ScanPack.h

Go to the documentation of this file.

// ScanPack.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_SCANPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_SCANPACK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/Tools/ForceInline.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/ScanPack.h>
#include <IPSDKUtil/InstructionSet/NbEltsPerReg.h>
#include <IPSDKUtil/InstructionSet/detail/BaseReg.h>
#include <IPSDKUtil/InstructionSet/detail/LoadReg.h>
#include <IPSDKUtil/InstructionSet/detail/UnloadReg.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <>
struct ScanPack<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Avx512, ipReal32>
    act(const BasePack<ePackType::ePT_Avx512, ipReal32>& in)
    {
        return ScanPack<eInstructionSet::eIS_Avx512, ipReal32>::act(in);
    }

    static IPSDK_FORCEINLINE
    void
    act(const BasePack<ePackType::ePT_Avx512, ipReal32>& in,
        BasePack<ePackType::ePT_Avx512, ipReal32>& out)
    {
        ScanPack<eInstructionSet::eIS_Avx512, ipReal32>::act(in, out);
    }
};

template <>
struct ScanFunc<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    void act(const ipReal32* in, ipReal32* out, ipUInt64 nbElts)
    {        
        __m512 regOffset = _mm512_setzero_ps();

        const ipUInt64 nbAlignedElts = nbElts - (nbElts%simd::NbEltsPerReg<ePackType::ePT_Avx512, ipReal32>::Value);

        ipUInt64 i = 0;

        simd::BaseReg<ePackType::ePT_Avx512, ipReal32>::Type regIn, regOut;

        for (; i < nbAlignedElts; i += simd::NbEltsPerReg<ePackType::ePT_Avx512, ipReal32>::Value) {

            simd::detail::LoadReg<eInstructionSet::eIS_Avx512, ipReal32>::act(regIn, in + i);
            simd::detail::ScanReg<eInstructionSet::eIS_Avx512, ipReal32>::act(regIn, regOut);

            regOut = _mm512_add_ps(regOut, regOffset);
            simd::detail::UnloadReg<eInstructionSet::eIS_Avx512, ipReal32>::act(regOut, out + i);
            regOffset = _mm512_permutex2var_ps(regOut, _mm512_set1_epi32(15), regOut);
        }
        for (; i < nbElts; ++i) {

            // update of mean
            out[i] = out[i - 1] + in[i];
        }
    }
};

template <>
struct ScanFunc<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
    void act(const ipReal64* in, ipReal64* out, ipUInt64 nbElts)
    {
        __m512d regOffset = _mm512_setzero_pd();

        const ipUInt64 nbAlignedElts = nbElts - (nbElts%simd::NbEltsPerReg<ePackType::ePT_Avx512, ipReal64>::Value);

        ipUInt64 i = 0;

        simd::BaseReg<ePackType::ePT_Avx512, ipReal64>::Type regIn, regOut;

        for (; i < nbAlignedElts; i += simd::NbEltsPerReg<ePackType::ePT_Avx512, ipReal64>::Value) {

            simd::detail::LoadReg<eInstructionSet::eIS_Avx512, ipReal64>::act(regIn, in + i);
            simd::detail::ScanReg<eInstructionSet::eIS_Avx512, ipReal64>::act(regIn, regOut);

            regOut = _mm512_add_pd(regOut, regOffset);
            simd::detail::UnloadReg<eInstructionSet::eIS_Avx512, ipReal64>::act(regOut, out + i);
            regOffset = _mm512_permutex2var_pd(regOut, _mm512_set1_epi64(63), regOut);
        }
        for (; i < nbElts; ++i) {

            // update of mean
            out[i] = out[i - 1] + in[i];
        }
    }
};

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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_SCANPACK_H__