CastReg.h

Go to the documentation of this file.

// CastReg.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_CASTREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_CASTREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/detail/CastReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {

template <typename TIn, typename TOut>
struct CastReg<eInstructionSet::eIS_Avx, TIn, TOut,
    typename boost::enable_if_c<
        boost::is_same<TIn, TOut>::value || (boost::is_integral<TIn>::value
        && boost::is_integral<TOut>::value && sizeof(TIn)==sizeof(TOut))
    >::type
>
{
    static IPSDK_FORCEINLINE
    void act(const typename AvxType<TIn>::Type& in, 
             typename AvxType<TOut>::Type& out)
    {
        out = in;
    }
};

template <>
struct CastReg<eInstructionSet::eIS_Avx, ipInt32, ipReal32>
{
    static IPSDK_FORCEINLINE
    void act(const AvxType<ipInt32>::Type& in,
             AvxType<ipReal32>::Type& out)
    {
        out = _mm256_cvtepi32_ps(in);
    }
};

template <>
struct CastReg<eInstructionSet::eIS_Avx, ipReal32, ipInt32>
{
    static IPSDK_FORCEINLINE
    void act(const AvxType<ipReal32>::Type& in,
             AvxType<ipInt32>::Type& out)
    {
        out = _mm256_cvttps_epi32(in);
    }
};

template <>
struct CastReg<eInstructionSet::eIS_Avx, ipReal64, ipReal32>
{
    static IPSDK_FORCEINLINE
    void act(const AvxType<ipReal64>::Type& inl,
        const AvxType<ipReal64>::Type& inh,
        AvxType<ipReal32>::Type& out)
    {
        out = _mm256_castps128_ps256(_mm256_cvtpd_ps(inl));
        out = _mm256_insertf128_ps(out, _mm256_cvtpd_ps(inh), 1);
    }
};

template <>
struct CastReg<eInstructionSet::eIS_Avx, ipReal32, ipReal64>
{
    static IPSDK_FORCEINLINE
    void act(const AvxType<ipReal32>::Type& in,
        AvxType<ipReal64>::Type& outl,
        AvxType<ipReal64>::Type& outh)
    {
        outl = _mm256_cvtps_pd(_mm256_extractf128_ps(in, 0));
        outh = _mm256_cvtps_pd(_mm256_extractf128_ps(in, 1));
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_CASTREG_H__