MulReg.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_MULREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_MULREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/AVX/AvxTypes.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/MulReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


static IPSDK_FORCEINLINE
__m256i selectb (const __m256i& s, const __m256i& a, const __m256i& b) {
    return _mm256_blendv_epi8 (b, a, s);
}

static IPSDK_FORCEINLINE
void selectb (const __m256i& s, const __m256i& a, const __m256i& b, __m256i& res) {
    res = _mm256_blendv_epi8 (b, a, s);
}

static IPSDK_FORCEINLINE
__m256i
_custom_mm256_mullo_epi8(__m256i a, __m256i b)
{
    __m256i aodd    = _mm256_srli_epi16(a,8);                 // odd numbered elements of a
    __m256i bodd    = _mm256_srli_epi16(b,8);                 // odd numbered elements of b
    __m256i muleven = _mm256_mullo_epi16(a,b);                // product of even numbered elements
    __m256i mulodd  = _mm256_mullo_epi16(aodd,bodd);          // product of odd  numbered elements
            mulodd  = _mm256_slli_epi16(mulodd,8);            // put odd numbered elements back in place
    __m256i mask    = _mm256_set1_epi32(0x00FF00FF);          // mask for even positions
    __m256i product = selectb(mask,muleven,mulodd);           // interleave even and odd
    return product;
}

static IPSDK_FORCEINLINE
void
_custom_mm256_mullo_epi8(__m256i a, __m256i b, __m256i& res)
{
    __m256i aodd    = _mm256_srli_epi16(a,8);                 // odd numbered elements of a
    __m256i bodd    = _mm256_srli_epi16(b,8);                 // odd numbered elements of b
    __m256i muleven = _mm256_mullo_epi16(a,b);                // product of even numbered elements
    __m256i mulodd  = _mm256_mullo_epi16(aodd,bodd);          // product of odd  numbered elements
            mulodd  = _mm256_slli_epi16(mulodd,8);            // put odd numbered elements back in place
    __m256i mask    = _mm256_set1_epi32(0x00FF00FF);          // mask for even positions
    selectb(mask,muleven,mulodd, res);                        // interleave even and odd
}


template <typename T>
struct MulReg<eInstructionSet::eIS_Avx2, T, 
    typename boost::enable_if_c<boost::is_integral<T>::value 
        && sizeof(T)==1>::type>
{
    static IPSDK_FORCEINLINE
    typename AvxType<T>::Type
    act(const typename AvxType<T>::Type& in1,
        const typename AvxType<T>::Type& in2)
    {
        return _custom_mm256_mullo_epi8(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename AvxType<T>::Type& in1,
        const typename AvxType<T>::Type& in2,
        typename AvxType<T>::Type& out)
    {
        _custom_mm256_mullo_epi8(in1, in2, out);
    }
};

template <typename T>
struct MulReg<eInstructionSet::eIS_Avx2, T,
    typename boost::enable_if_c<boost::is_integral<T>::value
        && sizeof(T)==2>::type>
{
    static IPSDK_FORCEINLINE
    typename AvxType<T>::Type
    act(const typename AvxType<T>::Type& in1,
        const typename AvxType<T>::Type& in2)
    {
        return _mm256_mullo_epi16(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename AvxType<T>::Type& in1,
        const typename AvxType<T>::Type& in2,
        typename AvxType<T>::Type& out)
    {
        out = _mm256_mullo_epi16(in1, in2);
    }
};

template <typename T>
struct MulReg<eInstructionSet::eIS_Avx2, T, 
    typename boost::enable_if_c<boost::is_integral<T>::value 
        && sizeof(T)==4>::type>
{
    static IPSDK_FORCEINLINE
    typename AvxType<T>::Type
    act(const typename AvxType<T>::Type& in1,
        const typename AvxType<T>::Type& in2)
    {
        return _mm256_mullo_epi32(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename AvxType<T>::Type& in1,
        const typename AvxType<T>::Type& in2,
        typename AvxType<T>::Type& out)
    {
        out = _mm256_mullo_epi32(in1, in2);
    }
};

template <>
struct MulReg<eInstructionSet::eIS_Avx2, ipReal32>
{
    static IPSDK_FORCEINLINE
    AvxType<ipReal32>::Type
    act(const AvxType<ipReal32>::Type& in1,
        const AvxType<ipReal32>::Type& in2)
    {
        return MulReg<eInstructionSet::eIS_Avx, ipReal32>::act(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal32>::Type& in1,
        const AvxType<ipReal32>::Type& in2,
        AvxType<ipReal32>::Type& out)
    {
        MulReg<eInstructionSet::eIS_Avx, ipReal32>::act(in1, in2, out);
    }
};

template <>
struct MulReg<eInstructionSet::eIS_Avx2, ipReal64>
{
    static IPSDK_FORCEINLINE
    AvxType<ipReal64>::Type
    act(const AvxType<ipReal64>::Type& in1,
        const AvxType<ipReal64>::Type& in2)
    {
        return MulReg<eInstructionSet::eIS_Avx, ipReal64>::act(in1, in2);
    }

    static IPSDK_FORCEINLINE
    void
    act(const AvxType<ipReal64>::Type& in1,
        const AvxType<ipReal64>::Type& in2,
        AvxType<ipReal64>::Type& out)
    {
        MulReg<eInstructionSet::eIS_Avx, ipReal64>::act(in1, in2, out);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_MULREG_H__