SumPack.h

Go to the documentation of this file.

// SumPack.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX2_SUMPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX2_SUMPACK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/cast.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/AVX2/AddReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

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


template <typename T>
struct SumPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 1>::type>
{
    static IPSDK_FORCEINLINE
    typename UpperType<T>::Type
    act(const BasePack<ePackType::ePT_Avx, T>& in)
    {
        typedef BasePack<ePackType::ePT_Avx, typename UpperType<T>::Type> Pack16;
        typedef AddReg<eInstructionSet::eIS_Avx2, typename UpperType<T>::Type> AddR;
        Pack16 in16 = cast<eInstructionSet::eIS_Avx2, typename UpperType<T>::Type>(in);

        in16._val[0] = AddR::act(in16._val[0], in16._val[1]);
        in16._val[0] = _mm256_hadd_epi16(in16._val[0], in16._val[0]);
        in16._val[0] = _mm256_hadd_epi16(in16._val[0], in16._val[0]);
        in16._val[0] = _mm256_hadd_epi16(in16._val[0], in16._val[0]);

        ipInt16 tab[16];
        _mm256_storeu_si256(reinterpret_cast<typename AvxType<T>::Type*>(tab), in16._val[0]);
        return static_cast<typename UpperType<T>::Type>(tab[0] + tab[8]);
    }
};

template <typename T>
struct SumPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 2>::type>
{
    static IPSDK_FORCEINLINE
    typename UpperType<T>::Type
    act(const BasePack<ePackType::ePT_Avx, T>& in)
    {
        typedef BasePack<ePackType::ePT_Avx, typename UpperType<T>::Type> Pack32;
        typedef AddReg<eInstructionSet::eIS_Avx2, typename UpperType<T>::Type> AddR;

        Pack32 in32 =
            cast<eInstructionSet::eIS_Avx2, typename UpperType<T>::Type>(in);

        in32._val[0] = AddR::act(in32._val[0], in32._val[1]);
        in32._val[0] = AddR::act(in32._val[0], in32._val[2]);
        in32._val[0] = AddR::act(in32._val[0], in32._val[3]);

        in32._val[0] = _mm256_hadd_epi32(in32._val[0], in32._val[0]);
        in32._val[0] = _mm256_hadd_epi32(in32._val[0], in32._val[0]);
        
        ipInt32 tab[16];
        _mm256_storeu_si256(reinterpret_cast<typename AvxType<T>::Type*>(tab),
            in32._val[0]);
        
        return static_cast<typename UpperType<T>::Type>(tab[0] + tab[4]);
    }
};

template <typename T>
struct SumPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 4>::type>
{
    static IPSDK_FORCEINLINE
    typename UpperType<T>::Type
    act(const BasePack<ePackType::ePT_Avx, T>& in)
    {
        return SumPack<eInstructionSet::eIS_Avx2, typename UpperType<T>::Type>::act(
            cast<eInstructionSet::eIS_Avx2, typename UpperType<T>::Type>(in));
    }
};

template <typename T>
struct SumPack<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T) == 8>::type>
{
    static IPSDK_FORCEINLINE
    typename UpperType<T>::Type
    act(const BasePack<ePackType::ePT_Avx, T>& in)
    {
        BasePack<ePackType::ePT_Avx, T> res;

        res._val[0] = _mm256_add_epi64(in._val[0], in._val[1]);
        res._val[0] = _mm256_add_epi64(res._val[0], in._val[2]);
        res._val[0] = _mm256_add_epi64(res._val[0], in._val[3]);
        res._val[0] = _mm256_add_epi64(res._val[0], in._val[4]);
        res._val[0] = _mm256_add_epi64(res._val[0], in._val[5]);
        res._val[0] = _mm256_add_epi64(res._val[0], in._val[6]);
        res._val[0] = _mm256_add_epi64(res._val[0], in._val[7]);

        res._val[0] =
            _mm256_add_epi64(res._val[0], _mm256_srli_si256(res._val[0], 8));

        ipInt64 tab[16];
        _mm256_storeu_si256(
            reinterpret_cast<typename AvxType<T>::Type*>(tab), res._val[0]);
        
        return tab[0] + tab[2];
    }
};

template <>
struct SumPack<eInstructionSet::eIS_Avx2, ipReal64>
{
    static IPSDK_FORCEINLINE
    UpperType<ipReal64>::Type
    act(const BasePack<ePackType::ePT_Avx, ipReal64>& in)
    {
        BasePack<ePackType::ePT_Avx, ipReal64> res;
        res._val[0] = _mm256_add_pd(in._val[0], in._val[1]);
        res._val[0] = _mm256_add_pd(res._val[0], in._val[2]);
        res._val[0] = _mm256_add_pd(res._val[0], in._val[3]);
        res._val[0] = _mm256_add_pd(res._val[0], in._val[4]);
        res._val[0] = _mm256_add_pd(res._val[0], in._val[5]);
        res._val[0] = _mm256_add_pd(res._val[0], in._val[6]);
        res._val[0] = _mm256_add_pd(res._val[0], in._val[7]);

        UpperType<ipReal64>::Type tab[4];
        _mm256_storeu_pd(tab, res._val[0]);
        
        return tab[0] + tab[1] + tab[2] + tab[3];
    }
};

template <>
struct SumPack<eInstructionSet::eIS_Avx2, ipReal32>
{
    static IPSDK_FORCEINLINE
    UpperType<ipReal32>::Type
    act(const BasePack<ePackType::ePT_Avx, ipReal32>& in)
    {
        return SumPack<eInstructionSet::eIS_Avx2, ipReal64>::act(
            cast<eInstructionSet::eIS_Avx2, ipReal64>(in));
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX2_SUMPACK_H__