int &touch_xmax,
                                         int &touch_ymin, int &touch_ymax,
                                         int &touch_zmin, int &touch_zmax)
{
    const double dX = X[1] - X[0];
    const double dY = Y[1] - Y[0];
    const double dZ = Z[1] - Z[0];
    const int NO_SYMM = 0;
    const int OCTANT = 2;

    touch_xmax = (std::fabs(X[ex[0] - 1] - bbox[3]) < dX) ? 1 : 0;
    touch_ymax = (std::fabs(Y[ex[1] - 1] - bbox[4]) < dY) ? 1 : 0;
    touch_zmax = (std::fabs(Z[ex[2] - 1] - bbox[5]) < dZ) ? 1 : 0;

    touch_xmin = (std::fabs(X[0] - bbox[0]) < dX &&
                  !(Symmetry == OCTANT && std::fabs(bbox[0]) < dX / 2.0)) ? 1 : 0;
    touch_ymin = (std::fabs(Y[0] - bbox[1]) < dY &&
                  !(Symmetry == OCTANT && std::fabs(bbox[1]) < dY / 2.0)) ? 1 : 0;
    touch_zmin = (std::fabs(Z[0] - bbox[2]) < dZ &&
                  !(Symmetry > NO_SYMM && std::fabs(bbox[2]) < dZ / 2.0)) ? 1 : 0;
}

static void upload_state_inputs(double **state_host, size_t all)
{
    const size_t bytes = all * sizeof(double);
    const bool profile = cuda_profile_enabled();
    const double t0 = profile ? cuda_profile_now_ms() : 0.0;
    static int direct_upload = -1;
    if (direct_upload < 0) {
        const char *env = getenv("AMSS_CUDA_DIRECT_STATE_UPLOAD");
        direct_upload = env ? ((atoi(env) != 0) ? 1 : 0) : 1;
    }
    if (direct_upload) {
        for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
            CUDA_CHECK(cudaMemcpyAsync(g_buf.slot[k_state_input_slots[i]], state_host[i],
                                       bytes, cudaMemcpyHostToDevice));
        }
        if (profile) {
            cuda_profile_sync();
            CudaProfileStats &stats = cuda_profile_stats();
            stats.upload_calls++;
            stats.upload_ms += cuda_profile_now_ms() - t0;
            stats.upload_gb += (double)((size_t)BSSN_STATE_COUNT * bytes) / 1.0e9;
        }
        return;
    }
    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
        std::memcpy(g_buf.h_stage + (size_t)i * all, state_host[i], bytes);
    }
    CUDA_CHECK(cudaMemcpy(g_buf.slot[S_chi], g_buf.h_stage,
                          (size_t)BSSN_STATE_COUNT * bytes,
                          cudaMemcpyHostToDevice));
    if (profile) {
        CudaProfileStats &stats = cuda_profile_stats();
        stats.upload_calls++;
        stats.upload_ms += cuda_profile_now_ms() - t0;
        stats.upload_gb += (double)((size_t)BSSN_STATE_COUNT * bytes) / 1.0e9;
    }
}

static void upload_matter_cache(StepContext &ctx,
                                double **matter_host,
                                size_t all)
{
    const size_t bytes = all * sizeof(double);
    for (int i = 0; i < BSSN_MATTER_COUNT; ++i) {
        std::memcpy(g_buf.h_stage + (size_t)i * all, matter_host[i], bytes);
    }
    CUDA_CHECK(cudaMemcpy(ctx.d_matter_mem, g_buf.h_stage,
                          (size_t)BSSN_MATTER_COUNT * bytes,
                          cudaMemcpyHostToDevice));
    ctx.matter_ready = true;
}

static void zero_matter_cache(StepContext &ctx, size_t all)
{
    CUDA_CHECK(cudaMemset(ctx.d_matter_mem, 0,
                          (size_t)BSSN_MATTER_COUNT * all * sizeof(double)));
    ctx.matter_ready = true;
}

static void bind_matter_slots(const StepContext &ctx)
{
    for (int i = 0; i < BSSN_MATTER_COUNT; ++i) {
        g_buf.slot[k_matter_slots[i]] = ctx.d_matter[i];
    }
}

static void bind_state_input_slots(const std::array<double *, BSSN_STATE_COUNT> &state)
{
    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
        g_buf.slot[k_state_input_slots[i]] = state[i];
    }
}

static void bind_state_output_slots(const std::array<double *, BSSN_STATE_COUNT> &state)
{
    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
        g_buf.slot[k_state_rhs_slots[i]] = state[i];
    }
}

static void launch_rhs_pipeline(int all, double eps, int co)
{
    const double SYM = 1.0;
    const double ANTI = -1.0;
    const bool stage_timing = rhs_stage_timing_enabled();
    double stage_ms[RHS_STAGE_COUNT] = {};
    double stage_t0 = stage_timing ? cuda_profile_now_ms() : 0.0;

    #define D(s) g_buf.slot[s]
    #define MARK_RHS_STAGE(stage_id) do {                             \
        if (stage_timing) {                                            \
            cuda_profile_sync();                                       \
            const double stage_t1 = cuda_profile_now_ms();             \
            stage_ms[(stage_id)] += stage_t1 - stage_t0;               \
            stage_t0 = stage_t1;                                       \
        }                                                             \
    } while (0)

    kern_phase1_prep<<<grid(all),BLK>>>(
        D(S_Lap), D(S_chi), D(S_dxx), D(S_dyy), D(S_dzz),
        D(S_alpn1), D(S_chin1), D(S_gxx), D(S_gyy), D(S_gzz));
    MARK_RHS_STAGE(RHS_STAGE_PREP);

    {
        double *src_fields[] = {
            D(S_betax), D(S_betay), D(S_betaz), D(S_chi),
            D(S_dxx), D(S_gxy), D(S_gxz), D(S_dyy),
            D(S_gyz), D(S_dzz), D(S_Lap), D(S_trK)
        };
        double *fx_fields[] = {
            D(S_betaxx), D(S_betayx), D(S_betazx), D(S_chix),
            D(S_gxxx), D(S_gxyx), D(S_gxzx), D(S_gyyx),
            D(S_gyzx), D(S_gzzx), D(S_Lapx), D(S_Kx)
        };
        double *fy_fields[] = {
            D(S_betaxy), D(S_betayy), D(S_betazy), D(S_chiy),
            D(S_gxxy), D(S_gxyy), D(S_gxzy), D(S_gyyy),
            D(S_gyzy), D(S_gzzy), D(S_Lapy), D(S_Ky)
        };
        double *fz_fields[] = {
            D(S_betaxz), D(S_betayz), D(S_betazz), D(S_chiz),
            D(S_gxxz), D(S_gxyz), D(S_gxzz), D(S_gyyz),
            D(S_gyzz), D(S_gzzz), D(S_Lapz), D(S_Kz)
        };
        const int soa_signs[] = {
            (int)ANTI, (int)SYM,  (int)SYM,
            (int)SYM,  (int)ANTI, (int)SYM,
            (int)SYM,  (int)SYM,  (int)ANTI,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)ANTI, (int)ANTI, (int)SYM,
            (int)ANTI, (int)SYM,  (int)ANTI,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)SYM,  (int)ANTI, (int)ANTI,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)SYM,  (int)SYM,  (int)SYM
        };
        gpu_fderivs_batch((int)(sizeof(src_fields) / sizeof(src_fields[0])),
                          src_fields, fx_fields, fy_fields, fz_fields,
                          soa_signs, all);
    }
    MARK_RHS_STAGE(RHS_STAGE_DERIV1);

    kern_phase2_metric_rhs<<<grid(all),BLK>>>(
        D(S_alpn1), D(S_chin1),
        D(S_gxx), D(S_gxy), D(S_gxz), D(S_gyy), D(S_gyz), D(S_gzz),
        D(S_trK),
        D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz),
        D(S_betaxx), D(S_betaxy), D(S_betaxz),
        D(S_betayx), D(S_betayy), D(S_betayz),
        D(S_betazx), D(S_betazy), D(S_betazz),
        D(S_chi_rhs), D(S_gxx_rhs), D(S_gyy_rhs), D(S_gzz_rhs),
        D(S_gxy_rhs), D(S_gyz_rhs), D(S_gxz_rhs));

    kern_phase2_inverse<<<grid(all),BLK>>>(
        D(S_gxx), D(S_gxy), D(S_gxz), D(S_gyy), D(S_gyz), D(S_gzz),
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz));

    if (co == 0) {
        kern_phase3_gamma_constraint<<<grid(all),BLK>>>(
            D(S_Gamx), D(S_Gamy), D(S_Gamz),
            D(S_gupxx), D(S_gupxy), D(S_gupxz),
            D(S_gupyy), D(S_gupyz), D(S_gupzz),
            D(S_gxxx), D(S_gxyx), D(S_gxzx), D(S_gyyx), D(S_gyzx), D(S_gzzx),
            D(S_gxxy), D(S_gxyy), D(S_gxzy), D(S_gyyy), D(S_gyzy), D(S_gzzy),
            D(S_gxxz), D(S_gxyz), D(S_gxzz), D(S_gyyz), D(S_gyzz), D(S_gzzz),
            D(S_Gmx_Res), D(S_Gmy_Res), D(S_Gmz_Res));
    }

    kern_phase4_christoffel<<<grid(all),BLK>>>(
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_gxxx), D(S_gxyx), D(S_gxzx), D(S_gyyx), D(S_gyzx), D(S_gzzx),
        D(S_gxxy), D(S_gxyy), D(S_gxzy), D(S_gyyy), D(S_gyzy), D(S_gzzy),
        D(S_gxxz), D(S_gxyz), D(S_gxzz), D(S_gyyz), D(S_gyzz), D(S_gzzz),
        D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
        D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
        D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
        D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
        D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
        D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz));

    kern_phase5_6_gamma_rhs_part1_fused<<<grid(all),BLK>>>(
        D(S_Lapx), D(S_Lapy), D(S_Lapz),
        D(S_alpn1), D(S_chin1),
        D(S_chix), D(S_chiy), D(S_chiz),
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz),
        D(S_Kx), D(S_Ky), D(S_Kz),
        D(S_Sx), D(S_Sy), D(S_Sz),
        D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
        D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
        D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
        D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
        D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
        D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz),
        D(S_Gamx_rhs), D(S_Gamy_rhs), D(S_Gamz_rhs));
    MARK_RHS_STAGE(RHS_STAGE_METRIC);

    {
        double *src_fields[] = {D(S_betax), D(S_betay), D(S_betaz)};
        double *fxx_fields[] = {D(S_gxxx), D(S_gxxy), D(S_gxxz)};
        double *fxy_fields[] = {D(S_gxyx), D(S_gxyy), D(S_gxyz)};
        double *fxz_fields[] = {D(S_gxzx), D(S_gxzy), D(S_gxzz)};
        double *fyy_fields[] = {D(S_gyyx), D(S_gyyy), D(S_gyyz)};
        double *fyz_fields[] = {D(S_gyzx), D(S_gyzy), D(S_gyzz)};
        double *fzz_fields[] = {D(S_gzzx), D(S_gzzy), D(S_gzzz)};
        const int soa_signs[] = {
            (int)ANTI, (int)SYM,  (int)SYM,
            (int)SYM,  (int)ANTI, (int)SYM,
            (int)SYM,  (int)SYM,  (int)ANTI
        };
        gpu_fdderivs_batch((int)(sizeof(src_fields) / sizeof(src_fields[0])),
                           src_fields, fxx_fields, fxy_fields, fxz_fields,
                           fyy_fields, fyz_fields, fzz_fields,
                           soa_signs, all);
    }
    {
        double *src_fields[] = {D(S_Gamx), D(S_Gamy), D(S_Gamz)};
        double *fx_fields[] = {D(S_Gamxx), D(S_Gamyx), D(S_Gamzx)};
        double *fy_fields[] = {D(S_Gamxy), D(S_Gamyy_t), D(S_Gamzy)};
        double *fz_fields[] = {D(S_Gamxz), D(S_Gamyz_t), D(S_Gamzz_t)};
        const int soa_signs[] = {
            (int)ANTI, (int)SYM,  (int)SYM,
            (int)SYM,  (int)ANTI, (int)SYM,
            (int)SYM,  (int)SYM,  (int)ANTI
        };
        gpu_fderivs_batch((int)(sizeof(src_fields) / sizeof(src_fields[0])),
                          src_fields, fx_fields, fy_fields, fz_fields,
                          soa_signs, all);
    }
    MARK_RHS_STAGE(RHS_STAGE_GAUGE_DERIV);

    kern_phase8_9_gamma_rhs_contract_fused<<<grid(all),BLK>>>(
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_gxxx),D(S_gxyx),D(S_gxzx),D(S_gyyx),D(S_gyzx),D(S_gzzx),
        D(S_gxxy),D(S_gxyy),D(S_gxzy),D(S_gyyy),D(S_gyzy),D(S_gzzy),
        D(S_gxxz),D(S_gxyz),D(S_gxzz),D(S_gyyz),D(S_gyzz),D(S_gzzz),
        D(S_Gamxxx),D(S_Gamxxy),D(S_Gamxxz),
        D(S_Gamxyy),D(S_Gamxyz),D(S_Gamxzz),
        D(S_Gamyxx),D(S_Gamyxy),D(S_Gamyxz),
        D(S_Gamyyy),D(S_Gamyyz),D(S_Gamyzz),
        D(S_Gamzxx),D(S_Gamzxy),D(S_Gamzxz),
        D(S_Gamzyy),D(S_Gamzyz),D(S_Gamzzz),
        D(S_betaxx),D(S_betaxy),D(S_betaxz),
        D(S_betayx),D(S_betayy),D(S_betayz),
        D(S_betazx),D(S_betazy),D(S_betazz),
        D(S_gxx),D(S_gxy),D(S_gxz),D(S_gyy),D(S_gyz),D(S_gzz),
        D(S_Gamx_rhs),D(S_Gamy_rhs),D(S_Gamz_rhs),
        D(S_Gamxa),D(S_Gamya),D(S_Gamza),
        D(S_gxxx),D(S_gxyx),D(S_gxzx),D(S_gyyx),D(S_gyzx),D(S_gzzx),
        D(S_gxxy),D(S_gxyy),D(S_gxzy),D(S_gyyy),D(S_gyzy),D(S_gzzy),
        D(S_gxxz),D(S_gxyz),D(S_gxzz),D(S_gyyz),D(S_gyzz),D(S_gzzz));
    MARK_RHS_STAGE(RHS_STAGE_GAMMA_CONTRACT);

    {
        double *src_fields[] = {D(S_dxx), D(S_dyy), D(S_dzz), D(S_gxy), D(S_gxz), D(S_gyz)};
        double *dst_fields[] = {D(S_Rxx), D(S_Ryy), D(S_Rzz), D(S_Rxy), D(S_Rxz), D(S_Ryz)};
        const int soa_signs[] = {
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)ANTI, (int)ANTI, (int)SYM,
            (int)ANTI, (int)SYM,  (int)ANTI,
            (int)SYM,  (int)ANTI, (int)ANTI
        };
        gpu_phase10_ricci_batch(D(S_gupxx), D(S_gupxy), D(S_gupxz),
                                D(S_gupyy), D(S_gupyz), D(S_gupzz),
                                src_fields, dst_fields, soa_signs, all);
    }
    MARK_RHS_STAGE(RHS_STAGE_RICCI_DIFF);

    kern_phase11_ricci_fused<<<grid(all),BLK>>>(
        D(S_gxx),D(S_gxy),D(S_gxz),D(S_gyy),D(S_gyz),D(S_gzz),
        D(S_gupxx),D(S_gupxy),D(S_gupxz),D(S_gupyy),D(S_gupyz),D(S_gupzz),
        D(S_Gamxa),D(S_Gamya),D(S_Gamza),
        D(S_Gamxx),D(S_Gamxy),D(S_Gamxz),
        D(S_Gamyx),D(S_Gamyy_t),D(S_Gamyz_t),
        D(S_Gamzx),D(S_Gamzy),D(S_Gamzz_t),
        D(S_Gamxxx),D(S_Gamxxy),D(S_Gamxxz),
        D(S_Gamxyy),D(S_Gamxyz),D(S_Gamxzz),
        D(S_Gamyxx),D(S_Gamyxy),D(S_Gamyxz),
        D(S_Gamyyy),D(S_Gamyyz),D(S_Gamyzz),
        D(S_Gamzxx),D(S_Gamzxy),D(S_Gamzxz),
        D(S_Gamzyy),D(S_Gamzyz),D(S_Gamzzz),
        D(S_gxxx),D(S_gxyx),D(S_gxzx),D(S_gyyx),D(S_gyzx),D(S_gzzx),
        D(S_gxxy),D(S_gxyy),D(S_gxzy),D(S_gyyy),D(S_gyzy),D(S_gzzy),
        D(S_gxxz),D(S_gxyz),D(S_gxzz),D(S_gyyz),D(S_gyzz),D(S_gzzz),
        D(S_Rxx),D(S_Rxy),D(S_Rxz),
        D(S_Ryy),D(S_Ryz),D(S_Rzz));
    MARK_RHS_STAGE(RHS_STAGE_RICCI_FUSED);

    kern_phase12_13_chi_correction_fused<<<grid((size_t)all),BLK>>>(
        D(S_chi), D(S_chin1),
        D(S_chix), D(S_chiy), D(S_chiz),
        D(S_gxx), D(S_gxy), D(S_gxz), D(S_gyy), D(S_gyz), D(S_gzz),
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
        D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
        D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
        D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
        D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
        D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz),
        D(S_Rxx), D(S_Rxy), D(S_Rxz),
        D(S_Ryy), D(S_Ryz), D(S_Rzz));
    MARK_RHS_STAGE(RHS_STAGE_CHI);

    kern_phase15_trK_Aij_gauge<<<grid(all),BLK>>>(
        D(S_alpn1), D(S_chin1),
        D(S_chix), D(S_chiy), D(S_chiz),
        D(S_gxx), D(S_gxy), D(S_gxz), D(S_gyy), D(S_gyz), D(S_gzz),
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_trK),
        D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz),
        D(S_Lapx), D(S_Lapy), D(S_Lapz),
        D(S_betaxx), D(S_betaxy), D(S_betaxz),
        D(S_betayx), D(S_betayy), D(S_betayz),
        D(S_betazx), D(S_betazy), D(S_betazz),
        D(S_rho),
        D(S_Sx), D(S_Sy), D(S_Sz),
        D(S_Sxx), D(S_Sxy), D(S_Sxz), D(S_Syy), D(S_Syz), D(S_Szz),
        D(S_dtSfx), D(S_dtSfy), D(S_dtSfz),
        D(S_Rxx), D(S_Rxy), D(S_Rxz), D(S_Ryy), D(S_Ryz), D(S_Rzz),
        D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
        D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
        D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
        D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
        D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
        D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz),
        D(S_dtSfx_rhs), D(S_dtSfy_rhs), D(S_dtSfz_rhs),
        D(S_trK_rhs),
        D(S_Axx_rhs), D(S_Axy_rhs), D(S_Axz_rhs),
        D(S_Ayy_rhs), D(S_Ayz_rhs), D(S_Azz_rhs),
        D(S_Lap_rhs),
        D(S_betax_rhs), D(S_betay_rhs), D(S_betaz_rhs),
        D(S_Gamx_rhs), D(S_Gamy_rhs), D(S_Gamz_rhs),
        D(S_f_arr), D(S_S_arr));
    MARK_RHS_STAGE(RHS_STAGE_GAUGE_RHS);

    gpu_lopsided_kodis_state_batch(eps, all);
    MARK_RHS_STAGE(RHS_STAGE_KODIS);

    if (co == 0) {
        {
            double *src_fields[] = {D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz)};
            double *fx_fields[] = {D(S_gxxx), D(S_gxyx), D(S_gxzx), D(S_gyyx), D(S_gyzx), D(S_gzzx)};
            double *fy_fields[] = {D(S_gxxy), D(S_gxyy), D(S_gxzy), D(S_gyyy), D(S_gyzy), D(S_gzzy)};
            double *fz_fields[] = {D(S_gxxz), D(S_gxyz), D(S_gxzz), D(S_gyyz), D(S_gyzz), D(S_gzzz)};
            const int soa_signs[] = {
                (int)SYM,  (int)SYM,  (int)SYM,
                (int)ANTI, (int)ANTI, (int)SYM,
                (int)ANTI, (int)SYM,  (int)ANTI,
                (int)SYM,  (int)SYM,  (int)SYM,
                (int)SYM,  (int)ANTI, (int)ANTI,
                (int)SYM,  (int)SYM,  (int)SYM
            };
            gpu_fderivs_batch((int)(sizeof(src_fields) / sizeof(src_fields[0])),
                              src_fields, fx_fields, fy_fields, fz_fields,
                              soa_signs, all);
        }

        kern_phase18_constraints<<<grid(all),BLK>>>(
            D(S_chin1),
            D(S_chix), D(S_chiy), D(S_chiz),
            D(S_gupxx), D(S_gupxy), D(S_gupxz),
            D(S_gupyy), D(S_gupyz), D(S_gupzz),
            D(S_trK),
            D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz),
            D(S_Rxx), D(S_Rxy), D(S_Rxz), D(S_Ryy), D(S_Ryz), D(S_Rzz),
            D(S_rho), D(S_Sx), D(S_Sy), D(S_Sz),
            D(S_Kx), D(S_Ky), D(S_Kz),
            D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
            D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
            D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
            D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
            D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
            D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz),
            D(S_gxxx), D(S_gxxy), D(S_gxxz),
            D(S_gxyx), D(S_gxyy), D(S_gxyz),
            D(S_gxzx), D(S_gxzy), D(S_gxzz),
            D(S_gyyx), D(S_gyyy), D(S_gyyz),
            D(S_gyzx), D(S_gyzy), D(S_gyzz),
            D(S_gzzx), D(S_gzzy), D(S_gzzz),
            D(S_ham_Res), D(S_movx_Res), D(S_movy_Res), D(S_movz_Res));
    }
    MARK_RHS_STAGE(RHS_STAGE_CONSTRAINTS);

    rhs_stage_profile_accumulate(stage_ms);
    #undef MARK_RHS_STAGE
    #undef D
}

static void download_state_outputs(double **state_host_out, size_t all)
{
    const size_t bytes = all * sizeof(double);
    CUDA_CHECK(cudaMemcpy(g_buf.h_stage, g_buf.slot[S_chi_rhs],
                          (size_t)BSSN_STATE_COUNT * bytes,
                          cudaMemcpyDeviceToHost));
    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
        std::memcpy(state_host_out[i], g_buf.h_stage + (size_t)i * all, bytes);
    }
}

static void download_constraint_outputs(double **constraint_host_out, size_t all)
{
    const size_t bytes = all * sizeof(double);
    CUDA_CHECK(cudaMemcpy(g_buf.h_stage, g_buf.slot[S_ham_Res],
                          (size_t)D2H_CONSTRAINT_SLOT_COUNT * bytes,
                          cudaMemcpyDeviceToHost));
    for (int i = 0; i < D2H_CONSTRAINT_SLOT_COUNT; ++i) {
        std::memcpy(constraint_host_out[i], g_buf.h_stage + (size_t)i * all, bytes);
    }
}

__global__ void kern_pack_state_region_batch(const double * __restrict__ src_mem,
                                             double * __restrict__ dst,
                                             int nx, int ny,
                                             int i0, int j0, int k0,
                                             int sx, int sy, int sz,
                                             int region_all,
                                             int state_count,
                                             int all)
{
    const int state_index = blockIdx.y;
    if (state_index >= state_count) return;
    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 src = (i0 + ii) + (j0 + jj) * nx + (k0 + kk) * nx * ny;
        dst[(size_t)state_index * region_all + local] =
            src_mem[(size_t)state_index * all + src];
    }
}

__global__ void kern_unpack_state_region_batch(double * __restrict__ dst_mem,
                                               const double * __restrict__ src,
                                               int nx, int ny,
                                               int i0, int j0, int k0,
                                               int sx, int sy, int sz,
                                               int region_all,
                                               int state_count,
                                               int all)
{
    const int state_index = blockIdx.y;
    if (state_index >= state_count) return;
    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 dst = (i0 + ii) + (j0 + jj) * nx + (k0 + kk) * nx * ny;
        dst_mem[(size_t)state_index * all + dst] =
            src[(size_t)state_index * region_all + local];
    }
}

__global__ void kern_pack_state_segments_batch(const double * __restrict__ src_mem,
                                               double * __restrict__ dst,
                                               int nx, int ny,
                                               const int * __restrict__ meta,
                                               int state_count,
                                               int all)
{
    const int segment = blockIdx.z;
    const int state_index = blockIdx.y;
    const int *m = meta + segment * 8;
    const int i0 = m[0], j0 = m[1], k0 = m[2];
    const int sx = m[3], sy = m[4];
    const int region_all = m[6];
    const int offset = m[7];
    if (state_index >= state_count) return;
    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 src = (i0 + ii) + (j0 + jj) * nx + (k0 + kk) * nx * ny;
        dst[(size_t)offset + (size_t)state_index * region_all + local] =
            src_mem[(size_t)state_index * all + src];
    }
}

__global__ void kern_unpack_state_segments_batch(double * __restrict__ dst_mem,
                                                 const double * __restrict__ src,
                                                 int nx, int ny,
                                                 const int * __restrict__ meta,
                                                 int state_count,
                                                 int all)
{
    const int segment = blockIdx.z;
    const int state_index = blockIdx.y;
    const int *m = meta + segment * 8;
    const int i0 = m[0], j0 = m[1], k0 = m[2];
    const int sx = m[3], sy = m[4];
    const int region_all = m[6];
    const int offset = m[7];
    if (state_index >= state_count) return;
    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 dst = (i0 + ii) + (j0 + jj) * nx + (k0 + kk) * nx * ny;
        dst_mem[(size_t)state_index * all + dst] =
            src[(size_t)offset + (size_t)state_index * region_all + local];
    }
}

__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;
    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};

    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 < 6; ++oz)
        {
            const int z = fc_k + offs[oz];
            const double wz = w[oz];
            for (int oy = 0; oy < 6; ++oy)
            {
                const int y = fc_j + offs[oy];
                const double wyz = wz * w[oy];
                for (int ox = 0; ox < 6; ++ox)
                {
                    const int x = fc_i + offs[ox];
                    const int src = x + y * nx + z * nx * ny;
                    sum += wyz * w[ox] * src_mem[(size_t)state_index * all + src];
                }
            }
        }
        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;
    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};

    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 < 6; ++oz)
        {
            const int z = ck + offs[oz];
            const double wzv = wz[oz];
            for (int oy = 0; oy < 6; ++oy)
            {
                const int y = cj + offs[oy];
                const double wyz = wzv * wy[oy];
                for (int ox = 0; ox < 6; ++ox)
                {
                    const int x = ci + offs[ox];
                    const int src = x + y * nx + z * nx * ny;
                    sum += wyz * wx[ox] * src_mem[(size_t)state_index * all + src];
                }
            }
        }
        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,
                                       int all)
{
    const int subset_slot = blockIdx.y;
    if (subset_slot >= subset_count) return;
    const int state_index = d_subset_state_indices[subset_slot];
    for (int src = blockIdx.x * blockDim.x + threadIdx.x;
         src < all;
         src += blockDim.x * gridDim.x)
    {
        dst[(size_t)subset_slot * all + src] =
            src_mem[(size_t)state_index * all + src];
    }
}

__global__ void kern_unpack_state_subset(double * __restrict__ dst_mem,
                                         const double * __restrict__ src,
                                         int subset_count,
                                         int all)
{
    const int subset_slot = blockIdx.y;
    if (subset_slot >= subset_count) return;
    const int state_index = d_subset_state_indices[subset_slot];
    for (int dst = blockIdx.x * blockDim.x + threadIdx.x;
         dst < all;
         dst += blockDim.x * gridDim.x)
    {
        dst_mem[(size_t)state_index * all + dst] =
            src[(size_t)subset_slot * all + dst];
    }
}

static void copy_state_region_cuda(void *block_tag,
                                   int state_index,
                                   double *host_state,
                                   const int *ex,
                                   int i0, int j0, int k0,
                                   int sx, int sy, int sz,
                                   cudaMemcpyKind kind)
{
    if (state_index < 0 || state_index >= BSSN_STATE_COUNT) return;
    if (sx <= 0 || sy <= 0 || sz <= 0) return;

    const size_t pitch = (size_t)ex[0] * sizeof(double);
    StepContext &ctx = ensure_step_ctx(block_tag, (size_t)ex[0] * ex[1] * ex[2]);

    cudaMemcpy3DParms p = {};
    p.extent = make_cudaExtent((size_t)sx * sizeof(double), (size_t)sy, (size_t)sz);
    p.srcPos = make_cudaPos((size_t)i0 * sizeof(double), j0, k0);
    p.dstPos = make_cudaPos((size_t)i0 * sizeof(double), j0, k0);

    if (kind == cudaMemcpyDeviceToHost) {
        p.srcPtr = make_cudaPitchedPtr((void *)ctx.d_state_curr[state_index], pitch, ex[0], ex[1]);
        p.dstPtr = make_cudaPitchedPtr((void *)host_state, pitch, ex[0], ex[1]);
    } else {
        p.srcPtr = make_cudaPitchedPtr((void *)host_state, pitch, ex[0], ex[1]);
        p.dstPtr = make_cudaPitchedPtr((void *)ctx.d_state_curr[state_index], pitch, ex[0], ex[1]);
    }
    CUDA_CHECK(cudaMemcpy3D(&p));
}

static void copy_state_region_packed_cuda(void *block_tag,
                                          int state_index,
                                          double *host_buffer,
                                          const int *ex,
                                          int i0, int j0, int k0,
                                          int sx, int sy, int sz,
                                          cudaMemcpyKind kind)
{
    if (state_index < 0 || state_index >= BSSN_STATE_COUNT) return;
    if (sx <= 0 || sy <= 0 || sz <= 0) return;

    const size_t src_pitch = (size_t)ex[0] * sizeof(double);
    const size_t dst_pitch = (size_t)sx * sizeof(double);
    StepContext &ctx = ensure_step_ctx(block_tag, (size_t)ex[0] * ex[1] * ex[2]);

    cudaMemcpy3DParms p = {};
    p.extent = make_cudaExtent((size_t)sx * sizeof(double), (size_t)sy, (size_t)sz);

    if (kind == cudaMemcpyDeviceToHost) {
        p.srcPtr = make_cudaPitchedPtr((void *)ctx.d_state_curr[state_index], src_pitch, ex[0], ex[1]);
        p.srcPos = make_cudaPos((size_t)i0 * sizeof(double), j0, k0);
        p.dstPtr = make_cudaPitchedPtr((void *)host_buffer, dst_pitch, sx, sy);
        p.dstPos = make_cudaPos(0, 0, 0);
    } else {
        p.srcPtr = make_cudaPitchedPtr((void *)host_buffer, dst_pitch, sx, sy);
        p.srcPos = make_cudaPos(0, 0, 0);
        p.dstPtr = make_cudaPitchedPtr((void *)ctx.d_state_curr[state_index], src_pitch, ex[0], ex[1]);
        p.dstPos = make_cudaPos((size_t)i0 * sizeof(double), j0, k0);
    }

    CUDA_CHECK(cudaMemcpy3D(&p));
}

static void copy_state_region_packed_batch_cuda(void *block_tag,
                                                int state_count,
                                                double *host_buffer,
                                                const int *ex,
                                                int i0, int j0, int k0,
                                                int sx, int sy, int sz,
                                                cudaMemcpyKind kind)
{
    if (state_count <= 0 || state_count > BSSN_STATE_COUNT) return;
    if (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;
    const size_t total_doubles = (size_t)state_count * (size_t)region_all;
    double *d_comm = ensure_step_comm_buffer(ctx, total_doubles);

    if (kind == cudaMemcpyDeviceToHost) {
        dim3 launch_grid((unsigned int)grid((size_t)region_all),
                         (unsigned int)state_count);
        kern_pack_state_region_batch<<<launch_grid, BLK>>>(
            ctx.d_state_curr_mem, d_comm, ex[0], ex[1],
            i0, j0, k0, sx, sy, sz, region_all, state_count,
            ex[0] * ex[1] * ex[2]);
        CUDA_CHECK(cudaMemcpy(host_buffer, d_comm,
                              total_doubles * sizeof(double),
                              cudaMemcpyDeviceToHost));
    } else {
        CUDA_CHECK(cudaMemcpy(d_comm, host_buffer,
                              total_doubles * sizeof(double),
                              cudaMemcpyHostToDevice));
        dim3 launch_grid((unsigned int)grid((size_t)region_all),
                         (unsigned int)state_count);
        kern_unpack_state_region_batch<<<launch_grid, BLK>>>(
            ctx.d_state_curr_mem, d_comm, ex[0], ex[1],
            i0, j0, k0, sx, sy, sz, region_all, state_count,
            ex[0] * ex[1] * ex[2]);
    }
}

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];
    const size_t bytes = all * sizeof(double);
    StepContext &ctx = ensure_step_ctx(block_tag, all);
    const bool profile = cuda_profile_enabled();
    const double t0 = profile ? cuda_profile_now_ms() : 0.0;
    static int direct_download = -1;
    if (direct_download < 0) {
        const char *env = getenv("AMSS_CUDA_DIRECT_STATE_DOWNLOAD");
        direct_download = env ? ((atoi(env) != 0) ? 1 : 0) : 1;
    }
    if (direct_download) {
        for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
            CUDA_CHECK(cudaMemcpyAsync(state_host_out[i], ctx.d_state_curr[i],
                                       bytes, cudaMemcpyDeviceToHost));
        }
        CUDA_CHECK(cudaDeviceSynchronize());
        if (profile) {
            CudaProfileStats &stats = cuda_profile_stats();
            stats.resident_download_calls++;
            stats.resident_download_ms += cuda_profile_now_ms() - t0;
            stats.resident_download_gb += (double)((size_t)BSSN_STATE_COUNT * bytes) / 1.0e9;
        }
        return;
    }
    CUDA_CHECK(cudaMemcpy(g_buf.h_stage, ctx.d_state_curr_mem,
                          (size_t)BSSN_STATE_COUNT * bytes,
                          cudaMemcpyDeviceToHost));
    for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
        std::memcpy(state_host_out[i], g_buf.h_stage + (size_t)i * all, bytes);
    }
    if (profile) {
        CudaProfileStats &stats = cuda_profile_stats();
        stats.resident_download_calls++;
        stats.resident_download_ms += cuda_profile_now_ms() - t0;
        stats.resident_download_gb += (double)((size_t)BSSN_STATE_COUNT * bytes) / 1.0e9;
    }
}

static void copy_state_subset(void *block_tag,
                              int *ex,
                              int subset_count,
                              const int *state_indices,
                              double **state_host,
                              cudaMemcpyKind kind)
{
    if (subset_count <= 0) return;
    const size_t all = (size_t)ex[0] * ex[1] * ex[2];
    const size_t bytes = all * sizeof(double);
    StepContext &ctx = ensure_step_ctx(block_tag, all);
    int active_state_indices[BSSN_STATE_COUNT];
    double *active_state_host[BSSN_STATE_COUNT];
    int active_count = 0;

    for (int i = 0; i < subset_count; ++i) {
        const int state_index = state_indices[i];
        if (state_index < 0 || state_index >= BSSN_STATE_COUNT) continue;
        if (!state_host[i]) continue;
        active_state_indices[active_count] = state_index;
        active_state_host[active_count] = state_host[i];
        ++active_count;
    }
    if (active_count <= 0) return;

    const size_t total_doubles = (size_t)active_count * all;
    double *d_comm = ensure_step_comm_buffer(ctx, total_doubles);
    double *h_comm = ensure_step_host_comm_buffer(ctx, total_doubles);

    CUDA_CHECK(cudaMemcpyToSymbol(d_subset_state_indices, active_state_indices,
                                  (size_t)active_count * sizeof(int),
                                  0, cudaMemcpyHostToDevice));
    if (kind == cudaMemcpyDeviceToHost) {
        dim3 launch_grid((unsigned int)grid(all), (unsigned int)active_count);
        kern_pack_state_subset<<<launch_grid, BLK>>>(
            ctx.d_state_curr_mem, d_comm, active_count, (int)all);
        CUDA_CHECK(cudaMemcpy(h_comm, d_comm,
                              total_doubles * sizeof(double),
                              cudaMemcpyDeviceToHost));
        for (int i = 0; i < active_count; ++i) {
            std::memcpy(active_state_host[i],
                        h_comm + (size_t)i * all,
                        bytes);
        }
    } else {
        for (int i = 0; i < active_count; ++i) {
            std::memcpy(h_comm + (size_t)i * all,
                        active_state_host[i],
                        bytes);
        }
        CUDA_CHECK(cudaMemcpy(d_comm, h_comm,
                              total_doubles * sizeof(double),
                              cudaMemcpyHostToDevice));
        dim3 launch_grid((unsigned int)grid(all), (unsigned int)active_count);
        kern_unpack_state_subset<<<launch_grid, BLK>>>(
            ctx.d_state_curr_mem, d_comm, active_count, (int)all);
    }
}

static bool has_resident_state(void *block_tag)
{
    auto it = g_step_ctx.find(block_tag);
    return it != g_step_ctx.end() && it->second.state_ready;
}

/* ================================================================== */
/*  Main host function — drop-in replacement for bssn_rhs_c.C          */
/* ================================================================== */

extern "C"
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_m, double *Sxz, double *Syy, double *Syz_m, 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)
{
    /* --- Multi-GPU: select device --- */
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));

    const int nx = ex[0], ny = ex[1], nz = ex[2];
    const int all = nx * ny * nz;
    const double SYM = 1.0, ANTI = -1.0;

    setup_grid_params(ex, X, Y, Z, Symmetry, eps, co);

    /* --- Shorthand for device slot pointers --- */
    #define D(s) g_buf.slot[s]
    const size_t bytes = (size_t)all * sizeof(double);

    /* --- H2D: stage all inputs, then one bulk copy --- */
    double *h2d_src[] = {
        chi, trK, dxx, gxy, gxz, dyy, gyz, dzz,
        Axx, Axy, Axz, Ayy, Ayz, Azz,
        Gamx, Gamy, Gamz,
        Lap, betax, betay, betaz,
        dtSfx, dtSfy, dtSfz,
        rho, Sx, Sy, Sz,
        Sxx, Sxy_m, Sxz, Syy, Syz_m, Szz
    };
    static_assert((int)(sizeof(h2d_src) / sizeof(h2d_src[0])) == H2D_INPUT_SLOT_COUNT,
                  "h2d_src list must match H2D_INPUT_SLOT_COUNT");
    for (int s = 0; s < H2D_INPUT_SLOT_COUNT; ++s) {
        std::memcpy(g_buf.h_stage + (size_t)s * all, h2d_src[s], bytes);
    }
    CUDA_CHECK(cudaMemcpy(D(S_chi), g_buf.h_stage,
                          (size_t)H2D_INPUT_SLOT_COUNT * bytes,
                          cudaMemcpyHostToDevice));

    /* ============================================================ */
    /*  Phase 1: prep — alpn1, chin1, gxx, gyy, gzz                 */
    /* ============================================================ */
    kern_phase1_prep<<<grid(all),BLK>>>(
        D(S_Lap), D(S_chi), D(S_dxx), D(S_dyy), D(S_dzz),
        D(S_alpn1), D(S_chin1), D(S_gxx), D(S_gyy), D(S_gzz));

    /* 12x fderivs */
    gpu_fderivs(D(S_betax), D(S_betaxx),D(S_betaxy),D(S_betaxz), ANTI,SYM,SYM, all);
    gpu_fderivs(D(S_betay), D(S_betayx),D(S_betayy),D(S_betayz), SYM,ANTI,SYM, all);
    gpu_fderivs(D(S_betaz), D(S_betazx),D(S_betazy),D(S_betazz), SYM,SYM,ANTI, all);
    gpu_fderivs(D(S_chi),   D(S_chix),D(S_chiy),D(S_chiz),       SYM,SYM,SYM, all);
    gpu_fderivs(D(S_dxx),   D(S_gxxx),D(S_gxxy),D(S_gxxz),       SYM,SYM,SYM, all);
    gpu_fderivs(D(S_gxy),   D(S_gxyx),D(S_gxyy),D(S_gxyz),       ANTI,ANTI,SYM, all);
    gpu_fderivs(D(S_gxz),   D(S_gxzx),D(S_gxzy),D(S_gxzz),       ANTI,SYM,ANTI, all);
    gpu_fderivs(D(S_dyy),   D(S_gyyx),D(S_gyyy),D(S_gyyz),       SYM,SYM,SYM, all);
    gpu_fderivs(D(S_gyz),   D(S_gyzx),D(S_gyzy),D(S_gyzz),       SYM,ANTI,ANTI, all);
    gpu_fderivs(D(S_dzz),   D(S_gzzx),D(S_gzzy),D(S_gzzz),       SYM,SYM,SYM, all);
    gpu_fderivs(D(S_Lap),   D(S_Lapx),D(S_Lapy),D(S_Lapz),       SYM,SYM,SYM, all);
    gpu_fderivs(D(S_trK),   D(S_Kx),D(S_Ky),D(S_Kz),             SYM,SYM,SYM, all);

    /* ============================================================ */
    /*  Phase 2: metric RHS + inverse                                */
    /* ============================================================ */
    kern_phase2_metric_rhs<<<grid(all),BLK>>>(
        D(S_alpn1), D(S_chin1),
        D(S_gxx), D(S_gxy), D(S_gxz), D(S_gyy), D(S_gyz), D(S_gzz),
        D(S_trK),
        D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz),
        D(S_betaxx), D(S_betaxy), D(S_betaxz),
        D(S_betayx), D(S_betayy), D(S_betayz),
        D(S_betazx), D(S_betazy), D(S_betazz),
        D(S_chi_rhs), D(S_gxx_rhs), D(S_gyy_rhs), D(S_gzz_rhs),
        D(S_gxy_rhs), D(S_gyz_rhs), D(S_gxz_rhs));

    kern_phase2_inverse<<<grid(all),BLK>>>(
        D(S_gxx), D(S_gxy), D(S_gxz), D(S_gyy), D(S_gyz), D(S_gzz),
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz));

    /* Phase 3: Gamma constraint (co==0) */
    if (co == 0) {
        kern_phase3_gamma_constraint<<<grid(all),BLK>>>(
            D(S_Gamx), D(S_Gamy), D(S_Gamz),
            D(S_gupxx), D(S_gupxy), D(S_gupxz),
            D(S_gupyy), D(S_gupyz), D(S_gupzz),
            D(S_gxxx), D(S_gxyx), D(S_gxzx), D(S_gyyx), D(S_gyzx), D(S_gzzx),
            D(S_gxxy), D(S_gxyy), D(S_gxzy), D(S_gyyy), D(S_gyzy), D(S_gzzy),
            D(S_gxxz), D(S_gxyz), D(S_gxzz), D(S_gyyz), D(S_gyzz), D(S_gzzz),
            D(S_Gmx_Res), D(S_Gmy_Res), D(S_Gmz_Res));
    }

    /* Phase 4: Christoffel symbols */
    kern_phase4_christoffel<<<grid(all),BLK>>>(
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_gxxx), D(S_gxyx), D(S_gxzx), D(S_gyyx), D(S_gyzx), D(S_gzzx),
        D(S_gxxy), D(S_gxyy), D(S_gxzy), D(S_gyyy), D(S_gyzy), D(S_gzzy),
        D(S_gxxz), D(S_gxyz), D(S_gxzz), D(S_gyyz), D(S_gyzz), D(S_gzzz),
        D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
        D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
        D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
        D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
        D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
        D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz));

    /* Phase 5+6: raise A in registers, then build Gamma_rhs part 1 */
    kern_phase5_6_gamma_rhs_part1_fused<<<grid(all),BLK>>>(
        D(S_Lapx), D(S_Lapy), D(S_Lapz),
        D(S_alpn1), D(S_chin1),
        D(S_chix), D(S_chiy), D(S_chiz),
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz),
        D(S_Kx), D(S_Ky), D(S_Kz),
        D(S_Sx), D(S_Sy), D(S_Sz),
        D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
        D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
        D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
        D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
        D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
        D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz),
        D(S_Gamx_rhs), D(S_Gamy_rhs), D(S_Gamz_rhs));

    /* Phase 7: fdderivs(beta) + fderivs(Gamma) */
    gpu_fdderivs(D(S_betax), D(S_gxxx),D(S_gxyx),D(S_gxzx),
                 D(S_gyyx),D(S_gyzx),D(S_gzzx), ANTI,SYM,SYM, all);
    gpu_fdderivs(D(S_betay), D(S_gxxy),D(S_gxyy),D(S_gxzy),
                 D(S_gyyy),D(S_gyzy),D(S_gzzy), SYM,ANTI,SYM, all);
    gpu_fdderivs(D(S_betaz), D(S_gxxz),D(S_gxyz),D(S_gxzz),
                 D(S_gyyz),D(S_gyzz),D(S_gzzz), SYM,SYM,ANTI, all);
    gpu_fderivs(D(S_Gamx), D(S_Gamxx),D(S_Gamxy),D(S_Gamxz), ANTI,SYM,SYM, all);
    gpu_fderivs(D(S_Gamy), D(S_Gamyx),D(S_Gamyy_t),D(S_Gamyz_t), SYM,ANTI,SYM, all);
    gpu_fderivs(D(S_Gamz), D(S_Gamzx),D(S_Gamzy),D(S_Gamzz_t), SYM,SYM,ANTI, all);

    /* Phase 8: Gamma_rhs part 2 */
    kern_phase8_9_gamma_rhs_contract_fused<<<grid(all),BLK>>>(
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_gxxx),D(S_gxyx),D(S_gxzx),D(S_gyyx),D(S_gyzx),D(S_gzzx),
        D(S_gxxy),D(S_gxyy),D(S_gxzy),D(S_gyyy),D(S_gyzy),D(S_gzzy),
        D(S_gxxz),D(S_gxyz),D(S_gxzz),D(S_gyyz),D(S_gyzz),D(S_gzzz),
        D(S_Gamxxx),D(S_Gamxxy),D(S_Gamxxz),
        D(S_Gamxyy),D(S_Gamxyz),D(S_Gamxzz),
        D(S_Gamyxx),D(S_Gamyxy),D(S_Gamyxz),
        D(S_Gamyyy),D(S_Gamyyz),D(S_Gamyzz),
        D(S_Gamzxx),D(S_Gamzxy),D(S_Gamzxz),
        D(S_Gamzyy),D(S_Gamzyz),D(S_Gamzzz),
        D(S_betaxx),D(S_betaxy),D(S_betaxz),
        D(S_betayx),D(S_betayy),D(S_betayz),
        D(S_betazx),D(S_betazy),D(S_betazz),
        D(S_gxx),D(S_gxy),D(S_gxz),D(S_gyy),D(S_gyz),D(S_gzz),
        D(S_Gamx_rhs),D(S_Gamy_rhs),D(S_Gamz_rhs),
        D(S_Gamxa),D(S_Gamya),D(S_Gamza),
        D(S_gxxx),D(S_gxyx),D(S_gxzx),D(S_gyyx),D(S_gyzx),D(S_gzzx),
        D(S_gxxy),D(S_gxyy),D(S_gxzy),D(S_gyyy),D(S_gyzy),D(S_gzzy),
        D(S_gxxz),D(S_gxyz),D(S_gxzz),D(S_gyyz),D(S_gyzz),D(S_gzzz));

    /* Phase 10: 6x fdderivs(metric) + Ricci contract */
    {
        double *src_fields[] = {D(S_dxx), D(S_dyy), D(S_dzz), D(S_gxy), D(S_gxz), D(S_gyz)};
        double *dst_fields[] = {D(S_Rxx), D(S_Ryy), D(S_Rzz), D(S_Rxy), D(S_Rxz), D(S_Ryz)};
        const int soa_signs[] = {
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)SYM,  (int)SYM,  (int)SYM,
            (int)ANTI, (int)ANTI, (int)SYM,
            (int)ANTI, (int)SYM,  (int)ANTI,
            (int)SYM,  (int)ANTI, (int)ANTI
        };
        gpu_phase10_ricci_batch(D(S_gupxx), D(S_gupxy), D(S_gupxz),
                                D(S_gupyy), D(S_gupyz), D(S_gupzz),
                                src_fields, dst_fields, soa_signs, all);
    }

    /* Phase 11: fused Ricci assembly */
    kern_phase11_ricci_fused<<<grid(all),BLK>>>(
        D(S_gxx),D(S_gxy),D(S_gxz),D(S_gyy),D(S_gyz),D(S_gzz),
        D(S_gupxx),D(S_gupxy),D(S_gupxz),D(S_gupyy),D(S_gupyz),D(S_gupzz),
        D(S_Gamxa),D(S_Gamya),D(S_Gamza),
        D(S_Gamxx),D(S_Gamxy),D(S_Gamxz),
        D(S_Gamyx),D(S_Gamyy_t),D(S_Gamyz_t),
        D(S_Gamzx),D(S_Gamzy),D(S_Gamzz_t),
        D(S_Gamxxx),D(S_Gamxxy),D(S_Gamxxz),
        D(S_Gamxyy),D(S_Gamxyz),D(S_Gamxzz),
        D(S_Gamyxx),D(S_Gamyxy),D(S_Gamyxz),
        D(S_Gamyyy),D(S_Gamyyz),D(S_Gamyzz),
        D(S_Gamzxx),D(S_Gamzxy),D(S_Gamzxz),
        D(S_Gamzyy),D(S_Gamzyz),D(S_Gamzzz),
        D(S_gxxx),D(S_gxyx),D(S_gxzx),D(S_gyyx),D(S_gyzx),D(S_gzzx),
        D(S_gxxy),D(S_gxyy),D(S_gxzy),D(S_gyyy),D(S_gyzy),D(S_gzzy),
        D(S_gxxz),D(S_gxyz),D(S_gxzz),D(S_gyyz),D(S_gyzz),D(S_gzzz),
        D(S_Rxx),D(S_Rxy),D(S_Rxz),
        D(S_Ryy),D(S_Ryz),D(S_Rzz));

    /* ============================================================ */
    /*  Phase 12/13: chi fdderivs + chi correction                   */
    /* ============================================================ */
    kern_phase12_13_chi_correction_fused<<<grid((size_t)all),BLK>>>(
        D(S_chi), D(S_chin1),
        D(S_chix), D(S_chiy), D(S_chiz),
        D(S_gxx), D(S_gxy), D(S_gxz), D(S_gyy), D(S_gyz), D(S_gzz),
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
        D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
        D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
        D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
        D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
        D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz),
        D(S_Rxx), D(S_Rxy), D(S_Rxz),
        D(S_Ryy), D(S_Ryz), D(S_Rzz));

    /* ============================================================ */
    /*  Phase 14/15: fused trK_rhs, Aij_rhs, gauge                   */
    /* ============================================================ */
    kern_phase15_trK_Aij_gauge<<<grid(all),BLK>>>(
        D(S_alpn1), D(S_chin1),
        D(S_chix), D(S_chiy), D(S_chiz),
        D(S_gxx), D(S_gxy), D(S_gxz), D(S_gyy), D(S_gyz), D(S_gzz),
        D(S_gupxx), D(S_gupxy), D(S_gupxz),
        D(S_gupyy), D(S_gupyz), D(S_gupzz),
        D(S_trK),
        D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz),
        D(S_Lapx), D(S_Lapy), D(S_Lapz),
        D(S_betaxx), D(S_betaxy), D(S_betaxz),
        D(S_betayx), D(S_betayy), D(S_betayz),
        D(S_betazx), D(S_betazy), D(S_betazz),
        D(S_rho),
        D(S_Sx), D(S_Sy), D(S_Sz),
        D(S_Sxx), D(S_Sxy), D(S_Sxz), D(S_Syy), D(S_Syz), D(S_Szz),
        D(S_dtSfx), D(S_dtSfy), D(S_dtSfz),
        D(S_Rxx), D(S_Rxy), D(S_Rxz), D(S_Ryy), D(S_Ryz), D(S_Rzz),
        D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
        D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
        D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
        D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
        D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
        D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz),
        D(S_dtSfx_rhs), D(S_dtSfy_rhs), D(S_dtSfz_rhs),
        D(S_trK_rhs),
        D(S_Axx_rhs), D(S_Axy_rhs), D(S_Axz_rhs),
        D(S_Ayy_rhs), D(S_Ayz_rhs), D(S_Azz_rhs),
        D(S_Lap_rhs),
        D(S_betax_rhs), D(S_betay_rhs), D(S_betaz_rhs),
        D(S_Gamx_rhs), D(S_Gamy_rhs), D(S_Gamz_rhs),
        D(S_f_arr), D(S_S_arr));

    /* ============================================================ */
    /*  Phase 16/17: advection + KO dissipation (shared ord=3 pack)  */
    /* ============================================================ */
    gpu_lopsided_kodis_state_batch(eps, all);

    /* ============================================================ */
    /*  Phase 18: Hamilton & momentum constraints (co==0)             */
    /* ============================================================ */
    if (co == 0) {
        /* 6x fderivs on Aij — reuse gxxx..gzzz slots for dA/dx output */
        gpu_fderivs(D(S_Axx), D(S_gxxx),D(S_gxxy),D(S_gxxz), SYM,SYM,SYM,      all);
        gpu_fderivs(D(S_Axy), D(S_gxyx),D(S_gxyy),D(S_gxyz), ANTI,ANTI,SYM,     all);
        gpu_fderivs(D(S_Axz), D(S_gxzx),D(S_gxzy),D(S_gxzz), ANTI,SYM,ANTI,     all);
        gpu_fderivs(D(S_Ayy), D(S_gyyx),D(S_gyyy),D(S_gyyz), SYM,SYM,SYM,       all);
        gpu_fderivs(D(S_Ayz), D(S_gyzx),D(S_gyzy),D(S_gyzz), SYM,ANTI,ANTI,     all);
        gpu_fderivs(D(S_Azz), D(S_gzzx),D(S_gzzy),D(S_gzzz), SYM,SYM,SYM,       all);

        kern_phase18_constraints<<<grid(all),BLK>>>(
            D(S_chin1),
            D(S_chix), D(S_chiy), D(S_chiz),
            D(S_gupxx), D(S_gupxy), D(S_gupxz),
            D(S_gupyy), D(S_gupyz), D(S_gupzz),
            D(S_trK),
            D(S_Axx), D(S_Axy), D(S_Axz), D(S_Ayy), D(S_Ayz), D(S_Azz),
            D(S_Rxx), D(S_Rxy), D(S_Rxz), D(S_Ryy), D(S_Ryz), D(S_Rzz),
            D(S_rho), D(S_Sx), D(S_Sy), D(S_Sz),
            D(S_Kx), D(S_Ky), D(S_Kz),
            D(S_Gamxxx), D(S_Gamxxy), D(S_Gamxxz),
            D(S_Gamxyy), D(S_Gamxyz), D(S_Gamxzz),
            D(S_Gamyxx), D(S_Gamyxy), D(S_Gamyxz),
            D(S_Gamyyy), D(S_Gamyyz), D(S_Gamyzz),
            D(S_Gamzxx), D(S_Gamzxy), D(S_Gamzxz),
            D(S_Gamzyy), D(S_Gamzyz), D(S_Gamzzz),
            /* dA/dx arrays */
            D(S_gxxx), D(S_gxxy), D(S_gxxz),
            D(S_gxyx), D(S_gxyy), D(S_gxyz),
            D(S_gxzx), D(S_gxzy), D(S_gxzz),
            D(S_gyyx), D(S_gyyy), D(S_gyyz),
            D(S_gyzx), D(S_gyzy), D(S_gyzz),
            D(S_gzzx), D(S_gzzy), D(S_gzzz),
            D(S_ham_Res), D(S_movx_Res), D(S_movy_Res), D(S_movz_Res));
    }

    /* ============================================================ */
    /*  D2H: copy all output arrays back to host                     */
    /* ============================================================ */
    const int d2h_slot_count = D2H_BASE_SLOT_COUNT +
                               ((co == 0) ? D2H_CONSTRAINT_SLOT_COUNT : 0);
    CUDA_CHECK(cudaMemcpy(g_buf.h_stage, D(S_chi_rhs),
                          (size_t)d2h_slot_count * bytes,
                          cudaMemcpyDeviceToHost));

    double *d2h_dst[] = {
        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,
        Gamxxx, Gamxxy, Gamxxz, Gamxyy, Gamxyz, Gamxzz,
        Gamyxx, Gamyxy, Gamyxz, Gamyyy, Gamyyz, Gamyzz,
        Gamzxx, Gamzxy, Gamzxz, Gamzyy, Gamzyz, Gamzzz,
        Rxx, Rxy, Rxz, Ryy, Ryz, Rzz
    };
    static_assert((int)(sizeof(d2h_dst) / sizeof(d2h_dst[0])) == D2H_BASE_SLOT_COUNT,
                  "d2h_dst list must match D2H_BASE_SLOT_COUNT");
    for (int s = 0; s < D2H_BASE_SLOT_COUNT; ++s) {
        std::memcpy(d2h_dst[s], g_buf.h_stage + (size_t)s * all, bytes);
    }
    if (co == 0) {
        double *d2h_dst_co[] = {
            ham_Res, movx_Res, movy_Res, movz_Res, Gmx_Res, Gmy_Res, Gmz_Res
        };
        static_assert((int)(sizeof(d2h_dst_co) / sizeof(d2h_dst_co[0])) ==
                          D2H_CONSTRAINT_SLOT_COUNT,
                      "d2h_dst_co list must match D2H_CONSTRAINT_SLOT_COUNT");
        for (int s = 0; s < D2H_CONSTRAINT_SLOT_COUNT; ++s) {
            std::memcpy(d2h_dst_co[s],
                        g_buf.h_stage + (size_t)(D2H_BASE_SLOT_COUNT + s) * all,
                        bytes);
        }
    }

    #undef D
    return 0;
}

extern "C"
int bssn_cuda_rk4_substep(void *block_tag,
                          int *ex, double *X, double *Y, double *Z,
                          double **state_host_in,
                          double **state_host_out,
                          double **matter_host,
                          const double *propspeed,
                          const double *soa_flat,
                          const double *bbox,
                          double &dT,
                          double &T,
                          int &RK4,
                          int &apply_bam_bc,
                          int &Symmetry,
                          int &Lev,
                          double &eps,
                          int &co,
                          int &use_zero_matter,
                          int &keep_resident_state,
                          int &apply_enforce_ga,
                          double &chitiny)
{
    (void)T;
    (void)state_host_out;

    if (RK4 < 0 || RK4 > 3) return 1;

    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    const bool profile = cuda_profile_enabled();
    const double t_total0 = profile ? cuda_profile_now_ms() : 0.0;
    double state_ms = 0.0;
    double matter_ms = 0.0;
    double rhs_ms = 0.0;
    double bc_ms = 0.0;
    double finalize_ms = 0.0;
    double output_ms = 0.0;

    const size_t all = (size_t)ex[0] * ex[1] * ex[2];
    const size_t bytes = all * sizeof(double);
    int touch_xmin = 0, touch_xmax = 0;
    int touch_ymin = 0, touch_ymax = 0;
    int touch_zmin = 0, touch_zmax = 0;

    setup_grid_params(ex, X, Y, Z, Symmetry, eps, co);
    if (Lev > 0) {
        compute_patch_boundary_flags(ex, X, Y, Z, bbox, Symmetry,
                                     touch_xmin, touch_xmax,
                                     touch_ymin, touch_ymax,
                                     touch_zmin, touch_zmax);
    }
    StepContext &ctx = ensure_step_ctx(block_tag, all);
    const bool use_resident_state = (keep_resident_state != 0);
    if (use_resident_state) {
        bind_state_input_slots(ctx.d_state_curr);
        bind_state_output_slots(ctx.d_state_next);
    }
    double t0 = profile ? cuda_profile_now_ms() : 0.0;
    if (!use_resident_state || !ctx.state_ready) {
        upload_state_inputs(state_host_in, all);
    }
    if (apply_enforce_ga) {
        kern_enforce_ga_cuda<<<grid(all), BLK>>>(g_buf.slot[S_dxx], g_buf.slot[S_gxy], g_buf.slot[S_gxz],
                                                 g_buf.slot[S_dyy], g_buf.slot[S_gyz], g_buf.slot[S_dzz],
                                                 g_buf.slot[S_Axx], g_buf.slot[S_Axy], g_buf.slot[S_Axz],
                                                 g_buf.slot[S_Ayy], g_buf.slot[S_Ayz], g_buf.slot[S_Azz]);
    }
    if (profile) {
        cuda_profile_sync();
        state_ms += cuda_profile_now_ms() - t0;
    }

    t0 = profile ? cuda_profile_now_ms() : 0.0;
    if (RK4 == 0) {
        if (use_zero_matter) {
            if (!ctx.matter_ready) zero_matter_cache(ctx, all);
        } else {
            upload_matter_cache(ctx, matter_host, all);
        }
        CUDA_CHECK(cudaMemcpy(ctx.d_state0_mem, g_buf.slot[S_chi],
                              (size_t)BSSN_STATE_COUNT * bytes,
                              cudaMemcpyDeviceToDevice));
    } else if (!ctx.matter_ready) {
        if (use_zero_matter) zero_matter_cache(ctx, all);
        else upload_matter_cache(ctx, matter_host, all);
    }
    bind_matter_slots(ctx);
    if (profile) {
        cuda_profile_sync();
        matter_ms += cuda_profile_now_ms() - t0;
    }

    t0 = profile ? cuda_profile_now_ms() : 0.0;
    launch_rhs_pipeline((int)all, eps, co);
    if (profile) {
        cuda_profile_sync();
        rhs_ms += cuda_profile_now_ms() - t0;
    }

    t0 = profile ? cuda_profile_now_ms() : 0.0;
    if (apply_bam_bc) {
        for (int i = 0; i < BSSN_STATE_COUNT; ++i) {
            gpu_sommerfeld_routbam(g_buf.slot[k_state_input_slots[i]],
                                   g_buf.slot[k_state_rhs_slots[i]],
                                   propspeed[i],
                                   soa_flat[3 * i + 0],
                                   soa_flat[3 * i + 1],
                                   soa_flat[3 * i + 2],
                                   X, Y, Z, bbox, Symmetry);
        }
    }
    if (profile) {
        cuda_profile_sync();
        bc_ms += cuda_profile_now_ms() - t0;
    }

    t0 = profile ? cuda_profile_now_ms() : 0.0;
    gpu_rk4_finalize_batch(ctx, all, dT, RK4, chitiny);
    if (Lev > 0) {
        gpu_copy_patch_boundary_batch((int)all,
                                      touch_xmin, touch_xmax,
                                      touch_ymin, touch_ymax,
                                      touch_zmin, touch_zmax);
    }
    if (profile) {
        cuda_profile_sync();
        finalize_ms += cuda_profile_now_ms() - t0;
    }

    t0 = profile ? cuda_profile_now_ms() : 0.0;
    if (use_resident_state) {
        std::swap(ctx.d_state_curr_mem, ctx.d_state_next_mem);
        ctx.d_state_curr.swap(ctx.d_state_next);
        ctx.state_ready = true;
    } else {
        download_state_outputs(state_host_out, all);
    }
    if (RK4 == 3) {
        ctx.matter_ready = false;  /* invalidate matter cache for next timestep */
    }
    if (profile) {
        cuda_profile_sync();
        output_ms += cuda_profile_now_ms() - t0;
        CudaProfileStats &stats = cuda_profile_stats();
        stats.calls++;
        stats.total_ms += cuda_profile_now_ms() - t_total0;
        stats.state_ms += state_ms;
        stats.matter_ms += matter_ms;
        stats.rhs_ms += rhs_ms;
        stats.bc_ms += bc_ms;
        stats.finalize_ms += finalize_ms;
        stats.output_ms += output_ms;
        cuda_profile_maybe_log();
    }
    return 0;
}

extern "C"
int bssn_cuda_copy_state_region_to_host(void *block_tag,
                                        int state_index,
                                        double *host_state,
                                        int *ex,
                                        int i0, int j0, int k0,
                                        int sx, int sy, int sz)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_region_cuda(block_tag, state_index, host_state, ex,
                           i0, j0, k0, sx, sy, sz, cudaMemcpyDeviceToHost);
    return 0;
}

extern "C"
int bssn_cuda_copy_state_region_from_host(void *block_tag,
                                          int state_index,
                                          double *host_state,
                                          int *ex,
                                          int i0, int j0, int k0,
                                          int sx, int sy, int sz)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_region_cuda(block_tag, state_index, host_state, ex,
                           i0, j0, k0, sx, sy, sz, cudaMemcpyHostToDevice);
    return 0;
}

extern "C"
int bssn_cuda_download_resident_state(void *block_tag,
                                      int *ex,
                                      double **state_host_out)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    download_resident_state(block_tag, ex, state_host_out);
    return 0;
}

extern "C"
int bssn_cuda_download_constraint_outputs(int *ex,
                                          double **constraint_host_out)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    const size_t all = (size_t)ex[0] * ex[1] * ex[2];
    download_constraint_outputs(constraint_host_out, all);
    return 0;
}

extern "C"
int bssn_cuda_pack_state_region_to_host_buffer(void *block_tag,
                                               int state_index,
                                               double *host_buffer,
                                               int *ex,
                                               int i0, int j0, int k0,
                                               int sx, int sy, int sz)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_region_packed_cuda(block_tag, state_index, host_buffer, ex,
                                  i0, j0, k0, sx, sy, sz, cudaMemcpyDeviceToHost);
    return 0;
}

extern "C"
int bssn_cuda_unpack_state_region_from_host_buffer(void *block_tag,
                                                   int state_index,
                                                   double *host_buffer,
                                                   int *ex,
                                                   int i0, int j0, int k0,
                                                   int sx, int sy, int sz)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_region_packed_cuda(block_tag, state_index, host_buffer, ex,
                                  i0, j0, k0, sx, sy, sz, cudaMemcpyHostToDevice);
    return 0;
}

extern "C"
int bssn_cuda_pack_state_batch_to_host_buffer(void *block_tag,
                                              int state_count,
                                              double *host_buffer,
                                              int *ex,
                                              int i0, int j0, int k0,
                                              int sx, int sy, int sz)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_region_packed_batch_cuda(block_tag, state_count, host_buffer, ex,
                                        i0, j0, k0, sx, sy, sz,
                                        cudaMemcpyDeviceToHost);
    return 0;
}

extern "C"
int bssn_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
                                                  int state_count,
                                                  double *host_buffer,
                                                  int *ex,
                                                  int i0, int j0, int k0,
                                                  int sx, int sy, int sz)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_region_packed_batch_cuda(block_tag, state_count, host_buffer, ex,
                                        i0, j0, k0, sx, sy, sz,
                                        cudaMemcpyHostToDevice);
    return 0;
}

static void copy_state_device_batch(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 (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]);
    }
}

static void copy_state_device_segments(void *block_tag,
                                       int state_count,
                                       double *device_buffer,
                                       const int *ex,
                                       int segment_count,
                                       const int *segment_meta,
                                       int pack_not_unpack)
{
    if (state_count <= 0 || state_count > BSSN_STATE_COUNT) return;
    if (segment_count <= 0 || !segment_meta) return;

    int max_region_all = 0;
    for (int s = 0; s < segment_count; ++s) {
        const int *m = segment_meta + s * 8;
        if (m[3] <= 0 || m[4] <= 0 || m[5] <= 0 || m[6] <= 0) return;
        if (m[6] > max_region_all) max_region_all = m[6];
    }
    if (max_region_all <= 0) return;

    StepContext &ctx = ensure_step_ctx(block_tag, (size_t)ex[0] * ex[1] * ex[2]);
    int *d_meta = ensure_comm_segment_meta_buffer((size_t)segment_count * 8);
    CUDA_CHECK(cudaMemcpy(d_meta, segment_meta,
                          (size_t)segment_count * 8 * sizeof(int),
                          cudaMemcpyHostToDevice));

    dim3 launch_grid((unsigned int)grid((size_t)max_region_all),
                     (unsigned int)state_count,
                     (unsigned int)segment_count);
    if (pack_not_unpack) {
        kern_pack_state_segments_batch<<<launch_grid, BLK>>>(
            ctx.d_state_curr_mem, device_buffer,
            ex[0], ex[1], d_meta, state_count,
            ex[0] * ex[1] * ex[2]);
    } else {
        kern_unpack_state_segments_batch<<<launch_grid, BLK>>>(
            ctx.d_state_curr_mem, device_buffer,
            ex[0], ex[1], d_meta, state_count,
            ex[0] * ex[1] * ex[2]);
    }
}

extern "C"
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)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_device_batch(block_tag, state_count, device_buffer, ex,
                            i0, j0, k0, sx, sy, sz, 1);
    return 0;
}

extern "C"
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)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_device_batch(block_tag, state_count, device_buffer, ex,
                            i0, j0, k0, sx, sy, sz, 0);
    return 0;
}

extern "C"
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)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_device_segments(block_tag, state_count, device_buffer, ex,
                               segment_count, segment_meta, 1);
    return 0;
}

extern "C"
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)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_device_segments(block_tag, state_count, device_buffer, ex,
                               segment_count, segment_meta, 0);
    return 0;
}

extern "C"
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)
{
    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;
    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 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)
{
    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;
    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 bssn_cuda_download_state_subset(void *block_tag,
                                    int *ex,
                                    int subset_count,
                                    const int *state_indices,
                                    double **state_host_out)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_subset(block_tag, ex, subset_count, state_indices, state_host_out,
                      cudaMemcpyDeviceToHost);
    return 0;
}

extern "C"
int bssn_cuda_upload_state_subset(void *block_tag,
                                  int *ex,
                                  int subset_count,
                                  const int *state_indices,
                                  double **state_host_in)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    copy_state_subset(block_tag, ex, subset_count, state_indices, state_host_in,
                      cudaMemcpyHostToDevice);
    return 0;
}

extern "C"
int bssn_cuda_has_resident_state(void *block_tag)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    return has_resident_state(block_tag) ? 1 : 0;
}

extern "C"
void bssn_cuda_release_step_ctx(void *block_tag)
{
    init_gpu_dispatch();
    CUDA_CHECK(cudaSetDevice(g_dispatch.my_device));
    release_step_ctx(block_tag);
}