CeilReg.h

// CeilReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_CEILREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_CEILREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/BasePack.h>
#include <IPSDKUtil/InstructionSet/AVX512/Avx512Types.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/CeilReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#include "immintrin.h"

namespace ipsdk {
namespace simd {
namespace detail {


template <>
struct CeilReg<eInstructionSet::eIS_Avx512, ipReal32>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipReal32>::Type
    act(const Avx512Type<ipReal32>::Type& in)
    {
        Avx512Type<ipReal32>::Type out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal32>::Type& in,
        Avx512Type<ipReal32>::Type& out)
    {
        out = _mm512_roundscale_ps(in, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC);
    }
};

template <>
struct CeilReg<eInstructionSet::eIS_Avx512, ipReal64>
{
    static IPSDK_FORCEINLINE
    Avx512Type<ipReal64>::Type
    act(const Avx512Type<ipReal64>::Type& in)
    {
        Avx512Type<ipReal64>::Type out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const Avx512Type<ipReal64>::Type& in,
        Avx512Type<ipReal64>::Type& out)
    {
        out = _mm512_roundscale_pd(in, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_CEILREG_H__