SqrtReg.h

Go to the documentation of this file.

// SqrtReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_SQRTREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_SQRTREG_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


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

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipReal32>::Type& in,
        Sse2Type<ipReal32>::Type& out)
    {
        out = _mm_sqrt_ps(in);
    }
};

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

    static IPSDK_FORCEINLINE
    void
    act(const Sse2Type<ipReal64>::Type& in,
        Sse2Type<ipReal64>::Type& out)
    {
        out = _mm_sqrt_pd(in);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_SQRTREG_H__