AlignedBasic.h

Go to the documentation of this file.

// AlignedBasic.h:
// ---------------
//

#ifndef __IPSDKUTIL_ALIGNEDBASIC_H__
#define __IPSDKUTIL_ALIGNEDBASIC_H__

#include <IPSDKUtil/InstructionSet/Pack.h>
#include <IPSDKUtil/InstructionSet/load.h>
#include <IPSDKUtil/InstructionSet/unload.h>
#include <IPSDKUtil/InstructionSet/assign.h>

namespace ipsdk {


template <ipsdk::eInstructionSet::domain IS,
          typename BufType>
static IPSDK_FORCEINLINE void
alignedCopy(const BufType* pInputBuf,
            const ipUInt64 size,
            BufType* pOutputBuf)
{
    simd::Pack<IS, BufType> inputPack;
    for(ipUInt64 k=0; k<size; k+=simd::NbEltsPerOp<IS>::Value) {

        simd::load<IS>(inputPack, pInputBuf + k);
        simd::unload<IS>(inputPack, pOutputBuf + k);
    }
}

template <ipsdk::eInstructionSet::domain IS,
          typename BufType>
static IPSDK_FORCEINLINE void
alignedSet(const BufType& value,
           const ipUInt64 size,
           BufType* pOutputBuf)
{
    simd::Pack<IS, BufType> inputPack;
    simd::assign<IS>(inputPack, value);
    for(ipUInt64 k=0; k<size; k+=simd::NbEltsPerOp<IS>::Value)
        simd::unload<IS>(inputPack, pOutputBuf + k);
}


} // end of namespace ipsdk

#endif // __IPSDKUTIL_ALIGNEDBASIC_H__