PERF-2 P3 bench: per-level comparison of the device V-cycle's work vectors against the CPU's (vcycle_f32_work) to locate the disagreement

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

@@ -69,7 +69,7 @@ pub use piso::{PisoParameters, PisoResult, PisoSolver};
 // pub use piso_gpu::PisoGpuSolver;
 pub use poisson::{
     LevelExport, MgPrecision, MgSmoother, MultigridParameters, PcgCache, PoissonProblem,
-    PoissonSolution, export_hierarchy, vcycle_f32_reference,
+    PoissonSolution, export_hierarchy, vcycle_f32_reference, vcycle_f32_work,
     PoissonSolverKind, configure_threads, solve_multigrid_pcg, solve_multigrid_pcg_cached,
 };
 pub use polygon_sdf::PolygonSdf;

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

@@ -1094,6 +1094,24 @@ pub fn vcycle_f32_reference(
     hier.apply_preconditioner(r, z);
 }
 
+/// The f32 V-cycle's per-level work vectors after one application on `r`
+/// (for locating where a device V-cycle departs from the CPU one): per
+/// level `(b, x, r)` — the restricted right-hand side, the level's final
+/// correction (after its up-leg smoothing), and its down-leg residual.
+pub fn vcycle_f32_work(
+    problem: &PoissonProblem,
+    params: &MultigridParameters,
+    r: &[f64],
+) -> Vec<(Vec<f32>, Vec<f32>, Vec<f32>)> {
+    let mut hier = Hierarchy::<f32>::build(problem, params);
+    let mut z = vec![0.0; r.len()];
+    hier.apply_preconditioner(r, &mut z);
+    hier.work
+        .iter()
+        .map(|w| (w.b.clone(), w.x.clone(), w.r.clone()))
+        .collect()
+}
+
 /// A reusable prepared solver per V-cycle precision (PERF-2 P1.1): the
 /// operator's hierarchy is rebuilt only when the operator changes.
 #[derive(Default)]