PolynomPack.h

Go to the documentation of this file.

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

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_POLYNOMPACK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_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_AVX(N)   \
template <typename T>   \
struct PolynomPack<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Avx, 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_Avx, T>  \
    act(const BasePack<ePackType::ePT_Avx, 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_Avx, T>& in,   \
        BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), const T& c), \
        BasePack<ePackType::ePT_Avx, T>& out)    \
    {   \
        PolynomReg<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Avx, N, T>::act(in._val[0], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[0]);   \
        PolynomReg<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Avx, N, T>::act(in._val[1], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[1]);   \
        PolynomReg<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Avx, N, T>::act(in._val[2], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[2]);   \
        PolynomReg<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Avx, N, T>::act(in._val[3], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[3]);   \
    }   \
};  \
    \
template <typename T>   \
struct PolynomPack<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Fma3, 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_Avx, T>  \
    act(const BasePack<ePackType::ePT_Avx, 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_Avx, T>& in,   \
        BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), const T& c), \
        BasePack<ePackType::ePT_Avx, T>& out)    \
    {   \
        PolynomReg<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Fma3, N, T>::act(in._val[0], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[0]);   \
        PolynomReg<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Fma3, N, T>::act(in._val[1], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[1]);   \
        PolynomReg<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Fma3, N, T>::act(in._val[2], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[2]);   \
        PolynomReg<eInstructionSet::eIS_Avx, eInstructionSet::eIS_Fma3, N, T>::act(in._val[3], BOOST_PP_ENUM_PARAMS(BOOST_PP_INC(N), c), out._val[3]);   \
    }   \
};

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


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX_POLYNOMPACK_H__