ScanReg.h

Go to the documentation of this file.

// ScanReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_SCANREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_SCANREG_H__

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

namespace ipsdk {
namespace simd {
namespace detail {

template <>
struct ScanReg<eInstructionSet::eIS_Avx2, ipReal32>
{
    static IPSDK_FORCEINLINE
    AvxType<ipReal32>::Type
    act(const AvxType<ipReal32>::Type& in)
    {
        return ScanReg<eInstructionSet::eIS_Avx, ipReal32>::act(in);
    }

    static IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal32>::Type& in,
        AvxType<ipReal32>::Type& out)
    {
        ScanReg<eInstructionSet::eIS_Avx, ipReal32>::act(in, out);
    }
};

template <>
struct ScanReg<eInstructionSet::eIS_Avx2, ipReal64>
{
    static IPSDK_FORCEINLINE
        AvxType<ipReal64>::Type
        act(const AvxType<ipReal64>::Type& in)
    {
        return ScanReg<eInstructionSet::eIS_Avx, ipReal64>::act(in);
    }

    static IPSDK_FORCEINLINE
        void
        act(const AvxType<ipReal64>::Type& in,
            AvxType<ipReal64>::Type& out)
    {
        __m256d t0, t1;
        //shift1_AVX + add
        t0 = _mm256_permute4x64_pd(in, 0x90);
        t1 = _mm256_permute2f128_pd(t0, t0, 41);
        __m256d t2;
        t2 = _mm256_blend_pd(t0, t1, 0x8);
        out = _mm256_add_pd(in, _mm256_blend_pd(t0, t1, 0x1));

        //shift2_AVX + add
        t1 = _mm256_permute2f128_pd(out, out, 41);
        out = _mm256_add_pd(out, t1);
    }
};



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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_SCANREG_H__