MaxReg.h

Go to the documentation of this file.

// MaxReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX2_MAXREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX2_MAXREG_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 MaxReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T) == 1>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const typename AvxType<T>::Type& in)
    {
        T unloaded[32];
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(in, unloaded);
        T tmax = unloaded[0];
        for(int i=1; i<32; ++i) {
            tmax = (unloaded[i] > tmax ? unloaded[i] : tmax);
        }
        return tmax;
    }
};

template <typename T>
struct MaxReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const typename AvxType<T>::Type& in)
    {
        T unloaded[16];
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(in, unloaded);
        T tmax = unloaded[0];
        for(int i=1; i<16; ++i) {
            tmax = (unloaded[i] > tmax ? unloaded[i] : tmax);
        }
        return tmax;
    }
};

template <typename T>
struct MaxReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T) == 4>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const typename AvxType<T>::Type& in)
    {
        T unloaded[8];
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(in, unloaded);
        T tmax = unloaded[0];
        for(int i=1; i<8; ++i) {
            tmax = (unloaded[i] > tmax ? unloaded[i] : tmax);
        }
        return tmax;
    }
};

template <typename T>
struct MaxReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<sizeof(T) == 8>::type>
{
    static IPSDK_FORCEINLINE
    T
    act(const typename AvxType<T>::Type& in)
    {
        T unloaded[4];
        UnloadReg<eInstructionSet::eIS_Avx2, T>::act(in, unloaded);
        T tmax = unloaded[0];
        for(int i=1; i<4; ++i) {
            tmax = (unloaded[i] > tmax ? unloaded[i] : tmax);
        }
        return tmax;
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_REDUCTION_DETAIL_AVX2_MAXREGIMPL_H__