FmaddReg.h

Go to the documentation of this file.

// FmaddReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_FMA3_FMADDREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_FMA3_FMADDREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/InstructionSetTypes.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/FmaddReg.h>
#include <IPSDKUtil/InstructionSet/AVX/AvxTypes.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#if defined __linux__
#include <fmaintrin.h>
#endif

namespace ipsdk {
namespace simd {
namespace detail {


template <>
struct FmaddReg<eInstructionSet::eIS_Fma3, ipReal32>
{
    static
    IPSDK_FORCEINLINE
    AvxType<ipReal32>::Type
    act(const AvxType<ipReal32>::Type& in1,
        const AvxType<ipReal32>::Type& in2,
        const AvxType<ipReal32>::Type& in3)
    {
        return _mm256_fmadd_ps(in1, in2, in3);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal32>::Type& in1,
        const AvxType<ipReal32>::Type& in2,
        const AvxType<ipReal32>::Type& in3,
        AvxType<ipReal32>::Type& out)
    {
        out = _mm256_fmadd_ps(in1, in2, in3);
    }
};

template <>
struct FmaddReg<eInstructionSet::eIS_Fma3, ipReal64>
{
    static
    IPSDK_FORCEINLINE
    AvxType<ipReal64>::Type
    act(const AvxType<ipReal64>::Type& in1,
        const AvxType<ipReal64>::Type& in2,
        const AvxType<ipReal64>::Type& in3)
    {
        return _mm256_fmadd_pd(in1, in2, in3);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal64>::Type& in1,
        const AvxType<ipReal64>::Type& in2,
        const AvxType<ipReal64>::Type& in3,
        AvxType<ipReal64>::Type& out)
    {
        out = _mm256_fmadd_pd(in1, in2, in3);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_FMA3_FMADDREG_H__