MaskCastReg.h

Go to the documentation of this file.

// MaskCastReg.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_MASKCASTREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_MASKCASTREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/detail/MaskCastReg.h>
#include <IPSDKUtil/InstructionSet/detail/CastReg.h>
#include <IPSDKUtil/InstructionSet/Comparison/detail/IsNotEqualReg.h>
#include <IPSDKUtil/InstructionSet/RegMaskType.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {

template <typename TIn, typename TOut>
struct MaskCastReg<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 RegMaskType<eInstructionSet::eIS_Avx, TIn>::Type& in, 
             typename RegMaskType<eInstructionSet::eIS_Avx, TOut>::Type& out)
    {
        out = in;
    }
};

template <>
struct MaskCastReg<eInstructionSet::eIS_Avx, ipInt32, ipReal32>
{
    static IPSDK_FORCEINLINE
    void act(const RegMaskType<eInstructionSet::eIS_Avx, ipInt32>::Type& in,
             RegMaskType<eInstructionSet::eIS_Avx, ipReal32>::Type& out)
    {
        BitwiseCastReg<eInstructionSet::eIS_Avx, ipInt32, ipReal32>::act(in, out);
    }
};

template <>
struct MaskCastReg<eInstructionSet::eIS_Avx, ipReal32, ipInt32>
{
    static IPSDK_FORCEINLINE
    void act(const RegMaskType<eInstructionSet::eIS_Avx, ipReal32>::Type& in,
             RegMaskType<eInstructionSet::eIS_Avx, ipInt32>::Type& out)
    {
        BitwiseCastReg<eInstructionSet::eIS_Avx, ipReal32, ipInt32>::act(in, out);
    }
};

template <>
struct MaskCastReg<eInstructionSet::eIS_Avx, ipReal64, ipReal32>
{
    static IPSDK_FORCEINLINE
    void act(
        const RegMaskType<eInstructionSet::eIS_Avx, ipReal64>::Type& in0,
        const RegMaskType<eInstructionSet::eIS_Avx, ipReal64>::Type& in1,
        RegMaskType<eInstructionSet::eIS_Avx, ipReal32>::Type& out)
    {
        CastReg<eInstructionSet::eIS_Avx, ipReal64, ipReal32>::act(in0, in1, out);

        RegType<eInstructionSet::eIS_Avx, ipReal32>::Type zero;
        AssignReg<eInstructionSet::eIS_Avx, ipReal32>::act(zero, 0);
        IsNotEqualReg<eInstructionSet::eIS_Avx, ipReal32>::act(out, zero, out);
    }
};

template <>
struct MaskCastReg<eInstructionSet::eIS_Avx, ipReal32, ipReal64>
{
    static IPSDK_FORCEINLINE
    void act(
        const RegMaskType<eInstructionSet::eIS_Avx, ipReal32>::Type& in,
        RegMaskType<eInstructionSet::eIS_Avx, ipReal64>::Type& out0,
        RegMaskType<eInstructionSet::eIS_Avx, ipReal64>::Type& out1)
    {
        CastReg<eInstructionSet::eIS_Avx, ipReal32, ipReal64>::act(in, out0, out1);

        RegType<eInstructionSet::eIS_Avx, ipReal64>::Type zero;
        AssignReg<eInstructionSet::eIS_Avx, ipReal64>::act(zero, 0);
        IsNotEqualReg<eInstructionSet::eIS_Avx, ipReal64>::act(out0, zero, out0);
        IsNotEqualReg<eInstructionSet::eIS_Avx, ipReal64>::act(out1, zero, out1);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_MASKCASTREG_H__