Merge R6-2: the moving mask's classification, merging, face shifts and GCL table on the device (RTX_E3_MASK_DEVICE=1, default off); on top of R6-1's default-on device geometry. Merge fix: the mask knob requires the geometry not switched off (it required RTX_E3_GEOM_DEVICE=1 before the geometry became the default). Gate on the merged tree: slab ny 62 one period default and mask-on+BAND_CHECK both byte-identical to main's base, 0 differs; device suites 2/2, 3/3

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

crates/specialized/rtx-cfd/src/kernels/cuda/e3_mask.cu

@@ -0,0 +1,428 @@
+/**
+ * embedded3 R6-2: the moving mask's classification on the device, behind
+ * `RTX_E3_MASK_DEVICE=1` (with `RTX_E3_GEOM_DEVICE=1`). From the device
+ * geometry of the step (R6-1: the end-of-step apertures in DeviceCut's
+ * predictor set, the volumes and wall vectors in DeviceGeom) and the
+ * previous step's (snapshots taken before the geometry kernels overwrite
+ * them), the host `Mask` expressions in fp64 with FMA contraction off:
+ *
+ *   e3_mask_changed   the changed set (touched by either build, dilated by
+ *                     one; `set_step_apertures_with`);
+ *   e3_mask_face_flags the faces of the changed cells per component;
+ *   e3_mask_cells     per changed cell: fluid (vol > 0), the space-time
+ *                     activity, the GCL entry (`gcl_flux_table`);
+ *   e3_mask_merge     per changed cell: the virtual-merging master
+ *                     (`compute_merging` with the old mask);
+ *   e3_mask_faces     per face of a changed cell: the kind, the step
+ *                     aperture and open flag (`set_step_apertures`), the
+ *                     open-part centroid shift (`compute_face_shifts`);
+ *   e3_mask_wall_flags / e3_mask_wall_area / e3_mask_wall_flux
+ *                     the GCL table's wall cells, areas and the table
+ *                     with the compatibility correction;
+ *   e3_mask_slave_flags / e3_mask_fold_count / e3_mask_fold_fill
+ *                     the merged cells' CSR (slaves ascending per master);
+ *   e3_mask_band_flags the imposition band's faces (|d| <= band);
+ *   e3_scan_*         block scans: stream compaction (ascending) and the
+ *                     exclusive scan of counts.
+ * Every per-entry value depends on its own inputs alone; the two serial
+ * sums of the GCL table (in ascending order) are the host's.
+ */
+
+typedef unsigned int u32;
+typedef unsigned char u8;
+
+struct MaskGrid {
+    int nx, ny, nz, periodic;
+    double dx, dy, dz;
+};
+
+#define SCAN_BLOCK 1024
+
+/* Exclusive block scan of v over a 1024-thread block; *total = the block sum. */
+__device__ u32 block_excl_scan(u32 v, u32* total)
+{
+    __shared__ u32 warp_sums[32];
+    int lane = threadIdx.x & 31, wid = threadIdx.x >> 5;
+    u32 x = v;
+    for (int o = 1; o < 32; o <<= 1) {
+        u32 y = __shfl_up_sync(0xffffffffu, x, o);
+        if (lane >= o) x += y;
+    }
+    if (lane == 31) warp_sums[wid] = x;
+    __syncthreads();
+    if (wid == 0) {
+        u32 s = warp_sums[lane];
+        for (int o = 1; o < 32; o <<= 1) {
+            u32 y = __shfl_up_sync(0xffffffffu, s, o);
+            if (lane >= o) s += y;
+        }
+        warp_sums[lane] = s;
+    }
+    __syncthreads();
+    u32 incl = x + (wid > 0 ? warp_sums[wid - 1] : 0u);
+    *total = warp_sums[31];
+    __syncthreads();
+    return incl - v;
+}
+
+/* Per block of 1024: the number of flags set. */
+extern "C" __global__ void e3_scan_count_flags(long long n, const u8* __restrict__ flags, u32* __restrict__ sums)
+{
+    long long i = (long long) blockIdx.x * SCAN_BLOCK + threadIdx.x;
+    u32 total;
+    block_excl_scan(i < n ? (u32) (flags[i] != 0) : 0u, &total);
+    if (threadIdx.x == 0) sums[blockIdx.x] = total;
+}
+
+/* Per block of 1024: the sum of the values. */
+extern "C" __global__ void e3_scan_count_vals(long long n, const u32* __restrict__ vals, u32* __restrict__ sums)
+{
+    long long i = (long long) blockIdx.x * SCAN_BLOCK + threadIdx.x;
+    u32 total;
+    block_excl_scan(i < n ? vals[i] : 0u, &total);
+    if (threadIdx.x == 0) sums[blockIdx.x] = total;
+}
+
+/* One block: the block sums scanned exclusively in place; sums[nb] = the total. */
+extern "C" __global__ void e3_scan_top(int nb, u32* __restrict__ sums)
+{
+    __shared__ u32 carry;
+    if (threadIdx.x == 0) carry = 0u;
+    __syncthreads();
+    for (int base = 0; base < nb; base += SCAN_BLOCK) {
+        int i = base + threadIdx.x;
+        u32 v = i < nb ? sums[i] : 0u;
+        u32 total;
+        u32 ex = block_excl_scan(v, &total);
+        u32 c = carry;
+        if (i < nb) sums[i] = c + ex;
+        __syncthreads();
+        if (threadIdx.x == 0) carry = c + total;
+        __syncthreads();
+    }
+    if (threadIdx.x == 0) sums[nb] = carry;
+}
+
+/* Stream compaction: the flagged indices, ascending, at their scanned offsets. */
+extern "C" __global__ void e3_scan_compact(long long n, const u8* __restrict__ flags, const u32* __restrict__ offsets, u32* __restrict__ out)
+{
+    long long i = (long long) blockIdx.x * SCAN_BLOCK + threadIdx.x;
+    u32 v = i < n ? (u32) (flags[i] != 0) : 0u;
+    u32 total;
+    u32 ex = block_excl_scan(v, &total);
+    if (v) out[offsets[blockIdx.x] + ex] = (u32) i;
+}
+
+/* The exclusive scan of the values. */
+extern "C" __global__ void e3_scan_values(long long n, const u32* __restrict__ vals, const u32* __restrict__ offsets, u32* __restrict__ out)
+{
+    long long i = (long long) blockIdx.x * SCAN_BLOCK + threadIdx.x;
+    u32 total;
+    u32 ex = block_excl_scan(i < n ? vals[i] : 0u, &total);
+    if (i < n) out[i] = offsets[blockIdx.x] + ex;
+}
+
+/* The changed set: touched by either build, dilated by one (periodic z wraps). */
+extern "C" __global__ void e3_mask_changed(MaskGrid g, const u8* __restrict__ t_new, const u8* __restrict__ t_old, u8* __restrict__ flag)
+{
+    long long idx = (long long) blockIdx.x * blockDim.x + threadIdx.x;
+    long long nx = g.nx, ny = g.ny, nz = g.nz, nxy = nx * ny;
+    if (idx >= nxy * nz) return;
+    long long k = idx / nxy, j = (idx % nxy) / nx, i = idx % nx;
+#define T(q) (t_new[q] | t_old[q])
+    u8 f = T(idx)
+        || (i + 1 < nx && T(idx + 1)) || (i > 0 && T(idx - 1))
+        || (j + 1 < ny && T(idx + nx)) || (j > 0 && T(idx - nx))
+        || (k + 1 < nz && T(idx + nxy)) || (k > 0 && T(idx - nxy))
+        || (g.periodic && nz > 1 && k + 1 == nz && T(idx - (nz - 1) * nxy))
+        || (g.periodic && nz > 1 && k == 0 && T(idx + (nz - 1) * nxy));
+#undef T
+    flag[idx] = f;
+}
+
+/* The faces of the changed cells, component c (a face is its two cells' face). */
+extern "C" __global__ void e3_mask_face_flags(MaskGrid g, int c, const u8* __restrict__ changed, u8* __restrict__ flag)
+{
+    long long f = (long long) blockIdx.x * blockDim.x + threadIdx.x;
+    long long ni = g.nx + (c == 0), nj = g.ny + (c == 1), nk = g.nz + (c == 2);
+    if (f >= ni * nj * nk) return;
+    long long k = f / (nj * ni), j = (f / ni) % nj, i = f % ni;
+    long long nx = g.nx, nxy = (long long) g.nx * g.ny;
+    long long cell = (k * g.ny + j) * nx + i;
+    long long own = c == 0 ? i : (c == 1 ? j : k);
+    long long lim = c == 0 ? g.nx : (c == 1 ? g.ny : g.nz);
+    long long step = c == 0 ? 1 : (c == 1 ? nx : nxy);
+    u8 m = 0;
+    if (own < lim) m |= changed[cell];
+    if (own > 0) m |= changed[cell - step];
+    flag[f] = m;
+}
+
+/* Per changed cell: fluid at the new geometry, the space-time activity (fluid
+ * at either end), the GCL entry `(V^{n+1} − V^n) dv / dt` of an active cell;
+ * the device tables written (activity: step and predictor sets). Out flags:
+ * bit 0 fluid, bit 1 active. */
+extern "C" __global__ void e3_mask_cells(
+    int n, const u32* __restrict__ list, const double* __restrict__ vol, const double* __restrict__ vol_prev,
+    double dv, double dt, int* __restrict__ active, int* __restrict__ active_pred,
+    u8* __restrict__ out_flags, double* __restrict__ out_entry)
+{
+    int t = blockIdx.x * blockDim.x + threadIdx.x;
+    if (t >= n) return;
+    u32 idx = list[t];
+    int fn = vol[idx] > 0.0;
+    int fo = vol_prev[idx] > 0.0;
+    int act = fn || fo;
+    active[idx] = act;
+    active_pred[idx] = fn;
+    out_flags[t] = (u8) (fn | (act << 1));
+    out_entry[t] = act ? (vol[idx] - vol_prev[idx]) * dv / dt : 0.0;
+}
+
+/* lat.cell: the cell at (i, j, k) with the periodic wrap in z; -1 outside. */
+__device__ __forceinline__ long long lat_cell(const MaskGrid& g, long long i, long long j, long long k)
+{
+    if (i < 0 || i >= g.nx || j < 0 || j >= g.ny) return -1;
+    if (g.periodic) {
+        k %= g.nz;
+        if (k < 0) k += g.nz;
+    } else if (k < 0 || k >= g.nz) {
+        return -1;
+    }
+    return (k * g.ny + j) * g.nx + i;
+}
+
+__device__ __forceinline__ int is_small(long long q, const double* vol, const double* vol_prev, const int* active, double threshold)
+{
+    return active[q] && fmax(vol[q], vol_prev[q]) < threshold;
+}
+
+/* Per changed cell: the master (the active, not small face neighbour of
+ * largest new volume, the first on ties) of a small cell. */
+extern "C" __global__ void e3_mask_merge(
+    MaskGrid g, int n, const u32* __restrict__ list, const double* __restrict__ vol,
+    const double* __restrict__ vol_prev, const int* __restrict__ active, double threshold,
+    u32* __restrict__ owner, u32* __restrict__ out_master)
+{
+    int t = blockIdx.x * blockDim.x + threadIdx.x;
+    if (t >= n) return;
+    long long idx = list[t];
+    long long nxy = (long long) g.nx * g.ny;
+    long long k = idx / nxy, j = (idx % nxy) / g.nx, i = idx % g.nx;
+    u32 master = 0xffffffffu;
+    if (is_small(idx, vol, vol_prev, active, threshold)) {
+        int have = 0;
+        double best = 0.0;
+        long long bi = 0;
+        for (int d = 0; d < 3; ++d) {
+            for (int side = -1; side <= 1; side += 2) {
+                long long q = lat_cell(g, i + (d == 0) * side, j + (d == 1) * side, k + (d == 2) * side);
+                if (q < 0) continue;
+                double v = vol[q];
+                if (active[q] && !is_small(q, vol, vol_prev, active, threshold) && (!have || v > best)) {
+                    have = 1;
+                    best = v;
+                    bi = q;
+                }
+            }
+        }
+        if (have) master = (u32) bi;
+    }
+    owner[idx] = master == 0xffffffffu ? (u32) idx : master;
+    out_master[t] = master;
+}
+
+/* lat.face: the index of the face of component c at (i, j, k), -1 outside
+ * (periodic z wraps by nz whatever the component). */
+__device__ __forceinline__ long long lat_face(const MaskGrid& g, int c, long long i, long long j, long long k)
+{
+    long long ni = g.nx + (c == 0), nj = g.ny + (c == 1), nk = g.nz + (c == 2);
+    if (i < 0 || i >= ni || j < 0 || j >= nj) return -1;
+    if (g.periodic) {
+        k %= g.nz;
+        if (k < 0) k += g.nz;
+    } else if (k < 0 || k >= nk) {
+        return -1;
+    }
+    if (c == 0) return (k * g.ny + j) * (g.nx + 1) + i;
+    if (c == 1) return (k * (g.ny + 1) + j) * g.nx + i;
+    return (k * g.ny + j) * g.nx + i;
+}
+
+struct Apertures {
+    const double* a[3];
+};
+
+/* aperture(c, p).unwrap_or(dflt) */
+__device__ __forceinline__ double ap_or(const MaskGrid& g, const Apertures& A, int c, const long long p[3], double dflt)
+{
+    long long f = lat_face(g, c, p[0], p[1], p[2]);
+    return f < 0 ? dflt : A.a[c][f];
+}
+
+/* Per face of a changed cell, component c: the kind (bit 0: fluid), the step
+ * aperture ½(αⁿ⁺¹ + αⁿ) and its open flag (bit 1), the open-part centroid
+ * shift (`compute_face_shifts`, with `has_shift`). */
+extern "C" __global__ void e3_mask_faces(
+    MaskGrid g, int c, int n, const u32* __restrict__ list, Apertures A, const double* __restrict__ a_prev,
+    double* __restrict__ a_step, int* __restrict__ open, int* __restrict__ open_pred, int has_shift,
+    double* __restrict__ shift, u8* __restrict__ out_flags, double* __restrict__ out_a, double* __restrict__ out_shift)
+{
+    int t = blockIdx.x * blockDim.x + threadIdx.x;
+    if (t >= n) return;
+    long long f = list[t];
+    long long ni = g.nx + (c == 0), nj = g.ny + (c == 1);
+    long long k = f / (nj * ni), j = (f / ni) % nj, i = f % ni;
+    double a_new = A.a[c][f];
+    int side;
+    if (c == 0) side = i == 0 || i == g.nx;
+    else if (c == 1) side = j == 0 || j == g.ny;
+    else side = !g.periodic && (k == 0 || k == g.nz);
+    int fluid = side || a_new > 0.0;
+    double as = 0.5 * (a_new + a_prev[f]);
+    int op = as > 0.0;
+    a_step[f] = as;
+    open[f] = op;
+    open_pred[f] = fluid;
+    out_flags[t] = (u8) (fluid | (op << 1));
+    out_a[t] = as;
+    if (!has_shift) return;
+    double out[3] = {0.0, 0.0, 0.0};
+    long long p[3] = {i, j, k};
+    double h[3] = {g.dx, g.dy, g.dz};
+    long long nlim[3] = {g.nx, g.ny, g.nz};
+    long long fa = lat_face(g, c, p[0], p[1], p[2]);
+    /* `self.aperture(c, p)` is Some for a face inside the grid */
+    double alpha = A.a[c][fa];
+    int on_side = p[c] == 0 || p[c] == nlim[c];
+    if (alpha > 0.0 && alpha < 1.0 && !(on_side && !(c == 2 && g.periodic))) {
+        /* cv_geometry(c, p): the control volume's side apertures and wall */
+        double area[3] = {g.dy * g.dz, g.dx * g.dz, g.dx * g.dy};
+        long long cm[3] = {p[0], p[1], p[2]};
+        cm[c] -= 1;
+        double ap[3][2];
+        for (int d = 0; d < 3; ++d) {
+            if (d == c) {
+                long long q[3] = {p[0], p[1], p[2]};
+                q[c] -= 1;
+                double am = ap_or(g, A, c, q, alpha);
+                q[c] += 2;
+                double apl = ap_or(g, A, c, q, alpha);
+                ap[d][0] = 0.5 * (am + alpha);
+                ap[d][1] = 0.5 * (alpha + apl);
+            } else {
+                long long cmd[3] = {cm[0], cm[1], cm[2]};
+                long long cpd[3] = {p[0], p[1], p[2]};
+                cmd[d] += 1;
+                cpd[d] += 1;
+                double minus = 0.5 * (ap_or(g, A, d, cm, 1.0) + ap_or(g, A, d, p, 1.0));
+                double plus = 0.5 * (ap_or(g, A, d, cmd, 1.0) + ap_or(g, A, d, cpd, 1.0));
+                ap[d][0] = minus;
+                ap[d][1] = plus;
+            }
+        }
+        double nw[3];
+        for (int d = 0; d < 3; ++d) nw[d] = -(ap[d][1] - ap[d][0]) * area[d];
+        nw[c] = 0.0;
+        double a = sqrt(nw[0] * nw[0] + nw[1] * nw[1] + nw[2] * nw[2]);
+        if (a != 0.0) {
+            for (int d = 0; d < 3; ++d) out[d] = -0.5 * h[d] * (1.0 - alpha) * nw[d] / a;
+        }
+    }
+    for (int d = 0; d < 3; ++d) {
+        shift[3 * f + d] = out[d];
+        out_shift[3 * t + d] = out[d];
+    }
+}
+
+/* The GCL table's wall cells: active with a nonzero wall vector. */
+extern "C" __global__ void e3_mask_wall_flags(long long nc, const int* __restrict__ active, const double* __restrict__ wall, u8* __restrict__ flag)
+{
+    long long idx = (long long) blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= nc) return;
+    const double* w = wall + 3 * idx;
+    flag[idx] = active[idx] && (w[0] != 0.0 || w[1] != 0.0 || w[2] != 0.0);
+}
+
+/* The wall areas of the listed cells. */
+extern "C" __global__ void e3_mask_wall_area(int n, const u32* __restrict__ list, const double* __restrict__ wall, double* __restrict__ out)
+{
+    int t = blockIdx.x * blockDim.x + threadIdx.x;
+    if (t >= n) return;
+    const double* w = wall + 3 * (long long) list[t];
+    out[t] = sqrt(w[0] * w[0] + w[1] * w[1] + w[2] * w[2]);
+}
+
+/* The GCL wall-flux table: the entries of the changed active cells, less the
+ * correction per unit wall area on every cell with a wall. */
+extern "C" __global__ void e3_mask_wall_flux(
+    long long nc, const u8* __restrict__ changed, const int* __restrict__ active, const double* __restrict__ vol,
+    const double* __restrict__ vol_prev, const double* __restrict__ wall, double dv, double dt, double correction,
+    double* __restrict__ table)
+{
+    long long idx = (long long) blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= nc) return;
+    double v = 0.0;
+    if (changed[idx] && active[idx]) v = (vol[idx] - vol_prev[idx]) * dv / dt;
+    if (correction != 0.0) {
+        const double* w = wall + 3 * idx;
+        if (w[0] != 0.0 || w[1] != 0.0 || w[2] != 0.0) {
+            double a = sqrt(w[0] * w[0] + w[1] * w[1] + w[2] * w[2]);
+            v -= correction * a;
+        }
+    }
+    table[idx] = v;
+}
+
+/* The merged cells: owner != self. */
+extern "C" __global__ void e3_mask_slave_flags(long long nc, const u32* __restrict__ owner, u8* __restrict__ flag)
+{
+    long long idx = (long long) blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= nc) return;
+    flag[idx] = owner[idx] != (u32) idx;
+}
+
+/* The slave count per master (counts zeroed by the caller). */
+extern "C" __global__ void e3_mask_fold_count(int n, const u32* __restrict__ slaves, const u32* __restrict__ owner, u32* __restrict__ counts)
+{
+    int t = blockIdx.x * blockDim.x + threadIdx.x;
+    if (t >= n) return;
+    atomicAdd(&counts[owner[slaves[t]]], 1u);
+}
+
+/* The CSR's slave lists, ascending per master: a slave's rank = its master's
+ * slaves (face neighbours of the master, distinct) of smaller index. */
+extern "C" __global__ void e3_mask_fold_fill(
+    MaskGrid g, int n, const u32* __restrict__ slaves, const u32* __restrict__ owner, const u32* __restrict__ fold_ptr,
+    u32* __restrict__ fold_idx)
+{
+    int t = blockIdx.x * blockDim.x + threadIdx.x;
+    if (t >= n) return;
+    long long s = slaves[t];
+    long long m = owner[s];
+    long long nxy = (long long) g.nx * g.ny;
+    long long k = m / nxy, j = (m % nxy) / g.nx, i = m % g.nx;
+    long long seen[6];
+    int ns = 0;
+    u32 rank = 0;
+    for (int d = 0; d < 3; ++d) {
+        for (int side = -1; side <= 1; side += 2) {
+            long long q = lat_cell(g, i + (d == 0) * side, j + (d == 1) * side, k + (d == 2) * side);
+            if (q < 0 || q == m) continue;
+            int dup = 0;
+            for (int r = 0; r < ns; ++r) dup |= seen[r] == q;
+            if (dup) continue;
+            seen[ns++] = q;
+            if (q < s && owner[q] == (u32) m) ++rank;
+        }
+    }
+    fold_idx[fold_ptr[m] + rank] = (u32) s;
+}
+
+/* The imposition band's faces: |d| <= band. */
+extern "C" __global__ void e3_mask_band_flags(long long n, const double* __restrict__ d, double band, u8* __restrict__ flag)
+{
+    long long f = (long long) blockIdx.x * blockDim.x + threadIdx.x;
+    if (f >= n) return;
+    flag[f] = fabs(d[f]) <= band;
+}

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/cutwall.rs

@@ -13,13 +13,13 @@
 //! the computed one as the aperture closes (the shear coefficient grows as
 //! `1/A_f`), which is what makes the wall smooth in the interface position.
 
-use super::Grid;
 use super::body::Body;
 use super::cut::CutGeometry;
 use super::exchange::in_load_window;
 use super::field::Field;
 use super::step::{Boundaries, Side};
 use super::wall::{FaceKind, Mask};
+use super::Grid;
 
 /// The inertia floor: the momentum volume's fraction in the time
 /// derivative is at least this.
@@ -32,6 +32,12 @@ pub(super) const DISTANCE_FLOOR_FINE: f64 = 0.01;
 /// step) stays below this shares its pressure unknown with a neighbour.
 pub(super) const MERGE_FRACTION: f64 = 0.1;
 
+/// `RTX_E3_MOVING_PROFILE` (read once).
+pub(super) fn profile_on() -> bool {
+    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
+    *ON.get_or_init(|| std::env::var("RTX_E3_MOVING_PROFILE").is_ok())
+}
+
 /// Lattice addressing of faces and cells with the periodic wrap in z as
 /// data: a face of component `c` at `p = [i, j, k]` (its own coordinate is
 /// the face index, the others the cell's), a cell at `[i, j, k]`.
@@ -138,6 +144,7 @@ impl Mask {
     /// R6-1: the mask of a cut geometry built elsewhere (the device's
     /// `DeviceGeom` mirror); `build_cut_from` is this after `CutGeometry::build_from`.
     pub fn from_cut(cut: CutGeometry, g: Grid, b: Boundaries) -> Result<Self, String> {
+        let lap = std::time::Instant::now();
         let (nx, ny, nz) = (g.nx, g.ny, g.nz);
         let periodic = b.z0 == Side::Periodic;
         let allowed = |side: Side| matches!(side, Side::Velocity | Side::Periodic | Side::SlipWall);
@@ -148,7 +155,9 @@ impl Mask {
         let nxy = nx * ny;
         let cell_fluid: Vec<bool> = cut.vol.par_iter().map(|&v| v > 0.0).collect();
         let fluid_cells = cell_fluid.par_iter().filter(|&&f| f).count();
-        let anchor = (0..g.cells()).into_par_iter().find_first(|&idx| cell_fluid[idx]);
+        let anchor = (0..g.cells())
+            .into_par_iter()
+            .find_first(|&idx| cell_fluid[idx]);
         let touching = (0..g.cells()).into_par_iter().find_first(|&idx| {
             if cell_fluid[idx] {
                 return false;
@@ -182,24 +191,69 @@ impl Mask {
             .into_par_iter()
             .map(|f| {
                 let i = f % (nx + 1);
-                if i == 0 || i == nx { FaceKind::Fluid } else { kind(cut.a_u[f]) }
+                if i == 0 || i == nx {
+                    FaceKind::Fluid
+                } else {
+                    kind(cut.a_u[f])
+                }
             })
             .collect();
         let v_kind: Vec<FaceKind> = (0..g.n_vfaces())
             .into_par_iter()
             .map(|f| {
                 let j = (f / nx) % (ny + 1);
-                if j == 0 || j == ny { FaceKind::Fluid } else { kind(cut.a_v[f]) }
+                if j == 0 || j == ny {
+                    FaceKind::Fluid
+                } else {
+                    kind(cut.a_v[f])
+                }
             })
             .collect();
         let w_kind: Vec<FaceKind> = (0..g.n_wfaces())
             .into_par_iter()
             .map(|f| {
                 let k = f / nxy;
-                if !periodic && (k == 0 || k == nz) { FaceKind::Fluid } else { kind(cut.a_w[f]) }
+                if !periodic && (k == 0 || k == nz) {
+                    FaceKind::Fluid
+                } else {
+                    kind(cut.a_w[f])
+                }
             })
             .collect();
-        let mut mask = Self {
+        let mut mask = Self::from_parts(
+            g,
+            periodic,
+            cell_fluid,
+            [u_kind, v_kind, w_kind],
+            anchor,
+            fluid_cells,
+            cut,
+        );
+        let l_class = lap.elapsed();
+        mask.compute_merging(None);
+        if profile_on() {
+            eprintln!(
+                "    from_cut laps: classification {:.0} ms, merging {:.0} ms",
+                l_class.as_secs_f64() * 1e3,
+                (lap.elapsed() - l_class).as_secs_f64() * 1e3
+            );
+        }
+        Ok(mask)
+    }
+
+    /// The mask struct of a cut classification (every closure flag at its
+    /// default; the caller sets them).
+    pub(super) fn from_parts(
+        g: Grid,
+        periodic: bool,
+        cell_fluid: Vec<bool>,
+        kinds: [Vec<FaceKind>; 3],
+        anchor: usize,
+        fluid_cells: usize,
+        cut: CutGeometry,
+    ) -> Self {
+        let [u_kind, v_kind, w_kind] = kinds;
+        Self {
             grid: g,
             periodic_z: periodic,
             cell_fluid,
@@ -234,9 +288,7 @@ impl Mask {
             grad_weights: None,
             diffusion_centroid: false,
             face_shifts: None,
-        };
-        mask.compute_merging(None);
-        Ok(mask)
+        }
     }
 
     /// The virtual merging map: a small cell (fraction < `MERGE_FRACTION`
@@ -306,6 +358,7 @@ impl Mask {
         let (Some(cut), Some(old_cut)) = (self.cut.as_ref(), old.cut.as_ref()) else {
             return;
         };
+        let lap = std::time::Instant::now();
         let n = inner.len() + 1;
         let w_end = 0.5 / n as f64;
         let w_in = 1.0 / n as f64;
@@ -336,12 +389,17 @@ impl Mask {
                     || (periodic && nz > 1 && k == 0 && t(idx + (nz - 1) * nxy))
             })
             .collect();
+        let l_changed = lap.elapsed();
         // P1-5 (j): with the trapezoid alone and a previous step's arrays, a
         // face of no changed cell keeps its step aperture (its corners were
         // untouched in both builds, so αⁿ⁻¹ = αⁿ = αⁿ⁺¹): the old mask's
         // arrays move over and only the changed cells' faces and the
         // changed cells' activity are recomputed, with the same expressions.
-        let prev = if inner.is_empty() { old.step_apertures.take().zip(old.step_open.take()) } else { None };
+        let prev = if inner.is_empty() {
+            old.step_apertures.take().zip(old.step_open.take())
+        } else {
+            None
+        };
         if let Some(((mut au, mut av, mut aw), (mut ou, mut ov, mut ow, mut active))) = prev {
             let mix = |a: &[f64], b: &[f64], f: usize| w_end * (a[f] + b[f]);
             for &idx in &changed {
@@ -387,7 +445,16 @@ impl Mask {
             self.step_open = Some((open(&au), open(&av), open(&aw), active));
             self.step_apertures = Some((au, av, aw));
         }
+        let l_apert = lap.elapsed();
         self.compute_merging(Some(old));
+        if profile_on() {
+            eprintln!(
+                "    step-aperture laps: changed set {:.0} ms, apertures {:.0} ms, merging {:.0} ms",
+                l_changed.as_secs_f64() * 1e3,
+                (l_apert - l_changed).as_secs_f64() * 1e3,
+                (lap.elapsed() - l_apert).as_secs_f64() * 1e3
+            );
+        }
         self.changed_cells = Some(changed);
     }
 
@@ -449,7 +516,11 @@ impl Mask {
         }
         // S2-7: the sides' own apertures instead of the whole-face averages.
         if self.cv_sides_exact {
-            if let Some(exact) = self.cut.as_ref().and_then(|cut| self.exact_cv_sides(cut, c, p)) {
+            if let Some(exact) = self
+                .cut
+                .as_ref()
+                .and_then(|cut| self.exact_cv_sides(cut, c, p))
+            {
                 ap = exact;
             }
         }
@@ -715,6 +786,7 @@ impl Mask {
         let (Some(cut), Some(old_cut)) = (self.cut.as_ref(), old.cut.as_ref()) else {
             return (table, 0.0);
         };
+        let lap = std::time::Instant::now();
         let g = self.grid;
         let dv = g.dx * g.dy * g.dz;
         let (mut net, mut area) = (0.0, 0.0);
@@ -759,6 +831,7 @@ impl Mask {
                 vn1 - vn
             );
         }
+        let l_sums = lap.elapsed();
         let correction = if area > 0.0 { net / area } else { 0.0 };
         if correction != 0.0 {
             // Every cell with a wall (the others subtract exactly 0).
@@ -770,6 +843,13 @@ impl Mask {
                 }
             }
         }
+        if profile_on() {
+            eprintln!(
+                "    gcl laps: lists + sums {:.0} ms, correction {:.0} ms",
+                l_sums.as_secs_f64() * 1e3,
+                (lap.elapsed() - l_sums).as_secs_f64() * 1e3
+            );
+        }
         (table, correction)
     }
 

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/maskupdate.rs

@@ -0,0 +1,374 @@
+//! R6-2 (`RTX_E3_MASK_DEVICE=1`): the host mirror of a moving mask whose
+//! classification ran on the device. The device computes, on the changed set
+//! of the step (the cells touched by either build, dilated by one) and the
+//! faces of those cells, exactly the values the host rebuild computes over
+//! the whole grid (`Mask::from_cut`, `compute_face_shifts`,
+//! `set_step_apertures`, `compute_merging`, `gcl_flux_table`) and hands them
+//! over compactly; the host mask is the previous one's arrays with those
+//! entries replaced. Outside the changed set every one of these arrays is
+//! unchanged over two steps (their inputs are the apertures and volumes of
+//! the touched cells' neighbourhoods), so the instantaneous arrays come from
+//! the mask retired one step earlier and the step arrays from the old mask.
+
+use super::cut::CutGeometry;
+use super::step::{Boundaries, Side};
+use super::wall::{FaceKind, Mask};
+use super::Grid;
+use rayon::prelude::*;
+
+/// `RTX_E3_MASK_DEVICE=1`, with the device geometry it requires (default
+/// ON since R6-1's gate; `RTX_E3_GEOM_DEVICE=0` turns both off).
+#[must_use]
+pub fn mask_device_enabled() -> bool {
+    std::env::var("RTX_E3_MASK_DEVICE").is_ok_and(|v| v == "1")
+        && !std::env::var("RTX_E3_GEOM_DEVICE").is_ok_and(|v| v == "0")
+}
+
+/// The device classification of one moving step, compact.
+#[derive(Debug, Default)]
+pub struct MaskUpdate {
+    /// The changed cells, ascending.
+    pub changed: Vec<usize>,
+    /// Per changed cell: bit 0 fluid at the new geometry, bit 1 active over the step.
+    pub cell_flags: Vec<u8>,
+    /// Per changed cell: the merging master (`u32::MAX` = its own row).
+    pub master: Vec<u32>,
+    /// The faces of the changed cells per component, ascending.
+    pub faces: [Vec<u32>; 3],
+    /// Per face: bit 0 kind fluid, bit 1 step-open.
+    pub face_flags: [Vec<u8>; 3],
+    /// Per face: the step-averaged aperture.
+    pub face_a: [Vec<f64>; 3],
+    /// Per face: the centroid shift (three interleaved; empty without the centroid diffusion).
+    pub face_shift: [Vec<f64>; 3],
+    /// The GCL table's compatibility correction (the table itself stays on the device).
+    pub correction: f64,
+    /// The merged cells.
+    pub merged: usize,
+}
+
+/// The instantaneous arrays of a retired mask, kept for the mask two steps later.
+#[derive(Default)]
+pub struct MaskArrays {
+    pub(super) generation: u64,
+    pub(super) cell_fluid: Vec<bool>,
+    pub(super) kinds: [Vec<FaceKind>; 3],
+    pub(super) face_shifts: Option<[Vec<f64>; 3]>,
+}
+
+/// A raw pointer that crosses threads for the disjoint writes of a scatter.
+struct Shared<T>(*mut T);
+impl<T> Clone for Shared<T> {
+    fn clone(&self) -> Self {
+        *self
+    }
+}
+impl<T> Copy for Shared<T> {}
+unsafe impl<T> Send for Shared<T> {}
+unsafe impl<T> Sync for Shared<T> {}
+impl<T> Shared<T> {
+    /// SAFETY: the caller writes each in-range index from one thread only.
+    unsafe fn write(self, i: usize, v: T) {
+        unsafe { *self.0.add(i) = v };
+    }
+}
+
+/// An index of a compact list.
+trait Ix: Copy + Sync {
+    fn ix(self) -> usize;
+}
+impl Ix for u32 {
+    fn ix(self) -> usize {
+        self as usize
+    }
+}
+impl Ix for usize {
+    fn ix(self) -> usize {
+        self
+    }
+}
+
+/// `dst[idx[t]] = f(t)` in parallel; `idx` holds distinct in-range indices.
+fn scatter<T: Send, I: Ix>(dst: &mut [T], idx: &[I], f: impl Fn(usize) -> T + Sync) {
+    let n = dst.len();
+    assert!(
+        idx.iter().all(|&i| i.ix() < n),
+        "scatter index out of range"
+    );
+    let p = Shared(dst.as_mut_ptr());
+    idx.par_iter()
+        .enumerate()
+        .with_min_len(4096)
+        .for_each(|(t, &i)| {
+            // SAFETY: the indices are distinct (ascending compaction output) and in range.
+            unsafe { p.write(i.ix(), f(t)) };
+        });
+}
+
+fn allowed(side: Side) -> bool {
+    matches!(side, Side::Velocity | Side::Periodic | Side::SlipWall)
+}
+
+impl Mask {
+    /// The mask at `cut` from the device classification `upd`, the old
+    /// mask (generation n − 1: its step arrays and merging map move over)
+    /// and `pool` (generation n − 2's instantaneous arrays; the old mask's
+    /// are cloned without it). Classification flags (scheme, closures) are
+    /// the caller's, as for [`Mask::from_cut`].
+    pub(crate) fn from_update(
+        cut: CutGeometry,
+        g: Grid,
+        b: Boundaries,
+        upd: &MaskUpdate,
+        old: &mut Mask,
+        pool: Option<MaskArrays>,
+    ) -> Result<Self, String> {
+        let (nx, ny, nz) = (g.nx, g.ny, g.nz);
+        let nxy = nx * ny;
+        let periodic = b.z0 == Side::Periodic;
+        let (mut cell_fluid, kinds, mut face_shifts) = match pool {
+            Some(p) => (p.cell_fluid, p.kinds, p.face_shifts),
+            None => (
+                old.cell_fluid.clone(),
+                [old.u_kind.clone(), old.v_kind.clone(), old.w_kind.clone()],
+                old.face_shifts.clone(),
+            ),
+        };
+        let [mut u_kind, mut v_kind, mut w_kind] = kinds;
+        let ch = &upd.changed;
+        scatter(&mut cell_fluid, ch, |t| upd.cell_flags[t] & 1 != 0);
+        // The touching check: only a changed cell can newly reach a side.
+        let touches = |idx: usize| {
+            let (k, j, i) = (idx / nxy, (idx % nxy) / nx, idx % nx);
+            (i == 0 && !allowed(b.x0))
+                || (i + 1 == nx && !allowed(b.x1))
+                || (j == 0 && !allowed(b.y0))
+                || (j + 1 == ny && !allowed(b.y1))
+                || (k == 0 && !allowed(b.z0))
+                || (k + 1 == nz && !allowed(b.z1))
+        };
+        if ch.iter().any(|&idx| !cell_fluid[idx] && touches(idx)) {
+            let first = (0..g.cells())
+                .into_par_iter()
+                .find_first(|&idx| !cell_fluid[idx] && touches(idx))
+                .expect("touching cell");
+            let (k, j, i) = (first / nxy, (first % nxy) / nx, first % nx);
+            return Err(format!(
+                "embedded body reaches a domain side that is not a Velocity/Periodic side at cell ({k}, {j}, {i})"
+            ));
+        }
+        // The fluid count and the anchor (the smallest fluid index).
+        let delta: i64 = ch
+            .iter()
+            .map(|&idx| i64::from(cell_fluid[idx]) - i64::from(old.cell_fluid[idx]))
+            .sum();
+        let fluid_cells = (old.fluid_cells as i64 + delta) as usize;
+        let anchor = if cell_fluid[old.anchor] {
+            match ch.iter().find(|&&idx| cell_fluid[idx]) {
+                Some(&idx) if idx < old.anchor => Some(idx),
+                _ => Some(old.anchor),
+            }
+        } else {
+            (0..g.cells())
+                .into_par_iter()
+                .find_first(|&idx| cell_fluid[idx])
+        };
+        let Some(anchor) = anchor else {
+            return Err("embedded body covers the whole domain".into());
+        };
+        // The faces: kinds, the step arrays (moved from the old mask), the shifts.
+        let (Some((mut au, mut av, mut aw)), Some((mut ou, mut ov, mut ow, mut active))) =
+            (old.step_apertures.take(), old.step_open.take())
+        else {
+            return Err("R6-2: the old mask has no step arrays".into());
+        };
+        let kind_of = |bits: u8| {
+            if bits & 1 != 0 {
+                FaceKind::Fluid
+            } else {
+                FaceKind::Solid
+            }
+        };
+        {
+            let kinds: [&mut Vec<FaceKind>; 3] = [&mut u_kind, &mut v_kind, &mut w_kind];
+            let steps: [&mut Vec<f64>; 3] = [&mut au, &mut av, &mut aw];
+            let opens: [&mut Vec<bool>; 3] = [&mut ou, &mut ov, &mut ow];
+            for (c, ((kd, st), op)) in kinds.into_iter().zip(steps).zip(opens).enumerate() {
+                let (list, flags, a) = (&upd.faces[c], &upd.face_flags[c], &upd.face_a[c]);
+                scatter(kd, list.as_slice(), |t| kind_of(flags[t]));
+                scatter(st, list.as_slice(), |t| a[t]);
+                scatter(op, list.as_slice(), |t| flags[t] & 2 != 0);
+            }
+        }
+        scatter(&mut active, ch, |t| upd.cell_flags[t] & 2 != 0);
+        if let Some(tables) = face_shifts.as_mut() {
+            for c in 0..3 {
+                let (list, sh) = (&upd.faces[c], &upd.face_shift[c]);
+                assert_eq!(sh.len(), 3 * list.len(), "R6-2: face shifts missing");
+                let tab = &mut tables[c];
+                let n = tab.len();
+                assert!(
+                    list.iter().all(|&f| 3 * f as usize + 2 < n),
+                    "shift index out of range"
+                );
+                let p = Shared(tab.as_mut_ptr());
+                list.par_iter()
+                    .enumerate()
+                    .with_min_len(4096)
+                    .for_each(|(t, &f)| {
+                        let f = f as usize;
+                        // SAFETY: distinct faces, three entries each, in range.
+                        unsafe {
+                            p.write(3 * f, sh[3 * t]);
+                            p.write(3 * f + 1, sh[3 * t + 1]);
+                            p.write(3 * f + 2, sh[3 * t + 2]);
+                        }
+                    });
+            }
+        }
+        let mut merge_master = std::mem::take(&mut old.merge_master);
+        if merge_master.len() != g.cells() {
+            return Err("R6-2: the old mask has no merging map".into());
+        }
+        scatter(&mut merge_master, ch, |t| match upd.master[t] {
+            u32::MAX => usize::MAX,
+            m => m as usize,
+        });
+        let mut mask = Self::from_parts(
+            g,
+            periodic,
+            cell_fluid,
+            [u_kind, v_kind, w_kind],
+            anchor,
+            fluid_cells,
+            cut,
+        );
+        mask.step_apertures = Some((au, av, aw));
+        mask.step_open = Some((ou, ov, ow, active));
+        mask.merge_master = merge_master;
+        mask.face_shifts = face_shifts;
+        mask.changed_cells = Some(upd.changed.clone());
+        Ok(mask)
+    }
+
+    /// The instantaneous arrays of this (retiring) mask for the pool.
+    pub(crate) fn into_pool(&mut self, generation: u64, cell_fluid: Vec<bool>) -> MaskArrays {
+        MaskArrays {
+            generation,
+            cell_fluid,
+            kinds: [
+                std::mem::take(&mut self.u_kind),
+                std::mem::take(&mut self.v_kind),
+                std::mem::take(&mut self.w_kind),
+            ],
+            face_shifts: self.face_shifts.take(),
+        }
+    }
+
+    /// `RTX_E3_BAND_CHECK=1`: this mask's classification arrays against the
+    /// host rebuild's `reference`, bit for bit; the differences named.
+    pub(crate) fn compare_classification(&self, reference: &Mask) -> Vec<String> {
+        let mut report = Vec::new();
+        let mut eq = |name: &str, ok: bool| {
+            if !ok {
+                report.push(format!("{name} differs"));
+            }
+        };
+        let bits = |a: &[f64], b: &[f64]| {
+            a.len() == b.len() && a.iter().zip(b).all(|(x, y)| x.to_bits() == y.to_bits())
+        };
+        eq("cell_fluid", self.cell_fluid == reference.cell_fluid);
+        eq("u_kind", self.u_kind == reference.u_kind);
+        eq("v_kind", self.v_kind == reference.v_kind);
+        eq("w_kind", self.w_kind == reference.w_kind);
+        eq("anchor", self.anchor == reference.anchor);
+        eq("fluid_cells", self.fluid_cells == reference.fluid_cells);
+        eq("merge_master", self.merge_master == reference.merge_master);
+        eq(
+            "changed_cells",
+            self.changed_cells == reference.changed_cells,
+        );
+        match (&self.step_apertures, &reference.step_apertures) {
+            (Some(a), Some(r)) => {
+                eq("step a_u", bits(&a.0, &r.0));
+                eq("step a_v", bits(&a.1, &r.1));
+                eq("step a_w", bits(&a.2, &r.2));
+            }
+            (a, r) => eq("step apertures presence", a.is_some() == r.is_some()),
+        }
+        match (&self.step_open, &reference.step_open) {
+            (Some(a), Some(r)) => {
+                eq("step open u", a.0 == r.0);
+                eq("step open v", a.1 == r.1);
+                eq("step open w", a.2 == r.2);
+                eq("step active", a.3 == r.3);
+            }
+            (a, r) => eq("step open presence", a.is_some() == r.is_some()),
+        }
+        match (&self.face_shifts, &reference.face_shifts) {
+            (Some(a), Some(r)) => {
+                for c in 0..3 {
+                    eq(&format!("face shifts[{c}]"), bits(&a[c], &r[c]));
+                }
+            }
+            (a, r) => eq("face shifts presence", a.is_some() == r.is_some()),
+        }
+        report
+    }
+}
+
+/// Refill the pressure of the cells fluid in `new` and not in `old` from
+/// their face neighbours fluid in both; returns their count — the host
+/// `refill_fresh_cells` over the changed cells (a fresh cell is touched),
+/// in the same ascending order.
+pub(crate) fn refill_fresh_in(old: &Mask, new: &Mask, p: &mut [f64], changed: &[usize]) -> usize {
+    let g = new.grid;
+    let (nx, ny, nz) = (g.nx, g.ny, g.nz);
+    let periodic = new.periodic_z;
+    let mut fresh = 0;
+    let mut refills = Vec::new();
+    for &idx in changed {
+        if !(new.cell_fluid[idx] && !old.cell_fluid[idx]) {
+            continue;
+        }
+        fresh += 1;
+        let (k, j, i) = g.kji(idx);
+        let mut sum = 0.0;
+        let mut count = 0usize;
+        let mut visit = |nb: usize| {
+            if new.cell_fluid[nb] && old.cell_fluid[nb] {
+                sum += p[nb];
+                count += 1;
+            }
+        };
+        if i + 1 < nx {
+            visit(g.cell(k, j, i + 1));
+        }
+        if i > 0 {
+            visit(g.cell(k, j, i - 1));
+        }
+        if j + 1 < ny {
+            visit(g.cell(k, j + 1, i));
+        }
+        if j > 0 {
+            visit(g.cell(k, j - 1, i));
+        }
+        if k + 1 < nz {
+            visit(g.cell(k + 1, j, i));
+        } else if periodic && nz > 1 {
+            visit(g.cell(0, j, i));
+        }
+        if k > 0 {
+            visit(g.cell(k - 1, j, i));
+        } else if periodic && nz > 1 {
+            visit(g.cell(nz - 1, j, i));
+        }
+        if count > 0 {
+            refills.push((idx, sum / count as f64));
+        }
+    }
+    for (idx, v) in refills {
+        p[idx] = v;
+    }
+    fresh
+}

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/mod.rs

@@ -16,6 +16,7 @@ pub mod field;
 pub mod grid;
 pub mod impose;
 pub mod loads;
+pub mod maskupdate;
 pub mod poisson;
 pub mod reconstruct;
 pub mod step;

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/device.rs

@@ -7,12 +7,13 @@
 
 mod cut;
 mod geom;
+mod mask;
 
 use super::{Side, Solver, StepResult};
-use crate::solvers::incompressible::embedded3::Grid;
 use crate::solvers::incompressible::embedded3::field::Field;
 use crate::solvers::incompressible::embedded3::poisson::device::{cfg, load_module, runtime};
 use crate::solvers::incompressible::embedded3::poisson::device_cg::DeviceCg;
+use crate::solvers::incompressible::embedded3::Grid;
 use crate::solvers::incompressible::poisson::MultigridParameters;
 use crate::solvers::incompressible::simple::ConvectionScheme;
 use cudarc::driver::{
@@ -165,6 +166,8 @@ pub struct DeviceStep {
     steps_since_hierarchy: usize,
     /// R6-1: the persistent device cut geometry (default ON; `RTX_E3_GEOM_DEVICE=0` off).
     geom: Option<geom::DeviceGeom>,
+    /// R6-2: the device classification (`RTX_E3_MASK_DEVICE=1`).
+    dmask: Option<mask::DeviceMask>,
 }
 
 impl DeviceStep {
@@ -244,6 +247,7 @@ impl DeviceStep {
             cut,
             steps_since_hierarchy: 0,
             geom: None,
+            dmask: None,
         }
     }
 

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/device/cut.rs

@@ -5,13 +5,13 @@
 //! kernels (`e3_cut.cu`, appended to `e3_step.cu` at load).
 
 use super::{DeviceStep, E3Params, E3Ptrs, StepResult};
-use crate::solvers::incompressible::embedded3::Grid;
 use crate::solvers::incompressible::embedded3::field::Field;
-use crate::solvers::incompressible::embedded3::poisson::GuessBasis;
 use crate::solvers::incompressible::embedded3::poisson::device::{cfg, load_module, runtime};
 use crate::solvers::incompressible::embedded3::poisson::device_cg::DeviceCg;
+use crate::solvers::incompressible::embedded3::poisson::GuessBasis;
 use crate::solvers::incompressible::embedded3::step::Solver;
 use crate::solvers::incompressible::embedded3::wall::FaceKind;
+use crate::solvers::incompressible::embedded3::Grid;
 use crate::solvers::incompressible::poisson::MultigridParameters;
 use cudarc::driver::{
     CudaFunction, CudaModule, CudaSlice, DevicePtr, DeviceRepr, PushKernelArg, ValidAsZeroBits,
@@ -64,27 +64,27 @@ unsafe impl ValidAsZeroBits for E3CutPtrs {}
 /// The static cut-cell mask on the device.
 pub(super) struct DeviceCut {
     /// The surface-velocity tables (kept to share between the phases).
-    ub_host: [Vec<f64>; 3],
+    pub(super) ub_host: [Vec<f64>; 3],
     /// Which phase's apertures / open flags / activity `ptrs` hands out
     /// (P1-5: both sets live in the struct and persist across steps).
-    phase: Phase,
+    pub(super) phase: Phase,
     /// The predictor phase's instantaneous apertures, open flags (fluid
     /// kind) and activity (fluid cells).
-    a_pred: [CudaSlice<f64>; 3],
-    open_pred: [CudaSlice<i32>; 3],
-    active_pred: CudaSlice<i32>,
-    a: [CudaSlice<f64>; 3],
-    d: [CudaSlice<f64>; 3],
-    ub: [CudaSlice<f64>; 3],
-    wall_flux: CudaSlice<f64>,
-    open: [CudaSlice<i32>; 3],
-    active: CudaSlice<i32>,
-    owner: CudaSlice<u32>,
-    fold_ptr: CudaSlice<u32>,
-    fold_idx: CudaSlice<u32>,
+    pub(super) a_pred: [CudaSlice<f64>; 3],
+    pub(super) open_pred: [CudaSlice<i32>; 3],
+    pub(super) active_pred: CudaSlice<i32>,
+    pub(super) a: [CudaSlice<f64>; 3],
+    pub(super) d: [CudaSlice<f64>; 3],
+    pub(super) ub: [CudaSlice<f64>; 3],
+    pub(super) wall_flux: CudaSlice<f64>,
+    pub(super) open: [CudaSlice<i32>; 3],
+    pub(super) active: CudaSlice<i32>,
+    pub(super) owner: CudaSlice<u32>,
+    pub(super) fold_ptr: CudaSlice<u32>,
+    pub(super) fold_idx: CudaSlice<u32>,
     cell_flux: CudaSlice<f64>,
     /// The open-part centroid shifts per face (S2-5; one dummy entry when off).
-    shift: [CudaSlice<f64>; 3],
+    pub(super) shift: [CudaSlice<f64>; 3],
     /// The wall's normal velocity into the fluid per face (A3-i; one dummy entry when off).
     vn: [CudaSlice<f64>; 3],
     /// The wall velocity at the two axis feet along each direction per face,
@@ -234,7 +234,8 @@ impl DeviceCut {
                                         let delta = mask.exchange_delta(&cv, d);
                                         let mut xf = x;
                                         xf[d] += sign * delta;
-                                        out[2 * d + side] = mask.surface_velocity_at(body, xf, c, t);
+                                        out[2 * d + side] =
+                                            mask.surface_velocity_at(body, xf, c, t);
                                     }
                                 }
                             }
@@ -333,11 +334,18 @@ impl DeviceCut {
                 })
                 .collect()
         };
-        let active_pred: Vec<i32> = (0..nc).into_par_iter().map(|i| i32::from(mask.is_fluid_cell(i))).collect();
+        let active_pred: Vec<i32> = (0..nc)
+            .into_par_iter()
+            .map(|i| i32::from(mask.is_fluid_cell(i)))
+            .collect();
         let (a_pred, open_pred_dev, active_pred_dev) = if projection {
             (
                 [up_f(&cut.a_u), up_f(&cut.a_v), up_f(&cut.a_w)],
-                [up_i(&open_pred(0)), up_i(&open_pred(1)), up_i(&open_pred(2))],
+                [
+                    up_i(&open_pred(0)),
+                    up_i(&open_pred(1)),
+                    up_i(&open_pred(2)),
+                ],
                 up_i(&active_pred),
             )
         } else {
@@ -410,12 +418,20 @@ impl DeviceCut {
     /// global). `false` when a full build is needed (no changed set, the
     /// advancing-wall closure, a size change).
     pub(super) fn update(&mut self, solver: &Solver, g: Grid, t: f64) -> bool {
-        use crate::solvers::incompressible::embedded3::poisson::device_cg::{scatter_f64, scatter_i32, scatter_u32};
+        use crate::solvers::incompressible::embedded3::poisson::device_cg::{
+            scatter_f64, scatter_i32, scatter_u32,
+        };
         use rayon::prelude::*;
-        let Some(mask) = solver.mask() else { return false };
+        let Some(mask) = solver.mask() else {
+            return false;
+        };
         let Some(cut) = mask.cut() else { return false };
-        let Some(body) = solver.body() else { return false };
-        let Some(changed) = mask.changed_cells() else { return false };
+        let Some(body) = solver.body() else {
+            return false;
+        };
+        let Some(changed) = mask.changed_cells() else {
+            return false;
+        };
         if mask.wall_advancing || mask.wall_exchange_foot {
             return false;
         }
@@ -429,7 +445,11 @@ impl DeviceCut {
         let lap = Instant::now();
         let profile = std::env::var("RTX_E3_MOVING_PROFILE").is_ok();
         // The faces of the changed cells, per component, ascending.
-        let mut marks: [Vec<bool>; 3] = [vec![false; counts[0]], vec![false; counts[1]], vec![false; counts[2]]];
+        let mut marks: [Vec<bool>; 3] = [
+            vec![false; counts[0]],
+            vec![false; counts[1]],
+            vec![false; counts[2]],
+        ];
         for &idx in changed {
             let (k, j, i) = (idx / nxy, (idx % nxy) / nx, idx % nx);
             marks[0][g.uface(k, j, i)] = true;
@@ -440,19 +460,30 @@ impl DeviceCut {
             marks[2][g.wface(k + 1, j, i)] = true;
         }
         let touched: [Vec<u32>; 3] = [0, 1, 2].map(|c| {
-            (0..counts[c]).into_par_iter().filter(|&f| marks[c][f]).map(|f| f as u32).collect()
+            (0..counts[c])
+                .into_par_iter()
+                .filter(|&f| marks[c][f])
+                .map(|f| f as u32)
+                .collect()
         });
         let cells_u32: Vec<u32> = changed.iter().map(|&i| i as u32).collect();
+        let l_touched = lap.elapsed();
         // The imposition band's faces (their surface velocity moves with t).
         let band = mask.impose_band().unwrap_or(f64::INFINITY);
         let dists: [&[f64]; 3] = [&cut.d_u, &cut.d_v, &cut.d_w];
         let band_faces: [Vec<u32>; 3] = [0, 1, 2].map(|c| {
-            (0..counts[c]).into_par_iter().filter(|&f| dists[c][f].abs() <= band).map(|f| f as u32).collect()
+            (0..counts[c])
+                .into_par_iter()
+                .filter(|&f| dists[c][f].abs() <= band)
+                .map(|f| f as u32)
+                .collect()
         });
         let h = [g.dx, g.dy, g.dz];
         let shifts = mask.face_shift_tables();
         let step = mask.step_apertures();
         let step_open = mask.step_open_flags();
+        let l_band = lap.elapsed();
+        let mut l_ub = std::time::Duration::ZERO;
         for c in 0..3 {
             let (ni, nj) = (nx + usize::from(c == 0), ny + usize::from(c == 1));
             let pos = |idx: usize| -> [f64; 3] {
@@ -463,7 +494,11 @@ impl DeviceCut {
                     (k as f64 + if c == 2 { 0.0 } else { 0.5 }) * h[2],
                 ]
             };
-            let a_inst: &[f64] = match c { 0 => &cut.a_u, 1 => &cut.a_v, _ => &cut.a_w };
+            let a_inst: &[f64] = match c {
+                0 => &cut.a_u,
+                1 => &cut.a_v,
+                _ => &cut.a_w,
+            };
             let a_step: &[f64] = match (step, c) {
                 (Some(a), 0) => &a.0,
                 (Some(a), 1) => &a.1,
@@ -477,7 +512,12 @@ impl DeviceCut {
                 scatter_f64(tf, &pick(a_inst), &mut self.a_pred[c]);
                 scatter_f64(tf, &pick(dists[c]), &mut self.d[c]);
             }
-            let open_step: Vec<i32> = tf.iter().map(|&f| i32::from(step_open.map_or(a_inst[f as usize] > 0.0, |o| o[c][f as usize]))).collect();
+            let open_step: Vec<i32> = tf
+                .iter()
+                .map(|&f| {
+                    i32::from(step_open.map_or(a_inst[f as usize] > 0.0, |o| o[c][f as usize]))
+                })
+                .collect();
             scatter_i32(tf, &open_step, &mut self.open[c]);
             let open_inst: Vec<i32> = tf
                 .iter()
@@ -492,18 +532,35 @@ impl DeviceCut {
                 .collect();
             scatter_i32(tf, &open_inst, &mut self.open_pred[c]);
             if let Some(sh) = shifts {
-                let idx3: Vec<u32> = tf.iter().flat_map(|&f| [3 * f, 3 * f + 1, 3 * f + 2]).collect();
-                let val3: Vec<f64> = tf.iter().flat_map(|&f| [sh[c][3 * f as usize], sh[c][3 * f as usize + 1], sh[c][3 * f as usize + 2]]).collect();
+                let idx3: Vec<u32> = tf
+                    .iter()
+                    .flat_map(|&f| [3 * f, 3 * f + 1, 3 * f + 2])
+                    .collect();
+                let val3: Vec<f64> = tf
+                    .iter()
+                    .flat_map(|&f| {
+                        [
+                            sh[c][3 * f as usize],
+                            sh[c][3 * f as usize + 1],
+                            sh[c][3 * f as usize + 2],
+                        ]
+                    })
+                    .collect();
                 scatter_f64(&idx3, &val3, &mut self.shift[c]);
             }
             // The surface velocity at the foot for the band's faces.
+            let lu = Instant::now();
             let bf = &band_faces[c];
             let ubv: Vec<f64> = bf
                 .par_iter()
                 .map(|&f| {
                     let x = pos(f as usize);
                     let xf = match (shifts, mask.wall_foot_centroid) {
-                        (Some(sh), true) => [x[0] + sh[c][3 * f as usize], x[1] + sh[c][3 * f as usize + 1], x[2] + sh[c][3 * f as usize + 2]],
+                        (Some(sh), true) => [
+                            x[0] + sh[c][3 * f as usize],
+                            x[1] + sh[c][3 * f as usize + 1],
+                            x[2] + sh[c][3 * f as usize + 2],
+                        ],
                         _ => x,
                     };
                     mask.surface_velocity_at(body, xf, c, t)
@@ -513,15 +570,26 @@ impl DeviceCut {
                 self.ub_host[c][f as usize] = v;
             }
             scatter_f64(bf, &ubv, &mut self.ub[c]);
+            l_ub += lu.elapsed();
         }
         let l_faces = lap.elapsed();
         // Per-cell tables for the changed cells.
-        let active_step: Vec<i32> = changed.iter().map(|&i| i32::from(mask.cell_active(i))).collect();
-        let active_inst: Vec<i32> = changed.iter().map(|&i| i32::from(mask.is_fluid_cell(i))).collect();
-        let owner: Vec<u32> = changed.iter().map(|&i| mask.master(i).unwrap_or(i) as u32).collect();
+        let active_step: Vec<i32> = changed
+            .iter()
+            .map(|&i| i32::from(mask.cell_active(i)))
+            .collect();
+        let active_inst: Vec<i32> = changed
+            .iter()
+            .map(|&i| i32::from(mask.is_fluid_cell(i)))
+            .collect();
+        let owner: Vec<u32> = changed
+            .iter()
+            .map(|&i| mask.master(i).unwrap_or(i) as u32)
+            .collect();
         scatter_i32(&cells_u32, &active_step, &mut self.active);
         scatter_i32(&cells_u32, &active_inst, &mut self.active_pred);
         scatter_u32(&cells_u32, &owner, &mut self.owner);
+        let l_owner = lap.elapsed();
         // The wall flux: the changed cells and every wall cell (the correction).
         let wall_flux: &[f64] = solver.wall_fluxes();
         if wall_flux.len() != nc {
@@ -538,6 +606,7 @@ impl DeviceCut {
             .collect();
         let wvals: Vec<f64> = wcells.iter().map(|&i| wall_flux[i as usize]).collect();
         scatter_f64(&wcells, &wvals, &mut self.wall_flux);
+        let l_wflux = lap.elapsed();
         // The merged-cell CSR whole (its lengths change).
         let mut fold_ptr = vec![0u32; nc + 1];
         let mut merged = 0;
@@ -584,6 +653,16 @@ impl DeviceCut {
                 ms(lap.elapsed() - l_faces),
                 changed.len()
             );
+            eprintln!(
+                "    update laps: touched lists {:.0} ms, band scan {:.0} ms, face scatters {:.0} ms, ub {:.0} ms, cell scatters {:.0} ms, wall flux {:.0} ms, fold CSR {:.0} ms",
+                ms(l_touched),
+                ms(l_band - l_touched),
+                ms(l_faces - l_band - l_ub),
+                ms(l_ub),
+                ms(l_owner - l_faces),
+                ms(l_wflux - l_owner),
+                ms(lap.elapsed() - l_wflux)
+            );
         }
         true
     }
@@ -600,15 +679,27 @@ impl DeviceCut {
     pub(super) fn check_against(&self, full: &Self) {
         let rt = runtime();
         let same_f = |name: &str, a: &CudaSlice<f64>, b: &CudaSlice<f64>| {
-            let (x, y) = (rt.stream.memcpy_dtov(a).expect("dl"), rt.stream.memcpy_dtov(b).expect("dl"));
-            assert!(x.len() == y.len() && x.iter().zip(&y).all(|(p, q)| p.to_bits() == q.to_bits()), "band tables: {name} differs");
+            let (x, y) = (
+                rt.stream.memcpy_dtov(a).expect("dl"),
+                rt.stream.memcpy_dtov(b).expect("dl"),
+            );
+            assert!(
+                x.len() == y.len() && x.iter().zip(&y).all(|(p, q)| p.to_bits() == q.to_bits()),
+                "band tables: {name} differs"
+            );
         };
         let same_i = |name: &str, a: &CudaSlice<i32>, b: &CudaSlice<i32>| {
-            let (x, y) = (rt.stream.memcpy_dtov(a).expect("dl"), rt.stream.memcpy_dtov(b).expect("dl"));
+            let (x, y) = (
+                rt.stream.memcpy_dtov(a).expect("dl"),
+                rt.stream.memcpy_dtov(b).expect("dl"),
+            );
             assert!(x == y, "band tables: {name} differs");
         };
         let same_u = |name: &str, a: &CudaSlice<u32>, b: &CudaSlice<u32>| {
-            let (x, y) = (rt.stream.memcpy_dtov(a).expect("dl"), rt.stream.memcpy_dtov(b).expect("dl"));
+            let (x, y) = (
+                rt.stream.memcpy_dtov(a).expect("dl"),
+                rt.stream.memcpy_dtov(b).expect("dl"),
+            );
             assert!(x == y, "band tables: {name} differs");
         };
         for c in 0..3 {
@@ -619,8 +710,15 @@ impl DeviceCut {
             same_i("open_pred", &self.open_pred[c], &full.open_pred[c]);
             same_f("shift", &self.shift[c], &full.shift[c]);
             // ub: the band's faces only (beyond it the full build writes 0, ours keeps stale — never read).
-            let (x, y) = (rt.stream.memcpy_dtov(&self.ub[c]).expect("dl"), rt.stream.memcpy_dtov(&full.ub[c]).expect("dl"));
-            let bad = x.iter().zip(&y).filter(|(p, q)| **q != 0.0 && p.to_bits() != q.to_bits()).count();
+            let (x, y) = (
+                rt.stream.memcpy_dtov(&self.ub[c]).expect("dl"),
+                rt.stream.memcpy_dtov(&full.ub[c]).expect("dl"),
+            );
+            let bad = x
+                .iter()
+                .zip(&y)
+                .filter(|(p, q)| **q != 0.0 && p.to_bits() != q.to_bits())
+                .count();
             assert!(bad == 0, "band tables: ub[{c}] differs on {bad} band faces");
         }
         same_f("wall_flux", &self.wall_flux, &full.wall_flux);
@@ -759,17 +857,52 @@ impl DeviceStep {
                     }
                     None => true,
                 };
+            let mut device_mask: Option<usize> = None;
             if geom_on {
                 if let Some(dc) = self.cut.as_mut() {
+                    let fresh_geom = self.geom.is_none();
                     let geom = self
                         .geom
                         .get_or_insert_with(|| super::geom::DeviceGeom::new(g));
+                    // R6-2: the previous step's apertures, volumes and
+                    // evaluated cells, before the geometry kernels overwrite them.
+                    let dm = if super::mask::enabled() && !fresh_geom {
+                        let dm = self
+                            .dmask
+                            .get_or_insert_with(|| super::mask::DeviceMask::new(g));
+                        dm.snapshot(dc, geom);
+                        Some(dm)
+                    } else {
+                        None
+                    };
                     let mut pool = std::mem::take(&mut self.solver.geom_pool);
                     let built = geom.build(&self.solver, &mut pool, g, t_new, dt, dc);
                     self.solver.geom_pool = pool;
                     if let Some(mirror) = built {
                         self.solver.pending_cut = Some(mirror);
                         dc.geom_written = true;
+                        if let Some(dm) = dm {
+                            if dm.ready(&self.solver, geom) {
+                                let lm = Instant::now();
+                                let upd = dm.run(&self.solver, g, dt, dc, geom);
+                                device_mask = Some(upd.merged);
+                                if profile {
+                                    let l = dm.laps;
+                                    eprintln!(
+                                        "    moving laps: device mask {:.0} ms (lists {:.0}, cells + merging {:.0}, faces {:.0}, GCL {:.0}, fold {:.0}; {} changed, {} merged)",
+                                        lm.elapsed().as_secs_f64() * 1e3,
+                                        l[0],
+                                        l[1],
+                                        l[2],
+                                        l[3],
+                                        l[4],
+                                        upd.changed.len(),
+                                        upd.merged
+                                    );
+                                }
+                                self.solver.pending_mask = Some(upd);
+                            }
+                        }
                     }
                 }
                 if profile {
@@ -791,10 +924,25 @@ impl DeviceStep {
             let rebuild_all = std::env::var("RTX_E3_TABLES_REBUILD").is_ok();
             let check = std::env::var("RTX_E3_BAND_CHECK").is_ok();
             self.cut = match self.cut.take() {
+                // R6-2: the device classification wrote the tables; the
+                // surface velocities remain.
+                Some(mut prev) if device_mask.is_some() => {
+                    let dm = self.dmask.as_mut().expect("device mask");
+                    let merged = device_mask.expect("merged");
+                    assert!(prev.update_after_device_mask(&self.solver, g, t_new, dm, merged));
+                    if check {
+                        let full = DeviceCut::build(&self.solver, g, Phase::Projection, t_new)
+                            .expect("full");
+                        prev.check_against(&full);
+                        eprintln!("    table check t {t_new:.6}: device tables IDENTICAL to the full build");
+                    }
+                    Some(prev)
+                }
                 Some(mut prev) if !rebuild_all => {
                     if prev.update(&self.solver, g, t_new) {
                         if check {
-                            let full = DeviceCut::build(&self.solver, g, Phase::Projection, t_new).expect("full");
+                            let full = DeviceCut::build(&self.solver, g, Phase::Projection, t_new)
+                                .expect("full");
                             prev.check_against(&full);
                         }
                         Some(prev)
@@ -839,7 +987,10 @@ impl DeviceStep {
                     smoother: self.solver.params.poisson_smoother,
                     ..MultigridParameters::default()
                 };
-                let changed: Option<Vec<usize>> = self.solver.mask().and_then(|m| m.changed_cells().map(|c| c.to_vec()));
+                let changed: Option<Vec<usize>> = self
+                    .solver
+                    .mask()
+                    .and_then(|m| m.changed_cells().map(|c| c.to_vec()));
                 cg.refresh_with(problem, &params, changed.as_deref());
             }
             rt.stream

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/device/geom.rs

@@ -12,11 +12,11 @@
 //! down.
 
 use super::cut::DeviceCut;
-use crate::solvers::incompressible::embedded3::Grid;
 use crate::solvers::incompressible::embedded3::cut::{CutGeometry, GeomPool};
 use crate::solvers::incompressible::embedded3::poisson::device::{cfg, load_module, runtime};
 use crate::solvers::incompressible::embedded3::step::Solver;
 use crate::solvers::incompressible::embedded3::wall::WallScheme;
+use crate::solvers::incompressible::embedded3::Grid;
 use cudarc::driver::{
     CudaFunction, CudaModule, CudaSlice, DeviceRepr, PushKernelArg, ValidAsZeroBits,
 };
@@ -135,6 +135,8 @@ pub(super) struct DeviceGeom {
     generation: u64,
     chain_start: u64,
     history: std::collections::VecDeque<(u64, [Vec<u32>; 3], Vec<u32>)>,
+    /// R6-2: the last build was a full pass (the device state was re-synced).
+    pub(super) last_full: bool,
 }
 
 /// Generations of band lists kept (a recycled array older than this is copied instead).
@@ -160,9 +162,16 @@ impl DeviceGeom {
             generation: 0,
             chain_start: 1,
             history: std::collections::VecDeque::new(),
+            last_full: true,
         }
     }
 
+    /// R6-2: the cell volumes, wall vectors (3 interleaved) and evaluated
+    /// cells of the last build.
+    pub(super) fn cell_tables(&self) -> (&CudaSlice<f64>, &CudaSlice<f64>, &CudaSlice<u8>) {
+        (&self.vol, &self.wall, &self.touched_cell)
+    }
+
     /// The cut geometry of the solver's body at `t` (the moving rebuild's
     /// end-of-step geometry, the band of the current mask moved by the
     /// step `dt`) on the device; the face tables written into `dc`'s
@@ -241,6 +250,7 @@ impl DeviceGeom {
             None => (0i32, 0.0, 0.0),
         };
         let full_i = i32::from(full);
+        self.last_full = full;
         unsafe {
             rt.stream
                 .launch_builder(&k.phi)

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/device/mask.rs

@@ -0,0 +1,699 @@
+//! R6-2: the moving mask's classification on the device (`e3_mask.cu`),
+//! behind `RTX_E3_MASK_DEVICE=1` (which needs `RTX_E3_GEOM_DEVICE=1`). After
+//! the device geometry of the step (R6-1), the changed set, the faces of the
+//! changed cells, the kinds, the step apertures and open flags, the space-time
+//! activity, the merging masters, the centroid shifts, the GCL wall-flux
+//! table and the merged cells' CSR are computed on the device and written
+//! straight into `DeviceCut`'s persistent tables; the host receives the same
+//! values compactly (`MaskUpdate`) for its mirror. The previous step's
+//! apertures, volumes and evaluated cells are snapshotted before the geometry
+//! kernels overwrite them.
+
+use super::cut::{DeviceCut, Phase};
+use super::geom::DeviceGeom;
+use crate::solvers::incompressible::embedded3::cutwall::MERGE_FRACTION;
+use crate::solvers::incompressible::embedded3::maskupdate::MaskUpdate;
+use crate::solvers::incompressible::embedded3::poisson::device::{cfg, load_module, runtime};
+use crate::solvers::incompressible::embedded3::step::Solver;
+use crate::solvers::incompressible::embedded3::wall::WallScheme;
+use crate::solvers::incompressible::embedded3::Grid;
+use cudarc::driver::{
+    CudaFunction, CudaModule, CudaSlice, DevicePtr, DeviceRepr, LaunchConfig, PushKernelArg,
+    ValidAsZeroBits,
+};
+use std::sync::{Arc, OnceLock};
+use std::time::Instant;
+
+const MASK_KERNELS: &str = include_str!("../../../../../kernels/cuda/e3_mask.cu");
+const SCAN_BLOCK: usize = 1024;
+
+struct MaskKernels {
+    _module: Arc<CudaModule>,
+    count_flags: CudaFunction,
+    count_vals: CudaFunction,
+    top: CudaFunction,
+    compact: CudaFunction,
+    values: CudaFunction,
+    changed: CudaFunction,
+    face_flags: CudaFunction,
+    cells: CudaFunction,
+    merge: CudaFunction,
+    faces: CudaFunction,
+    wall_flags: CudaFunction,
+    wall_area: CudaFunction,
+    wall_flux: CudaFunction,
+    slave_flags: CudaFunction,
+    fold_count: CudaFunction,
+    fold_fill: CudaFunction,
+    band_flags: CudaFunction,
+}
+
+static MASK_ONCE: OnceLock<MaskKernels> = OnceLock::new();
+
+fn kernels() -> &'static MaskKernels {
+    MASK_ONCE.get_or_init(|| {
+        // FMA contraction off: the host's roundings.
+        let module = load_module(MASK_KERNELS, "e3_mask.cu", true);
+        let f = |name: &str| module.load_function(name).expect(name);
+        MaskKernels {
+            count_flags: f("e3_scan_count_flags"),
+            count_vals: f("e3_scan_count_vals"),
+            top: f("e3_scan_top"),
+            compact: f("e3_scan_compact"),
+            values: f("e3_scan_values"),
+            changed: f("e3_mask_changed"),
+            face_flags: f("e3_mask_face_flags"),
+            cells: f("e3_mask_cells"),
+            merge: f("e3_mask_merge"),
+            faces: f("e3_mask_faces"),
+            wall_flags: f("e3_mask_wall_flags"),
+            wall_area: f("e3_mask_wall_area"),
+            wall_flux: f("e3_mask_wall_flux"),
+            slave_flags: f("e3_mask_slave_flags"),
+            fold_count: f("e3_mask_fold_count"),
+            fold_fill: f("e3_mask_fold_fill"),
+            band_flags: f("e3_mask_band_flags"),
+            _module: module,
+        }
+    })
+}
+
+/// `struct MaskGrid` in e3_mask.cu.
+#[repr(C)]
+#[derive(Clone, Copy)]
+struct MaskGrid {
+    nx: i32,
+    ny: i32,
+    nz: i32,
+    periodic: i32,
+    dx: f64,
+    dy: f64,
+    dz: f64,
+}
+unsafe impl DeviceRepr for MaskGrid {}
+unsafe impl ValidAsZeroBits for MaskGrid {}
+
+/// `struct Apertures` in e3_mask.cu: the three instantaneous aperture tables.
+#[repr(C)]
+#[derive(Clone, Copy)]
+struct Apertures {
+    a: [u64; 3],
+}
+unsafe impl DeviceRepr for Apertures {}
+unsafe impl ValidAsZeroBits for Apertures {}
+
+/// `RTX_E3_MASK_DEVICE=1` with `RTX_E3_GEOM_DEVICE=1`.
+pub(super) fn enabled() -> bool {
+    crate::solvers::incompressible::embedded3::maskupdate::mask_device_enabled()
+}
+
+fn scan_cfg(n: usize) -> LaunchConfig {
+    LaunchConfig {
+        grid_dim: (n.div_ceil(SCAN_BLOCK).max(1) as u32, 1, 1),
+        block_dim: (SCAN_BLOCK as u32, 1, 1),
+        shared_mem_bytes: 0,
+    }
+}
+
+/// A device buffer grown on demand (never shrunk).
+fn ensure<T: DeviceRepr + ValidAsZeroBits>(buf: &mut Option<CudaSlice<T>>, n: usize) {
+    let n = n.max(1);
+    if buf.as_ref().is_none_or(|b| b.len() < n) {
+        let cap = n + n / 4;
+        *buf = Some(runtime().stream.alloc_zeros::<T>(cap).expect("alloc"));
+    }
+}
+
+/// Download the first `n` entries.
+fn head<T: DeviceRepr>(buf: &CudaSlice<T>, n: usize) -> Vec<T> {
+    if n == 0 {
+        return Vec::new();
+    }
+    runtime()
+        .stream
+        .memcpy_dtov(&buf.slice(0..n))
+        .expect("download")
+}
+
+/// The persistent state of the device classification.
+pub(super) struct DeviceMask {
+    /// The previous step's instantaneous apertures, volumes and evaluated cells.
+    a_prev: [CudaSlice<f64>; 3],
+    vol_prev: CudaSlice<f64>,
+    touched_prev: CudaSlice<u8>,
+    /// The snapshot was taken this step (before the geometry kernels).
+    snapped: bool,
+    changed_flag: CudaSlice<u8>,
+    flags: CudaSlice<u8>,
+    sums: CudaSlice<u32>,
+    counts: CudaSlice<u32>,
+    changed_list: Option<CudaSlice<u32>>,
+    face_list: [Option<CudaSlice<u32>>; 3],
+    list_tmp: Option<CudaSlice<u32>>,
+    out_cell_flags: Option<CudaSlice<u8>>,
+    out_entry: Option<CudaSlice<f64>>,
+    out_master: Option<CudaSlice<u32>>,
+    out_face_flags: [Option<CudaSlice<u8>>; 3],
+    out_face_a: [Option<CudaSlice<f64>>; 3],
+    out_face_shift: [Option<CudaSlice<f64>>; 3],
+    out_wall_a: Option<CudaSlice<f64>>,
+    /// The laps of the last run (ms): lists, cells + merging, faces, GCL, fold.
+    pub(super) laps: [f64; 6],
+}
+
+impl DeviceMask {
+    pub(super) fn new(g: Grid) -> Self {
+        let rt = runtime();
+        let nc = g.cells();
+        let counts = [g.n_ufaces(), g.n_vfaces(), g.n_wfaces()];
+        let nmax = nc.max(counts[0]).max(counts[1]).max(counts[2]) + 1;
+        let z8 = |n: usize| rt.stream.alloc_zeros::<u8>(n).expect("alloc");
+        Self {
+            a_prev: counts.map(|n| rt.stream.alloc_zeros::<f64>(n).expect("alloc")),
+            vol_prev: rt.stream.alloc_zeros::<f64>(nc).expect("alloc"),
+            touched_prev: z8(nc),
+            snapped: false,
+            changed_flag: z8(nc),
+            flags: z8(nmax),
+            sums: rt
+                .stream
+                .alloc_zeros::<u32>(nmax.div_ceil(SCAN_BLOCK) + 1)
+                .expect("alloc"),
+            counts: rt.stream.alloc_zeros::<u32>(nc).expect("alloc"),
+            changed_list: None,
+            face_list: [None, None, None],
+            list_tmp: None,
+            out_cell_flags: None,
+            out_entry: None,
+            out_master: None,
+            out_face_flags: [None, None, None],
+            out_face_a: [None, None, None],
+            out_face_shift: [None, None, None],
+            out_wall_a: None,
+            laps: [0.0; 6],
+        }
+    }
+
+    /// Before the geometry kernels: the current instantaneous apertures,
+    /// volumes and evaluated cells become the previous step's.
+    pub(super) fn snapshot(&mut self, dc: &DeviceCut, geom: &DeviceGeom) {
+        let rt = runtime();
+        let (vol, _, touched) = geom.cell_tables();
+        for c in 0..3 {
+            rt.stream
+                .memcpy_dtod(&dc.a_pred[c], &mut self.a_prev[c])
+                .expect("a_prev");
+        }
+        rt.stream
+            .memcpy_dtod(vol, &mut self.vol_prev)
+            .expect("vol_prev");
+        rt.stream
+            .memcpy_dtod(touched, &mut self.touched_prev)
+            .expect("touched_prev");
+        self.snapped = true;
+    }
+
+    /// Whether the classification can run on the device this step: the
+    /// snapshot is of the geometry the host holds as the previous one (no
+    /// re-sync), the old mask is a moving one (step arrays, merging map),
+    /// and no closure the device does not carry is on.
+    pub(super) fn ready(&self, solver: &Solver, geom: &DeviceGeom) -> bool {
+        let p = &solver.params;
+        let Some(m) = solver.mask() else { return false };
+        self.snapped
+            && !geom.last_full
+            && p.wall_scheme == WallScheme::CutCell
+            && p.aperture_substeps == 0
+            && !p.wall_advancing
+            && !p.wall_exchange_foot
+            && !p.pressure_centroid
+            && std::env::var("RTX_E3_TABLES_REBUILD").is_err()
+            && m.step_apertures().is_some()
+            && m.step_open_flags().is_some()
+            && m.changed_cells().is_some()
+            && m.face_shift_tables().is_some() == p.diffusion_centroid
+    }
+
+    /// Compact the first `n` flags of `self.flags` into `out` (ascending); the count.
+    fn compact(&mut self, n: usize, out: &mut Option<CudaSlice<u32>>) -> usize {
+        let rt = runtime();
+        let k = kernels();
+        let nb = n.div_ceil(SCAN_BLOCK).max(1);
+        let n64 = n as i64;
+        let nb32 = nb as i32;
+        unsafe {
+            rt.stream
+                .launch_builder(&k.count_flags)
+                .arg(&n64)
+                .arg(&self.flags)
+                .arg(&mut self.sums)
+                .launch(scan_cfg(n))
+                .expect("e3_scan_count_flags");
+            rt.stream
+                .launch_builder(&k.top)
+                .arg(&nb32)
+                .arg(&mut self.sums)
+                .launch(LaunchConfig {
+                    grid_dim: (1, 1, 1),
+                    block_dim: (SCAN_BLOCK as u32, 1, 1),
+                    shared_mem_bytes: 0,
+                })
+                .expect("e3_scan_top");
+        }
+        let total = rt
+            .stream
+            .memcpy_dtov(&self.sums.slice(nb..nb + 1))
+            .expect("count")[0] as usize;
+        ensure(out, total);
+        let dst = out.as_mut().expect("list");
+        unsafe {
+            rt.stream
+                .launch_builder(&k.compact)
+                .arg(&n64)
+                .arg(&self.flags)
+                .arg(&self.sums)
+                .arg(dst)
+                .launch(scan_cfg(n))
+                .expect("e3_scan_compact");
+        }
+        total
+    }
+
+    /// The device classification of the step (the geometry just built, the
+    /// snapshot the previous one): `dc`'s projection tables written, the
+    /// compact values returned for the host mirror.
+    pub(super) fn run(
+        &mut self,
+        solver: &Solver,
+        g: Grid,
+        dt: f64,
+        dc: &mut DeviceCut,
+        geom: &DeviceGeom,
+    ) -> MaskUpdate {
+        let rt = runtime();
+        let k = kernels();
+        let lap = Instant::now();
+        self.snapped = false;
+        let nc = g.cells();
+        let counts = [g.n_ufaces(), g.n_vfaces(), g.n_wfaces()];
+        let mg = MaskGrid {
+            nx: g.nx as i32,
+            ny: g.ny as i32,
+            nz: g.nz as i32,
+            periodic: i32::from(solver.params.boundaries.periodic_z()),
+            dx: g.dx,
+            dy: g.dy,
+            dz: g.dz,
+        };
+        let (vol, wall, touched) = geom.cell_tables();
+        // The changed set.
+        unsafe {
+            rt.stream
+                .launch_builder(&k.changed)
+                .arg(&mg)
+                .arg(touched)
+                .arg(&self.touched_prev)
+                .arg(&mut self.changed_flag)
+                .launch(cfg(nc))
+                .expect("e3_mask_changed");
+        }
+        rt.stream
+            .memcpy_dtod(&self.changed_flag, &mut self.flags.slice_mut(0..nc))
+            .expect("flags");
+        let mut changed_list = self.changed_list.take();
+        let n_ch = self.compact(nc, &mut changed_list);
+        self.changed_list = changed_list;
+        // The faces of the changed cells.
+        let mut n_faces = [0usize; 3];
+        for c in 0..3 {
+            let ci = c as i32;
+            unsafe {
+                rt.stream
+                    .launch_builder(&k.face_flags)
+                    .arg(&mg)
+                    .arg(&ci)
+                    .arg(&self.changed_flag)
+                    .arg(&mut self.flags)
+                    .launch(cfg(counts[c]))
+                    .expect("e3_mask_face_flags");
+            }
+            let mut fl = self.face_list[c].take();
+            n_faces[c] = self.compact(counts[c], &mut fl);
+            self.face_list[c] = fl;
+        }
+        let ch_list = self.changed_list.as_ref().expect("changed");
+        let changed: Vec<usize> = head(ch_list, n_ch)
+            .into_iter()
+            .map(|i| i as usize)
+            .collect();
+        let faces: [Vec<u32>; 3] =
+            [0, 1, 2].map(|c| head(self.face_list[c].as_ref().expect("faces"), n_faces[c]));
+        let l_lists = lap.elapsed();
+        // Per changed cell: classification, activity, GCL entry; then the masters.
+        ensure(&mut self.out_cell_flags, n_ch);
+        ensure(&mut self.out_entry, n_ch);
+        ensure(&mut self.out_master, n_ch);
+        let dv = g.dx * g.dy * g.dz;
+        let n_ch32 = n_ch as i32;
+        let threshold = std::env::var("RTX_E3_MERGE_FRACTION")
+            .ok()
+            .and_then(|v| v.parse().ok())
+            .unwrap_or(MERGE_FRACTION);
+        if n_ch > 0 {
+            unsafe {
+                rt.stream
+                    .launch_builder(&k.cells)
+                    .arg(&n_ch32)
+                    .arg(ch_list)
+                    .arg(vol)
+                    .arg(&self.vol_prev)
+                    .arg(&dv)
+                    .arg(&dt)
+                    .arg(&mut dc.active)
+                    .arg(&mut dc.active_pred)
+                    .arg(self.out_cell_flags.as_mut().expect("out"))
+                    .arg(self.out_entry.as_mut().expect("out"))
+                    .launch(cfg(n_ch))
+                    .expect("e3_mask_cells");
+                rt.stream
+                    .launch_builder(&k.merge)
+                    .arg(&mg)
+                    .arg(&n_ch32)
+                    .arg(ch_list)
+                    .arg(vol)
+                    .arg(&self.vol_prev)
+                    .arg(&dc.active)
+                    .arg(&threshold)
+                    .arg(&mut dc.owner)
+                    .arg(self.out_master.as_mut().expect("out"))
+                    .launch(cfg(n_ch))
+                    .expect("e3_mask_merge");
+            }
+        }
+        let cell_flags = head(self.out_cell_flags.as_ref().expect("out"), n_ch);
+        let entries = head(self.out_entry.as_ref().expect("out"), n_ch);
+        let master = head(self.out_master.as_ref().expect("out"), n_ch);
+        let l_cells = lap.elapsed();
+        // Per face of a changed cell: kind, step aperture and open flag, shift.
+        let s = &rt.stream;
+        let ap = Apertures {
+            a: [0, 1, 2].map(|c| dc.a_pred[c].device_ptr(s).0),
+        };
+        let has_shift =
+            i32::from(solver.params.diffusion_centroid && dc.shift[0].len() == 3 * counts[0]);
+        let mut face_flags: [Vec<u8>; 3] = Default::default();
+        let mut face_a: [Vec<f64>; 3] = Default::default();
+        let mut face_shift: [Vec<f64>; 3] = Default::default();
+        for c in 0..3 {
+            let n = n_faces[c];
+            ensure(&mut self.out_face_flags[c], n);
+            ensure(&mut self.out_face_a[c], n);
+            ensure(&mut self.out_face_shift[c], 3 * n);
+            if n > 0 {
+                let (ci, n32) = (c as i32, n as i32);
+                unsafe {
+                    rt.stream
+                        .launch_builder(&k.faces)
+                        .arg(&mg)
+                        .arg(&ci)
+                        .arg(&n32)
+                        .arg(self.face_list[c].as_ref().expect("faces"))
+                        .arg(&ap)
+                        .arg(&self.a_prev[c])
+                        .arg(&mut dc.a[c])
+                        .arg(&mut dc.open[c])
+                        .arg(&mut dc.open_pred[c])
+                        .arg(&has_shift)
+                        .arg(&mut dc.shift[c])
+                        .arg(self.out_face_flags[c].as_mut().expect("out"))
+                        .arg(self.out_face_a[c].as_mut().expect("out"))
+                        .arg(self.out_face_shift[c].as_mut().expect("out"))
+                        .launch(cfg(n))
+                        .expect("e3_mask_faces");
+                }
+            }
+            face_flags[c] = head(self.out_face_flags[c].as_ref().expect("out"), n);
+            face_a[c] = head(self.out_face_a[c].as_ref().expect("out"), n);
+            if has_shift != 0 {
+                face_shift[c] = head(self.out_face_shift[c].as_ref().expect("out"), 3 * n);
+            }
+        }
+        let l_faces = lap.elapsed();
+        // The GCL table: the entries' and the wall areas' sums in ascending
+        // order on the host (the serial loops' order), the table on the device.
+        let nc64 = nc as i64;
+        unsafe {
+            rt.stream
+                .launch_builder(&k.wall_flags)
+                .arg(&nc64)
+                .arg(&dc.active)
+                .arg(wall)
+                .arg(&mut self.flags)
+                .launch(cfg(nc))
+                .expect("e3_mask_wall_flags");
+        }
+        let mut tmp = self.list_tmp.take();
+        let n_wall = self.compact(nc, &mut tmp);
+        self.list_tmp = tmp;
+        ensure(&mut self.out_wall_a, n_wall);
+        if n_wall > 0 {
+            let nw32 = n_wall as i32;
+            unsafe {
+                rt.stream
+                    .launch_builder(&k.wall_area)
+                    .arg(&nw32)
+                    .arg(self.list_tmp.as_ref().expect("list"))
+                    .arg(wall)
+                    .arg(self.out_wall_a.as_mut().expect("out"))
+                    .launch(cfg(n_wall))
+                    .expect("e3_mask_wall_area");
+            }
+        }
+        let areas = head(self.out_wall_a.as_ref().expect("out"), n_wall);
+        let mut net = 0.0;
+        for (t, &e) in entries.iter().enumerate() {
+            if cell_flags[t] & 2 != 0 {
+                net += e;
+            }
+        }
+        let mut area = 0.0;
+        for &a in &areas {
+            area += a;
+        }
+        let correction = if area > 0.0 { net / area } else { 0.0 };
+        unsafe {
+            rt.stream
+                .launch_builder(&k.wall_flux)
+                .arg(&nc64)
+                .arg(&self.changed_flag)
+                .arg(&dc.active)
+                .arg(vol)
+                .arg(&self.vol_prev)
+                .arg(wall)
+                .arg(&dv)
+                .arg(&dt)
+                .arg(&correction)
+                .arg(&mut dc.wall_flux)
+                .launch(cfg(nc))
+                .expect("e3_mask_wall_flux");
+        }
+        let l_gcl = lap.elapsed();
+        // The merged cells' CSR.
+        unsafe {
+            rt.stream
+                .launch_builder(&k.slave_flags)
+                .arg(&nc64)
+                .arg(&dc.owner)
+                .arg(&mut self.flags)
+                .launch(cfg(nc))
+                .expect("e3_mask_slave_flags");
+        }
+        let mut tmp = self.list_tmp.take();
+        let merged = self.compact(nc, &mut tmp);
+        self.list_tmp = tmp;
+        let slaves = self.list_tmp.as_ref().expect("slaves");
+        rt.stream.memset_zeros(&mut self.counts).expect("counts");
+        if dc.fold_ptr.len() != nc + 1 {
+            dc.fold_ptr = rt.stream.alloc_zeros::<u32>(nc + 1).expect("alloc");
+        }
+        if dc.fold_idx.len() != merged.max(1) {
+            dc.fold_idx = rt.stream.alloc_zeros::<u32>(merged.max(1)).expect("alloc");
+        }
+        let m32 = merged as i32;
+        let nb = nc.div_ceil(SCAN_BLOCK).max(1);
+        let nb32 = nb as i32;
+        unsafe {
+            if merged > 0 {
+                rt.stream
+                    .launch_builder(&k.fold_count)
+                    .arg(&m32)
+                    .arg(slaves)
+                    .arg(&dc.owner)
+                    .arg(&mut self.counts)
+                    .launch(cfg(merged))
+                    .expect("e3_mask_fold_count");
+            }
+            rt.stream
+                .launch_builder(&k.count_vals)
+                .arg(&nc64)
+                .arg(&self.counts)
+                .arg(&mut self.sums)
+                .launch(scan_cfg(nc))
+                .expect("e3_scan_count_vals");
+            rt.stream
+                .launch_builder(&k.top)
+                .arg(&nb32)
+                .arg(&mut self.sums)
+                .launch(LaunchConfig {
+                    grid_dim: (1, 1, 1),
+                    block_dim: (SCAN_BLOCK as u32, 1, 1),
+                    shared_mem_bytes: 0,
+                })
+                .expect("e3_scan_top");
+            rt.stream
+                .launch_builder(&k.values)
+                .arg(&nc64)
+                .arg(&self.counts)
+                .arg(&self.sums)
+                .arg(&mut dc.fold_ptr)
+                .launch(scan_cfg(nc))
+                .expect("e3_scan_values");
+        }
+        rt.stream
+            .memcpy_htod(&[merged as u32][..], &mut dc.fold_ptr.slice_mut(nc..nc + 1))
+            .expect("fold_ptr end");
+        if merged > 0 {
+            unsafe {
+                rt.stream
+                    .launch_builder(&k.fold_fill)
+                    .arg(&mg)
+                    .arg(&m32)
+                    .arg(slaves)
+                    .arg(&dc.owner)
+                    .arg(&dc.fold_ptr)
+                    .arg(&mut dc.fold_idx)
+                    .launch(cfg(merged))
+                    .expect("e3_mask_fold_fill");
+            }
+        }
+        rt.stream.synchronize().expect("sync");
+        let l_fold = lap.elapsed();
+        let ms = |d: std::time::Duration| d.as_secs_f64() * 1e3;
+        self.laps = [
+            ms(l_lists),
+            ms(l_cells - l_lists),
+            ms(l_faces - l_cells),
+            ms(l_gcl - l_faces),
+            ms(l_fold - l_gcl),
+            ms(l_fold),
+        ];
+        MaskUpdate {
+            changed,
+            cell_flags,
+            master,
+            faces,
+            face_flags,
+            face_a,
+            face_shift,
+            correction,
+            merged,
+        }
+    }
+
+    /// The imposition band's faces per component (|d| <= band), ascending.
+    pub(super) fn band_faces(&mut self, dc: &DeviceCut, g: Grid, band: f64) -> [Vec<u32>; 3] {
+        let rt = runtime();
+        let k = kernels();
+        let counts = [g.n_ufaces(), g.n_vfaces(), g.n_wfaces()];
+        [0usize, 1, 2].map(|c| {
+            let n64 = counts[c] as i64;
+            unsafe {
+                rt.stream
+                    .launch_builder(&k.band_flags)
+                    .arg(&n64)
+                    .arg(&dc.d[c])
+                    .arg(&band)
+                    .arg(&mut self.flags)
+                    .launch(cfg(counts[c]))
+                    .expect("e3_mask_band_flags");
+            }
+            let mut tmp = self.list_tmp.take();
+            let n = self.compact(counts[c], &mut tmp);
+            self.list_tmp = tmp;
+            head(self.list_tmp.as_ref().expect("band"), n)
+        })
+    }
+}
+
+impl DeviceCut {
+    /// R6-2: finish the projection tables after the device classification
+    /// (`DeviceMask::run` wrote every table but the surface velocities): the
+    /// surface velocity at the foot for the imposition band's faces (listed
+    /// on the device), the merged count, the phase.
+    pub(super) fn update_after_device_mask(
+        &mut self,
+        solver: &Solver,
+        g: Grid,
+        t: f64,
+        dm: &mut DeviceMask,
+        merged: usize,
+    ) -> bool {
+        use crate::solvers::incompressible::embedded3::poisson::device_cg::scatter_f64;
+        use rayon::prelude::*;
+        let (Some(mask), Some(body)) = (solver.mask(), solver.body()) else {
+            return false;
+        };
+        let lap = Instant::now();
+        let band = mask.impose_band().unwrap_or(f64::INFINITY);
+        let band_faces = dm.band_faces(self, g, band);
+        let l_band = lap.elapsed();
+        let h = [g.dx, g.dy, g.dz];
+        let shifts = mask.face_shift_tables();
+        let (nx, ny) = (g.nx, g.ny);
+        for (c, bf) in band_faces.iter().enumerate() {
+            let (ni, nj) = (nx + usize::from(c == 0), ny + usize::from(c == 1));
+            let pos = |idx: usize| -> [f64; 3] {
+                let (k, j, i) = (idx / (nj * ni), (idx / ni) % nj, idx % ni);
+                [
+                    (i as f64 + if c == 0 { 0.0 } else { 0.5 }) * h[0],
+                    (j as f64 + if c == 1 { 0.0 } else { 0.5 }) * h[1],
+                    (k as f64 + if c == 2 { 0.0 } else { 0.5 }) * h[2],
+                ]
+            };
+            let ubv: Vec<f64> = bf
+                .par_iter()
+                .map(|&f| {
+                    let x = pos(f as usize);
+                    let xf = match (shifts, mask.wall_foot_centroid) {
+                        (Some(sh), true) => [
+                            x[0] + sh[c][3 * f as usize],
+                            x[1] + sh[c][3 * f as usize + 1],
+                            x[2] + sh[c][3 * f as usize + 2],
+                        ],
+                        _ => x,
+                    };
+                    mask.surface_velocity_at(body, xf, c, t)
+                })
+                .collect();
+            for (&f, &v) in bf.iter().zip(&ubv) {
+                self.ub_host[c][f as usize] = v;
+            }
+            scatter_f64(bf, &ubv, &mut self.ub[c]);
+        }
+        self.merged = merged;
+        self.phase = Phase::Projection;
+        self.geom_written = false;
+        runtime().stream.synchronize().expect("sync");
+        if std::env::var("RTX_E3_MOVING_PROFILE").is_ok() {
+            eprintln!(
+                "    table laps (device mask): band list {:.0} ms, ub {:.0} ms ({} / {} / {} band faces)",
+                l_band.as_secs_f64() * 1e3,
+                (lap.elapsed() - l_band).as_secs_f64() * 1e3,
+                band_faces[0].len(),
+                band_faces[1].len(),
+                band_faces[2].len()
+            );
+        }
+        true
+    }
+}

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/mod.rs

@@ -11,12 +11,12 @@ mod moving;
 mod predictor;
 mod projection;
 
-use super::Grid;
 use super::body::Body;
 use super::cut::CutGeometry;
 use super::field::Field;
-use super::poisson::{PcgCache, Problem, solve_pcg_cached};
+use super::poisson::{solve_pcg_cached, PcgCache, Problem};
 use super::wall::{FaceKind, Mask, WallScheme};
+use super::Grid;
 use crate::solvers::incompressible::poisson::{
     MgPrecision, MgSmoother, MultigridParameters, PoissonSolution,
 };
@@ -304,6 +304,14 @@ pub struct Solver {
     /// and the generation of `apertures_old`.
     pub(super) geom_pool: super::cut::GeomPool,
     pub(super) apertures_old_gen: u64,
+    /// R6-2: the next moving rebuild's classification, computed on the
+    /// device (`RTX_E3_MASK_DEVICE=1`); `rebuild_moving_mask` assembles the
+    /// mask from it instead of classifying the whole grid.
+    pub(super) pending_mask: Option<super::maskupdate::MaskUpdate>,
+    /// R6-2: the instantaneous arrays of the mask retired at the previous
+    /// rebuild, and the rebuild counter (the mask generation).
+    pub(super) mask_pool: Option<super::maskupdate::MaskArrays>,
+    pub(super) mask_gen: u64,
 }
 
 impl Solver {
@@ -340,6 +348,9 @@ impl Solver {
             pending_cut: None,
             geom_pool: super::cut::GeomPool::default(),
             apertures_old_gen: 0,
+            pending_mask: None,
+            mask_pool: None,
+            mask_gen: 0,
         }
     }
 
@@ -403,23 +414,7 @@ impl Solver {
         }
         .map(|mut m| {
             let lap_closures = std::time::Instant::now();
-            m.scheme = self.params.convection_scheme;
-            m.density = self.fluid.density;
-            m.wall_order = self.params.wall_order;
-            m.wall_distance_oblique = self.params.wall_distance_oblique;
-            m.wall_exchange_axis = self.params.wall_exchange_axis;
-            m.diffusion_transverse =
-                self.params.diffusion_transverse && self.params.diffusion_centroid;
-            m.distance_floor_fine = self.params.distance_floor_fine;
-            m.wall_advancing = self.params.wall_advancing;
-            m.exchange_convection_off = self.params.exchange_convection_off;
-            m.cv_sides_exact = self.params.cv_sides_exact;
-            m.wall_order2_centroid = self.params.wall_order2_centroid;
-            m.wall_exchange_foot = self.params.wall_exchange_foot;
-            m.conv_sides_exact = self.params.conv_sides_exact;
-            m.wall_flux_true_normal = self.params.wall_flux_true_normal;
-            m.wall_foot_centroid = self.params.wall_foot_centroid;
-            m.diffusion_centroid = self.params.diffusion_centroid;
+            self.configure_mask(&mut m);
             if self.params.diffusion_centroid {
                 m.compute_face_shifts();
             }
@@ -437,6 +432,26 @@ impl Solver {
         .expect("embedded mask")
     }
 
+    /// The solver's closure flags on a freshly classified mask.
+    pub(super) fn configure_mask(&self, m: &mut Mask) {
+        m.scheme = self.params.convection_scheme;
+        m.density = self.fluid.density;
+        m.wall_order = self.params.wall_order;
+        m.wall_distance_oblique = self.params.wall_distance_oblique;
+        m.wall_exchange_axis = self.params.wall_exchange_axis;
+        m.diffusion_transverse = self.params.diffusion_transverse && self.params.diffusion_centroid;
+        m.distance_floor_fine = self.params.distance_floor_fine;
+        m.wall_advancing = self.params.wall_advancing;
+        m.exchange_convection_off = self.params.exchange_convection_off;
+        m.cv_sides_exact = self.params.cv_sides_exact;
+        m.wall_order2_centroid = self.params.wall_order2_centroid;
+        m.wall_exchange_foot = self.params.wall_exchange_foot;
+        m.conv_sides_exact = self.params.conv_sides_exact;
+        m.wall_flux_true_normal = self.params.wall_flux_true_normal;
+        m.wall_foot_centroid = self.params.wall_foot_centroid;
+        m.diffusion_centroid = self.params.diffusion_centroid;
+    }
+
     /// R5 instrument: record the cut predictor's momentum terms per face
     /// from the next step on (host path only; a no-op on the device path).
     pub fn enable_term_probe(&self) {

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/moving.rs

@@ -14,6 +14,9 @@ impl Solver {
     /// step-averaged apertures and the GCL wall-flux table (cut wall).
     /// Returns the fresh-cell count. `field` holds the predicted field.
     pub fn rebuild_moving_mask(&mut self, field: &mut Field, dt: f64, t_new: f64) -> usize {
+        if let Some(upd) = self.pending_mask.take() {
+            return self.rebuild_from_update(field, dt, t_new, upd);
+        }
         let pending = self.pending_cut.take();
         let Some(body) = &self.body else {
             return 0;
@@ -22,8 +25,11 @@ impl Solver {
         let lap = std::time::Instant::now();
         let mut new_mask = self.build_mask(body, field.grid, t_new, dt, pending);
         let l_build = lap.elapsed();
+        let mut l_refill = std::time::Duration::ZERO;
         if let Some(old_mask) = self.mask.as_mut() {
+            let lr = std::time::Instant::now();
             fresh_cells = refill_fresh_cells(old_mask, &new_mask, field);
+            l_refill = lr.elapsed();
             let n_in = self.params.aperture_substeps;
             if n_in == 0 {
                 new_mask.set_step_apertures(old_mask);
@@ -73,64 +79,205 @@ impl Solver {
             self.last_ghost_correction = correction;
         }
         let s_gcl = sub.elapsed();
-        // P1-5 (a): the old mask is replaced below — its arrays MOVE into
-        // the previous-step records instead of being cloned (1 GB per step
-        // at ny 124); the values are the same.
-        if let Some(mut old) = self.mask.take() {
-            let fluid = std::mem::take(&mut old.cell_fluid);
-            if let Some(c) = old.cut.as_mut() {
-                let retired = self.apertures_old.replace([
-                    std::mem::take(&mut c.a_u),
-                    std::mem::take(&mut c.a_v),
-                    std::mem::take(&mut c.a_w),
-                ]);
-                // R6-1: a device-built generation's arrays go back to the
-                // pool (the host path's generation is 0: dropped as before).
-                if let Some(r) = retired {
-                    for (kind, v) in super::super::cut::GeomPool::A.into_iter().zip(r) {
-                        self.geom_pool.put(self.apertures_old_gen, kind, v);
-                    }
-                }
-                self.apertures_old_gen = c.generation;
-                let mut vol = std::mem::take(&mut c.vol);
-                use rayon::prelude::*;
-                vol.par_iter_mut()
-                    .zip(fluid.par_iter())
-                    .for_each(|(v, &f)| {
-                        if !f {
-                            *v = 0.0;
-                        }
-                    });
-                self.vol_old = vol;
-                if let Some(c) = old.cut.take() {
-                    self.geom_pool.put_cut(c);
-                }
-            } else {
-                self.apertures_old = None;
-                self.vol_old = vec![0.0; field.grid.cells()];
-            }
+        if let Some(old) = self.mask.take() {
+            self.retire_mask(old, field.grid);
         }
+        self.mask_gen += 1;
         if std::env::var("RTX_E3_MOVING_PROFILE").is_ok() {
             let ms = |d: std::time::Duration| d.as_secs_f64() * 1e3;
             eprintln!(
-                "    mask laps: build_mask {:.0} ms, refill + step apertures + merging {:.0} ms, impose + GCL + volumes {:.0} ms (impose {:.0}, GCL table {:.0}, old apertures + volumes {:.0})",
+                "    mask laps: build_mask {:.0} ms, refill + step apertures + merging {:.0} ms, impose + GCL + volumes {:.0} ms (impose {:.0}, GCL table {:.0}, old apertures + volumes {:.0}); refill {:.0} ms",
                 ms(l_build),
                 ms(l_step - l_build),
                 ms(lap.elapsed() - l_step),
                 ms(s_impose),
                 ms(s_gcl - s_impose),
-                ms(sub.elapsed() - s_gcl)
+                ms(sub.elapsed() - s_gcl),
+                ms(l_refill)
             );
         }
         self.mask = Some(new_mask);
         fresh_cells
     }
+
+    /// R6-2: the moving rebuild from the device classification `upd` (the
+    /// geometry's mirror in `pending_cut`): the mask assembled from the old
+    /// one's arrays and the compact device values, the fresh cells refilled
+    /// over the changed set, the imposition as on the host path; the GCL
+    /// table stays on the device (its correction comes back). With
+    /// `RTX_E3_BAND_CHECK=1` the host path runs on copies and every array is
+    /// compared bit for bit.
+    fn rebuild_from_update(
+        &mut self,
+        field: &mut Field,
+        dt: f64,
+        t_new: f64,
+        upd: super::super::maskupdate::MaskUpdate,
+    ) -> usize {
+        use super::super::maskupdate::refill_fresh_in;
+        let g = field.grid;
+        let b = self.params.boundaries;
+        let cut = self
+            .pending_cut
+            .take()
+            .expect("R6-2: the device geometry's mirror");
+        let mut old_mask = self.mask.take().expect("R6-2: the previous mask");
+        let lap = std::time::Instant::now();
+        let check = std::env::var("RTX_E3_BAND_CHECK").is_ok();
+        // The reference: the host rebuild on copies.
+        let reference = check.then(|| {
+            let mut r = Mask::from_cut(cut.clone(), g, b).expect("reference mask");
+            self.configure_mask(&mut r);
+            if self.params.diffusion_centroid {
+                r.compute_face_shifts();
+            }
+            let mut p_ref = field.p.clone();
+            let fresh_ref = refill_fresh_cells_p(&old_mask, &r, g, &mut p_ref);
+            let mut old_copy = old_mask.clone();
+            r.set_step_apertures(&mut old_copy);
+            let (table, correction) = r.gcl_flux_table(&old_mask, dt);
+            (r, p_ref, fresh_ref, table, correction)
+        });
+        let l_ref = lap.elapsed();
+        let generation = self.mask_gen + 1;
+        let pool = self
+            .mask_pool
+            .take()
+            .filter(|p| p.generation + 2 == generation);
+        let pooled = pool.is_some();
+        let mut new_mask =
+            Mask::from_update(cut, g, b, &upd, &mut old_mask, pool).expect("embedded mask");
+        self.configure_mask(&mut new_mask);
+        let l_build = lap.elapsed();
+        let fresh_cells = refill_fresh_in(&old_mask, &new_mask, &mut field.p, &upd.changed);
+        let l_step = lap.elapsed();
+        let sub = std::time::Instant::now();
+        let body = self.body.as_ref().expect("body");
+        new_mask.impose_from(
+            body,
+            &field.u_old,
+            &field.v_old,
+            &field.w_old,
+            &mut field.u,
+            &mut field.v,
+            &mut field.w,
+            t_new,
+        );
+        let s_impose = sub.elapsed();
+        self.last_ghost_correction = upd.correction;
+        if let Some((r, p_ref, fresh_ref, table, correction)) = reference {
+            let mut report = new_mask.compare_classification(&r);
+            if correction.to_bits() != upd.correction.to_bits() {
+                report.push(format!(
+                    "GCL correction differs ({correction:e} vs {:e})",
+                    upd.correction
+                ));
+            }
+            if fresh_ref != fresh_cells
+                || p_ref
+                    .iter()
+                    .zip(&field.p)
+                    .any(|(a, b)| a.to_bits() != b.to_bits())
+            {
+                report.push("fresh-cell refill differs".into());
+            }
+            // The device tables' full reference build reads the host table.
+            self.wall_fluxes = table;
+            if report.is_empty() {
+                eprintln!(
+                    "    mask check t {t_new:.6}: device classification IDENTICAL to the host rebuild ({} changed cells, {} / {} / {} faces, pool {})",
+                    upd.changed.len(),
+                    upd.faces[0].len(),
+                    upd.faces[1].len(),
+                    upd.faces[2].len(),
+                    pooled
+                );
+            } else {
+                eprintln!(
+                    "    mask check t {t_new:.6}: DIFFERS — {}",
+                    report.join("; ")
+                );
+                if std::env::var("RTX_E3_GEOM_CHECK_SOFT").is_err() {
+                    panic!("mask check: device classification differs from the host rebuild");
+                }
+            }
+        }
+        let s_gcl = sub.elapsed();
+        self.retire_mask(old_mask, g);
+        self.mask_gen = generation;
+        if std::env::var("RTX_E3_MOVING_PROFILE").is_ok() {
+            let ms = |d: std::time::Duration| d.as_secs_f64() * 1e3;
+            eprintln!(
+                "    mask laps: build_mask {:.0} ms, refill + step apertures + merging {:.0} ms, impose + GCL + volumes {:.0} ms (impose {:.0}, GCL table {:.0}, old apertures + volumes {:.0}); refill {:.0} ms; device-mask assembly (pool {}), check reference {:.0} ms",
+                ms(l_build - l_ref),
+                ms(l_step - l_build),
+                ms(lap.elapsed() - l_step),
+                ms(s_impose),
+                ms(s_gcl - s_impose),
+                ms(sub.elapsed() - s_gcl),
+                ms(l_step - l_build),
+                pooled,
+                ms(l_ref)
+            );
+        }
+        self.mask = Some(new_mask);
+        fresh_cells
+    }
+    /// Retire the previous mask: its apertures and volumes MOVE into the
+    /// previous-step records (P1-5 (a): no 1 GB clone per step at ny 124),
+    /// its geometry arrays to the pool.
+    fn retire_mask(&mut self, mut old: Mask, g: super::super::Grid) {
+        // P1-5 (a): the old mask is replaced below — its arrays MOVE into
+        // the previous-step records instead of being cloned (1 GB per step
+        // at ny 124); the values are the same.
+        let fluid = std::mem::take(&mut old.cell_fluid);
+        let keep = super::super::maskupdate::mask_device_enabled();
+        if let Some(c) = old.cut.as_mut() {
+            let retired = self.apertures_old.replace([
+                std::mem::take(&mut c.a_u),
+                std::mem::take(&mut c.a_v),
+                std::mem::take(&mut c.a_w),
+            ]);
+            // R6-1: a device-built generation's arrays go back to the
+            // pool (the host path's generation is 0: dropped as before).
+            if let Some(r) = retired {
+                for (kind, v) in super::super::cut::GeomPool::A.into_iter().zip(r) {
+                    self.geom_pool.put(self.apertures_old_gen, kind, v);
+                }
+            }
+            self.apertures_old_gen = c.generation;
+            let mut vol = std::mem::take(&mut c.vol);
+            use rayon::prelude::*;
+            vol.par_iter_mut()
+                .zip(fluid.par_iter())
+                .for_each(|(v, &f)| {
+                    if !f {
+                        *v = 0.0;
+                    }
+                });
+            self.vol_old = vol;
+            if let Some(c) = old.cut.take() {
+                self.geom_pool.put_cut(c);
+            }
+        } else {
+            self.apertures_old = None;
+            self.vol_old = vec![0.0; g.cells()];
+        }
+        // R6-2: the instantaneous arrays wait for the mask two steps on.
+        if keep {
+            self.mask_pool = Some(old.into_pool(self.mask_gen, fluid));
+        }
+    }
 }
 
 /// Refill the pressure of the cells fluid in `new` and not in `old` from
 /// their face neighbours fluid in both; returns their count.
 fn refill_fresh_cells(old: &Mask, new: &Mask, field: &mut Field) -> usize {
     let g = field.grid;
+    refill_fresh_cells_p(old, new, g, &mut field.p)
+}
+
+fn refill_fresh_cells_p(old: &Mask, new: &Mask, g: super::super::Grid, p: &mut [f64]) -> usize {
     let (nx, ny, nz) = (g.nx, g.ny, g.nz);
     let periodic = new.periodic_z();
     let mut fresh = 0;
@@ -147,7 +294,7 @@ fn refill_fresh_cells(old: &Mask, new: &Mask, field: &mut Field) -> usize {
                 let mut count = 0usize;
                 let mut visit = |nb: usize| {
                     if new.is_fluid_cell(nb) && old.is_fluid_cell(nb) {
-                        sum += field.p[nb];
+                        sum += p[nb];
                         count += 1;
                     }
                 };
@@ -179,8 +326,8 @@ fn refill_fresh_cells(old: &Mask, new: &Mask, field: &mut Field) -> usize {
             }
         }
     }
-    for (idx, p) in refills {
-        field.p[idx] = p;
+    for (idx, v) in refills {
+        p[idx] = v;
     }
     fresh
 }