AddPack.h

Go to the documentation of this file.

// AddPack.h:
// -----------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_ADDPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_ADDPACK_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/AddReg.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 <eInstructionSet::domain instructionSet,
          typename TIn1,
          typename TIn2,
          typename Enable=void>
struct AddPack
{
};


// addition of 2 packs of same type
template <eInstructionSet::domain instructionSet, typename TIn1, typename TIn2>
struct AddPack<instructionSet, TIn1, TIn2,
    typename boost::enable_if_c<boost::is_same<TIn1, TIn2>::value>::type>
{
    static
    IPSDK_FORCEINLINE
    const typename IS2Pack<instructionSet, TIn1>::Type
    act(const typename IS2Pack<instructionSet, TIn1>::Type& in1,
        const typename IS2Pack<instructionSet, TIn2>::Type& in2)
    {
        return BinaryPackOp<instructionSet, TIn1,
            AddReg<instructionSet, TIn1> >::act(in1, in2);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const typename IS2Pack<instructionSet, TIn1>::Type& in1,
        const typename IS2Pack<instructionSet, TIn2>::Type& in2,
        typename IS2Pack<instructionSet, TIn1>::Type& out)
    {
        BinaryPackOp<instructionSet, TIn1, AddReg<instructionSet, TIn1> >::act(
            in1, in2, out);
    }
};

// addition of 2 packs of different types; TIn1 is greater than TIn2
template <eInstructionSet::domain instructionSet, typename TIn1, typename TIn2>
struct AddPack<instructionSet, 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
    typename IS2Pack<instructionSet, TIn1>::Type
    act(const typename IS2Pack<instructionSet, TIn1>::Type& in1,
        const typename IS2Pack<instructionSet, TIn2>::Type& in2)
    {
        typename IS2Pack<instructionSet, TIn1>::Type in2Promoted =
            cast<instructionSet, TIn2, TIn1>(in2);
        return BinaryPackOp<instructionSet, TIn1,
            AddReg<instructionSet, TIn1> >::act(in1, in2Promoted);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const typename IS2Pack<instructionSet, TIn1>::Type& in1,
        const typename IS2Pack<instructionSet, TIn2>::Type& in2,
        typename IS2Pack<instructionSet, TIn1>::Type& out)
    {
        typename IS2Pack<instructionSet, TIn1>::Type in2Promoted;
        cast<instructionSet, TIn2, TIn1>(in2, in2Promoted);
        BinaryPackOp<instructionSet, TIn1, AddReg<instructionSet, TIn1> >::act(
            in1, in2Promoted, out);
    }
};

// addition of 2 packs of different types; TIn2 is greater than TIn1
template <eInstructionSet::domain instructionSet, typename TIn1, typename TIn2>
struct AddPack<instructionSet, 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
    typename IS2Pack<instructionSet, TIn2>::Type
    act(const typename IS2Pack<instructionSet, TIn1>::Type& in1,
        const typename IS2Pack<instructionSet, TIn2>::Type& in2)
    {
        return AddPack<instructionSet, TIn2, TIn1>::act(in2, in1);
    }

    static
    IPSDK_FORCEINLINE
    void
    act(const typename IS2Pack<instructionSet, TIn1>::Type& in1,
        const typename IS2Pack<instructionSet, TIn2>::Type& in2,
        typename IS2Pack<instructionSet, TIn2>::Type& out)
    {
        AddPack<instructionSet, TIn2, TIn1>::act(in2, in1, out);
    }
};



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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_ADDPACK_H__