SumPack.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_SSE2_SUMPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_SSE2_SUMPACK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/cast.h>
#include <IPSDKUtil/InstructionSet/detail/UnloadReg.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/SSE2/AddReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

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


template <typename T>
struct SumPack<eInstructionSet::eIS_Sse2, 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_Sse, T>& in)
    {
        typedef BasePack<ePackType::ePT_Sse, typename UpperType<T>::Type> Pack16;
        typedef AddReg<eInstructionSet::eIS_Sse2, typename UpperType<T>::Type> AddR;
        Pack16 p16 =
            cast<eInstructionSet::eIS_Sse2, typename UpperType<T>::Type>(in);

        p16._val[0] = AddR::act(p16._val[0], p16._val[1]);
        p16._val[0] = AddR::act(p16._val[0], _mm_srli_si128(p16._val[0], 8));
        p16._val[0] = AddR::act(p16._val[0], _mm_srli_si128(p16._val[0], 4));
        p16._val[0] = AddR::act(p16._val[0], _mm_srli_si128(p16._val[0], 2));

        return static_cast<typename UpperType<T>::Type>(
            _mm_extract_epi16(p16._val[0], 0));
    }
};

template <typename T>
struct SumPack<eInstructionSet::eIS_Sse2, 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_Sse, T>& in)
    {
        typedef BasePack<ePackType::ePT_Sse, typename UpperType<T>::Type> Pack32;
        typedef AddReg<eInstructionSet::eIS_Sse2, typename UpperType<T>::Type> AddR;

        Pack32 in32 =
            cast<eInstructionSet::eIS_Sse2, 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] = AddR::act(in32._val[0], _mm_srli_si128(in32._val[0], 16));
        in32._val[0] = AddR::act(in32._val[0], _mm_srli_si128(in32._val[0], 8));
        in32._val[0] = AddR::act(in32._val[0], _mm_srli_si128(in32._val[0], 4));

        typename UpperType<T>::Type tab[4];
        UnloadReg<eInstructionSet::eIS_Sse2, typename UpperType<T>::Type>::act(in32._val[0],
            tab);

        return tab[0];
    }
};

template <typename T>
struct SumPack<eInstructionSet::eIS_Sse2, 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_Sse, T>& in)
    {
        return SumPack<eInstructionSet::eIS_Sse2, typename UpperType<T>::Type>::act(
            cast<eInstructionSet::eIS_Sse2, typename UpperType<T>::Type>(in));
    }
};

template <typename T>
struct SumPack<eInstructionSet::eIS_Sse2, 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_Sse, T>& in)
    {
        BasePack<ePackType::ePT_Sse, T> res;
        res._val[0] = _mm_add_epi64(in._val[0], in._val[1]);
        res._val[0] = _mm_add_epi64(res._val[0], in._val[2]);
        res._val[0] = _mm_add_epi64(res._val[0], in._val[3]);
        res._val[0] = _mm_add_epi64(res._val[0], in._val[4]);
        res._val[0] = _mm_add_epi64(res._val[0], in._val[5]);
        res._val[0] = _mm_add_epi64(res._val[0], in._val[6]);
        res._val[0] = _mm_add_epi64(res._val[0], in._val[7]);

        res._val[0] =
            _mm_add_epi64(res._val[0], _mm_srli_si128(res._val[0], 32));
        res._val[0] =
            _mm_add_epi64(res._val[0], _mm_srli_si128(res._val[0], 16));
        res._val[0] =
            _mm_add_epi64(res._val[0], _mm_srli_si128(res._val[0], 8));

        T tab[2];
        UnloadReg<eInstructionSet::eIS_Sse2, T>::act(res._val[0],
            tab);

        return tab[0];
    }
};

template <>
struct SumPack<eInstructionSet::eIS_Sse2, ipReal64>
{
    static IPSDK_FORCEINLINE
    UpperType<ipReal64>::Type
    act(const BasePack<ePackType::ePT_Sse, ipReal64>& in)
    {
        BasePack<ePackType::ePT_Sse, ipReal64> res;
        res._val[0] = _mm_add_pd(in._val[0], in._val[1]);
        res._val[0] = _mm_add_pd(res._val[0], in._val[2]);
        res._val[0] = _mm_add_pd(res._val[0], in._val[3]);
        res._val[0] = _mm_add_pd(res._val[0], in._val[4]);
        res._val[0] = _mm_add_pd(res._val[0], in._val[5]);
        res._val[0] = _mm_add_pd(res._val[0], in._val[6]);
        res._val[0] = _mm_add_pd(res._val[0], in._val[7]);

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

template <>
struct SumPack<eInstructionSet::eIS_Sse2, ipReal32>
{
    static IPSDK_FORCEINLINE
    UpperType<ipReal32>::Type
    act(const BasePack<ePackType::ePT_Sse, ipReal32>& in)
    {
        return SumPack<eInstructionSet::eIS_Sse2, ipReal64>::act(
            cast<eInstructionSet::eIS_Sse2, ipReal64>(in));
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_SSE2_SUMPACK_H__