policy.hpp File Reference

Backend enum for linear algebra operations. More...

Go to the source code of this file.

Namespaces
namespace	num

Enumerations
enum class	num::Backend {   num::seq , num::blocked , num::simd , num::blas ,   num::omp , num::gpu }
	Selects which backend handles a linalg operation. More...

Variables
constexpr Backend	num::seq = Backend::seq
constexpr Backend	num::blocked = Backend::blocked
constexpr Backend	num::simd = Backend::simd
constexpr Backend	num::blas = Backend::blas
constexpr Backend	num::omp = Backend::omp
constexpr Backend	num::gpu = Backend::gpu
constexpr bool	num::has_blas
	True when a BLAS/cblas library was found at configure time.
constexpr bool	num::has_omp
	True when OpenMP was found at configure time.
constexpr Backend	num::default_backend
constexpr Backend	num::best_backend
	Best backend for memory-bound vector ops: blas > omp > blocked.

Detailed Description

Backend enum for linear algebra operations.

Each module defines its own backend enum for the choices relevant to it. This enum covers linalg (vector, matrix, solvers, factorization, eigen, svd). Other modules define their own – e.g. spectral/ uses FFTBackend.

num::seq – naive serial C++ (always available) num::blocked – cache-blocked, no intrinsics (compiler auto-vectorizes) num::simd – hand-written SIMD (AVX2 on x86, NEON on AArch64) num::blas – cblas / LAPACKE (optional, NUMERICS_HAS_BLAS) num::omp – OpenMP parallel (optional, NUMERICS_HAS_OMP) num::gpu – CUDA (optional, NUMERICS_HAS_CUDA)

Definition in file policy.hpp.