SIMDVector.md

SIMDVector (concept)

Describes all SIMD vector types provided by Simdee, regardless of underlying datatype. A particular vector type always satisfies one of the extended concepts SIMDVectorB, SIMDVectorF, SIMDVectorU, SIMDVectorS.

Implementations

• sd::sseb, sd::ssef, sd::sseu, sd::sses
• sd::avxb, sd::avxf, sd::avxu, sd::avxs
• sd::neonb, sd::neonf, sd::neonu, sd::neons
• sd::dumb, sd::dumf, sd::dumu, sd::dums
• sd::dual<T>
• sd::vec4b, sd::vec4f, sd::vec4u, sd::vec4s
• sd::vec8b, sd::vec8f, sd::vec8u, sd::vec8s

Requirements

Member values

Type that satisfies SIMDVector must provide the following member values:

value	description
width	number of scalars in the vector

Member types

Type that satisfies SIMDVector must provide the following member types:

type	description	additional requirements
vec_b	related boolean SIMD vector	satisfies SIMDVectorB
vec_f	related floating-point SIMD vector	satisfies SIMDVectorF
vec_u	related unsigned integral SIMD vector	satisfies SIMDVectorU
vec_s	related signed integral SIMD vector	satisfies SIMDVectorS
vector_t	low-level underlying vector type
scalar_t	scalar type
storage_t	type for storing vector in memory

Construction and assignment

Type T that satisfies SIMDVector must be implicitly constructible and assignable from:

• T
• vector_t, scalar_t, storage_t
• result of sd::zero(), sd::all_bits(), sd::sign_bit(), sd::abs_mask()
• result of sd::aligned(), sd::unaligned()

Additionally, T must be assignable to:

• storage_t
• result of sd::aligned(), sd::unaligned()

Operations

For a type T that satisfies SIMDVector, the following operations must be provided:

syntax	description
x.eval()	no-op, used to enforce evaluation of expression templates
x.data()	provides access to the underlying low-level register value
x.broadcast<N>()	produces a vector with all scalars set to N-th scalar of x
x.aligned_load(ptr)	load vector from a memory location aligned to alignof(T) bytes
x.aligned_store(ptr)	store vector to a memory location aligned to alignof(T) bytes
x.unaligned_load(ptr)	load vector from an arbitrary memory location
x.unaligned_store(ptr)	store vector to an arbitrary memory location
cond(b, x, y)	based on values in b, select scalars from x (if true) or y (if false)
first_scalar(x)	retrieve the value of the first scalar in vector
reduce(x, f)	apply reduction f to x, storing the result in each scalar

where:

• x, y are values of type T
• N is a compile-time constant of type unsigned int
• ptr is a value of type scalar_t*
• step is a value of type int
• b is a value of type vec_b
• f is a function template or function object with a templated operator() that has a signature equivalent to S(const S&, const S&), where S is a template parameter.