SumReg.h

Go to the documentation of this file.

// SumReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX512_SUMREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX512_SUMREG_H__

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

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

/*
template <typename T>
struct SumReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 1>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const typename Avx512Type<T>::Type& in)
    {
        T unloaded[64];
        UnloadReg<eInstructionSet::eIS_Avx512, T>::act(in, unloaded);
        T tsum = unloaded[0];
        for(int i=1; i<64; ++i) {
            tsum += unloaded[i] ;
        }
        return tsum;
    }
};

template <typename T>
struct SumReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const typename Avx512Type<T>::Type& in)
    {
        T unloaded[32];
        UnloadReg<eInstructionSet::eIS_Avx512, T>::act(in, unloaded);
        T tsum = unloaded[0];
        for(int i=1; i<32; ++i) {
            tsum += unloaded[i];
        }
        return tsum;
    }
};*/

template <>
struct SumReg<eInstructionSet::eIS_Avx512, ipInt32>
{
    static IPSDK_FORCEINLINE
    ipInt32
    act(const Avx512Type<ipInt32>::Type& in)
    {
        return _mm512_reduce_add_epi32(in);
    }
};

template <>
struct SumReg<eInstructionSet::eIS_Avx512, ipUInt32>
{
    static IPSDK_FORCEINLINE
    ipUInt32
    act(const Avx512Type<ipUInt32>::Type& in)
    {
        return _mm512_reduce_add_epi32(in);
    }
};

template <>
struct SumReg<eInstructionSet::eIS_Avx512, ipInt64>
{
    static IPSDK_FORCEINLINE
    ipInt64
    act(const Avx512Type<ipInt64>::Type& in)
    {
        return _mm512_reduce_add_epi64(in);
    }
};

template <>
struct SumReg<eInstructionSet::eIS_Avx512, ipUInt64>
{
    static IPSDK_FORCEINLINE
    ipUInt64
    act(const Avx512Type<ipUInt64>::Type& in)
    {
        return _mm512_reduce_add_epi64(in);
    }
};

template <>
struct SumReg<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    ipReal32
    act(const Avx512Type<ipReal32>::Type& in)
    {
        return _mm512_reduce_add_ps(in);
    }
};

template <>
struct SumReg<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
    ipReal64
    act(const Avx512Type<ipReal64>::Type& in)
    {
        return _mm512_reduce_add_pd(in);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX512_SUMREG_H__