NormalRandomPack.h

Go to the documentation of this file.

// NormalRandomPack.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX512_NORMALRANDOMPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX512_NORMALRANDOMPACK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/saturated_round.h>
#include <IPSDKUtil/InstructionSet/detail/NormalRandomPack.h>
#include <IPSDKUtil/InstructionSet/detail/NormalRandomReg.h>
#include <IPSDKUtil/InstructionSet/BasePack.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <eInstructionSet::domain ISFma>
struct NormalRandomPack<eInstructionSet::eIS_Avx512, ISFma, ipReal32>
{
    static
    IPSDK_FORCEINLINE
    void
    act(const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& rMean,
        const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& rSigma,
        const BasePack<ePackType::ePT_Avx512, ipReal32>& rUnifRand1,
        const BasePack<ePackType::ePT_Avx512, ipReal32>& rUnifRand2,
        BasePack<ePackType::ePT_Avx512, ipReal32>& out1,
        BasePack<ePackType::ePT_Avx512, ipReal32>& out2)
    {
        RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type rOut1, rOut2;
        NormalRandomReg<eInstructionSet::eIS_Avx512, ISFma>::act(
            rMean, rSigma, rUnifRand1._val[0], rUnifRand2._val[0], rOut1, rOut2);
        out1._val[0] = rOut1;
        out2._val[0] = rOut2;

        NormalRandomReg<eInstructionSet::eIS_Avx512, ISFma>::act(
            rMean, rSigma, rUnifRand1._val[1], rUnifRand2._val[1], rOut1, rOut2);
        out1._val[1] = rOut1;
        out2._val[1] = rOut2;

        NormalRandomReg<eInstructionSet::eIS_Avx512, ISFma>::act(
            rMean, rSigma, rUnifRand1._val[2], rUnifRand2._val[2], rOut1, rOut2);
        out1._val[2] = rOut1;
        out2._val[2] = rOut2;

        NormalRandomReg<eInstructionSet::eIS_Avx512, ISFma>::act(
            rMean, rSigma, rUnifRand1._val[3], rUnifRand2._val[3], rOut1, rOut2);
        out1._val[3] = rOut1;
        out2._val[3] = rOut2;
    }
};

template <eInstructionSet::domain ISFma, typename T>
struct NormalRandomPack<eInstructionSet::eIS_Avx512, ISFma, T,
    typename boost::enable_if_c<boost::is_integral<T>::value>::type>
{
    static
    IPSDK_FORCEINLINE
    void
    act(const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& rMean,
        const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& rSigma,
        const BasePack<ePackType::ePT_Avx512, ipReal32>& rUnifRand1,
        const BasePack<ePackType::ePT_Avx512, ipReal32>& rUnifRand2,
        BasePack<ePackType::ePT_Avx512, T>& out1,
        BasePack<ePackType::ePT_Avx512, T>& out2)
    {
        RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type rOut1, rOut2;
        NormalRandomReg<eInstructionSet::eIS_Avx512, ISFma>::act(
            rMean, rSigma, rUnifRand1._val[0], rUnifRand2._val[0], rOut1, rOut2);
        BasePack<ePackType::ePT_Avx512, ipReal32> out1F32, out2F32;
        NormalRandomPack<eInstructionSet::eIS_Avx512, ISFma, ipReal32>::act(rMean, rSigma, rUnifRand1, rUnifRand2, out1F32, out2F32);
        SaturatedRoundPack<eInstructionSet::eIS_Avx512, ipReal32, T>::act(out1F32, out1);
        SaturatedRoundPack<eInstructionSet::eIS_Avx512, ipReal32, T>::act(out2F32, out2);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_NORMALRANDOMPACK_H__