LoadPack.h

Go to the documentation of this file.

// LoadPack.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_LOADPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_LOADPACK_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>
struct LoadPack<eInstructionSet::eIS_Avx, T,
    typename boost::enable_if_c<sizeof(T) == 4>::type>
{
    static IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Avx, T> act(const T* pT)
    {
        BasePack<ePackType::ePT_Avx, T> pack;
        pack._val[0] = LoadReg<eInstructionSet::eIS_Avx, T>::act(pT);
        pack._val[1] = LoadReg<eInstructionSet::eIS_Avx, T>::act(
            pT + AvxStep<T>::Value);
        pack._val[2] = LoadReg<eInstructionSet::eIS_Avx, T>::act(
            pT + 2*AvxStep<T>::Value);
        pack._val[3] = LoadReg<eInstructionSet::eIS_Avx, T>::act(
            pT + 3*AvxStep<T>::Value);
        return pack;
    }

    static IPSDK_FORCEINLINE
    void
    act(BasePack<ePackType::ePT_Avx, T>& pack, const T* pT)
    {
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[0], pT);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[1], pT + AvxStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[2], pT + 2*AvxStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[3], pT + 3*AvxStep<T>::Value);
    }
};

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

    static IPSDK_FORCEINLINE
    void
    act(BasePack<ePackType::ePT_Avx, T>& pack, const T* pT)
    {
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[0], pT);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[1], pT + AvxStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[2], pT + 2*AvxStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[3], pT + 3*AvxStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[4], pT + 4*AvxStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[5], pT + 5*AvxStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Avx, T>::act(pack._val[6], pT + 6*AvxStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Avx, 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_AVX_LOADPACK_H__