DivPack.h

Go to the documentation of this file.

// DivPack.h:
// -----------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_DIVPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_DIVPACK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/cast.h>
#include <IPSDKUtil/InstructionSet/InstructionSetTypes.h>
#include <IPSDKUtil/InstructionSet/PromotedType.h>
#include <IPSDKUtil/InstructionSet/BasePack.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/DivReg.h>
#include <IPSDKUtil/InstructionSet/detail/BinaryPackOp.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#include <boost/type_traits/is_same.hpp>
#include <boost/utility/enable_if.hpp>

namespace ipsdk {
namespace simd {
namespace detail {


template <>
struct DivPack<eInstructionSet::eIS_Sse2, ipReal32, ipReal32>
{
    static
    IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, ipReal32>
    act(const BasePack<ePackType::ePT_Sse, ipReal32>& in1,
        const BasePack<ePackType::ePT_Sse, ipReal32>& in2)
    {
        return BinaryPackOp<eInstructionSet::eIS_Sse2, ipReal32, 
            DivReg<eInstructionSet::eIS_Sse2, ipReal32> >::act(in1, in2);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const BasePack<ePackType::ePT_Sse, ipReal32>& in1,
        const BasePack<ePackType::ePT_Sse, ipReal32>& in2,
        BasePack<ePackType::ePT_Sse, ipReal32>& out)
    {
        BinaryPackOp<eInstructionSet::eIS_Sse2, ipReal32, 
            DivReg<eInstructionSet::eIS_Sse2, ipReal32> >::act(in1, in2, out);
    }
};

template <>
struct DivPack<eInstructionSet::eIS_Sse2, ipReal64, ipReal64>
{
    static
    IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, ipReal64>
    act(const BasePack<ePackType::ePT_Sse, ipReal64>& in1,
        const BasePack<ePackType::ePT_Sse, ipReal64>& in2)
    {
        return BinaryPackOp<eInstructionSet::eIS_Sse2, ipReal64, 
            DivReg<eInstructionSet::eIS_Sse2, ipReal64> >::act(in1, in2);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const BasePack<ePackType::ePT_Sse, ipReal64>& in1,
        const BasePack<ePackType::ePT_Sse, ipReal64>& in2,
        BasePack<ePackType::ePT_Sse, ipReal64>& out)
    {
        BinaryPackOp<eInstructionSet::eIS_Sse2, ipReal64, 
            DivReg<eInstructionSet::eIS_Sse2, ipReal64> >::act(in1, in2, out);
    }
};

template <typename TIn1, typename TIn2>
struct DivPack<eInstructionSet::eIS_Sse2, TIn1, TIn2,
    typename boost::enable_if_c<!boost::is_same<TIn1, TIn2>::value 
        && boost::is_same<TIn1, typename PromotedType<TIn1, TIn2>::Type>::value
    >::type>
{
    static
    IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, TIn1>
    act(const BasePack<ePackType::ePT_Sse, TIn1>& in1,
        const BasePack<ePackType::ePT_Sse, TIn2>& in2)
    {
        BasePack<ePackType::ePT_Sse, TIn1> in2Promoted =
            cast<eInstructionSet::eIS_Sse2, TIn2, TIn1>(in2);
        return DivPack<eInstructionSet::eIS_Sse2, TIn1, TIn1>::act(in1, in2Promoted);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const BasePack<ePackType::ePT_Sse, TIn1>& in1,
        const BasePack<ePackType::ePT_Sse, TIn2>& in2,
        BasePack<ePackType::ePT_Sse, TIn1>& out)
    {
        BasePack<ePackType::ePT_Sse, TIn1> in2Promoted;
        cast<eInstructionSet::eIS_Sse2, TIn2, TIn1>(in2, in2Promoted);
        DivPack<eInstructionSet::eIS_Sse2, TIn1, TIn1>::act(in1, in2Promoted, out);
    }
};

template <typename TIn1, typename TIn2>
struct DivPack<eInstructionSet::eIS_Sse2, TIn1, TIn2,
    typename boost::enable_if_c<!boost::is_same<TIn1, TIn2>::value
        && !boost::is_same<TIn1, typename PromotedType<TIn1, TIn2>::Type>::value
    >::type>
{
    static
    IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, TIn2>
    act(const BasePack<ePackType::ePT_Sse, TIn1>& in1,
        const BasePack<ePackType::ePT_Sse, TIn2>& in2)
    {
        BasePack<ePackType::ePT_Sse, TIn2> in1Promoted =
            cast<eInstructionSet::eIS_Sse2, TIn1, TIn2>(in1);
        return DivPack<eInstructionSet::eIS_Sse2, TIn2, TIn2>::act(
            in1Promoted, in2);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const BasePack<ePackType::ePT_Sse, TIn1>& in1,
        const BasePack<ePackType::ePT_Sse, TIn2>& in2,
        BasePack<ePackType::ePT_Sse, TIn2>& out)
    {
        BasePack<ePackType::ePT_Sse, TIn2> in1Promoted;
        cast<eInstructionSet::eIS_Sse2, TIn1, TIn2>(in1, in1Promoted);
        DivPack<eInstructionSet::eIS_Sse2, TIn2, TIn2>::act(
            in1Promoted, in2, out);
    }
};

template <typename TIn>
struct DivPack<eInstructionSet::eIS_Sse2, TIn, TIn,
    typename boost::enable_if_c<boost::is_integral<TIn>::value>::type>
{
    static
    IPSDK_FORCEINLINE
    BasePack<ePackType::ePT_Sse, TIn>
    act(const BasePack<ePackType::ePT_Sse, TIn>& in1,
        const BasePack<ePackType::ePT_Sse, TIn>& in2)
    {
        const BasePack<ePackType::ePT_Sse, ipReal32> in1Cast =
            cast<eInstructionSet::eIS_Sse2, TIn, ipReal32>(in1);
        const BasePack<ePackType::ePT_Sse, ipReal32> in2Cast =
            cast<eInstructionSet::eIS_Sse2, TIn, ipReal32>(in2);
        const BasePack<ePackType::ePT_Sse, ipReal32> outCast =
            DivPack<eInstructionSet::eIS_Sse2, ipReal32, ipReal32>::act(in1Cast, in2Cast);

        return cast<eInstructionSet::eIS_Sse2, ipReal32, TIn>(outCast);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const BasePack<ePackType::ePT_Sse, TIn>& in1,
        const BasePack<ePackType::ePT_Sse, TIn>& in2,
        BasePack<ePackType::ePT_Sse, TIn>& out)
    {
        BasePack<ePackType::ePT_Sse, ipReal32> in1Cast;
        cast<eInstructionSet::eIS_Sse2>(in1, in1Cast);
        BasePack<ePackType::ePT_Sse, ipReal32> in2Cast;
        cast<eInstructionSet::eIS_Sse2>(in2, in2Cast);
        BasePack<ePackType::ePT_Sse, ipReal32> outCast;
        DivPack<eInstructionSet::eIS_Sse2, ipReal32, ipReal32>::act(in1Cast, in2Cast, outCast);

        cast<eInstructionSet::eIS_Sse2>(outCast, out);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_DIVPACK_H__