policy.hpp File Reference

flow — the multidimensional iteration policies: iteration order follows storage. More...

#include <Kokkos_Core.hpp>

Go to the source code of this file.

Namespaces
namespace	peclet
namespace	peclet::flow

Typedefs
template<class Exec >
using	peclet::flow::MDRange2 = Kokkos::MDRangePolicy< Exec, Kokkos::Rank< 2, Kokkos::Iterate::Left, Kokkos::Iterate::Left > >
	2-D range, first index fastest (both the tile order and the order within a tile).
template<class Exec >
using	peclet::flow::MDRange3 = Kokkos::MDRangePolicy< Exec, Kokkos::Rank< 3, Kokkos::Iterate::Left, Kokkos::Iterate::Left > >
	3-D range, x (the first index) fastest (both the tile order and the order within a tile).

Detailed Description

flow — the multidimensional iteration policies: iteration order follows storage.

Every flow field is x-fastest: cell (x, y, z) of an nx*ny*nz block lives at I = x + y*nx + z*nx*ny (Kokkos LayoutLeft, Fortran order in Python, x-fastest in VTI; suite/docs/CONVENTIONS.md). Every multidimensional kernel therefore iterates x FASTEST on EVERY backend, never Kokkos' backend default: the default is Iterate::Left on CUDA/HIP (first index fastest – already right) but Iterate::Right on the host backends (LAST index fastest), which made every OpenMP/Serial kernel stride nx*ny doubles between consecutive iterations.

The rule: an MDRange kernel is written MDRange3<Exec>(space, {x0, y0, z0}, {x1, y1, z1}) with the lambda taking (x, y, z) – the storage-fastest index FIRST; a 2-D plane kernel should take its two in-plane indices in storage order too (the cyclic (a+1)%3, (a+2)%3 face planes do not for a = y: O(N^2) work, left as they are). Kokkos::Rank< and a bare Kokkos::MDRangePolicy< may appear only in this header; tests/python/test_iteration_order.py (ctest iteration_order) fails otherwise, in src/ and tests/ alike.

Numerics: on CUDA/HIP these are the defaults already, so the device code is bit-identical. On a host backend a kernel that writes each cell from exactly one iteration is bit-identical as well; an MDRange REDUCTION, or an atomic scatter into a shared cell, may change its summation order (it follows the iteration order).

Definition in file policy.hpp.