ScanReg.h

Go to the documentation of this file.

// ScanReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_SCANREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_SCANREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/ScanReg.h>
#include <IPSDKUtil/InstructionSet/SSE2/Sse2Types.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {

template <>
struct ScanReg<eInstructionSet::eIS_Sse2, ipReal32>
{
    static IPSDK_FORCEINLINE
    Sse2Type<ipReal32>::Type
    act(const Sse2Type<ipReal32>::Type& in)
    {
        Sse2Type<ipReal32>::Type out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipReal32>::Type& in,
        Sse2Type<ipReal32>::Type& out)
    {
        out = _mm_add_ps(in, _mm_castsi128_ps(_mm_slli_si128(_mm_castps_si128(in), 4)));
        out = _mm_add_ps(out, _mm_shuffle_ps(_mm_setzero_ps(), out, 0x40));
    }
};

template <>
struct ScanReg<eInstructionSet::eIS_Sse2, ipReal64>
{
    static IPSDK_FORCEINLINE
        Sse2Type<ipReal64>::Type
        act(const Sse2Type<ipReal64>::Type& in)
    {
        Sse2Type<ipReal64>::Type out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
        void
        act(const Sse2Type<ipReal64>::Type& in,
            Sse2Type<ipReal64>::Type& out)
    {
        out = _mm_add_pd(in, _mm_castsi128_pd(_mm_slli_si128(_mm_castpd_si128(in), 8)));
    }
};

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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_SCANREG_H__