std::experi...:resize_simd(3)

📦 version 2024.07.05, in nixpkgs unstable, opensuse-leap 16.0

💅 NAME 💅

okay so this page is std::experi...:resize_simd(3) — that's the header title, ellipsis and all, bestie — from the C++ Standard Libary manual (yes, “Libary”, not a typo from me 💀). The NAME line is literally:

std::experimental::rebind_simd,std::experimental::resize_simd — std::experimental::rebind_simd,std::experimental::resize_simd

It's giving template-alias realness for the parallelism TS v2 girlies 💅✨

✨ Synopsis ✨

Defined in header <experimental/simd>:

template< class T, class V >          (1) (parallelism TS v2)
struct rebind_simd;
template< int N, class V >            (2) (parallelism TS v2)
struct resize_simd;

These bad boys create a simd or simd_mask type with a different element type or size. The new type likely uses an ABI tag type different from V::abi_type — yes, “likely” is literally the word on the page 👀.

(1) Changes the element type to T and keeps the size unchanged.
(2) Changes the size to N and keeps the element type unchanged.

🧚 Template parameters 🧚

T — the new element type; an arithmetic type other than bool
N — the new number of elements
V — a simd or simd_mask type

💖 Member types 💖

Name → Definition tea ☕: type is a simd or simd_mask type with a different element type (1) or size (2).

🦋 Helper types 🦋

The helper aliases are here and they are iconic:

template< class T, class V >                             (parallelism TS v2)
using rebind_simd_t = typename rebind_simd<T, V>::type;
template< int N, class V >                               (parallelism TS v2)
using resize_simd_t = typename resize_simd<N, V>::type;

🍓 Example 🍓

The original says // Run this code — so run it, bestie 😭:

#include <experimental/simd>
#include <iostream>

namespace stdx = std::experimental;
using floatv = stdx::native_simd<float>;

// use double precision internally
floatv dp(floatv x)
{
    using doublev = stdx::rebind_simd_t<double, floatv>;
    return stdx::static_simd_cast<floatv>(stdx::simd_cast<doublev>(x) - 1.234);
}

template<class T>
stdx::resize_simd_t<T::size() / 2, T> partial_reduction(T x)
{
    auto [lo, hi] = stdx::split<stdx::resize_simd_t<T::size() / 2, T>>(x);
    return lo + hi;
}

int main()
{
    floatv x([](auto i) { return 1.234f + std::numeric_limits<float>::epsilon() * i; });
    x = dp(x);
    const auto y = partial_reduction(x);
    for (unsigned i = 0; i < y.size(); ++i)
        std::cout << y[i] << ' ';
    std::cout << '\n';
}

⭐ Possible output ⭐

1.73569e-07 4.11987e-07