PERF-2 P3: the batched device V-cycle microbenchmark — mg_vcycle.cu (masked variable-coefficient five-point red-black half-sweeps, residual, CSR restriction in a fixed order, prolongation, coarsest sweeps; [K][n] layout, blockIdx.y = march), LevelExport/export_hierarchy and vcycle_f32_reference on the CPU side, and the ignored cuda-feature test that checks the K = 1 device V-cycle against the CPU f32 reference and times K = 1 / 8 / 16 per march
CI / Build (macos-latest) (push) Waiting to run
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 / 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
Performance Benchmarks / Run Benchmarks (push) Failing after 6s
CI / Build (ubuntu-latest) (push) Failing after 5s
Documentation / Build API Documentation (push) Failing after 6s
Documentation / Build User Guide (push) Successful in 7s
CI / Format Check (push) Failing after 12s
CI / Clippy Check (push) Failing after 36s
CI / Build CPU-Only (Explicit) (push) Failing after 1m37s

Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01YJPeT6WA2e7YvAnS875AHL
This commit is contained in:
Omar Sobh
2026-09-16 01:03:02 -05:00
co-authored by Claude Fable 5.1
parent 53078c2aa7
commit 0dc95a8de5
4 changed files with 571 additions and 1 deletions
@@ -65,7 +65,8 @@ pub use piso::{PisoParameters, PisoResult, PisoSolver};
#[cfg(feature = "cuda")]
pub use piso_gpu::PisoGpuSolver;
pub use poisson::{
MgPrecision, MgSmoother, MultigridParameters, PcgCache, PoissonProblem, PoissonSolution,
LevelExport, MgPrecision, MgSmoother, MultigridParameters, PcgCache, PoissonProblem,
PoissonSolution, export_hierarchy, vcycle_f32_reference,
PoissonSolverKind, configure_threads, solve_multigrid_pcg, solve_multigrid_pcg_cached,
};
pub use polygon_sdf::PolygonSdf;
@@ -1001,6 +1001,99 @@ pub fn configure_threads(threads: usize) -> usize {
n
}
/// PERF-2 P3 (`docs/perf2_campaign.md`): one hierarchy level exported for a
/// device V-cycle — the sanitised f32 coefficients, the active-cell and
/// colour index lists, the parent map, and the coarse cells' children in
/// CSR form (so a restriction can sum in a fixed order on the device).
#[derive(Debug, Clone)]
pub struct LevelExport {
pub nx: usize,
pub ny: usize,
pub cells: Vec<u32>,
pub red: Vec<u32>,
pub black: Vec<u32>,
/// Fine cell → coarse cell (`u32::MAX` without an equation; empty on
/// the coarsest level).
pub coarse_of: Vec<u32>,
/// For the NEXT level's cells, in its `cells` order: the fine cells
/// restricting into each (CSR: `children_ptr[c]..children_ptr[c + 1]`).
pub children_ptr: Vec<u32>,
pub children_idx: Vec<u32>,
pub ae: Vec<f32>,
pub aw: Vec<f32>,
pub an: Vec<f32>,
pub as_: Vec<f32>,
pub ap: Vec<f32>,
}
/// The f32 hierarchy of `problem` (red-black colour lists included), level
/// 0 fine, for a device implementation of [`Hierarchy::apply_preconditioner`].
pub fn export_hierarchy(problem: &PoissonProblem, params: &MultigridParameters) -> Vec<LevelExport> {
let hier = Hierarchy::<f32>::build(problem, params);
let depth = hier.levels.len();
(0..depth)
.map(|l| {
let lv = &hier.levels[l];
let to_u32 = |v: &[usize]| v.iter().map(|&i| i as u32).collect::<Vec<u32>>();
let (children_ptr, children_idx) = if l + 1 < depth {
let coarse = &hier.levels[l + 1];
let mut pos = vec![usize::MAX; coarse.problem.nx * coarse.problem.ny];
for (k, &c) in coarse.cells.iter().enumerate() {
pos[c] = k;
}
let mut lists: Vec<Vec<u32>> = vec![Vec::new(); coarse.cells.len()];
for &idx in &lv.cells {
let c = lv.coarse_of[idx];
if coarse.active[c] {
lists[pos[c]].push(idx as u32);
}
}
let mut ptr = Vec::with_capacity(lists.len() + 1);
let mut flat = Vec::new();
ptr.push(0u32);
for list in &lists {
flat.extend_from_slice(list);
ptr.push(flat.len() as u32);
}
(ptr, flat)
} else {
(Vec::new(), Vec::new())
};
LevelExport {
nx: lv.problem.nx,
ny: lv.problem.ny,
cells: to_u32(&lv.cells),
red: to_u32(&lv.red),
black: to_u32(&lv.black),
coarse_of: lv
.coarse_of
.iter()
.map(|&c| if c == usize::MAX { u32::MAX } else { c as u32 })
.collect(),
children_ptr,
children_idx,
ae: lv.ae.clone(),
aw: lv.aw.clone(),
an: lv.an.clone(),
as_: lv.as_.clone(),
ap: lv.ap.clone(),
}
})
.collect()
}
/// One f32 V-cycle of `problem`'s hierarchy on `r` (the CPU reference for a
/// device V-cycle): `z = M⁻¹ r` exactly as the preconditioner computes it.
pub fn vcycle_f32_reference(
problem: &PoissonProblem,
params: &MultigridParameters,
r: &[f64],
z: &mut [f64],
) {
let mut hier = Hierarchy::<f32>::build(problem, params);
hier.apply_preconditioner(r, z);
}
/// 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)]