LoadReg.h

Go to the documentation of this file.

// LoadReg.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_LOADREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_LOADREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/InstructionSetTypes.h>
#include <IPSDKUtil/InstructionSet/BasePack.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


// load implementation for AVX for ipReal32
template<>
struct LoadReg<eInstructionSet::eIS_Avx, ipReal32>
{
    static IPSDK_FORCEINLINE
    AvxType<ipReal32>::Type act(const ipReal32* pT)
    {
        return _mm256_loadu_ps(pT);
    }

    static IPSDK_FORCEINLINE
    void
    act(AvxType<ipReal32>::Type& reg, const ipReal32* pT)
    {
        reg = _mm256_loadu_ps(pT);
    }
};

// load implementation for AVX for ipReal64
template<>
struct LoadReg<eInstructionSet::eIS_Avx, ipReal64>
{
    static IPSDK_FORCEINLINE
    AvxType<ipReal64>::Type act(const ipReal64* pT)
    {
        return _mm256_loadu_pd(pT);
    }

    static IPSDK_FORCEINLINE
    void act(AvxType<ipReal64>::Type& reg, const ipReal64* pT)
    {
        reg = _mm256_loadu_pd(pT);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_LOADREG_H__