AssignRegImpl.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_ASSIGNREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_ASSIGNREGIMPL_H__

#include <IPSDKUtil/InstructionSet/detail/AVX2/AssignRegDecl.h>
#include <IPSDKUtil/InstructionSet/detail/AVX/AssignRegDecl.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
IPSDK_FORCEINLINE
typename AvxType<T>::Type
AssignReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T) == 1>::type>::act(const T& value)
{
    return _mm256_set1_epi8(value);
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T) == 1>::type>::act(typename AvxType<T>::Type& reg, const T& value)
{
    reg = _mm256_set1_epi8(value);
}

template <typename T>
IPSDK_FORCEINLINE
typename AvxType<T>::Type
AssignReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type>::act(const T& value)
{
    return _mm256_set1_epi16(value);
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type>::act(typename AvxType<T>::Type& reg, const T& value)
{
    reg = _mm256_set1_epi16(value);
}

template <typename T>
IPSDK_FORCEINLINE
typename AvxType<T>::Type
AssignReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 4>::type>::act(const T& value)
{
    return _mm256_set1_epi32(value);
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 4>::type>::act(typename AvxType<T>::Type& reg, const T& value)
{
    reg = _mm256_set1_epi32(value);
}

template <typename T>
IPSDK_FORCEINLINE
typename AvxType<T>::Type
AssignReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 8>::type>::act(const T& value)
{
    // there's no instruction such as "_mm256_set1_epi32" for 64 bits integers,
    // so we use a work-around...
    T values[4];
    values[0] = values[1] = values[2] = values[3] = value;
    return _mm256_loadu_si256(
        reinterpret_cast<const typename AvxType<T>::Type*>(values));
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 8>::type>::act(typename AvxType<T>::Type& reg, const T& value)
{
    // there's no instruction such as "_mm256_set1_epi32" for 64 bits integers,
    // so we use a work-around...
    T values[4];
    values[0] = values[1] = values[2] = values[3] = value;
    reg = _mm256_loadu_si256(
        reinterpret_cast<const typename AvxType<T>::Type*>(values));
}

// load implementation for AVX2 for real32 types
IPSDK_FORCEINLINE
AvxType<ipReal32>::Type
AssignReg<eInstructionSet::eIS_Avx2, ipReal32>::act(const ipReal32& value)
{
    return AssignReg<eInstructionSet::eIS_Avx, ipReal32>::act(value);
}

IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx2, ipReal32>::act(AvxType<ipReal32>::Type& reg, const ipReal32& value)
{
    AssignReg<eInstructionSet::eIS_Avx, ipReal32>::act(reg, value);
}

// load implementation for AVX2 for real64 types
IPSDK_FORCEINLINE
AvxType<ipReal64>::Type
AssignReg<eInstructionSet::eIS_Avx2, ipReal64>::act(const ipReal64& value)
{
    return AssignReg<eInstructionSet::eIS_Avx, ipReal64>::act(value);
}

IPSDK_FORCEINLINE
void
AssignReg<eInstructionSet::eIS_Avx2, ipReal64>::act(AvxType<ipReal64>::Type& reg, const ipReal64& value)
{
    AssignReg<eInstructionSet::eIS_Avx, ipReal64>::act(reg, value);
}


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

#include <IPSDKUtil/InstructionSet/detail/AVX/AssignRegImpl.h>

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_ASSIGNREGIMPL_H__