LoadRegImpl.h

// LoadReg.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_IMPL_LOADREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_IMPL_LOADREGIMPL_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/InstructionSetTypes.h>
#include <IPSDKUtil/InstructionSet/BasePack.h>
#include <IPSDKUtil/InstructionSet/detail/LoadReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
typename AvxType<T>::Type
LoadReg<
    eInstructionSet::eIS_Avx2,
    T,
    typename boost::enable_if_c<
        boost::is_integral<T>::value>::type
    >::act(const T* pT)
{
    return _mm256_loadu_si256(
        reinterpret_cast<const typename AvxType<T>::Type*>(pT));
}

template <typename T>
void
LoadReg<
    eInstructionSet::eIS_Avx2,
    T,
    typename boost::enable_if_c<
        boost::is_integral<T>::value>::type
    >::act(typename AvxType<T>::Type& reg, const T* pT)
{
    reg = _mm256_loadu_si256(
        reinterpret_cast<const typename AvxType<T>::Type*>(pT));
}

AvxType<ipReal32>::Type
LoadReg<eInstructionSet::eIS_Avx2, ipReal32>::act(const ipReal32* pT)
{
    return LoadReg<eInstructionSet::eIS_Avx, ipReal32>::act(pT);
}

void
LoadReg<eInstructionSet::eIS_Avx2, ipReal32>::act(AvxType<ipReal32>::Type& reg, const ipReal32* pT)
{
    LoadReg<eInstructionSet::eIS_Avx, ipReal32>::act(reg, pT);
}

AvxType<ipReal64>::Type
LoadReg<eInstructionSet::eIS_Avx2, ipReal64>::act(const ipReal64* pT)
{
    return LoadReg<eInstructionSet::eIS_Avx, ipReal64>::act(pT);
}

void
LoadReg<eInstructionSet::eIS_Avx2, ipReal64>::act(
    AvxType<ipReal64>::Type& reg,
    const ipReal64* pT)
{
    LoadReg<eInstructionSet::eIS_Avx, ipReal64>::act(reg, pT);
}


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_LOADREGIMPL_H__