MaxRegImpl.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_MAXREGIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_MAXREGIMPL_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/Tools/ForceInline.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/MaxReg.h>
#include <IPSDKUtil/InstructionSet/detail/AVX/Reg128_256ConvertTools.h>

namespace ipsdk {
namespace simd {
namespace detail {


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

void
MaxReg<eInstructionSet::eIS_Avx, ipInt32>::act(
    const AvxType<ipInt32>::Type& in1,
    const AvxType<ipInt32>::Type& in2,
    AvxType<ipInt32>::Type& out)
{
    __m128i in11, in12;
    __m128i in21, in22;
    COPY_IMM_TO_XMM(in1, in11, in12);
    COPY_IMM_TO_XMM(in2, in21, in22);
    __m128i out1 = MaxReg<eInstructionSet::eIS_Sse2, ipInt32>::act(in11, in21);
    __m128i out2 = MaxReg<eInstructionSet::eIS_Sse2, ipInt32>::act(in12, in22);
    COPY_XMM_TO_IMM(out1, out2, out);
}

AvxType<ipReal32>::Type
MaxReg<eInstructionSet::eIS_Avx, ipReal32>::act(
    const AvxType<ipReal32>::Type& in1,
    const AvxType<ipReal32>::Type& in2)
{
    return _mm256_max_ps(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx, ipReal32>::act(
    const AvxType<ipReal32>::Type& in1,
    const AvxType<ipReal32>::Type& in2,
    AvxType<ipReal32>::Type& out)
{
    out = _mm256_max_ps(in1, in2);
}

AvxType<ipReal64>::Type
MaxReg<eInstructionSet::eIS_Avx, ipReal64>::act(
    const AvxType<ipReal64>::Type& in1,
    const AvxType<ipReal64>::Type& in2)
{
    return _mm256_max_pd(in1, in2);
}

void
MaxReg<eInstructionSet::eIS_Avx, ipReal64>::act(
    const AvxType<ipReal64>::Type& in1,
    const AvxType<ipReal64>::Type& in2,
    AvxType<ipReal64>::Type& out)
{
    out = _mm256_max_pd(in1, in2);
}


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_MAXREGIMPL_H__