IfElseReg.h

Go to the documentation of this file.

// IfElseReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX2_IFELSEREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX2_IFELSEREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/Logical/detail/BitwiseSelectReg.h>
#include <IPSDKUtil/InstructionSet/Logical/detail/IfElseReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
struct IfElseReg<
    eInstructionSet::eIS_Avx2,
    T,
    typename boost::enable_if_c<boost::is_integral<T>::value>::type>
{
    static IPSDK_FORCEINLINE
    typename AvxType<T>::Type
    act(const typename AvxMaskType<T>::Type& mask,
        const typename AvxType<T>::Type& in1,
        const typename AvxType<T>::Type& in2)
    {
        typename AvxType<T>::Type regFromMask;
        CastReg<eInstructionSet::eIS_Avx2, T, T>::act(mask, regFromMask);
        typename AvxType<T>::Type out;
        
        BitwiseSelectReg<eInstructionSet::eIS_Avx2, T>::act(regFromMask, in1, in2, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename AvxMaskType<T>::Type& mask,
        const typename AvxType<T>::Type& in1,
        const typename AvxType<T>::Type& in2,
        typename AvxType<T>::Type& out)
    {
        typename AvxType<T>::Type regFromMask;
        CastReg<eInstructionSet::eIS_Avx2, T, T>::act(mask, regFromMask);
        BitwiseSelectReg<eInstructionSet::eIS_Avx2, T>::act(regFromMask, in1, in2, out);
    }
};

template <>
struct IfElseReg<eInstructionSet::eIS_Avx2, ipReal32>
{
    static IPSDK_FORCEINLINE
        AvxType<ipReal32>::Type
        act(const AvxMaskType<ipReal32>::Type& mask,
            const AvxType<ipReal32>::Type& in1,
            const AvxType<ipReal32>::Type& in2)
    {
        AvxType<ipReal32>::Type regFromMask;
        CastReg<eInstructionSet::eIS_Avx2, ipReal32, ipReal32>::act(mask, regFromMask);
        AvxType<ipReal32>::Type out;

        out = _mm256_blendv_ps(in2, in1, regFromMask);
    }

    static IPSDK_FORCEINLINE
        void
        act(const AvxMaskType<ipReal32>::Type& mask,
            const AvxType<ipReal32>::Type& in1,
            const AvxType<ipReal32>::Type& in2,
            AvxType<ipReal32>::Type& out)
    {
        AvxType<ipReal32>::Type regFromMask;
        CastReg<eInstructionSet::eIS_Avx2, ipReal32, ipReal32>::act(mask, regFromMask);

        out = _mm256_blendv_ps(in2, in1, regFromMask);
    }
};

template <>
struct IfElseReg<eInstructionSet::eIS_Avx2, ipReal64>
{
    static IPSDK_FORCEINLINE
        AvxType<ipReal64>::Type
        act(const AvxMaskType<ipReal64>::Type& mask,
            const AvxType<ipReal64>::Type& in1,
            const AvxType<ipReal64>::Type& in2)
    {
        AvxType<ipReal64>::Type regFromMask;
        CastReg<eInstructionSet::eIS_Avx2, ipReal64, ipReal64>::act(mask, regFromMask);
        AvxType<ipReal64>::Type out;

        out = _mm256_blendv_pd(in2, in1, regFromMask);
    }

    static IPSDK_FORCEINLINE
        void
        act(const AvxMaskType<ipReal64>::Type& mask,
            const AvxType<ipReal64>::Type& in1,
            const AvxType<ipReal64>::Type& in2,
            AvxType<ipReal64>::Type& out)
    {
        AvxType<ipReal64>::Type regFromMask;
        CastReg<eInstructionSet::eIS_Avx2, ipReal64, ipReal64>::act(mask, regFromMask);

        out = _mm256_blendv_pd(in2, in1, regFromMask);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX2_IFELSEREG_H__