SumPack.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX_SUMPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX_SUMPACK_H__

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

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

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


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX_SUMPACK_H__