GatherPackImpl.h

Go to the documentation of this file.

// GatherPackImpl.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_GATHERPACKIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_AVX_GATHERPACKIMPL_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/Gather/detail/GatherPackImpl.h>
#include <IPSDKUtil/InstructionSet/Gather/detail/GatherReg.h>
#include <IPSDKUtil/InstructionSet/detail/AVX/BasePack.h>
#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx, ipReal32>::act(
    const ipReal32* baseAddress,
    const ipUInt32* indexes,
    BasePack<ePackType::ePT_Avx, ipReal32>& out)
{
    GatherReg<eInstructionSet::eIS_Avx, ipReal32>::act(
        baseAddress, indexes, out._val[0]);
    GatherReg<eInstructionSet::eIS_Avx, ipReal32>::act(
        baseAddress, indexes+4, out._val[1]);
    GatherReg<eInstructionSet::eIS_Avx, ipReal32>::act(
        baseAddress, indexes+8, out._val[2]);
    GatherReg<eInstructionSet::eIS_Avx, ipReal32>::act(
        baseAddress, indexes+12, out._val[3]);
}

IPSDK_FORCEINLINE
void
GatherPack<eInstructionSet::eIS_Avx, ipReal32>::act(
    const ipReal32* baseAddress,
    const BasePack<ePackType::ePT_Avx, ipUInt32>& packIdx,
    BasePack<ePackType::ePT_Avx, ipReal32>& out)
{
    GatherReg<eInstructionSet::eIS_Avx, ipReal32>::act(
        baseAddress, packIdx._val[0], out._val[0]);
    GatherReg<eInstructionSet::eIS_Avx, ipReal32>::act(
        baseAddress, packIdx._val[1], out._val[1]);
    GatherReg<eInstructionSet::eIS_Avx, ipReal32>::act(
        baseAddress, packIdx._val[2], out._val[2]);
    GatherReg<eInstructionSet::eIS_Avx, ipReal32>::act(
        baseAddress, packIdx._val[3], out._val[3]);
}


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_GATHER_DETAIL_AVX_GATHERPACKIMPL_H__