FmaddReg.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_FMADDREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_FMADDREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/InstructionSetTypes.h>
#include <IPSDKUtil/InstructionSet/AVX512/Avx512Types.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/FmaddReg.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/AddReg.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/MulReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
struct FmaddReg<
    eInstructionSet::eIS_Avx512,
    T,
    typename boost::enable_if<typename boost::is_integral<T>::type>::type>
{
    static
    IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        const typename Avx512Type<T>::Type& in3)
    {
        return AddReg<eInstructionSet::eIS_Avx512, T>::act(
            MulReg<eInstructionSet::eIS_Avx512, T>::act(in1, in2),
            in3);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        const typename Avx512Type<T>::Type& in3,
        typename Avx512Type<T>::Type& out)
    {
        typename Avx512Type<T>::Type prod;
        MulReg<eInstructionSet::eIS_Avx512, T>::act(in1, in2, prod);
        AddReg<eInstructionSet::eIS_Avx512, T>::act(prod, in3, out);
    }
};

template <>
struct FmaddReg<eInstructionSet::eIS_Avx512, ipReal32>
{
    static
    IPSDK_FORCEINLINE
    Avx512Type<ipReal32>::Type
    act(const Avx512Type<ipReal32>::Type& in1,
        const Avx512Type<ipReal32>::Type& in2,
        const Avx512Type<ipReal32>::Type& in3)
    {
        return _mm512_fmadd_ps(in1, in2, in3);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal32>::Type& in1,
        const Avx512Type<ipReal32>::Type& in2,
        const Avx512Type<ipReal32>::Type& in3,
        Avx512Type<ipReal32>::Type& out)
    {
        out = _mm512_fmadd_ps(in1, in2, in3);
    }
};

template <>
struct FmaddReg<eInstructionSet::eIS_Avx512, ipReal64>
{
    static
    IPSDK_FORCEINLINE
    Avx512Type<ipReal64>::Type
    act(const Avx512Type<ipReal64>::Type& in1,
        const Avx512Type<ipReal64>::Type& in2,
        const Avx512Type<ipReal64>::Type& in3)
    {
        return _mm512_fmadd_pd(in1, in2, in3);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal64>::Type& in1,
        const Avx512Type<ipReal64>::Type& in2,
        const Avx512Type<ipReal64>::Type& in3,
        Avx512Type<ipReal64>::Type& out)
    {
        out = _mm512_fmadd_pd(in1, in2, in3);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_FMADDREG_H__