embedded3 item 9b: the device step carries a static cut-cell mask (e3_cut.cu: cut predictor, apertured merged continuity with the fold, owner-read corrections; device CG off-stencil links) — host = device to 2e-10 (CFD1 cylinder nz 4) and 4e-14 (sphere) under tight tolerances

Co-Authored-By: Claude Fable 5.1 <[email protected]>

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

@@ -70,6 +70,10 @@ pub struct DeviceCg {
     at: CudaSlice<f64>,
     ab: CudaSlice<f64>,
     ap: CudaSlice<f64>,
+    /// Off-stencil links per cell (CSR; the merged small cells).
+    link_ptr: CudaSlice<u32>,
+    link_idx: CudaSlice<u32>,
+    link_coef: CudaSlice<f64>,
     b: CudaSlice<f64>,
     r: CudaSlice<f64>,
     z: CudaSlice<f64>,
@@ -118,6 +122,18 @@ impl DeviceCg {
         };
         let up_f = |v: &[f64]| -> CudaSlice<f64> { rt.stream.memcpy_stod(v).expect("upload") };
         let zeros = || rt.stream.alloc_zeros::<f64>(n).expect("alloc");
+        let lists = problem.link_lists();
+        let mut link_ptr = Vec::with_capacity(n + 1);
+        let mut link_idx = Vec::new();
+        let mut link_coef = Vec::new();
+        link_ptr.push(0u32);
+        for list in &lists {
+            for &(other, c) in list {
+                link_idx.push(other as u32);
+                link_coef.push(c);
+            }
+            link_ptr.push(link_idx.len() as u32);
+        }
         Self {
             key: OperatorKey::of(problem, params),
             n,
@@ -134,6 +150,13 @@ impl DeviceCg {
             at: up_f(&fine.at),
             ab: up_f(&fine.ab),
             ap: up_f(&fine.ap),
+            link_ptr: up_u(&link_ptr),
+            link_idx: up_u(&link_idx),
+            link_coef: up_f(if link_coef.is_empty() {
+                &[0.0]
+            } else {
+                &link_coef
+            }),
             b: zeros(),
             r: zeros(),
             z: zeros(),
@@ -324,6 +347,9 @@ impl DeviceCg {
                 .arg(&self.ab)
                 .arg(&self.top)
                 .arg(&self.bot)
+                .arg(&self.link_ptr)
+                .arg(&self.link_idx)
+                .arg(&self.link_coef)
                 .arg(&self.ap)
                 .arg(&self.b)
                 .arg(p)
@@ -481,6 +507,9 @@ impl DeviceCg {
                     .arg(&self.ab)
                     .arg(&self.top)
                     .arg(&self.bot)
+                    .arg(&self.link_ptr)
+                    .arg(&self.link_idx)
+                    .arg(&self.link_coef)
                     .arg(&self.ap)
                     .arg(&self.d)
                     .arg(&mut self.q)

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

@@ -208,7 +208,14 @@ impl Problem {
             }
         }
         for &(a, b, c) in &self.links {
-            if a >= n || b >= n || a == b || !self.active[a] || !self.active[b] || c <= 0.0 || c.is_nan() {
+            if a >= n
+                || b >= n
+                || a == b
+                || !self.active[a]
+                || !self.active[b]
+                || c <= 0.0
+                || c.is_nan()
+            {
                 return Err(format!("invalid link ({a}, {b}, {c})"));
             }
         }

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

@@ -5,6 +5,8 @@
 //! fields into a `Field` at instants. Compiled with FMA contraction
 //! off so the predictors are the host's arithmetic to the bit.
 
+mod cut;
+
 use super::{Side, Solver, StepResult};
 use crate::solvers::incompressible::embedded3::Grid;
 use crate::solvers::incompressible::embedded3::field::Field;
@@ -150,6 +152,8 @@ pub struct DeviceStep {
     cg_dt: f64,
     timers: Option<StepTimers>,
     initialized: bool,
+    /// A static cut-cell mask's tables (item 9b), when the solver has one.
+    cut: Option<cut::DeviceCut>,
 }
 
 impl DeviceStep {
@@ -177,6 +181,7 @@ impl DeviceStep {
         let timers = std::env::var("RTX_PROFILE")
             .is_ok()
             .then(StepTimers::default);
+        let cut = cut::DeviceCut::build(&solver, grid);
         Self {
             solver,
             grid,
@@ -203,9 +208,15 @@ impl DeviceStep {
             cg_dt: 0.0,
             timers,
             initialized: false,
+            cut,
         }
     }
 
+    /// The number of virtually merged cells on the device mask (0 without one).
+    pub fn merged_cells(&self) -> usize {
+        self.cut.as_ref().map_or(0, |c| c.merged)
+    }
+
     pub fn timers(&self) -> Option<StepTimers> {
         self.timers
     }
@@ -385,6 +396,9 @@ impl DeviceStep {
         if !self.initialized {
             self.initialize();
         }
+        if self.cut.is_some() {
+            return self.advance_cut(dt);
+        }
         let rt = runtime();
         let k = kernels();
         let g = self.grid;

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

@@ -0,0 +1,380 @@
+//! Item 9b: the device step carrying a static cut-cell mask — the mask's
+//! tables uploaded once (apertures, face distances, surface velocities at
+//! the feet, the compatible wall flux, the open faces, the active cells,
+//! the merged cells' owners and the fold lists) and the step's cut
+//! 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::poisson::device::{cfg, load_module, runtime};
+use crate::solvers::incompressible::embedded3::poisson::device_cg::DeviceCg;
+use crate::solvers::incompressible::embedded3::step::Solver;
+use crate::solvers::incompressible::poisson::MultigridParameters;
+use cudarc::driver::{
+    CudaFunction, CudaModule, CudaSlice, DevicePtr, DeviceRepr, PushKernelArg, ValidAsZeroBits,
+};
+use std::sync::{Arc, OnceLock};
+use std::time::Instant;
+
+const CUT_KERNELS: &str = concat!(
+    include_str!("../../../../../kernels/cuda/e3_step.cu"),
+    include_str!("../../../../../kernels/cuda/e3_cut.cu")
+);
+
+struct CutKernels {
+    _module: Arc<CudaModule>,
+    predict: CudaFunction,
+    divergence: CudaFunction,
+    fold: CudaFunction,
+    correct: CudaFunction,
+    add_p: CudaFunction,
+}
+
+static CUT_KERNELS_ONCE: OnceLock<CutKernels> = OnceLock::new();
+
+fn cut_kernels() -> &'static CutKernels {
+    CUT_KERNELS_ONCE.get_or_init(|| {
+        let module = load_module(CUT_KERNELS, "e3_cut.cu", true);
+        let f = |name: &str| module.load_function(name).expect(name);
+        CutKernels {
+            predict: f("e3_cut_predict"),
+            divergence: f("e3_cut_divergence"),
+            fold: f("e3_cut_fold"),
+            correct: f("e3_cut_correct"),
+            add_p: f("e3_cut_add_p"),
+            _module: module,
+        }
+    })
+}
+
+/// `struct E3Cut` in e3_cut.cu: 18 device pointers.
+#[repr(C)]
+#[derive(Clone, Copy)]
+struct E3CutPtrs {
+    ptrs: [u64; 18],
+}
+unsafe impl DeviceRepr for E3CutPtrs {}
+unsafe impl ValidAsZeroBits for E3CutPtrs {}
+
+/// The static cut-cell mask on the device.
+pub(super) struct DeviceCut {
+    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>,
+    cell_flux: CudaSlice<f64>,
+    pub(super) merged: usize,
+}
+
+impl DeviceCut {
+    /// The tables of the solver's cut mask (`None` without one).
+    pub(super) fn build(solver: &Solver, g: Grid) -> Option<Self> {
+        let mask = solver.mask()?;
+        let cut = mask.cut()?;
+        let body = solver.body()?;
+        let rt = runtime();
+        let (nx, ny, nz) = (g.nx, g.ny, g.nz);
+        let h = [g.dx, g.dy, g.dz];
+        let counts = [(nx + 1) * ny * nz, nx * (ny + 1) * nz, nx * ny * (nz + 1)];
+        let up_f = |v: &[f64]| -> CudaSlice<f64> {
+            rt.stream
+                .memcpy_stod(if v.is_empty() { &[0.0f64][..] } else { v })
+                .expect("upload")
+        };
+        let up_i = |v: &[i32]| -> CudaSlice<i32> { rt.stream.memcpy_stod(v).expect("upload") };
+        let up_u = |v: &[u32]| -> CudaSlice<u32> {
+            rt.stream
+                .memcpy_stod(if v.is_empty() { &[0u32][..] } else { v })
+                .expect("upload")
+        };
+        // Surface velocity at the foot per face, and the open flags.
+        let mut ub: [Vec<f64>; 3] = [Vec::new(), Vec::new(), Vec::new()];
+        let mut open: [Vec<i32>; 3] = [Vec::new(), Vec::new(), Vec::new()];
+        for c in 0..3 {
+            let (ni, nj, nk) = match c {
+                0 => (nx + 1, ny, nz),
+                1 => (nx, ny + 1, nz),
+                _ => (nx, ny, nz + 1),
+            };
+            let mut ubc = vec![0.0; counts[c]];
+            let mut opc = vec![0i32; counts[c]];
+            for k in 0..nk {
+                for j in 0..nj {
+                    for i in 0..ni {
+                        let idx = (k * nj + j) * ni + i;
+                        let x = [
+                            (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],
+                        ];
+                        ubc[idx] = mask.surface_velocity_at(body, x, c, 0.0);
+                        opc[idx] = i32::from(match c {
+                            0 => mask.u_open(idx),
+                            1 => mask.v_open(idx),
+                            _ => mask.w_open(idx),
+                        });
+                    }
+                }
+            }
+            ub[c] = ubc;
+            open[c] = opc;
+        }
+        let (wall_flux, _) = mask.wall_flux_table(body, 0.0);
+        let nc = g.cells();
+        let active: Vec<i32> = (0..nc).map(|i| i32::from(mask.cell_active(i))).collect();
+        let owner: Vec<u32> = (0..nc)
+            .map(|i| mask.master(i).unwrap_or(i) as u32)
+            .collect();
+        let mut fold_ptr = Vec::with_capacity(nc + 1);
+        let mut fold_idx = Vec::new();
+        let mut slaves_of: Vec<Vec<u32>> = vec![Vec::new(); nc];
+        let mut merged = 0;
+        for i in 0..nc {
+            if let Some(m) = mask.master(i) {
+                slaves_of[m].push(i as u32);
+                merged += 1;
+            }
+        }
+        fold_ptr.push(0u32);
+        for list in &slaves_of {
+            fold_idx.extend_from_slice(list);
+            fold_ptr.push(fold_idx.len() as u32);
+        }
+        Some(Self {
+            a: [up_f(&cut.a_u), up_f(&cut.a_v), up_f(&cut.a_w)],
+            d: [up_f(&cut.d_u), up_f(&cut.d_v), up_f(&cut.d_w)],
+            ub: [up_f(&ub[0]), up_f(&ub[1]), up_f(&ub[2])],
+            wall_flux: up_f(&wall_flux),
+            open: [up_i(&open[0]), up_i(&open[1]), up_i(&open[2])],
+            active: up_i(&active),
+            owner: up_u(&owner),
+            fold_ptr: up_u(&fold_ptr),
+            fold_idx: up_u(&fold_idx),
+            cell_flux: rt.stream.alloc_zeros::<f64>(nc).expect("alloc"),
+            merged,
+        })
+    }
+
+    fn ptrs(&self) -> E3CutPtrs {
+        let rt = runtime();
+        let s = &rt.stream;
+        let pf = |x: &CudaSlice<f64>| x.device_ptr(s).0;
+        let pi = |x: &CudaSlice<i32>| x.device_ptr(s).0;
+        let pu = |x: &CudaSlice<u32>| x.device_ptr(s).0;
+        E3CutPtrs {
+            ptrs: [
+                pf(&self.a[0]),
+                pf(&self.a[1]),
+                pf(&self.a[2]),
+                pf(&self.d[0]),
+                pf(&self.d[1]),
+                pf(&self.d[2]),
+                pf(&self.ub[0]),
+                pf(&self.ub[1]),
+                pf(&self.ub[2]),
+                pf(&self.wall_flux),
+                pi(&self.open[0]),
+                pi(&self.open[1]),
+                pi(&self.open[2]),
+                pi(&self.active),
+                pu(&self.owner),
+                pu(&self.fold_ptr),
+                pu(&self.fold_idx),
+                pf(&self.cell_flux),
+            ],
+        }
+    }
+}
+
+impl DeviceStep {
+    /// One step on the cut-cell mask (the host `advance` with the cut
+    /// predictor, the apertured merged continuity and the owner-read
+    /// corrections).
+    pub(super) fn advance_cut(&mut self, dt: f64) -> StepResult {
+        let rt = runtime();
+        let k = cut_kernels();
+        let g = self.grid;
+        let t_old = self.solver.time();
+        let t_new = t_old + dt;
+        let t0 = Instant::now();
+        rt.stream
+            .memcpy_dtod(&self.u, &mut self.u_old)
+            .expect("u_old");
+        rt.stream
+            .memcpy_dtod(&self.v, &mut self.v_old)
+            .expect("v_old");
+        rt.stream
+            .memcpy_dtod(&self.w, &mut self.w_old)
+            .expect("w_old");
+        self.upload_tables(t_old);
+        let prm: E3Params = self.params(dt);
+        let ptrs: E3Ptrs = self.ptrs();
+        let cptrs = self.cut.as_ref().expect("cut").ptrs();
+        let counts = [
+            (g.nx + 1) * g.ny * g.nz,
+            g.nx * (g.ny + 1) * g.nz,
+            g.nx * g.ny * (g.nz + 1),
+        ];
+        for c in 0..3i32 {
+            unsafe {
+                rt.stream
+                    .launch_builder(&k.predict)
+                    .arg(&prm)
+                    .arg(&ptrs)
+                    .arg(&cptrs)
+                    .arg(&c)
+                    .launch(cfg(counts[c as usize]))
+                    .expect("e3_cut_predict");
+            }
+        }
+        self.launch_sides(prm, &ptrs, 0);
+        self.upload_tables(t_new);
+        self.launch_sides(prm, &ptrs, 1);
+        rt.stream
+            .memcpy_dtod(&self.u, &mut self.u_star)
+            .expect("u*");
+        rt.stream
+            .memcpy_dtod(&self.v, &mut self.v_star)
+            .expect("v*");
+        rt.stream
+            .memcpy_dtod(&self.w, &mut self.w_star)
+            .expect("w*");
+        rt.stream.synchronize().expect("sync");
+        let t_pred = t0.elapsed();
+        if self.cg.is_none() || self.cg_dt != dt {
+            let problem = self.solver.poisson_operator(g, dt);
+            let params = MultigridParameters {
+                precision: self.solver.params.poisson_precision,
+                smoother: self.solver.params.poisson_smoother,
+                ..MultigridParameters::default()
+            };
+            self.cg = Some(DeviceCg::new(&problem, &params));
+            self.cg_dt = dt;
+        }
+        let anchor = self.solver.anchor_cell(g);
+        let mut total = 0;
+        let mut final_residual = f64::INFINITY;
+        let mut cg_iterations = 0;
+        let mut t_poisson = std::time::Duration::ZERO;
+        let mut t_apply = std::time::Duration::ZERO;
+        let one = 1i32;
+        let zero = 0i32;
+        for corrector in 0..self.solver.params.corrector_steps.max(1) {
+            let tp = Instant::now();
+            unsafe {
+                rt.stream
+                    .launch_builder(&k.divergence)
+                    .arg(&prm)
+                    .arg(&ptrs)
+                    .arg(&cptrs)
+                    .arg(&one)
+                    .launch(cfg(g.cells()))
+                    .expect("e3_cut_divergence");
+                rt.stream
+                    .launch_builder(&k.fold)
+                    .arg(&prm)
+                    .arg(&ptrs)
+                    .arg(&cptrs)
+                    .arg(&one)
+                    .arg(&mut self.partial)
+                    .launch(cfg(g.cells()))
+                    .expect("e3_cut_fold");
+            }
+            let source_scale = self.reduce();
+            let inner_stop = self.solver.inner_stop(g, source_scale);
+            if corrector > 0 {
+                rt.stream.memset_zeros(&mut self.p_prime).expect("p' = 0");
+            }
+            let sol = {
+                let cg = self.cg.as_mut().expect("cg");
+                cg.solve_device(&self.sp, &mut self.p_prime, inner_stop, anchor, 0)
+            };
+            cg_iterations += sol.iterations;
+            t_poisson += tp.elapsed();
+            let ta = Instant::now();
+            for c in 0..3i32 {
+                unsafe {
+                    rt.stream
+                        .launch_builder(&k.correct)
+                        .arg(&prm)
+                        .arg(&ptrs)
+                        .arg(&cptrs)
+                        .arg(&c)
+                        .launch(cfg(counts[c as usize]))
+                        .expect("e3_cut_correct");
+                }
+            }
+            if prm.periodic_z != 0 {
+                self.launch_sides(prm, &ptrs, 0);
+            }
+            unsafe {
+                rt.stream
+                    .launch_builder(&k.add_p)
+                    .arg(&prm)
+                    .arg(&ptrs)
+                    .arg(&cptrs)
+                    .launch(cfg(g.cells()))
+                    .expect("e3_cut_add_p");
+                rt.stream
+                    .launch_builder(&k.divergence)
+                    .arg(&prm)
+                    .arg(&ptrs)
+                    .arg(&cptrs)
+                    .arg(&zero)
+                    .launch(cfg(g.cells()))
+                    .expect("e3_cut_divergence");
+                rt.stream
+                    .launch_builder(&k.fold)
+                    .arg(&prm)
+                    .arg(&ptrs)
+                    .arg(&cptrs)
+                    .arg(&zero)
+                    .arg(&mut self.partial)
+                    .launch(cfg(g.cells()))
+                    .expect("e3_cut_fold");
+            }
+            let imbalance = self.reduce();
+            let reference_flux = self.solver.reference_flux(g);
+            let mass_residual = if reference_flux > 0.0 {
+                imbalance / reference_flux
+            } else {
+                imbalance
+            };
+            final_residual = mass_residual;
+            total += 1;
+            t_apply += ta.elapsed();
+            if mass_residual < self.solver.params.tolerance {
+                break;
+            }
+            rt.stream
+                .memcpy_dtod(&self.u, &mut self.u_star)
+                .expect("u*");
+            rt.stream
+                .memcpy_dtod(&self.v, &mut self.v_star)
+                .expect("v*");
+            rt.stream
+                .memcpy_dtod(&self.w, &mut self.w_star)
+                .expect("w*");
+        }
+        self.solver.set_time(t_new);
+        if let Some(tm) = self.timers.as_mut() {
+            tm.predictor_ns += t_pred.as_nanos() as u64;
+            tm.poisson_ns += t_poisson.as_nanos() as u64;
+            tm.apply_ns += t_apply.as_nanos() as u64;
+            tm.steps += 1;
+            tm.cg_iterations += cg_iterations as u64;
+        }
+        StepResult {
+            fresh_cells: 0,
+            converged: final_residual < self.solver.params.tolerance,
+            corrector_steps_performed: total,
+            final_residual,
+            poisson_iterations: cg_iterations,
+        }
+    }
+}