IsLessEqualRegImpl.h

Go to the documentation of this file.

// IsLessEqualRegImpl.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISLESSEQUALREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISLESSEQUALREGIMPL_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


#define DEFINE_ISLESSEQUALREG_INT(nbBitsInInt, nbBitsInMask)                        \
template <>                                                                         \
struct IsLessEqualReg<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_LE);     \
    }                                                                               \
    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_LE);      \
    }                                                                               \
};

#define DEFINE_ISLESSEQUALREG_UINT(nbBitsInInt, nbBitsInMask)                       \
template <>                                                                         \
struct IsLessEqualReg<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_LE);     \
    }                                                                               \
    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_LE);      \
    }                                                                               \
};

DEFINE_ISLESSEQUALREG_INT(8, 64)
DEFINE_ISLESSEQUALREG_UINT(8, 64)
DEFINE_ISLESSEQUALREG_INT(16, 32)
DEFINE_ISLESSEQUALREG_UINT(16, 32)
DEFINE_ISLESSEQUALREG_INT(32, 16)
DEFINE_ISLESSEQUALREG_UINT(32, 16)
DEFINE_ISLESSEQUALREG_INT(64, 8)
DEFINE_ISLESSEQUALREG_UINT(64, 8)

template <>
struct IsLessEqualReg<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_LE);
    }

    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_LE);
    }

    static IPSDK_FORCEINLINE
        void
        act(const Avx512Type<ipReal32>::Type& in1,
            const ipReal32& in2,
            __mmask16& out)
    {
        Avx512Type<ipReal32>::Type regIn2;
        AssignReg<eInstructionSet::eIS_Avx512, ipReal32>::act(regIn2, in2);
        out = _mm512_cmp_ps_mask(in1, regIn2, _MM_CMPINT_LE);
    }
};

template <>
struct IsLessEqualReg<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_LE);
    }

    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_LE);
    }
};


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

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

#endif // __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_AVX512_ISLESSEQUALREGIMPL_H__