MaxPack.h

Go to the documentation of this file.

// MaxPack.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX512_MAXPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX512_MAXPACK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/cast.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {
namespace reduction {


template <typename T>
struct MaxPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 1>::type>
{    
    static IPSDK_FORCEINLINE
    T
    act(const BasePack<ePackType::ePT_Avx512, T>& in)
    {
        typedef typename UpperType<typename UpperType<T>::Type>::Type UpperT;
        BasePack<ePackType::ePT_Avx512, UpperT> inUpper;
        cast<eInstructionSet::eIS_Avx512>(in, inUpper);
        return static_cast<T>(reduction::MaxPack<eInstructionSet::eIS_Avx512, UpperT>::act(inUpper));
    }
};

template <typename T>
struct MaxPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 2>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const BasePack<ePackType::ePT_Avx512, T>& in)
    {
        typedef typename UpperType<T>::Type UpperT;
        BasePack<ePackType::ePT_Avx512, UpperT> inUpper;
        cast<eInstructionSet::eIS_Avx512>(in, inUpper);
        return static_cast<T>(reduction::MaxPack<eInstructionSet::eIS_Avx512, UpperT>::act(inUpper));
    }
};

template <typename T>
struct MaxPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 4>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const BasePack<ePackType::ePT_Avx512, T>& in)
    {
        const typename Avx512Type<T>::Type res0 =
            detail::MaxReg<eInstructionSet::eIS_Avx512, T>::act(
            in._val[0], in._val[1]);
        const typename Avx512Type<T>::Type res1 =
            detail::MaxReg<eInstructionSet::eIS_Avx512, T>::act(
            res0, in._val[2]);
        const typename Avx512Type<T>::Type res =
            detail::MaxReg<eInstructionSet::eIS_Avx512, T>::act(
            res1, in._val[3]);
        return reduction::MaxReg<eInstructionSet::eIS_Avx512, T>::act(res);
    }
};

template <typename T>
struct MaxPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 8>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const BasePack<ePackType::ePT_Avx512, T>& in)
    {
        typedef detail::MaxReg<eInstructionSet::eIS_Avx512, T> MaxR;
        typename Avx512Type<T>::Type res = MaxR::act(in._val[0], in._val[1]);
        res = MaxR::act(res, in._val[2]);
        res = MaxR::act(res, in._val[3]);
        res = MaxR::act(res, in._val[4]);
        res = MaxR::act(res, in._val[5]);
        res = MaxR::act(res, in._val[6]);
        res = MaxR::act(res, in._val[7]);
        
        return reduction::MaxReg<eInstructionSet::eIS_Avx512, T>::act(res);
    }
};

template <>
struct MaxPack<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    ipReal32
    act(const BasePack<ePackType::ePT_Avx512, ipReal32>& in)
    {
        typedef detail::MaxReg<eInstructionSet::eIS_Avx512, ipReal32> MaxR;
        typedef Avx512Type<ipReal32>::Type RegF;
        const RegF res0 = MaxR::act(in._val[0], in._val[1]);
        const RegF res1 = MaxR::act(res0, in._val[2]);
        const RegF res = MaxR::act(res1, in._val[3]);
        return reduction::MaxReg<eInstructionSet::eIS_Avx512, ipReal32>::act(res);
    }
};

template <>
struct MaxPack<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
    ipReal64
    act(const BasePack<ePackType::ePT_Avx512, ipReal64>& in)
    {
        typedef detail::MaxReg<eInstructionSet::eIS_Avx512, ipReal64> MaxR;
        Avx512Type<ipReal64>::Type res = MaxR::act(in._val[0], in._val[1]);
        res = MaxR::act(res, in._val[2]);
        res = MaxR::act(res, in._val[3]);
        res = MaxR::act(res, in._val[4]);
        res = MaxR::act(res, in._val[5]);
        res = MaxR::act(res, in._val[6]);
        res = MaxR::act(res, in._val[7]);

        return reduction::MaxReg<eInstructionSet::eIS_Avx512, ipReal64>::act(res);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX512_MAXREGIMPL_H__