AddReg.h

Go to the documentation of this file.

// AddReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_ADDREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_ADDREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/AVX/AvxTypes.h>
#include <IPSDKUtil/InstructionSet/detail/AVX/Reg128_256ConvertTools.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <>
struct AddReg<eInstructionSet::eIS_Avx, ipInt32>
{
    static IPSDK_FORCEINLINE
    AvxType<ipInt32>::Type
    act(const AvxType<ipInt32>::Type& in1,
        const AvxType<ipInt32>::Type& in2)
    {
        AvxType<ipInt32>::Type out;
        act(in1, in2, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const AvxType<ipInt32>::Type& in1,
        const AvxType<ipInt32>::Type& in2,
        AvxType<ipInt32>::Type& out)
    {
        __m128i in11, in12;
        __m128i in21, in22;
        COPY_IMM_TO_XMM(in1, in11, in12);
        COPY_IMM_TO_XMM(in2, in21, in22);
        __m128i out1 = _mm_add_epi32(in11, in21);
        __m128i out2 = _mm_add_epi32(in12, in22);
        COPY_XMM_TO_IMM(out1, out2, out);
    }
};

template <>
struct AddReg<eInstructionSet::eIS_Avx, ipReal32>
{
    static IPSDK_FORCEINLINE
    const AvxType<ipReal32>::Type
    act(const AvxType<ipReal32>::Type& in1, const AvxType<ipReal32>::Type& in2)
    {
        return _mm256_add_ps(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal32>::Type& in1,
        const AvxType<ipReal32>::Type& in2,
        AvxType<ipReal32>::Type& out)
    {
        out = _mm256_add_ps(in1, in2);
    }
};

template <>
struct AddReg<eInstructionSet::eIS_Avx, ipReal64>
{
    static IPSDK_FORCEINLINE
    AvxType<ipReal64>::Type
    act(const AvxType<ipReal64>::Type& in1, const AvxType<ipReal64>::Type& in2)
    {
        return _mm256_add_pd(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal64>::Type& in1,
        const AvxType<ipReal64>::Type& in2,
        AvxType<ipReal64>::Type& out)
    {
        out = _mm256_add_pd(in1, in2);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_ADDREG_H__