CopyImpl.h

Go to the documentation of this file.

// CopyImpl.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_COPYIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_COPYIMPL_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/detail/LoadReg.h>
#include <IPSDKUtil/InstructionSet/detail/UnloadReg.h>
#include <IPSDKUtil/InstructionSet/NbEltsPerReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
struct CopyImpl<eInstructionSet::eIS_Avx2, T> {
    static IPSDK_FORCEINLINE
    void act(T* dest, const T* src, ipUInt64 nbElts)
    {
        const ipUInt64 nbAlignedElts =
            nbElts - (nbElts%NbEltsPerReg<ePackType::ePT_Avx, T>::Value);

        typename AvxType<T>::Type reg;

        for(ipUInt64 i=0; i<nbAlignedElts; i+=NbEltsPerReg<ePackType::ePT_Avx, T>::Value) {
            LoadReg<eInstructionSet::eIS_Avx2, T>::act(reg, src+i);
            UnloadReg<eInstructionSet::eIS_Avx2, T>::act(reg, dest+i);
        }

        memcpy(reinterpret_cast<void*>(dest+nbAlignedElts),
               reinterpret_cast<const void*>(src +nbAlignedElts),
               (nbElts-nbAlignedElts)*sizeof(T));
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX2_COPYIMPL_H__