MaxRegImpl.h

Go to the documentation of this file.

// MaxRegImpl.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_MAXREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_MAXREGIMPL_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


AvxType<ipUInt8>::Type
MaxReg<eInstructionSet::eIS_Avx2, ipUInt8>::act(
    const AvxType<ipUInt8>::Type& in1,
    const AvxType<ipUInt8>::Type& in2)
{
    return _mm256_max_epu8(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx2, ipUInt8>::act(
    const AvxType<ipUInt8>::Type& in1,
    const AvxType<ipUInt8>::Type& in2,
    AvxType<ipUInt8>::Type& out)
{
    out = _mm256_max_epu8(in1, in2);
}

AvxType<ipInt8>::Type
MaxReg<eInstructionSet::eIS_Avx2, ipInt8>::act(
    const AvxType<ipInt8>::Type& in1,
    const AvxType<ipInt8>::Type& in2)
{
    return _mm256_max_epi8(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx2, ipInt8>::act(
    const AvxType<ipInt8>::Type& in1,
    const AvxType<ipInt8>::Type& in2,
    AvxType<ipInt8>::Type& out)
{
    out = _mm256_max_epi8(in1, in2);
}

AvxType<ipUInt16>::Type
MaxReg<eInstructionSet::eIS_Avx2, ipUInt16>::act(
    const AvxType<ipUInt16>::Type& in1,
    const AvxType<ipUInt16>::Type& in2)
{
    return _mm256_max_epu16(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx2, ipUInt16>::act(
    const AvxType<ipUInt16>::Type& in1,
    const AvxType<ipUInt16>::Type& in2,
    AvxType<ipUInt16>::Type& out)
{
    out = _mm256_max_epu16(in1, in2);
}

AvxType<ipInt16>::Type
MaxReg<eInstructionSet::eIS_Avx2, ipInt16>::act(
    const AvxType<ipInt16>::Type& in1,
    const AvxType<ipInt16>::Type& in2)
{
    return _mm256_max_epi16(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx2, ipInt16>::act(
    const AvxType<ipInt16>::Type& in1,
    const AvxType<ipInt16>::Type& in2,
    AvxType<ipInt16>::Type& out)
{
    out = _mm256_max_epi16(in1, in2);
}

AvxType<ipUInt32>::Type
MaxReg<eInstructionSet::eIS_Avx2, ipUInt32>::act(
    const AvxType<ipUInt32>::Type& in1,
    const AvxType<ipUInt32>::Type& in2)
{
    return _mm256_max_epu32(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx2, ipUInt32>::act(
    const AvxType<ipUInt32>::Type& in1,
    const AvxType<ipUInt32>::Type& in2,
    AvxType<ipUInt32>::Type& out)
{
    out = _mm256_max_epu32(in1, in2);
}

AvxType<ipInt32>::Type
MaxReg<eInstructionSet::eIS_Avx2, ipInt32>::act(
    const AvxType<ipInt32>::Type& in1,
    const AvxType<ipInt32>::Type& in2)
{
    return _mm256_max_epi32(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx2, ipInt32>::act(
    const AvxType<ipInt32>::Type& in1,
    const AvxType<ipInt32>::Type& in2,
    AvxType<ipInt32>::Type& out)
{
    out = _mm256_max_epi32(in1, in2);
}

AvxType<ipReal32>::Type
MaxReg<eInstructionSet::eIS_Avx2, ipReal32>::act(
    const AvxType<ipReal32>::Type& in1,
    const AvxType<ipReal32>::Type& in2)
{
    return MaxReg<eInstructionSet::eIS_Avx, ipReal32>::act(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx2, ipReal32>::act(
    const AvxType<ipReal32>::Type& in1,
    const AvxType<ipReal32>::Type& in2,
    AvxType<ipReal32>::Type& out)
{
    MaxReg<eInstructionSet::eIS_Avx, ipReal32>::act(in1, in2, out);
}

AvxType<ipReal64>::Type
MaxReg<eInstructionSet::eIS_Avx2, ipReal64>::act(
    const AvxType<ipReal64>::Type& in1,
    const AvxType<ipReal64>::Type& in2)
{
    return MaxReg<eInstructionSet::eIS_Avx, ipReal64>::act(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx2, ipReal64>::act(
    const AvxType<ipReal64>::Type& in1,
    const AvxType<ipReal64>::Type& in2,
    AvxType<ipReal64>::Type& out)
{
    MaxReg<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_MAXREGIMPL_H__