ScanReg.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_SCANREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_SCANREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/ScanReg.h>
#include <IPSDKUtil/InstructionSet/detail/LoadReg.h>
#include <IPSDKUtil/InstructionSet/detail/UnloadReg.h>
#include <IPSDKUtil/InstructionSet/AVX/AvxTypes.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {

template <>
struct ScanReg<eInstructionSet::eIS_Avx, ipReal32>
{
    static IPSDK_FORCEINLINE
    AvxType<ipReal32>::Type
    act(const AvxType<ipReal32>::Type& in)
    {
        __m256 out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal32>::Type& in,
        AvxType<ipReal32>::Type& out)
    {
        __m256 t0, t1;
        //shift1_AVX + add
        t0 = _mm256_permute_ps(in, _MM_SHUFFLE(2, 1, 0, 3));
        t1 = _mm256_permute2f128_ps(t0, t0, 41);
        out = _mm256_add_ps(in, _mm256_blend_ps(t0, t1, 0x11));
        //shift2_AVX + add
        t0 = _mm256_permute_ps(out, _MM_SHUFFLE(1, 0, 3, 2));
        t1 = _mm256_permute2f128_ps(t0, t0, 41);
        out = _mm256_add_ps(out, _mm256_blend_ps(t0, t1, 0x33));
        //shift3_AVX + add
        out = _mm256_add_ps(out,_mm256_permute2f128_ps(out, out, 41));
    }
};

template <>
struct ScanReg<eInstructionSet::eIS_Avx, ipReal64>
{
    static IPSDK_FORCEINLINE
    AvxType<ipReal64>::Type
    act(const AvxType<ipReal64>::Type& in)
    {
        __m256d out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal64>::Type& in,
        AvxType<ipReal64>::Type& out)
    {
        ipReal64 pIn[4];
        detail::UnloadReg<eInstructionSet::eIS_Avx, ipReal64>::act(in, pIn);
        pIn[1] += pIn[0];
        pIn[2] += pIn[1];
        pIn[3] += pIn[2];
        detail::LoadReg<eInstructionSet::eIS_Avx, ipReal64>::act(out, pIn);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_SCANREG_H__