IsLessRegImpl.h

// IsLessRegImpl.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISLESSREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISLESSREGIMPL_H__

#include <IPSDKUtil/InstructionSet/Comparison/detail/AVX512/IsLessReg.h>

namespace ipsdk {
namespace simd {
namespace detail {


#define DEFINE_ISLESSREG_INT(nbBitsInInt)                                       \
template <>                                                                     \
IPSDK_FORCEINLINE                                                               \
RegMaskType<eInstructionSet::eIS_Avx512, ipInt ## nbBitsInInt>::Type            \
IsLessReg<eInstructionSet::eIS_Avx512, ipInt ## nbBitsInInt>::act(              \
    const Avx512Type<ipInt ## nbBitsInInt>::Type& in1,                          \
    const Avx512Type<ipInt ## nbBitsInInt>::Type& in2)                          \
{                                                                               \
    return _mm512_cmp_epi ## nbBitsInInt ## _mask(in1, in2, _MM_CMPINT_LT);     \
}                                                                               \
template <>                                                                     \
IPSDK_FORCEINLINE                                                               \
void                                                                            \
IsLessReg<eInstructionSet::eIS_Avx512, ipInt ## nbBitsInInt>::act(              \
    const Avx512Type<ipInt ## nbBitsInInt>::Type& in1,                          \
    const Avx512Type<ipInt ## nbBitsInInt>::Type& in2,                          \
    RegMaskType<eInstructionSet::eIS_Avx512, ipInt ## nbBitsInInt>::Type& out)  \
{                                                                               \
    out = _mm512_cmp_epi ## nbBitsInInt ## _mask(in1, in2, _MM_CMPINT_LT);      \
}

#define DEFINE_ISLESSREG_UINT(nbBitsInInt)                                      \
template <>                                                                     \
IPSDK_FORCEINLINE                                                               \
RegMaskType<eInstructionSet::eIS_Avx512, ipUInt ## nbBitsInInt>::Type           \
IsLessReg<eInstructionSet::eIS_Avx512, ipUInt ## nbBitsInInt>::act(             \
    const Avx512Type<ipInt ## nbBitsInInt>::Type& in1,                          \
    const Avx512Type<ipInt ## nbBitsInInt>::Type& in2)                          \
{                                                                               \
    return _mm512_cmp_epu ## nbBitsInInt ## _mask(in1, in2, _MM_CMPINT_LT);     \
}                                                                               \
template <>                                                                     \
IPSDK_FORCEINLINE                                                               \
void                                                                            \
IsLessReg<eInstructionSet::eIS_Avx512, ipUInt ## nbBitsInInt>::act(             \
    const Avx512Type<ipUInt ## nbBitsInInt>::Type& in1,                         \
    const Avx512Type<ipUInt ## nbBitsInInt>::Type& in2,                         \
    RegMaskType<eInstructionSet::eIS_Avx512, ipUInt ## nbBitsInInt>::Type& out) \
{                                                                               \
    out = _mm512_cmp_epu ## nbBitsInInt ## _mask(in1, in2, _MM_CMPINT_LT);      \
}

DEFINE_ISLESSREG_INT(8)
DEFINE_ISLESSREG_UINT(8)
DEFINE_ISLESSREG_INT(16)
DEFINE_ISLESSREG_UINT(16)
DEFINE_ISLESSREG_INT(32)
DEFINE_ISLESSREG_UINT(32)
DEFINE_ISLESSREG_INT(64)
DEFINE_ISLESSREG_UINT(64)

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

template <>
IPSDK_FORCEINLINE
void
IsLessReg<eInstructionSet::eIS_Avx512, ipReal32>::act(
    const Avx512Type<ipReal32>::Type& in1,
    const Avx512Type<ipReal32>::Type& in2,
    __mmask16& out)
{
    out = _mm512_cmp_ps_mask(in1, in2, _MM_CMPINT_LT);
}

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

template <>
IPSDK_FORCEINLINE
void
IsLessReg<eInstructionSet::eIS_Avx512, ipReal64>::act(
    const Avx512Type<ipReal64>::Type& in1,
    const Avx512Type<ipReal64>::Type& in2,
    __mmask8& out)
{
    out = _mm512_cmp_pd_mask(in1, in2, _MM_CMPINT_LT);
}


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISLESSREGIMPL_H__