AssignRegImpl.h

Go to the documentation of this file.

// AssignRegImpl.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX512_ASSIGNREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX512_ASSIGNREGIMPL_H__

#include <IPSDKUtil/InstructionSet/detail/AVX512/AssignRegDecl.h>
#include <IPSDKUtil/InstructionSet/detail/AVX512/AssignRegDecl.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
IPSDK_FORCEINLINE
typename Avx512Type<T>::Type
AssignReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 1>::type>::act(const T& value)
{
    return _mm512_set1_epi8(value);
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 1>::type>::act(typename Avx512Type<T>::Type& reg, const T& value)
{
    reg = _mm512_set1_epi8(value);
}

template <typename T>
IPSDK_FORCEINLINE
typename Avx512Type<T>::Type
AssignReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type>::act(const T& value)
{
    return _mm512_set1_epi16(value);
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type>::act(typename Avx512Type<T>::Type& reg, const T& value)
{
    reg = _mm512_set1_epi16(value);
}

template <typename T>
IPSDK_FORCEINLINE
typename Avx512Type<T>::Type
AssignReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 4>::type>::act(const T& value)
{
    return _mm512_set1_epi32(value);
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 4>::type>::act(typename Avx512Type<T>::Type& reg, const T& value)
{
    reg = _mm512_set1_epi32(value);
}

template <typename T>
IPSDK_FORCEINLINE
typename Avx512Type<T>::Type
AssignReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 8>::type>::act(const T& value)
{
    return _mm512_set1_epi64(value);
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 8>::type>::act(typename Avx512Type<T>::Type& reg, const T& value)
{
    reg = _mm512_set1_epi64(value);
}

// load implementation for AVX512 for real32 types
IPSDK_FORCEINLINE
Avx512Type<ipReal32>::Type
AssignReg<eInstructionSet::eIS_Avx512, ipReal32>::act(const ipReal32& value)
{
    return _mm512_set1_ps(value);
}

IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx512, ipReal32>::act(Avx512Type<ipReal32>::Type& reg, const ipReal32& value)
{
    reg = _mm512_set1_ps(value);
}

// load implementation for AVX512 for real64 types
IPSDK_FORCEINLINE
Avx512Type<ipReal64>::Type
AssignReg<eInstructionSet::eIS_Avx512, ipReal64>::act(const ipReal64& value)
{
    return _mm512_set1_pd(value);
}

IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx512, ipReal64>::act(Avx512Type<ipReal64>::Type& reg, const ipReal64& value)
{
    reg = _mm512_set1_pd(value);
}


} // end of namespace detail
} // end of namespace simd
} // end of namespace ipsdk

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX512_ASSIGNREGIMPL_H__