#include <cstddef>
#include <iterator>
#include <type_traits>
#include "corecel/Macros.hh"
#include "corecel/cont/Span.hh"
#include "LdgTraits.hh"

This graph shows which files directly or indirectly include this file:

Classes
class	celeritas::LdgIterator< T >
	Iterator for read-only device data in global memory. More...

struct	celeritas::LdgValue< T >
	Wrapper struct for specializing on types supported by LdgIterator. More...

Typedefs
template<class T , std::size_t Extent = dynamic_extent>
using	celeritas::LdgSpan = Span< LdgValue< T >, Extent >
	Alias for a Span iterating over const values read using __ldg.

Functions
template<class T >
CELER_CONSTEXPR_FUNCTION T	celeritas::ldg (T const *ptr)
	Wrap the low-level CUDA/HIP "load read-only global memory" function.

template<class T >
	celeritas::LdgIterator (T *) -> LdgIterator< std::add_const_t< T > >

template<class T , std::size_t N>
CELER_CONSTEXPR_FUNCTION auto	celeritas::make_array (LdgSpan< T, N > const &s)
	Construct an array from a fixed-size span, removing LdgValue marker.


template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator== (LdgIterator< T > const &lhs, LdgIterator< T > const &rhs) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator!= (LdgIterator< T > const &lhs, LdgIterator< T > const &rhs) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator== (LdgIterator< T > const &it, std::nullptr_t) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator!= (LdgIterator< T > const &it, std::nullptr_t) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator== (std::nullptr_t, LdgIterator< T > const &it) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator!= (std::nullptr_t, LdgIterator< T > const &it) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator< (LdgIterator< T > const &lhs, LdgIterator< T > const &rhs) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator> (LdgIterator< T > const &lhs, LdgIterator< T > const &rhs) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator<= (LdgIterator< T > const &lhs, LdgIterator< T > const &rhs) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION bool	celeritas::operator>= (LdgIterator< T > const &lhs, LdgIterator< T > const &rhs) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION LdgIterator< T >	celeritas::operator+ (LdgIterator< T > const &it, typename LdgIterator< T >::difference_type const n) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION LdgIterator< T >	celeritas::operator+ (typename LdgIterator< T >::difference_type const n, LdgIterator< T > const &it) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION LdgIterator< T >	celeritas::operator- (LdgIterator< T > const &it, typename LdgIterator< T >::difference_type const n) noexcept

template<class T >
CELER_CONSTEXPR_FUNCTION auto	celeritas::operator- (LdgIterator< T > const &lhs, LdgIterator< T > const &rhs) noexcept -> typename LdgIterator< T >::difference_type

Variables
template<class T >
constexpr bool	celeritas::is_ldg_supported_v = !std::is_void_v<typename LdgTraits<T>::underlying_type>
	Whether a type is supported by __ldg.

Function Documentation

◆ ldg()

template<class T >

CELER_CONSTEXPR_FUNCTION T celeritas::ldg ( T const * ptr )

Wrap the low-level CUDA/HIP "load read-only global memory" function.

This low-level capability allows improved caching because we're promising that the data is not mem. For CUDA the load is cached in L1/texture memory, theoretically improving performance if repeatedly accessed.

Warning: The target address must be read-only for the lifetime of the kernel. This is generally true for Params data but not State data.

◆ make_array()

template<class T , std::size_t N>

CELER_CONSTEXPR_FUNCTION auto celeritas::make_array ( LdgSpan< T, N > const & s )

Construct an array from a fixed-size span, removing LdgValue marker.

Note:

make_array(Span<T,N> const&)

is not reused because:

Using this overload reads input data using __ldg
return make_array<T, N>(s)

results in segfault (gcc 11.3). This might be a compiler bug because temporary lifetime should be extended until the end of the expression and we return a copy...