TernaryPackOpRetMask.h

Go to the documentation of this file.

// TernaryPackOpRetMask.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_TERNARYPACKOPRETMASK_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_TERNARYPACKOPRETMASK_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/InstructionSetTypes.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T, class OpClass>
struct TernaryPackOpRetMask<eInstructionSet::eIS_Avx, T, OpClass,
    typename boost::enable_if_c<sizeof(T) == 4>::type>
{
    static IPSDK_FORCEINLINE
    BaseMaskPack<ePackType::ePT_Avx, T>
    act(const BasePack<ePackType::ePT_Avx, T>& in1,
        const BasePack<ePackType::ePT_Avx, T>& in2,
        const BasePack<ePackType::ePT_Avx, T>& in3)
    {
        BaseMaskPack<ePackType::ePT_Avx, T> out;
        out._val[0] = OpClass::act(in1._val[0], in2._val[0], in3._val[0]);
        out._val[1] = OpClass::act(in1._val[1], in2._val[1], in3._val[1]);
        out._val[2] = OpClass::act(in1._val[2], in2._val[2], in3._val[2]);
        out._val[3] = OpClass::act(in1._val[3], in2._val[3], in3._val[3]);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const BasePack<ePackType::ePT_Avx, T>& in1,
        const BasePack<ePackType::ePT_Avx, T>& in2,
        const BasePack<ePackType::ePT_Avx, T>& in3,
        BaseMaskPack<ePackType::ePT_Avx, T>& out)
    {
        OpClass::act(in1._val[0], in2._val[0], in3._val[0], out._val[0]);
        OpClass::act(in1._val[1], in2._val[1], in3._val[1], out._val[1]);
        OpClass::act(in1._val[2], in2._val[2], in3._val[2], out._val[2]);
        OpClass::act(in1._val[3], in2._val[3], in3._val[3], out._val[3]);
    }
};

template <typename T, class OpClass>
struct TernaryPackOpRetMask<eInstructionSet::eIS_Avx, T, OpClass,
    typename boost::enable_if_c<sizeof(T) == 8>::type>
{
    static IPSDK_FORCEINLINE
    BaseMaskPack<ePackType::ePT_Avx, T>
    act(const BasePack<ePackType::ePT_Avx, T>& in1,
        const BasePack<ePackType::ePT_Avx, T>& in2,
        const BasePack<ePackType::ePT_Avx, T>& in3)
    {
        BaseMaskPack<ePackType::ePT_Avx, T> out;
        out._val[0] = OpClass::act(in1._val[0], in2._val[0], in3._val[0]);
        out._val[1] = OpClass::act(in1._val[1], in2._val[1], in3._val[1]);
        out._val[2] = OpClass::act(in1._val[2], in2._val[2], in3._val[2]);
        out._val[3] = OpClass::act(in1._val[3], in2._val[3], in3._val[3]);
        out._val[4] = OpClass::act(in1._val[0], in2._val[4], in3._val[4]);
        out._val[5] = OpClass::act(in1._val[1], in2._val[5], in3._val[5]);
        out._val[6] = OpClass::act(in1._val[2], in2._val[6], in3._val[6]);
        out._val[7] = OpClass::act(in1._val[3], in2._val[7], in3._val[7]);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const BasePack<ePackType::ePT_Avx, T>& in1,
        const BasePack<ePackType::ePT_Avx, T>& in2,
        const BasePack<ePackType::ePT_Avx, T>& in3,
        BaseMaskPack<ePackType::ePT_Avx, T>& out)
    {
        OpClass::act(in1._val[0], in2._val[0], in3._val[0], out._val[0]);
        OpClass::act(in1._val[1], in2._val[1], in3._val[1], out._val[1]);
        OpClass::act(in1._val[2], in2._val[2], in3._val[2], out._val[2]);
        OpClass::act(in1._val[3], in2._val[3], in3._val[3], out._val[3]);
        OpClass::act(in1._val[4], in2._val[4], in3._val[4], out._val[4]);
        OpClass::act(in1._val[5], in2._val[5], in3._val[5], out._val[5]);
        OpClass::act(in1._val[6], in2._val[6], in3._val[6], out._val[6]);
        OpClass::act(in1._val[7], in2._val[7], in3._val[7], out._val[7]);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_TERNARYPACKOPRETMASK_H__