simdize<T>

Detailed Description

Automatic type vectorization.

Struct Vectorization

The Vc::simdize<T> expression transforms the type T to a vectorized type. This requires the type T to be a class template instance or an arithmetic type.

Example: First, we declare a class template for a three-dimensional point. The template parameter T determines the type of the members and is float in the scalar (classical) case.

template <typename T> struct PointTemplate
{
  T x, y, z;
 
  // Declares tuple_size and makes the members accessible via get<N>(point), allowing
  // the simdize implementation to convert between Point and PointV (see below).
  Vc_SIMDIZE_INTERFACE((x, y, z));
 
  PointTemplate(T xx, T yy, T zz) : x{xx}, y{yy}, z{zz} {};
 
  // The following function will automatically be vectorized in the PointV type.
  T distance_to_origin() const {
    using std::sqrt;
    return sqrt(x * x + y * y + z * z);
  }
};

In the following we create a type alias for the scalar type, which simply means instantiating PointTemplate with float. The resulting type can then be transformed with simdize.

using Point  = PointTemplate<float>;  // A simple struct with three floats and two functions.
using PointV = Vc::simdize<Point>;    // The vectorization of Point stores three float_v and thus
                                      // float_v::size() Points.

The following shows a code example using the above Point and PointV types.

PointV pv = Point{0.f, 1.f, 2.f};  // Constructs a PointV containing PointV::size()
                                   // copies of Point{0, 1, 2}.
for (int i = 1; i < int(pv.size()); ++i) {
  assign(pv, i, {i + 0.f, i + 1.f, i + 2.f});
}
 
const Vc::float_v l = pv.distance_to_origin();
std::cout << l << '\n';
// prints [2.23607, 3.74166, 5.38516, 7.07107, 8.77496, 10.4881, 12.2066, 13.9284] with
// float_v::size() == 8
 
const Point most_distant = extract(pv, (l.max() == l).firstOne());
std::cout << '(' << most_distant.x << ", " << most_distant.y << ", " << most_distant.z << ")\n";
// prints (7, 8, 9) with float_v::size() == 8

Iterator Vectorization

Vc::simdize<Iterator> can also be used to turn an iterator type into a new iterator type with Vc::simdize<Iterator::value_type> as its value_type. Note that Vc::simdize<double> turns into Vc::Vector<double>, which makes it easy to iterate over a given container of builtin arithmetics using Vc::Vector.

void classic(const std::vector<Point> &data) {
  using It = std::vector<Point>::const_iterator;
  const It end = data.end();
  for (It it = data.begin(); it != end; ++it) {
    Point x = *it;
    do_something(x);
  }
}
 
void vectorized(const std::vector<float> &data) {
  using It = Vc::simdize<std::vector<Point>::const_iterator>;
  const It end = data.end();
  for (It it = data.begin(); it != end; ++it) {
    Vc::simdize<Point> x = *it;  // i.e. PointV
    do_something(x);
  }
}

Macros
#define	Vc_SIMDIZE_INTERFACE(MEMBERS_)

Typedefs
template<typename T , size_t N = 0, typename MT = void>
using	simdize = SimdizeDetail::simdize<T, N, MT>

Functions
template<typename S , typename T , size_t N>
Adapter< S, T, N >	shifted (const Adapter< S, T, N > &a, int shift)
	Returns a new vectorized object where each entry is shifted by shift.
template<typename S , typename T , std::size_t N>
void	swap (Adapter< S, T, N > &a, std::size_t i, S &x)
	Swaps one scalar object x with a SIMD slot at offset i in the simdized object a.
template<typename A >
void	swap (Scalar< A > &&a, typename A::scalar_type &b)
	std::swap interface to swapping one scalar object with a (virtual) reference to another object inside a vectorized object
template<typename A >
void	swap (typename A::scalar_type &b, Scalar< A > &&a)
	std::swap interface to swapping one scalar object with a (virtual) reference to another object inside a vectorized object
template<class F , class = decltype(static_cast<Scalar>(std::declval<F>()( size_t())))>
	Adapter (F &&fun)
	Generator constructor {{{.

Macro Definition Documentation

Vc_SIMDIZE_INTERFACE

#define Vc_SIMDIZE_INTERFACE

(

MEMBERS_

)

Value:

    template <std::size_t N_>                                                            \
    inline auto vc_get_()->decltype(std::get<N_>(std::tie MEMBERS_))                     \
    {                                                                                    \
        return std::get<N_>(std::tie MEMBERS_);                                          \
    }                                                                                    \
    template <std::size_t N_>                                                            \
    inline auto vc_get_() const->decltype(std::get<N_>(std::tie MEMBERS_))               \
    {                                                                                    \
        return std::get<N_>(std::tie MEMBERS_);                                          \
    }                                                                                    \
    enum : std::size_t {                                                                 \
        tuple_size = std::tuple_size<decltype(std::make_tuple MEMBERS_)>::value          \
    }

Declares functions and constants for introspection by the simdize functions. This allows e.g. conversion between scalar T and simdize<T>.

Parameters

MEMBERS_

The data members of this struct/class listed inside extra parenthesis. The extra parenthesis are required because the macro would otherwise see a variable number of arguments.

Example:

template <typename T, typename U> struct X {
  T a;
  U b;
  Vc_SIMDIZE_INTERFACE((a, b));
};

Note

You must use this macros in the public section of a class.

Definition at line 1931 of file simdize.h.

Typedef Documentation

simdize

template<typename T , size_t N = 0, typename MT = void>

using simdize = SimdizeDetail::simdize<T, N, MT>

Vectorize/Simdize the given type T.

Template Parameters

T	This type must be a class template instance where the template arguments can be recursively replaced with their vectorized variant. If the type implements a specific interface for introspection and member modification, the resulting type can easily be constructed from objects of type T and scalar objects of type T can be extracted from it.
N	This value determines the width of the vectorization. Per default it is set to 0 making the implementation choose the value considering the compilation target and the given type T.
MT	This type determines the type to be used when replacing bool with Mask<MT>. If it is set to void the implementation choosed the type as smart as possible.

See also

Vc_SIMDIZE_STRUCT, Vc_SIMDIZE_MEMBER

Definition at line 1910 of file simdize.h.

Function Documentation

shifted()

template<typename S , typename T , size_t N>

Adapter< S, T, N > shifted ( const Adapter< S, T, N > & a, int shift )

inline

Returns a new vectorized object where each entry is shifted by shift.

This basically calls Vector<T>::shifted on every entry.

Parameters

a	The object to apply the shift on.
shift	The number of entries to shift by.

Returns

a copy of a shifted by shift.

Definition at line 1069 of file simdize.h.

Referenced by Vc::SimdizeDetail::shifted().