AddReg.h

Go to the documentation of this file.

// AddReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_ADDREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_ADDREG_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
struct AddReg<eInstructionSet::eIS_Avx512, T, 
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T)==1>::type>
{
    static IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& in1, const typename Avx512Type<T>::Type& in2)
    {
        return _mm512_add_epi8(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        out = _mm512_add_epi8(in1, in2);
    }
};

template <typename T>
struct AddReg<eInstructionSet::eIS_Avx512, T, 
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T)==2>::type>
{
    static IPSDK_FORCEINLINE
    typename AvxType<T>::Type
    act(const typename Avx512Type<T>::Type& in1, const typename Avx512Type<T>::Type& in2)
    {
        return _mm512_add_epi16(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        out = _mm512_add_epi16(in1, in2);
    }
};

template <typename T>
struct AddReg<eInstructionSet::eIS_Avx512, T, 
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T)==4>::type>
{
    static IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& in1, const typename Avx512Type<T>::Type& in2)
    {
        return _mm512_add_epi32(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        out = _mm512_add_epi32(in1, in2);
    }
};

template <typename T>
struct AddReg<eInstructionSet::eIS_Avx512, T, 
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T)==8>::type>
{
    static IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2)
    {
        return _mm512_add_epi64(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        out = _mm512_add_epi64(in1, in2);
    }
};


template <>
struct AddReg<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipReal32>::Type
    act(const Avx512Type<ipReal32>::Type& in1,
        const Avx512Type<ipReal32>::Type& in2)
    {
        return _mm512_add_ps(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal32>::Type& in1,
        const Avx512Type<ipReal32>::Type& in2,
        Avx512Type<ipReal32>::Type& out)
    {
        out = _mm512_add_ps(in1, in2);
    }
};

template <>
struct AddReg<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipReal64>::Type
    act(const Avx512Type<ipReal64>::Type& in1, const Avx512Type<ipReal64>::Type& in2)
    {
        return _mm512_add_pd(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal64>::Type& in1,
        const Avx512Type<ipReal64>::Type& in2,
        Avx512Type<ipReal64>::Type& out)
    {
        out = _mm512_add_pd(in1, in2);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_ADDREG_H__