UnloadPack.h

Go to the documentation of this file.

// UnloadPack.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_UNLOADPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_UNLOADPACK_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


// unload implementation for AVX2 for integral types
template <typename T>
struct UnloadPack<eInstructionSet::eIS_Avx2, T, 
    typename boost::enable_if_c<sizeof(T) == 1>::type>
{
    static IPSDK_FORCEINLINE
    void act(const BasePack<ePackType::ePT_Avx, T>& pack, T* pT)
    {
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[0], pT);
    }
};

template <typename T>
struct UnloadPack<eInstructionSet::eIS_Avx2, T, 
    typename boost::enable_if_c<sizeof(T) == 2 >::type>
{
    static IPSDK_FORCEINLINE
    void act(const BasePack<ePackType::ePT_Avx, T>& pack, T* pT)
    {
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[0], pT);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[1],
            pT + AvxStep<T>::Value);
    }
};

template <typename T>
struct UnloadPack<eInstructionSet::eIS_Avx2, T, 
    typename boost::enable_if_c<sizeof(T) == 4>::type>
{
    static IPSDK_FORCEINLINE
    void act(const BasePack<ePackType::ePT_Avx, T>& pack, T* pT)
    {
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[0], pT);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[1],
            pT + AvxStep<T>::Value);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[2],
            pT + 2*AvxStep<T>::Value);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[3],
            pT + 3*AvxStep<T>::Value);
    }
};

template <typename T>
struct UnloadPack<eInstructionSet::eIS_Avx2, T, 
    typename boost::enable_if_c<sizeof(T) == 8>::type>
{
    static IPSDK_FORCEINLINE
    void act(const BasePack<ePackType::ePT_Avx, T>& pack, T* pT)
    {
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[0], pT);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[1],
            pT + AvxStep<T>::Value);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[2],
            pT + 2*AvxStep<T>::Value);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[3],
            pT + 3*AvxStep<T>::Value);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[4],
            pT + 4*AvxStep<T>::Value);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[5],
            pT + 5*AvxStep<T>::Value);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[6],
            pT + 6*AvxStep<T>::Value);
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(pack._val[7],
            pT + 7*AvxStep<T>::Value);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_UNLOADPACK_H__