GatherPackImpl.h

Go to the documentation of this file.

// GatherPackImpl.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX512_GATHERPACKIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX512_GATHERPACKIMPL_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/Gather/detail/GatherPackImpl.h>
#include <IPSDKUtil/InstructionSet/Gather/detail/GatherReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 1>::type
>::act(
    const T* baseAddress,
    const ipUInt32* indexes,
    BasePack<ePackType::ePT_Avx512, T>& out)
{
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress,
        indexes,
        out._val[0]);
}

template <typename T>
IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 1>::type
>::act(
    const T* baseAddress,
    const BasePack<ePackType::ePT_Avx512, ipUInt32>& packIdx,
    BasePack<ePackType::ePT_Avx512, T>& out)
{
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress,
        packIdx._val[0], packIdx._val[1],
        packIdx._val[2], packIdx._val[3],
        out._val[0]);
}

template <typename T>
IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type
>::act(
    const T* baseAddress,
    const ipUInt32* indexes,
    BasePack<ePackType::ePT_Avx512, T>& out)
{
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes, out._val[0]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+32, out._val[1]);
}

template <typename T>
IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 2>::type
>::act(
    const T* baseAddress,
    const BasePack<ePackType::ePT_Avx512, ipUInt32>& packIdx,
    BasePack<ePackType::ePT_Avx512, T>& out)
{
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, packIdx._val[0], packIdx._val[1], out._val[0]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, packIdx._val[2], packIdx._val[3], out._val[1]);
}

template <typename T>
IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 4>::type
>::act(
    const T* baseAddress,
    const ipUInt32* indexes,
    BasePack<ePackType::ePT_Avx512, T>& out)
{
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes, out._val[0]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+16, out._val[1]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+32, out._val[2]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+48, out._val[3]);
}

template <typename T>
IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 4>::type
>::act(
    const T* baseAddress,
    const BasePack<ePackType::ePT_Avx512, ipUInt32>& packIdx,
    BasePack<ePackType::ePT_Avx512, T>& out)
{
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, packIdx._val[0], out._val[0]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, packIdx._val[1], out._val[1]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, packIdx._val[2], out._val[2]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, packIdx._val[3], out._val[3]);
}

template <typename T>
IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 8>::type
>::act(
    const T* baseAddress,
    const ipUInt32* indexes,
    BasePack<ePackType::ePT_Avx512, T>& out)
{
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes, out._val[0]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+8, out._val[1]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+16, out._val[2]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+24, out._val[3]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+32, out._val[4]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+40, out._val[5]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+48, out._val[6]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, indexes+56, out._val[7]);
}

template <typename T>
IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx512, T,
    typename boost::enable_if_c<sizeof(T) == 8>::type
>::act(
    const T* baseAddress,
    const BasePack<ePackType::ePT_Avx512, ipUInt32>& packIdx,
    BasePack<ePackType::ePT_Avx512, T>& out)
{
    const __m512i duplicateHiMask = _mm512_setr_epi32(
        8, 9, 10, 11, 12, 13, 14, 15, 8, 9, 10, 11, 12, 13, 14, 15);

    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, _mm512_castsi512_si256(packIdx._val[0]), out._val[0]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, 
        _mm512_castsi512_si256(_mm512_permutex2var_epi32(
            packIdx._val[0], duplicateHiMask, packIdx._val[0])),
        out._val[1]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, _mm512_castsi512_si256(packIdx._val[1]), out._val[2]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress,
        _mm512_castsi512_si256(_mm512_permutex2var_epi32(
            packIdx._val[1], duplicateHiMask, packIdx._val[1])),
        out._val[3]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, _mm512_castsi512_si256(packIdx._val[2]), out._val[4]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress,
        _mm512_castsi512_si256(_mm512_permutex2var_epi32(
            packIdx._val[2], duplicateHiMask, packIdx._val[2])),
        out._val[5]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress, _mm512_castsi512_si256(packIdx._val[3]), out._val[6]);
    GatherReg<eInstructionSet::eIS_Avx512, T>::act(
        baseAddress,
        _mm512_castsi512_si256(_mm512_permutex2var_epi32(
            packIdx._val[3], duplicateHiMask, packIdx._val[3])),
        out._val[7]);
}


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_GATHER_DETAIL_AVX512_GATHERPACKIMPL_H__