UnloadReg.h

Go to the documentation of this file.

// UnloadReg.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_UNLOADREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_UNLOADREG_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


// unload implementation for SSE2 for integral types
template <typename T>
struct UnloadReg<eInstructionSet::eIS_Sse2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value>::type>
{
    static IPSDK_FORCEINLINE
    void act(const typename Sse2Type<T>::Type& reg, T* pT)
    {
        _mm_storeu_si128(reinterpret_cast<typename Sse2Type<T>::Type*>(pT),
            reg);
    }
};

// unload implementation for SSE2 for real32 types
template <>
struct UnloadReg<eInstructionSet::eIS_Sse2, ipReal32>
{
    static IPSDK_FORCEINLINE
    void act(const Sse2Type<ipReal32>::Type& reg, ipReal32* pT)
    {
        _mm_storeu_ps(pT, reg);
    }
};

// unload implementation for SSE2 for real64 types
template <>
struct UnloadReg<eInstructionSet::eIS_Sse2, ipReal64>
{
    static IPSDK_FORCEINLINE
    void act(const Sse2Type<ipReal64>::Type& reg, ipReal64* pT)
    {
        _mm_storeu_pd(pT, reg);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_UNLOADREG_H__