PolynomPack.h

Go to the documentation of this file.

// PolynomPack.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_POLYNOMPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE_POLYNOMPACK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/BasePack.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/PolynomReg.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#include <boost/mpl/not_equal_to.hpp>

namespace ipsdk {
namespace simd {
namespace detail {


#define DEFINE_POLYNOMPACK_SSE2(N)   \
template <typename T>   \
struct PolynomPack<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T, typename boost::enable_if_c<boost::mpl::equal_to<boost::mpl::int_<sizeof(T)>, boost::mpl::int_<1> >::value>::type> \
{   \
    static IPSDK_FORCEINLINE    \
    BasePack<ePackType::ePT_Sse, T>  \
    act(const BasePack<ePackType::ePT_Sse, T>& in,    \
        BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), const T& c))  \
    {   \
        BasePack<ePackType::ePT_Scalar, T> out;    \
        act(in, BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out); \
        return out; \
    }   \
    \
    static IPSDK_FORCEINLINE    \
    void    \
    act(const BasePack<ePackType::ePT_Sse, T>& in,   \
        BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), const T& c), \
        BasePack<ePackType::ePT_Sse, T>& out)    \
    {   \
        PolynomReg<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T>::act(in._val[0], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[0]);   \
    }   \
};  \
    \
template <typename T>   \
struct PolynomPack<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T, typename boost::enable_if_c<boost::mpl::equal_to<boost::mpl::int_<sizeof(T)>, boost::mpl::int_<2> >::value>::type> \
{   \
    static IPSDK_FORCEINLINE    \
    BasePack<ePackType::ePT_Sse, T>  \
    act(const BasePack<ePackType::ePT_Sse, T>& in,    \
        BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), const T& c))  \
    {   \
        BasePack<ePackType::ePT_Scalar, T> out;    \
        act(in, BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out); \
        return out; \
    }   \
    \
    static IPSDK_FORCEINLINE    \
    void    \
    act(const BasePack<ePackType::ePT_Sse, T>& in,   \
        BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), const T& c), \
        BasePack<ePackType::ePT_Sse, T>& out)    \
    {   \
        PolynomReg<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T>::act(in._val[0], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[0]);   \
        PolynomReg<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T>::act(in._val[1], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[1]);   \
    }   \
};  \
    \
template <typename T>   \
struct PolynomPack<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T, typename boost::enable_if_c<boost::mpl::equal_to<boost::mpl::int_<sizeof(T)>, boost::mpl::int_<4> >::value>::type> \
{   \
    static IPSDK_FORCEINLINE    \
    BasePack<ePackType::ePT_Sse, T>  \
    act(const BasePack<ePackType::ePT_Sse, T>& in,    \
        BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), const T& c))  \
    {   \
        BasePack<ePackType::ePT_Scalar, T> out;    \
        act(in, BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out); \
        return out; \
    }   \
    \
    static IPSDK_FORCEINLINE    \
    void    \
    act(const BasePack<ePackType::ePT_Sse, T>& in,   \
        BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), const T& c), \
        BasePack<ePackType::ePT_Sse, T>& out)    \
    {   \
        PolynomReg<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T>::act(in._val[0], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[0]);   \
        PolynomReg<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T>::act(in._val[1], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[1]);   \
        PolynomReg<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T>::act(in._val[2], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[2]);   \
        PolynomReg<eInstructionSet::eIS_Sse2, eInstructionSet::eIS_Sse2, N, T>::act(in._val[3], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[3]);   \
    }   \
};  \

DEFINE_POLYNOMPACK_SSE2(20)
DEFINE_POLYNOMPACK_SSE2(19)
DEFINE_POLYNOMPACK_SSE2(18)
DEFINE_POLYNOMPACK_SSE2(17)
DEFINE_POLYNOMPACK_SSE2(16)
DEFINE_POLYNOMPACK_SSE2(15)
DEFINE_POLYNOMPACK_SSE2(14)
DEFINE_POLYNOMPACK_SSE2(13)
DEFINE_POLYNOMPACK_SSE2(12)
DEFINE_POLYNOMPACK_SSE2(11)
DEFINE_POLYNOMPACK_SSE2(10)
DEFINE_POLYNOMPACK_SSE2(9)
DEFINE_POLYNOMPACK_SSE2(8)
DEFINE_POLYNOMPACK_SSE2(7)
DEFINE_POLYNOMPACK_SSE2(6)
DEFINE_POLYNOMPACK_SSE2(5)
DEFINE_POLYNOMPACK_SSE2(4)
DEFINE_POLYNOMPACK_SSE2(3)
DEFINE_POLYNOMPACK_SSE2(2)
DEFINE_POLYNOMPACK_SSE2(1)


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_SSE2_POLYNOMPACK_H__