AlignedArithmetic.h

// AlignedArithmetic.h:
// --------------------
//

#ifndef __IPSDKUTIL_ALIGNEDARITHMETIC_H__
#define __IPSDKUTIL_ALIGNEDARITHMETIC_H__

#include <IPSDKUtil/InstructionSet/Pack.h>
#include <IPSDKUtil/InstructionSet/load.h>
#include <IPSDKUtil/InstructionSet/unload.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/add.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/sub.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/round.h>

namespace ipsdk {


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

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

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

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

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

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



} // end of namespace ipsdk

#endif // __IPSDKUTIL_ALIGNEDARITHMETIC_H__