IsNotEqualRegImpl.h

Go to the documentation of this file.

// IsNotEqualRegImpl.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX_ISNOTEQUALREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX_ISNOTEQUALREGIMPL_H__

#include <IPSDKUtil/InstructionSet/Comparison/detail/IsNotEqualReg.h>
#include <IPSDKUtil/InstructionSet/detail/AVX/Reg128_256ConvertTools.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
IPSDK_FORCEINLINE
typename RegMaskType<eInstructionSet::eIS_Avx, T>::Type
IsNotEqualReg<eInstructionSet::eIS_Avx, T,
    typename boost::enable_if<
        typename boost::mpl::and_<
            typename boost::is_integral<T>::type,
            typename boost::mpl::equal_to<boost::mpl::int_<sizeof(T)>, boost::mpl::int_<4> >::type
        >::type
    >::type
>::act(
    const typename AvxType<T>::Type& in1,
    const typename AvxType<T>::Type& in2)
{
    typename RegMaskType<eInstructionSet::eIS_Avx, T>::Type out;
    act(in1, in2, out);
    return out;
}

template <typename T>
IPSDK_FORCEINLINE
void
IsNotEqualReg<eInstructionSet::eIS_Avx, T,
    typename boost::enable_if<
        typename boost::mpl::and_<
            typename boost::is_integral<T>::type,
            typename boost::mpl::equal_to<boost::mpl::int_<sizeof(T)>, boost::mpl::int_<4> >::type
        >::type
    >::type
>::act(
    const typename AvxType<T>::Type& in1,
    const typename AvxType<T>::Type& in2,
    typename RegMaskType<eInstructionSet::eIS_Avx, T>::Type& out)
{
    __m128i in11, in12;
    __m128i in21, in22;
    COPY_IMM_TO_XMM(in1, in11, in12);
    COPY_IMM_TO_XMM(in2, in21, in22);
    typename RegMaskType<eInstructionSet::eIS_Sse2, T>::Type out1 =
        IsNotEqualReg<eInstructionSet::eIS_Sse2, T>::act(in11, in21);
    typename RegMaskType<eInstructionSet::eIS_Sse2, T>::Type out2 =
        IsNotEqualReg<eInstructionSet::eIS_Sse2, T>::act(in12, in22);
    COPY_XMM_TO_IMM(out1, out2, out);
}

IPSDK_FORCEINLINE
RegMaskType<eInstructionSet::eIS_Avx, ipReal32>::Type
IsNotEqualReg<eInstructionSet::eIS_Avx, ipReal32>::act(
    const AvxType<ipReal32>::Type& in1,
    const AvxType<ipReal32>::Type& in2)
{
    return _mm256_cmp_ps(in1, in2, _CMP_NEQ_UQ);
}

IPSDK_FORCEINLINE
void
IsNotEqualReg<eInstructionSet::eIS_Avx, ipReal32>::act(
    const AvxType<ipReal32>::Type& in1,
    const AvxType<ipReal32>::Type& in2,
    RegMaskType<eInstructionSet::eIS_Avx, ipReal32>::Type& out)
{
    out = _mm256_cmp_ps(in1, in2, _CMP_NEQ_UQ);
}

IPSDK_FORCEINLINE
RegMaskType<eInstructionSet::eIS_Avx, ipReal64>::Type
IsNotEqualReg<eInstructionSet::eIS_Avx, ipReal64>::act(
    const AvxType<ipReal64>::Type& in1,
    const AvxType<ipReal64>::Type& in2)
{
    return _mm256_cmp_pd(in1, in2, _CMP_NEQ_UQ);
}

IPSDK_FORCEINLINE
void
IsNotEqualReg<eInstructionSet::eIS_Avx, ipReal64>::act(
    const AvxType<ipReal64>::Type& in1,
    const AvxType<ipReal64>::Type& in2,
    RegMaskType<eInstructionSet::eIS_Avx, ipReal64>::Type& out)
{
    out = _mm256_cmp_pd(in1, in2, _CMP_NEQ_UQ);
}


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX_ISNOTEQUALREGIMPL_H__