embedded3 S2-4: quadratic wall gradient behind Parameters::wall_order (host predictor, operator route, e3_cut.cu; RTX_E3_WALL_ORDER); density pin holds at order 2
CI / WASM Build + Size Check (push) Blocked by required conditions
CI / Distributed Training Tests (push) Blocked by required conditions
CI / CI Success (push) Blocked by required conditions
CI / Test (macos-latest) (push) Blocked by required conditions
CI / Test (ubuntu-latest) (push) Blocked by required conditions
CI / Python Bindings (maturin) (macos-latest) (push) Blocked by required conditions
CI / Python Bindings (maturin) (ubuntu-latest) (push) Blocked by required conditions
CI / Build (macos-latest) (push) Waiting to run
CI / Clippy Check (push) Failing after 3s
CI / Build (ubuntu-latest) (push) Failing after 4s
CI / Build CPU-Only (Explicit) (push) Failing after 4s
CI / Format Check (push) Failing after 5s
Documentation / Build API Documentation (push) Failing after 4s
Documentation / Build User Guide (push) Successful in 4s
Performance Benchmarks / Run Benchmarks (push) Successful in 2m40s

Co-Authored-By: Claude Fable 5.1 <[email protected]>
This commit is contained in:
Omar Sobh
2026-09-18 07:21:57 -05:00
co-authored by Claude Fable 5.1
parent 53b1babb91
commit cda973df26
7 changed files with 87 additions and 9 deletions
@@ -187,10 +187,28 @@ __device__ double cut_face_update(const E3Params& g, const E3Ptrs& f, const E3Cu
const double* src = c == 0 ? f.su : (c == 1 ? f.sv : f.sw); const double* src = c == 0 ? f.su : (c == 1 ? f.sv : f.sw);
double source = src[fidx] * v_u; double source = src[fidx] * v_u;
double a_w = sqrt(wall[0] * wall[0] + wall[1] * wall[1] + wall[2] * wall[2]); double a_w = sqrt(wall[0] * wall[0] + wall[1] * wall[1] + wall[2] * wall[2]);
double shear = mu * a_w / distance; /* The wall gradient: one-point (order 1) or quadratic through the next
open face away from the body along the wall normal's dominant axis
(order 2; the neighbour's old value explicit). */
double c1 = 1.0 / distance, shear_explicit = 0.0;
if (g.wall_order >= 2 && a_w > 0.0) {
double nw[3] = { wall[0] / a_w, wall[1] / a_w, wall[2] / a_w };
int d = 0;
for (int kk = 1; kk < 3; ++kk) if (fabs(nw[kk]) > fabs(nw[d])) d = kk;
int q[3] = { i, j, k };
q[d] -= nw[d] > 0.0 ? 1 : -1;
int fn = cut_face(g, c, q[0], q[1], q[2]);
if (fn >= 0 && cut_ap(m, c)[fn] > 0.0) {
double d1 = distance, d2 = d1 + h[d] * fabs(nw[d]);
c1 = d2 / (d1 * (d2 - d1));
double c2 = -d1 / (d2 * (d2 - d1));
shear_explicit = mu * a_w * c2 * (old_c[fn] - ub);
}
}
double shear = mu * a_w * c1;
double v_eff = fmax(alpha, CUT_INERTIA_FLOOR) * h[c] * area[c]; double v_eff = fmax(alpha, CUT_INERTIA_FLOOR) * h[c] * area[c];
double inertia = rho * v_eff / g.dt; double inertia = rho * v_eff / g.dt;
return (inertia * u0 - conv + diff + pressure + source + shear * ub) / (inertia + shear); return (inertia * u0 - conv + diff + pressure + source + shear * ub - shear_explicit) / (inertia + shear);
} }
/* The predictor on the open interior faces of component c. */ /* The predictor on the open interior faces of component c. */
@@ -19,7 +19,7 @@ struct E3Params {
int periodic_z; int periodic_z;
int bx0, bx1, by0, by1, bz0, bz1; int bx0, bx1, by0, by1, bz0, bz1;
int scheme; int scheme;
int pad; int wall_order;
double dx, dy, dz, dt, rho, nu; double dx, dy, dz, dt, rho, nu;
}; };
@@ -209,6 +209,7 @@ impl Mask {
merge_master: Vec::new(), merge_master: Vec::new(),
scheme: crate::solvers::incompressible::ConvectionScheme::Upwind, scheme: crate::solvers::incompressible::ConvectionScheme::Upwind,
density: 1.0, density: 1.0,
wall_order: 1,
}; };
mask.compute_merging(None); mask.compute_merging(None);
Ok(mask) Ok(mask)
@@ -306,6 +307,47 @@ impl Mask {
self.compute_merging(Some(old)); self.compute_merging(Some(old));
} }
/// The wall-gradient coefficients of the unknown face of component
/// `c` at `p` with control volume `cv`: `u'(0) = c_1 (u_f U_b) + c_2
/// (u_n U_b)` with `u_n` the face returned (one lattice step away
/// from the body along the wall normal's dominant axis). Order 1, or
/// no open neighbour: `(1/d_f, 0, None)`.
pub(super) fn wall_gradient(
&self,
c: usize,
p: [i64; 3],
cv: &CvGeometry,
) -> (f64, f64, Option<usize>) {
let linear = (1.0 / cv.distance, 0.0, None);
if self.wall_order < 2 {
return linear;
}
let a_w =
(cv.wall[0] * cv.wall[0] + cv.wall[1] * cv.wall[1] + cv.wall[2] * cv.wall[2]).sqrt();
if a_w == 0.0 {
return linear;
}
let n = [cv.wall[0] / a_w, cv.wall[1] / a_w, cv.wall[2] / a_w];
let mut d = 0;
for k in 1..3 {
if n[k].abs() > n[d].abs() {
d = k;
}
}
// `n` points into the body: step the other way.
let mut q = p;
q[d] -= if n[d] > 0.0 { 1 } else { -1 };
let open = self.aperture(c, q).is_some_and(|a| a > 0.0);
if !open {
return linear;
}
let f = self.lattice().face(c, q).expect("open face");
let h = [self.grid.dx, self.grid.dy, self.grid.dz];
let d1 = cv.distance;
let d2 = d1 + h[d] * n[d].abs();
(d2 / (d1 * (d2 - d1)), -d1 / (d2 * (d2 - d1)), Some(f))
}
pub(super) fn lattice(&self) -> Lattice { pub(super) fn lattice(&self) -> Lattice {
Lattice { Lattice {
g: self.grid, g: self.grid,
@@ -659,7 +701,9 @@ impl Mask {
continue; continue;
} }
let ub = self.surface_velocity_at(body, lat.face_position(c, p), c, t); let ub = self.surface_velocity_at(body, lat.face_position(c, p), c, t);
force[c] += mu * a_w * (values[c][idx] - ub) / cv.distance; let (c1, c2, nb) = self.wall_gradient(c, p, &cv);
let un = nb.map_or(ub, |f| values[c][f]);
force[c] += mu * a_w * (c1 * (values[c][idx] - ub) + c2 * (un - ub));
} }
} }
} }
@@ -226,7 +226,10 @@ impl Solver {
}); });
let a_w = let a_w =
(cv.wall[0] * cv.wall[0] + cv.wall[1] * cv.wall[1] + cv.wall[2] * cv.wall[2]).sqrt(); (cv.wall[0] * cv.wall[0] + cv.wall[1] * cv.wall[1] + cv.wall[2] * cv.wall[2]).sqrt();
let shear = mu * a_w / cv.distance; let (c1, c2, nb) = mask.wall_gradient(c, p, &cv);
let shear = mu * a_w * c1;
// The explicit part of the quadratic wall gradient (order 2).
let shear_explicit = nb.map_or(0.0, |f| mu * a_w * c2 * (old[c][f] - ub));
let fraction = if self.params.momentum_volume_cell_mean { let fraction = if self.params.momentum_volume_cell_mean {
let vol = |q: [i64; 3]| lat.cell(q).map_or(cv.alpha, |ci| mask.vol(ci)); let vol = |q: [i64; 3]| lat.cell(q).map_or(cv.alpha, |ci| mask.vol(ci));
0.5 * (vol(cell_minus) + vol(cell_plus)) 0.5 * (vol(cell_minus) + vol(cell_plus))
@@ -235,8 +238,8 @@ impl Solver {
}; };
let v_eff = fraction.max(INERTIA_FLOOR) * h[c] * area[c]; let v_eff = fraction.max(INERTIA_FLOOR) * h[c] * area[c];
let inertia = rho * v_eff / dt; let inertia = rho * v_eff / dt;
let u_star = let u_star = (inertia * u0 - conv + diff + pressure + source + shear * ub - shear_explicit)
(inertia * u0 - conv + diff + pressure + source + shear * ub) / (inertia + shear); / (inertia + shear);
let v_alpha = fraction * h[c] * area[c]; let v_alpha = fraction * h[c] * area[c];
(u_star, rho * (v_eff - v_alpha) * (u_star - u0) / dt) (u_star, rho * (v_eff - v_alpha) * (u_star - u0) / dt)
} }
@@ -78,7 +78,7 @@ struct E3Params {
bz0: i32, bz0: i32,
bz1: i32, bz1: i32,
scheme: i32, scheme: i32,
pad: i32, wall_order: i32,
dx: f64, dx: f64,
dy: f64, dy: f64,
dz: f64, dz: f64,
@@ -284,7 +284,7 @@ impl DeviceStep {
bz0: side_code(b.z0), bz0: side_code(b.z0),
bz1: side_code(b.z1), bz1: side_code(b.z1),
scheme: scheme_code(self.solver.params.convection_scheme), scheme: scheme_code(self.solver.params.convection_scheme),
pad: 0, wall_order: i32::from(self.solver.params.wall_order),
dx: g.dx, dx: g.dx,
dy: g.dy, dy: g.dy,
dz: g.dz, dz: g.dz,
@@ -91,6 +91,11 @@ pub struct Parameters {
/// inertia of the fluid half of its volume instead of a tenth of a /// inertia of the fluid half of its volume instead of a tenth of a
/// cell against O(1) fluxes). /// cell against O(1) fluxes).
pub momentum_volume_cell_mean: bool, pub momentum_volume_cell_mean: bool,
/// The cut wall's shear closure: 1 = one-point `μ A_w (u_f U_b)/d_f`
/// (the recorded form), 2 = the quadratic wall gradient through the
/// next open face along the wall normal's dominant axis (S2-4).
/// `Parameters::default()` reads `RTX_E3_WALL_ORDER` (default 1).
pub wall_order: u8,
} }
impl Default for Parameters { impl Default for Parameters {
@@ -107,6 +112,10 @@ impl Default for Parameters {
max_surface_speed: None, max_surface_speed: None,
aperture_substeps: 0, aperture_substeps: 0,
momentum_volume_cell_mean: false, momentum_volume_cell_mean: false,
wall_order: std::env::var("RTX_E3_WALL_ORDER")
.ok()
.and_then(|v| v.parse().ok())
.unwrap_or(1),
} }
} }
} }
@@ -231,6 +240,7 @@ impl Solver {
.map(|mut m| { .map(|mut m| {
m.scheme = self.params.convection_scheme; m.scheme = self.params.convection_scheme;
m.density = self.fluid.density; m.density = self.fluid.density;
m.wall_order = self.params.wall_order;
m m
}) })
.expect("embedded mask") .expect("embedded mask")
@@ -97,6 +97,8 @@ pub struct Mask {
pub(super) scheme: crate::solvers::incompressible::ConvectionScheme, pub(super) scheme: crate::solvers::incompressible::ConvectionScheme,
/// The fluid's density (the exchange route's convective flux). /// The fluid's density (the exchange route's convective flux).
pub(super) density: f64, pub(super) density: f64,
/// The cut wall's shear closure order (S2-4).
pub(super) wall_order: u8,
} }
/// The z lattice position of a query: the lower plane index, the upper /// The z lattice position of a query: the lower plane index, the upper
@@ -509,6 +511,7 @@ impl Mask {
merge_master: Vec::new(), merge_master: Vec::new(),
scheme: crate::solvers::incompressible::ConvectionScheme::Upwind, scheme: crate::solvers::incompressible::ConvectionScheme::Upwind,
density: 1.0, density: 1.0,
wall_order: 1,
}) })
} }