Performance-portable RHS pipeline for the 2D surface router (Phase 1). More...

#include <Kokkos_Core.hpp>

Include dependency graph for KokkosSurfaceKernels.hpp:

This graph shows which files directly or indirectly include this file:

Classes
struct	openswmm::twoD::gpu::MeshViews
	Immutable mesh geometry/topology mirrored to device once at initialize(). More...

struct	openswmm::twoD::gpu::StateViews
	Mutable per-step state mirrored on device. More...

Namespaces
namespace	openswmm

namespace	openswmm::twoD

namespace	openswmm::twoD::gpu

Typedefs
using	openswmm::twoD::gpu::ExecSpace = Kokkos::DefaultExecutionSpace

using	openswmm::twoD::gpu::MemSpace = ExecSpace::memory_space

using	openswmm::twoD::gpu::DView = Kokkos::View<double*, MemSpace>
	mutable double array

using	openswmm::twoD::gpu::CDView = Kokkos::View<const double*, MemSpace>
	const double array

using	openswmm::twoD::gpu::IView = Kokkos::View<int*, MemSpace>
	mutable int array

using	openswmm::twoD::gpu::CIView = Kokkos::View<const int*, MemSpace>
	const int array

Enumerations
enum	: int { openswmm::twoD::gpu::BC_WALL = 0 , openswmm::twoD::gpu::BC_NORMAL_FLOW = 1 , openswmm::twoD::gpu::BC_SPECIFIED_STAGE = 2 , openswmm::twoD::gpu::BC_SPECIFIED_FLOW = 3 , openswmm::twoD::gpu::BC_RATING_CURVE = 4 }

Functions
KOKKOS_INLINE_FUNCTION double	openswmm::twoD::gpu::evapSink (double rate, double depth, double dry_depth)
	Depth-limited evaporation sink (mirrors evapSink()).

KOKKOS_INLINE_FUNCTION double	openswmm::twoD::gpu::sq (double x)

KOKKOS_INLINE_FUNCTION int	openswmm::twoD::gpu::nbr_of (int e, int n0, int n1, int n2)

KOKKOS_INLINE_FUNCTION double	openswmm::twoD::gpu::regSqrt (double x, double eps)

KOKKOS_INLINE_FUNCTION double	openswmm::twoD::gpu::boundaryEdgeFlux (int bc, double slope, double h_bc, double q_flow, double depth, double head, double tri_cz, double area, double L, double n, double dh_eps)

void	openswmm::twoD::gpu::evaluateRhs (const MeshViews &m, const StateViews &s, DView y, DView ydot, double dry_depth, double limiter_eps, double dh_eps)

void	openswmm::twoD::gpu::precondSetup (const MeshViews &m, const StateViews &s)
	Build the per-cell diagonal D from the most recent edge fluxes/heads.

void	openswmm::twoD::gpu::precondSolve (const StateViews &s, DView r, DView z, double gamma)
	Apply (I − γD)^{-1} element-wise: z = r / (1 − γ·D).

Detailed Description

Performance-portable RHS pipeline for the 2D surface router (Phase 1).

Faithful Kokkos translation of the serial finite-volume RHS in src/engine/2d/solver/SurfaceFluxCalculator.cpp and the Jacobi preconditioner in CvodeSurfaceSolver.cpp. The math is byte-for-byte the same as the CPU reference; only the loop structure changes from serial for to Kokkos::parallel_for, so the same source compiles to OpenMP (Phase 1), CUDA / HIP / SYCL (Phase 2+) by selecting the Kokkos execution space at build time.

IMPORTANT — this layout has NO cross-cell scatter. Each stage writes only the index it owns:

per-cell stages write cell i's slot,
the edge-flux stage writes cell i's own three edge slots [i*3+e] (edge flux is stored redundantly per incident cell),
assembly gathers cell i's own three edge slots. Gradient/limiter stages only read neighbours. So the pipeline is embarrassingly parallel and needs no atomics — the §2.4 scatter concern in docs/2D_GPU_PORTABLE_CVODE_STRATEGY.md does not arise for this data structure.

NOTE on stages: the vertex-head reconstruction and the gradient/limiter stages are recomputed each RHS call (as in the CPU path) to keep the diagnostic state fields current, even though the collapsed-FD edge flux does not consume them. Omitting them would still give an identical ydot, but we mirror the reference exactly.

Author: Caleb Buahin caleb.nosp@m..bua.nosp@m.hin@g.nosp@m.mail.nosp@m..com

License\n MIT License

Classes

Namespaces

Typedefs

Enumerations

Functions

Detailed Description