Data-parallel types (SIMD) (since C++26)

The library provides data-parallel types and operations on these types: portable types for explicitly stating data-parallelism and structuring data through data-parallel execution resources where available, such as SIMD registers and instructions or execution units that are driven by a common instruction decoder.

The set of vectorizable types comprises:

• all standard integers and character types;
• most floating-point types including float, double, and the selected extended floating-point types: std::float16_t, std::float32_t, and std::float64_t if defined; and
• std::complex<T> where T is a vectorizable floating-point type.

A data-parallel type consists of one or more elements of an underlying vectorizable type, called the element type . The number of elements, called the width , is constant for each data-parallel type.

The data-parallel type refers to all enabled specializations of the class templates basic_vec and basic_mask.

A data-parallel object of data-parallel type behaves analogously to objects of type T. But while T stores and manipulates a single value, the data-parallel type with the element type T stores and manipulates multiple values.

Every operation on a data-parallel object acts element-wise (except for horizontal operations, such as reductions, which are clearly marked as such) applying to each element of the object or to corresponding elements of two objects. Each such application is unsequenced with respect to the others. This simple rule expresses data-parallelism and will be used by the compiler to generate SIMD instructions and/or independent execution streams.

All operations (except non-constexpr math function overloads) on data-parallel objects are constexpr: it is possible to create and use data-parallel objects in the evaluation of a constant expression.

Alias templates vec and mask are defined to allow users to specify the width to a certain size. The default width is determined by the implementation at compile-time.

Main classes

Load and store flags

Load and store operations

Casts

Algorithms

Reductions

Permutations

Traits

Math functions

All functions in <cmath> and <complex> are overloaded for basic_vec.

Bit manipulation functions

All bit manipulation functions in <bit> except for std::bit_cast are overloaded for basic_vec.

Implementation details

ABI tags

The data-parallel types basic_vec and basic_mask are associated with ABI tags . These tags are types that specify the size and binary representation of data-parallel objects. The design intends the size and binary representation to vary based on target architecture and compiler flags. The ABI tag, together with the element type, determines the width.

The ABI tag remains independent of machine instruction set selection. The chosen machine instruction set limits the usable ABI tag types. The ABI tags enable users to safely pass objects of data-parallel type across translation unit boundaries.

Exposition-only entities

using /*simd-size-type*/ = /* see description */;

template< std::size_t Bytes >
using /*integer-from*/ = /* see description */;

template< class T, class Abi >
constexpr /*simd-size-type*/ /*simd-size-v*/ = /* see description */;

template< class T >
constexpr std::size_t /*mask-element-size*/ = /* see description */;

template< class T >
concept /*constexpr-wrapper-like*/ = /* see description */;

template< class T >
using /*deduced-simd-t*/ = /* see description */;

template< class V, class T >
using /*make-compatible-simd-t*/ = /* see description */;

template< class T >
concept /*constexpr-wrapper-like*/ =
    std::convertible_to<T, decltype(T::value)> &&
    std::equality_comparable_with<T, decltype(T::value)> &&
    std::bool_constant<T() == T::value>::value &&
    std::bool_constant<static_cast<decltype(T::value)>(T()) == T::value>::value;

• decltype(x+x), if the type of x + x is an enabled specialization of basic_vec; otherwise
• void.

• /*deduced-simd-t*/<T>, if that type is not void, otherwise
• std::simd::vec<decltype(x+x),V​::​size()>.

template< class V >
concept /*simd-floating-point*/ = /* see description */;

template< class... Ts >
concept /*math-floating-point*/ = /* see description */;

template< class... Ts >
  requires /*math-floating-point*/<Ts...>
using /*math-common-simd-t*/ = /* see description */;

template< class BinaryOp, class T >
concept /*reduction-binary-operation*/ = /* see description */;

template< class V >
concept /*simd-floating-point*/ =
    std::same_as<V,
                 std::simd::basic_vec<typename V::value_type,
                 typename V::abi_type>> &&
    std::is_default_constructible_v<V> && 
    std::floating_point<typename V::value_type>;

template< class... Ts >
concept /*math-floating-point*/ =
    (/*simd-floating-point*/</*deduced-simd-t*/<Ts>> || ...);

• /*deduced-simd-t*/<T0>, if sizeof...(Ts) == 1 is true
• otherwise, std::common_type_t</*deduced-simd-t*/<T0>,/*deduced-simd-t*/<T1>>, if sizeof...(Ts) == 2 is true and /*math-floating-point*/<T0> && /*math-floating-point*/<T1> is true,
• otherwise, std::common_type_t</*deduced-simd-t*/<T0>,T1> if sizeof...(Ts) == 2 is true and /*math-floating-point*/<T0> is true,
• otherwise, std::common_type_t<T0,/*deduced-simd-t*/<T1>> if sizeof...(Ts) == 2 is true,
• otherwise, std::common_type_t</*math-common-simd-t*/<T0,T1>,TRest...>, if /*math-common-simd-t*/<T0,T1> is a valid type,
• otherwise, std::common_type_t</*math-common-simd-t*/<TRest...>,T0,T1>.

template< class BinaryOp, class T >
concept /*reduction-binary-operation*/ =
    requires (const BinaryOp binary_op, const std::simd::vec<T, 1> v) {
        { binary_op(v, v) } -> std::same_as<std::simd::vec<T, 1>>;
    };

template< class T >
using /*native-abi*/ = /* see description */;

template< class T, /*simd-size-type*/ N >
using /*deduce-abi-t*/ = /* see description */;

• /*simd-size-v*/<T,/*deduce-abi-t*/<T,N>> equals N,
• std::simd::basic_vec<T,/*deduce-abi-t*/<T,N>> is an enabled specialization, and
• std::simd::basic_mask<sizeof(T),/*deduce-abi-t*/</*integer-from*/<sizeof(T)>,N>> is an enabled specialization.

struct /*convert-flag*/;

struct /*aligned-flag*/;

template< std::size_t N >
struct /*overaligned-flag*/;

Notes

Example

#include <iostream>
#include <simd>
#include <string_view>

namespace simd = std::simd;

void println(std::string_view name, auto const& a)
{
    std::cout << name << ": ";
    for (std::size_t i{}; i != a.size(); ++i)
        std::cout << a[i] << ' ';
    std::cout << '\n';
}

template<class A>
constexpr simd::basic_vec<int, A> my_abs(simd::basic_vec<int, A> x)
{
    return simd::select(x < 0, -x, x);
}

int main()
{
    constexpr simd::vec<int> a = 1;
    println("a", a);

    constexpr simd::vec<int> b([](int i) { return i - 2; });
    println("b", b);

    constexpr auto c = a + b;
    println("c", c);

    constexpr auto d = my_abs(c);
    println("d", d);

    constexpr auto e = d * d;
    println("e", e);

    constexpr auto inner_product = simd::reduce(e);
    std::cout << "inner product: " << inner_product << '\n';

    constexpr simd::vec<double, 16> x([](int i) { return i; });
    println("x", x);
    // overloaded math functions are defined in <simd>
    println("cos²(x) + sin²(x)", std::pow(std::cos(x), 2) + std::pow(std::sin(x), 2));
}

a: 1 1 1 1 
b: -2 -1 0 1 
c: -1 0 1 2 
d: 1 0 1 2 
e: 1 0 1 4 
inner product: 6
x: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 
cos²(x) + sin²(x): 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

See also