MulReg.h

Go to the documentation of this file.

// MulReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_MULREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_MULREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/AVX512/Avx512Types.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/MulReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


static IPSDK_FORCEINLINE
__m512i selectb (const __mmask64& s, const __m512i& a, const __m512i& b) {
    return _mm512_mask_blend_epi8(s, b, a);
}

static IPSDK_FORCEINLINE
void selectb (const __mmask64& s, const __m512i& a, const __m512i& b, __m512i& res) {
    res = _mm512_mask_blend_epi8(s, b, a);
}

static IPSDK_FORCEINLINE
__m512i
_custom_mm512_mullo_epi8(__m512i a, __m512i b)
{
    __m512i aodd    = _mm512_srli_epi16(a,8);                 // odd numbered elements of a
    __m512i bodd    = _mm512_srli_epi16(b,8);                 // odd numbered elements of b
    __m512i muleven = _mm512_mullo_epi16(a,b);                // product of even numbered elements
    __m512i mulodd  = _mm512_mullo_epi16(aodd,bodd);          // product of odd  numbered elements
            mulodd  = _mm512_slli_epi16(mulodd,8);            // put odd numbered elements back in place
    __mmask64 mask = 0x5555555555555555;                      // mask for even positions
    __m512i product = selectb(mask, muleven, mulodd);         // interleave even and odd
    return product;
}

static IPSDK_FORCEINLINE
void
_custom_mm512_mullo_epi8(__m512i a, __m512i b, __m512i& res)
{
    __m512i aodd    = _mm512_srli_epi16(a,8);                 // odd numbered elements of a
    __m512i bodd    = _mm512_srli_epi16(b,8);                 // odd numbered elements of b
    __m512i muleven = _mm512_mullo_epi16(a,b);                // product of even numbered elements
    __m512i mulodd  = _mm512_mullo_epi16(aodd,bodd);          // product of odd  numbered elements
            mulodd  = _mm512_slli_epi16(mulodd,8);            // put odd numbered elements back in place
    __mmask64 mask = 0x5555555555555555;                      // mask for even positions
    selectb(mask,muleven,mulodd, res);                        // interleave even and odd
}


template <typename T>
struct MulReg<eInstructionSet::eIS_Avx512, T, 
    typename boost::enable_if_c<boost::is_integral<T>::value 
        && sizeof(T)==1>::type>
{
    static IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2)
    {
        return _custom_mm512_mullo_epi8(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        _custom_mm512_mullo_epi8(in1, in2, out);
    }
};

template <typename T>
struct MulReg<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T)==2>::type>
{
    static IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2)
    {
        return _mm512_mullo_epi16(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        out = _mm512_mullo_epi16(in1, in2);
    }
};

template <typename T>
struct MulReg<eInstructionSet::eIS_Avx512, T, 
    typename boost::enable_if_c<boost::is_integral<T>::value 
        && sizeof(T)==4>::type>
{
    static IPSDK_FORCEINLINE
    typename Avx512Type<T>::Type
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2)
    {
        return _mm512_mullo_epi32(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename Avx512Type<T>::Type& in1,
        const typename Avx512Type<T>::Type& in2,
        typename Avx512Type<T>::Type& out)
    {
        out = _mm512_mullo_epi32(in1, in2);
    }
};

template <>
struct MulReg<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipReal32>::Type
    act(const Avx512Type<ipReal32>::Type& in1,
        const Avx512Type<ipReal32>::Type& in2)
    {
        return _mm512_mul_ps(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal32>::Type& in1,
        const Avx512Type<ipReal32>::Type& in2,
        Avx512Type<ipReal32>::Type& out)
    {
        out = _mm512_mul_ps(in1, in2);
    }
};

template <>
struct MulReg<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipReal64>::Type
    act(const Avx512Type<ipReal64>::Type& in1,
        const Avx512Type<ipReal64>::Type& in2)
    {
        return _mm512_mul_pd(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal64>::Type& in1,
        const Avx512Type<ipReal64>::Type& in2,
        Avx512Type<ipReal64>::Type& out)
    {
        out = _mm512_mul_pd(in1, in2);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_MULREG_H__