LoadPack.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_LOADPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_LOADPACK_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


// load implementation for SSE2 for integral types
template <typename T>
struct LoadPack<eInstructionSet::eIS_Sse2, T, 
    typename boost::enable_if_c<sizeof(T) == 1>::type>
{
    static IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, T> act(const T* pT)
    {
        BasePack<ePackType::ePT_Sse, T> pack;
        pack._val[0] = LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT);
        return pack;
    }

    static IPSDK_FORCEINLINE
    void act(BasePack<ePackType::ePT_Sse, T>& pack, const T* pT)
    {
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[0], pT);
    }
};

template <typename T>
struct LoadPack<eInstructionSet::eIS_Sse2, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type>
{
    static IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, T> act(const T* pT)
    {
        BasePack<ePackType::ePT_Sse, T> pack;
        pack._val[0] = LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT);
        pack._val[1] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+SseStep<T>::Value);
        return pack;
    }

    static IPSDK_FORCEINLINE
    void
    act(BasePack<ePackType::ePT_Sse, T>& pack, const T* pT)
    {
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[0], pT);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[1], pT+SseStep<T>::Value);
    }
};

template <typename T>
struct LoadPack<eInstructionSet::eIS_Sse2, T,
    typename boost::enable_if_c<sizeof(T) == 4>::type>
{
    static IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, T> act(const T* pT)
    {
        BasePack<ePackType::ePT_Sse, T> regs;
        regs._val[0] = LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT);
        regs._val[1] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+SseStep<T>::Value);
        regs._val[2] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+2*SseStep<T>::Value);
        regs._val[3] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+3*SseStep<T>::Value);
        return regs;
    }

    static IPSDK_FORCEINLINE
    void act(BasePack<ePackType::ePT_Sse, T>& pack, const T* pT)
    {
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[0], pT);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[1], pT+SseStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[2], pT+2*SseStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[3], pT+3*SseStep<T>::Value);
    }
};

template <typename T>
struct LoadPack<eInstructionSet::eIS_Sse2, T,
    typename boost::enable_if_c<sizeof(T) == 8>::type>
{
    static IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, T> act(const T* pT)
    {
        BasePack<ePackType::ePT_Sse, T> regs;
        regs._val[0] = LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT);
        regs._val[1] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+SseStep<T>::Value);
        regs._val[2] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+2*SseStep<T>::Value);
        regs._val[3] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+3*SseStep<T>::Value);
        regs._val[4] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+4*SseStep<T>::Value);
        regs._val[5] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+5*SseStep<T>::Value);
        regs._val[6] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+6*SseStep<T>::Value);
        regs._val[7] =
            LoadReg<eInstructionSet::eIS_Sse2, T>::act(pT+7*SseStep<T>::Value);
        return regs;
    }

    static IPSDK_FORCEINLINE
    void
    act(BasePack<ePackType::ePT_Sse, T>& pack, const T* pT)
    {
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[0], pT);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[1], pT+SseStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[2], pT+2*SseStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[3], pT+3*SseStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[4], pT+4*SseStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[5], pT+5*SseStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[6], pT+6*SseStep<T>::Value);
        LoadReg<eInstructionSet::eIS_Sse2, T>::act(pack._val[7], pT+7*SseStep<T>::Value);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_LOADPACK_H__