AssignPackImpl.h

Go to the documentation of this file.

// AssignPackImpl.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_ASSIGNPACKIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_ASSIGNPACKIMPL_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/detail/AssignPackDecl.h>
#include <IPSDKUtil/InstructionSet/detail/AVX/BasePack.h>
#include <IPSDKUtil/InstructionSet/detail/AVX/BasePackImpl.h>
#include <IPSDKUtil/InstructionSet/detail/AVX2/AssignRegDecl.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#include <boost/type_traits/is_same.hpp>
#include <boost/type_traits/is_signed.hpp>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
IPSDK_FORCEINLINE
BasePack<ePackType::ePT_Avx, T>
AssignPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T)==1>::type>::act(const T& value)
{
    BasePack<ePackType::ePT_Avx, T> out;
    out._val[0] = AssignReg<eInstructionSet::eIS_Avx2, T>::act(value);
    return out;
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T)==1>::type>::act(
    BasePack<ePackType::ePT_Avx, T>& pack, const T& value)
{
    AssignReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[0], value);
}

template <typename T>
IPSDK_FORCEINLINE
BasePack<ePackType::ePT_Avx, T>
AssignPack<eInstructionSet::eIS_Avx2, T, 
    typename boost::enable_if_c<sizeof(T)==2>::type>::act(const T& value)
{
    const typename AvxType<T>::Type value256 = AssignReg<eInstructionSet::eIS_Avx2, T>::act(value);
    BasePack<ePackType::ePT_Avx, T> out;
    out._val[0] = value256;
    out._val[1] = value256;
    return out;
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignPack<eInstructionSet::eIS_Avx2, T, 
    typename boost::enable_if_c<sizeof(T)==2>::type>::act(
    BasePack<ePackType::ePT_Avx, T>& pack, const T& value)
{
    typename AvxType<T>::Type value256;
    AssignReg<eInstructionSet::eIS_Avx2, T>::act(value256, value);
    pack._val[0] = value256;
    pack._val[1] = value256;
}

template <typename T>
IPSDK_FORCEINLINE
BasePack<ePackType::ePT_Avx, T>
AssignPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T)==4>::type>::act(const T& value)
{
    const typename AvxType<T>::Type value256 = AssignReg<eInstructionSet::eIS_Avx2, T>::act(value);
    BasePack<ePackType::ePT_Avx, T> out;
    out._val[0] = value256;
    out._val[1] = value256;
    out._val[2] = value256;
    out._val[3] = value256;
    return out;
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T)==4>::type>::act(
    BasePack<ePackType::ePT_Avx, T>& pack, const T& value)
{
    typename AvxType<T>::Type value256;
    AssignReg<eInstructionSet::eIS_Avx2, T>::act(value256, value);
    pack._val[0] = value256;
    pack._val[1] = value256;
    pack._val[2] = value256;
    pack._val[3] = value256;
}

template <typename T>
IPSDK_FORCEINLINE
BasePack<ePackType::ePT_Avx, T>
AssignPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T)==8>::type>::act(const T& value)
{
    const typename AvxType<T>::Type value256 =
        AssignReg<eInstructionSet::eIS_Avx2, T>::act(value);

    BasePack<ePackType::ePT_Avx, T> out;
    out._val[0] = value256;
    out._val[1] = value256;
    out._val[2] = value256;
    out._val[3] = value256;
    out._val[4] = value256;
    out._val[5] = value256;
    out._val[6] = value256;
    out._val[7] = value256;
    return out;
}

template <typename T>
IPSDK_FORCEINLINE
void
AssignPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T)==8>::type>::act(
    BasePack<ePackType::ePT_Avx, T>& pack, const T& value)
{
    typename AvxType<T>::Type value256;
    AssignReg<eInstructionSet::eIS_Avx2, T>::act(value256, value);
    pack._val[0] = value256;
    pack._val[1] = value256;
    pack._val[2] = value256;
    pack._val[3] = value256;
    pack._val[4] = value256;
    pack._val[5] = value256;
    pack._val[6] = value256;
    pack._val[7] = value256;
}


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

#include <IPSDKUtil/InstructionSet/detail/AVX2/AssignRegImpl.h>

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_ASSIGNPACKIMPL_H__