ShiftRightReg.h

Go to the documentation of this file.

// ShiftRightReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_SSE2_SHIFTRIGHTREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_SSE2_SHIFTRIGHTREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/Tools/ForceInline.h>
#include <IPSDKUtil/InstructionSet/detail/CastReg.h>
#include <IPSDKUtil/InstructionSet/Logical/detail/ShiftRightReg.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <>
struct ShiftRightReg<eInstructionSet::eIS_Sse2, ipUInt16>
{
    static IPSDK_FORCEINLINE
    Sse2Type<ipUInt16>::Type
    act(const Sse2Type<ipUInt16>::Type& in,
        ipUInt8 nbBits)
    {
        return _mm_srli_epi16(in, nbBits);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipUInt16>::Type& in,
        ipUInt8 nbBits,
        Sse2Type<ipUInt16>::Type& out)
    {
        out = _mm_srli_epi16(in, nbBits);
    }
};

template <>
struct ShiftRightReg<eInstructionSet::eIS_Sse2, ipInt16>
{
    static IPSDK_FORCEINLINE
    Sse2Type<ipInt16>::Type
    act(const Sse2Type<ipInt16>::Type& in,
        ipUInt8 nbBits)
    {
        return _mm_srai_epi16(in, nbBits);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipInt16>::Type& in,
        ipUInt8 nbBits,
        Sse2Type<ipInt16>::Type& out)
    {
        out = _mm_srai_epi16(in, nbBits);
    }
};

template <>
struct ShiftRightReg<eInstructionSet::eIS_Sse2, ipUInt32>
{
    static IPSDK_FORCEINLINE
    Sse2Type<ipUInt32>::Type
    act(const Sse2Type<ipUInt32>::Type& in,
        ipUInt8 nbBits)
    {
        return _mm_srli_epi32(in, nbBits);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipUInt32>::Type& in,
        ipUInt8 nbBits,
        Sse2Type<ipUInt32>::Type& out)
    {
        out = _mm_srli_epi32(in, nbBits);
    }
};

template <>
struct ShiftRightReg<eInstructionSet::eIS_Sse2, ipInt32>
{
    static IPSDK_FORCEINLINE
    Sse2Type<ipInt32>::Type
    act(const Sse2Type<ipInt32>::Type& in,
        ipUInt8 nbBits)
    {
        return _mm_srai_epi32(in, nbBits);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipInt32>::Type& in,
        ipUInt8 nbBits,
        Sse2Type<ipInt32>::Type& out)
    {
        out = _mm_srai_epi32(in, nbBits);
    }
};

template <>
struct ShiftRightReg<eInstructionSet::eIS_Sse2, ipUInt8>
{
    static IPSDK_FORCEINLINE
    Sse2Type<ipUInt8>::Type
    act(const Sse2Type<ipUInt8>::Type& in,
        ipUInt8 nbBits)
    {
        Sse2Type<ipUInt16>::Type in16l, in16h;
        CastReg<eInstructionSet::eIS_Sse2, ipUInt8, ipUInt16>::act(in, in16l, in16h);
        ShiftRightReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(in16l, nbBits, in16l);
        ShiftRightReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(in16h, nbBits, in16h);
        Sse2Type<ipUInt8>::Type out;
        CastReg<eInstructionSet::eIS_Sse2, ipUInt16, ipUInt8>::act(in16l, in16h, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipUInt8>::Type& in,
        ipUInt8 nbBits,
        Sse2Type<ipUInt8>::Type& out)
    {
        Sse2Type<ipUInt16>::Type in16l, in16h;
        CastReg<eInstructionSet::eIS_Sse2, ipUInt8, ipUInt16>::act(in, in16l, in16h);
        ShiftRightReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(in16l, nbBits, in16l);
        ShiftRightReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(in16h, nbBits, in16h);
        CastReg<eInstructionSet::eIS_Sse2, ipUInt16, ipUInt8>::act(in16l, in16h, out);
    }
};

template <>
struct ShiftRightReg<eInstructionSet::eIS_Sse2, ipInt8>
{
    static IPSDK_FORCEINLINE
    Sse2Type<ipInt8>::Type
    act(const Sse2Type<ipInt8>::Type& in,
        ipUInt8 nbBits)
    {
        Sse2Type<ipInt16>::Type in16l, in16h;
        CastReg<eInstructionSet::eIS_Sse2, ipInt8, ipInt16>::act(in, in16l, in16h);
        ShiftRightReg<eInstructionSet::eIS_Sse2, ipInt16>::act(in16l, nbBits, in16l);
        ShiftRightReg<eInstructionSet::eIS_Sse2, ipInt16>::act(in16h, nbBits, in16h);
        Sse2Type<ipInt8>::Type out;
        CastReg<eInstructionSet::eIS_Sse2, ipInt16, ipInt8>::act(in16l, in16h, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipInt8>::Type& in,
        ipUInt8 nbBits,
        Sse2Type<ipInt8>::Type& out)
    {
        Sse2Type<ipInt16>::Type in16l, in16h;
        CastReg<eInstructionSet::eIS_Sse2, ipInt8, ipInt16>::act(in, in16l, in16h);
        ShiftRightReg<eInstructionSet::eIS_Sse2, ipInt16>::act(in16l, nbBits, in16l);
        ShiftRightReg<eInstructionSet::eIS_Sse2, ipInt16>::act(in16h, nbBits, in16h);
        return CastReg<eInstructionSet::eIS_Sse2, ipInt16, ipInt8>::act(in16l, in16h, out);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_SSE2_SHIFTRIGHTPACK_H__