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_simd and basic_simd_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 simd and simd_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

Traits

Math functions

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

Bit manipulation functions

All bit manipulation functions in <bit> are overloaded for basic_simd.

Implementation details

ABI tags

The data-parallel types basic_simd and basic_simd_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

1) /*simd-size-type*/ is an alias for for a signed integer type. The implementation is free to choose any signed integer type.

2) /*integer-from*/<Bytes> is an alias for a signed integer type T such that sizeof(T) equals Bytes.

3) /*simd-size-v*/<T, Abi> denotes the width of the enabled specialization basic_simd<T, Abi>, or 0 otherwise.

4) If T denotes std::datapar::basic_simd_mask<Bytes, Abi>, /*mask-element-size*/<T> equals Bytes.

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;

template< class V >
concept /*simd-floating-point*/ =
    std::same_as<V,
                 std::datapar::basic_simd<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>> || ...);

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

12) /*native-abi*/<T> is an implementation-defined alias for an ABI tag. This is the primary ABI tag to use for efficient explicit vectorization. As a result, basic_simd<T, /*native-abi*/<T>> is an enabled specialization.

14-16) These tag types are used as a template argument of std::datapar::flags. See load and store flags for their corresponding uses.

Notes

Example

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

namespace dp = std::datapar;

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 dp::basic_simd<int, A> my_abs(dp::basic_simd<int, A> x)
{
    return dp::select(x < 0, -x, x);
}

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

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

    constexpr dp::simd<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

External links