IsNotEqualRegImpl.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISNOTEQUALREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISNOTEQUALREGIMPL_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


#define DEFINE_ISNOTEQUALREG_INT(nbBitsInInt, nbBitsInMask)                     \
template <>                                                                     \
struct IsNotEqualReg<eInstructionSet::eIS_Avx512, ipInt ## nbBitsInInt>         \
{                                                                               \
    static IPSDK_FORCEINLINE                                                    \
    __mmask ## nbBitsInMask                                                     \
    act(const Avx512Type<ipInt ## nbBitsInInt>::Type& in1,                      \
        const Avx512Type<ipInt ## nbBitsInInt>::Type& in2)                      \
    {                                                                           \
        return _mm512_cmp_epi ## nbBitsInInt ## _mask(in1, in2, _MM_CMPINT_NE); \
    }                                                                           \
    static IPSDK_FORCEINLINE                                                    \
    void                                                                        \
    act(const Avx512Type<ipInt ## nbBitsInInt>::Type& in1,                      \
        const Avx512Type<ipInt ## nbBitsInInt>::Type& in2,                      \
        __mmask ## nbBitsInMask &out)                                           \
    {                                                                           \
        out = _mm512_cmp_epi ## nbBitsInInt ## _mask(in1, in2, _MM_CMPINT_NE);  \
    }                                                                           \
};
    
#define DEFINE_ISNOTEQUALREG_UINT(nbBitsInInt, nbBitsInMask)                    \
template <>                                                                     \
struct IsNotEqualReg<eInstructionSet::eIS_Avx512, ipUInt ## nbBitsInInt>        \
{                                                                               \
    static IPSDK_FORCEINLINE                                                    \
    __mmask ## nbBitsInMask                                                     \
    act(const Avx512Type<ipUInt ## nbBitsInInt>::Type& in1,                     \
        const Avx512Type<ipUInt ## nbBitsInInt>::Type& in2)                     \
    {                                                                           \
        return _mm512_cmp_epu ## nbBitsInInt ## _mask(in1, in2, _MM_CMPINT_NE); \
    }                                                                           \
    static IPSDK_FORCEINLINE                                                    \
    void                                                                        \
    act(const Avx512Type<ipUInt ## nbBitsInInt>::Type& in1,                     \
        const Avx512Type<ipUInt ## nbBitsInInt>::Type& in2,                     \
        __mmask ## nbBitsInMask &out)                                           \
    {                                                                           \
        out = _mm512_cmp_epu ## nbBitsInInt ## _mask(in1, in2, _MM_CMPINT_NE);  \
    }                                                                           \
};

DEFINE_ISNOTEQUALREG_INT(8, 64)
DEFINE_ISNOTEQUALREG_UINT(8, 64)
DEFINE_ISNOTEQUALREG_INT(16, 32)
DEFINE_ISNOTEQUALREG_UINT(16, 32)
DEFINE_ISNOTEQUALREG_INT(32, 16)
DEFINE_ISNOTEQUALREG_UINT(32, 16)
DEFINE_ISNOTEQUALREG_INT(64, 8)
DEFINE_ISNOTEQUALREG_UINT(64, 8)

template <>
struct IsNotEqualReg<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    __mmask16
    act(const Avx512Type<ipReal32>::Type& in1,
        const Avx512Type<ipReal32>::Type& in2)
    {
        return _mm512_cmp_ps_mask(in1, in2, _MM_CMPINT_NE);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal32>::Type& in1,
        const Avx512Type<ipReal32>::Type& in2,
        __mmask16& out)
    {
        out = _mm512_cmp_ps_mask(in1, in2, _MM_CMPINT_NE);
    }
};

template <>
struct IsNotEqualReg<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
    __mmask8
    act(const Avx512Type<ipReal64>::Type& in1,
        const Avx512Type<ipReal64>::Type& in2)
    {
        return _mm512_cmp_pd_mask(in1, in2, _MM_CMPINT_NE);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal64>::Type& in1,
        const Avx512Type<ipReal64>::Type& in2,
        __mmask8& out)
    {
        out = _mm512_cmp_pd_mask(in1, in2, _MM_CMPINT_NE);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISNOTEQUALREGIMPL_H__