MaskCastPack.h

// MaskCastPack.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_MASKCASTPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_MASKCASTPACK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/BaseMaskPack.h>
#include <IPSDKUtil/InstructionSet/BasePack.h>
#include <IPSDKUtil/InstructionSet/detail/MaskCastPack.h>
#include <IPSDKUtil/Tools/arrayUtils.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#include <boost/type_traits/is_signed.hpp>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename TIn, typename TOut>
struct MaskCastPack<eInstructionSet::eIS_Avx, TIn, TOut,
                typename boost::enable_if_c<boost::is_same<TIn, TOut>::value>::type>
{ 
    static IPSDK_FORCEINLINE BaseMaskPack<ePackType::ePT_Avx, TOut> 
    act(const BaseMaskPack<ePackType::ePT_Avx, TIn>& in)
    {
        return in;
    }

    static IPSDK_FORCEINLINE
    void
    act(const BaseMaskPack<ePackType::ePT_Avx, TIn>& in,
        BaseMaskPack<ePackType::ePT_Avx, TOut> & out)
    {
        typedef typename ipsdk::AvxType<TIn>::Type RegType;
        arrayop::copyArray<sizeof(TIn), RegType>(in._val, out._val);
    }

};

template <>
struct MaskCastPack<eInstructionSet::eIS_Avx, ipReal32, ipReal64>
{ 
    static IPSDK_FORCEINLINE BaseMaskPack<ePackType::ePT_Avx, ipReal64> 
    act(const BaseMaskPack<ePackType::ePT_Avx, ipReal32>& in)
    {
        BaseMaskPack<ePackType::ePT_Avx, ipReal64> out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const BaseMaskPack<ePackType::ePT_Avx, ipReal32>& in,
         BaseMaskPack<ePackType::ePT_Avx, ipReal64>& out)
    {
        MaskCastReg<eInstructionSet::eIS_Avx, ipReal32, ipReal64>::act(in._val[0], out._val[0], out._val[1]);
        MaskCastReg<eInstructionSet::eIS_Avx, ipReal32, ipReal64>::act(in._val[1], out._val[2], out._val[3]);
        MaskCastReg<eInstructionSet::eIS_Avx, ipReal32, ipReal64>::act(in._val[2], out._val[4], out._val[5]);
        MaskCastReg<eInstructionSet::eIS_Avx, ipReal32, ipReal64>::act(in._val[3], out._val[6], out._val[7]);
    }
};

template <>
struct MaskCastPack<eInstructionSet::eIS_Avx, ipReal64, ipReal32>
{ 
    static IPSDK_FORCEINLINE BaseMaskPack<ePackType::ePT_Avx, ipReal32> 
    act(const BaseMaskPack<ePackType::ePT_Avx, ipReal64>& in)
    {
        BaseMaskPack<ePackType::ePT_Avx, ipReal32> out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const BaseMaskPack<ePackType::ePT_Avx, ipReal64>& in,
         BaseMaskPack<ePackType::ePT_Avx, ipReal32>& out)
    {
        MaskCastReg<eInstructionSet::eIS_Avx, ipReal64, ipReal32>::act(in._val[0], in._val[1], out._val[0]);
        MaskCastReg<eInstructionSet::eIS_Avx, ipReal64, ipReal32>::act(in._val[2], in._val[3], out._val[1]);
        MaskCastReg<eInstructionSet::eIS_Avx, ipReal64, ipReal32>::act(in._val[4], in._val[5], out._val[2]);
        MaskCastReg<eInstructionSet::eIS_Avx, ipReal64, ipReal32>::act(in._val[6], in._val[7], out._val[3]);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_MASKCASTPACK_H__