BinaryMaskPackOp.h

Go to the documentation of this file.

// BinaryMaskPackOp.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_BINARYMASKPACKOP_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_BINARYMASKPACKOP_H__

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

namespace ipsdk {
namespace simd {
namespace detail {


template <class OpClass, typename Scalar>
struct BinaryMaskPackOp<eInstructionSet::eIS_Avx, ipReal32, OpClass, Scalar>
{
    static IPSDK_FORCEINLINE
    const BaseMaskPack<ePackType::ePT_Avx, ipReal32>
    act(const BaseMaskPack<ePackType::ePT_Avx, ipReal32>& in1,
        const BaseMaskPack<ePackType::ePT_Avx, ipReal32>& in2)
    {
        BaseMaskPack<ePackType::ePT_Avx, ipReal32> out;
        out._val[0] = OpClass::act(in1._val[0], in2._val[0]);
        out._val[1] = OpClass::act(in1._val[1], in2._val[1]);
        out._val[2] = OpClass::act(in1._val[2], in2._val[2]);
        out._val[3] = OpClass::act(in1._val[3], in2._val[3]);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const BaseMaskPack<ePackType::ePT_Avx, ipReal32>& in1,
        const BaseMaskPack<ePackType::ePT_Avx, ipReal32>& in2,
        BaseMaskPack<ePackType::ePT_Avx, ipReal32>& out)
    {
        OpClass::act(in1._val[0], in2._val[0], out._val[0]);
        OpClass::act(in1._val[1], in2._val[1], out._val[1]);
        OpClass::act(in1._val[2], in2._val[2], out._val[2]);
        OpClass::act(in1._val[3], in2._val[3], out._val[3]);
    }

    static IPSDK_FORCEINLINE
    void
    act_io_io(BaseMaskPack<ePackType::ePT_Avx, ipReal32>& inOut1,
              BaseMaskPack<ePackType::ePT_Avx, ipReal32>& inOut2)
    {
        OpClass::act(inOut1._val[0], inOut2._val[0]);
        OpClass::act(inOut1._val[1], inOut2._val[1]);
        OpClass::act(inOut1._val[2], inOut2._val[2]);
        OpClass::act(inOut1._val[3], inOut2._val[3]);
    }
};

template <class OpClass, typename Scalar>
struct BinaryMaskPackOp<eInstructionSet::eIS_Avx, ipReal64, OpClass, Scalar>
{
    static IPSDK_FORCEINLINE
    BaseMaskPack<ePackType::ePT_Avx, ipReal64>
    act(const BaseMaskPack<ePackType::ePT_Avx, ipReal64>& in1,
        const BaseMaskPack<ePackType::ePT_Avx, ipReal64>& in2)
    {
        BaseMaskPack<ePackType::ePT_Avx, ipReal64> out;
        out._val[0] = OpClass::act(in1._val[0], in2._val[0]);
        out._val[1] = OpClass::act(in1._val[1], in2._val[1]);
        out._val[2] = OpClass::act(in1._val[2], in2._val[2]);
        out._val[3] = OpClass::act(in1._val[3], in2._val[3]);
        out._val[4] = OpClass::act(in1._val[4], in2._val[4]);
        out._val[5] = OpClass::act(in1._val[5], in2._val[5]);
        out._val[6] = OpClass::act(in1._val[6], in2._val[6]);
        out._val[7] = OpClass::act(in1._val[7], in2._val[7]);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const BaseMaskPack<ePackType::ePT_Avx, ipReal64>& in1,
        const BaseMaskPack<ePackType::ePT_Avx, ipReal64>& in2,
        BaseMaskPack<ePackType::ePT_Avx, ipReal64>& out)
    {
        OpClass::act(in1._val[0], in2._val[0], out._val[0]);
        OpClass::act(in1._val[1], in2._val[1], out._val[1]);
        OpClass::act(in1._val[2], in2._val[2], out._val[2]);
        OpClass::act(in1._val[3], in2._val[3], out._val[3]);
        OpClass::act(in1._val[4], in2._val[4], out._val[4]);
        OpClass::act(in1._val[5], in2._val[5], out._val[5]);
        OpClass::act(in1._val[6], in2._val[6], out._val[6]);
        OpClass::act(in1._val[7], in2._val[7], out._val[7]);
    }

    static IPSDK_FORCEINLINE
    void
    act_io_io(BaseMaskPack<ePackType::ePT_Avx, ipReal64>& inOut1,
              BaseMaskPack<ePackType::ePT_Avx, ipReal64>& inOut2)
    {
        OpClass::act(inOut1._val[0], inOut2._val[0]);
        OpClass::act(inOut1._val[1], inOut2._val[1]);
        OpClass::act(inOut1._val[2], inOut2._val[2]);
        OpClass::act(inOut1._val[3], inOut2._val[3]);
        OpClass::act(inOut1._val[4], inOut2._val[4]);
        OpClass::act(inOut1._val[5], inOut2._val[5]);
        OpClass::act(inOut1._val[6], inOut2._val[6]);
        OpClass::act(inOut1._val[7], inOut2._val[7]);
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_BINARYMASKPACKOP_H__