BitwiseSelectReg.h

Go to the documentation of this file.

// BitwiseSelectReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX512_BITWISESELECTREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX512_BITWISESELECTREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/Logical/bitwiseAnd.h>
#include <IPSDKUtil/InstructionSet/Logical/bitwiseAndNot.h>
#include <IPSDKUtil/InstructionSet/Logical/bitwiseOr.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#include <boost/type_traits/is_integral.hpp>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
struct BitwiseSelectReg<eInstructionSet::eIS_Avx512, T>
{
    static IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& mask,
        const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2)
    {
        typename Avx512Type<T>::Type in2AndNotMask =
            BitwiseAndNotReg<eInstructionSet::eIS_Avx512, T>::act(in2, mask);

        typename Avx512Type<T>::Type out =
            BitwiseOrReg<eInstructionSet::eIS_Avx512, T>::act(
                BitwiseAndReg<eInstructionSet::eIS_Avx512, T>::act(in1, mask),
                in2AndNotMask);

        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& mask,
        const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        typename Avx512Type<T>::Type in2AndNotMask, in1AndMask;
        BitwiseAndNotReg<eInstructionSet::eIS_Avx512, T>::act(in2, mask, in2AndNotMask);
        BitwiseAndReg<eInstructionSet::eIS_Avx512, T>::act(in1, mask, in1AndMask);
        BitwiseOrReg<eInstructionSet::eIS_Avx512, T>::act(in1AndMask, in2AndNotMask, out);
    }
};
/*template <typename T>
struct BitwiseSelectReg<eInstructionSet::eIS_Avx512, T, typename boost::enable_if<typename boost::is_integral<T>::type>::type>
{
    static IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& mask,
        const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2)
    {
        return _mm512_ternarylogic_epi32(in2, in1, mask, 0xe4);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& mask,
        const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        out = _mm512_ternarylogic_epi32(in2, in1, mask, 0xe4);
    }
};

template <>
struct BitwiseSelectReg<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
        Avx512Type<ipReal32>::Type
        act(const Avx512Type<ipReal32>::Type& mask,
            const Avx512Type<ipReal32>::Type& in1,
            const Avx512Type<ipReal32>::Type& in2)
    {
        return _mm512_mask_blend_ps(
            _mm512_cmp_ps_mask(in1, in2, _MM_CMPINT_EQ),
            in2,
            in1);
    }

    static IPSDK_FORCEINLINE
        void
        act(const Avx512Type<ipReal32>::Type& mask,
            const Avx512Type<ipReal32>::Type& in1,
            const Avx512Type<ipReal32>::Type& in2,
            Avx512Type<ipReal32>::Type& out)
    {
        out = _mm512_mask_blend_ps(
            _mm512_cmp_ps_mask(in1, in2, _MM_CMPINT_EQ),
            in2,
            in1);
    }
};

template <>
struct BitwiseSelectReg<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
        Avx512Type<ipReal64>::Type
        act(const Avx512Type<ipReal64>::Type& mask,
            const Avx512Type<ipReal64>::Type& in1,
            const Avx512Type<ipReal64>::Type& in2)
    {
        return _mm512_mask_blend_pd(
            _mm512_cmp_pd_mask(in1, in2, _MM_CMPINT_EQ),
            in2,
            in1);
    }

    static IPSDK_FORCEINLINE
        void
        act(const Avx512Type<ipReal64>::Type& mask,
            const Avx512Type<ipReal64>::Type& in1,
            const Avx512Type<ipReal64>::Type& in2,
            Avx512Type<ipReal64>::Type& out)
    {
        out = _mm512_mask_blend_pd(
            _mm512_cmp_pd_mask(in1, in2, _MM_CMPINT_EQ),
            in2,
            in1);
    }
};*/


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX512_BITWISESELECTREG_H__