MaxRegImpl.h

Go to the documentation of this file.

// MaxRegImpl.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_MAXREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_MAXREGIMPL_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/SSE2/Sse2Types.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/MaxReg.h>
#include <IPSDKUtil/InstructionSet/Logical/detail/BitwiseSelectReg.h>
#include <IPSDKUtil/InstructionSet/Comparison/detail/IsLessReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#include <boost/mpl/contains.hpp>
#include <boost/mpl/vector.hpp>

namespace ipsdk {
namespace simd {
namespace detail {



template <typename T>
typename Sse2Type<T>::Type
MaxReg<
    eInstructionSet::eIS_Sse2,
    T,
    typename boost::enable_if<typename boost::mpl::not_<typename boost::mpl::contains<SpecificTypesForSSE2MaxRegImpl, T>::type>::type>::type
>::act(const typename Sse2Type<T>::Type& in1,
        const typename Sse2Type<T>::Type& in2)
{
    return BitwiseSelectReg<eInstructionSet::eIS_Sse2, T>::act(
        IsLessReg<eInstructionSet::eIS_Sse2, T>::act(in1, in2),
        in2, in1);
}

template <typename T>
void
MaxReg<
    eInstructionSet::eIS_Sse2,
    T,
    typename boost::enable_if<typename boost::mpl::not_<typename boost::mpl::contains<SpecificTypesForSSE2MaxRegImpl, T>::type>::type>::type
>::act(const typename Sse2Type<T>::Type& in1,
        const typename Sse2Type<T>::Type& in2,
        typename Sse2Type<T>::Type& out)
{
    typename Sse2Type<T>::Type lessReg;
    IsLessReg<eInstructionSet::eIS_Sse2, T>::act(in1, in2, lessReg);
    BitwiseSelectReg<eInstructionSet::eIS_Sse2, T>::act(
        lessReg, in2, in1, out);
}

Sse2Type<ipUInt8>::Type
MaxReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(
    const Sse2Type<ipUInt8>::Type& in1,
    const Sse2Type<ipUInt8>::Type& in2)
{
    return _mm_max_epu8(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(
    const Sse2Type<ipUInt8>::Type& in1,
    const Sse2Type<ipUInt8>::Type& in2,
    Sse2Type<ipUInt8>::Type& out)
{
    out = _mm_max_epu8(in1, in2);
}

Sse2Type<ipUInt16>::Type
MaxReg<eInstructionSet::eIS_Sse2, ipInt16>::act(
    const Sse2Type<ipUInt16>::Type& in1,
    const Sse2Type<ipUInt16>::Type& in2)
{
    return _mm_max_epi16(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Sse2, ipInt16>::act(
    const Sse2Type<ipUInt16>::Type& in1,
    const Sse2Type<ipUInt16>::Type& in2,
    Sse2Type<ipUInt16>::Type& out)
{
    out = _mm_max_epi16(in1, in2);
}

Sse2Type<ipReal32>::Type
MaxReg<eInstructionSet::eIS_Sse2, ipReal32>::act(
    const Sse2Type<ipReal32>::Type& in1,
    const Sse2Type<ipReal32>::Type& in2)
{
    return _mm_max_ps(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Sse2, ipReal32>::act(
    const Sse2Type<ipReal32>::Type& in1,
    const Sse2Type<ipReal32>::Type& in2,
    Sse2Type<ipReal32>::Type& out)
{
    out = _mm_max_ps(in1, in2);
}


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_MAXREGIMPL_H__