LoadRegImpl.h

Go to the documentation of this file.

// LoadRegImpl.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_LOADREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_LOADREGIMPL_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


// load implementation for SSE2 for integral types
template <typename T>
typename Sse2Type<T>::Type
LoadReg<
    eInstructionSet::eIS_Sse2,
    T,
    typename boost::enable_if_c<
        boost::is_integral<T>::value
    >::type
>::act(const T* pT)
{
    return _mm_loadu_si128(
        reinterpret_cast<const typename Sse2Type<T>::Type*>(pT));
}

template <typename T>
void
LoadReg<
    eInstructionSet::eIS_Sse2,
    T,
    typename boost::enable_if_c<
        boost::is_integral<T>::value
    >::type
>::act(typename Sse2Type<T>::Type& reg, const T* pT)
{
    reg = _mm_loadu_si128(
        reinterpret_cast<const typename Sse2Type<T>::Type*>(pT));
}

Sse2Type<ipReal32>::Type
LoadReg<eInstructionSet::eIS_Sse2, ipReal32>::act(const ipReal32* pT)
{
    return _mm_loadu_ps(pT);
}

void
LoadReg<eInstructionSet::eIS_Sse2, ipReal32>::act(
    Sse2Type<ipReal32>::Type& reg,
    const ipReal32* pT)
{
    reg = _mm_loadu_ps(pT);
}

Sse2Type<ipReal64>::Type
LoadReg<eInstructionSet::eIS_Sse2, ipReal64>::act(const ipReal64* pT)
{
    return _mm_loadu_pd(pT);
}

void
LoadReg<eInstructionSet::eIS_Sse2, ipReal64>::act(
    Sse2Type<ipReal64>::Type& reg,
    const ipReal64* pT)
{
    reg = _mm_loadu_pd(pT);
}


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_LOADREGIMPL_H__