AbsReg.h

Go to the documentation of this file.

// AbsReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_ABS512_ADDREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_ABS512_ADDREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/AVX512/Avx512Types.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/AbsReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <>
struct AbsReg<eInstructionSet::eIS_Avx512, ipInt8>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipInt8>::Type
    act(const Avx512Type<ipInt8>::Type& in)
    {
        return _mm512_abs_epi8(in);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipInt8>::Type& in,
        Avx512Type<ipInt8>::Type& out)
    {
        out = _mm512_abs_epi8(in);
    }
};

template <>
struct AbsReg<eInstructionSet::eIS_Avx512, ipInt16>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipInt16>::Type
    act(const Avx512Type<ipInt16>::Type& in)
    {
        return _mm512_abs_epi16(in);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipInt16>::Type& in,
        Avx512Type<ipInt16>::Type& out)
    {
        out = _mm512_abs_epi16(in);
    }
};

template <>
struct AbsReg<eInstructionSet::eIS_Avx512, ipInt32>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipInt32>::Type
    act(const Avx512Type<ipInt32>::Type& in)
    {
        return _mm512_abs_epi32(in);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipInt32>::Type& in,
        Avx512Type<ipInt32>::Type& out)
    {
        out = _mm512_abs_epi32(in);
    }
};

template <>
struct AbsReg<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipReal32>::Type
    act(const Avx512Type<ipReal32>::Type& in)
    {
        return _mm512_abs_ps(in);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal32>::Type& in,
        Avx512Type<ipReal32>::Type& out)
    {
        out = _mm512_abs_ps(in);
    }
};

template <>
struct AbsReg<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipReal64>::Type
    act(const Avx512Type<ipReal64>::Type& in)
    {
#ifdef IPSDK_WINDOWS
        return _mm512_abs_pd(in);
#else
        Avx512Type<ipReal64>::Type mzero, notmzero;
        AssignReg<eInstructionSet::eIS_Avx512, ipReal64>::act(mzero, -.0);
        BitwiseNotReg<eInstructionSet::eIS_Avx512, ipReal64>::act(mzero, notmzero);
        return BitwiseAndReg<eInstructionSet::eIS_Avx512, ipReal64>::act(in, notmzero);
#endif
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal64>::Type& in,
        Avx512Type<ipReal64>::Type& out)
    {
#ifdef IPSDK_WINDOWS
        out = _mm512_abs_pd(in);
#else
        Avx512Type<ipReal64>::Type mzero, notmzero;
        AssignReg<eInstructionSet::eIS_Avx512, ipReal64>::act(mzero, -.0);
        BitwiseNotReg<eInstructionSet::eIS_Avx512, ipReal64>::act(mzero, notmzero);
        BitwiseAndReg<eInstructionSet::eIS_Avx512, ipReal64>::act(in, notmzero, out);
#endif
    }
};

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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_ABSREG_H__